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/*
2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
4 * Leo Duran <leo.duran@amd.com>
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20#include <linux/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 <linux/export.h>
29#include <linux/iommu.h>
30#include <asm/pci-direct.h>
31#include <asm/iommu.h>
32#include <asm/gart.h>
33#include <asm/x86_init.h>
34#include <asm/iommu_table.h>
35#include <asm/io_apic.h>
36#include <asm/irq_remapping.h>
37
38#include "amd_iommu_proto.h"
39#include "amd_iommu_types.h"
40#include "irq_remapping.h"
41
42/*
43 * definitions for the ACPI scanning code
44 */
45#define IVRS_HEADER_LENGTH 48
46
47#define ACPI_IVHD_TYPE 0x10
48#define ACPI_IVMD_TYPE_ALL 0x20
49#define ACPI_IVMD_TYPE 0x21
50#define ACPI_IVMD_TYPE_RANGE 0x22
51
52#define IVHD_DEV_ALL 0x01
53#define IVHD_DEV_SELECT 0x02
54#define IVHD_DEV_SELECT_RANGE_START 0x03
55#define IVHD_DEV_RANGE_END 0x04
56#define IVHD_DEV_ALIAS 0x42
57#define IVHD_DEV_ALIAS_RANGE 0x43
58#define IVHD_DEV_EXT_SELECT 0x46
59#define IVHD_DEV_EXT_SELECT_RANGE 0x47
60#define IVHD_DEV_SPECIAL 0x48
61
62#define IVHD_SPECIAL_IOAPIC 1
63#define IVHD_SPECIAL_HPET 2
64
65#define IVHD_FLAG_HT_TUN_EN_MASK 0x01
66#define IVHD_FLAG_PASSPW_EN_MASK 0x02
67#define IVHD_FLAG_RESPASSPW_EN_MASK 0x04
68#define IVHD_FLAG_ISOC_EN_MASK 0x08
69
70#define IVMD_FLAG_EXCL_RANGE 0x08
71#define IVMD_FLAG_UNITY_MAP 0x01
72
73#define ACPI_DEVFLAG_INITPASS 0x01
74#define ACPI_DEVFLAG_EXTINT 0x02
75#define ACPI_DEVFLAG_NMI 0x04
76#define ACPI_DEVFLAG_SYSMGT1 0x10
77#define ACPI_DEVFLAG_SYSMGT2 0x20
78#define ACPI_DEVFLAG_LINT0 0x40
79#define ACPI_DEVFLAG_LINT1 0x80
80#define ACPI_DEVFLAG_ATSDIS 0x10000000
81
82/*
83 * ACPI table definitions
84 *
85 * These data structures are laid over the table to parse the important values
86 * out of it.
87 */
88
89/*
90 * structure describing one IOMMU in the ACPI table. Typically followed by one
91 * or more ivhd_entrys.
92 */
93struct ivhd_header {
94 u8 type;
95 u8 flags;
96 u16 length;
97 u16 devid;
98 u16 cap_ptr;
99 u64 mmio_phys;
100 u16 pci_seg;
101 u16 info;
102 u32 efr;
103} __attribute__((packed));
104
105/*
106 * A device entry describing which devices a specific IOMMU translates and
107 * which requestor ids they use.
108 */
109struct ivhd_entry {
110 u8 type;
111 u16 devid;
112 u8 flags;
113 u32 ext;
114} __attribute__((packed));
115
116/*
117 * An AMD IOMMU memory definition structure. It defines things like exclusion
118 * ranges for devices and regions that should be unity mapped.
119 */
120struct ivmd_header {
121 u8 type;
122 u8 flags;
123 u16 length;
124 u16 devid;
125 u16 aux;
126 u64 resv;
127 u64 range_start;
128 u64 range_length;
129} __attribute__((packed));
130
131bool amd_iommu_dump;
132bool amd_iommu_irq_remap __read_mostly;
133
134static bool amd_iommu_detected;
135static bool __initdata amd_iommu_disabled;
136
137u16 amd_iommu_last_bdf; /* largest PCI device id we have
138 to handle */
139LIST_HEAD(amd_iommu_unity_map); /* a list of required unity mappings
140 we find in ACPI */
141bool amd_iommu_unmap_flush; /* if true, flush on every unmap */
142
143LIST_HEAD(amd_iommu_list); /* list of all AMD IOMMUs in the
144 system */
145
146/* Array to assign indices to IOMMUs*/
147struct amd_iommu *amd_iommus[MAX_IOMMUS];
148int amd_iommus_present;
149
150/* IOMMUs have a non-present cache? */
151bool amd_iommu_np_cache __read_mostly;
152bool amd_iommu_iotlb_sup __read_mostly = true;
153
154u32 amd_iommu_max_pasid __read_mostly = ~0;
155
156bool amd_iommu_v2_present __read_mostly;
157static bool amd_iommu_pc_present __read_mostly;
158
159bool amd_iommu_force_isolation __read_mostly;
160
161/*
162 * List of protection domains - used during resume
163 */
164LIST_HEAD(amd_iommu_pd_list);
165spinlock_t amd_iommu_pd_lock;
166
167/*
168 * Pointer to the device table which is shared by all AMD IOMMUs
169 * it is indexed by the PCI device id or the HT unit id and contains
170 * information about the domain the device belongs to as well as the
171 * page table root pointer.
172 */
173struct dev_table_entry *amd_iommu_dev_table;
174
175/*
176 * The alias table is a driver specific data structure which contains the
177 * mappings of the PCI device ids to the actual requestor ids on the IOMMU.
178 * More than one device can share the same requestor id.
179 */
180u16 *amd_iommu_alias_table;
181
182/*
183 * The rlookup table is used to find the IOMMU which is responsible
184 * for a specific device. It is also indexed by the PCI device id.
185 */
186struct amd_iommu **amd_iommu_rlookup_table;
187
188/*
189 * This table is used to find the irq remapping table for a given device id
190 * quickly.
191 */
192struct irq_remap_table **irq_lookup_table;
193
194/*
195 * AMD IOMMU allows up to 2^16 different protection domains. This is a bitmap
196 * to know which ones are already in use.
197 */
198unsigned long *amd_iommu_pd_alloc_bitmap;
199
200static u32 dev_table_size; /* size of the device table */
201static u32 alias_table_size; /* size of the alias table */
202static u32 rlookup_table_size; /* size if the rlookup table */
203
204enum iommu_init_state {
205 IOMMU_START_STATE,
206 IOMMU_IVRS_DETECTED,
207 IOMMU_ACPI_FINISHED,
208 IOMMU_ENABLED,
209 IOMMU_PCI_INIT,
210 IOMMU_INTERRUPTS_EN,
211 IOMMU_DMA_OPS,
212 IOMMU_INITIALIZED,
213 IOMMU_NOT_FOUND,
214 IOMMU_INIT_ERROR,
215};
216
217/* Early ioapic and hpet maps from kernel command line */
218#define EARLY_MAP_SIZE 4
219static struct devid_map __initdata early_ioapic_map[EARLY_MAP_SIZE];
220static struct devid_map __initdata early_hpet_map[EARLY_MAP_SIZE];
221static int __initdata early_ioapic_map_size;
222static int __initdata early_hpet_map_size;
223static bool __initdata cmdline_maps;
224
225static enum iommu_init_state init_state = IOMMU_START_STATE;
226
227static int amd_iommu_enable_interrupts(void);
228static int __init iommu_go_to_state(enum iommu_init_state state);
229static void init_device_table_dma(void);
230
231static int iommu_pc_get_set_reg_val(struct amd_iommu *iommu,
232 u8 bank, u8 cntr, u8 fxn,
233 u64 *value, bool is_write);
234
235static inline void update_last_devid(u16 devid)
236{
237 if (devid > amd_iommu_last_bdf)
238 amd_iommu_last_bdf = devid;
239}
240
241static inline unsigned long tbl_size(int entry_size)
242{
243 unsigned shift = PAGE_SHIFT +
244 get_order(((int)amd_iommu_last_bdf + 1) * entry_size);
245
246 return 1UL << shift;
247}
248
249/* Access to l1 and l2 indexed register spaces */
250
251static u32 iommu_read_l1(struct amd_iommu *iommu, u16 l1, u8 address)
252{
253 u32 val;
254
255 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16));
256 pci_read_config_dword(iommu->dev, 0xfc, &val);
257 return val;
258}
259
260static void iommu_write_l1(struct amd_iommu *iommu, u16 l1, u8 address, u32 val)
261{
262 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16 | 1 << 31));
263 pci_write_config_dword(iommu->dev, 0xfc, val);
264 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16));
265}
266
267static u32 iommu_read_l2(struct amd_iommu *iommu, u8 address)
268{
269 u32 val;
270
271 pci_write_config_dword(iommu->dev, 0xf0, address);
272 pci_read_config_dword(iommu->dev, 0xf4, &val);
273 return val;
274}
275
276static void iommu_write_l2(struct amd_iommu *iommu, u8 address, u32 val)
277{
278 pci_write_config_dword(iommu->dev, 0xf0, (address | 1 << 8));
279 pci_write_config_dword(iommu->dev, 0xf4, val);
280}
281
282/****************************************************************************
283 *
284 * AMD IOMMU MMIO register space handling functions
285 *
286 * These functions are used to program the IOMMU device registers in
287 * MMIO space required for that driver.
