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 list_head alias_list; /* Link alias-groups together */
93 struct iommu_dev_data *alias_data;/* The alias dev_data */
94 struct protection_domain *domain; /* Domain the device is bound to */
95 u16 devid; /* PCI Device ID */
96 bool iommu_v2; /* Device can make use of IOMMUv2 */
97 bool passthrough; /* Device is identity mapped */
101 } ats; /* ATS state */
102 bool pri_tlp; /* PASID TLB required for
104 u32 errata; /* Bitmap for errata to apply */
108 * general struct to manage commands send to an IOMMU
114 struct kmem_cache *amd_iommu_irq_cache;
116 static void update_domain(struct protection_domain *domain);
117 static int protection_domain_init(struct protection_domain *domain);
119 /****************************************************************************
123 ****************************************************************************/
125 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
127 return container_of(dom, struct protection_domain, domain);
130 static struct iommu_dev_data *alloc_dev_data(u16 devid)
132 struct iommu_dev_data *dev_data;
135 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
139 INIT_LIST_HEAD(&dev_data->alias_list);
141 dev_data->devid = devid;
143 spin_lock_irqsave(&dev_data_list_lock, flags);
144 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
145 spin_unlock_irqrestore(&dev_data_list_lock, flags);
150 static void free_dev_data(struct iommu_dev_data *dev_data)
154 spin_lock_irqsave(&dev_data_list_lock, flags);
155 list_del(&dev_data->dev_data_list);
156 spin_unlock_irqrestore(&dev_data_list_lock, flags);
161 static struct iommu_dev_data *search_dev_data(u16 devid)
163 struct iommu_dev_data *dev_data;
166 spin_lock_irqsave(&dev_data_list_lock, flags);
167 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
168 if (dev_data->devid == devid)
175 spin_unlock_irqrestore(&dev_data_list_lock, flags);
180 static struct iommu_dev_data *find_dev_data(u16 devid)
182 struct iommu_dev_data *dev_data;
184 dev_data = search_dev_data(devid);
186 if (dev_data == NULL)
187 dev_data = alloc_dev_data(devid);
192 static inline u16 get_device_id(struct device *dev)
194 struct pci_dev *pdev = to_pci_dev(dev);
196 return PCI_DEVID(pdev->bus->number, pdev->devfn);
199 static struct iommu_dev_data *get_dev_data(struct device *dev)
201 return dev->archdata.iommu;
204 static bool pci_iommuv2_capable(struct pci_dev *pdev)
206 static const int caps[] = {
209 PCI_EXT_CAP_ID_PASID,
213 for (i = 0; i < 3; ++i) {
214 pos = pci_find_ext_capability(pdev, caps[i]);
222 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
224 struct iommu_dev_data *dev_data;
226 dev_data = get_dev_data(&pdev->dev);
228 return dev_data->errata & (1 << erratum) ? true : false;
232 * This function actually applies the mapping to the page table of the
235 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
236 struct unity_map_entry *e)
240 for (addr = e->address_start; addr < e->address_end;
242 if (addr < dma_dom->aperture_size)
243 __set_bit(addr >> PAGE_SHIFT,
244 dma_dom->aperture[0]->bitmap);
249 * Inits the unity mappings required for a specific device
251 static void init_unity_mappings_for_device(struct device *dev,
252 struct dma_ops_domain *dma_dom)
254 struct unity_map_entry *e;
257 devid = get_device_id(dev);
259 list_for_each_entry(e, &amd_iommu_unity_map, list) {
260 if (!(devid >= e->devid_start && devid <= e->devid_end))
262 alloc_unity_mapping(dma_dom, e);
267 * This function checks if the driver got a valid device from the caller to
268 * avoid dereferencing invalid pointers.
270 static bool check_device(struct device *dev)
274 if (!dev || !dev->dma_mask)
278 if (!dev_is_pci(dev))
281 devid = get_device_id(dev);
283 /* Out of our scope? */
284 if (devid > amd_iommu_last_bdf)
287 if (amd_iommu_rlookup_table[devid] == NULL)
293 static void init_iommu_group(struct device *dev)
295 struct dma_ops_domain *dma_domain;
296 struct iommu_domain *domain;
297 struct iommu_group *group;
299 group = iommu_group_get_for_dev(dev);
303 domain = iommu_group_default_domain(group);
307 dma_domain = to_pdomain(domain)->priv;
309 init_unity_mappings_for_device(dev, dma_domain);
311 iommu_group_put(group);
314 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
316 *(u16 *)data = alias;
320 static u16 get_alias(struct device *dev)
322 struct pci_dev *pdev = to_pci_dev(dev);
323 u16 devid, ivrs_alias, pci_alias;
325 devid = get_device_id(dev);
326 ivrs_alias = amd_iommu_alias_table[devid];
327 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
329 if (ivrs_alias == pci_alias)
335 * The IVRS is fairly reliable in telling us about aliases, but it
336 * can't know about every screwy device. If we don't have an IVRS
337 * reported alias, use the PCI reported alias. In that case we may
338 * still need to initialize the rlookup and dev_table entries if the
339 * alias is to a non-existent device.
341 if (ivrs_alias == devid) {
342 if (!amd_iommu_rlookup_table[pci_alias]) {
343 amd_iommu_rlookup_table[pci_alias] =
344 amd_iommu_rlookup_table[devid];
345 memcpy(amd_iommu_dev_table[pci_alias].data,
346 amd_iommu_dev_table[devid].data,
347 sizeof(amd_iommu_dev_table[pci_alias].data));
353 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
354 "for device %s[%04x:%04x], kernel reported alias "
355 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
356 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
357 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
358 PCI_FUNC(pci_alias));
361 * If we don't have a PCI DMA alias and the IVRS alias is on the same
362 * bus, then the IVRS table may know about a quirk that we don't.
364 if (pci_alias == devid &&
365 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
366 pdev->dev_flags |= PCI_DEV_FLAGS_DMA_ALIAS_DEVFN;
367 pdev->dma_alias_devfn = ivrs_alias & 0xff;
368 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
369 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
376 static int iommu_init_device(struct device *dev)
378 struct pci_dev *pdev = to_pci_dev(dev);
379 struct iommu_dev_data *dev_data;
382 if (dev->archdata.iommu)
385 dev_data = find_dev_data(get_device_id(dev));
389 alias = get_alias(dev);
391 if (alias != dev_data->devid) {
392 struct iommu_dev_data *alias_data;
394 alias_data = find_dev_data(alias);
395 if (alias_data == NULL) {
396 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
398 free_dev_data(dev_data);
401 dev_data->alias_data = alias_data;
403 /* Add device to the alias_list */
404 list_add(&dev_data->alias_list, &alias_data->alias_list);
407 if (pci_iommuv2_capable(pdev)) {
408 struct amd_iommu *iommu;
410 iommu = amd_iommu_rlookup_table[dev_data->devid];
411 dev_data->iommu_v2 = iommu->is_iommu_v2;
414 dev->archdata.iommu = dev_data;
416 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
422 static void iommu_ignore_device(struct device *dev)
426 devid = get_device_id(dev);
427 alias = amd_iommu_alias_table[devid];
429 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
430 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
432 amd_iommu_rlookup_table[devid] = NULL;
433 amd_iommu_rlookup_table[alias] = NULL;
436 static void iommu_uninit_device(struct device *dev)
438 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
443 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
446 iommu_group_remove_device(dev);
448 /* Unlink from alias, it may change if another device is re-plugged */
449 dev_data->alias_data = NULL;
452 dev->archdata.dma_ops = NULL;
455 * We keep dev_data around for unplugged devices and reuse it when the
456 * device is re-plugged - not doing so would introduce a ton of races.
