Linux Audio

Check our new training course

Loading...
Note: File does not exist in v5.9.
   1/*
   2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
   3 * Author: Joerg Roedel <jroedel@suse.de>
   4 *         Leo Duran <leo.duran@amd.com>
   5 *
   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.
   9 *
  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.
  14 *
  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
  18 */
  19
  20#include <linux/ratelimit.h>
  21#include <linux/pci.h>
  22#include <linux/acpi.h>
  23#include <linux/amba/bus.h>
  24#include <linux/platform_device.h>
  25#include <linux/pci-ats.h>
  26#include <linux/bitmap.h>
  27#include <linux/slab.h>
  28#include <linux/debugfs.h>
  29#include <linux/scatterlist.h>
  30#include <linux/dma-mapping.h>
  31#include <linux/dma-direct.h>
  32#include <linux/iommu-helper.h>
  33#include <linux/iommu.h>
  34#include <linux/delay.h>
  35#include <linux/amd-iommu.h>
  36#include <linux/notifier.h>
  37#include <linux/export.h>
  38#include <linux/irq.h>
  39#include <linux/msi.h>
  40#include <linux/dma-contiguous.h>
  41#include <linux/irqdomain.h>
  42#include <linux/percpu.h>
  43#include <linux/iova.h>
  44#include <asm/irq_remapping.h>
  45#include <asm/io_apic.h>
  46#include <asm/apic.h>
  47#include <asm/hw_irq.h>
  48#include <asm/msidef.h>
  49#include <asm/proto.h>
  50#include <asm/iommu.h>
  51#include <asm/gart.h>
  52#include <asm/dma.h>
  53
  54#include "amd_iommu_proto.h"
  55#include "amd_iommu_types.h"
  56#include "irq_remapping.h"
  57
  58#define AMD_IOMMU_MAPPING_ERROR	0
  59
  60#define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
  61
  62#define LOOP_TIMEOUT	100000
  63
  64/* IO virtual address start page frame number */
  65#define IOVA_START_PFN		(1)
  66#define IOVA_PFN(addr)		((addr) >> PAGE_SHIFT)
  67
  68/* Reserved IOVA ranges */
  69#define MSI_RANGE_START		(0xfee00000)
  70#define MSI_RANGE_END		(0xfeefffff)
  71#define HT_RANGE_START		(0xfd00000000ULL)
  72#define HT_RANGE_END		(0xffffffffffULL)
  73
  74/*
  75 * This bitmap is used to advertise the page sizes our hardware support
  76 * to the IOMMU core, which will then use this information to split
  77 * physically contiguous memory regions it is mapping into page sizes
  78 * that we support.
  79 *
  80 * 512GB Pages are not supported due to a hardware bug
  81 */
  82#define AMD_IOMMU_PGSIZES	((~0xFFFUL) & ~(2ULL << 38))
  83
  84static DEFINE_SPINLOCK(amd_iommu_devtable_lock);
  85static DEFINE_SPINLOCK(pd_bitmap_lock);
  86
  87/* List of all available dev_data structures */
  88static LLIST_HEAD(dev_data_list);
  89
  90LIST_HEAD(ioapic_map);
  91LIST_HEAD(hpet_map);
  92LIST_HEAD(acpihid_map);
  93
  94/*
  95 * Domain for untranslated devices - only allocated
  96 * if iommu=pt passed on kernel cmd line.
  97 */
  98const struct iommu_ops amd_iommu_ops;
  99
 100static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
 101int amd_iommu_max_glx_val = -1;
 102
 103static const struct dma_map_ops amd_iommu_dma_ops;
 104
 105/*
 106 * general struct to manage commands send to an IOMMU
 107 */
 108struct iommu_cmd {
 109	u32 data[4];
 110};
 111
 112struct kmem_cache *amd_iommu_irq_cache;
 113
 114static void update_domain(struct protection_domain *domain);
 115static int protection_domain_init(struct protection_domain *domain);
 116static void detach_device(struct device *dev);
 117static void iova_domain_flush_tlb(struct iova_domain *iovad);
 118
 119/*
 120 * Data container for a dma_ops specific protection domain
 121 */
 122struct dma_ops_domain {
 123	/* generic protection domain information */
 124	struct protection_domain domain;
 125
 126	/* IOVA RB-Tree */
 127	struct iova_domain iovad;
 128};
 129
 130static struct iova_domain reserved_iova_ranges;
 131static struct lock_class_key reserved_rbtree_key;
 132
 133/****************************************************************************
 134 *
 135 * Helper functions
 136 *
 137 ****************************************************************************/
 138
 139static inline int match_hid_uid(struct device *dev,
 140				struct acpihid_map_entry *entry)
 141{
 142	const char *hid, *uid;
 143
 144	hid = acpi_device_hid(ACPI_COMPANION(dev));
 145	uid = acpi_device_uid(ACPI_COMPANION(dev));
 146
 147	if (!hid || !(*hid))
 148		return -ENODEV;
 149
 150	if (!uid || !(*uid))
 151		return strcmp(hid, entry->hid);
 152
 153	if (!(*entry->uid))
 154		return strcmp(hid, entry->hid);
 155
 156	return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
 157}
 158
 159static inline u16 get_pci_device_id(struct device *dev)
 160{
 161	struct pci_dev *pdev = to_pci_dev(dev);
 162
 163	return PCI_DEVID(pdev->bus->number, pdev->devfn);
 164}
 165
 166static inline int get_acpihid_device_id(struct device *dev,
 167					struct acpihid_map_entry **entry)
 168{
 169	struct acpihid_map_entry *p;
 170
 171	list_for_each_entry(p, &acpihid_map, list) {
 172		if (!match_hid_uid(dev, p)) {
 173			if (entry)
 174				*entry = p;
 175			return p->devid;
 176		}
 177	}
 178	return -EINVAL;
 179}
 180
 181static inline int get_device_id(struct device *dev)
 182{
 183	int devid;
 184
 185	if (dev_is_pci(dev))
 186		devid = get_pci_device_id(dev);
 187	else
 188		devid = get_acpihid_device_id(dev, NULL);
 189
 190	return devid;
 191}
 192
 193static struct protection_domain *to_pdomain(struct iommu_domain *dom)
 194{
 195	return container_of(dom, struct protection_domain, domain);
 196}
 197
 198static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
 199{
 200	BUG_ON(domain->flags != PD_DMA_OPS_MASK);
 201	return container_of(domain, struct dma_ops_domain, domain);
 202}
 203
 204static struct iommu_dev_data *alloc_dev_data(u16 devid)
 205{
 206	struct iommu_dev_data *dev_data;
 207
 208	dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
 209	if (!dev_data)
 210		return NULL;
 211
 212	dev_data->devid = devid;
 213	ratelimit_default_init(&dev_data->rs);
 214
 215	llist_add(&dev_data->dev_data_list, &dev_data_list);
 216	return dev_data;
 217}
 218
 219static struct iommu_dev_data *search_dev_data(u16 devid)
 220{
 221	struct iommu_dev_data *dev_data;
 222	struct llist_node *node;
 223
 224	if (llist_empty(&dev_data_list))
 225		return NULL;
 226
 227	node = dev_data_list.first;
 228	llist_for_each_entry(dev_data, node, dev_data_list) {
 229		if (dev_data->devid == devid)
 230			return dev_data;
 231	}
 232
 233	return NULL;
 234}
 235
 236static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
 237{
 238	*(u16 *)data = alias;
 239	return 0;
 240}
 241
 242static u16 get_alias(struct device *dev)
 243{
 244	struct pci_dev *pdev = to_pci_dev(dev);
 245	u16 devid, ivrs_alias, pci_alias;
 246
 247	/* The callers make sure that get_device_id() does not fail here */
 248	devid = get_device_id(dev);
 249	ivrs_alias = amd_iommu_alias_table[devid];
 250	pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
 251
 252	if (ivrs_alias == pci_alias)
 253		return ivrs_alias;
 254
 255	/*
 256	 * DMA alias showdown
 257	 *
 258	 * The IVRS is fairly reliable in telling us about aliases, but it
 259	 * can't know about every screwy device.  If we don't have an IVRS
 260	 * reported alias, use the PCI reported alias.  In that case we may
 261	 * still need to initialize the rlookup and dev_table entries if the
 262	 * alias is to a non-existent device.
 263	 */
 264	if (ivrs_alias == devid) {
 265		if (!amd_iommu_rlookup_table[pci_alias]) {
 266			amd_iommu_rlookup_table[pci_alias] =
 267				amd_iommu_rlookup_table[devid];
 268			memcpy(amd_iommu_dev_table[pci_alias].data,
 269			       amd_iommu_dev_table[devid].data,
 270			       sizeof(amd_iommu_dev_table[pci_alias].data));
 271		}
 272
 273		return pci_alias;
 274	}
 275
 276	pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
 277		"for device %s[%04x:%04x], kernel reported alias "
 278		"%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
 279		PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
 280		PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
 281		PCI_FUNC(pci_alias));
 282
 283	/*
 284	 * If we don't have a PCI DMA alias and the IVRS alias is on the same
 285	 * bus, then the IVRS table may know about a quirk that we don't.
 286	 */
 287	if (pci_alias == devid &&
 288	    PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
 289		pci_add_dma_alias(pdev, ivrs_alias & 0xff);
 290		pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
 291			PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
 292			dev_name(dev));
 293	}
 294
 295	return ivrs_alias;
 296}
 297
 298static struct iommu_dev_data *find_dev_data(u16 devid)
 299{
 300	struct iommu_dev_data *dev_data;
 301	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
 302
 303	dev_data = search_dev_data(devid);
 304
 305	if (dev_data == NULL) {
 306		dev_data = alloc_dev_data(devid);
 307		if (!dev_data)
 308			return NULL;
 309
 310		if (translation_pre_enabled(iommu))
 311			dev_data->defer_attach = true;
 312	}
 313
 314	return dev_data;
 315}
 316
 317struct iommu_dev_data *get_dev_data(struct device *dev)
 318{
 319	return dev->archdata.iommu;
 320}
 321EXPORT_SYMBOL(get_dev_data);
 322
 323/*
 324* Find or create an IOMMU group for a acpihid device.
 325*/
 326static struct iommu_group *acpihid_device_group(struct device *dev)
 327{
 328	struct acpihid_map_entry *p, *entry = NULL;
 329	int devid;
 330
 331	devid = get_acpihid_device_id(dev, &entry);
 332	if (devid < 0)
 333		return ERR_PTR(devid);
 334
 335	list_for_each_entry(p, &acpihid_map, list) {
 336		if ((devid == p->devid) && p->group)
 337			entry->group = p->group;
 338	}
 339
 340	if (!entry->group)
 341		entry->group = generic_device_group(dev);
 342	else
 343		iommu_group_ref_get(entry->group);
 344
 345	return entry->group;
 346}
 347
 348static bool pci_iommuv2_capable(struct pci_dev *pdev)
 349{
 350	static const int caps[] = {
 351		PCI_EXT_CAP_ID_ATS,
 352		PCI_EXT_CAP_ID_PRI,
 353		PCI_EXT_CAP_ID_PASID,
 354	};
 355	int i, pos;
 356
 357	for (i = 0; i < 3; ++i) {
 358		pos = pci_find_ext_capability(pdev, caps[i]);
 359		if (pos == 0)
 360			return false;
 361	}
 362
 363	return true;
 364}
 365
 366static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
 367{
 368	struct iommu_dev_data *dev_data;
 369
 370	dev_data = get_dev_data(&pdev->dev);
 371
 372	return dev_data->errata & (1 << erratum) ? true : false;
 373}
 374
 375/*
 376 * This function checks if the driver got a valid device from the caller to
 377 * avoid dereferencing invalid pointers.
 378 */
 379static bool check_device(struct device *dev)
 380{
 381	int devid;
 382
 383	if (!dev || !dev->dma_mask)
 384		return false;
 385
 386	devid = get_device_id(dev);
 387	if (devid < 0)
 388		return false;
 389
 390	/* Out of our scope? */
 391	if (devid > amd_iommu_last_bdf)
 392		return false;
 393
 394	if (amd_iommu_rlookup_table[devid] == NULL)
 395		return false;
 396
 397	return true;
 398}
 399
 400static void init_iommu_group(struct device *dev)
 401{
 402	struct iommu_group *group;
 403
 404	group = iommu_group_get_for_dev(dev);
 405	if (IS_ERR(group))
 406		return;
 407
 408	iommu_group_put(group);
 409}
 410
 411static int iommu_init_device(struct device *dev)
 412{
 413	struct iommu_dev_data *dev_data;
 414	struct amd_iommu *iommu;
 415	int devid;
 416
 417	if (dev->archdata.iommu)
 418		return 0;
 419
 420	devid = get_device_id(dev);
 421	if (devid < 0)
 422		return devid;
 423
 424	iommu = amd_iommu_rlookup_table[devid];
 425
 426	dev_data = find_dev_data(devid);
 427	if (!dev_data)
 428		return -ENOMEM;
 429
 430	dev_data->alias = get_alias(dev);
 431
 432	if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
 433		struct amd_iommu *iommu;
 434
 435		iommu = amd_iommu_rlookup_table[dev_data->devid];
 436		dev_data->iommu_v2 = iommu->is_iommu_v2;
 437	}
 438
 439	dev->archdata.iommu = dev_data;
 440
 441	iommu_device_link(&iommu->iommu, dev);
 442
 443	return 0;
 444}
 445
 446static void iommu_ignore_device(struct device *dev)
 447{
 448	u16 alias;
 449	int devid;
 450
 451	devid = get_device_id(dev);
 452	if (devid < 0)
 453		return;
 454
 455	alias = get_alias(dev);
 456
 457	memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
 458	memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
 459
 460	amd_iommu_rlookup_table[devid] = NULL;
 461	amd_iommu_rlookup_table[alias] = NULL;
 462}
 463
 464static void iommu_uninit_device(struct device *dev)
 465{
 466	struct iommu_dev_data *dev_data;
 467	struct amd_iommu *iommu;
 468	int devid;
 469
 470	devid = get_device_id(dev);
 471	if (devid < 0)
 472		return;
 473
 474	iommu = amd_iommu_rlookup_table[devid];
 475
 476	dev_data = search_dev_data(devid);
 477	if (!dev_data)
 478		return;
 479
 480	if (dev_data->domain)
 481		detach_device(dev);
 482
 483	iommu_device_unlink(&iommu->iommu, dev);
 484
 485	iommu_group_remove_device(dev);
 486
 487	/* Remove dma-ops */
 488	dev->dma_ops = NULL;
 489
 490	/*
 491	 * We keep dev_data around for unplugged devices and reuse it when the
 492	 * device is re-plugged - not doing so would introduce a ton of races.
