?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge git://git.kernel.org/pub/scm/linux/kernel/git/kyle/parisc-2.6
[linux-2.6/openmoko-kernel/knife-kernel.git] / drivers / pci / intel-iommu.c
blob66c0fd21894b155e9db690b65a86bfddffca29c4
1 /*
2 * Copyright (c) 2006, 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 * You should have received a copy of the GNU General Public License along with
14 * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
15 * Place - Suite 330, Boston, MA 02111-1307 USA.
16 *
17 * Copyright (C) 2006-2008 Intel Corporation
18 * Author: Ashok Raj <ashok.raj@intel.com>
19 * Author: Shaohua Li <shaohua.li@intel.com>
20 * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
21 */
22
23 #include <linux/init.h>
24 #include <linux/bitmap.h>
25 #include <linux/debugfs.h>
26 #include <linux/slab.h>
27 #include <linux/irq.h>
28 #include <linux/interrupt.h>
29 #include <linux/sysdev.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/timer.h>
36 #include "iova.h"
37 #include "intel-iommu.h"
38 #include <asm/proto.h> /* force_iommu in this header in x86-64*/
39 #include <asm/cacheflush.h>
40 #include <asm/gart.h>
41 #include "pci.h"
42
43 #define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
44 #define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
45
46 #define IOAPIC_RANGE_START (0xfee00000)
47 #define IOAPIC_RANGE_END (0xfeefffff)
48 #define IOVA_START_ADDR (0x1000)
49
50 #define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
51
52 #define DMAR_OPERATION_TIMEOUT ((cycles_t) tsc_khz*10*1000) /* 10sec */
53
54 #define DOMAIN_MAX_ADDR(gaw) ((((u64)1) << gaw) - 1)
55
56
57 static void flush_unmaps_timeout(unsigned long data);
58
59 DEFINE_TIMER(unmap_timer, flush_unmaps_timeout, 0, 0);
60
61 static struct intel_iommu *g_iommus;
62
63 #define HIGH_WATER_MARK 250
64 struct deferred_flush_tables {
65 int next;
66 struct iova *iova[HIGH_WATER_MARK];
67 struct dmar_domain *domain[HIGH_WATER_MARK];
68 };
69
70 static struct deferred_flush_tables *deferred_flush;
71
72 /* bitmap for indexing intel_iommus */
73 static int g_num_of_iommus;
74
75 static DEFINE_SPINLOCK(async_umap_flush_lock);
76 static LIST_HEAD(unmaps_to_do);
77
78 static int timer_on;
79 static long list_size;
80
81 static void domain_remove_dev_info(struct dmar_domain *domain);
82
83 static int dmar_disabled;
84 static int __initdata dmar_map_gfx = 1;
85 static int dmar_forcedac;
86 static int intel_iommu_strict;
87
88 #define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
89 static DEFINE_SPINLOCK(device_domain_lock);
90 static LIST_HEAD(device_domain_list);
91
92 static int __init intel_iommu_setup(char *str)
93 {
94 if (!str)
95 return -EINVAL;
96 while (*str) {
97 if (!strncmp(str, "off", 3)) {
98 dmar_disabled = 1;
99 printk(KERN_INFO"Intel-IOMMU: disabled\n");
100 } else if (!strncmp(str, "igfx_off", 8)) {
101 dmar_map_gfx = 0;
102 printk(KERN_INFO
103 "Intel-IOMMU: disable GFX device mapping\n");
104 } else if (!strncmp(str, "forcedac", 8)) {
105 printk(KERN_INFO
106 "Intel-IOMMU: Forcing DAC for PCI devices\n");
107 dmar_forcedac = 1;
108 } else if (!strncmp(str, "strict", 6)) {
109 printk(KERN_INFO
110 "Intel-IOMMU: disable batched IOTLB flush\n");
111 intel_iommu_strict = 1;
112 }
113
114 str += strcspn(str, ",");
115 while (*str == ',')
116 str++;
117 }
118 return 0;
119 }
120 __setup("intel_iommu=", intel_iommu_setup);
121
122 static struct kmem_cache *iommu_domain_cache;
123 static struct kmem_cache *iommu_devinfo_cache;
124 static struct kmem_cache *iommu_iova_cache;
125
126 static inline void *iommu_kmem_cache_alloc(struct kmem_cache *cachep)
127 {
128 unsigned int flags;
129 void *vaddr;
130
131 /* trying to avoid low memory issues */
132 flags = current->flags & PF_MEMALLOC;
133 current->flags |= PF_MEMALLOC;
134 vaddr = kmem_cache_alloc(cachep, GFP_ATOMIC);
135 current->flags &= (~PF_MEMALLOC | flags);
136 return vaddr;
137 }
138
139
140 static inline void *alloc_pgtable_page(void)
141 {
142 unsigned int flags;
143 void *vaddr;
144
145 /* trying to avoid low memory issues */
146 flags = current->flags & PF_MEMALLOC;
147 current->flags |= PF_MEMALLOC;
148 vaddr = (void *)get_zeroed_page(GFP_ATOMIC);
149 current->flags &= (~PF_MEMALLOC | flags);
150 return vaddr;
151 }
152
153 static inline void free_pgtable_page(void *vaddr)
154 {
155 free_page((unsigned long)vaddr);
156 }
157
158 static inline void *alloc_domain_mem(void)
159 {
160 return iommu_kmem_cache_alloc(iommu_domain_cache);
161 }
162
163 static inline void free_domain_mem(void *vaddr)
164 {
165 kmem_cache_free(iommu_domain_cache, vaddr);
166 }
167
168 static inline void * alloc_devinfo_mem(void)
169 {
170 return iommu_kmem_cache_alloc(iommu_devinfo_cache);
171 }
172
173 static inline void free_devinfo_mem(void *vaddr)
174 {
175 kmem_cache_free(iommu_devinfo_cache, vaddr);
176 }
177
178 struct iova *alloc_iova_mem(void)
179 {
180 return iommu_kmem_cache_alloc(iommu_iova_cache);
181 }
182
183 void free_iova_mem(struct iova *iova)
184 {
185 kmem_cache_free(iommu_iova_cache, iova);
186 }
187
188 static inline void __iommu_flush_cache(
189 struct intel_iommu *iommu, void *addr, int size)
190 {
191 if (!ecap_coherent(iommu->ecap))
192 clflush_cache_range(addr, size);
193 }
194
195 /* Gets context entry for a given bus and devfn */
196 static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
197 u8 bus, u8 devfn)
198 {
199 struct root_entry *root;
200 struct context_entry *context;
201 unsigned long phy_addr;
202 unsigned long flags;
203
204 spin_lock_irqsave(&iommu->lock, flags);
205 root = &iommu->root_entry[bus];
206 context = get_context_addr_from_root(root);
207 if (!context) {
208 context = (struct context_entry *)alloc_pgtable_page();
209 if (!context) {
210 spin_unlock_irqrestore(&iommu->lock, flags);
211 return NULL;
212 }
213 __iommu_flush_cache(iommu, (void *)context, PAGE_SIZE_4K);
214 phy_addr = virt_to_phys((void *)context);
215 set_root_value(root, phy_addr);
216 set_root_present(root);
217 __iommu_flush_cache(iommu, root, sizeof(*root));
218 }
219 spin_unlock_irqrestore(&iommu->lock, flags);
220 return &context[devfn];
221 }
222
223 static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
224 {
225 struct root_entry *root;
226 struct context_entry *context;
227 int ret;
228 unsigned long flags;
229
230 spin_lock_irqsave(&iommu->lock, flags);
231 root = &iommu->root_entry[bus];
232 context = get_context_addr_from_root(root);
233 if (!context) {
234 ret = 0;
235 goto out;
236 }
237 ret = context_present(context[devfn]);
238 out:
239 spin_unlock_irqrestore(&iommu->lock, flags);
240 return ret;
241 }
242
243 static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
244 {
245 struct root_entry *root;
246 struct context_entry *context;
247 unsigned long flags;
248
249 spin_lock_irqsave(&iommu->lock, flags);
250 root = &iommu->root_entry[bus];
251 context = get_context_addr_from_root(root);
252 if (context) {
253 context_clear_entry(context[devfn]);
254 __iommu_flush_cache(iommu, &context[devfn], \
255 sizeof(*context));
256 }
257 spin_unlock_irqrestore(&iommu->lock, flags);
258 }
259
260 static void free_context_table(struct intel_iommu *iommu)
261 {
262 struct root_entry *root;
263 int i;
264 unsigned long flags;
265 struct context_entry *context;
266
267 spin_lock_irqsave(&iommu->lock, flags);
268 if (!iommu->root_entry) {
269 goto out;
270 }
271 for (i = 0; i < ROOT_ENTRY_NR; i++) {
272 root = &iommu->root_entry[i];
273 context = get_context_addr_from_root(root);
274 if (context)
275 free_pgtable_page(context);
276 }
277 free_pgtable_page(iommu->root_entry);
278 iommu->root_entry = NULL;
279 out:
280 spin_unlock_irqrestore(&iommu->lock, flags);
281 }
282
283 /* page table handling */
284 #define LEVEL_STRIDE (9)
285 #define LEVEL_MASK (((u64)1 << LEVEL_STRIDE) - 1)
286
287 static inline int agaw_to_level(int agaw)
288 {
289 return agaw + 2;
290 }
291
292 static inline int agaw_to_width(int agaw)
293 {
294 return 30 + agaw * LEVEL_STRIDE;
295
296 }
297
298 static inline int width_to_agaw(int width)
299 {
300 return (width - 30) / LEVEL_STRIDE;
301 }
302
303 static inline unsigned int level_to_offset_bits(int level)
304 {
305 return (12 + (level - 1) * LEVEL_STRIDE);
306 }
307
308 static inline int address_level_offset(u64 addr, int level)
309 {
310 return ((addr >> level_to_offset_bits(level)) & LEVEL_MASK);
311 }
312
313 static inline u64 level_mask(int level)
314 {
315 return ((u64)-1 << level_to_offset_bits(level));
316 }
317
318 static inline u64 level_size(int level)
319 {
320 return ((u64)1 << level_to_offset_bits(level));
321 }
322
323 static inline u64 align_to_level(u64 addr, int level)
324 {
325 return ((addr + level_size(level) - 1) & level_mask(level));
326 }
327
328 static struct dma_pte * addr_to_dma_pte(struct dmar_domain *domain, u64 addr)
329 {
330 int addr_width = agaw_to_width(domain->agaw);
331 struct dma_pte *parent, *pte = NULL;
332 int level = agaw_to_level(domain->agaw);
333 int offset;
334 unsigned long flags;
335
336 BUG_ON(!domain->pgd);
