?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
perf tools: Fix find tids routine by excluding "." and ".."
[cris-mirror.git] / arch / x86 / kernel / amd_iommu.c
blobfa5a1474cd182db250e189cfc8b262683f2585ba
1 /*
2 * Copyright (C) 2007-2009 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20 #include <linux/pci.h>
21 #include <linux/bitmap.h>
22 #include <linux/slab.h>
23 #include <linux/debugfs.h>
24 #include <linux/scatterlist.h>
25 #include <linux/dma-mapping.h>
26 #include <linux/iommu-helper.h>
27 #include <linux/iommu.h>
28 #include <asm/proto.h>
29 #include <asm/iommu.h>
30 #include <asm/gart.h>
31 #include <asm/amd_iommu_proto.h>
32 #include <asm/amd_iommu_types.h>
33 #include <asm/amd_iommu.h>
34
35 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
36
37 #define EXIT_LOOP_COUNT 10000000
38
39 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
40
41 /* A list of preallocated protection domains */
42 static LIST_HEAD(iommu_pd_list);
43 static DEFINE_SPINLOCK(iommu_pd_list_lock);
44
45 /*
46 * Domain for untranslated devices - only allocated
47 * if iommu=pt passed on kernel cmd line.
48 */
49 static struct protection_domain *pt_domain;
50
51 static struct iommu_ops amd_iommu_ops;
52
53 /*
54 * general struct to manage commands send to an IOMMU
55 */
56 struct iommu_cmd {
57 u32 data[4];
58 };
59
60 static void reset_iommu_command_buffer(struct amd_iommu *iommu);
61 static void update_domain(struct protection_domain *domain);
62
63 /****************************************************************************
64 *
65 * Helper functions
66 *
67 ****************************************************************************/
68
69 static inline u16 get_device_id(struct device *dev)
70 {
71 struct pci_dev *pdev = to_pci_dev(dev);
72
73 return calc_devid(pdev->bus->number, pdev->devfn);
74 }
75
76 static struct iommu_dev_data *get_dev_data(struct device *dev)
77 {
78 return dev->archdata.iommu;
79 }
80
81 /*
82 * In this function the list of preallocated protection domains is traversed to
83 * find the domain for a specific device
84 */
85 static struct dma_ops_domain *find_protection_domain(u16 devid)
86 {
87 struct dma_ops_domain *entry, *ret = NULL;
88 unsigned long flags;
89 u16 alias = amd_iommu_alias_table[devid];
90
91 if (list_empty(&iommu_pd_list))
92 return NULL;
93
94 spin_lock_irqsave(&iommu_pd_list_lock, flags);
95
96 list_for_each_entry(entry, &iommu_pd_list, list) {
97 if (entry->target_dev == devid ||
98 entry->target_dev == alias) {
99 ret = entry;
100 break;
101 }
102 }
103
104 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
105
106 return ret;
107 }
108
109 /*
110 * This function checks if the driver got a valid device from the caller to
111 * avoid dereferencing invalid pointers.
112 */
113 static bool check_device(struct device *dev)
114 {
115 u16 devid;
116
117 if (!dev || !dev->dma_mask)
118 return false;
119
120 /* No device or no PCI device */
121 if (dev->bus != &pci_bus_type)
122 return false;
123
124 devid = get_device_id(dev);
125
126 /* Out of our scope? */
127 if (devid > amd_iommu_last_bdf)
128 return false;
129
130 if (amd_iommu_rlookup_table[devid] == NULL)
131 return false;
132
133 return true;
134 }
135
136 static int iommu_init_device(struct device *dev)
137 {
138 struct iommu_dev_data *dev_data;
139 struct pci_dev *pdev;
140 u16 devid, alias;
141
142 if (dev->archdata.iommu)
143 return 0;
144
145 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
146 if (!dev_data)
147 return -ENOMEM;
148
149 dev_data->dev = dev;
150
151 devid = get_device_id(dev);
152 alias = amd_iommu_alias_table[devid];
153 pdev = pci_get_bus_and_slot(PCI_BUS(alias), alias & 0xff);
154 if (pdev)
155 dev_data->alias = &pdev->dev;
156
157 atomic_set(&dev_data->bind, 0);
158
159 dev->archdata.iommu = dev_data;
160
161
162 return 0;
163 }
164
165 static void iommu_uninit_device(struct device *dev)
166 {
167 kfree(dev->archdata.iommu);
168 }
169
170 void __init amd_iommu_uninit_devices(void)
171 {
172 struct pci_dev *pdev = NULL;
173
174 for_each_pci_dev(pdev) {
175
176 if (!check_device(&pdev->dev))
177 continue;
178
179 iommu_uninit_device(&pdev->dev);
180 }
181 }
182
183 int __init amd_iommu_init_devices(void)
184 {
185 struct pci_dev *pdev = NULL;
186 int ret = 0;
187
188 for_each_pci_dev(pdev) {
189
190 if (!check_device(&pdev->dev))
191 continue;
192
193 ret = iommu_init_device(&pdev->dev);
194 if (ret)
195 goto out_free;
196 }
197
198 return 0;
199
200 out_free:
201
202 amd_iommu_uninit_devices();
203
204 return ret;
205 }
206 #ifdef CONFIG_AMD_IOMMU_STATS
207
208 /*
209 * Initialization code for statistics collection
210 */
211
212 DECLARE_STATS_COUNTER(compl_wait);
213 DECLARE_STATS_COUNTER(cnt_map_single);
214 DECLARE_STATS_COUNTER(cnt_unmap_single);
215 DECLARE_STATS_COUNTER(cnt_map_sg);
216 DECLARE_STATS_COUNTER(cnt_unmap_sg);
217 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
218 DECLARE_STATS_COUNTER(cnt_free_coherent);
219 DECLARE_STATS_COUNTER(cross_page);
220 DECLARE_STATS_COUNTER(domain_flush_single);
221 DECLARE_STATS_COUNTER(domain_flush_all);
222 DECLARE_STATS_COUNTER(alloced_io_mem);
223 DECLARE_STATS_COUNTER(total_map_requests);
224
225 static struct dentry *stats_dir;
226 static struct dentry *de_fflush;
227
228 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
229 {
230 if (stats_dir == NULL)
231 return;
232
233 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
234 &cnt->value);
235 }
236
237 static void amd_iommu_stats_init(void)
238 {
239 stats_dir = debugfs_create_dir("amd-iommu", NULL);
240 if (stats_dir == NULL)
241 return;
242
243 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
244 (u32 *)&amd_iommu_unmap_flush);
245
246 amd_iommu_stats_add(&compl_wait);
247 amd_iommu_stats_add(&cnt_map_single);
248 amd_iommu_stats_add(&cnt_unmap_single);
249 amd_iommu_stats_add(&cnt_map_sg);
250 amd_iommu_stats_add(&cnt_unmap_sg);
251 amd_iommu_stats_add(&cnt_alloc_coherent);
252 amd_iommu_stats_add(&cnt_free_coherent);
253 amd_iommu_stats_add(&cross_page);
254 amd_iommu_stats_add(&domain_flush_single);
255 amd_iommu_stats_add(&domain_flush_all);
256 amd_iommu_stats_add(&alloced_io_mem);
257 amd_iommu_stats_add(&total_map_requests);
258 }
259
260 #endif
261
262 /****************************************************************************
263 *
264 * Interrupt handling functions
265 *
266 ****************************************************************************/
267
268 static void dump_dte_entry(u16 devid)
269 {
270 int i;
271
272 for (i = 0; i < 8; ++i)
273 pr_err("AMD-Vi: DTE[%d]: %08x\n", i,
274 amd_iommu_dev_table[devid].data[i]);
275 }
276
277 static void dump_command(unsigned long phys_addr)
278 {
279 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
280 int i;
281
282 for (i = 0; i < 4; ++i)
283 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
284 }
285
286 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
287 {
288 u32 *event = __evt;
289 int type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
290 int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
291 int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
292 int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
293 u64 address = (u64)(((u64)event[3]) << 32) | event[2];
294
295 printk(KERN_ERR "AMD-Vi: Event logged [");
296
297 switch (type) {
298 case EVENT_TYPE_ILL_DEV:
299 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
300 "address=0x%016llx flags=0x%04x]\n",
301 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
302 address, flags);
303 dump_dte_entry(devid);
