2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
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 struct protection_domain *pt_domain;
81 static const struct iommu_ops amd_iommu_ops;
83 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
84 int amd_iommu_max_glx_val = -1;
86 static struct dma_map_ops amd_iommu_dma_ops;
89 * This struct contains device specific data for the IOMMU
91 struct iommu_dev_data {
92 struct list_head list; /* For domain->dev_list */
93 struct list_head dev_data_list; /* For global dev_data_list */
94 struct list_head alias_list; /* Link alias-groups together */
95 struct iommu_dev_data *alias_data;/* The alias dev_data */
96 struct protection_domain *domain; /* Domain the device is bound to */
97 u16 devid; /* PCI Device ID */
98 bool iommu_v2; /* Device can make use of IOMMUv2 */
99 bool passthrough; /* Default for device is pt_domain */
103 } ats; /* ATS state */
104 bool pri_tlp; /* PASID TLB required for
106 u32 errata; /* Bitmap for errata to apply */
110 * general struct to manage commands send to an IOMMU
116 struct kmem_cache *amd_iommu_irq_cache;
118 static void update_domain(struct protection_domain *domain);
119 static int alloc_passthrough_domain(void);
120 static int protection_domain_init(struct protection_domain *domain);
122 /****************************************************************************
126 ****************************************************************************/
128 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
130 return container_of(dom, struct protection_domain, domain);
133 static struct iommu_dev_data *alloc_dev_data(u16 devid)
135 struct iommu_dev_data *dev_data;
138 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
142 INIT_LIST_HEAD(&dev_data->alias_list);
144 dev_data->devid = devid;
146 spin_lock_irqsave(&dev_data_list_lock, flags);
147 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
148 spin_unlock_irqrestore(&dev_data_list_lock, flags);
153 static void free_dev_data(struct iommu_dev_data *dev_data)
157 spin_lock_irqsave(&dev_data_list_lock, flags);
158 list_del(&dev_data->dev_data_list);
159 spin_unlock_irqrestore(&dev_data_list_lock, flags);
164 static struct iommu_dev_data *search_dev_data(u16 devid)
166 struct iommu_dev_data *dev_data;
169 spin_lock_irqsave(&dev_data_list_lock, flags);
170 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
171 if (dev_data->devid == devid)
178 spin_unlock_irqrestore(&dev_data_list_lock, flags);
183 static struct iommu_dev_data *find_dev_data(u16 devid)
185 struct iommu_dev_data *dev_data;
187 dev_data = search_dev_data(devid);
189 if (dev_data == NULL)
190 dev_data = alloc_dev_data(devid);
195 static inline u16 get_device_id(struct device *dev)
197 struct pci_dev *pdev = to_pci_dev(dev);
199 return PCI_DEVID(pdev->bus->number, pdev->devfn);
202 static struct iommu_dev_data *get_dev_data(struct device *dev)
204 return dev->archdata.iommu;
207 static bool pci_iommuv2_capable(struct pci_dev *pdev)
209 static const int caps[] = {
212 PCI_EXT_CAP_ID_PASID,
216 for (i = 0; i < 3; ++i) {
217 pos = pci_find_ext_capability(pdev, caps[i]);
225 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
227 struct iommu_dev_data *dev_data;
229 dev_data = get_dev_data(&pdev->dev);
231 return dev_data->errata & (1 << erratum) ? true : false;
235 * This function actually applies the mapping to the page table of the
238 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
239 struct unity_map_entry *e)
243 for (addr = e->address_start; addr < e->address_end;
245 if (addr < dma_dom->aperture_size)
246 __set_bit(addr >> PAGE_SHIFT,
247 dma_dom->aperture[0]->bitmap);
252 * Inits the unity mappings required for a specific device
254 static void init_unity_mappings_for_device(struct device *dev,
255 struct dma_ops_domain *dma_dom)
257 struct unity_map_entry *e;
260 devid = get_device_id(dev);
262 list_for_each_entry(e, &amd_iommu_unity_map, list) {
263 if (!(devid >= e->devid_start && devid <= e->devid_end))
265 alloc_unity_mapping(dma_dom, e);
270 * This function checks if the driver got a valid device from the caller to
271 * avoid dereferencing invalid pointers.
273 static bool check_device(struct device *dev)
277 if (!dev || !dev->dma_mask)
281 if (!dev_is_pci(dev))
284 devid = get_device_id(dev);
286 /* Out of our scope? */
287 if (devid > amd_iommu_last_bdf)
290 if (amd_iommu_rlookup_table[devid] == NULL)
296 static void init_iommu_group(struct device *dev)
298 struct dma_ops_domain *dma_domain;
299 struct iommu_domain *domain;
300 struct iommu_group *group;
302 group = iommu_group_get_for_dev(dev);
306 domain = iommu_group_default_domain(group);
310 dma_domain = to_pdomain(domain)->priv;
312 init_unity_mappings_for_device(dev, dma_domain);
314 iommu_group_put(group);
317 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
319 *(u16 *)data = alias;
323 static u16 get_alias(struct device *dev)
325 struct pci_dev *pdev = to_pci_dev(dev);
326 u16 devid, ivrs_alias, pci_alias;
328 devid = get_device_id(dev);
329 ivrs_alias = amd_iommu_alias_table[devid];
330 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
332 if (ivrs_alias == pci_alias)
338 * The IVRS is fairly reliable in telling us about aliases, but it
339 * can't know about every screwy device. If we don't have an IVRS
340 * reported alias, use the PCI reported alias. In that case we may
341 * still need to initialize the rlookup and dev_table entries if the
342 * alias is to a non-existent device.
344 if (ivrs_alias == devid) {
345 if (!amd_iommu_rlookup_table[pci_alias]) {
346 amd_iommu_rlookup_table[pci_alias] =
347 amd_iommu_rlookup_table[devid];
348 memcpy(amd_iommu_dev_table[pci_alias].data,
349 amd_iommu_dev_table[devid].data,
350 sizeof(amd_iommu_dev_table[pci_alias].data));
356 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
357 "for device %s[%04x:%04x], kernel reported alias "
358 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
359 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
360 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
361 PCI_FUNC(pci_alias));
364 * If we don't have a PCI DMA alias and the IVRS alias is on the same
365 * bus, then the IVRS table may know about a quirk that we don't.
367 if (pci_alias == devid &&
368 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
369 pdev->dev_flags |= PCI_DEV_FLAGS_DMA_ALIAS_DEVFN;
370 pdev->dma_alias_devfn = ivrs_alias & 0xff;
371 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
372 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
379 static int iommu_init_device(struct device *dev)
381 struct pci_dev *pdev = to_pci_dev(dev);
382 struct iommu_dev_data *dev_data;
385 if (dev->archdata.iommu)
388 dev_data = find_dev_data(get_device_id(dev));
392 alias = get_alias(dev);
394 if (alias != dev_data->devid) {
395 struct iommu_dev_data *alias_data;
397 alias_data = find_dev_data(alias);
398 if (alias_data == NULL) {
399 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
401 free_dev_data(dev_data);
404 dev_data->alias_data = alias_data;
406 /* Add device to the alias_list */
407 list_add(&dev_data->alias_list, &alias_data->alias_list);
410 if (pci_iommuv2_capable(pdev)) {
411 struct amd_iommu *iommu;
413 iommu = amd_iommu_rlookup_table[dev_data->devid];
414 dev_data->iommu_v2 = iommu->is_iommu_v2;
417 dev->archdata.iommu = dev_data;
419 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
425 static void iommu_ignore_device(struct device *dev)
429 devid = get_device_id(dev);
430 alias = amd_iommu_alias_table[devid];
432 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
433 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
435 amd_iommu_rlookup_table[devid] = NULL;
436 amd_iommu_rlookup_table[alias] = NULL;
439 static void iommu_uninit_device(struct device *dev)
441 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
446 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
449 iommu_group_remove_device(dev);
451 /* Unlink from alias, it may change if another device is re-plugged */
452 dev_data->alias_data = NULL;
455 dev->archdata.dma_ops = NULL;
458 * We keep dev_data around for unplugged devices and reuse it when the
459 * device is re-plugged - not doing so would introduce a ton of races.