337
338 addr &= (((u64)1) << addr_width) - 1;
339 parent = domain->pgd;
340
341 spin_lock_irqsave(&domain->mapping_lock, flags);
342 while (level > 0) {
343 void *tmp_page;
344
345 offset = address_level_offset(addr, level);
346 pte = &parent[offset];
347 if (level == 1)
348 break;
349
350 if (!dma_pte_present(*pte)) {
351 tmp_page = alloc_pgtable_page();
352
353 if (!tmp_page) {
354 spin_unlock_irqrestore(&domain->mapping_lock,
355 flags);
356 return NULL;
357 }
358 __iommu_flush_cache(domain->iommu, tmp_page,
359 PAGE_SIZE_4K);
360 dma_set_pte_addr(*pte, virt_to_phys(tmp_page));
361 /*
362 * high level table always sets r/w, last level page
363 * table control read/write
364 */
365 dma_set_pte_readable(*pte);
366 dma_set_pte_writable(*pte);
367 __iommu_flush_cache(domain->iommu, pte, sizeof(*pte));
368 }
369 parent = phys_to_virt(dma_pte_addr(*pte));
370 level--;
371 }
372
373 spin_unlock_irqrestore(&domain->mapping_lock, flags);
374 return pte;
375 }
376
377 /* return address's pte at specific level */
378 static struct dma_pte *dma_addr_level_pte(struct dmar_domain *domain, u64 addr,
379 int level)
380 {
381 struct dma_pte *parent, *pte = NULL;
382 int total = agaw_to_level(domain->agaw);
383 int offset;
384
385 parent = domain->pgd;
386 while (level <= total) {
387 offset = address_level_offset(addr, total);
388 pte = &parent[offset];
389 if (level == total)
390 return pte;
391
392 if (!dma_pte_present(*pte))
393 break;
394 parent = phys_to_virt(dma_pte_addr(*pte));
395 total--;
396 }
397 return NULL;
398 }
399
400 /* clear one page's page table */
401 static void dma_pte_clear_one(struct dmar_domain *domain, u64 addr)
402 {
403 struct dma_pte *pte = NULL;
404
405 /* get last level pte */
406 pte = dma_addr_level_pte(domain, addr, 1);
407
408 if (pte) {
409 dma_clear_pte(*pte);
410 __iommu_flush_cache(domain->iommu, pte, sizeof(*pte));
411 }
412 }
413
414 /* clear last level pte, a tlb flush should be followed */
415 static void dma_pte_clear_range(struct dmar_domain *domain, u64 start, u64 end)
416 {
417 int addr_width = agaw_to_width(domain->agaw);
418
419 start &= (((u64)1) << addr_width) - 1;
420 end &= (((u64)1) << addr_width) - 1;
421 /* in case it's partial page */
422 start = PAGE_ALIGN_4K(start);
423 end &= PAGE_MASK_4K;
424
425 /* we don't need lock here, nobody else touches the iova range */
426 while (start < end) {
427 dma_pte_clear_one(domain, start);
428 start += PAGE_SIZE_4K;
429 }
430 }
431
432 /* free page table pages. last level pte should already be cleared */
433 static void dma_pte_free_pagetable(struct dmar_domain *domain,
434 u64 start, u64 end)
435 {
436 int addr_width = agaw_to_width(domain->agaw);
437 struct dma_pte *pte;
438 int total = agaw_to_level(domain->agaw);
439 int level;
440 u64 tmp;
441
442 start &= (((u64)1) << addr_width) - 1;
443 end &= (((u64)1) << addr_width) - 1;
444
445 /* we don't need lock here, nobody else touches the iova range */
446 level = 2;
447 while (level <= total) {
448 tmp = align_to_level(start, level);
449 if (tmp >= end || (tmp + level_size(level) > end))
450 return;
451
452 while (tmp < end) {
453 pte = dma_addr_level_pte(domain, tmp, level);
454 if (pte) {
455 free_pgtable_page(
456 phys_to_virt(dma_pte_addr(*pte)));
457 dma_clear_pte(*pte);
458 __iommu_flush_cache(domain->iommu,
459 pte, sizeof(*pte));
460 }
461 tmp += level_size(level);
462 }
463 level++;
464 }
465 /* free pgd */
466 if (start == 0 && end >= ((((u64)1) << addr_width) - 1)) {
467 free_pgtable_page(domain->pgd);
468 domain->pgd = NULL;
469 }
470 }
471
472 /* iommu handling */
473 static int iommu_alloc_root_entry(struct intel_iommu *iommu)
474 {
475 struct root_entry *root;
476 unsigned long flags;
477
478 root = (struct root_entry *)alloc_pgtable_page();
479 if (!root)
480 return -ENOMEM;
481
482 __iommu_flush_cache(iommu, root, PAGE_SIZE_4K);
483
484 spin_lock_irqsave(&iommu->lock, flags);
485 iommu->root_entry = root;
486 spin_unlock_irqrestore(&iommu->lock, flags);
487
488 return 0;
489 }
490
491 #define IOMMU_WAIT_OP(iommu, offset, op, cond, sts) \
492 {\
493 cycles_t start_time = get_cycles();\
494 while (1) {\
495 sts = op (iommu->reg + offset);\
496 if (cond)\
497 break;\
498 if (DMAR_OPERATION_TIMEOUT < (get_cycles() - start_time))\
499 panic("DMAR hardware is malfunctioning\n");\
500 cpu_relax();\
501 }\
502 }
503
504 static void iommu_set_root_entry(struct intel_iommu *iommu)
505 {
506 void *addr;
507 u32 cmd, sts;
508 unsigned long flag;
509
510 addr = iommu->root_entry;
511
512 spin_lock_irqsave(&iommu->register_lock, flag);
513 dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
514
515 cmd = iommu->gcmd | DMA_GCMD_SRTP;
516 writel(cmd, iommu->reg + DMAR_GCMD_REG);
517
518 /* Make sure hardware complete it */
519 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
520 readl, (sts & DMA_GSTS_RTPS), sts);
521
522 spin_unlock_irqrestore(&iommu->register_lock, flag);
523 }
524
525 static void iommu_flush_write_buffer(struct intel_iommu *iommu)
526 {
527 u32 val;
528 unsigned long flag;
529
530 if (!cap_rwbf(iommu->cap))
531 return;
532 val = iommu->gcmd | DMA_GCMD_WBF;
533
534 spin_lock_irqsave(&iommu->register_lock, flag);
535 writel(val, iommu->reg + DMAR_GCMD_REG);
536
537 /* Make sure hardware complete it */
538 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
539 readl, (!(val & DMA_GSTS_WBFS)), val);
540
541 spin_unlock_irqrestore(&iommu->register_lock, flag);
542 }
543
544 /* return value determine if we need a write buffer flush */
545 static int __iommu_flush_context(struct intel_iommu *iommu,
546 u16 did, u16 source_id, u8 function_mask, u64 type,
547 int non_present_entry_flush)
548 {
549 u64 val = 0;
550 unsigned long flag;
551
552 /*
553 * In the non-present entry flush case, if hardware doesn't cache
554 * non-present entry we do nothing and if hardware cache non-present
555 * entry, we flush entries of domain 0 (the domain id is used to cache
556 * any non-present entries)
557 */
558 if (non_present_entry_flush) {
559 if (!cap_caching_mode(iommu->cap))
560 return 1;
561 else
562 did = 0;
563 }
564
565 switch (type) {
566 case DMA_CCMD_GLOBAL_INVL:
567 val = DMA_CCMD_GLOBAL_INVL;
568 break;
569 case DMA_CCMD_DOMAIN_INVL:
570 val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
571 break;
572 case DMA_CCMD_DEVICE_INVL:
573 val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
574 | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
575 break;
576 default:
577 BUG();
578 }
579 val |= DMA_CCMD_ICC;
580
581 spin_lock_irqsave(&iommu->register_lock, flag);
582 dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
583
584 /* Make sure hardware complete it */
585 IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
586 dmar_readq, (!(val & DMA_CCMD_ICC)), val);
587
588 spin_unlock_irqrestore(&iommu->register_lock, flag);
589
590 /* flush context entry will implictly flush write buffer */
591 return 0;
592 }
593
594 static int inline iommu_flush_context_global(struct intel_iommu *iommu,
595 int non_present_entry_flush)
596 {
597 return __iommu_flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL,
598 non_present_entry_flush);
599 }
600
601 static int inline iommu_flush_context_domain(struct intel_iommu *iommu, u16 did,
602 int non_present_entry_flush)
603 {
604 return __iommu_flush_context(iommu, did, 0, 0, DMA_CCMD_DOMAIN_INVL,
605 non_present_entry_flush);
606 }
607
608 static int inline iommu_flush_context_device(struct intel_iommu *iommu,
609 u16 did, u16 source_id, u8 function_mask, int non_present_entry_flush)
610 {
611 return __iommu_flush_context(iommu, did, source_id, function_mask,
612 DMA_CCMD_DEVICE_INVL, non_present_entry_flush);
613 }
614
615 /* return value determine if we need a write buffer flush */
616 static int __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
617 u64 addr, unsigned int size_order, u64 type,
618 int non_present_entry_flush)
619 {
620 int tlb_offset = ecap_iotlb_offset(iommu->ecap);
621 u64 val = 0, val_iva = 0;
622 unsigned long flag;
623
624 /*
625 * In the non-present entry flush case, if hardware doesn't cache
626 * non-present entry we do nothing and if hardware cache non-present
627 * entry, we flush entries of domain 0 (the domain id is used to cache
628 * any non-present entries)
629 */
630 if (non_present_entry_flush) {
631 if (!cap_caching_mode(iommu->cap))
632 return 1;
633 else
634 did = 0;
635 }
636
637 switch (type) {
638 case DMA_TLB_GLOBAL_FLUSH:
639 /* global flush doesn't need set IVA_REG */
640 val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
641 break;
642 case DMA_TLB_DSI_FLUSH:
643 val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
644 break;
645 case DMA_TLB_PSI_FLUSH:
646 val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
647 /* Note: always flush non-leaf currently */
648 val_iva = size_order | addr;
649 break;
650 default:
651 BUG();
652 }
653 /* Note: set drain read/write */
654 #if 0
655 /*
656 * This is probably to be super secure.. Looks like we can
657 * ignore it without any impact.