288 *
289 ****************************************************************************/
290
291/*
292 * This function set the exclusion range in the IOMMU. DMA accesses to the
293 * exclusion range are passed through untranslated
294 */
295static void iommu_set_exclusion_range(struct amd_iommu *iommu)
296{
297 u64 start = iommu->exclusion_start & PAGE_MASK;
298 u64 limit = (start + iommu->exclusion_length) & PAGE_MASK;
299 u64 entry;
300
301 if (!iommu->exclusion_start)
302 return;
303
304 entry = start | MMIO_EXCL_ENABLE_MASK;
305 memcpy_toio(iommu->mmio_base + MMIO_EXCL_BASE_OFFSET,
306 &entry, sizeof(entry));
307
308 entry = limit;
309 memcpy_toio(iommu->mmio_base + MMIO_EXCL_LIMIT_OFFSET,
310 &entry, sizeof(entry));
311}
312
313/* Programs the physical address of the device table into the IOMMU hardware */
314static void iommu_set_device_table(struct amd_iommu *iommu)
315{
316 u64 entry;
317
318 BUG_ON(iommu->mmio_base == NULL);
319
320 entry = virt_to_phys(amd_iommu_dev_table);
321 entry |= (dev_table_size >> 12) - 1;
322 memcpy_toio(iommu->mmio_base + MMIO_DEV_TABLE_OFFSET,
323 &entry, sizeof(entry));
324}
325
326/* Generic functions to enable/disable certain features of the IOMMU. */
327static void iommu_feature_enable(struct amd_iommu *iommu, u8 bit)
328{
329 u32 ctrl;
330
331 ctrl = readl(iommu->mmio_base + MMIO_CONTROL_OFFSET);
332 ctrl |= (1 << bit);
333 writel(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET);
334}
335
336static void iommu_feature_disable(struct amd_iommu *iommu, u8 bit)
337{
338 u32 ctrl;
339
340 ctrl = readl(iommu->mmio_base + MMIO_CONTROL_OFFSET);
341 ctrl &= ~(1 << bit);
342 writel(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET);
343}
344
345static void iommu_set_inv_tlb_timeout(struct amd_iommu *iommu, int timeout)
346{
347 u32 ctrl;
348
349 ctrl = readl(iommu->mmio_base + MMIO_CONTROL_OFFSET);
350 ctrl &= ~CTRL_INV_TO_MASK;
351 ctrl |= (timeout << CONTROL_INV_TIMEOUT) & CTRL_INV_TO_MASK;
352 writel(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET);
353}
354
355/* Function to enable the hardware */
356static void iommu_enable(struct amd_iommu *iommu)
357{
358 iommu_feature_enable(iommu, CONTROL_IOMMU_EN);
359}
360
361static void iommu_disable(struct amd_iommu *iommu)
362{
363 /* Disable command buffer */
364 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN);
365
366 /* Disable event logging and event interrupts */
367 iommu_feature_disable(iommu, CONTROL_EVT_INT_EN);
368 iommu_feature_disable(iommu, CONTROL_EVT_LOG_EN);
369
370 /* Disable IOMMU hardware itself */
371 iommu_feature_disable(iommu, CONTROL_IOMMU_EN);
372}
373
374/*
375 * mapping and unmapping functions for the IOMMU MMIO space. Each AMD IOMMU in
376 * the system has one.
377 */
378static u8 __iomem * __init iommu_map_mmio_space(u64 address, u64 end)
379{
380 if (!request_mem_region(address, end, "amd_iommu")) {
381 pr_err("AMD-Vi: Can not reserve memory region %llx-%llx for mmio\n",
382 address, end);
383 pr_err("AMD-Vi: This is a BIOS bug. Please contact your hardware vendor\n");
384 return NULL;
385 }
386
387 return (u8 __iomem *)ioremap_nocache(address, end);
388}
389
390static void __init iommu_unmap_mmio_space(struct amd_iommu *iommu)
391{
392 if (iommu->mmio_base)
393 iounmap(iommu->mmio_base);
394 release_mem_region(iommu->mmio_phys, iommu->mmio_phys_end);
395}
396
397/****************************************************************************
398 *
399 * The functions below belong to the first pass of AMD IOMMU ACPI table
400 * parsing. In this pass we try to find out the highest device id this
401 * code has to handle. Upon this information the size of the shared data
402 * structures is determined later.
403 *
404 ****************************************************************************/
405
406/*
407 * This function calculates the length of a given IVHD entry
408 */
409static inline int ivhd_entry_length(u8 *ivhd)
410{
411 return 0x04 << (*ivhd >> 6);
412}
413
414/*
415 * After reading the highest device id from the IOMMU PCI capability header
416 * this function looks if there is a higher device id defined in the ACPI table
417 */
418static int __init find_last_devid_from_ivhd(struct ivhd_header *h)
419{
420 u8 *p = (void *)h, *end = (void *)h;
421 struct ivhd_entry *dev;
422
423 p += sizeof(*h);
424 end += h->length;
425
426 while (p < end) {
427 dev = (struct ivhd_entry *)p;
428 switch (dev->type) {
429 case IVHD_DEV_ALL:
430 /* Use maximum BDF value for DEV_ALL */
431 update_last_devid(0xffff);
432 break;
433 case IVHD_DEV_SELECT:
434 case IVHD_DEV_RANGE_END:
435 case IVHD_DEV_ALIAS:
436 case IVHD_DEV_EXT_SELECT:
437 /* all the above subfield types refer to device ids */
438 update_last_devid(dev->devid);
439 break;
440 default:
441 break;
442 }
443 p += ivhd_entry_length(p);
444 }
445
446 WARN_ON(p != end);
447
448 return 0;
449}
450
451/*
452 * Iterate over all IVHD entries in the ACPI table and find the highest device
453 * id which we need to handle. This is the first of three functions which parse
454 * the ACPI table. So we check the checksum here.
455 */
456static int __init find_last_devid_acpi(struct acpi_table_header *table)
457{
458 int i;
459 u8 checksum = 0, *p = (u8 *)table, *end = (u8 *)table;
460 struct ivhd_header *h;
461
462 /*
463 * Validate checksum here so we don't need to do it when
464 * we actually parse the table
465 */
466 for (i = 0; i < table->length; ++i)
467 checksum += p[i];
468 if (checksum != 0)
469 /* ACPI table corrupt */
470 return -ENODEV;
471
472 p += IVRS_HEADER_LENGTH;
473
474 end += table->length;
475 while (p < end) {
476 h = (struct ivhd_header *)p;
477 switch (h->type) {
478 case ACPI_IVHD_TYPE:
479 find_last_devid_from_ivhd(h);
480 break;
481 default:
482 break;
483 }
484 p += h->length;
485 }
486 WARN_ON(p != end);
487
488 return 0;
489}
490
491/****************************************************************************
492 *
493 * The following functions belong to the code path which parses the ACPI table
494 * the second time. In this ACPI parsing iteration we allocate IOMMU specific
495 * data structures, initialize the device/alias/rlookup table and also
496 * basically initialize the hardware.
497 *
498 ****************************************************************************/
499
500/*
501 * Allocates the command buffer. This buffer is per AMD IOMMU. We can
502 * write commands to that buffer later and the IOMMU will execute them
503 * asynchronously
504 */
505static int __init alloc_command_buffer(struct amd_iommu *iommu)
506{
507 iommu->cmd_buf = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
508 get_order(CMD_BUFFER_SIZE));
509
510 return iommu->cmd_buf ? 0 : -ENOMEM;
511}
512
513/*
514 * This function resets the command buffer if the IOMMU stopped fetching
515 * commands from it.
516 */
517void amd_iommu_reset_cmd_buffer(struct amd_iommu *iommu)
518{
519 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN);
520
521 writel(0x00, iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
522 writel(0x00, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
523
524 iommu_feature_enable(iommu, CONTROL_CMDBUF_EN);
525}
526
527/*
528 * This function writes the command buffer address to the hardware and
529 * enables it.
530 */
531static void iommu_enable_command_buffer(struct amd_iommu *iommu)
532{
533 u64 entry;
534
535 BUG_ON(iommu->cmd_buf == NULL);
536
537 entry = (u64)virt_to_phys(iommu->cmd_buf);
538 entry |= MMIO_CMD_SIZE_512;
539
540 memcpy_toio(iommu->mmio_base + MMIO_CMD_BUF_OFFSET,
541 &entry, sizeof(entry));
542
543 amd_iommu_reset_cmd_buffer(iommu);
544}
545
546static void __init free_command_buffer(struct amd_iommu *iommu)
547{
548 free_pages((unsigned long)iommu->cmd_buf, get_order(CMD_BUFFER_SIZE));
549}
550
551/* allocates the memory where the IOMMU will log its events to */
552static int __init alloc_event_buffer(struct amd_iommu *iommu)
553{
554 iommu->evt_buf = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
555 get_order(EVT_BUFFER_SIZE));
556
557 return iommu->evt_buf ? 0 : -ENOMEM;
558}
559
560static void iommu_enable_event_buffer(struct amd_iommu *iommu)
561{
562 u64 entry;
563
564 BUG_ON(iommu->evt_buf == NULL);
565
566 entry = (u64)virt_to_phys(iommu->evt_buf) | EVT_LEN_MASK;
567
568 memcpy_toio(iommu->mmio_base + MMIO_EVT_BUF_OFFSET,
569 &entry, sizeof(entry));
570
571 /* set head and tail to zero manually */
572 writel(0x00, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
573 writel(0x00, iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
574
575 iommu_feature_enable(iommu, CONTROL_EVT_LOG_EN);
576}
577
578static void __init free_event_buffer(struct amd_iommu *iommu)
579{
580 free_pages((unsigned long)iommu->evt_buf, get_order(EVT_BUFFER_SIZE));
581}
582
583/* allocates the memory where the IOMMU will log its events to */
584static int __init alloc_ppr_log(struct amd_iommu *iommu)
585{
586 iommu->ppr_log = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
587 get_order(PPR_LOG_SIZE));
588
589 return iommu->ppr_log ? 0 : -ENOMEM;
590}
591
592static void iommu_enable_ppr_log(struct amd_iommu *iommu)
593{
594 u64 entry;
595
596 if (iommu->ppr_log == NULL)
597 return;
598
599 entry = (u64)virt_to_phys(iommu->ppr_log) | PPR_LOG_SIZE_512;
