?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
of: MSI: Simplify irqdomain lookup
[linux/fpc-iii.git] / drivers / iommu / intel-iommu.c
blobac7387686ddc7b2a7c7757f2cb3fbd003c8a23af
1 /*
2 * Copyright © 2006-2014 Intel Corporation.
3 *
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
11 * more details.
12 *
13 * Authors: David Woodhouse <dwmw2@infradead.org>,
14 * Ashok Raj <ashok.raj@intel.com>,
15 * Shaohua Li <shaohua.li@intel.com>,
16 * Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>,
17 * Fenghua Yu <fenghua.yu@intel.com>
18 * Joerg Roedel <jroedel@suse.de>
19 */
20
21 #define pr_fmt(fmt) "DMAR: " fmt
22
23 #include <linux/init.h>
24 #include <linux/bitmap.h>
25 #include <linux/debugfs.h>
26 #include <linux/export.h>
27 #include <linux/slab.h>
28 #include <linux/irq.h>
29 #include <linux/interrupt.h>
30 #include <linux/spinlock.h>
31 #include <linux/pci.h>
32 #include <linux/dmar.h>
33 #include <linux/dma-mapping.h>
34 #include <linux/mempool.h>
35 #include <linux/memory.h>
36 #include <linux/timer.h>
37 #include <linux/io.h>
38 #include <linux/iova.h>
39 #include <linux/iommu.h>
40 #include <linux/intel-iommu.h>
41 #include <linux/syscore_ops.h>
42 #include <linux/tboot.h>
43 #include <linux/dmi.h>
44 #include <linux/pci-ats.h>
45 #include <linux/memblock.h>
46 #include <linux/dma-contiguous.h>
47 #include <linux/crash_dump.h>
48 #include <asm/irq_remapping.h>
49 #include <asm/cacheflush.h>
50 #include <asm/iommu.h>
51
52 #include "irq_remapping.h"
53
54 #define ROOT_SIZE VTD_PAGE_SIZE
55 #define CONTEXT_SIZE VTD_PAGE_SIZE
56
57 #define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
58 #define IS_USB_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_SERIAL_USB)
59 #define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
60 #define IS_AZALIA(pdev) ((pdev)->vendor == 0x8086 && (pdev)->device == 0x3a3e)
61
62 #define IOAPIC_RANGE_START (0xfee00000)
63 #define IOAPIC_RANGE_END (0xfeefffff)
64 #define IOVA_START_ADDR (0x1000)
65
66 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
67
68 #define MAX_AGAW_WIDTH 64
69 #define MAX_AGAW_PFN_WIDTH (MAX_AGAW_WIDTH - VTD_PAGE_SHIFT)
70
71 #define __DOMAIN_MAX_PFN(gaw) ((((uint64_t)1) << (gaw-VTD_PAGE_SHIFT)) - 1)
72 #define __DOMAIN_MAX_ADDR(gaw) ((((uint64_t)1) << gaw) - 1)
73
74 /* We limit DOMAIN_MAX_PFN to fit in an unsigned long, and DOMAIN_MAX_ADDR
75 to match. That way, we can use 'unsigned long' for PFNs with impunity. */
76 #define DOMAIN_MAX_PFN(gaw) ((unsigned long) min_t(uint64_t, \
77 __DOMAIN_MAX_PFN(gaw), (unsigned long)-1))
78 #define DOMAIN_MAX_ADDR(gaw) (((uint64_t)__DOMAIN_MAX_PFN(gaw)) << VTD_PAGE_SHIFT)
79
80 /* IO virtual address start page frame number */
81 #define IOVA_START_PFN (1)
82
83 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
84 #define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
85 #define DMA_64BIT_PFN IOVA_PFN(DMA_BIT_MASK(64))
86
87 /* page table handling */
88 #define LEVEL_STRIDE (9)
89 #define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
90
91 /*
92 * This bitmap is used to advertise the page sizes our hardware support
93 * to the IOMMU core, which will then use this information to split
94 * physically contiguous memory regions it is mapping into page sizes
95 * that we support.
96 *
97 * Traditionally the IOMMU core just handed us the mappings directly,
98 * after making sure the size is an order of a 4KiB page and that the
99 * mapping has natural alignment.
100 *
101 * To retain this behavior, we currently advertise that we support
102 * all page sizes that are an order of 4KiB.
103 *
104 * If at some point we'd like to utilize the IOMMU core's new behavior,
105 * we could change this to advertise the real page sizes we support.
106 */
107 #define INTEL_IOMMU_PGSIZES (~0xFFFUL)
108
109 static inline int agaw_to_level(int agaw)
110 {
111 return agaw + 2;
112 }
113
114 static inline int agaw_to_width(int agaw)
115 {
116 return min_t(int, 30 + agaw * LEVEL_STRIDE, MAX_AGAW_WIDTH);
117 }
118
119 static inline int width_to_agaw(int width)
120 {
121 return DIV_ROUND_UP(width - 30, LEVEL_STRIDE);
122 }
123
124 static inline unsigned int level_to_offset_bits(int level)
125 {
126 return (level - 1) * LEVEL_STRIDE;
127 }
128
129 static inline int pfn_level_offset(unsigned long pfn, int level)
130 {
131 return (pfn >> level_to_offset_bits(level)) & LEVEL_MASK;
132 }
133
134 static inline unsigned long level_mask(int level)
135 {
136 return -1UL << level_to_offset_bits(level);
137 }
138
139 static inline unsigned long level_size(int level)
140 {
141 return 1UL << level_to_offset_bits(level);
142 }
143
144 static inline unsigned long align_to_level(unsigned long pfn, int level)
145 {
146 return (pfn + level_size(level) - 1) & level_mask(level);
147 }
148
149 static inline unsigned long lvl_to_nr_pages(unsigned int lvl)
150 {
151 return 1 << min_t(int, (lvl - 1) * LEVEL_STRIDE, MAX_AGAW_PFN_WIDTH);
152 }
153
154 /* VT-d pages must always be _smaller_ than MM pages. Otherwise things
155 are never going to work. */
156 static inline unsigned long dma_to_mm_pfn(unsigned long dma_pfn)
157 {
158 return dma_pfn >> (PAGE_SHIFT - VTD_PAGE_SHIFT);
159 }
160
161 static inline unsigned long mm_to_dma_pfn(unsigned long mm_pfn)
162 {
163 return mm_pfn << (PAGE_SHIFT - VTD_PAGE_SHIFT);
164 }
165 static inline unsigned long page_to_dma_pfn(struct page *pg)
166 {
167 return mm_to_dma_pfn(page_to_pfn(pg));
168 }
169 static inline unsigned long virt_to_dma_pfn(void *p)
170 {
171 return page_to_dma_pfn(virt_to_page(p));
172 }
173
174 /* global iommu list, set NULL for ignored DMAR units */
175 static struct intel_iommu **g_iommus;
176
177 static void __init check_tylersburg_isoch(void);
178 static int rwbf_quirk;
179
180 /*
181 * set to 1 to panic kernel if can't successfully enable VT-d
182 * (used when kernel is launched w/ TXT)
183 */
184 static int force_on = 0;
185
186 /*
187 * 0: Present
188 * 1-11: Reserved
189 * 12-63: Context Ptr (12 - (haw-1))
190 * 64-127: Reserved
191 */
192 struct root_entry {
193 u64 lo;
194 u64 hi;
195 };
196 #define ROOT_ENTRY_NR (VTD_PAGE_SIZE/sizeof(struct root_entry))
197
198 /*
199 * Take a root_entry and return the Lower Context Table Pointer (LCTP)
200 * if marked present.
201 */
202 static phys_addr_t root_entry_lctp(struct root_entry *re)
203 {
204 if (!(re->lo & 1))
205 return 0;
206
207 return re->lo & VTD_PAGE_MASK;
208 }
209
210 /*
211 * Take a root_entry and return the Upper Context Table Pointer (UCTP)
212 * if marked present.
213 */
214 static phys_addr_t root_entry_uctp(struct root_entry *re)
215 {
216 if (!(re->hi & 1))
217 return 0;
218
219 return re->hi & VTD_PAGE_MASK;
220 }
221 /*
222 * low 64 bits:
223 * 0: present
224 * 1: fault processing disable
225 * 2-3: translation type
226 * 12-63: address space root
227 * high 64 bits:
228 * 0-2: address width
229 * 3-6: aval
230 * 8-23: domain id
231 */
232 struct context_entry {
233 u64 lo;
234 u64 hi;
235 };
236
237 static inline void context_clear_pasid_enable(struct context_entry *context)
238 {
239 context->lo &= ~(1ULL << 11);
240 }
241
242 static inline bool context_pasid_enabled(struct context_entry *context)
243 {
244 return !!(context->lo & (1ULL << 11));
245 }
246
247 static inline void context_set_copied(struct context_entry *context)
248 {
249 context->hi |= (1ull << 3);
250 }
251
252 static inline bool context_copied(struct context_entry *context)
253 {
254 return !!(context->hi & (1ULL << 3));
255 }
256
257 static inline bool __context_present(struct context_entry *context)
258 {
259 return (context->lo & 1);
260 }
261
262 static inline bool context_present(struct context_entry *context)
263 {
264 return context_pasid_enabled(context) ?
265 __context_present(context) :
266 __context_present(context) && !context_copied(context);
267 }
268
269 static inline void context_set_present(struct context_entry *context)
270 {
271 context->lo |= 1;
272 }
273
274 static inline void context_set_fault_enable(struct context_entry *context)
275 {
276 context->lo &= (((u64)-1) << 2) | 1;
277 }
278
279 static inline void context_set_translation_type(struct context_entry *context,
280 unsigned long value)
281 {
282 context->lo &= (((u64)-1) << 4) | 3;
283 context->lo |= (value & 3) << 2;
284 }
285
286 static inline void context_set_address_root(struct context_entry *context,
287 unsigned long value)
288 {
289 context->lo &= ~VTD_PAGE_MASK;
290 context->lo |= value & VTD_PAGE_MASK;
291 }
292
293 static inline void context_set_address_width(struct context_entry *context,
294 unsigned long value)
295 {
296 context->hi |= value & 7;
297 }
298
299 static inline void context_set_domain_id(struct context_entry *context,
300 unsigned long value)
301 {
302 context->hi |= (value & ((1 << 16) - 1)) << 8;
303 }
304
305 static inline int context_domain_id(struct context_entry *c)
306 {
307 return((c->hi >> 8) & 0xffff);
308 }
309
310 static inline void context_clear_entry(struct context_entry *context)
311 {
312 context->lo = 0;
313 context->hi = 0;
314 }
315
316 /*
317 * 0: readable
318 * 1: writable
319 * 2-6: reserved
320 * 7: super page
321 * 8-10: available
322 * 11: snoop behavior
323 * 12-63: Host physcial address
324 */
325 struct dma_pte {
326 u64 val;
327 };
328
329 static inline void dma_clear_pte(struct dma_pte *pte)
330 {
331 pte->val = 0;
332 }
333
334 static inline u64 dma_pte_addr(struct dma_pte *pte)
335 {
336 #ifdef CONFIG_64BIT
337 return pte->val & VTD_PAGE_MASK;
338 #else
339 /* Must have a full atomic 64-bit read */
340 return __cmpxchg64(&pte->val, 0ULL, 0ULL) & VTD_PAGE_MASK;
341 #endif
342 }
343
344 static inline bool dma_pte_present(struct dma_pte *pte)
345 {
346 return (pte->val & 3) != 0;
347 }
348
349 static inline bool dma_pte_superpage(struct dma_pte *pte)
350 {
351 return (pte->val & DMA_PTE_LARGE_PAGE);
352 }
353
354 static inline int first_pte_in_page(struct dma_pte *pte)
355 {
356 return !((unsigned long)pte & ~VTD_PAGE_MASK);
357 }
358
359 /*
360 * This domain is a statically identity mapping domain.
361 * 1. This domain creats a static 1:1 mapping to all usable memory.
362 * 2. It maps to each iommu if successful.
363 * 3. Each iommu mapps to this domain if successful.
364 */
365 static struct dmar_domain *si_domain;
366 static int hw_pass_through = 1;
367
368 /*
369 * Domain represents a virtual machine, more than one devices
370 * across iommus may be owned in one domain, e.g. kvm guest.
371 */
372 #define DOMAIN_FLAG_VIRTUAL_MACHINE (1 << 0)
373
374 /* si_domain contains mulitple devices */
375 #define DOMAIN_FLAG_STATIC_IDENTITY (1 << 1)
376
377 #define for_each_domain_iommu(idx, domain) \
378 for (idx = 0; idx < g_num_of_iommus; idx++) \
379 if (domain->iommu_refcnt[idx])
380
381 struct dmar_domain {
382 int nid; /* node id */
383
384 unsigned iommu_refcnt[DMAR_UNITS_SUPPORTED];
385 /* Refcount of devices per iommu */
386
387
388 u16 iommu_did[DMAR_UNITS_SUPPORTED];
389 /* Domain ids per IOMMU. Use u16 since
390 * domain ids are 16 bit wide according
391 * to VT-d spec, section 9.3 */
392
393 struct list_head devices; /* all devices' list */
394 struct iova_domain iovad; /* iova's that belong to this domain */
395
396 struct dma_pte *pgd; /* virtual address */
397 int gaw; /* max guest address width */
398
399 /* adjusted guest address width, 0 is level 2 30-bit */
400 int agaw;
401
402 int flags; /* flags to find out type of domain */
403
404 int iommu_coherency;/* indicate coherency of iommu access */
405 int iommu_snooping; /* indicate snooping control feature*/
406 int iommu_count; /* reference count of iommu */
407 int iommu_superpage;/* Level of superpages supported:
408 0 == 4KiB (no superpages), 1 == 2MiB,
409 2 == 1GiB, 3 == 512GiB, 4 == 1TiB */
410 u64 max_addr; /* maximum mapped address */
411
412 struct iommu_domain domain; /* generic domain data structure for
413 iommu core */
414 };
415
416 /* PCI domain-device relationship */
417 struct device_domain_info {
418 struct list_head link; /* link to domain siblings */
419 struct list_head global; /* link to global list */
420 u8 bus; /* PCI bus number */
421 u8 devfn; /* PCI devfn number */
422 u8 pasid_supported:3;
423 u8 pasid_enabled:1;
424 u8 pri_supported:1;
425 u8 pri_enabled:1;
426 u8 ats_supported:1;
427 u8 ats_enabled:1;
428 u8 ats_qdep;
429 struct device *dev; /* it's NULL for PCIe-to-PCI bridge */
430 struct intel_iommu *iommu; /* IOMMU used by this device */
431 struct dmar_domain *domain; /* pointer to domain */
432 };
433
434 struct dmar_rmrr_unit {
435 struct list_head list; /* list of rmrr units */
436 struct acpi_dmar_header *hdr; /* ACPI header */
437 u64 base_address; /* reserved base address*/
438 u64 end_address; /* reserved end address */
439 struct dmar_dev_scope *devices; /* target devices */
440 int devices_cnt; /* target device count */
441 };
442
443 struct dmar_atsr_unit {
444 struct list_head list; /* list of ATSR units */
445 struct acpi_dmar_header *hdr; /* ACPI header */
446 struct dmar_dev_scope *devices; /* target devices */
447 int devices_cnt; /* target device count */
448 u8 include_all:1; /* include all ports */
449 };
450
451 static LIST_HEAD(dmar_atsr_units);
452 static LIST_HEAD(dmar_rmrr_units);
453
454 #define for_each_rmrr_units(rmrr) \
455 list_for_each_entry(rmrr, &dmar_rmrr_units, list)
456
457 static void flush_unmaps_timeout(unsigned long data);
458
459 static DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
460
461 #define HIGH_WATER_MARK 250
462 struct deferred_flush_tables {
463 int next;
464 struct iova *iova[HIGH_WATER_MARK];
465 struct dmar_domain *domain[HIGH_WATER_MARK];
466 struct page *freelist[HIGH_WATER_MARK];
467 };
468
469 static struct deferred_flush_tables *deferred_flush;
470
471 /* bitmap for indexing intel_iommus */
472 static int g_num_of_iommus;
473
474 static DEFINE_SPINLOCK(async_umap_flush_lock);
475 static LIST_HEAD(unmaps_to_do);
476
477 static int timer_on;
478 static long list_size;
479
480 static void domain_exit(struct dmar_domain *domain);
481 static void domain_remove_dev_info(struct dmar_domain *domain);
482 static void dmar_remove_one_dev_info(struct dmar_domain *domain,
483 struct device *dev);
484 static void __dmar_remove_one_dev_info(struct device_domain_info *info);
485 static void domain_context_clear(struct intel_iommu *iommu,
486 struct device *dev);
487 static int domain_detach_iommu(struct dmar_domain *domain,
488 struct intel_iommu *iommu);
489
490 #ifdef CONFIG_INTEL_IOMMU_DEFAULT_ON
491 int dmar_disabled = 0;
492 #else
493 int dmar_disabled = 1;
494 #endif /*CONFIG_INTEL_IOMMU_DEFAULT_ON*/
495
496 int intel_iommu_enabled = 0;
497 EXPORT_SYMBOL_GPL(intel_iommu_enabled);
498
499 static int dmar_map_gfx = 1;
500 static int dmar_forcedac;
501 static int intel_iommu_strict;
502 static int intel_iommu_superpage = 1;
503 static int intel_iommu_ecs = 1;
504 static int intel_iommu_pasid28;
505 static int iommu_identity_mapping;
506
507 #define IDENTMAP_ALL 1
508 #define IDENTMAP_GFX 2
509 #define IDENTMAP_AZALIA 4
510
511 /* Broadwell and Skylake have broken ECS support — normal so-called "second
512 * level" translation of DMA requests-without-PASID doesn't actually happen
513 * unless you also set the NESTE bit in an extended context-entry. Which of
514 * course means that SVM doesn't work because it's trying to do nested
515 * translation of the physical addresses it finds in the process page tables,
516 * through the IOVA->phys mapping found in the "second level" page tables.
517 *
518 * The VT-d specification was retroactively changed to change the definition
519 * of the capability bits and pretend that Broadwell/Skylake never happened...
520 * but unfortunately the wrong bit was changed. It's ECS which is broken, but
521 * for some reason it was the PASID capability bit which was redefined (from
522 * bit 28 on BDW/SKL to bit 40 in future).
523 *
524 * So our test for ECS needs to eschew those implementations which set the old
525 * PASID capabiity bit 28, since those are the ones on which ECS is broken.
526 * Unless we are working around the 'pasid28' limitations, that is, by putting
527 * the device into passthrough mode for normal DMA and thus masking the bug.
528 */
529 #define ecs_enabled(iommu) (intel_iommu_ecs && ecap_ecs(iommu->ecap) && \
530 (intel_iommu_pasid28 || !ecap_broken_pasid(iommu->ecap)))
531 /* PASID support is thus enabled if ECS is enabled and *either* of the old
532 * or new capability bits are set. */
533 #define pasid_enabled(iommu) (ecs_enabled(iommu) && \
534 (ecap_pasid(iommu->ecap) || ecap_broken_pasid(iommu->ecap)))
535
536 int intel_iommu_gfx_mapped;
537 EXPORT_SYMBOL_GPL(intel_iommu_gfx_mapped);
538
539 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
540 static DEFINE_SPINLOCK(device_domain_lock);
541 static LIST_HEAD(device_domain_list);
542
543 static const struct iommu_ops intel_iommu_ops;
544
545 static bool translation_pre_enabled(struct intel_iommu *iommu)
546 {
547 return (iommu->flags & VTD_FLAG_TRANS_PRE_ENABLED);
548 }
549
550 static void clear_translation_pre_enabled(struct intel_iommu *iommu)
551 {
552 iommu->flags &= ~VTD_FLAG_TRANS_PRE_ENABLED;
553 }
554
555 static void init_translation_status(struct intel_iommu *iommu)
556 {
557 u32 gsts;
558
559 gsts = readl(iommu->reg + DMAR_GSTS_REG);
560 if (gsts & DMA_GSTS_TES)
561 iommu->flags |= VTD_FLAG_TRANS_PRE_ENABLED;
562 }
563
564 /* Convert generic 'struct iommu_domain to private struct dmar_domain */
565 static struct dmar_domain *to_dmar_domain(struct iommu_domain *dom)
566 {
567 return container_of(dom, struct dmar_domain, domain);
568 }
569
570 static int __init intel_iommu_setup(char *str)
571 {
572 if (!str)
573 return -EINVAL;
574 while (*str) {
575 if (!strncmp(str, "on", 2)) {
576 dmar_disabled = 0;
577 pr_info("IOMMU enabled\n");
578 } else if (!strncmp(str, "off", 3)) {
579 dmar_disabled = 1;
580 pr_info("IOMMU disabled\n");
581 } else if (!strncmp(str, "igfx_off", 8)) {
582 dmar_map_gfx = 0;
583 pr_info("Disable GFX device mapping\n");
584 } else if (!strncmp(str, "forcedac", 8)) {
585 pr_info("Forcing DAC for PCI devices\n");
586 dmar_forcedac = 1;
587 } else if (!strncmp(str, "strict", 6)) {
588 pr_info("Disable batched IOTLB flush\n");
589 intel_iommu_strict = 1;
590 } else if (!strncmp(str, "sp_off", 6)) {
591 pr_info("Disable supported super page\n");
592 intel_iommu_superpage = 0;
593 } else if (!strncmp(str, "ecs_off", 7)) {
594 printk(KERN_INFO
595 "Intel-IOMMU: disable extended context table support\n");
596 intel_iommu_ecs = 0;
597 } else if (!strncmp(str, "pasid28", 7)) {
598 printk(KERN_INFO
599 "Intel-IOMMU: enable pre-production PASID support\n");
600 intel_iommu_pasid28 = 1;
601 iommu_identity_mapping |= IDENTMAP_GFX;
602 }
603
604 str += strcspn(str, ",");
605 while (*str == ',')
606 str++;
607 }
608 return 0;
609 }
610 __setup("intel_iommu=", intel_iommu_setup);
611
612 static struct kmem_cache *iommu_domain_cache;
613 static struct kmem_cache *iommu_devinfo_cache;
614
615 static struct dmar_domain* get_iommu_domain(struct intel_iommu *iommu, u16 did)
616 {
617 struct dmar_domain **domains;
618 int idx = did >> 8;
619
620 domains = iommu->domains[idx];
621 if (!domains)
622 return NULL;
623
624 return domains[did & 0xff];
625 }
626
627 static void set_iommu_domain(struct intel_iommu *iommu, u16 did,
628 struct dmar_domain *domain)
629 {
630 struct dmar_domain **domains;
631 int idx = did >> 8;
632
633 if (!iommu->domains[idx]) {
634 size_t size = 256 * sizeof(struct dmar_domain *);
635 iommu->domains[idx] = kzalloc(size, GFP_ATOMIC);
636 }
637
638 domains = iommu->domains[idx];
639 if (WARN_ON(!domains))
640 return;
641 else
642 domains[did & 0xff] = domain;
643 }
644
645 static inline void *alloc_pgtable_page(int node)
646 {
647 struct page *page;
648 void *vaddr = NULL;
649
650 page = alloc_pages_node(node, GFP_ATOMIC | __GFP_ZERO, 0);
651 if (page)
652 vaddr = page_address(page);
653 return vaddr;
654 }
655
656 static inline void free_pgtable_page(void *vaddr)
657 {
658 free_page((unsigned long)vaddr);
659 }
660
661 static inline void *alloc_domain_mem(void)
662 {
663 return kmem_cache_alloc(iommu_domain_cache, GFP_ATOMIC);
664 }
665
666 static void free_domain_mem(void *vaddr)
667 {
668 kmem_cache_free(iommu_domain_cache, vaddr);
669 }
670
671 static inline void * alloc_devinfo_mem(void)
672 {
673 return kmem_cache_alloc(iommu_devinfo_cache, GFP_ATOMIC);
674 }
675
676 static inline void free_devinfo_mem(void *vaddr)
677 {
678 kmem_cache_free(iommu_devinfo_cache, vaddr);
679 }
680
681 static inline int domain_type_is_vm(struct dmar_domain *domain)
682 {
683 return domain->flags & DOMAIN_FLAG_VIRTUAL_MACHINE;
684 }
685
686 static inline int domain_type_is_si(struct dmar_domain *domain)
687 {
688 return domain->flags & DOMAIN_FLAG_STATIC_IDENTITY;
689 }
690
691 static inline int domain_type_is_vm_or_si(struct dmar_domain *domain)
692 {
693 return domain->flags & (DOMAIN_FLAG_VIRTUAL_MACHINE |
694 DOMAIN_FLAG_STATIC_IDENTITY);
695 }
696
697 static inline int domain_pfn_supported(struct dmar_domain *domain,
698 unsigned long pfn)
699 {
700 int addr_width = agaw_to_width(domain->agaw) - VTD_PAGE_SHIFT;
701
702 return !(addr_width < BITS_PER_LONG && pfn >> addr_width);
703 }
704
705 static int __iommu_calculate_agaw(struct intel_iommu *iommu, int max_gaw)
706 {
707 unsigned long sagaw;
708 int agaw = -1;
709
710 sagaw = cap_sagaw(iommu->cap);
711 for (agaw = width_to_agaw(max_gaw);
712 agaw >= 0; agaw--) {
713 if (test_bit(agaw, &sagaw))
714 break;
715 }
716
717 return agaw;
718 }
719
720 /*
721 * Calculate max SAGAW for each iommu.