658 */
659 if (cap_read_drain(iommu->cap))
660 val |= DMA_TLB_READ_DRAIN;
661 #endif
662 if (cap_write_drain(iommu->cap))
663 val |= DMA_TLB_WRITE_DRAIN;
664
665 spin_lock_irqsave(&iommu->register_lock, flag);
666 /* Note: Only uses first TLB reg currently */
667 if (val_iva)
668 dmar_writeq(iommu->reg + tlb_offset, val_iva);
669 dmar_writeq(iommu->reg + tlb_offset + 8, val);
670
671 /* Make sure hardware complete it */
672 IOMMU_WAIT_OP(iommu, tlb_offset + 8,
673 dmar_readq, (!(val & DMA_TLB_IVT)), val);
674
675 spin_unlock_irqrestore(&iommu->register_lock, flag);
676
677 /* check IOTLB invalidation granularity */
678 if (DMA_TLB_IAIG(val) == 0)
679 printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
680 if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
681 pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
682 DMA_TLB_IIRG(type), DMA_TLB_IAIG(val));
683 /* flush context entry will implictly flush write buffer */
684 return 0;
685 }
686
687 static int inline iommu_flush_iotlb_global(struct intel_iommu *iommu,
688 int non_present_entry_flush)
689 {
690 return __iommu_flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH,
691 non_present_entry_flush);
692 }
693
694 static int inline iommu_flush_iotlb_dsi(struct intel_iommu *iommu, u16 did,
695 int non_present_entry_flush)
696 {
697 return __iommu_flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH,
698 non_present_entry_flush);
699 }
700
701 static int iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
702 u64 addr, unsigned int pages, int non_present_entry_flush)
703 {
704 unsigned int mask;
705
706 BUG_ON(addr & (~PAGE_MASK_4K));
707 BUG_ON(pages == 0);
708
709 /* Fallback to domain selective flush if no PSI support */
710 if (!cap_pgsel_inv(iommu->cap))
711 return iommu_flush_iotlb_dsi(iommu, did,
712 non_present_entry_flush);
713
714 /*
715 * PSI requires page size to be 2 ^ x, and the base address is naturally
716 * aligned to the size
717 */
718 mask = ilog2(__roundup_pow_of_two(pages));
719 /* Fallback to domain selective flush if size is too big */
720 if (mask > cap_max_amask_val(iommu->cap))
721 return iommu_flush_iotlb_dsi(iommu, did,
722 non_present_entry_flush);
723
724 return __iommu_flush_iotlb(iommu, did, addr, mask,
725 DMA_TLB_PSI_FLUSH, non_present_entry_flush);
726 }
727
728 static void iommu_disable_protect_mem_regions(struct intel_iommu *iommu)
729 {
730 u32 pmen;
731 unsigned long flags;
732
733 spin_lock_irqsave(&iommu->register_lock, flags);
734 pmen = readl(iommu->reg + DMAR_PMEN_REG);
735 pmen &= ~DMA_PMEN_EPM;
736 writel(pmen, iommu->reg + DMAR_PMEN_REG);
737
738 /* wait for the protected region status bit to clear */
739 IOMMU_WAIT_OP(iommu, DMAR_PMEN_REG,
740 readl, !(pmen & DMA_PMEN_PRS), pmen);
741
742 spin_unlock_irqrestore(&iommu->register_lock, flags);
743 }
744
745 static int iommu_enable_translation(struct intel_iommu *iommu)
746 {
747 u32 sts;
748 unsigned long flags;
749
750 spin_lock_irqsave(&iommu->register_lock, flags);
751 writel(iommu->gcmd|DMA_GCMD_TE, iommu->reg + DMAR_GCMD_REG);
752
753 /* Make sure hardware complete it */
754 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
755 readl, (sts & DMA_GSTS_TES), sts);
756
757 iommu->gcmd |= DMA_GCMD_TE;
758 spin_unlock_irqrestore(&iommu->register_lock, flags);
759 return 0;
760 }
761
762 static int iommu_disable_translation(struct intel_iommu *iommu)
763 {
764 u32 sts;
765 unsigned long flag;
766
767 spin_lock_irqsave(&iommu->register_lock, flag);
768 iommu->gcmd &= ~DMA_GCMD_TE;
769 writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
770
771 /* Make sure hardware complete it */
772 IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
773 readl, (!(sts & DMA_GSTS_TES)), sts);
774
775 spin_unlock_irqrestore(&iommu->register_lock, flag);
776 return 0;
777 }
778
779 /* iommu interrupt handling. Most stuff are MSI-like. */
780
781 static const char *fault_reason_strings[] =
782 {
783 "Software",
784 "Present bit in root entry is clear",
785 "Present bit in context entry is clear",
786 "Invalid context entry",
787 "Access beyond MGAW",
788 "PTE Write access is not set",
789 "PTE Read access is not set",
790 "Next page table ptr is invalid",
791 "Root table address invalid",
792 "Context table ptr is invalid",
793 "non-zero reserved fields in RTP",
794 "non-zero reserved fields in CTP",
795 "non-zero reserved fields in PTE",
796 };
797 #define MAX_FAULT_REASON_IDX (ARRAY_SIZE(fault_reason_strings) - 1)
798
799 const char *dmar_get_fault_reason(u8 fault_reason)
800 {
801 if (fault_reason > MAX_FAULT_REASON_IDX)
802 return "Unknown";
803 else
804 return fault_reason_strings[fault_reason];
805 }
806
807 void dmar_msi_unmask(unsigned int irq)
808 {
809 struct intel_iommu *iommu = get_irq_data(irq);
810 unsigned long flag;
811
812 /* unmask it */
813 spin_lock_irqsave(&iommu->register_lock, flag);
814 writel(0, iommu->reg + DMAR_FECTL_REG);
815 /* Read a reg to force flush the post write */
816 readl(iommu->reg + DMAR_FECTL_REG);
817 spin_unlock_irqrestore(&iommu->register_lock, flag);
818 }
819
820 void dmar_msi_mask(unsigned int irq)
821 {
822 unsigned long flag;
823 struct intel_iommu *iommu = get_irq_data(irq);
824
825 /* mask it */
826 spin_lock_irqsave(&iommu->register_lock, flag);
827 writel(DMA_FECTL_IM, iommu->reg + DMAR_FECTL_REG);
828 /* Read a reg to force flush the post write */
829 readl(iommu->reg + DMAR_FECTL_REG);
830 spin_unlock_irqrestore(&iommu->register_lock, flag);
831 }
832
833 void dmar_msi_write(int irq, struct msi_msg *msg)
834 {
835 struct intel_iommu *iommu = get_irq_data(irq);
836 unsigned long flag;
837
838 spin_lock_irqsave(&iommu->register_lock, flag);
839 writel(msg->data, iommu->reg + DMAR_FEDATA_REG);
840 writel(msg->address_lo, iommu->reg + DMAR_FEADDR_REG);
841 writel(msg->address_hi, iommu->reg + DMAR_FEUADDR_REG);
842 spin_unlock_irqrestore(&iommu->register_lock, flag);
843 }
844
845 void dmar_msi_read(int irq, struct msi_msg *msg)
846 {
847 struct intel_iommu *iommu = get_irq_data(irq);
848 unsigned long flag;
849
850 spin_lock_irqsave(&iommu->register_lock, flag);
851 msg->data = readl(iommu->reg + DMAR_FEDATA_REG);
852 msg->address_lo = readl(iommu->reg + DMAR_FEADDR_REG);
853 msg->address_hi = readl(iommu->reg + DMAR_FEUADDR_REG);
854 spin_unlock_irqrestore(&iommu->register_lock, flag);
855 }
856
857 static int iommu_page_fault_do_one(struct intel_iommu *iommu, int type,
858 u8 fault_reason, u16 source_id, u64 addr)
859 {
860 const char *reason;
861
862 reason = dmar_get_fault_reason(fault_reason);
863
864 printk(KERN_ERR
865 "DMAR:[%s] Request device [%02x:%02x.%d] "
866 "fault addr %llx \n"
867 "DMAR:[fault reason %02d] %s\n",
868 (type ? "DMA Read" : "DMA Write"),
869 (source_id >> 8), PCI_SLOT(source_id & 0xFF),
870 PCI_FUNC(source_id & 0xFF), addr, fault_reason, reason);
871 return 0;
872 }
873
874 #define PRIMARY_FAULT_REG_LEN (16)
875 static irqreturn_t iommu_page_fault(int irq, void *dev_id)
876 {
877 struct intel_iommu *iommu = dev_id;
878 int reg, fault_index;
879 u32 fault_status;
880 unsigned long flag;
881
882 spin_lock_irqsave(&iommu->register_lock, flag);
883 fault_status = readl(iommu->reg + DMAR_FSTS_REG);
884
885 /* TBD: ignore advanced fault log currently */
886 if (!(fault_status & DMA_FSTS_PPF))
887 goto clear_overflow;
888
889 fault_index = dma_fsts_fault_record_index(fault_status);
890 reg = cap_fault_reg_offset(iommu->cap);
891 while (1) {
892 u8 fault_reason;
893 u16 source_id;
894 u64 guest_addr;
895 int type;
896 u32 data;
897
898 /* highest 32 bits */
899 data = readl(iommu->reg + reg +
900 fault_index * PRIMARY_FAULT_REG_LEN + 12);
901 if (!(data & DMA_FRCD_F))
902 break;
903
904 fault_reason = dma_frcd_fault_reason(data);
905 type = dma_frcd_type(data);
906
907 data = readl(iommu->reg + reg +
908 fault_index * PRIMARY_FAULT_REG_LEN + 8);
909 source_id = dma_frcd_source_id(data);
910
911 guest_addr = dmar_readq(iommu->reg + reg +
912 fault_index * PRIMARY_FAULT_REG_LEN);
913 guest_addr = dma_frcd_page_addr(guest_addr);
914 /* clear the fault */
915 writel(DMA_FRCD_F, iommu->reg + reg +
916 fault_index * PRIMARY_FAULT_REG_LEN + 12);
917
918 spin_unlock_irqrestore(&iommu->register_lock, flag);
919
920 iommu_page_fault_do_one(iommu, type, fault_reason,
921 source_id, guest_addr);
922
923 fault_index++;
924 if (fault_index > cap_num_fault_regs(iommu->cap))
925 fault_index = 0;
926 spin_lock_irqsave(&iommu->register_lock, flag);
927 }
928 clear_overflow:
929 /* clear primary fault overflow */
930 fault_status = readl(iommu->reg + DMAR_FSTS_REG);
931 if (fault_status & DMA_FSTS_PFO)
932 writel(DMA_FSTS_PFO, iommu->reg + DMAR_FSTS_REG);
933
934 spin_unlock_irqrestore(&iommu->register_lock, flag);
935 return IRQ_HANDLED;
936 }
937
938 int dmar_set_interrupt(struct intel_iommu *iommu)
939 {
940 int irq, ret;
941
942 irq = create_irq();
943 if (!irq) {
944 printk(KERN_ERR "IOMMU: no free vectors\n");
945 return -EINVAL;
946 }
947
948 set_irq_data(irq, iommu);
949 iommu->irq = irq;
950
951 ret = arch_setup_dmar_msi(irq);
952 if (ret) {
953 set_irq_data(irq, NULL);
954 iommu->irq = 0;
955 destroy_irq(irq);
956 return 0;
957 }
958
959 /* Force fault register is cleared */
960 iommu_page_fault(irq, iommu);
961
962 ret = request_irq(irq, iommu_page_fault, 0, iommu->name, iommu);
963 if (ret)
964 printk(KERN_ERR "IOMMU: can't request irq\n");
965 return ret;
966 }
967
968 static int iommu_init_domains(struct intel_iommu *iommu)
969 {
970 unsigned long ndomains;
971 unsigned long nlongs;
972
973 ndomains = cap_ndoms(iommu->cap);
974 pr_debug("Number of Domains supportd <%ld>\n", ndomains);
975 nlongs = BITS_TO_LONGS(ndomains);
976
977 /* TBD: there might be 64K domains,
978 * consider other allocation for future chip
979 */
980 iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
981 if (!iommu->domain_ids) {
982 printk(KERN_ERR "Allocating domain id array failed\n");
983 return -ENOMEM;
984 }
985 iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
986 GFP_KERNEL);
987 if (!iommu->domains) {
988 printk(KERN_ERR "Allocating domain array failed\n");
989 kfree(iommu->domain_ids);
990 return -ENOMEM;
991 }
992
993 /*
994 * if Caching mode is set, then invalid translations are tagged
995 * with domainid 0. Hence we need to pre-allocate it.