600
601 memcpy_toio(iommu->mmio_base + MMIO_PPR_LOG_OFFSET,
602 &entry, sizeof(entry));
603
604 /* set head and tail to zero manually */
605 writel(0x00, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
606 writel(0x00, iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
607
608 iommu_feature_enable(iommu, CONTROL_PPFLOG_EN);
609 iommu_feature_enable(iommu, CONTROL_PPR_EN);
610}
611
612static void __init free_ppr_log(struct amd_iommu *iommu)
613{
614 if (iommu->ppr_log == NULL)
615 return;
616
617 free_pages((unsigned long)iommu->ppr_log, get_order(PPR_LOG_SIZE));
618}
619
620static void iommu_enable_gt(struct amd_iommu *iommu)
621{
622 if (!iommu_feature(iommu, FEATURE_GT))
623 return;
624
625 iommu_feature_enable(iommu, CONTROL_GT_EN);
626}
627
628/* sets a specific bit in the device table entry. */
629static void set_dev_entry_bit(u16 devid, u8 bit)
630{
631 int i = (bit >> 6) & 0x03;
632 int _bit = bit & 0x3f;
633
634 amd_iommu_dev_table[devid].data[i] |= (1UL << _bit);
635}
636
637static int get_dev_entry_bit(u16 devid, u8 bit)
638{
639 int i = (bit >> 6) & 0x03;
640 int _bit = bit & 0x3f;
641
642 return (amd_iommu_dev_table[devid].data[i] & (1UL << _bit)) >> _bit;
643}
644
645
646void amd_iommu_apply_erratum_63(u16 devid)
647{
648 int sysmgt;
649
650 sysmgt = get_dev_entry_bit(devid, DEV_ENTRY_SYSMGT1) |
651 (get_dev_entry_bit(devid, DEV_ENTRY_SYSMGT2) << 1);
652
653 if (sysmgt == 0x01)
654 set_dev_entry_bit(devid, DEV_ENTRY_IW);
655}
656
657/* Writes the specific IOMMU for a device into the rlookup table */
658static void __init set_iommu_for_device(struct amd_iommu *iommu, u16 devid)
659{
660 amd_iommu_rlookup_table[devid] = iommu;
661}
662
663/*
664 * This function takes the device specific flags read from the ACPI
665 * table and sets up the device table entry with that information
666 */
667static void __init set_dev_entry_from_acpi(struct amd_iommu *iommu,
668 u16 devid, u32 flags, u32 ext_flags)
669{
670 if (flags & ACPI_DEVFLAG_INITPASS)
671 set_dev_entry_bit(devid, DEV_ENTRY_INIT_PASS);
672 if (flags & ACPI_DEVFLAG_EXTINT)
673 set_dev_entry_bit(devid, DEV_ENTRY_EINT_PASS);
674 if (flags & ACPI_DEVFLAG_NMI)
675 set_dev_entry_bit(devid, DEV_ENTRY_NMI_PASS);
676 if (flags & ACPI_DEVFLAG_SYSMGT1)
677 set_dev_entry_bit(devid, DEV_ENTRY_SYSMGT1);
678 if (flags & ACPI_DEVFLAG_SYSMGT2)
679 set_dev_entry_bit(devid, DEV_ENTRY_SYSMGT2);
680 if (flags & ACPI_DEVFLAG_LINT0)
681 set_dev_entry_bit(devid, DEV_ENTRY_LINT0_PASS);
682 if (flags & ACPI_DEVFLAG_LINT1)
683 set_dev_entry_bit(devid, DEV_ENTRY_LINT1_PASS);
684
685 amd_iommu_apply_erratum_63(devid);
686
687 set_iommu_for_device(iommu, devid);
688}
689
690static int __init add_special_device(u8 type, u8 id, u16 *devid, bool cmd_line)
691{
692 struct devid_map *entry;
693 struct list_head *list;
694
695 if (type == IVHD_SPECIAL_IOAPIC)
696 list = &ioapic_map;
697 else if (type == IVHD_SPECIAL_HPET)
698 list = &hpet_map;
699 else
700 return -EINVAL;
701
702 list_for_each_entry(entry, list, list) {
703 if (!(entry->id == id && entry->cmd_line))
704 continue;
705
706 pr_info("AMD-Vi: Command-line override present for %s id %d - ignoring\n",
707 type == IVHD_SPECIAL_IOAPIC ? "IOAPIC" : "HPET", id);
708
709 *devid = entry->devid;
710
711 return 0;
712 }
713
714 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
715 if (!entry)
716 return -ENOMEM;
717
718 entry->id = id;
719 entry->devid = *devid;
720 entry->cmd_line = cmd_line;
721
722 list_add_tail(&entry->list, list);
723
724 return 0;
725}
726
727static int __init add_early_maps(void)
728{
729 int i, ret;
730
731 for (i = 0; i < early_ioapic_map_size; ++i) {
732 ret = add_special_device(IVHD_SPECIAL_IOAPIC,
733 early_ioapic_map[i].id,
734 &early_ioapic_map[i].devid,
735 early_ioapic_map[i].cmd_line);
736 if (ret)
737 return ret;
738 }
739
740 for (i = 0; i < early_hpet_map_size; ++i) {
741 ret = add_special_device(IVHD_SPECIAL_HPET,
742 early_hpet_map[i].id,
743 &early_hpet_map[i].devid,
744 early_hpet_map[i].cmd_line);
745 if (ret)
746 return ret;
747 }
748
749 return 0;
750}
751
752/*
753 * Reads the device exclusion range from ACPI and initializes the IOMMU with
754 * it
755 */
756static void __init set_device_exclusion_range(u16 devid, struct ivmd_header *m)
757{
758 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
759
760 if (!(m->flags & IVMD_FLAG_EXCL_RANGE))
761 return;
762
763 if (iommu) {
764 /*
765 * We only can configure exclusion ranges per IOMMU, not
766 * per device. But we can enable the exclusion range per
767 * device. This is done here
768 */
769 set_dev_entry_bit(devid, DEV_ENTRY_EX);
770 iommu->exclusion_start = m->range_start;
771 iommu->exclusion_length = m->range_length;
772 }
773}
774
775/*
776 * Takes a pointer to an AMD IOMMU entry in the ACPI table and
777 * initializes the hardware and our data structures with it.
778 */
779static int __init init_iommu_from_acpi(struct amd_iommu *iommu,
780 struct ivhd_header *h)
781{
782 u8 *p = (u8 *)h;
783 u8 *end = p, flags = 0;
784 u16 devid = 0, devid_start = 0, devid_to = 0;
785 u32 dev_i, ext_flags = 0;
786 bool alias = false;
787 struct ivhd_entry *e;
788 int ret;
789
790
791 ret = add_early_maps();
792 if (ret)
793 return ret;
794
795 /*
796 * First save the recommended feature enable bits from ACPI
797 */
798 iommu->acpi_flags = h->flags;
799
800 /*
801 * Done. Now parse the device entries
802 */
803 p += sizeof(struct ivhd_header);
804 end += h->length;
805
806
807 while (p < end) {
808 e = (struct ivhd_entry *)p;
809 switch (e->type) {
810 case IVHD_DEV_ALL:
811
812 DUMP_printk(" DEV_ALL\t\t\tflags: %02x\n", e->flags);
813
814 for (dev_i = 0; dev_i <= amd_iommu_last_bdf; ++dev_i)
815 set_dev_entry_from_acpi(iommu, dev_i, e->flags, 0);
816 break;
817 case IVHD_DEV_SELECT:
818
819 DUMP_printk(" DEV_SELECT\t\t\t devid: %02x:%02x.%x "
820 "flags: %02x\n",
821 PCI_BUS_NUM(e->devid),
822 PCI_SLOT(e->devid),
823 PCI_FUNC(e->devid),
824 e->flags);
825
826 devid = e->devid;
827 set_dev_entry_from_acpi(iommu, devid, e->flags, 0);
828 break;
829 case IVHD_DEV_SELECT_RANGE_START:
830
831 DUMP_printk(" DEV_SELECT_RANGE_START\t "
832 "devid: %02x:%02x.%x flags: %02x\n",
833 PCI_BUS_NUM(e->devid),
834 PCI_SLOT(e->devid),
835 PCI_FUNC(e->devid),
836 e->flags);
837
838 devid_start = e->devid;
839 flags = e->flags;
840 ext_flags = 0;
841 alias = false;
842 break;
843 case IVHD_DEV_ALIAS:
844
845 DUMP_printk(" DEV_ALIAS\t\t\t devid: %02x:%02x.%x "
846 "flags: %02x devid_to: %02x:%02x.%x\n",
847 PCI_BUS_NUM(e->devid),
848 PCI_SLOT(e->devid),
849 PCI_FUNC(e->devid),
850 e->flags,
851 PCI_BUS_NUM(e->ext >> 8),
852 PCI_SLOT(e->ext >> 8),
853 PCI_FUNC(e->ext >> 8));
854
855 devid = e->devid;
856 devid_to = e->ext >> 8;
857 set_dev_entry_from_acpi(iommu, devid , e->flags, 0);
858 set_dev_entry_from_acpi(iommu, devid_to, e->flags, 0);
859 amd_iommu_alias_table[devid] = devid_to;
860 break;
861 case IVHD_DEV_ALIAS_RANGE:
862
863 DUMP_printk(" DEV_ALIAS_RANGE\t\t "
864 "devid: %02x:%02x.%x flags: %02x "
865 "devid_to: %02x:%02x.%x\n",
866 PCI_BUS_NUM(e->devid),
867 PCI_SLOT(e->devid),
868 PCI_FUNC(e->devid),
869 e->flags,
870 PCI_BUS_NUM(e->ext >> 8),
871 PCI_SLOT(e->ext >> 8),
872 PCI_FUNC(e->ext >> 8));
873
874 devid_start = e->devid;
875 flags = e->flags;
876 devid_to = e->ext >> 8;
877 ext_flags = 0;
878 alias = true;
879 break;
880 case IVHD_DEV_EXT_SELECT:
881
882 DUMP_printk(" DEV_EXT_SELECT\t\t devid: %02x:%02x.%x "
883 "flags: %02x ext: %08x\n",
884 PCI_BUS_NUM(e->devid),
885 PCI_SLOT(e->devid),
886 PCI_FUNC(e->devid),
887 e->flags, e->ext);
888
889 devid = e->devid;
890 set_dev_entry_from_acpi(iommu, devid, e->flags,
891 e->ext);
892 break;
893 case IVHD_DEV_EXT_SELECT_RANGE:
894
895 DUMP_printk(" DEV_EXT_SELECT_RANGE\t devid: "
896 "%02x:%02x.%x flags: %02x ext: %08x\n",
897 PCI_BUS_NUM(e->devid),
898 PCI_SLOT(e->devid),
899 PCI_FUNC(e->devid),
900 e->flags, e->ext);
901
902 devid_start = e->devid;
903 flags = e->flags;
904 ext_flags = e->ext;
905 alias = false;
906 break;
907 case IVHD_DEV_RANGE_END:
908
909 DUMP_printk(" DEV_RANGE_END\t\t devid: %02x:%02x.%x\n",
910 PCI_BUS_NUM(e->devid),
911 PCI_SLOT(e->devid),
912 PCI_FUNC(e->devid));
913
914 devid = e->devid;
915 for (dev_i = devid_start; dev_i <= devid; ++dev_i) {
916 if (alias) {
917 amd_iommu_alias_table[dev_i] = devid_to;
918 set_dev_entry_from_acpi(iommu,
919 devid_to, flags, ext_flags);
920 }
921 set_dev_entry_from_acpi(iommu, dev_i,
922 flags, ext_flags);
923 }
924 break;
925 case IVHD_DEV_SPECIAL: {
926 u8 handle, type;
927 const char *var;
928 u16 devid;
929 int ret;
930
931 handle = e->ext & 0xff;
932 devid = (e->ext >> 8) & 0xffff;
933 type = (e->ext >> 24) & 0xff;
934
935 if (type == IVHD_SPECIAL_IOAPIC)
936 var = "IOAPIC";
937 else if (type == IVHD_SPECIAL_HPET)
938 var = "HPET";
939 else
940 var = "UNKNOWN";
941
942 DUMP_printk(" DEV_SPECIAL(%s[%d])\t\tdevid: %02x:%02x.%x\n",
943 var, (int)handle,
944 PCI_BUS_NUM(devid),
945 PCI_SLOT(devid),
946 PCI_FUNC(devid));
947
948 ret = add_special_device(type, handle, &devid, false);
949 if (ret)
950 return ret;
951
952 /*
953 * add_special_device might update the devid in case a
954 * command-line override is present. So call
955 * set_dev_entry_from_acpi after add_special_device.
