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 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);
121 /****************************************************************************
125 ****************************************************************************/
127 static struct protection_domain *to_pdomain(struct iommu_domain *dom)
129 return container_of(dom, struct protection_domain, domain);
132 static struct iommu_dev_data *alloc_dev_data(u16 devid)
134 struct iommu_dev_data *dev_data;
137 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
141 INIT_LIST_HEAD(&dev_data->alias_list);
143 dev_data->devid = devid;
145 spin_lock_irqsave(&dev_data_list_lock, flags);
146 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
147 spin_unlock_irqrestore(&dev_data_list_lock, flags);
152 static void free_dev_data(struct iommu_dev_data *dev_data)
156 spin_lock_irqsave(&dev_data_list_lock, flags);
157 list_del(&dev_data->dev_data_list);
158 spin_unlock_irqrestore(&dev_data_list_lock, flags);
163 static struct iommu_dev_data *search_dev_data(u16 devid)
165 struct iommu_dev_data *dev_data;
168 spin_lock_irqsave(&dev_data_list_lock, flags);
169 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
170 if (dev_data->devid == devid)
177 spin_unlock_irqrestore(&dev_data_list_lock, flags);
182 static struct iommu_dev_data *find_dev_data(u16 devid)
184 struct iommu_dev_data *dev_data;
186 dev_data = search_dev_data(devid);
188 if (dev_data == NULL)
189 dev_data = alloc_dev_data(devid);
194 static inline u16 get_device_id(struct device *dev)
196 struct pci_dev *pdev = to_pci_dev(dev);
198 return PCI_DEVID(pdev->bus->number, pdev->devfn);
201 static struct iommu_dev_data *get_dev_data(struct device *dev)
203 return dev->archdata.iommu;
206 static bool pci_iommuv2_capable(struct pci_dev *pdev)
208 static const int caps[] = {
211 PCI_EXT_CAP_ID_PASID,
215 for (i = 0; i < 3; ++i) {
216 pos = pci_find_ext_capability(pdev, caps[i]);
224 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
226 struct iommu_dev_data *dev_data;
228 dev_data = get_dev_data(&pdev->dev);
230 return dev_data->errata & (1 << erratum) ? true : false;
234 * This function actually applies the mapping to the page table of the
237 static void alloc_unity_mapping(struct dma_ops_domain *dma_dom,
238 struct unity_map_entry *e)
242 for (addr = e->address_start; addr < e->address_end;
244 if (addr < dma_dom->aperture_size)
245 __set_bit(addr >> PAGE_SHIFT,
246 dma_dom->aperture[0]->bitmap);
251 * Inits the unity mappings required for a specific device
253 static void init_unity_mappings_for_device(struct device *dev,
254 struct dma_ops_domain *dma_dom)
256 struct unity_map_entry *e;
259 devid = get_device_id(dev);
261 list_for_each_entry(e, &amd_iommu_unity_map, list) {
262 if (!(devid >= e->devid_start && devid <= e->devid_end))
264 alloc_unity_mapping(dma_dom, e);
269 * This function checks if the driver got a valid device from the caller to
270 * avoid dereferencing invalid pointers.
272 static bool check_device(struct device *dev)
276 if (!dev || !dev->dma_mask)
280 if (!dev_is_pci(dev))
283 devid = get_device_id(dev);
285 /* Out of our scope? */
286 if (devid > amd_iommu_last_bdf)
289 if (amd_iommu_rlookup_table[devid] == NULL)
295 static void init_iommu_group(struct device *dev)
297 struct dma_ops_domain *dma_domain;
298 struct iommu_domain *domain;
299 struct iommu_group *group;
301 group = iommu_group_get_for_dev(dev);
305 domain = iommu_group_default_domain(group);
309 dma_domain = to_pdomain(domain)->priv;
311 init_unity_mappings_for_device(dev, dma_domain);
313 iommu_group_put(group);
316 static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
318 *(u16 *)data = alias;
322 static u16 get_alias(struct device *dev)
324 struct pci_dev *pdev = to_pci_dev(dev);
325 u16 devid, ivrs_alias, pci_alias;
327 devid = get_device_id(dev);
328 ivrs_alias = amd_iommu_alias_table[devid];
329 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
331 if (ivrs_alias == pci_alias)
337 * The IVRS is fairly reliable in telling us about aliases, but it
338 * can't know about every screwy device. If we don't have an IVRS
339 * reported alias, use the PCI reported alias. In that case we may
340 * still need to initialize the rlookup and dev_table entries if the
341 * alias is to a non-existent device.
343 if (ivrs_alias == devid) {
344 if (!amd_iommu_rlookup_table[pci_alias]) {
345 amd_iommu_rlookup_table[pci_alias] =
346 amd_iommu_rlookup_table[devid];
347 memcpy(amd_iommu_dev_table[pci_alias].data,
348 amd_iommu_dev_table[devid].data,
349 sizeof(amd_iommu_dev_table[pci_alias].data));
355 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
356 "for device %s[%04x:%04x], kernel reported alias "
357 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
358 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
359 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
360 PCI_FUNC(pci_alias));
363 * If we don't have a PCI DMA alias and the IVRS alias is on the same
364 * bus, then the IVRS table may know about a quirk that we don't.
366 if (pci_alias == devid &&
367 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
368 pdev->dev_flags |= PCI_DEV_FLAGS_DMA_ALIAS_DEVFN;
369 pdev->dma_alias_devfn = ivrs_alias & 0xff;
370 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
371 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
378 static int iommu_init_device(struct device *dev)
380 struct pci_dev *pdev = to_pci_dev(dev);
381 struct iommu_dev_data *dev_data;
384 if (dev->archdata.iommu)
387 dev_data = find_dev_data(get_device_id(dev));
391 alias = get_alias(dev);
393 if (alias != dev_data->devid) {
394 struct iommu_dev_data *alias_data;
396 alias_data = find_dev_data(alias);
397 if (alias_data == NULL) {
398 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
400 free_dev_data(dev_data);
403 dev_data->alias_data = alias_data;
405 /* Add device to the alias_list */
406 list_add(&dev_data->alias_list, &alias_data->alias_list);
409 if (pci_iommuv2_capable(pdev)) {
410 struct amd_iommu *iommu;
412 iommu = amd_iommu_rlookup_table[dev_data->devid];
413 dev_data->iommu_v2 = iommu->is_iommu_v2;
416 dev->archdata.iommu = dev_data;
418 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
424 static void iommu_ignore_device(struct device *dev)
428 devid = get_device_id(dev);
429 alias = amd_iommu_alias_table[devid];
431 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
432 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
434 amd_iommu_rlookup_table[devid] = NULL;
435 amd_iommu_rlookup_table[alias] = NULL;
438 static void iommu_uninit_device(struct device *dev)
440 struct iommu_dev_data *dev_data = search_dev_data(get_device_id(dev));
445 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
448 iommu_group_remove_device(dev);
450 /* Unlink from alias, it may change if another device is re-plugged */
451 dev_data->alias_data = NULL;
454 dev->archdata.dma_ops = NULL;
457 * We keep dev_data around for unplugged devices and reuse it when the
458 * device is re-plugged - not doing so would introduce a ton of races.
462 #ifdef CONFIG_AMD_IOMMU_STATS
465 * Initialization code for statistics collection
468 DECLARE_STATS_COUNTER(compl_wait);
469 DECLARE_STATS_COUNTER(cnt_map_single);
470 DECLARE_STATS_COUNTER(cnt_unmap_single);
471 DECLARE_STATS_COUNTER(cnt_map_sg);
472 DECLARE_STATS_COUNTER(cnt_unmap_sg);
473 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
474 DECLARE_STATS_COUNTER(cnt_free_coherent);
475 DECLARE_STATS_COUNTER(cross_page);
476 DECLARE_STATS_COUNTER(domain_flush_single);
477 DECLARE_STATS_COUNTER(domain_flush_all);
478 DECLARE_STATS_COUNTER(alloced_io_mem);
479 DECLARE_STATS_COUNTER(total_map_requests);
480 DECLARE_STATS_COUNTER(complete_ppr);
481 DECLARE_STATS_COUNTER(invalidate_iotlb);
482 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
483 DECLARE_STATS_COUNTER(pri_requests);
485 static struct dentry *stats_dir;
486 static struct dentry *de_fflush;
488 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
490 if (stats_dir == NULL)
493 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
497 static void amd_iommu_stats_init(void)
499 stats_dir = debugfs_create_dir("amd-iommu", NULL);
500 if (stats_dir == NULL)
503 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
504 &amd_iommu_unmap_flush);
506 amd_iommu_stats_add(&compl_wait);
507 amd_iommu_stats_add(&cnt_map_single);
508 amd_iommu_stats_add(&cnt_unmap_single);
509 amd_iommu_stats_add(&cnt_map_sg);
510 amd_iommu_stats_add(&cnt_unmap_sg);
511 amd_iommu_stats_add(&cnt_alloc_coherent);
512 amd_iommu_stats_add(&cnt_free_coherent);
513 amd_iommu_stats_add(&cross_page);
514 amd_iommu_stats_add(&domain_flush_single);
515 amd_iommu_stats_add(&domain_flush_all);
516 amd_iommu_stats_add(&alloced_io_mem);
517 amd_iommu_stats_add(&total_map_requests);
518 amd_iommu_stats_add(&complete_ppr);
519 amd_iommu_stats_add(&invalidate_iotlb);
520 amd_iommu_stats_add(&invalidate_iotlb_all);
521 amd_iommu_stats_add(&pri_requests);
526 /****************************************************************************
528 * Interrupt handling functions
530 ****************************************************************************/
532 static void dump_dte_entry(u16 devid)
536 for (i = 0; i < 4; ++i)
537 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
538 amd_iommu_dev_table[devid].data[i]);
541 static void dump_command(unsigned long phys_addr)
543 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
546 for (i = 0; i < 4; ++i)
547 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
550 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
552 int type, devid, domid, flags;
553 volatile u32 *event = __evt;
558 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
559 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
560 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
561 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
562 address = (u64)(((u64)event[3]) << 32) | event[2];
565 /* Did we hit the erratum? */
566 if (++count == LOOP_TIMEOUT) {
567 pr_err("AMD-Vi: No event written to event log\n");
574 printk(KERN_ERR "AMD-Vi: Event logged [");
577 case EVENT_TYPE_ILL_DEV:
578 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
579 "address=0x%016llx flags=0x%04x]\n",
580 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
582 dump_dte_entry(devid);
584 case EVENT_TYPE_IO_FAULT:
585 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
586 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
587 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
588 domid, address, flags);
590 case EVENT_TYPE_DEV_TAB_ERR:
591 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
592 "address=0x%016llx flags=0x%04x]\n",
593 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
596 case EVENT_TYPE_PAGE_TAB_ERR:
597 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
598 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
599 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
600 domid, address, flags);
602 case EVENT_TYPE_ILL_CMD:
603 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
604 dump_command(address);
606 case EVENT_TYPE_CMD_HARD_ERR:
607 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
608 "flags=0x%04x]\n", address, flags);
610 case EVENT_TYPE_IOTLB_INV_TO:
611 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
612 "address=0x%016llx]\n",
613 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
616 case EVENT_TYPE_INV_DEV_REQ:
617 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