304 break;
305 case EVENT_TYPE_IO_FAULT:
306 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
307 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
308 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
309 domid, address, flags);
310 break;
311 case EVENT_TYPE_DEV_TAB_ERR:
312 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
313 "address=0x%016llx flags=0x%04x]\n",
314 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
315 address, flags);
316 break;
317 case EVENT_TYPE_PAGE_TAB_ERR:
318 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
319 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
320 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
321 domid, address, flags);
322 break;
323 case EVENT_TYPE_ILL_CMD:
324 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
325 iommu->reset_in_progress = true;
326 reset_iommu_command_buffer(iommu);
327 dump_command(address);
328 break;
329 case EVENT_TYPE_CMD_HARD_ERR:
330 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
331 "flags=0x%04x]\n", address, flags);
332 break;
333 case EVENT_TYPE_IOTLB_INV_TO:
334 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
335 "address=0x%016llx]\n",
336 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
337 address);
338 break;
339 case EVENT_TYPE_INV_DEV_REQ:
340 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
341 "address=0x%016llx flags=0x%04x]\n",
342 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
343 address, flags);
344 break;
345 default:
346 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
347 }
348 }
349
350 static void iommu_poll_events(struct amd_iommu *iommu)
351 {
352 u32 head, tail;
353 unsigned long flags;
354
355 spin_lock_irqsave(&iommu->lock, flags);
356
357 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
358 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
359
360 while (head != tail) {
361 iommu_print_event(iommu, iommu->evt_buf + head);
362 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
363 }
364
365 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
366
367 spin_unlock_irqrestore(&iommu->lock, flags);
368 }
369
370 irqreturn_t amd_iommu_int_handler(int irq, void *data)
371 {
372 struct amd_iommu *iommu;
373
374 for_each_iommu(iommu)
375 iommu_poll_events(iommu);
376
377 return IRQ_HANDLED;
378 }
379
380 /****************************************************************************
381 *
382 * IOMMU command queuing functions
383 *
384 ****************************************************************************/
385
386 /*
387 * Writes the command to the IOMMUs command buffer and informs the
388 * hardware about the new command. Must be called with iommu->lock held.
389 */
390 static int __iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
391 {
392 u32 tail, head;
393 u8 *target;
394
395 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
396 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
397 target = iommu->cmd_buf + tail;
398 memcpy_toio(target, cmd, sizeof(*cmd));
399 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
400 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
401 if (tail == head)
402 return -ENOMEM;
403 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
404
405 return 0;
406 }
407
408 /*
409 * General queuing function for commands. Takes iommu->lock and calls
410 * __iommu_queue_command().
411 */
412 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
413 {
414 unsigned long flags;
415 int ret;
416
417 spin_lock_irqsave(&iommu->lock, flags);
418 ret = __iommu_queue_command(iommu, cmd);
419 if (!ret)
420 iommu->need_sync = true;
421 spin_unlock_irqrestore(&iommu->lock, flags);
422
423 return ret;
424 }
425
426 /*
427 * This function waits until an IOMMU has completed a completion
428 * wait command
429 */
430 static void __iommu_wait_for_completion(struct amd_iommu *iommu)
431 {
432 int ready = 0;
433 unsigned status = 0;
434 unsigned long i = 0;
435
436 INC_STATS_COUNTER(compl_wait);
437
438 while (!ready && (i < EXIT_LOOP_COUNT)) {
439 ++i;
440 /* wait for the bit to become one */
441 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
442 ready = status & MMIO_STATUS_COM_WAIT_INT_MASK;
443 }
444
445 /* set bit back to zero */
446 status &= ~MMIO_STATUS_COM_WAIT_INT_MASK;
447 writel(status, iommu->mmio_base + MMIO_STATUS_OFFSET);
448
449 if (unlikely(i == EXIT_LOOP_COUNT))
450 iommu->reset_in_progress = true;
451 }
452
453 /*
454 * This function queues a completion wait command into the command
455 * buffer of an IOMMU
456 */
457 static int __iommu_completion_wait(struct amd_iommu *iommu)
458 {
459 struct iommu_cmd cmd;
460
461 memset(&cmd, 0, sizeof(cmd));
462 cmd.data[0] = CMD_COMPL_WAIT_INT_MASK;
463 CMD_SET_TYPE(&cmd, CMD_COMPL_WAIT);
464
465 return __iommu_queue_command(iommu, &cmd);
466 }
467
468 /*
469 * This function is called whenever we need to ensure that the IOMMU has
470 * completed execution of all commands we sent. It sends a
471 * COMPLETION_WAIT command and waits for it to finish. The IOMMU informs
472 * us about that by writing a value to a physical address we pass with
473 * the command.
474 */
475 static int iommu_completion_wait(struct amd_iommu *iommu)
476 {
477 int ret = 0;
478 unsigned long flags;
479
480 spin_lock_irqsave(&iommu->lock, flags);
481
482 if (!iommu->need_sync)
483 goto out;
484
485 ret = __iommu_completion_wait(iommu);
486
487 iommu->need_sync = false;
488
489 if (ret)
490 goto out;
491
492 __iommu_wait_for_completion(iommu);
493
494 out:
495 spin_unlock_irqrestore(&iommu->lock, flags);
496
497 if (iommu->reset_in_progress)
498 reset_iommu_command_buffer(iommu);
499
500 return 0;
501 }
502
503 static void iommu_flush_complete(struct protection_domain *domain)
504 {
505 int i;
506
507 for (i = 0; i < amd_iommus_present; ++i) {
508 if (!domain->dev_iommu[i])
509 continue;
510
511 /*
512 * Devices of this domain are behind this IOMMU
513 * We need to wait for completion of all commands.
514 */
515 iommu_completion_wait(amd_iommus[i]);
516 }
517 }
518
519 /*
520 * Command send function for invalidating a device table entry
521 */
522 static int iommu_flush_device(struct device *dev)
523 {
524 struct amd_iommu *iommu;
525 struct iommu_cmd cmd;
526 u16 devid;
527
528 devid = get_device_id(dev);
529 iommu = amd_iommu_rlookup_table[devid];
530
531 /* Build command */
532 memset(&cmd, 0, sizeof(cmd));
533 CMD_SET_TYPE(&cmd, CMD_INV_DEV_ENTRY);
534 cmd.data[0] = devid;
535
536 return iommu_queue_command(iommu, &cmd);
537 }
538
539 static void __iommu_build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
540 u16 domid, int pde, int s)
541 {
542 memset(cmd, 0, sizeof(*cmd));
543 address &= PAGE_MASK;
544 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
545 cmd->data[1] |= domid;
546 cmd->data[2] = lower_32_bits(address);
547 cmd->data[3] = upper_32_bits(address);
548 if (s) /* size bit - we flush more than one 4kb page */
549 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
550 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
551 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
552 }
553
554 /*
555 * Generic command send function for invalidaing TLB entries
556 */
557 static int iommu_queue_inv_iommu_pages(struct amd_iommu *iommu,
558 u64 address, u16 domid, int pde, int s)
559 {
560 struct iommu_cmd cmd;
561 int ret;
562
563 __iommu_build_inv_iommu_pages(&cmd, address, domid, pde, s);
564
565 ret = iommu_queue_command(iommu, &cmd);
566
567 return ret;
568 }
569
570 /*
571 * TLB invalidation function which is called from the mapping functions.