722 */
723 int iommu_calculate_max_sagaw(struct intel_iommu *iommu)
724 {
725 return __iommu_calculate_agaw(iommu, MAX_AGAW_WIDTH);
726 }
727
728 /*
729 * calculate agaw for each iommu.
730 * "SAGAW" may be different across iommus, use a default agaw, and
731 * get a supported less agaw for iommus that don't support the default agaw.
732 */
733 int iommu_calculate_agaw(struct intel_iommu *iommu)
734 {
735 return __iommu_calculate_agaw(iommu, DEFAULT_DOMAIN_ADDRESS_WIDTH);
736 }
737
738 /* This functionin only returns single iommu in a domain */
739 static struct intel_iommu *domain_get_iommu(struct dmar_domain *domain)
740 {
741 int iommu_id;
742
743 /* si_domain and vm domain should not get here. */
744 BUG_ON(domain_type_is_vm_or_si(domain));
745 for_each_domain_iommu(iommu_id, domain)
746 break;
747
748 if (iommu_id < 0 || iommu_id >= g_num_of_iommus)
749 return NULL;
750
751 return g_iommus[iommu_id];
752 }
753
754 static void domain_update_iommu_coherency(struct dmar_domain *domain)
755 {
756 struct dmar_drhd_unit *drhd;
757 struct intel_iommu *iommu;
758 bool found = false;
759 int i;
760
761 domain->iommu_coherency = 1;
762
763 for_each_domain_iommu(i, domain) {
764 found = true;
765 if (!ecap_coherent(g_iommus[i]->ecap)) {
766 domain->iommu_coherency = 0;
767 break;
768 }
769 }
770 if (found)
771 return;
772
773 /* No hardware attached; use lowest common denominator */
774 rcu_read_lock();
775 for_each_active_iommu(iommu, drhd) {
776 if (!ecap_coherent(iommu->ecap)) {
777 domain->iommu_coherency = 0;
778 break;
779 }
780 }
781 rcu_read_unlock();
782 }
783
784 static int domain_update_iommu_snooping(struct intel_iommu *skip)
785 {
786 struct dmar_drhd_unit *drhd;
787 struct intel_iommu *iommu;
788 int ret = 1;
789
790 rcu_read_lock();
791 for_each_active_iommu(iommu, drhd) {
792 if (iommu != skip) {
793 if (!ecap_sc_support(iommu->ecap)) {
794 ret = 0;
795 break;
796 }
797 }
798 }
799 rcu_read_unlock();
800
801 return ret;
802 }
803
804 static int domain_update_iommu_superpage(struct intel_iommu *skip)
805 {
806 struct dmar_drhd_unit *drhd;
807 struct intel_iommu *iommu;
808 int mask = 0xf;
809
810 if (!intel_iommu_superpage) {
811 return 0;
812 }
813
814 /* set iommu_superpage to the smallest common denominator */
815 rcu_read_lock();
816 for_each_active_iommu(iommu, drhd) {
817 if (iommu != skip) {
818 mask &= cap_super_page_val(iommu->cap);
819 if (!mask)
820 break;
821 }
822 }
823 rcu_read_unlock();
824
825 return fls(mask);
826 }
827
828 /* Some capabilities may be different across iommus */
829 static void domain_update_iommu_cap(struct dmar_domain *domain)
830 {
831 domain_update_iommu_coherency(domain);
832 domain->iommu_snooping = domain_update_iommu_snooping(NULL);
833 domain->iommu_superpage = domain_update_iommu_superpage(NULL);
834 }
835
836 static inline struct context_entry *iommu_context_addr(struct intel_iommu *iommu,
837 u8 bus, u8 devfn, int alloc)
838 {
839 struct root_entry *root = &iommu->root_entry[bus];
840 struct context_entry *context;
841 u64 *entry;
842
843 entry = &root->lo;
844 if (ecs_enabled(iommu)) {
845 if (devfn >= 0x80) {
846 devfn -= 0x80;
847 entry = &root->hi;
848 }
849 devfn *= 2;
850 }
851 if (*entry & 1)
852 context = phys_to_virt(*entry & VTD_PAGE_MASK);
853 else {
854 unsigned long phy_addr;
855 if (!alloc)
856 return NULL;
857
858 context = alloc_pgtable_page(iommu->node);
859 if (!context)
860 return NULL;
861
862 __iommu_flush_cache(iommu, (void *)context, CONTEXT_SIZE);
863 phy_addr = virt_to_phys((void *)context);
864 *entry = phy_addr | 1;
865 __iommu_flush_cache(iommu, entry, sizeof(*entry));
866 }
867 return &context[devfn];
868 }
869
870 static int iommu_dummy(struct device *dev)
871 {
872 return dev->archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO;
873 }
874
875 static struct intel_iommu *device_to_iommu(struct device *dev, u8 *bus, u8 *devfn)
876 {
877 struct dmar_drhd_unit *drhd = NULL;
878 struct intel_iommu *iommu;
879 struct device *tmp;
880 struct pci_dev *ptmp, *pdev = NULL;
881 u16 segment = 0;
882 int i;
883
884 if (iommu_dummy(dev))
885 return NULL;
886
887 if (dev_is_pci(dev)) {
888 pdev = to_pci_dev(dev);
889 segment = pci_domain_nr(pdev->bus);
890 } else if (has_acpi_companion(dev))
891 dev = &ACPI_COMPANION(dev)->dev;
892
893 rcu_read_lock();
894 for_each_active_iommu(iommu, drhd) {
895 if (pdev && segment != drhd->segment)
896 continue;
897
898 for_each_active_dev_scope(drhd->devices,
899 drhd->devices_cnt, i, tmp) {
900 if (tmp == dev) {
901 *bus = drhd->devices[i].bus;
902 *devfn = drhd->devices[i].devfn;
903 goto out;
904 }
905
906 if (!pdev || !dev_is_pci(tmp))
907 continue;
908
909 ptmp = to_pci_dev(tmp);
910 if (ptmp->subordinate &&
911 ptmp->subordinate->number <= pdev->bus->number &&
912 ptmp->subordinate->busn_res.end >= pdev->bus->number)
913 goto got_pdev;
914 }
915
916 if (pdev && drhd->include_all) {
917 got_pdev:
918 *bus = pdev->bus->number;
919 *devfn = pdev->devfn;
920 goto out;
921 }
922 }
923 iommu = NULL;
924 out:
925 rcu_read_unlock();
926
927 return iommu;
928 }
929
930 static void domain_flush_cache(struct dmar_domain *domain,
931 void *addr, int size)
932 {
933 if (!domain->iommu_coherency)
934 clflush_cache_range(addr, size);
935 }
936
937 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
938 {
939 struct context_entry *context;
940 int ret = 0;
941 unsigned long flags;
942
943 spin_lock_irqsave(&iommu->lock, flags);
944 context = iommu_context_addr(iommu, bus, devfn, 0);
945 if (context)
946 ret = context_present(context);
947 spin_unlock_irqrestore(&iommu->lock, flags);
948 return ret;
949 }
950
951 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
952 {
953 struct context_entry *context;
954 unsigned long flags;
955
956 spin_lock_irqsave(&iommu->lock, flags);
957 context = iommu_context_addr(iommu, bus, devfn, 0);
958 if (context) {
959 context_clear_entry(context);
960 __iommu_flush_cache(iommu, context, sizeof(*context));
961 }
962 spin_unlock_irqrestore(&iommu->lock, flags);
963 }
964
965 static void free_context_table(struct intel_iommu *iommu)
966 {
967 int i;
968 unsigned long flags;
969 struct context_entry *context;
970
971 spin_lock_irqsave(&iommu->lock, flags);
972 if (!iommu->root_entry) {
973 goto out;
974 }
975 for (i = 0; i < ROOT_ENTRY_NR; i++) {
976 context = iommu_context_addr(iommu, i, 0, 0);
977 if (context)
978 free_pgtable_page(context);
979
980 if (!ecs_enabled(iommu))
981 continue;
982
983 context = iommu_context_addr(iommu, i, 0x80, 0);
984 if (context)
985 free_pgtable_page(context);
986
987 }
988 free_pgtable_page(iommu->root_entry);
989 iommu->root_entry = NULL;
990 out:
991 spin_unlock_irqrestore(&iommu->lock, flags);
992 }
993
994 static struct dma_pte *pfn_to_dma_pte(struct dmar_domain *domain,
995 unsigned long pfn, int *target_level)
996 {
997 struct dma_pte *parent, *pte = NULL;
998 int level = agaw_to_level(domain->agaw);
999 int offset;
1000
1001 BUG_ON(!domain->pgd);
1002
1003 if (!domain_pfn_supported(domain, pfn))
1004 /* Address beyond IOMMU's addressing capabilities. */
1005 return NULL;
1006
1007 parent = domain->pgd;
1008
1009 while (1) {
1010 void *tmp_page;
1011
1012 offset = pfn_level_offset(pfn, level);
1013 pte = &parent[offset];
1014 if (!*target_level && (dma_pte_superpage(pte) || !dma_pte_present(pte)))
1015 break;
1016 if (level == *target_level)
1017 break;
1018
1019 if (!dma_pte_present(pte)) {
1020 uint64_t pteval;
1021
1022 tmp_page = alloc_pgtable_page(domain->nid);
1023
1024 if (!tmp_page)
1025 return NULL;
1026
1027 domain_flush_cache(domain, tmp_page, VTD_PAGE_SIZE);
1028 pteval = ((uint64_t)virt_to_dma_pfn(tmp_page) << VTD_PAGE_SHIFT) | DMA_PTE_READ | DMA_PTE_WRITE;
1029 if (cmpxchg64(&pte->val, 0ULL, pteval))
1030 /* Someone else set it while we were thinking; use theirs. */
1031 free_pgtable_page(tmp_page);
1032 else
1033 domain_flush_cache(domain, pte, sizeof(*pte));
1034 }
1035 if (level == 1)
1036 break;
1037
1038 parent = phys_to_virt(dma_pte_addr(pte));
1039 level--;
1040 }
1041
1042 if (!*target_level)
1043 *target_level = level;
1044
1045 return pte;
1046 }
1047
1048
1049 /* return address's pte at specific level */
1050 static struct dma_pte *dma_pfn_level_pte(struct dmar_domain *domain,
1051 unsigned long pfn,
1052 int level, int *large_page)
1053 {
1054 struct dma_pte *parent, *pte = NULL;
1055 int total = agaw_to_level(domain->agaw);
1056 int offset;
1057
1058 parent = domain->pgd;
1059 while (level <= total) {
1060 offset = pfn_level_offset(pfn, total);
1061 pte = &parent[offset];
1062 if (level == total)
1063 return pte;
1064
1065 if (!dma_pte_present(pte)) {
1066 *large_page = total;
1067 break;
1068 }
1069
1070 if (dma_pte_superpage(pte)) {
1071 *large_page = total;
1072 return pte;
1073 }
1074
1075 parent = phys_to_virt(dma_pte_addr(pte));
1076 total--;
1077 }
1078 return NULL;
1079 }
1080
1081 /* clear last level pte, a tlb flush should be followed */
1082 static void dma_pte_clear_range(struct dmar_domain *domain,
1083 unsigned long start_pfn,
1084 unsigned long last_pfn)
1085 {
1086 unsigned int large_page = 1;
1087 struct dma_pte *first_pte, *pte;
1088
1089 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1090 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1091 BUG_ON(start_pfn > last_pfn);
1092
1093 /* we don't need lock here; nobody else touches the iova range */
1094 do {
1095 large_page = 1;
1096 first_pte = pte = dma_pfn_level_pte(domain, start_pfn, 1, &large_page);
1097 if (!pte) {
1098 start_pfn = align_to_level(start_pfn + 1, large_page + 1);
1099 continue;
1100 }
1101 do {
1102 dma_clear_pte(pte);
1103 start_pfn += lvl_to_nr_pages(large_page);
1104 pte++;
1105 } while (start_pfn <= last_pfn && !first_pte_in_page(pte));
1106
1107 domain_flush_cache(domain, first_pte,
1108 (void *)pte - (void *)first_pte);
1109
1110 } while (start_pfn && start_pfn <= last_pfn);
1111 }
1112
1113 static void dma_pte_free_level(struct dmar_domain *domain, int level,
1114 struct dma_pte *pte, unsigned long pfn,
1115 unsigned long start_pfn, unsigned long last_pfn)
1116 {
1117 pfn = max(start_pfn, pfn);
1118 pte = &pte[pfn_level_offset(pfn, level)];
1119
1120 do {
1121 unsigned long level_pfn;
1122 struct dma_pte *level_pte;
1123
1124 if (!dma_pte_present(pte) || dma_pte_superpage(pte))
1125 goto next;
1126
1127 level_pfn = pfn & level_mask(level - 1);
1128 level_pte = phys_to_virt(dma_pte_addr(pte));
1129
1130 if (level > 2)
1131 dma_pte_free_level(domain, level - 1, level_pte,
1132 level_pfn, start_pfn, last_pfn);
1133
1134 /* If range covers entire pagetable, free it */
1135 if (!(start_pfn > level_pfn ||
1136 last_pfn < level_pfn + level_size(level) - 1)) {
1137 dma_clear_pte(pte);
1138 domain_flush_cache(domain, pte, sizeof(*pte));
1139 free_pgtable_page(level_pte);
1140 }
1141 next:
1142 pfn += level_size(level);
1143 } while (!first_pte_in_page(++pte) && pfn <= last_pfn);
1144 }
1145
1146 /* free page table pages. last level pte should already be cleared */
1147 static void dma_pte_free_pagetable(struct dmar_domain *domain,
1148 unsigned long start_pfn,
1149 unsigned long last_pfn)
1150 {
1151 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1152 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1153 BUG_ON(start_pfn > last_pfn);
1154
1155 dma_pte_clear_range(domain, start_pfn, last_pfn);
1156
1157 /* We don't need lock here; nobody else touches the iova range */
1158 dma_pte_free_level(domain, agaw_to_level(domain->agaw),
1159 domain->pgd, 0, start_pfn, last_pfn);
1160
1161 /* free pgd */
1162 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
1163 free_pgtable_page(domain->pgd);
1164 domain->pgd = NULL;
1165 }
1166 }
1167
1168 /* When a page at a given level is being unlinked from its parent, we don't
1169 need to *modify* it at all. All we need to do is make a list of all the
1170 pages which can be freed just as soon as we've flushed the IOTLB and we
1171 know the hardware page-walk will no longer touch them.
1172 The 'pte' argument is the *parent* PTE, pointing to the page that is to
1173 be freed. */
1174 static struct page *dma_pte_list_pagetables(struct dmar_domain *domain,
1175 int level, struct dma_pte *pte,
1176 struct page *freelist)
1177 {
1178 struct page *pg;
1179
1180 pg = pfn_to_page(dma_pte_addr(pte) >> PAGE_SHIFT);
1181 pg->freelist = freelist;
1182 freelist = pg;
1183
1184 if (level == 1)
1185 return freelist;
1186
1187 pte = page_address(pg);
1188 do {
1189 if (dma_pte_present(pte) && !dma_pte_superpage(pte))
1190 freelist = dma_pte_list_pagetables(domain, level - 1,
1191 pte, freelist);
1192 pte++;
1193 } while (!first_pte_in_page(pte));
1194
1195 return freelist;
1196 }
1197
1198 static struct page *dma_pte_clear_level(struct dmar_domain *domain, int level,
1199 struct dma_pte *pte, unsigned long pfn,
1200 unsigned long start_pfn,
1201 unsigned long last_pfn,
1202 struct page *freelist)
1203 {
1204 struct dma_pte *first_pte = NULL, *last_pte = NULL;
1205
1206 pfn = max(start_pfn, pfn);
1207 pte = &pte[pfn_level_offset(pfn, level)];
1208
1209 do {
1210 unsigned long level_pfn;
1211
1212 if (!dma_pte_present(pte))
1213 goto next;
1214
1215 level_pfn = pfn & level_mask(level);
1216
1217 /* If range covers entire pagetable, free it */
1218 if (start_pfn <= level_pfn &&
1219 last_pfn >= level_pfn + level_size(level) - 1) {
1220 /* These suborbinate page tables are going away entirely. Don't
1221 bother to clear them; we're just going to *free* them. */
1222 if (level > 1 && !dma_pte_superpage(pte))
1223 freelist = dma_pte_list_pagetables(domain, level - 1, pte, freelist);
1224
1225 dma_clear_pte(pte);
1226 if (!first_pte)
1227 first_pte = pte;
1228 last_pte = pte;
1229 } else if (level > 1) {
1230 /* Recurse down into a level that isn't *entirely* obsolete */
1231 freelist = dma_pte_clear_level(domain, level - 1,
1232 phys_to_virt(dma_pte_addr(pte)),
1233 level_pfn, start_pfn, last_pfn,
1234 freelist);
1235 }
1236 next:
1237 pfn += level_size(level);
1238 } while (!first_pte_in_page(++pte) && pfn <= last_pfn);
1239
1240 if (first_pte)
1241 domain_flush_cache(domain, first_pte,
1242 (void *)++last_pte - (void *)first_pte);
1243
1244 return freelist;
1245 }
1246
1247 /* We can't just free the pages because the IOMMU may still be walking
1248 the page tables, and may have cached the intermediate levels. The
1249 pages can only be freed after the IOTLB flush has been done. */
1250 static struct page *domain_unmap(struct dmar_domain *domain,
1251 unsigned long start_pfn,
1252 unsigned long last_pfn)
1253 {
1254 struct page *freelist = NULL;
1255
1256 BUG_ON(!domain_pfn_supported(domain, start_pfn));
1257 BUG_ON(!domain_pfn_supported(domain, last_pfn));
1258 BUG_ON(start_pfn > last_pfn);
1259
1260 /* we don't need lock here; nobody else touches the iova range */
1261 freelist = dma_pte_clear_level(domain, agaw_to_level(domain->agaw),
1262 domain->pgd, 0, start_pfn, last_pfn, NULL);
1263
1264 /* free pgd */
1265 if (start_pfn == 0 && last_pfn == DOMAIN_MAX_PFN(domain->gaw)) {
1266 struct page *pgd_page = virt_to_page(domain->pgd);
1267 pgd_page->freelist = freelist;
1268 freelist = pgd_page;
1269
1270 domain->pgd = NULL;
1271 }
1272
1273 return freelist;
1274 }
1275
1276 static void dma_free_pagelist(struct page *freelist)
1277 {
1278 struct page *pg;
1279
1280 while ((pg = freelist)) {
1281 freelist = pg->freelist;
1282 free_pgtable_page(page_address(pg));
1283 }
1284 }
1285
1286 /* iommu handling */
1287 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
1288 {
1289 struct root_entry *root;
1290 unsigned long flags;
1291
1292 root = (struct root_entry *)alloc_pgtable_page(iommu->node);
1293 if (!root) {
1294 pr_err("Allocating root entry for %s failed\n",
1295 iommu->name);
1296 return -ENOMEM;
1297 }
1298
1299 __iommu_flush_cache(iommu, root, ROOT_SIZE);
1300
1301 spin_lock_irqsave(&iommu->lock, flags);
1302 iommu->root_entry = root;
1303 spin_unlock_irqrestore(&iommu->lock, flags);
1304
1305 return 0;
1306 }
1307
1308 static void iommu_set_root_entry(struct intel_iommu *iommu)
1309 {
1310 u64 addr;
1311 u32 sts;
1312 unsigned long flag;
1313
1314 addr = virt_to_phys(iommu->root_entry);
1315 if (ecs_enabled(iommu))
1316 addr |= DMA_RTADDR_RTT;
1317
1318 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1319 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, addr);
1320
1321 writel(iommu->gcmd | DMA_GCMD_SRTP, iommu->reg + DMAR_GCMD_REG);
1322
1323 /* Make sure hardware complete it */
1324 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1325 readl, (sts & DMA_GSTS_RTPS), sts);
1326
1327 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1328 }
1329
1330 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
1331 {
1332 u32 val;
1333 unsigned long flag;
1334
1335 if (!rwbf_quirk && !cap_rwbf(iommu->cap))
1336 return;
1337
1338 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1339 writel(iommu->gcmd | DMA_GCMD_WBF, iommu->reg + DMAR_GCMD_REG);
1340
1341 /* Make sure hardware complete it */
1342 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1343 readl, (!(val & DMA_GSTS_WBFS)), val);
1344
1345 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1346 }
1347
1348 /* return value determine if we need a write buffer flush */
1349 static void __iommu_flush_context(struct intel_iommu *iommu,
1350 u16 did, u16 source_id, u8 function_mask,
1351 u64 type)
1352 {
1353 u64 val = 0;
1354 unsigned long flag;
1355
1356 switch (type) {
1357 case DMA_CCMD_GLOBAL_INVL:
1358 val = DMA_CCMD_GLOBAL_INVL;
1359 break;
1360 case DMA_CCMD_DOMAIN_INVL:
1361 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
1362 break;
1363 case DMA_CCMD_DEVICE_INVL:
1364 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
1365 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
1366 break;
1367 default:
1368 BUG();
1369 }
1370 val |= DMA_CCMD_ICC;
1371
1372 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1373 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
1374
1375 /* Make sure hardware complete it */
1376 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
1377 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
1378
1379 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1380 }
1381
1382 /* return value determine if we need a write buffer flush */
1383 static void __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
1384 u64 addr, unsigned int size_order, u64 type)
1385 {
1386 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
1387 u64 val = 0, val_iva = 0;
1388 unsigned long flag;
1389
1390 switch (type) {
1391 case DMA_TLB_GLOBAL_FLUSH:
1392 /* global flush doesn't need set IVA_REG */
1393 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
1394 break;
1395 case DMA_TLB_DSI_FLUSH:
1396 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
1397 break;
1398 case DMA_TLB_PSI_FLUSH:
1399 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
1400 /* IH bit is passed in as part of address */
1401 val_iva = size_order | addr;
1402 break;
1403 default:
1404 BUG();
1405 }
1406 /* Note: set drain read/write */
1407 #if 0
1408 /*
1409 * This is probably to be super secure.. Looks like we can
1410 * ignore it without any impact.