618 "address=0x%016llx flags=0x%04x]\n",
619 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
623 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
626 memset(__evt, 0, 4 * sizeof(u32));
629 static void iommu_poll_events(struct amd_iommu *iommu)
633 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
634 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
636 while (head != tail) {
637 iommu_print_event(iommu, iommu->evt_buf + head);
638 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
641 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
644 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
646 struct amd_iommu_fault fault;
648 INC_STATS_COUNTER(pri_requests);
650 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
651 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
655 fault.address = raw[1];
656 fault.pasid = PPR_PASID(raw[0]);
657 fault.device_id = PPR_DEVID(raw[0]);
658 fault.tag = PPR_TAG(raw[0]);
659 fault.flags = PPR_FLAGS(raw[0]);
661 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
664 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
668 if (iommu->ppr_log == NULL)
671 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
672 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
674 while (head != tail) {
679 raw = (u64 *)(iommu->ppr_log + head);
682 * Hardware bug: Interrupt may arrive before the entry is
683 * written to memory. If this happens we need to wait for the
686 for (i = 0; i < LOOP_TIMEOUT; ++i) {
687 if (PPR_REQ_TYPE(raw[0]) != 0)
692 /* Avoid memcpy function-call overhead */
697 * To detect the hardware bug we need to clear the entry
700 raw[0] = raw[1] = 0UL;
702 /* Update head pointer of hardware ring-buffer */
703 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
704 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
706 /* Handle PPR entry */
707 iommu_handle_ppr_entry(iommu, entry);
709 /* Refresh ring-buffer information */
710 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
711 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
715 irqreturn_t amd_iommu_int_thread(int irq, void *data)
717 struct amd_iommu *iommu = (struct amd_iommu *) data;
718 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
720 while (status & (MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK)) {
721 /* Enable EVT and PPR interrupts again */
722 writel((MMIO_STATUS_EVT_INT_MASK | MMIO_STATUS_PPR_INT_MASK),
723 iommu->mmio_base + MMIO_STATUS_OFFSET);
725 if (status & MMIO_STATUS_EVT_INT_MASK) {
726 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
727 iommu_poll_events(iommu);
730 if (status & MMIO_STATUS_PPR_INT_MASK) {
731 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
732 iommu_poll_ppr_log(iommu);
736 * Hardware bug: ERBT1312
737 * When re-enabling interrupt (by writing 1
738 * to clear the bit), the hardware might also try to set
739 * the interrupt bit in the event status register.
740 * In this scenario, the bit will be set, and disable
741 * subsequent interrupts.
743 * Workaround: The IOMMU driver should read back the
744 * status register and check if the interrupt bits are cleared.
745 * If not, driver will need to go through the interrupt handler
746 * again and re-clear the bits
748 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
753 irqreturn_t amd_iommu_int_handler(int irq, void *data)
755 return IRQ_WAKE_THREAD;
758 /****************************************************************************
760 * IOMMU command queuing functions
762 ****************************************************************************/
764 static int wait_on_sem(volatile u64 *sem)
768 while (*sem == 0 && i < LOOP_TIMEOUT) {
773 if (i == LOOP_TIMEOUT) {
774 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
781 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
782 struct iommu_cmd *cmd,
787 target = iommu->cmd_buf + tail;
788 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
790 /* Copy command to buffer */
791 memcpy(target, cmd, sizeof(*cmd));
793 /* Tell the IOMMU about it */
794 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
797 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
799 WARN_ON(address & 0x7ULL);
801 memset(cmd, 0, sizeof(*cmd));
802 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
803 cmd->data[1] = upper_32_bits(__pa(address));
805 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
808 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
810 memset(cmd, 0, sizeof(*cmd));
811 cmd->data[0] = devid;
812 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
815 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
816 size_t size, u16 domid, int pde)
821 pages = iommu_num_pages(address, size, PAGE_SIZE);
826 * If we have to flush more than one page, flush all
827 * TLB entries for this domain
829 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
833 address &= PAGE_MASK;
835 memset(cmd, 0, sizeof(*cmd));
836 cmd->data[1] |= domid;
837 cmd->data[2] = lower_32_bits(address);
838 cmd->data[3] = upper_32_bits(address);
839 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
840 if (s) /* size bit - we flush more than one 4kb page */
841 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
842 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
843 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
846 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
847 u64 address, size_t size)
852 pages = iommu_num_pages(address, size, PAGE_SIZE);
857 * If we have to flush more than one page, flush all
858 * TLB entries for this domain
860 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
864 address &= PAGE_MASK;
866 memset(cmd, 0, sizeof(*cmd));
867 cmd->data[0] = devid;
868 cmd->data[0] |= (qdep & 0xff) << 24;
869 cmd->data[1] = devid;
870 cmd->data[2] = lower_32_bits(address);
871 cmd->data[3] = upper_32_bits(address);
872 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
874 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
877 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
878 u64 address, bool size)
880 memset(cmd, 0, sizeof(*cmd));
882 address &= ~(0xfffULL);
884 cmd->data[0] = pasid;
885 cmd->data[1] = domid;
886 cmd->data[2] = lower_32_bits(address);
887 cmd->data[3] = upper_32_bits(address);
888 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
889 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
891 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
892 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
895 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
896 int qdep, u64 address, bool size)
898 memset(cmd, 0, sizeof(*cmd));
900 address &= ~(0xfffULL);
902 cmd->data[0] = devid;
903 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
904 cmd->data[0] |= (qdep & 0xff) << 24;
905 cmd->data[1] = devid;
906 cmd->data[1] |= (pasid & 0xff) << 16;
907 cmd->data[2] = lower_32_bits(address);
908 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
909 cmd->data[3] = upper_32_bits(address);
911 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
912 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
915 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
916 int status, int tag, bool gn)
918 memset(cmd, 0, sizeof(*cmd));
920 cmd->data[0] = devid;
922 cmd->data[1] = pasid;
923 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
925 cmd->data[3] = tag & 0x1ff;
926 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
928 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
931 static void build_inv_all(struct iommu_cmd *cmd)
933 memset(cmd, 0, sizeof(*cmd));
934 CMD_SET_TYPE(cmd, CMD_INV_ALL);
937 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
939 memset(cmd, 0, sizeof(*cmd));
940 cmd->data[0] = devid;
941 CMD_SET_TYPE(cmd, CMD_INV_IRT);
945 * Writes the command to the IOMMUs command buffer and informs the
946 * hardware about the new command.
948 static int iommu_queue_command_sync(struct amd_iommu *iommu,
949 struct iommu_cmd *cmd,
952 u32 left, tail, head, next_tail;
955 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
958 spin_lock_irqsave(&iommu->lock, flags);
960 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
961 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
962 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
963 left = (head - next_tail) % iommu->cmd_buf_size;
966 struct iommu_cmd sync_cmd;
967 volatile u64 sem = 0;
970 build_completion_wait(&sync_cmd, (u64)&sem);
971 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
973 spin_unlock_irqrestore(&iommu->lock, flags);
975 if ((ret = wait_on_sem(&sem)) != 0)
981 copy_cmd_to_buffer(iommu, cmd, tail);
983 /* We need to sync now to make sure all commands are processed */
984 iommu->need_sync = sync;
986 spin_unlock_irqrestore(&iommu->lock, flags);
991 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
993 return iommu_queue_command_sync(iommu, cmd, true);
997 * This function queues a completion wait command into the command
1000 static int iommu_completion_wait(struct amd_iommu *iommu)
1002 struct iommu_cmd cmd;
1003 volatile u64 sem = 0;
1006 if (!iommu->need_sync)
1009 build_completion_wait(&cmd, (u64)&sem);
1011 ret = iommu_queue_command_sync(iommu, &cmd, false);
1015 return wait_on_sem(&sem);
1018 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1020 struct iommu_cmd cmd;
1022 build_inv_dte(&cmd, devid);
1024 return iommu_queue_command(iommu, &cmd);
1027 static void iommu_flush_dte_all(struct amd_iommu *iommu)
1031 for (devid = 0; devid <= 0xffff; ++devid)
1032 iommu_flush_dte(iommu, devid);
1034 iommu_completion_wait(iommu);
1038 * This function uses heavy locking and may disable irqs for some time. But
1039 * this is no issue because it is only called during resume.
1041 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1045 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1046 struct iommu_cmd cmd;
1047 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1049 iommu_queue_command(iommu, &cmd);
1052 iommu_completion_wait(iommu);
1055 static void iommu_flush_all(struct amd_iommu *iommu)
1057 struct iommu_cmd cmd;
1059 build_inv_all(&cmd);
1061 iommu_queue_command(iommu, &cmd);
1062 iommu_completion_wait(iommu);
1065 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1067 struct iommu_cmd cmd;
1069 build_inv_irt(&cmd, devid);
1071 iommu_queue_command(iommu, &cmd);
1074 static void iommu_flush_irt_all(struct amd_iommu *iommu)
1078 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1079 iommu_flush_irt(iommu, devid);
1081 iommu_completion_wait(iommu);
1084 void iommu_flush_all_caches(struct amd_iommu *iommu)
1086 if (iommu_feature(iommu, FEATURE_IA)) {
1087 iommu_flush_all(iommu);
1089 iommu_flush_dte_all(iommu);
1090 iommu_flush_irt_all(iommu);
1091 iommu_flush_tlb_all(iommu);
1096 * Command send function for flushing on-device TLB
1098 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1099 u64 address, size_t size)
1101 struct amd_iommu *iommu;
1102 struct iommu_cmd cmd;
1105 qdep = dev_data->ats.qdep;
1106 iommu = amd_iommu_rlookup_table[dev_data->devid];
1108 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1110 return iommu_queue_command(iommu, &cmd);
1114 * Command send function for invalidating a device table entry
1116 static int device_flush_dte(struct iommu_dev_data *dev_data)
1118 struct amd_iommu *iommu;
1121 iommu = amd_iommu_rlookup_table[dev_data->devid];
1123 ret = iommu_flush_dte(iommu, dev_data->devid);
1127 if (dev_data->ats.enabled)
1128 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1134 * TLB invalidation function which is called from the mapping functions.