572 * It invalidates a single PTE if the range to flush is within a single
573 * page. Otherwise it flushes the whole TLB of the IOMMU.
574 */
575 static void __iommu_flush_pages(struct protection_domain *domain,
576 u64 address, size_t size, int pde)
577 {
578 int s = 0, i;
579 unsigned long pages = iommu_num_pages(address, size, PAGE_SIZE);
580
581 address &= PAGE_MASK;
582
583 if (pages > 1) {
584 /*
585 * If we have to flush more than one page, flush all
586 * TLB entries for this domain
587 */
588 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
589 s = 1;
590 }
591
592
593 for (i = 0; i < amd_iommus_present; ++i) {
594 if (!domain->dev_iommu[i])
595 continue;
596
597 /*
598 * Devices of this domain are behind this IOMMU
599 * We need a TLB flush
600 */
601 iommu_queue_inv_iommu_pages(amd_iommus[i], address,
602 domain->id, pde, s);
603 }
604
605 return;
606 }
607
608 static void iommu_flush_pages(struct protection_domain *domain,
609 u64 address, size_t size)
610 {
611 __iommu_flush_pages(domain, address, size, 0);
612 }
613
614 /* Flush the whole IO/TLB for a given protection domain */
615 static void iommu_flush_tlb(struct protection_domain *domain)
616 {
617 __iommu_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
618 }
619
620 /* Flush the whole IO/TLB for a given protection domain - including PDE */
621 static void iommu_flush_tlb_pde(struct protection_domain *domain)
622 {
623 __iommu_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
624 }
625
626
627 /*
628 * This function flushes the DTEs for all devices in domain
629 */
630 static void iommu_flush_domain_devices(struct protection_domain *domain)
631 {
632 struct iommu_dev_data *dev_data;
633 unsigned long flags;
634
635 spin_lock_irqsave(&domain->lock, flags);
636
637 list_for_each_entry(dev_data, &domain->dev_list, list)
638 iommu_flush_device(dev_data->dev);
639
640 spin_unlock_irqrestore(&domain->lock, flags);
641 }
642
643 static void iommu_flush_all_domain_devices(void)
644 {
645 struct protection_domain *domain;
646 unsigned long flags;
647
648 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
649
650 list_for_each_entry(domain, &amd_iommu_pd_list, list) {
651 iommu_flush_domain_devices(domain);
652 iommu_flush_complete(domain);
653 }
654
655 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
656 }
657
658 void amd_iommu_flush_all_devices(void)
659 {
660 iommu_flush_all_domain_devices();
661 }
662
663 /*
664 * This function uses heavy locking and may disable irqs for some time. But
665 * this is no issue because it is only called during resume.
666 */
667 void amd_iommu_flush_all_domains(void)
668 {
669 struct protection_domain *domain;
670 unsigned long flags;
671
672 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
673
674 list_for_each_entry(domain, &amd_iommu_pd_list, list) {
675 spin_lock(&domain->lock);
676 iommu_flush_tlb_pde(domain);
677 iommu_flush_complete(domain);
678 spin_unlock(&domain->lock);
679 }
680
681 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
682 }
683
684 static void reset_iommu_command_buffer(struct amd_iommu *iommu)
685 {
686 pr_err("AMD-Vi: Resetting IOMMU command buffer\n");
687
688 if (iommu->reset_in_progress)
689 panic("AMD-Vi: ILLEGAL_COMMAND_ERROR while resetting command buffer\n");
690
691 amd_iommu_reset_cmd_buffer(iommu);
692 amd_iommu_flush_all_devices();
693 amd_iommu_flush_all_domains();
694
695 iommu->reset_in_progress = false;
696 }
697
698 /****************************************************************************
699 *
700 * The functions below are used the create the page table mappings for
701 * unity mapped regions.
702 *
703 ****************************************************************************/
704
705 /*
706 * This function is used to add another level to an IO page table. Adding
707 * another level increases the size of the address space by 9 bits to a size up
708 * to 64 bits.
709 */
710 static bool increase_address_space(struct protection_domain *domain,
711 gfp_t gfp)
712 {
713 u64 *pte;
714
715 if (domain->mode == PAGE_MODE_6_LEVEL)
716 /* address space already 64 bit large */
717 return false;
718
719 pte = (void *)get_zeroed_page(gfp);
720 if (!pte)
721 return false;
722
723 *pte = PM_LEVEL_PDE(domain->mode,
724 virt_to_phys(domain->pt_root));
725 domain->pt_root = pte;
726 domain->mode += 1;
727 domain->updated = true;
728
729 return true;
730 }
731
732 static u64 *alloc_pte(struct protection_domain *domain,
733 unsigned long address,
734 unsigned long page_size,
735 u64 **pte_page,
736 gfp_t gfp)
737 {
738 int level, end_lvl;
739 u64 *pte, *page;
740
741 BUG_ON(!is_power_of_2(page_size));
742
743 while (address > PM_LEVEL_SIZE(domain->mode))
744 increase_address_space(domain, gfp);
745
746 level = domain->mode - 1;
747 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
748 address = PAGE_SIZE_ALIGN(address, page_size);
749 end_lvl = PAGE_SIZE_LEVEL(page_size);
750
751 while (level > end_lvl) {
752 if (!IOMMU_PTE_PRESENT(*pte)) {
753 page = (u64 *)get_zeroed_page(gfp);
754 if (!page)
755 return NULL;
756 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
757 }
758
759 /* No level skipping support yet */
760 if (PM_PTE_LEVEL(*pte) != level)
761 return NULL;
762
763 level -= 1;
764
765 pte = IOMMU_PTE_PAGE(*pte);
766
767 if (pte_page && level == end_lvl)
768 *pte_page = pte;
769
770 pte = &pte[PM_LEVEL_INDEX(level, address)];
771 }
772
773 return pte;
774 }
775
776 /*
777 * This function checks if there is a PTE for a given dma address. If
778 * there is one, it returns the pointer to it.
779 */
780 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
781 {
782 int level;
783 u64 *pte;
784
785 if (address > PM_LEVEL_SIZE(domain->mode))
786 return NULL;
787
788 level = domain->mode - 1;
789 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
790
791 while (level > 0) {
792
793 /* Not Present */
794 if (!IOMMU_PTE_PRESENT(*pte))
795 return NULL;
796
797 /* Large PTE */
798 if (PM_PTE_LEVEL(*pte) == 0x07) {
799 unsigned long pte_mask, __pte;
800
801 /*
802 * If we have a series of large PTEs, make
803 * sure to return a pointer to the first one.
804 */
805 pte_mask = PTE_PAGE_SIZE(*pte);
806 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
807 __pte = ((unsigned long)pte) & pte_mask;
808
809 return (u64 *)__pte;
810 }
811
812 /* No level skipping support yet */
813 if (PM_PTE_LEVEL(*pte) != level)
814 return NULL;
815
816 level -= 1;
817
818 /* Walk to the next level */
819 pte = IOMMU_PTE_PAGE(*pte);
820 pte = &pte[PM_LEVEL_INDEX(level, address)];
821 }
822
823 return pte;
824 }
825
826 /*
827 * Generic mapping functions. It maps a physical address into a DMA
828 * address space. It allocates the page table pages if necessary.
829 * In the future it can be extended to a generic mapping function
830 * supporting all features of AMD IOMMU page tables like level skipping
831 * and full 64 bit address spaces.
