Loading...
1/*
2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
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/pci.h>
21#include <linux/acpi.h>
22#include <linux/list.h>
23#include <linux/slab.h>
24#include <linux/syscore_ops.h>
25#include <linux/interrupt.h>
26#include <linux/msi.h>
27#include <linux/amd-iommu.h>
28#include <asm/pci-direct.h>
29#include <asm/iommu.h>
30#include <asm/gart.h>
31#include <asm/x86_init.h>
32#include <asm/iommu_table.h>
33
34#include "amd_iommu_proto.h"
35#include "amd_iommu_types.h"
36
37/*
38 * definitions for the ACPI scanning code
39 */
40#define IVRS_HEADER_LENGTH 48
41
42#define ACPI_IVHD_TYPE 0x10
43#define ACPI_IVMD_TYPE_ALL 0x20
44#define ACPI_IVMD_TYPE 0x21
45#define ACPI_IVMD_TYPE_RANGE 0x22
46
47#define IVHD_DEV_ALL 0x01
48#define IVHD_DEV_SELECT 0x02
49#define IVHD_DEV_SELECT_RANGE_START 0x03
50#define IVHD_DEV_RANGE_END 0x04
51#define IVHD_DEV_ALIAS 0x42
52#define IVHD_DEV_ALIAS_RANGE 0x43
53#define IVHD_DEV_EXT_SELECT 0x46
54#define IVHD_DEV_EXT_SELECT_RANGE 0x47
55
56#define IVHD_FLAG_HT_TUN_EN_MASK 0x01
57#define IVHD_FLAG_PASSPW_EN_MASK 0x02
58#define IVHD_FLAG_RESPASSPW_EN_MASK 0x04
59#define IVHD_FLAG_ISOC_EN_MASK 0x08
60
61#define IVMD_FLAG_EXCL_RANGE 0x08
62#define IVMD_FLAG_UNITY_MAP 0x01
63
64#define ACPI_DEVFLAG_INITPASS 0x01
65#define ACPI_DEVFLAG_EXTINT 0x02
66#define ACPI_DEVFLAG_NMI 0x04
67#define ACPI_DEVFLAG_SYSMGT1 0x10
68#define ACPI_DEVFLAG_SYSMGT2 0x20
69#define ACPI_DEVFLAG_LINT0 0x40
70#define ACPI_DEVFLAG_LINT1 0x80
71#define ACPI_DEVFLAG_ATSDIS 0x10000000
72
73/*
74 * ACPI table definitions
75 *
76 * These data structures are laid over the table to parse the important values
77 * out of it.
78 */
79
80/*
81 * structure describing one IOMMU in the ACPI table. Typically followed by one
82 * or more ivhd_entrys.
83 */
84struct ivhd_header {
85 u8 type;
86 u8 flags;
87 u16 length;
88 u16 devid;
89 u16 cap_ptr;
90 u64 mmio_phys;
91 u16 pci_seg;
92 u16 info;
93 u32 reserved;
94} __attribute__((packed));
95
96/*
97 * A device entry describing which devices a specific IOMMU translates and
98 * which requestor ids they use.
99 */
100struct ivhd_entry {
101 u8 type;
102 u16 devid;
103 u8 flags;
104 u32 ext;
105} __attribute__((packed));
106
107/*
108 * An AMD IOMMU memory definition structure. It defines things like exclusion
109 * ranges for devices and regions that should be unity mapped.
110 */
111struct ivmd_header {
112 u8 type;
113 u8 flags;
114 u16 length;
115 u16 devid;
116 u16 aux;
117 u64 resv;
118 u64 range_start;
119 u64 range_length;
120} __attribute__((packed));
121
122bool amd_iommu_dump;
123
124static int __initdata amd_iommu_detected;
125static bool __initdata amd_iommu_disabled;
126
127u16 amd_iommu_last_bdf; /* largest PCI device id we have
128 to handle */
129LIST_HEAD(amd_iommu_unity_map); /* a list of required unity mappings
130 we find in ACPI */
131bool amd_iommu_unmap_flush; /* if true, flush on every unmap */
132
133LIST_HEAD(amd_iommu_list); /* list of all AMD IOMMUs in the
134 system */
135
136/* Array to assign indices to IOMMUs*/
137struct amd_iommu *amd_iommus[MAX_IOMMUS];
138int amd_iommus_present;
139
140/* IOMMUs have a non-present cache? */
141bool amd_iommu_np_cache __read_mostly;
142bool amd_iommu_iotlb_sup __read_mostly = true;
143
144/*
145 * The ACPI table parsing functions set this variable on an error
146 */
147static int __initdata amd_iommu_init_err;
148
149/*
150 * List of protection domains - used during resume
151 */
152LIST_HEAD(amd_iommu_pd_list);
153spinlock_t amd_iommu_pd_lock;
154
155/*
156 * Pointer to the device table which is shared by all AMD IOMMUs
157 * it is indexed by the PCI device id or the HT unit id and contains
158 * information about the domain the device belongs to as well as the
159 * page table root pointer.
160 */
161struct dev_table_entry *amd_iommu_dev_table;
162
163/*
164 * The alias table is a driver specific data structure which contains the
165 * mappings of the PCI device ids to the actual requestor ids on the IOMMU.
166 * More than one device can share the same requestor id.
167 */
168u16 *amd_iommu_alias_table;
169
170/*
171 * The rlookup table is used to find the IOMMU which is responsible
172 * for a specific device. It is also indexed by the PCI device id.
173 */
174struct amd_iommu **amd_iommu_rlookup_table;
175
176/*
177 * AMD IOMMU allows up to 2^16 differend protection domains. This is a bitmap
178 * to know which ones are already in use.
179 */
180unsigned long *amd_iommu_pd_alloc_bitmap;
181
182static u32 dev_table_size; /* size of the device table */
183static u32 alias_table_size; /* size of the alias table */
184static u32 rlookup_table_size; /* size if the rlookup table */
185
186/*
187 * This function flushes all internal caches of
188 * the IOMMU used by this driver.
189 */
190extern void iommu_flush_all_caches(struct amd_iommu *iommu);
191
192static inline void update_last_devid(u16 devid)
193{
194 if (devid > amd_iommu_last_bdf)
195 amd_iommu_last_bdf = devid;
196}
197
198static inline unsigned long tbl_size(int entry_size)
199{
200 unsigned shift = PAGE_SHIFT +
201 get_order(((int)amd_iommu_last_bdf + 1) * entry_size);
202
203 return 1UL << shift;
204}
205
206/* Access to l1 and l2 indexed register spaces */
207
208static u32 iommu_read_l1(struct amd_iommu *iommu, u16 l1, u8 address)
209{
210 u32 val;
211
212 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16));
213 pci_read_config_dword(iommu->dev, 0xfc, &val);
214 return val;
215}
216
217static void iommu_write_l1(struct amd_iommu *iommu, u16 l1, u8 address, u32 val)
218{
219 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16 | 1 << 31));
220 pci_write_config_dword(iommu->dev, 0xfc, val);
221 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16));
222}
223
224static u32 iommu_read_l2(struct amd_iommu *iommu, u8 address)
225{
226 u32 val;
227
228 pci_write_config_dword(iommu->dev, 0xf0, address);
229 pci_read_config_dword(iommu->dev, 0xf4, &val);
230 return val;
231}
232
233static void iommu_write_l2(struct amd_iommu *iommu, u8 address, u32 val)
234{
235 pci_write_config_dword(iommu->dev, 0xf0, (address | 1 << 8));
236 pci_write_config_dword(iommu->dev, 0xf4, val);
237}
238
239/****************************************************************************
240 *
241 * AMD IOMMU MMIO register space handling functions
242 *
243 * These functions are used to program the IOMMU device registers in
244 * MMIO space required for that driver.
245 *
246 ****************************************************************************/
247
248/*
249 * This function set the exclusion range in the IOMMU. DMA accesses to the
250 * exclusion range are passed through untranslated
251 */
252static void iommu_set_exclusion_range(struct amd_iommu *iommu)
253{
254 u64 start = iommu->exclusion_start & PAGE_MASK;
255 u64 limit = (start + iommu->exclusion_length) & PAGE_MASK;
256 u64 entry;
257
258 if (!iommu->exclusion_start)
259 return;
260
261 entry = start | MMIO_EXCL_ENABLE_MASK;
262 memcpy_toio(iommu->mmio_base + MMIO_EXCL_BASE_OFFSET,
263 &entry, sizeof(entry));
264
265 entry = limit;
266 memcpy_toio(iommu->mmio_base + MMIO_EXCL_LIMIT_OFFSET,
267 &entry, sizeof(entry));
268}
269
270/* Programs the physical address of the device table into the IOMMU hardware */
271static void __init iommu_set_device_table(struct amd_iommu *iommu)
272{
273 u64 entry;
274
275 BUG_ON(iommu->mmio_base == NULL);
276
277 entry = virt_to_phys(amd_iommu_dev_table);
278 entry |= (dev_table_size >> 12) - 1;
279 memcpy_toio(iommu->mmio_base + MMIO_DEV_TABLE_OFFSET,
280 &entry, sizeof(entry));
281}
282
283/* Generic functions to enable/disable certain features of the IOMMU. */
284static void iommu_feature_enable(struct amd_iommu *iommu, u8 bit)
285{
286 u32 ctrl;
287
288 ctrl = readl(iommu->mmio_base + MMIO_CONTROL_OFFSET);
289 ctrl |= (1 << bit);
290 writel(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET);
291}
292
293static void iommu_feature_disable(struct amd_iommu *iommu, u8 bit)
294{
295 u32 ctrl;
296
297 ctrl = readl(iommu->mmio_base + MMIO_CONTROL_OFFSET);
298 ctrl &= ~(1 << bit);
299 writel(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET);
300}
301
302/* Function to enable the hardware */
303static void iommu_enable(struct amd_iommu *iommu)
304{
305 static const char * const feat_str[] = {
306 "PreF", "PPR", "X2APIC", "NX", "GT", "[5]",
307 "IA", "GA", "HE", "PC", NULL
308 };
309 int i;
310
311 printk(KERN_INFO "AMD-Vi: Enabling IOMMU at %s cap 0x%hx",
312 dev_name(&iommu->dev->dev), iommu->cap_ptr);
313
314 if (iommu->cap & (1 << IOMMU_CAP_EFR)) {
315 printk(KERN_CONT " extended features: ");
316 for (i = 0; feat_str[i]; ++i)
317 if (iommu_feature(iommu, (1ULL << i)))
318 printk(KERN_CONT " %s", feat_str[i]);
319 }
320 printk(KERN_CONT "\n");
321
322 iommu_feature_enable(iommu, CONTROL_IOMMU_EN);
323}
324
325static void iommu_disable(struct amd_iommu *iommu)
326{
327 /* Disable command buffer */
328 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN);
329
330 /* Disable event logging and event interrupts */
331 iommu_feature_disable(iommu, CONTROL_EVT_INT_EN);
332 iommu_feature_disable(iommu, CONTROL_EVT_LOG_EN);
333
334 /* Disable IOMMU hardware itself */
335 iommu_feature_disable(iommu, CONTROL_IOMMU_EN);
336}
337
338/*
339 * mapping and unmapping functions for the IOMMU MMIO space. Each AMD IOMMU in
340 * the system has one.
341 */
342static u8 * __init iommu_map_mmio_space(u64 address)
343{
344 u8 *ret;
345
346 if (!request_mem_region(address, MMIO_REGION_LENGTH, "amd_iommu")) {
347 pr_err("AMD-Vi: Can not reserve memory region %llx for mmio\n",
348 address);
349 pr_err("AMD-Vi: This is a BIOS bug. Please contact your hardware vendor\n");
350 return NULL;
351 }
352
353 ret = ioremap_nocache(address, MMIO_REGION_LENGTH);
354 if (ret != NULL)
355 return ret;
356
357 release_mem_region(address, MMIO_REGION_LENGTH);
358
359 return NULL;
360}
361
362static void __init iommu_unmap_mmio_space(struct amd_iommu *iommu)
363{
364 if (iommu->mmio_base)
365 iounmap(iommu->mmio_base);
366 release_mem_region(iommu->mmio_phys, MMIO_REGION_LENGTH);
367}
368
369/****************************************************************************
370 *
371 * The functions below belong to the first pass of AMD IOMMU ACPI table
372 * parsing. In this pass we try to find out the highest device id this
373 * code has to handle. Upon this information the size of the shared data
374 * structures is determined later.
375 *
376 ****************************************************************************/
377
378/*
379 * This function calculates the length of a given IVHD entry
380 */
381static inline int ivhd_entry_length(u8 *ivhd)
382{
383 return 0x04 << (*ivhd >> 6);
384}
385
386/*
387 * This function reads the last device id the IOMMU has to handle from the PCI
388 * capability header for this IOMMU
389 */
390static int __init find_last_devid_on_pci(int bus, int dev, int fn, int cap_ptr)
391{
392 u32 cap;
393
394 cap = read_pci_config(bus, dev, fn, cap_ptr+MMIO_RANGE_OFFSET);
395 update_last_devid(calc_devid(MMIO_GET_BUS(cap), MMIO_GET_LD(cap)));
396
397 return 0;
398}
399
400/*
401 * After reading the highest device id from the IOMMU PCI capability header
402 * this function looks if there is a higher device id defined in the ACPI table
403 */
404static int __init find_last_devid_from_ivhd(struct ivhd_header *h)
405{
406 u8 *p = (void *)h, *end = (void *)h;
407 struct ivhd_entry *dev;
408
409 p += sizeof(*h);
410 end += h->length;
411
412 find_last_devid_on_pci(PCI_BUS(h->devid),
413 PCI_SLOT(h->devid),
414 PCI_FUNC(h->devid),
415 h->cap_ptr);
416
417 while (p < end) {
418 dev = (struct ivhd_entry *)p;
419 switch (dev->type) {
420 case IVHD_DEV_SELECT:
421 case IVHD_DEV_RANGE_END:
422 case IVHD_DEV_ALIAS:
423 case IVHD_DEV_EXT_SELECT:
424 /* all the above subfield types refer to device ids */
425 update_last_devid(dev->devid);
426 break;
427 default:
428 break;
429 }
430 p += ivhd_entry_length(p);
431 }
432
433 WARN_ON(p != end);
434
435 return 0;
436}
437
438/*
439 * Iterate over all IVHD entries in the ACPI table and find the highest device
440 * id which we need to handle. This is the first of three functions which parse
441 * the ACPI table. So we check the checksum here.
442 */
443static int __init find_last_devid_acpi(struct acpi_table_header *table)
444{
445 int i;
446 u8 checksum = 0, *p = (u8 *)table, *end = (u8 *)table;
447 struct ivhd_header *h;
448
449 /*
450 * Validate checksum here so we don't need to do it when
451 * we actually parse the table
452 */
453 for (i = 0; i < table->length; ++i)
454 checksum += p[i];
455 if (checksum != 0) {
456 /* ACPI table corrupt */
457 amd_iommu_init_err = -ENODEV;
458 return 0;
459 }
460
461 p += IVRS_HEADER_LENGTH;
462
463 end += table->length;
464 while (p < end) {
465 h = (struct ivhd_header *)p;
466 switch (h->type) {
467 case ACPI_IVHD_TYPE:
468 find_last_devid_from_ivhd(h);
469 break;
470 default:
471 break;
472 }
473 p += h->length;
474 }
475 WARN_ON(p != end);
476
477 return 0;
478}
479
480/****************************************************************************
481 *
482 * The following functions belong the the code path which parses the ACPI table
483 * the second time. In this ACPI parsing iteration we allocate IOMMU specific
484 * data structures, initialize the device/alias/rlookup table and also
485 * basically initialize the hardware.
486 *
487 ****************************************************************************/
488
489/*
490 * Allocates the command buffer. This buffer is per AMD IOMMU. We can
491 * write commands to that buffer later and the IOMMU will execute them
492 * asynchronously
493 */
494static u8 * __init alloc_command_buffer(struct amd_iommu *iommu)
495{
496 u8 *cmd_buf = (u8 *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
497 get_order(CMD_BUFFER_SIZE));
498
499 if (cmd_buf == NULL)
500 return NULL;
501
502 iommu->cmd_buf_size = CMD_BUFFER_SIZE | CMD_BUFFER_UNINITIALIZED;
503
504 return cmd_buf;
505}
506
507/*
508 * This function resets the command buffer if the IOMMU stopped fetching
509 * commands from it.
510 */
511void amd_iommu_reset_cmd_buffer(struct amd_iommu *iommu)
512{
513 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN);
514
515 writel(0x00, iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
516 writel(0x00, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
517
518 iommu_feature_enable(iommu, CONTROL_CMDBUF_EN);
519}
520
521/*
522 * This function writes the command buffer address to the hardware and
523 * enables it.
524 */
525static void iommu_enable_command_buffer(struct amd_iommu *iommu)
526{
527 u64 entry;
528
529 BUG_ON(iommu->cmd_buf == NULL);
530
531 entry = (u64)virt_to_phys(iommu->cmd_buf);
532 entry |= MMIO_CMD_SIZE_512;
533
534 memcpy_toio(iommu->mmio_base + MMIO_CMD_BUF_OFFSET,
535 &entry, sizeof(entry));
536
537 amd_iommu_reset_cmd_buffer(iommu);
538 iommu->cmd_buf_size &= ~(CMD_BUFFER_UNINITIALIZED);
539}
540
541static void __init free_command_buffer(struct amd_iommu *iommu)
542{
543 free_pages((unsigned long)iommu->cmd_buf,
544 get_order(iommu->cmd_buf_size & ~(CMD_BUFFER_UNINITIALIZED)));
545}
546
547/* allocates the memory where the IOMMU will log its events to */
548static u8 * __init alloc_event_buffer(struct amd_iommu *iommu)
549{
550 iommu->evt_buf = (u8 *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
551 get_order(EVT_BUFFER_SIZE));
552
553 if (iommu->evt_buf == NULL)
554 return NULL;
555
556 iommu->evt_buf_size = EVT_BUFFER_SIZE;
557
558 return iommu->evt_buf;
559}
560
561static void iommu_enable_event_buffer(struct amd_iommu *iommu)
562{
563 u64 entry;
564
565 BUG_ON(iommu->evt_buf == NULL);
566
567 entry = (u64)virt_to_phys(iommu->evt_buf) | EVT_LEN_MASK;
568
569 memcpy_toio(iommu->mmio_base + MMIO_EVT_BUF_OFFSET,
570 &entry, sizeof(entry));
571
572 /* set head and tail to zero manually */
573 writel(0x00, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
574 writel(0x00, iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
575
576 iommu_feature_enable(iommu, CONTROL_EVT_LOG_EN);
577}
578
579static void __init free_event_buffer(struct amd_iommu *iommu)
580{
581 free_pages((unsigned long)iommu->evt_buf, get_order(EVT_BUFFER_SIZE));
582}
583
584/* sets a specific bit in the device table entry. */
585static void set_dev_entry_bit(u16 devid, u8 bit)
586{
587 int i = (bit >> 5) & 0x07;
588 int _bit = bit & 0x1f;
589
590 amd_iommu_dev_table[devid].data[i] |= (1 << _bit);
591}
592
593static int get_dev_entry_bit(u16 devid, u8 bit)
594{
595 int i = (bit >> 5) & 0x07;
596 int _bit = bit & 0x1f;
597
598 return (amd_iommu_dev_table[devid].data[i] & (1 << _bit)) >> _bit;
599}
600
601
602void amd_iommu_apply_erratum_63(u16 devid)
603{
604 int sysmgt;
605
606 sysmgt = get_dev_entry_bit(devid, DEV_ENTRY_SYSMGT1) |
607 (get_dev_entry_bit(devid, DEV_ENTRY_SYSMGT2) << 1);
608
609 if (sysmgt == 0x01)
610 set_dev_entry_bit(devid, DEV_ENTRY_IW);
611}
612
613/* Writes the specific IOMMU for a device into the rlookup table */
614static void __init set_iommu_for_device(struct amd_iommu *iommu, u16 devid)
615{
616 amd_iommu_rlookup_table[devid] = iommu;
617}
618
619/*
620 * This function takes the device specific flags read from the ACPI
621 * table and sets up the device table entry with that information
622 */
623static void __init set_dev_entry_from_acpi(struct amd_iommu *iommu,
624 u16 devid, u32 flags, u32 ext_flags)
625{
626 if (flags & ACPI_DEVFLAG_INITPASS)
627 set_dev_entry_bit(devid, DEV_ENTRY_INIT_PASS);
628 if (flags & ACPI_DEVFLAG_EXTINT)
629 set_dev_entry_bit(devid, DEV_ENTRY_EINT_PASS);
630 if (flags & ACPI_DEVFLAG_NMI)
631 set_dev_entry_bit(devid, DEV_ENTRY_NMI_PASS);
632 if (flags & ACPI_DEVFLAG_SYSMGT1)
633 set_dev_entry_bit(devid, DEV_ENTRY_SYSMGT1);
634 if (flags & ACPI_DEVFLAG_SYSMGT2)
635 set_dev_entry_bit(devid, DEV_ENTRY_SYSMGT2);
636 if (flags & ACPI_DEVFLAG_LINT0)
637 set_dev_entry_bit(devid, DEV_ENTRY_LINT0_PASS);
638 if (flags & ACPI_DEVFLAG_LINT1)
639 set_dev_entry_bit(devid, DEV_ENTRY_LINT1_PASS);
640
641 amd_iommu_apply_erratum_63(devid);
642
643 set_iommu_for_device(iommu, devid);
644}
645
646/*
647 * Reads the device exclusion range from ACPI and initialize IOMMU with
648 * it
649 */
650static void __init set_device_exclusion_range(u16 devid, struct ivmd_header *m)
651{
652 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
653
654 if (!(m->flags & IVMD_FLAG_EXCL_RANGE))
655 return;
656
657 if (iommu) {
658 /*
659 * We only can configure exclusion ranges per IOMMU, not
660 * per device. But we can enable the exclusion range per
661 * device. This is done here
662 */
663 set_dev_entry_bit(m->devid, DEV_ENTRY_EX);
664 iommu->exclusion_start = m->range_start;
665 iommu->exclusion_length = m->range_length;
666 }
667}
668
669/*
670 * This function reads some important data from the IOMMU PCI space and
671 * initializes the driver data structure with it. It reads the hardware
672 * capabilities and the first/last device entries
673 */
674static void __init init_iommu_from_pci(struct amd_iommu *iommu)
675{
676 int cap_ptr = iommu->cap_ptr;
677 u32 range, misc, low, high;
678 int i, j;
679
680 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_CAP_HDR_OFFSET,
681 &iommu->cap);
682 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_RANGE_OFFSET,
683 &range);
684 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_MISC_OFFSET,
685 &misc);
686
687 iommu->first_device = calc_devid(MMIO_GET_BUS(range),
688 MMIO_GET_FD(range));
689 iommu->last_device = calc_devid(MMIO_GET_BUS(range),
690 MMIO_GET_LD(range));
691 iommu->evt_msi_num = MMIO_MSI_NUM(misc);
692
693 if (!(iommu->cap & (1 << IOMMU_CAP_IOTLB)))
694 amd_iommu_iotlb_sup = false;
695
696 /* read extended feature bits */
697 low = readl(iommu->mmio_base + MMIO_EXT_FEATURES);
698 high = readl(iommu->mmio_base + MMIO_EXT_FEATURES + 4);
699
700 iommu->features = ((u64)high << 32) | low;
701
702 if (!is_rd890_iommu(iommu->dev))
703 return;
704
705 /*
706 * Some rd890 systems may not be fully reconfigured by the BIOS, so
707 * it's necessary for us to store this information so it can be
708 * reprogrammed on resume
709 */
710
711 pci_read_config_dword(iommu->dev, iommu->cap_ptr + 4,
712 &iommu->stored_addr_lo);
713 pci_read_config_dword(iommu->dev, iommu->cap_ptr + 8,
714 &iommu->stored_addr_hi);
715
716 /* Low bit locks writes to configuration space */
717 iommu->stored_addr_lo &= ~1;
718
719 for (i = 0; i < 6; i++)
720 for (j = 0; j < 0x12; j++)
721 iommu->stored_l1[i][j] = iommu_read_l1(iommu, i, j);
722
723 for (i = 0; i < 0x83; i++)
724 iommu->stored_l2[i] = iommu_read_l2(iommu, i);
725}
726
727/*
728 * Takes a pointer to an AMD IOMMU entry in the ACPI table and
729 * initializes the hardware and our data structures with it.