460 #ifdef CONFIG_AMD_IOMMU_STATS
463 * Initialization code for statistics collection
466 DECLARE_STATS_COUNTER(compl_wait);
467 DECLARE_STATS_COUNTER(cnt_map_single);
468 DECLARE_STATS_COUNTER(cnt_unmap_single);
469 DECLARE_STATS_COUNTER(cnt_map_sg);
470 DECLARE_STATS_COUNTER(cnt_unmap_sg);
471 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
472 DECLARE_STATS_COUNTER(cnt_free_coherent);
473 DECLARE_STATS_COUNTER(cross_page);
474 DECLARE_STATS_COUNTER(domain_flush_single);
475 DECLARE_STATS_COUNTER(domain_flush_all);
476 DECLARE_STATS_COUNTER(alloced_io_mem);
477 DECLARE_STATS_COUNTER(total_map_requests);
478 DECLARE_STATS_COUNTER(complete_ppr);
479 DECLARE_STATS_COUNTER(invalidate_iotlb);
480 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
481 DECLARE_STATS_COUNTER(pri_requests);
483 static struct dentry *stats_dir;
484 static struct dentry *de_fflush;
486 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
488 if (stats_dir == NULL)
491 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
495 static void amd_iommu_stats_init(void)
497 stats_dir = debugfs_create_dir("amd-iommu", NULL);
498 if (stats_dir == NULL)
501 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
502 &amd_iommu_unmap_flush);
504 amd_iommu_stats_add(&compl_wait);
505 amd_iommu_stats_add(&cnt_map_single);
506 amd_iommu_stats_add(&cnt_unmap_single);
507 amd_iommu_stats_add(&cnt_map_sg);
508 amd_iommu_stats_add(&cnt_unmap_sg);
509 amd_iommu_stats_add(&cnt_alloc_coherent);
510 amd_iommu_stats_add(&cnt_free_coherent);
511 amd_iommu_stats_add(&cross_page);
512 amd_iommu_stats_add(&domain_flush_single);
513 amd_iommu_stats_add(&domain_flush_all);
514 amd_iommu_stats_add(&alloced_io_mem);
515 amd_iommu_stats_add(&total_map_requests);
516 amd_iommu_stats_add(&complete_ppr);
517 amd_iommu_stats_add(&invalidate_iotlb);
518 amd_iommu_stats_add(&invalidate_iotlb_all);
519 amd_iommu_stats_add(&pri_requests);
524 /****************************************************************************
526 * Interrupt handling functions
528 ****************************************************************************/
530 static void dump_dte_entry(u16 devid)
534 for (i = 0; i < 4; ++i)
535 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
536 amd_iommu_dev_table[devid].data[i]);
539 static void dump_command(unsigned long phys_addr)
541 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
544 for (i = 0; i < 4; ++i)
545 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
548 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
550 int type, devid, domid, flags;
551 volatile u32 *event = __evt;
556 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
557 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
558 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
559 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
560 address = (u64)(((u64)event[3]) << 32) | event[2];
563 /* Did we hit the erratum? */
564 if (++count == LOOP_TIMEOUT) {
565 pr_err("AMD-Vi: No event written to event log\n");
572 printk(KERN_ERR "AMD-Vi: Event logged [");
575 case EVENT_TYPE_ILL_DEV:
576 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
577 "address=0x%016llx flags=0x%04x]\n",
578 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
580 dump_dte_entry(devid);
582 case EVENT_TYPE_IO_FAULT:
583 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
584 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
585 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
586 domid, address, flags);
588 case EVENT_TYPE_DEV_TAB_ERR:
589 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
590 "address=0x%016llx flags=0x%04x]\n",
591 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
594 case EVENT_TYPE_PAGE_TAB_ERR:
595 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
596 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
597 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
598 domid, address, flags);
600 case EVENT_TYPE_ILL_CMD:
601 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
602 dump_command(address);
604 case EVENT_TYPE_CMD_HARD_ERR:
605 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
606 "flags=0x%04x]\n", address, flags);
608 case EVENT_TYPE_IOTLB_INV_TO:
609 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
610 "address=0x%016llx]\n",
611 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
614 case EVENT_TYPE_INV_DEV_REQ:
615 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
616 "address=0x%016llx flags=0x%04x]\n",
617 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
621 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
624 memset(__evt, 0, 4 * sizeof(u32));
627 static void iommu_poll_events(struct amd_iommu *iommu)
631 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
632 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
634 while (head != tail) {
635 iommu_print_event(iommu, iommu->evt_buf + head);
636 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
639 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
642 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
644 struct amd_iommu_fault fault;
646 INC_STATS_COUNTER(pri_requests);
648 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
649 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
653 fault.address = raw[1];
654 fault.pasid = PPR_PASID(raw[0]);
655 fault.device_id = PPR_DEVID(raw[0]);
656 fault.tag = PPR_TAG(raw[0]);
657 fault.flags = PPR_FLAGS(raw[0]);
659 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
662 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
666 if (iommu->ppr_log == NULL)
669 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
670 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
672 while (head != tail) {
677 raw = (u64 *)(iommu->ppr_log + head);
680 * Hardware bug: Interrupt may arrive before the entry is
681 * written to memory. If this happens we need to wait for the
684 for (i = 0; i < LOOP_TIMEOUT; ++i) {
685 if (PPR_REQ_TYPE(raw[0]) != 0)
690 /* Avoid memcpy function-call overhead */
695 * To detect the hardware bug we need to clear the entry
698 raw[0] = raw[1] = 0UL;
700 /* Update head pointer of hardware ring-buffer */
701 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
702 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
704 /* Handle PPR entry */
705 iommu_handle_ppr_entry(iommu, entry);
707 /* Refresh ring-buffer information */
708 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
709 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
713 irqreturn_t amd_iommu_int_thread(int irq, void *data)
715 struct amd_iommu *iommu = (struct amd_iommu *) data;
716 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
718 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
719 /* Enable EVT and PPR interrupts again */
720 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
721 iommu->mmio_base + MMIO_STATUS_OFFSET);
723 if (status & MMIO_STATUS_EVT_INT_MASK) {
724 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
725 iommu_poll_events(iommu);
728 if (status & MMIO_STATUS_PPR_INT_MASK) {
729 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
730 iommu_poll_ppr_log(iommu);
734 * Hardware bug: ERBT1312
735 * When re-enabling interrupt (by writing 1
736 * to clear the bit), the hardware might also try to set
737 * the interrupt bit in the event status register.
738 * In this scenario, the bit will be set, and disable
739 * subsequent interrupts.
741 * Workaround: The IOMMU driver should read back the
742 * status register and check if the interrupt bits are cleared.
743 * If not, driver will need to go through the interrupt handler
744 * again and re-clear the bits
746 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
751 irqreturn_t amd_iommu_int_handler(int irq, void *data)
753 return IRQ_WAKE_THREAD;
756 /****************************************************************************
758 * IOMMU command queuing functions
760 ****************************************************************************/
762 static int wait_on_sem(volatile u64 *sem)
766 while (*sem == 0 && i < LOOP_TIMEOUT) {
771 if (i == LOOP_TIMEOUT) {
772 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
779 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
780 struct iommu_cmd *cmd,
785 target = iommu->cmd_buf + tail;
786 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
788 /* Copy command to buffer */
789 memcpy(target, cmd, sizeof(*cmd));
791 /* Tell the IOMMU about it */
792 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
795 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
797 WARN_ON(address & 0x7ULL);
799 memset(cmd, 0, sizeof(*cmd));
800 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
801 cmd->data[1] = upper_32_bits(__pa(address));
803 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
806 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
808 memset(cmd, 0, sizeof(*cmd));
809 cmd->data[0] = devid;
810 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
813 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
814 size_t size, u16 domid, int pde)
819 pages = iommu_num_pages(address, size, PAGE_SIZE);
824 * If we have to flush more than one page, flush all
825 * TLB entries for this domain
827 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
831 address &= PAGE_MASK;
833 memset(cmd, 0, sizeof(*cmd));
834 cmd->data[1] |= domid;
835 cmd->data[2] = lower_32_bits(address);
836 cmd->data[3] = upper_32_bits(address);
837 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
838 if (s) /* size bit - we flush more than one 4kb page */
839 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
840 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
841 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
844 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
845 u64 address, size_t size)
850 pages = iommu_num_pages(address, size, PAGE_SIZE);
855 * If we have to flush more than one page, flush all
856 * TLB entries for this domain
858 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
862 address &= PAGE_MASK;
864 memset(cmd, 0, sizeof(*cmd));
865 cmd->data[0] = devid;
866 cmd->data[0] |= (qdep & 0xff) << 24;
867 cmd->data[1] = devid;
868 cmd->data[2] = lower_32_bits(address);
869 cmd->data[3] = upper_32_bits(address);
870 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
872 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
875 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
876 u64 address, bool size)
878 memset(cmd, 0, sizeof(*cmd));
880 address &= ~(0xfffULL);
882 cmd->data[0] = pasid;
883 cmd->data[1] = domid;
884 cmd->data[2] = lower_32_bits(address);
885 cmd->data[3] = upper_32_bits(address);
886 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
887 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
889 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
890 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
893 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
894 int qdep, u64 address, bool size)
896 memset(cmd, 0, sizeof(*cmd));
898 address &= ~(0xfffULL);
900 cmd->data[0] = devid;
901 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
902 cmd->data[0] |= (qdep & 0xff) << 24;
903 cmd->data[1] = devid;
904 cmd->data[1] |= (pasid & 0xff) << 16;
905 cmd->data[2] = lower_32_bits(address);
906 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
907 cmd->data[3] = upper_32_bits(address);
909 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
910 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
913 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
914 int status, int tag, bool gn)
916 memset(cmd, 0, sizeof(*cmd));
918 cmd->data[0] = devid;
920 cmd->data[1] = pasid;
921 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
923 cmd->data[3] = tag & 0x1ff;
924 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
926 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
929 static void build_inv_all(struct iommu_cmd *cmd)
931 memset(cmd, 0, sizeof(*cmd));
932 CMD_SET_TYPE(cmd, CMD_INV_ALL);
935 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
937 memset(cmd, 0, sizeof(*cmd));
938 cmd->data[0] = devid;
939 CMD_SET_TYPE(cmd, CMD_INV_IRT);
943 * Writes the command to the IOMMUs command buffer and informs the
944 * hardware about the new command.
946 static int iommu_queue_command_sync(struct amd_iommu *iommu,
947 struct iommu_cmd *cmd,
950 u32 left, tail, head, next_tail;
953 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
956 spin_lock_irqsave(&iommu->lock, flags);
958 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
959 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
960 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
961 left = (head - next_tail) % iommu->cmd_buf_size;
964 struct iommu_cmd sync_cmd;
965 volatile u64 sem = 0;
968 build_completion_wait(&sync_cmd, (u64)&sem);
969 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
971 spin_unlock_irqrestore(&iommu->lock, flags);
973 if ((ret = wait_on_sem(&sem)) != 0)
979 copy_cmd_to_buffer(iommu, cmd, tail);
981 /* We need to sync now to make sure all commands are processed */
982 iommu->need_sync = sync;
984 spin_unlock_irqrestore(&iommu->lock, flags);
989 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
991 return iommu_queue_command_sync(iommu, cmd, true);
995 * This function queues a completion wait command into the command
998 static int iommu_completion_wait(struct amd_iommu *iommu)
1000 struct iommu_cmd cmd;
1001 volatile u64 sem = 0;
1004 if (!iommu->need_sync)
1007 build_completion_wait(&cmd, (u64)&sem);
1009 ret = iommu_queue_command_sync(iommu, &cmd, false);
1013 return wait_on_sem(&sem);
1016 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1018 struct iommu_cmd cmd;
1020 build_inv_dte(&cmd, devid);
1022 return iommu_queue_command(iommu, &cmd);
1025 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1029 for (devid = 0; devid <= 0xffff; ++devid)
1030 iommu_flush_dte(iommu, devid);
1032 iommu_completion_wait(iommu);
1036 * This function uses heavy locking and may disable irqs for some time. But
1037 * this is no issue because it is only called during resume.