832 */
833 static int iommu_map_page(struct protection_domain *dom,
834 unsigned long bus_addr,
835 unsigned long phys_addr,
836 int prot,
837 unsigned long page_size)
838 {
839 u64 __pte, *pte;
840 int i, count;
841
842 if (!(prot & IOMMU_PROT_MASK))
843 return -EINVAL;
844
845 bus_addr = PAGE_ALIGN(bus_addr);
846 phys_addr = PAGE_ALIGN(phys_addr);
847 count = PAGE_SIZE_PTE_COUNT(page_size);
848 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
849
850 for (i = 0; i < count; ++i)
851 if (IOMMU_PTE_PRESENT(pte[i]))
852 return -EBUSY;
853
854 if (page_size > PAGE_SIZE) {
855 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
856 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
857 } else
858 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
859
860 if (prot & IOMMU_PROT_IR)
861 __pte |= IOMMU_PTE_IR;
862 if (prot & IOMMU_PROT_IW)
863 __pte |= IOMMU_PTE_IW;
864
865 for (i = 0; i < count; ++i)
866 pte[i] = __pte;
867
868 update_domain(dom);
869
870 return 0;
871 }
872
873 static unsigned long iommu_unmap_page(struct protection_domain *dom,
874 unsigned long bus_addr,
875 unsigned long page_size)
876 {
877 unsigned long long unmap_size, unmapped;
878 u64 *pte;
879
880 BUG_ON(!is_power_of_2(page_size));
881
882 unmapped = 0;
883
884 while (unmapped < page_size) {
885
886 pte = fetch_pte(dom, bus_addr);
887
888 if (!pte) {
889 /*
890 * No PTE for this address
891 * move forward in 4kb steps
892 */
893 unmap_size = PAGE_SIZE;
894 } else if (PM_PTE_LEVEL(*pte) == 0) {
895 /* 4kb PTE found for this address */
896 unmap_size = PAGE_SIZE;
897 *pte = 0ULL;
898 } else {
899 int count, i;
900
901 /* Large PTE found which maps this address */
902 unmap_size = PTE_PAGE_SIZE(*pte);
903 count = PAGE_SIZE_PTE_COUNT(unmap_size);
904 for (i = 0; i < count; i++)
905 pte[i] = 0ULL;
906 }
907
908 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
909 unmapped += unmap_size;
910 }
911
912 BUG_ON(!is_power_of_2(unmapped));
913
914 return unmapped;
915 }
916
917 /*
918 * This function checks if a specific unity mapping entry is needed for
919 * this specific IOMMU.
920 */
921 static int iommu_for_unity_map(struct amd_iommu *iommu,
922 struct unity_map_entry *entry)
923 {
924 u16 bdf, i;
925
926 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
927 bdf = amd_iommu_alias_table[i];
928 if (amd_iommu_rlookup_table[bdf] == iommu)
929 return 1;
930 }
931
932 return 0;
933 }
934
935 /*
936 * This function actually applies the mapping to the page table of the
937 * dma_ops domain.
938 */
939 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
940 struct unity_map_entry *e)
941 {
942 u64 addr;
943 int ret;
944
945 for (addr = e->address_start; addr < e->address_end;
946 addr += PAGE_SIZE) {
947 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
948 PAGE_SIZE);
949 if (ret)
950 return ret;
951 /*
952 * if unity mapping is in aperture range mark the page
953 * as allocated in the aperture
954 */
955 if (addr < dma_dom->aperture_size)
956 __set_bit(addr >> PAGE_SHIFT,
957 dma_dom->aperture[0]->bitmap);
958 }
959
960 return 0;
961 }
962
963 /*
964 * Init the unity mappings for a specific IOMMU in the system
965 *
966 * Basically iterates over all unity mapping entries and applies them to
967 * the default domain DMA of that IOMMU if necessary.
968 */
969 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
970 {
971 struct unity_map_entry *entry;
972 int ret;
973
974 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
975 if (!iommu_for_unity_map(iommu, entry))
976 continue;
977 ret = dma_ops_unity_map(iommu->default_dom, entry);
978 if (ret)
979 return ret;
980 }
981
982 return 0;
983 }
984
985 /*
986 * Inits the unity mappings required for a specific device
987 */
988 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
989 u16 devid)
990 {
991 struct unity_map_entry *e;
992 int ret;
993
994 list_for_each_entry(e, &amd_iommu_unity_map, list) {
995 if (!(devid >= e->devid_start && devid <= e->devid_end))
996 continue;
997 ret = dma_ops_unity_map(dma_dom, e);
998 if (ret)
999 return ret;
1000 }
1001
1002 return 0;
1003 }
1004
1005 /****************************************************************************
1006 *
1007 * The next functions belong to the address allocator for the dma_ops
1008 * interface functions. They work like the allocators in the other IOMMU
1009 * drivers. Its basically a bitmap which marks the allocated pages in
1010 * the aperture. Maybe it could be enhanced in the future to a more
1011 * efficient allocator.
1012 *
1013 ****************************************************************************/
1014
1015 /*
1016 * The address allocator core functions.
1017 *
1018 * called with domain->lock held
1019 */
1020
1021 /*
1022 * Used to reserve address ranges in the aperture (e.g. for exclusion
1023 * ranges.
1024 */
1025 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1026 unsigned long start_page,
1027 unsigned int pages)
1028 {
1029 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1030
1031 if (start_page + pages > last_page)
1032 pages = last_page - start_page;
1033
1034 for (i = start_page; i < start_page + pages; ++i) {
1035 int index = i / APERTURE_RANGE_PAGES;
1036 int page = i % APERTURE_RANGE_PAGES;
1037 __set_bit(page, dom->aperture[index]->bitmap);
1038 }
1039 }
1040
1041 /*
1042 * This function is used to add a new aperture range to an existing
1043 * aperture in case of dma_ops domain allocation or address allocation
1044 * failure.
1045 */
1046 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1047 bool populate, gfp_t gfp)
1048 {
1049 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1050 struct amd_iommu *iommu;
1051 unsigned long i;
1052
1053 #ifdef CONFIG_IOMMU_STRESS
1054 populate = false;
1055 #endif
1056
1057 if (index >= APERTURE_MAX_RANGES)
1058 return -ENOMEM;
1059
1060 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1061 if (!dma_dom->aperture[index])
1062 return -ENOMEM;
1063
1064 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1065 if (!dma_dom->aperture[index]->bitmap)
1066 goto out_free;
1067
1068 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1069
1070 if (populate) {
1071 unsigned long address = dma_dom->aperture_size;
1072 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1073 u64 *pte, *pte_page;
1074
1075 for (i = 0; i < num_ptes; ++i) {
1076 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1077 &pte_page, gfp);
1078 if (!pte)
1079 goto out_free;
1080
1081 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1082
1083 address += APERTURE_RANGE_SIZE / 64;
1084 }
1085 }
1086
1087 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1088
1089 /* Intialize the exclusion range if necessary */
1090 for_each_iommu(iommu) {
1091 if (iommu->exclusion_start &&
1092 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1093 && iommu->exclusion_start < dma_dom->aperture_size) {
1094 unsigned long startpage;
1095 int pages = iommu_num_pages(iommu->exclusion_start,
1096 iommu->exclusion_length,
1097 PAGE_SIZE);
1098 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1099 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1100 }
1101 }
1102
1103 /*
1104 * Check for areas already mapped as present in the new aperture
1105 * range and mark those pages as reserved in the allocator. Such
1106 * mappings may already exist as a result of requested unity
1107 * mappings for devices.