1411 */
1412 if (cap_read_drain(iommu->cap))
1413 val |= DMA_TLB_READ_DRAIN;
1414 #endif
1415 if (cap_write_drain(iommu->cap))
1416 val |= DMA_TLB_WRITE_DRAIN;
1417
1418 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1419 /* Note: Only uses first TLB reg currently */
1420 if (val_iva)
1421 dmar_writeq(iommu->reg + tlb_offset, val_iva);
1422 dmar_writeq(iommu->reg + tlb_offset + 8, val);
1423
1424 /* Make sure hardware complete it */
1425 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
1426 dmar_readq, (!(val & DMA_TLB_IVT)), val);
1427
1428 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1429
1430 /* check IOTLB invalidation granularity */
1431 if (DMA_TLB_IAIG(val) == 0)
1432 pr_err("Flush IOTLB failed\n");
1433 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
1434 pr_debug("TLB flush request %Lx, actual %Lx\n",
1435 (unsigned long long)DMA_TLB_IIRG(type),
1436 (unsigned long long)DMA_TLB_IAIG(val));
1437 }
1438
1439 static struct device_domain_info *
1440 iommu_support_dev_iotlb (struct dmar_domain *domain, struct intel_iommu *iommu,
1441 u8 bus, u8 devfn)
1442 {
1443 struct device_domain_info *info;
1444
1445 assert_spin_locked(&device_domain_lock);
1446
1447 if (!iommu->qi)
1448 return NULL;
1449
1450 list_for_each_entry(info, &domain->devices, link)
1451 if (info->iommu == iommu && info->bus == bus &&
1452 info->devfn == devfn) {
1453 if (info->ats_supported && info->dev)
1454 return info;
1455 break;
1456 }
1457
1458 return NULL;
1459 }
1460
1461 static void iommu_enable_dev_iotlb(struct device_domain_info *info)
1462 {
1463 struct pci_dev *pdev;
1464
1465 if (!info || !dev_is_pci(info->dev))
1466 return;
1467
1468 pdev = to_pci_dev(info->dev);
1469
1470 #ifdef CONFIG_INTEL_IOMMU_SVM
1471 /* The PCIe spec, in its wisdom, declares that the behaviour of
1472 the device if you enable PASID support after ATS support is
1473 undefined. So always enable PASID support on devices which
1474 have it, even if we can't yet know if we're ever going to
1475 use it. */
1476 if (info->pasid_supported && !pci_enable_pasid(pdev, info->pasid_supported & ~1))
1477 info->pasid_enabled = 1;
1478
1479 if (info->pri_supported && !pci_reset_pri(pdev) && !pci_enable_pri(pdev, 32))
1480 info->pri_enabled = 1;
1481 #endif
1482 if (info->ats_supported && !pci_enable_ats(pdev, VTD_PAGE_SHIFT)) {
1483 info->ats_enabled = 1;
1484 info->ats_qdep = pci_ats_queue_depth(pdev);
1485 }
1486 }
1487
1488 static void iommu_disable_dev_iotlb(struct device_domain_info *info)
1489 {
1490 struct pci_dev *pdev;
1491
1492 if (dev_is_pci(info->dev))
1493 return;
1494
1495 pdev = to_pci_dev(info->dev);
1496
1497 if (info->ats_enabled) {
1498 pci_disable_ats(pdev);
1499 info->ats_enabled = 0;
1500 }
1501 #ifdef CONFIG_INTEL_IOMMU_SVM
1502 if (info->pri_enabled) {
1503 pci_disable_pri(pdev);
1504 info->pri_enabled = 0;
1505 }
1506 if (info->pasid_enabled) {
1507 pci_disable_pasid(pdev);
1508 info->pasid_enabled = 0;
1509 }
1510 #endif
1511 }
1512
1513 static void iommu_flush_dev_iotlb(struct dmar_domain *domain,
1514 u64 addr, unsigned mask)
1515 {
1516 u16 sid, qdep;
1517 unsigned long flags;
1518 struct device_domain_info *info;
1519
1520 spin_lock_irqsave(&device_domain_lock, flags);
1521 list_for_each_entry(info, &domain->devices, link) {
1522 if (!info->ats_enabled)
1523 continue;
1524
1525 sid = info->bus << 8 | info->devfn;
1526 qdep = info->ats_qdep;
1527 qi_flush_dev_iotlb(info->iommu, sid, qdep, addr, mask);
1528 }
1529 spin_unlock_irqrestore(&device_domain_lock, flags);
1530 }
1531
1532 static void iommu_flush_iotlb_psi(struct intel_iommu *iommu,
1533 struct dmar_domain *domain,
1534 unsigned long pfn, unsigned int pages,
1535 int ih, int map)
1536 {
1537 unsigned int mask = ilog2(__roundup_pow_of_two(pages));
1538 uint64_t addr = (uint64_t)pfn << VTD_PAGE_SHIFT;
1539 u16 did = domain->iommu_did[iommu->seq_id];
1540
1541 BUG_ON(pages == 0);
1542
1543 if (ih)
1544 ih = 1 << 6;
1545 /*
1546 * Fallback to domain selective flush if no PSI support or the size is
1547 * too big.
1548 * PSI requires page size to be 2 ^ x, and the base address is naturally
1549 * aligned to the size
1550 */
1551 if (!cap_pgsel_inv(iommu->cap) || mask > cap_max_amask_val(iommu->cap))
1552 iommu->flush.flush_iotlb(iommu, did, 0, 0,
1553 DMA_TLB_DSI_FLUSH);
1554 else
1555 iommu->flush.flush_iotlb(iommu, did, addr | ih, mask,
1556 DMA_TLB_PSI_FLUSH);
1557
1558 /*
1559 * In caching mode, changes of pages from non-present to present require
1560 * flush. However, device IOTLB doesn't need to be flushed in this case.
1561 */
1562 if (!cap_caching_mode(iommu->cap) || !map)
1563 iommu_flush_dev_iotlb(get_iommu_domain(iommu, did),
1564 addr, mask);
1565 }
1566
1567 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
1568 {
1569 u32 pmen;
1570 unsigned long flags;
1571
1572 raw_spin_lock_irqsave(&iommu->register_lock, flags);
1573 pmen = readl(iommu->reg + DMAR_PMEN_REG);
1574 pmen &= ~DMA_PMEN_EPM;
1575 writel(pmen, iommu->reg + DMAR_PMEN_REG);
1576
1577 /* wait for the protected region status bit to clear */
1578 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
1579 readl, !(pmen & DMA_PMEN_PRS), pmen);
1580
1581 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1582 }
1583
1584 static void iommu_enable_translation(struct intel_iommu *iommu)
1585 {
1586 u32 sts;
1587 unsigned long flags;
1588
1589 raw_spin_lock_irqsave(&iommu->register_lock, flags);
1590 iommu->gcmd |= DMA_GCMD_TE;
1591 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1592
1593 /* Make sure hardware complete it */
1594 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1595 readl, (sts & DMA_GSTS_TES), sts);
1596
1597 raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1598 }
1599
1600 static void iommu_disable_translation(struct intel_iommu *iommu)
1601 {
1602 u32 sts;
1603 unsigned long flag;
1604
1605 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1606 iommu->gcmd &= ~DMA_GCMD_TE;
1607 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1608
1609 /* Make sure hardware complete it */
1610 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
1611 readl, (!(sts & DMA_GSTS_TES)), sts);
1612
1613 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1614 }
1615
1616
1617 static int iommu_init_domains(struct intel_iommu *iommu)
1618 {
1619 u32 ndomains, nlongs;
1620 size_t size;
1621
1622 ndomains = cap_ndoms(iommu->cap);
1623 pr_debug("%s: Number of Domains supported <%d>\n",
1624 iommu->name, ndomains);
1625 nlongs = BITS_TO_LONGS(ndomains);
1626
1627 spin_lock_init(&iommu->lock);
1628
1629 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
1630 if (!iommu->domain_ids) {
1631 pr_err("%s: Allocating domain id array failed\n",
1632 iommu->name);
1633 return -ENOMEM;
1634 }
1635
1636 size = ((ndomains >> 8) + 1) * sizeof(struct dmar_domain **);
1637 iommu->domains = kzalloc(size, GFP_KERNEL);
1638
1639 if (iommu->domains) {
1640 size = 256 * sizeof(struct dmar_domain *);
1641 iommu->domains[0] = kzalloc(size, GFP_KERNEL);
1642 }
1643
1644 if (!iommu->domains || !iommu->domains[0]) {
1645 pr_err("%s: Allocating domain array failed\n",
1646 iommu->name);
1647 kfree(iommu->domain_ids);
1648 kfree(iommu->domains);
1649 iommu->domain_ids = NULL;
1650 iommu->domains = NULL;
1651 return -ENOMEM;
1652 }
1653
1654
1655
1656 /*
1657 * If Caching mode is set, then invalid translations are tagged
1658 * with domain-id 0, hence we need to pre-allocate it. We also
1659 * use domain-id 0 as a marker for non-allocated domain-id, so
1660 * make sure it is not used for a real domain.
1661 */
1662 set_bit(0, iommu->domain_ids);
1663
1664 return 0;
1665 }
1666
1667 static void disable_dmar_iommu(struct intel_iommu *iommu)
1668 {
1669 struct device_domain_info *info, *tmp;
1670 unsigned long flags;
1671
1672 if (!iommu->domains || !iommu->domain_ids)
1673 return;
1674
1675 spin_lock_irqsave(&device_domain_lock, flags);
1676 list_for_each_entry_safe(info, tmp, &device_domain_list, global) {
1677 struct dmar_domain *domain;
1678
1679 if (info->iommu != iommu)
1680 continue;
1681
1682 if (!info->dev || !info->domain)
1683 continue;
1684
1685 domain = info->domain;
1686
1687 dmar_remove_one_dev_info(domain, info->dev);
1688
1689 if (!domain_type_is_vm_or_si(domain))
1690 domain_exit(domain);
1691 }
1692 spin_unlock_irqrestore(&device_domain_lock, flags);
1693
1694 if (iommu->gcmd & DMA_GCMD_TE)
1695 iommu_disable_translation(iommu);
1696 }
1697
1698 static void free_dmar_iommu(struct intel_iommu *iommu)
1699 {
1700 if ((iommu->domains) && (iommu->domain_ids)) {
1701 int elems = (cap_ndoms(iommu->cap) >> 8) + 1;
1702 int i;
1703
1704 for (i = 0; i < elems; i++)
1705 kfree(iommu->domains[i]);
1706 kfree(iommu->domains);
1707 kfree(iommu->domain_ids);
1708 iommu->domains = NULL;
1709 iommu->domain_ids = NULL;
1710 }
1711
1712 g_iommus[iommu->seq_id] = NULL;
1713
1714 /* free context mapping */
1715 free_context_table(iommu);
1716
1717 #ifdef CONFIG_INTEL_IOMMU_SVM
1718 if (pasid_enabled(iommu)) {
1719 if (ecap_prs(iommu->ecap))
1720 intel_svm_finish_prq(iommu);
1721 intel_svm_free_pasid_tables(iommu);
1722 }
1723 #endif
1724 }
1725
1726 static struct dmar_domain *alloc_domain(int flags)
1727 {
1728 struct dmar_domain *domain;
1729
1730 domain = alloc_domain_mem();
1731 if (!domain)
1732 return NULL;
1733
1734 memset(domain, 0, sizeof(*domain));
1735 domain->nid = -1;
1736 domain->flags = flags;
1737 INIT_LIST_HEAD(&domain->devices);
1738
1739 return domain;
1740 }
1741
1742 /* Must be called with iommu->lock */
1743 static int domain_attach_iommu(struct dmar_domain *domain,
1744 struct intel_iommu *iommu)
1745 {
1746 unsigned long ndomains;
1747 int num;
1748
1749 assert_spin_locked(&device_domain_lock);
1750 assert_spin_locked(&iommu->lock);
1751
1752 domain->iommu_refcnt[iommu->seq_id] += 1;
1753 domain->iommu_count += 1;
1754 if (domain->iommu_refcnt[iommu->seq_id] == 1) {
1755 ndomains = cap_ndoms(iommu->cap);
1756 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1757
1758 if (num >= ndomains) {
1759 pr_err("%s: No free domain ids\n", iommu->name);
1760 domain->iommu_refcnt[iommu->seq_id] -= 1;
1761 domain->iommu_count -= 1;
1762 return -ENOSPC;
1763 }
1764
1765 set_bit(num, iommu->domain_ids);
1766 set_iommu_domain(iommu, num, domain);
1767
1768 domain->iommu_did[iommu->seq_id] = num;
1769 domain->nid = iommu->node;
1770
1771 domain_update_iommu_cap(domain);
1772 }
1773
1774 return 0;
1775 }
1776
1777 static int domain_detach_iommu(struct dmar_domain *domain,
1778 struct intel_iommu *iommu)
1779 {
1780 int num, count = INT_MAX;
1781
1782 assert_spin_locked(&device_domain_lock);
1783 assert_spin_locked(&iommu->lock);
1784
1785 domain->iommu_refcnt[iommu->seq_id] -= 1;
1786 count = --domain->iommu_count;
1787 if (domain->iommu_refcnt[iommu->seq_id] == 0) {
1788 num = domain->iommu_did[iommu->seq_id];
1789 clear_bit(num, iommu->domain_ids);
1790 set_iommu_domain(iommu, num, NULL);
1791
1792 domain_update_iommu_cap(domain);
1793 domain->iommu_did[iommu->seq_id] = 0;
1794 }
1795
1796 return count;
1797 }
1798
1799 static struct iova_domain reserved_iova_list;
1800 static struct lock_class_key reserved_rbtree_key;
1801
1802 static int dmar_init_reserved_ranges(void)
1803 {
1804 struct pci_dev *pdev = NULL;
1805 struct iova *iova;
1806 int i;
1807
1808 init_iova_domain(&reserved_iova_list, VTD_PAGE_SIZE, IOVA_START_PFN,
1809 DMA_32BIT_PFN);
1810
1811 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1812 &reserved_rbtree_key);
1813
1814 /* IOAPIC ranges shouldn't be accessed by DMA */
1815 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1816 IOVA_PFN(IOAPIC_RANGE_END));
1817 if (!iova) {
1818 pr_err("Reserve IOAPIC range failed\n");
1819 return -ENODEV;
1820 }
1821
1822 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1823 for_each_pci_dev(pdev) {
1824 struct resource *r;
1825
1826 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1827 r = &pdev->resource[i];
1828 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1829 continue;
1830 iova = reserve_iova(&reserved_iova_list,
1831 IOVA_PFN(r->start),
1832 IOVA_PFN(r->end));
1833 if (!iova) {
1834 pr_err("Reserve iova failed\n");
1835 return -ENODEV;
1836 }
1837 }
1838 }
1839 return 0;
1840 }
1841
1842 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1843 {
1844 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1845 }
1846
1847 static inline int guestwidth_to_adjustwidth(int gaw)
1848 {
1849 int agaw;
1850 int r = (gaw - 12) % 9;
1851
1852 if (r == 0)
1853 agaw = gaw;
1854 else
1855 agaw = gaw + 9 - r;
1856 if (agaw > 64)
1857 agaw = 64;
1858 return agaw;
1859 }
1860
1861 static int domain_init(struct dmar_domain *domain, struct intel_iommu *iommu,
1862 int guest_width)
1863 {
1864 int adjust_width, agaw;
1865 unsigned long sagaw;
1866
1867 init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN,
1868 DMA_32BIT_PFN);
1869 domain_reserve_special_ranges(domain);
1870
1871 /* calculate AGAW */
1872 if (guest_width > cap_mgaw(iommu->cap))
1873 guest_width = cap_mgaw(iommu->cap);
1874 domain->gaw = guest_width;
1875 adjust_width = guestwidth_to_adjustwidth(guest_width);
1876 agaw = width_to_agaw(adjust_width);
1877 sagaw = cap_sagaw(iommu->cap);
1878 if (!test_bit(agaw, &sagaw)) {
1879 /* hardware doesn't support it, choose a bigger one */
1880 pr_debug("Hardware doesn't support agaw %d\n", agaw);
1881 agaw = find_next_bit(&sagaw, 5, agaw);
1882 if (agaw >= 5)
1883 return -ENODEV;
1884 }
1885 domain->agaw = agaw;
1886
1887 if (ecap_coherent(iommu->ecap))
1888 domain->iommu_coherency = 1;
1889 else
1890 domain->iommu_coherency = 0;
1891
1892 if (ecap_sc_support(iommu->ecap))
1893 domain->iommu_snooping = 1;
1894 else
1895 domain->iommu_snooping = 0;
1896
1897 if (intel_iommu_superpage)
1898 domain->iommu_superpage = fls(cap_super_page_val(iommu->cap));
1899 else
1900 domain->iommu_superpage = 0;
1901
1902 domain->nid = iommu->node;
1903
1904 /* always allocate the top pgd */
1905 domain->pgd = (struct dma_pte *)alloc_pgtable_page(domain->nid);
1906 if (!domain->pgd)
1907 return -ENOMEM;
1908 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE);
1909 return 0;
1910 }
1911
1912 static void domain_exit(struct dmar_domain *domain)
1913 {
1914 struct page *freelist = NULL;
1915
1916 /* Domain 0 is reserved, so dont process it */
1917 if (!domain)
1918 return;
1919
1920 /* Flush any lazy unmaps that may reference this domain */
1921 if (!intel_iommu_strict)
1922 flush_unmaps_timeout(0);
1923
1924 /* Remove associated devices and clear attached or cached domains */
1925 rcu_read_lock();
1926 domain_remove_dev_info(domain);
1927 rcu_read_unlock();
1928
1929 /* destroy iovas */
1930 put_iova_domain(&domain->iovad);
1931
1932 freelist = domain_unmap(domain, 0, DOMAIN_MAX_PFN(domain->gaw));
1933
1934 dma_free_pagelist(freelist);
1935
1936 free_domain_mem(domain);
1937 }
1938
1939 static int domain_context_mapping_one(struct dmar_domain *domain,
1940 struct intel_iommu *iommu,
1941 u8 bus, u8 devfn)
1942 {
1943 u16 did = domain->iommu_did[iommu->seq_id];
1944 int translation = CONTEXT_TT_MULTI_LEVEL;
1945 struct device_domain_info *info = NULL;
1946 struct context_entry *context;
1947 unsigned long flags;
1948 struct dma_pte *pgd;
1949 int ret, agaw;
1950
1951 WARN_ON(did == 0);
1952
1953 if (hw_pass_through && domain_type_is_si(domain))
1954 translation = CONTEXT_TT_PASS_THROUGH;
1955
1956 pr_debug("Set context mapping for %02x:%02x.%d\n",
1957 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1958
1959 BUG_ON(!domain->pgd);
1960
1961 spin_lock_irqsave(&device_domain_lock, flags);
1962 spin_lock(&iommu->lock);
1963
1964 ret = -ENOMEM;
1965 context = iommu_context_addr(iommu, bus, devfn, 1);
1966 if (!context)
1967 goto out_unlock;
1968
1969 ret = 0;
1970 if (context_present(context))
1971 goto out_unlock;
1972
1973 pgd = domain->pgd;
1974
1975 context_clear_entry(context);
1976 context_set_domain_id(context, did);
1977
1978 /*
1979 * Skip top levels of page tables for iommu which has less agaw
1980 * than default. Unnecessary for PT mode.
1981 */
1982 if (translation != CONTEXT_TT_PASS_THROUGH) {
1983 for (agaw = domain->agaw; agaw != iommu->agaw; agaw--) {
1984 ret = -ENOMEM;
1985 pgd = phys_to_virt(dma_pte_addr(pgd));
1986 if (!dma_pte_present(pgd))
1987 goto out_unlock;
1988 }
1989
1990 info = iommu_support_dev_iotlb(domain, iommu, bus, devfn);
1991 if (info && info->ats_supported)
1992 translation = CONTEXT_TT_DEV_IOTLB;
1993 else
1994 translation = CONTEXT_TT_MULTI_LEVEL;
1995
1996 context_set_address_root(context, virt_to_phys(pgd));
1997 context_set_address_width(context, iommu->agaw);
1998 } else {
1999 /*
2000 * In pass through mode, AW must be programmed to
2001 * indicate the largest AGAW value supported by
2002 * hardware. And ASR is ignored by hardware.