1135 * It invalidates a single PTE if the range to flush is within a single
1136 * page. Otherwise it flushes the whole TLB of the IOMMU.
1138 static void __domain_flush_pages(struct protection_domain *domain,
1139 u64 address, size_t size, int pde)
1141 struct iommu_dev_data *dev_data;
1142 struct iommu_cmd cmd;
1145 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1147 for (i = 0; i < amd_iommus_present; ++i) {
1148 if (!domain->dev_iommu[i])
1152 * Devices of this domain are behind this IOMMU
1153 * We need a TLB flush
1155 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1158 list_for_each_entry(dev_data, &domain->dev_list, list) {
1160 if (!dev_data->ats.enabled)
1163 ret |= device_flush_iotlb(dev_data, address, size);
1169 static void domain_flush_pages(struct protection_domain *domain,
1170 u64 address, size_t size)
1172 __domain_flush_pages(domain, address, size, 0);
1175 /* Flush the whole IO/TLB for a given protection domain */
1176 static void domain_flush_tlb(struct protection_domain *domain)
1178 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1181 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1182 static void domain_flush_tlb_pde(struct protection_domain *domain)
1184 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1187 static void domain_flush_complete(struct protection_domain *domain)
1191 for (i = 0; i < amd_iommus_present; ++i) {
1192 if (!domain->dev_iommu[i])
1196 * Devices of this domain are behind this IOMMU
1197 * We need to wait for completion of all commands.
1199 iommu_completion_wait(amd_iommus[i]);
1205 * This function flushes the DTEs for all devices in domain
1207 static void domain_flush_devices(struct protection_domain *domain)
1209 struct iommu_dev_data *dev_data;
1211 list_for_each_entry(dev_data, &domain->dev_list, list)
1212 device_flush_dte(dev_data);
1215 /****************************************************************************
1217 * The functions below are used the create the page table mappings for
1218 * unity mapped regions.
1220 ****************************************************************************/
1223 * This function is used to add another level to an IO page table. Adding
1224 * another level increases the size of the address space by 9 bits to a size up
1227 static bool increase_address_space(struct protection_domain *domain,
1232 if (domain->mode == PAGE_MODE_6_LEVEL)
1233 /* address space already 64 bit large */
1236 pte = (void *)get_zeroed_page(gfp);
1240 *pte = PM_LEVEL_PDE(domain->mode,
1241 virt_to_phys(domain->pt_root));
1242 domain->pt_root = pte;
1244 domain->updated = true;
1249 static u64 *alloc_pte(struct protection_domain *domain,
1250 unsigned long address,
1251 unsigned long page_size,
1258 BUG_ON(!is_power_of_2(page_size));
1260 while (address > PM_LEVEL_SIZE(domain->mode))
1261 increase_address_space(domain, gfp);
1263 level = domain->mode - 1;
1264 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1265 address = PAGE_SIZE_ALIGN(address, page_size);
1266 end_lvl = PAGE_SIZE_LEVEL(page_size);
1268 while (level > end_lvl) {
1269 if (!IOMMU_PTE_PRESENT(*pte)) {
1270 page = (u64 *)get_zeroed_page(gfp);
1273 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1276 /* No level skipping support yet */
1277 if (PM_PTE_LEVEL(*pte) != level)
1282 pte = IOMMU_PTE_PAGE(*pte);
1284 if (pte_page && level == end_lvl)
1287 pte = &pte[PM_LEVEL_INDEX(level, address)];
1294 * This function checks if there is a PTE for a given dma address. If
1295 * there is one, it returns the pointer to it.
1297 static u64 *fetch_pte(struct protection_domain *domain,
1298 unsigned long address,
1299 unsigned long *page_size)
1304 if (address > PM_LEVEL_SIZE(domain->mode))
1307 level = domain->mode - 1;
1308 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1309 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1314 if (!IOMMU_PTE_PRESENT(*pte))
1318 if (PM_PTE_LEVEL(*pte) == 7 ||
1319 PM_PTE_LEVEL(*pte) == 0)
1322 /* No level skipping support yet */
1323 if (PM_PTE_LEVEL(*pte) != level)
1328 /* Walk to the next level */
1329 pte = IOMMU_PTE_PAGE(*pte);
1330 pte = &pte[PM_LEVEL_INDEX(level, address)];
1331 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1334 if (PM_PTE_LEVEL(*pte) == 0x07) {
1335 unsigned long pte_mask;
1338 * If we have a series of large PTEs, make
1339 * sure to return a pointer to the first one.
1341 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1342 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1343 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1350 * Generic mapping functions. It maps a physical address into a DMA
1351 * address space. It allocates the page table pages if necessary.
1352 * In the future it can be extended to a generic mapping function
1353 * supporting all features of AMD IOMMU page tables like level skipping
1354 * and full 64 bit address spaces.
1356 static int iommu_map_page(struct protection_domain *dom,
1357 unsigned long bus_addr,
1358 unsigned long phys_addr,
1360 unsigned long page_size)
1365 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1366 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1368 if (!(prot & IOMMU_PROT_MASK))
1371 count = PAGE_SIZE_PTE_COUNT(page_size);
1372 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1377 for (i = 0; i < count; ++i)
1378 if (IOMMU_PTE_PRESENT(pte[i]))
1382 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1383 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1385 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1387 if (prot & IOMMU_PROT_IR)
1388 __pte |= IOMMU_PTE_IR;
1389 if (prot & IOMMU_PROT_IW)
1390 __pte |= IOMMU_PTE_IW;
1392 for (i = 0; i < count; ++i)
1400 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1401 unsigned long bus_addr,
1402 unsigned long page_size)
1404 unsigned long long unmapped;
1405 unsigned long unmap_size;
1408 BUG_ON(!is_power_of_2(page_size));
1412 while (unmapped < page_size) {
1414 pte = fetch_pte(dom, bus_addr, &unmap_size);
1419 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1420 for (i = 0; i < count; i++)
1424 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1425 unmapped += unmap_size;
1428 BUG_ON(unmapped && !is_power_of_2(unmapped));
1433 /****************************************************************************
1435 * The next functions belong to the address allocator for the dma_ops
1436 * interface functions. They work like the allocators in the other IOMMU
1437 * drivers. Its basically a bitmap which marks the allocated pages in
1438 * the aperture. Maybe it could be enhanced in the future to a more
1439 * efficient allocator.
1441 ****************************************************************************/
1444 * The address allocator core functions.
1446 * called with domain->lock held
1450 * Used to reserve address ranges in the aperture (e.g. for exclusion
1453 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1454 unsigned long start_page,
1457 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1459 if (start_page + pages > last_page)
1460 pages = last_page - start_page;
1462 for (i = start_page; i < start_page + pages; ++i) {
1463 int index = i / APERTURE_RANGE_PAGES;
1464 int page = i % APERTURE_RANGE_PAGES;
1465 __set_bit(page, dom->aperture[index]->bitmap);
1470 * This function is used to add a new aperture range to an existing
1471 * aperture in case of dma_ops domain allocation or address allocation
1474 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1475 bool populate, gfp_t gfp)
1477 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1478 struct amd_iommu *iommu;
1479 unsigned long i, old_size, pte_pgsize;
1481 #ifdef CONFIG_IOMMU_STRESS
1485 if (index >= APERTURE_MAX_RANGES)
1488 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1489 if (!dma_dom->aperture[index])
1492 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1493 if (!dma_dom->aperture[index]->bitmap)
1496 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1499 unsigned long address = dma_dom->aperture_size;
1500 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1501 u64 *pte, *pte_page;
1503 for (i = 0; i < num_ptes; ++i) {
1504 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1509 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1511 address += APERTURE_RANGE_SIZE / 64;
1515 old_size = dma_dom->aperture_size;
1516 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1518 /* Reserve address range used for MSI messages */
1519 if (old_size < MSI_ADDR_BASE_LO &&
1520 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1521 unsigned long spage;
1524 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1525 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1527 dma_ops_reserve_addresses(dma_dom, spage, pages);
1530 /* Initialize the exclusion range if necessary */
1531 for_each_iommu(iommu) {
1532 if (iommu->exclusion_start &&
1533 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1534 && iommu->exclusion_start < dma_dom->aperture_size) {
1535 unsigned long startpage;
1536 int pages = iommu_num_pages(iommu->exclusion_start,
1537 iommu->exclusion_length,
1539 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1540 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1545 * Check for areas already mapped as present in the new aperture
1546 * range and mark those pages as reserved in the allocator. Such
1547 * mappings may already exist as a result of requested unity
1548 * mappings for devices.
1550 for (i = dma_dom->aperture[index]->offset;
1551 i < dma_dom->aperture_size;
1553 u64 *pte = fetch_pte(&dma_dom->domain, i, &pte_pgsize);
1554 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1557 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT,
1561 update_domain(&dma_dom->domain);
1566 update_domain(&dma_dom->domain);
1568 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1570 kfree(dma_dom->aperture[index]);
1571 dma_dom->aperture[index] = NULL;
1576 static unsigned long dma_ops_area_alloc(struct device *dev,
1577 struct dma_ops_domain *dom,
1579 unsigned long align_mask,
1581 unsigned long start)
1583 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1584 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1585 int i = start >> APERTURE_RANGE_SHIFT;
1586 unsigned long boundary_size, mask;
1587 unsigned long address = -1;
1588 unsigned long limit;
1590 next_bit >>= PAGE_SHIFT;
1592 mask = dma_get_seg_boundary(dev);
1594 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
1595 1UL << (BITS_PER_LONG - PAGE_SHIFT);
1597 for (;i < max_index; ++i) {
1598 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1600 if (dom->aperture[i]->offset >= dma_mask)
1603 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1604 dma_mask >> PAGE_SHIFT);
1606 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1607 limit, next_bit, pages, 0,
1608 boundary_size, align_mask);
1609 if (address != -1) {
1610 address = dom->aperture[i]->offset +
1611 (address << PAGE_SHIFT);
1612 dom->next_address = address + (pages << PAGE_SHIFT);
1622 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1623 struct dma_ops_domain *dom,
1625 unsigned long align_mask,
1628 unsigned long address;
1630 #ifdef CONFIG_IOMMU_STRESS
1631 dom->next_address = 0;
1632 dom->need_flush = true;
1635 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1636 dma_mask, dom->next_address);
1638 if (address == -1) {
1639 dom->next_address = 0;
1640 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1642 dom->need_flush = true;
1645 if (unlikely(address == -1))
1646 address = DMA_ERROR_CODE;
1648 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1654 * The address free function.