1108 */
1109 for (i = dma_dom->aperture[index]->offset;
1110 i < dma_dom->aperture_size;
1111 i += PAGE_SIZE) {
1112 u64 *pte = fetch_pte(&dma_dom->domain, i);
1113 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1114 continue;
1115
1116 dma_ops_reserve_addresses(dma_dom, i << PAGE_SHIFT, 1);
1117 }
1118
1119 update_domain(&dma_dom->domain);
1120
1121 return 0;
1122
1123 out_free:
1124 update_domain(&dma_dom->domain);
1125
1126 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1127
1128 kfree(dma_dom->aperture[index]);
1129 dma_dom->aperture[index] = NULL;
1130
1131 return -ENOMEM;
1132 }
1133
1134 static unsigned long dma_ops_area_alloc(struct device *dev,
1135 struct dma_ops_domain *dom,
1136 unsigned int pages,
1137 unsigned long align_mask,
1138 u64 dma_mask,
1139 unsigned long start)
1140 {
1141 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1142 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1143 int i = start >> APERTURE_RANGE_SHIFT;
1144 unsigned long boundary_size;
1145 unsigned long address = -1;
1146 unsigned long limit;
1147
1148 next_bit >>= PAGE_SHIFT;
1149
1150 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1151 PAGE_SIZE) >> PAGE_SHIFT;
1152
1153 for (;i < max_index; ++i) {
1154 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1155
1156 if (dom->aperture[i]->offset >= dma_mask)
1157 break;
1158
1159 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1160 dma_mask >> PAGE_SHIFT);
1161
1162 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1163 limit, next_bit, pages, 0,
1164 boundary_size, align_mask);
1165 if (address != -1) {
1166 address = dom->aperture[i]->offset +
1167 (address << PAGE_SHIFT);
1168 dom->next_address = address + (pages << PAGE_SHIFT);
1169 break;
1170 }
1171
1172 next_bit = 0;
1173 }
1174
1175 return address;
1176 }
1177
1178 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1179 struct dma_ops_domain *dom,
1180 unsigned int pages,
1181 unsigned long align_mask,
1182 u64 dma_mask)
1183 {
1184 unsigned long address;
1185
1186 #ifdef CONFIG_IOMMU_STRESS
1187 dom->next_address = 0;
1188 dom->need_flush = true;
1189 #endif
1190
1191 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1192 dma_mask, dom->next_address);
1193
1194 if (address == -1) {
1195 dom->next_address = 0;
1196 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1197 dma_mask, 0);
1198 dom->need_flush = true;
1199 }
1200
1201 if (unlikely(address == -1))
1202 address = DMA_ERROR_CODE;
1203
1204 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1205
1206 return address;
1207 }
1208
1209 /*
1210 * The address free function.
1211 *
1212 * called with domain->lock held
1213 */
1214 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1215 unsigned long address,
1216 unsigned int pages)
1217 {
1218 unsigned i = address >> APERTURE_RANGE_SHIFT;
1219 struct aperture_range *range = dom->aperture[i];
1220
1221 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1222
1223 #ifdef CONFIG_IOMMU_STRESS
1224 if (i < 4)
1225 return;
1226 #endif
1227
1228 if (address >= dom->next_address)
1229 dom->need_flush = true;
1230
1231 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1232
1233 bitmap_clear(range->bitmap, address, pages);
1234
1235 }
1236
1237 /****************************************************************************
1238 *
1239 * The next functions belong to the domain allocation. A domain is
1240 * allocated for every IOMMU as the default domain. If device isolation
1241 * is enabled, every device get its own domain. The most important thing
1242 * about domains is the page table mapping the DMA address space they
1243 * contain.
1244 *
1245 ****************************************************************************/
1246
1247 /*
1248 * This function adds a protection domain to the global protection domain list
1249 */
1250 static void add_domain_to_list(struct protection_domain *domain)
1251 {
1252 unsigned long flags;
1253
1254 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1255 list_add(&domain->list, &amd_iommu_pd_list);
1256 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1257 }
1258
1259 /*
1260 * This function removes a protection domain to the global
1261 * protection domain list
1262 */
1263 static void del_domain_from_list(struct protection_domain *domain)
1264 {
1265 unsigned long flags;
1266
1267 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1268 list_del(&domain->list);
1269 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1270 }
1271
1272 static u16 domain_id_alloc(void)
1273 {
1274 unsigned long flags;
1275 int id;
1276
1277 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1278 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1279 BUG_ON(id == 0);
1280 if (id > 0 && id < MAX_DOMAIN_ID)
1281 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1282 else
1283 id = 0;
1284 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1285
1286 return id;
1287 }
1288
1289 static void domain_id_free(int id)
1290 {
1291 unsigned long flags;
1292
1293 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1294 if (id > 0 && id < MAX_DOMAIN_ID)
1295 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1296 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1297 }
1298
1299 static void free_pagetable(struct protection_domain *domain)
1300 {
1301 int i, j;
1302 u64 *p1, *p2, *p3;
1303
1304 p1 = domain->pt_root;
1305
1306 if (!p1)
1307 return;
1308
1309 for (i = 0; i < 512; ++i) {
1310 if (!IOMMU_PTE_PRESENT(p1[i]))
1311 continue;
1312
1313 p2 = IOMMU_PTE_PAGE(p1[i]);
1314 for (j = 0; j < 512; ++j) {
1315 if (!IOMMU_PTE_PRESENT(p2[j]))
1316 continue;
1317 p3 = IOMMU_PTE_PAGE(p2[j]);
1318 free_page((unsigned long)p3);
1319 }
1320
1321 free_page((unsigned long)p2);
1322 }
1323
1324 free_page((unsigned long)p1);
1325
1326 domain->pt_root = NULL;
1327 }
1328
1329 /*
1330 * Free a domain, only used if something went wrong in the
1331 * allocation path and we need to free an already allocated page table
1332 */
1333 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1334 {
1335 int i;
1336
1337 if (!dom)
1338 return;
1339
1340 del_domain_from_list(&dom->domain);
1341
1342 free_pagetable(&dom->domain);
1343
1344 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1345 if (!dom->aperture[i])
1346 continue;
1347 free_page((unsigned long)dom->aperture[i]->bitmap);
1348 kfree(dom->aperture[i]);
1349 }
1350
1351 kfree(dom);
1352 }
1353
1354 /*
1355 * Allocates a new protection domain usable for the dma_ops functions.