956 */
957 set_dev_entry_from_acpi(iommu, devid, e->flags, 0);
958
959 break;
960 }
961 default:
962 break;
963 }
964
965 p += ivhd_entry_length(p);
966 }
967
968 return 0;
969}
970
971static void __init free_iommu_one(struct amd_iommu *iommu)
972{
973 free_command_buffer(iommu);
974 free_event_buffer(iommu);
975 free_ppr_log(iommu);
976 iommu_unmap_mmio_space(iommu);
977}
978
979static void __init free_iommu_all(void)
980{
981 struct amd_iommu *iommu, *next;
982
983 for_each_iommu_safe(iommu, next) {
984 list_del(&iommu->list);
985 free_iommu_one(iommu);
986 kfree(iommu);
987 }
988}
989
990/*
991 * Family15h Model 10h-1fh erratum 746 (IOMMU Logging May Stall Translations)
992 * Workaround:
993 * BIOS should disable L2B micellaneous clock gating by setting
994 * L2_L2B_CK_GATE_CONTROL[CKGateL2BMiscDisable](D0F2xF4_x90[2]) = 1b
995 */
996static void amd_iommu_erratum_746_workaround(struct amd_iommu *iommu)
997{
998 u32 value;
999
1000 if ((boot_cpu_data.x86 != 0x15) ||
1001 (boot_cpu_data.x86_model < 0x10) ||
1002 (boot_cpu_data.x86_model > 0x1f))
1003 return;
1004
1005 pci_write_config_dword(iommu->dev, 0xf0, 0x90);
1006 pci_read_config_dword(iommu->dev, 0xf4, &value);
1007
1008 if (value & BIT(2))
1009 return;
1010
1011 /* Select NB indirect register 0x90 and enable writing */
1012 pci_write_config_dword(iommu->dev, 0xf0, 0x90 | (1 << 8));
1013
1014 pci_write_config_dword(iommu->dev, 0xf4, value | 0x4);
1015 pr_info("AMD-Vi: Applying erratum 746 workaround for IOMMU at %s\n",
1016 dev_name(&iommu->dev->dev));
1017
1018 /* Clear the enable writing bit */
1019 pci_write_config_dword(iommu->dev, 0xf0, 0x90);
1020}
1021
1022/*
1023 * Family15h Model 30h-3fh (IOMMU Mishandles ATS Write Permission)
1024 * Workaround:
1025 * BIOS should enable ATS write permission check by setting
1026 * L2_DEBUG_3[AtsIgnoreIWDis](D0F2xF4_x47[0]) = 1b
1027 */
1028static void amd_iommu_ats_write_check_workaround(struct amd_iommu *iommu)
1029{
1030 u32 value;
1031
1032 if ((boot_cpu_data.x86 != 0x15) ||
1033 (boot_cpu_data.x86_model < 0x30) ||
1034 (boot_cpu_data.x86_model > 0x3f))
1035 return;
1036
1037 /* Test L2_DEBUG_3[AtsIgnoreIWDis] == 1 */
1038 value = iommu_read_l2(iommu, 0x47);
1039
1040 if (value & BIT(0))
1041 return;
1042
1043 /* Set L2_DEBUG_3[AtsIgnoreIWDis] = 1 */
1044 iommu_write_l2(iommu, 0x47, value | BIT(0));
1045
1046 pr_info("AMD-Vi: Applying ATS write check workaround for IOMMU at %s\n",
1047 dev_name(&iommu->dev->dev));
1048}
1049
1050/*
1051 * This function clues the initialization function for one IOMMU
1052 * together and also allocates the command buffer and programs the
1053 * hardware. It does NOT enable the IOMMU. This is done afterwards.
1054 */
1055static int __init init_iommu_one(struct amd_iommu *iommu, struct ivhd_header *h)
1056{
1057 int ret;
1058
1059 spin_lock_init(&iommu->lock);
1060
1061 /* Add IOMMU to internal data structures */
1062 list_add_tail(&iommu->list, &amd_iommu_list);
1063 iommu->index = amd_iommus_present++;
1064
1065 if (unlikely(iommu->index >= MAX_IOMMUS)) {
1066 WARN(1, "AMD-Vi: System has more IOMMUs than supported by this driver\n");
1067 return -ENOSYS;
1068 }
1069
1070 /* Index is fine - add IOMMU to the array */
1071 amd_iommus[iommu->index] = iommu;
1072
1073 /*
1074 * Copy data from ACPI table entry to the iommu struct
1075 */
1076 iommu->devid = h->devid;
1077 iommu->cap_ptr = h->cap_ptr;
1078 iommu->pci_seg = h->pci_seg;
1079 iommu->mmio_phys = h->mmio_phys;
1080
1081 /* Check if IVHD EFR contains proper max banks/counters */
1082 if ((h->efr != 0) &&
1083 ((h->efr & (0xF << 13)) != 0) &&
1084 ((h->efr & (0x3F << 17)) != 0)) {
1085 iommu->mmio_phys_end = MMIO_REG_END_OFFSET;
1086 } else {
1087 iommu->mmio_phys_end = MMIO_CNTR_CONF_OFFSET;
1088 }
1089
1090 iommu->mmio_base = iommu_map_mmio_space(iommu->mmio_phys,
1091 iommu->mmio_phys_end);
1092 if (!iommu->mmio_base)
1093 return -ENOMEM;
1094
1095 if (alloc_command_buffer(iommu))
1096 return -ENOMEM;
1097
1098 if (alloc_event_buffer(iommu))
1099 return -ENOMEM;
1100
1101 iommu->int_enabled = false;
1102
1103 ret = init_iommu_from_acpi(iommu, h);
1104 if (ret)
1105 return ret;
1106
1107 ret = amd_iommu_create_irq_domain(iommu);
1108 if (ret)
1109 return ret;
1110
1111 /*
1112 * Make sure IOMMU is not considered to translate itself. The IVRS
1113 * table tells us so, but this is a lie!