1656 * called with domain->lock held
1658 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1659 unsigned long address,
1662 unsigned i = address >> APERTURE_RANGE_SHIFT;
1663 struct aperture_range *range = dom->aperture[i];
1665 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1667 #ifdef CONFIG_IOMMU_STRESS
1672 if (address >= dom->next_address)
1673 dom->need_flush = true;
1675 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1677 bitmap_clear(range->bitmap, address, pages);
1681 /****************************************************************************
1683 * The next functions belong to the domain allocation. A domain is
1684 * allocated for every IOMMU as the default domain. If device isolation
1685 * is enabled, every device get its own domain. The most important thing
1686 * about domains is the page table mapping the DMA address space they
1689 ****************************************************************************/
1692 * This function adds a protection domain to the global protection domain list
1694 static void add_domain_to_list(struct protection_domain *domain)
1696 unsigned long flags;
1698 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1699 list_add(&domain->list, &amd_iommu_pd_list);
1700 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1704 * This function removes a protection domain to the global
1705 * protection domain list
1707 static void del_domain_from_list(struct protection_domain *domain)
1709 unsigned long flags;
1711 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1712 list_del(&domain->list);
1713 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1716 static u16 domain_id_alloc(void)
1718 unsigned long flags;
1721 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1722 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1724 if (id > 0 && id < MAX_DOMAIN_ID)
1725 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1728 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1733 static void domain_id_free(int id)
1735 unsigned long flags;
1737 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1738 if (id > 0 && id < MAX_DOMAIN_ID)
1739 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1740 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1743 #define DEFINE_FREE_PT_FN(LVL, FN) \
1744 static void free_pt_##LVL (unsigned long __pt) \
1752 for (i = 0; i < 512; ++i) { \
1753 /* PTE present? */ \
1754 if (!IOMMU_PTE_PRESENT(pt[i])) \
1758 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1759 PM_PTE_LEVEL(pt[i]) == 7) \
1762 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1765 free_page((unsigned long)pt); \
1768 DEFINE_FREE_PT_FN(l2, free_page)
1769 DEFINE_FREE_PT_FN(l3, free_pt_l2)
1770 DEFINE_FREE_PT_FN(l4, free_pt_l3)
1771 DEFINE_FREE_PT_FN(l5, free_pt_l4)
1772 DEFINE_FREE_PT_FN(l6, free_pt_l5)
1774 static void free_pagetable(struct protection_domain *domain)
1776 unsigned long root = (unsigned long)domain->pt_root;
1778 switch (domain->mode) {
1779 case PAGE_MODE_NONE:
1781 case PAGE_MODE_1_LEVEL:
1784 case PAGE_MODE_2_LEVEL:
1787 case PAGE_MODE_3_LEVEL:
1790 case PAGE_MODE_4_LEVEL:
1793 case PAGE_MODE_5_LEVEL:
1796 case PAGE_MODE_6_LEVEL:
1804 static void free_gcr3_tbl_level1(u64 *tbl)
1809 for (i = 0; i < 512; ++i) {
1810 if (!(tbl[i] & GCR3_VALID))
1813 ptr = __va(tbl[i] & PAGE_MASK);
1815 free_page((unsigned long)ptr);
1819 static void free_gcr3_tbl_level2(u64 *tbl)
1824 for (i = 0; i < 512; ++i) {
1825 if (!(tbl[i] & GCR3_VALID))
1828 ptr = __va(tbl[i] & PAGE_MASK);
1830 free_gcr3_tbl_level1(ptr);
1834 static void free_gcr3_table(struct protection_domain *domain)
1836 if (domain->glx == 2)
1837 free_gcr3_tbl_level2(domain->gcr3_tbl);
1838 else if (domain->glx == 1)
1839 free_gcr3_tbl_level1(domain->gcr3_tbl);
1840 else if (domain->glx != 0)
1843 free_page((unsigned long)domain->gcr3_tbl);
1847 * Free a domain, only used if something went wrong in the
1848 * allocation path and we need to free an already allocated page table
1850 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1857 del_domain_from_list(&dom->domain);
1859 free_pagetable(&dom->domain);
1861 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1862 if (!dom->aperture[i])
1864 free_page((unsigned long)dom->aperture[i]->bitmap);
1865 kfree(dom->aperture[i]);
1872 * Allocates a new protection domain usable for the dma_ops functions.
1873 * It also initializes the page table and the address allocator data
1874 * structures required for the dma_ops interface
1876 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1878 struct dma_ops_domain *dma_dom;
1880 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1884 spin_lock_init(&dma_dom->domain.lock);
1886 dma_dom->domain.id = domain_id_alloc();
1887 if (dma_dom->domain.id == 0)
1889 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1890 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1891 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1892 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1893 dma_dom->domain.priv = dma_dom;
1894 if (!dma_dom->domain.pt_root)
1897 dma_dom->need_flush = false;
1899 add_domain_to_list(&dma_dom->domain);
1901 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1905 * mark the first page as allocated so we never return 0 as
1906 * a valid dma-address. So we can use 0 as error value
1908 dma_dom->aperture[0]->bitmap[0] = 1;
1909 dma_dom->next_address = 0;
1915 dma_ops_domain_free(dma_dom);
1921 * little helper function to check whether a given protection domain is a
1924 static bool dma_ops_domain(struct protection_domain *domain)
1926 return domain->flags & PD_DMA_OPS_MASK;
1929 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1934 if (domain->mode != PAGE_MODE_NONE)
1935 pte_root = virt_to_phys(domain->pt_root);
1937 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1938 << DEV_ENTRY_MODE_SHIFT;
1939 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1941 flags = amd_iommu_dev_table[devid].data[1];
1944 flags |= DTE_FLAG_IOTLB;
1946 if (domain->flags & PD_IOMMUV2_MASK) {
1947 u64 gcr3 = __pa(domain->gcr3_tbl);
1948 u64 glx = domain->glx;
1951 pte_root |= DTE_FLAG_GV;
1952 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1954 /* First mask out possible old values for GCR3 table */
1955 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1958 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1961 /* Encode GCR3 table into DTE */
1962 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1965 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1968 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1972 flags &= ~(0xffffUL);
1973 flags |= domain->id;
1975 amd_iommu_dev_table[devid].data[1] = flags;
1976 amd_iommu_dev_table[devid].data[0] = pte_root;
1979 static void clear_dte_entry(u16 devid)
1981 /* remove entry from the device table seen by the hardware */
1982 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1983 amd_iommu_dev_table[devid].data[1] = 0;
1985 amd_iommu_apply_erratum_63(devid);
1988 static void do_attach(struct iommu_dev_data *dev_data,
1989 struct protection_domain *domain)
1991 struct amd_iommu *iommu;
1994 iommu = amd_iommu_rlookup_table[dev_data->devid];
1995 ats = dev_data->ats.enabled;
1997 /* Update data structures */
1998 dev_data->domain = domain;
1999 list_add(&dev_data->list, &domain->dev_list);
2000 set_dte_entry(dev_data->devid, domain, ats);
2002 /* Do reference counting */
2003 domain->dev_iommu[iommu->index] += 1;
2004 domain->dev_cnt += 1;
2006 /* Flush the DTE entry */
2007 device_flush_dte(dev_data);
2010 static void do_detach(struct iommu_dev_data *dev_data)
2012 struct amd_iommu *iommu;
2014 iommu = amd_iommu_rlookup_table[dev_data->devid];
2016 /* decrease reference counters */
2017 dev_data->domain->dev_iommu[iommu->index] -= 1;
2018 dev_data->domain->dev_cnt -= 1;
2020 /* Update data structures */
2021 dev_data->domain = NULL;
2022 list_del(&dev_data->list);
2023 clear_dte_entry(dev_data->devid);
2025 /* Flush the DTE entry */
2026 device_flush_dte(dev_data);
2030 * If a device is not yet associated with a domain, this function does
2031 * assigns it visible for the hardware
2033 static int __attach_device(struct iommu_dev_data *dev_data,
2034 struct protection_domain *domain)
2036 struct iommu_dev_data *head, *entry;
2040 spin_lock(&domain->lock);
2044 if (head->alias_data != NULL)
2045 head = head->alias_data;
2047 /* Now we have the root of the alias group, if any */
2050 if (head->domain != NULL)
2053 /* Attach alias group root */
2054 do_attach(head, domain);
2056 /* Attach other devices in the alias group */
2057 list_for_each_entry(entry, &head->alias_list, alias_list)
2058 do_attach(entry, domain);
2065 spin_unlock(&domain->lock);
2071 static void pdev_iommuv2_disable(struct pci_dev *pdev)
2073 pci_disable_ats(pdev);
2074 pci_disable_pri(pdev);
2075 pci_disable_pasid(pdev);
2078 /* FIXME: Change generic reset-function to do the same */
2079 static int pri_reset_while_enabled(struct pci_dev *pdev)
2084 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2088 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2089 control |= PCI_PRI_CTRL_RESET;
2090 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
2095 static int pdev_iommuv2_enable(struct pci_dev *pdev)
2100 /* FIXME: Hardcode number of outstanding requests for now */
2102 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2104 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2106 /* Only allow access to user-accessible pages */
2107 ret = pci_enable_pasid(pdev, 0);
2111 /* First reset the PRI state of the device */
2112 ret = pci_reset_pri(pdev);
2117 ret = pci_enable_pri(pdev, reqs);
2122 ret = pri_reset_while_enabled(pdev);
2127 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2134 pci_disable_pri(pdev);
2135 pci_disable_pasid(pdev);
2140 /* FIXME: Move this to PCI code */
2141 #define PCI_PRI_TLP_OFF (1 << 15)
2143 static bool pci_pri_tlp_required(struct pci_dev *pdev)
2148 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2152 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2154 return (status & PCI_PRI_TLP_OFF) ? true : false;
2158 * If a device is not yet associated with a domain, this function
2159 * assigns it visible for the hardware
2161 static int attach_device(struct device *dev,
2162 struct protection_domain *domain)
2164 struct pci_dev *pdev = to_pci_dev(dev);
2165 struct iommu_dev_data *dev_data;
2166 unsigned long flags;
2169 dev_data = get_dev_data(dev);
2171 if (domain->flags & PD_IOMMUV2_MASK) {
2172 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2175 if (pdev_iommuv2_enable(pdev) != 0)
2178 dev_data->ats.enabled = true;
2179 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2180 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2181 } else if (amd_iommu_iotlb_sup &&
2182 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2183 dev_data->ats.enabled = true;
2184 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2187 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2188 ret = __attach_device(dev_data, domain);
2189 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2192 * We might boot into a crash-kernel here. The crashed kernel
2193 * left the caches in the IOMMU dirty. So we have to flush
2194 * here to evict all dirty stuff.