730 */
731static void __init init_iommu_from_acpi(struct amd_iommu *iommu,
732 struct ivhd_header *h)
733{
734 u8 *p = (u8 *)h;
735 u8 *end = p, flags = 0;
736 u16 devid = 0, devid_start = 0, devid_to = 0;
737 u32 dev_i, ext_flags = 0;
738 bool alias = false;
739 struct ivhd_entry *e;
740
741 /*
742 * First save the recommended feature enable bits from ACPI
743 */
744 iommu->acpi_flags = h->flags;
745
746 /*
747 * Done. Now parse the device entries
748 */
749 p += sizeof(struct ivhd_header);
750 end += h->length;
751
752
753 while (p < end) {
754 e = (struct ivhd_entry *)p;
755 switch (e->type) {
756 case IVHD_DEV_ALL:
757
758 DUMP_printk(" DEV_ALL\t\t\t first devid: %02x:%02x.%x"
759 " last device %02x:%02x.%x flags: %02x\n",
760 PCI_BUS(iommu->first_device),
761 PCI_SLOT(iommu->first_device),
762 PCI_FUNC(iommu->first_device),
763 PCI_BUS(iommu->last_device),
764 PCI_SLOT(iommu->last_device),
765 PCI_FUNC(iommu->last_device),
766 e->flags);
767
768 for (dev_i = iommu->first_device;
769 dev_i <= iommu->last_device; ++dev_i)
770 set_dev_entry_from_acpi(iommu, dev_i,
771 e->flags, 0);
772 break;
773 case IVHD_DEV_SELECT:
774
775 DUMP_printk(" DEV_SELECT\t\t\t devid: %02x:%02x.%x "
776 "flags: %02x\n",
777 PCI_BUS(e->devid),
778 PCI_SLOT(e->devid),
779 PCI_FUNC(e->devid),
780 e->flags);
781
782 devid = e->devid;
783 set_dev_entry_from_acpi(iommu, devid, e->flags, 0);
784 break;
785 case IVHD_DEV_SELECT_RANGE_START:
786
787 DUMP_printk(" DEV_SELECT_RANGE_START\t "
788 "devid: %02x:%02x.%x flags: %02x\n",
789 PCI_BUS(e->devid),
790 PCI_SLOT(e->devid),
791 PCI_FUNC(e->devid),
792 e->flags);
793
794 devid_start = e->devid;
795 flags = e->flags;
796 ext_flags = 0;
797 alias = false;
798 break;
799 case IVHD_DEV_ALIAS:
800
801 DUMP_printk(" DEV_ALIAS\t\t\t devid: %02x:%02x.%x "
802 "flags: %02x devid_to: %02x:%02x.%x\n",
803 PCI_BUS(e->devid),
804 PCI_SLOT(e->devid),
805 PCI_FUNC(e->devid),
806 e->flags,
807 PCI_BUS(e->ext >> 8),
808 PCI_SLOT(e->ext >> 8),
809 PCI_FUNC(e->ext >> 8));
810
811 devid = e->devid;
812 devid_to = e->ext >> 8;
813 set_dev_entry_from_acpi(iommu, devid , e->flags, 0);
814 set_dev_entry_from_acpi(iommu, devid_to, e->flags, 0);
815 amd_iommu_alias_table[devid] = devid_to;
816 break;
817 case IVHD_DEV_ALIAS_RANGE:
818
819 DUMP_printk(" DEV_ALIAS_RANGE\t\t "
820 "devid: %02x:%02x.%x flags: %02x "
821 "devid_to: %02x:%02x.%x\n",
822 PCI_BUS(e->devid),
823 PCI_SLOT(e->devid),
824 PCI_FUNC(e->devid),
825 e->flags,
826 PCI_BUS(e->ext >> 8),
827 PCI_SLOT(e->ext >> 8),
828 PCI_FUNC(e->ext >> 8));
829
830 devid_start = e->devid;
831 flags = e->flags;
832 devid_to = e->ext >> 8;
833 ext_flags = 0;
834 alias = true;
835 break;
836 case IVHD_DEV_EXT_SELECT:
837
838 DUMP_printk(" DEV_EXT_SELECT\t\t devid: %02x:%02x.%x "
839 "flags: %02x ext: %08x\n",
840 PCI_BUS(e->devid),
841 PCI_SLOT(e->devid),
842 PCI_FUNC(e->devid),
843 e->flags, e->ext);
844
845 devid = e->devid;
846 set_dev_entry_from_acpi(iommu, devid, e->flags,
847 e->ext);
848 break;
849 case IVHD_DEV_EXT_SELECT_RANGE:
850
851 DUMP_printk(" DEV_EXT_SELECT_RANGE\t devid: "
852 "%02x:%02x.%x flags: %02x ext: %08x\n",
853 PCI_BUS(e->devid),
854 PCI_SLOT(e->devid),
855 PCI_FUNC(e->devid),
856 e->flags, e->ext);
857
858 devid_start = e->devid;
859 flags = e->flags;
860 ext_flags = e->ext;
861 alias = false;
862 break;
863 case IVHD_DEV_RANGE_END:
864
865 DUMP_printk(" DEV_RANGE_END\t\t devid: %02x:%02x.%x\n",
866 PCI_BUS(e->devid),
867 PCI_SLOT(e->devid),
868 PCI_FUNC(e->devid));
869
870 devid = e->devid;
871 for (dev_i = devid_start; dev_i <= devid; ++dev_i) {
872 if (alias) {
873 amd_iommu_alias_table[dev_i] = devid_to;
874 set_dev_entry_from_acpi(iommu,
875 devid_to, flags, ext_flags);
876 }
877 set_dev_entry_from_acpi(iommu, dev_i,
878 flags, ext_flags);
879 }
880 break;
881 default:
882 break;
883 }
884
885 p += ivhd_entry_length(p);
886 }
887}
888
889/* Initializes the device->iommu mapping for the driver */
890static int __init init_iommu_devices(struct amd_iommu *iommu)
891{
892 u32 i;
893
894 for (i = iommu->first_device; i <= iommu->last_device; ++i)
895 set_iommu_for_device(iommu, i);
896
897 return 0;
898}
899
900static void __init free_iommu_one(struct amd_iommu *iommu)
901{
902 free_command_buffer(iommu);
903 free_event_buffer(iommu);
904 iommu_unmap_mmio_space(iommu);
905}
906
907static void __init free_iommu_all(void)
908{
909 struct amd_iommu *iommu, *next;
910
911 for_each_iommu_safe(iommu, next) {
912 list_del(&iommu->list);
913 free_iommu_one(iommu);
914 kfree(iommu);
915 }
916}
917
918/*
919 * This function clues the initialization function for one IOMMU
920 * together and also allocates the command buffer and programs the
921 * hardware. It does NOT enable the IOMMU. This is done afterwards.
922 */
923static int __init init_iommu_one(struct amd_iommu *iommu, struct ivhd_header *h)
924{
925 spin_lock_init(&iommu->lock);
926
927 /* Add IOMMU to internal data structures */
928 list_add_tail(&iommu->list, &amd_iommu_list);
929 iommu->index = amd_iommus_present++;
930
931 if (unlikely(iommu->index >= MAX_IOMMUS)) {
932 WARN(1, "AMD-Vi: System has more IOMMUs than supported by this driver\n");
933 return -ENOSYS;
934 }
935
936 /* Index is fine - add IOMMU to the array */
937 amd_iommus[iommu->index] = iommu;
938
939 /*
940 * Copy data from ACPI table entry to the iommu struct
941 */
942 iommu->dev = pci_get_bus_and_slot(PCI_BUS(h->devid), h->devid & 0xff);
943 if (!iommu->dev)
944 return 1;
945
946 iommu->cap_ptr = h->cap_ptr;
947 iommu->pci_seg = h->pci_seg;
948 iommu->mmio_phys = h->mmio_phys;
949 iommu->mmio_base = iommu_map_mmio_space(h->mmio_phys);
950 if (!iommu->mmio_base)
951 return -ENOMEM;
952
953 iommu->cmd_buf = alloc_command_buffer(iommu);
954 if (!iommu->cmd_buf)
955 return -ENOMEM;
956
957 iommu->evt_buf = alloc_event_buffer(iommu);
958 if (!iommu->evt_buf)
959 return -ENOMEM;
960
961 iommu->int_enabled = false;
962
963 init_iommu_from_pci(iommu);
964 init_iommu_from_acpi(iommu, h);
965 init_iommu_devices(iommu);
966
967 if (iommu->cap & (1UL << IOMMU_CAP_NPCACHE))
968 amd_iommu_np_cache = true;
969
970 return pci_enable_device(iommu->dev);
971}
972
973/*
974 * Iterates over all IOMMU entries in the ACPI table, allocates the
975 * IOMMU structure and initializes it with init_iommu_one()
976 */
977static int __init init_iommu_all(struct acpi_table_header *table)
978{
979 u8 *p = (u8 *)table, *end = (u8 *)table;
980 struct ivhd_header *h;
981 struct amd_iommu *iommu;
982 int ret;
983
984 end += table->length;
985 p += IVRS_HEADER_LENGTH;
986
987 while (p < end) {
988 h = (struct ivhd_header *)p;
989 switch (*p) {
990 case ACPI_IVHD_TYPE:
991
992 DUMP_printk("device: %02x:%02x.%01x cap: %04x "
993 "seg: %d flags: %01x info %04x\n",
994 PCI_BUS(h->devid), PCI_SLOT(h->devid),
995 PCI_FUNC(h->devid), h->cap_ptr,
996 h->pci_seg, h->flags, h->info);
997 DUMP_printk(" mmio-addr: %016llx\n",
998 h->mmio_phys);
999
1000 iommu = kzalloc(sizeof(struct amd_iommu), GFP_KERNEL);
1001 if (iommu == NULL) {
1002 amd_iommu_init_err = -ENOMEM;
1003 return 0;
1004 }
1005
1006 ret = init_iommu_one(iommu, h);
1007 if (ret) {
1008 amd_iommu_init_err = ret;
1009 return 0;
1010 }
1011 break;
1012 default:
1013 break;
1014 }
1015 p += h->length;
1016
1017 }
1018 WARN_ON(p != end);
1019
1020 return 0;
1021}
1022
1023/****************************************************************************
1024 *
1025 * The following functions initialize the MSI interrupts for all IOMMUs
1026 * in the system. Its a bit challenging because there could be multiple
1027 * IOMMUs per PCI BDF but we can call pci_enable_msi(x) only once per
1028 * pci_dev.
1029 *
1030 ****************************************************************************/
1031
1032static int iommu_setup_msi(struct amd_iommu *iommu)
1033{
1034 int r;
1035
1036 if (pci_enable_msi(iommu->dev))
1037 return 1;
1038
1039 r = request_threaded_irq(iommu->dev->irq,
1040 amd_iommu_int_handler,
1041 amd_iommu_int_thread,
1042 0, "AMD-Vi",
1043 iommu->dev);
1044
1045 if (r) {
1046 pci_disable_msi(iommu->dev);
1047 return 1;
1048 }
1049
1050 iommu->int_enabled = true;
1051 iommu_feature_enable(iommu, CONTROL_EVT_INT_EN);
1052
1053 return 0;
1054}
1055
1056static int iommu_init_msi(struct amd_iommu *iommu)
1057{
1058 if (iommu->int_enabled)
1059 return 0;
1060
1061 if (pci_find_capability(iommu->dev, PCI_CAP_ID_MSI))
1062 return iommu_setup_msi(iommu);
1063
1064 return 1;
1065}
1066
1067/****************************************************************************
1068 *
1069 * The next functions belong to the third pass of parsing the ACPI
1070 * table. In this last pass the memory mapping requirements are
1071 * gathered (like exclusion and unity mapping reanges).
1072 *
1073 ****************************************************************************/
1074
1075static void __init free_unity_maps(void)
1076{
1077 struct unity_map_entry *entry, *next;
1078
1079 list_for_each_entry_safe(entry, next, &amd_iommu_unity_map, list) {
1080 list_del(&entry->list);
1081 kfree(entry);
1082 }
1083}
1084
1085/* called when we find an exclusion range definition in ACPI */
1086static int __init init_exclusion_range(struct ivmd_header *m)
1087{
1088 int i;
1089
1090 switch (m->type) {
1091 case ACPI_IVMD_TYPE:
1092 set_device_exclusion_range(m->devid, m);
1093 break;
1094 case ACPI_IVMD_TYPE_ALL:
1095 for (i = 0; i <= amd_iommu_last_bdf; ++i)
1096 set_device_exclusion_range(i, m);
1097 break;
1098 case ACPI_IVMD_TYPE_RANGE:
1099 for (i = m->devid; i <= m->aux; ++i)
1100 set_device_exclusion_range(i, m);
1101 break;
1102 default:
1103 break;
1104 }
1105
1106 return 0;
1107}
1108
1109/* called for unity map ACPI definition */
1110static int __init init_unity_map_range(struct ivmd_header *m)
1111{
1112 struct unity_map_entry *e = 0;
1113 char *s;
1114
1115 e = kzalloc(sizeof(*e), GFP_KERNEL);
1116 if (e == NULL)
1117 return -ENOMEM;
1118
1119 switch (m->type) {
1120 default:
1121 kfree(e);
1122 return 0;
1123 case ACPI_IVMD_TYPE:
1124 s = "IVMD_TYPEi\t\t\t";
1125 e->devid_start = e->devid_end = m->devid;
1126 break;
1127 case ACPI_IVMD_TYPE_ALL:
1128 s = "IVMD_TYPE_ALL\t\t";
1129 e->devid_start = 0;
1130 e->devid_end = amd_iommu_last_bdf;
1131 break;
1132 case ACPI_IVMD_TYPE_RANGE:
1133 s = "IVMD_TYPE_RANGE\t\t";
1134 e->devid_start = m->devid;
1135 e->devid_end = m->aux;
1136 break;
1137 }
1138 e->address_start = PAGE_ALIGN(m->range_start);
1139 e->address_end = e->address_start + PAGE_ALIGN(m->range_length);
1140 e->prot = m->flags >> 1;
1141
1142 DUMP_printk("%s devid_start: %02x:%02x.%x devid_end: %02x:%02x.%x"
1143 " range_start: %016llx range_end: %016llx flags: %x\n", s,
1144 PCI_BUS(e->devid_start), PCI_SLOT(e->devid_start),
1145 PCI_FUNC(e->devid_start), PCI_BUS(e->devid_end),
1146 PCI_SLOT(e->devid_end), PCI_FUNC(e->devid_end),
1147 e->address_start, e->address_end, m->flags);
1148
1149 list_add_tail(&e->list, &amd_iommu_unity_map);
1150
1151 return 0;
1152}
1153
1154/* iterates over all memory definitions we find in the ACPI table */
1155static int __init init_memory_definitions(struct acpi_table_header *table)
1156{
1157 u8 *p = (u8 *)table, *end = (u8 *)table;
1158 struct ivmd_header *m;
1159
1160 end += table->length;
1161 p += IVRS_HEADER_LENGTH;
1162
1163 while (p < end) {
1164 m = (struct ivmd_header *)p;
1165 if (m->flags & IVMD_FLAG_EXCL_RANGE)
1166 init_exclusion_range(m);
1167 else if (m->flags & IVMD_FLAG_UNITY_MAP)
1168 init_unity_map_range(m);
1169
1170 p += m->length;
1171 }
1172
1173 return 0;
1174}
1175
1176/*
1177 * Init the device table to not allow DMA access for devices and
1178 * suppress all page faults
1179 */
1180static void init_device_table(void)
1181{
1182 u32 devid;
1183
1184 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid) {
1185 set_dev_entry_bit(devid, DEV_ENTRY_VALID);
1186 set_dev_entry_bit(devid, DEV_ENTRY_TRANSLATION);
1187 }
1188}
1189
1190static void iommu_init_flags(struct amd_iommu *iommu)
1191{
1192 iommu->acpi_flags & IVHD_FLAG_HT_TUN_EN_MASK ?
1193 iommu_feature_enable(iommu, CONTROL_HT_TUN_EN) :
1194 iommu_feature_disable(iommu, CONTROL_HT_TUN_EN);
1195
1196 iommu->acpi_flags & IVHD_FLAG_PASSPW_EN_MASK ?
1197 iommu_feature_enable(iommu, CONTROL_PASSPW_EN) :
1198 iommu_feature_disable(iommu, CONTROL_PASSPW_EN);
1199
1200 iommu->acpi_flags & IVHD_FLAG_RESPASSPW_EN_MASK ?
1201 iommu_feature_enable(iommu, CONTROL_RESPASSPW_EN) :
1202 iommu_feature_disable(iommu, CONTROL_RESPASSPW_EN);
1203
1204 iommu->acpi_flags & IVHD_FLAG_ISOC_EN_MASK ?