1114 */
1115 amd_iommu_rlookup_table[iommu->devid] = NULL;
1116
1117 return 0;
1118}
1119
1120/*
1121 * Iterates over all IOMMU entries in the ACPI table, allocates the
1122 * IOMMU structure and initializes it with init_iommu_one()
1123 */
1124static int __init init_iommu_all(struct acpi_table_header *table)
1125{
1126 u8 *p = (u8 *)table, *end = (u8 *)table;
1127 struct ivhd_header *h;
1128 struct amd_iommu *iommu;
1129 int ret;
1130
1131 end += table->length;
1132 p += IVRS_HEADER_LENGTH;
1133
1134 while (p < end) {
1135 h = (struct ivhd_header *)p;
1136 switch (*p) {
1137 case ACPI_IVHD_TYPE:
1138
1139 DUMP_printk("device: %02x:%02x.%01x cap: %04x "
1140 "seg: %d flags: %01x info %04x\n",
1141 PCI_BUS_NUM(h->devid), PCI_SLOT(h->devid),
1142 PCI_FUNC(h->devid), h->cap_ptr,
1143 h->pci_seg, h->flags, h->info);
1144 DUMP_printk(" mmio-addr: %016llx\n",
1145 h->mmio_phys);
1146
1147 iommu = kzalloc(sizeof(struct amd_iommu), GFP_KERNEL);
1148 if (iommu == NULL)
1149 return -ENOMEM;
1150
1151 ret = init_iommu_one(iommu, h);
1152 if (ret)
1153 return ret;
1154 break;
1155 default:
1156 break;
1157 }
1158 p += h->length;
1159
1160 }
1161 WARN_ON(p != end);
1162
1163 return 0;
1164}
1165
1166
1167static void init_iommu_perf_ctr(struct amd_iommu *iommu)
1168{
1169 u64 val = 0xabcd, val2 = 0;
1170
1171 if (!iommu_feature(iommu, FEATURE_PC))
1172 return;
1173
1174 amd_iommu_pc_present = true;
1175
1176 /* Check if the performance counters can be written to */
1177 if ((0 != iommu_pc_get_set_reg_val(iommu, 0, 0, 0, &val, true)) ||
1178 (0 != iommu_pc_get_set_reg_val(iommu, 0, 0, 0, &val2, false)) ||
1179 (val != val2)) {
1180 pr_err("AMD-Vi: Unable to write to IOMMU perf counter.\n");
1181 amd_iommu_pc_present = false;
1182 return;
1183 }
1184
1185 pr_info("AMD-Vi: IOMMU performance counters supported\n");
1186
1187 val = readl(iommu->mmio_base + MMIO_CNTR_CONF_OFFSET);
1188 iommu->max_banks = (u8) ((val >> 12) & 0x3f);
1189 iommu->max_counters = (u8) ((val >> 7) & 0xf);
1190}
1191
1192static ssize_t amd_iommu_show_cap(struct device *dev,
1193 struct device_attribute *attr,
1194 char *buf)
1195{
1196 struct amd_iommu *iommu = dev_get_drvdata(dev);
1197 return sprintf(buf, "%x\n", iommu->cap);
1198}
1199static DEVICE_ATTR(cap, S_IRUGO, amd_iommu_show_cap, NULL);
1200
1201static ssize_t amd_iommu_show_features(struct device *dev,
1202 struct device_attribute *attr,
1203 char *buf)
1204{
1205 struct amd_iommu *iommu = dev_get_drvdata(dev);
1206 return sprintf(buf, "%llx\n", iommu->features);
1207}
1208static DEVICE_ATTR(features, S_IRUGO, amd_iommu_show_features, NULL);
1209
1210static struct attribute *amd_iommu_attrs[] = {
1211 &dev_attr_cap.attr,
1212 &dev_attr_features.attr,
1213 NULL,
1214};
1215
1216static struct attribute_group amd_iommu_group = {
1217 .name = "amd-iommu",
1218 .attrs = amd_iommu_attrs,
1219};
1220
1221static const struct attribute_group *amd_iommu_groups[] = {
1222 &amd_iommu_group,
1223 NULL,
1224};
1225
1226static int iommu_init_pci(struct amd_iommu *iommu)
1227{
1228 int cap_ptr = iommu->cap_ptr;
1229 u32 range, misc, low, high;
1230
1231 iommu->dev = pci_get_bus_and_slot(PCI_BUS_NUM(iommu->devid),
1232 iommu->devid & 0xff);
1233 if (!iommu->dev)
1234 return -ENODEV;
1235
1236 /* Prevent binding other PCI device drivers to IOMMU devices */
1237 iommu->dev->match_driver = false;
1238
1239 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_CAP_HDR_OFFSET,
1240 &iommu->cap);
1241 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_RANGE_OFFSET,
1242 &range);
1243 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_MISC_OFFSET,
1244 &misc);
1245
1246 if (!(iommu->cap & (1 << IOMMU_CAP_IOTLB)))
1247 amd_iommu_iotlb_sup = false;
1248
1249 /* read extended feature bits */
1250 low = readl(iommu->mmio_base + MMIO_EXT_FEATURES);
1251 high = readl(iommu->mmio_base + MMIO_EXT_FEATURES + 4);
1252
1253 iommu->features = ((u64)high << 32) | low;
1254
1255 if (iommu_feature(iommu, FEATURE_GT)) {
1256 int glxval;
1257 u32 max_pasid;
1258 u64 pasmax;
1259
1260 pasmax = iommu->features & FEATURE_PASID_MASK;
1261 pasmax >>= FEATURE_PASID_SHIFT;
1262 max_pasid = (1 << (pasmax + 1)) - 1;
1263
1264 amd_iommu_max_pasid = min(amd_iommu_max_pasid, max_pasid);
1265
1266 BUG_ON(amd_iommu_max_pasid & ~PASID_MASK);
1267
1268 glxval = iommu->features & FEATURE_GLXVAL_MASK;
1269 glxval >>= FEATURE_GLXVAL_SHIFT;
1270
1271 if (amd_iommu_max_glx_val == -1)
1272 amd_iommu_max_glx_val = glxval;
1273 else
1274 amd_iommu_max_glx_val = min(amd_iommu_max_glx_val, glxval);
1275 }
1276
1277 if (iommu_feature(iommu, FEATURE_GT) &&
1278 iommu_feature(iommu, FEATURE_PPR)) {
1279 iommu->is_iommu_v2 = true;
1280 amd_iommu_v2_present = true;
1281 }
1282
1283 if (iommu_feature(iommu, FEATURE_PPR) && alloc_ppr_log(iommu))
1284 return -ENOMEM;
1285
1286 if (iommu->cap & (1UL << IOMMU_CAP_NPCACHE))
1287 amd_iommu_np_cache = true;
1288
1289 init_iommu_perf_ctr(iommu);
1290
1291 if (is_rd890_iommu(iommu->dev)) {
1292 int i, j;
1293
1294 iommu->root_pdev = pci_get_bus_and_slot(iommu->dev->bus->number,
1295 PCI_DEVFN(0, 0));
1296
1297 /*
1298 * Some rd890 systems may not be fully reconfigured by the
1299 * BIOS, so it's necessary for us to store this information so
1300 * it can be reprogrammed on resume
1301 */
1302 pci_read_config_dword(iommu->dev, iommu->cap_ptr + 4,
1303 &iommu->stored_addr_lo);
1304 pci_read_config_dword(iommu->dev, iommu->cap_ptr + 8,
1305 &iommu->stored_addr_hi);
1306
1307 /* Low bit locks writes to configuration space */
1308 iommu->stored_addr_lo &= ~1;
1309
1310 for (i = 0; i < 6; i++)
1311 for (j = 0; j < 0x12; j++)
1312 iommu->stored_l1[i][j] = iommu_read_l1(iommu, i, j);
1313
1314 for (i = 0; i < 0x83; i++)
1315 iommu->stored_l2[i] = iommu_read_l2(iommu, i);
1316 }
1317
1318 amd_iommu_erratum_746_workaround(iommu);
1319 amd_iommu_ats_write_check_workaround(iommu);
1320
1321 iommu->iommu_dev = iommu_device_create(&iommu->dev->dev, iommu,
1322 amd_iommu_groups, "ivhd%d",
1323 iommu->index);
1324
1325 return pci_enable_device(iommu->dev);
1326}
1327
1328static void print_iommu_info(void)
1329{
1330 static const char * const feat_str[] = {
1331 "PreF", "PPR", "X2APIC", "NX", "GT", "[5]",
1332 "IA", "GA", "HE", "PC"
1333 };
1334 struct amd_iommu *iommu;
1335
1336 for_each_iommu(iommu) {
1337 int i;
1338
1339 pr_info("AMD-Vi: Found IOMMU at %s cap 0x%hx\n",
1340 dev_name(&iommu->dev->dev), iommu->cap_ptr);
1341
1342 if (iommu->cap & (1 << IOMMU_CAP_EFR)) {
1343 pr_info("AMD-Vi: Extended features: ");
1344 for (i = 0; i < ARRAY_SIZE(feat_str); ++i) {
1345 if (iommu_feature(iommu, (1ULL << i)))
1346 pr_cont(" %s", feat_str[i]);
1347 }
1348 pr_cont("\n");
1349 }
1350 }
1351 if (irq_remapping_enabled)
1352 pr_info("AMD-Vi: Interrupt remapping enabled\n");
1353}
1354
1355static int __init amd_iommu_init_pci(void)
1356{
1357 struct amd_iommu *iommu;
1358 int ret = 0;
1359
1360 for_each_iommu(iommu) {
1361 ret = iommu_init_pci(iommu);
1362 if (ret)
1363 break;
1364 }
1365
1366 init_device_table_dma();
1367
1368 for_each_iommu(iommu)
1369 iommu_flush_all_caches(iommu);
1370
1371 ret = amd_iommu_init_api();
1372
1373 if (!ret)
1374 print_iommu_info();
1375
1376 return ret;
1377}
1378
1379/****************************************************************************
1380 *
1381 * The following functions initialize the MSI interrupts for all IOMMUs
1382 * in the system. It's a bit challenging because there could be multiple
1383 * IOMMUs per PCI BDF but we can call pci_enable_msi(x) only once per
1384 * pci_dev.
1385 *
1386 ****************************************************************************/
1387
1388static int iommu_setup_msi(struct amd_iommu *iommu)
1389{
1390 int r;
1391
1392 r = pci_enable_msi(iommu->dev);
1393 if (r)
1394 return r;
1395
1396 r = request_threaded_irq(iommu->dev->irq,
1397 amd_iommu_int_handler,
1398 amd_iommu_int_thread,
1399 0, "AMD-Vi",
1400 iommu);
1401
1402 if (r) {
1403 pci_disable_msi(iommu->dev);
1404 return r;
1405 }
1406
1407 iommu->int_enabled = true;
1408
1409 return 0;
1410}
1411
1412static int iommu_init_msi(struct amd_iommu *iommu)
1413{
1414 int ret;
1415
1416 if (iommu->int_enabled)
1417 goto enable_faults;
1418
1419 if (iommu->dev->msi_cap)
1420 ret = iommu_setup_msi(iommu);
1421 else
1422 ret = -ENODEV;
1423
1424 if (ret)
1425 return ret;
1426
1427enable_faults:
1428 iommu_feature_enable(iommu, CONTROL_EVT_INT_EN);
1429
1430 if (iommu->ppr_log != NULL)
1431 iommu_feature_enable(iommu, CONTROL_PPFINT_EN);
1432
1433 return 0;
1434}
1435
1436/****************************************************************************
1437 *
1438 * The next functions belong to the third pass of parsing the ACPI
1439 * table. In this last pass the memory mapping requirements are
1440 * gathered (like exclusion and unity mapping ranges).