2003 */
2004 context_set_address_width(context, iommu->msagaw);
2005 }
2006
2007 context_set_translation_type(context, translation);
2008 context_set_fault_enable(context);
2009 context_set_present(context);
2010 domain_flush_cache(domain, context, sizeof(*context));
2011
2012 /*
2013 * It's a non-present to present mapping. If hardware doesn't cache
2014 * non-present entry we only need to flush the write-buffer. If the
2015 * _does_ cache non-present entries, then it does so in the special
2016 * domain #0, which we have to flush:
2017 */
2018 if (cap_caching_mode(iommu->cap)) {
2019 iommu->flush.flush_context(iommu, 0,
2020 (((u16)bus) << 8) | devfn,
2021 DMA_CCMD_MASK_NOBIT,
2022 DMA_CCMD_DEVICE_INVL);
2023 iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
2024 } else {
2025 iommu_flush_write_buffer(iommu);
2026 }
2027 iommu_enable_dev_iotlb(info);
2028
2029 ret = 0;
2030
2031 out_unlock:
2032 spin_unlock(&iommu->lock);
2033 spin_unlock_irqrestore(&device_domain_lock, flags);
2034
2035 return 0;
2036 }
2037
2038 struct domain_context_mapping_data {
2039 struct dmar_domain *domain;
2040 struct intel_iommu *iommu;
2041 };
2042
2043 static int domain_context_mapping_cb(struct pci_dev *pdev,
2044 u16 alias, void *opaque)
2045 {
2046 struct domain_context_mapping_data *data = opaque;
2047
2048 return domain_context_mapping_one(data->domain, data->iommu,
2049 PCI_BUS_NUM(alias), alias & 0xff);
2050 }
2051
2052 static int
2053 domain_context_mapping(struct dmar_domain *domain, struct device *dev)
2054 {
2055 struct intel_iommu *iommu;
2056 u8 bus, devfn;
2057 struct domain_context_mapping_data data;
2058
2059 iommu = device_to_iommu(dev, &bus, &devfn);
2060 if (!iommu)
2061 return -ENODEV;
2062
2063 if (!dev_is_pci(dev))
2064 return domain_context_mapping_one(domain, iommu, bus, devfn);
2065
2066 data.domain = domain;
2067 data.iommu = iommu;
2068
2069 return pci_for_each_dma_alias(to_pci_dev(dev),
2070 &domain_context_mapping_cb, &data);
2071 }
2072
2073 static int domain_context_mapped_cb(struct pci_dev *pdev,
2074 u16 alias, void *opaque)
2075 {
2076 struct intel_iommu *iommu = opaque;
2077
2078 return !device_context_mapped(iommu, PCI_BUS_NUM(alias), alias & 0xff);
2079 }
2080
2081 static int domain_context_mapped(struct device *dev)
2082 {
2083 struct intel_iommu *iommu;
2084 u8 bus, devfn;
2085
2086 iommu = device_to_iommu(dev, &bus, &devfn);
2087 if (!iommu)
2088 return -ENODEV;
2089
2090 if (!dev_is_pci(dev))
2091 return device_context_mapped(iommu, bus, devfn);
2092
2093 return !pci_for_each_dma_alias(to_pci_dev(dev),
2094 domain_context_mapped_cb, iommu);
2095 }
2096
2097 /* Returns a number of VTD pages, but aligned to MM page size */
2098 static inline unsigned long aligned_nrpages(unsigned long host_addr,
2099 size_t size)
2100 {
2101 host_addr &= ~PAGE_MASK;
2102 return PAGE_ALIGN(host_addr + size) >> VTD_PAGE_SHIFT;
2103 }
2104
2105 /* Return largest possible superpage level for a given mapping */
2106 static inline int hardware_largepage_caps(struct dmar_domain *domain,
2107 unsigned long iov_pfn,
2108 unsigned long phy_pfn,
2109 unsigned long pages)
2110 {
2111 int support, level = 1;
2112 unsigned long pfnmerge;
2113
2114 support = domain->iommu_superpage;
2115
2116 /* To use a large page, the virtual *and* physical addresses
2117 must be aligned to 2MiB/1GiB/etc. Lower bits set in either
2118 of them will mean we have to use smaller pages. So just
2119 merge them and check both at once. */
2120 pfnmerge = iov_pfn | phy_pfn;
2121
2122 while (support && !(pfnmerge & ~VTD_STRIDE_MASK)) {
2123 pages >>= VTD_STRIDE_SHIFT;
2124 if (!pages)
2125 break;
2126 pfnmerge >>= VTD_STRIDE_SHIFT;
2127 level++;
2128 support--;
2129 }
2130 return level;
2131 }
2132
2133 static int __domain_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2134 struct scatterlist *sg, unsigned long phys_pfn,
2135 unsigned long nr_pages, int prot)
2136 {
2137 struct dma_pte *first_pte = NULL, *pte = NULL;
2138 phys_addr_t uninitialized_var(pteval);
2139 unsigned long sg_res = 0;
2140 unsigned int largepage_lvl = 0;
2141 unsigned long lvl_pages = 0;
2142
2143 BUG_ON(!domain_pfn_supported(domain, iov_pfn + nr_pages - 1));
2144
2145 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
2146 return -EINVAL;
2147
2148 prot &= DMA_PTE_READ | DMA_PTE_WRITE | DMA_PTE_SNP;
2149
2150 if (!sg) {
2151 sg_res = nr_pages;
2152 pteval = ((phys_addr_t)phys_pfn << VTD_PAGE_SHIFT) | prot;
2153 }
2154
2155 while (nr_pages > 0) {
2156 uint64_t tmp;
2157
2158 if (!sg_res) {
2159 sg_res = aligned_nrpages(sg->offset, sg->length);
2160 sg->dma_address = ((dma_addr_t)iov_pfn << VTD_PAGE_SHIFT) + sg->offset;
2161 sg->dma_length = sg->length;
2162 pteval = page_to_phys(sg_page(sg)) | prot;
2163 phys_pfn = pteval >> VTD_PAGE_SHIFT;
2164 }
2165
2166 if (!pte) {
2167 largepage_lvl = hardware_largepage_caps(domain, iov_pfn, phys_pfn, sg_res);
2168
2169 first_pte = pte = pfn_to_dma_pte(domain, iov_pfn, &largepage_lvl);
2170 if (!pte)
2171 return -ENOMEM;
2172 /* It is large page*/
2173 if (largepage_lvl > 1) {
2174 unsigned long nr_superpages, end_pfn;
2175
2176 pteval |= DMA_PTE_LARGE_PAGE;
2177 lvl_pages = lvl_to_nr_pages(largepage_lvl);
2178
2179 nr_superpages = sg_res / lvl_pages;
2180 end_pfn = iov_pfn + nr_superpages * lvl_pages - 1;
2181
2182 /*
2183 * Ensure that old small page tables are
2184 * removed to make room for superpage(s).
2185 */
2186 dma_pte_free_pagetable(domain, iov_pfn, end_pfn);
2187 } else {
2188 pteval &= ~(uint64_t)DMA_PTE_LARGE_PAGE;
2189 }
2190
2191 }
2192 /* We don't need lock here, nobody else
2193 * touches the iova range
2194 */
2195 tmp = cmpxchg64_local(&pte->val, 0ULL, pteval);
2196 if (tmp) {
2197 static int dumps = 5;
2198 pr_crit("ERROR: DMA PTE for vPFN 0x%lx already set (to %llx not %llx)\n",
2199 iov_pfn, tmp, (unsigned long long)pteval);
2200 if (dumps) {
2201 dumps--;
2202 debug_dma_dump_mappings(NULL);
2203 }
2204 WARN_ON(1);
2205 }
2206
2207 lvl_pages = lvl_to_nr_pages(largepage_lvl);
2208
2209 BUG_ON(nr_pages < lvl_pages);
2210 BUG_ON(sg_res < lvl_pages);
2211
2212 nr_pages -= lvl_pages;
2213 iov_pfn += lvl_pages;
2214 phys_pfn += lvl_pages;
2215 pteval += lvl_pages * VTD_PAGE_SIZE;
2216 sg_res -= lvl_pages;
2217
2218 /* If the next PTE would be the first in a new page, then we
2219 need to flush the cache on the entries we've just written.
2220 And then we'll need to recalculate 'pte', so clear it and
2221 let it get set again in the if (!pte) block above.
2222
2223 If we're done (!nr_pages) we need to flush the cache too.
2224
2225 Also if we've been setting superpages, we may need to
2226 recalculate 'pte' and switch back to smaller pages for the
2227 end of the mapping, if the trailing size is not enough to
2228 use another superpage (i.e. sg_res < lvl_pages). */
2229 pte++;
2230 if (!nr_pages || first_pte_in_page(pte) ||
2231 (largepage_lvl > 1 && sg_res < lvl_pages)) {
2232 domain_flush_cache(domain, first_pte,
2233 (void *)pte - (void *)first_pte);
2234 pte = NULL;
2235 }
2236
2237 if (!sg_res && nr_pages)
2238 sg = sg_next(sg);
2239 }
2240 return 0;
2241 }
2242
2243 static inline int domain_sg_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2244 struct scatterlist *sg, unsigned long nr_pages,
2245 int prot)
2246 {
2247 return __domain_mapping(domain, iov_pfn, sg, 0, nr_pages, prot);
2248 }
2249
2250 static inline int domain_pfn_mapping(struct dmar_domain *domain, unsigned long iov_pfn,
2251 unsigned long phys_pfn, unsigned long nr_pages,
2252 int prot)
2253 {
2254 return __domain_mapping(domain, iov_pfn, NULL, phys_pfn, nr_pages, prot);
2255 }
2256
2257 static void domain_context_clear_one(struct intel_iommu *iommu, u8 bus, u8 devfn)
2258 {
2259 if (!iommu)
2260 return;
2261
2262 clear_context_table(iommu, bus, devfn);
2263 iommu->flush.flush_context(iommu, 0, 0, 0,
2264 DMA_CCMD_GLOBAL_INVL);
2265 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
2266 }
2267
2268 static inline void unlink_domain_info(struct device_domain_info *info)
2269 {
2270 assert_spin_locked(&device_domain_lock);
2271 list_del(&info->link);
2272 list_del(&info->global);
2273 if (info->dev)
2274 info->dev->archdata.iommu = NULL;
2275 }
2276
2277 static void domain_remove_dev_info(struct dmar_domain *domain)
2278 {
2279 struct device_domain_info *info, *tmp;
2280 unsigned long flags;
2281
2282 spin_lock_irqsave(&device_domain_lock, flags);
2283 list_for_each_entry_safe(info, tmp, &domain->devices, link)
2284 __dmar_remove_one_dev_info(info);
2285 spin_unlock_irqrestore(&device_domain_lock, flags);
2286 }
2287
2288 /*
2289 * find_domain
2290 * Note: we use struct device->archdata.iommu stores the info
2291 */
2292 static struct dmar_domain *find_domain(struct device *dev)
2293 {
2294 struct device_domain_info *info;
2295
2296 /* No lock here, assumes no domain exit in normal case */
2297 info = dev->archdata.iommu;
2298 if (info)
2299 return info->domain;
2300 return NULL;
2301 }
2302
2303 static inline struct device_domain_info *
2304 dmar_search_domain_by_dev_info(int segment, int bus, int devfn)
2305 {
2306 struct device_domain_info *info;
2307
2308 list_for_each_entry(info, &device_domain_list, global)
2309 if (info->iommu->segment == segment && info->bus == bus &&
2310 info->devfn == devfn)
2311 return info;
2312
2313 return NULL;
2314 }
2315
2316 static struct dmar_domain *dmar_insert_one_dev_info(struct intel_iommu *iommu,
2317 int bus, int devfn,
2318 struct device *dev,
2319 struct dmar_domain *domain)
2320 {
2321 struct dmar_domain *found = NULL;
2322 struct device_domain_info *info;
2323 unsigned long flags;
2324 int ret;
2325
2326 info = alloc_devinfo_mem();
2327 if (!info)
2328 return NULL;
2329
2330 info->bus = bus;
2331 info->devfn = devfn;
2332 info->ats_supported = info->pasid_supported = info->pri_supported = 0;
2333 info->ats_enabled = info->pasid_enabled = info->pri_enabled = 0;
2334 info->ats_qdep = 0;
2335 info->dev = dev;
2336 info->domain = domain;
2337 info->iommu = iommu;
2338
2339 if (dev && dev_is_pci(dev)) {
2340 struct pci_dev *pdev = to_pci_dev(info->dev);
2341
2342 if (ecap_dev_iotlb_support(iommu->ecap) &&
2343 pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS) &&
2344 dmar_find_matched_atsr_unit(pdev))
2345 info->ats_supported = 1;
2346
2347 if (ecs_enabled(iommu)) {
2348 if (pasid_enabled(iommu)) {
2349 int features = pci_pasid_features(pdev);
2350 if (features >= 0)
2351 info->pasid_supported = features | 1;
2352 }
2353
2354 if (info->ats_supported && ecap_prs(iommu->ecap) &&
2355 pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI))
2356 info->pri_supported = 1;
2357 }
2358 }
2359
2360 spin_lock_irqsave(&device_domain_lock, flags);
2361 if (dev)
2362 found = find_domain(dev);
2363
2364 if (!found) {
2365 struct device_domain_info *info2;
2366 info2 = dmar_search_domain_by_dev_info(iommu->segment, bus, devfn);
2367 if (info2) {
2368 found = info2->domain;
2369 info2->dev = dev;
2370 }
2371 }
2372
2373 if (found) {
2374 spin_unlock_irqrestore(&device_domain_lock, flags);
2375 free_devinfo_mem(info);
2376 /* Caller must free the original domain */
2377 return found;
2378 }
2379
2380 spin_lock(&iommu->lock);
2381 ret = domain_attach_iommu(domain, iommu);
2382 spin_unlock(&iommu->lock);
2383
2384 if (ret) {
2385 spin_unlock_irqrestore(&device_domain_lock, flags);
2386 free_devinfo_mem(info);
2387 return NULL;
2388 }
2389
2390 list_add(&info->link, &domain->devices);
2391 list_add(&info->global, &device_domain_list);
2392 if (dev)
2393 dev->archdata.iommu = info;
2394 spin_unlock_irqrestore(&device_domain_lock, flags);
2395
2396 if (dev && domain_context_mapping(domain, dev)) {
2397 pr_err("Domain context map for %s failed\n", dev_name(dev));
2398 dmar_remove_one_dev_info(domain, dev);
2399 return NULL;
2400 }
2401
2402 return domain;
2403 }
2404
2405 static int get_last_alias(struct pci_dev *pdev, u16 alias, void *opaque)
2406 {
2407 *(u16 *)opaque = alias;
2408 return 0;
2409 }
2410
2411 /* domain is initialized */
2412 static struct dmar_domain *get_domain_for_dev(struct device *dev, int gaw)
2413 {
2414 struct device_domain_info *info = NULL;
2415 struct dmar_domain *domain, *tmp;
2416 struct intel_iommu *iommu;
2417 u16 req_id, dma_alias;
2418 unsigned long flags;
2419 u8 bus, devfn;
2420
2421 domain = find_domain(dev);
2422 if (domain)
2423 return domain;
2424
2425 iommu = device_to_iommu(dev, &bus, &devfn);
2426 if (!iommu)
2427 return NULL;
2428
2429 req_id = ((u16)bus << 8) | devfn;
2430
2431 if (dev_is_pci(dev)) {
2432 struct pci_dev *pdev = to_pci_dev(dev);
2433
2434 pci_for_each_dma_alias(pdev, get_last_alias, &dma_alias);
2435
2436 spin_lock_irqsave(&device_domain_lock, flags);
2437 info = dmar_search_domain_by_dev_info(pci_domain_nr(pdev->bus),
2438 PCI_BUS_NUM(dma_alias),
2439 dma_alias & 0xff);
2440 if (info) {
2441 iommu = info->iommu;
2442 domain = info->domain;
2443 }
2444 spin_unlock_irqrestore(&device_domain_lock, flags);
2445
2446 /* DMA alias already has a domain, uses it */
2447 if (info)
2448 goto found_domain;
2449 }
2450
2451 /* Allocate and initialize new domain for the device */
2452 domain = alloc_domain(0);
2453 if (!domain)
2454 return NULL;
2455 if (domain_init(domain, iommu, gaw)) {
2456 domain_exit(domain);
2457 return NULL;
2458 }
2459
2460 /* register PCI DMA alias device */
2461 if (req_id != dma_alias && dev_is_pci(dev)) {
2462 tmp = dmar_insert_one_dev_info(iommu, PCI_BUS_NUM(dma_alias),
2463 dma_alias & 0xff, NULL, domain);
2464
2465 if (!tmp || tmp != domain) {
2466 domain_exit(domain);
2467 domain = tmp;
2468 }
2469
2470 if (!domain)
2471 return NULL;
2472 }
2473
2474 found_domain:
2475 tmp = dmar_insert_one_dev_info(iommu, bus, devfn, dev, domain);
2476
2477 if (!tmp || tmp != domain) {
2478 domain_exit(domain);
2479 domain = tmp;
2480 }
2481
2482 return domain;
2483 }
2484
2485 static int iommu_domain_identity_map(struct dmar_domain *domain,
2486 unsigned long long start,
2487 unsigned long long end)
2488 {
2489 unsigned long first_vpfn = start >> VTD_PAGE_SHIFT;
2490 unsigned long last_vpfn = end >> VTD_PAGE_SHIFT;
2491
2492 if (!reserve_iova(&domain->iovad, dma_to_mm_pfn(first_vpfn),
2493 dma_to_mm_pfn(last_vpfn))) {
2494 pr_err("Reserving iova failed\n");
2495 return -ENOMEM;
2496 }
2497
2498 pr_debug("Mapping reserved region %llx-%llx\n", start, end);
2499 /*
2500 * RMRR range might have overlap with physical memory range,
2501 * clear it first
2502 */
2503 dma_pte_clear_range(domain, first_vpfn, last_vpfn);
2504
2505 return domain_pfn_mapping(domain, first_vpfn, first_vpfn,
2506 last_vpfn - first_vpfn + 1,
2507 DMA_PTE_READ|DMA_PTE_WRITE);
2508 }
2509
2510 static int domain_prepare_identity_map(struct device *dev,
2511 struct dmar_domain *domain,
2512 unsigned long long start,
2513 unsigned long long end)
2514 {
2515 /* For _hardware_ passthrough, don't bother. But for software
2516 passthrough, we do it anyway -- it may indicate a memory
2517 range which is reserved in E820, so which didn't get set
2518 up to start with in si_domain */
2519 if (domain == si_domain && hw_pass_through) {
2520 pr_warn("Ignoring identity map for HW passthrough device %s [0x%Lx - 0x%Lx]\n",
2521 dev_name(dev), start, end);
2522 return 0;
2523 }
2524
2525 pr_info("Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
2526 dev_name(dev), start, end);
2527
2528 if (end < start) {
2529 WARN(1, "Your BIOS is broken; RMRR ends before it starts!\n"
2530 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
2531 dmi_get_system_info(DMI_BIOS_VENDOR),
2532 dmi_get_system_info(DMI_BIOS_VERSION),
2533 dmi_get_system_info(DMI_PRODUCT_VERSION));
2534 return -EIO;
2535 }
2536
2537 if (end >> agaw_to_width(domain->agaw)) {
2538 WARN(1, "Your BIOS is broken; RMRR exceeds permitted address width (%d bits)\n"
2539 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
2540 agaw_to_width(domain->agaw),
2541 dmi_get_system_info(DMI_BIOS_VENDOR),
2542 dmi_get_system_info(DMI_BIOS_VERSION),
2543 dmi_get_system_info(DMI_PRODUCT_VERSION));
2544 return -EIO;
2545 }
2546
2547 return iommu_domain_identity_map(domain, start, end);
2548 }
2549
2550 static int iommu_prepare_identity_map(struct device *dev,
2551 unsigned long long start,
2552 unsigned long long end)
2553 {
2554 struct dmar_domain *domain;
2555 int ret;
2556
2557 domain = get_domain_for_dev(dev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
2558 if (!domain)
2559 return -ENOMEM;
2560
2561 ret = domain_prepare_identity_map(dev, domain, start, end);
2562 if (ret)
2563 domain_exit(domain);
2564
2565 return ret;
2566 }
2567
2568 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
2569 struct device *dev)
2570 {
2571 if (dev->archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2572 return 0;
2573 return iommu_prepare_identity_map(dev, rmrr->base_address,
2574 rmrr->end_address);
2575 }
2576
2577 #ifdef CONFIG_INTEL_IOMMU_FLOPPY_WA
2578 static inline void iommu_prepare_isa(void)
2579 {
2580 struct pci_dev *pdev;
2581 int ret;
2582
2583 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
2584 if (!pdev)
2585 return;
2586
2587 pr_info("Prepare 0-16MiB unity mapping for LPC\n");
2588 ret = iommu_prepare_identity_map(&pdev->dev, 0, 16*1024*1024 - 1);
2589
2590 if (ret)
2591 pr_err("Failed to create 0-16MiB identity map - floppy might not work\n");
2592
2593 pci_dev_put(pdev);
2594 }
2595 #else
2596 static inline void iommu_prepare_isa(void)
2597 {
2598 return;
2599 }
2600 #endif /* !CONFIG_INTEL_IOMMU_FLPY_WA */
2601
2602 static int md_domain_init(struct dmar_domain *domain, int guest_width);
2603
2604 static int __init si_domain_init(int hw)
2605 {
2606 int nid, ret = 0;
2607
2608 si_domain = alloc_domain(DOMAIN_FLAG_STATIC_IDENTITY);
2609 if (!si_domain)
2610 return -EFAULT;
2611
2612 if (md_domain_init(si_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
2613 domain_exit(si_domain);
2614 return -EFAULT;
2615 }
2616
2617 pr_debug("Identity mapping domain allocated\n");
2618
2619 if (hw)
2620 return 0;
2621
2622 for_each_online_node(nid) {
2623 unsigned long start_pfn, end_pfn;
2624 int i;
2625
2626 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
2627 ret = iommu_domain_identity_map(si_domain,
2628 PFN_PHYS(start_pfn), PFN_PHYS(end_pfn));
2629 if (ret)
2630 return ret;
2631 }
2632 }
2633
2634 return 0;
2635 }
2636
2637 static int identity_mapping(struct device *dev)
2638 {
2639 struct device_domain_info *info;
2640
2641 if (likely(!iommu_identity_mapping))
2642 return 0;
2643
2644 info = dev->archdata.iommu;
2645 if (info && info != DUMMY_DEVICE_DOMAIN_INFO)
2646 return (info->domain == si_domain);
2647
2648 return 0;
2649 }
2650
2651 static int domain_add_dev_info(struct dmar_domain *domain, struct device *dev)
2652 {
2653 struct dmar_domain *ndomain;
2654 struct intel_iommu *iommu;
2655 u8 bus, devfn;
2656
2657 iommu = device_to_iommu(dev, &bus, &devfn);
2658 if (!iommu)
2659 return -ENODEV;
2660
2661 ndomain = dmar_insert_one_dev_info(iommu, bus, devfn, dev, domain);
2662 if (ndomain != domain)
2663 return -EBUSY;
2664
2665 return 0;
2666 }
2667
2668 static bool device_has_rmrr(struct device *dev)
2669 {
2670 struct dmar_rmrr_unit *rmrr;
2671 struct device *tmp;
2672 int i;
2673
2674 rcu_read_lock();
2675 for_each_rmrr_units(rmrr) {
2676 /*
2677 * Return TRUE if this RMRR contains the device that
2678 * is passed in.
2679 */
2680 for_each_active_dev_scope(rmrr->devices,
2681 rmrr->devices_cnt, i, tmp)
2682 if (tmp == dev) {
2683 rcu_read_unlock();
2684 return true;
2685 }
2686 }
2687 rcu_read_unlock();
2688 return false;
2689 }
2690
2691 /*
2692 * There are a couple cases where we need to restrict the functionality of
2693 * devices associated with RMRRs. The first is when evaluating a device for
2694 * identity mapping because problems exist when devices are moved in and out
2695 * of domains and their respective RMRR information is lost. This means that
2696 * a device with associated RMRRs will never be in a "passthrough" domain.
2697 * The second is use of the device through the IOMMU API. This interface
2698 * expects to have full control of the IOVA space for the device. We cannot
2699 * satisfy both the requirement that RMRR access is maintained and have an
2700 * unencumbered IOVA space. We also have no ability to quiesce the device's
2701 * use of the RMRR space or even inform the IOMMU API user of the restriction.