1205 iommu_feature_enable(iommu, CONTROL_ISOC_EN) :
1206 iommu_feature_disable(iommu, CONTROL_ISOC_EN);
1207
1208 /*
1209 * make IOMMU memory accesses cache coherent
1210 */
1211 iommu_feature_enable(iommu, CONTROL_COHERENT_EN);
1212}
1213
1214static void iommu_apply_resume_quirks(struct amd_iommu *iommu)
1215{
1216 int i, j;
1217 u32 ioc_feature_control;
1218 struct pci_dev *pdev = NULL;
1219
1220 /* RD890 BIOSes may not have completely reconfigured the iommu */
1221 if (!is_rd890_iommu(iommu->dev))
1222 return;
1223
1224 /*
1225 * First, we need to ensure that the iommu is enabled. This is
1226 * controlled by a register in the northbridge
1227 */
1228 pdev = pci_get_bus_and_slot(iommu->dev->bus->number, PCI_DEVFN(0, 0));
1229
1230 if (!pdev)
1231 return;
1232
1233 /* Select Northbridge indirect register 0x75 and enable writing */
1234 pci_write_config_dword(pdev, 0x60, 0x75 | (1 << 7));
1235 pci_read_config_dword(pdev, 0x64, &ioc_feature_control);
1236
1237 /* Enable the iommu */
1238 if (!(ioc_feature_control & 0x1))
1239 pci_write_config_dword(pdev, 0x64, ioc_feature_control | 1);
1240
1241 pci_dev_put(pdev);
1242
1243 /* Restore the iommu BAR */
1244 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 4,
1245 iommu->stored_addr_lo);
1246 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 8,
1247 iommu->stored_addr_hi);
1248
1249 /* Restore the l1 indirect regs for each of the 6 l1s */
1250 for (i = 0; i < 6; i++)
1251 for (j = 0; j < 0x12; j++)
1252 iommu_write_l1(iommu, i, j, iommu->stored_l1[i][j]);
1253
1254 /* Restore the l2 indirect regs */
1255 for (i = 0; i < 0x83; i++)
1256 iommu_write_l2(iommu, i, iommu->stored_l2[i]);
1257
1258 /* Lock PCI setup registers */
1259 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 4,
1260 iommu->stored_addr_lo | 1);
1261}
1262
1263/*
1264 * This function finally enables all IOMMUs found in the system after
1265 * they have been initialized
1266 */
1267static void enable_iommus(void)
1268{
1269 struct amd_iommu *iommu;
1270
1271 for_each_iommu(iommu) {
1272 iommu_disable(iommu);
1273 iommu_init_flags(iommu);
1274 iommu_set_device_table(iommu);
1275 iommu_enable_command_buffer(iommu);
1276 iommu_enable_event_buffer(iommu);
1277 iommu_set_exclusion_range(iommu);
1278 iommu_init_msi(iommu);
1279 iommu_enable(iommu);
1280 iommu_flush_all_caches(iommu);
1281 }
1282}
1283
1284static void disable_iommus(void)
1285{
1286 struct amd_iommu *iommu;
1287
1288 for_each_iommu(iommu)
1289 iommu_disable(iommu);
1290}
1291
1292/*
1293 * Suspend/Resume support
1294 * disable suspend until real resume implemented
1295 */
1296
1297static void amd_iommu_resume(void)
1298{
1299 struct amd_iommu *iommu;
1300
1301 for_each_iommu(iommu)
1302 iommu_apply_resume_quirks(iommu);
1303
1304 /* re-load the hardware */
1305 enable_iommus();
1306
1307 /*
1308 * we have to flush after the IOMMUs are enabled because a
1309 * disabled IOMMU will never execute the commands we send
1310 */
1311 for_each_iommu(iommu)
1312 iommu_flush_all_caches(iommu);
1313}
1314
1315static int amd_iommu_suspend(void)
1316{
1317 /* disable IOMMUs to go out of the way for BIOS */
1318 disable_iommus();
1319
1320 return 0;
1321}
1322
1323static struct syscore_ops amd_iommu_syscore_ops = {
1324 .suspend = amd_iommu_suspend,
1325 .resume = amd_iommu_resume,
1326};
1327
1328/*
1329 * This is the core init function for AMD IOMMU hardware in the system.
1330 * This function is called from the generic x86 DMA layer initialization
1331 * code.
1332 *
1333 * This function basically parses the ACPI table for AMD IOMMU (IVRS)
1334 * three times:
1335 *
1336 * 1 pass) Find the highest PCI device id the driver has to handle.
1337 * Upon this information the size of the data structures is
1338 * determined that needs to be allocated.
1339 *
1340 * 2 pass) Initialize the data structures just allocated with the
1341 * information in the ACPI table about available AMD IOMMUs
1342 * in the system. It also maps the PCI devices in the
1343 * system to specific IOMMUs
1344 *
1345 * 3 pass) After the basic data structures are allocated and
1346 * initialized we update them with information about memory
1347 * remapping requirements parsed out of the ACPI table in
1348 * this last pass.
1349 *
1350 * After that the hardware is initialized and ready to go. In the last
1351 * step we do some Linux specific things like registering the driver in
1352 * the dma_ops interface and initializing the suspend/resume support
1353 * functions. Finally it prints some information about AMD IOMMUs and
1354 * the driver state and enables the hardware.
1355 */
1356static int __init amd_iommu_init(void)
1357{
1358 int i, ret = 0;
1359
1360 /*
1361 * First parse ACPI tables to find the largest Bus/Dev/Func
1362 * we need to handle. Upon this information the shared data
1363 * structures for the IOMMUs in the system will be allocated
1364 */
1365 if (acpi_table_parse("IVRS", find_last_devid_acpi) != 0)
1366 return -ENODEV;
1367
1368 ret = amd_iommu_init_err;
1369 if (ret)
1370 goto out;
1371
1372 dev_table_size = tbl_size(DEV_TABLE_ENTRY_SIZE);
1373 alias_table_size = tbl_size(ALIAS_TABLE_ENTRY_SIZE);
1374 rlookup_table_size = tbl_size(RLOOKUP_TABLE_ENTRY_SIZE);
1375
1376 ret = -ENOMEM;
1377
1378 /* Device table - directly used by all IOMMUs */
1379 amd_iommu_dev_table = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1380 get_order(dev_table_size));
1381 if (amd_iommu_dev_table == NULL)
1382 goto out;
1383
1384 /*
1385 * Alias table - map PCI Bus/Dev/Func to Bus/Dev/Func the
1386 * IOMMU see for that device
1387 */
1388 amd_iommu_alias_table = (void *)__get_free_pages(GFP_KERNEL,
1389 get_order(alias_table_size));
1390 if (amd_iommu_alias_table == NULL)
1391 goto free;
1392
1393 /* IOMMU rlookup table - find the IOMMU for a specific device */
1394 amd_iommu_rlookup_table = (void *)__get_free_pages(
1395 GFP_KERNEL | __GFP_ZERO,
1396 get_order(rlookup_table_size));
1397 if (amd_iommu_rlookup_table == NULL)
1398 goto free;
1399
1400 amd_iommu_pd_alloc_bitmap = (void *)__get_free_pages(
1401 GFP_KERNEL | __GFP_ZERO,
1402 get_order(MAX_DOMAIN_ID/8));
1403 if (amd_iommu_pd_alloc_bitmap == NULL)
1404 goto free;
1405
1406 /* init the device table */
1407 init_device_table();
1408
1409 /*
1410 * let all alias entries point to itself
1411 */
1412 for (i = 0; i <= amd_iommu_last_bdf; ++i)
1413 amd_iommu_alias_table[i] = i;
1414
1415 /*
1416 * never allocate domain 0 because its used as the non-allocated and
1417 * error value placeholder
1418 */
1419 amd_iommu_pd_alloc_bitmap[0] = 1;
1420
1421 spin_lock_init(&amd_iommu_pd_lock);
1422
1423 /*
1424 * now the data structures are allocated and basically initialized
1425 * start the real acpi table scan
1426 */
1427 ret = -ENODEV;
1428 if (acpi_table_parse("IVRS", init_iommu_all) != 0)
1429 goto free;
1430
1431 if (amd_iommu_init_err) {
1432 ret = amd_iommu_init_err;
1433 goto free;
1434 }
1435
1436 if (acpi_table_parse("IVRS", init_memory_definitions) != 0)
1437 goto free;
1438
1439 if (amd_iommu_init_err) {
1440 ret = amd_iommu_init_err;
1441 goto free;
1442 }
1443
1444 ret = amd_iommu_init_devices();
1445 if (ret)
1446 goto free;
1447
1448 enable_iommus();
1449
1450 if (iommu_pass_through)
1451 ret = amd_iommu_init_passthrough();
1452 else
1453 ret = amd_iommu_init_dma_ops();
1454
1455 if (ret)
1456 goto free_disable;
1457
1458 amd_iommu_init_api();
1459
1460 amd_iommu_init_notifier();
1461
1462 register_syscore_ops(&amd_iommu_syscore_ops);
1463
1464 if (iommu_pass_through)
1465 goto out;
1466
1467 if (amd_iommu_unmap_flush)
1468 printk(KERN_INFO "AMD-Vi: IO/TLB flush on unmap enabled\n");
1469 else
1470 printk(KERN_INFO "AMD-Vi: Lazy IO/TLB flushing enabled\n");
1471
1472 x86_platform.iommu_shutdown = disable_iommus;
1473out:
1474 return ret;
1475
1476free_disable:
1477 disable_iommus();
1478
1479free:
1480 amd_iommu_uninit_devices();
1481
1482 free_pages((unsigned long)amd_iommu_pd_alloc_bitmap,
1483 get_order(MAX_DOMAIN_ID/8));
1484
1485 free_pages((unsigned long)amd_iommu_rlookup_table,
1486 get_order(rlookup_table_size));
1487
1488 free_pages((unsigned long)amd_iommu_alias_table,
1489 get_order(alias_table_size));
1490
1491 free_pages((unsigned long)amd_iommu_dev_table,
1492 get_order(dev_table_size));
1493
1494 free_iommu_all();
1495
1496 free_unity_maps();
1497
1498#ifdef CONFIG_GART_IOMMU
1499 /*
1500 * We failed to initialize the AMD IOMMU - try fallback to GART
1501 * if possible.
1502 */
1503 gart_iommu_init();
1504
1505#endif
1506
1507 goto out;
1508}
1509
1510/****************************************************************************
1511 *
1512 * Early detect code. This code runs at IOMMU detection time in the DMA
1513 * layer. It just looks if there is an IVRS ACPI table to detect AMD
1514 * IOMMUs
1515 *
1516 ****************************************************************************/
1517static int __init early_amd_iommu_detect(struct acpi_table_header *table)
1518{
1519 return 0;
1520}
1521
1522int __init amd_iommu_detect(void)
1523{
1524 if (no_iommu || (iommu_detected && !gart_iommu_aperture))
1525 return -ENODEV;
1526
1527 if (amd_iommu_disabled)
1528 return -ENODEV;
1529
1530 if (acpi_table_parse("IVRS", early_amd_iommu_detect) == 0) {
1531 iommu_detected = 1;
1532 amd_iommu_detected = 1;
1533 x86_init.iommu.iommu_init = amd_iommu_init;
1534
1535 /* Make sure ACS will be enabled */
1536 pci_request_acs();
1537 return 1;
1538 }
1539 return -ENODEV;
1540}
1541
1542/****************************************************************************
1543 *
1544 * Parsing functions for the AMD IOMMU specific kernel command line
1545 * options.
1546 *
1547 ****************************************************************************/
1548
1549static int __init parse_amd_iommu_dump(char *str)
1550{
1551 amd_iommu_dump = true;
1552
1553 return 1;
1554}
1555
1556static int __init parse_amd_iommu_options(char *str)
1557{
1558 for (; *str; ++str) {
1559 if (strncmp(str, "fullflush", 9) == 0)
1560 amd_iommu_unmap_flush = true;
1561 if (strncmp(str, "off", 3) == 0)
1562 amd_iommu_disabled = true;
1563 }
1564
1565 return 1;
1566}
1567
1568__setup("amd_iommu_dump", parse_amd_iommu_dump);
1569__setup("amd_iommu=", parse_amd_iommu_options);
1570
1571IOMMU_INIT_FINISH(amd_iommu_detect,
1572 gart_iommu_hole_init,
1573 0,
1574 0);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
4 * Author: Joerg Roedel <jroedel@suse.de>
5 * Leo Duran <leo.duran@amd.com>
6 */
7
8#define pr_fmt(fmt) "AMD-Vi: " fmt
9#define dev_fmt(fmt) pr_fmt(fmt)
10
11#include <linux/pci.h>
12#include <linux/acpi.h>
13#include <linux/list.h>
14#include <linux/bitmap.h>
15#include <linux/slab.h>
16#include <linux/syscore_ops.h>
17#include <linux/interrupt.h>
18#include <linux/msi.h>
19#include <linux/amd-iommu.h>
20#include <linux/export.h>
21#include <linux/iommu.h>
22#include <linux/kmemleak.h>
23#include <linux/mem_encrypt.h>
24#include <asm/pci-direct.h>
25#include <asm/iommu.h>
26#include <asm/apic.h>
27#include <asm/msidef.h>
28#include <asm/gart.h>
29#include <asm/x86_init.h>
30#include <asm/iommu_table.h>
31#include <asm/io_apic.h>
32#include <asm/irq_remapping.h>
33
34#include <linux/crash_dump.h>
35#include "amd_iommu.h"
36#include "amd_iommu_proto.h"
37#include "amd_iommu_types.h"
38#include "irq_remapping.h"
39
40/*
41 * definitions for the ACPI scanning code
42 */
43#define IVRS_HEADER_LENGTH 48
44
45#define ACPI_IVHD_TYPE_MAX_SUPPORTED 0x40
46#define ACPI_IVMD_TYPE_ALL 0x20
47#define ACPI_IVMD_TYPE 0x21
48#define ACPI_IVMD_TYPE_RANGE 0x22
49
50#define IVHD_DEV_ALL 0x01
51#define IVHD_DEV_SELECT 0x02
52#define IVHD_DEV_SELECT_RANGE_START 0x03
53#define IVHD_DEV_RANGE_END 0x04
54#define IVHD_DEV_ALIAS 0x42
55#define IVHD_DEV_ALIAS_RANGE 0x43
56#define IVHD_DEV_EXT_SELECT 0x46
57#define IVHD_DEV_EXT_SELECT_RANGE 0x47
58#define IVHD_DEV_SPECIAL 0x48
59#define IVHD_DEV_ACPI_HID 0xf0
60
61#define UID_NOT_PRESENT 0
62#define UID_IS_INTEGER 1
63#define UID_IS_CHARACTER 2
64
65#define IVHD_SPECIAL_IOAPIC 1
66#define IVHD_SPECIAL_HPET 2
67
68#define IVHD_FLAG_HT_TUN_EN_MASK 0x01
69#define IVHD_FLAG_PASSPW_EN_MASK 0x02
70#define IVHD_FLAG_RESPASSPW_EN_MASK 0x04
71#define IVHD_FLAG_ISOC_EN_MASK 0x08
72
73#define IVMD_FLAG_EXCL_RANGE 0x08
74#define IVMD_FLAG_UNITY_MAP 0x01
75
76#define ACPI_DEVFLAG_INITPASS 0x01
77#define ACPI_DEVFLAG_EXTINT 0x02
78#define ACPI_DEVFLAG_NMI 0x04
79#define ACPI_DEVFLAG_SYSMGT1 0x10
80#define ACPI_DEVFLAG_SYSMGT2 0x20
81#define ACPI_DEVFLAG_LINT0 0x40
82#define ACPI_DEVFLAG_LINT1 0x80
83#define ACPI_DEVFLAG_ATSDIS 0x10000000
84
85#define LOOP_TIMEOUT 100000
86/*
87 * ACPI table definitions
88 *
89 * These data structures are laid over the table to parse the important values
90 * out of it.
91 */
92
93extern const struct iommu_ops amd_iommu_ops;
94
95/*
96 * structure describing one IOMMU in the ACPI table. Typically followed by one
97 * or more ivhd_entrys.
98 */
99struct ivhd_header {
100 u8 type;
101 u8 flags;
102 u16 length;
103 u16 devid;
104 u16 cap_ptr;
105 u64 mmio_phys;
106 u16 pci_seg;
107 u16 info;
108 u32 efr_attr;
109
110 /* Following only valid on IVHD type 11h and 40h */
111 u64 efr_reg; /* Exact copy of MMIO_EXT_FEATURES */
112 u64 res;
113} __attribute__((packed));
114
115/*
116 * A device entry describing which devices a specific IOMMU translates and
117 * which requestor ids they use.
118 */
119struct ivhd_entry {
120 u8 type;
121 u16 devid;
122 u8 flags;
123 u32 ext;
124 u32 hidh;
125 u64 cid;
126 u8 uidf;
127 u8 uidl;
128 u8 uid;
129} __attribute__((packed));
130
131/*
132 * An AMD IOMMU memory definition structure. It defines things like exclusion
133 * ranges for devices and regions that should be unity mapped.
134 */
135struct ivmd_header {
136 u8 type;
137 u8 flags;
138 u16 length;
139 u16 devid;
140 u16 aux;
141 u64 resv;
142 u64 range_start;
143 u64 range_length;
144} __attribute__((packed));
145
146bool amd_iommu_dump;
147bool amd_iommu_irq_remap __read_mostly;
148
149int amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_VAPIC;
150static int amd_iommu_xt_mode = IRQ_REMAP_X2APIC_MODE;
151
152static bool amd_iommu_detected;
153static bool __initdata amd_iommu_disabled;
154static int amd_iommu_target_ivhd_type;
155
156u16 amd_iommu_last_bdf; /* largest PCI device id we have
157 to handle */
158LIST_HEAD(amd_iommu_unity_map); /* a list of required unity mappings
159 we find in ACPI */
160bool amd_iommu_unmap_flush; /* if true, flush on every unmap */
161
162LIST_HEAD(amd_iommu_list); /* list of all AMD IOMMUs in the
163 system */
164
165/* Array to assign indices to IOMMUs*/
166struct amd_iommu *amd_iommus[MAX_IOMMUS];
167
168/* Number of IOMMUs present in the system */
169static int amd_iommus_present;
170
171/* IOMMUs have a non-present cache? */
172bool amd_iommu_np_cache __read_mostly;
173bool amd_iommu_iotlb_sup __read_mostly = true;
174
175u32 amd_iommu_max_pasid __read_mostly = ~0;
176
177bool amd_iommu_v2_present __read_mostly;
178static bool amd_iommu_pc_present __read_mostly;
179
180bool amd_iommu_force_isolation __read_mostly;
181
182/*
183 * Pointer to the device table which is shared by all AMD IOMMUs
184 * it is indexed by the PCI device id or the HT unit id and contains
185 * information about the domain the device belongs to as well as the
186 * page table root pointer.
187 */
188struct dev_table_entry *amd_iommu_dev_table;
189/*
190 * Pointer to a device table which the content of old device table
191 * will be copied to. It's only be used in kdump kernel.
192 */
193static struct dev_table_entry *old_dev_tbl_cpy;
194
195/*
196 * The alias table is a driver specific data structure which contains the
197 * mappings of the PCI device ids to the actual requestor ids on the IOMMU.
198 * More than one device can share the same requestor id.
199 */
200u16 *amd_iommu_alias_table;
201
202/*
203 * The rlookup table is used to find the IOMMU which is responsible
204 * for a specific device. It is also indexed by the PCI device id.
205 */
206struct amd_iommu **amd_iommu_rlookup_table;
207EXPORT_SYMBOL(amd_iommu_rlookup_table);
208
209/*
210 * This table is used to find the irq remapping table for a given device id
211 * quickly.
212 */
213struct irq_remap_table **irq_lookup_table;
214
215/*
216 * AMD IOMMU allows up to 2^16 different protection domains. This is a bitmap
217 * to know which ones are already in use.