1039 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1043 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1044 struct iommu_cmd cmd;
1045 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1047 iommu_queue_command(iommu, &cmd);
1050 iommu_completion_wait(iommu);
1053 static void iommu_flush_all(struct amd_iommu *iommu)
1055 struct iommu_cmd cmd;
1057 build_inv_all(&cmd);
1059 iommu_queue_command(iommu, &cmd);
1060 iommu_completion_wait(iommu);
1063 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1065 struct iommu_cmd cmd;
1067 build_inv_irt(&cmd, devid);
1069 iommu_queue_command(iommu, &cmd);
1072 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1076 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1077 iommu_flush_irt(iommu, devid);
1079 iommu_completion_wait(iommu);
1082 void iommu_flush_all_caches(struct amd_iommu *iommu)
1084 if (iommu_feature(iommu, FEATURE_IA)) {
1085 iommu_flush_all(iommu);
1087 iommu_flush_dte_all(iommu);
1088 iommu_flush_irt_all(iommu);
1089 iommu_flush_tlb_all(iommu);
1094 * Command send function for flushing on-device TLB
1096 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1097 u64 address, size_t size)
1099 struct amd_iommu *iommu;
1100 struct iommu_cmd cmd;
1103 qdep = dev_data->ats.qdep;
1104 iommu = amd_iommu_rlookup_table[dev_data->devid];
1106 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1108 return iommu_queue_command(iommu, &cmd);
1112 * Command send function for invalidating a device table entry
1114 static int device_flush_dte(struct iommu_dev_data *dev_data)
1116 struct amd_iommu *iommu;
1119 iommu = amd_iommu_rlookup_table[dev_data->devid];
1121 ret = iommu_flush_dte(iommu, dev_data->devid);
1125 if (dev_data->ats.enabled)
1126 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1132 * TLB invalidation function which is called from the mapping functions.
1133 * It invalidates a single PTE if the range to flush is within a single
1134 * page. Otherwise it flushes the whole TLB of the IOMMU.
1136 static void __domain_flush_pages(struct protection_domain *domain,
1137 u64 address, size_t size, int pde)
1139 struct iommu_dev_data *dev_data;
1140 struct iommu_cmd cmd;
1143 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1145 for (i = 0; i < amd_iommus_present; ++i) {
1146 if (!domain->dev_iommu[i])
1150 * Devices of this domain are behind this IOMMU
1151 * We need a TLB flush
1153 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1156 list_for_each_entry(dev_data, &domain->dev_list, list) {
1158 if (!dev_data->ats.enabled)
1161 ret |= device_flush_iotlb(dev_data, address, size);
1167 static void domain_flush_pages(struct protection_domain *domain,
1168 u64 address, size_t size)
1170 __domain_flush_pages(domain, address, size, 0);
1173 /* Flush the whole IO/TLB for a given protection domain */
1174 static void domain_flush_tlb(struct protection_domain *domain)
1176 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1179 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1180 static void domain_flush_tlb_pde(struct protection_domain *domain)
1182 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1185 static void domain_flush_complete(struct protection_domain *domain)
1189 for (i = 0; i < amd_iommus_present; ++i) {
1190 if (!domain->dev_iommu[i])
1194 * Devices of this domain are behind this IOMMU
1195 * We need to wait for completion of all commands.
1197 iommu_completion_wait(amd_iommus[i]);
1203 * This function flushes the DTEs for all devices in domain
1205 static void domain_flush_devices(struct protection_domain *domain)
1207 struct iommu_dev_data *dev_data;
1209 list_for_each_entry(dev_data, &domain->dev_list, list)
1210 device_flush_dte(dev_data);
1213 /****************************************************************************
1215 * The functions below are used the create the page table mappings for
1216 * unity mapped regions.
1218 ****************************************************************************/
1221 * This function is used to add another level to an IO page table. Adding
1222 * another level increases the size of the address space by 9 bits to a size up
1225 static bool increase_address_space(struct protection_domain *domain,
1230 if (domain->mode == PAGE_MODE_6_LEVEL)
1231 /* address space already 64 bit large */
1234 pte = (void *)get_zeroed_page(gfp);
1238 *pte = PM_LEVEL_PDE(domain->mode,
1239 virt_to_phys(domain->pt_root));
1240 domain->pt_root = pte;
1242 domain->updated = true;
1247 static u64 *alloc_pte(struct protection_domain *domain,
1248 unsigned long address,
1249 unsigned long page_size,
1256 BUG_ON(!is_power_of_2(page_size));
1258 while (address > PM_LEVEL_SIZE(domain->mode))
1259 increase_address_space(domain, gfp);
1261 level = domain->mode - 1;
1262 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1263 address = PAGE_SIZE_ALIGN(address, page_size);
1264 end_lvl = PAGE_SIZE_LEVEL(page_size);
1266 while (level > end_lvl) {
1267 if (!IOMMU_PTE_PRESENT(*pte)) {
1268 page = (u64 *)get_zeroed_page(gfp);
1271 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1274 /* No level skipping support yet */
1275 if (PM_PTE_LEVEL(*pte) != level)
1280 pte = IOMMU_PTE_PAGE(*pte);
1282 if (pte_page && level == end_lvl)
1285 pte = &pte[PM_LEVEL_INDEX(level, address)];
1292 * This function checks if there is a PTE for a given dma address. If
1293 * there is one, it returns the pointer to it.
1295 static u64 *fetch_pte(struct protection_domain *domain,
1296 unsigned long address,
1297 unsigned long *page_size)
1302 if (address > PM_LEVEL_SIZE(domain->mode))
1305 level = domain->mode - 1;
1306 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1307 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1312 if (!IOMMU_PTE_PRESENT(*pte))
1316 if (PM_PTE_LEVEL(*pte) == 7 ||
1317 PM_PTE_LEVEL(*pte) == 0)
1320 /* No level skipping support yet */
1321 if (PM_PTE_LEVEL(*pte) != level)
1326 /* Walk to the next level */
1327 pte = IOMMU_PTE_PAGE(*pte);
1328 pte = &pte[PM_LEVEL_INDEX(level, address)];
1329 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1332 if (PM_PTE_LEVEL(*pte) == 0x07) {
1333 unsigned long pte_mask;
1336 * If we have a series of large PTEs, make
1337 * sure to return a pointer to the first one.
1339 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1340 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1341 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1348 * Generic mapping functions. It maps a physical address into a DMA
1349 * address space. It allocates the page table pages if necessary.
1350 * In the future it can be extended to a generic mapping function
1351 * supporting all features of AMD IOMMU page tables like level skipping
1352 * and full 64 bit address spaces.
1354 static int iommu_map_page(struct protection_domain *dom,
1355 unsigned long bus_addr,
1356 unsigned long phys_addr,
1358 unsigned long page_size)
1363 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1364 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1366 if (!(prot & IOMMU_PROT_MASK))
1369 count = PAGE_SIZE_PTE_COUNT(page_size);
1370 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1375 for (i = 0; i < count; ++i)
1376 if (IOMMU_PTE_PRESENT(pte[i]))
1380 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1381 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1383 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1385 if (prot & IOMMU_PROT_IR)
1386 __pte |= IOMMU_PTE_IR;
1387 if (prot & IOMMU_PROT_IW)
1388 __pte |= IOMMU_PTE_IW;
1390 for (i = 0; i < count; ++i)
1398 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1399 unsigned long bus_addr,
1400 unsigned long page_size)
1402 unsigned long long unmapped;
1403 unsigned long unmap_size;
1406 BUG_ON(!is_power_of_2(page_size));
1410 while (unmapped < page_size) {
1412 pte = fetch_pte(dom, bus_addr, &unmap_size);
1417 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1418 for (i = 0; i < count; i++)
1422 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1423 unmapped += unmap_size;
1426 BUG_ON(unmapped && !is_power_of_2(unmapped));
1431 /****************************************************************************
1433 * The next functions belong to the address allocator for the dma_ops
1434 * interface functions. They work like the allocators in the other IOMMU
1435 * drivers. Its basically a bitmap which marks the allocated pages in
1436 * the aperture. Maybe it could be enhanced in the future to a more
1437 * efficient allocator.
1439 ****************************************************************************/
1442 * The address allocator core functions.
1444 * called with domain->lock held
1448 * Used to reserve address ranges in the aperture (e.g. for exclusion
1451 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1452 unsigned long start_page,
1455 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1457 if (start_page + pages > last_page)
1458 pages = last_page - start_page;
1460 for (i = start_page; i < start_page + pages; ++i) {
1461 int index = i / APERTURE_RANGE_PAGES;
1462 int page = i % APERTURE_RANGE_PAGES;
1463 __set_bit(page, dom->aperture[index]->bitmap);
1468 * This function is used to add a new aperture range to an existing
1469 * aperture in case of dma_ops domain allocation or address allocation
1472 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1473 bool populate, gfp_t gfp)
1475 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1476 struct amd_iommu *iommu;
1477 unsigned long i, old_size, pte_pgsize;
1479 #ifdef CONFIG_IOMMU_STRESS
1483 if (index >= APERTURE_MAX_RANGES)
1486 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1487 if (!dma_dom->aperture[index])
1490 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1491 if (!dma_dom->aperture[index]->bitmap)
1494 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1497 unsigned long address = dma_dom->aperture_size;
1498 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1499 u64 *pte, *pte_page;
1501 for (i = 0; i < num_ptes; ++i) {
1502 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1507 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1509 address += APERTURE_RANGE_SIZE / 64;
1513 old_size = dma_dom->aperture_size;
1514 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1516 /* Reserve address range used for MSI messages */
1517 if (old_size < MSI_ADDR_BASE_LO &&
1518 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1519 unsigned long spage;
1522 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1523 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1525 dma_ops_reserve_addresses(dma_dom, spage, pages);
1528 /* Initialize the exclusion range if necessary */
1529 for_each_iommu(iommu) {
1530 if (iommu->exclusion_start &&
1531 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1532 && iommu->exclusion_start < dma_dom->aperture_size) {
1533 unsigned long startpage;
1534 int pages = iommu_num_pages(iommu->exclusion_start,
1535 iommu->exclusion_length,
1537 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1538 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1543 * Check for areas already mapped as present in the new aperture
1544 * range and mark those pages as reserved in the allocator. Such
1545 * mappings may already exist as a result of requested unity
1546 * mappings for devices.