1441 *
1442 ****************************************************************************/
1443
1444static void __init free_unity_maps(void)
1445{
1446 struct unity_map_entry *entry, *next;
1447
1448 list_for_each_entry_safe(entry, next, &amd_iommu_unity_map, list) {
1449 list_del(&entry->list);
1450 kfree(entry);
1451 }
1452}
1453
1454/* called when we find an exclusion range definition in ACPI */
1455static int __init init_exclusion_range(struct ivmd_header *m)
1456{
1457 int i;
1458
1459 switch (m->type) {
1460 case ACPI_IVMD_TYPE:
1461 set_device_exclusion_range(m->devid, m);
1462 break;
1463 case ACPI_IVMD_TYPE_ALL:
1464 for (i = 0; i <= amd_iommu_last_bdf; ++i)
1465 set_device_exclusion_range(i, m);
1466 break;
1467 case ACPI_IVMD_TYPE_RANGE:
1468 for (i = m->devid; i <= m->aux; ++i)
1469 set_device_exclusion_range(i, m);
1470 break;
1471 default:
1472 break;
1473 }
1474
1475 return 0;
1476}
1477
1478/* called for unity map ACPI definition */
1479static int __init init_unity_map_range(struct ivmd_header *m)
1480{
1481 struct unity_map_entry *e = NULL;
1482 char *s;
1483
1484 e = kzalloc(sizeof(*e), GFP_KERNEL);
1485 if (e == NULL)
1486 return -ENOMEM;
1487
1488 switch (m->type) {
1489 default:
1490 kfree(e);
1491 return 0;
1492 case ACPI_IVMD_TYPE:
1493 s = "IVMD_TYPEi\t\t\t";
1494 e->devid_start = e->devid_end = m->devid;
1495 break;
1496 case ACPI_IVMD_TYPE_ALL:
1497 s = "IVMD_TYPE_ALL\t\t";
1498 e->devid_start = 0;
1499 e->devid_end = amd_iommu_last_bdf;
1500 break;
1501 case ACPI_IVMD_TYPE_RANGE:
1502 s = "IVMD_TYPE_RANGE\t\t";
1503 e->devid_start = m->devid;
1504 e->devid_end = m->aux;
1505 break;
1506 }
1507 e->address_start = PAGE_ALIGN(m->range_start);
1508 e->address_end = e->address_start + PAGE_ALIGN(m->range_length);
1509 e->prot = m->flags >> 1;
1510
1511 DUMP_printk("%s devid_start: %02x:%02x.%x devid_end: %02x:%02x.%x"
1512 " range_start: %016llx range_end: %016llx flags: %x\n", s,
1513 PCI_BUS_NUM(e->devid_start), PCI_SLOT(e->devid_start),
1514 PCI_FUNC(e->devid_start), PCI_BUS_NUM(e->devid_end),
1515 PCI_SLOT(e->devid_end), PCI_FUNC(e->devid_end),
1516 e->address_start, e->address_end, m->flags);
1517
1518 list_add_tail(&e->list, &amd_iommu_unity_map);
1519
1520 return 0;
1521}
1522
1523/* iterates over all memory definitions we find in the ACPI table */
1524static int __init init_memory_definitions(struct acpi_table_header *table)
1525{
1526 u8 *p = (u8 *)table, *end = (u8 *)table;
1527 struct ivmd_header *m;
1528
1529 end += table->length;
1530 p += IVRS_HEADER_LENGTH;
1531
1532 while (p < end) {
1533 m = (struct ivmd_header *)p;
1534 if (m->flags & IVMD_FLAG_EXCL_RANGE)
1535 init_exclusion_range(m);
1536 else if (m->flags & IVMD_FLAG_UNITY_MAP)
1537 init_unity_map_range(m);
1538
1539 p += m->length;
1540 }
1541
1542 return 0;
1543}
1544
1545/*
1546 * Init the device table to not allow DMA access for devices and
1547 * suppress all page faults
1548 */
1549static void init_device_table_dma(void)
1550{
1551 u32 devid;
1552
1553 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid) {
1554 set_dev_entry_bit(devid, DEV_ENTRY_VALID);
1555 set_dev_entry_bit(devid, DEV_ENTRY_TRANSLATION);
1556 }
1557}
1558
1559static void __init uninit_device_table_dma(void)
1560{
1561 u32 devid;
1562
1563 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid) {
1564 amd_iommu_dev_table[devid].data[0] = 0ULL;
1565 amd_iommu_dev_table[devid].data[1] = 0ULL;
1566 }
1567}
1568
1569static void init_device_table(void)
1570{
1571 u32 devid;
1572
1573 if (!amd_iommu_irq_remap)
1574 return;
1575
1576 for (devid = 0; devid <= amd_iommu_last_bdf; ++devid)
1577 set_dev_entry_bit(devid, DEV_ENTRY_IRQ_TBL_EN);
1578}
1579
1580static void iommu_init_flags(struct amd_iommu *iommu)
1581{
1582 iommu->acpi_flags & IVHD_FLAG_HT_TUN_EN_MASK ?
1583 iommu_feature_enable(iommu, CONTROL_HT_TUN_EN) :
1584 iommu_feature_disable(iommu, CONTROL_HT_TUN_EN);
1585
1586 iommu->acpi_flags & IVHD_FLAG_PASSPW_EN_MASK ?
1587 iommu_feature_enable(iommu, CONTROL_PASSPW_EN) :
1588 iommu_feature_disable(iommu, CONTROL_PASSPW_EN);
1589
1590 iommu->acpi_flags & IVHD_FLAG_RESPASSPW_EN_MASK ?
1591 iommu_feature_enable(iommu, CONTROL_RESPASSPW_EN) :
1592 iommu_feature_disable(iommu, CONTROL_RESPASSPW_EN);
1593
1594 iommu->acpi_flags & IVHD_FLAG_ISOC_EN_MASK ?
1595 iommu_feature_enable(iommu, CONTROL_ISOC_EN) :
1596 iommu_feature_disable(iommu, CONTROL_ISOC_EN);
1597
1598 /*
1599 * make IOMMU memory accesses cache coherent
1600 */
1601 iommu_feature_enable(iommu, CONTROL_COHERENT_EN);
1602
1603 /* Set IOTLB invalidation timeout to 1s */
1604 iommu_set_inv_tlb_timeout(iommu, CTRL_INV_TO_1S);
1605}
1606
1607static void iommu_apply_resume_quirks(struct amd_iommu *iommu)
1608{
1609 int i, j;
1610 u32 ioc_feature_control;
1611 struct pci_dev *pdev = iommu->root_pdev;
1612
1613 /* RD890 BIOSes may not have completely reconfigured the iommu */
1614 if (!is_rd890_iommu(iommu->dev) || !pdev)
1615 return;
1616
1617 /*
1618 * First, we need to ensure that the iommu is enabled. This is
1619 * controlled by a register in the northbridge
1620 */
1621
1622 /* Select Northbridge indirect register 0x75 and enable writing */
1623 pci_write_config_dword(pdev, 0x60, 0x75 | (1 << 7));
1624 pci_read_config_dword(pdev, 0x64, &ioc_feature_control);
1625
1626 /* Enable the iommu */
1627 if (!(ioc_feature_control & 0x1))
1628 pci_write_config_dword(pdev, 0x64, ioc_feature_control | 1);
1629
1630 /* Restore the iommu BAR */
1631 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 4,
1632 iommu->stored_addr_lo);
1633 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 8,
1634 iommu->stored_addr_hi);
1635
1636 /* Restore the l1 indirect regs for each of the 6 l1s */
1637 for (i = 0; i < 6; i++)
1638 for (j = 0; j < 0x12; j++)
1639 iommu_write_l1(iommu, i, j, iommu->stored_l1[i][j]);
1640
1641 /* Restore the l2 indirect regs */
1642 for (i = 0; i < 0x83; i++)
1643 iommu_write_l2(iommu, i, iommu->stored_l2[i]);
1644
1645 /* Lock PCI setup registers */
1646 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 4,
1647 iommu->stored_addr_lo | 1);
1648}
1649
1650/*
1651 * This function finally enables all IOMMUs found in the system after
1652 * they have been initialized
1653 */
1654static void early_enable_iommus(void)
1655{
1656 struct amd_iommu *iommu;
1657
1658 for_each_iommu(iommu) {
1659 iommu_disable(iommu);
1660 iommu_init_flags(iommu);
1661 iommu_set_device_table(iommu);
1662 iommu_enable_command_buffer(iommu);
1663 iommu_enable_event_buffer(iommu);
1664 iommu_set_exclusion_range(iommu);
1665 iommu_enable(iommu);
1666 iommu_flush_all_caches(iommu);
1667 }
1668}
1669
1670static void enable_iommus_v2(void)
1671{
1672 struct amd_iommu *iommu;
1673
1674 for_each_iommu(iommu) {
1675 iommu_enable_ppr_log(iommu);
1676 iommu_enable_gt(iommu);
1677 }
1678}
1679
1680static void enable_iommus(void)
1681{
1682 early_enable_iommus();
1683
1684 enable_iommus_v2();
1685}
1686
1687static void disable_iommus(void)
1688{
1689 struct amd_iommu *iommu;
1690
1691 for_each_iommu(iommu)
1692 iommu_disable(iommu);
1693}
1694
1695/*
1696 * Suspend/Resume support
1697 * disable suspend until real resume implemented
1698 */
1699
1700static void amd_iommu_resume(void)
1701{
1702 struct amd_iommu *iommu;
1703
1704 for_each_iommu(iommu)
1705 iommu_apply_resume_quirks(iommu);
1706
1707 /* re-load the hardware */
1708 enable_iommus();
1709
1710 amd_iommu_enable_interrupts();
1711}
1712
1713static int amd_iommu_suspend(void)
1714{
1715 /* disable IOMMUs to go out of the way for BIOS */
1716 disable_iommus();
1717
1718 return 0;
1719}
1720
1721static struct syscore_ops amd_iommu_syscore_ops = {
1722 .suspend = amd_iommu_suspend,
1723 .resume = amd_iommu_resume,
1724};
1725
1726static void __init free_on_init_error(void)
1727{
1728 free_pages((unsigned long)irq_lookup_table,
1729 get_order(rlookup_table_size));
1730
1731 kmem_cache_destroy(amd_iommu_irq_cache);
1732 amd_iommu_irq_cache = NULL;
1733
1734 free_pages((unsigned long)amd_iommu_rlookup_table,
1735 get_order(rlookup_table_size));
1736
1737 free_pages((unsigned long)amd_iommu_alias_table,
1738 get_order(alias_table_size));
1739
1740 free_pages((unsigned long)amd_iommu_dev_table,
1741 get_order(dev_table_size));
1742
1743 free_iommu_all();
1744
1745#ifdef CONFIG_GART_IOMMU
1746 /*
1747 * We failed to initialize the AMD IOMMU - try fallback to GART
1748 * if possible.