2702 * We therefore prevent devices associated with an RMRR from participating in
2703 * the IOMMU API, which eliminates them from device assignment.
2704 *
2705 * In both cases we assume that PCI USB devices with RMRRs have them largely
2706 * for historical reasons and that the RMRR space is not actively used post
2707 * boot. This exclusion may change if vendors begin to abuse it.
2708 *
2709 * The same exception is made for graphics devices, with the requirement that
2710 * any use of the RMRR regions will be torn down before assigning the device
2711 * to a guest.
2712 */
2713 static bool device_is_rmrr_locked(struct device *dev)
2714 {
2715 if (!device_has_rmrr(dev))
2716 return false;
2717
2718 if (dev_is_pci(dev)) {
2719 struct pci_dev *pdev = to_pci_dev(dev);
2720
2721 if (IS_USB_DEVICE(pdev) || IS_GFX_DEVICE(pdev))
2722 return false;
2723 }
2724
2725 return true;
2726 }
2727
2728 static int iommu_should_identity_map(struct device *dev, int startup)
2729 {
2730
2731 if (dev_is_pci(dev)) {
2732 struct pci_dev *pdev = to_pci_dev(dev);
2733
2734 if (device_is_rmrr_locked(dev))
2735 return 0;
2736
2737 if ((iommu_identity_mapping & IDENTMAP_AZALIA) && IS_AZALIA(pdev))
2738 return 1;
2739
2740 if ((iommu_identity_mapping & IDENTMAP_GFX) && IS_GFX_DEVICE(pdev))
2741 return 1;
2742
2743 if (!(iommu_identity_mapping & IDENTMAP_ALL))
2744 return 0;
2745
2746 /*
2747 * We want to start off with all devices in the 1:1 domain, and
2748 * take them out later if we find they can't access all of memory.
2749 *
2750 * However, we can't do this for PCI devices behind bridges,
2751 * because all PCI devices behind the same bridge will end up
2752 * with the same source-id on their transactions.
2753 *
2754 * Practically speaking, we can't change things around for these
2755 * devices at run-time, because we can't be sure there'll be no
2756 * DMA transactions in flight for any of their siblings.
2757 *
2758 * So PCI devices (unless they're on the root bus) as well as
2759 * their parent PCI-PCI or PCIe-PCI bridges must be left _out_ of
2760 * the 1:1 domain, just in _case_ one of their siblings turns out
2761 * not to be able to map all of memory.
2762 */
2763 if (!pci_is_pcie(pdev)) {
2764 if (!pci_is_root_bus(pdev->bus))
2765 return 0;
2766 if (pdev->class >> 8 == PCI_CLASS_BRIDGE_PCI)
2767 return 0;
2768 } else if (pci_pcie_type(pdev) == PCI_EXP_TYPE_PCI_BRIDGE)
2769 return 0;
2770 } else {
2771 if (device_has_rmrr(dev))
2772 return 0;
2773 }
2774
2775 /*
2776 * At boot time, we don't yet know if devices will be 64-bit capable.
2777 * Assume that they will — if they turn out not to be, then we can
2778 * take them out of the 1:1 domain later.
2779 */
2780 if (!startup) {
2781 /*
2782 * If the device's dma_mask is less than the system's memory
2783 * size then this is not a candidate for identity mapping.
2784 */
2785 u64 dma_mask = *dev->dma_mask;
2786
2787 if (dev->coherent_dma_mask &&
2788 dev->coherent_dma_mask < dma_mask)
2789 dma_mask = dev->coherent_dma_mask;
2790
2791 return dma_mask >= dma_get_required_mask(dev);
2792 }
2793
2794 return 1;
2795 }
2796
2797 static int __init dev_prepare_static_identity_mapping(struct device *dev, int hw)
2798 {
2799 int ret;
2800
2801 if (!iommu_should_identity_map(dev, 1))
2802 return 0;
2803
2804 ret = domain_add_dev_info(si_domain, dev);
2805 if (!ret)
2806 pr_info("%s identity mapping for device %s\n",
2807 hw ? "Hardware" : "Software", dev_name(dev));
2808 else if (ret == -ENODEV)
2809 /* device not associated with an iommu */
2810 ret = 0;
2811
2812 return ret;
2813 }
2814
2815
2816 static int __init iommu_prepare_static_identity_mapping(int hw)
2817 {
2818 struct pci_dev *pdev = NULL;
2819 struct dmar_drhd_unit *drhd;
2820 struct intel_iommu *iommu;
2821 struct device *dev;
2822 int i;
2823 int ret = 0;
2824
2825 for_each_pci_dev(pdev) {
2826 ret = dev_prepare_static_identity_mapping(&pdev->dev, hw);
2827 if (ret)
2828 return ret;
2829 }
2830
2831 for_each_active_iommu(iommu, drhd)
2832 for_each_active_dev_scope(drhd->devices, drhd->devices_cnt, i, dev) {
2833 struct acpi_device_physical_node *pn;
2834 struct acpi_device *adev;
2835
2836 if (dev->bus != &acpi_bus_type)
2837 continue;
2838
2839 adev= to_acpi_device(dev);
2840 mutex_lock(&adev->physical_node_lock);
2841 list_for_each_entry(pn, &adev->physical_node_list, node) {
2842 ret = dev_prepare_static_identity_mapping(pn->dev, hw);
2843 if (ret)
2844 break;
2845 }
2846 mutex_unlock(&adev->physical_node_lock);
2847 if (ret)
2848 return ret;
2849 }
2850
2851 return 0;
2852 }
2853
2854 static void intel_iommu_init_qi(struct intel_iommu *iommu)
2855 {
2856 /*
2857 * Start from the sane iommu hardware state.
2858 * If the queued invalidation is already initialized by us
2859 * (for example, while enabling interrupt-remapping) then
2860 * we got the things already rolling from a sane state.
2861 */
2862 if (!iommu->qi) {
2863 /*
2864 * Clear any previous faults.
2865 */
2866 dmar_fault(-1, iommu);
2867 /*
2868 * Disable queued invalidation if supported and already enabled
2869 * before OS handover.
2870 */
2871 dmar_disable_qi(iommu);
2872 }
2873
2874 if (dmar_enable_qi(iommu)) {
2875 /*
2876 * Queued Invalidate not enabled, use Register Based Invalidate
2877 */
2878 iommu->flush.flush_context = __iommu_flush_context;
2879 iommu->flush.flush_iotlb = __iommu_flush_iotlb;
2880 pr_info("%s: Using Register based invalidation\n",
2881 iommu->name);
2882 } else {
2883 iommu->flush.flush_context = qi_flush_context;
2884 iommu->flush.flush_iotlb = qi_flush_iotlb;
2885 pr_info("%s: Using Queued invalidation\n", iommu->name);
2886 }
2887 }
2888
2889 static int copy_context_table(struct intel_iommu *iommu,
2890 struct root_entry *old_re,
2891 struct context_entry **tbl,
2892 int bus, bool ext)
2893 {
2894 int tbl_idx, pos = 0, idx, devfn, ret = 0, did;
2895 struct context_entry *new_ce = NULL, ce;
2896 struct context_entry *old_ce = NULL;
2897 struct root_entry re;
2898 phys_addr_t old_ce_phys;
2899
2900 tbl_idx = ext ? bus * 2 : bus;
2901 memcpy(&re, old_re, sizeof(re));
2902
2903 for (devfn = 0; devfn < 256; devfn++) {
2904 /* First calculate the correct index */
2905 idx = (ext ? devfn * 2 : devfn) % 256;
2906
2907 if (idx == 0) {
2908 /* First save what we may have and clean up */
2909 if (new_ce) {
2910 tbl[tbl_idx] = new_ce;
2911 __iommu_flush_cache(iommu, new_ce,
2912 VTD_PAGE_SIZE);
2913 pos = 1;
2914 }
2915
2916 if (old_ce)
2917 iounmap(old_ce);
2918
2919 ret = 0;
2920 if (devfn < 0x80)
2921 old_ce_phys = root_entry_lctp(&re);
2922 else
2923 old_ce_phys = root_entry_uctp(&re);
2924
2925 if (!old_ce_phys) {
2926 if (ext && devfn == 0) {
2927 /* No LCTP, try UCTP */
2928 devfn = 0x7f;
2929 continue;
2930 } else {
2931 goto out;
2932 }
2933 }
2934
2935 ret = -ENOMEM;
2936 old_ce = memremap(old_ce_phys, PAGE_SIZE,
2937 MEMREMAP_WB);
2938 if (!old_ce)
2939 goto out;
2940
2941 new_ce = alloc_pgtable_page(iommu->node);
2942 if (!new_ce)
2943 goto out_unmap;
2944
2945 ret = 0;
2946 }
2947
2948 /* Now copy the context entry */
2949 memcpy(&ce, old_ce + idx, sizeof(ce));
2950
2951 if (!__context_present(&ce))
2952 continue;
2953
2954 did = context_domain_id(&ce);
2955 if (did >= 0 && did < cap_ndoms(iommu->cap))
2956 set_bit(did, iommu->domain_ids);
2957
2958 /*
2959 * We need a marker for copied context entries. This
2960 * marker needs to work for the old format as well as
2961 * for extended context entries.
2962 *
2963 * Bit 67 of the context entry is used. In the old
2964 * format this bit is available to software, in the
2965 * extended format it is the PGE bit, but PGE is ignored
2966 * by HW if PASIDs are disabled (and thus still
2967 * available).
2968 *
2969 * So disable PASIDs first and then mark the entry
2970 * copied. This means that we don't copy PASID
2971 * translations from the old kernel, but this is fine as
2972 * faults there are not fatal.
2973 */
2974 context_clear_pasid_enable(&ce);
2975 context_set_copied(&ce);
2976
2977 new_ce[idx] = ce;
2978 }
2979
2980 tbl[tbl_idx + pos] = new_ce;
2981
2982 __iommu_flush_cache(iommu, new_ce, VTD_PAGE_SIZE);
2983
2984 out_unmap:
2985 memunmap(old_ce);
2986
2987 out:
2988 return ret;
2989 }
2990
2991 static int copy_translation_tables(struct intel_iommu *iommu)
2992 {
2993 struct context_entry **ctxt_tbls;
2994 struct root_entry *old_rt;
2995 phys_addr_t old_rt_phys;
2996 int ctxt_table_entries;
2997 unsigned long flags;
2998 u64 rtaddr_reg;
2999 int bus, ret;
3000 bool new_ext, ext;
3001
3002 rtaddr_reg = dmar_readq(iommu->reg + DMAR_RTADDR_REG);
3003 ext = !!(rtaddr_reg & DMA_RTADDR_RTT);
3004 new_ext = !!ecap_ecs(iommu->ecap);
3005
3006 /*
3007 * The RTT bit can only be changed when translation is disabled,
3008 * but disabling translation means to open a window for data
3009 * corruption. So bail out and don't copy anything if we would
3010 * have to change the bit.
3011 */
3012 if (new_ext != ext)
3013 return -EINVAL;
3014
3015 old_rt_phys = rtaddr_reg & VTD_PAGE_MASK;
3016 if (!old_rt_phys)
3017 return -EINVAL;
3018
3019 old_rt = memremap(old_rt_phys, PAGE_SIZE, MEMREMAP_WB);
3020 if (!old_rt)
3021 return -ENOMEM;
3022
3023 /* This is too big for the stack - allocate it from slab */
3024 ctxt_table_entries = ext ? 512 : 256;
3025 ret = -ENOMEM;
3026 ctxt_tbls = kzalloc(ctxt_table_entries * sizeof(void *), GFP_KERNEL);
3027 if (!ctxt_tbls)
3028 goto out_unmap;
3029
3030 for (bus = 0; bus < 256; bus++) {
3031 ret = copy_context_table(iommu, &old_rt[bus],
3032 ctxt_tbls, bus, ext);
3033 if (ret) {
3034 pr_err("%s: Failed to copy context table for bus %d\n",
3035 iommu->name, bus);
3036 continue;
3037 }
3038 }
3039
3040 spin_lock_irqsave(&iommu->lock, flags);
3041
3042 /* Context tables are copied, now write them to the root_entry table */
3043 for (bus = 0; bus < 256; bus++) {
3044 int idx = ext ? bus * 2 : bus;
3045 u64 val;
3046
3047 if (ctxt_tbls[idx]) {
3048 val = virt_to_phys(ctxt_tbls[idx]) | 1;
3049 iommu->root_entry[bus].lo = val;
3050 }
3051
3052 if (!ext || !ctxt_tbls[idx + 1])
3053 continue;
3054
3055 val = virt_to_phys(ctxt_tbls[idx + 1]) | 1;
3056 iommu->root_entry[bus].hi = val;
3057 }
3058
3059 spin_unlock_irqrestore(&iommu->lock, flags);
3060
3061 kfree(ctxt_tbls);
3062
3063 __iommu_flush_cache(iommu, iommu->root_entry, PAGE_SIZE);
3064
3065 ret = 0;
3066
3067 out_unmap:
3068 memunmap(old_rt);
3069
3070 return ret;
3071 }
3072
3073 static int __init init_dmars(void)
3074 {
3075 struct dmar_drhd_unit *drhd;
3076 struct dmar_rmrr_unit *rmrr;
3077 bool copied_tables = false;
3078 struct device *dev;
3079 struct intel_iommu *iommu;
3080 int i, ret;
3081
3082 /*
3083 * for each drhd
3084 * allocate root
3085 * initialize and program root entry to not present
3086 * endfor
3087 */
3088 for_each_drhd_unit(drhd) {
3089 /*
3090 * lock not needed as this is only incremented in the single
3091 * threaded kernel __init code path all other access are read
3092 * only
3093 */
3094 if (g_num_of_iommus < DMAR_UNITS_SUPPORTED) {
3095 g_num_of_iommus++;
3096 continue;
3097 }
3098 pr_err_once("Exceeded %d IOMMUs\n", DMAR_UNITS_SUPPORTED);
3099 }
3100
3101 /* Preallocate enough resources for IOMMU hot-addition */
3102 if (g_num_of_iommus < DMAR_UNITS_SUPPORTED)
3103 g_num_of_iommus = DMAR_UNITS_SUPPORTED;
3104
3105 g_iommus = kcalloc(g_num_of_iommus, sizeof(struct intel_iommu *),
3106 GFP_KERNEL);
3107 if (!g_iommus) {
3108 pr_err("Allocating global iommu array failed\n");
3109 ret = -ENOMEM;
3110 goto error;
3111 }
3112
3113 deferred_flush = kzalloc(g_num_of_iommus *
3114 sizeof(struct deferred_flush_tables), GFP_KERNEL);
3115 if (!deferred_flush) {
3116 ret = -ENOMEM;
3117 goto free_g_iommus;
3118 }
3119
3120 for_each_active_iommu(iommu, drhd) {
3121 g_iommus[iommu->seq_id] = iommu;
3122
3123 intel_iommu_init_qi(iommu);
3124
3125 ret = iommu_init_domains(iommu);
3126 if (ret)
3127 goto free_iommu;
3128
3129 init_translation_status(iommu);
3130
3131 if (translation_pre_enabled(iommu) && !is_kdump_kernel()) {
3132 iommu_disable_translation(iommu);
3133 clear_translation_pre_enabled(iommu);
3134 pr_warn("Translation was enabled for %s but we are not in kdump mode\n",
3135 iommu->name);
3136 }
3137
3138 /*
3139 * TBD:
3140 * we could share the same root & context tables
3141 * among all IOMMU's. Need to Split it later.
3142 */
3143 ret = iommu_alloc_root_entry(iommu);
3144 if (ret)
3145 goto free_iommu;
3146
3147 if (translation_pre_enabled(iommu)) {
3148 pr_info("Translation already enabled - trying to copy translation structures\n");
3149
3150 ret = copy_translation_tables(iommu);
3151 if (ret) {
3152 /*
3153 * We found the IOMMU with translation
3154 * enabled - but failed to copy over the
3155 * old root-entry table. Try to proceed
3156 * by disabling translation now and
3157 * allocating a clean root-entry table.
3158 * This might cause DMAR faults, but
3159 * probably the dump will still succeed.
3160 */
3161 pr_err("Failed to copy translation tables from previous kernel for %s\n",
3162 iommu->name);
3163 iommu_disable_translation(iommu);
3164 clear_translation_pre_enabled(iommu);
3165 } else {
3166 pr_info("Copied translation tables from previous kernel for %s\n",
3167 iommu->name);
3168 copied_tables = true;
3169 }
3170 }
3171
3172 iommu_flush_write_buffer(iommu);
3173 iommu_set_root_entry(iommu);
3174 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
3175 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
3176
3177 if (!ecap_pass_through(iommu->ecap))
3178 hw_pass_through = 0;
3179 #ifdef CONFIG_INTEL_IOMMU_SVM
3180 if (pasid_enabled(iommu))
3181 intel_svm_alloc_pasid_tables(iommu);
3182 #endif
3183 }
3184
3185 if (iommu_pass_through)
3186 iommu_identity_mapping |= IDENTMAP_ALL;
3187
3188 #ifdef CONFIG_INTEL_IOMMU_BROKEN_GFX_WA
3189 iommu_identity_mapping |= IDENTMAP_GFX;
3190 #endif
3191
3192 if (iommu_identity_mapping) {
3193 ret = si_domain_init(hw_pass_through);
3194 if (ret)
3195 goto free_iommu;
3196 }
3197
3198 check_tylersburg_isoch();
3199
3200 /*
3201 * If we copied translations from a previous kernel in the kdump
3202 * case, we can not assign the devices to domains now, as that
3203 * would eliminate the old mappings. So skip this part and defer
3204 * the assignment to device driver initialization time.
3205 */
3206 if (copied_tables)
3207 goto domains_done;
3208
3209 /*
3210 * If pass through is not set or not enabled, setup context entries for
3211 * identity mappings for rmrr, gfx, and isa and may fall back to static
3212 * identity mapping if iommu_identity_mapping is set.
3213 */
3214 if (iommu_identity_mapping) {
3215 ret = iommu_prepare_static_identity_mapping(hw_pass_through);
3216 if (ret) {
3217 pr_crit("Failed to setup IOMMU pass-through\n");
3218 goto free_iommu;
3219 }
3220 }
3221 /*
3222 * For each rmrr
3223 * for each dev attached to rmrr
3224 * do
3225 * locate drhd for dev, alloc domain for dev
3226 * allocate free domain
3227 * allocate page table entries for rmrr
3228 * if context not allocated for bus
3229 * allocate and init context
3230 * set present in root table for this bus
3231 * init context with domain, translation etc
3232 * endfor
3233 * endfor
3234 */
3235 pr_info("Setting RMRR:\n");
3236 for_each_rmrr_units(rmrr) {
3237 /* some BIOS lists non-exist devices in DMAR table. */
3238 for_each_active_dev_scope(rmrr->devices, rmrr->devices_cnt,
3239 i, dev) {
3240 ret = iommu_prepare_rmrr_dev(rmrr, dev);
3241 if (ret)
3242 pr_err("Mapping reserved region failed\n");
3243 }
3244 }
3245
3246 iommu_prepare_isa();
3247
3248 domains_done:
3249
3250 /*
3251 * for each drhd
3252 * enable fault log
3253 * global invalidate context cache
3254 * global invalidate iotlb
3255 * enable translation
3256 */
3257 for_each_iommu(iommu, drhd) {
3258 if (drhd->ignored) {
3259 /*
3260 * we always have to disable PMRs or DMA may fail on
3261 * this device
3262 */
3263 if (force_on)
3264 iommu_disable_protect_mem_regions(iommu);
3265 continue;
3266 }
3267
3268 iommu_flush_write_buffer(iommu);
3269
3270 #ifdef CONFIG_INTEL_IOMMU_SVM
3271 if (pasid_enabled(iommu) && ecap_prs(iommu->ecap)) {
3272 ret = intel_svm_enable_prq(iommu);
3273 if (ret)
3274 goto free_iommu;
3275 }
3276 #endif
3277 ret = dmar_set_interrupt(iommu);
3278 if (ret)
3279 goto free_iommu;
3280
3281 if (!translation_pre_enabled(iommu))
3282 iommu_enable_translation(iommu);
3283
3284 iommu_disable_protect_mem_regions(iommu);
3285 }
3286
3287 return 0;
3288
3289 free_iommu:
3290 for_each_active_iommu(iommu, drhd) {
3291 disable_dmar_iommu(iommu);
3292 free_dmar_iommu(iommu);
3293 }
3294 kfree(deferred_flush);
3295 free_g_iommus:
3296 kfree(g_iommus);
3297 error:
3298 return ret;
3299 }
3300
3301 /* This takes a number of _MM_ pages, not VTD pages */
3302 static struct iova *intel_alloc_iova(struct device *dev,
3303 struct dmar_domain *domain,
3304 unsigned long nrpages, uint64_t dma_mask)
3305 {
3306 struct iova *iova = NULL;
3307
3308 /* Restrict dma_mask to the width that the iommu can handle */
3309 dma_mask = min_t(uint64_t, DOMAIN_MAX_ADDR(domain->gaw), dma_mask);
3310 /* Ensure we reserve the whole size-aligned region */
3311 nrpages = __roundup_pow_of_two(nrpages);
3312
3313 if (!dmar_forcedac && dma_mask > DMA_BIT_MASK(32)) {
3314 /*
3315 * First try to allocate an io virtual address in
3316 * DMA_BIT_MASK(32) and if that fails then try allocating
3317 * from higher range
3318 */
3319 iova = alloc_iova(&domain->iovad, nrpages,
3320 IOVA_PFN(DMA_BIT_MASK(32)), 1);
3321 if (iova)
3322 return iova;
3323 }
3324 iova = alloc_iova(&domain->iovad, nrpages, IOVA_PFN(dma_mask), 1);
3325 if (unlikely(!iova)) {
3326 pr_err("Allocating %ld-page iova for %s failed",
3327 nrpages, dev_name(dev));
3328 return NULL;
3329 }
3330
3331 return iova;
3332 }
3333
3334 static struct dmar_domain *__get_valid_domain_for_dev(struct device *dev)
3335 {
3336 struct dmar_rmrr_unit *rmrr;
3337 struct dmar_domain *domain;
3338 struct device *i_dev;
3339 int i, ret;
3340
3341 domain = get_domain_for_dev(dev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
3342 if (!domain) {
3343 pr_err("Allocating domain for %s failed\n",
3344 dev_name(dev));
3345 return NULL;
3346 }
3347
3348 /* We have a new domain - setup possible RMRRs for the device */
3349 rcu_read_lock();
3350 for_each_rmrr_units(rmrr) {
3351 for_each_active_dev_scope(rmrr->devices, rmrr->devices_cnt,
3352 i, i_dev) {
3353 if (i_dev != dev)
3354 continue;
3355
3356 ret = domain_prepare_identity_map(dev, domain,
3357 rmrr->base_address,
3358 rmrr->end_address);
3359 if (ret)
3360 dev_err(dev, "Mapping reserved region failed\n");
3361 }
3362 }
3363 rcu_read_unlock();
3364
3365 return domain;
3366 }
3367
3368 static inline struct dmar_domain *get_valid_domain_for_dev(struct device *dev)
3369 {
3370 struct device_domain_info *info;
3371
3372 /* No lock here, assumes no domain exit in normal case */
3373 info = dev->archdata.iommu;
3374 if (likely(info))
3375 return info->domain;
3376
3377 return __get_valid_domain_for_dev(dev);
3378 }
3379
3380 /* Check if the dev needs to go through non-identity map and unmap process.*/
3381 static int iommu_no_mapping(struct device *dev)
3382 {
3383 int found;
3384
3385 if (iommu_dummy(dev))
3386 return 1;
3387
3388 if (!iommu_identity_mapping)
3389 return 0;
3390
3391 found = identity_mapping(dev);
3392 if (found) {
3393 if (iommu_should_identity_map(dev, 0))
3394 return 1;
3395 else {
3396 /*
3397 * 32 bit DMA is removed from si_domain and fall back
3398 * to non-identity mapping.