1548 for (i = dma_dom->aperture[index]->offset;
1549 i < dma_dom->aperture_size;
1551 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1552 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1555 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1559 update_domain(&dma_dom->domain);
1564 update_domain(&dma_dom->domain);
1566 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1568 kfree(dma_dom->aperture[index]);
1569 dma_dom->aperture[index] = NULL;
1574 static unsigned long dma_ops_area_alloc(struct device *dev,
1575 struct dma_ops_domain *dom,
1577 unsigned long align_mask,
1579 unsigned long start)
1581 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1582 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1583 int i = start >> APERTURE_RANGE_SHIFT;
1584 unsigned long boundary_size, mask;
1585 unsigned long address = -1;
1586 unsigned long limit;
1588 next_bit >>= PAGE_SHIFT;
1590 mask = dma_get_seg_boundary(dev);
1592 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1593 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1595 for (;i < max_index; ++i) {
1596 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1598 if (dom->aperture[i]->offset >= dma_mask)
1601 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1602 dma_mask >> PAGE_SHIFT);
1604 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1605 limit, next_bit, pages, 0,
1606 boundary_size, align_mask);
1607 if (address != -1) {
1608 address = dom->aperture[i]->offset +
1609 (address << PAGE_SHIFT);
1610 dom->next_address = address + (pages << PAGE_SHIFT);
1620 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1621 struct dma_ops_domain *dom,
1623 unsigned long align_mask,
1626 unsigned long address;
1628 #ifdef CONFIG_IOMMU_STRESS
1629 dom->next_address = 0;
1630 dom->need_flush = true;
1633 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1634 dma_mask, dom->next_address);
1636 if (address == -1) {
1637 dom->next_address = 0;
1638 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1640 dom->need_flush = true;
1643 if (unlikely(address == -1))
1644 address = DMA_ERROR_CODE;
1646 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1652 * The address free function.
1654 * called with domain->lock held
1656 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1657 unsigned long address,
1660 unsigned i = address >> APERTURE_RANGE_SHIFT;
1661 struct aperture_range *range = dom->aperture[i];
1663 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1665 #ifdef CONFIG_IOMMU_STRESS
1670 if (address >= dom->next_address)
1671 dom->need_flush = true;
1673 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1675 bitmap_clear(range->bitmap, address, pages);
1679 /****************************************************************************
1681 * The next functions belong to the domain allocation. A domain is
1682 * allocated for every IOMMU as the default domain. If device isolation
1683 * is enabled, every device get its own domain. The most important thing
1684 * about domains is the page table mapping the DMA address space they
1687 ****************************************************************************/
1690 * This function adds a protection domain to the global protection domain list
1692 static void add_domain_to_list(struct protection_domain *domain)
1694 unsigned long flags;
1696 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1697 list_add(&domain->list, &amd_iommu_pd_list);
1698 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1702 * This function removes a protection domain to the global
1703 * protection domain list
1705 static void del_domain_from_list(struct protection_domain *domain)
1707 unsigned long flags;
1709 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1710 list_del(&domain->list);
1711 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1714 static u16 domain_id_alloc(void)
1716 unsigned long flags;
1719 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1720 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1722 if (id > 0 && id < MAX_DOMAIN_ID)
1723 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1726 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1731 static void domain_id_free(int id)
1733 unsigned long flags;
1735 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1736 if (id > 0 && id < MAX_DOMAIN_ID)
1737 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1738 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1741 #define DEFINE_FREE_PT_FN(LVL, FN) \
1742 static void free_pt_##LVL (unsigned long __pt) \
1750 for (i = 0; i < 512; ++i) { \
1751 /* PTE present? */ \
1752 if (!IOMMU_PTE_PRESENT(pt[i])) \
1756 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1757 PM_PTE_LEVEL(pt[i]) == 7) \
1760 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1763 free_page((unsigned long)pt); \
1766 DEFINE_FREE_PT_FN(l2, free_page)
1767 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1768 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1769 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1770 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1772 static void free_pagetable(struct protection_domain *domain)
1774 unsigned long root = (unsigned long)domain->pt_root;
1776 switch (domain->mode) {
1777 case PAGE_MODE_NONE:
1779 case PAGE_MODE_1_LEVEL:
1782 case PAGE_MODE_2_LEVEL:
1785 case PAGE_MODE_3_LEVEL:
1788 case PAGE_MODE_4_LEVEL:
1791 case PAGE_MODE_5_LEVEL:
1794 case PAGE_MODE_6_LEVEL:
1802 static void free_gcr3_tbl_level1(u64 *tbl)
1807 for (i = 0; i < 512; ++i) {
1808 if (!(tbl[i] & GCR3_VALID))
1811 ptr = __va(tbl[i] & PAGE_MASK);
1813 free_page((unsigned long)ptr);
1817 static void free_gcr3_tbl_level2(u64 *tbl)
1822 for (i = 0; i < 512; ++i) {
1823 if (!(tbl[i] & GCR3_VALID))
1826 ptr = __va(tbl[i] & PAGE_MASK);
1828 free_gcr3_tbl_level1(ptr);
1832 static void free_gcr3_table(struct protection_domain *domain)
1834 if (domain->glx == 2)
1835 free_gcr3_tbl_level2(domain->gcr3_tbl);
1836 else if (domain->glx == 1)
1837 free_gcr3_tbl_level1(domain->gcr3_tbl);
1838 else if (domain->glx != 0)
1841 free_page((unsigned long)domain->gcr3_tbl);
1845 * Free a domain, only used if something went wrong in the
1846 * allocation path and we need to free an already allocated page table
1848 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1855 del_domain_from_list(&dom->domain);
1857 free_pagetable(&dom->domain);
1859 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1860 if (!dom->aperture[i])
1862 free_page((unsigned long)dom->aperture[i]->bitmap);
1863 kfree(dom->aperture[i]);
1870 * Allocates a new protection domain usable for the dma_ops functions.
1871 * It also initializes the page table and the address allocator data
1872 * structures required for the dma_ops interface
1874 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1876 struct dma_ops_domain *dma_dom;
1878 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1882 if (protection_domain_init(&dma_dom->domain))
1885 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1886 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1887 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1888 dma_dom->domain.priv = dma_dom;
1889 if (!dma_dom->domain.pt_root)
1892 dma_dom->need_flush = false;
1894 add_domain_to_list(&dma_dom->domain);
1896 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1900 * mark the first page as allocated so we never return 0 as
1901 * a valid dma-address. So we can use 0 as error value
1903 dma_dom->aperture[0]->bitmap[0] = 1;
1904 dma_dom->next_address = 0;
1910 dma_ops_domain_free(dma_dom);
1916 * little helper function to check whether a given protection domain is a
1919 static bool dma_ops_domain(struct protection_domain *domain)
1921 return domain->flags & PD_DMA_OPS_MASK;
1924 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1929 if (domain->mode != PAGE_MODE_NONE)
1930 pte_root = virt_to_phys(domain->pt_root);
1932 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1933 << DEV_ENTRY_MODE_SHIFT;
1934 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1936 flags = amd_iommu_dev_table[devid].data[1];
1939 flags |= DTE_FLAG_IOTLB;
1941 if (domain->flags & PD_IOMMUV2_MASK) {
1942 u64 gcr3 = __pa(domain->gcr3_tbl);
1943 u64 glx = domain->glx;
1946 pte_root |= DTE_FLAG_GV;
1947 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1949 /* First mask out possible old values for GCR3 table */
1950 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1953 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1956 /* Encode GCR3 table into DTE */
1957 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1960 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1963 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1967 flags &= ~(0xffffUL);
1968 flags |= domain->id;
1970 amd_iommu_dev_table[devid].data[1] = flags;
1971 amd_iommu_dev_table[devid].data[0] = pte_root;
1974 static void clear_dte_entry(u16 devid)
1976 /* remove entry from the device table seen by the hardware */
1977 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1978 amd_iommu_dev_table[devid].data[1] = 0;
1980 amd_iommu_apply_erratum_63(devid);
1983 static void do_attach(struct iommu_dev_data *dev_data,
1984 struct protection_domain *domain)
1986 struct amd_iommu *iommu;
1989 iommu = amd_iommu_rlookup_table[dev_data->devid];
1990 ats = dev_data->ats.enabled;
1992 /* Update data structures */
1993 dev_data->domain = domain;
1994 list_add(&dev_data->list, &domain->dev_list);
1995 set_dte_entry(dev_data->devid, domain, ats);
1997 /* Do reference counting */
1998 domain->dev_iommu[iommu->index] += 1;
1999 domain->dev_cnt += 1;
2001 /* Flush the DTE entry */
2002 device_flush_dte(dev_data);
2005 static void do_detach(struct iommu_dev_data *dev_data)
2007 struct amd_iommu *iommu;
2009 iommu = amd_iommu_rlookup_table[dev_data->devid];
2011 /* decrease reference counters */
2012 dev_data->domain->dev_iommu[iommu->index] -= 1;
2013 dev_data->domain->dev_cnt -= 1;
2015 /* Update data structures */
2016 dev_data->domain = NULL;
2017 list_del(&dev_data->list);
2018 clear_dte_entry(dev_data->devid);
2020 /* Flush the DTE entry */
2021 device_flush_dte(dev_data);
2025 * If a device is not yet associated with a domain, this function does
2026 * assigns it visible for the hardware
2028 static int __attach_device(struct iommu_dev_data *dev_data,
2029 struct protection_domain *domain)
2031 struct iommu_dev_data *head, *entry;
2035 spin_lock(&domain->lock);
2039 if (head->alias_data != NULL)
2040 head = head->alias_data;
2042 /* Now we have the root of the alias group, if any */
2045 if (head->domain != NULL)
2048 /* Attach alias group root */
2049 do_attach(head, domain);
2051 /* Attach other devices in the alias group */
2052 list_for_each_entry(entry, &head->alias_list, alias_list)
2053 do_attach(entry, domain);
2060 spin_unlock(&domain->lock);
2066 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2068 pci_disable_ats(pdev);
2069 pci_disable_pri(pdev);
2070 pci_disable_pasid(pdev);
2073 /* FIXME: Change generic reset-function to do the same */
2074 static int pri_reset_while_enabled(struct pci_dev *pdev)
2079 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2083 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2084 control |= PCI_PRI_CTRL_RESET;
2085 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2090 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2095 /* FIXME: Hardcode number of outstanding requests for now */
2097 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2099 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2101 /* Only allow access to user-accessible pages */
2102 ret = pci_enable_pasid(pdev, 0);
2106 /* First reset the PRI state of the device */
2107 ret = pci_reset_pri(pdev);
2112 ret = pci_enable_pri(pdev, reqs);
2117 ret = pri_reset_while_enabled(pdev);
2122 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2129 pci_disable_pri(pdev);
2130 pci_disable_pasid(pdev);
2135 /* FIXME: Move this to PCI code */
2136 #define PCI_PRI_TLP_OFF (1 << 15)
2138 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2143 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2147 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2149 return (status & PCI_PRI_TLP_OFF) ? true : false;
2153 * If a device is not yet associated with a domain, this function
2154 * assigns it visible for the hardware
2156 static int attach_device(struct device *dev,
2157 struct protection_domain *domain)
2159 struct pci_dev *pdev = to_pci_dev(dev);
2160 struct iommu_dev_data *dev_data;
2161 unsigned long flags;
2164 dev_data = get_dev_data(dev);
2166 if (domain->flags & PD_IOMMUV2_MASK) {
2167 if (!dev_data->passthrough)
2170 if (dev_data->iommu_v2) {
2171 if (pdev_iommuv2_enable(pdev) != 0)
2174 dev_data->ats.enabled = true;
2175 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2176 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2178 } else if (amd_iommu_iotlb_sup &&
2179 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2180 dev_data->ats.enabled = true;
2181 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2184 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2185 ret = __attach_device(dev_data, domain);
2186 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2189 * We might boot into a crash-kernel here. The crashed kernel
2190 * left the caches in the IOMMU dirty. So we have to flush
2191 * here to evict all dirty stuff.