996 */
997 if (cap_caching_mode(iommu->cap))
998 set_bit(0, iommu->domain_ids);
999 return 0;
1000 }
1001 static struct intel_iommu *alloc_iommu(struct intel_iommu *iommu,
1002 struct dmar_drhd_unit *drhd)
1003 {
1004 int ret;
1005 int map_size;
1006 u32 ver;
1007
1008 iommu->reg = ioremap(drhd->reg_base_addr, PAGE_SIZE_4K);
1009 if (!iommu->reg) {
1010 printk(KERN_ERR "IOMMU: can't map the region\n");
1011 goto error;
1012 }
1013 iommu->cap = dmar_readq(iommu->reg + DMAR_CAP_REG);
1014 iommu->ecap = dmar_readq(iommu->reg + DMAR_ECAP_REG);
1015
1016 /* the registers might be more than one page */
1017 map_size = max_t(int, ecap_max_iotlb_offset(iommu->ecap),
1018 cap_max_fault_reg_offset(iommu->cap));
1019 map_size = PAGE_ALIGN_4K(map_size);
1020 if (map_size > PAGE_SIZE_4K) {
1021 iounmap(iommu->reg);
1022 iommu->reg = ioremap(drhd->reg_base_addr, map_size);
1023 if (!iommu->reg) {
1024 printk(KERN_ERR "IOMMU: can't map the region\n");
1025 goto error;
1026 }
1027 }
1028
1029 ver = readl(iommu->reg + DMAR_VER_REG);
1030 pr_debug("IOMMU %llx: ver %d:%d cap %llx ecap %llx\n",
1031 drhd->reg_base_addr, DMAR_VER_MAJOR(ver), DMAR_VER_MINOR(ver),
1032 iommu->cap, iommu->ecap);
1033 ret = iommu_init_domains(iommu);
1034 if (ret)
1035 goto error_unmap;
1036 spin_lock_init(&iommu->lock);
1037 spin_lock_init(&iommu->register_lock);
1038
1039 drhd->iommu = iommu;
1040 return iommu;
1041 error_unmap:
1042 iounmap(iommu->reg);
1043 error:
1044 kfree(iommu);
1045 return NULL;
1046 }
1047
1048 static void domain_exit(struct dmar_domain *domain);
1049 static void free_iommu(struct intel_iommu *iommu)
1050 {
1051 struct dmar_domain *domain;
1052 int i;
1053
1054 if (!iommu)
1055 return;
1056
1057 i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
1058 for (; i < cap_ndoms(iommu->cap); ) {
1059 domain = iommu->domains[i];
1060 clear_bit(i, iommu->domain_ids);
1061 domain_exit(domain);
1062 i = find_next_bit(iommu->domain_ids,
1063 cap_ndoms(iommu->cap), i+1);
1064 }
1065
1066 if (iommu->gcmd & DMA_GCMD_TE)
1067 iommu_disable_translation(iommu);
1068
1069 if (iommu->irq) {
1070 set_irq_data(iommu->irq, NULL);
1071 /* This will mask the irq */
1072 free_irq(iommu->irq, iommu);
1073 destroy_irq(iommu->irq);
1074 }
1075
1076 kfree(iommu->domains);
1077 kfree(iommu->domain_ids);
1078
1079 /* free context mapping */
1080 free_context_table(iommu);
1081
1082 if (iommu->reg)
1083 iounmap(iommu->reg);
1084 kfree(iommu);
1085 }
1086
1087 static struct dmar_domain * iommu_alloc_domain(struct intel_iommu *iommu)
1088 {
1089 unsigned long num;
1090 unsigned long ndomains;
1091 struct dmar_domain *domain;
1092 unsigned long flags;
1093
1094 domain = alloc_domain_mem();
1095 if (!domain)
1096 return NULL;
1097
1098 ndomains = cap_ndoms(iommu->cap);
1099
1100 spin_lock_irqsave(&iommu->lock, flags);
1101 num = find_first_zero_bit(iommu->domain_ids, ndomains);
1102 if (num >= ndomains) {
1103 spin_unlock_irqrestore(&iommu->lock, flags);
1104 free_domain_mem(domain);
1105 printk(KERN_ERR "IOMMU: no free domain ids\n");
1106 return NULL;
1107 }
1108
1109 set_bit(num, iommu->domain_ids);
1110 domain->id = num;
1111 domain->iommu = iommu;
1112 iommu->domains[num] = domain;
1113 spin_unlock_irqrestore(&iommu->lock, flags);
1114
1115 return domain;
1116 }
1117
1118 static void iommu_free_domain(struct dmar_domain *domain)
1119 {
1120 unsigned long flags;
1121
1122 spin_lock_irqsave(&domain->iommu->lock, flags);
1123 clear_bit(domain->id, domain->iommu->domain_ids);
1124 spin_unlock_irqrestore(&domain->iommu->lock, flags);
1125 }
1126
1127 static struct iova_domain reserved_iova_list;
1128 static struct lock_class_key reserved_alloc_key;
1129 static struct lock_class_key reserved_rbtree_key;
1130
1131 static void dmar_init_reserved_ranges(void)
1132 {
1133 struct pci_dev *pdev = NULL;
1134 struct iova *iova;
1135 int i;
1136 u64 addr, size;
1137
1138 init_iova_domain(&reserved_iova_list, DMA_32BIT_PFN);
1139
1140 lockdep_set_class(&reserved_iova_list.iova_alloc_lock,
1141 &reserved_alloc_key);
1142 lockdep_set_class(&reserved_iova_list.iova_rbtree_lock,
1143 &reserved_rbtree_key);
1144
1145 /* IOAPIC ranges shouldn't be accessed by DMA */
1146 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
1147 IOVA_PFN(IOAPIC_RANGE_END));
1148 if (!iova)
1149 printk(KERN_ERR "Reserve IOAPIC range failed\n");
1150
1151 /* Reserve all PCI MMIO to avoid peer-to-peer access */
1152 for_each_pci_dev(pdev) {
1153 struct resource *r;
1154
1155 for (i = 0; i < PCI_NUM_RESOURCES; i++) {
1156 r = &pdev->resource[i];
1157 if (!r->flags || !(r->flags & IORESOURCE_MEM))
1158 continue;
1159 addr = r->start;
1160 addr &= PAGE_MASK_4K;
1161 size = r->end - addr;
1162 size = PAGE_ALIGN_4K(size);
1163 iova = reserve_iova(&reserved_iova_list, IOVA_PFN(addr),
1164 IOVA_PFN(size + addr) - 1);
1165 if (!iova)
1166 printk(KERN_ERR "Reserve iova failed\n");
1167 }
1168 }
1169
1170 }
1171
1172 static void domain_reserve_special_ranges(struct dmar_domain *domain)
1173 {
1174 copy_reserved_iova(&reserved_iova_list, &domain->iovad);
1175 }
1176
1177 static inline int guestwidth_to_adjustwidth(int gaw)
1178 {
1179 int agaw;
1180 int r = (gaw - 12) % 9;
1181
1182 if (r == 0)
1183 agaw = gaw;
1184 else
1185 agaw = gaw + 9 - r;
1186 if (agaw > 64)
1187 agaw = 64;
1188 return agaw;
1189 }
1190
1191 static int domain_init(struct dmar_domain *domain, int guest_width)
1192 {
1193 struct intel_iommu *iommu;
1194 int adjust_width, agaw;
1195 unsigned long sagaw;
1196
1197 init_iova_domain(&domain->iovad, DMA_32BIT_PFN);
1198 spin_lock_init(&domain->mapping_lock);
1199
1200 domain_reserve_special_ranges(domain);
1201
1202 /* calculate AGAW */
1203 iommu = domain->iommu;
1204 if (guest_width > cap_mgaw(iommu->cap))
1205 guest_width = cap_mgaw(iommu->cap);
1206 domain->gaw = guest_width;
1207 adjust_width = guestwidth_to_adjustwidth(guest_width);
1208 agaw = width_to_agaw(adjust_width);
1209 sagaw = cap_sagaw(iommu->cap);
1210 if (!test_bit(agaw, &sagaw)) {
1211 /* hardware doesn't support it, choose a bigger one */
1212 pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
1213 agaw = find_next_bit(&sagaw, 5, agaw);
1214 if (agaw >= 5)
1215 return -ENODEV;
1216 }
1217 domain->agaw = agaw;
1218 INIT_LIST_HEAD(&domain->devices);
1219
1220 /* always allocate the top pgd */
1221 domain->pgd = (struct dma_pte *)alloc_pgtable_page();
1222 if (!domain->pgd)
1223 return -ENOMEM;
1224 __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE_4K);
1225 return 0;
1226 }
1227
1228 static void domain_exit(struct dmar_domain *domain)
1229 {
1230 u64 end;
1231
1232 /* Domain 0 is reserved, so dont process it */
1233 if (!domain)
1234 return;
1235
1236 domain_remove_dev_info(domain);
1237 /* destroy iovas */
1238 put_iova_domain(&domain->iovad);
1239 end = DOMAIN_MAX_ADDR(domain->gaw);
1240 end = end & (~PAGE_MASK_4K);
1241
1242 /* clear ptes */
1243 dma_pte_clear_range(domain, 0, end);
1244
1245 /* free page tables */