3399 */
3400 dmar_remove_one_dev_info(si_domain, dev);
3401 pr_info("32bit %s uses non-identity mapping\n",
3402 dev_name(dev));
3403 return 0;
3404 }
3405 } else {
3406 /*
3407 * In case of a detached 64 bit DMA device from vm, the device
3408 * is put into si_domain for identity mapping.
3409 */
3410 if (iommu_should_identity_map(dev, 0)) {
3411 int ret;
3412 ret = domain_add_dev_info(si_domain, dev);
3413 if (!ret) {
3414 pr_info("64bit %s uses identity mapping\n",
3415 dev_name(dev));
3416 return 1;
3417 }
3418 }
3419 }
3420
3421 return 0;
3422 }
3423
3424 static dma_addr_t __intel_map_single(struct device *dev, phys_addr_t paddr,
3425 size_t size, int dir, u64 dma_mask)
3426 {
3427 struct dmar_domain *domain;
3428 phys_addr_t start_paddr;
3429 struct iova *iova;
3430 int prot = 0;
3431 int ret;
3432 struct intel_iommu *iommu;
3433 unsigned long paddr_pfn = paddr >> PAGE_SHIFT;
3434
3435 BUG_ON(dir == DMA_NONE);
3436
3437 if (iommu_no_mapping(dev))
3438 return paddr;
3439
3440 domain = get_valid_domain_for_dev(dev);
3441 if (!domain)
3442 return 0;
3443
3444 iommu = domain_get_iommu(domain);
3445 size = aligned_nrpages(paddr, size);
3446
3447 iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size), dma_mask);
3448 if (!iova)
3449 goto error;
3450
3451 /*
3452 * Check if DMAR supports zero-length reads on write only
3453 * mappings..
3454 */
3455 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
3456 !cap_zlr(iommu->cap))
3457 prot |= DMA_PTE_READ;
3458 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
3459 prot |= DMA_PTE_WRITE;
3460 /*
3461 * paddr - (paddr + size) might be partial page, we should map the whole
3462 * page. Note: if two part of one page are separately mapped, we
3463 * might have two guest_addr mapping to the same host paddr, but this
3464 * is not a big problem
3465 */
3466 ret = domain_pfn_mapping(domain, mm_to_dma_pfn(iova->pfn_lo),
3467 mm_to_dma_pfn(paddr_pfn), size, prot);
3468 if (ret)
3469 goto error;
3470
3471 /* it's a non-present to present mapping. Only flush if caching mode */
3472 if (cap_caching_mode(iommu->cap))
3473 iommu_flush_iotlb_psi(iommu, domain,
3474 mm_to_dma_pfn(iova->pfn_lo),
3475 size, 0, 1);
3476 else
3477 iommu_flush_write_buffer(iommu);
3478
3479 start_paddr = (phys_addr_t)iova->pfn_lo << PAGE_SHIFT;
3480 start_paddr += paddr & ~PAGE_MASK;
3481 return start_paddr;
3482
3483 error:
3484 if (iova)
3485 __free_iova(&domain->iovad, iova);
3486 pr_err("Device %s request: %zx@%llx dir %d --- failed\n",
3487 dev_name(dev), size, (unsigned long long)paddr, dir);
3488 return 0;
3489 }
3490
3491 static dma_addr_t intel_map_page(struct device *dev, struct page *page,
3492 unsigned long offset, size_t size,
3493 enum dma_data_direction dir,
3494 struct dma_attrs *attrs)
3495 {
3496 return __intel_map_single(dev, page_to_phys(page) + offset, size,
3497 dir, *dev->dma_mask);
3498 }
3499
3500 static void flush_unmaps(void)
3501 {
3502 int i, j;
3503
3504 timer_on = 0;
3505
3506 /* just flush them all */
3507 for (i = 0; i < g_num_of_iommus; i++) {
3508 struct intel_iommu *iommu = g_iommus[i];
3509 if (!iommu)
3510 continue;
3511
3512 if (!deferred_flush[i].next)
3513 continue;
3514
3515 /* In caching mode, global flushes turn emulation expensive */
3516 if (!cap_caching_mode(iommu->cap))
3517 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3518 DMA_TLB_GLOBAL_FLUSH);
3519 for (j = 0; j < deferred_flush[i].next; j++) {
3520 unsigned long mask;
3521 struct iova *iova = deferred_flush[i].iova[j];
3522 struct dmar_domain *domain = deferred_flush[i].domain[j];
3523
3524 /* On real hardware multiple invalidations are expensive */
3525 if (cap_caching_mode(iommu->cap))
3526 iommu_flush_iotlb_psi(iommu, domain,
3527 iova->pfn_lo, iova_size(iova),
3528 !deferred_flush[i].freelist[j], 0);
3529 else {
3530 mask = ilog2(mm_to_dma_pfn(iova_size(iova)));
3531 iommu_flush_dev_iotlb(deferred_flush[i].domain[j],
3532 (uint64_t)iova->pfn_lo << PAGE_SHIFT, mask);
3533 }
3534 __free_iova(&deferred_flush[i].domain[j]->iovad, iova);
3535 if (deferred_flush[i].freelist[j])
3536 dma_free_pagelist(deferred_flush[i].freelist[j]);
3537 }
3538 deferred_flush[i].next = 0;
3539 }
3540
3541 list_size = 0;
3542 }
3543
3544 static void flush_unmaps_timeout(unsigned long data)
3545 {
3546 unsigned long flags;
3547
3548 spin_lock_irqsave(&async_umap_flush_lock, flags);
3549 flush_unmaps();
3550 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
3551 }
3552
3553 static void add_unmap(struct dmar_domain *dom, struct iova *iova, struct page *freelist)
3554 {
3555 unsigned long flags;
3556 int next, iommu_id;
3557 struct intel_iommu *iommu;
3558
3559 spin_lock_irqsave(&async_umap_flush_lock, flags);
3560 if (list_size == HIGH_WATER_MARK)
3561 flush_unmaps();
3562
3563 iommu = domain_get_iommu(dom);
3564 iommu_id = iommu->seq_id;
3565
3566 next = deferred_flush[iommu_id].next;
3567 deferred_flush[iommu_id].domain[next] = dom;
3568 deferred_flush[iommu_id].iova[next] = iova;
3569 deferred_flush[iommu_id].freelist[next] = freelist;
3570 deferred_flush[iommu_id].next++;
3571
3572 if (!timer_on) {
3573 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
3574 timer_on = 1;
3575 }
3576 list_size++;
3577 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
3578 }
3579
3580 static void intel_unmap(struct device *dev, dma_addr_t dev_addr)
3581 {
3582 struct dmar_domain *domain;
3583 unsigned long start_pfn, last_pfn;
3584 struct iova *iova;
3585 struct intel_iommu *iommu;
3586 struct page *freelist;
3587
3588 if (iommu_no_mapping(dev))
3589 return;
3590
3591 domain = find_domain(dev);
3592 BUG_ON(!domain);
3593
3594 iommu = domain_get_iommu(domain);
3595
3596 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
3597 if (WARN_ONCE(!iova, "Driver unmaps unmatched page at PFN %llx\n",
3598 (unsigned long long)dev_addr))
3599 return;
3600
3601 start_pfn = mm_to_dma_pfn(iova->pfn_lo);
3602 last_pfn = mm_to_dma_pfn(iova->pfn_hi + 1) - 1;
3603
3604 pr_debug("Device %s unmapping: pfn %lx-%lx\n",
3605 dev_name(dev), start_pfn, last_pfn);
3606
3607 freelist = domain_unmap(domain, start_pfn, last_pfn);
3608
3609 if (intel_iommu_strict) {
3610 iommu_flush_iotlb_psi(iommu, domain, start_pfn,
3611 last_pfn - start_pfn + 1, !freelist, 0);
3612 /* free iova */
3613 __free_iova(&domain->iovad, iova);
3614 dma_free_pagelist(freelist);
3615 } else {
3616 add_unmap(domain, iova, freelist);
3617 /*
3618 * queue up the release of the unmap to save the 1/6th of the
3619 * cpu used up by the iotlb flush operation...
3620 */
3621 }
3622 }
3623
3624 static void intel_unmap_page(struct device *dev, dma_addr_t dev_addr,
3625 size_t size, enum dma_data_direction dir,
3626 struct dma_attrs *attrs)
3627 {
3628 intel_unmap(dev, dev_addr);
3629 }
3630
3631 static void *intel_alloc_coherent(struct device *dev, size_t size,
3632 dma_addr_t *dma_handle, gfp_t flags,
3633 struct dma_attrs *attrs)
3634 {
3635 struct page *page = NULL;
3636 int order;
3637
3638 size = PAGE_ALIGN(size);
3639 order = get_order(size);
3640
3641 if (!iommu_no_mapping(dev))
3642 flags &= ~(GFP_DMA | GFP_DMA32);
3643 else if (dev->coherent_dma_mask < dma_get_required_mask(dev)) {
3644 if (dev->coherent_dma_mask < DMA_BIT_MASK(32))
3645 flags |= GFP_DMA;
3646 else
3647 flags |= GFP_DMA32;
3648 }
3649
3650 if (gfpflags_allow_blocking(flags)) {
3651 unsigned int count = size >> PAGE_SHIFT;
3652
3653 page = dma_alloc_from_contiguous(dev, count, order);
3654 if (page && iommu_no_mapping(dev) &&
3655 page_to_phys(page) + size > dev->coherent_dma_mask) {
3656 dma_release_from_contiguous(dev, page, count);
3657 page = NULL;
3658 }
3659 }
3660
3661 if (!page)
3662 page = alloc_pages(flags, order);
3663 if (!page)
3664 return NULL;
3665 memset(page_address(page), 0, size);
3666
3667 *dma_handle = __intel_map_single(dev, page_to_phys(page), size,
3668 DMA_BIDIRECTIONAL,
3669 dev->coherent_dma_mask);
3670 if (*dma_handle)
3671 return page_address(page);
3672 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3673 __free_pages(page, order);
3674
3675 return NULL;
3676 }
3677
3678 static void intel_free_coherent(struct device *dev, size_t size, void *vaddr,
3679 dma_addr_t dma_handle, struct dma_attrs *attrs)
3680 {
3681 int order;
3682 struct page *page = virt_to_page(vaddr);
3683
3684 size = PAGE_ALIGN(size);
3685 order = get_order(size);
3686
3687 intel_unmap(dev, dma_handle);
3688 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
3689 __free_pages(page, order);
3690 }
3691
3692 static void intel_unmap_sg(struct device *dev, struct scatterlist *sglist,
3693 int nelems, enum dma_data_direction dir,
3694 struct dma_attrs *attrs)
3695 {
3696 intel_unmap(dev, sglist[0].dma_address);
3697 }
3698
3699 static int intel_nontranslate_map_sg(struct device *hddev,
3700 struct scatterlist *sglist, int nelems, int dir)
3701 {
3702 int i;
3703 struct scatterlist *sg;
3704
3705 for_each_sg(sglist, sg, nelems, i) {
3706 BUG_ON(!sg_page(sg));
3707 sg->dma_address = page_to_phys(sg_page(sg)) + sg->offset;
3708 sg->dma_length = sg->length;
3709 }
3710 return nelems;
3711 }
3712
3713 static int intel_map_sg(struct device *dev, struct scatterlist *sglist, int nelems,
3714 enum dma_data_direction dir, struct dma_attrs *attrs)
3715 {
3716 int i;
3717 struct dmar_domain *domain;
3718 size_t size = 0;
3719 int prot = 0;
3720 struct iova *iova = NULL;
3721 int ret;
3722 struct scatterlist *sg;
3723 unsigned long start_vpfn;
3724 struct intel_iommu *iommu;
3725
3726 BUG_ON(dir == DMA_NONE);
3727 if (iommu_no_mapping(dev))
3728 return intel_nontranslate_map_sg(dev, sglist, nelems, dir);
3729
3730 domain = get_valid_domain_for_dev(dev);
3731 if (!domain)
3732 return 0;
3733
3734 iommu = domain_get_iommu(domain);
3735
3736 for_each_sg(sglist, sg, nelems, i)
3737 size += aligned_nrpages(sg->offset, sg->length);
3738
3739 iova = intel_alloc_iova(dev, domain, dma_to_mm_pfn(size),
3740 *dev->dma_mask);
3741 if (!iova) {
3742 sglist->dma_length = 0;
3743 return 0;
3744 }
3745
3746 /*
3747 * Check if DMAR supports zero-length reads on write only
3748 * mappings..
3749 */
3750 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
3751 !cap_zlr(iommu->cap))
3752 prot |= DMA_PTE_READ;
3753 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
3754 prot |= DMA_PTE_WRITE;
3755
3756 start_vpfn = mm_to_dma_pfn(iova->pfn_lo);
3757
3758 ret = domain_sg_mapping(domain, start_vpfn, sglist, size, prot);
3759 if (unlikely(ret)) {
3760 dma_pte_free_pagetable(domain, start_vpfn,
3761 start_vpfn + size - 1);
3762 __free_iova(&domain->iovad, iova);
3763 return 0;
3764 }
3765
3766 /* it's a non-present to present mapping. Only flush if caching mode */
3767 if (cap_caching_mode(iommu->cap))
3768 iommu_flush_iotlb_psi(iommu, domain, start_vpfn, size, 0, 1);
3769 else
3770 iommu_flush_write_buffer(iommu);
3771
3772 return nelems;
3773 }
3774
3775 static int intel_mapping_error(struct device *dev, dma_addr_t dma_addr)
3776 {
3777 return !dma_addr;
3778 }
3779
3780 struct dma_map_ops intel_dma_ops = {
3781 .alloc = intel_alloc_coherent,
3782 .free = intel_free_coherent,
3783 .map_sg = intel_map_sg,
3784 .unmap_sg = intel_unmap_sg,
3785 .map_page = intel_map_page,
3786 .unmap_page = intel_unmap_page,
3787 .mapping_error = intel_mapping_error,
3788 };
3789
3790 static inline int iommu_domain_cache_init(void)
3791 {
3792 int ret = 0;
3793
3794 iommu_domain_cache = kmem_cache_create("iommu_domain",
3795 sizeof(struct dmar_domain),
3796 0,
3797 SLAB_HWCACHE_ALIGN,
3798
3799 NULL);
3800 if (!iommu_domain_cache) {
3801 pr_err("Couldn't create iommu_domain cache\n");
3802 ret = -ENOMEM;
3803 }
3804
3805 return ret;
3806 }
3807
3808 static inline int iommu_devinfo_cache_init(void)
3809 {
3810 int ret = 0;
3811
3812 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
3813 sizeof(struct device_domain_info),
3814 0,
3815 SLAB_HWCACHE_ALIGN,
3816 NULL);
3817 if (!iommu_devinfo_cache) {
3818 pr_err("Couldn't create devinfo cache\n");
3819 ret = -ENOMEM;
3820 }
3821
3822 return ret;
3823 }
3824
3825 static int __init iommu_init_mempool(void)
3826 {
3827 int ret;
3828 ret = iova_cache_get();
3829 if (ret)
3830 return ret;
3831
3832 ret = iommu_domain_cache_init();
3833 if (ret)
3834 goto domain_error;
3835
3836 ret = iommu_devinfo_cache_init();
3837 if (!ret)
3838 return ret;
3839
3840 kmem_cache_destroy(iommu_domain_cache);
3841 domain_error:
3842 iova_cache_put();
3843
3844 return -ENOMEM;
3845 }
3846
3847 static void __init iommu_exit_mempool(void)
3848 {
3849 kmem_cache_destroy(iommu_devinfo_cache);
3850 kmem_cache_destroy(iommu_domain_cache);
3851 iova_cache_put();
3852 }
3853
3854 static void quirk_ioat_snb_local_iommu(struct pci_dev *pdev)
3855 {
3856 struct dmar_drhd_unit *drhd;
3857 u32 vtbar;
3858 int rc;
3859
3860 /* We know that this device on this chipset has its own IOMMU.
3861 * If we find it under a different IOMMU, then the BIOS is lying
3862 * to us. Hope that the IOMMU for this device is actually
3863 * disabled, and it needs no translation...
3864 */
3865 rc = pci_bus_read_config_dword(pdev->bus, PCI_DEVFN(0, 0), 0xb0, &vtbar);
3866 if (rc) {
3867 /* "can't" happen */
3868 dev_info(&pdev->dev, "failed to run vt-d quirk\n");
3869 return;
3870 }
3871 vtbar &= 0xffff0000;
3872
3873 /* we know that the this iommu should be at offset 0xa000 from vtbar */
3874 drhd = dmar_find_matched_drhd_unit(pdev);
3875 if (WARN_TAINT_ONCE(!drhd || drhd->reg_base_addr - vtbar != 0xa000,
3876 TAINT_FIRMWARE_WORKAROUND,
3877 "BIOS assigned incorrect VT-d unit for Intel(R) QuickData Technology device\n"))
3878 pdev->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3879 }
3880 DECLARE_PCI_FIXUP_ENABLE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_IOAT_SNB, quirk_ioat_snb_local_iommu);
3881
3882 static void __init init_no_remapping_devices(void)
3883 {
3884 struct dmar_drhd_unit *drhd;
3885 struct device *dev;
3886 int i;
3887
3888 for_each_drhd_unit(drhd) {
3889 if (!drhd->include_all) {
3890 for_each_active_dev_scope(drhd->devices,
3891 drhd->devices_cnt, i, dev)
3892 break;
3893 /* ignore DMAR unit if no devices exist */
3894 if (i == drhd->devices_cnt)
3895 drhd->ignored = 1;
3896 }
3897 }
3898
3899 for_each_active_drhd_unit(drhd) {
3900 if (drhd->include_all)
3901 continue;
3902
3903 for_each_active_dev_scope(drhd->devices,
3904 drhd->devices_cnt, i, dev)
3905 if (!dev_is_pci(dev) || !IS_GFX_DEVICE(to_pci_dev(dev)))
3906 break;
3907 if (i < drhd->devices_cnt)
3908 continue;
3909
3910 /* This IOMMU has *only* gfx devices. Either bypass it or
3911 set the gfx_mapped flag, as appropriate */
3912 if (dmar_map_gfx) {
3913 intel_iommu_gfx_mapped = 1;
3914 } else {
3915 drhd->ignored = 1;
3916 for_each_active_dev_scope(drhd->devices,
3917 drhd->devices_cnt, i, dev)
3918 dev->archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
3919 }
3920 }
3921 }
3922
3923 #ifdef CONFIG_SUSPEND
3924 static int init_iommu_hw(void)
3925 {
3926 struct dmar_drhd_unit *drhd;
3927 struct intel_iommu *iommu = NULL;
3928
3929 for_each_active_iommu(iommu, drhd)
3930 if (iommu->qi)
3931 dmar_reenable_qi(iommu);
3932
3933 for_each_iommu(iommu, drhd) {
3934 if (drhd->ignored) {
3935 /*
3936 * we always have to disable PMRs or DMA may fail on
3937 * this device
3938 */
3939 if (force_on)
3940 iommu_disable_protect_mem_regions(iommu);
3941 continue;
3942 }
3943
3944 iommu_flush_write_buffer(iommu);
3945
3946 iommu_set_root_entry(iommu);
3947
3948 iommu->flush.flush_context(iommu, 0, 0, 0,
3949 DMA_CCMD_GLOBAL_INVL);
3950 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
3951 iommu_enable_translation(iommu);
3952 iommu_disable_protect_mem_regions(iommu);
3953 }
3954
3955 return 0;
3956 }
3957
3958 static void iommu_flush_all(void)
3959 {
3960 struct dmar_drhd_unit *drhd;
3961 struct intel_iommu *iommu;
3962
3963 for_each_active_iommu(iommu, drhd) {
3964 iommu->flush.flush_context(iommu, 0, 0, 0,
3965 DMA_CCMD_GLOBAL_INVL);
3966 iommu->flush.flush_iotlb(iommu, 0, 0, 0,
3967 DMA_TLB_GLOBAL_FLUSH);
3968 }
3969 }
3970
3971 static int iommu_suspend(void)
3972 {
3973 struct dmar_drhd_unit *drhd;
3974 struct intel_iommu *iommu = NULL;
3975 unsigned long flag;
3976
3977 for_each_active_iommu(iommu, drhd) {
3978 iommu->iommu_state = kzalloc(sizeof(u32) * MAX_SR_DMAR_REGS,
3979 GFP_ATOMIC);
3980 if (!iommu->iommu_state)
3981 goto nomem;
3982 }
3983
3984 iommu_flush_all();
3985
3986 for_each_active_iommu(iommu, drhd) {
3987 iommu_disable_translation(iommu);
3988
3989 raw_spin_lock_irqsave(&iommu->register_lock, flag);
3990
3991 iommu->iommu_state[SR_DMAR_FECTL_REG] =
3992 readl(iommu->reg + DMAR_FECTL_REG);
3993 iommu->iommu_state[SR_DMAR_FEDATA_REG] =
3994 readl(iommu->reg + DMAR_FEDATA_REG);
3995 iommu->iommu_state[SR_DMAR_FEADDR_REG] =
3996 readl(iommu->reg + DMAR_FEADDR_REG);
3997 iommu->iommu_state[SR_DMAR_FEUADDR_REG] =
3998 readl(iommu->reg + DMAR_FEUADDR_REG);
3999
4000 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
4001 }
4002 return 0;
4003
4004 nomem:
4005 for_each_active_iommu(iommu, drhd)
4006 kfree(iommu->iommu_state);
4007
4008 return -ENOMEM;
4009 }
4010
4011 static void iommu_resume(void)
4012 {
4013 struct dmar_drhd_unit *drhd;
4014 struct intel_iommu *iommu = NULL;
4015 unsigned long flag;
4016
4017 if (init_iommu_hw()) {
4018 if (force_on)
4019 panic("tboot: IOMMU setup failed, DMAR can not resume!\n");
4020 else
4021 WARN(1, "IOMMU setup failed, DMAR can not resume!\n");
4022 return;
4023 }
4024
4025 for_each_active_iommu(iommu, drhd) {
4026
4027 raw_spin_lock_irqsave(&iommu->register_lock, flag);
4028
4029 writel(iommu->iommu_state[SR_DMAR_FECTL_REG],
4030 iommu->reg + DMAR_FECTL_REG);
4031 writel(iommu->iommu_state[SR_DMAR_FEDATA_REG],
4032 iommu->reg + DMAR_FEDATA_REG);
4033 writel(iommu->iommu_state[SR_DMAR_FEADDR_REG],
4034 iommu->reg + DMAR_FEADDR_REG);
4035 writel(iommu->iommu_state[SR_DMAR_FEUADDR_REG],
4036 iommu->reg + DMAR_FEUADDR_REG);
4037
4038 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
4039 }
4040
4041 for_each_active_iommu(iommu, drhd)
4042 kfree(iommu->iommu_state);
4043 }
4044
4045 static struct syscore_ops iommu_syscore_ops = {
4046 .resume = iommu_resume,
4047 .suspend = iommu_suspend,
4048 };
4049
4050 static void __init init_iommu_pm_ops(void)
4051 {
4052 register_syscore_ops(&iommu_syscore_ops);
4053 }
4054
4055 #else
4056 static inline void init_iommu_pm_ops(void) {}
4057 #endif /* CONFIG_PM */
4058
4059
4060 int __init dmar_parse_one_rmrr(struct acpi_dmar_header *header, void *arg)
4061 {
4062 struct acpi_dmar_reserved_memory *rmrr;
4063 struct dmar_rmrr_unit *rmrru;
4064
4065 rmrru = kzalloc(sizeof(*rmrru), GFP_KERNEL);
4066 if (!rmrru)
4067 return -ENOMEM;
4068
4069 rmrru->hdr = header;
4070 rmrr = (struct acpi_dmar_reserved_memory *)header;
4071 rmrru->base_address = rmrr->base_address;
4072 rmrru->end_address = rmrr->end_address;
4073 rmrru->devices = dmar_alloc_dev_scope((void *)(rmrr + 1),
4074 ((void *)rmrr) + rmrr->header.length,
4075 &rmrru->devices_cnt);
4076 if (rmrru->devices_cnt && rmrru->devices == NULL) {
4077 kfree(rmrru);
4078 return -ENOMEM;
4079 }
4080
4081 list_add(&rmrru->list, &dmar_rmrr_units);
4082
4083 return 0;
4084 }
4085
4086 static struct dmar_atsr_unit *dmar_find_atsr(struct acpi_dmar_atsr *atsr)
4087 {
4088 struct dmar_atsr_unit *atsru;
4089 struct acpi_dmar_atsr *tmp;
4090
4091 list_for_each_entry_rcu(atsru, &dmar_atsr_units, list) {
4092 tmp = (struct acpi_dmar_atsr *)atsru->hdr;
4093 if (atsr->segment != tmp->segment)
4094 continue;
4095 if (atsr->header.length != tmp->header.length)
4096 continue;
4097 if (memcmp(atsr, tmp, atsr->header.length) == 0)
4098 return atsru;
4099 }
4100
4101 return NULL;
4102 }
4103
4104 int dmar_parse_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4105 {
4106 struct acpi_dmar_atsr *atsr;
4107 struct dmar_atsr_unit *atsru;
4108
4109 if (system_state != SYSTEM_BOOTING && !intel_iommu_enabled)
4110 return 0;
4111
4112 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4113 atsru = dmar_find_atsr(atsr);
4114 if (atsru)
4115 return 0;
4116
4117 atsru = kzalloc(sizeof(*atsru) + hdr->length, GFP_KERNEL);
4118 if (!atsru)
4119 return -ENOMEM;
4120
4121 /*
4122 * If memory is allocated from slab by ACPI _DSM method, we need to
4123 * copy the memory content because the memory buffer will be freed
4124 * on return.