2193 domain_flush_tlb_pde(domain);
2199 * Removes a device from a protection domain (unlocked)
2201 static void __detach_device(struct iommu_dev_data *dev_data)
2203 struct iommu_dev_data *head, *entry;
2204 struct protection_domain *domain;
2205 unsigned long flags;
2207 BUG_ON(!dev_data->domain);
2209 domain = dev_data->domain;
2211 spin_lock_irqsave(&domain->lock, flags);
2214 if (head->alias_data != NULL)
2215 head = head->alias_data;
2217 list_for_each_entry(entry, &head->alias_list, alias_list)
2222 spin_unlock_irqrestore(&domain->lock, flags);
2226 * Removes a device from a protection domain (with devtable_lock held)
2228 static void detach_device(struct device *dev)
2230 struct protection_domain *domain;
2231 struct iommu_dev_data *dev_data;
2232 unsigned long flags;
2234 dev_data = get_dev_data(dev);
2235 domain = dev_data->domain;
2237 /* lock device table */
2238 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2239 __detach_device(dev_data);
2240 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2242 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2243 pdev_iommuv2_disable(to_pci_dev(dev));
2244 else if (dev_data->ats.enabled)
2245 pci_disable_ats(to_pci_dev(dev));
2247 dev_data->ats.enabled = false;
2250 static int amd_iommu_add_device(struct device *dev)
2252 struct iommu_dev_data *dev_data;
2253 struct iommu_domain *domain;
2254 struct amd_iommu *iommu;
2258 if (!check_device(dev) || get_dev_data(dev))
2261 devid = get_device_id(dev);
2262 iommu = amd_iommu_rlookup_table[devid];
2264 ret = iommu_init_device(dev);
2266 if (ret != -ENOTSUPP)
2267 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2270 iommu_ignore_device(dev);
2271 dev->archdata.dma_ops = &nommu_dma_ops;
2274 init_iommu_group(dev);
2276 dev_data = get_dev_data(dev);
2280 if (iommu_pass_through || dev_data->iommu_v2)
2281 iommu_request_dm_for_dev(dev);
2283 /* Domains are initialized for this device - have a look what we ended up with */
2284 domain = iommu_get_domain_for_dev(dev);
2285 if (domain->type == IOMMU_DOMAIN_IDENTITY) {
2286 dev_data->passthrough = true;
2287 dev->archdata.dma_ops = &nommu_dma_ops;
2289 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2293 iommu_completion_wait(iommu);
2298 static void amd_iommu_remove_device(struct device *dev)
2300 struct amd_iommu *iommu;
2303 if (!check_device(dev))
2306 devid = get_device_id(dev);
2307 iommu = amd_iommu_rlookup_table[devid];
2309 iommu_uninit_device(dev);
2310 iommu_completion_wait(iommu);
2313 /*****************************************************************************
2315 * The next functions belong to the dma_ops mapping/unmapping code.
2317 *****************************************************************************/
2320 * In the dma_ops path we only have the struct device. This function
2321 * finds the corresponding IOMMU, the protection domain and the
2322 * requestor id for a given device.
2323 * If the device is not yet associated with a domain this is also done
2326 static struct protection_domain *get_domain(struct device *dev)
2328 struct protection_domain *domain;
2329 struct iommu_domain *io_domain;
2331 if (!check_device(dev))
2332 return ERR_PTR(-EINVAL);
2334 io_domain = iommu_get_domain_for_dev(dev);
2338 domain = to_pdomain(io_domain);
2339 if (!dma_ops_domain(domain))
2340 return ERR_PTR(-EBUSY);
2345 static void update_device_table(struct protection_domain *domain)
2347 struct iommu_dev_data *dev_data;
2349 list_for_each_entry(dev_data, &domain->dev_list, list)
2350 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2353 static void update_domain(struct protection_domain *domain)
2355 if (!domain->updated)
2358 update_device_table(domain);
2360 domain_flush_devices(domain);
2361 domain_flush_tlb_pde(domain);
2363 domain->updated = false;
2367 * This function fetches the PTE for a given address in the aperture
2369 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2370 unsigned long address)
2372 struct aperture_range *aperture;
2373 u64 *pte, *pte_page;
2375 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2379 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2381 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2383 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2385 pte += PM_LEVEL_INDEX(0, address);
2387 update_domain(&dom->domain);
2393 * This is the generic map function. It maps one 4kb page at paddr to
2394 * the given address in the DMA address space for the domain.
2396 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2397 unsigned long address,
2403 WARN_ON(address > dom->aperture_size);
2407 pte = dma_ops_get_pte(dom, address);
2409 return DMA_ERROR_CODE;
2411 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2413 if (direction == DMA_TO_DEVICE)
2414 __pte |= IOMMU_PTE_IR;
2415 else if (direction == DMA_FROM_DEVICE)
2416 __pte |= IOMMU_PTE_IW;
2417 else if (direction == DMA_BIDIRECTIONAL)
2418 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2424 return (dma_addr_t)address;
2428 * The generic unmapping function for on page in the DMA address space.
2430 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2431 unsigned long address)
2433 struct aperture_range *aperture;
2436 if (address >= dom->aperture_size)
2439 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2443 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2447 pte += PM_LEVEL_INDEX(0, address);
2455 * This function contains common code for mapping of a physically
2456 * contiguous memory region into DMA address space. It is used by all
2457 * mapping functions provided with this IOMMU driver.
2458 * Must be called with the domain lock held.
2460 static dma_addr_t __map_single(struct device *dev,
2461 struct dma_ops_domain *dma_dom,
2468 dma_addr_t offset = paddr & ~PAGE_MASK;
2469 dma_addr_t address, start, ret;
2471 unsigned long align_mask = 0;
2474 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2477 INC_STATS_COUNTER(total_map_requests);
2480 INC_STATS_COUNTER(cross_page);
2483 align_mask = (1UL << get_order(size)) - 1;
2486 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2488 if (unlikely(address == DMA_ERROR_CODE)) {
2490 * setting next_address here will let the address
2491 * allocator only scan the new allocated range in the
2492 * first run. This is a small optimization.
2494 dma_dom->next_address = dma_dom->aperture_size;
2496 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2500 * aperture was successfully enlarged by 128 MB, try
2507 for (i = 0; i < pages; ++i) {
2508 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2509 if (ret == DMA_ERROR_CODE)
2517 ADD_STATS_COUNTER(alloced_io_mem, size);
2519 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2520 domain_flush_tlb(&dma_dom->domain);
2521 dma_dom->need_flush = false;
2522 } else if (unlikely(amd_iommu_np_cache))
2523 domain_flush_pages(&dma_dom->domain, address, size);
2530 for (--i; i >= 0; --i) {
2532 dma_ops_domain_unmap(dma_dom, start);
2535 dma_ops_free_addresses(dma_dom, address, pages);
2537 return DMA_ERROR_CODE;
2541 * Does the reverse of the __map_single function. Must be called with
2542 * the domain lock held too
2544 static void __unmap_single(struct dma_ops_domain *dma_dom,
2545 dma_addr_t dma_addr,
2549 dma_addr_t flush_addr;
2550 dma_addr_t i, start;
2553 if ((dma_addr == DMA_ERROR_CODE) ||
2554 (dma_addr + size > dma_dom->aperture_size))
2557 flush_addr = dma_addr;
2558 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2559 dma_addr &= PAGE_MASK;
2562 for (i = 0; i < pages; ++i) {
2563 dma_ops_domain_unmap(dma_dom, start);
2567 SUB_STATS_COUNTER(alloced_io_mem, size);
2569 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2571 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2572 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2573 dma_dom->need_flush = false;
2578 * The exported map_single function for dma_ops.
2580 static dma_addr_t map_page(struct device *dev, struct page *page,
2581 unsigned long offset, size_t size,
2582 enum dma_data_direction dir,
2583 struct dma_attrs *attrs)
2585 unsigned long flags;
2586 struct protection_domain *domain;
2589 phys_addr_t paddr = page_to_phys(page) + offset;
2591 INC_STATS_COUNTER(cnt_map_single);
2593 domain = get_domain(dev);
2594 if (PTR_ERR(domain) == -EINVAL)
2595 return (dma_addr_t)paddr;
2596 else if (IS_ERR(domain))
2597 return DMA_ERROR_CODE;
2599 dma_mask = *dev->dma_mask;
2601 spin_lock_irqsave(&domain->lock, flags);
2603 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2605 if (addr == DMA_ERROR_CODE)
2608 domain_flush_complete(domain);
2611 spin_unlock_irqrestore(&domain->lock, flags);
2617 * The exported unmap_single function for dma_ops.