463 #ifdef CONFIG_AMD_IOMMU_STATS
466 * Initialization code for statistics collection
469 DECLARE_STATS_COUNTER(compl_wait);
470 DECLARE_STATS_COUNTER(cnt_map_single);
471 DECLARE_STATS_COUNTER(cnt_unmap_single);
472 DECLARE_STATS_COUNTER(cnt_map_sg);
473 DECLARE_STATS_COUNTER(cnt_unmap_sg);
474 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
475 DECLARE_STATS_COUNTER(cnt_free_coherent);
476 DECLARE_STATS_COUNTER(cross_page);
477 DECLARE_STATS_COUNTER(domain_flush_single);
478 DECLARE_STATS_COUNTER(domain_flush_all);
479 DECLARE_STATS_COUNTER(alloced_io_mem);
480 DECLARE_STATS_COUNTER(total_map_requests);
481 DECLARE_STATS_COUNTER(complete_ppr);
482 DECLARE_STATS_COUNTER(invalidate_iotlb);
483 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
484 DECLARE_STATS_COUNTER(pri_requests);
486 static struct dentry *stats_dir;
487 static struct dentry *de_fflush;
489 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
491 if (stats_dir == NULL)
494 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
498 static void amd_iommu_stats_init(void)
500 stats_dir = debugfs_create_dir("amd-iommu", NULL);
501 if (stats_dir == NULL)
504 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
505 &amd_iommu_unmap_flush);
507 amd_iommu_stats_add(&compl_wait);
508 amd_iommu_stats_add(&cnt_map_single);
509 amd_iommu_stats_add(&cnt_unmap_single);
510 amd_iommu_stats_add(&cnt_map_sg);
511 amd_iommu_stats_add(&cnt_unmap_sg);
512 amd_iommu_stats_add(&cnt_alloc_coherent);
513 amd_iommu_stats_add(&cnt_free_coherent);
514 amd_iommu_stats_add(&cross_page);
515 amd_iommu_stats_add(&domain_flush_single);
516 amd_iommu_stats_add(&domain_flush_all);
517 amd_iommu_stats_add(&alloced_io_mem);
518 amd_iommu_stats_add(&total_map_requests);
519 amd_iommu_stats_add(&complete_ppr);
520 amd_iommu_stats_add(&invalidate_iotlb);
521 amd_iommu_stats_add(&invalidate_iotlb_all);
522 amd_iommu_stats_add(&pri_requests);
527 /****************************************************************************
529 * Interrupt handling functions
531 ****************************************************************************/
533 static void dump_dte_entry(u16 devid)
537 for (i = 0; i < 4; ++i)
538 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
539 amd_iommu_dev_table[devid].data[i]);
542 static void dump_command(unsigned long phys_addr)
544 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
547 for (i = 0; i < 4; ++i)
548 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
551 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
553 int type, devid, domid, flags;
554 volatile u32 *event = __evt;
559 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
560 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
561 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
562 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
563 address = (u64)(((u64)event[3]) << 32) | event[2];
566 /* Did we hit the erratum? */
567 if (++count == LOOP_TIMEOUT) {
568 pr_err("AMD-Vi: No event written to event log\n");
575 printk(KERN_ERR "AMD-Vi: Event logged [");
578 case EVENT_TYPE_ILL_DEV:
579 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
580 "address=0x%016llx flags=0x%04x]\n",
581 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
583 dump_dte_entry(devid);
585 case EVENT_TYPE_IO_FAULT:
586 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
587 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
588 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
589 domid, address, flags);
591 case EVENT_TYPE_DEV_TAB_ERR:
592 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
593 "address=0x%016llx flags=0x%04x]\n",
594 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
597 case EVENT_TYPE_PAGE_TAB_ERR:
598 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
599 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
600 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
601 domid, address, flags);
603 case EVENT_TYPE_ILL_CMD:
604 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
605 dump_command(address);
607 case EVENT_TYPE_CMD_HARD_ERR:
608 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
609 "flags=0x%04x]\n", address, flags);
611 case EVENT_TYPE_IOTLB_INV_TO:
612 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
613 "address=0x%016llx]\n",
614 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
617 case EVENT_TYPE_INV_DEV_REQ:
618 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
619 "address=0x%016llx flags=0x%04x]\n",
620 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
624 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
627 memset(__evt, 0, 4 * sizeof(u32));
630 static void iommu_poll_events(struct amd_iommu *iommu)
634 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
635 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
637 while (head != tail) {
638 iommu_print_event(iommu, iommu->evt_buf + head);
639 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
642 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
645 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
647 struct amd_iommu_fault fault;
649 INC_STATS_COUNTER(pri_requests);
651 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
652 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
656 fault.address = raw[1];
657 fault.pasid = PPR_PASID(raw[0]);
658 fault.device_id = PPR_DEVID(raw[0]);
659 fault.tag = PPR_TAG(raw[0]);
660 fault.flags = PPR_FLAGS(raw[0]);
662 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
665 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
669 if (iommu->ppr_log == NULL)
672 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
673 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
675 while (head != tail) {
680 raw = (u64 *)(iommu->ppr_log + head);
683 * Hardware bug: Interrupt may arrive before the entry is
684 * written to memory. If this happens we need to wait for the
687 for (i = 0; i < LOOP_TIMEOUT; ++i) {
688 if (PPR_REQ_TYPE(raw[0]) != 0)
693 /* Avoid memcpy function-call overhead */
698 * To detect the hardware bug we need to clear the entry
701 raw[0] = raw[1] = 0UL;
703 /* Update head pointer of hardware ring-buffer */
704 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
705 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
707 /* Handle PPR entry */
708 iommu_handle_ppr_entry(iommu, entry);
710 /* Refresh ring-buffer information */
711 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
712 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
716 irqreturn_t amd_iommu_int_thread(int irq, void *data)
718 struct amd_iommu *iommu = (struct amd_iommu *) data;
719 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
721 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
722 /* Enable EVT and PPR interrupts again */
723 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
724 iommu->mmio_base + MMIO_STATUS_OFFSET);
726 if (status & MMIO_STATUS_EVT_INT_MASK) {
727 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
728 iommu_poll_events(iommu);
731 if (status & MMIO_STATUS_PPR_INT_MASK) {
732 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
733 iommu_poll_ppr_log(iommu);
737 * Hardware bug: ERBT1312
738 * When re-enabling interrupt (by writing 1
739 * to clear the bit), the hardware might also try to set
740 * the interrupt bit in the event status register.
741 * In this scenario, the bit will be set, and disable
742 * subsequent interrupts.
744 * Workaround: The IOMMU driver should read back the
745 * status register and check if the interrupt bits are cleared.
746 * If not, driver will need to go through the interrupt handler
747 * again and re-clear the bits
749 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
754 irqreturn_t amd_iommu_int_handler(int irq, void *data)
756 return IRQ_WAKE_THREAD;
759 /****************************************************************************
761 * IOMMU command queuing functions
763 ****************************************************************************/
765 static int wait_on_sem(volatile u64 *sem)
769 while (*sem == 0 && i < LOOP_TIMEOUT) {
774 if (i == LOOP_TIMEOUT) {
775 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
782 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
783 struct iommu_cmd *cmd,
788 target = iommu->cmd_buf + tail;
789 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
791 /* Copy command to buffer */
792 memcpy(target, cmd, sizeof(*cmd));
794 /* Tell the IOMMU about it */
795 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
798 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
800 WARN_ON(address & 0x7ULL);
802 memset(cmd, 0, sizeof(*cmd));
803 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
804 cmd->data[1] = upper_32_bits(__pa(address));
806 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
809 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
811 memset(cmd, 0, sizeof(*cmd));
812 cmd->data[0] = devid;
813 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
816 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
817 size_t size, u16 domid, int pde)
822 pages = iommu_num_pages(address, size, PAGE_SIZE);
827 * If we have to flush more than one page, flush all
828 * TLB entries for this domain
830 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
834 address &= PAGE_MASK;
836 memset(cmd, 0, sizeof(*cmd));
837 cmd->data[1] |= domid;
838 cmd->data[2] = lower_32_bits(address);
839 cmd->data[3] = upper_32_bits(address);
840 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
841 if (s) /* size bit - we flush more than one 4kb page */
842 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
843 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
844 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
847 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
848 u64 address, size_t size)
853 pages = iommu_num_pages(address, size, PAGE_SIZE);
858 * If we have to flush more than one page, flush all
859 * TLB entries for this domain
861 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
865 address &= PAGE_MASK;
867 memset(cmd, 0, sizeof(*cmd));
868 cmd->data[0] = devid;
869 cmd->data[0] |= (qdep & 0xff) << 24;
870 cmd->data[1] = devid;
871 cmd->data[2] = lower_32_bits(address);
872 cmd->data[3] = upper_32_bits(address);
873 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
875 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
878 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
879 u64 address, bool size)
881 memset(cmd, 0, sizeof(*cmd));
883 address &= ~(0xfffULL);
885 cmd->data[0] = pasid;
886 cmd->data[1] = domid;
887 cmd->data[2] = lower_32_bits(address);
888 cmd->data[3] = upper_32_bits(address);
889 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
890 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
892 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
893 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
896 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
897 int qdep, u64 address, bool size)
899 memset(cmd, 0, sizeof(*cmd));
901 address &= ~(0xfffULL);
903 cmd->data[0] = devid;
904 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
905 cmd->data[0] |= (qdep & 0xff) << 24;
906 cmd->data[1] = devid;
907 cmd->data[1] |= (pasid & 0xff) << 16;
908 cmd->data[2] = lower_32_bits(address);
909 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
910 cmd->data[3] = upper_32_bits(address);
912 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
913 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
916 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
917 int status, int tag, bool gn)
919 memset(cmd, 0, sizeof(*cmd));
921 cmd->data[0] = devid;
923 cmd->data[1] = pasid;
924 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
926 cmd->data[3] = tag & 0x1ff;
927 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
929 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
932 static void build_inv_all(struct iommu_cmd *cmd)
934 memset(cmd, 0, sizeof(*cmd));
935 CMD_SET_TYPE(cmd, CMD_INV_ALL);
938 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
940 memset(cmd, 0, sizeof(*cmd));
941 cmd->data[0] = devid;
942 CMD_SET_TYPE(cmd, CMD_INV_IRT);
946 * Writes the command to the IOMMUs command buffer and informs the
947 * hardware about the new command.
949 static int iommu_queue_command_sync(struct amd_iommu *iommu,
950 struct iommu_cmd *cmd,
953 u32 left, tail, head, next_tail;
956 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
959 spin_lock_irqsave(&iommu->lock, flags);
961 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
962 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
963 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
964 left = (head - next_tail) % iommu->cmd_buf_size;
967 struct iommu_cmd sync_cmd;
968 volatile u64 sem = 0;
971 build_completion_wait(&sync_cmd, (u64)&sem);
972 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
974 spin_unlock_irqrestore(&iommu->lock, flags);
976 if ((ret = wait_on_sem(&sem)) != 0)
982 copy_cmd_to_buffer(iommu, cmd, tail);
984 /* We need to sync now to make sure all commands are processed */
985 iommu->need_sync = sync;
987 spin_unlock_irqrestore(&iommu->lock, flags);
992 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
994 return iommu_queue_command_sync(iommu, cmd, true);
998 * This function queues a completion wait command into the command
1001 static int iommu_completion_wait(struct amd_iommu *iommu)
1003 struct iommu_cmd cmd;
1004 volatile u64 sem = 0;
1007 if (!iommu->need_sync)
1010 build_completion_wait(&cmd, (u64)&sem);
1012 ret = iommu_queue_command_sync(iommu, &cmd, false);
1016 return wait_on_sem(&sem);
1019 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1021 struct iommu_cmd cmd;
1023 build_inv_dte(&cmd, devid);
1025 return iommu_queue_command(iommu, &cmd);
1028 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1032 for (devid = 0; devid <= 0xffff; ++devid)
1033 iommu_flush_dte(iommu, devid);
1035 iommu_completion_wait(iommu);
1039 * This function uses heavy locking and may disable irqs for some time. But
1040 * this is no issue because it is only called during resume.