4125 */
4126 atsru->hdr = (void *)(atsru + 1);
4127 memcpy(atsru->hdr, hdr, hdr->length);
4128 atsru->include_all = atsr->flags & 0x1;
4129 if (!atsru->include_all) {
4130 atsru->devices = dmar_alloc_dev_scope((void *)(atsr + 1),
4131 (void *)atsr + atsr->header.length,
4132 &atsru->devices_cnt);
4133 if (atsru->devices_cnt && atsru->devices == NULL) {
4134 kfree(atsru);
4135 return -ENOMEM;
4136 }
4137 }
4138
4139 list_add_rcu(&atsru->list, &dmar_atsr_units);
4140
4141 return 0;
4142 }
4143
4144 static void intel_iommu_free_atsr(struct dmar_atsr_unit *atsru)
4145 {
4146 dmar_free_dev_scope(&atsru->devices, &atsru->devices_cnt);
4147 kfree(atsru);
4148 }
4149
4150 int dmar_release_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4151 {
4152 struct acpi_dmar_atsr *atsr;
4153 struct dmar_atsr_unit *atsru;
4154
4155 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4156 atsru = dmar_find_atsr(atsr);
4157 if (atsru) {
4158 list_del_rcu(&atsru->list);
4159 synchronize_rcu();
4160 intel_iommu_free_atsr(atsru);
4161 }
4162
4163 return 0;
4164 }
4165
4166 int dmar_check_one_atsr(struct acpi_dmar_header *hdr, void *arg)
4167 {
4168 int i;
4169 struct device *dev;
4170 struct acpi_dmar_atsr *atsr;
4171 struct dmar_atsr_unit *atsru;
4172
4173 atsr = container_of(hdr, struct acpi_dmar_atsr, header);
4174 atsru = dmar_find_atsr(atsr);
4175 if (!atsru)
4176 return 0;
4177
4178 if (!atsru->include_all && atsru->devices && atsru->devices_cnt)
4179 for_each_active_dev_scope(atsru->devices, atsru->devices_cnt,
4180 i, dev)
4181 return -EBUSY;
4182
4183 return 0;
4184 }
4185
4186 static int intel_iommu_add(struct dmar_drhd_unit *dmaru)
4187 {
4188 int sp, ret = 0;
4189 struct intel_iommu *iommu = dmaru->iommu;
4190
4191 if (g_iommus[iommu->seq_id])
4192 return 0;
4193
4194 if (hw_pass_through && !ecap_pass_through(iommu->ecap)) {
4195 pr_warn("%s: Doesn't support hardware pass through.\n",
4196 iommu->name);
4197 return -ENXIO;
4198 }
4199 if (!ecap_sc_support(iommu->ecap) &&
4200 domain_update_iommu_snooping(iommu)) {
4201 pr_warn("%s: Doesn't support snooping.\n",
4202 iommu->name);
4203 return -ENXIO;
4204 }
4205 sp = domain_update_iommu_superpage(iommu) - 1;
4206 if (sp >= 0 && !(cap_super_page_val(iommu->cap) & (1 << sp))) {
4207 pr_warn("%s: Doesn't support large page.\n",
4208 iommu->name);
4209 return -ENXIO;
4210 }
4211
4212 /*
4213 * Disable translation if already enabled prior to OS handover.
4214 */
4215 if (iommu->gcmd & DMA_GCMD_TE)
4216 iommu_disable_translation(iommu);
4217
4218 g_iommus[iommu->seq_id] = iommu;
4219 ret = iommu_init_domains(iommu);
4220 if (ret == 0)
4221 ret = iommu_alloc_root_entry(iommu);
4222 if (ret)
4223 goto out;
4224
4225 #ifdef CONFIG_INTEL_IOMMU_SVM
4226 if (pasid_enabled(iommu))
4227 intel_svm_alloc_pasid_tables(iommu);
4228 #endif
4229
4230 if (dmaru->ignored) {
4231 /*
4232 * we always have to disable PMRs or DMA may fail on this device
4233 */
4234 if (force_on)
4235 iommu_disable_protect_mem_regions(iommu);
4236 return 0;
4237 }
4238
4239 intel_iommu_init_qi(iommu);
4240 iommu_flush_write_buffer(iommu);
4241
4242 #ifdef CONFIG_INTEL_IOMMU_SVM
4243 if (pasid_enabled(iommu) && ecap_prs(iommu->ecap)) {
4244 ret = intel_svm_enable_prq(iommu);
4245 if (ret)
4246 goto disable_iommu;
4247 }
4248 #endif
4249 ret = dmar_set_interrupt(iommu);
4250 if (ret)
4251 goto disable_iommu;
4252
4253 iommu_set_root_entry(iommu);
4254 iommu->flush.flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL);
4255 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
4256 iommu_enable_translation(iommu);
4257
4258 iommu_disable_protect_mem_regions(iommu);
4259 return 0;
4260
4261 disable_iommu:
4262 disable_dmar_iommu(iommu);
4263 out:
4264 free_dmar_iommu(iommu);
4265 return ret;
4266 }
4267
4268 int dmar_iommu_hotplug(struct dmar_drhd_unit *dmaru, bool insert)
4269 {
4270 int ret = 0;
4271 struct intel_iommu *iommu = dmaru->iommu;
4272
4273 if (!intel_iommu_enabled)
4274 return 0;
4275 if (iommu == NULL)
4276 return -EINVAL;
4277
4278 if (insert) {
4279 ret = intel_iommu_add(dmaru);
4280 } else {
4281 disable_dmar_iommu(iommu);
4282 free_dmar_iommu(iommu);
4283 }
4284
4285 return ret;
4286 }
4287
4288 static void intel_iommu_free_dmars(void)
4289 {
4290 struct dmar_rmrr_unit *rmrru, *rmrr_n;
4291 struct dmar_atsr_unit *atsru, *atsr_n;
4292
4293 list_for_each_entry_safe(rmrru, rmrr_n, &dmar_rmrr_units, list) {
4294 list_del(&rmrru->list);
4295 dmar_free_dev_scope(&rmrru->devices, &rmrru->devices_cnt);
4296 kfree(rmrru);
4297 }
4298
4299 list_for_each_entry_safe(atsru, atsr_n, &dmar_atsr_units, list) {
4300 list_del(&atsru->list);
4301 intel_iommu_free_atsr(atsru);
4302 }
4303 }
4304
4305 int dmar_find_matched_atsr_unit(struct pci_dev *dev)
4306 {
4307 int i, ret = 1;
4308 struct pci_bus *bus;
4309 struct pci_dev *bridge = NULL;
4310 struct device *tmp;
4311 struct acpi_dmar_atsr *atsr;
4312 struct dmar_atsr_unit *atsru;
4313
4314 dev = pci_physfn(dev);
4315 for (bus = dev->bus; bus; bus = bus->parent) {
4316 bridge = bus->self;
4317 /* If it's an integrated device, allow ATS */
4318 if (!bridge)
4319 return 1;
4320 /* Connected via non-PCIe: no ATS */
4321 if (!pci_is_pcie(bridge) ||
4322 pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE)
4323 return 0;
4324 /* If we found the root port, look it up in the ATSR */
4325 if (pci_pcie_type(bridge) == PCI_EXP_TYPE_ROOT_PORT)
4326 break;
4327 }
4328
4329 rcu_read_lock();
4330 list_for_each_entry_rcu(atsru, &dmar_atsr_units, list) {
4331 atsr = container_of(atsru->hdr, struct acpi_dmar_atsr, header);
4332 if (atsr->segment != pci_domain_nr(dev->bus))
4333 continue;
4334
4335 for_each_dev_scope(atsru->devices, atsru->devices_cnt, i, tmp)
4336 if (tmp == &bridge->dev)
4337 goto out;
4338
4339 if (atsru->include_all)
4340 goto out;
4341 }
4342 ret = 0;
4343 out:
4344 rcu_read_unlock();
4345
4346 return ret;
4347 }
4348
4349 int dmar_iommu_notify_scope_dev(struct dmar_pci_notify_info *info)
4350 {
4351 int ret = 0;
4352 struct dmar_rmrr_unit *rmrru;
4353 struct dmar_atsr_unit *atsru;
4354 struct acpi_dmar_atsr *atsr;
4355 struct acpi_dmar_reserved_memory *rmrr;
4356
4357 if (!intel_iommu_enabled && system_state != SYSTEM_BOOTING)
4358 return 0;
4359
4360 list_for_each_entry(rmrru, &dmar_rmrr_units, list) {
4361 rmrr = container_of(rmrru->hdr,
4362 struct acpi_dmar_reserved_memory, header);
4363 if (info->event == BUS_NOTIFY_ADD_DEVICE) {
4364 ret = dmar_insert_dev_scope(info, (void *)(rmrr + 1),
4365 ((void *)rmrr) + rmrr->header.length,
4366 rmrr->segment, rmrru->devices,
4367 rmrru->devices_cnt);
4368 if(ret < 0)
4369 return ret;
4370 } else if (info->event == BUS_NOTIFY_DEL_DEVICE) {
4371 dmar_remove_dev_scope(info, rmrr->segment,
4372 rmrru->devices, rmrru->devices_cnt);
4373 }
4374 }
4375
4376 list_for_each_entry(atsru, &dmar_atsr_units, list) {
4377 if (atsru->include_all)
4378 continue;
4379
4380 atsr = container_of(atsru->hdr, struct acpi_dmar_atsr, header);
4381 if (info->event == BUS_NOTIFY_ADD_DEVICE) {
4382 ret = dmar_insert_dev_scope(info, (void *)(atsr + 1),
4383 (void *)atsr + atsr->header.length,
4384 atsr->segment, atsru->devices,
4385 atsru->devices_cnt);
4386 if (ret > 0)
4387 break;
4388 else if(ret < 0)
4389 return ret;
4390 } else if (info->event == BUS_NOTIFY_DEL_DEVICE) {
4391 if (dmar_remove_dev_scope(info, atsr->segment,
4392 atsru->devices, atsru->devices_cnt))
4393 break;
4394 }
4395 }
4396
4397 return 0;
4398 }
4399
4400 /*
4401 * Here we only respond to action of unbound device from driver.
4402 *
4403 * Added device is not attached to its DMAR domain here yet. That will happen
4404 * when mapping the device to iova.
4405 */
4406 static int device_notifier(struct notifier_block *nb,
4407 unsigned long action, void *data)
4408 {
4409 struct device *dev = data;
4410 struct dmar_domain *domain;
4411
4412 if (iommu_dummy(dev))
4413 return 0;
4414
4415 if (action != BUS_NOTIFY_REMOVED_DEVICE)
4416 return 0;
4417
4418 domain = find_domain(dev);
4419 if (!domain)
4420 return 0;
4421
4422 dmar_remove_one_dev_info(domain, dev);
4423 if (!domain_type_is_vm_or_si(domain) && list_empty(&domain->devices))
4424 domain_exit(domain);
4425
4426 return 0;
4427 }
4428
4429 static struct notifier_block device_nb = {
4430 .notifier_call = device_notifier,
4431 };
4432
4433 static int intel_iommu_memory_notifier(struct notifier_block *nb,
4434 unsigned long val, void *v)
4435 {
4436 struct memory_notify *mhp = v;
4437 unsigned long long start, end;
4438 unsigned long start_vpfn, last_vpfn;
4439
4440 switch (val) {
4441 case MEM_GOING_ONLINE:
4442 start = mhp->start_pfn << PAGE_SHIFT;
4443 end = ((mhp->start_pfn + mhp->nr_pages) << PAGE_SHIFT) - 1;
4444 if (iommu_domain_identity_map(si_domain, start, end)) {
4445 pr_warn("Failed to build identity map for [%llx-%llx]\n",
4446 start, end);
4447 return NOTIFY_BAD;
4448 }
4449 break;
4450
4451 case MEM_OFFLINE:
4452 case MEM_CANCEL_ONLINE:
4453 start_vpfn = mm_to_dma_pfn(mhp->start_pfn);
4454 last_vpfn = mm_to_dma_pfn(mhp->start_pfn + mhp->nr_pages - 1);
4455 while (start_vpfn <= last_vpfn) {
4456 struct iova *iova;
4457 struct dmar_drhd_unit *drhd;
4458 struct intel_iommu *iommu;
4459 struct page *freelist;
4460
4461 iova = find_iova(&si_domain->iovad, start_vpfn);
4462 if (iova == NULL) {
4463 pr_debug("Failed get IOVA for PFN %lx\n",
4464 start_vpfn);
4465 break;
4466 }
4467
4468 iova = split_and_remove_iova(&si_domain->iovad, iova,
4469 start_vpfn, last_vpfn);
4470 if (iova == NULL) {
4471 pr_warn("Failed to split IOVA PFN [%lx-%lx]\n",
4472 start_vpfn, last_vpfn);
4473 return NOTIFY_BAD;
4474 }
4475
4476 freelist = domain_unmap(si_domain, iova->pfn_lo,
4477 iova->pfn_hi);
4478
4479 rcu_read_lock();
4480 for_each_active_iommu(iommu, drhd)
4481 iommu_flush_iotlb_psi(iommu, si_domain,
4482 iova->pfn_lo, iova_size(iova),
4483 !freelist, 0);
4484 rcu_read_unlock();
4485 dma_free_pagelist(freelist);
4486
4487 start_vpfn = iova->pfn_hi + 1;
4488 free_iova_mem(iova);
4489 }
4490 break;
4491 }
4492
4493 return NOTIFY_OK;
4494 }
4495
4496 static struct notifier_block intel_iommu_memory_nb = {
4497 .notifier_call = intel_iommu_memory_notifier,
4498 .priority = 0
4499 };
4500
4501
4502 static ssize_t intel_iommu_show_version(struct device *dev,
4503 struct device_attribute *attr,
4504 char *buf)
4505 {
4506 struct intel_iommu *iommu = dev_get_drvdata(dev);
4507 u32 ver = readl(iommu->reg + DMAR_VER_REG);
4508 return sprintf(buf, "%d:%d\n",
4509 DMAR_VER_MAJOR(ver), DMAR_VER_MINOR(ver));
4510 }
4511 static DEVICE_ATTR(version, S_IRUGO, intel_iommu_show_version, NULL);
4512
4513 static ssize_t intel_iommu_show_address(struct device *dev,
4514 struct device_attribute *attr,
4515 char *buf)
4516 {
4517 struct intel_iommu *iommu = dev_get_drvdata(dev);
4518 return sprintf(buf, "%llx\n", iommu->reg_phys);
4519 }
4520 static DEVICE_ATTR(address, S_IRUGO, intel_iommu_show_address, NULL);
4521
4522 static ssize_t intel_iommu_show_cap(struct device *dev,
4523 struct device_attribute *attr,
4524 char *buf)
4525 {
4526 struct intel_iommu *iommu = dev_get_drvdata(dev);
4527 return sprintf(buf, "%llx\n", iommu->cap);
4528 }
4529 static DEVICE_ATTR(cap, S_IRUGO, intel_iommu_show_cap, NULL);
4530
4531 static ssize_t intel_iommu_show_ecap(struct device *dev,
4532 struct device_attribute *attr,
4533 char *buf)
4534 {
4535 struct intel_iommu *iommu = dev_get_drvdata(dev);
4536 return sprintf(buf, "%llx\n", iommu->ecap);
4537 }
4538 static DEVICE_ATTR(ecap, S_IRUGO, intel_iommu_show_ecap, NULL);
4539
4540 static ssize_t intel_iommu_show_ndoms(struct device *dev,
4541 struct device_attribute *attr,
4542 char *buf)
4543 {
4544 struct intel_iommu *iommu = dev_get_drvdata(dev);
4545 return sprintf(buf, "%ld\n", cap_ndoms(iommu->cap));
4546 }
4547 static DEVICE_ATTR(domains_supported, S_IRUGO, intel_iommu_show_ndoms, NULL);
4548
4549 static ssize_t intel_iommu_show_ndoms_used(struct device *dev,
4550 struct device_attribute *attr,
4551 char *buf)
4552 {
4553 struct intel_iommu *iommu = dev_get_drvdata(dev);
4554 return sprintf(buf, "%d\n", bitmap_weight(iommu->domain_ids,
4555 cap_ndoms(iommu->cap)));
4556 }
4557 static DEVICE_ATTR(domains_used, S_IRUGO, intel_iommu_show_ndoms_used, NULL);
4558
4559 static struct attribute *intel_iommu_attrs[] = {
4560 &dev_attr_version.attr,
4561 &dev_attr_address.attr,
4562 &dev_attr_cap.attr,
4563 &dev_attr_ecap.attr,
4564 &dev_attr_domains_supported.attr,
4565 &dev_attr_domains_used.attr,
4566 NULL,
4567 };
4568
4569 static struct attribute_group intel_iommu_group = {
4570 .name = "intel-iommu",
4571 .attrs = intel_iommu_attrs,
4572 };
4573
4574 const struct attribute_group *intel_iommu_groups[] = {
4575 &intel_iommu_group,
4576 NULL,
4577 };
4578
4579 int __init intel_iommu_init(void)
4580 {
4581 int ret = -ENODEV;
4582 struct dmar_drhd_unit *drhd;
4583 struct intel_iommu *iommu;
4584
4585 /* VT-d is required for a TXT/tboot launch, so enforce that */
4586 force_on = tboot_force_iommu();
4587
4588 if (iommu_init_mempool()) {
4589 if (force_on)
4590 panic("tboot: Failed to initialize iommu memory\n");
4591 return -ENOMEM;
4592 }
4593
4594 down_write(&dmar_global_lock);
4595 if (dmar_table_init()) {
4596 if (force_on)
4597 panic("tboot: Failed to initialize DMAR table\n");
4598 goto out_free_dmar;
4599 }
4600
4601 if (dmar_dev_scope_init() < 0) {
4602 if (force_on)
4603 panic("tboot: Failed to initialize DMAR device scope\n");
4604 goto out_free_dmar;
4605 }
4606
4607 if (no_iommu || dmar_disabled)
4608 goto out_free_dmar;
4609
4610 if (list_empty(&dmar_rmrr_units))
4611 pr_info("No RMRR found\n");
4612
4613 if (list_empty(&dmar_atsr_units))
4614 pr_info("No ATSR found\n");
4615
4616 if (dmar_init_reserved_ranges()) {
4617 if (force_on)
4618 panic("tboot: Failed to reserve iommu ranges\n");
4619 goto out_free_reserved_range;
4620 }
4621
4622 init_no_remapping_devices();
4623
4624 ret = init_dmars();
4625 if (ret) {
4626 if (force_on)
4627 panic("tboot: Failed to initialize DMARs\n");
4628 pr_err("Initialization failed\n");
4629 goto out_free_reserved_range;
4630 }
4631 up_write(&dmar_global_lock);
4632 pr_info("Intel(R) Virtualization Technology for Directed I/O\n");
4633
4634 init_timer(&unmap_timer);
4635 #ifdef CONFIG_SWIOTLB
4636 swiotlb = 0;
4637 #endif
4638 dma_ops = &intel_dma_ops;
4639
4640 init_iommu_pm_ops();
4641
4642 for_each_active_iommu(iommu, drhd)
4643 iommu->iommu_dev = iommu_device_create(NULL, iommu,
4644 intel_iommu_groups,
4645 "%s", iommu->name);
4646
4647 bus_set_iommu(&pci_bus_type, &intel_iommu_ops);
4648 bus_register_notifier(&pci_bus_type, &device_nb);
4649 if (si_domain && !hw_pass_through)
4650 register_memory_notifier(&intel_iommu_memory_nb);
4651
4652 intel_iommu_enabled = 1;
4653
4654 return 0;
4655
4656 out_free_reserved_range:
4657 put_iova_domain(&reserved_iova_list);
4658 out_free_dmar:
4659 intel_iommu_free_dmars();
4660 up_write(&dmar_global_lock);
4661 iommu_exit_mempool();
4662 return ret;
4663 }
4664
4665 static int domain_context_clear_one_cb(struct pci_dev *pdev, u16 alias, void *opaque)
4666 {
4667 struct intel_iommu *iommu = opaque;
4668
4669 domain_context_clear_one(iommu, PCI_BUS_NUM(alias), alias & 0xff);
4670 return 0;
4671 }
4672
4673 /*
4674 * NB - intel-iommu lacks any sort of reference counting for the users of
4675 * dependent devices. If multiple endpoints have intersecting dependent
4676 * devices, unbinding the driver from any one of them will possibly leave
4677 * the others unable to operate.