2619 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2620 enum dma_data_direction dir, struct dma_attrs *attrs)
2622 unsigned long flags;
2623 struct protection_domain *domain;
2625 INC_STATS_COUNTER(cnt_unmap_single);
2627 domain = get_domain(dev);
2631 spin_lock_irqsave(&domain->lock, flags);
2633 __unmap_single(domain->priv, dma_addr, size, dir);
2635 domain_flush_complete(domain);
2637 spin_unlock_irqrestore(&domain->lock, flags);
2641 * The exported map_sg function for dma_ops (handles scatter-gather
2644 static int map_sg(struct device *dev, struct scatterlist *sglist,
2645 int nelems, enum dma_data_direction dir,
2646 struct dma_attrs *attrs)
2648 unsigned long flags;
2649 struct protection_domain *domain;
2651 struct scatterlist *s;
2653 int mapped_elems = 0;
2656 INC_STATS_COUNTER(cnt_map_sg);
2658 domain = get_domain(dev);
2662 dma_mask = *dev->dma_mask;
2664 spin_lock_irqsave(&domain->lock, flags);
2666 for_each_sg(sglist, s, nelems, i) {
2669 s->dma_address = __map_single(dev, domain->priv,
2670 paddr, s->length, dir, false,
2673 if (s->dma_address) {
2674 s->dma_length = s->length;
2680 domain_flush_complete(domain);
2683 spin_unlock_irqrestore(&domain->lock, flags);
2685 return mapped_elems;
2687 for_each_sg(sglist, s, mapped_elems, i) {
2689 __unmap_single(domain->priv, s->dma_address,
2690 s->dma_length, dir);
2691 s->dma_address = s->dma_length = 0;
2700 * The exported map_sg function for dma_ops (handles scatter-gather
2703 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2704 int nelems, enum dma_data_direction dir,
2705 struct dma_attrs *attrs)
2707 unsigned long flags;
2708 struct protection_domain *domain;
2709 struct scatterlist *s;
2712 INC_STATS_COUNTER(cnt_unmap_sg);
2714 domain = get_domain(dev);
2718 spin_lock_irqsave(&domain->lock, flags);
2720 for_each_sg(sglist, s, nelems, i) {
2721 __unmap_single(domain->priv, s->dma_address,
2722 s->dma_length, dir);
2723 s->dma_address = s->dma_length = 0;
2726 domain_flush_complete(domain);
2728 spin_unlock_irqrestore(&domain->lock, flags);
2732 * The exported alloc_coherent function for dma_ops.
2734 static void *alloc_coherent(struct device *dev, size_t size,
2735 dma_addr_t *dma_addr, gfp_t flag,
2736 struct dma_attrs *attrs)
2738 u64 dma_mask = dev->coherent_dma_mask;
2739 struct protection_domain *domain;
2740 unsigned long flags;
2743 INC_STATS_COUNTER(cnt_alloc_coherent);
2745 domain = get_domain(dev);
2746 if (PTR_ERR(domain) == -EINVAL) {
2747 page = alloc_pages(flag, get_order(size));
2748 *dma_addr = page_to_phys(page);
2749 return page_address(page);
2750 } else if (IS_ERR(domain))
2753 size = PAGE_ALIGN(size);
2754 dma_mask = dev->coherent_dma_mask;
2755 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2758 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2760 if (!(flag & __GFP_WAIT))
2763 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2770 dma_mask = *dev->dma_mask;
2772 spin_lock_irqsave(&domain->lock, flags);
2774 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2775 size, DMA_BIDIRECTIONAL, true, dma_mask);
2777 if (*dma_addr == DMA_ERROR_CODE) {
2778 spin_unlock_irqrestore(&domain->lock, flags);
2782 domain_flush_complete(domain);
2784 spin_unlock_irqrestore(&domain->lock, flags);
2786 return page_address(page);
2790 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2791 __free_pages(page, get_order(size));
2797 * The exported free_coherent function for dma_ops.
2799 static void free_coherent(struct device *dev, size_t size,
2800 void *virt_addr, dma_addr_t dma_addr,
2801 struct dma_attrs *attrs)
2803 struct protection_domain *domain;
2804 unsigned long flags;
2807 INC_STATS_COUNTER(cnt_free_coherent);
2809 page = virt_to_page(virt_addr);
2810 size = PAGE_ALIGN(size);
2812 domain = get_domain(dev);
2816 spin_lock_irqsave(&domain->lock, flags);
2818 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2820 domain_flush_complete(domain);
2822 spin_unlock_irqrestore(&domain->lock, flags);
2825 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2826 __free_pages(page, get_order(size));
2830 * This function is called by the DMA layer to find out if we can handle a
2831 * particular device. It is part of the dma_ops.
2833 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2835 return check_device(dev);
2838 static struct dma_map_ops amd_iommu_dma_ops = {
2839 .alloc = alloc_coherent,
2840 .free = free_coherent,
2841 .map_page = map_page,
2842 .unmap_page = unmap_page,
2844 .unmap_sg = unmap_sg,
2845 .dma_supported = amd_iommu_dma_supported,
2848 int __init amd_iommu_init_api(void)
2850 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2853 int __init amd_iommu_init_dma_ops(void)
2858 amd_iommu_stats_init();
2860 if (amd_iommu_unmap_flush)
2861 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2863 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2868 /*****************************************************************************
2870 * The following functions belong to the exported interface of AMD IOMMU
2872 * This interface allows access to lower level functions of the IOMMU
2873 * like protection domain handling and assignement of devices to domains
2874 * which is not possible with the dma_ops interface.
2876 *****************************************************************************/
2878 static void cleanup_domain(struct protection_domain *domain)
2880 struct iommu_dev_data *entry;
2881 unsigned long flags;
2883 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2885 while (!list_empty(&domain->dev_list)) {
2886 entry = list_first_entry(&domain->dev_list,
2887 struct iommu_dev_data, list);
2888 __detach_device(entry);
2891 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2894 static void protection_domain_free(struct protection_domain *domain)
2899 del_domain_from_list(domain);
2902 domain_id_free(domain->id);
2907 static int protection_domain_init(struct protection_domain *domain)
2909 spin_lock_init(&domain->lock);
2910 mutex_init(&domain->api_lock);
2911 domain->id = domain_id_alloc();
2914 INIT_LIST_HEAD(&domain->dev_list);
2919 static struct protection_domain *protection_domain_alloc(void)
2921 struct protection_domain *domain;
2923 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2927 if (protection_domain_init(domain))
2930 add_domain_to_list(domain);
2940 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2942 struct protection_domain *pdomain;
2943 struct dma_ops_domain *dma_domain;
2946 case IOMMU_DOMAIN_UNMANAGED:
2947 pdomain = protection_domain_alloc();
2951 pdomain->mode = PAGE_MODE_3_LEVEL;
2952 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2953 if (!pdomain->pt_root) {
2954 protection_domain_free(pdomain);
2958 pdomain->domain.geometry.aperture_start = 0;
2959 pdomain->domain.geometry.aperture_end = ~0ULL;
2960 pdomain->domain.geometry.force_aperture = true;
2963 case IOMMU_DOMAIN_DMA:
2964 dma_domain = dma_ops_domain_alloc();
2966 pr_err("AMD-Vi: Failed to allocate\n");
2969 pdomain = &dma_domain->domain;
2971 case IOMMU_DOMAIN_IDENTITY:
2972 pdomain = protection_domain_alloc();
2976 pdomain->mode = PAGE_MODE_NONE;
2982 return &pdomain->domain;
2985 static void amd_iommu_domain_free(struct iommu_domain *dom)
2987 struct protection_domain *domain;
2992 domain = to_pdomain(dom);
2994 if (domain->dev_cnt > 0)
2995 cleanup_domain(domain);
2997 BUG_ON(domain->dev_cnt != 0);
2999 if (domain->mode != PAGE_MODE_NONE)
3000 free_pagetable(domain);
3002 if (domain->flags & PD_IOMMUV2_MASK)
3003 free_gcr3_table(domain);
3005 protection_domain_free(domain);
3008 static void amd_iommu_detach_device(struct iommu_domain *dom,
3011 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3012 struct amd_iommu *iommu;
3015 if (!check_device(dev))
3018 devid = get_device_id(dev);
3020 if (dev_data->domain != NULL)
3023 iommu = amd_iommu_rlookup_table[devid];
3027 iommu_completion_wait(iommu);
3030 static int amd_iommu_attach_device(struct iommu_domain *dom,
3033 struct protection_domain *domain = to_pdomain(dom);
3034 struct iommu_dev_data *dev_data;
3035 struct amd_iommu *iommu;
3038 if (!check_device(dev))
3041 dev_data = dev->archdata.iommu;
3043 iommu = amd_iommu_rlookup_table[dev_data->devid];
3047 if (dev_data->domain)
3050 ret = attach_device(dev, domain);
3052 iommu_completion_wait(iommu);
3057 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3058 phys_addr_t paddr, size_t page_size, int iommu_prot)
3060 struct protection_domain *domain = to_pdomain(dom);
3064 if (domain->mode == PAGE_MODE_NONE)
3067 if (iommu_prot & IOMMU_READ)
3068 prot |= IOMMU_PROT_IR;
3069 if (iommu_prot & IOMMU_WRITE)
3070 prot |= IOMMU_PROT_IW;
3072 mutex_lock(&domain->api_lock);
3073 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3074 mutex_unlock(&domain->api_lock);
3079 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3082 struct protection_domain *domain = to_pdomain(dom);
3085 if (domain->mode == PAGE_MODE_NONE)
3088 mutex_lock(&domain->api_lock);
3089 unmap_size = iommu_unmap_page(domain, iova, page_size);
3090 mutex_unlock(&domain->api_lock);
3092 domain_flush_tlb_pde(domain);
3097 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3100 struct protection_domain *domain = to_pdomain(dom);
3101 unsigned long offset_mask, pte_pgsize;
3104 if (domain->mode == PAGE_MODE_NONE)
3107 pte = fetch_pte(domain, iova, &pte_pgsize);
3109 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3112 offset_mask = pte_pgsize - 1;
3113 __pte = *pte & PM_ADDR_MASK;
3115 return (__pte & ~offset_mask) | (iova & offset_mask);
3118 static bool amd_iommu_capable(enum iommu_cap cap)
3121 case IOMMU_CAP_CACHE_COHERENCY:
3123 case IOMMU_CAP_INTR_REMAP:
3124 return (irq_remapping_enabled == 1);
3125 case IOMMU_CAP_NOEXEC:
3132 static void amd_iommu_get_dm_regions(struct device *dev,
3133 struct list_head *head)
3135 struct unity_map_entry *entry;
3138 devid = get_device_id(dev);
3140 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3141 struct iommu_dm_region *region;
3143 if (devid < entry->devid_start || devid > entry->devid_end)
3146 region = kzalloc(sizeof(*region), GFP_KERNEL);
3148 pr_err("Out of memory allocating dm-regions for %s\n",
3153 region->start = entry->address_start;
3154 region->length = entry->address_end - entry->address_start;
3155 if (entry->prot & IOMMU_PROT_IR)
3156 region->prot |= IOMMU_READ;
3157 if (entry->prot & IOMMU_PROT_IW)
3158 region->prot |= IOMMU_WRITE;
3160 list_add_tail(®ion->list, head);
3164 static void amd_iommu_put_dm_regions(struct device *dev,
3165 struct list_head *head)
3167 struct iommu_dm_region *entry, *next;
3169 list_for_each_entry_safe(entry, next, head, list)
3173 static const struct iommu_ops amd_iommu_ops = {
3174 .capable = amd_iommu_capable,
3175 .domain_alloc = amd_iommu_domain_alloc,
3176 .domain_free = amd_iommu_domain_free,
3177 .attach_dev = amd_iommu_attach_device,
3178 .detach_dev = amd_iommu_detach_device,
3179 .map = amd_iommu_map,
3180 .unmap = amd_iommu_unmap,
3181 .map_sg = default_iommu_map_sg,
3182 .iova_to_phys = amd_iommu_iova_to_phys,
3183 .add_device = amd_iommu_add_device,
3184 .remove_device = amd_iommu_remove_device,
3185 .get_dm_regions = amd_iommu_get_dm_regions,
3186 .put_dm_regions = amd_iommu_put_dm_regions,
3187 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3190 /*****************************************************************************
3192 * The next functions do a basic initialization of IOMMU for pass through
3195 * In passthrough mode the IOMMU is initialized and enabled but not used for
3196 * DMA-API translation.