1356 * It also intializes the page table and the address allocator data
1357 * structures required for the dma_ops interface
1358 */
1359 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1360 {
1361 struct dma_ops_domain *dma_dom;
1362
1363 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1364 if (!dma_dom)
1365 return NULL;
1366
1367 spin_lock_init(&dma_dom->domain.lock);
1368
1369 dma_dom->domain.id = domain_id_alloc();
1370 if (dma_dom->domain.id == 0)
1371 goto free_dma_dom;
1372 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1373 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1374 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1375 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1376 dma_dom->domain.priv = dma_dom;
1377 if (!dma_dom->domain.pt_root)
1378 goto free_dma_dom;
1379
1380 dma_dom->need_flush = false;
1381 dma_dom->target_dev = 0xffff;
1382
1383 add_domain_to_list(&dma_dom->domain);
1384
1385 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1386 goto free_dma_dom;
1387
1388 /*
1389 * mark the first page as allocated so we never return 0 as
1390 * a valid dma-address. So we can use 0 as error value
1391 */
1392 dma_dom->aperture[0]->bitmap[0] = 1;
1393 dma_dom->next_address = 0;
1394
1395
1396 return dma_dom;
1397
1398 free_dma_dom:
1399 dma_ops_domain_free(dma_dom);
1400
1401 return NULL;
1402 }
1403
1404 /*
1405 * little helper function to check whether a given protection domain is a
1406 * dma_ops domain
1407 */
1408 static bool dma_ops_domain(struct protection_domain *domain)
1409 {
1410 return domain->flags & PD_DMA_OPS_MASK;
1411 }
1412
1413 static void set_dte_entry(u16 devid, struct protection_domain *domain)
1414 {
1415 u64 pte_root = virt_to_phys(domain->pt_root);
1416
1417 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1418 << DEV_ENTRY_MODE_SHIFT;
1419 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1420
1421 amd_iommu_dev_table[devid].data[2] = domain->id;
1422 amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
1423 amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
1424 }
1425
1426 static void clear_dte_entry(u16 devid)
1427 {
1428 /* remove entry from the device table seen by the hardware */
1429 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1430 amd_iommu_dev_table[devid].data[1] = 0;
1431 amd_iommu_dev_table[devid].data[2] = 0;
1432
1433 amd_iommu_apply_erratum_63(devid);
1434 }
1435
1436 static void do_attach(struct device *dev, struct protection_domain *domain)
1437 {
1438 struct iommu_dev_data *dev_data;
1439 struct amd_iommu *iommu;
1440 u16 devid;
1441
1442 devid = get_device_id(dev);
1443 iommu = amd_iommu_rlookup_table[devid];
1444 dev_data = get_dev_data(dev);
1445
1446 /* Update data structures */
1447 dev_data->domain = domain;
1448 list_add(&dev_data->list, &domain->dev_list);
1449 set_dte_entry(devid, domain);
1450
1451 /* Do reference counting */
1452 domain->dev_iommu[iommu->index] += 1;
1453 domain->dev_cnt += 1;
1454
1455 /* Flush the DTE entry */
1456 iommu_flush_device(dev);
1457 }
1458
1459 static void do_detach(struct device *dev)
1460 {
1461 struct iommu_dev_data *dev_data;
1462 struct amd_iommu *iommu;
1463 u16 devid;
1464
1465 devid = get_device_id(dev);
1466 iommu = amd_iommu_rlookup_table[devid];
1467 dev_data = get_dev_data(dev);
1468
1469 /* decrease reference counters */
1470 dev_data->domain->dev_iommu[iommu->index] -= 1;
1471 dev_data->domain->dev_cnt -= 1;
1472
1473 /* Update data structures */
1474 dev_data->domain = NULL;
1475 list_del(&dev_data->list);
1476 clear_dte_entry(devid);
1477
1478 /* Flush the DTE entry */
1479 iommu_flush_device(dev);
1480 }
1481
1482 /*
1483 * If a device is not yet associated with a domain, this function does
1484 * assigns it visible for the hardware
1485 */
1486 static int __attach_device(struct device *dev,
1487 struct protection_domain *domain)
1488 {
1489 struct iommu_dev_data *dev_data, *alias_data;
1490
1491 dev_data = get_dev_data(dev);
1492 alias_data = get_dev_data(dev_data->alias);
1493
1494 if (!alias_data)
1495 return -EINVAL;
1496
1497 /* lock domain */
1498 spin_lock(&domain->lock);
1499
1500 /* Some sanity checks */
1501 if (alias_data->domain != NULL &&
1502 alias_data->domain != domain)
1503 return -EBUSY;
1504
1505 if (dev_data->domain != NULL &&
1506 dev_data->domain != domain)
1507 return -EBUSY;
1508
1509 /* Do real assignment */
1510 if (dev_data->alias != dev) {
1511 alias_data = get_dev_data(dev_data->alias);
1512 if (alias_data->domain == NULL)
1513 do_attach(dev_data->alias, domain);
1514
1515 atomic_inc(&alias_data->bind);
1516 }
1517
1518 if (dev_data->domain == NULL)
1519 do_attach(dev, domain);
1520
1521 atomic_inc(&dev_data->bind);
1522
1523 /* ready */
1524 spin_unlock(&domain->lock);
1525
1526 return 0;
1527 }
1528
1529 /*
1530 * If a device is not yet associated with a domain, this function does
1531 * assigns it visible for the hardware
1532 */
1533 static int attach_device(struct device *dev,
1534 struct protection_domain *domain)
1535 {
1536 unsigned long flags;
1537 int ret;
1538
1539 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1540 ret = __attach_device(dev, domain);
1541 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1542
1543 /*
1544 * We might boot into a crash-kernel here. The crashed kernel
1545 * left the caches in the IOMMU dirty. So we have to flush
1546 * here to evict all dirty stuff.
1547 */
1548 iommu_flush_tlb_pde(domain);
1549
1550 return ret;
1551 }
1552
1553 /*
1554 * Removes a device from a protection domain (unlocked)
1555 */
1556 static void __detach_device(struct device *dev)
1557 {
1558 struct iommu_dev_data *dev_data = get_dev_data(dev);
1559 struct iommu_dev_data *alias_data;
1560 struct protection_domain *domain;
1561 unsigned long flags;
1562
1563 BUG_ON(!dev_data->domain);
1564
1565 domain = dev_data->domain;
1566
1567 spin_lock_irqsave(&domain->lock, flags);
1568
1569 if (dev_data->alias != dev) {
1570 alias_data = get_dev_data(dev_data->alias);
1571 if (atomic_dec_and_test(&alias_data->bind))
1572 do_detach(dev_data->alias);
1573 }
1574
1575 if (atomic_dec_and_test(&dev_data->bind))
1576 do_detach(dev);
1577
1578 spin_unlock_irqrestore(&domain->lock, flags);
1579
1580 /*
1581 * If we run in passthrough mode the device must be assigned to the
1582 * passthrough domain if it is detached from any other domain.
1583 * Make sure we can deassign from the pt_domain itself.
1584 */
1585 if (iommu_pass_through &&
1586 (dev_data->domain == NULL && domain != pt_domain))
1587 __attach_device(dev, pt_domain);
1588 }
1589
1590 /*
1591 * Removes a device from a protection domain (with devtable_lock held)
1592 */
1593 static void detach_device(struct device *dev)
1594 {
1595 unsigned long flags;
1596
1597 /* lock device table */
1598 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1599 __detach_device(dev);
1600 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1601 }
1602
1603 /*
1604 * Find out the protection domain structure for a given PCI device. This
1605 * will give us the pointer to the page table root for example.
1606 */
1607 static struct protection_domain *domain_for_device(struct device *dev)
1608 {
1609 struct protection_domain *dom;
1610 struct iommu_dev_data *dev_data, *alias_data;
1611 unsigned long flags;
1612 u16 devid, alias;
1613
1614 devid = get_device_id(dev);
1615 alias = amd_iommu_alias_table[devid];
1616 dev_data = get_dev_data(dev);
1617 alias_data = get_dev_data(dev_data->alias);
1618 if (!alias_data)
1619 return NULL;
1620
1621 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
1622 dom = dev_data->domain;
1623 if (dom == NULL &&
1624 alias_data->domain != NULL) {
1625 __attach_device(dev, alias_data->domain);
1626 dom = alias_data->domain;
1627 }
1628
1629 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1630
1631 return dom;
1632 }
1633
1634 static int device_change_notifier(struct notifier_block *nb,
1635 unsigned long action, void *data)
1636 {
1637 struct device *dev = data;
1638 u16 devid;
1639 struct protection_domain *domain;
1640 struct dma_ops_domain *dma_domain;
1641 struct amd_iommu *iommu;
1642 unsigned long flags;
1643
1644 if (!check_device(dev))
1645 return 0;
1646
1647 devid = get_device_id(dev);
1648 iommu = amd_iommu_rlookup_table[devid];
1649
1650 switch (action) {
1651 case BUS_NOTIFY_UNBOUND_DRIVER:
1652
1653 domain = domain_for_device(dev);
1654
1655 if (!domain)
1656 goto out;
1657 if (iommu_pass_through)
1658 break;
1659 detach_device(dev);
1660 break;
1661 case BUS_NOTIFY_ADD_DEVICE:
1662
1663 iommu_init_device(dev);
1664
1665 domain = domain_for_device(dev);
1666
1667 /* allocate a protection domain if a device is added */
1668 dma_domain = find_protection_domain(devid);
1669 if (dma_domain)
1670 goto out;
1671 dma_domain = dma_ops_domain_alloc();
1672 if (!dma_domain)
1673 goto out;
1674 dma_domain->target_dev = devid;
1675
1676 spin_lock_irqsave(&iommu_pd_list_lock, flags);
1677 list_add_tail(&dma_domain->list, &iommu_pd_list);
1678 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
1679
1680 break;
1681 case BUS_NOTIFY_DEL_DEVICE:
1682
1683 iommu_uninit_device(dev);
1684
1685 default:
1686 goto out;
1687 }
1688
1689 iommu_flush_device(dev);
1690 iommu_completion_wait(iommu);
1691
1692 out:
1693 return 0;
1694 }
1695
1696 static struct notifier_block device_nb = {
1697 .notifier_call = device_change_notifier,
1698 };
1699
1700 void amd_iommu_init_notifier(void)
1701 {
1702 bus_register_notifier(&pci_bus_type, &device_nb);
1703 }
1704
1705 /*****************************************************************************
1706 *
1707 * The next functions belong to the dma_ops mapping/unmapping code.