 493	 */
 494}
 495
 496/****************************************************************************
 497 *
 498 * Interrupt handling functions
 499 *
 500 ****************************************************************************/
 501
 502static void dump_dte_entry(u16 devid)
 503{
 504	int i;
 505
 506	for (i = 0; i < 4; ++i)
 507		pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
 508			amd_iommu_dev_table[devid].data[i]);
 509}
 510
 511static void dump_command(unsigned long phys_addr)
 512{
 513	struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr);
 514	int i;
 515
 516	for (i = 0; i < 4; ++i)
 517		pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
 518}
 519
 520static void amd_iommu_report_page_fault(u16 devid, u16 domain_id,
 521					u64 address, int flags)
 522{
 523	struct iommu_dev_data *dev_data = NULL;
 524	struct pci_dev *pdev;
 525
 526	pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
 527					   devid & 0xff);
 528	if (pdev)
 529		dev_data = get_dev_data(&pdev->dev);
 530
 531	if (dev_data && __ratelimit(&dev_data->rs)) {
 532		dev_err(&pdev->dev, "AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%016llx flags=0x%04x]\n",
 533			domain_id, address, flags);
 534	} else if (printk_ratelimit()) {
 535		pr_err("AMD-Vi: Event logged [IO_PAGE_FAULT device=%02x:%02x.%x domain=0x%04x address=0x%016llx flags=0x%04x]\n",
 536			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
 537			domain_id, address, flags);
 538	}
 539
 540	if (pdev)
 541		pci_dev_put(pdev);
 542}
 543
 544static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
 545{
 546	struct device *dev = iommu->iommu.dev;
 547	int type, devid, domid, flags;
 548	volatile u32 *event = __evt;
 549	int count = 0;
 550	u64 address;
 551
 552retry:
 553	type    = (event[1] >> EVENT_TYPE_SHIFT)  & EVENT_TYPE_MASK;
 554	devid   = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
 555	domid   = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
 556	flags   = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
 557	address = (u64)(((u64)event[3]) << 32) | event[2];
 558
 559	if (type == 0) {
 560		/* Did we hit the erratum? */
 561		if (++count == LOOP_TIMEOUT) {
 562			pr_err("AMD-Vi: No event written to event log\n");
 563			return;
 564		}
 565		udelay(1);
 566		goto retry;
 567	}
 568
 569	if (type == EVENT_TYPE_IO_FAULT) {
 570		amd_iommu_report_page_fault(devid, domid, address, flags);
 571		return;
 572	} else {
 573		dev_err(dev, "AMD-Vi: Event logged [");
 574	}
 575
 576	switch (type) {
 577	case EVENT_TYPE_ILL_DEV:
 578		dev_err(dev, "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),
 581			address, flags);
 582		dump_dte_entry(devid);
 583		break;
 584	case EVENT_TYPE_DEV_TAB_ERR:
 585		dev_err(dev, "DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
 586			"address=0x%016llx flags=0x%04x]\n",
 587			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
 588			address, flags);
 589		break;
 590	case EVENT_TYPE_PAGE_TAB_ERR:
 591		dev_err(dev, "PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
 592			"domain=0x%04x address=0x%016llx flags=0x%04x]\n",
 593			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
 594			domid, address, flags);
 595		break;
 596	case EVENT_TYPE_ILL_CMD:
 597		dev_err(dev, "ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
 598		dump_command(address);
 599		break;
 600	case EVENT_TYPE_CMD_HARD_ERR:
 601		dev_err(dev, "COMMAND_HARDWARE_ERROR address=0x%016llx "
 602			"flags=0x%04x]\n", address, flags);
 603		break;
 604	case EVENT_TYPE_IOTLB_INV_TO:
 605		dev_err(dev, "IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
 606			"address=0x%016llx]\n",
 607			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
 608			address);
 609		break;
 610	case EVENT_TYPE_INV_DEV_REQ:
 611		dev_err(dev, "INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
 612			"address=0x%016llx flags=0x%04x]\n",
 613			PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
 614			address, flags);
 615		break;
 616	default:
 617		dev_err(dev, KERN_ERR "UNKNOWN event[0]=0x%08x event[1]=0x%08x "
 618			"event[2]=0x%08x event[3]=0x%08x\n",
 619			event[0], event[1], event[2], event[3]);
 620	}
 621
 622	memset(__evt, 0, 4 * sizeof(u32));
 623}
 624
 625static void iommu_poll_events(struct amd_iommu *iommu)
 626{
 627	u32 head, tail;
 628
 629	head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
 630	tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
 631
 632	while (head != tail) {
 633		iommu_print_event(iommu, iommu->evt_buf + head);
 634		head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
 635	}
 636
 637	writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
 638}
 639
 640static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
 641{
 642	struct amd_iommu_fault fault;
 643
 644	if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
 645		pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
 646		return;
 647	}
 648
 649	fault.address   = raw[1];
 650	fault.pasid     = PPR_PASID(raw[0]);
 651	fault.device_id = PPR_DEVID(raw[0]);
 652	fault.tag       = PPR_TAG(raw[0]);
 653	fault.flags     = PPR_FLAGS(raw[0]);
 654
 655	atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
 656}
 657
 658static void iommu_poll_ppr_log(struct amd_iommu *iommu)
 659{
 660	u32 head, tail;
 661
 662	if (iommu->ppr_log == NULL)
 663		return;
 664
 665	head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
 666	tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
 667
 668	while (head != tail) {
 669		volatile u64 *raw;
 670		u64 entry[2];
 671		int i;
 672
 673		raw = (u64 *)(iommu->ppr_log + head);
 674
 675		/*
 676		 * Hardware bug: Interrupt may arrive before the entry is
 677		 * written to memory. If this happens we need to wait for the
 678		 * entry to arrive.
 679		 */
 680		for (i = 0; i < LOOP_TIMEOUT; ++i) {
 681			if (PPR_REQ_TYPE(raw[0]) != 0)
 682				break;
 683			udelay(1);
 684		}
 685
 686		/* Avoid memcpy function-call overhead */
 687		entry[0] = raw[0];
 688		entry[1] = raw[1];
 689
 690		/*
 691		 * To detect the hardware bug we need to clear the entry
 692		 * back to zero.
 693		 */
 694		raw[0] = raw[1] = 0UL;
 695
 696		/* Update head pointer of hardware ring-buffer */
 697		head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
 698		writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
 699
 700		/* Handle PPR entry */
 701		iommu_handle_ppr_entry(iommu, entry);
 702
 703		/* Refresh ring-buffer information */
 704		head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
 705		tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
 706	}
 707}
 708
 709#ifdef CONFIG_IRQ_REMAP
 710static int (*iommu_ga_log_notifier)(u32);
 711
 712int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
 713{
 714	iommu_ga_log_notifier = notifier;
 715
 716	return 0;
 717}
 718EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
 719
 720static void iommu_poll_ga_log(struct amd_iommu *iommu)
 721{
 722	u32 head, tail, cnt = 0;
 723
 724	if (iommu->ga_log == NULL)
 725		return;
 726
 727	head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
 728	tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
 729
 730	while (head != tail) {
 731		volatile u64 *raw;
 732		u64 log_entry;
 733
 734		raw = (u64 *)(iommu->ga_log + head);
 735		cnt++;
 736
 737		/* Avoid memcpy function-call overhead */
 738		log_entry = *raw;
 739
 740		/* Update head pointer of hardware ring-buffer */
 741		head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
 742		writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
 743
 744		/* Handle GA entry */
 745		switch (GA_REQ_TYPE(log_entry)) {
 746		case GA_GUEST_NR:
 747			if (!iommu_ga_log_notifier)
 748				break;
 749
 750			pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
 751				 __func__, GA_DEVID(log_entry),
 752				 GA_TAG(log_entry));
 753
 754			if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
 755				pr_err("AMD-Vi: GA log notifier failed.\n");
 756			break;
 757		default:
 758			break;
 759		}
 760	}
 761}
 762#endif /* CONFIG_IRQ_REMAP */
 763
 764#define AMD_IOMMU_INT_MASK	\
 765	(MMIO_STATUS_EVT_INT_MASK | \
 766	 MMIO_STATUS_PPR_INT_MASK | \
 767	 MMIO_STATUS_GALOG_INT_MASK)
 768
 769irqreturn_t amd_iommu_int_thread(int irq, void *data)
 770{
 771	struct amd_iommu *iommu = (struct amd_iommu *) data;
 772	u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
 773
 774	while (status & AMD_IOMMU_INT_MASK) {
 775		/* Enable EVT and PPR and GA interrupts again */
 776		writel(AMD_IOMMU_INT_MASK,
 777			iommu->mmio_base + MMIO_STATUS_OFFSET);
 778
 779		if (status & MMIO_STATUS_EVT_INT_MASK) {
 780			pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
 781			iommu_poll_events(iommu);
 782		}
 783
 784		if (status & MMIO_STATUS_PPR_INT_MASK) {
 785			pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
 786			iommu_poll_ppr_log(iommu);
 787		}
 788
 789#ifdef CONFIG_IRQ_REMAP
 790		if (status & MMIO_STATUS_GALOG_INT_MASK) {
 791			pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
 792			iommu_poll_ga_log(iommu);
 793		}
 794#endif
 795
 796		/*
 797		 * Hardware bug: ERBT1312
 798		 * When re-enabling interrupt (by writing 1
 799		 * to clear the bit), the hardware might also try to set
 800		 * the interrupt bit in the event status register.
 801		 * In this scenario, the bit will be set, and disable
 802		 * subsequent interrupts.
 803		 *
 804		 * Workaround: The IOMMU driver should read back the
 805		 * status register and check if the interrupt bits are cleared.
 806		 * If not, driver will need to go through the interrupt handler
 807		 * again and re-clear the bits
 808		 */
 809		status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
 810	}
 811	return IRQ_HANDLED;
 812}
 813
 814irqreturn_t amd_iommu_int_handler(int irq, void *data)
 815{
 816	return IRQ_WAKE_THREAD;
 817}
 818
 819/****************************************************************************
 820 *
 821 * IOMMU command queuing functions
 822 *
 823 ****************************************************************************/
 824
 825static int wait_on_sem(volatile u64 *sem)
 826{
 827	int i = 0;
 828
 829	while (*sem == 0 && i < LOOP_TIMEOUT) {
 830		udelay(1);
 831		i += 1;
 832	}
 833
 834	if (i == LOOP_TIMEOUT) {
 835		pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
 836		return -EIO;
 837	}
 838
 839	return 0;
 840}
 841
 842static void copy_cmd_to_buffer(struct amd_iommu *iommu,
 843			       struct iommu_cmd *cmd)
 844{
 845	u8 *target;
 846
 847	target = iommu->cmd_buf + iommu->cmd_buf_tail;
 848
 849	iommu->cmd_buf_tail += sizeof(*cmd);
 850	iommu->cmd_buf_tail %= CMD_BUFFER_SIZE;
 851
 852	/* Copy command to buffer */
 853	memcpy(target, cmd, sizeof(*cmd));
 854
 855	/* Tell the IOMMU about it */
 856	writel(iommu->cmd_buf_tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
 857}
 858
 859static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
 860{
 861	u64 paddr = iommu_virt_to_phys((void *)address);
 862
 863	WARN_ON(address & 0x7ULL);
 864
 865	memset(cmd, 0, sizeof(*cmd));
 866	cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK;
 867	cmd->data[1] = upper_32_bits(paddr);
 868	cmd->data[2] = 1;
 869	CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
 870}
 871
 872static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
 873{
 874	memset(cmd, 0, sizeof(*cmd));
 875	cmd->data[0] = devid;
 876	CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
 877}
 878
 879static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
 880				  size_t size, u16 domid, int pde)
 881{
 882	u64 pages;
 883	bool s;
 884
 885	pages = iommu_num_pages(address, size, PAGE_SIZE);
 886	s     = false;
 887
 888	if (pages > 1) {
 889		/*
 890		 * If we have to flush more than one page, flush all
 891		 * TLB entries for this domain
 892		 */
 893		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
 894		s = true;
 895	}
 896
 897	address &= PAGE_MASK;
 898
 899	memset(cmd, 0, sizeof(*cmd));
 900	cmd->data[1] |= domid;
 901	cmd->data[2]  = lower_32_bits(address);
 902	cmd->data[3]  = upper_32_bits(address);
 903	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
 904	if (s) /* size bit - we flush more than one 4kb page */
 905		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
 906	if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
 907		cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
 908}
 909
 910static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
 911				  u64 address, size_t size)
 912{
 913	u64 pages;
 914	bool s;
 915
 916	pages = iommu_num_pages(address, size, PAGE_SIZE);
 917	s     = false;
 918
 919	if (pages > 1) {
 920		/*
 921		 * If we have to flush more than one page, flush all
 922		 * TLB entries for this domain
 923		 */
 924		address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
 925		s = true;
 926	}
 927
 928	address &= PAGE_MASK;
 929
 930	memset(cmd, 0, sizeof(*cmd));
 931	cmd->data[0]  = devid;
 932	cmd->data[0] |= (qdep & 0xff) << 24;
 933	cmd->data[1]  = devid;
 934	cmd->data[2]  = lower_32_bits(address);
 935	cmd->data[3]  = upper_32_bits(address);
 936	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
 937	if (s)
 938		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
 939}
 940
 941static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
 942				  u64 address, bool size)
 943{
 944	memset(cmd, 0, sizeof(*cmd));
 945
 946	address &= ~(0xfffULL);
 947
 948	cmd->data[0]  = pasid;
 949	cmd->data[1]  = domid;
 950	cmd->data[2]  = lower_32_bits(address);
 951	cmd->data[3]  = upper_32_bits(address);
 952	cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
 953	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
 954	if (size)
 955		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
 956	CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
 957}
 958
 959static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
 960				  int qdep, u64 address, bool size)
 961{
 962	memset(cmd, 0, sizeof(*cmd));
 963
 964	address &= ~(0xfffULL);
 965
 966	cmd->data[0]  = devid;
 967	cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
 968	cmd->data[0] |= (qdep  & 0xff) << 24;
 969	cmd->data[1]  = devid;
 970	cmd->data[1] |= (pasid & 0xff) << 16;
 971	cmd->data[2]  = lower_32_bits(address);
 972	cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
 973	cmd->data[3]  = upper_32_bits(address);
 974	if (size)
 975		cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
 976	CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
 977}
 978
 979static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
 980			       int status, int tag, bool gn)
 981{
 982	memset(cmd, 0, sizeof(*cmd));
 983
 984	cmd->data[0]  = devid;
 985	if (gn) {
 986		cmd->data[1]  = pasid;
 987		cmd->data[2]  = CMD_INV_IOMMU_PAGES_GN_MASK;
 988	}
 989	cmd->data[3]  = tag & 0x1ff;
 990	cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
 991
 992	CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
 993}
 994
 995static void build_inv_all(struct iommu_cmd *cmd)
 996{
 997	memset(cmd, 0, sizeof(*cmd));
 998	CMD_SET_TYPE(cmd, CMD_INV_ALL);
 999}
1000
1001static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1002{
1003	memset(cmd, 0, sizeof(*cmd));
1004	cmd->data[0] = devid;
1005	CMD_SET_TYPE(cmd, CMD_INV_IRT);
1006}
1007
1008/*
1009 * Writes the command to the IOMMUs command buffer and informs the
1010 * hardware about the new command.
1011 */
1012static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1013				      struct iommu_cmd *cmd,
1014				      bool sync)
1015{
1016	unsigned int count = 0;
1017	u32 left, next_tail;
1018
1019	next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1020again:
1021	left      = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE;
1022
1023	if (left <= 0x20) {
1024		/* Skip udelay() the first time around */
1025		if (count++) {
1026			if (count == LOOP_TIMEOUT) {
1027				pr_err("AMD-Vi: Command buffer timeout\n");
1028				return -EIO;
1029			}
1030
1031			udelay(1);
1032		}
1033
1034		/* Update head and recheck remaining space */
1035		iommu->cmd_buf_head = readl(iommu->mmio_base +
1036					    MMIO_CMD_HEAD_OFFSET);
1037
1038		goto again;
1039	}
1040
1041	copy_cmd_to_buffer(iommu, cmd);
1042
1043	/* Do we need to make sure all commands are processed? */
1044	iommu->need_sync = sync;
1045
1046	return 0;
1047}
1048
1049static int iommu_queue_command_sync(struct amd_iommu *iommu,
1050				    struct iommu_cmd *cmd,
1051				    bool sync)
1052{
1053	unsigned long flags;
1054	int ret;
1055
1056	raw_spin_lock_irqsave(&iommu->lock, flags);
1057	ret = __iommu_queue_command_sync(iommu, cmd, sync);
1058	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1059
1060	return ret;
1061}
1062
1063static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1064{
1065	return iommu_queue_command_sync(iommu, cmd, true);
1066}
1067
1068/*
1069 * This function queues a completion wait command into the command
1070 * buffer of an IOMMU
1071 */
1072static int iommu_completion_wait(struct amd_iommu *iommu)
1073{
1074	struct iommu_cmd cmd;
1075	unsigned long flags;
1076	int ret;
1077
1078	if (!iommu->need_sync)
1079		return 0;
1080
1081
1082	build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1083
1084	raw_spin_lock_irqsave(&iommu->lock, flags);
1085
1086	iommu->cmd_sem = 0;
1087
1088	ret = __iommu_queue_command_sync(iommu, &cmd, false);
1089	if (ret)
1090		goto out_unlock;
1091
1092	ret = wait_on_sem(&iommu->cmd_sem);
1093
1094out_unlock:
1095	raw_spin_unlock_irqrestore(&iommu->lock, flags);
1096
1097	return ret;
1098}
1099
1100static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1101{
1102	struct iommu_cmd cmd;
1103
1104	build_inv_dte(&cmd, devid);
1105
1106	return iommu_queue_command(iommu, &cmd);
1107}
1108
1109static void amd_iommu_flush_dte_all(struct amd_iommu *iommu)
1110{
1111	u32 devid;
1112
1113	for (devid = 0; devid <= 0xffff; ++devid)
1114		iommu_flush_dte(iommu, devid);
1115
1116	iommu_completion_wait(iommu);
1117}
1118
1119/*
1120 * This function uses heavy locking and may disable irqs for some time. But
1121 * this is no issue because it is only called during resume.