3198 *****************************************************************************/
3200 /* IOMMUv2 specific functions */
3201 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3203 return atomic_notifier_chain_register(&ppr_notifier, nb);
3205 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3207 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3209 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3211 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3213 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3215 struct protection_domain *domain = to_pdomain(dom);
3216 unsigned long flags;
3218 spin_lock_irqsave(&domain->lock, flags);
3220 /* Update data structure */
3221 domain->mode = PAGE_MODE_NONE;
3222 domain->updated = true;
3224 /* Make changes visible to IOMMUs */
3225 update_domain(domain);
3227 /* Page-table is not visible to IOMMU anymore, so free it */
3228 free_pagetable(domain);
3230 spin_unlock_irqrestore(&domain->lock, flags);
3232 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3234 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3236 struct protection_domain *domain = to_pdomain(dom);
3237 unsigned long flags;
3240 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3243 /* Number of GCR3 table levels required */
3244 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3247 if (levels > amd_iommu_max_glx_val)
3250 spin_lock_irqsave(&domain->lock, flags);
3253 * Save us all sanity checks whether devices already in the
3254 * domain support IOMMUv2. Just force that the domain has no
3255 * devices attached when it is switched into IOMMUv2 mode.
3258 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3262 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3263 if (domain->gcr3_tbl == NULL)
3266 domain->glx = levels;
3267 domain->flags |= PD_IOMMUV2_MASK;
3268 domain->updated = true;
3270 update_domain(domain);
3275 spin_unlock_irqrestore(&domain->lock, flags);
3279 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3281 static int __flush_pasid(struct protection_domain *domain, int pasid,
3282 u64 address, bool size)
3284 struct iommu_dev_data *dev_data;
3285 struct iommu_cmd cmd;
3288 if (!(domain->flags & PD_IOMMUV2_MASK))
3291 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3294 * IOMMU TLB needs to be flushed before Device TLB to
3295 * prevent device TLB refill from IOMMU TLB
3297 for (i = 0; i < amd_iommus_present; ++i) {
3298 if (domain->dev_iommu[i] == 0)
3301 ret = iommu_queue_command(amd_iommus[i], &cmd);
3306 /* Wait until IOMMU TLB flushes are complete */
3307 domain_flush_complete(domain);
3309 /* Now flush device TLBs */
3310 list_for_each_entry(dev_data, &domain->dev_list, list) {
3311 struct amd_iommu *iommu;
3314 BUG_ON(!dev_data->ats.enabled);
3316 qdep = dev_data->ats.qdep;
3317 iommu = amd_iommu_rlookup_table[dev_data->devid];
3319 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3320 qdep, address, size);
3322 ret = iommu_queue_command(iommu, &cmd);
3327 /* Wait until all device TLBs are flushed */
3328 domain_flush_complete(domain);
3337 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3340 INC_STATS_COUNTER(invalidate_iotlb);
3342 return __flush_pasid(domain, pasid, address, false);
3345 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3348 struct protection_domain *domain = to_pdomain(dom);
3349 unsigned long flags;
3352 spin_lock_irqsave(&domain->lock, flags);
3353 ret = __amd_iommu_flush_page(domain, pasid, address);
3354 spin_unlock_irqrestore(&domain->lock, flags);
3358 EXPORT_SYMBOL(amd_iommu_flush_page);
3360 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3362 INC_STATS_COUNTER(invalidate_iotlb_all);
3364 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3368 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3370 struct protection_domain *domain = to_pdomain(dom);
3371 unsigned long flags;
3374 spin_lock_irqsave(&domain->lock, flags);
3375 ret = __amd_iommu_flush_tlb(domain, pasid);
3376 spin_unlock_irqrestore(&domain->lock, flags);
3380 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3382 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3389 index = (pasid >> (9 * level)) & 0x1ff;
3395 if (!(*pte & GCR3_VALID)) {
3399 root = (void *)get_zeroed_page(GFP_ATOMIC);
3403 *pte = __pa(root) | GCR3_VALID;
3406 root = __va(*pte & PAGE_MASK);
3414 static int __set_gcr3(struct protection_domain *domain, int pasid,
3419 if (domain->mode != PAGE_MODE_NONE)
3422 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3426 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3428 return __amd_iommu_flush_tlb(domain, pasid);
3431 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3435 if (domain->mode != PAGE_MODE_NONE)
3438 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3444 return __amd_iommu_flush_tlb(domain, pasid);
3447 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3450 struct protection_domain *domain = to_pdomain(dom);
3451 unsigned long flags;
3454 spin_lock_irqsave(&domain->lock, flags);
3455 ret = __set_gcr3(domain, pasid, cr3);
3456 spin_unlock_irqrestore(&domain->lock, flags);
3460 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3462 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3464 struct protection_domain *domain = to_pdomain(dom);
3465 unsigned long flags;
3468 spin_lock_irqsave(&domain->lock, flags);
3469 ret = __clear_gcr3(domain, pasid);
3470 spin_unlock_irqrestore(&domain->lock, flags);
3474 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3476 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3477 int status, int tag)
3479 struct iommu_dev_data *dev_data;
3480 struct amd_iommu *iommu;
3481 struct iommu_cmd cmd;
3483 INC_STATS_COUNTER(complete_ppr);
3485 dev_data = get_dev_data(&pdev->dev);
3486 iommu = amd_iommu_rlookup_table[dev_data->devid];
3488 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3489 tag, dev_data->pri_tlp);
3491 return iommu_queue_command(iommu, &cmd);
3493 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3495 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3497 struct protection_domain *pdomain;
3499 pdomain = get_domain(&pdev->dev);
3500 if (IS_ERR(pdomain))
3503 /* Only return IOMMUv2 domains */
3504 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3507 return &pdomain->domain;
3509 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3511 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3513 struct iommu_dev_data *dev_data;
3515 if (!amd_iommu_v2_supported())
3518 dev_data = get_dev_data(&pdev->dev);
3519 dev_data->errata |= (1 << erratum);
3521 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3523 int amd_iommu_device_info(struct pci_dev *pdev,
3524 struct amd_iommu_device_info *info)
3529 if (pdev == NULL || info == NULL)
3532 if (!amd_iommu_v2_supported())
3535 memset(info, 0, sizeof(*info));
3537 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3539 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3541 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3543 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3545 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3549 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3550 max_pasids = min(max_pasids, (1 << 20));
3552 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3553 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3555 features = pci_pasid_features(pdev);
3556 if (features & PCI_PASID_CAP_EXEC)
3557 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3558 if (features & PCI_PASID_CAP_PRIV)
3559 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3564 EXPORT_SYMBOL(amd_iommu_device_info);
3566 #ifdef CONFIG_IRQ_REMAP
3568 /*****************************************************************************
3570 * Interrupt Remapping Implementation
3572 *****************************************************************************/
3590 u16 devid; /* Device ID for IRTE table */
3591 u16 index; /* Index into IRTE table*/
3594 struct amd_ir_data {
3595 struct irq_2_irte irq_2_irte;
3596 union irte irte_entry;
3598 struct msi_msg msi_entry;
3602 static struct irq_chip amd_ir_chip;
3604 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3605 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3606 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3607 #define DTE_IRQ_REMAP_ENABLE 1ULL
3609 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3613 dte = amd_iommu_dev_table[devid].data[2];
3614 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3615 dte |= virt_to_phys(table->table);
3616 dte |= DTE_IRQ_REMAP_INTCTL;
3617 dte |= DTE_IRQ_TABLE_LEN;
3618 dte |= DTE_IRQ_REMAP_ENABLE;
3620 amd_iommu_dev_table[devid].data[2] = dte;
3623 #define IRTE_ALLOCATED (~1U)
3625 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3627 struct irq_remap_table *table = NULL;
3628 struct amd_iommu *iommu;
3629 unsigned long flags;
3632 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3634 iommu = amd_iommu_rlookup_table[devid];
3638 table = irq_lookup_table[devid];
3642 alias = amd_iommu_alias_table[devid];
3643 table = irq_lookup_table[alias];
3645 irq_lookup_table[devid] = table;
3646 set_dte_irq_entry(devid, table);