1246 dma_pte_free_pagetable(domain, 0, end);
1247
1248 iommu_free_domain(domain);
1249 free_domain_mem(domain);
1250 }
1251
1252 static int domain_context_mapping_one(struct dmar_domain *domain,
1253 u8 bus, u8 devfn)
1254 {
1255 struct context_entry *context;
1256 struct intel_iommu *iommu = domain->iommu;
1257 unsigned long flags;
1258
1259 pr_debug("Set context mapping for %02x:%02x.%d\n",
1260 bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
1261 BUG_ON(!domain->pgd);
1262 context = device_to_context_entry(iommu, bus, devfn);
1263 if (!context)
1264 return -ENOMEM;
1265 spin_lock_irqsave(&iommu->lock, flags);
1266 if (context_present(*context)) {
1267 spin_unlock_irqrestore(&iommu->lock, flags);
1268 return 0;
1269 }
1270
1271 context_set_domain_id(*context, domain->id);
1272 context_set_address_width(*context, domain->agaw);
1273 context_set_address_root(*context, virt_to_phys(domain->pgd));
1274 context_set_translation_type(*context, CONTEXT_TT_MULTI_LEVEL);
1275 context_set_fault_enable(*context);
1276 context_set_present(*context);
1277 __iommu_flush_cache(iommu, context, sizeof(*context));
1278
1279 /* it's a non-present to present mapping */
1280 if (iommu_flush_context_device(iommu, domain->id,
1281 (((u16)bus) << 8) | devfn, DMA_CCMD_MASK_NOBIT, 1))
1282 iommu_flush_write_buffer(iommu);
1283 else
1284 iommu_flush_iotlb_dsi(iommu, 0, 0);
1285 spin_unlock_irqrestore(&iommu->lock, flags);
1286 return 0;
1287 }
1288
1289 static int
1290 domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev)
1291 {
1292 int ret;
1293 struct pci_dev *tmp, *parent;
1294
1295 ret = domain_context_mapping_one(domain, pdev->bus->number,
1296 pdev->devfn);
1297 if (ret)
1298 return ret;
1299
1300 /* dependent device mapping */
1301 tmp = pci_find_upstream_pcie_bridge(pdev);
1302 if (!tmp)
1303 return 0;
1304 /* Secondary interface's bus number and devfn 0 */
1305 parent = pdev->bus->self;
1306 while (parent != tmp) {
1307 ret = domain_context_mapping_one(domain, parent->bus->number,
1308 parent->devfn);
1309 if (ret)
1310 return ret;
1311 parent = parent->bus->self;
1312 }
1313 if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
1314 return domain_context_mapping_one(domain,
1315 tmp->subordinate->number, 0);
1316 else /* this is a legacy PCI bridge */
1317 return domain_context_mapping_one(domain,
1318 tmp->bus->number, tmp->devfn);
1319 }
1320
1321 static int domain_context_mapped(struct dmar_domain *domain,
1322 struct pci_dev *pdev)
1323 {
1324 int ret;
1325 struct pci_dev *tmp, *parent;
1326
1327 ret = device_context_mapped(domain->iommu,
1328 pdev->bus->number, pdev->devfn);
1329 if (!ret)
1330 return ret;
1331 /* dependent device mapping */
1332 tmp = pci_find_upstream_pcie_bridge(pdev);
1333 if (!tmp)
1334 return ret;
1335 /* Secondary interface's bus number and devfn 0 */
1336 parent = pdev->bus->self;
1337 while (parent != tmp) {
1338 ret = device_context_mapped(domain->iommu, parent->bus->number,
1339 parent->devfn);
1340 if (!ret)
1341 return ret;
1342 parent = parent->bus->self;
1343 }
1344 if (tmp->is_pcie)
1345 return device_context_mapped(domain->iommu,
1346 tmp->subordinate->number, 0);
1347 else
1348 return device_context_mapped(domain->iommu,
1349 tmp->bus->number, tmp->devfn);
1350 }
1351
1352 static int
1353 domain_page_mapping(struct dmar_domain *domain, dma_addr_t iova,
1354 u64 hpa, size_t size, int prot)
1355 {
1356 u64 start_pfn, end_pfn;
1357 struct dma_pte *pte;
1358 int index;
1359
1360 if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
1361 return -EINVAL;
1362 iova &= PAGE_MASK_4K;
1363 start_pfn = ((u64)hpa) >> PAGE_SHIFT_4K;
1364 end_pfn = (PAGE_ALIGN_4K(((u64)hpa) + size)) >> PAGE_SHIFT_4K;
1365 index = 0;
1366 while (start_pfn < end_pfn) {
1367 pte = addr_to_dma_pte(domain, iova + PAGE_SIZE_4K * index);
1368 if (!pte)
1369 return -ENOMEM;
1370 /* We don't need lock here, nobody else
1371 * touches the iova range
1372 */
1373 BUG_ON(dma_pte_addr(*pte));
1374 dma_set_pte_addr(*pte, start_pfn << PAGE_SHIFT_4K);
1375 dma_set_pte_prot(*pte, prot);
1376 __iommu_flush_cache(domain->iommu, pte, sizeof(*pte));
1377 start_pfn++;
1378 index++;
1379 }
1380 return 0;
1381 }
1382
1383 static void detach_domain_for_dev(struct dmar_domain *domain, u8 bus, u8 devfn)
1384 {
1385 clear_context_table(domain->iommu, bus, devfn);
1386 iommu_flush_context_global(domain->iommu, 0);
1387 iommu_flush_iotlb_global(domain->iommu, 0);
1388 }
1389
1390 static void domain_remove_dev_info(struct dmar_domain *domain)
1391 {
1392 struct device_domain_info *info;
1393 unsigned long flags;
1394
1395 spin_lock_irqsave(&device_domain_lock, flags);
1396 while (!list_empty(&domain->devices)) {
1397 info = list_entry(domain->devices.next,
1398 struct device_domain_info, link);
1399 list_del(&info->link);
1400 list_del(&info->global);
1401 if (info->dev)
1402 info->dev->dev.archdata.iommu = NULL;
1403 spin_unlock_irqrestore(&device_domain_lock, flags);
1404
1405 detach_domain_for_dev(info->domain, info->bus, info->devfn);
1406 free_devinfo_mem(info);
1407
1408 spin_lock_irqsave(&device_domain_lock, flags);
1409 }
1410 spin_unlock_irqrestore(&device_domain_lock, flags);
1411 }
1412
1413 /*
1414 * find_domain
1415 * Note: we use struct pci_dev->dev.archdata.iommu stores the info
1416 */
1417 struct dmar_domain *
1418 find_domain(struct pci_dev *pdev)
1419 {
1420 struct device_domain_info *info;
1421
1422 /* No lock here, assumes no domain exit in normal case */
1423 info = pdev->dev.archdata.iommu;
1424 if (info)
1425 return info->domain;
1426 return NULL;
1427 }
1428
1429 static int dmar_pci_device_match(struct pci_dev *devices[], int cnt,
1430 struct pci_dev *dev)
1431 {
1432 int index;
1433
1434 while (dev) {
1435 for (index = 0; index < cnt; index++)
1436 if (dev == devices[index])
1437 return 1;
1438
1439 /* Check our parent */
1440 dev = dev->bus->self;
1441 }
1442
1443 return 0;
1444 }
1445
1446 static struct dmar_drhd_unit *
1447 dmar_find_matched_drhd_unit(struct pci_dev *dev)
1448 {
1449 struct dmar_drhd_unit *drhd = NULL;
1450
1451 list_for_each_entry(drhd, &dmar_drhd_units, list) {
1452 if (drhd->include_all || dmar_pci_device_match(drhd->devices,
1453 drhd->devices_cnt, dev))
1454 return drhd;
1455 }
1456
1457 return NULL;
1458 }
1459
1460 /* domain is initialized */
1461 static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
1462 {
1463 struct dmar_domain *domain, *found = NULL;
1464 struct intel_iommu *iommu;
1465 struct dmar_drhd_unit *drhd;
1466 struct device_domain_info *info, *tmp;
1467 struct pci_dev *dev_tmp;
1468 unsigned long flags;
1469 int bus = 0, devfn = 0;
1470
1471 domain = find_domain(pdev);
1472 if (domain)
1473 return domain;
1474
1475 dev_tmp = pci_find_upstream_pcie_bridge(pdev);
1476 if (dev_tmp) {
1477 if (dev_tmp->is_pcie) {
1478 bus = dev_tmp->subordinate->number;
1479 devfn = 0;
1480 } else {
1481 bus = dev_tmp->bus->number;
1482 devfn = dev_tmp->devfn;
1483 }