218 */
219unsigned long *amd_iommu_pd_alloc_bitmap;
220
221static u32 dev_table_size; /* size of the device table */
222static u32 alias_table_size; /* size of the alias table */
223static u32 rlookup_table_size; /* size if the rlookup table */
224
225enum iommu_init_state {
226 IOMMU_START_STATE,
227 IOMMU_IVRS_DETECTED,
228 IOMMU_ACPI_FINISHED,
229 IOMMU_ENABLED,
230 IOMMU_PCI_INIT,
231 IOMMU_INTERRUPTS_EN,
232 IOMMU_DMA_OPS,
233 IOMMU_INITIALIZED,
234 IOMMU_NOT_FOUND,
235 IOMMU_INIT_ERROR,
236 IOMMU_CMDLINE_DISABLED,
237};
238
239/* Early ioapic and hpet maps from kernel command line */
240#define EARLY_MAP_SIZE 4
241static struct devid_map __initdata early_ioapic_map[EARLY_MAP_SIZE];
242static struct devid_map __initdata early_hpet_map[EARLY_MAP_SIZE];
243static struct acpihid_map_entry __initdata early_acpihid_map[EARLY_MAP_SIZE];
244
245static int __initdata early_ioapic_map_size;
246static int __initdata early_hpet_map_size;
247static int __initdata early_acpihid_map_size;
248
249static bool __initdata cmdline_maps;
250
251static enum iommu_init_state init_state = IOMMU_START_STATE;
252
253static int amd_iommu_enable_interrupts(void);
254static int __init iommu_go_to_state(enum iommu_init_state state);
255static void init_device_table_dma(void);
256
257static bool amd_iommu_pre_enabled = true;
258
259bool translation_pre_enabled(struct amd_iommu *iommu)
260{
261 return (iommu->flags & AMD_IOMMU_FLAG_TRANS_PRE_ENABLED);
262}
263EXPORT_SYMBOL(translation_pre_enabled);
264
265static void clear_translation_pre_enabled(struct amd_iommu *iommu)
266{
267 iommu->flags &= ~AMD_IOMMU_FLAG_TRANS_PRE_ENABLED;
268}
269
270static void init_translation_status(struct amd_iommu *iommu)
271{
272 u64 ctrl;
273
274 ctrl = readq(iommu->mmio_base + MMIO_CONTROL_OFFSET);
275 if (ctrl & (1<<CONTROL_IOMMU_EN))
276 iommu->flags |= AMD_IOMMU_FLAG_TRANS_PRE_ENABLED;
277}
278
279static inline void update_last_devid(u16 devid)
280{
281 if (devid > amd_iommu_last_bdf)
282 amd_iommu_last_bdf = devid;
283}
284
285static inline unsigned long tbl_size(int entry_size)
286{
287 unsigned shift = PAGE_SHIFT +
288 get_order(((int)amd_iommu_last_bdf + 1) * entry_size);
289
290 return 1UL << shift;
291}
292
293int amd_iommu_get_num_iommus(void)
294{
295 return amd_iommus_present;
296}
297
298/* Access to l1 and l2 indexed register spaces */
299
300static u32 iommu_read_l1(struct amd_iommu *iommu, u16 l1, u8 address)
301{
302 u32 val;
303
304 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16));
305 pci_read_config_dword(iommu->dev, 0xfc, &val);
306 return val;
307}
308
309static void iommu_write_l1(struct amd_iommu *iommu, u16 l1, u8 address, u32 val)
310{
311 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16 | 1 << 31));
312 pci_write_config_dword(iommu->dev, 0xfc, val);
313 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16));
314}
315
316static u32 iommu_read_l2(struct amd_iommu *iommu, u8 address)
317{
318 u32 val;
319
320 pci_write_config_dword(iommu->dev, 0xf0, address);
321 pci_read_config_dword(iommu->dev, 0xf4, &val);
322 return val;
323}
324
325static void iommu_write_l2(struct amd_iommu *iommu, u8 address, u32 val)
326{
327 pci_write_config_dword(iommu->dev, 0xf0, (address | 1 << 8));
328 pci_write_config_dword(iommu->dev, 0xf4, val);
329}
330
331/****************************************************************************
332 *
333 * AMD IOMMU MMIO register space handling functions
334 *
335 * These functions are used to program the IOMMU device registers in
336 * MMIO space required for that driver.
337 *
338 ****************************************************************************/
339
340/*
341 * This function set the exclusion range in the IOMMU. DMA accesses to the
342 * exclusion range are passed through untranslated
343 */
344static void iommu_set_exclusion_range(struct amd_iommu *iommu)
345{
346 u64 start = iommu->exclusion_start & PAGE_MASK;
347 u64 limit = (start + iommu->exclusion_length - 1) & PAGE_MASK;
348 u64 entry;
349
350 if (!iommu->exclusion_start)
351 return;
352
353 entry = start | MMIO_EXCL_ENABLE_MASK;
354 memcpy_toio(iommu->mmio_base + MMIO_EXCL_BASE_OFFSET,
355 &entry, sizeof(entry));
356
357 entry = limit;
358 memcpy_toio(iommu->mmio_base + MMIO_EXCL_LIMIT_OFFSET,
359 &entry, sizeof(entry));
360}
361
362/* Programs the physical address of the device table into the IOMMU hardware */
363static void iommu_set_device_table(struct amd_iommu *iommu)
364{
365 u64 entry;
366
367 BUG_ON(iommu->mmio_base == NULL);
368
369 entry = iommu_virt_to_phys(amd_iommu_dev_table);
370 entry |= (dev_table_size >> 12) - 1;
371 memcpy_toio(iommu->mmio_base + MMIO_DEV_TABLE_OFFSET,
372 &entry, sizeof(entry));
373}
374
375/* Generic functions to enable/disable certain features of the IOMMU. */
376static void iommu_feature_enable(struct amd_iommu *iommu, u8 bit)
377{
378 u64 ctrl;
379
380 ctrl = readq(iommu->mmio_base + MMIO_CONTROL_OFFSET);
381 ctrl |= (1ULL << bit);
382 writeq(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET);
383}
384
385static void iommu_feature_disable(struct amd_iommu *iommu, u8 bit)
386{
387 u64 ctrl;
388
389 ctrl = readq(iommu->mmio_base + MMIO_CONTROL_OFFSET);
390 ctrl &= ~(1ULL << bit);
391 writeq(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET);
392}
393
394static void iommu_set_inv_tlb_timeout(struct amd_iommu *iommu, int timeout)
395{
396 u64 ctrl;
397
398 ctrl = readq(iommu->mmio_base + MMIO_CONTROL_OFFSET);
399 ctrl &= ~CTRL_INV_TO_MASK;
400 ctrl |= (timeout << CONTROL_INV_TIMEOUT) & CTRL_INV_TO_MASK;
401 writeq(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET);
402}
403
404/* Function to enable the hardware */
405static void iommu_enable(struct amd_iommu *iommu)
406{
407 iommu_feature_enable(iommu, CONTROL_IOMMU_EN);
408}
409
410static void iommu_disable(struct amd_iommu *iommu)
411{
412 if (!iommu->mmio_base)
413 return;
414
415 /* Disable command buffer */
416 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN);
417
418 /* Disable event logging and event interrupts */
419 iommu_feature_disable(iommu, CONTROL_EVT_INT_EN);
420 iommu_feature_disable(iommu, CONTROL_EVT_LOG_EN);
421
422 /* Disable IOMMU GA_LOG */
423 iommu_feature_disable(iommu, CONTROL_GALOG_EN);
424 iommu_feature_disable(iommu, CONTROL_GAINT_EN);
425
426 /* Disable IOMMU hardware itself */
427 iommu_feature_disable(iommu, CONTROL_IOMMU_EN);
428}
429
430/*
431 * mapping and unmapping functions for the IOMMU MMIO space. Each AMD IOMMU in
432 * the system has one.
433 */
434static u8 __iomem * __init iommu_map_mmio_space(u64 address, u64 end)
435{
436 if (!request_mem_region(address, end, "amd_iommu")) {
437 pr_err("Can not reserve memory region %llx-%llx for mmio\n",
438 address, end);
439 pr_err("This is a BIOS bug. Please contact your hardware vendor\n");
440 return NULL;
441 }
442
443 return (u8 __iomem *)ioremap_nocache(address, end);
444}
445
446static void __init iommu_unmap_mmio_space(struct amd_iommu *iommu)
447{
448 if (iommu->mmio_base)
449 iounmap(iommu->mmio_base);
450 release_mem_region(iommu->mmio_phys, iommu->mmio_phys_end);
451}
452
453static inline u32 get_ivhd_header_size(struct ivhd_header *h)
454{
455 u32 size = 0;
456
457 switch (h->type) {
458 case 0x10:
459 size = 24;
460 break;
461 case 0x11:
462 case 0x40:
463 size = 40;
464 break;
465 }
466 return size;
467}
468
469/****************************************************************************
470 *
471 * The functions below belong to the first pass of AMD IOMMU ACPI table
472 * parsing. In this pass we try to find out the highest device id this
473 * code has to handle. Upon this information the size of the shared data
474 * structures is determined later.
475 *
476 ****************************************************************************/
477
478/*
479 * This function calculates the length of a given IVHD entry
480 */
481static inline int ivhd_entry_length(u8 *ivhd)
482{
483 u32 type = ((struct ivhd_entry *)ivhd)->type;
484
485 if (type < 0x80) {
486 return 0x04 << (*ivhd >> 6);
487 } else if (type == IVHD_DEV_ACPI_HID) {
488 /* For ACPI_HID, offset 21 is uid len */
489 return *((u8 *)ivhd + 21) + 22;
490 }
491 return 0;
492}
493
494/*
495 * After reading the highest device id from the IOMMU PCI capability header
496 * this function looks if there is a higher device id defined in the ACPI table
497 */
498static int __init find_last_devid_from_ivhd(struct ivhd_header *h)
499{
500 u8 *p = (void *)h, *end = (void *)h;
501 struct ivhd_entry *dev;
502
503 u32 ivhd_size = get_ivhd_header_size(h);
504
505 if (!ivhd_size) {
506 pr_err("Unsupported IVHD type %#x\n", h->type);
507 return -EINVAL;
508 }
509
510 p += ivhd_size;
511 end += h->length;
512
513 while (p < end) {
514 dev = (struct ivhd_entry *)p;
515 switch (dev->type) {
516 case IVHD_DEV_ALL:
517 /* Use maximum BDF value for DEV_ALL */
518 update_last_devid(0xffff);
519 break;
520 case IVHD_DEV_SELECT:
521 case IVHD_DEV_RANGE_END:
522 case IVHD_DEV_ALIAS:
523 case IVHD_DEV_EXT_SELECT:
524 /* all the above subfield types refer to device ids */
525 update_last_devid(dev->devid);
526 break;
527 default:
528 break;
529 }
530 p += ivhd_entry_length(p);
531 }
532
533 WARN_ON(p != end);
534
535 return 0;
536}
537
538static int __init check_ivrs_checksum(struct acpi_table_header *table)
539{
540 int i;
541 u8 checksum = 0, *p = (u8 *)table;
542
543 for (i = 0; i < table->length; ++i)
544 checksum += p[i];
545 if (checksum != 0) {
546 /* ACPI table corrupt */
547 pr_err(FW_BUG "IVRS invalid checksum\n");
548 return -ENODEV;
549 }
550
551 return 0;
552}
553
554/*
555 * Iterate over all IVHD entries in the ACPI table and find the highest device
556 * id which we need to handle. This is the first of three functions which parse
557 * the ACPI table. So we check the checksum here.
558 */
559static int __init find_last_devid_acpi(struct acpi_table_header *table)
560{
561 u8 *p = (u8 *)table, *end = (u8 *)table;
562 struct ivhd_header *h;
563
564 p += IVRS_HEADER_LENGTH;
565
566 end += table->length;
567 while (p < end) {
568 h = (struct ivhd_header *)p;
569 if (h->type == amd_iommu_target_ivhd_type) {
570 int ret = find_last_devid_from_ivhd(h);
571
572 if (ret)
573 return ret;
574 }
575 p += h->length;
576 }
577 WARN_ON(p != end);
578
579 return 0;
580}
581
582/****************************************************************************
583 *
584 * The following functions belong to the code path which parses the ACPI table
585 * the second time. In this ACPI parsing iteration we allocate IOMMU specific
586 * data structures, initialize the device/alias/rlookup table and also
587 * basically initialize the hardware.
588 *
589 ****************************************************************************/
590
591/*
592 * Allocates the command buffer. This buffer is per AMD IOMMU. We can
593 * write commands to that buffer later and the IOMMU will execute them
594 * asynchronously
595 */
596static int __init alloc_command_buffer(struct amd_iommu *iommu)
597{
598 iommu->cmd_buf = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
599 get_order(CMD_BUFFER_SIZE));
600
601 return iommu->cmd_buf ? 0 : -ENOMEM;
602}
603
604/*
605 * This function resets the command buffer if the IOMMU stopped fetching
606 * commands from it.
607 */
608void amd_iommu_reset_cmd_buffer(struct amd_iommu *iommu)
609{
610 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN);
611
612 writel(0x00, iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
613 writel(0x00, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
614 iommu->cmd_buf_head = 0;
615 iommu->cmd_buf_tail = 0;
616
617 iommu_feature_enable(iommu, CONTROL_CMDBUF_EN);
618}
619
620/*
621 * This function writes the command buffer address to the hardware and
622 * enables it.
623 */
624static void iommu_enable_command_buffer(struct amd_iommu *iommu)
625{
626 u64 entry;
627
628 BUG_ON(iommu->cmd_buf == NULL);
629
630 entry = iommu_virt_to_phys(iommu->cmd_buf);
631 entry |= MMIO_CMD_SIZE_512;
632
633 memcpy_toio(iommu->mmio_base + MMIO_CMD_BUF_OFFSET,
634 &entry, sizeof(entry));
635
636 amd_iommu_reset_cmd_buffer(iommu);
637}
638
639/*
640 * This function disables the command buffer
641 */
642static void iommu_disable_command_buffer(struct amd_iommu *iommu)
643{
644 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN);
645}
646
647static void __init free_command_buffer(struct amd_iommu *iommu)
648{
649 free_pages((unsigned long)iommu->cmd_buf, get_order(CMD_BUFFER_SIZE));
650}
651
652/* allocates the memory where the IOMMU will log its events to */
653static int __init alloc_event_buffer(struct amd_iommu *iommu)
654{
655 iommu->evt_buf = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
656 get_order(EVT_BUFFER_SIZE));
657
658 return iommu->evt_buf ? 0 : -ENOMEM;
659}
660
661static void iommu_enable_event_buffer(struct amd_iommu *iommu)
662{
663 u64 entry;
664
665 BUG_ON(iommu->evt_buf == NULL);
666
667 entry = iommu_virt_to_phys(iommu->evt_buf) | EVT_LEN_MASK;
668
669 memcpy_toio(iommu->mmio_base + MMIO_EVT_BUF_OFFSET,
670 &entry, sizeof(entry));
671
672 /* set head and tail to zero manually */
673 writel(0x00, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
674 writel(0x00, iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
675
676 iommu_feature_enable(iommu, CONTROL_EVT_LOG_EN);
677}
678
679/*
680 * This function disables the event log buffer
681 */
682static void iommu_disable_event_buffer(struct amd_iommu *iommu)
683{
684 iommu_feature_disable(iommu, CONTROL_EVT_LOG_EN);
685}
686
687static void __init free_event_buffer(struct amd_iommu *iommu)
688{
689 free_pages((unsigned long)iommu->evt_buf, get_order(EVT_BUFFER_SIZE));
690}
691
692/* allocates the memory where the IOMMU will log its events to */
693static int __init alloc_ppr_log(struct amd_iommu *iommu)
694{
695 iommu->ppr_log = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
696 get_order(PPR_LOG_SIZE));
697
698 return iommu->ppr_log ? 0 : -ENOMEM;
699}
700
701static void iommu_enable_ppr_log(struct amd_iommu *iommu)
702{
703 u64 entry;
704
705 if (iommu->ppr_log == NULL)
706 return;
707
708 entry = iommu_virt_to_phys(iommu->ppr_log) | PPR_LOG_SIZE_512;
709
710 memcpy_toio(iommu->mmio_base + MMIO_PPR_LOG_OFFSET,
711 &entry, sizeof(entry));
712
713 /* set head and tail to zero manually */
714 writel(0x00, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
715 writel(0x00, iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
716
717 iommu_feature_enable(iommu, CONTROL_PPFLOG_EN);
718 iommu_feature_enable(iommu, CONTROL_PPR_EN);
719}
720
721static void __init free_ppr_log(struct amd_iommu *iommu)
722{
723 if (iommu->ppr_log == NULL)
724 return;
725
726 free_pages((unsigned long)iommu->ppr_log, get_order(PPR_LOG_SIZE));
727}
728
729static void free_ga_log(struct amd_iommu *iommu)
730{
731#ifdef CONFIG_IRQ_REMAP
732 if (iommu->ga_log)
733 free_pages((unsigned long)iommu->ga_log,
734 get_order(GA_LOG_SIZE));
735 if (iommu->ga_log_tail)
736 free_pages((unsigned long)iommu->ga_log_tail,
737 get_order(8));
738#endif
739}
740
741static int iommu_ga_log_enable(struct amd_iommu *iommu)
742{
743#ifdef CONFIG_IRQ_REMAP
744 u32 status, i;
745
746 if (!iommu->ga_log)
747 return -EINVAL;
748
749 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
750
751 /* Check if already running */
752 if (status & (MMIO_STATUS_GALOG_RUN_MASK))
753 return 0;
754
755 iommu_feature_enable(iommu, CONTROL_GAINT_EN);
756 iommu_feature_enable(iommu, CONTROL_GALOG_EN);
757
758 for (i = 0; i < LOOP_TIMEOUT; ++i) {
759 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
760 if (status & (MMIO_STATUS_GALOG_RUN_MASK))
761 break;
762 }
763
764 if (i >= LOOP_TIMEOUT)
765 return -EINVAL;
766#endif /* CONFIG_IRQ_REMAP */
767 return 0;
768}
769
770#ifdef CONFIG_IRQ_REMAP
771static int iommu_init_ga_log(struct amd_iommu *iommu)
772{
773 u64 entry;
774
775 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir))
776 return 0;
777
778 iommu->ga_log = (u8 *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
779 get_order(GA_LOG_SIZE));
780 if (!iommu->ga_log)
781 goto err_out;
782
783 iommu->ga_log_tail = (u8 *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
784 get_order(8));
785 if (!iommu->ga_log_tail)
786 goto err_out;
787
788 entry = iommu_virt_to_phys(iommu->ga_log) | GA_LOG_SIZE_512;
789 memcpy_toio(iommu->mmio_base + MMIO_GA_LOG_BASE_OFFSET,
790 &entry, sizeof(entry));
791 entry = (iommu_virt_to_phys(iommu->ga_log_tail) &
792 (BIT_ULL(52)-1)) & ~7ULL;
793 memcpy_toio(iommu->mmio_base + MMIO_GA_LOG_TAIL_OFFSET,
794 &entry, sizeof(entry));
795 writel(0x00, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
796 writel(0x00, iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
797
798 return 0;
799err_out:
800 free_ga_log(iommu);
801 return -EINVAL;
802}
803#endif /* CONFIG_IRQ_REMAP */
804
805static int iommu_init_ga(struct amd_iommu *iommu)
806{
807 int ret = 0;
808
809#ifdef CONFIG_IRQ_REMAP
810 /* Note: We have already checked GASup from IVRS table.
811 * Now, we need to make sure that GAMSup is set.
812 */
813 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
814 !iommu_feature(iommu, FEATURE_GAM_VAPIC))
815 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY_GA;
816
817 ret = iommu_init_ga_log(iommu);
818#endif /* CONFIG_IRQ_REMAP */
819
820 return ret;
821}
822
823static void iommu_enable_xt(struct amd_iommu *iommu)
824{
825#ifdef CONFIG_IRQ_REMAP
826 /*
827 * XT mode (32-bit APIC destination ID) requires
828 * GA mode (128-bit IRTE support) as a prerequisite.
829 */
830 if (AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir) &&
831 amd_iommu_xt_mode == IRQ_REMAP_X2APIC_MODE)
832 iommu_feature_enable(iommu, CONTROL_XT_EN);
833#endif /* CONFIG_IRQ_REMAP */
834}
835
836static void iommu_enable_gt(struct amd_iommu *iommu)
837{
838 if (!iommu_feature(iommu, FEATURE_GT))
839 return;
840
841 iommu_feature_enable(iommu, CONTROL_GT_EN);
842}
843
844/* sets a specific bit in the device table entry. */
845static void set_dev_entry_bit(u16 devid, u8 bit)
846{
847 int i = (bit >> 6) & 0x03;
848 int _bit = bit & 0x3f;
849
850 amd_iommu_dev_table[devid].data[i] |= (1UL << _bit);
851}
852
853static int get_dev_entry_bit(u16 devid, u8 bit)
854{
855 int i = (bit >> 6) & 0x03;
856 int _bit = bit & 0x3f;
857
858 return (amd_iommu_dev_table[devid].data[i] & (1UL << _bit)) >> _bit;
859}
860
861
862static bool copy_device_table(void)
863{
864 u64 int_ctl, int_tab_len, entry = 0, last_entry = 0;
865 struct dev_table_entry *old_devtb = NULL;
866 u32 lo, hi, devid, old_devtb_size;
867 phys_addr_t old_devtb_phys;
868 struct amd_iommu *iommu;
869 u16 dom_id, dte_v, irq_v;
870 gfp_t gfp_flag;
871 u64 tmp;
872
873 if (!amd_iommu_pre_enabled)
874 return false;
875
876 pr_warn("Translation is already enabled - trying to copy translation structures\n");
877 for_each_iommu(iommu) {
878 /* All IOMMUs should use the same device table with the same size */
879 lo = readl(iommu->mmio_base + MMIO_DEV_TABLE_OFFSET);
880 hi = readl(iommu->mmio_base + MMIO_DEV_TABLE_OFFSET + 4);
881 entry = (((u64) hi) << 32) + lo;
882 if (last_entry && last_entry != entry) {
883 pr_err("IOMMU:%d should use the same dev table as others!\n",
884 iommu->index);
885 return false;
886 }
887 last_entry = entry;
888
889 old_devtb_size = ((entry & ~PAGE_MASK) + 1) << 12;
890 if (old_devtb_size != dev_table_size) {
891 pr_err("The device table size of IOMMU:%d is not expected!\n",
892 iommu->index);
893 return false;
894 }
895 }
896
897 /*
898 * When SME is enabled in the first kernel, the entry includes the
899 * memory encryption mask(sme_me_mask), we must remove the memory
900 * encryption mask to obtain the true physical address in kdump kernel.