1749 */
1750 gart_iommu_init();
1751
1752#endif
1753}
1754
1755/* SB IOAPIC is always on this device in AMD systems */
1756#define IOAPIC_SB_DEVID ((0x00 << 8) | PCI_DEVFN(0x14, 0))
1757
1758static bool __init check_ioapic_information(void)
1759{
1760 const char *fw_bug = FW_BUG;
1761 bool ret, has_sb_ioapic;
1762 int idx;
1763
1764 has_sb_ioapic = false;
1765 ret = false;
1766
1767 /*
1768 * If we have map overrides on the kernel command line the
1769 * messages in this function might not describe firmware bugs
1770 * anymore - so be careful
1771 */
1772 if (cmdline_maps)
1773 fw_bug = "";
1774
1775 for (idx = 0; idx < nr_ioapics; idx++) {
1776 int devid, id = mpc_ioapic_id(idx);
1777
1778 devid = get_ioapic_devid(id);
1779 if (devid < 0) {
1780 pr_err("%sAMD-Vi: IOAPIC[%d] not in IVRS table\n",
1781 fw_bug, id);
1782 ret = false;
1783 } else if (devid == IOAPIC_SB_DEVID) {
1784 has_sb_ioapic = true;
1785 ret = true;
1786 }
1787 }
1788
1789 if (!has_sb_ioapic) {
1790 /*
1791 * We expect the SB IOAPIC to be listed in the IVRS
1792 * table. The system timer is connected to the SB IOAPIC
1793 * and if we don't have it in the list the system will
1794 * panic at boot time. This situation usually happens
1795 * when the BIOS is buggy and provides us the wrong
1796 * device id for the IOAPIC in the system.
1797 */
1798 pr_err("%sAMD-Vi: No southbridge IOAPIC found\n", fw_bug);
1799 }
1800
1801 if (!ret)
1802 pr_err("AMD-Vi: Disabling interrupt remapping\n");
1803
1804 return ret;
1805}
1806
1807static void __init free_dma_resources(void)
1808{
1809 free_pages((unsigned long)amd_iommu_pd_alloc_bitmap,
1810 get_order(MAX_DOMAIN_ID/8));
1811
1812 free_unity_maps();
1813}
1814
1815/*
1816 * This is the hardware init function for AMD IOMMU in the system.
1817 * This function is called either from amd_iommu_init or from the interrupt
1818 * remapping setup code.
1819 *
1820 * This function basically parses the ACPI table for AMD IOMMU (IVRS)
1821 * three times:
1822 *
1823 * 1 pass) Find the highest PCI device id the driver has to handle.
1824 * Upon this information the size of the data structures is
1825 * determined that needs to be allocated.
1826 *
1827 * 2 pass) Initialize the data structures just allocated with the
1828 * information in the ACPI table about available AMD IOMMUs
1829 * in the system. It also maps the PCI devices in the
1830 * system to specific IOMMUs
1831 *
1832 * 3 pass) After the basic data structures are allocated and
1833 * initialized we update them with information about memory
1834 * remapping requirements parsed out of the ACPI table in
1835 * this last pass.
1836 *
1837 * After everything is set up the IOMMUs are enabled and the necessary
1838 * hotplug and suspend notifiers are registered.
1839 */
1840static int __init early_amd_iommu_init(void)
1841{
1842 struct acpi_table_header *ivrs_base;
1843 acpi_size ivrs_size;
1844 acpi_status status;
1845 int i, ret = 0;
1846
1847 if (!amd_iommu_detected)
1848 return -ENODEV;
1849
1850 status = acpi_get_table_with_size("IVRS", 0, &ivrs_base, &ivrs_size);
1851 if (status == AE_NOT_FOUND)
1852 return -ENODEV;
1853 else if (ACPI_FAILURE(status)) {
1854 const char *err = acpi_format_exception(status);
1855 pr_err("AMD-Vi: IVRS table error: %s\n", err);
1856 return -EINVAL;
1857 }
1858
1859 /*
1860 * First parse ACPI tables to find the largest Bus/Dev/Func
1861 * we need to handle. Upon this information the shared data
1862 * structures for the IOMMUs in the system will be allocated
1863 */
1864 ret = find_last_devid_acpi(ivrs_base);
1865 if (ret)
1866 goto out;
1867
1868 dev_table_size = tbl_size(DEV_TABLE_ENTRY_SIZE);
1869 alias_table_size = tbl_size(ALIAS_TABLE_ENTRY_SIZE);
1870 rlookup_table_size = tbl_size(RLOOKUP_TABLE_ENTRY_SIZE);
1871
1872 /* Device table - directly used by all IOMMUs */
1873 ret = -ENOMEM;
1874 amd_iommu_dev_table = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1875 get_order(dev_table_size));
1876 if (amd_iommu_dev_table == NULL)
1877 goto out;
1878
1879 /*
1880 * Alias table - map PCI Bus/Dev/Func to Bus/Dev/Func the
1881 * IOMMU see for that device
1882 */
1883 amd_iommu_alias_table = (void *)__get_free_pages(GFP_KERNEL,
1884 get_order(alias_table_size));
1885 if (amd_iommu_alias_table == NULL)
1886 goto out;
1887
1888 /* IOMMU rlookup table - find the IOMMU for a specific device */
1889 amd_iommu_rlookup_table = (void *)__get_free_pages(
1890 GFP_KERNEL | __GFP_ZERO,
1891 get_order(rlookup_table_size));
1892 if (amd_iommu_rlookup_table == NULL)
1893 goto out;
1894
1895 amd_iommu_pd_alloc_bitmap = (void *)__get_free_pages(
1896 GFP_KERNEL | __GFP_ZERO,
1897 get_order(MAX_DOMAIN_ID/8));
1898 if (amd_iommu_pd_alloc_bitmap == NULL)
1899 goto out;
1900
1901 /*
1902 * let all alias entries point to itself
1903 */
1904 for (i = 0; i <= amd_iommu_last_bdf; ++i)
1905 amd_iommu_alias_table[i] = i;
1906
1907 /*
1908 * never allocate domain 0 because its used as the non-allocated and
1909 * error value placeholder
1910 */
1911 amd_iommu_pd_alloc_bitmap[0] = 1;
1912
1913 spin_lock_init(&amd_iommu_pd_lock);
1914
1915 /*
1916 * now the data structures are allocated and basically initialized
1917 * start the real acpi table scan
1918 */
1919 ret = init_iommu_all(ivrs_base);
1920 if (ret)
1921 goto out;
1922
1923 if (amd_iommu_irq_remap)
1924 amd_iommu_irq_remap = check_ioapic_information();
1925
1926 if (amd_iommu_irq_remap) {
1927 /*
1928 * Interrupt remapping enabled, create kmem_cache for the
1929 * remapping tables.
1930 */
1931 ret = -ENOMEM;
1932 amd_iommu_irq_cache = kmem_cache_create("irq_remap_cache",
1933 MAX_IRQS_PER_TABLE * sizeof(u32),
1934 IRQ_TABLE_ALIGNMENT,
1935 0, NULL);
1936 if (!amd_iommu_irq_cache)
1937 goto out;
1938
1939 irq_lookup_table = (void *)__get_free_pages(
1940 GFP_KERNEL | __GFP_ZERO,
1941 get_order(rlookup_table_size));
1942 if (!irq_lookup_table)
1943 goto out;
1944 }
1945
1946 ret = init_memory_definitions(ivrs_base);
1947 if (ret)
1948 goto out;
1949
1950 /* init the device table */
1951 init_device_table();
1952
1953out:
1954 /* Don't leak any ACPI memory */
1955 early_acpi_os_unmap_memory((char __iomem *)ivrs_base, ivrs_size);
1956 ivrs_base = NULL;
1957
1958 return ret;
1959}
1960
1961static int amd_iommu_enable_interrupts(void)
1962{
1963 struct amd_iommu *iommu;
1964 int ret = 0;
1965
1966 for_each_iommu(iommu) {
1967 ret = iommu_init_msi(iommu);
1968 if (ret)
1969 goto out;
1970 }
1971
1972out:
1973 return ret;
1974}
1975
1976static bool detect_ivrs(void)
1977{
1978 struct acpi_table_header *ivrs_base;
1979 acpi_size ivrs_size;
1980 acpi_status status;
1981
1982 status = acpi_get_table_with_size("IVRS", 0, &ivrs_base, &ivrs_size);
1983 if (status == AE_NOT_FOUND)
1984 return false;
1985 else if (ACPI_FAILURE(status)) {
1986 const char *err = acpi_format_exception(status);
1987 pr_err("AMD-Vi: IVRS table error: %s\n", err);
1988 return false;
1989 }
1990
1991 early_acpi_os_unmap_memory((char __iomem *)ivrs_base, ivrs_size);
1992
1993 /* Make sure ACS will be enabled during PCI probe */
1994 pci_request_acs();
1995
1996 return true;
1997}
1998
1999/****************************************************************************
2000 *
2001 * AMD IOMMU Initialization State Machine
2002 *
2003 ****************************************************************************/
2004
2005static int __init state_next(void)
2006{
2007 int ret = 0;
2008
2009 switch (init_state) {
2010 case IOMMU_START_STATE:
2011 if (!detect_ivrs()) {
2012 init_state = IOMMU_NOT_FOUND;
2013 ret = -ENODEV;
2014 } else {
2015 init_state = IOMMU_IVRS_DETECTED;
2016 }
2017 break;
2018 case IOMMU_IVRS_DETECTED:
2019 ret = early_amd_iommu_init();
2020 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_ACPI_FINISHED;
2021 break;
2022 case IOMMU_ACPI_FINISHED:
2023 early_enable_iommus();
2024 register_syscore_ops(&amd_iommu_syscore_ops);
2025 x86_platform.iommu_shutdown = disable_iommus;
2026 init_state = IOMMU_ENABLED;
2027 break;
2028 case IOMMU_ENABLED:
2029 ret = amd_iommu_init_pci();
2030 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_PCI_INIT;
2031 enable_iommus_v2();
2032 break;
2033 case IOMMU_PCI_INIT:
2034 ret = amd_iommu_enable_interrupts();
2035 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_INTERRUPTS_EN;
2036 break;
2037 case IOMMU_INTERRUPTS_EN:
2038 ret = amd_iommu_init_dma_ops();
2039 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_DMA_OPS;
2040 break;
2041 case IOMMU_DMA_OPS:
2042 init_state = IOMMU_INITIALIZED;
2043 break;
2044 case IOMMU_INITIALIZED:
2045 /* Nothing to do */
2046 break;
2047 case IOMMU_NOT_FOUND:
2048 case IOMMU_INIT_ERROR:
2049 /* Error states => do nothing */
2050 ret = -EINVAL;
2051 break;
2052 default:
2053 /* Unknown state */
2054 BUG();
2055 }
2056
2057 return ret;
2058}
2059
2060static int __init iommu_go_to_state(enum iommu_init_state state)
2061{
2062 int ret = 0;
2063
2064 while (init_state != state) {
2065 ret = state_next();
2066 if (init_state == IOMMU_NOT_FOUND ||
2067 init_state == IOMMU_INIT_ERROR)
2068 break;
2069 }
2070
2071 return ret;
2072}
2073
2074#ifdef CONFIG_IRQ_REMAP
2075int __init amd_iommu_prepare(void)
2076{
2077 int ret;
2078
2079 amd_iommu_irq_remap = true;
2080
2081 ret = iommu_go_to_state(IOMMU_ACPI_FINISHED);
2082 if (ret)
2083 return ret;
2084 return amd_iommu_irq_remap ? 0 : -ENODEV;
2085}
2086
2087int __init amd_iommu_enable(void)
2088{
2089 int ret;
2090
2091 ret = iommu_go_to_state(IOMMU_ENABLED);
2092 if (ret)
2093 return ret;
2094
2095 irq_remapping_enabled = 1;
2096
2097 return 0;
2098}
2099
2100void amd_iommu_disable(void)
2101{
2102 amd_iommu_suspend();
2103}
2104
2105int amd_iommu_reenable(int mode)
2106{
2107 amd_iommu_resume();
2108
2109 return 0;
2110}
2111
2112int __init amd_iommu_enable_faulting(void)
2113{
2114 /* We enable MSI later when PCI is initialized */
2115 return 0;
2116}
2117#endif
2118
2119/*
2120 * This is the core init function for AMD IOMMU hardware in the system.