1484 spin_lock_irqsave(&device_domain_lock, flags);
1485 list_for_each_entry(info, &device_domain_list, global) {
1486 if (info->bus == bus && info->devfn == devfn) {
1487 found = info->domain;
1488 break;
1489 }
1490 }
1491 spin_unlock_irqrestore(&device_domain_lock, flags);
1492 /* pcie-pci bridge already has a domain, uses it */
1493 if (found) {
1494 domain = found;
1495 goto found_domain;
1496 }
1497 }
1498
1499 /* Allocate new domain for the device */
1500 drhd = dmar_find_matched_drhd_unit(pdev);
1501 if (!drhd) {
1502 printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
1503 pci_name(pdev));
1504 return NULL;
1505 }
1506 iommu = drhd->iommu;
1507
1508 domain = iommu_alloc_domain(iommu);
1509 if (!domain)
1510 goto error;
1511
1512 if (domain_init(domain, gaw)) {
1513 domain_exit(domain);
1514 goto error;
1515 }
1516
1517 /* register pcie-to-pci device */
1518 if (dev_tmp) {
1519 info = alloc_devinfo_mem();
1520 if (!info) {
1521 domain_exit(domain);
1522 goto error;
1523 }
1524 info->bus = bus;
1525 info->devfn = devfn;
1526 info->dev = NULL;
1527 info->domain = domain;
1528 /* This domain is shared by devices under p2p bridge */
1529 domain->flags |= DOMAIN_FLAG_MULTIPLE_DEVICES;
1530
1531 /* pcie-to-pci bridge already has a domain, uses it */
1532 found = NULL;
1533 spin_lock_irqsave(&device_domain_lock, flags);
1534 list_for_each_entry(tmp, &device_domain_list, global) {
1535 if (tmp->bus == bus && tmp->devfn == devfn) {
1536 found = tmp->domain;
1537 break;
1538 }
1539 }
1540 if (found) {
1541 free_devinfo_mem(info);
1542 domain_exit(domain);
1543 domain = found;
1544 } else {
1545 list_add(&info->link, &domain->devices);
1546 list_add(&info->global, &device_domain_list);
1547 }
1548 spin_unlock_irqrestore(&device_domain_lock, flags);
1549 }
1550
1551 found_domain:
1552 info = alloc_devinfo_mem();
1553 if (!info)
1554 goto error;
1555 info->bus = pdev->bus->number;
1556 info->devfn = pdev->devfn;
1557 info->dev = pdev;
1558 info->domain = domain;
1559 spin_lock_irqsave(&device_domain_lock, flags);
1560 /* somebody is fast */
1561 found = find_domain(pdev);
1562 if (found != NULL) {
1563 spin_unlock_irqrestore(&device_domain_lock, flags);
1564 if (found != domain) {
1565 domain_exit(domain);
1566 domain = found;
1567 }
1568 free_devinfo_mem(info);
1569 return domain;
1570 }
1571 list_add(&info->link, &domain->devices);
1572 list_add(&info->global, &device_domain_list);
1573 pdev->dev.archdata.iommu = info;
1574 spin_unlock_irqrestore(&device_domain_lock, flags);
1575 return domain;
1576 error:
1577 /* recheck it here, maybe others set it */
1578 return find_domain(pdev);
1579 }
1580
1581 static int iommu_prepare_identity_map(struct pci_dev *pdev, u64 start, u64 end)
1582 {
1583 struct dmar_domain *domain;
1584 unsigned long size;
1585 u64 base;
1586 int ret;
1587
1588 printk(KERN_INFO
1589 "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
1590 pci_name(pdev), start, end);
1591 /* page table init */
1592 domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
1593 if (!domain)
1594 return -ENOMEM;
1595
1596 /* The address might not be aligned */
1597 base = start & PAGE_MASK_4K;
1598 size = end - base;
1599 size = PAGE_ALIGN_4K(size);
1600 if (!reserve_iova(&domain->iovad, IOVA_PFN(base),
1601 IOVA_PFN(base + size) - 1)) {
1602 printk(KERN_ERR "IOMMU: reserve iova failed\n");
1603 ret = -ENOMEM;
1604 goto error;
1605 }
1606
1607 pr_debug("Mapping reserved region %lx@%llx for %s\n",
1608 size, base, pci_name(pdev));
1609 /*
1610 * RMRR range might have overlap with physical memory range,
1611 * clear it first
1612 */
1613 dma_pte_clear_range(domain, base, base + size);
1614
1615 ret = domain_page_mapping(domain, base, base, size,
1616 DMA_PTE_READ|DMA_PTE_WRITE);
1617 if (ret)
1618 goto error;
1619
1620 /* context entry init */
1621 ret = domain_context_mapping(domain, pdev);
1622 if (!ret)
1623 return 0;
1624 error:
1625 domain_exit(domain);
1626 return ret;
1627
1628 }
1629
1630 static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
1631 struct pci_dev *pdev)
1632 {
1633 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
1634 return 0;
1635 return iommu_prepare_identity_map(pdev, rmrr->base_address,
1636 rmrr->end_address + 1);
1637 }
1638
1639 #ifdef CONFIG_DMAR_GFX_WA
1640 extern int arch_get_ram_range(int slot, u64 *addr, u64 *size);
1641 static void __init iommu_prepare_gfx_mapping(void)
1642 {
1643 struct pci_dev *pdev = NULL;
1644 u64 base, size;
1645 int slot;
1646 int ret;
1647
1648 for_each_pci_dev(pdev) {
1649 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO ||
1650 !IS_GFX_DEVICE(pdev))
1651 continue;
1652 printk(KERN_INFO "IOMMU: gfx device %s 1-1 mapping\n",
1653 pci_name(pdev));
1654 slot = arch_get_ram_range(0, &base, &size);
1655 while (slot >= 0) {
1656 ret = iommu_prepare_identity_map(pdev,
1657 base, base + size);
1658 if (ret)
1659 goto error;
1660 slot = arch_get_ram_range(slot, &base, &size);
1661 }
1662 continue;
1663 error:
1664 printk(KERN_ERR "IOMMU: mapping reserved region failed\n");
1665 }
1666 }
1667 #endif
1668
1669 #ifdef CONFIG_DMAR_FLOPPY_WA
1670 static inline void iommu_prepare_isa(void)
1671 {
1672 struct pci_dev *pdev;
1673 int ret;
1674
1675 pdev = pci_get_class(PCI_CLASS_BRIDGE_ISA << 8, NULL);
1676 if (!pdev)
1677 return;
1678
1679 printk(KERN_INFO "IOMMU: Prepare 0-16M unity mapping for LPC\n");
1680 ret = iommu_prepare_identity_map(pdev, 0, 16*1024*1024);
1681
1682 if (ret)
1683 printk("IOMMU: Failed to create 0-64M identity map, "
1684 "floppy might not work\n");
1685
1686 }
1687 #else
1688 static inline void iommu_prepare_isa(void)
1689 {
1690 return;
1691 }
1692 #endif /* !CONFIG_DMAR_FLPY_WA */
1693
1694 int __init init_dmars(void)
1695 {
1696 struct dmar_drhd_unit *drhd;
1697 struct dmar_rmrr_unit *rmrr;
1698 struct pci_dev *pdev;
1699 struct intel_iommu *iommu;
1700 int i, ret, unit = 0;
1701
1702 /*
1703 * for each drhd
1704 * allocate root
1705 * initialize and program root entry to not present
1706 * endfor
1707 */
1708 for_each_drhd_unit(drhd) {
1709 if (drhd->ignored)
1710 continue;
1711 g_num_of_iommus++;
1712 /*
1713 * lock not needed as this is only incremented in the single
1714 * threaded kernel __init code path all other access are read
1715 * only
1716 */
1717 }
1718
1719 g_iommus = kzalloc(g_num_of_iommus * sizeof(*iommu), GFP_KERNEL);
1720 if (!g_iommus) {
1721 ret = -ENOMEM;
1722 goto error;
1723 }
1724
1725 deferred_flush = kzalloc(g_num_of_iommus *
1726 sizeof(struct deferred_flush_tables), GFP_KERNEL);
1727 if (!deferred_flush) {
1728 kfree(g_iommus);
1729 ret = -ENOMEM;
1730 goto error;
1731 }
1732
1733 i = 0;
1734 for_each_drhd_unit(drhd) {
1735 if (drhd->ignored)
1736 continue;
1737 iommu = alloc_iommu(&g_iommus[i], drhd);
1738 i++;
1739 if (!iommu) {
1740 ret = -ENOMEM;
1741 goto error;
1742 }
1743
1744 /*
1745 * TBD:
1746 * we could share the same root & context tables
1747 * amoung all IOMMU's. Need to Split it later.