2196 domain_flush_tlb_pde(domain);
2202 * Removes a device from a protection domain (unlocked)
2204 static void __detach_device(struct iommu_dev_data *dev_data)
2206 struct iommu_dev_data *head, *entry;
2207 struct protection_domain *domain;
2208 unsigned long flags;
2210 BUG_ON(!dev_data->domain);
2212 domain = dev_data->domain;
2214 spin_lock_irqsave(&domain->lock, flags);
2217 if (head->alias_data != NULL)
2218 head = head->alias_data;
2220 list_for_each_entry(entry, &head->alias_list, alias_list)
2225 spin_unlock_irqrestore(&domain->lock, flags);
2228 * If we run in passthrough mode the device must be assigned to the
2229 * passthrough domain if it is detached from any other domain.
2230 * Make sure we can deassign from the pt_domain itself.
2232 if (dev_data->passthrough &&
2233 (dev_data->domain == NULL && domain != pt_domain))
2234 __attach_device(dev_data, pt_domain);
2238 * Removes a device from a protection domain (with devtable_lock held)
2240 static void detach_device(struct device *dev)
2242 struct protection_domain *domain;
2243 struct iommu_dev_data *dev_data;
2244 unsigned long flags;
2246 dev_data = get_dev_data(dev);
2247 domain = dev_data->domain;
2249 /* lock device table */
2250 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2251 __detach_device(dev_data);
2252 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2254 if (domain->flags & PD_IOMMUV2_MASK)
2255 pdev_iommuv2_disable(to_pci_dev(dev));
2256 else if (dev_data->ats.enabled)
2257 pci_disable_ats(to_pci_dev(dev));
2259 dev_data->ats.enabled = false;
2262 static int amd_iommu_add_device(struct device *dev)
2264 struct iommu_dev_data *dev_data;
2265 struct iommu_domain *domain;
2266 struct amd_iommu *iommu;
2270 if (!check_device(dev) || get_dev_data(dev))
2273 devid = get_device_id(dev);
2274 iommu = amd_iommu_rlookup_table[devid];
2276 ret = iommu_init_device(dev);
2278 if (ret != -ENOTSUPP)
2279 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2282 iommu_ignore_device(dev);
2283 dev->archdata.dma_ops = &nommu_dma_ops;
2286 init_iommu_group(dev);
2288 dev_data = get_dev_data(dev);
2292 if (dev_data->iommu_v2)
2293 iommu_request_dm_for_dev(dev);
2295 /* Domains are initialized for this device - have a look what we ended up with */
2296 domain = iommu_get_domain_for_dev(dev);
2297 if (domain->type == IOMMU_DOMAIN_IDENTITY) {
2298 dev_data->passthrough = true;
2299 dev->archdata.dma_ops = &nommu_dma_ops;
2301 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2305 iommu_completion_wait(iommu);
2310 static void amd_iommu_remove_device(struct device *dev)
2312 struct amd_iommu *iommu;
2315 if (!check_device(dev))
2318 devid = get_device_id(dev);
2319 iommu = amd_iommu_rlookup_table[devid];
2321 iommu_uninit_device(dev);
2322 iommu_completion_wait(iommu);
2325 /*****************************************************************************
2327 * The next functions belong to the dma_ops mapping/unmapping code.
2329 *****************************************************************************/
2332 * In the dma_ops path we only have the struct device. This function
2333 * finds the corresponding IOMMU, the protection domain and the
2334 * requestor id for a given device.
2335 * If the device is not yet associated with a domain this is also done
2338 static struct protection_domain *get_domain(struct device *dev)
2340 struct protection_domain *domain;
2341 struct iommu_domain *io_domain;
2343 if (!check_device(dev))
2344 return ERR_PTR(-EINVAL);
2346 io_domain = iommu_get_domain_for_dev(dev);
2350 domain = to_pdomain(io_domain);
2351 if (!dma_ops_domain(domain))
2352 return ERR_PTR(-EBUSY);
2357 static void update_device_table(struct protection_domain *domain)
2359 struct iommu_dev_data *dev_data;
2361 list_for_each_entry(dev_data, &domain->dev_list, list)
2362 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2365 static void update_domain(struct protection_domain *domain)
2367 if (!domain->updated)
2370 update_device_table(domain);
2372 domain_flush_devices(domain);
2373 domain_flush_tlb_pde(domain);
2375 domain->updated = false;
2379 * This function fetches the PTE for a given address in the aperture
2381 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2382 unsigned long address)
2384 struct aperture_range *aperture;
2385 u64 *pte, *pte_page;
2387 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2391 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2393 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2395 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2397 pte += PM_LEVEL_INDEX(0, address);
2399 update_domain(&dom->domain);
2405 * This is the generic map function. It maps one 4kb page at paddr to
2406 * the given address in the DMA address space for the domain.
2408 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2409 unsigned long address,
2415 WARN_ON(address > dom->aperture_size);
2419 pte = dma_ops_get_pte(dom, address);
2421 return DMA_ERROR_CODE;
2423 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2425 if (direction == DMA_TO_DEVICE)
2426 __pte |= IOMMU_PTE_IR;
2427 else if (direction == DMA_FROM_DEVICE)
2428 __pte |= IOMMU_PTE_IW;
2429 else if (direction == DMA_BIDIRECTIONAL)
2430 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2436 return (dma_addr_t)address;
2440 * The generic unmapping function for on page in the DMA address space.
2442 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2443 unsigned long address)
2445 struct aperture_range *aperture;
2448 if (address >= dom->aperture_size)
2451 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2455 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2459 pte += PM_LEVEL_INDEX(0, address);
2467 * This function contains common code for mapping of a physically
2468 * contiguous memory region into DMA address space. It is used by all
2469 * mapping functions provided with this IOMMU driver.
2470 * Must be called with the domain lock held.
2472 static dma_addr_t __map_single(struct device *dev,
2473 struct dma_ops_domain *dma_dom,
2480 dma_addr_t offset = paddr & ~PAGE_MASK;
2481 dma_addr_t address, start, ret;
2483 unsigned long align_mask = 0;
2486 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2489 INC_STATS_COUNTER(total_map_requests);
2492 INC_STATS_COUNTER(cross_page);
2495 align_mask = (1UL << get_order(size)) - 1;
2498 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2500 if (unlikely(address == DMA_ERROR_CODE)) {
2502 * setting next_address here will let the address
2503 * allocator only scan the new allocated range in the
2504 * first run. This is a small optimization.
2506 dma_dom->next_address = dma_dom->aperture_size;
2508 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2512 * aperture was successfully enlarged by 128 MB, try
2519 for (i = 0; i < pages; ++i) {
2520 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2521 if (ret == DMA_ERROR_CODE)
2529 ADD_STATS_COUNTER(alloced_io_mem, size);
2531 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2532 domain_flush_tlb(&dma_dom->domain);
2533 dma_dom->need_flush = false;
2534 } else if (unlikely(amd_iommu_np_cache))
2535 domain_flush_pages(&dma_dom->domain, address, size);
2542 for (--i; i >= 0; --i) {
2544 dma_ops_domain_unmap(dma_dom, start);
2547 dma_ops_free_addresses(dma_dom, address, pages);
2549 return DMA_ERROR_CODE;
2553 * Does the reverse of the __map_single function. Must be called with
2554 * the domain lock held too
2556 static void __unmap_single(struct dma_ops_domain *dma_dom,
2557 dma_addr_t dma_addr,
2561 dma_addr_t flush_addr;
2562 dma_addr_t i, start;
2565 if ((dma_addr == DMA_ERROR_CODE) ||
2566 (dma_addr + size > dma_dom->aperture_size))
2569 flush_addr = dma_addr;
2570 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2571 dma_addr &= PAGE_MASK;
2574 for (i = 0; i < pages; ++i) {
2575 dma_ops_domain_unmap(dma_dom, start);
2579 SUB_STATS_COUNTER(alloced_io_mem, size);
2581 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2583 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2584 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2585 dma_dom->need_flush = false;
2590 * The exported map_single function for dma_ops.
2592 static dma_addr_t map_page(struct device *dev, struct page *page,
2593 unsigned long offset, size_t size,
2594 enum dma_data_direction dir,
2595 struct dma_attrs *attrs)
2597 unsigned long flags;
2598 struct protection_domain *domain;
2601 phys_addr_t paddr = page_to_phys(page) + offset;
2603 INC_STATS_COUNTER(cnt_map_single);
2605 domain = get_domain(dev);
2606 if (PTR_ERR(domain) == -EINVAL)
2607 return (dma_addr_t)paddr;
2608 else if (IS_ERR(domain))
2609 return DMA_ERROR_CODE;
2611 dma_mask = *dev->dma_mask;
2613 spin_lock_irqsave(&domain->lock, flags);
2615 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2617 if (addr == DMA_ERROR_CODE)
2620 domain_flush_complete(domain);
2623 spin_unlock_irqrestore(&domain->lock, flags);
2629 * The exported unmap_single function for dma_ops.