1122 */
1123static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu)
1124{
1125	u32 dom_id;
1126
1127	for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1128		struct iommu_cmd cmd;
1129		build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1130				      dom_id, 1);
1131		iommu_queue_command(iommu, &cmd);
1132	}
1133
1134	iommu_completion_wait(iommu);
1135}
1136
1137static void amd_iommu_flush_all(struct amd_iommu *iommu)
1138{
1139	struct iommu_cmd cmd;
1140
1141	build_inv_all(&cmd);
1142
1143	iommu_queue_command(iommu, &cmd);
1144	iommu_completion_wait(iommu);
1145}
1146
1147static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1148{
1149	struct iommu_cmd cmd;
1150
1151	build_inv_irt(&cmd, devid);
1152
1153	iommu_queue_command(iommu, &cmd);
1154}
1155
1156static void amd_iommu_flush_irt_all(struct amd_iommu *iommu)
1157{
1158	u32 devid;
1159
1160	for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1161		iommu_flush_irt(iommu, devid);
1162
1163	iommu_completion_wait(iommu);
1164}
1165
1166void iommu_flush_all_caches(struct amd_iommu *iommu)
1167{
1168	if (iommu_feature(iommu, FEATURE_IA)) {
1169		amd_iommu_flush_all(iommu);
1170	} else {
1171		amd_iommu_flush_dte_all(iommu);
1172		amd_iommu_flush_irt_all(iommu);
1173		amd_iommu_flush_tlb_all(iommu);
1174	}
1175}
1176
1177/*
1178 * Command send function for flushing on-device TLB
1179 */
1180static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1181			      u64 address, size_t size)
1182{
1183	struct amd_iommu *iommu;
1184	struct iommu_cmd cmd;
1185	int qdep;
1186
1187	qdep     = dev_data->ats.qdep;
1188	iommu    = amd_iommu_rlookup_table[dev_data->devid];
1189
1190	build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1191
1192	return iommu_queue_command(iommu, &cmd);
1193}
1194
1195/*
1196 * Command send function for invalidating a device table entry
1197 */
1198static int device_flush_dte(struct iommu_dev_data *dev_data)
1199{
1200	struct amd_iommu *iommu;
1201	u16 alias;
1202	int ret;
1203
1204	iommu = amd_iommu_rlookup_table[dev_data->devid];
1205	alias = dev_data->alias;
1206
1207	ret = iommu_flush_dte(iommu, dev_data->devid);
1208	if (!ret && alias != dev_data->devid)
1209		ret = iommu_flush_dte(iommu, alias);
1210	if (ret)
1211		return ret;
1212
1213	if (dev_data->ats.enabled)
1214		ret = device_flush_iotlb(dev_data, 0, ~0UL);
1215
1216	return ret;
1217}
1218
1219/*
1220 * TLB invalidation function which is called from the mapping functions.
1221 * It invalidates a single PTE if the range to flush is within a single
1222 * page. Otherwise it flushes the whole TLB of the IOMMU.
1223 */
1224static void __domain_flush_pages(struct protection_domain *domain,
1225				 u64 address, size_t size, int pde)
1226{
1227	struct iommu_dev_data *dev_data;
1228	struct iommu_cmd cmd;
1229	int ret = 0, i;
1230
1231	build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1232
1233	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1234		if (!domain->dev_iommu[i])
1235			continue;
1236
1237		/*
1238		 * Devices of this domain are behind this IOMMU
1239		 * We need a TLB flush
1240		 */
1241		ret |= iommu_queue_command(amd_iommus[i], &cmd);
1242	}
1243
1244	list_for_each_entry(dev_data, &domain->dev_list, list) {
1245
1246		if (!dev_data->ats.enabled)
1247			continue;
1248
1249		ret |= device_flush_iotlb(dev_data, address, size);
1250	}
1251
1252	WARN_ON(ret);
1253}
1254
1255static void domain_flush_pages(struct protection_domain *domain,
1256			       u64 address, size_t size)
1257{
1258	__domain_flush_pages(domain, address, size, 0);
1259}
1260
1261/* Flush the whole IO/TLB for a given protection domain */
1262static void domain_flush_tlb(struct protection_domain *domain)
1263{
1264	__domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1265}
1266
1267/* Flush the whole IO/TLB for a given protection domain - including PDE */
1268static void domain_flush_tlb_pde(struct protection_domain *domain)
1269{
1270	__domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1271}
1272
1273static void domain_flush_complete(struct protection_domain *domain)
1274{
1275	int i;
1276
1277	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
1278		if (domain && !domain->dev_iommu[i])
1279			continue;
1280
1281		/*
1282		 * Devices of this domain are behind this IOMMU
1283		 * We need to wait for completion of all commands.
1284		 */
1285		iommu_completion_wait(amd_iommus[i]);
1286	}
1287}
1288
1289
1290/*
1291 * This function flushes the DTEs for all devices in domain
1292 */
1293static void domain_flush_devices(struct protection_domain *domain)
1294{
1295	struct iommu_dev_data *dev_data;
1296
1297	list_for_each_entry(dev_data, &domain->dev_list, list)
1298		device_flush_dte(dev_data);
1299}
1300
1301/****************************************************************************
1302 *
1303 * The functions below are used the create the page table mappings for
1304 * unity mapped regions.
1305 *
1306 ****************************************************************************/
1307
1308/*
1309 * This function is used to add another level to an IO page table. Adding
1310 * another level increases the size of the address space by 9 bits to a size up
1311 * to 64 bits.
1312 */
1313static bool increase_address_space(struct protection_domain *domain,
1314				   gfp_t gfp)
1315{
1316	u64 *pte;
1317
1318	if (domain->mode == PAGE_MODE_6_LEVEL)
1319		/* address space already 64 bit large */
1320		return false;
1321
1322	pte = (void *)get_zeroed_page(gfp);
1323	if (!pte)
1324		return false;
1325
1326	*pte             = PM_LEVEL_PDE(domain->mode,
1327					iommu_virt_to_phys(domain->pt_root));
1328	domain->pt_root  = pte;
1329	domain->mode    += 1;
1330	domain->updated  = true;
1331
1332	return true;
1333}
1334
1335static u64 *alloc_pte(struct protection_domain *domain,
1336		      unsigned long address,
1337		      unsigned long page_size,
1338		      u64 **pte_page,
1339		      gfp_t gfp)
1340{
1341	int level, end_lvl;
1342	u64 *pte, *page;
1343
1344	BUG_ON(!is_power_of_2(page_size));
1345
1346	while (address > PM_LEVEL_SIZE(domain->mode))
1347		increase_address_space(domain, gfp);
1348
1349	level   = domain->mode - 1;
1350	pte     = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1351	address = PAGE_SIZE_ALIGN(address, page_size);
1352	end_lvl = PAGE_SIZE_LEVEL(page_size);
1353
1354	while (level > end_lvl) {
1355		u64 __pte, __npte;
1356
1357		__pte = *pte;
1358
1359		if (!IOMMU_PTE_PRESENT(__pte)) {
1360			page = (u64 *)get_zeroed_page(gfp);
1361			if (!page)
1362				return NULL;
1363
1364			__npte = PM_LEVEL_PDE(level, iommu_virt_to_phys(page));
1365
1366			/* pte could have been changed somewhere. */
1367			if (cmpxchg64(pte, __pte, __npte) != __pte) {
1368				free_page((unsigned long)page);
1369				continue;
1370			}
1371		}
1372
1373		/* No level skipping support yet */
1374		if (PM_PTE_LEVEL(*pte) != level)
1375			return NULL;
1376
1377		level -= 1;
1378
1379		pte = IOMMU_PTE_PAGE(*pte);
1380
1381		if (pte_page && level == end_lvl)
1382			*pte_page = pte;
1383
1384		pte = &pte[PM_LEVEL_INDEX(level, address)];
1385	}
1386
1387	return pte;
1388}
1389
1390/*
1391 * This function checks if there is a PTE for a given dma address. If
1392 * there is one, it returns the pointer to it.
1393 */
1394static u64 *fetch_pte(struct protection_domain *domain,
1395		      unsigned long address,
1396		      unsigned long *page_size)
1397{
1398	int level;
1399	u64 *pte;
1400
1401	if (address > PM_LEVEL_SIZE(domain->mode))
1402		return NULL;
1403
1404	level	   =  domain->mode - 1;
1405	pte	   = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1406	*page_size =  PTE_LEVEL_PAGE_SIZE(level);
1407
1408	while (level > 0) {
1409
1410		/* Not Present */
1411		if (!IOMMU_PTE_PRESENT(*pte))
1412			return NULL;
1413
1414		/* Large PTE */
1415		if (PM_PTE_LEVEL(*pte) == 7 ||
1416		    PM_PTE_LEVEL(*pte) == 0)
1417			break;
1418
1419		/* No level skipping support yet */
1420		if (PM_PTE_LEVEL(*pte) != level)
1421			return NULL;
1422
1423		level -= 1;
1424
1425		/* Walk to the next level */
1426		pte	   = IOMMU_PTE_PAGE(*pte);
1427		pte	   = &pte[PM_LEVEL_INDEX(level, address)];
1428		*page_size = PTE_LEVEL_PAGE_SIZE(level);
1429	}
1430
1431	if (PM_PTE_LEVEL(*pte) == 0x07) {
1432		unsigned long pte_mask;
1433
1434		/*
1435		 * If we have a series of large PTEs, make
1436		 * sure to return a pointer to the first one.
1437		 */
1438		*page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1439		pte_mask   = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1440		pte        = (u64 *)(((unsigned long)pte) & pte_mask);
1441	}
1442
1443	return pte;
1444}
1445
1446/*
1447 * Generic mapping functions. It maps a physical address into a DMA
1448 * address space. It allocates the page table pages if necessary.
1449 * In the future it can be extended to a generic mapping function
1450 * supporting all features of AMD IOMMU page tables like level skipping
1451 * and full 64 bit address spaces.
1452 */
1453static int iommu_map_page(struct protection_domain *dom,
1454			  unsigned long bus_addr,
1455			  unsigned long phys_addr,
1456			  unsigned long page_size,
1457			  int prot,
1458			  gfp_t gfp)
1459{
1460	u64 __pte, *pte;
1461	int i, count;
1462
1463	BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1464	BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1465
1466	if (!(prot & IOMMU_PROT_MASK))
1467		return -EINVAL;
1468
1469	count = PAGE_SIZE_PTE_COUNT(page_size);
1470	pte   = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1471
1472	if (!pte)
1473		return -ENOMEM;
1474
1475	for (i = 0; i < count; ++i)
1476		if (IOMMU_PTE_PRESENT(pte[i]))
1477			return -EBUSY;
1478
1479	if (count > 1) {
1480		__pte = PAGE_SIZE_PTE(__sme_set(phys_addr), page_size);
1481		__pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1482	} else
1483		__pte = __sme_set(phys_addr) | IOMMU_PTE_PR | IOMMU_PTE_FC;
1484
1485	if (prot & IOMMU_PROT_IR)
1486		__pte |= IOMMU_PTE_IR;
1487	if (prot & IOMMU_PROT_IW)
1488		__pte |= IOMMU_PTE_IW;
1489
1490	for (i = 0; i < count; ++i)
1491		pte[i] = __pte;
1492
1493	update_domain(dom);
1494
1495	return 0;
1496}
1497
1498static unsigned long iommu_unmap_page(struct protection_domain *dom,
1499				      unsigned long bus_addr,
1500				      unsigned long page_size)
1501{
1502	unsigned long long unmapped;
1503	unsigned long unmap_size;
1504	u64 *pte;
1505
1506	BUG_ON(!is_power_of_2(page_size));
1507
1508	unmapped = 0;
1509
1510	while (unmapped < page_size) {
1511
1512		pte = fetch_pte(dom, bus_addr, &unmap_size);
1513
1514		if (pte) {
1515			int i, count;
1516
1517			count = PAGE_SIZE_PTE_COUNT(unmap_size);
1518			for (i = 0; i < count; i++)
1519				pte[i] = 0ULL;
1520		}
1521
1522		bus_addr  = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1523		unmapped += unmap_size;
1524	}
1525
1526	BUG_ON(unmapped && !is_power_of_2(unmapped));
1527
1528	return unmapped;
1529}
1530
1531/****************************************************************************
1532 *
1533 * The next functions belong to the address allocator for the dma_ops
1534 * interface functions.
1535 *
1536 ****************************************************************************/
1537
1538
1539static unsigned long dma_ops_alloc_iova(struct device *dev,
1540					struct dma_ops_domain *dma_dom,
1541					unsigned int pages, u64 dma_mask)
1542{
1543	unsigned long pfn = 0;
1544
1545	pages = __roundup_pow_of_two(pages);
1546
1547	if (dma_mask > DMA_BIT_MASK(32))
1548		pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1549				      IOVA_PFN(DMA_BIT_MASK(32)), false);
1550
1551	if (!pfn)
1552		pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1553				      IOVA_PFN(dma_mask), true);
1554
1555	return (pfn << PAGE_SHIFT);
1556}
1557
1558static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1559			      unsigned long address,
1560			      unsigned int pages)
1561{
1562	pages = __roundup_pow_of_two(pages);
1563	address >>= PAGE_SHIFT;
1564
1565	free_iova_fast(&dma_dom->iovad, address, pages);
1566}
1567
1568/****************************************************************************
1569 *
1570 * The next functions belong to the domain allocation. A domain is
1571 * allocated for every IOMMU as the default domain. If device isolation
1572 * is enabled, every device get its own domain. The most important thing
1573 * about domains is the page table mapping the DMA address space they
1574 * contain.
1575 *
1576 ****************************************************************************/
1577
1578/*
1579 * This function adds a protection domain to the global protection domain list
1580 */
1581static void add_domain_to_list(struct protection_domain *domain)
1582{
1583	unsigned long flags;
1584
1585	spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1586	list_add(&domain->list, &amd_iommu_pd_list);
1587	spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1588}
1589
1590/*
1591 * This function removes a protection domain to the global
1592 * protection domain list
1593 */
1594static void del_domain_from_list(struct protection_domain *domain)
1595{
1596	unsigned long flags;
1597
1598	spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1599	list_del(&domain->list);
1600	spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1601}
1602
1603static u16 domain_id_alloc(void)
1604{
1605	int id;
1606
1607	spin_lock(&pd_bitmap_lock);
1608	id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1609	BUG_ON(id == 0);
1610	if (id > 0 && id < MAX_DOMAIN_ID)
1611		__set_bit(id, amd_iommu_pd_alloc_bitmap);
1612	else
1613		id = 0;
1614	spin_unlock(&pd_bitmap_lock);
1615
1616	return id;
1617}
1618
1619static void domain_id_free(int id)
1620{
1621	spin_lock(&pd_bitmap_lock);
1622	if (id > 0 && id < MAX_DOMAIN_ID)
1623		__clear_bit(id, amd_iommu_pd_alloc_bitmap);
1624	spin_unlock(&pd_bitmap_lock);
1625}
1626
1627#define DEFINE_FREE_PT_FN(LVL, FN)				\
1628static void free_pt_##LVL (unsigned long __pt)			\
1629{								\
1630	unsigned long p;					\
1631	u64 *pt;						\
1632	int i;							\
1633								\
1634	pt = (u64 *)__pt;					\
1635								\
1636	for (i = 0; i < 512; ++i) {				\
1637		/* PTE present? */				\
1638		if (!IOMMU_PTE_PRESENT(pt[i]))			\
1639			continue;				\
1640								\
1641		/* Large PTE? */				\
1642		if (PM_PTE_LEVEL(pt[i]) == 0 ||			\
1643		    PM_PTE_LEVEL(pt[i]) == 7)			\
1644			continue;				\
1645								\
1646		p = (unsigned long)IOMMU_PTE_PAGE(pt[i]);	\
1647		FN(p);						\
1648	}							\
1649	free_page((unsigned long)pt);				\
1650}
1651
1652DEFINE_FREE_PT_FN(l2, free_page)
1653DEFINE_FREE_PT_FN(l3, free_pt_l2)
1654DEFINE_FREE_PT_FN(l4, free_pt_l3)
1655DEFINE_FREE_PT_FN(l5, free_pt_l4)
1656DEFINE_FREE_PT_FN(l6, free_pt_l5)
1657
1658static void free_pagetable(struct protection_domain *domain)
1659{
1660	unsigned long root = (unsigned long)domain->pt_root;
1661
1662	switch (domain->mode) {
1663	case PAGE_MODE_NONE:
1664		break;
1665	case PAGE_MODE_1_LEVEL:
1666		free_page(root);
1667		break;
1668	case PAGE_MODE_2_LEVEL:
1669		free_pt_l2(root);
1670		break;
1671	case PAGE_MODE_3_LEVEL:
1672		free_pt_l3(root);
1673		break;
1674	case PAGE_MODE_4_LEVEL:
1675		free_pt_l4(root);
1676		break;
1677	case PAGE_MODE_5_LEVEL:
1678		free_pt_l5(root);
1679		break;
1680	case PAGE_MODE_6_LEVEL:
1681		free_pt_l6(root);
1682		break;
1683	default:
1684		BUG();
1685	}
1686}
1687
1688static void free_gcr3_tbl_level1(u64 *tbl)
1689{
1690	u64 *ptr;
1691	int i;
1692
1693	for (i = 0; i < 512; ++i) {
1694		if (!(tbl[i] & GCR3_VALID))
1695			continue;
1696
1697		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1698
1699		free_page((unsigned long)ptr);
1700	}
1701}
1702
1703static void free_gcr3_tbl_level2(u64 *tbl)
1704{
1705	u64 *ptr;
1706	int i;
1707
1708	for (i = 0; i < 512; ++i) {
1709		if (!(tbl[i] & GCR3_VALID))
1710			continue;
1711
1712		ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK);
1713
1714		free_gcr3_tbl_level1(ptr);
1715	}
1716}
1717
1718static void free_gcr3_table(struct protection_domain *domain)
1719{
1720	if (domain->glx == 2)
1721		free_gcr3_tbl_level2(domain->gcr3_tbl);
1722	else if (domain->glx == 1)
1723		free_gcr3_tbl_level1(domain->gcr3_tbl);
1724	else
1725		BUG_ON(domain->glx != 0);
1726
1727	free_page((unsigned long)domain->gcr3_tbl);
1728}
1729
1730static void dma_ops_domain_flush_tlb(struct dma_ops_domain *dom)
1731{
1732	domain_flush_tlb(&dom->domain);
1733	domain_flush_complete(&dom->domain);
1734}
1735
1736static void iova_domain_flush_tlb(struct iova_domain *iovad)
1737{
1738	struct dma_ops_domain *dom;
1739
1740	dom = container_of(iovad, struct dma_ops_domain, iovad);
1741
1742	dma_ops_domain_flush_tlb(dom);
1743}
1744
1745/*
1746 * Free a domain, only used if something went wrong in the
1747 * allocation path and we need to free an already allocated page table
1748 */
1749static void dma_ops_domain_free(struct dma_ops_domain *dom)
1750{
1751	if (!dom)
1752		return;
1753
1754	del_domain_from_list(&dom->domain);
1755
1756	put_iova_domain(&dom->iovad);
1757
1758	free_pagetable(&dom->domain);
1759
1760	if (dom->domain.id)
1761		domain_id_free(dom->domain.id);
1762
1763	kfree(dom);
1764}
1765
1766/*
1767 * Allocates a new protection domain usable for the dma_ops functions.