4678 */
4679 static void domain_context_clear(struct intel_iommu *iommu, struct device *dev)
4680 {
4681 if (!iommu || !dev || !dev_is_pci(dev))
4682 return;
4683
4684 pci_for_each_dma_alias(to_pci_dev(dev), &domain_context_clear_one_cb, iommu);
4685 }
4686
4687 static void __dmar_remove_one_dev_info(struct device_domain_info *info)
4688 {
4689 struct intel_iommu *iommu;
4690 unsigned long flags;
4691
4692 assert_spin_locked(&device_domain_lock);
4693
4694 if (WARN_ON(!info))
4695 return;
4696
4697 iommu = info->iommu;
4698
4699 if (info->dev) {
4700 iommu_disable_dev_iotlb(info);
4701 domain_context_clear(iommu, info->dev);
4702 }
4703
4704 unlink_domain_info(info);
4705
4706 spin_lock_irqsave(&iommu->lock, flags);
4707 domain_detach_iommu(info->domain, iommu);
4708 spin_unlock_irqrestore(&iommu->lock, flags);
4709
4710 free_devinfo_mem(info);
4711 }
4712
4713 static void dmar_remove_one_dev_info(struct dmar_domain *domain,
4714 struct device *dev)
4715 {
4716 struct device_domain_info *info;
4717 unsigned long flags;
4718
4719 spin_lock_irqsave(&device_domain_lock, flags);
4720 info = dev->archdata.iommu;
4721 __dmar_remove_one_dev_info(info);
4722 spin_unlock_irqrestore(&device_domain_lock, flags);
4723 }
4724
4725 static int md_domain_init(struct dmar_domain *domain, int guest_width)
4726 {
4727 int adjust_width;
4728
4729 init_iova_domain(&domain->iovad, VTD_PAGE_SIZE, IOVA_START_PFN,
4730 DMA_32BIT_PFN);
4731 domain_reserve_special_ranges(domain);
4732
4733 /* calculate AGAW */
4734 domain->gaw = guest_width;
4735 adjust_width = guestwidth_to_adjustwidth(guest_width);
4736 domain->agaw = width_to_agaw(adjust_width);
4737
4738 domain->iommu_coherency = 0;
4739 domain->iommu_snooping = 0;
4740 domain->iommu_superpage = 0;
4741 domain->max_addr = 0;
4742
4743 /* always allocate the top pgd */
4744 domain->pgd = (struct dma_pte *)alloc_pgtable_page(domain->nid);
4745 if (!domain->pgd)
4746 return -ENOMEM;
4747 domain_flush_cache(domain, domain->pgd, PAGE_SIZE);
4748 return 0;
4749 }
4750
4751 static struct iommu_domain *intel_iommu_domain_alloc(unsigned type)
4752 {
4753 struct dmar_domain *dmar_domain;
4754 struct iommu_domain *domain;
4755
4756 if (type != IOMMU_DOMAIN_UNMANAGED)
4757 return NULL;
4758
4759 dmar_domain = alloc_domain(DOMAIN_FLAG_VIRTUAL_MACHINE);
4760 if (!dmar_domain) {
4761 pr_err("Can't allocate dmar_domain\n");
4762 return NULL;
4763 }
4764 if (md_domain_init(dmar_domain, DEFAULT_DOMAIN_ADDRESS_WIDTH)) {
4765 pr_err("Domain initialization failed\n");
4766 domain_exit(dmar_domain);
4767 return NULL;
4768 }
4769 domain_update_iommu_cap(dmar_domain);
4770
4771 domain = &dmar_domain->domain;
4772 domain->geometry.aperture_start = 0;
4773 domain->geometry.aperture_end = __DOMAIN_MAX_ADDR(dmar_domain->gaw);
4774 domain->geometry.force_aperture = true;
4775
4776 return domain;
4777 }
4778
4779 static void intel_iommu_domain_free(struct iommu_domain *domain)
4780 {
4781 domain_exit(to_dmar_domain(domain));
4782 }
4783
4784 static int intel_iommu_attach_device(struct iommu_domain *domain,
4785 struct device *dev)
4786 {
4787 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4788 struct intel_iommu *iommu;
4789 int addr_width;
4790 u8 bus, devfn;
4791
4792 if (device_is_rmrr_locked(dev)) {
4793 dev_warn(dev, "Device is ineligible for IOMMU domain attach due to platform RMRR requirement. Contact your platform vendor.\n");
4794 return -EPERM;
4795 }
4796
4797 /* normally dev is not mapped */
4798 if (unlikely(domain_context_mapped(dev))) {
4799 struct dmar_domain *old_domain;
4800
4801 old_domain = find_domain(dev);
4802 if (old_domain) {
4803 rcu_read_lock();
4804 dmar_remove_one_dev_info(old_domain, dev);
4805 rcu_read_unlock();
4806
4807 if (!domain_type_is_vm_or_si(old_domain) &&
4808 list_empty(&old_domain->devices))
4809 domain_exit(old_domain);
4810 }
4811 }
4812
4813 iommu = device_to_iommu(dev, &bus, &devfn);
4814 if (!iommu)
4815 return -ENODEV;
4816
4817 /* check if this iommu agaw is sufficient for max mapped address */
4818 addr_width = agaw_to_width(iommu->agaw);
4819 if (addr_width > cap_mgaw(iommu->cap))
4820 addr_width = cap_mgaw(iommu->cap);
4821
4822 if (dmar_domain->max_addr > (1LL << addr_width)) {
4823 pr_err("%s: iommu width (%d) is not "
4824 "sufficient for the mapped address (%llx)\n",
4825 __func__, addr_width, dmar_domain->max_addr);
4826 return -EFAULT;
4827 }
4828 dmar_domain->gaw = addr_width;
4829
4830 /*
4831 * Knock out extra levels of page tables if necessary
4832 */
4833 while (iommu->agaw < dmar_domain->agaw) {
4834 struct dma_pte *pte;
4835
4836 pte = dmar_domain->pgd;
4837 if (dma_pte_present(pte)) {
4838 dmar_domain->pgd = (struct dma_pte *)
4839 phys_to_virt(dma_pte_addr(pte));
4840 free_pgtable_page(pte);
4841 }
4842 dmar_domain->agaw--;
4843 }
4844
4845 return domain_add_dev_info(dmar_domain, dev);
4846 }
4847
4848 static void intel_iommu_detach_device(struct iommu_domain *domain,
4849 struct device *dev)
4850 {
4851 dmar_remove_one_dev_info(to_dmar_domain(domain), dev);
4852 }
4853
4854 static int intel_iommu_map(struct iommu_domain *domain,
4855 unsigned long iova, phys_addr_t hpa,
4856 size_t size, int iommu_prot)
4857 {
4858 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4859 u64 max_addr;
4860 int prot = 0;
4861 int ret;
4862
4863 if (iommu_prot & IOMMU_READ)
4864 prot |= DMA_PTE_READ;
4865 if (iommu_prot & IOMMU_WRITE)
4866 prot |= DMA_PTE_WRITE;
4867 if ((iommu_prot & IOMMU_CACHE) && dmar_domain->iommu_snooping)
4868 prot |= DMA_PTE_SNP;
4869
4870 max_addr = iova + size;
4871 if (dmar_domain->max_addr < max_addr) {
4872 u64 end;
4873
4874 /* check if minimum agaw is sufficient for mapped address */
4875 end = __DOMAIN_MAX_ADDR(dmar_domain->gaw) + 1;
4876 if (end < max_addr) {
4877 pr_err("%s: iommu width (%d) is not "
4878 "sufficient for the mapped address (%llx)\n",
4879 __func__, dmar_domain->gaw, max_addr);
4880 return -EFAULT;
4881 }
4882 dmar_domain->max_addr = max_addr;
4883 }
4884 /* Round up size to next multiple of PAGE_SIZE, if it and
4885 the low bits of hpa would take us onto the next page */
4886 size = aligned_nrpages(hpa, size);
4887 ret = domain_pfn_mapping(dmar_domain, iova >> VTD_PAGE_SHIFT,
4888 hpa >> VTD_PAGE_SHIFT, size, prot);
4889 return ret;
4890 }
4891
4892 static size_t intel_iommu_unmap(struct iommu_domain *domain,
4893 unsigned long iova, size_t size)
4894 {
4895 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4896 struct page *freelist = NULL;
4897 struct intel_iommu *iommu;
4898 unsigned long start_pfn, last_pfn;
4899 unsigned int npages;
4900 int iommu_id, level = 0;
4901
4902 /* Cope with horrid API which requires us to unmap more than the
4903 size argument if it happens to be a large-page mapping. */
4904 BUG_ON(!pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level));
4905
4906 if (size < VTD_PAGE_SIZE << level_to_offset_bits(level))
4907 size = VTD_PAGE_SIZE << level_to_offset_bits(level);
4908
4909 start_pfn = iova >> VTD_PAGE_SHIFT;
4910 last_pfn = (iova + size - 1) >> VTD_PAGE_SHIFT;
4911
4912 freelist = domain_unmap(dmar_domain, start_pfn, last_pfn);
4913
4914 npages = last_pfn - start_pfn + 1;
4915
4916 for_each_domain_iommu(iommu_id, dmar_domain) {
4917 iommu = g_iommus[iommu_id];
4918
4919 iommu_flush_iotlb_psi(g_iommus[iommu_id], dmar_domain,
4920 start_pfn, npages, !freelist, 0);
4921 }
4922
4923 dma_free_pagelist(freelist);
4924
4925 if (dmar_domain->max_addr == iova + size)
4926 dmar_domain->max_addr = iova;
4927
4928 return size;
4929 }
4930
4931 static phys_addr_t intel_iommu_iova_to_phys(struct iommu_domain *domain,
4932 dma_addr_t iova)
4933 {
4934 struct dmar_domain *dmar_domain = to_dmar_domain(domain);
4935 struct dma_pte *pte;
4936 int level = 0;
4937 u64 phys = 0;
4938
4939 pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT, &level);
4940 if (pte)
4941 phys = dma_pte_addr(pte);
4942
4943 return phys;
4944 }
4945
4946 static bool intel_iommu_capable(enum iommu_cap cap)
4947 {
4948 if (cap == IOMMU_CAP_CACHE_COHERENCY)
4949 return domain_update_iommu_snooping(NULL) == 1;
4950 if (cap == IOMMU_CAP_INTR_REMAP)
4951 return irq_remapping_enabled == 1;
4952
4953 return false;
4954 }
4955
4956 static int intel_iommu_add_device(struct device *dev)
4957 {
4958 struct intel_iommu *iommu;
4959 struct iommu_group *group;
4960 u8 bus, devfn;
4961
4962 iommu = device_to_iommu(dev, &bus, &devfn);
4963 if (!iommu)
4964 return -ENODEV;
4965
4966 iommu_device_link(iommu->iommu_dev, dev);
4967
4968 group = iommu_group_get_for_dev(dev);
4969
4970 if (IS_ERR(group))
4971 return PTR_ERR(group);
4972
4973 iommu_group_put(group);
4974 return 0;
4975 }
4976
4977 static void intel_iommu_remove_device(struct device *dev)
4978 {
4979 struct intel_iommu *iommu;
4980 u8 bus, devfn;
4981
4982 iommu = device_to_iommu(dev, &bus, &devfn);
4983 if (!iommu)
4984 return;
4985
4986 iommu_group_remove_device(dev);
4987
4988 iommu_device_unlink(iommu->iommu_dev, dev);
4989 }
4990
4991 #ifdef CONFIG_INTEL_IOMMU_SVM
4992 int intel_iommu_enable_pasid(struct intel_iommu *iommu, struct intel_svm_dev *sdev)
4993 {
4994 struct device_domain_info *info;
4995 struct context_entry *context;
4996 struct dmar_domain *domain;
4997 unsigned long flags;
4998 u64 ctx_lo;
4999 int ret;
5000
5001 domain = get_valid_domain_for_dev(sdev->dev);
5002 if (!domain)
5003 return -EINVAL;
5004
5005 spin_lock_irqsave(&device_domain_lock, flags);
5006 spin_lock(&iommu->lock);
5007
5008 ret = -EINVAL;
5009 info = sdev->dev->archdata.iommu;
5010 if (!info || !info->pasid_supported)
5011 goto out;
5012
5013 context = iommu_context_addr(iommu, info->bus, info->devfn, 0);
5014 if (WARN_ON(!context))
5015 goto out;
5016
5017 ctx_lo = context[0].lo;
5018
5019 sdev->did = domain->iommu_did[iommu->seq_id];
5020 sdev->sid = PCI_DEVID(info->bus, info->devfn);
5021
5022 if (!(ctx_lo & CONTEXT_PASIDE)) {
5023 context[1].hi = (u64)virt_to_phys(iommu->pasid_state_table);
5024 context[1].lo = (u64)virt_to_phys(iommu->pasid_table) | ecap_pss(iommu->ecap);
5025 wmb();
5026 /* CONTEXT_TT_MULTI_LEVEL and CONTEXT_TT_DEV_IOTLB are both
5027 * extended to permit requests-with-PASID if the PASIDE bit
5028 * is set. which makes sense. For CONTEXT_TT_PASS_THROUGH,
5029 * however, the PASIDE bit is ignored and requests-with-PASID
5030 * are unconditionally blocked. Which makes less sense.
5031 * So convert from CONTEXT_TT_PASS_THROUGH to one of the new
5032 * "guest mode" translation types depending on whether ATS
5033 * is available or not. Annoyingly, we can't use the new
5034 * modes *unless* PASIDE is set. */
5035 if ((ctx_lo & CONTEXT_TT_MASK) == (CONTEXT_TT_PASS_THROUGH << 2)) {
5036 ctx_lo &= ~CONTEXT_TT_MASK;
5037 if (info->ats_supported)
5038 ctx_lo |= CONTEXT_TT_PT_PASID_DEV_IOTLB << 2;
5039 else
5040 ctx_lo |= CONTEXT_TT_PT_PASID << 2;
5041 }
5042 ctx_lo |= CONTEXT_PASIDE;
5043 if (iommu->pasid_state_table)
5044 ctx_lo |= CONTEXT_DINVE;
5045 if (info->pri_supported)
5046 ctx_lo |= CONTEXT_PRS;
5047 context[0].lo = ctx_lo;
5048 wmb();
5049 iommu->flush.flush_context(iommu, sdev->did, sdev->sid,
5050 DMA_CCMD_MASK_NOBIT,
5051 DMA_CCMD_DEVICE_INVL);
5052 }
5053
5054 /* Enable PASID support in the device, if it wasn't already */
5055 if (!info->pasid_enabled)
5056 iommu_enable_dev_iotlb(info);
5057
5058 if (info->ats_enabled) {
5059 sdev->dev_iotlb = 1;
5060 sdev->qdep = info->ats_qdep;
5061 if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
5062 sdev->qdep = 0;
5063 }
5064 ret = 0;
5065
5066 out:
5067 spin_unlock(&iommu->lock);
5068 spin_unlock_irqrestore(&device_domain_lock, flags);
5069
5070 return ret;
5071 }
5072
5073 struct intel_iommu *intel_svm_device_to_iommu(struct device *dev)
5074 {
5075 struct intel_iommu *iommu;
5076 u8 bus, devfn;
5077
5078 if (iommu_dummy(dev)) {
5079 dev_warn(dev,
5080 "No IOMMU translation for device; cannot enable SVM\n");
5081 return NULL;
5082 }
5083
5084 iommu = device_to_iommu(dev, &bus, &devfn);
5085 if ((!iommu)) {
5086 dev_err(dev, "No IOMMU for device; cannot enable SVM\n");
5087 return NULL;
5088 }
5089
5090 if (!iommu->pasid_table) {
5091 dev_err(dev, "PASID not enabled on IOMMU; cannot enable SVM\n");
5092 return NULL;
5093 }
5094
5095 return iommu;
5096 }
5097 #endif /* CONFIG_INTEL_IOMMU_SVM */
5098
5099 static const struct iommu_ops intel_iommu_ops = {
5100 .capable = intel_iommu_capable,
5101 .domain_alloc = intel_iommu_domain_alloc,
5102 .domain_free = intel_iommu_domain_free,
5103 .attach_dev = intel_iommu_attach_device,
5104 .detach_dev = intel_iommu_detach_device,
5105 .map = intel_iommu_map,
5106 .unmap = intel_iommu_unmap,
5107 .map_sg = default_iommu_map_sg,
5108 .iova_to_phys = intel_iommu_iova_to_phys,
5109 .add_device = intel_iommu_add_device,
5110 .remove_device = intel_iommu_remove_device,
5111 .device_group = pci_device_group,
5112 .pgsize_bitmap = INTEL_IOMMU_PGSIZES,
5113 };
5114
5115 static void quirk_iommu_g4x_gfx(struct pci_dev *dev)
5116 {
5117 /* G4x/GM45 integrated gfx dmar support is totally busted. */
5118 pr_info("Disabling IOMMU for graphics on this chipset\n");
5119 dmar_map_gfx = 0;
5120 }
5121
5122 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_g4x_gfx);
5123 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e00, quirk_iommu_g4x_gfx);
5124 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e10, quirk_iommu_g4x_gfx);
5125 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e20, quirk_iommu_g4x_gfx);
5126 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e30, quirk_iommu_g4x_gfx);
5127 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e40, quirk_iommu_g4x_gfx);
5128 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e90, quirk_iommu_g4x_gfx);
5129
5130 static void quirk_iommu_rwbf(struct pci_dev *dev)
5131 {
5132 /*
5133 * Mobile 4 Series Chipset neglects to set RWBF capability,
5134 * but needs it. Same seems to hold for the desktop versions.
5135 */
5136 pr_info("Forcing write-buffer flush capability\n");
5137 rwbf_quirk = 1;
5138 }
5139
5140 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2a40, quirk_iommu_rwbf);
5141 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e00, quirk_iommu_rwbf);
5142 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e10, quirk_iommu_rwbf);
5143 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e20, quirk_iommu_rwbf);
5144 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e30, quirk_iommu_rwbf);
5145 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e40, quirk_iommu_rwbf);
5146 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x2e90, quirk_iommu_rwbf);
5147
5148 #define GGC 0x52
5149 #define GGC_MEMORY_SIZE_MASK (0xf << 8)
5150 #define GGC_MEMORY_SIZE_NONE (0x0 << 8)
5151 #define GGC_MEMORY_SIZE_1M (0x1 << 8)
5152 #define GGC_MEMORY_SIZE_2M (0x3 << 8)
5153 #define GGC_MEMORY_VT_ENABLED (0x8 << 8)
5154 #define GGC_MEMORY_SIZE_2M_VT (0x9 << 8)
5155 #define GGC_MEMORY_SIZE_3M_VT (0xa << 8)
5156 #define GGC_MEMORY_SIZE_4M_VT (0xb << 8)
5157
5158 static void quirk_calpella_no_shadow_gtt(struct pci_dev *dev)
5159 {
5160 unsigned short ggc;
5161
5162 if (pci_read_config_word(dev, GGC, &ggc))
5163 return;
5164
5165 if (!(ggc & GGC_MEMORY_VT_ENABLED)) {
5166 pr_info("BIOS has allocated no shadow GTT; disabling IOMMU for graphics\n");
5167 dmar_map_gfx = 0;
5168 } else if (dmar_map_gfx) {
5169 /* we have to ensure the gfx device is idle before we flush */
5170 pr_info("Disabling batched IOTLB flush on Ironlake\n");
5171 intel_iommu_strict = 1;
5172 }
5173 }
5174 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0040, quirk_calpella_no_shadow_gtt);
5175 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0044, quirk_calpella_no_shadow_gtt);
5176 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x0062, quirk_calpella_no_shadow_gtt);
5177 DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_INTEL, 0x006a, quirk_calpella_no_shadow_gtt);
5178
5179 /* On Tylersburg chipsets, some BIOSes have been known to enable the
5180 ISOCH DMAR unit for the Azalia sound device, but not give it any
5181 TLB entries, which causes it to deadlock. Check for that. We do
5182 this in a function called from init_dmars(), instead of in a PCI
5183 quirk, because we don't want to print the obnoxious "BIOS broken"
5184 message if VT-d is actually disabled.
5185 */
5186 static void __init check_tylersburg_isoch(void)
5187 {
5188 struct pci_dev *pdev;
5189 uint32_t vtisochctrl;
5190
5191 /* If there's no Azalia in the system anyway, forget it. */
5192 pdev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x3a3e, NULL);
5193 if (!pdev)
5194 return;
5195 pci_dev_put(pdev);
5196
5197 /* System Management Registers. Might be hidden, in which case
5198 we can't do the sanity check. But that's OK, because the
5199 known-broken BIOSes _don't_ actually hide it, so far. */
5200 pdev = pci_get_device(PCI_VENDOR_ID_INTEL, 0x342e, NULL);
5201 if (!pdev)
5202 return;
5203
5204 if (pci_read_config_dword(pdev, 0x188, &vtisochctrl)) {
5205 pci_dev_put(pdev);
5206 return;
5207 }
5208
5209 pci_dev_put(pdev);
5210
5211 /* If Azalia DMA is routed to the non-isoch DMAR unit, fine. */
5212 if (vtisochctrl & 1)
5213 return;
5214
5215 /* Drop all bits other than the number of TLB entries */
5216 vtisochctrl &= 0x1c;
5217
5218 /* If we have the recommended number of TLB entries (16), fine. */
5219 if (vtisochctrl == 0x10)
5220 return;
5221
5222 /* Zero TLB entries? You get to ride the short bus to school. */
5223 if (!vtisochctrl) {
5224 WARN(1, "Your BIOS is broken; DMA routed to ISOCH DMAR unit but no TLB space.\n"
5225 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
5226 dmi_get_system_info(DMI_BIOS_VENDOR),
5227 dmi_get_system_info(DMI_BIOS_VERSION),
5228 dmi_get_system_info(DMI_PRODUCT_VERSION));
5229 iommu_identity_mapping |= IDENTMAP_AZALIA;
5230 return;
5231 }
5232
5233 pr_warn("Recommended TLB entries for ISOCH unit is 16; your BIOS set %d\n",
5234 vtisochctrl);
5235 }
Linux with added FPC-III support