2631 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2632 enum dma_data_direction dir, struct dma_attrs *attrs)
2634 unsigned long flags;
2635 struct protection_domain *domain;
2637 INC_STATS_COUNTER(cnt_unmap_single);
2639 domain = get_domain(dev);
2643 spin_lock_irqsave(&domain->lock, flags);
2645 __unmap_single(domain->priv, dma_addr, size, dir);
2647 domain_flush_complete(domain);
2649 spin_unlock_irqrestore(&domain->lock, flags);
2653 * The exported map_sg function for dma_ops (handles scatter-gather
2656 static int map_sg(struct device *dev, struct scatterlist *sglist,
2657 int nelems, enum dma_data_direction dir,
2658 struct dma_attrs *attrs)
2660 unsigned long flags;
2661 struct protection_domain *domain;
2663 struct scatterlist *s;
2665 int mapped_elems = 0;
2668 INC_STATS_COUNTER(cnt_map_sg);
2670 domain = get_domain(dev);
2674 dma_mask = *dev->dma_mask;
2676 spin_lock_irqsave(&domain->lock, flags);
2678 for_each_sg(sglist, s, nelems, i) {
2681 s->dma_address = __map_single(dev, domain->priv,
2682 paddr, s->length, dir, false,
2685 if (s->dma_address) {
2686 s->dma_length = s->length;
2692 domain_flush_complete(domain);
2695 spin_unlock_irqrestore(&domain->lock, flags);
2697 return mapped_elems;
2699 for_each_sg(sglist, s, mapped_elems, i) {
2701 __unmap_single(domain->priv, s->dma_address,
2702 s->dma_length, dir);
2703 s->dma_address = s->dma_length = 0;
2712 * The exported map_sg function for dma_ops (handles scatter-gather
2715 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2716 int nelems, enum dma_data_direction dir,
2717 struct dma_attrs *attrs)
2719 unsigned long flags;
2720 struct protection_domain *domain;
2721 struct scatterlist *s;
2724 INC_STATS_COUNTER(cnt_unmap_sg);
2726 domain = get_domain(dev);
2730 spin_lock_irqsave(&domain->lock, flags);
2732 for_each_sg(sglist, s, nelems, i) {
2733 __unmap_single(domain->priv, s->dma_address,
2734 s->dma_length, dir);
2735 s->dma_address = s->dma_length = 0;
2738 domain_flush_complete(domain);
2740 spin_unlock_irqrestore(&domain->lock, flags);
2744 * The exported alloc_coherent function for dma_ops.
2746 static void *alloc_coherent(struct device *dev, size_t size,
2747 dma_addr_t *dma_addr, gfp_t flag,
2748 struct dma_attrs *attrs)
2750 u64 dma_mask = dev->coherent_dma_mask;
2751 struct protection_domain *domain;
2752 unsigned long flags;
2755 INC_STATS_COUNTER(cnt_alloc_coherent);
2757 domain = get_domain(dev);
2758 if (PTR_ERR(domain) == -EINVAL) {
2759 page = alloc_pages(flag, get_order(size));
2760 *dma_addr = page_to_phys(page);
2761 return page_address(page);
2762 } else if (IS_ERR(domain))
2765 size = PAGE_ALIGN(size);
2766 dma_mask = dev->coherent_dma_mask;
2767 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2770 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2772 if (!(flag & __GFP_WAIT))
2775 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2782 dma_mask = *dev->dma_mask;
2784 spin_lock_irqsave(&domain->lock, flags);
2786 *dma_addr = __map_single(dev, domain->priv, page_to_phys(page),
2787 size, DMA_BIDIRECTIONAL, true, dma_mask);
2789 if (*dma_addr == DMA_ERROR_CODE) {
2790 spin_unlock_irqrestore(&domain->lock, flags);
2794 domain_flush_complete(domain);
2796 spin_unlock_irqrestore(&domain->lock, flags);
2798 return page_address(page);
2802 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2803 __free_pages(page, get_order(size));
2809 * The exported free_coherent function for dma_ops.
2811 static void free_coherent(struct device *dev, size_t size,
2812 void *virt_addr, dma_addr_t dma_addr,
2813 struct dma_attrs *attrs)
2815 struct protection_domain *domain;
2816 unsigned long flags;
2819 INC_STATS_COUNTER(cnt_free_coherent);
2821 page = virt_to_page(virt_addr);
2822 size = PAGE_ALIGN(size);
2824 domain = get_domain(dev);
2828 spin_lock_irqsave(&domain->lock, flags);
2830 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2832 domain_flush_complete(domain);
2834 spin_unlock_irqrestore(&domain->lock, flags);
2837 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2838 __free_pages(page, get_order(size));
2842 * This function is called by the DMA layer to find out if we can handle a
2843 * particular device. It is part of the dma_ops.
2845 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2847 return check_device(dev);
2850 static struct dma_map_ops amd_iommu_dma_ops = {
2851 .alloc = alloc_coherent,
2852 .free = free_coherent,
2853 .map_page = map_page,
2854 .unmap_page = unmap_page,
2856 .unmap_sg = unmap_sg,
2857 .dma_supported = amd_iommu_dma_supported,
2860 int __init amd_iommu_init_api(void)
2862 return bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2865 int __init amd_iommu_init_dma_ops(void)
2870 amd_iommu_stats_init();
2872 if (amd_iommu_unmap_flush)
2873 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2875 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2880 /*****************************************************************************
2882 * The following functions belong to the exported interface of AMD IOMMU
2884 * This interface allows access to lower level functions of the IOMMU
2885 * like protection domain handling and assignement of devices to domains
2886 * which is not possible with the dma_ops interface.
2888 *****************************************************************************/
2890 static void cleanup_domain(struct protection_domain *domain)
2892 struct iommu_dev_data *entry;
2893 unsigned long flags;
2895 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2897 while (!list_empty(&domain->dev_list)) {
2898 entry = list_first_entry(&domain->dev_list,
2899 struct iommu_dev_data, list);
2900 __detach_device(entry);
2903 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2906 static void protection_domain_free(struct protection_domain *domain)
2911 del_domain_from_list(domain);
2914 domain_id_free(domain->id);
2919 static struct protection_domain *protection_domain_alloc(void)
2921 struct protection_domain *domain;
2923 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2927 spin_lock_init(&domain->lock);
2928 mutex_init(&domain->api_lock);
2929 domain->id = domain_id_alloc();
2932 INIT_LIST_HEAD(&domain->dev_list);
2934 add_domain_to_list(domain);
2944 static int alloc_passthrough_domain(void)
2946 if (pt_domain != NULL)
2949 /* allocate passthrough domain */
2950 pt_domain = protection_domain_alloc();
2954 pt_domain->mode = PAGE_MODE_NONE;
2959 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2961 struct protection_domain *pdomain;
2962 struct dma_ops_domain *dma_domain;
2965 case IOMMU_DOMAIN_UNMANAGED:
2966 pdomain = protection_domain_alloc();
2970 pdomain->mode = PAGE_MODE_3_LEVEL;
2971 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2972 if (!pdomain->pt_root) {
2973 protection_domain_free(pdomain);
2977 pdomain->domain.geometry.aperture_start = 0;
2978 pdomain->domain.geometry.aperture_end = ~0ULL;
2979 pdomain->domain.geometry.force_aperture = true;
2982 case IOMMU_DOMAIN_DMA:
2983 dma_domain = dma_ops_domain_alloc();
2985 pr_err("AMD-Vi: Failed to allocate\n");
2988 pdomain = &dma_domain->domain;
2990 case IOMMU_DOMAIN_IDENTITY:
2991 pdomain = protection_domain_alloc();
2995 pdomain->mode = PAGE_MODE_NONE;
3001 return &pdomain->domain;
3004 static void amd_iommu_domain_free(struct iommu_domain *dom)
3006 struct protection_domain *domain;
3011 domain = to_pdomain(dom);
3013 if (domain->dev_cnt > 0)
3014 cleanup_domain(domain);
3016 BUG_ON(domain->dev_cnt != 0);
3018 if (domain->mode != PAGE_MODE_NONE)
3019 free_pagetable(domain);
3021 if (domain->flags & PD_IOMMUV2_MASK)
3022 free_gcr3_table(domain);
3024 protection_domain_free(domain);
3027 static void amd_iommu_detach_device(struct iommu_domain *dom,
3030 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3031 struct amd_iommu *iommu;
3034 if (!check_device(dev))
3037 devid = get_device_id(dev);
3039 if (dev_data->domain != NULL)
3042 iommu = amd_iommu_rlookup_table[devid];
3046 iommu_completion_wait(iommu);
3049 static int amd_iommu_attach_device(struct iommu_domain *dom,
3052 struct protection_domain *domain = to_pdomain(dom);
3053 struct iommu_dev_data *dev_data;
3054 struct amd_iommu *iommu;
3057 if (!check_device(dev))
3060 dev_data = dev->archdata.iommu;
3062 iommu = amd_iommu_rlookup_table[dev_data->devid];
3066 if (dev_data->domain)
3069 ret = attach_device(dev, domain);
3071 iommu_completion_wait(iommu);
3076 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3077 phys_addr_t paddr, size_t page_size, int iommu_prot)
3079 struct protection_domain *domain = to_pdomain(dom);
3083 if (domain->mode == PAGE_MODE_NONE)
3086 if (iommu_prot & IOMMU_READ)
3087 prot |= IOMMU_PROT_IR;
3088 if (iommu_prot & IOMMU_WRITE)
3089 prot |= IOMMU_PROT_IW;
3091 mutex_lock(&domain->api_lock);
3092 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3093 mutex_unlock(&domain->api_lock);
3098 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3101 struct protection_domain *domain = to_pdomain(dom);
3104 if (domain->mode == PAGE_MODE_NONE)
3107 mutex_lock(&domain->api_lock);
3108 unmap_size = iommu_unmap_page(domain, iova, page_size);
3109 mutex_unlock(&domain->api_lock);
3111 domain_flush_tlb_pde(domain);
3116 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3119 struct protection_domain *domain = to_pdomain(dom);
3120 unsigned long offset_mask, pte_pgsize;
3123 if (domain->mode == PAGE_MODE_NONE)
3126 pte = fetch_pte(domain, iova, &pte_pgsize);
3128 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3131 offset_mask = pte_pgsize - 1;
3132 __pte = *pte & PM_ADDR_MASK;
3134 return (__pte & ~offset_mask) | (iova & offset_mask);
3137 static bool amd_iommu_capable(enum iommu_cap cap)
3140 case IOMMU_CAP_CACHE_COHERENCY:
3142 case IOMMU_CAP_INTR_REMAP:
3143 return (irq_remapping_enabled == 1);
3144 case IOMMU_CAP_NOEXEC:
3151 static void amd_iommu_get_dm_regions(struct device *dev,
3152 struct list_head *head)
3154 struct unity_map_entry *entry;
3157 devid = get_device_id(dev);
3159 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3160 struct iommu_dm_region *region;
3162 if (devid < entry->devid_start || devid > entry->devid_end)
3165 region = kzalloc(sizeof(*region), GFP_KERNEL);
3167 pr_err("Out of memory allocating dm-regions for %s\n",
3172 region->start = entry->address_start;
3173 region->length = entry->address_end - entry->address_start;
3174 if (entry->prot & IOMMU_PROT_IR)
3175 region->prot |= IOMMU_READ;
3176 if (entry->prot & IOMMU_PROT_IW)
3177 region->prot |= IOMMU_WRITE;
3179 list_add_tail(®ion->list, head);
3183 static void amd_iommu_put_dm_regions(struct device *dev,
3184 struct list_head *head)
3186 struct iommu_dm_region *entry, *next;
3188 list_for_each_entry_safe(entry, next, head, list)
3192 static const struct iommu_ops amd_iommu_ops = {
3193 .capable = amd_iommu_capable,
3194 .domain_alloc = amd_iommu_domain_alloc,
3195 .domain_free = amd_iommu_domain_free,
3196 .attach_dev = amd_iommu_attach_device,
3197 .detach_dev = amd_iommu_detach_device,
3198 .map = amd_iommu_map,
3199 .unmap = amd_iommu_unmap,
3200 .map_sg = default_iommu_map_sg,
3201 .iova_to_phys = amd_iommu_iova_to_phys,
3202 .add_device = amd_iommu_add_device,
3203 .remove_device = amd_iommu_remove_device,
3204 .get_dm_regions = amd_iommu_get_dm_regions,
3205 .put_dm_regions = amd_iommu_put_dm_regions,
3206 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3209 /*****************************************************************************
3211 * The next functions do a basic initialization of IOMMU for pass through
3214 * In passthrough mode the IOMMU is initialized and enabled but not used for
3215 * DMA-API translation.