1748 */
1749 ret = iommu_alloc_root_entry(iommu);
1750 if (ret) {
1751 printk(KERN_ERR "IOMMU: allocate root entry failed\n");
1752 goto error;
1753 }
1754 }
1755
1756 /*
1757 * For each rmrr
1758 * for each dev attached to rmrr
1759 * do
1760 * locate drhd for dev, alloc domain for dev
1761 * allocate free domain
1762 * allocate page table entries for rmrr
1763 * if context not allocated for bus
1764 * allocate and init context
1765 * set present in root table for this bus
1766 * init context with domain, translation etc
1767 * endfor
1768 * endfor
1769 */
1770 for_each_rmrr_units(rmrr) {
1771 for (i = 0; i < rmrr->devices_cnt; i++) {
1772 pdev = rmrr->devices[i];
1773 /* some BIOS lists non-exist devices in DMAR table */
1774 if (!pdev)
1775 continue;
1776 ret = iommu_prepare_rmrr_dev(rmrr, pdev);
1777 if (ret)
1778 printk(KERN_ERR
1779 "IOMMU: mapping reserved region failed\n");
1780 }
1781 }
1782
1783 iommu_prepare_gfx_mapping();
1784
1785 iommu_prepare_isa();
1786
1787 /*
1788 * for each drhd
1789 * enable fault log
1790 * global invalidate context cache
1791 * global invalidate iotlb
1792 * enable translation
1793 */
1794 for_each_drhd_unit(drhd) {
1795 if (drhd->ignored)
1796 continue;
1797 iommu = drhd->iommu;
1798 sprintf (iommu->name, "dmar%d", unit++);
1799
1800 iommu_flush_write_buffer(iommu);
1801
1802 ret = dmar_set_interrupt(iommu);
1803 if (ret)
1804 goto error;
1805
1806 iommu_set_root_entry(iommu);
1807
1808 iommu_flush_context_global(iommu, 0);
1809 iommu_flush_iotlb_global(iommu, 0);
1810
1811 iommu_disable_protect_mem_regions(iommu);
1812
1813 ret = iommu_enable_translation(iommu);
1814 if (ret)
1815 goto error;
1816 }
1817
1818 return 0;
1819 error:
1820 for_each_drhd_unit(drhd) {
1821 if (drhd->ignored)
1822 continue;
1823 iommu = drhd->iommu;
1824 free_iommu(iommu);
1825 }
1826 kfree(g_iommus);
1827 return ret;
1828 }
1829
1830 static inline u64 aligned_size(u64 host_addr, size_t size)
1831 {
1832 u64 addr;
1833 addr = (host_addr & (~PAGE_MASK_4K)) + size;
1834 return PAGE_ALIGN_4K(addr);
1835 }
1836
1837 struct iova *
1838 iommu_alloc_iova(struct dmar_domain *domain, size_t size, u64 end)
1839 {
1840 struct iova *piova;
1841
1842 /* Make sure it's in range */
1843 end = min_t(u64, DOMAIN_MAX_ADDR(domain->gaw), end);
1844 if (!size || (IOVA_START_ADDR + size > end))
1845 return NULL;
1846
1847 piova = alloc_iova(&domain->iovad,
1848 size >> PAGE_SHIFT_4K, IOVA_PFN(end), 1);
1849 return piova;
1850 }
1851
1852 static struct iova *
1853 __intel_alloc_iova(struct device *dev, struct dmar_domain *domain,
1854 size_t size)
1855 {
1856 struct pci_dev *pdev = to_pci_dev(dev);
1857 struct iova *iova = NULL;
1858
1859 if ((pdev->dma_mask <= DMA_32BIT_MASK) || (dmar_forcedac)) {
1860 iova = iommu_alloc_iova(domain, size, pdev->dma_mask);
1861 } else {
1862 /*
1863 * First try to allocate an io virtual address in
1864 * DMA_32BIT_MASK and if that fails then try allocating
1865 * from higher range
1866 */
1867 iova = iommu_alloc_iova(domain, size, DMA_32BIT_MASK);
1868 if (!iova)
1869 iova = iommu_alloc_iova(domain, size, pdev->dma_mask);
1870 }
1871
1872 if (!iova) {
1873 printk(KERN_ERR"Allocating iova for %s failed", pci_name(pdev));
1874 return NULL;
1875 }
1876
1877 return iova;
1878 }
1879
1880 static struct dmar_domain *
1881 get_valid_domain_for_dev(struct pci_dev *pdev)
1882 {
1883 struct dmar_domain *domain;
1884 int ret;
1885
1886 domain = get_domain_for_dev(pdev,
1887 DEFAULT_DOMAIN_ADDRESS_WIDTH);
1888 if (!domain) {
1889 printk(KERN_ERR
1890 "Allocating domain for %s failed", pci_name(pdev));
1891 return NULL;
1892 }
1893
1894 /* make sure context mapping is ok */
1895 if (unlikely(!domain_context_mapped(domain, pdev))) {
1896 ret = domain_context_mapping(domain, pdev);
1897 if (ret) {
1898 printk(KERN_ERR
1899 "Domain context map for %s failed",
1900 pci_name(pdev));
1901 return NULL;
1902 }
1903 }
1904
1905 return domain;
1906 }
1907
1908 static dma_addr_t
1909 intel_map_single(struct device *hwdev, phys_addr_t paddr, size_t size, int dir)
1910 {
1911 struct pci_dev *pdev = to_pci_dev(hwdev);
1912 struct dmar_domain *domain;
1913 unsigned long start_paddr;
1914 struct iova *iova;
1915 int prot = 0;
1916 int ret;
1917
1918 BUG_ON(dir == DMA_NONE);
1919 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
1920 return paddr;
1921
1922 domain = get_valid_domain_for_dev(pdev);
1923 if (!domain)
1924 return 0;
1925
1926 size = aligned_size((u64)paddr, size);
1927
1928 iova = __intel_alloc_iova(hwdev, domain, size);
1929 if (!iova)
1930 goto error;
1931
1932 start_paddr = iova->pfn_lo << PAGE_SHIFT_4K;
1933
1934 /*
1935 * Check if DMAR supports zero-length reads on write only
1936 * mappings..
1937 */
1938 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
1939 !cap_zlr(domain->iommu->cap))
1940 prot |= DMA_PTE_READ;
1941 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
1942 prot |= DMA_PTE_WRITE;
1943 /*
1944 * paddr - (paddr + size) might be partial page, we should map the whole
1945 * page. Note: if two part of one page are separately mapped, we
1946 * might have two guest_addr mapping to the same host paddr, but this
1947 * is not a big problem
1948 */
1949 ret = domain_page_mapping(domain, start_paddr,
1950 ((u64)paddr) & PAGE_MASK_4K, size, prot);
1951 if (ret)
1952 goto error;
1953
1954 pr_debug("Device %s request: %lx@%llx mapping: %lx@%llx, dir %d\n",
1955 pci_name(pdev), size, (u64)paddr,
1956 size, (u64)start_paddr, dir);
1957
1958 /* it's a non-present to present mapping */
1959 ret = iommu_flush_iotlb_psi(domain->iommu, domain->id,
1960 start_paddr, size >> PAGE_SHIFT_4K, 1);
1961 if (ret)
1962 iommu_flush_write_buffer(domain->iommu);
1963
1964 return (start_paddr + ((u64)paddr & (~PAGE_MASK_4K)));
1965
1966 error:
1967 if (iova)
1968 __free_iova(&domain->iovad, iova);
1969 printk(KERN_ERR"Device %s request: %lx@%llx dir %d --- failed\n",
1970 pci_name(pdev), size, (u64)paddr, dir);
1971 return 0;
1972 }
1973
1974 static void flush_unmaps(void)
1975 {
1976 int i, j;
1977
1978 timer_on = 0;
1979
1980 /* just flush them all */
1981 for (i = 0; i < g_num_of_iommus; i++) {
1982 if (deferred_flush[i].next) {
1983 iommu_flush_iotlb_global(&g_iommus[i], 0);
1984 for (j = 0; j < deferred_flush[i].next; j++) {
1985 __free_iova(&deferred_flush[i].domain[j]->iovad,
1986 deferred_flush[i].iova[j]);
1987 }
1988 deferred_flush[i].next = 0;
1989 }
1990 }
1991
1992 list_size = 0;
1993 }
1994
1995 static void flush_unmaps_timeout(unsigned long data)
1996 {
1997 unsigned long flags;
1998
1999 spin_lock_irqsave(&async_umap_flush_lock, flags);
2000 flush_unmaps();
2001 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2002 }
2003
2004 static void add_unmap(struct dmar_domain *dom, struct iova *iova)
2005 {
2006 unsigned long flags;
2007 int next, iommu_id;
2008
2009 spin_lock_irqsave(&async_umap_flush_lock, flags);
2010 if (list_size == HIGH_WATER_MARK)
2011 flush_unmaps();
2012
2013 iommu_id = dom->iommu - g_iommus;
2014 next = deferred_flush[iommu_id].next;
2015 deferred_flush[iommu_id].domain[next] = dom;
2016 deferred_flush[iommu_id].iova[next] = iova;
2017 deferred_flush[iommu_id].next++;
2018
2019 if (!timer_on) {
2020 mod_timer(&unmap_timer, jiffies + msecs_to_jiffies(10));
2021 timer_on = 1;
2022 }
2023 list_size++;
2024 spin_unlock_irqrestore(&async_umap_flush_lock, flags);
2025 }
2026
2027 static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr,
2028 size_t size, int dir)
2029 {
2030 struct pci_dev *pdev = to_pci_dev(dev);
2031 struct dmar_domain *domain;
2032 unsigned long start_addr;
2033 struct iova *iova;
2034
2035 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2036 return;
2037 domain = find_domain(pdev);
2038 BUG_ON(!domain);
2039
2040 iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
2041 if (!iova)
2042 return;
2043
2044 start_addr = iova->pfn_lo << PAGE_SHIFT_4K;
2045 size = aligned_size((u64)dev_addr, size);
2046
2047 pr_debug("Device %s unmapping: %lx@%llx\n",
2048 pci_name(pdev), size, (u64)start_addr);
2049
2050 /* clear the whole page */
2051 dma_pte_clear_range(domain, start_addr, start_addr + size);
2052 /* free page tables */
2053 dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2054 if (intel_iommu_strict) {
2055 if (iommu_flush_iotlb_psi(domain->iommu,
2056 domain->id, start_addr, size >> PAGE_SHIFT_4K, 0))
2057 iommu_flush_write_buffer(domain->iommu);
2058 /* free iova */
2059 __free_iova(&domain->iovad, iova);
2060 } else {
2061 add_unmap(domain, iova);
2062 /*
2063 * queue up the release of the unmap to save the 1/6th of the
2064 * cpu used up by the iotlb flush operation...