901 */
902 old_devtb_phys = __sme_clr(entry) & PAGE_MASK;
903
904 if (old_devtb_phys >= 0x100000000ULL) {
905 pr_err("The address of old device table is above 4G, not trustworthy!\n");
906 return false;
907 }
908 old_devtb = (sme_active() && is_kdump_kernel())
909 ? (__force void *)ioremap_encrypted(old_devtb_phys,
910 dev_table_size)
911 : memremap(old_devtb_phys, dev_table_size, MEMREMAP_WB);
912
913 if (!old_devtb)
914 return false;
915
916 gfp_flag = GFP_KERNEL | __GFP_ZERO | GFP_DMA32;
917 old_dev_tbl_cpy = (void *)__get_free_pages(gfp_flag,
918 get_order(dev_table_size));
919 if (old_dev_tbl_cpy == NULL) {
920 pr_err("Failed to allocate memory for copying old device table!\n");
921 return false;
922 }
923
924 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid) {
925 old_dev_tbl_cpy[devid] = old_devtb[devid];
926 dom_id = old_devtb[devid].data[1] & DEV_DOMID_MASK;
927 dte_v = old_devtb[devid].data[0] & DTE_FLAG_V;
928
929 if (dte_v && dom_id) {
930 old_dev_tbl_cpy[devid].data[0] = old_devtb[devid].data[0];
931 old_dev_tbl_cpy[devid].data[1] = old_devtb[devid].data[1];
932 __set_bit(dom_id, amd_iommu_pd_alloc_bitmap);
933 /* If gcr3 table existed, mask it out */
934 if (old_devtb[devid].data[0] & DTE_FLAG_GV) {
935 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
936 tmp |= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
937 old_dev_tbl_cpy[devid].data[1] &= ~tmp;
938 tmp = DTE_GCR3_VAL_A(~0ULL) << DTE_GCR3_SHIFT_A;
939 tmp |= DTE_FLAG_GV;
940 old_dev_tbl_cpy[devid].data[0] &= ~tmp;
941 }
942 }
943
944 irq_v = old_devtb[devid].data[2] & DTE_IRQ_REMAP_ENABLE;
945 int_ctl = old_devtb[devid].data[2] & DTE_IRQ_REMAP_INTCTL_MASK;
946 int_tab_len = old_devtb[devid].data[2] & DTE_IRQ_TABLE_LEN_MASK;
947 if (irq_v && (int_ctl || int_tab_len)) {
948 if ((int_ctl != DTE_IRQ_REMAP_INTCTL) ||
949 (int_tab_len != DTE_IRQ_TABLE_LEN)) {
950 pr_err("Wrong old irq remapping flag: %#x\n", devid);
951 return false;
952 }
953
954 old_dev_tbl_cpy[devid].data[2] = old_devtb[devid].data[2];
955 }
956 }
957 memunmap(old_devtb);
958
959 return true;
960}
961
962void amd_iommu_apply_erratum_63(u16 devid)
963{
964 int sysmgt;
965
966 sysmgt = get_dev_entry_bit(devid, DEV_ENTRY_SYSMGT1) |
967 (get_dev_entry_bit(devid, DEV_ENTRY_SYSMGT2) << 1);
968
969 if (sysmgt == 0x01)
970 set_dev_entry_bit(devid, DEV_ENTRY_IW);
971}
972
973/* Writes the specific IOMMU for a device into the rlookup table */
974static void __init set_iommu_for_device(struct amd_iommu *iommu, u16 devid)
975{
976 amd_iommu_rlookup_table[devid] = iommu;
977}
978
979/*
980 * This function takes the device specific flags read from the ACPI
981 * table and sets up the device table entry with that information
982 */
983static void __init set_dev_entry_from_acpi(struct amd_iommu *iommu,
984 u16 devid, u32 flags, u32 ext_flags)
985{
986 if (flags & ACPI_DEVFLAG_INITPASS)
987 set_dev_entry_bit(devid, DEV_ENTRY_INIT_PASS);
988 if (flags & ACPI_DEVFLAG_EXTINT)
989 set_dev_entry_bit(devid, DEV_ENTRY_EINT_PASS);
990 if (flags & ACPI_DEVFLAG_NMI)
991 set_dev_entry_bit(devid, DEV_ENTRY_NMI_PASS);
992 if (flags & ACPI_DEVFLAG_SYSMGT1)
993 set_dev_entry_bit(devid, DEV_ENTRY_SYSMGT1);
994 if (flags & ACPI_DEVFLAG_SYSMGT2)
995 set_dev_entry_bit(devid, DEV_ENTRY_SYSMGT2);
996 if (flags & ACPI_DEVFLAG_LINT0)
997 set_dev_entry_bit(devid, DEV_ENTRY_LINT0_PASS);
998 if (flags & ACPI_DEVFLAG_LINT1)
999 set_dev_entry_bit(devid, DEV_ENTRY_LINT1_PASS);
1000
1001 amd_iommu_apply_erratum_63(devid);
1002
1003 set_iommu_for_device(iommu, devid);
1004}
1005
1006int __init add_special_device(u8 type, u8 id, u16 *devid, bool cmd_line)
1007{
1008 struct devid_map *entry;
1009 struct list_head *list;
1010
1011 if (type == IVHD_SPECIAL_IOAPIC)
1012 list = &ioapic_map;
1013 else if (type == IVHD_SPECIAL_HPET)
1014 list = &hpet_map;
1015 else
1016 return -EINVAL;
1017
1018 list_for_each_entry(entry, list, list) {
1019 if (!(entry->id == id && entry->cmd_line))
1020 continue;
1021
1022 pr_info("Command-line override present for %s id %d - ignoring\n",
1023 type == IVHD_SPECIAL_IOAPIC ? "IOAPIC" : "HPET", id);
1024
1025 *devid = entry->devid;
1026
1027 return 0;
1028 }
1029
1030 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1031 if (!entry)
1032 return -ENOMEM;
1033
1034 entry->id = id;
1035 entry->devid = *devid;
1036 entry->cmd_line = cmd_line;
1037
1038 list_add_tail(&entry->list, list);
1039
1040 return 0;
1041}
1042
1043static int __init add_acpi_hid_device(u8 *hid, u8 *uid, u16 *devid,
1044 bool cmd_line)
1045{
1046 struct acpihid_map_entry *entry;
1047 struct list_head *list = &acpihid_map;
1048
1049 list_for_each_entry(entry, list, list) {
1050 if (strcmp(entry->hid, hid) ||
1051 (*uid && *entry->uid && strcmp(entry->uid, uid)) ||
1052 !entry->cmd_line)
1053 continue;
1054
1055 pr_info("Command-line override for hid:%s uid:%s\n",
1056 hid, uid);
1057 *devid = entry->devid;
1058 return 0;
1059 }
1060
1061 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
1062 if (!entry)
1063 return -ENOMEM;
1064
1065 memcpy(entry->uid, uid, strlen(uid));
1066 memcpy(entry->hid, hid, strlen(hid));
1067 entry->devid = *devid;
1068 entry->cmd_line = cmd_line;
1069 entry->root_devid = (entry->devid & (~0x7));
1070
1071 pr_info("%s, add hid:%s, uid:%s, rdevid:%d\n",
1072 entry->cmd_line ? "cmd" : "ivrs",
1073 entry->hid, entry->uid, entry->root_devid);
1074
1075 list_add_tail(&entry->list, list);
1076 return 0;
1077}
1078
1079static int __init add_early_maps(void)
1080{
1081 int i, ret;
1082
1083 for (i = 0; i < early_ioapic_map_size; ++i) {
1084 ret = add_special_device(IVHD_SPECIAL_IOAPIC,
1085 early_ioapic_map[i].id,
1086 &early_ioapic_map[i].devid,
1087 early_ioapic_map[i].cmd_line);
1088 if (ret)
1089 return ret;
1090 }
1091
1092 for (i = 0; i < early_hpet_map_size; ++i) {
1093 ret = add_special_device(IVHD_SPECIAL_HPET,
1094 early_hpet_map[i].id,
1095 &early_hpet_map[i].devid,
1096 early_hpet_map[i].cmd_line);
1097 if (ret)
1098 return ret;
1099 }
1100
1101 for (i = 0; i < early_acpihid_map_size; ++i) {
1102 ret = add_acpi_hid_device(early_acpihid_map[i].hid,
1103 early_acpihid_map[i].uid,
1104 &early_acpihid_map[i].devid,
1105 early_acpihid_map[i].cmd_line);
1106 if (ret)
1107 return ret;
1108 }
1109
1110 return 0;
1111}
1112
1113/*
1114 * Reads the device exclusion range from ACPI and initializes the IOMMU with
1115 * it
1116 */
1117static void __init set_device_exclusion_range(u16 devid, struct ivmd_header *m)
1118{
1119 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
1120
1121 if (!(m->flags & IVMD_FLAG_EXCL_RANGE))
1122 return;
1123
1124 if (iommu) {
1125 /*
1126 * We only can configure exclusion ranges per IOMMU, not
1127 * per device. But we can enable the exclusion range per
1128 * device. This is done here
1129 */
1130 set_dev_entry_bit(devid, DEV_ENTRY_EX);
1131 iommu->exclusion_start = m->range_start;
1132 iommu->exclusion_length = m->range_length;
1133 }
1134}
1135
1136/*
1137 * Takes a pointer to an AMD IOMMU entry in the ACPI table and
1138 * initializes the hardware and our data structures with it.
1139 */
1140static int __init init_iommu_from_acpi(struct amd_iommu *iommu,
1141 struct ivhd_header *h)
1142{
1143 u8 *p = (u8 *)h;
1144 u8 *end = p, flags = 0;
1145 u16 devid = 0, devid_start = 0, devid_to = 0;
1146 u32 dev_i, ext_flags = 0;
1147 bool alias = false;
1148 struct ivhd_entry *e;
1149 u32 ivhd_size;
1150 int ret;
1151
1152
1153 ret = add_early_maps();
1154 if (ret)
1155 return ret;
1156
1157 amd_iommu_apply_ivrs_quirks();
1158
1159 /*
1160 * First save the recommended feature enable bits from ACPI
1161 */
1162 iommu->acpi_flags = h->flags;
1163
1164 /*
1165 * Done. Now parse the device entries
1166 */
1167 ivhd_size = get_ivhd_header_size(h);
1168 if (!ivhd_size) {
1169 pr_err("Unsupported IVHD type %#x\n", h->type);
1170 return -EINVAL;
1171 }
1172
1173 p += ivhd_size;
1174
1175 end += h->length;
1176
1177
1178 while (p < end) {
1179 e = (struct ivhd_entry *)p;
1180 switch (e->type) {
1181 case IVHD_DEV_ALL:
1182
1183 DUMP_printk(" DEV_ALL\t\t\tflags: %02x\n", e->flags);
1184
1185 for (dev_i = 0; dev_i <= amd_iommu_last_bdf; ++dev_i)
1186 set_dev_entry_from_acpi(iommu, dev_i, e->flags, 0);
1187 break;
1188 case IVHD_DEV_SELECT:
1189
1190 DUMP_printk(" DEV_SELECT\t\t\t devid: %02x:%02x.%x "
1191 "flags: %02x\n",
1192 PCI_BUS_NUM(e->devid),
1193 PCI_SLOT(e->devid),
1194 PCI_FUNC(e->devid),
1195 e->flags);
1196
1197 devid = e->devid;
1198 set_dev_entry_from_acpi(iommu, devid, e->flags, 0);
1199 break;
1200 case IVHD_DEV_SELECT_RANGE_START:
1201
1202 DUMP_printk(" DEV_SELECT_RANGE_START\t "
1203 "devid: %02x:%02x.%x flags: %02x\n",
1204 PCI_BUS_NUM(e->devid),
1205 PCI_SLOT(e->devid),
1206 PCI_FUNC(e->devid),
1207 e->flags);
1208
1209 devid_start = e->devid;
1210 flags = e->flags;
1211 ext_flags = 0;
1212 alias = false;
1213 break;
1214 case IVHD_DEV_ALIAS:
1215
1216 DUMP_printk(" DEV_ALIAS\t\t\t devid: %02x:%02x.%x "
1217 "flags: %02x devid_to: %02x:%02x.%x\n",
1218 PCI_BUS_NUM(e->devid),
1219 PCI_SLOT(e->devid),
1220 PCI_FUNC(e->devid),
1221 e->flags,
1222 PCI_BUS_NUM(e->ext >> 8),
1223 PCI_SLOT(e->ext >> 8),
1224 PCI_FUNC(e->ext >> 8));
1225
1226 devid = e->devid;
1227 devid_to = e->ext >> 8;
1228 set_dev_entry_from_acpi(iommu, devid , e->flags, 0);
1229 set_dev_entry_from_acpi(iommu, devid_to, e->flags, 0);
1230 amd_iommu_alias_table[devid] = devid_to;
1231 break;
1232 case IVHD_DEV_ALIAS_RANGE:
1233
1234 DUMP_printk(" DEV_ALIAS_RANGE\t\t "
1235 "devid: %02x:%02x.%x flags: %02x "
1236 "devid_to: %02x:%02x.%x\n",
1237 PCI_BUS_NUM(e->devid),
1238 PCI_SLOT(e->devid),
1239 PCI_FUNC(e->devid),
1240 e->flags,
1241 PCI_BUS_NUM(e->ext >> 8),
1242 PCI_SLOT(e->ext >> 8),
1243 PCI_FUNC(e->ext >> 8));
1244
1245 devid_start = e->devid;
1246 flags = e->flags;
1247 devid_to = e->ext >> 8;
1248 ext_flags = 0;
1249 alias = true;
1250 break;
1251 case IVHD_DEV_EXT_SELECT:
1252
1253 DUMP_printk(" DEV_EXT_SELECT\t\t devid: %02x:%02x.%x "
1254 "flags: %02x ext: %08x\n",
1255 PCI_BUS_NUM(e->devid),
1256 PCI_SLOT(e->devid),
1257 PCI_FUNC(e->devid),
1258 e->flags, e->ext);
1259
1260 devid = e->devid;
1261 set_dev_entry_from_acpi(iommu, devid, e->flags,
1262 e->ext);
1263 break;
1264 case IVHD_DEV_EXT_SELECT_RANGE:
1265
1266 DUMP_printk(" DEV_EXT_SELECT_RANGE\t devid: "
1267 "%02x:%02x.%x flags: %02x ext: %08x\n",
1268 PCI_BUS_NUM(e->devid),
1269 PCI_SLOT(e->devid),
1270 PCI_FUNC(e->devid),
1271 e->flags, e->ext);
1272
1273 devid_start = e->devid;
1274 flags = e->flags;
1275 ext_flags = e->ext;
1276 alias = false;
1277 break;
1278 case IVHD_DEV_RANGE_END:
1279
1280 DUMP_printk(" DEV_RANGE_END\t\t devid: %02x:%02x.%x\n",
1281 PCI_BUS_NUM(e->devid),
1282 PCI_SLOT(e->devid),
1283 PCI_FUNC(e->devid));
1284
1285 devid = e->devid;
1286 for (dev_i = devid_start; dev_i <= devid; ++dev_i) {
1287 if (alias) {
1288 amd_iommu_alias_table[dev_i] = devid_to;
1289 set_dev_entry_from_acpi(iommu,
1290 devid_to, flags, ext_flags);
1291 }
1292 set_dev_entry_from_acpi(iommu, dev_i,
1293 flags, ext_flags);
1294 }
1295 break;
1296 case IVHD_DEV_SPECIAL: {
1297 u8 handle, type;
1298 const char *var;
1299 u16 devid;
1300 int ret;
1301
1302 handle = e->ext & 0xff;
1303 devid = (e->ext >> 8) & 0xffff;
1304 type = (e->ext >> 24) & 0xff;
1305
1306 if (type == IVHD_SPECIAL_IOAPIC)
1307 var = "IOAPIC";
1308 else if (type == IVHD_SPECIAL_HPET)
1309 var = "HPET";
1310 else
1311 var = "UNKNOWN";
1312
1313 DUMP_printk(" DEV_SPECIAL(%s[%d])\t\tdevid: %02x:%02x.%x\n",
1314 var, (int)handle,
1315 PCI_BUS_NUM(devid),
1316 PCI_SLOT(devid),
1317 PCI_FUNC(devid));
1318
1319 ret = add_special_device(type, handle, &devid, false);
1320 if (ret)
1321 return ret;
1322
1323 /*
1324 * add_special_device might update the devid in case a
1325 * command-line override is present. So call
1326 * set_dev_entry_from_acpi after add_special_device.
1327 */
1328 set_dev_entry_from_acpi(iommu, devid, e->flags, 0);
1329
1330 break;
1331 }
1332 case IVHD_DEV_ACPI_HID: {
1333 u16 devid;
1334 u8 hid[ACPIHID_HID_LEN] = {0};
1335 u8 uid[ACPIHID_UID_LEN] = {0};
1336 int ret;
1337
1338 if (h->type != 0x40) {
1339 pr_err(FW_BUG "Invalid IVHD device type %#x\n",
1340 e->type);
1341 break;
1342 }
1343
1344 memcpy(hid, (u8 *)(&e->ext), ACPIHID_HID_LEN - 1);
1345 hid[ACPIHID_HID_LEN - 1] = '\0';
1346
1347 if (!(*hid)) {
1348 pr_err(FW_BUG "Invalid HID.\n");
1349 break;
1350 }
1351
1352 switch (e->uidf) {
1353 case UID_NOT_PRESENT:
1354
1355 if (e->uidl != 0)
1356 pr_warn(FW_BUG "Invalid UID length.\n");
1357
1358 break;
1359 case UID_IS_INTEGER:
1360
1361 sprintf(uid, "%d", e->uid);
1362
1363 break;
1364 case UID_IS_CHARACTER:
1365
1366 memcpy(uid, (u8 *)(&e->uid), ACPIHID_UID_LEN - 1);
1367 uid[ACPIHID_UID_LEN - 1] = '\0';
1368
1369 break;
1370 default:
1371 break;
1372 }
1373
1374 devid = e->devid;
1375 DUMP_printk(" DEV_ACPI_HID(%s[%s])\t\tdevid: %02x:%02x.%x\n",
1376 hid, uid,
1377 PCI_BUS_NUM(devid),
1378 PCI_SLOT(devid),
1379 PCI_FUNC(devid));
1380
1381 flags = e->flags;
1382
1383 ret = add_acpi_hid_device(hid, uid, &devid, false);
1384 if (ret)
1385 return ret;
1386
1387 /*
1388 * add_special_device might update the devid in case a
1389 * command-line override is present. So call
1390 * set_dev_entry_from_acpi after add_special_device.