3217 *****************************************************************************/
3219 int __init amd_iommu_init_passthrough(void)
3221 struct iommu_dev_data *dev_data;
3222 struct pci_dev *dev = NULL;
3225 ret = alloc_passthrough_domain();
3229 for_each_pci_dev(dev) {
3230 if (!check_device(&dev->dev))
3233 dev_data = get_dev_data(&dev->dev);
3234 dev_data->passthrough = true;
3236 attach_device(&dev->dev, pt_domain);
3239 amd_iommu_stats_init();
3241 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3246 /* IOMMUv2 specific functions */
3247 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3249 return atomic_notifier_chain_register(&ppr_notifier, nb);
3251 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3253 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3255 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3257 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3259 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3261 struct protection_domain *domain = to_pdomain(dom);
3262 unsigned long flags;
3264 spin_lock_irqsave(&domain->lock, flags);
3266 /* Update data structure */
3267 domain->mode = PAGE_MODE_NONE;
3268 domain->updated = true;
3270 /* Make changes visible to IOMMUs */
3271 update_domain(domain);
3273 /* Page-table is not visible to IOMMU anymore, so free it */
3274 free_pagetable(domain);
3276 spin_unlock_irqrestore(&domain->lock, flags);
3278 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3280 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3282 struct protection_domain *domain = to_pdomain(dom);
3283 unsigned long flags;
3286 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3289 /* Number of GCR3 table levels required */
3290 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3293 if (levels > amd_iommu_max_glx_val)
3296 spin_lock_irqsave(&domain->lock, flags);
3299 * Save us all sanity checks whether devices already in the
3300 * domain support IOMMUv2. Just force that the domain has no
3301 * devices attached when it is switched into IOMMUv2 mode.
3304 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3308 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3309 if (domain->gcr3_tbl == NULL)
3312 domain->glx = levels;
3313 domain->flags |= PD_IOMMUV2_MASK;
3314 domain->updated = true;
3316 update_domain(domain);
3321 spin_unlock_irqrestore(&domain->lock, flags);
3325 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3327 static int __flush_pasid(struct protection_domain *domain, int pasid,
3328 u64 address, bool size)
3330 struct iommu_dev_data *dev_data;
3331 struct iommu_cmd cmd;
3334 if (!(domain->flags & PD_IOMMUV2_MASK))
3337 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3340 * IOMMU TLB needs to be flushed before Device TLB to
3341 * prevent device TLB refill from IOMMU TLB
3343 for (i = 0; i < amd_iommus_present; ++i) {
3344 if (domain->dev_iommu[i] == 0)
3347 ret = iommu_queue_command(amd_iommus[i], &cmd);
3352 /* Wait until IOMMU TLB flushes are complete */
3353 domain_flush_complete(domain);
3355 /* Now flush device TLBs */
3356 list_for_each_entry(dev_data, &domain->dev_list, list) {
3357 struct amd_iommu *iommu;
3360 BUG_ON(!dev_data->ats.enabled);
3362 qdep = dev_data->ats.qdep;
3363 iommu = amd_iommu_rlookup_table[dev_data->devid];
3365 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3366 qdep, address, size);
3368 ret = iommu_queue_command(iommu, &cmd);
3373 /* Wait until all device TLBs are flushed */
3374 domain_flush_complete(domain);
3383 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3386 INC_STATS_COUNTER(invalidate_iotlb);
3388 return __flush_pasid(domain, pasid, address, false);
3391 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3394 struct protection_domain *domain = to_pdomain(dom);
3395 unsigned long flags;
3398 spin_lock_irqsave(&domain->lock, flags);
3399 ret = __amd_iommu_flush_page(domain, pasid, address);
3400 spin_unlock_irqrestore(&domain->lock, flags);
3404 EXPORT_SYMBOL(amd_iommu_flush_page);
3406 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3408 INC_STATS_COUNTER(invalidate_iotlb_all);
3410 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3414 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3416 struct protection_domain *domain = to_pdomain(dom);
3417 unsigned long flags;
3420 spin_lock_irqsave(&domain->lock, flags);
3421 ret = __amd_iommu_flush_tlb(domain, pasid);
3422 spin_unlock_irqrestore(&domain->lock, flags);
3426 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3428 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3435 index = (pasid >> (9 * level)) & 0x1ff;
3441 if (!(*pte & GCR3_VALID)) {
3445 root = (void *)get_zeroed_page(GFP_ATOMIC);
3449 *pte = __pa(root) | GCR3_VALID;
3452 root = __va(*pte & PAGE_MASK);
3460 static int __set_gcr3(struct protection_domain *domain, int pasid,
3465 if (domain->mode != PAGE_MODE_NONE)
3468 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3472 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3474 return __amd_iommu_flush_tlb(domain, pasid);
3477 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3481 if (domain->mode != PAGE_MODE_NONE)
3484 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3490 return __amd_iommu_flush_tlb(domain, pasid);
3493 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3496 struct protection_domain *domain = to_pdomain(dom);
3497 unsigned long flags;
3500 spin_lock_irqsave(&domain->lock, flags);
3501 ret = __set_gcr3(domain, pasid, cr3);
3502 spin_unlock_irqrestore(&domain->lock, flags);
3506 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3508 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3510 struct protection_domain *domain = to_pdomain(dom);
3511 unsigned long flags;
3514 spin_lock_irqsave(&domain->lock, flags);
3515 ret = __clear_gcr3(domain, pasid);
3516 spin_unlock_irqrestore(&domain->lock, flags);
3520 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3522 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3523 int status, int tag)
3525 struct iommu_dev_data *dev_data;
3526 struct amd_iommu *iommu;
3527 struct iommu_cmd cmd;
3529 INC_STATS_COUNTER(complete_ppr);
3531 dev_data = get_dev_data(&pdev->dev);
3532 iommu = amd_iommu_rlookup_table[dev_data->devid];
3534 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3535 tag, dev_data->pri_tlp);
3537 return iommu_queue_command(iommu, &cmd);
3539 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3541 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3543 struct protection_domain *pdomain;
3545 pdomain = get_domain(&pdev->dev);
3546 if (IS_ERR(pdomain))
3549 /* Only return IOMMUv2 domains */
3550 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3553 return &pdomain->domain;
3555 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3557 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3559 struct iommu_dev_data *dev_data;
3561 if (!amd_iommu_v2_supported())
3564 dev_data = get_dev_data(&pdev->dev);
3565 dev_data->errata |= (1 << erratum);
3567 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3569 int amd_iommu_device_info(struct pci_dev *pdev,
3570 struct amd_iommu_device_info *info)
3575 if (pdev == NULL || info == NULL)
3578 if (!amd_iommu_v2_supported())
3581 memset(info, 0, sizeof(*info));
3583 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3585 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3587 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3589 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3591 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3595 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3596 max_pasids = min(max_pasids, (1 << 20));
3598 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3599 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3601 features = pci_pasid_features(pdev);
3602 if (features & PCI_PASID_CAP_EXEC)
3603 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3604 if (features & PCI_PASID_CAP_PRIV)
3605 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3610 EXPORT_SYMBOL(amd_iommu_device_info);
3612 #ifdef CONFIG_IRQ_REMAP
3614 /*****************************************************************************
3616 * Interrupt Remapping Implementation
3618 *****************************************************************************/
3636 u16 devid; /* Device ID for IRTE table */
3637 u16 index; /* Index into IRTE table*/
3640 struct amd_ir_data {
3641 struct irq_2_irte irq_2_irte;
3642 union irte irte_entry;
3644 struct msi_msg msi_entry;
3648 static struct irq_chip amd_ir_chip;
3650 #define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3651 #define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3652 #define DTE_IRQ_TABLE_LEN (8ULL << 1)
3653 #define DTE_IRQ_REMAP_ENABLE 1ULL
3655 static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3659 dte = amd_iommu_dev_table[devid].data[2];
3660 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3661 dte |= virt_to_phys(table->table);
3662 dte |= DTE_IRQ_REMAP_INTCTL;
3663 dte |= DTE_IRQ_TABLE_LEN;
3664 dte |= DTE_IRQ_REMAP_ENABLE;
3666 amd_iommu_dev_table[devid].data[2] = dte;
3669 #define IRTE_ALLOCATED (~1U)
3671 static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3673 struct irq_remap_table *table = NULL;
3674 struct amd_iommu *iommu;
3675 unsigned long flags;
3678 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3680 iommu = amd_iommu_rlookup_table[devid];