1042 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1046 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1047 struct iommu_cmd cmd;
1048 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1050 iommu_queue_command(iommu, &cmd);
1053 iommu_completion_wait(iommu);
1056 static void iommu_flush_all(struct amd_iommu *iommu)
1058 struct iommu_cmd cmd;
1060 build_inv_all(&cmd);
1062 iommu_queue_command(iommu, &cmd);
1063 iommu_completion_wait(iommu);
1066 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1068 struct iommu_cmd cmd;
1070 build_inv_irt(&cmd, devid);
1072 iommu_queue_command(iommu, &cmd);
1075 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1079 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1080 iommu_flush_irt(iommu, devid);
1082 iommu_completion_wait(iommu);
1085 void iommu_flush_all_caches(struct amd_iommu *iommu)
1087 if (iommu_feature(iommu, FEATURE_IA)) {
1088 iommu_flush_all(iommu);
1090 iommu_flush_dte_all(iommu);
1091 iommu_flush_irt_all(iommu);
1092 iommu_flush_tlb_all(iommu);
1097 * Command send function for flushing on-device TLB
1099 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1100 u64 address, size_t size)
1102 struct amd_iommu *iommu;
1103 struct iommu_cmd cmd;
1106 qdep = dev_data->ats.qdep;
1107 iommu = amd_iommu_rlookup_table[dev_data->devid];
1109 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1111 return iommu_queue_command(iommu, &cmd);
1115 * Command send function for invalidating a device table entry
1117 static int device_flush_dte(struct iommu_dev_data *dev_data)
1119 struct amd_iommu *iommu;
1122 iommu = amd_iommu_rlookup_table[dev_data->devid];
1124 ret = iommu_flush_dte(iommu, dev_data->devid);
1128 if (dev_data->ats.enabled)
1129 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1135 * TLB invalidation function which is called from the mapping functions.
1136 * It invalidates a single PTE if the range to flush is within a single
1137 * page. Otherwise it flushes the whole TLB of the IOMMU.
1139 static void __domain_flush_pages(struct protection_domain *domain,
1140 u64 address, size_t size, int pde)
1142 struct iommu_dev_data *dev_data;
1143 struct iommu_cmd cmd;
1146 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1148 for (i = 0; i < amd_iommus_present; ++i) {
1149 if (!domain->dev_iommu[i])
1153 * Devices of this domain are behind this IOMMU
1154 * We need a TLB flush
1156 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1159 list_for_each_entry(dev_data, &domain->dev_list, list) {
1161 if (!dev_data->ats.enabled)
1164 ret |= device_flush_iotlb(dev_data, address, size);
1170 static void domain_flush_pages(struct protection_domain *domain,
1171 u64 address, size_t size)
1173 __domain_flush_pages(domain, address, size, 0);
1176 /* Flush the whole IO/TLB for a given protection domain */
1177 static void domain_flush_tlb(struct protection_domain *domain)
1179 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1182 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1183 static void domain_flush_tlb_pde(struct protection_domain *domain)
1185 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1188 static void domain_flush_complete(struct protection_domain *domain)
1192 for (i = 0; i < amd_iommus_present; ++i) {
1193 if (!domain->dev_iommu[i])
1197 * Devices of this domain are behind this IOMMU
1198 * We need to wait for completion of all commands.
1200 iommu_completion_wait(amd_iommus[i]);
1206 * This function flushes the DTEs for all devices in domain
1208 static void domain_flush_devices(struct protection_domain *domain)
1210 struct iommu_dev_data *dev_data;
1212 list_for_each_entry(dev_data, &domain->dev_list, list)
1213 device_flush_dte(dev_data);
1216 /****************************************************************************
1218 * The functions below are used the create the page table mappings for
1219 * unity mapped regions.
1221 ****************************************************************************/
1224 * This function is used to add another level to an IO page table. Adding
1225 * another level increases the size of the address space by 9 bits to a size up
1228 static bool increase_address_space(struct protection_domain *domain,
1233 if (domain->mode == PAGE_MODE_6_LEVEL)
1234 /* address space already 64 bit large */
1237 pte = (void *)get_zeroed_page(gfp);
1241 *pte = PM_LEVEL_PDE(domain->mode,
1242 virt_to_phys(domain->pt_root));
1243 domain->pt_root = pte;
1245 domain->updated = true;
1250 static u64 *alloc_pte(struct protection_domain *domain,
1251 unsigned long address,
1252 unsigned long page_size,
1259 BUG_ON(!is_power_of_2(page_size));
1261 while (address > PM_LEVEL_SIZE(domain->mode))
1262 increase_address_space(domain, gfp);
1264 level = domain->mode - 1;
1265 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1266 address = PAGE_SIZE_ALIGN(address, page_size);
1267 end_lvl = PAGE_SIZE_LEVEL(page_size);
1269 while (level > end_lvl) {
1270 if (!IOMMU_PTE_PRESENT(*pte)) {
1271 page = (u64 *)get_zeroed_page(gfp);
1274 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1277 /* No level skipping support yet */
1278 if (PM_PTE_LEVEL(*pte) != level)
1283 pte = IOMMU_PTE_PAGE(*pte);
1285 if (pte_page && level == end_lvl)
1288 pte = &pte[PM_LEVEL_INDEX(level, address)];
1295 * This function checks if there is a PTE for a given dma address. If
1296 * there is one, it returns the pointer to it.
1298 static u64 *fetch_pte(struct protection_domain *domain,
1299 unsigned long address,
1300 unsigned long *page_size)
1305 if (address > PM_LEVEL_SIZE(domain->mode))
1308 level = domain->mode - 1;
1309 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1310 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1315 if (!IOMMU_PTE_PRESENT(*pte))
1319 if (PM_PTE_LEVEL(*pte) == 7 ||
1320 PM_PTE_LEVEL(*pte) == 0)
1323 /* No level skipping support yet */
1324 if (PM_PTE_LEVEL(*pte) != level)
1329 /* Walk to the next level */
1330 pte = IOMMU_PTE_PAGE(*pte);
1331 pte = &pte[PM_LEVEL_INDEX(level, address)];
1332 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1335 if (PM_PTE_LEVEL(*pte) == 0x07) {
1336 unsigned long pte_mask;
1339 * If we have a series of large PTEs, make
1340 * sure to return a pointer to the first one.
1342 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1343 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1344 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1351 * Generic mapping functions. It maps a physical address into a DMA
1352 * address space. It allocates the page table pages if necessary.
1353 * In the future it can be extended to a generic mapping function
1354 * supporting all features of AMD IOMMU page tables like level skipping
1355 * and full 64 bit address spaces.
1357 static int iommu_map_page(struct protection_domain *dom,
1358 unsigned long bus_addr,
1359 unsigned long phys_addr,
1361 unsigned long page_size)
1366 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1367 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1369 if (!(prot & IOMMU_PROT_MASK))
1372 count = PAGE_SIZE_PTE_COUNT(page_size);
1373 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1378 for (i = 0; i < count; ++i)
1379 if (IOMMU_PTE_PRESENT(pte[i]))
1383 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1384 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1386 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1388 if (prot & IOMMU_PROT_IR)
1389 __pte |= IOMMU_PTE_IR;
1390 if (prot & IOMMU_PROT_IW)
1391 __pte |= IOMMU_PTE_IW;
1393 for (i = 0; i < count; ++i)
1401 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1402 unsigned long bus_addr,
1403 unsigned long page_size)
1405 unsigned long long unmapped;
1406 unsigned long unmap_size;
1409 BUG_ON(!is_power_of_2(page_size));
1413 while (unmapped < page_size) {
1415 pte = fetch_pte(dom, bus_addr, &unmap_size);
1420 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1421 for (i = 0; i < count; i++)
1425 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1426 unmapped += unmap_size;
1429 BUG_ON(unmapped && !is_power_of_2(unmapped));
1434 /****************************************************************************
1436 * The next functions belong to the address allocator for the dma_ops
1437 * interface functions. They work like the allocators in the other IOMMU
1438 * drivers. Its basically a bitmap which marks the allocated pages in
1439 * the aperture. Maybe it could be enhanced in the future to a more
1440 * efficient allocator.
1442 ****************************************************************************/
1445 * The address allocator core functions.
1447 * called with domain->lock held
1451 * Used to reserve address ranges in the aperture (e.g. for exclusion
1454 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1455 unsigned long start_page,
1458 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1460 if (start_page + pages > last_page)
1461 pages = last_page - start_page;
1463 for (i = start_page; i < start_page + pages; ++i) {
1464 int index = i / APERTURE_RANGE_PAGES;
1465 int page = i % APERTURE_RANGE_PAGES;
1466 __set_bit(page, dom->aperture[index]->bitmap);
1471 * This function is used to add a new aperture range to an existing
1472 * aperture in case of dma_ops domain allocation or address allocation
1475 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1476 bool populate, gfp_t gfp)
1478 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1479 struct amd_iommu *iommu;
1480 unsigned long i, old_size, pte_pgsize;
1482 #ifdef CONFIG_IOMMU_STRESS
1486 if (index >= APERTURE_MAX_RANGES)
1489 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1490 if (!dma_dom->aperture[index])
1493 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1494 if (!dma_dom->aperture[index]->bitmap)
1497 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1500 unsigned long address = dma_dom->aperture_size;
1501 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1502 u64 *pte, *pte_page;
1504 for (i = 0; i < num_ptes; ++i) {
1505 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1510 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1512 address += APERTURE_RANGE_SIZE / 64;
1516 old_size = dma_dom->aperture_size;
1517 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1519 /* Reserve address range used for MSI messages */
1520 if (old_size < MSI_ADDR_BASE_LO &&
1521 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1522 unsigned long spage;
1525 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1526 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1528 dma_ops_reserve_addresses(dma_dom, spage, pages);
1531 /* Initialize the exclusion range if necessary */
1532 for_each_iommu(iommu) {
1533 if (iommu->exclusion_start &&
1534 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1535 && iommu->exclusion_start < dma_dom->aperture_size) {
1536 unsigned long startpage;
1537 int pages = iommu_num_pages(iommu->exclusion_start,
1538 iommu->exclusion_length,
1540 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1541 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1546 * Check for areas already mapped as present in the new aperture
1547 * range and mark those pages as reserved in the allocator. Such
1548 * mappings may already exist as a result of requested unity
1549 * mappings for devices.