1391 */
1392 set_dev_entry_from_acpi(iommu, devid, e->flags, 0);
1393
1394 break;
1395 }
1396 default:
1397 break;
1398 }
1399
1400 p += ivhd_entry_length(p);
1401 }
1402
1403 return 0;
1404}
1405
1406static void __init free_iommu_one(struct amd_iommu *iommu)
1407{
1408 free_command_buffer(iommu);
1409 free_event_buffer(iommu);
1410 free_ppr_log(iommu);
1411 free_ga_log(iommu);
1412 iommu_unmap_mmio_space(iommu);
1413}
1414
1415static void __init free_iommu_all(void)
1416{
1417 struct amd_iommu *iommu, *next;
1418
1419 for_each_iommu_safe(iommu, next) {
1420 list_del(&iommu->list);
1421 free_iommu_one(iommu);
1422 kfree(iommu);
1423 }
1424}
1425
1426/*
1427 * Family15h Model 10h-1fh erratum 746 (IOMMU Logging May Stall Translations)
1428 * Workaround:
1429 * BIOS should disable L2B micellaneous clock gating by setting
1430 * L2_L2B_CK_GATE_CONTROL[CKGateL2BMiscDisable](D0F2xF4_x90[2]) = 1b
1431 */
1432static void amd_iommu_erratum_746_workaround(struct amd_iommu *iommu)
1433{
1434 u32 value;
1435
1436 if ((boot_cpu_data.x86 != 0x15) ||
1437 (boot_cpu_data.x86_model < 0x10) ||
1438 (boot_cpu_data.x86_model > 0x1f))
1439 return;
1440
1441 pci_write_config_dword(iommu->dev, 0xf0, 0x90);
1442 pci_read_config_dword(iommu->dev, 0xf4, &value);
1443
1444 if (value & BIT(2))
1445 return;
1446
1447 /* Select NB indirect register 0x90 and enable writing */
1448 pci_write_config_dword(iommu->dev, 0xf0, 0x90 | (1 << 8));
1449
1450 pci_write_config_dword(iommu->dev, 0xf4, value | 0x4);
1451 pci_info(iommu->dev, "Applying erratum 746 workaround\n");
1452
1453 /* Clear the enable writing bit */
1454 pci_write_config_dword(iommu->dev, 0xf0, 0x90);
1455}
1456
1457/*
1458 * Family15h Model 30h-3fh (IOMMU Mishandles ATS Write Permission)
1459 * Workaround:
1460 * BIOS should enable ATS write permission check by setting
1461 * L2_DEBUG_3[AtsIgnoreIWDis](D0F2xF4_x47[0]) = 1b
1462 */
1463static void amd_iommu_ats_write_check_workaround(struct amd_iommu *iommu)
1464{
1465 u32 value;
1466
1467 if ((boot_cpu_data.x86 != 0x15) ||
1468 (boot_cpu_data.x86_model < 0x30) ||
1469 (boot_cpu_data.x86_model > 0x3f))
1470 return;
1471
1472 /* Test L2_DEBUG_3[AtsIgnoreIWDis] == 1 */
1473 value = iommu_read_l2(iommu, 0x47);
1474
1475 if (value & BIT(0))
1476 return;
1477
1478 /* Set L2_DEBUG_3[AtsIgnoreIWDis] = 1 */
1479 iommu_write_l2(iommu, 0x47, value | BIT(0));
1480
1481 pci_info(iommu->dev, "Applying ATS write check workaround\n");
1482}
1483
1484/*
1485 * This function clues the initialization function for one IOMMU
1486 * together and also allocates the command buffer and programs the
1487 * hardware. It does NOT enable the IOMMU. This is done afterwards.
1488 */
1489static int __init init_iommu_one(struct amd_iommu *iommu, struct ivhd_header *h)
1490{
1491 int ret;
1492
1493 raw_spin_lock_init(&iommu->lock);
1494
1495 /* Add IOMMU to internal data structures */
1496 list_add_tail(&iommu->list, &amd_iommu_list);
1497 iommu->index = amd_iommus_present++;
1498
1499 if (unlikely(iommu->index >= MAX_IOMMUS)) {
1500 WARN(1, "System has more IOMMUs than supported by this driver\n");
1501 return -ENOSYS;
1502 }
1503
1504 /* Index is fine - add IOMMU to the array */
1505 amd_iommus[iommu->index] = iommu;
1506
1507 /*
1508 * Copy data from ACPI table entry to the iommu struct
1509 */
1510 iommu->devid = h->devid;
1511 iommu->cap_ptr = h->cap_ptr;
1512 iommu->pci_seg = h->pci_seg;
1513 iommu->mmio_phys = h->mmio_phys;
1514
1515 switch (h->type) {
1516 case 0x10:
1517 /* Check if IVHD EFR contains proper max banks/counters */
1518 if ((h->efr_attr != 0) &&
1519 ((h->efr_attr & (0xF << 13)) != 0) &&
1520 ((h->efr_attr & (0x3F << 17)) != 0))
1521 iommu->mmio_phys_end = MMIO_REG_END_OFFSET;
1522 else
1523 iommu->mmio_phys_end = MMIO_CNTR_CONF_OFFSET;
1524 if (((h->efr_attr & (0x1 << IOMMU_FEAT_GASUP_SHIFT)) == 0))
1525 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY;
1526 if (((h->efr_attr & (0x1 << IOMMU_FEAT_XTSUP_SHIFT)) == 0))
1527 amd_iommu_xt_mode = IRQ_REMAP_XAPIC_MODE;
1528 break;
1529 case 0x11:
1530 case 0x40:
1531 if (h->efr_reg & (1 << 9))
1532 iommu->mmio_phys_end = MMIO_REG_END_OFFSET;
1533 else
1534 iommu->mmio_phys_end = MMIO_CNTR_CONF_OFFSET;
1535 if (((h->efr_reg & (0x1 << IOMMU_EFR_GASUP_SHIFT)) == 0))
1536 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY;
1537 if (((h->efr_reg & (0x1 << IOMMU_EFR_XTSUP_SHIFT)) == 0))
1538 amd_iommu_xt_mode = IRQ_REMAP_XAPIC_MODE;
1539 break;
1540 default:
1541 return -EINVAL;
1542 }
1543
1544 iommu->mmio_base = iommu_map_mmio_space(iommu->mmio_phys,
1545 iommu->mmio_phys_end);
1546 if (!iommu->mmio_base)
1547 return -ENOMEM;
1548
1549 if (alloc_command_buffer(iommu))
1550 return -ENOMEM;
1551
1552 if (alloc_event_buffer(iommu))
1553 return -ENOMEM;
1554
1555 iommu->int_enabled = false;
1556
1557 init_translation_status(iommu);
1558 if (translation_pre_enabled(iommu) && !is_kdump_kernel()) {
1559 iommu_disable(iommu);
1560 clear_translation_pre_enabled(iommu);
1561 pr_warn("Translation was enabled for IOMMU:%d but we are not in kdump mode\n",
1562 iommu->index);
1563 }
1564 if (amd_iommu_pre_enabled)
1565 amd_iommu_pre_enabled = translation_pre_enabled(iommu);
1566
1567 ret = init_iommu_from_acpi(iommu, h);
1568 if (ret)
1569 return ret;
1570
1571 ret = amd_iommu_create_irq_domain(iommu);
1572 if (ret)
1573 return ret;
1574
1575 /*
1576 * Make sure IOMMU is not considered to translate itself. The IVRS
1577 * table tells us so, but this is a lie!
1578 */
1579 amd_iommu_rlookup_table[iommu->devid] = NULL;
1580
1581 return 0;
1582}
1583
1584/**
1585 * get_highest_supported_ivhd_type - Look up the appropriate IVHD type
1586 * @ivrs Pointer to the IVRS header
1587 *
1588 * This function search through all IVDB of the maximum supported IVHD
1589 */
1590static u8 get_highest_supported_ivhd_type(struct acpi_table_header *ivrs)
1591{
1592 u8 *base = (u8 *)ivrs;
1593 struct ivhd_header *ivhd = (struct ivhd_header *)
1594 (base + IVRS_HEADER_LENGTH);
1595 u8 last_type = ivhd->type;
1596 u16 devid = ivhd->devid;
1597
1598 while (((u8 *)ivhd - base < ivrs->length) &&
1599 (ivhd->type <= ACPI_IVHD_TYPE_MAX_SUPPORTED)) {
1600 u8 *p = (u8 *) ivhd;
1601
1602 if (ivhd->devid == devid)
1603 last_type = ivhd->type;
1604 ivhd = (struct ivhd_header *)(p + ivhd->length);
1605 }
1606
1607 return last_type;
1608}
1609
1610/*
1611 * Iterates over all IOMMU entries in the ACPI table, allocates the
1612 * IOMMU structure and initializes it with init_iommu_one()
1613 */
1614static int __init init_iommu_all(struct acpi_table_header *table)
1615{
1616 u8 *p = (u8 *)table, *end = (u8 *)table;
1617 struct ivhd_header *h;
1618 struct amd_iommu *iommu;
1619 int ret;
1620
1621 end += table->length;
1622 p += IVRS_HEADER_LENGTH;
1623
1624 while (p < end) {
1625 h = (struct ivhd_header *)p;
1626 if (*p == amd_iommu_target_ivhd_type) {
1627
1628 DUMP_printk("device: %02x:%02x.%01x cap: %04x "
1629 "seg: %d flags: %01x info %04x\n",
1630 PCI_BUS_NUM(h->devid), PCI_SLOT(h->devid),
1631 PCI_FUNC(h->devid), h->cap_ptr,
1632 h->pci_seg, h->flags, h->info);
1633 DUMP_printk(" mmio-addr: %016llx\n",
1634 h->mmio_phys);
1635
1636 iommu = kzalloc(sizeof(struct amd_iommu), GFP_KERNEL);
1637 if (iommu == NULL)
1638 return -ENOMEM;
1639
1640 ret = init_iommu_one(iommu, h);
1641 if (ret)
1642 return ret;
1643 }
1644 p += h->length;
1645
1646 }
1647 WARN_ON(p != end);
1648
1649 return 0;
1650}
1651
1652static int iommu_pc_get_set_reg(struct amd_iommu *iommu, u8 bank, u8 cntr,
1653 u8 fxn, u64 *value, bool is_write);
1654
1655static void init_iommu_perf_ctr(struct amd_iommu *iommu)
1656{
1657 struct pci_dev *pdev = iommu->dev;
1658 u64 val = 0xabcd, val2 = 0;
1659
1660 if (!iommu_feature(iommu, FEATURE_PC))
1661 return;
1662
1663 amd_iommu_pc_present = true;
1664
1665 /* Check if the performance counters can be written to */
1666 if ((iommu_pc_get_set_reg(iommu, 0, 0, 0, &val, true)) ||
1667 (iommu_pc_get_set_reg(iommu, 0, 0, 0, &val2, false)) ||
1668 (val != val2)) {
1669 pci_err(pdev, "Unable to write to IOMMU perf counter.\n");
1670 amd_iommu_pc_present = false;
1671 return;
1672 }
1673
1674 pci_info(pdev, "IOMMU performance counters supported\n");
1675
1676 val = readl(iommu->mmio_base + MMIO_CNTR_CONF_OFFSET);
1677 iommu->max_banks = (u8) ((val >> 12) & 0x3f);
1678 iommu->max_counters = (u8) ((val >> 7) & 0xf);
1679}
1680
1681static ssize_t amd_iommu_show_cap(struct device *dev,
1682 struct device_attribute *attr,
1683 char *buf)
1684{
1685 struct amd_iommu *iommu = dev_to_amd_iommu(dev);
1686 return sprintf(buf, "%x\n", iommu->cap);
1687}
1688static DEVICE_ATTR(cap, S_IRUGO, amd_iommu_show_cap, NULL);
1689
1690static ssize_t amd_iommu_show_features(struct device *dev,
1691 struct device_attribute *attr,
1692 char *buf)
1693{
1694 struct amd_iommu *iommu = dev_to_amd_iommu(dev);
1695 return sprintf(buf, "%llx\n", iommu->features);
1696}
1697static DEVICE_ATTR(features, S_IRUGO, amd_iommu_show_features, NULL);
1698
1699static struct attribute *amd_iommu_attrs[] = {
1700 &dev_attr_cap.attr,
1701 &dev_attr_features.attr,
1702 NULL,
1703};
1704
1705static struct attribute_group amd_iommu_group = {
1706 .name = "amd-iommu",
1707 .attrs = amd_iommu_attrs,
1708};
1709
1710static const struct attribute_group *amd_iommu_groups[] = {
1711 &amd_iommu_group,
1712 NULL,
1713};
1714
1715static int __init iommu_init_pci(struct amd_iommu *iommu)
1716{
1717 int cap_ptr = iommu->cap_ptr;
1718 u32 range, misc, low, high;
1719 int ret;
1720
1721 iommu->dev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(iommu->devid),
1722 iommu->devid & 0xff);
1723 if (!iommu->dev)
1724 return -ENODEV;
1725
1726 /* Prevent binding other PCI device drivers to IOMMU devices */
1727 iommu->dev->match_driver = false;
1728
1729 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_CAP_HDR_OFFSET,
1730 &iommu->cap);
1731 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_RANGE_OFFSET,
1732 &range);
1733 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_MISC_OFFSET,
1734 &misc);
1735
1736 if (!(iommu->cap & (1 << IOMMU_CAP_IOTLB)))
1737 amd_iommu_iotlb_sup = false;
1738
1739 /* read extended feature bits */
1740 low = readl(iommu->mmio_base + MMIO_EXT_FEATURES);
1741 high = readl(iommu->mmio_base + MMIO_EXT_FEATURES + 4);
1742
1743 iommu->features = ((u64)high << 32) | low;
1744
1745 if (iommu_feature(iommu, FEATURE_GT)) {
1746 int glxval;
1747 u32 max_pasid;
1748 u64 pasmax;
1749
1750 pasmax = iommu->features & FEATURE_PASID_MASK;
1751 pasmax >>= FEATURE_PASID_SHIFT;
1752 max_pasid = (1 << (pasmax + 1)) - 1;
1753
1754 amd_iommu_max_pasid = min(amd_iommu_max_pasid, max_pasid);
1755
1756 BUG_ON(amd_iommu_max_pasid & ~PASID_MASK);
1757
1758 glxval = iommu->features & FEATURE_GLXVAL_MASK;
1759 glxval >>= FEATURE_GLXVAL_SHIFT;
1760
1761 if (amd_iommu_max_glx_val == -1)
1762 amd_iommu_max_glx_val = glxval;
1763 else
1764 amd_iommu_max_glx_val = min(amd_iommu_max_glx_val, glxval);
1765 }
1766
1767 if (iommu_feature(iommu, FEATURE_GT) &&
1768 iommu_feature(iommu, FEATURE_PPR)) {
1769 iommu->is_iommu_v2 = true;
1770 amd_iommu_v2_present = true;
1771 }
1772
1773 if (iommu_feature(iommu, FEATURE_PPR) && alloc_ppr_log(iommu))
1774 return -ENOMEM;
1775
1776 ret = iommu_init_ga(iommu);
1777 if (ret)
1778 return ret;
1779
1780 if (iommu->cap & (1UL << IOMMU_CAP_NPCACHE))
1781 amd_iommu_np_cache = true;
1782
1783 init_iommu_perf_ctr(iommu);
1784
1785 if (is_rd890_iommu(iommu->dev)) {
1786 int i, j;
1787
1788 iommu->root_pdev =
1789 pci_get_domain_bus_and_slot(0, iommu->dev->bus->number,
1790 PCI_DEVFN(0, 0));
1791
1792 /*
1793 * Some rd890 systems may not be fully reconfigured by the
1794 * BIOS, so it's necessary for us to store this information so
1795 * it can be reprogrammed on resume
1796 */
1797 pci_read_config_dword(iommu->dev, iommu->cap_ptr + 4,
1798 &iommu->stored_addr_lo);
1799 pci_read_config_dword(iommu->dev, iommu->cap_ptr + 8,
1800 &iommu->stored_addr_hi);
1801
1802 /* Low bit locks writes to configuration space */
1803 iommu->stored_addr_lo &= ~1;
1804
1805 for (i = 0; i < 6; i++)
1806 for (j = 0; j < 0x12; j++)
1807 iommu->stored_l1[i][j] = iommu_read_l1(iommu, i, j);
1808
1809 for (i = 0; i < 0x83; i++)
1810 iommu->stored_l2[i] = iommu_read_l2(iommu, i);
1811 }
1812
1813 amd_iommu_erratum_746_workaround(iommu);
1814 amd_iommu_ats_write_check_workaround(iommu);
1815
1816 iommu_device_sysfs_add(&iommu->iommu, &iommu->dev->dev,
1817 amd_iommu_groups, "ivhd%d", iommu->index);
1818 iommu_device_set_ops(&iommu->iommu, &amd_iommu_ops);
1819 iommu_device_register(&iommu->iommu);
1820
1821 return pci_enable_device(iommu->dev);
1822}
1823
1824static void print_iommu_info(void)
1825{
1826 static const char * const feat_str[] = {
1827 "PreF", "PPR", "X2APIC", "NX", "GT", "[5]",
1828 "IA", "GA", "HE", "PC"
1829 };
1830 struct amd_iommu *iommu;
1831
1832 for_each_iommu(iommu) {
1833 struct pci_dev *pdev = iommu->dev;
1834 int i;
1835
1836 pci_info(pdev, "Found IOMMU cap 0x%hx\n", iommu->cap_ptr);
1837
1838 if (iommu->cap & (1 << IOMMU_CAP_EFR)) {
1839 pci_info(pdev, "Extended features (%#llx):\n",
1840 iommu->features);
1841 for (i = 0; i < ARRAY_SIZE(feat_str); ++i) {
1842 if (iommu_feature(iommu, (1ULL << i)))
1843 pr_cont(" %s", feat_str[i]);
1844 }
1845
1846 if (iommu->features & FEATURE_GAM_VAPIC)
1847 pr_cont(" GA_vAPIC");
1848
1849 pr_cont("\n");
1850 }
1851 }
1852 if (irq_remapping_enabled) {
1853 pr_info("Interrupt remapping enabled\n");
1854 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir))
1855 pr_info("Virtual APIC enabled\n");
1856 if (amd_iommu_xt_mode == IRQ_REMAP_X2APIC_MODE)
1857 pr_info("X2APIC enabled\n");
1858 }
1859}
1860
1861static int __init amd_iommu_init_pci(void)
1862{
1863 struct amd_iommu *iommu;
1864 int ret = 0;
1865
1866 for_each_iommu(iommu) {
1867 ret = iommu_init_pci(iommu);
1868 if (ret)
1869 break;
1870 }
1871
1872 /*
1873 * Order is important here to make sure any unity map requirements are
1874 * fulfilled. The unity mappings are created and written to the device
1875 * table during the amd_iommu_init_api() call.
1876 *
1877 * After that we call init_device_table_dma() to make sure any
1878 * uninitialized DTE will block DMA, and in the end we flush the caches
1879 * of all IOMMUs to make sure the changes to the device table are
1880 * active.
1881 */
1882 ret = amd_iommu_init_api();
1883
1884 init_device_table_dma();
1885
1886 for_each_iommu(iommu)
1887 iommu_flush_all_caches(iommu);
1888
1889 if (!ret)
1890 print_iommu_info();
1891
1892 return ret;
1893}
1894
1895/****************************************************************************
1896 *
1897 * The following functions initialize the MSI interrupts for all IOMMUs
1898 * in the system. It's a bit challenging because there could be multiple
1899 * IOMMUs per PCI BDF but we can call pci_enable_msi(x) only once per
1900 * pci_dev.
1901 *
1902 ****************************************************************************/
1903
1904static int iommu_setup_msi(struct amd_iommu *iommu)
1905{
1906 int r;
1907
1908 r = pci_enable_msi(iommu->dev);
1909 if (r)
1910 return r;
1911
1912 r = request_threaded_irq(iommu->dev->irq,
1913 amd_iommu_int_handler,
1914 amd_iommu_int_thread,
1915 0, "AMD-Vi",
1916 iommu);
1917
1918 if (r) {
1919 pci_disable_msi(iommu->dev);
1920 return r;
1921 }
1922
1923 iommu->int_enabled = true;
1924
1925 return 0;
1926}
1927
1928#define XT_INT_DEST_MODE(x) (((x) & 0x1ULL) << 2)
1929#define XT_INT_DEST_LO(x) (((x) & 0xFFFFFFULL) << 8)
1930#define XT_INT_VEC(x) (((x) & 0xFFULL) << 32)
1931#define XT_INT_DEST_HI(x) ((((x) >> 24) & 0xFFULL) << 56)
1932
1933/**
1934 * Setup the IntCapXT registers with interrupt routing information
1935 * based on the PCI MSI capability block registers, accessed via
1936 * MMIO MSI address low/hi and MSI data registers.
1937 */
1938static void iommu_update_intcapxt(struct amd_iommu *iommu)
1939{
1940 u64 val;
1941 u32 addr_lo = readl(iommu->mmio_base + MMIO_MSI_ADDR_LO_OFFSET);
1942 u32 addr_hi = readl(iommu->mmio_base + MMIO_MSI_ADDR_HI_OFFSET);
1943 u32 data = readl(iommu->mmio_base + MMIO_MSI_DATA_OFFSET);
1944 bool dm = (addr_lo >> MSI_ADDR_DEST_MODE_SHIFT) & 0x1;
1945 u32 dest = ((addr_lo >> MSI_ADDR_DEST_ID_SHIFT) & 0xFF);
1946
1947 if (x2apic_enabled())
1948 dest |= MSI_ADDR_EXT_DEST_ID(addr_hi);
1949
1950 val = XT_INT_VEC(data & 0xFF) |
1951 XT_INT_DEST_MODE(dm) |
1952 XT_INT_DEST_LO(dest) |
1953 XT_INT_DEST_HI(dest);
1954
1955 /**
1956 * Current IOMMU implemtation uses the same IRQ for all
1957 * 3 IOMMU interrupts.