3684 table = irq_lookup_table[devid];
3688 alias = amd_iommu_alias_table[devid];
3689 table = irq_lookup_table[alias];
3691 irq_lookup_table[devid] = table;
3692 set_dte_irq_entry(devid, table);
3693 iommu_flush_dte(iommu, devid);
3697 /* Nothing there yet, allocate new irq remapping table */
3698 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3702 /* Initialize table spin-lock */
3703 spin_lock_init(&table->lock);
3706 /* Keep the first 32 indexes free for IOAPIC interrupts */
3707 table->min_index = 32;
3709 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3710 if (!table->table) {
3716 memset(table->table, 0, MAX_IRQS_PER_TABLE * sizeof(u32));
3721 for (i = 0; i < 32; ++i)
3722 table->table[i] = IRTE_ALLOCATED;
3725 irq_lookup_table[devid] = table;
3726 set_dte_irq_entry(devid, table);
3727 iommu_flush_dte(iommu, devid);
3728 if (devid != alias) {
3729 irq_lookup_table[alias] = table;
3730 set_dte_irq_entry(alias, table);
3731 iommu_flush_dte(iommu, alias);
3735 iommu_completion_wait(iommu);
3738 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3743 static int alloc_irq_index(u16 devid, int count)
3745 struct irq_remap_table *table;
3746 unsigned long flags;
3749 table = get_irq_table(devid, false);
3753 spin_lock_irqsave(&table->lock, flags);
3755 /* Scan table for free entries */
3756 for (c = 0, index = table->min_index;
3757 index < MAX_IRQS_PER_TABLE;
3759 if (table->table[index] == 0)
3766 table->table[index - c + 1] = IRTE_ALLOCATED;
3776 spin_unlock_irqrestore(&table->lock, flags);
3781 static int modify_irte(u16 devid, int index, union irte irte)
3783 struct irq_remap_table *table;
3784 struct amd_iommu *iommu;
3785 unsigned long flags;
3787 iommu = amd_iommu_rlookup_table[devid];
3791 table = get_irq_table(devid, false);
3795 spin_lock_irqsave(&table->lock, flags);
3796 table->table[index] = irte.val;
3797 spin_unlock_irqrestore(&table->lock, flags);
3799 iommu_flush_irt(iommu, devid);
3800 iommu_completion_wait(iommu);
3805 static void free_irte(u16 devid, int index)
3807 struct irq_remap_table *table;
3808 struct amd_iommu *iommu;
3809 unsigned long flags;
3811 iommu = amd_iommu_rlookup_table[devid];
3815 table = get_irq_table(devid, false);
3819 spin_lock_irqsave(&table->lock, flags);
3820 table->table[index] = 0;
3821 spin_unlock_irqrestore(&table->lock, flags);
3823 iommu_flush_irt(iommu, devid);
3824 iommu_completion_wait(iommu);
3827 static int get_devid(struct irq_alloc_info *info)
3831 switch (info->type) {
3832 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3833 devid = get_ioapic_devid(info->ioapic_id);
3835 case X86_IRQ_ALLOC_TYPE_HPET:
3836 devid = get_hpet_devid(info->hpet_id);
3838 case X86_IRQ_ALLOC_TYPE_MSI:
3839 case X86_IRQ_ALLOC_TYPE_MSIX:
3840 devid = get_device_id(&info->msi_dev->dev);
3850 static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3852 struct amd_iommu *iommu;
3858 devid = get_devid(info);
3860 iommu = amd_iommu_rlookup_table[devid];
3862 return iommu->ir_domain;
3868 static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3870 struct amd_iommu *iommu;
3876 switch (info->type) {
3877 case X86_IRQ_ALLOC_TYPE_MSI:
3878 case X86_IRQ_ALLOC_TYPE_MSIX:
3879 devid = get_device_id(&info->msi_dev->dev);
3881 iommu = amd_iommu_rlookup_table[devid];
3883 return iommu->msi_domain;
3893 struct irq_remap_ops amd_iommu_irq_ops = {
3894 .prepare = amd_iommu_prepare,
3895 .enable = amd_iommu_enable,
3896 .disable = amd_iommu_disable,
3897 .reenable = amd_iommu_reenable,
3898 .enable_faulting = amd_iommu_enable_faulting,
3899 .get_ir_irq_domain = get_ir_irq_domain,
3900 .get_irq_domain = get_irq_domain,
3903 static void irq_remapping_prepare_irte(struct amd_ir_data *data,
3904 struct irq_cfg *irq_cfg,
3905 struct irq_alloc_info *info,
3906 int devid, int index, int sub_handle)
3908 struct irq_2_irte *irte_info = &data->irq_2_irte;
3909 struct msi_msg *msg = &data->msi_entry;
3910 union irte *irte = &data->irte_entry;
3911 struct IO_APIC_route_entry *entry;
3913 data->irq_2_irte.devid = devid;
3914 data->irq_2_irte.index = index + sub_handle;
3916 /* Setup IRTE for IOMMU */
3918 irte->fields.vector = irq_cfg->vector;
3919 irte->fields.int_type = apic->irq_delivery_mode;
3920 irte->fields.destination = irq_cfg->dest_apicid;
3921 irte->fields.dm = apic->irq_dest_mode;
3922 irte->fields.valid = 1;
3924 switch (info->type) {
3925 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3926 /* Setup IOAPIC entry */
3927 entry = info->ioapic_entry;
3928 info->ioapic_entry = NULL;
3929 memset(entry, 0, sizeof(*entry));
3930 entry->vector = index;
3932 entry->trigger = info->ioapic_trigger;
3933 entry->polarity = info->ioapic_polarity;
3934 /* Mask level triggered irqs. */
3935 if (info->ioapic_trigger)
3939 case X86_IRQ_ALLOC_TYPE_HPET:
3940 case X86_IRQ_ALLOC_TYPE_MSI:
3941 case X86_IRQ_ALLOC_TYPE_MSIX:
3942 msg->address_hi = MSI_ADDR_BASE_HI;
3943 msg->address_lo = MSI_ADDR_BASE_LO;
3944 msg->data = irte_info->index;
3953 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
3954 unsigned int nr_irqs, void *arg)
3956 struct irq_alloc_info *info = arg;
3957 struct irq_data *irq_data;
3958 struct amd_ir_data *data;
3959 struct irq_cfg *cfg;
3965 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
3966 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
3970 * With IRQ remapping enabled, don't need contiguous CPU vectors
3971 * to support multiple MSI interrupts.
3973 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
3974 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
3976 devid = get_devid(info);
3980 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
3985 data = kzalloc(sizeof(*data), GFP_KERNEL);
3987 goto out_free_parent;
3989 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
3990 if (get_irq_table(devid, true))
3991 index = info->ioapic_pin;
3995 index = alloc_irq_index(devid, nr_irqs);
3998 pr_warn("Failed to allocate IRTE\n");
4000 goto out_free_parent;
4003 for (i = 0; i < nr_irqs; i++) {
4004 irq_data = irq_domain_get_irq_data(domain, virq + i);
4005 cfg = irqd_cfg(irq_data);
4006 if (!irq_data || !cfg) {
4012 data = kzalloc(sizeof(*data), GFP_KERNEL);
4016 irq_data->hwirq = (devid << 16) + i;
4017 irq_data->chip_data = data;
4018 irq_data->chip = &amd_ir_chip;
4019 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4020 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4025 for (i--; i >= 0; i--) {
4026 irq_data = irq_domain_get_irq_data(domain, virq + i);
4028 kfree(irq_data->chip_data);
4030 for (i = 0; i < nr_irqs; i++)
4031 free_irte(devid, index + i);
4033 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4037 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4038 unsigned int nr_irqs)
4040 struct irq_2_irte *irte_info;
4041 struct irq_data *irq_data;
4042 struct amd_ir_data *data;
4045 for (i = 0; i < nr_irqs; i++) {
4046 irq_data = irq_domain_get_irq_data(domain, virq + i);
4047 if (irq_data && irq_data->chip_data) {
4048 data = irq_data->chip_data;
4049 irte_info = &data->irq_2_irte;
4050 free_irte(irte_info->devid, irte_info->index);
4054 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4057 static void irq_remapping_activate(struct irq_domain *domain,
4058 struct irq_data *irq_data)
4060 struct amd_ir_data *data = irq_data->chip_data;
4061 struct irq_2_irte *irte_info = &data->irq_2_irte;
4063 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4066 static void irq_remapping_deactivate(struct irq_domain *domain,
4067 struct irq_data *irq_data)
4069 struct amd_ir_data *data = irq_data->chip_data;
4070 struct irq_2_irte *irte_info = &data->irq_2_irte;
4074 modify_irte(irte_info->devid, irte_info->index, data->irte_entry);
4077 static struct irq_domain_ops amd_ir_domain_ops = {
4078 .alloc = irq_remapping_alloc,
4079 .free = irq_remapping_free,
4080 .activate = irq_remapping_activate,
4081 .deactivate = irq_remapping_deactivate,
4084 static int amd_ir_set_affinity(struct irq_data *data,
4085 const struct cpumask *mask, bool force)
4087 struct amd_ir_data *ir_data = data->chip_data;
4088 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4089 struct irq_cfg *cfg = irqd_cfg(data);
4090 struct irq_data *parent = data->parent_data;
4093 ret = parent->chip->irq_set_affinity(parent, mask, force);
4094 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4098 * Atomically updates the IRTE with the new destination, vector
4099 * and flushes the interrupt entry cache.
4101 ir_data->irte_entry.fields.vector = cfg->vector;
4102 ir_data->irte_entry.fields.destination = cfg->dest_apicid;
4103 modify_irte(irte_info->devid, irte_info->index, ir_data->irte_entry);
4106 * After this point, all the interrupts will start arriving
4107 * at the new destination. So, time to cleanup the previous
4108 * vector allocation.
4110 send_cleanup_vector(cfg);
4112 return IRQ_SET_MASK_OK_DONE;
4115 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4117 struct amd_ir_data *ir_data = irq_data->chip_data;
4119 *msg = ir_data->msi_entry;
4122 static struct irq_chip amd_ir_chip = {
4123 .irq_ack = ir_ack_apic_edge,
4124 .irq_set_affinity = amd_ir_set_affinity,
4125 .irq_compose_msi_msg = ir_compose_msi_msg,
4128 int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4130 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4131 if (!iommu->ir_domain)
4134 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4135 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);