1768 * It also initializes the page table and the address allocator data
1769 * structures required for the dma_ops interface
1770 */
1771static struct dma_ops_domain *dma_ops_domain_alloc(void)
1772{
1773	struct dma_ops_domain *dma_dom;
1774
1775	dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1776	if (!dma_dom)
1777		return NULL;
1778
1779	if (protection_domain_init(&dma_dom->domain))
1780		goto free_dma_dom;
1781
1782	dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1783	dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1784	dma_dom->domain.flags = PD_DMA_OPS_MASK;
1785	if (!dma_dom->domain.pt_root)
1786		goto free_dma_dom;
1787
1788	init_iova_domain(&dma_dom->iovad, PAGE_SIZE, IOVA_START_PFN);
1789
1790	if (init_iova_flush_queue(&dma_dom->iovad, iova_domain_flush_tlb, NULL))
1791		goto free_dma_dom;
1792
1793	/* Initialize reserved ranges */
1794	copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1795
1796	add_domain_to_list(&dma_dom->domain);
1797
1798	return dma_dom;
1799
1800free_dma_dom:
1801	dma_ops_domain_free(dma_dom);
1802
1803	return NULL;
1804}
1805
1806/*
1807 * little helper function to check whether a given protection domain is a
1808 * dma_ops domain
1809 */
1810static bool dma_ops_domain(struct protection_domain *domain)
1811{
1812	return domain->flags & PD_DMA_OPS_MASK;
1813}
1814
1815static void set_dte_entry(u16 devid, struct protection_domain *domain,
1816			  bool ats, bool ppr)
1817{
1818	u64 pte_root = 0;
1819	u64 flags = 0;
1820
1821	if (domain->mode != PAGE_MODE_NONE)
1822		pte_root = iommu_virt_to_phys(domain->pt_root);
1823
1824	pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1825		    << DEV_ENTRY_MODE_SHIFT;
1826	pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V | DTE_FLAG_TV;
1827
1828	flags = amd_iommu_dev_table[devid].data[1];
1829
1830	if (ats)
1831		flags |= DTE_FLAG_IOTLB;
1832
1833	if (ppr) {
1834		struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1835
1836		if (iommu_feature(iommu, FEATURE_EPHSUP))
1837			pte_root |= 1ULL << DEV_ENTRY_PPR;
1838	}
1839
1840	if (domain->flags & PD_IOMMUV2_MASK) {
1841		u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl);
1842		u64 glx  = domain->glx;
1843		u64 tmp;
1844
1845		pte_root |= DTE_FLAG_GV;
1846		pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1847
1848		/* First mask out possible old values for GCR3 table */
1849		tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1850		flags    &= ~tmp;
1851
1852		tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1853		flags    &= ~tmp;
1854
1855		/* Encode GCR3 table into DTE */
1856		tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1857		pte_root |= tmp;
1858
1859		tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1860		flags    |= tmp;
1861
1862		tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1863		flags    |= tmp;
1864	}
1865
1866	flags &= ~DEV_DOMID_MASK;
1867	flags |= domain->id;
1868
1869	amd_iommu_dev_table[devid].data[1]  = flags;
1870	amd_iommu_dev_table[devid].data[0]  = pte_root;
1871}
1872
1873static void clear_dte_entry(u16 devid)
1874{
1875	/* remove entry from the device table seen by the hardware */
1876	amd_iommu_dev_table[devid].data[0]  = DTE_FLAG_V | DTE_FLAG_TV;
1877	amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1878
1879	amd_iommu_apply_erratum_63(devid);
1880}
1881
1882static void do_attach(struct iommu_dev_data *dev_data,
1883		      struct protection_domain *domain)
1884{
1885	struct amd_iommu *iommu;
1886	u16 alias;
1887	bool ats;
1888
1889	iommu = amd_iommu_rlookup_table[dev_data->devid];
1890	alias = dev_data->alias;
1891	ats   = dev_data->ats.enabled;
1892
1893	/* Update data structures */
1894	dev_data->domain = domain;
1895	list_add(&dev_data->list, &domain->dev_list);
1896
1897	/* Do reference counting */
1898	domain->dev_iommu[iommu->index] += 1;
1899	domain->dev_cnt                 += 1;
1900
1901	/* Update device table */
1902	set_dte_entry(dev_data->devid, domain, ats, dev_data->iommu_v2);
1903	if (alias != dev_data->devid)
1904		set_dte_entry(alias, domain, ats, dev_data->iommu_v2);
1905
1906	device_flush_dte(dev_data);
1907}
1908
1909static void do_detach(struct iommu_dev_data *dev_data)
1910{
1911	struct amd_iommu *iommu;
1912	u16 alias;
1913
1914	/*
1915	 * First check if the device is still attached. It might already
1916	 * be detached from its domain because the generic
1917	 * iommu_detach_group code detached it and we try again here in
1918	 * our alias handling.
1919	 */
1920	if (!dev_data->domain)
1921		return;
1922
1923	iommu = amd_iommu_rlookup_table[dev_data->devid];
1924	alias = dev_data->alias;
1925
1926	/* decrease reference counters */
1927	dev_data->domain->dev_iommu[iommu->index] -= 1;
1928	dev_data->domain->dev_cnt                 -= 1;
1929
1930	/* Update data structures */
1931	dev_data->domain = NULL;
1932	list_del(&dev_data->list);
1933	clear_dte_entry(dev_data->devid);
1934	if (alias != dev_data->devid)
1935		clear_dte_entry(alias);
1936
1937	/* Flush the DTE entry */
1938	device_flush_dte(dev_data);
1939}
1940
1941/*
1942 * If a device is not yet associated with a domain, this function does
1943 * assigns it visible for the hardware
1944 */
1945static int __attach_device(struct iommu_dev_data *dev_data,
1946			   struct protection_domain *domain)
1947{
1948	int ret;
1949
1950	/*
1951	 * Must be called with IRQs disabled. Warn here to detect early
1952	 * when its not.
1953	 */
1954	WARN_ON(!irqs_disabled());
1955
1956	/* lock domain */
1957	spin_lock(&domain->lock);
1958
1959	ret = -EBUSY;
1960	if (dev_data->domain != NULL)
1961		goto out_unlock;
1962
1963	/* Attach alias group root */
1964	do_attach(dev_data, domain);
1965
1966	ret = 0;
1967
1968out_unlock:
1969
1970	/* ready */
1971	spin_unlock(&domain->lock);
1972
1973	return ret;
1974}
1975
1976
1977static void pdev_iommuv2_disable(struct pci_dev *pdev)
1978{
1979	pci_disable_ats(pdev);
1980	pci_disable_pri(pdev);
1981	pci_disable_pasid(pdev);
1982}
1983
1984/* FIXME: Change generic reset-function to do the same */
1985static int pri_reset_while_enabled(struct pci_dev *pdev)
1986{
1987	u16 control;
1988	int pos;
1989
1990	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1991	if (!pos)
1992		return -EINVAL;
1993
1994	pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1995	control |= PCI_PRI_CTRL_RESET;
1996	pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1997
1998	return 0;
1999}
2000
2001static int pdev_iommuv2_enable(struct pci_dev *pdev)
2002{
2003	bool reset_enable;
2004	int reqs, ret;
2005
2006	/* FIXME: Hardcode number of outstanding requests for now */
2007	reqs = 32;
2008	if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2009		reqs = 1;
2010	reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2011
2012	/* Only allow access to user-accessible pages */
2013	ret = pci_enable_pasid(pdev, 0);
2014	if (ret)
2015		goto out_err;
2016
2017	/* First reset the PRI state of the device */
2018	ret = pci_reset_pri(pdev);
2019	if (ret)
2020		goto out_err;
2021
2022	/* Enable PRI */
2023	ret = pci_enable_pri(pdev, reqs);
2024	if (ret)
2025		goto out_err;
2026
2027	if (reset_enable) {
2028		ret = pri_reset_while_enabled(pdev);
2029		if (ret)
2030			goto out_err;
2031	}
2032
2033	ret = pci_enable_ats(pdev, PAGE_SHIFT);
2034	if (ret)
2035		goto out_err;
2036
2037	return 0;
2038
2039out_err:
2040	pci_disable_pri(pdev);
2041	pci_disable_pasid(pdev);
2042
2043	return ret;
2044}
2045
2046/* FIXME: Move this to PCI code */
2047#define PCI_PRI_TLP_OFF		(1 << 15)
2048
2049static bool pci_pri_tlp_required(struct pci_dev *pdev)
2050{
2051	u16 status;
2052	int pos;
2053
2054	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2055	if (!pos)
2056		return false;
2057
2058	pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2059
2060	return (status & PCI_PRI_TLP_OFF) ? true : false;
2061}
2062
2063/*
2064 * If a device is not yet associated with a domain, this function
2065 * assigns it visible for the hardware
2066 */
2067static int attach_device(struct device *dev,
2068			 struct protection_domain *domain)
2069{
2070	struct pci_dev *pdev;
2071	struct iommu_dev_data *dev_data;
2072	unsigned long flags;
2073	int ret;
2074
2075	dev_data = get_dev_data(dev);
2076
2077	if (!dev_is_pci(dev))
2078		goto skip_ats_check;
2079
2080	pdev = to_pci_dev(dev);
2081	if (domain->flags & PD_IOMMUV2_MASK) {
2082		if (!dev_data->passthrough)
2083			return -EINVAL;
2084
2085		if (dev_data->iommu_v2) {
2086			if (pdev_iommuv2_enable(pdev) != 0)
2087				return -EINVAL;
2088
2089			dev_data->ats.enabled = true;
2090			dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
2091			dev_data->pri_tlp     = pci_pri_tlp_required(pdev);
2092		}
2093	} else if (amd_iommu_iotlb_sup &&
2094		   pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2095		dev_data->ats.enabled = true;
2096		dev_data->ats.qdep    = pci_ats_queue_depth(pdev);
2097	}
2098
2099skip_ats_check:
2100	spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2101	ret = __attach_device(dev_data, domain);
2102	spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2103
2104	/*
2105	 * We might boot into a crash-kernel here. The crashed kernel
2106	 * left the caches in the IOMMU dirty. So we have to flush
2107	 * here to evict all dirty stuff.
2108	 */
2109	domain_flush_tlb_pde(domain);
2110
2111	return ret;
2112}
2113
2114/*
2115 * Removes a device from a protection domain (unlocked)
2116 */
2117static void __detach_device(struct iommu_dev_data *dev_data)
2118{
2119	struct protection_domain *domain;
2120
2121	/*
2122	 * Must be called with IRQs disabled. Warn here to detect early
2123	 * when its not.
2124	 */
2125	WARN_ON(!irqs_disabled());
2126
2127	if (WARN_ON(!dev_data->domain))
2128		return;
2129
2130	domain = dev_data->domain;
2131
2132	spin_lock(&domain->lock);
2133
2134	do_detach(dev_data);
2135
2136	spin_unlock(&domain->lock);
2137}
2138
2139/*
2140 * Removes a device from a protection domain (with devtable_lock held)
2141 */
2142static void detach_device(struct device *dev)
2143{
2144	struct protection_domain *domain;
2145	struct iommu_dev_data *dev_data;
2146	unsigned long flags;
2147
2148	dev_data = get_dev_data(dev);
2149	domain   = dev_data->domain;
2150
2151	/* lock device table */
2152	spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2153	__detach_device(dev_data);
2154	spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2155
2156	if (!dev_is_pci(dev))
2157		return;
2158
2159	if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2160		pdev_iommuv2_disable(to_pci_dev(dev));
2161	else if (dev_data->ats.enabled)
2162		pci_disable_ats(to_pci_dev(dev));
2163
2164	dev_data->ats.enabled = false;
2165}
2166
2167static int amd_iommu_add_device(struct device *dev)
2168{
2169	struct iommu_dev_data *dev_data;
2170	struct iommu_domain *domain;
2171	struct amd_iommu *iommu;
2172	int ret, devid;
2173
2174	if (!check_device(dev) || get_dev_data(dev))
2175		return 0;
2176
2177	devid = get_device_id(dev);
2178	if (devid < 0)
2179		return devid;
2180
2181	iommu = amd_iommu_rlookup_table[devid];
2182
2183	ret = iommu_init_device(dev);
2184	if (ret) {
2185		if (ret != -ENOTSUPP)
2186			pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2187				dev_name(dev));
2188
2189		iommu_ignore_device(dev);
2190		dev->dma_ops = &dma_direct_ops;
2191		goto out;
2192	}
2193	init_iommu_group(dev);
2194
2195	dev_data = get_dev_data(dev);
2196
2197	BUG_ON(!dev_data);
2198
2199	if (iommu_pass_through || dev_data->iommu_v2)
2200		iommu_request_dm_for_dev(dev);
2201
2202	/* Domains are initialized for this device - have a look what we ended up with */
2203	domain = iommu_get_domain_for_dev(dev);
2204	if (domain->type == IOMMU_DOMAIN_IDENTITY)
2205		dev_data->passthrough = true;
2206	else
2207		dev->dma_ops = &amd_iommu_dma_ops;
2208
2209out:
2210	iommu_completion_wait(iommu);
2211
2212	return 0;
2213}
2214
2215static void amd_iommu_remove_device(struct device *dev)
2216{
2217	struct amd_iommu *iommu;
2218	int devid;
2219
2220	if (!check_device(dev))
2221		return;
2222
2223	devid = get_device_id(dev);
2224	if (devid < 0)
2225		return;
2226
2227	iommu = amd_iommu_rlookup_table[devid];
2228
2229	iommu_uninit_device(dev);
2230	iommu_completion_wait(iommu);
2231}
2232
2233static struct iommu_group *amd_iommu_device_group(struct device *dev)
2234{
2235	if (dev_is_pci(dev))
2236		return pci_device_group(dev);
2237
2238	return acpihid_device_group(dev);
2239}
2240
2241/*****************************************************************************
2242 *
2243 * The next functions belong to the dma_ops mapping/unmapping code.
2244 *
2245 *****************************************************************************/
2246
2247/*
2248 * In the dma_ops path we only have the struct device. This function
2249 * finds the corresponding IOMMU, the protection domain and the
2250 * requestor id for a given device.
2251 * If the device is not yet associated with a domain this is also done
2252 * in this function.