3647 iommu_flush_dte(iommu, devid);
3651 /* Nothing there yet, allocate new irq remapping table */
3652 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3656 /* Initialize table spin-lock */
3657 spin_lock_init(&table->lock);
3660 /* Keep the first 32 indexes free for IOAPIC interrupts */
3661 table->min_index = 32;
3663 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3664 if (!table->table) {
3670 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3675 for (i = 0; i < 32; ++i)
3676 table->table[i] = IRTE_ALLOCATED;
3679 irq_lookup_table[devid] = table;
3680 set_dte_irq_entry(devid, table);
3681 iommu_flush_dte(iommu, devid);
3682 if (devid != alias) {
3683 irq_lookup_table[alias] = table;
3684 set_dte_irq_entry(alias, table);
3685 iommu_flush_dte(iommu, alias);
3689 iommu_completion_wait(iommu);
3692 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3697 static int alloc_irq_index(u16 devid, int count)
3699 struct irq_remap_table *table;
3700 unsigned long flags;
3703 table = get_irq_table(devid, false);
3707 spin_lock_irqsave(&table->lock, flags);
3709 /* Scan table for free entries */
3710 for (c = 0, index = table->min_index;
3711 index < MAX_IRQS_PER_TABLE;
3713 if (table->table[index] == 0)
3720 table->table[index - c + 1] = IRTE_ALLOCATED;
3730 spin_unlock_irqrestore(&table->lock, flags);
3735 static int modify_irte(u16 devid, int index, union irte irte)
3737 struct irq_remap_table *table;
3738 struct amd_iommu *iommu;
3739 unsigned long flags;
3741 iommu = amd_iommu_rlookup_table[devid];
3745 table = get_irq_table(devid, false);
3749 spin_lock_irqsave(&table->lock, flags);
3750 table->table[index] = irte.val;
3751 spin_unlock_irqrestore(&table->lock, flags);
3753 iommu_flush_irt(iommu, devid);
3754 iommu_completion_wait(iommu);
3759 static void free_irte(u16 devid, int index)
3761 struct irq_remap_table *table;
3762 struct amd_iommu *iommu;
3763 unsigned long flags;
3765 iommu = amd_iommu_rlookup_table[devid];
3769 table = get_irq_table(devid, false);
3773 spin_lock_irqsave(&table->lock, flags);
3774 table->table[index] = 0;
3775 spin_unlock_irqrestore(&table->lock, flags);
3777 iommu_flush_irt(iommu, devid);
3778 iommu_completion_wait(iommu);
3781 static int get_devid(struct irq_alloc_info *info)
3785 switch (info->type) {
3786 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3787 devid = get_ioapic_devid(info->ioapic_id);
3789 case X86_IRQ_ALLOC_TYPE_HPET:
3790 devid = get_hpet_devid(info->hpet_id);
3792 case X86_IRQ_ALLOC_TYPE_MSI:
3793 case X86_IRQ_ALLOC_TYPE_MSIX:
3794 devid = get_device_id(&info->msi_dev->dev);
3804 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3806 struct amd_iommu *iommu;
3812 devid = get_devid(info);
3814 iommu = amd_iommu_rlookup_table[devid];
3816 return iommu->ir_domain;
3822 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3824 struct amd_iommu *iommu;
3830 switch (info->type) {
3831 case X86_IRQ_ALLOC_TYPE_MSI:
3832 case X86_IRQ_ALLOC_TYPE_MSIX:
3833 devid = get_device_id(&info->msi_dev->dev);
3835 iommu = amd_iommu_rlookup_table[devid];
3837 return iommu->msi_domain;
3847 struct irq_remap_ops amd_iommu_irq_ops = {
3848 .prepare = amd_iommu_prepare,
3849 .enable = amd_iommu_enable,
3850 .disable = amd_iommu_disable,
3851 .reenable = amd_iommu_reenable,
3852 .enable_faulting = amd_iommu_enable_faulting,
3853 .get_ir_irq_domain = get_ir_irq_domain,
3854 .get_irq_domain = get_irq_domain,
3857 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3858 struct irq_cfg *irq_cfg,
3859 struct irq_alloc_info *info,
3860 int devid, int index, int sub_handle)
3862 struct irq_2_irte *irte_info = &data->irq_2_irte;
3863 struct msi_msg *msg = &data->msi_entry;
3864 union irte *irte = &data->irte_entry;
3865 struct IO_APIC_route_entry *entry;
3867 data->irq_2_irte.devid = devid;
3868 data->irq_2_irte.index = index + sub_handle;
3870 /* Setup IRTE for IOMMU */
3872 irte->fields.vector = irq_cfg->vector;
3873 irte->fields.int_type = apic->irq_delivery_mode;
3874 irte->fields.destination = irq_cfg->dest_apicid;
3875 irte->fields.dm = apic->irq_dest_mode;
3876 irte->fields.valid = 1;
3878 switch (info->type) {
3879 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3880 /* Setup IOAPIC entry */
3881 entry = info->ioapic_entry;
3882 info->ioapic_entry = NULL;
3883 memset(entry, 0, sizeof(*entry));
3884 entry->vector = index;
3886 entry->trigger = info->ioapic_trigger;
3887 entry->polarity = info->ioapic_polarity;
3888 /* Mask level triggered irqs. */
3889 if (info->ioapic_trigger)
3893 case X86_IRQ_ALLOC_TYPE_HPET:
3894 case X86_IRQ_ALLOC_TYPE_MSI:
3895 case X86_IRQ_ALLOC_TYPE_MSIX:
3896 msg->address_hi = MSI_ADDR_BASE_HI;
3897 msg->address_lo = MSI_ADDR_BASE_LO;
3898 msg->data = irte_info->index;
3907 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3908 unsigned int nr_irqs, void *arg)
3910 struct irq_alloc_info *info = arg;
3911 struct irq_data *irq_data;
3912 struct amd_ir_data *data;
3913 struct irq_cfg *cfg;
3919 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3920 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3924 * With IRQ remapping enabled, don't need contiguous CPU vectors
3925 * to support multiple MSI interrupts.
3927 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3928 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3930 devid = get_devid(info);
3934 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3939 data = kzalloc(sizeof(*data), GFP_KERNEL);
3941 goto out_free_parent;
3943 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3944 if (get_irq_table(devid, true))
3945 index = info->ioapic_pin;
3949 index = alloc_irq_index(devid, nr_irqs);
3952 pr_warn("Failed to allocate IRTE\n");
3954 goto out_free_parent;
3957 for (i = 0; i < nr_irqs; i++) {
3958 irq_data = irq_domain_get_irq_data(domain, virq + i);
3959 cfg = irqd_cfg(irq_data);
3960 if (!irq_data || !cfg) {
3966 data = kzalloc(sizeof(*data), GFP_KERNEL);
3970 irq_data->hwirq = (devid << 16) + i;
3971 irq_data->chip_data = data;
3972 irq_data->chip = &amd_ir_chip;
3973 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
3974 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
3979 for (i--; i >= 0; i--) {
3980 irq_data = irq_domain_get_irq_data(domain, virq + i);
3982 kfree(irq_data->chip_data);
3984 for (i = 0; i < nr_irqs; i++)
3985 free_irte(devid, index + i);
3987 irq_domain_free_irqs_common(domain, virq, nr_irqs);
3991 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
3992 unsigned int nr_irqs)
3994 struct irq_2_irte *irte_info;
3995 struct irq_data *irq_data;
3996 struct amd_ir_data *data;
3999 for (i = 0; i < nr_irqs; i++) {
4000 irq_data = irq_domain_get_irq_data(domain, virq + i);
4001 if (irq_data && irq_data->chip_data) {
4002 data = irq_data->chip_data;
4003 irte_info = &data->irq_2_irte;
4004 free_irte(irte_info->devid, irte_info->index);
4008 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4011 static void irq_remapping_activate(struct irq_domain *domain,
4012 struct irq_data *irq_data)
4014 struct amd_ir_data *data = irq_data->chip_data;
4015 struct irq_2_irte *irte_info = &data->irq_2_irte;
4017 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4020 static void irq_remapping_deactivate(struct irq_domain *domain,
4021 struct irq_data *irq_data)
4023 struct amd_ir_data *data = irq_data->chip_data;
4024 struct irq_2_irte *irte_info = &data->irq_2_irte;
4028 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4031 static struct irq_domain_ops amd_ir_domain_ops = {
4032 .alloc = irq_remapping_alloc,
4033 .free = irq_remapping_free,
4034 .activate = irq_remapping_activate,
4035 .deactivate = irq_remapping_deactivate,
4038 static int amd_ir_set_affinity(struct irq_data *data,
4039 const struct cpumask *mask, bool force)
4041 struct amd_ir_data *ir_data = data->chip_data;
4042 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4043 struct irq_cfg *cfg = irqd_cfg(data);
4044 struct irq_data *parent = data->parent_data;
4047 ret = parent->chip->irq_set_affinity(parent, mask, force);
4048 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4052 * Atomically updates the IRTE with the new destination, vector
4053 * and flushes the interrupt entry cache.
4055 ir_data->irte_entry.fields.vector = cfg->vector;
4056 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4057 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4060 * After this point, all the interrupts will start arriving
4061 * at the new destination. So, time to cleanup the previous
4062 * vector allocation.
4064 send_cleanup_vector(cfg);
4066 return IRQ_SET_MASK_OK_DONE;
4069 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4071 struct amd_ir_data *ir_data = irq_data->chip_data;
4073 *msg = ir_data->msi_entry;
4076 static struct irq_chip amd_ir_chip = {
4077 .irq_ack = ir_ack_apic_edge,
4078 .irq_set_affinity = amd_ir_set_affinity,
4079 .irq_compose_msi_msg = ir_compose_msi_msg,
4082 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4084 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4085 if (!iommu->ir_domain)
4088 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4089 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);