1551 for (i = dma_dom->aperture[index]->offset;
1552 i < dma_dom->aperture_size;
1554 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1555 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1558 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1562 update_domain(&dma_dom->domain);
1567 update_domain(&dma_dom->domain);
1569 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1571 kfree(dma_dom->aperture[index]);
1572 dma_dom->aperture[index] = NULL;
1577 static unsigned long dma_ops_area_alloc(struct device *dev,
1578 struct dma_ops_domain *dom,
1580 unsigned long align_mask,
1582 unsigned long start)
1584 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1585 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1586 int i = start >> APERTURE_RANGE_SHIFT;
1587 unsigned long boundary_size, mask;
1588 unsigned long address = -1;
1589 unsigned long limit;
1591 next_bit >>= PAGE_SHIFT;
1593 mask = dma_get_seg_boundary(dev);
1595 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1596 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1598 for (;i < max_index; ++i) {
1599 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1601 if (dom->aperture[i]->offset >= dma_mask)
1604 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1605 dma_mask >> PAGE_SHIFT);
1607 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1608 limit, next_bit, pages, 0,
1609 boundary_size, align_mask);
1610 if (address != -1) {
1611 address = dom->aperture[i]->offset +
1612 (address << PAGE_SHIFT);
1613 dom->next_address = address + (pages << PAGE_SHIFT);
1623 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1624 struct dma_ops_domain *dom,
1626 unsigned long align_mask,
1629 unsigned long address;
1631 #ifdef CONFIG_IOMMU_STRESS
1632 dom->next_address = 0;
1633 dom->need_flush = true;
1636 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1637 dma_mask, dom->next_address);
1639 if (address == -1) {
1640 dom->next_address = 0;
1641 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1643 dom->need_flush = true;
1646 if (unlikely(address == -1))
1647 address = DMA_ERROR_CODE;
1649 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1655 * The address free function.
1657 * called with domain->lock held
1659 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1660 unsigned long address,
1663 unsigned i = address >> APERTURE_RANGE_SHIFT;
1664 struct aperture_range *range = dom->aperture[i];
1666 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1668 #ifdef CONFIG_IOMMU_STRESS
1673 if (address >= dom->next_address)
1674 dom->need_flush = true;
1676 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1678 bitmap_clear(range->bitmap, address, pages);
1682 /****************************************************************************
1684 * The next functions belong to the domain allocation. A domain is
1685 * allocated for every IOMMU as the default domain. If device isolation
1686 * is enabled, every device get its own domain. The most important thing
1687 * about domains is the page table mapping the DMA address space they
1690 ****************************************************************************/
1693 * This function adds a protection domain to the global protection domain list
1695 static void add_domain_to_list(struct protection_domain *domain)
1697 unsigned long flags;
1699 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1700 list_add(&domain->list, &amd_iommu_pd_list);
1701 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1705 * This function removes a protection domain to the global
1706 * protection domain list
1708 static void del_domain_from_list(struct protection_domain *domain)
1710 unsigned long flags;
1712 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1713 list_del(&domain->list);
1714 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1717 static u16 domain_id_alloc(void)
1719 unsigned long flags;
1722 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1723 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1725 if (id > 0 && id < MAX_DOMAIN_ID)
1726 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1729 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1734 static void domain_id_free(int id)
1736 unsigned long flags;
1738 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1739 if (id > 0 && id < MAX_DOMAIN_ID)
1740 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1741 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1744 #define DEFINE_FREE_PT_FN(LVL, FN) \
1745 static void free_pt_##LVL (unsigned long __pt) \
1753 for (i = 0; i < 512; ++i) { \
1754 /* PTE present? */ \
1755 if (!IOMMU_PTE_PRESENT(pt[i])) \
1759 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1760 PM_PTE_LEVEL(pt[i]) == 7) \
1763 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1766 free_page((unsigned long)pt); \
1769 DEFINE_FREE_PT_FN(l2, free_page)
1770 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1771 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1772 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1773 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1775 static void free_pagetable(struct protection_domain *domain)
1777 unsigned long root = (unsigned long)domain->pt_root;
1779 switch (domain->mode) {
1780 case PAGE_MODE_NONE:
1782 case PAGE_MODE_1_LEVEL:
1785 case PAGE_MODE_2_LEVEL:
1788 case PAGE_MODE_3_LEVEL:
1791 case PAGE_MODE_4_LEVEL:
1794 case PAGE_MODE_5_LEVEL:
1797 case PAGE_MODE_6_LEVEL:
1805 static void free_gcr3_tbl_level1(u64 *tbl)
1810 for (i = 0; i < 512; ++i) {
1811 if (!(tbl[i] & GCR3_VALID))
1814 ptr = __va(tbl[i] & PAGE_MASK);
1816 free_page((unsigned long)ptr);
1820 static void free_gcr3_tbl_level2(u64 *tbl)
1825 for (i = 0; i < 512; ++i) {
1826 if (!(tbl[i] & GCR3_VALID))
1829 ptr = __va(tbl[i] & PAGE_MASK);
1831 free_gcr3_tbl_level1(ptr);
1835 static void free_gcr3_table(struct protection_domain *domain)
1837 if (domain->glx == 2)
1838 free_gcr3_tbl_level2(domain->gcr3_tbl);
1839 else if (domain->glx == 1)
1840 free_gcr3_tbl_level1(domain->gcr3_tbl);
1841 else if (domain->glx != 0)
1844 free_page((unsigned long)domain->gcr3_tbl);
1848 * Free a domain, only used if something went wrong in the
1849 * allocation path and we need to free an already allocated page table
1851 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1858 del_domain_from_list(&dom->domain);
1860 free_pagetable(&dom->domain);
1862 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1863 if (!dom->aperture[i])
1865 free_page((unsigned long)dom->aperture[i]->bitmap);
1866 kfree(dom->aperture[i]);
1873 * Allocates a new protection domain usable for the dma_ops functions.
1874 * It also initializes the page table and the address allocator data
1875 * structures required for the dma_ops interface
1877 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1879 struct dma_ops_domain *dma_dom;
1881 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1885 if (protection_domain_init(&dma_dom->domain))
1888 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1889 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1890 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1891 dma_dom->domain.priv = dma_dom;
1892 if (!dma_dom->domain.pt_root)
1895 dma_dom->need_flush = false;
1897 add_domain_to_list(&dma_dom->domain);
1899 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1903 * mark the first page as allocated so we never return 0 as
1904 * a valid dma-address. So we can use 0 as error value
1906 dma_dom->aperture[0]->bitmap[0] = 1;
1907 dma_dom->next_address = 0;
1913 dma_ops_domain_free(dma_dom);
1919 * little helper function to check whether a given protection domain is a
1922 static bool dma_ops_domain(struct protection_domain *domain)
1924 return domain->flags & PD_DMA_OPS_MASK;
1927 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1932 if (domain->mode != PAGE_MODE_NONE)
1933 pte_root = virt_to_phys(domain->pt_root);
1935 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1936 << DEV_ENTRY_MODE_SHIFT;
1937 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1939 flags = amd_iommu_dev_table[devid].data[1];
1942 flags |= DTE_FLAG_IOTLB;
1944 if (domain->flags & PD_IOMMUV2_MASK) {
1945 u64 gcr3 = __pa(domain->gcr3_tbl);
1946 u64 glx = domain->glx;
1949 pte_root |= DTE_FLAG_GV;
1950 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1952 /* First mask out possible old values for GCR3 table */
1953 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1956 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1959 /* Encode GCR3 table into DTE */
1960 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1963 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1966 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1970 flags &= ~(0xffffUL);
1971 flags |= domain->id;
1973 amd_iommu_dev_table[devid].data[1] = flags;
1974 amd_iommu_dev_table[devid].data[0] = pte_root;
1977 static void clear_dte_entry(u16 devid)
1979 /* remove entry from the device table seen by the hardware */
1980 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1981 amd_iommu_dev_table[devid].data[1] = 0;
1983 amd_iommu_apply_erratum_63(devid);
1986 static void do_attach(struct iommu_dev_data *dev_data,
1987 struct protection_domain *domain)
1989 struct amd_iommu *iommu;
1992 iommu = amd_iommu_rlookup_table[dev_data->devid];
1993 ats = dev_data->ats.enabled;
1995 /* Update data structures */
1996 dev_data->domain = domain;
1997 list_add(&dev_data->list, &domain->dev_list);
1998 set_dte_entry(dev_data->devid, domain, ats);
2000 /* Do reference counting */
2001 domain->dev_iommu[iommu->index] += 1;
2002 domain->dev_cnt += 1;
2004 /* Flush the DTE entry */
2005 device_flush_dte(dev_data);
2008 static void do_detach(struct iommu_dev_data *dev_data)
2010 struct amd_iommu *iommu;
2012 iommu = amd_iommu_rlookup_table[dev_data->devid];
2014 /* decrease reference counters */
2015 dev_data->domain->dev_iommu[iommu->index] -= 1;
2016 dev_data->domain->dev_cnt -= 1;
2018 /* Update data structures */
2019 dev_data->domain = NULL;
2020 list_del(&dev_data->list);
2021 clear_dte_entry(dev_data->devid);
2023 /* Flush the DTE entry */
2024 device_flush_dte(dev_data);
2028 * If a device is not yet associated with a domain, this function does
2029 * assigns it visible for the hardware
2031 static int __attach_device(struct iommu_dev_data *dev_data,
2032 struct protection_domain *domain)
2034 struct iommu_dev_data *head, *entry;
2038 spin_lock(&domain->lock);
2042 if (head->alias_data != NULL)
2043 head = head->alias_data;
2045 /* Now we have the root of the alias group, if any */
2048 if (head->domain != NULL)
2051 /* Attach alias group root */
2052 do_attach(head, domain);
2054 /* Attach other devices in the alias group */
2055 list_for_each_entry(entry, &head->alias_list, alias_list)
2056 do_attach(entry, domain);
2063 spin_unlock(&domain->lock);
2069 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2071 pci_disable_ats(pdev);
2072 pci_disable_pri(pdev);
2073 pci_disable_pasid(pdev);
2076 /* FIXME: Change generic reset-function to do the same */
2077 static int pri_reset_while_enabled(struct pci_dev *pdev)
2082 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2086 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2087 control |= PCI_PRI_CTRL_RESET;
2088 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2093 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2098 /* FIXME: Hardcode number of outstanding requests for now */
2100 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2102 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2104 /* Only allow access to user-accessible pages */
2105 ret = pci_enable_pasid(pdev, 0);
2109 /* First reset the PRI state of the device */
2110 ret = pci_reset_pri(pdev);
2115 ret = pci_enable_pri(pdev, reqs);
2120 ret = pri_reset_while_enabled(pdev);
2125 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2132 pci_disable_pri(pdev);
2133 pci_disable_pasid(pdev);
2138 /* FIXME: Move this to PCI code */
2139 #define PCI_PRI_TLP_OFF (1 << 15)
2141 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2146 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2150 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2152 return (status & PCI_PRI_TLP_OFF) ? true : false;
2156 * If a device is not yet associated with a domain, this function
2157 * assigns it visible for the hardware
2159 static int attach_device(struct device *dev,
2160 struct protection_domain *domain)
2162 struct pci_dev *pdev = to_pci_dev(dev);
2163 struct iommu_dev_data *dev_data;
2164 unsigned long flags;
2167 dev_data = get_dev_data(dev);
2169 if (domain->flags & PD_IOMMUV2_MASK) {
2170 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2173 if (pdev_iommuv2_enable(pdev) != 0)
2176 dev_data->ats.enabled = true;
2177 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2178 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2179 } else if (amd_iommu_iotlb_sup &&
2180 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2181 dev_data->ats.enabled = true;
2182 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2185 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2186 ret = __attach_device(dev_data, domain);
2187 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2190 * We might boot into a crash-kernel here. The crashed kernel
2191 * left the caches in the IOMMU dirty. So we have to flush
2192 * here to evict all dirty stuff.