2253 */
2254static struct protection_domain *get_domain(struct device *dev)
2255{
2256	struct protection_domain *domain;
2257	struct iommu_domain *io_domain;
2258
2259	if (!check_device(dev))
2260		return ERR_PTR(-EINVAL);
2261
2262	domain = get_dev_data(dev)->domain;
2263	if (domain == NULL && get_dev_data(dev)->defer_attach) {
2264		get_dev_data(dev)->defer_attach = false;
2265		io_domain = iommu_get_domain_for_dev(dev);
2266		domain = to_pdomain(io_domain);
2267		attach_device(dev, domain);
2268	}
2269	if (domain == NULL)
2270		return ERR_PTR(-EBUSY);
2271
2272	if (!dma_ops_domain(domain))
2273		return ERR_PTR(-EBUSY);
2274
2275	return domain;
2276}
2277
2278static void update_device_table(struct protection_domain *domain)
2279{
2280	struct iommu_dev_data *dev_data;
2281
2282	list_for_each_entry(dev_data, &domain->dev_list, list) {
2283		set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled,
2284			      dev_data->iommu_v2);
2285
2286		if (dev_data->devid == dev_data->alias)
2287			continue;
2288
2289		/* There is an alias, update device table entry for it */
2290		set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled,
2291			      dev_data->iommu_v2);
2292	}
2293}
2294
2295static void update_domain(struct protection_domain *domain)
2296{
2297	if (!domain->updated)
2298		return;
2299
2300	update_device_table(domain);
2301
2302	domain_flush_devices(domain);
2303	domain_flush_tlb_pde(domain);
2304
2305	domain->updated = false;
2306}
2307
2308static int dir2prot(enum dma_data_direction direction)
2309{
2310	if (direction == DMA_TO_DEVICE)
2311		return IOMMU_PROT_IR;
2312	else if (direction == DMA_FROM_DEVICE)
2313		return IOMMU_PROT_IW;
2314	else if (direction == DMA_BIDIRECTIONAL)
2315		return IOMMU_PROT_IW | IOMMU_PROT_IR;
2316	else
2317		return 0;
2318}
2319
2320/*
2321 * This function contains common code for mapping of a physically
2322 * contiguous memory region into DMA address space. It is used by all
2323 * mapping functions provided with this IOMMU driver.
2324 * Must be called with the domain lock held.
2325 */
2326static dma_addr_t __map_single(struct device *dev,
2327			       struct dma_ops_domain *dma_dom,
2328			       phys_addr_t paddr,
2329			       size_t size,
2330			       enum dma_data_direction direction,
2331			       u64 dma_mask)
2332{
2333	dma_addr_t offset = paddr & ~PAGE_MASK;
2334	dma_addr_t address, start, ret;
2335	unsigned int pages;
2336	int prot = 0;
2337	int i;
2338
2339	pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2340	paddr &= PAGE_MASK;
2341
2342	address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2343	if (address == AMD_IOMMU_MAPPING_ERROR)
2344		goto out;
2345
2346	prot = dir2prot(direction);
2347
2348	start = address;
2349	for (i = 0; i < pages; ++i) {
2350		ret = iommu_map_page(&dma_dom->domain, start, paddr,
2351				     PAGE_SIZE, prot, GFP_ATOMIC);
2352		if (ret)
2353			goto out_unmap;
2354
2355		paddr += PAGE_SIZE;
2356		start += PAGE_SIZE;
2357	}
2358	address += offset;
2359
2360	if (unlikely(amd_iommu_np_cache)) {
2361		domain_flush_pages(&dma_dom->domain, address, size);
2362		domain_flush_complete(&dma_dom->domain);
2363	}
2364
2365out:
2366	return address;
2367
2368out_unmap:
2369
2370	for (--i; i >= 0; --i) {
2371		start -= PAGE_SIZE;
2372		iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2373	}
2374
2375	domain_flush_tlb(&dma_dom->domain);
2376	domain_flush_complete(&dma_dom->domain);
2377
2378	dma_ops_free_iova(dma_dom, address, pages);
2379
2380	return AMD_IOMMU_MAPPING_ERROR;
2381}
2382
2383/*
2384 * Does the reverse of the __map_single function. Must be called with
2385 * the domain lock held too
2386 */
2387static void __unmap_single(struct dma_ops_domain *dma_dom,
2388			   dma_addr_t dma_addr,
2389			   size_t size,
2390			   int dir)
2391{
2392	dma_addr_t i, start;
2393	unsigned int pages;
2394
2395	pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2396	dma_addr &= PAGE_MASK;
2397	start = dma_addr;
2398
2399	for (i = 0; i < pages; ++i) {
2400		iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2401		start += PAGE_SIZE;
2402	}
2403
2404	if (amd_iommu_unmap_flush) {
2405		dma_ops_free_iova(dma_dom, dma_addr, pages);
2406		domain_flush_tlb(&dma_dom->domain);
2407		domain_flush_complete(&dma_dom->domain);
2408	} else {
2409		pages = __roundup_pow_of_two(pages);
2410		queue_iova(&dma_dom->iovad, dma_addr >> PAGE_SHIFT, pages, 0);
2411	}
2412}
2413
2414/*
2415 * The exported map_single function for dma_ops.
2416 */
2417static dma_addr_t map_page(struct device *dev, struct page *page,
2418			   unsigned long offset, size_t size,
2419			   enum dma_data_direction dir,
2420			   unsigned long attrs)
2421{
2422	phys_addr_t paddr = page_to_phys(page) + offset;
2423	struct protection_domain *domain;
2424	struct dma_ops_domain *dma_dom;
2425	u64 dma_mask;
2426
2427	domain = get_domain(dev);
2428	if (PTR_ERR(domain) == -EINVAL)
2429		return (dma_addr_t)paddr;
2430	else if (IS_ERR(domain))
2431		return AMD_IOMMU_MAPPING_ERROR;
2432
2433	dma_mask = *dev->dma_mask;
2434	dma_dom = to_dma_ops_domain(domain);
2435
2436	return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2437}
2438
2439/*
2440 * The exported unmap_single function for dma_ops.
2441 */
2442static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2443		       enum dma_data_direction dir, unsigned long attrs)
2444{
2445	struct protection_domain *domain;
2446	struct dma_ops_domain *dma_dom;
2447
2448	domain = get_domain(dev);
2449	if (IS_ERR(domain))
2450		return;
2451
2452	dma_dom = to_dma_ops_domain(domain);
2453
2454	__unmap_single(dma_dom, dma_addr, size, dir);
2455}
2456
2457static int sg_num_pages(struct device *dev,
2458			struct scatterlist *sglist,
2459			int nelems)
2460{
2461	unsigned long mask, boundary_size;
2462	struct scatterlist *s;
2463	int i, npages = 0;
2464
2465	mask          = dma_get_seg_boundary(dev);
2466	boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2467				   1UL << (BITS_PER_LONG - PAGE_SHIFT);
2468
2469	for_each_sg(sglist, s, nelems, i) {
2470		int p, n;
2471
2472		s->dma_address = npages << PAGE_SHIFT;
2473		p = npages % boundary_size;
2474		n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2475		if (p + n > boundary_size)
2476			npages += boundary_size - p;
2477		npages += n;
2478	}
2479
2480	return npages;
2481}
2482
2483/*
2484 * The exported map_sg function for dma_ops (handles scatter-gather
2485 * lists).
2486 */
2487static int map_sg(struct device *dev, struct scatterlist *sglist,
2488		  int nelems, enum dma_data_direction direction,
2489		  unsigned long attrs)
2490{
2491	int mapped_pages = 0, npages = 0, prot = 0, i;
2492	struct protection_domain *domain;
2493	struct dma_ops_domain *dma_dom;
2494	struct scatterlist *s;
2495	unsigned long address;
2496	u64 dma_mask;
2497
2498	domain = get_domain(dev);
2499	if (IS_ERR(domain))
2500		return 0;
2501
2502	dma_dom  = to_dma_ops_domain(domain);
2503	dma_mask = *dev->dma_mask;
2504
2505	npages = sg_num_pages(dev, sglist, nelems);
2506
2507	address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2508	if (address == AMD_IOMMU_MAPPING_ERROR)
2509		goto out_err;
2510
2511	prot = dir2prot(direction);
2512
2513	/* Map all sg entries */
2514	for_each_sg(sglist, s, nelems, i) {
2515		int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2516
2517		for (j = 0; j < pages; ++j) {
2518			unsigned long bus_addr, phys_addr;
2519			int ret;
2520
2521			bus_addr  = address + s->dma_address + (j << PAGE_SHIFT);
2522			phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2523			ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2524			if (ret)
2525				goto out_unmap;
2526
2527			mapped_pages += 1;
2528		}
2529	}
2530
2531	/* Everything is mapped - write the right values into s->dma_address */
2532	for_each_sg(sglist, s, nelems, i) {
2533		s->dma_address += address + s->offset;
2534		s->dma_length   = s->length;
2535	}
2536
2537	return nelems;
2538
2539out_unmap:
2540	pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2541	       dev_name(dev), npages);
2542
2543	for_each_sg(sglist, s, nelems, i) {
2544		int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2545
2546		for (j = 0; j < pages; ++j) {
2547			unsigned long bus_addr;
2548
2549			bus_addr  = address + s->dma_address + (j << PAGE_SHIFT);
2550			iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2551
2552			if (--mapped_pages)
2553				goto out_free_iova;
2554		}
2555	}
2556
2557out_free_iova:
2558	free_iova_fast(&dma_dom->iovad, address, npages);
2559
2560out_err:
2561	return 0;
2562}
2563
2564/*
2565 * The exported map_sg function for dma_ops (handles scatter-gather
2566 * lists).
2567 */
2568static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2569		     int nelems, enum dma_data_direction dir,
2570		     unsigned long attrs)
2571{
2572	struct protection_domain *domain;
2573	struct dma_ops_domain *dma_dom;
2574	unsigned long startaddr;
2575	int npages = 2;
2576
2577	domain = get_domain(dev);
2578	if (IS_ERR(domain))
2579		return;
2580
2581	startaddr = sg_dma_address(sglist) & PAGE_MASK;
2582	dma_dom   = to_dma_ops_domain(domain);
2583	npages    = sg_num_pages(dev, sglist, nelems);
2584
2585	__unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2586}
2587
2588/*
2589 * The exported alloc_coherent function for dma_ops.
2590 */
2591static void *alloc_coherent(struct device *dev, size_t size,
2592			    dma_addr_t *dma_addr, gfp_t flag,
2593			    unsigned long attrs)
2594{
2595	u64 dma_mask = dev->coherent_dma_mask;
2596	struct protection_domain *domain = get_domain(dev);
2597	bool is_direct = false;
2598	void *virt_addr;
2599
2600	if (IS_ERR(domain)) {
2601		if (PTR_ERR(domain) != -EINVAL)
2602			return NULL;
2603		is_direct = true;
2604	}
2605
2606	virt_addr = dma_direct_alloc(dev, size, dma_addr, flag, attrs);
2607	if (!virt_addr || is_direct)
2608		return virt_addr;
2609
2610	if (!dma_mask)
2611		dma_mask = *dev->dma_mask;
2612
2613	*dma_addr = __map_single(dev, to_dma_ops_domain(domain),
2614			virt_to_phys(virt_addr), PAGE_ALIGN(size),
2615			DMA_BIDIRECTIONAL, dma_mask);
2616	if (*dma_addr == AMD_IOMMU_MAPPING_ERROR)
2617		goto out_free;
2618	return virt_addr;
2619
2620out_free:
2621	dma_direct_free(dev, size, virt_addr, *dma_addr, attrs);
2622	return NULL;
2623}
2624
2625/*
2626 * The exported free_coherent function for dma_ops.
2627 */
2628static void free_coherent(struct device *dev, size_t size,
2629			  void *virt_addr, dma_addr_t dma_addr,
2630			  unsigned long attrs)
2631{
2632	struct protection_domain *domain = get_domain(dev);
2633
2634	size = PAGE_ALIGN(size);
2635
2636	if (!IS_ERR(domain)) {
2637		struct dma_ops_domain *dma_dom = to_dma_ops_domain(domain);
2638
2639		__unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2640	}
2641
2642	dma_direct_free(dev, size, virt_addr, dma_addr, attrs);
2643}
2644
2645/*
2646 * This function is called by the DMA layer to find out if we can handle a
2647 * particular device. It is part of the dma_ops.
2648 */
2649static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2650{
2651	if (!dma_direct_supported(dev, mask))
2652		return 0;
2653	return check_device(dev);
2654}
2655
2656static int amd_iommu_mapping_error(struct device *dev, dma_addr_t dma_addr)
2657{
2658	return dma_addr == AMD_IOMMU_MAPPING_ERROR;
2659}
2660
2661static const struct dma_map_ops amd_iommu_dma_ops = {
2662	.alloc		= alloc_coherent,
2663	.free		= free_coherent,
2664	.map_page	= map_page,
2665	.unmap_page	= unmap_page,
2666	.map_sg		= map_sg,
2667	.unmap_sg	= unmap_sg,
2668	.dma_supported	= amd_iommu_dma_supported,
2669	.mapping_error	= amd_iommu_mapping_error,
2670};
2671
2672static int init_reserved_iova_ranges(void)
2673{
2674	struct pci_dev *pdev = NULL;
2675	struct iova *val;
2676
2677	init_iova_domain(&reserved_iova_ranges, PAGE_SIZE, IOVA_START_PFN);
2678
2679	lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2680			  &reserved_rbtree_key);
2681
2682	/* MSI memory range */
2683	val = reserve_iova(&reserved_iova_ranges,
2684			   IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2685	if (!val) {
2686		pr_err("Reserving MSI range failed\n");
2687		return -ENOMEM;
2688	}
2689
2690	/* HT memory range */
2691	val = reserve_iova(&reserved_iova_ranges,
2692			   IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2693	if (!val) {
2694		pr_err("Reserving HT range failed\n");
2695		return -ENOMEM;
2696	}
2697
2698	/*
2699	 * Memory used for PCI resources
2700	 * FIXME: Check whether we can reserve the PCI-hole completly
2701	 */
2702	for_each_pci_dev(pdev) {
2703		int i;
2704
2705		for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2706			struct resource *r = &pdev->resource[i];
2707
2708			if (!(r->flags & IORESOURCE_MEM))
2709				continue;
2710
2711			val = reserve_iova(&reserved_iova_ranges,
2712					   IOVA_PFN(r->start),
2713					   IOVA_PFN(r->end));
2714			if (!val) {
2715				pr_err("Reserve pci-resource range failed\n");
2716				return -ENOMEM;
2717			}
2718		}
2719	}
2720
2721	return 0;
2722}
2723
2724int __init amd_iommu_init_api(void)
2725{
2726	int ret, err = 0;
2727
2728	ret = iova_cache_get();
2729	if (ret)
2730		return ret;
2731
2732	ret = init_reserved_iova_ranges();
2733	if (ret)
2734		return ret;
2735
2736	err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2737	if (err)
2738		return err;
2739#ifdef CONFIG_ARM_AMBA
2740	err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2741	if (err)
2742		return err;
2743#endif
2744	err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2745	if (err)
2746		return err;
2747
2748	return 0;
2749}
2750
2751int __init amd_iommu_init_dma_ops(void)
2752{
2753	swiotlb        = (iommu_pass_through || sme_me_mask) ? 1 : 0;
2754	iommu_detected = 1;
2755
2756	/*
2757	 * In case we don't initialize SWIOTLB (actually the common case
2758	 * when AMD IOMMU is enabled and SME is not active), make sure there
2759	 * are global dma_ops set as a fall-back for devices not handled by
2760	 * this driver (for example non-PCI devices). When SME is active,
2761	 * make sure that swiotlb variable remains set so the global dma_ops
2762	 * continue to be SWIOTLB.
2763	 */
2764	if (!swiotlb)
2765		dma_ops = &dma_direct_ops;
2766
2767	if (amd_iommu_unmap_flush)
2768		pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2769	else
2770		pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2771
2772	return 0;
2773
2774}
2775
2776/*****************************************************************************
2777 *
2778 * The following functions belong to the exported interface of AMD IOMMU
2779 *
2780 * This interface allows access to lower level functions of the IOMMU
2781 * like protection domain handling and assignement of devices to domains
2782 * which is not possible with the dma_ops interface.