1708 *
1709 *****************************************************************************/
1710
1711 /*
1712 * In the dma_ops path we only have the struct device. This function
1713 * finds the corresponding IOMMU, the protection domain and the
1714 * requestor id for a given device.
1715 * If the device is not yet associated with a domain this is also done
1716 * in this function.
1717 */
1718 static struct protection_domain *get_domain(struct device *dev)
1719 {
1720 struct protection_domain *domain;
1721 struct dma_ops_domain *dma_dom;
1722 u16 devid = get_device_id(dev);
1723
1724 if (!check_device(dev))
1725 return ERR_PTR(-EINVAL);
1726
1727 domain = domain_for_device(dev);
1728 if (domain != NULL && !dma_ops_domain(domain))
1729 return ERR_PTR(-EBUSY);
1730
1731 if (domain != NULL)
1732 return domain;
1733
1734 /* Device not bount yet - bind it */
1735 dma_dom = find_protection_domain(devid);
1736 if (!dma_dom)
1737 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
1738 attach_device(dev, &dma_dom->domain);
1739 DUMP_printk("Using protection domain %d for device %s\n",
1740 dma_dom->domain.id, dev_name(dev));
1741
1742 return &dma_dom->domain;
1743 }
1744
1745 static void update_device_table(struct protection_domain *domain)
1746 {
1747 struct iommu_dev_data *dev_data;
1748
1749 list_for_each_entry(dev_data, &domain->dev_list, list) {
1750 u16 devid = get_device_id(dev_data->dev);
1751 set_dte_entry(devid, domain);
1752 }
1753 }
1754
1755 static void update_domain(struct protection_domain *domain)
1756 {
1757 if (!domain->updated)
1758 return;
1759
1760 update_device_table(domain);
1761 iommu_flush_domain_devices(domain);
1762 iommu_flush_tlb_pde(domain);
1763
1764 domain->updated = false;
1765 }
1766
1767 /*
1768 * This function fetches the PTE for a given address in the aperture
1769 */
1770 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
1771 unsigned long address)
1772 {
1773 struct aperture_range *aperture;
1774 u64 *pte, *pte_page;
1775
1776 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
1777 if (!aperture)
1778 return NULL;
1779
1780 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
1781 if (!pte) {
1782 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
1783 GFP_ATOMIC);
1784 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
1785 } else
1786 pte += PM_LEVEL_INDEX(0, address);
1787
1788 update_domain(&dom->domain);
1789
1790 return pte;
1791 }
1792
1793 /*
1794 * This is the generic map function. It maps one 4kb page at paddr to
1795 * the given address in the DMA address space for the domain.
1796 */
1797 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
1798 unsigned long address,
1799 phys_addr_t paddr,
1800 int direction)
1801 {
1802 u64 *pte, __pte;
1803
1804 WARN_ON(address > dom->aperture_size);
1805
1806 paddr &= PAGE_MASK;
1807
1808 pte = dma_ops_get_pte(dom, address);
1809 if (!pte)
1810 return DMA_ERROR_CODE;
1811
1812 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
1813
1814 if (direction == DMA_TO_DEVICE)
1815 __pte |= IOMMU_PTE_IR;
1816 else if (direction == DMA_FROM_DEVICE)
1817 __pte |= IOMMU_PTE_IW;
1818 else if (direction == DMA_BIDIRECTIONAL)
1819 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
1820
1821 WARN_ON(*pte);
1822
1823 *pte = __pte;
1824
1825 return (dma_addr_t)address;
1826 }
1827
1828 /*
1829 * The generic unmapping function for on page in the DMA address space.
1830 */
1831 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
1832 unsigned long address)
1833 {
1834 struct aperture_range *aperture;
1835 u64 *pte;
1836
1837 if (address >= dom->aperture_size)
1838 return;
1839
1840 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
1841 if (!aperture)
1842 return;
1843
1844 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
1845 if (!pte)
1846 return;
1847
1848 pte += PM_LEVEL_INDEX(0, address);
1849
1850 WARN_ON(!*pte);
1851
1852 *pte = 0ULL;
1853 }
1854
1855 /*
1856 * This function contains common code for mapping of a physically
1857 * contiguous memory region into DMA address space. It is used by all
1858 * mapping functions provided with this IOMMU driver.
1859 * Must be called with the domain lock held.
1860 */
1861 static dma_addr_t __map_single(struct device *dev,
1862 struct dma_ops_domain *dma_dom,
1863 phys_addr_t paddr,
1864 size_t size,
1865 int dir,
1866 bool align,
1867 u64 dma_mask)
1868 {
1869 dma_addr_t offset = paddr & ~PAGE_MASK;
1870 dma_addr_t address, start, ret;
1871 unsigned int pages;
1872 unsigned long align_mask = 0;
1873 int i;
1874
1875 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
1876 paddr &= PAGE_MASK;
1877
1878 INC_STATS_COUNTER(total_map_requests);
1879
1880 if (pages > 1)
1881 INC_STATS_COUNTER(cross_page);
1882
1883 if (align)
1884 align_mask = (1UL << get_order(size)) - 1;
1885
1886 retry:
1887 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
1888 dma_mask);
1889 if (unlikely(address == DMA_ERROR_CODE)) {
1890 /*
1891 * setting next_address here will let the address
1892 * allocator only scan the new allocated range in the
1893 * first run. This is a small optimization.
1894 */
1895 dma_dom->next_address = dma_dom->aperture_size;
1896
1897 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
1898 goto out;
1899
1900 /*
1901 * aperture was successfully enlarged by 128 MB, try
1902 * allocation again
1903 */
1904 goto retry;
1905 }
1906
1907 start = address;
1908 for (i = 0; i < pages; ++i) {
1909 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
1910 if (ret == DMA_ERROR_CODE)
1911 goto out_unmap;
1912
1913 paddr += PAGE_SIZE;
1914 start += PAGE_SIZE;
1915 }
1916 address += offset;
1917
1918 ADD_STATS_COUNTER(alloced_io_mem, size);
1919
1920 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
1921 iommu_flush_tlb(&dma_dom->domain);
1922 dma_dom->need_flush = false;
1923 } else if (unlikely(amd_iommu_np_cache))
1924 iommu_flush_pages(&dma_dom->domain, address, size);
1925
1926 out:
1927 return address;
1928
1929 out_unmap:
1930
1931 for (--i; i >= 0; --i) {
1932 start -= PAGE_SIZE;
1933 dma_ops_domain_unmap(dma_dom, start);
1934 }
1935
1936 dma_ops_free_addresses(dma_dom, address, pages);
1937
1938 return DMA_ERROR_CODE;
1939 }
1940
1941 /*
1942 * Does the reverse of the __map_single function. Must be called with
1943 * the domain lock held too
1944 */
1945 static void __unmap_single(struct dma_ops_domain *dma_dom,
1946 dma_addr_t dma_addr,
1947 size_t size,
1948 int dir)
1949 {
1950 dma_addr_t i, start;
1951 unsigned int pages;
1952
1953 if ((dma_addr == DMA_ERROR_CODE) ||
1954 (dma_addr + size > dma_dom->aperture_size))
1955 return;
1956
1957 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
1958 dma_addr &= PAGE_MASK;
1959 start = dma_addr;
1960
1961 for (i = 0; i < pages; ++i) {
1962 dma_ops_domain_unmap(dma_dom, start);
1963 start += PAGE_SIZE;
1964 }
1965
1966 SUB_STATS_COUNTER(alloced_io_mem, size);
1967
1968 dma_ops_free_addresses(dma_dom, dma_addr, pages);
1969
1970 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
1971 iommu_flush_pages(&dma_dom->domain, dma_addr, size);
1972 dma_dom->need_flush = false;
1973 }
1974 }
1975
1976 /*
1977 * The exported map_single function for dma_ops.