2194 domain_flush_tlb_pde(domain);
2200 * Removes a device from a protection domain (unlocked)
2202 static void __detach_device(struct iommu_dev_data *dev_data)
2204 struct iommu_dev_data *head, *entry;
2205 struct protection_domain *domain;
2206 unsigned long flags;
2208 BUG_ON(!dev_data->domain);
2210 domain = dev_data->domain;
2212 spin_lock_irqsave(&domain->lock, flags);
2215 if (head->alias_data != NULL)
2216 head = head->alias_data;
2218 list_for_each_entry(entry, &head->alias_list, alias_list)
2223 spin_unlock_irqrestore(&domain->lock, flags);
2226 * If we run in passthrough mode the device must be assigned to the
2227 * passthrough domain if it is detached from any other domain.
2228 * Make sure we can deassign from the pt_domain itself.
2230 if (dev_data->passthrough &&
2231 (dev_data->domain == NULL && domain != pt_domain))
2232 __attach_device(dev_data, pt_domain);
2236 * Removes a device from a protection domain (with devtable_lock held)
2238 static void detach_device(struct device *dev)
2240 struct protection_domain *domain;
2241 struct iommu_dev_data *dev_data;
2242 unsigned long flags;
2244 dev_data = get_dev_data(dev);
2245 domain = dev_data->domain;
2247 /* lock device table */
2248 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2249 __detach_device(dev_data);
2250 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2252 if (domain->flags & PD_IOMMUV2_MASK)
2253 pdev_iommuv2_disable(to_pci_dev(dev));
2254 else if (dev_data->ats.enabled)
2255 pci_disable_ats(to_pci_dev(dev));
2257 dev_data->ats.enabled = false;
2260 static int amd_iommu_add_device(struct device *dev)
2262 struct iommu_dev_data *dev_data;
2263 struct iommu_domain *domain;
2264 struct amd_iommu *iommu;
2268 if (!check_device(dev) || get_dev_data(dev))
2271 devid = get_device_id(dev);
2272 iommu = amd_iommu_rlookup_table[devid];
2274 ret = iommu_init_device(dev);
2276 if (ret != -ENOTSUPP)
2277 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2280 iommu_ignore_device(dev);
2281 dev->archdata.dma_ops = &nommu_dma_ops;
2284 init_iommu_group(dev);
2286 dev_data = get_dev_data(dev);
2290 if (dev_data->iommu_v2)
2291 iommu_request_dm_for_dev(dev);
2293 /* Domains are initialized for this device - have a look what we ended up with */
2294 domain = iommu_get_domain_for_dev(dev);
2295 if (domain->type == IOMMU_DOMAIN_IDENTITY) {
2296 dev_data->passthrough = true;
2297 dev->archdata.dma_ops = &nommu_dma_ops;
2299 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2303 iommu_completion_wait(iommu);
2308 static void amd_iommu_remove_device(struct device *dev)
2310 struct amd_iommu *iommu;
2313 if (!check_device(dev))
2316 devid = get_device_id(dev);
2317 iommu = amd_iommu_rlookup_table[devid];
2319 iommu_uninit_device(dev);
2320 iommu_completion_wait(iommu);
2323 /*****************************************************************************
2325 * The next functions belong to the dma_ops mapping/unmapping code.
2327 *****************************************************************************/
2330 * In the dma_ops path we only have the struct device. This function
2331 * finds the corresponding IOMMU, the protection domain and the
2332 * requestor id for a given device.
2333 * If the device is not yet associated with a domain this is also done
2336 static struct protection_domain *get_domain(struct device *dev)
2338 struct protection_domain *domain;
2339 struct iommu_domain *io_domain;
2341 if (!check_device(dev))
2342 return ERR_PTR(-EINVAL);
2344 io_domain = iommu_get_domain_for_dev(dev);
2348 domain = to_pdomain(io_domain);
2349 if (!dma_ops_domain(domain))
2350 return ERR_PTR(-EBUSY);
2355 static void update_device_table(struct protection_domain *domain)
2357 struct iommu_dev_data *dev_data;
2359 list_for_each_entry(dev_data, &domain->dev_list, list)
2360 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2363 static void update_domain(struct protection_domain *domain)
2365 if (!domain->updated)
2368 update_device_table(domain);
2370 domain_flush_devices(domain);
2371 domain_flush_tlb_pde(domain);
2373 domain->updated = false;
2377 * This function fetches the PTE for a given address in the aperture
2379 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2380 unsigned long address)
2382 struct aperture_range *aperture;
2383 u64 *pte, *pte_page;
2385 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2389 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2391 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2393 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2395 pte += PM_LEVEL_INDEX(0, address);
2397 update_domain(&dom->domain);
2403 * This is the generic map function. It maps one 4kb page at paddr to
2404 * the given address in the DMA address space for the domain.
2406 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2407 unsigned long address,
2413 WARN_ON(address > dom->aperture_size);
2417 pte = dma_ops_get_pte(dom, address);
2419 return DMA_ERROR_CODE;
2421 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2423 if (direction == DMA_TO_DEVICE)
2424 __pte |= IOMMU_PTE_IR;
2425 else if (direction == DMA_FROM_DEVICE)
2426 __pte |= IOMMU_PTE_IW;
2427 else if (direction == DMA_BIDIRECTIONAL)
2428 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2434 return (dma_addr_t)address;
2438 * The generic unmapping function for on page in the DMA address space.
2440 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2441 unsigned long address)
2443 struct aperture_range *aperture;
2446 if (address >= dom->aperture_size)
2449 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2453 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2457 pte += PM_LEVEL_INDEX(0, address);
2465 * This function contains common code for mapping of a physically
2466 * contiguous memory region into DMA address space. It is used by all
2467 * mapping functions provided with this IOMMU driver.
2468 * Must be called with the domain lock held.
2470 static dma_addr_t __map_single(struct device *dev,
2471 struct dma_ops_domain *dma_dom,
2478 dma_addr_t offset = paddr & ~PAGE_MASK;
2479 dma_addr_t address, start, ret;
2481 unsigned long align_mask = 0;
2484 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2487 INC_STATS_COUNTER(total_map_requests);
2490 INC_STATS_COUNTER(cross_page);
2493 align_mask = (1UL << get_order(size)) - 1;
2496 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2498 if (unlikely(address == DMA_ERROR_CODE)) {
2500 * setting next_address here will let the address
2501 * allocator only scan the new allocated range in the
2502 * first run. This is a small optimization.
2504 dma_dom->next_address = dma_dom->aperture_size;
2506 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2510 * aperture was successfully enlarged by 128 MB, try
2517 for (i = 0; i < pages; ++i) {
2518 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2519 if (ret == DMA_ERROR_CODE)
2527 ADD_STATS_COUNTER(alloced_io_mem, size);
2529 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2530 domain_flush_tlb(&dma_dom->domain);
2531 dma_dom->need_flush = false;
2532 } else if (unlikely(amd_iommu_np_cache))
2533 domain_flush_pages(&dma_dom->domain, address, size);
2540 for (--i; i >= 0; --i) {
2542 dma_ops_domain_unmap(dma_dom, start);
2545 dma_ops_free_addresses(dma_dom, address, pages);
2547 return DMA_ERROR_CODE;
2551 * Does the reverse of the __map_single function. Must be called with
2552 * the domain lock held too
2554 static void __unmap_single(struct dma_ops_domain *dma_dom,
2555 dma_addr_t dma_addr,
2559 dma_addr_t flush_addr;
2560 dma_addr_t i, start;
2563 if ((dma_addr == DMA_ERROR_CODE) ||
2564 (dma_addr + size > dma_dom->aperture_size))
2567 flush_addr = dma_addr;
2568 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2569 dma_addr &= PAGE_MASK;
2572 for (i = 0; i < pages; ++i) {
2573 dma_ops_domain_unmap(dma_dom, start);
2577 SUB_STATS_COUNTER(alloced_io_mem, size);
2579 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2581 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2582 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2583 dma_dom->need_flush = false;
2588 * The exported map_single function for dma_ops.
2590 static dma_addr_t map_page(struct device *dev, struct page *page,
2591 unsigned long offset, size_t size,
2592 enum dma_data_direction dir,
2593 struct dma_attrs *attrs)
2595 unsigned long flags;
2596 struct protection_domain *domain;
2599 phys_addr_t paddr = page_to_phys(page) + offset;
2601 INC_STATS_COUNTER(cnt_map_single);
2603 domain = get_domain(dev);
2604 if (PTR_ERR(domain) == -EINVAL)
2605 return (dma_addr_t)paddr;
2606 else if (IS_ERR(domain))
2607 return DMA_ERROR_CODE;
2609 dma_mask = *dev->dma_mask;
2611 spin_lock_irqsave(&domain->lock, flags);
2613 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2615 if (addr == DMA_ERROR_CODE)
2618 domain_flush_complete(domain);
2621 spin_unlock_irqrestore(&domain->lock, flags);
2627 * The exported unmap_single function for dma_ops.