2783 *
2784 *****************************************************************************/
2785
2786static void cleanup_domain(struct protection_domain *domain)
2787{
2788	struct iommu_dev_data *entry;
2789	unsigned long flags;
2790
2791	spin_lock_irqsave(&amd_iommu_devtable_lock, flags);
2792
2793	while (!list_empty(&domain->dev_list)) {
2794		entry = list_first_entry(&domain->dev_list,
2795					 struct iommu_dev_data, list);
2796		__detach_device(entry);
2797	}
2798
2799	spin_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2800}
2801
2802static void protection_domain_free(struct protection_domain *domain)
2803{
2804	if (!domain)
2805		return;
2806
2807	del_domain_from_list(domain);
2808
2809	if (domain->id)
2810		domain_id_free(domain->id);
2811
2812	kfree(domain);
2813}
2814
2815static int protection_domain_init(struct protection_domain *domain)
2816{
2817	spin_lock_init(&domain->lock);
2818	mutex_init(&domain->api_lock);
2819	domain->id = domain_id_alloc();
2820	if (!domain->id)
2821		return -ENOMEM;
2822	INIT_LIST_HEAD(&domain->dev_list);
2823
2824	return 0;
2825}
2826
2827static struct protection_domain *protection_domain_alloc(void)
2828{
2829	struct protection_domain *domain;
2830
2831	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2832	if (!domain)
2833		return NULL;
2834
2835	if (protection_domain_init(domain))
2836		goto out_err;
2837
2838	add_domain_to_list(domain);
2839
2840	return domain;
2841
2842out_err:
2843	kfree(domain);
2844
2845	return NULL;
2846}
2847
2848static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2849{
2850	struct protection_domain *pdomain;
2851	struct dma_ops_domain *dma_domain;
2852
2853	switch (type) {
2854	case IOMMU_DOMAIN_UNMANAGED:
2855		pdomain = protection_domain_alloc();
2856		if (!pdomain)
2857			return NULL;
2858
2859		pdomain->mode    = PAGE_MODE_3_LEVEL;
2860		pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2861		if (!pdomain->pt_root) {
2862			protection_domain_free(pdomain);
2863			return NULL;
2864		}
2865
2866		pdomain->domain.geometry.aperture_start = 0;
2867		pdomain->domain.geometry.aperture_end   = ~0ULL;
2868		pdomain->domain.geometry.force_aperture = true;
2869
2870		break;
2871	case IOMMU_DOMAIN_DMA:
2872		dma_domain = dma_ops_domain_alloc();
2873		if (!dma_domain) {
2874			pr_err("AMD-Vi: Failed to allocate\n");
2875			return NULL;
2876		}
2877		pdomain = &dma_domain->domain;
2878		break;
2879	case IOMMU_DOMAIN_IDENTITY:
2880		pdomain = protection_domain_alloc();
2881		if (!pdomain)
2882			return NULL;
2883
2884		pdomain->mode = PAGE_MODE_NONE;
2885		break;
2886	default:
2887		return NULL;
2888	}
2889
2890	return &pdomain->domain;
2891}
2892
2893static void amd_iommu_domain_free(struct iommu_domain *dom)
2894{
2895	struct protection_domain *domain;
2896	struct dma_ops_domain *dma_dom;
2897
2898	domain = to_pdomain(dom);
2899
2900	if (domain->dev_cnt > 0)
2901		cleanup_domain(domain);
2902
2903	BUG_ON(domain->dev_cnt != 0);
2904
2905	if (!dom)
2906		return;
2907
2908	switch (dom->type) {
2909	case IOMMU_DOMAIN_DMA:
2910		/* Now release the domain */
2911		dma_dom = to_dma_ops_domain(domain);
2912		dma_ops_domain_free(dma_dom);
2913		break;
2914	default:
2915		if (domain->mode != PAGE_MODE_NONE)
2916			free_pagetable(domain);
2917
2918		if (domain->flags & PD_IOMMUV2_MASK)
2919			free_gcr3_table(domain);
2920
2921		protection_domain_free(domain);
2922		break;
2923	}
2924}
2925
2926static void amd_iommu_detach_device(struct iommu_domain *dom,
2927				    struct device *dev)
2928{
2929	struct iommu_dev_data *dev_data = dev->archdata.iommu;
2930	struct amd_iommu *iommu;
2931	int devid;
2932
2933	if (!check_device(dev))
2934		return;
2935
2936	devid = get_device_id(dev);
2937	if (devid < 0)
2938		return;
2939
2940	if (dev_data->domain != NULL)
2941		detach_device(dev);
2942
2943	iommu = amd_iommu_rlookup_table[devid];
2944	if (!iommu)
2945		return;
2946
2947#ifdef CONFIG_IRQ_REMAP
2948	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
2949	    (dom->type == IOMMU_DOMAIN_UNMANAGED))
2950		dev_data->use_vapic = 0;
2951#endif
2952
2953	iommu_completion_wait(iommu);
2954}
2955
2956static int amd_iommu_attach_device(struct iommu_domain *dom,
2957				   struct device *dev)
2958{
2959	struct protection_domain *domain = to_pdomain(dom);
2960	struct iommu_dev_data *dev_data;
2961	struct amd_iommu *iommu;
2962	int ret;
2963
2964	if (!check_device(dev))
2965		return -EINVAL;
2966
2967	dev_data = dev->archdata.iommu;
2968
2969	iommu = amd_iommu_rlookup_table[dev_data->devid];
2970	if (!iommu)
2971		return -EINVAL;
2972
2973	if (dev_data->domain)
2974		detach_device(dev);
2975
2976	ret = attach_device(dev, domain);
2977
2978#ifdef CONFIG_IRQ_REMAP
2979	if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
2980		if (dom->type == IOMMU_DOMAIN_UNMANAGED)
2981			dev_data->use_vapic = 1;
2982		else
2983			dev_data->use_vapic = 0;
2984	}
2985#endif
2986
2987	iommu_completion_wait(iommu);
2988
2989	return ret;
2990}
2991
2992static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2993			 phys_addr_t paddr, size_t page_size, int iommu_prot)
2994{
2995	struct protection_domain *domain = to_pdomain(dom);
2996	int prot = 0;
2997	int ret;
2998
2999	if (domain->mode == PAGE_MODE_NONE)
3000		return -EINVAL;
3001
3002	if (iommu_prot & IOMMU_READ)
3003		prot |= IOMMU_PROT_IR;
3004	if (iommu_prot & IOMMU_WRITE)
3005		prot |= IOMMU_PROT_IW;
3006
3007	mutex_lock(&domain->api_lock);
3008	ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3009	mutex_unlock(&domain->api_lock);
3010
3011	return ret;
3012}
3013
3014static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3015			   size_t page_size)
3016{
3017	struct protection_domain *domain = to_pdomain(dom);
3018	size_t unmap_size;
3019
3020	if (domain->mode == PAGE_MODE_NONE)
3021		return 0;
3022
3023	mutex_lock(&domain->api_lock);
3024	unmap_size = iommu_unmap_page(domain, iova, page_size);
3025	mutex_unlock(&domain->api_lock);
3026
3027	return unmap_size;
3028}
3029
3030static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3031					  dma_addr_t iova)
3032{
3033	struct protection_domain *domain = to_pdomain(dom);
3034	unsigned long offset_mask, pte_pgsize;
3035	u64 *pte, __pte;
3036
3037	if (domain->mode == PAGE_MODE_NONE)
3038		return iova;
3039
3040	pte = fetch_pte(domain, iova, &pte_pgsize);
3041
3042	if (!pte || !IOMMU_PTE_PRESENT(*pte))
3043		return 0;
3044
3045	offset_mask = pte_pgsize - 1;
3046	__pte	    = *pte & PM_ADDR_MASK;
3047
3048	return (__pte & ~offset_mask) | (iova & offset_mask);
3049}
3050
3051static bool amd_iommu_capable(enum iommu_cap cap)
3052{
3053	switch (cap) {
3054	case IOMMU_CAP_CACHE_COHERENCY:
3055		return true;
3056	case IOMMU_CAP_INTR_REMAP:
3057		return (irq_remapping_enabled == 1);
3058	case IOMMU_CAP_NOEXEC:
3059		return false;
3060	}
3061
3062	return false;
3063}
3064
3065static void amd_iommu_get_resv_regions(struct device *dev,
3066				       struct list_head *head)
3067{
3068	struct iommu_resv_region *region;
3069	struct unity_map_entry *entry;
3070	int devid;
3071
3072	devid = get_device_id(dev);
3073	if (devid < 0)
3074		return;
3075
3076	list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3077		size_t length;
3078		int prot = 0;
3079
3080		if (devid < entry->devid_start || devid > entry->devid_end)
3081			continue;
3082
3083		length = entry->address_end - entry->address_start;
3084		if (entry->prot & IOMMU_PROT_IR)
3085			prot |= IOMMU_READ;
3086		if (entry->prot & IOMMU_PROT_IW)
3087			prot |= IOMMU_WRITE;
3088
3089		region = iommu_alloc_resv_region(entry->address_start,
3090						 length, prot,
3091						 IOMMU_RESV_DIRECT);
3092		if (!region) {
3093			pr_err("Out of memory allocating dm-regions for %s\n",
3094				dev_name(dev));
3095			return;
3096		}
3097		list_add_tail(&region->list, head);
3098	}
3099
3100	region = iommu_alloc_resv_region(MSI_RANGE_START,
3101					 MSI_RANGE_END - MSI_RANGE_START + 1,
3102					 0, IOMMU_RESV_MSI);
3103	if (!region)
3104		return;
3105	list_add_tail(&region->list, head);
3106
3107	region = iommu_alloc_resv_region(HT_RANGE_START,
3108					 HT_RANGE_END - HT_RANGE_START + 1,
3109					 0, IOMMU_RESV_RESERVED);
3110	if (!region)
3111		return;
3112	list_add_tail(&region->list, head);
3113}
3114
3115static void amd_iommu_put_resv_regions(struct device *dev,
3116				     struct list_head *head)
3117{
3118	struct iommu_resv_region *entry, *next;
3119
3120	list_for_each_entry_safe(entry, next, head, list)
3121		kfree(entry);
3122}
3123
3124static void amd_iommu_apply_resv_region(struct device *dev,
3125				      struct iommu_domain *domain,
3126				      struct iommu_resv_region *region)
3127{
3128	struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3129	unsigned long start, end;
3130
3131	start = IOVA_PFN(region->start);
3132	end   = IOVA_PFN(region->start + region->length - 1);
3133
3134	WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3135}
3136
3137static bool amd_iommu_is_attach_deferred(struct iommu_domain *domain,
3138					 struct device *dev)
3139{
3140	struct iommu_dev_data *dev_data = dev->archdata.iommu;
3141	return dev_data->defer_attach;
3142}
3143
3144static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain)
3145{
3146	struct protection_domain *dom = to_pdomain(domain);
3147
3148	domain_flush_tlb_pde(dom);
3149	domain_flush_complete(dom);
3150}
3151
3152static void amd_iommu_iotlb_range_add(struct iommu_domain *domain,
3153				      unsigned long iova, size_t size)
3154{
3155}
3156
3157const struct iommu_ops amd_iommu_ops = {
3158	.capable = amd_iommu_capable,
3159	.domain_alloc = amd_iommu_domain_alloc,
3160	.domain_free  = amd_iommu_domain_free,
3161	.attach_dev = amd_iommu_attach_device,
3162	.detach_dev = amd_iommu_detach_device,
3163	.map = amd_iommu_map,
3164	.unmap = amd_iommu_unmap,
3165	.map_sg = default_iommu_map_sg,
3166	.iova_to_phys = amd_iommu_iova_to_phys,
3167	.add_device = amd_iommu_add_device,
3168	.remove_device = amd_iommu_remove_device,
3169	.device_group = amd_iommu_device_group,
3170	.get_resv_regions = amd_iommu_get_resv_regions,
3171	.put_resv_regions = amd_iommu_put_resv_regions,
3172	.apply_resv_region = amd_iommu_apply_resv_region,
3173	.is_attach_deferred = amd_iommu_is_attach_deferred,
3174	.pgsize_bitmap	= AMD_IOMMU_PGSIZES,
3175	.flush_iotlb_all = amd_iommu_flush_iotlb_all,
3176	.iotlb_range_add = amd_iommu_iotlb_range_add,
3177	.iotlb_sync = amd_iommu_flush_iotlb_all,
3178};
3179
3180/*****************************************************************************
3181 *
3182 * The next functions do a basic initialization of IOMMU for pass through
3183 * mode
3184 *
3185 * In passthrough mode the IOMMU is initialized and enabled but not used for
3186 * DMA-API translation.
3187 *
3188 *****************************************************************************/
3189
3190/* IOMMUv2 specific functions */
3191int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3192{
3193	return atomic_notifier_chain_register(&ppr_notifier, nb);
3194}
3195EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3196
3197int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3198{
3199	return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3200}
3201EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3202
3203void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3204{
3205	struct protection_domain *domain = to_pdomain(dom);
3206	unsigned long flags;
3207
3208	spin_lock_irqsave(&domain->lock, flags);
3209
3210	/* Update data structure */
3211	domain->mode    = PAGE_MODE_NONE;
3212	domain->updated = true;
3213
3214	/* Make changes visible to IOMMUs */
3215	update_domain(domain);
3216
3217	/* Page-table is not visible to IOMMU anymore, so free it */
3218	free_pagetable(domain);
3219
3220	spin_unlock_irqrestore(&domain->lock, flags);
3221}
3222EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3223
3224int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3225{
3226	struct protection_domain *domain = to_pdomain(dom);
3227	unsigned long flags;
3228	int levels, ret;
3229
3230	if (pasids <= 0 || pasids > (PASID_MASK + 1))
3231		return -EINVAL;
3232
3233	/* Number of GCR3 table levels required */
3234	for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3235		levels += 1;
3236
3237	if (levels > amd_iommu_max_glx_val)
3238		return -EINVAL;
3239
3240	spin_lock_irqsave(&domain->lock, flags);
3241
3242	/*
3243	 * Save us all sanity checks whether devices already in the
3244	 * domain support IOMMUv2. Just force that the domain has no
3245	 * devices attached when it is switched into IOMMUv2 mode.
3246	 */
3247	ret = -EBUSY;
3248	if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3249		goto out;
3250
3251	ret = -ENOMEM;
3252	domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3253	if (domain->gcr3_tbl == NULL)
3254		goto out;
3255
3256	domain->glx      = levels;
3257	domain->flags   |= PD_IOMMUV2_MASK;
3258	domain->updated  = true;
3259
3260	update_domain(domain);
3261
3262	ret = 0;
3263
3264out:
3265	spin_unlock_irqrestore(&domain->lock, flags);
3266
3267	return ret;
3268}
3269EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3270
3271static int __flush_pasid(struct protection_domain *domain, int pasid,
3272			 u64 address, bool size)
3273{
3274	struct iommu_dev_data *dev_data;
3275	struct iommu_cmd cmd;
3276	int i, ret;
3277
3278	if (!(domain->flags & PD_IOMMUV2_MASK))
3279		return -EINVAL;
3280
3281	build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3282
3283	/*
3284	 * IOMMU TLB needs to be flushed before Device TLB to
3285	 * prevent device TLB refill from IOMMU TLB
3286	 */
3287	for (i = 0; i < amd_iommu_get_num_iommus(); ++i) {
3288		if (domain->dev_iommu[i] == 0)
3289			continue;
3290
3291		ret = iommu_queue_command(amd_iommus[i], &cmd);
3292		if (ret != 0)
3293			goto out;
3294	}
3295
3296	/* Wait until IOMMU TLB flushes are complete */
3297	domain_flush_complete(domain);
3298
3299	/* Now flush device TLBs */
3300	list_for_each_entry(dev_data, &domain->dev_list, list) {
3301		struct amd_iommu *iommu;
3302		int qdep;
3303
3304		/*
3305		   There might be non-IOMMUv2 capable devices in an IOMMUv2
3306		 * domain.