2121 * This function is called from the generic x86 DMA layer initialization
2122 * code.
2123 */
2124static int __init amd_iommu_init(void)
2125{
2126 int ret;
2127
2128 ret = iommu_go_to_state(IOMMU_INITIALIZED);
2129 if (ret) {
2130 free_dma_resources();
2131 if (!irq_remapping_enabled) {
2132 disable_iommus();
2133 free_on_init_error();
2134 } else {
2135 struct amd_iommu *iommu;
2136
2137 uninit_device_table_dma();
2138 for_each_iommu(iommu)
2139 iommu_flush_all_caches(iommu);
2140 }
2141 }
2142
2143 return ret;
2144}
2145
2146/****************************************************************************
2147 *
2148 * Early detect code. This code runs at IOMMU detection time in the DMA
2149 * layer. It just looks if there is an IVRS ACPI table to detect AMD
2150 * IOMMUs
2151 *
2152 ****************************************************************************/
2153int __init amd_iommu_detect(void)
2154{
2155 int ret;
2156
2157 if (no_iommu || (iommu_detected && !gart_iommu_aperture))
2158 return -ENODEV;
2159
2160 if (amd_iommu_disabled)
2161 return -ENODEV;
2162
2163 ret = iommu_go_to_state(IOMMU_IVRS_DETECTED);
2164 if (ret)
2165 return ret;
2166
2167 amd_iommu_detected = true;
2168 iommu_detected = 1;
2169 x86_init.iommu.iommu_init = amd_iommu_init;
2170
2171 return 1;
2172}
2173
2174/****************************************************************************
2175 *
2176 * Parsing functions for the AMD IOMMU specific kernel command line
2177 * options.
2178 *
2179 ****************************************************************************/
2180
2181static int __init parse_amd_iommu_dump(char *str)
2182{
2183 amd_iommu_dump = true;
2184
2185 return 1;
2186}
2187
2188static int __init parse_amd_iommu_options(char *str)
2189{
2190 for (; *str; ++str) {
2191 if (strncmp(str, "fullflush", 9) == 0)
2192 amd_iommu_unmap_flush = true;
2193 if (strncmp(str, "off", 3) == 0)
2194 amd_iommu_disabled = true;
2195 if (strncmp(str, "force_isolation", 15) == 0)
2196 amd_iommu_force_isolation = true;
2197 }
2198
2199 return 1;
2200}
2201
2202static int __init parse_ivrs_ioapic(char *str)
2203{
2204 unsigned int bus, dev, fn;
2205 int ret, id, i;
2206 u16 devid;
2207
2208 ret = sscanf(str, "[%d]=%x:%x.%x", &id, &bus, &dev, &fn);
2209
2210 if (ret != 4) {
2211 pr_err("AMD-Vi: Invalid command line: ivrs_ioapic%s\n", str);
2212 return 1;
2213 }
2214
2215 if (early_ioapic_map_size == EARLY_MAP_SIZE) {
2216 pr_err("AMD-Vi: Early IOAPIC map overflow - ignoring ivrs_ioapic%s\n",
2217 str);
2218 return 1;
2219 }
2220
2221 devid = ((bus & 0xff) << 8) | ((dev & 0x1f) << 3) | (fn & 0x7);
2222
2223 cmdline_maps = true;
2224 i = early_ioapic_map_size++;
2225 early_ioapic_map[i].id = id;
2226 early_ioapic_map[i].devid = devid;
2227 early_ioapic_map[i].cmd_line = true;
2228
2229 return 1;
2230}
2231
2232static int __init parse_ivrs_hpet(char *str)
2233{
2234 unsigned int bus, dev, fn;
2235 int ret, id, i;
2236 u16 devid;
2237
2238 ret = sscanf(str, "[%d]=%x:%x.%x", &id, &bus, &dev, &fn);
2239
2240 if (ret != 4) {
2241 pr_err("AMD-Vi: Invalid command line: ivrs_hpet%s\n", str);
2242 return 1;
2243 }
2244
2245 if (early_hpet_map_size == EARLY_MAP_SIZE) {
2246 pr_err("AMD-Vi: Early HPET map overflow - ignoring ivrs_hpet%s\n",
2247 str);
2248 return 1;
2249 }
2250
2251 devid = ((bus & 0xff) << 8) | ((dev & 0x1f) << 3) | (fn & 0x7);
2252
2253 cmdline_maps = true;
2254 i = early_hpet_map_size++;
2255 early_hpet_map[i].id = id;
2256 early_hpet_map[i].devid = devid;
2257 early_hpet_map[i].cmd_line = true;
2258
2259 return 1;
2260}
2261
2262__setup("amd_iommu_dump", parse_amd_iommu_dump);
2263__setup("amd_iommu=", parse_amd_iommu_options);
2264__setup("ivrs_ioapic", parse_ivrs_ioapic);
2265__setup("ivrs_hpet", parse_ivrs_hpet);
2266
2267IOMMU_INIT_FINISH(amd_iommu_detect,
2268 gart_iommu_hole_init,
2269 NULL,
2270 NULL);
2271
2272bool amd_iommu_v2_supported(void)
2273{
2274 return amd_iommu_v2_present;
2275}
2276EXPORT_SYMBOL(amd_iommu_v2_supported);
2277
2278/****************************************************************************
2279 *
2280 * IOMMU EFR Performance Counter support functionality. This code allows
2281 * access to the IOMMU PC functionality.
2282 *
2283 ****************************************************************************/
2284
2285u8 amd_iommu_pc_get_max_banks(u16 devid)
2286{
2287 struct amd_iommu *iommu;
2288 u8 ret = 0;
2289
2290 /* locate the iommu governing the devid */
2291 iommu = amd_iommu_rlookup_table[devid];
2292 if (iommu)
2293 ret = iommu->max_banks;
2294
2295 return ret;
2296}
2297EXPORT_SYMBOL(amd_iommu_pc_get_max_banks);
2298
2299bool amd_iommu_pc_supported(void)
2300{
2301 return amd_iommu_pc_present;
2302}
2303EXPORT_SYMBOL(amd_iommu_pc_supported);
2304
2305u8 amd_iommu_pc_get_max_counters(u16 devid)
2306{
2307 struct amd_iommu *iommu;
2308 u8 ret = 0;
2309
2310 /* locate the iommu governing the devid */
2311 iommu = amd_iommu_rlookup_table[devid];
2312 if (iommu)
2313 ret = iommu->max_counters;
2314
2315 return ret;
2316}
2317EXPORT_SYMBOL(amd_iommu_pc_get_max_counters);
2318
2319static int iommu_pc_get_set_reg_val(struct amd_iommu *iommu,
2320 u8 bank, u8 cntr, u8 fxn,
2321 u64 *value, bool is_write)
2322{
2323 u32 offset;
2324 u32 max_offset_lim;
2325
2326 /* Check for valid iommu and pc register indexing */
2327 if (WARN_ON((fxn > 0x28) || (fxn & 7)))
2328 return -ENODEV;
2329
2330 offset = (u32)(((0x40|bank) << 12) | (cntr << 8) | fxn);
2331
2332 /* Limit the offset to the hw defined mmio region aperture */
2333 max_offset_lim = (u32)(((0x40|iommu->max_banks) << 12) |
2334 (iommu->max_counters << 8) | 0x28);
2335 if ((offset < MMIO_CNTR_REG_OFFSET) ||
2336 (offset > max_offset_lim))
2337 return -EINVAL;
2338
2339 if (is_write) {
2340 writel((u32)*value, iommu->mmio_base + offset);
2341 writel((*value >> 32), iommu->mmio_base + offset + 4);
2342 } else {
2343 *value = readl(iommu->mmio_base + offset + 4);
2344 *value <<= 32;
2345 *value = readl(iommu->mmio_base + offset);
2346 }
2347
2348 return 0;
2349}
2350EXPORT_SYMBOL(amd_iommu_pc_get_set_reg_val);
2351
2352int amd_iommu_pc_get_set_reg_val(u16 devid, u8 bank, u8 cntr, u8 fxn,
2353 u64 *value, bool is_write)
2354{
2355 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
2356
2357 /* Make sure the IOMMU PC resource is available */
2358 if (!amd_iommu_pc_present || iommu == NULL)
2359 return -ENODEV;
2360
2361 return iommu_pc_get_set_reg_val(iommu, bank, cntr, fxn,
2362 value, is_write);
2363}