1958 */
1959 writeq(val, iommu->mmio_base + MMIO_INTCAPXT_EVT_OFFSET);
1960 writeq(val, iommu->mmio_base + MMIO_INTCAPXT_PPR_OFFSET);
1961 writeq(val, iommu->mmio_base + MMIO_INTCAPXT_GALOG_OFFSET);
1962}
1963
1964static void _irq_notifier_notify(struct irq_affinity_notify *notify,
1965 const cpumask_t *mask)
1966{
1967 struct amd_iommu *iommu;
1968
1969 for_each_iommu(iommu) {
1970 if (iommu->dev->irq == notify->irq) {
1971 iommu_update_intcapxt(iommu);
1972 break;
1973 }
1974 }
1975}
1976
1977static void _irq_notifier_release(struct kref *ref)
1978{
1979}
1980
1981static int iommu_init_intcapxt(struct amd_iommu *iommu)
1982{
1983 int ret;
1984 struct irq_affinity_notify *notify = &iommu->intcapxt_notify;
1985
1986 /**
1987 * IntCapXT requires XTSup=1, which can be inferred
1988 * amd_iommu_xt_mode.
1989 */
1990 if (amd_iommu_xt_mode != IRQ_REMAP_X2APIC_MODE)
1991 return 0;
1992
1993 /**
1994 * Also, we need to setup notifier to update the IntCapXT registers
1995 * whenever the irq affinity is changed from user-space.
1996 */
1997 notify->irq = iommu->dev->irq;
1998 notify->notify = _irq_notifier_notify,
1999 notify->release = _irq_notifier_release,
2000 ret = irq_set_affinity_notifier(iommu->dev->irq, notify);
2001 if (ret) {
2002 pr_err("Failed to register irq affinity notifier (devid=%#x, irq %d)\n",
2003 iommu->devid, iommu->dev->irq);
2004 return ret;
2005 }
2006
2007 iommu_update_intcapxt(iommu);
2008 iommu_feature_enable(iommu, CONTROL_INTCAPXT_EN);
2009 return ret;
2010}
2011
2012static int iommu_init_msi(struct amd_iommu *iommu)
2013{
2014 int ret;
2015
2016 if (iommu->int_enabled)
2017 goto enable_faults;
2018
2019 if (iommu->dev->msi_cap)
2020 ret = iommu_setup_msi(iommu);
2021 else
2022 ret = -ENODEV;
2023
2024 if (ret)
2025 return ret;
2026
2027enable_faults:
2028 ret = iommu_init_intcapxt(iommu);
2029 if (ret)
2030 return ret;
2031
2032 iommu_feature_enable(iommu, CONTROL_EVT_INT_EN);
2033
2034 if (iommu->ppr_log != NULL)
2035 iommu_feature_enable(iommu, CONTROL_PPFINT_EN);
2036
2037 iommu_ga_log_enable(iommu);
2038
2039 return 0;
2040}
2041
2042/****************************************************************************
2043 *
2044 * The next functions belong to the third pass of parsing the ACPI
2045 * table. In this last pass the memory mapping requirements are
2046 * gathered (like exclusion and unity mapping ranges).
2047 *
2048 ****************************************************************************/
2049
2050static void __init free_unity_maps(void)
2051{
2052 struct unity_map_entry *entry, *next;
2053
2054 list_for_each_entry_safe(entry, next, &amd_iommu_unity_map, list) {
2055 list_del(&entry->list);
2056 kfree(entry);
2057 }
2058}
2059
2060/* called when we find an exclusion range definition in ACPI */
2061static int __init init_exclusion_range(struct ivmd_header *m)
2062{
2063 int i;
2064
2065 switch (m->type) {
2066 case ACPI_IVMD_TYPE:
2067 set_device_exclusion_range(m->devid, m);
2068 break;
2069 case ACPI_IVMD_TYPE_ALL:
2070 for (i = 0; i <= amd_iommu_last_bdf; ++i)
2071 set_device_exclusion_range(i, m);
2072 break;
2073 case ACPI_IVMD_TYPE_RANGE:
2074 for (i = m->devid; i <= m->aux; ++i)
2075 set_device_exclusion_range(i, m);
2076 break;
2077 default:
2078 break;
2079 }
2080
2081 return 0;
2082}
2083
2084/* called for unity map ACPI definition */
2085static int __init init_unity_map_range(struct ivmd_header *m)
2086{
2087 struct unity_map_entry *e = NULL;
2088 char *s;
2089
2090 e = kzalloc(sizeof(*e), GFP_KERNEL);
2091 if (e == NULL)
2092 return -ENOMEM;
2093
2094 if (m->flags & IVMD_FLAG_EXCL_RANGE)
2095 init_exclusion_range(m);
2096
2097 switch (m->type) {
2098 default:
2099 kfree(e);
2100 return 0;
2101 case ACPI_IVMD_TYPE:
2102 s = "IVMD_TYPEi\t\t\t";
2103 e->devid_start = e->devid_end = m->devid;
2104 break;
2105 case ACPI_IVMD_TYPE_ALL:
2106 s = "IVMD_TYPE_ALL\t\t";
2107 e->devid_start = 0;
2108 e->devid_end = amd_iommu_last_bdf;
2109 break;
2110 case ACPI_IVMD_TYPE_RANGE:
2111 s = "IVMD_TYPE_RANGE\t\t";
2112 e->devid_start = m->devid;
2113 e->devid_end = m->aux;
2114 break;
2115 }
2116 e->address_start = PAGE_ALIGN(m->range_start);
2117 e->address_end = e->address_start + PAGE_ALIGN(m->range_length);
2118 e->prot = m->flags >> 1;
2119
2120 DUMP_printk("%s devid_start: %02x:%02x.%x devid_end: %02x:%02x.%x"
2121 " range_start: %016llx range_end: %016llx flags: %x\n", s,
2122 PCI_BUS_NUM(e->devid_start), PCI_SLOT(e->devid_start),
2123 PCI_FUNC(e->devid_start), PCI_BUS_NUM(e->devid_end),
2124 PCI_SLOT(e->devid_end), PCI_FUNC(e->devid_end),
2125 e->address_start, e->address_end, m->flags);
2126
2127 list_add_tail(&e->list, &amd_iommu_unity_map);
2128
2129 return 0;
2130}
2131
2132/* iterates over all memory definitions we find in the ACPI table */
2133static int __init init_memory_definitions(struct acpi_table_header *table)
2134{
2135 u8 *p = (u8 *)table, *end = (u8 *)table;
2136 struct ivmd_header *m;
2137
2138 end += table->length;
2139 p += IVRS_HEADER_LENGTH;
2140
2141 while (p < end) {
2142 m = (struct ivmd_header *)p;
2143 if (m->flags & (IVMD_FLAG_UNITY_MAP | IVMD_FLAG_EXCL_RANGE))
2144 init_unity_map_range(m);
2145
2146 p += m->length;
2147 }
2148
2149 return 0;
2150}
2151
2152/*
2153 * Init the device table to not allow DMA access for devices
2154 */
2155static void init_device_table_dma(void)
2156{
2157 u32 devid;
2158
2159 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid) {
2160 set_dev_entry_bit(devid, DEV_ENTRY_VALID);
2161 set_dev_entry_bit(devid, DEV_ENTRY_TRANSLATION);
2162 }
2163}
2164
2165static void __init uninit_device_table_dma(void)
2166{
2167 u32 devid;
2168
2169 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid) {
2170 amd_iommu_dev_table[devid].data[0] = 0ULL;
2171 amd_iommu_dev_table[devid].data[1] = 0ULL;
2172 }
2173}
2174
2175static void init_device_table(void)
2176{
2177 u32 devid;
2178
2179 if (!amd_iommu_irq_remap)
2180 return;
2181
2182 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid)
2183 set_dev_entry_bit(devid, DEV_ENTRY_IRQ_TBL_EN);
2184}
2185
2186static void iommu_init_flags(struct amd_iommu *iommu)
2187{
2188 iommu->acpi_flags & IVHD_FLAG_HT_TUN_EN_MASK ?
2189 iommu_feature_enable(iommu, CONTROL_HT_TUN_EN) :
2190 iommu_feature_disable(iommu, CONTROL_HT_TUN_EN);
2191
2192 iommu->acpi_flags & IVHD_FLAG_PASSPW_EN_MASK ?
2193 iommu_feature_enable(iommu, CONTROL_PASSPW_EN) :
2194 iommu_feature_disable(iommu, CONTROL_PASSPW_EN);
2195
2196 iommu->acpi_flags & IVHD_FLAG_RESPASSPW_EN_MASK ?
2197 iommu_feature_enable(iommu, CONTROL_RESPASSPW_EN) :
2198 iommu_feature_disable(iommu, CONTROL_RESPASSPW_EN);
2199
2200 iommu->acpi_flags & IVHD_FLAG_ISOC_EN_MASK ?
2201 iommu_feature_enable(iommu, CONTROL_ISOC_EN) :
2202 iommu_feature_disable(iommu, CONTROL_ISOC_EN);
2203
2204 /*
2205 * make IOMMU memory accesses cache coherent
2206 */
2207 iommu_feature_enable(iommu, CONTROL_COHERENT_EN);
2208
2209 /* Set IOTLB invalidation timeout to 1s */
2210 iommu_set_inv_tlb_timeout(iommu, CTRL_INV_TO_1S);
2211}
2212
2213static void iommu_apply_resume_quirks(struct amd_iommu *iommu)
2214{
2215 int i, j;
2216 u32 ioc_feature_control;
2217 struct pci_dev *pdev = iommu->root_pdev;
2218
2219 /* RD890 BIOSes may not have completely reconfigured the iommu */
2220 if (!is_rd890_iommu(iommu->dev) || !pdev)
2221 return;
2222
2223 /*
2224 * First, we need to ensure that the iommu is enabled. This is
2225 * controlled by a register in the northbridge
2226 */
2227
2228 /* Select Northbridge indirect register 0x75 and enable writing */
2229 pci_write_config_dword(pdev, 0x60, 0x75 | (1 << 7));
2230 pci_read_config_dword(pdev, 0x64, &ioc_feature_control);
2231
2232 /* Enable the iommu */
2233 if (!(ioc_feature_control & 0x1))
2234 pci_write_config_dword(pdev, 0x64, ioc_feature_control | 1);
2235
2236 /* Restore the iommu BAR */
2237 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 4,
2238 iommu->stored_addr_lo);
2239 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 8,
2240 iommu->stored_addr_hi);
2241
2242 /* Restore the l1 indirect regs for each of the 6 l1s */
2243 for (i = 0; i < 6; i++)
2244 for (j = 0; j < 0x12; j++)
2245 iommu_write_l1(iommu, i, j, iommu->stored_l1[i][j]);
2246
2247 /* Restore the l2 indirect regs */
2248 for (i = 0; i < 0x83; i++)
2249 iommu_write_l2(iommu, i, iommu->stored_l2[i]);
2250
2251 /* Lock PCI setup registers */
2252 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 4,
2253 iommu->stored_addr_lo | 1);
2254}
2255
2256static void iommu_enable_ga(struct amd_iommu *iommu)
2257{
2258#ifdef CONFIG_IRQ_REMAP
2259 switch (amd_iommu_guest_ir) {
2260 case AMD_IOMMU_GUEST_IR_VAPIC:
2261 iommu_feature_enable(iommu, CONTROL_GAM_EN);
2262 /* Fall through */
2263 case AMD_IOMMU_GUEST_IR_LEGACY_GA:
2264 iommu_feature_enable(iommu, CONTROL_GA_EN);
2265 iommu->irte_ops = &irte_128_ops;
2266 break;
2267 default:
2268 iommu->irte_ops = &irte_32_ops;
2269 break;
2270 }
2271#endif
2272}
2273
2274static void early_enable_iommu(struct amd_iommu *iommu)
2275{
2276 iommu_disable(iommu);
2277 iommu_init_flags(iommu);
2278 iommu_set_device_table(iommu);
2279 iommu_enable_command_buffer(iommu);
2280 iommu_enable_event_buffer(iommu);
2281 iommu_set_exclusion_range(iommu);
2282 iommu_enable_ga(iommu);
2283 iommu_enable_xt(iommu);
2284 iommu_enable(iommu);
2285 iommu_flush_all_caches(iommu);
2286}
2287
2288/*
2289 * This function finally enables all IOMMUs found in the system after
2290 * they have been initialized.
2291 *
2292 * Or if in kdump kernel and IOMMUs are all pre-enabled, try to copy
2293 * the old content of device table entries. Not this case or copy failed,
2294 * just continue as normal kernel does.
2295 */
2296static void early_enable_iommus(void)
2297{
2298 struct amd_iommu *iommu;
2299
2300
2301 if (!copy_device_table()) {
2302 /*
2303 * If come here because of failure in copying device table from old
2304 * kernel with all IOMMUs enabled, print error message and try to
2305 * free allocated old_dev_tbl_cpy.
2306 */
2307 if (amd_iommu_pre_enabled)
2308 pr_err("Failed to copy DEV table from previous kernel.\n");
2309 if (old_dev_tbl_cpy != NULL)
2310 free_pages((unsigned long)old_dev_tbl_cpy,
2311 get_order(dev_table_size));
2312
2313 for_each_iommu(iommu) {
2314 clear_translation_pre_enabled(iommu);
2315 early_enable_iommu(iommu);
2316 }
2317 } else {
2318 pr_info("Copied DEV table from previous kernel.\n");
2319 free_pages((unsigned long)amd_iommu_dev_table,
2320 get_order(dev_table_size));
2321 amd_iommu_dev_table = old_dev_tbl_cpy;
2322 for_each_iommu(iommu) {
2323 iommu_disable_command_buffer(iommu);
2324 iommu_disable_event_buffer(iommu);
2325 iommu_enable_command_buffer(iommu);
2326 iommu_enable_event_buffer(iommu);
2327 iommu_enable_ga(iommu);
2328 iommu_enable_xt(iommu);
2329 iommu_set_device_table(iommu);
2330 iommu_flush_all_caches(iommu);
2331 }
2332 }
2333
2334#ifdef CONFIG_IRQ_REMAP
2335 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir))
2336 amd_iommu_irq_ops.capability |= (1 << IRQ_POSTING_CAP);
2337#endif
2338}
2339
2340static void enable_iommus_v2(void)
2341{
2342 struct amd_iommu *iommu;
2343
2344 for_each_iommu(iommu) {
2345 iommu_enable_ppr_log(iommu);
2346 iommu_enable_gt(iommu);
2347 }
2348}
2349
2350static void enable_iommus(void)
2351{
2352 early_enable_iommus();
2353
2354 enable_iommus_v2();
2355}
2356
2357static void disable_iommus(void)
2358{
2359 struct amd_iommu *iommu;
2360
2361 for_each_iommu(iommu)
2362 iommu_disable(iommu);
2363
2364#ifdef CONFIG_IRQ_REMAP
2365 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir))
2366 amd_iommu_irq_ops.capability &= ~(1 << IRQ_POSTING_CAP);
2367#endif
2368}
2369
2370/*
2371 * Suspend/Resume support
2372 * disable suspend until real resume implemented
2373 */
2374
2375static void amd_iommu_resume(void)
2376{
2377 struct amd_iommu *iommu;
2378
2379 for_each_iommu(iommu)
2380 iommu_apply_resume_quirks(iommu);
2381
2382 /* re-load the hardware */
2383 enable_iommus();
2384
2385 amd_iommu_enable_interrupts();
2386}
2387
2388static int amd_iommu_suspend(void)
2389{
2390 /* disable IOMMUs to go out of the way for BIOS */
2391 disable_iommus();
2392
2393 return 0;
2394}
2395
2396static struct syscore_ops amd_iommu_syscore_ops = {
2397 .suspend = amd_iommu_suspend,
2398 .resume = amd_iommu_resume,
2399};
2400
2401static void __init free_iommu_resources(void)
2402{
2403 kmemleak_free(irq_lookup_table);
2404 free_pages((unsigned long)irq_lookup_table,
2405 get_order(rlookup_table_size));
2406 irq_lookup_table = NULL;
2407
2408 kmem_cache_destroy(amd_iommu_irq_cache);
2409 amd_iommu_irq_cache = NULL;
2410
2411 free_pages((unsigned long)amd_iommu_rlookup_table,
2412 get_order(rlookup_table_size));
2413 amd_iommu_rlookup_table = NULL;
2414
2415 free_pages((unsigned long)amd_iommu_alias_table,
2416 get_order(alias_table_size));
2417 amd_iommu_alias_table = NULL;
2418
2419 free_pages((unsigned long)amd_iommu_dev_table,
2420 get_order(dev_table_size));
2421 amd_iommu_dev_table = NULL;
2422
2423 free_iommu_all();
2424}
2425
2426/* SB IOAPIC is always on this device in AMD systems */
2427#define IOAPIC_SB_DEVID ((0x00 << 8) | PCI_DEVFN(0x14, 0))
2428
2429static bool __init check_ioapic_information(void)
2430{
2431 const char *fw_bug = FW_BUG;
2432 bool ret, has_sb_ioapic;
2433 int idx;
2434
2435 has_sb_ioapic = false;
2436 ret = false;
2437
2438 /*
2439 * If we have map overrides on the kernel command line the
2440 * messages in this function might not describe firmware bugs
2441 * anymore - so be careful
2442 */
2443 if (cmdline_maps)
2444 fw_bug = "";
2445
2446 for (idx = 0; idx < nr_ioapics; idx++) {
2447 int devid, id = mpc_ioapic_id(idx);
2448
2449 devid = get_ioapic_devid(id);
2450 if (devid < 0) {
2451 pr_err("%s: IOAPIC[%d] not in IVRS table\n",
2452 fw_bug, id);
2453 ret = false;
2454 } else if (devid == IOAPIC_SB_DEVID) {
2455 has_sb_ioapic = true;
2456 ret = true;
2457 }
2458 }
2459
2460 if (!has_sb_ioapic) {
2461 /*
2462 * We expect the SB IOAPIC to be listed in the IVRS
2463 * table. The system timer is connected to the SB IOAPIC
2464 * and if we don't have it in the list the system will
2465 * panic at boot time. This situation usually happens
2466 * when the BIOS is buggy and provides us the wrong
2467 * device id for the IOAPIC in the system.
2468 */
2469 pr_err("%s: No southbridge IOAPIC found\n", fw_bug);
2470 }
2471
2472 if (!ret)
2473 pr_err("Disabling interrupt remapping\n");
2474
2475 return ret;
2476}
2477
2478static void __init free_dma_resources(void)
2479{
2480 free_pages((unsigned long)amd_iommu_pd_alloc_bitmap,
2481 get_order(MAX_DOMAIN_ID/8));
2482 amd_iommu_pd_alloc_bitmap = NULL;
2483
2484 free_unity_maps();
2485}
2486
2487/*
2488 * This is the hardware init function for AMD IOMMU in the system.
2489 * This function is called either from amd_iommu_init or from the interrupt
2490 * remapping setup code.
2491 *
2492 * This function basically parses the ACPI table for AMD IOMMU (IVRS)
2493 * four times:
2494 *
2495 * 1 pass) Discover the most comprehensive IVHD type to use.
2496 *
2497 * 2 pass) Find the highest PCI device id the driver has to handle.
2498 * Upon this information the size of the data structures is
2499 * determined that needs to be allocated.
2500 *
2501 * 3 pass) Initialize the data structures just allocated with the
2502 * information in the ACPI table about available AMD IOMMUs
2503 * in the system. It also maps the PCI devices in the
2504 * system to specific IOMMUs
2505 *
2506 * 4 pass) After the basic data structures are allocated and
2507 * initialized we update them with information about memory
2508 * remapping requirements parsed out of the ACPI table in
2509 * this last pass.
2510 *
2511 * After everything is set up the IOMMUs are enabled and the necessary
2512 * hotplug and suspend notifiers are registered.