3307		 */
3308		if (!dev_data->ats.enabled)
3309			continue;
3310
3311		qdep  = dev_data->ats.qdep;
3312		iommu = amd_iommu_rlookup_table[dev_data->devid];
3313
3314		build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3315				      qdep, address, size);
3316
3317		ret = iommu_queue_command(iommu, &cmd);
3318		if (ret != 0)
3319			goto out;
3320	}
3321
3322	/* Wait until all device TLBs are flushed */
3323	domain_flush_complete(domain);
3324
3325	ret = 0;
3326
3327out:
3328
3329	return ret;
3330}
3331
3332static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3333				  u64 address)
3334{
3335	return __flush_pasid(domain, pasid, address, false);
3336}
3337
3338int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3339			 u64 address)
3340{
3341	struct protection_domain *domain = to_pdomain(dom);
3342	unsigned long flags;
3343	int ret;
3344
3345	spin_lock_irqsave(&domain->lock, flags);
3346	ret = __amd_iommu_flush_page(domain, pasid, address);
3347	spin_unlock_irqrestore(&domain->lock, flags);
3348
3349	return ret;
3350}
3351EXPORT_SYMBOL(amd_iommu_flush_page);
3352
3353static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3354{
3355	return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3356			     true);
3357}
3358
3359int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3360{
3361	struct protection_domain *domain = to_pdomain(dom);
3362	unsigned long flags;
3363	int ret;
3364
3365	spin_lock_irqsave(&domain->lock, flags);
3366	ret = __amd_iommu_flush_tlb(domain, pasid);
3367	spin_unlock_irqrestore(&domain->lock, flags);
3368
3369	return ret;
3370}
3371EXPORT_SYMBOL(amd_iommu_flush_tlb);
3372
3373static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3374{
3375	int index;
3376	u64 *pte;
3377
3378	while (true) {
3379
3380		index = (pasid >> (9 * level)) & 0x1ff;
3381		pte   = &root[index];
3382
3383		if (level == 0)
3384			break;
3385
3386		if (!(*pte & GCR3_VALID)) {
3387			if (!alloc)
3388				return NULL;
3389
3390			root = (void *)get_zeroed_page(GFP_ATOMIC);
3391			if (root == NULL)
3392				return NULL;
3393
3394			*pte = iommu_virt_to_phys(root) | GCR3_VALID;
3395		}
3396
3397		root = iommu_phys_to_virt(*pte & PAGE_MASK);
3398
3399		level -= 1;
3400	}
3401
3402	return pte;
3403}
3404
3405static int __set_gcr3(struct protection_domain *domain, int pasid,
3406		      unsigned long cr3)
3407{
3408	u64 *pte;
3409
3410	if (domain->mode != PAGE_MODE_NONE)
3411		return -EINVAL;
3412
3413	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3414	if (pte == NULL)
3415		return -ENOMEM;
3416
3417	*pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3418
3419	return __amd_iommu_flush_tlb(domain, pasid);
3420}
3421
3422static int __clear_gcr3(struct protection_domain *domain, int pasid)
3423{
3424	u64 *pte;
3425
3426	if (domain->mode != PAGE_MODE_NONE)
3427		return -EINVAL;
3428
3429	pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3430	if (pte == NULL)
3431		return 0;
3432
3433	*pte = 0;
3434
3435	return __amd_iommu_flush_tlb(domain, pasid);
3436}
3437
3438int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3439			      unsigned long cr3)
3440{
3441	struct protection_domain *domain = to_pdomain(dom);
3442	unsigned long flags;
3443	int ret;
3444
3445	spin_lock_irqsave(&domain->lock, flags);
3446	ret = __set_gcr3(domain, pasid, cr3);
3447	spin_unlock_irqrestore(&domain->lock, flags);
3448
3449	return ret;
3450}
3451EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3452
3453int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3454{
3455	struct protection_domain *domain = to_pdomain(dom);
3456	unsigned long flags;
3457	int ret;
3458
3459	spin_lock_irqsave(&domain->lock, flags);
3460	ret = __clear_gcr3(domain, pasid);
3461	spin_unlock_irqrestore(&domain->lock, flags);
3462
3463	return ret;
3464}
3465EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3466
3467int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3468			   int status, int tag)
3469{
3470	struct iommu_dev_data *dev_data;
3471	struct amd_iommu *iommu;
3472	struct iommu_cmd cmd;
3473
3474	dev_data = get_dev_data(&pdev->dev);
3475	iommu    = amd_iommu_rlookup_table[dev_data->devid];
3476
3477	build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3478			   tag, dev_data->pri_tlp);
3479
3480	return iommu_queue_command(iommu, &cmd);
3481}
3482EXPORT_SYMBOL(amd_iommu_complete_ppr);
3483
3484struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3485{
3486	struct protection_domain *pdomain;
3487
3488	pdomain = get_domain(&pdev->dev);
3489	if (IS_ERR(pdomain))
3490		return NULL;
3491
3492	/* Only return IOMMUv2 domains */
3493	if (!(pdomain->flags & PD_IOMMUV2_MASK))
3494		return NULL;
3495
3496	return &pdomain->domain;
3497}
3498EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3499
3500void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3501{
3502	struct iommu_dev_data *dev_data;
3503
3504	if (!amd_iommu_v2_supported())
3505		return;
3506
3507	dev_data = get_dev_data(&pdev->dev);
3508	dev_data->errata |= (1 << erratum);
3509}
3510EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3511
3512int amd_iommu_device_info(struct pci_dev *pdev,
3513                          struct amd_iommu_device_info *info)
3514{
3515	int max_pasids;
3516	int pos;
3517
3518	if (pdev == NULL || info == NULL)
3519		return -EINVAL;
3520
3521	if (!amd_iommu_v2_supported())
3522		return -EINVAL;
3523
3524	memset(info, 0, sizeof(*info));
3525
3526	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3527	if (pos)
3528		info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3529
3530	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3531	if (pos)
3532		info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3533
3534	pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3535	if (pos) {
3536		int features;
3537
3538		max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3539		max_pasids = min(max_pasids, (1 << 20));
3540
3541		info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3542		info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3543
3544		features = pci_pasid_features(pdev);
3545		if (features & PCI_PASID_CAP_EXEC)
3546			info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3547		if (features & PCI_PASID_CAP_PRIV)
3548			info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3549	}
3550
3551	return 0;
3552}
3553EXPORT_SYMBOL(amd_iommu_device_info);
3554
3555#ifdef CONFIG_IRQ_REMAP
3556
3557/*****************************************************************************
3558 *
3559 * Interrupt Remapping Implementation
3560 *
3561 *****************************************************************************/
3562
3563static struct irq_chip amd_ir_chip;
3564static DEFINE_SPINLOCK(iommu_table_lock);
3565
3566static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3567{
3568	u64 dte;
3569
3570	dte	= amd_iommu_dev_table[devid].data[2];
3571	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
3572	dte	|= iommu_virt_to_phys(table->table);
3573	dte	|= DTE_IRQ_REMAP_INTCTL;
3574	dte	|= DTE_IRQ_TABLE_LEN;
3575	dte	|= DTE_IRQ_REMAP_ENABLE;
3576
3577	amd_iommu_dev_table[devid].data[2] = dte;
3578}
3579
3580static struct irq_remap_table *get_irq_table(u16 devid)
3581{
3582	struct irq_remap_table *table;
3583
3584	if (WARN_ONCE(!amd_iommu_rlookup_table[devid],
3585		      "%s: no iommu for devid %x\n", __func__, devid))
3586		return NULL;
3587
3588	table = irq_lookup_table[devid];
3589	if (WARN_ONCE(!table, "%s: no table for devid %x\n", __func__, devid))
3590		return NULL;
3591
3592	return table;
3593}
3594
3595static struct irq_remap_table *__alloc_irq_table(void)
3596{
3597	struct irq_remap_table *table;
3598
3599	table = kzalloc(sizeof(*table), GFP_KERNEL);
3600	if (!table)
3601		return NULL;
3602
3603	table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL);
3604	if (!table->table) {
3605		kfree(table);
3606		return NULL;
3607	}
3608	raw_spin_lock_init(&table->lock);
3609
3610	if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3611		memset(table->table, 0,
3612		       MAX_IRQS_PER_TABLE * sizeof(u32));
3613	else
3614		memset(table->table, 0,
3615		       (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3616	return table;
3617}
3618
3619static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid,
3620				  struct irq_remap_table *table)
3621{
3622	irq_lookup_table[devid] = table;
3623	set_dte_irq_entry(devid, table);
3624	iommu_flush_dte(iommu, devid);
3625}
3626
3627static struct irq_remap_table *alloc_irq_table(u16 devid)
3628{
3629	struct irq_remap_table *table = NULL;
3630	struct irq_remap_table *new_table = NULL;
3631	struct amd_iommu *iommu;
3632	unsigned long flags;
3633	u16 alias;
3634
3635	spin_lock_irqsave(&iommu_table_lock, flags);
3636
3637	iommu = amd_iommu_rlookup_table[devid];
3638	if (!iommu)
3639		goto out_unlock;
3640
3641	table = irq_lookup_table[devid];
3642	if (table)
3643		goto out_unlock;
3644
3645	alias = amd_iommu_alias_table[devid];
3646	table = irq_lookup_table[alias];
3647	if (table) {
3648		set_remap_table_entry(iommu, devid, table);
3649		goto out_wait;
3650	}
3651	spin_unlock_irqrestore(&iommu_table_lock, flags);
3652
3653	/* Nothing there yet, allocate new irq remapping table */
3654	new_table = __alloc_irq_table();
3655	if (!new_table)
3656		return NULL;
3657
3658	spin_lock_irqsave(&iommu_table_lock, flags);
3659
3660	table = irq_lookup_table[devid];
3661	if (table)
3662		goto out_unlock;
3663
3664	table = irq_lookup_table[alias];
3665	if (table) {
3666		set_remap_table_entry(iommu, devid, table);
3667		goto out_wait;
3668	}
3669
3670	table = new_table;
3671	new_table = NULL;
3672
3673	set_remap_table_entry(iommu, devid, table);
3674	if (devid != alias)
3675		set_remap_table_entry(iommu, alias, table);
3676
3677out_wait:
3678	iommu_completion_wait(iommu);
3679
3680out_unlock:
3681	spin_unlock_irqrestore(&iommu_table_lock, flags);
3682
3683	if (new_table) {
3684		kmem_cache_free(amd_iommu_irq_cache, new_table->table);
3685		kfree(new_table);
3686	}
3687	return table;
3688}
3689
3690static int alloc_irq_index(u16 devid, int count, bool align)
3691{
3692	struct irq_remap_table *table;
3693	int index, c, alignment = 1;
3694	unsigned long flags;
3695	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3696
3697	if (!iommu)
3698		return -ENODEV;
3699
3700	table = alloc_irq_table(devid);
3701	if (!table)
3702		return -ENODEV;
3703
3704	if (align)
3705		alignment = roundup_pow_of_two(count);
3706
3707	raw_spin_lock_irqsave(&table->lock, flags);
3708
3709	/* Scan table for free entries */
3710	for (index = ALIGN(table->min_index, alignment), c = 0;
3711	     index < MAX_IRQS_PER_TABLE;) {
3712		if (!iommu->irte_ops->is_allocated(table, index)) {
3713			c += 1;
3714		} else {
3715			c     = 0;
3716			index = ALIGN(index + 1, alignment);
3717			continue;
3718		}
3719
3720		if (c == count)	{
3721			for (; c != 0; --c)
3722				iommu->irte_ops->set_allocated(table, index - c + 1);
3723
3724			index -= count - 1;
3725			goto out;
3726		}
3727
3728		index++;
3729	}
3730
3731	index = -ENOSPC;
3732
3733out:
3734	raw_spin_unlock_irqrestore(&table->lock, flags);
3735
3736	return index;
3737}
3738
3739static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3740			  struct amd_ir_data *data)
3741{
3742	struct irq_remap_table *table;
3743	struct amd_iommu *iommu;
3744	unsigned long flags;
3745	struct irte_ga *entry;
3746
3747	iommu = amd_iommu_rlookup_table[devid];
3748	if (iommu == NULL)
3749		return -EINVAL;
3750
3751	table = get_irq_table(devid);
3752	if (!table)
3753		return -ENOMEM;
3754
3755	raw_spin_lock_irqsave(&table->lock, flags);
3756
3757	entry = (struct irte_ga *)table->table;
3758	entry = &entry[index];
3759	entry->lo.fields_remap.valid = 0;
3760	entry->hi.val = irte->hi.val;
3761	entry->lo.val = irte->lo.val;
3762	entry->lo.fields_remap.valid = 1;
3763	if (data)
3764		data->ref = entry;
3765
3766	raw_spin_unlock_irqrestore(&table->lock, flags);
3767
3768	iommu_flush_irt(iommu, devid);
3769	iommu_completion_wait(iommu);
3770
3771	return 0;
3772}
3773
3774static int modify_irte(u16 devid, int index, union irte *irte)
3775{
3776	struct irq_remap_table *table;
3777	struct amd_iommu *iommu;
3778	unsigned long flags;
3779
3780	iommu = amd_iommu_rlookup_table[devid];
3781	if (iommu == NULL)
3782		return -EINVAL;
3783
3784	table = get_irq_table(devid);
3785	if (!table)
3786		return -ENOMEM;
3787
3788	raw_spin_lock_irqsave(&table->lock, flags);
3789	table->table[index] = irte->val;
3790	raw_spin_unlock_irqrestore(&table->lock, flags);
3791
3792	iommu_flush_irt(iommu, devid);
3793	iommu_completion_wait(iommu);
3794
3795	return 0;
3796}
3797
3798static void free_irte(u16 devid, int index)
3799{
3800	struct irq_remap_table *table;
3801	struct amd_iommu *iommu;
3802	unsigned long flags;
3803
3804	iommu = amd_iommu_rlookup_table[devid];
3805	if (iommu == NULL)
3806		return;
3807
3808	table = get_irq_table(devid);
3809	if (!table)
3810		return;
3811
3812	raw_spin_lock_irqsave(&table->lock, flags);
3813	iommu->irte_ops->clear_allocated(table, index);
3814	raw_spin_unlock_irqrestore(&table->lock, flags);
3815
3816	iommu_flush_irt(iommu, devid);
3817	iommu_completion_wait(iommu);
3818}
3819
3820static void irte_prepare(void *entry,
3821			 u32 delivery_mode, u32 dest_mode,
3822			 u8 vector, u32 dest_apicid, int devid)
3823{
3824	union irte *irte = (union irte *) entry;
3825
3826	irte->val                = 0;
3827	irte->fields.vector      = vector;
3828	irte->fields.int_type    = delivery_mode;
3829	irte->fields.destination = dest_apicid;
3830	irte->fields.dm          = dest_mode;
3831	irte->fields.valid       = 1;
3832}
3833
3834static void irte_ga_prepare(void *entry,
3835			    u32 delivery_mode, u32 dest_mode,
3836			    u8 vector, u32 dest_apicid, int devid)
3837{
3838	struct irte_ga *irte = (struct irte_ga *) entry;
3839
3840	irte->lo.val                      = 0;
3841	irte->hi.val                      = 0;
3842	irte->lo.fields_remap.int_type    = delivery_mode;
3843	irte->lo.fields_remap.dm          = dest_mode;
3844	irte->hi.fields.vector            = vector;
3845	irte->lo.fields_remap.destination = dest_apicid;
3846	irte->lo.fields_remap.valid       = 1;
3847}
3848
3849static void irte_activate(void *entry, u16 devid, u16 index)
3850{
3851	union irte *irte = (union irte *) entry;
3852
3853	irte->fields.valid = 1;
3854	modify_irte(devid, index, irte);
3855}
3856
3857static void irte_ga_activate(void *entry, u16 devid, u16 index)
3858{
3859	struct irte_ga *irte = (struct irte_ga *) entry;
3860
3861	irte->lo.fields_remap.valid = 1;
3862	modify_irte_ga(devid, index, irte, NULL);
3863}
3864
3865static void irte_deactivate(void *entry, u16 devid, u16 index)
3866{
3867	union irte *irte = (union irte *) entry;
3868
3869	irte->fields.valid = 0;
3870	modify_irte(devid, index, irte);
3871}
3872
3873static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3874{
3875	struct irte_ga *irte = (struct irte_ga *) entry;
3876
3877	irte->lo.fields_remap.valid = 0;
3878	modify_irte_ga(devid, index, irte, NULL);
3879}
3880
3881static void irte_set_affinity(void *entry, u16 devid, u16 index,
3882			      u8 vector, u32 dest_apicid)
3883{
3884	union irte *irte = (union irte *) entry;
3885
3886	irte->fields.vector = vector;
3887	irte->fields.destination = dest_apicid;
3888	modify_irte(devid, index, irte);
3889}
3890
3891static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3892				 u8 vector, u32 dest_apicid)
3893{
3894	struct irte_ga *irte = (struct irte_ga *) entry;
3895
3896	if (!irte->lo.fields_remap.guest_mode) {
3897		irte->hi.fields.vector = vector;
3898		irte->lo.fields_remap.destination = dest_apicid;
3899		modify_irte_ga(devid, index, irte, NULL);
3900	}
3901}
3902
3903#define IRTE_ALLOCATED (~1U)
3904static void irte_set_allocated(struct irq_remap_table *table, int index)
3905{
3906	table->table[index] = IRTE_ALLOCATED;