2065 */
2066 }
2067 }
2068
2069 static void * intel_alloc_coherent(struct device *hwdev, size_t size,
2070 dma_addr_t *dma_handle, gfp_t flags)
2071 {
2072 void *vaddr;
2073 int order;
2074
2075 size = PAGE_ALIGN_4K(size);
2076 order = get_order(size);
2077 flags &= ~(GFP_DMA | GFP_DMA32);
2078
2079 vaddr = (void *)__get_free_pages(flags, order);
2080 if (!vaddr)
2081 return NULL;
2082 memset(vaddr, 0, size);
2083
2084 *dma_handle = intel_map_single(hwdev, virt_to_bus(vaddr), size, DMA_BIDIRECTIONAL);
2085 if (*dma_handle)
2086 return vaddr;
2087 free_pages((unsigned long)vaddr, order);
2088 return NULL;
2089 }
2090
2091 static void intel_free_coherent(struct device *hwdev, size_t size,
2092 void *vaddr, dma_addr_t dma_handle)
2093 {
2094 int order;
2095
2096 size = PAGE_ALIGN_4K(size);
2097 order = get_order(size);
2098
2099 intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
2100 free_pages((unsigned long)vaddr, order);
2101 }
2102
2103 #define SG_ENT_VIRT_ADDRESS(sg) (sg_virt((sg)))
2104 static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sglist,
2105 int nelems, int dir)
2106 {
2107 int i;
2108 struct pci_dev *pdev = to_pci_dev(hwdev);
2109 struct dmar_domain *domain;
2110 unsigned long start_addr;
2111 struct iova *iova;
2112 size_t size = 0;
2113 void *addr;
2114 struct scatterlist *sg;
2115
2116 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2117 return;
2118
2119 domain = find_domain(pdev);
2120
2121 iova = find_iova(&domain->iovad, IOVA_PFN(sglist[0].dma_address));
2122 if (!iova)
2123 return;
2124 for_each_sg(sglist, sg, nelems, i) {
2125 addr = SG_ENT_VIRT_ADDRESS(sg);
2126 size += aligned_size((u64)addr, sg->length);
2127 }
2128
2129 start_addr = iova->pfn_lo << PAGE_SHIFT_4K;
2130
2131 /* clear the whole page */
2132 dma_pte_clear_range(domain, start_addr, start_addr + size);
2133 /* free page tables */
2134 dma_pte_free_pagetable(domain, start_addr, start_addr + size);
2135
2136 if (iommu_flush_iotlb_psi(domain->iommu, domain->id, start_addr,
2137 size >> PAGE_SHIFT_4K, 0))
2138 iommu_flush_write_buffer(domain->iommu);
2139
2140 /* free iova */
2141 __free_iova(&domain->iovad, iova);
2142 }
2143
2144 static int intel_nontranslate_map_sg(struct device *hddev,
2145 struct scatterlist *sglist, int nelems, int dir)
2146 {
2147 int i;
2148 struct scatterlist *sg;
2149
2150 for_each_sg(sglist, sg, nelems, i) {
2151 BUG_ON(!sg_page(sg));
2152 sg->dma_address = virt_to_bus(SG_ENT_VIRT_ADDRESS(sg));
2153 sg->dma_length = sg->length;
2154 }
2155 return nelems;
2156 }
2157
2158 static int intel_map_sg(struct device *hwdev, struct scatterlist *sglist,
2159 int nelems, int dir)
2160 {
2161 void *addr;
2162 int i;
2163 struct pci_dev *pdev = to_pci_dev(hwdev);
2164 struct dmar_domain *domain;
2165 size_t size = 0;
2166 int prot = 0;
2167 size_t offset = 0;
2168 struct iova *iova = NULL;
2169 int ret;
2170 struct scatterlist *sg;
2171 unsigned long start_addr;
2172
2173 BUG_ON(dir == DMA_NONE);
2174 if (pdev->dev.archdata.iommu == DUMMY_DEVICE_DOMAIN_INFO)
2175 return intel_nontranslate_map_sg(hwdev, sglist, nelems, dir);
2176
2177 domain = get_valid_domain_for_dev(pdev);
2178 if (!domain)
2179 return 0;
2180
2181 for_each_sg(sglist, sg, nelems, i) {
2182 addr = SG_ENT_VIRT_ADDRESS(sg);
2183 addr = (void *)virt_to_phys(addr);
2184 size += aligned_size((u64)addr, sg->length);
2185 }
2186
2187 iova = __intel_alloc_iova(hwdev, domain, size);
2188 if (!iova) {
2189 sglist->dma_length = 0;
2190 return 0;
2191 }
2192
2193 /*
2194 * Check if DMAR supports zero-length reads on write only
2195 * mappings..
2196 */
2197 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
2198 !cap_zlr(domain->iommu->cap))
2199 prot |= DMA_PTE_READ;
2200 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
2201 prot |= DMA_PTE_WRITE;
2202
2203 start_addr = iova->pfn_lo << PAGE_SHIFT_4K;
2204 offset = 0;
2205 for_each_sg(sglist, sg, nelems, i) {
2206 addr = SG_ENT_VIRT_ADDRESS(sg);
2207 addr = (void *)virt_to_phys(addr);
2208 size = aligned_size((u64)addr, sg->length);
2209 ret = domain_page_mapping(domain, start_addr + offset,
2210 ((u64)addr) & PAGE_MASK_4K,
2211 size, prot);
2212 if (ret) {
2213 /* clear the page */
2214 dma_pte_clear_range(domain, start_addr,
2215 start_addr + offset);
2216 /* free page tables */
2217 dma_pte_free_pagetable(domain, start_addr,
2218 start_addr + offset);
2219 /* free iova */
2220 __free_iova(&domain->iovad, iova);
2221 return 0;
2222 }
2223 sg->dma_address = start_addr + offset +
2224 ((u64)addr & (~PAGE_MASK_4K));
2225 sg->dma_length = sg->length;
2226 offset += size;
2227 }
2228
2229 /* it's a non-present to present mapping */
2230 if (iommu_flush_iotlb_psi(domain->iommu, domain->id,
2231 start_addr, offset >> PAGE_SHIFT_4K, 1))
2232 iommu_flush_write_buffer(domain->iommu);
2233 return nelems;
2234 }
2235
2236 static struct dma_mapping_ops intel_dma_ops = {
2237 .alloc_coherent = intel_alloc_coherent,
2238 .free_coherent = intel_free_coherent,
2239 .map_single = intel_map_single,
2240 .unmap_single = intel_unmap_single,
2241 .map_sg = intel_map_sg,
2242 .unmap_sg = intel_unmap_sg,
2243 };
2244
2245 static inline int iommu_domain_cache_init(void)
2246 {
2247 int ret = 0;
2248
2249 iommu_domain_cache = kmem_cache_create("iommu_domain",
2250 sizeof(struct dmar_domain),
2251 0,
2252 SLAB_HWCACHE_ALIGN,
2253
2254 NULL);
2255 if (!iommu_domain_cache) {
2256 printk(KERN_ERR "Couldn't create iommu_domain cache\n");
2257 ret = -ENOMEM;
2258 }
2259
2260 return ret;
2261 }
2262
2263 static inline int iommu_devinfo_cache_init(void)
2264 {
2265 int ret = 0;
2266
2267 iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
2268 sizeof(struct device_domain_info),
2269 0,
2270 SLAB_HWCACHE_ALIGN,
2271
2272 NULL);
2273 if (!iommu_devinfo_cache) {
2274 printk(KERN_ERR "Couldn't create devinfo cache\n");
2275 ret = -ENOMEM;
2276 }
2277
2278 return ret;
2279 }
2280
2281 static inline int iommu_iova_cache_init(void)
2282 {
2283 int ret = 0;
2284
2285 iommu_iova_cache = kmem_cache_create("iommu_iova",
2286 sizeof(struct iova),
2287 0,
2288 SLAB_HWCACHE_ALIGN,
2289
2290 NULL);
2291 if (!iommu_iova_cache) {
2292 printk(KERN_ERR "Couldn't create iova cache\n");
2293 ret = -ENOMEM;
2294 }
2295
2296 return ret;
2297 }
2298
2299 static int __init iommu_init_mempool(void)
2300 {
2301 int ret;
2302 ret = iommu_iova_cache_init();
2303 if (ret)
2304 return ret;
2305
2306 ret = iommu_domain_cache_init();
2307 if (ret)
2308 goto domain_error;
2309
2310 ret = iommu_devinfo_cache_init();
2311 if (!ret)
2312 return ret;
2313
2314 kmem_cache_destroy(iommu_domain_cache);
2315 domain_error:
2316 kmem_cache_destroy(iommu_iova_cache);
2317
2318 return -ENOMEM;
2319 }
2320
2321 static void __init iommu_exit_mempool(void)
2322 {
2323 kmem_cache_destroy(iommu_devinfo_cache);
2324 kmem_cache_destroy(iommu_domain_cache);
2325 kmem_cache_destroy(iommu_iova_cache);
2326
2327 }
2328
2329 void __init detect_intel_iommu(void)
2330 {
2331 if (swiotlb || no_iommu || iommu_detected || dmar_disabled)
2332 return;
2333 if (early_dmar_detect()) {
2334 iommu_detected = 1;
2335 }
2336 }
2337
2338 static void __init init_no_remapping_devices(void)
2339 {
2340 struct dmar_drhd_unit *drhd;
2341
2342 for_each_drhd_unit(drhd) {
2343 if (!drhd->include_all) {
2344 int i;
2345 for (i = 0; i < drhd->devices_cnt; i++)
2346 if (drhd->devices[i] != NULL)
2347 break;
2348 /* ignore DMAR unit if no pci devices exist */
2349 if (i == drhd->devices_cnt)
2350 drhd->ignored = 1;
2351 }
2352 }
2353
2354 if (dmar_map_gfx)
2355 return;
2356
2357 for_each_drhd_unit(drhd) {
2358 int i;
2359 if (drhd->ignored || drhd->include_all)
2360 continue;
2361
2362 for (i = 0; i < drhd->devices_cnt; i++)
2363 if (drhd->devices[i] &&
2364 !IS_GFX_DEVICE(drhd->devices[i]))
2365 break;
2366
2367 if (i < drhd->devices_cnt)
2368 continue;
2369
2370 /* bypass IOMMU if it is just for gfx devices */
2371 drhd->ignored = 1;
2372 for (i = 0; i < drhd->devices_cnt; i++) {
2373 if (!drhd->devices[i])
2374 continue;
2375 drhd->devices[i]->dev.archdata.iommu = DUMMY_DEVICE_DOMAIN_INFO;
2376 }
2377 }
2378 }
2379
2380 int __init intel_iommu_init(void)
2381 {
2382 int ret = 0;
2383
2384 if (no_iommu || swiotlb || dmar_disabled)
2385 return -ENODEV;
2386
2387 if (dmar_table_init())
2388 return -ENODEV;
2389
2390 iommu_init_mempool();
2391 dmar_init_reserved_ranges();
2392
2393 init_no_remapping_devices();
2394
2395 ret = init_dmars();
2396 if (ret) {
2397 printk(KERN_ERR "IOMMU: dmar init failed\n");
2398 put_iova_domain(&reserved_iova_list);
2399 iommu_exit_mempool();
2400 return ret;
2401 }
2402 printk(KERN_INFO
2403 "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
2404
2405 init_timer(&unmap_timer);
2406 force_iommu = 1;
2407 dma_ops = &intel_dma_ops;
2408 return 0;
2409 }
2410
Community contributions to the Openmoko Kernel