2629 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2630 enum dma_data_direction dir, struct dma_attrs *attrs)
2632 unsigned long flags;
2633 struct protection_domain *domain;
2635 INC_STATS_COUNTER(cnt_unmap_single);
2637 domain = get_domain(dev);
2641 spin_lock_irqsave(&domain->lock, flags);
2643 __unmap_single(domain->priv, dma_addr, size, dir);
2645 domain_flush_complete(domain);
2647 spin_unlock_irqrestore(&domain->lock, flags);
2651 * The exported map_sg function for dma_ops (handles scatter-gather
2654 static int map_sg(struct device *dev, struct scatterlist *sglist,
2655 int nelems, enum dma_data_direction dir,
2656 struct dma_attrs *attrs)
2658 unsigned long flags;
2659 struct protection_domain *domain;
2661 struct scatterlist *s;
2663 int mapped_elems = 0;
2666 INC_STATS_COUNTER(cnt_map_sg);
2668 domain = get_domain(dev);
2672 dma_mask = *dev->dma_mask;
2674 spin_lock_irqsave(&domain->lock, flags);
2676 for_each_sg(sglist, s, nelems, i) {
2679 s->dma_address = __map_single(dev, domain->priv,
2680 paddr, s->length, dir, false,
2683 if (s->dma_address) {
2684 s->dma_length = s->length;
2690 domain_flush_complete(domain);
2693 spin_unlock_irqrestore(&domain->lock, flags);
2695 return mapped_elems;
2697 for_each_sg(sglist, s, mapped_elems, i) {
2699 __unmap_single(domain->priv, s->dma_address,
2700 s->dma_length, dir);
2701 s->dma_address = s->dma_length = 0;
2710 * The exported map_sg function for dma_ops (handles scatter-gather
2713 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2714 int nelems, enum dma_data_direction dir,
2715 struct dma_attrs *attrs)
2717 unsigned long flags;
2718 struct protection_domain *domain;
2719 struct scatterlist *s;
2722 INC_STATS_COUNTER(cnt_unmap_sg);
2724 domain = get_domain(dev);
2728 spin_lock_irqsave(&domain->lock, flags);
2730 for_each_sg(sglist, s, nelems, i) {
2731 __unmap_single(domain->priv, s->dma_address,
2732 s->dma_length, dir);
2733 s->dma_address = s->dma_length = 0;
2736 domain_flush_complete(domain);
2738 spin_unlock_irqrestore(&domain->lock, flags);
2742 * The exported alloc_coherent function for dma_ops.
2744 static void *alloc_coherent(struct device *dev, size_t size,
2745 dma_addr_t *dma_addr, gfp_t flag,
2746 struct dma_attrs *attrs)
2748 u64 dma_mask = dev->coherent_dma_mask;
2749 struct protection_domain *domain;
2750 unsigned long flags;
2753 INC_STATS_COUNTER(cnt_alloc_coherent);
2755 domain = get_domain(dev);
2756 if (PTR_ERR(domain) == -EINVAL) {
2757 page = alloc_pages(flag, get_order(size));
2758 *dma_addr = page_to_phys(page);
2759 return page_address(page);
2760 } else if (IS_ERR(domain))
2763 size = PAGE_ALIGN(size);
2764 dma_mask = dev->coherent_dma_mask;
2765 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2768 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2770 if (!(flag & __GFP_WAIT))
2773 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2780 dma_mask = *dev->dma_mask;
2782 spin_lock_irqsave(&domain->lock, flags);
2784 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2785 size, DMA_BIDIRECTIONAL, true, dma_mask);
2787 if (*dma_addr == DMA_ERROR_CODE) {
2788 spin_unlock_irqrestore(&domain->lock, flags);
2792 domain_flush_complete(domain);
2794 spin_unlock_irqrestore(&domain->lock, flags);
2796 return page_address(page);
2800 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2801 __free_pages(page, get_order(size));
2807 * The exported free_coherent function for dma_ops.
2809 static void free_coherent(struct device *dev, size_t size,
2810 void *virt_addr, dma_addr_t dma_addr,
2811 struct dma_attrs *attrs)
2813 struct protection_domain *domain;
2814 unsigned long flags;
2817 INC_STATS_COUNTER(cnt_free_coherent);
2819 page = virt_to_page(virt_addr);
2820 size = PAGE_ALIGN(size);
2822 domain = get_domain(dev);
2826 spin_lock_irqsave(&domain->lock, flags);
2828 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2830 domain_flush_complete(domain);
2832 spin_unlock_irqrestore(&domain->lock, flags);
2835 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2836 __free_pages(page, get_order(size));
2840 * This function is called by the DMA layer to find out if we can handle a
2841 * particular device. It is part of the dma_ops.
2843 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2845 return check_device(dev);
2848 static struct dma_map_ops amd_iommu_dma_ops = {
2849 .alloc = alloc_coherent,
2850 .free = free_coherent,
2851 .map_page = map_page,
2852 .unmap_page = unmap_page,
2854 .unmap_sg = unmap_sg,
2855 .dma_supported = amd_iommu_dma_supported,
2858 int __init amd_iommu_init_api(void)
2860 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2863 int __init amd_iommu_init_dma_ops(void)
2868 amd_iommu_stats_init();
2870 if (amd_iommu_unmap_flush)
2871 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2873 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2878 /*****************************************************************************
2880 * The following functions belong to the exported interface of AMD IOMMU
2882 * This interface allows access to lower level functions of the IOMMU
2883 * like protection domain handling and assignement of devices to domains
2884 * which is not possible with the dma_ops interface.
2886 *****************************************************************************/
2888 static void cleanup_domain(struct protection_domain *domain)
2890 struct iommu_dev_data *entry;
2891 unsigned long flags;
2893 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2895 while (!list_empty(&domain->dev_list)) {
2896 entry = list_first_entry(&domain->dev_list,
2897 struct iommu_dev_data, list);
2898 __detach_device(entry);
2901 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2904 static void protection_domain_free(struct protection_domain *domain)
2909 del_domain_from_list(domain);
2912 domain_id_free(domain->id);
2917 static int protection_domain_init(struct protection_domain *domain)
2919 spin_lock_init(&domain->lock);
2920 mutex_init(&domain->api_lock);
2921 domain->id = domain_id_alloc();
2924 INIT_LIST_HEAD(&domain->dev_list);
2929 static struct protection_domain *protection_domain_alloc(void)
2931 struct protection_domain *domain;
2933 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2937 if (protection_domain_init(domain))
2940 add_domain_to_list(domain);
2950 static int alloc_passthrough_domain(void)
2952 if (pt_domain != NULL)
2955 /* allocate passthrough domain */
2956 pt_domain = protection_domain_alloc();
2960 pt_domain->mode = PAGE_MODE_NONE;
2965 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2967 struct protection_domain *pdomain;
2968 struct dma_ops_domain *dma_domain;
2971 case IOMMU_DOMAIN_UNMANAGED:
2972 pdomain = protection_domain_alloc();
2976 pdomain->mode = PAGE_MODE_3_LEVEL;
2977 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2978 if (!pdomain->pt_root) {
2979 protection_domain_free(pdomain);
2983 pdomain->domain.geometry.aperture_start = 0;
2984 pdomain->domain.geometry.aperture_end = ~0ULL;
2985 pdomain->domain.geometry.force_aperture = true;
2988 case IOMMU_DOMAIN_DMA:
2989 dma_domain = dma_ops_domain_alloc();
2991 pr_err("AMD-Vi: Failed to allocate\n");
2994 pdomain = &dma_domain->domain;
2996 case IOMMU_DOMAIN_IDENTITY:
2997 pdomain = protection_domain_alloc();
3001 pdomain->mode = PAGE_MODE_NONE;
3007 return &pdomain->domain;
3010 static void amd_iommu_domain_free(struct iommu_domain *dom)
3012 struct protection_domain *domain;
3017 domain = to_pdomain(dom);
3019 if (domain->dev_cnt > 0)
3020 cleanup_domain(domain);
3022 BUG_ON(domain->dev_cnt != 0);
3024 if (domain->mode != PAGE_MODE_NONE)
3025 free_pagetable(domain);
3027 if (domain->flags & PD_IOMMUV2_MASK)
3028 free_gcr3_table(domain);
3030 protection_domain_free(domain);
3033 static void amd_iommu_detach_device(struct iommu_domain *dom,
3036 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3037 struct amd_iommu *iommu;
3040 if (!check_device(dev))
3043 devid = get_device_id(dev);
3045 if (dev_data->domain != NULL)
3048 iommu = amd_iommu_rlookup_table[devid];
3052 iommu_completion_wait(iommu);
3055 static int amd_iommu_attach_device(struct iommu_domain *dom,
3058 struct protection_domain *domain = to_pdomain(dom);
3059 struct iommu_dev_data *dev_data;
3060 struct amd_iommu *iommu;
3063 if (!check_device(dev))
3066 dev_data = dev->archdata.iommu;
3068 iommu = amd_iommu_rlookup_table[dev_data->devid];
3072 if (dev_data->domain)
3075 ret = attach_device(dev, domain);
3077 iommu_completion_wait(iommu);
3082 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3083 phys_addr_t paddr, size_t page_size, int iommu_prot)
3085 struct protection_domain *domain = to_pdomain(dom);
3089 if (domain->mode == PAGE_MODE_NONE)
3092 if (iommu_prot & IOMMU_READ)
3093 prot |= IOMMU_PROT_IR;
3094 if (iommu_prot & IOMMU_WRITE)
3095 prot |= IOMMU_PROT_IW;
3097 mutex_lock(&domain->api_lock);
3098 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3099 mutex_unlock(&domain->api_lock);
3104 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3107 struct protection_domain *domain = to_pdomain(dom);
3110 if (domain->mode == PAGE_MODE_NONE)
3113 mutex_lock(&domain->api_lock);
3114 unmap_size = iommu_unmap_page(domain, iova, page_size);
3115 mutex_unlock(&domain->api_lock);
3117 domain_flush_tlb_pde(domain);
3122 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3125 struct protection_domain *domain = to_pdomain(dom);
3126 unsigned long offset_mask, pte_pgsize;
3129 if (domain->mode == PAGE_MODE_NONE)
3132 pte = fetch_pte(domain, iova, &pte_pgsize);
3134 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3137 offset_mask = pte_pgsize - 1;
3138 __pte = *pte & PM_ADDR_MASK;
3140 return (__pte & ~offset_mask) | (iova & offset_mask);
3143 static bool amd_iommu_capable(enum iommu_cap cap)
3146 case IOMMU_CAP_CACHE_COHERENCY:
3148 case IOMMU_CAP_INTR_REMAP:
3149 return (irq_remapping_enabled == 1);
3150 case IOMMU_CAP_NOEXEC:
3157 static void amd_iommu_get_dm_regions(struct device *dev,
3158 struct list_head *head)
3160 struct unity_map_entry *entry;
3163 devid = get_device_id(dev);
3165 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3166 struct iommu_dm_region *region;
3168 if (devid < entry->devid_start || devid > entry->devid_end)
3171 region = kzalloc(sizeof(*region), GFP_KERNEL);
3173 pr_err("Out of memory allocating dm-regions for %s\n",
3178 region->start = entry->address_start;
3179 region->length = entry->address_end - entry->address_start;
3180 if (entry->prot & IOMMU_PROT_IR)
3181 region->prot |= IOMMU_READ;
3182 if (entry->prot & IOMMU_PROT_IW)
3183 region->prot |= IOMMU_WRITE;
3185 list_add_tail(®ion->list, head);
3189 static void amd_iommu_put_dm_regions(struct device *dev,
3190 struct list_head *head)
3192 struct iommu_dm_region *entry, *next;
3194 list_for_each_entry_safe(entry, next, head, list)
3198 static const struct iommu_ops amd_iommu_ops = {
3199 .capable = amd_iommu_capable,
3200 .domain_alloc = amd_iommu_domain_alloc,
3201 .domain_free = amd_iommu_domain_free,
3202 .attach_dev = amd_iommu_attach_device,
3203 .detach_dev = amd_iommu_detach_device,
3204 .map = amd_iommu_map,
3205 .unmap = amd_iommu_unmap,
3206 .map_sg = default_iommu_map_sg,
3207 .iova_to_phys = amd_iommu_iova_to_phys,
3208 .add_device = amd_iommu_add_device,
3209 .remove_device = amd_iommu_remove_device,
3210 .get_dm_regions = amd_iommu_get_dm_regions,
3211 .put_dm_regions = amd_iommu_put_dm_regions,
3212 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3215 /*****************************************************************************
3217 * The next functions do a basic initialization of IOMMU for pass through
3220 * In passthrough mode the IOMMU is initialized and enabled but not used for
3221 * DMA-API translation.