3907}
3908
3909static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3910{
3911	struct irte_ga *ptr = (struct irte_ga *)table->table;
3912	struct irte_ga *irte = &ptr[index];
3913
3914	memset(&irte->lo.val, 0, sizeof(u64));
3915	memset(&irte->hi.val, 0, sizeof(u64));
3916	irte->hi.fields.vector = 0xff;
3917}
3918
3919static bool irte_is_allocated(struct irq_remap_table *table, int index)
3920{
3921	union irte *ptr = (union irte *)table->table;
3922	union irte *irte = &ptr[index];
3923
3924	return irte->val != 0;
3925}
3926
3927static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3928{
3929	struct irte_ga *ptr = (struct irte_ga *)table->table;
3930	struct irte_ga *irte = &ptr[index];
3931
3932	return irte->hi.fields.vector != 0;
3933}
3934
3935static void irte_clear_allocated(struct irq_remap_table *table, int index)
3936{
3937	table->table[index] = 0;
3938}
3939
3940static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3941{
3942	struct irte_ga *ptr = (struct irte_ga *)table->table;
3943	struct irte_ga *irte = &ptr[index];
3944
3945	memset(&irte->lo.val, 0, sizeof(u64));
3946	memset(&irte->hi.val, 0, sizeof(u64));
3947}
3948
3949static int get_devid(struct irq_alloc_info *info)
3950{
3951	int devid = -1;
3952
3953	switch (info->type) {
3954	case X86_IRQ_ALLOC_TYPE_IOAPIC:
3955		devid     = get_ioapic_devid(info->ioapic_id);
3956		break;
3957	case X86_IRQ_ALLOC_TYPE_HPET:
3958		devid     = get_hpet_devid(info->hpet_id);
3959		break;
3960	case X86_IRQ_ALLOC_TYPE_MSI:
3961	case X86_IRQ_ALLOC_TYPE_MSIX:
3962		devid = get_device_id(&info->msi_dev->dev);
3963		break;
3964	default:
3965		BUG_ON(1);
3966		break;
3967	}
3968
3969	return devid;
3970}
3971
3972static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
3973{
3974	struct amd_iommu *iommu;
3975	int devid;
3976
3977	if (!info)
3978		return NULL;
3979
3980	devid = get_devid(info);
3981	if (devid >= 0) {
3982		iommu = amd_iommu_rlookup_table[devid];
3983		if (iommu)
3984			return iommu->ir_domain;
3985	}
3986
3987	return NULL;
3988}
3989
3990static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
3991{
3992	struct amd_iommu *iommu;
3993	int devid;
3994
3995	if (!info)
3996		return NULL;
3997
3998	switch (info->type) {
3999	case X86_IRQ_ALLOC_TYPE_MSI:
4000	case X86_IRQ_ALLOC_TYPE_MSIX:
4001		devid = get_device_id(&info->msi_dev->dev);
4002		if (devid < 0)
4003			return NULL;
4004
4005		iommu = amd_iommu_rlookup_table[devid];
4006		if (iommu)
4007			return iommu->msi_domain;
4008		break;
4009	default:
4010		break;
4011	}
4012
4013	return NULL;
4014}
4015
4016struct irq_remap_ops amd_iommu_irq_ops = {
4017	.prepare		= amd_iommu_prepare,
4018	.enable			= amd_iommu_enable,
4019	.disable		= amd_iommu_disable,
4020	.reenable		= amd_iommu_reenable,
4021	.enable_faulting	= amd_iommu_enable_faulting,
4022	.get_ir_irq_domain	= get_ir_irq_domain,
4023	.get_irq_domain		= get_irq_domain,
4024};
4025
4026static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4027				       struct irq_cfg *irq_cfg,
4028				       struct irq_alloc_info *info,
4029				       int devid, int index, int sub_handle)
4030{
4031	struct irq_2_irte *irte_info = &data->irq_2_irte;
4032	struct msi_msg *msg = &data->msi_entry;
4033	struct IO_APIC_route_entry *entry;
4034	struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4035
4036	if (!iommu)
4037		return;
4038
4039	data->irq_2_irte.devid = devid;
4040	data->irq_2_irte.index = index + sub_handle;
4041	iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4042				 apic->irq_dest_mode, irq_cfg->vector,
4043				 irq_cfg->dest_apicid, devid);
4044
4045	switch (info->type) {
4046	case X86_IRQ_ALLOC_TYPE_IOAPIC:
4047		/* Setup IOAPIC entry */
4048		entry = info->ioapic_entry;
4049		info->ioapic_entry = NULL;
4050		memset(entry, 0, sizeof(*entry));
4051		entry->vector        = index;
4052		entry->mask          = 0;
4053		entry->trigger       = info->ioapic_trigger;
4054		entry->polarity      = info->ioapic_polarity;
4055		/* Mask level triggered irqs. */
4056		if (info->ioapic_trigger)
4057			entry->mask = 1;
4058		break;
4059
4060	case X86_IRQ_ALLOC_TYPE_HPET:
4061	case X86_IRQ_ALLOC_TYPE_MSI:
4062	case X86_IRQ_ALLOC_TYPE_MSIX:
4063		msg->address_hi = MSI_ADDR_BASE_HI;
4064		msg->address_lo = MSI_ADDR_BASE_LO;
4065		msg->data = irte_info->index;
4066		break;
4067
4068	default:
4069		BUG_ON(1);
4070		break;
4071	}
4072}
4073
4074struct amd_irte_ops irte_32_ops = {
4075	.prepare = irte_prepare,
4076	.activate = irte_activate,
4077	.deactivate = irte_deactivate,
4078	.set_affinity = irte_set_affinity,
4079	.set_allocated = irte_set_allocated,
4080	.is_allocated = irte_is_allocated,
4081	.clear_allocated = irte_clear_allocated,
4082};
4083
4084struct amd_irte_ops irte_128_ops = {
4085	.prepare = irte_ga_prepare,
4086	.activate = irte_ga_activate,
4087	.deactivate = irte_ga_deactivate,
4088	.set_affinity = irte_ga_set_affinity,
4089	.set_allocated = irte_ga_set_allocated,
4090	.is_allocated = irte_ga_is_allocated,
4091	.clear_allocated = irte_ga_clear_allocated,
4092};
4093
4094static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4095			       unsigned int nr_irqs, void *arg)
4096{
4097	struct irq_alloc_info *info = arg;
4098	struct irq_data *irq_data;
4099	struct amd_ir_data *data = NULL;
4100	struct irq_cfg *cfg;
4101	int i, ret, devid;
4102	int index;
4103
4104	if (!info)
4105		return -EINVAL;
4106	if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4107	    info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4108		return -EINVAL;
4109
4110	/*
4111	 * With IRQ remapping enabled, don't need contiguous CPU vectors
4112	 * to support multiple MSI interrupts.
4113	 */
4114	if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4115		info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4116
4117	devid = get_devid(info);
4118	if (devid < 0)
4119		return -EINVAL;
4120
4121	ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4122	if (ret < 0)
4123		return ret;
4124
4125	if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4126		struct irq_remap_table *table;
4127		struct amd_iommu *iommu;
4128
4129		table = alloc_irq_table(devid);
4130		if (table) {
4131			if (!table->min_index) {
4132				/*
4133				 * Keep the first 32 indexes free for IOAPIC
4134				 * interrupts.
4135				 */
4136				table->min_index = 32;
4137				iommu = amd_iommu_rlookup_table[devid];
4138				for (i = 0; i < 32; ++i)
4139					iommu->irte_ops->set_allocated(table, i);
4140			}
4141			WARN_ON(table->min_index != 32);
4142			index = info->ioapic_pin;
4143		} else {
4144			index = -ENOMEM;
4145		}
4146	} else {
4147		bool align = (info->type == X86_IRQ_ALLOC_TYPE_MSI);
4148
4149		index = alloc_irq_index(devid, nr_irqs, align);
4150	}
4151	if (index < 0) {
4152		pr_warn("Failed to allocate IRTE\n");
4153		ret = index;
4154		goto out_free_parent;
4155	}
4156
4157	for (i = 0; i < nr_irqs; i++) {
4158		irq_data = irq_domain_get_irq_data(domain, virq + i);
4159		cfg = irqd_cfg(irq_data);
4160		if (!irq_data || !cfg) {
4161			ret = -EINVAL;
4162			goto out_free_data;
4163		}
4164
4165		ret = -ENOMEM;
4166		data = kzalloc(sizeof(*data), GFP_KERNEL);
4167		if (!data)
4168			goto out_free_data;
4169
4170		if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4171			data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4172		else
4173			data->entry = kzalloc(sizeof(struct irte_ga),
4174						     GFP_KERNEL);
4175		if (!data->entry) {
4176			kfree(data);
4177			goto out_free_data;
4178		}
4179
4180		irq_data->hwirq = (devid << 16) + i;
4181		irq_data->chip_data = data;
4182		irq_data->chip = &amd_ir_chip;
4183		irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4184		irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4185	}
4186
4187	return 0;
4188
4189out_free_data:
4190	for (i--; i >= 0; i--) {
4191		irq_data = irq_domain_get_irq_data(domain, virq + i);
4192		if (irq_data)
4193			kfree(irq_data->chip_data);
4194	}
4195	for (i = 0; i < nr_irqs; i++)
4196		free_irte(devid, index + i);
4197out_free_parent:
4198	irq_domain_free_irqs_common(domain, virq, nr_irqs);
4199	return ret;
4200}
4201
4202static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4203			       unsigned int nr_irqs)
4204{
4205	struct irq_2_irte *irte_info;
4206	struct irq_data *irq_data;
4207	struct amd_ir_data *data;
4208	int i;
4209
4210	for (i = 0; i < nr_irqs; i++) {
4211		irq_data = irq_domain_get_irq_data(domain, virq  + i);
4212		if (irq_data && irq_data->chip_data) {
4213			data = irq_data->chip_data;
4214			irte_info = &data->irq_2_irte;
4215			free_irte(irte_info->devid, irte_info->index);
4216			kfree(data->entry);
4217			kfree(data);
4218		}
4219	}
4220	irq_domain_free_irqs_common(domain, virq, nr_irqs);
4221}
4222
4223static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4224			       struct amd_ir_data *ir_data,
4225			       struct irq_2_irte *irte_info,
4226			       struct irq_cfg *cfg);
4227
4228static int irq_remapping_activate(struct irq_domain *domain,
4229				  struct irq_data *irq_data, bool reserve)
4230{
4231	struct amd_ir_data *data = irq_data->chip_data;
4232	struct irq_2_irte *irte_info = &data->irq_2_irte;
4233	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4234	struct irq_cfg *cfg = irqd_cfg(irq_data);
4235
4236	if (!iommu)
4237		return 0;
4238
4239	iommu->irte_ops->activate(data->entry, irte_info->devid,
4240				  irte_info->index);
4241	amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg);
4242	return 0;
4243}
4244
4245static void irq_remapping_deactivate(struct irq_domain *domain,
4246				     struct irq_data *irq_data)
4247{
4248	struct amd_ir_data *data = irq_data->chip_data;
4249	struct irq_2_irte *irte_info = &data->irq_2_irte;
4250	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4251
4252	if (iommu)
4253		iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4254					    irte_info->index);
4255}
4256
4257static const struct irq_domain_ops amd_ir_domain_ops = {
4258	.alloc = irq_remapping_alloc,
4259	.free = irq_remapping_free,
4260	.activate = irq_remapping_activate,
4261	.deactivate = irq_remapping_deactivate,
4262};
4263
4264static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4265{
4266	struct amd_iommu *iommu;
4267	struct amd_iommu_pi_data *pi_data = vcpu_info;
4268	struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4269	struct amd_ir_data *ir_data = data->chip_data;
4270	struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4271	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4272	struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4273
4274	/* Note:
4275	 * This device has never been set up for guest mode.
4276	 * we should not modify the IRTE
4277	 */
4278	if (!dev_data || !dev_data->use_vapic)
4279		return 0;
4280
4281	pi_data->ir_data = ir_data;
4282
4283	/* Note:
4284	 * SVM tries to set up for VAPIC mode, but we are in
4285	 * legacy mode. So, we force legacy mode instead.
4286	 */
4287	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4288		pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4289			 __func__);
4290		pi_data->is_guest_mode = false;
4291	}
4292
4293	iommu = amd_iommu_rlookup_table[irte_info->devid];
4294	if (iommu == NULL)
4295		return -EINVAL;
4296
4297	pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4298	if (pi_data->is_guest_mode) {
4299		/* Setting */
4300		irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4301		irte->hi.fields.vector = vcpu_pi_info->vector;
4302		irte->lo.fields_vapic.ga_log_intr = 1;
4303		irte->lo.fields_vapic.guest_mode = 1;
4304		irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4305
4306		ir_data->cached_ga_tag = pi_data->ga_tag;
4307	} else {
4308		/* Un-Setting */
4309		struct irq_cfg *cfg = irqd_cfg(data);
4310
4311		irte->hi.val = 0;
4312		irte->lo.val = 0;
4313		irte->hi.fields.vector = cfg->vector;
4314		irte->lo.fields_remap.guest_mode = 0;
4315		irte->lo.fields_remap.destination = cfg->dest_apicid;
4316		irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4317		irte->lo.fields_remap.dm = apic->irq_dest_mode;
4318
4319		/*
4320		 * This communicates the ga_tag back to the caller
4321		 * so that it can do all the necessary clean up.
4322		 */
4323		ir_data->cached_ga_tag = 0;
4324	}
4325
4326	return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4327}
4328
4329
4330static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
4331			       struct amd_ir_data *ir_data,
4332			       struct irq_2_irte *irte_info,
4333			       struct irq_cfg *cfg)
4334{
4335
4336	/*
4337	 * Atomically updates the IRTE with the new destination, vector
4338	 * and flushes the interrupt entry cache.
4339	 */
4340	iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4341				      irte_info->index, cfg->vector,
4342				      cfg->dest_apicid);
4343}
4344
4345static int amd_ir_set_affinity(struct irq_data *data,
4346			       const struct cpumask *mask, bool force)
4347{
4348	struct amd_ir_data *ir_data = data->chip_data;
4349	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4350	struct irq_cfg *cfg = irqd_cfg(data);
4351	struct irq_data *parent = data->parent_data;
4352	struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4353	int ret;
4354
4355	if (!iommu)
4356		return -ENODEV;
4357
4358	ret = parent->chip->irq_set_affinity(parent, mask, force);
4359	if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4360		return ret;
4361
4362	amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg);
4363	/*
4364	 * After this point, all the interrupts will start arriving
4365	 * at the new destination. So, time to cleanup the previous
4366	 * vector allocation.
4367	 */
4368	send_cleanup_vector(cfg);
4369
4370	return IRQ_SET_MASK_OK_DONE;
4371}
4372
4373static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4374{
4375	struct amd_ir_data *ir_data = irq_data->chip_data;
4376
4377	*msg = ir_data->msi_entry;
4378}
4379
4380static struct irq_chip amd_ir_chip = {
4381	.name			= "AMD-IR",
4382	.irq_ack		= ir_ack_apic_edge,
4383	.irq_set_affinity	= amd_ir_set_affinity,
4384	.irq_set_vcpu_affinity	= amd_ir_set_vcpu_affinity,
4385	.irq_compose_msi_msg	= ir_compose_msi_msg,
4386};
4387
4388int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4389{
4390	struct fwnode_handle *fn;
4391
4392	fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index);
4393	if (!fn)
4394		return -ENOMEM;
4395	iommu->ir_domain = irq_domain_create_tree(fn, &amd_ir_domain_ops, iommu);
4396	irq_domain_free_fwnode(fn);
4397	if (!iommu->ir_domain)
4398		return -ENOMEM;
4399
4400	iommu->ir_domain->parent = arch_get_ir_parent_domain();
4401	iommu->msi_domain = arch_create_remap_msi_irq_domain(iommu->ir_domain,
4402							     "AMD-IR-MSI",
4403							     iommu->index);
4404	return 0;
4405}
4406
4407int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4408{
4409	unsigned long flags;
4410	struct amd_iommu *iommu;
4411	struct irq_remap_table *table;
4412	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4413	int devid = ir_data->irq_2_irte.devid;
4414	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4415	struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4416
4417	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4418	    !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4419		return 0;
4420
4421	iommu = amd_iommu_rlookup_table[devid];
4422	if (!iommu)
4423		return -ENODEV;
4424
4425	table = get_irq_table(devid);
4426	if (!table)
4427		return -ENODEV;
4428
4429	raw_spin_lock_irqsave(&table->lock, flags);
4430
4431	if (ref->lo.fields_vapic.guest_mode) {
4432		if (cpu >= 0)
4433			ref->lo.fields_vapic.destination = cpu;
4434		ref->lo.fields_vapic.is_run = is_run;
4435		barrier();
4436	}
4437
4438	raw_spin_unlock_irqrestore(&table->lock, flags);
4439
4440	iommu_flush_irt(iommu, devid);
4441	iommu_completion_wait(iommu);
4442	return 0;
4443}
4444EXPORT_SYMBOL(amd_iommu_update_ga);
4445#endif