2513 */
2514static int __init early_amd_iommu_init(void)
2515{
2516 struct acpi_table_header *ivrs_base;
2517 acpi_status status;
2518 int i, remap_cache_sz, ret = 0;
2519
2520 if (!amd_iommu_detected)
2521 return -ENODEV;
2522
2523 status = acpi_get_table("IVRS", 0, &ivrs_base);
2524 if (status == AE_NOT_FOUND)
2525 return -ENODEV;
2526 else if (ACPI_FAILURE(status)) {
2527 const char *err = acpi_format_exception(status);
2528 pr_err("IVRS table error: %s\n", err);
2529 return -EINVAL;
2530 }
2531
2532 /*
2533 * Validate checksum here so we don't need to do it when
2534 * we actually parse the table
2535 */
2536 ret = check_ivrs_checksum(ivrs_base);
2537 if (ret)
2538 goto out;
2539
2540 amd_iommu_target_ivhd_type = get_highest_supported_ivhd_type(ivrs_base);
2541 DUMP_printk("Using IVHD type %#x\n", amd_iommu_target_ivhd_type);
2542
2543 /*
2544 * First parse ACPI tables to find the largest Bus/Dev/Func
2545 * we need to handle. Upon this information the shared data
2546 * structures for the IOMMUs in the system will be allocated
2547 */
2548 ret = find_last_devid_acpi(ivrs_base);
2549 if (ret)
2550 goto out;
2551
2552 dev_table_size = tbl_size(DEV_TABLE_ENTRY_SIZE);
2553 alias_table_size = tbl_size(ALIAS_TABLE_ENTRY_SIZE);
2554 rlookup_table_size = tbl_size(RLOOKUP_TABLE_ENTRY_SIZE);
2555
2556 /* Device table - directly used by all IOMMUs */
2557 ret = -ENOMEM;
2558 amd_iommu_dev_table = (void *)__get_free_pages(
2559 GFP_KERNEL | __GFP_ZERO | GFP_DMA32,
2560 get_order(dev_table_size));
2561 if (amd_iommu_dev_table == NULL)
2562 goto out;
2563
2564 /*
2565 * Alias table - map PCI Bus/Dev/Func to Bus/Dev/Func the
2566 * IOMMU see for that device
2567 */
2568 amd_iommu_alias_table = (void *)__get_free_pages(GFP_KERNEL,
2569 get_order(alias_table_size));
2570 if (amd_iommu_alias_table == NULL)
2571 goto out;
2572
2573 /* IOMMU rlookup table - find the IOMMU for a specific device */
2574 amd_iommu_rlookup_table = (void *)__get_free_pages(
2575 GFP_KERNEL | __GFP_ZERO,
2576 get_order(rlookup_table_size));
2577 if (amd_iommu_rlookup_table == NULL)
2578 goto out;
2579
2580 amd_iommu_pd_alloc_bitmap = (void *)__get_free_pages(
2581 GFP_KERNEL | __GFP_ZERO,
2582 get_order(MAX_DOMAIN_ID/8));
2583 if (amd_iommu_pd_alloc_bitmap == NULL)
2584 goto out;
2585
2586 /*
2587 * let all alias entries point to itself
2588 */
2589 for (i = 0; i <= amd_iommu_last_bdf; ++i)
2590 amd_iommu_alias_table[i] = i;
2591
2592 /*
2593 * never allocate domain 0 because its used as the non-allocated and
2594 * error value placeholder
2595 */
2596 __set_bit(0, amd_iommu_pd_alloc_bitmap);
2597
2598 /*
2599 * now the data structures are allocated and basically initialized
2600 * start the real acpi table scan
2601 */
2602 ret = init_iommu_all(ivrs_base);
2603 if (ret)
2604 goto out;
2605
2606 /* Disable any previously enabled IOMMUs */
2607 if (!is_kdump_kernel() || amd_iommu_disabled)
2608 disable_iommus();
2609
2610 if (amd_iommu_irq_remap)
2611 amd_iommu_irq_remap = check_ioapic_information();
2612
2613 if (amd_iommu_irq_remap) {
2614 /*
2615 * Interrupt remapping enabled, create kmem_cache for the
2616 * remapping tables.
2617 */
2618 ret = -ENOMEM;
2619 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
2620 remap_cache_sz = MAX_IRQS_PER_TABLE * sizeof(u32);
2621 else
2622 remap_cache_sz = MAX_IRQS_PER_TABLE * (sizeof(u64) * 2);
2623 amd_iommu_irq_cache = kmem_cache_create("irq_remap_cache",
2624 remap_cache_sz,
2625 IRQ_TABLE_ALIGNMENT,
2626 0, NULL);
2627 if (!amd_iommu_irq_cache)
2628 goto out;
2629
2630 irq_lookup_table = (void *)__get_free_pages(
2631 GFP_KERNEL | __GFP_ZERO,
2632 get_order(rlookup_table_size));
2633 kmemleak_alloc(irq_lookup_table, rlookup_table_size,
2634 1, GFP_KERNEL);
2635 if (!irq_lookup_table)
2636 goto out;
2637 }
2638
2639 ret = init_memory_definitions(ivrs_base);
2640 if (ret)
2641 goto out;
2642
2643 /* init the device table */
2644 init_device_table();
2645
2646out:
2647 /* Don't leak any ACPI memory */
2648 acpi_put_table(ivrs_base);
2649 ivrs_base = NULL;
2650
2651 return ret;
2652}
2653
2654static int amd_iommu_enable_interrupts(void)
2655{
2656 struct amd_iommu *iommu;
2657 int ret = 0;
2658
2659 for_each_iommu(iommu) {
2660 ret = iommu_init_msi(iommu);
2661 if (ret)
2662 goto out;
2663 }
2664
2665out:
2666 return ret;
2667}
2668
2669static bool detect_ivrs(void)
2670{
2671 struct acpi_table_header *ivrs_base;
2672 acpi_status status;
2673
2674 status = acpi_get_table("IVRS", 0, &ivrs_base);
2675 if (status == AE_NOT_FOUND)
2676 return false;
2677 else if (ACPI_FAILURE(status)) {
2678 const char *err = acpi_format_exception(status);
2679 pr_err("IVRS table error: %s\n", err);
2680 return false;
2681 }
2682
2683 acpi_put_table(ivrs_base);
2684
2685 /* Make sure ACS will be enabled during PCI probe */
2686 pci_request_acs();
2687
2688 return true;
2689}
2690
2691/****************************************************************************
2692 *
2693 * AMD IOMMU Initialization State Machine
2694 *
2695 ****************************************************************************/
2696
2697static int __init state_next(void)
2698{
2699 int ret = 0;
2700
2701 switch (init_state) {
2702 case IOMMU_START_STATE:
2703 if (!detect_ivrs()) {
2704 init_state = IOMMU_NOT_FOUND;
2705 ret = -ENODEV;
2706 } else {
2707 init_state = IOMMU_IVRS_DETECTED;
2708 }
2709 break;
2710 case IOMMU_IVRS_DETECTED:
2711 ret = early_amd_iommu_init();
2712 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_ACPI_FINISHED;
2713 if (init_state == IOMMU_ACPI_FINISHED && amd_iommu_disabled) {
2714 pr_info("AMD IOMMU disabled on kernel command-line\n");
2715 init_state = IOMMU_CMDLINE_DISABLED;
2716 ret = -EINVAL;
2717 }
2718 break;
2719 case IOMMU_ACPI_FINISHED:
2720 early_enable_iommus();
2721 x86_platform.iommu_shutdown = disable_iommus;
2722 init_state = IOMMU_ENABLED;
2723 break;
2724 case IOMMU_ENABLED:
2725 register_syscore_ops(&amd_iommu_syscore_ops);
2726 ret = amd_iommu_init_pci();
2727 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_PCI_INIT;
2728 enable_iommus_v2();
2729 break;
2730 case IOMMU_PCI_INIT:
2731 ret = amd_iommu_enable_interrupts();
2732 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_INTERRUPTS_EN;
2733 break;
2734 case IOMMU_INTERRUPTS_EN:
2735 ret = amd_iommu_init_dma_ops();
2736 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_DMA_OPS;
2737 break;
2738 case IOMMU_DMA_OPS:
2739 init_state = IOMMU_INITIALIZED;
2740 break;
2741 case IOMMU_INITIALIZED:
2742 /* Nothing to do */
2743 break;
2744 case IOMMU_NOT_FOUND:
2745 case IOMMU_INIT_ERROR:
2746 case IOMMU_CMDLINE_DISABLED:
2747 /* Error states => do nothing */
2748 ret = -EINVAL;
2749 break;
2750 default:
2751 /* Unknown state */
2752 BUG();
2753 }
2754
2755 if (ret) {
2756 free_dma_resources();
2757 if (!irq_remapping_enabled) {
2758 disable_iommus();
2759 free_iommu_resources();
2760 } else {
2761 struct amd_iommu *iommu;
2762
2763 uninit_device_table_dma();
2764 for_each_iommu(iommu)
2765 iommu_flush_all_caches(iommu);
2766 }
2767 }
2768 return ret;
2769}
2770
2771static int __init iommu_go_to_state(enum iommu_init_state state)
2772{
2773 int ret = -EINVAL;
2774
2775 while (init_state != state) {
2776 if (init_state == IOMMU_NOT_FOUND ||
2777 init_state == IOMMU_INIT_ERROR ||
2778 init_state == IOMMU_CMDLINE_DISABLED)
2779 break;
2780 ret = state_next();
2781 }
2782
2783 return ret;
2784}
2785
2786#ifdef CONFIG_IRQ_REMAP
2787int __init amd_iommu_prepare(void)
2788{
2789 int ret;
2790
2791 amd_iommu_irq_remap = true;
2792
2793 ret = iommu_go_to_state(IOMMU_ACPI_FINISHED);
2794 if (ret)
2795 return ret;
2796 return amd_iommu_irq_remap ? 0 : -ENODEV;
2797}
2798
2799int __init amd_iommu_enable(void)
2800{
2801 int ret;
2802
2803 ret = iommu_go_to_state(IOMMU_ENABLED);
2804 if (ret)
2805 return ret;
2806
2807 irq_remapping_enabled = 1;
2808 return amd_iommu_xt_mode;
2809}
2810
2811void amd_iommu_disable(void)
2812{
2813 amd_iommu_suspend();
2814}
2815
2816int amd_iommu_reenable(int mode)
2817{
2818 amd_iommu_resume();
2819
2820 return 0;
2821}
2822
2823int __init amd_iommu_enable_faulting(void)
2824{
2825 /* We enable MSI later when PCI is initialized */
2826 return 0;
2827}
2828#endif
2829
2830/*
2831 * This is the core init function for AMD IOMMU hardware in the system.
2832 * This function is called from the generic x86 DMA layer initialization
2833 * code.
2834 */
2835static int __init amd_iommu_init(void)
2836{
2837 struct amd_iommu *iommu;
2838 int ret;
2839
2840 ret = iommu_go_to_state(IOMMU_INITIALIZED);
2841#ifdef CONFIG_GART_IOMMU
2842 if (ret && list_empty(&amd_iommu_list)) {
2843 /*
2844 * We failed to initialize the AMD IOMMU - try fallback
2845 * to GART if possible.
2846 */
2847 gart_iommu_init();
2848 }
2849#endif
2850
2851 for_each_iommu(iommu)
2852 amd_iommu_debugfs_setup(iommu);
2853
2854 return ret;
2855}
2856
2857static bool amd_iommu_sme_check(void)
2858{
2859 if (!sme_active() || (boot_cpu_data.x86 != 0x17))
2860 return true;
2861
2862 /* For Fam17h, a specific level of support is required */
2863 if (boot_cpu_data.microcode >= 0x08001205)
2864 return true;
2865
2866 if ((boot_cpu_data.microcode >= 0x08001126) &&
2867 (boot_cpu_data.microcode <= 0x080011ff))
2868 return true;
2869
2870 pr_notice("IOMMU not currently supported when SME is active\n");
2871
2872 return false;
2873}
2874
2875/****************************************************************************
2876 *
2877 * Early detect code. This code runs at IOMMU detection time in the DMA
2878 * layer. It just looks if there is an IVRS ACPI table to detect AMD
2879 * IOMMUs
2880 *
2881 ****************************************************************************/
2882int __init amd_iommu_detect(void)
2883{
2884 int ret;
2885
2886 if (no_iommu || (iommu_detected && !gart_iommu_aperture))
2887 return -ENODEV;
2888
2889 if (!amd_iommu_sme_check())
2890 return -ENODEV;
2891
2892 ret = iommu_go_to_state(IOMMU_IVRS_DETECTED);
2893 if (ret)
2894 return ret;
2895
2896 amd_iommu_detected = true;
2897 iommu_detected = 1;
2898 x86_init.iommu.iommu_init = amd_iommu_init;
2899
2900 return 1;
2901}
2902
2903/****************************************************************************
2904 *
2905 * Parsing functions for the AMD IOMMU specific kernel command line
2906 * options.
2907 *
2908 ****************************************************************************/
2909
2910static int __init parse_amd_iommu_dump(char *str)
2911{
2912 amd_iommu_dump = true;
2913
2914 return 1;
2915}
2916
2917static int __init parse_amd_iommu_intr(char *str)
2918{
2919 for (; *str; ++str) {
2920 if (strncmp(str, "legacy", 6) == 0) {
2921 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY;
2922 break;
2923 }
2924 if (strncmp(str, "vapic", 5) == 0) {
2925 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_VAPIC;
2926 break;
2927 }
2928 }
2929 return 1;
2930}
2931
2932static int __init parse_amd_iommu_options(char *str)
2933{
2934 for (; *str; ++str) {
2935 if (strncmp(str, "fullflush", 9) == 0)
2936 amd_iommu_unmap_flush = true;
2937 if (strncmp(str, "off", 3) == 0)
2938 amd_iommu_disabled = true;
2939 if (strncmp(str, "force_isolation", 15) == 0)
2940 amd_iommu_force_isolation = true;
2941 }
2942
2943 return 1;
2944}
2945
2946static int __init parse_ivrs_ioapic(char *str)
2947{
2948 unsigned int bus, dev, fn;
2949 int ret, id, i;
2950 u16 devid;
2951
2952 ret = sscanf(str, "[%d]=%x:%x.%x", &id, &bus, &dev, &fn);
2953
2954 if (ret != 4) {
2955 pr_err("Invalid command line: ivrs_ioapic%s\n", str);
2956 return 1;
2957 }
2958
2959 if (early_ioapic_map_size == EARLY_MAP_SIZE) {
2960 pr_err("Early IOAPIC map overflow - ignoring ivrs_ioapic%s\n",
2961 str);
2962 return 1;
2963 }
2964
2965 devid = ((bus & 0xff) << 8) | ((dev & 0x1f) << 3) | (fn & 0x7);
2966
2967 cmdline_maps = true;
2968 i = early_ioapic_map_size++;
2969 early_ioapic_map[i].id = id;
2970 early_ioapic_map[i].devid = devid;
2971 early_ioapic_map[i].cmd_line = true;
2972
2973 return 1;
2974}
2975
2976static int __init parse_ivrs_hpet(char *str)
2977{
2978 unsigned int bus, dev, fn;
2979 int ret, id, i;
2980 u16 devid;
2981
2982 ret = sscanf(str, "[%d]=%x:%x.%x", &id, &bus, &dev, &fn);
2983
2984 if (ret != 4) {
2985 pr_err("Invalid command line: ivrs_hpet%s\n", str);
2986 return 1;
2987 }
2988
2989 if (early_hpet_map_size == EARLY_MAP_SIZE) {
2990 pr_err("Early HPET map overflow - ignoring ivrs_hpet%s\n",
2991 str);
2992 return 1;
2993 }
2994
2995 devid = ((bus & 0xff) << 8) | ((dev & 0x1f) << 3) | (fn & 0x7);
2996
2997 cmdline_maps = true;
2998 i = early_hpet_map_size++;
2999 early_hpet_map[i].id = id;
3000 early_hpet_map[i].devid = devid;
3001 early_hpet_map[i].cmd_line = true;
3002
3003 return 1;
3004}
3005
3006static int __init parse_ivrs_acpihid(char *str)
3007{
3008 u32 bus, dev, fn;
3009 char *hid, *uid, *p;
3010 char acpiid[ACPIHID_UID_LEN + ACPIHID_HID_LEN] = {0};
3011 int ret, i;
3012
3013 ret = sscanf(str, "[%x:%x.%x]=%s", &bus, &dev, &fn, acpiid);
3014 if (ret != 4) {
3015 pr_err("Invalid command line: ivrs_acpihid(%s)\n", str);
3016 return 1;
3017 }
3018
3019 p = acpiid;
3020 hid = strsep(&p, ":");
3021 uid = p;
3022
3023 if (!hid || !(*hid) || !uid) {
3024 pr_err("Invalid command line: hid or uid\n");
3025 return 1;
3026 }
3027
3028 i = early_acpihid_map_size++;
3029 memcpy(early_acpihid_map[i].hid, hid, strlen(hid));
3030 memcpy(early_acpihid_map[i].uid, uid, strlen(uid));
3031 early_acpihid_map[i].devid =
3032 ((bus & 0xff) << 8) | ((dev & 0x1f) << 3) | (fn & 0x7);
3033 early_acpihid_map[i].cmd_line = true;
3034
3035 return 1;
3036}
3037
3038__setup("amd_iommu_dump", parse_amd_iommu_dump);
3039__setup("amd_iommu=", parse_amd_iommu_options);
3040__setup("amd_iommu_intr=", parse_amd_iommu_intr);
3041__setup("ivrs_ioapic", parse_ivrs_ioapic);
3042__setup("ivrs_hpet", parse_ivrs_hpet);
3043__setup("ivrs_acpihid", parse_ivrs_acpihid);
3044
3045IOMMU_INIT_FINISH(amd_iommu_detect,
3046 gart_iommu_hole_init,
3047 NULL,
3048 NULL);
3049
3050bool amd_iommu_v2_supported(void)
3051{
3052 return amd_iommu_v2_present;
3053}
3054EXPORT_SYMBOL(amd_iommu_v2_supported);
3055
3056struct amd_iommu *get_amd_iommu(unsigned int idx)
3057{
3058 unsigned int i = 0;
3059 struct amd_iommu *iommu;
3060
3061 for_each_iommu(iommu)
3062 if (i++ == idx)
3063 return iommu;
3064 return NULL;
3065}
3066EXPORT_SYMBOL(get_amd_iommu);
3067
3068/****************************************************************************
3069 *
3070 * IOMMU EFR Performance Counter support functionality. This code allows
3071 * access to the IOMMU PC functionality.
3072 *
3073 ****************************************************************************/
3074
3075u8 amd_iommu_pc_get_max_banks(unsigned int idx)
3076{
3077 struct amd_iommu *iommu = get_amd_iommu(idx);
3078
3079 if (iommu)
3080 return iommu->max_banks;
3081
3082 return 0;
3083}
3084EXPORT_SYMBOL(amd_iommu_pc_get_max_banks);
3085
3086bool amd_iommu_pc_supported(void)
3087{
3088 return amd_iommu_pc_present;
3089}
3090EXPORT_SYMBOL(amd_iommu_pc_supported);
3091
3092u8 amd_iommu_pc_get_max_counters(unsigned int idx)
3093{
3094 struct amd_iommu *iommu = get_amd_iommu(idx);
3095
3096 if (iommu)
3097 return iommu->max_counters;
3098
3099 return 0;
3100}
3101EXPORT_SYMBOL(amd_iommu_pc_get_max_counters);
3102
3103static int iommu_pc_get_set_reg(struct amd_iommu *iommu, u8 bank, u8 cntr,
3104 u8 fxn, u64 *value, bool is_write)
3105{
3106 u32 offset;
3107 u32 max_offset_lim;
3108
3109 /* Make sure the IOMMU PC resource is available */
3110 if (!amd_iommu_pc_present)
3111 return -ENODEV;
3112
3113 /* Check for valid iommu and pc register indexing */
3114 if (WARN_ON(!iommu || (fxn > 0x28) || (fxn & 7)))
3115 return -ENODEV;
3116
3117 offset = (u32)(((0x40 | bank) << 12) | (cntr << 8) | fxn);
3118
3119 /* Limit the offset to the hw defined mmio region aperture */
3120 max_offset_lim = (u32)(((0x40 | iommu->max_banks) << 12) |
3121 (iommu->max_counters << 8) | 0x28);
3122 if ((offset < MMIO_CNTR_REG_OFFSET) ||
3123 (offset > max_offset_lim))
3124 return -EINVAL;
3125
3126 if (is_write) {
3127 u64 val = *value & GENMASK_ULL(47, 0);
3128
3129 writel((u32)val, iommu->mmio_base + offset);
3130 writel((val >> 32), iommu->mmio_base + offset + 4);
3131 } else {
3132 *value = readl(iommu->mmio_base + offset + 4);
3133 *value <<= 32;
3134 *value |= readl(iommu->mmio_base + offset);
3135 *value &= GENMASK_ULL(47, 0);
3136 }
3137
3138 return 0;
3139}
3140
3141int amd_iommu_pc_get_reg(struct amd_iommu *iommu, u8 bank, u8 cntr, u8 fxn, u64 *value)
3142{
3143 if (!iommu)
3144 return -EINVAL;
3145
3146 return iommu_pc_get_set_reg(iommu, bank, cntr, fxn, value, false);
3147}
3148EXPORT_SYMBOL(amd_iommu_pc_get_reg);
3149
3150int amd_iommu_pc_set_reg(struct amd_iommu *iommu, u8 bank, u8 cntr, u8 fxn, u64 *value)
3151{
3152 if (!iommu)
3153 return -EINVAL;
3154
3155 return iommu_pc_get_set_reg(iommu, bank, cntr, fxn, value, true);
3156}
3157EXPORT_SYMBOL(amd_iommu_pc_set_reg);