3223 *****************************************************************************/
3225 int __init amd_iommu_init_passthrough(void)
3227 struct iommu_dev_data *dev_data;
3228 struct pci_dev *dev = NULL;
3231 ret = alloc_passthrough_domain();
3235 for_each_pci_dev(dev) {
3236 if (!check_device(&dev->dev))
3239 dev_data = get_dev_data(&dev->dev);
3240 dev_data->passthrough = true;
3242 attach_device(&dev->dev, pt_domain);
3245 amd_iommu_stats_init();
3247 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3252 /* IOMMUv2 specific functions */
3253 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3255 return atomic_notifier_chain_register(&ppr_notifier, nb);
3257 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3259 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3261 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3263 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3265 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3267 struct protection_domain *domain = to_pdomain(dom);
3268 unsigned long flags;
3270 spin_lock_irqsave(&domain->lock, flags);
3272 /* Update data structure */
3273 domain->mode = PAGE_MODE_NONE;
3274 domain->updated = true;
3276 /* Make changes visible to IOMMUs */
3277 update_domain(domain);
3279 /* Page-table is not visible to IOMMU anymore, so free it */
3280 free_pagetable(domain);
3282 spin_unlock_irqrestore(&domain->lock, flags);
3284 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3286 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3288 struct protection_domain *domain = to_pdomain(dom);
3289 unsigned long flags;
3292 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3295 /* Number of GCR3 table levels required */
3296 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3299 if (levels > amd_iommu_max_glx_val)
3302 spin_lock_irqsave(&domain->lock, flags);
3305 * Save us all sanity checks whether devices already in the
3306 * domain support IOMMUv2. Just force that the domain has no
3307 * devices attached when it is switched into IOMMUv2 mode.
3310 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3314 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3315 if (domain->gcr3_tbl == NULL)
3318 domain->glx = levels;
3319 domain->flags |= PD_IOMMUV2_MASK;
3320 domain->updated = true;
3322 update_domain(domain);
3327 spin_unlock_irqrestore(&domain->lock, flags);
3331 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3333 static int __flush_pasid(struct protection_domain *domain, int pasid,
3334 u64 address, bool size)
3336 struct iommu_dev_data *dev_data;
3337 struct iommu_cmd cmd;
3340 if (!(domain->flags & PD_IOMMUV2_MASK))
3343 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3346 * IOMMU TLB needs to be flushed before Device TLB to
3347 * prevent device TLB refill from IOMMU TLB
3349 for (i = 0; i < amd_iommus_present; ++i) {
3350 if (domain->dev_iommu[i] == 0)
3353 ret = iommu_queue_command(amd_iommus[i], &cmd);
3358 /* Wait until IOMMU TLB flushes are complete */
3359 domain_flush_complete(domain);
3361 /* Now flush device TLBs */
3362 list_for_each_entry(dev_data, &domain->dev_list, list) {
3363 struct amd_iommu *iommu;
3366 BUG_ON(!dev_data->ats.enabled);
3368 qdep = dev_data->ats.qdep;
3369 iommu = amd_iommu_rlookup_table[dev_data->devid];
3371 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3372 qdep, address, size);
3374 ret = iommu_queue_command(iommu, &cmd);
3379 /* Wait until all device TLBs are flushed */
3380 domain_flush_complete(domain);
3389 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3392 INC_STATS_COUNTER(invalidate_iotlb);
3394 return __flush_pasid(domain, pasid, address, false);
3397 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3400 struct protection_domain *domain = to_pdomain(dom);
3401 unsigned long flags;
3404 spin_lock_irqsave(&domain->lock, flags);
3405 ret = __amd_iommu_flush_page(domain, pasid, address);
3406 spin_unlock_irqrestore(&domain->lock, flags);
3410 EXPORT_SYMBOL(amd_iommu_flush_page);
3412 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3414 INC_STATS_COUNTER(invalidate_iotlb_all);
3416 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3420 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3422 struct protection_domain *domain = to_pdomain(dom);
3423 unsigned long flags;
3426 spin_lock_irqsave(&domain->lock, flags);
3427 ret = __amd_iommu_flush_tlb(domain, pasid);
3428 spin_unlock_irqrestore(&domain->lock, flags);
3432 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3434 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3441 index = (pasid >> (9 * level)) & 0x1ff;
3447 if (!(*pte & GCR3_VALID)) {
3451 root = (void *)get_zeroed_page(GFP_ATOMIC);
3455 *pte = __pa(root) | GCR3_VALID;
3458 root = __va(*pte & PAGE_MASK);
3466 static int __set_gcr3(struct protection_domain *domain, int pasid,
3471 if (domain->mode != PAGE_MODE_NONE)
3474 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3478 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3480 return __amd_iommu_flush_tlb(domain, pasid);
3483 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3487 if (domain->mode != PAGE_MODE_NONE)
3490 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3496 return __amd_iommu_flush_tlb(domain, pasid);
3499 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3502 struct protection_domain *domain = to_pdomain(dom);
3503 unsigned long flags;
3506 spin_lock_irqsave(&domain->lock, flags);
3507 ret = __set_gcr3(domain, pasid, cr3);
3508 spin_unlock_irqrestore(&domain->lock, flags);
3512 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3514 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3516 struct protection_domain *domain = to_pdomain(dom);
3517 unsigned long flags;
3520 spin_lock_irqsave(&domain->lock, flags);
3521 ret = __clear_gcr3(domain, pasid);
3522 spin_unlock_irqrestore(&domain->lock, flags);
3526 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3528 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3529 int status, int tag)
3531 struct iommu_dev_data *dev_data;
3532 struct amd_iommu *iommu;
3533 struct iommu_cmd cmd;
3535 INC_STATS_COUNTER(complete_ppr);
3537 dev_data = get_dev_data(&pdev->dev);
3538 iommu = amd_iommu_rlookup_table[dev_data->devid];
3540 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3541 tag, dev_data->pri_tlp);
3543 return iommu_queue_command(iommu, &cmd);
3545 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3547 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3549 struct protection_domain *pdomain;
3551 pdomain = get_domain(&pdev->dev);
3552 if (IS_ERR(pdomain))
3555 /* Only return IOMMUv2 domains */
3556 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3559 return &pdomain->domain;
3561 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3563 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3565 struct iommu_dev_data *dev_data;
3567 if (!amd_iommu_v2_supported())
3570 dev_data = get_dev_data(&pdev->dev);
3571 dev_data->errata |= (1 << erratum);
3573 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3575 int amd_iommu_device_info(struct pci_dev *pdev,
3576 struct amd_iommu_device_info *info)
3581 if (pdev == NULL || info == NULL)
3584 if (!amd_iommu_v2_supported())
3587 memset(info, 0, sizeof(*info));
3589 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3591 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3593 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3595 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3597 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3601 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3602 max_pasids = min(max_pasids, (1 << 20));
3604 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3605 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3607 features = pci_pasid_features(pdev);
3608 if (features & PCI_PASID_CAP_EXEC)
3609 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3610 if (features & PCI_PASID_CAP_PRIV)
3611 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3616 EXPORT_SYMBOL(amd_iommu_device_info);
3618 #ifdef CONFIG_IRQ_REMAP
3620 /*****************************************************************************
3622 * Interrupt Remapping Implementation
3624 *****************************************************************************/
3642 u16 devid; /* Device ID for IRTE table */
3643 u16 index; /* Index into IRTE table*/
3646 struct amd_ir_data {
3647 struct irq_2_irte irq_2_irte;
3648 union irte irte_entry;
3650 struct msi_msg msi_entry;
3654 static struct irq_chip amd_ir_chip;
3656 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3657 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3658 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3659 #define DTE_IRQ_REMAP_ENABLE 1ULL
3661 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3665 dte = amd_iommu_dev_table[devid].data[2];
3666 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3667 dte |= virt_to_phys(table->table);
3668 dte |= DTE_IRQ_REMAP_INTCTL;
3669 dte |= DTE_IRQ_TABLE_LEN;
3670 dte |= DTE_IRQ_REMAP_ENABLE;
3672 amd_iommu_dev_table[devid].data[2] = dte;
3675 #define IRTE_ALLOCATED (~1U)
3677 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3679 struct irq_remap_table *table = NULL;
3680 struct amd_iommu *iommu;
3681 unsigned long flags;
3684 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3686 iommu = amd_iommu_rlookup_table[devid];
3690 table = irq_lookup_table[devid];
3694 alias = amd_iommu_alias_table[devid];
3695 table = irq_lookup_table[alias];
3697 irq_lookup_table[devid] = table;
3698 set_dte_irq_entry(devid, table);
3699 iommu_flush_dte(iommu, devid);
3703 /* Nothing there yet, allocate new irq remapping table */
3704 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3708 /* Initialize table spin-lock */
3709 spin_lock_init(&table->lock);
3712 /* Keep the first 32 indexes free for IOAPIC interrupts */
3713 table->min_index = 32;
3715 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3716 if (!table->table) {
3722 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3727 for (i = 0; i < 32; ++i)
3728 table->table[i] = IRTE_ALLOCATED;
3731 irq_lookup_table[devid] = table;
3732 set_dte_irq_entry(devid, table);
3733 iommu_flush_dte(iommu, devid);
3734 if (devid != alias) {
3735 irq_lookup_table[alias] = table;
3736 set_dte_irq_entry(alias, table);
3737 iommu_flush_dte(iommu, alias);
3741 iommu_completion_wait(iommu);
3744 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3749 static int alloc_irq_index(u16 devid, int count)
3751 struct irq_remap_table *table;
3752 unsigned long flags;
3755 table = get_irq_table(devid, false);
3759 spin_lock_irqsave(&table->lock, flags);
3761 /* Scan table for free entries */
3762 for (c = 0, index = table->min_index;
3763 index < MAX_IRQS_PER_TABLE;
3765 if (table->table[index] == 0)
3772 table->table[index - c + 1] = IRTE_ALLOCATED;
3782 spin_unlock_irqrestore(&table->lock, flags);
3787 static int modify_irte(u16 devid, int index, union irte irte)
3789 struct irq_remap_table *table;
3790 struct amd_iommu *iommu;
3791 unsigned long flags;
3793 iommu = amd_iommu_rlookup_table[devid];
3797 table = get_irq_table(devid, false);
3801 spin_lock_irqsave(&table->lock, flags);
3802 table->table[index] = irte.val;
3803 spin_unlock_irqrestore(&table->lock, flags);
3805 iommu_flush_irt(iommu, devid);
3806 iommu_completion_wait(iommu);
3811 static void free_irte(u16 devid, int index)
3813 struct irq_remap_table *table;
3814 struct amd_iommu *iommu;
3815 unsigned long flags;
3817 iommu = amd_iommu_rlookup_table[devid];
3821 table = get_irq_table(devid, false);
3825 spin_lock_irqsave(&table->lock, flags);
3826 table->table[index] = 0;
3827 spin_unlock_irqrestore(&table->lock, flags);
3829 iommu_flush_irt(iommu, devid);
3830 iommu_completion_wait(iommu);
3833 static int get_devid(struct irq_alloc_info *info)
3837 switch (info->type) {
3838 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3839 devid = get_ioapic_devid(info->ioapic_id);
3841 case X86_IRQ_ALLOC_TYPE_HPET:
3842 devid = get_hpet_devid(info->hpet_id);
3844 case X86_IRQ_ALLOC_TYPE_MSI:
3845 case X86_IRQ_ALLOC_TYPE_MSIX:
3846 devid = get_device_id(&info->msi_dev->dev);
3856 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3858 struct amd_iommu *iommu;
3864 devid = get_devid(info);
3866 iommu = amd_iommu_rlookup_table[devid];
3868 return iommu->ir_domain;
3874 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3876 struct amd_iommu *iommu;
3882 switch (info->type) {
3883 case X86_IRQ_ALLOC_TYPE_MSI:
3884 case X86_IRQ_ALLOC_TYPE_MSIX:
3885 devid = get_device_id(&info->msi_dev->dev);
3887 iommu = amd_iommu_rlookup_table[devid];
3889 return iommu->msi_domain;
3899 struct irq_remap_ops amd_iommu_irq_ops = {
3900 .prepare = amd_iommu_prepare,
3901 .enable = amd_iommu_enable,
3902 .disable = amd_iommu_disable,
3903 .reenable = amd_iommu_reenable,
3904 .enable_faulting = amd_iommu_enable_faulting,
3905 .get_ir_irq_domain = get_ir_irq_domain,
3906 .get_irq_domain = get_irq_domain,
3909 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3910 struct irq_cfg *irq_cfg,
3911 struct irq_alloc_info *info,
3912 int devid, int index, int sub_handle)
3914 struct irq_2_irte *irte_info = &data->irq_2_irte;
3915 struct msi_msg *msg = &data->msi_entry;
3916 union irte *irte = &data->irte_entry;
3917 struct IO_APIC_route_entry *entry;
3919 data->irq_2_irte.devid = devid;
3920 data->irq_2_irte.index = index + sub_handle;
3922 /* Setup IRTE for IOMMU */
3924 irte->fields.vector = irq_cfg->vector;
3925 irte->fields.int_type = apic->irq_delivery_mode;
3926 irte->fields.destination = irq_cfg->dest_apicid;
3927 irte->fields.dm = apic->irq_dest_mode;
3928 irte->fields.valid = 1;
3930 switch (info->type) {
3931 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3932 /* Setup IOAPIC entry */
3933 entry = info->ioapic_entry;
3934 info->ioapic_entry = NULL;
3935 memset(entry, 0, sizeof(*entry));
3936 entry->vector = index;
3938 entry->trigger = info->ioapic_trigger;
3939 entry->polarity = info->ioapic_polarity;
3940 /* Mask level triggered irqs. */
3941 if (info->ioapic_trigger)
3945 case X86_IRQ_ALLOC_TYPE_HPET:
3946 case X86_IRQ_ALLOC_TYPE_MSI:
3947 case X86_IRQ_ALLOC_TYPE_MSIX:
3948 msg->address_hi = MSI_ADDR_BASE_HI;
3949 msg->address_lo = MSI_ADDR_BASE_LO;
3950 msg->data = irte_info->index;
3959 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3960 unsigned int nr_irqs, void *arg)
3962 struct irq_alloc_info *info = arg;
3963 struct irq_data *irq_data;
3964 struct amd_ir_data *data;
3965 struct irq_cfg *cfg;
3971 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3972 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3976 * With IRQ remapping enabled, don't need contiguous CPU vectors
3977 * to support multiple MSI interrupts.
3979 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3980 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3982 devid = get_devid(info);
3986 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3991 data = kzalloc(sizeof(*data), GFP_KERNEL);
3993 goto out_free_parent;
3995 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3996 if (get_irq_table(devid, true))
3997 index = info->ioapic_pin;
4001 index = alloc_irq_index(devid, nr_irqs);
4004 pr_warn("Failed to allocate IRTE\n");
4006 goto out_free_parent;
4009 for (i = 0; i < nr_irqs; i++) {
4010 irq_data = irq_domain_get_irq_data(domain, virq + i);
4011 cfg = irqd_cfg(irq_data);
4012 if (!irq_data || !cfg) {
4018 data = kzalloc(sizeof(*data), GFP_KERNEL);
4022 irq_data->hwirq = (devid << 16) + i;
4023 irq_data->chip_data = data;
4024 irq_data->chip = &amd_ir_chip;
4025 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4026 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4031 for (i--; i >= 0; i--) {
4032 irq_data = irq_domain_get_irq_data(domain, virq + i);
4034 kfree(irq_data->chip_data);
4036 for (i = 0; i < nr_irqs; i++)
4037 free_irte(devid, index + i);
4039 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4043 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4044 unsigned int nr_irqs)
4046 struct irq_2_irte *irte_info;
4047 struct irq_data *irq_data;
4048 struct amd_ir_data *data;
4051 for (i = 0; i < nr_irqs; i++) {
4052 irq_data = irq_domain_get_irq_data(domain, virq + i);
4053 if (irq_data && irq_data->chip_data) {
4054 data = irq_data->chip_data;
4055 irte_info = &data->irq_2_irte;
4056 free_irte(irte_info->devid, irte_info->index);
4060 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4063 static void irq_remapping_activate(struct irq_domain *domain,
4064 struct irq_data *irq_data)
4066 struct amd_ir_data *data = irq_data->chip_data;
4067 struct irq_2_irte *irte_info = &data->irq_2_irte;
4069 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4072 static void irq_remapping_deactivate(struct irq_domain *domain,
4073 struct irq_data *irq_data)
4075 struct amd_ir_data *data = irq_data->chip_data;
4076 struct irq_2_irte *irte_info = &data->irq_2_irte;
4080 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4083 static struct irq_domain_ops amd_ir_domain_ops = {
4084 .alloc = irq_remapping_alloc,
4085 .free = irq_remapping_free,
4086 .activate = irq_remapping_activate,
4087 .deactivate = irq_remapping_deactivate,
4090 static int amd_ir_set_affinity(struct irq_data *data,
4091 const struct cpumask *mask, bool force)
4093 struct amd_ir_data *ir_data = data->chip_data;
4094 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4095 struct irq_cfg *cfg = irqd_cfg(data);
4096 struct irq_data *parent = data->parent_data;
4099 ret = parent->chip->irq_set_affinity(parent, mask, force);
4100 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4104 * Atomically updates the IRTE with the new destination, vector
4105 * and flushes the interrupt entry cache.
4107 ir_data->irte_entry.fields.vector = cfg->vector;
4108 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4109 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4112 * After this point, all the interrupts will start arriving
4113 * at the new destination. So, time to cleanup the previous
4114 * vector allocation.
4116 send_cleanup_vector(cfg);
4118 return IRQ_SET_MASK_OK_DONE;
4121 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4123 struct amd_ir_data *ir_data = irq_data->chip_data;
4125 *msg = ir_data->msi_entry;
4128 static struct irq_chip amd_ir_chip = {
4129 .irq_ack = ir_ack_apic_edge,
4130 .irq_set_affinity = amd_ir_set_affinity,
4131 .irq_compose_msi_msg = ir_compose_msi_msg,
4134 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4136 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4137 if (!iommu->ir_domain)
4140 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4141 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);