1978 */
1979 static dma_addr_t map_page(struct device *dev, struct page *page,
1980 unsigned long offset, size_t size,
1981 enum dma_data_direction dir,
1982 struct dma_attrs *attrs)
1983 {
1984 unsigned long flags;
1985 struct protection_domain *domain;
1986 dma_addr_t addr;
1987 u64 dma_mask;
1988 phys_addr_t paddr = page_to_phys(page) + offset;
1989
1990 INC_STATS_COUNTER(cnt_map_single);
1991
1992 domain = get_domain(dev);
1993 if (PTR_ERR(domain) == -EINVAL)
1994 return (dma_addr_t)paddr;
1995 else if (IS_ERR(domain))
1996 return DMA_ERROR_CODE;
1997
1998 dma_mask = *dev->dma_mask;
1999
2000 spin_lock_irqsave(&domain->lock, flags);
2001
2002 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2003 dma_mask);
2004 if (addr == DMA_ERROR_CODE)
2005 goto out;
2006
2007 iommu_flush_complete(domain);
2008
2009 out:
2010 spin_unlock_irqrestore(&domain->lock, flags);
2011
2012 return addr;
2013 }
2014
2015 /*
2016 * The exported unmap_single function for dma_ops.
2017 */
2018 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2019 enum dma_data_direction dir, struct dma_attrs *attrs)
2020 {
2021 unsigned long flags;
2022 struct protection_domain *domain;
2023
2024 INC_STATS_COUNTER(cnt_unmap_single);
2025
2026 domain = get_domain(dev);
2027 if (IS_ERR(domain))
2028 return;
2029
2030 spin_lock_irqsave(&domain->lock, flags);
2031
2032 __unmap_single(domain->priv, dma_addr, size, dir);
2033
2034 iommu_flush_complete(domain);
2035
2036 spin_unlock_irqrestore(&domain->lock, flags);
2037 }
2038
2039 /*
2040 * This is a special map_sg function which is used if we should map a
2041 * device which is not handled by an AMD IOMMU in the system.
2042 */
2043 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2044 int nelems, int dir)
2045 {
2046 struct scatterlist *s;
2047 int i;
2048
2049 for_each_sg(sglist, s, nelems, i) {
2050 s->dma_address = (dma_addr_t)sg_phys(s);
2051 s->dma_length = s->length;
2052 }
2053
2054 return nelems;
2055 }
2056
2057 /*
2058 * The exported map_sg function for dma_ops (handles scatter-gather
2059 * lists).
2060 */
2061 static int map_sg(struct device *dev, struct scatterlist *sglist,
2062 int nelems, enum dma_data_direction dir,
2063 struct dma_attrs *attrs)
2064 {
2065 unsigned long flags;
2066 struct protection_domain *domain;
2067 int i;
2068 struct scatterlist *s;
2069 phys_addr_t paddr;
2070 int mapped_elems = 0;
2071 u64 dma_mask;
2072
2073 INC_STATS_COUNTER(cnt_map_sg);
2074
2075 domain = get_domain(dev);
2076 if (PTR_ERR(domain) == -EINVAL)
2077 return map_sg_no_iommu(dev, sglist, nelems, dir);
2078 else if (IS_ERR(domain))
2079 return 0;
2080
2081 dma_mask = *dev->dma_mask;
2082
2083 spin_lock_irqsave(&domain->lock, flags);
2084
2085 for_each_sg(sglist, s, nelems, i) {
2086 paddr = sg_phys(s);
2087
2088 s->dma_address = __map_single(dev, domain->priv,
2089 paddr, s->length, dir, false,
2090 dma_mask);
2091
2092 if (s->dma_address) {
2093 s->dma_length = s->length;
2094 mapped_elems++;
2095 } else
2096 goto unmap;
2097 }
2098
2099 iommu_flush_complete(domain);
2100
2101 out:
2102 spin_unlock_irqrestore(&domain->lock, flags);
2103
2104 return mapped_elems;
2105 unmap:
2106 for_each_sg(sglist, s, mapped_elems, i) {
2107 if (s->dma_address)
2108 __unmap_single(domain->priv, s->dma_address,
2109 s->dma_length, dir);
2110 s->dma_address = s->dma_length = 0;
2111 }
2112
2113 mapped_elems = 0;
2114
2115 goto out;
2116 }
2117
2118 /*
2119 * The exported map_sg function for dma_ops (handles scatter-gather
2120 * lists).
2121 */
2122 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2123 int nelems, enum dma_data_direction dir,
2124 struct dma_attrs *attrs)
2125 {
2126 unsigned long flags;
2127 struct protection_domain *domain;
2128 struct scatterlist *s;
2129 int i;
2130
2131 INC_STATS_COUNTER(cnt_unmap_sg);
2132
2133 domain = get_domain(dev);
2134 if (IS_ERR(domain))
2135 return;
2136
2137 spin_lock_irqsave(&domain->lock, flags);
2138
2139 for_each_sg(sglist, s, nelems, i) {
2140 __unmap_single(domain->priv, s->dma_address,
2141 s->dma_length, dir);
2142 s->dma_address = s->dma_length = 0;
2143 }
2144
2145 iommu_flush_complete(domain);
2146
2147 spin_unlock_irqrestore(&domain->lock, flags);
2148 }
2149
2150 /*
2151 * The exported alloc_coherent function for dma_ops.
2152 */
2153 static void *alloc_coherent(struct device *dev, size_t size,
2154 dma_addr_t *dma_addr, gfp_t flag)
2155 {
2156 unsigned long flags;
2157 void *virt_addr;
2158 struct protection_domain *domain;
2159 phys_addr_t paddr;
2160 u64 dma_mask = dev->coherent_dma_mask;
2161
2162 INC_STATS_COUNTER(cnt_alloc_coherent);
2163
2164 domain = get_domain(dev);
2165 if (PTR_ERR(domain) == -EINVAL) {
2166 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2167 *dma_addr = __pa(virt_addr);
2168 return virt_addr;
2169 } else if (IS_ERR(domain))
2170 return NULL;
2171
2172 dma_mask = dev->coherent_dma_mask;
2173 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2174 flag |= __GFP_ZERO;
2175
2176 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2177 if (!virt_addr)
2178 return NULL;
2179
2180 paddr = virt_to_phys(virt_addr);
2181
2182 if (!dma_mask)
2183 dma_mask = *dev->dma_mask;
2184
2185 spin_lock_irqsave(&domain->lock, flags);
2186
2187 *dma_addr = __map_single(dev, domain->priv, paddr,
2188 size, DMA_BIDIRECTIONAL, true, dma_mask);
2189
2190 if (*dma_addr == DMA_ERROR_CODE) {
2191 spin_unlock_irqrestore(&domain->lock, flags);
2192 goto out_free;
2193 }
2194
2195 iommu_flush_complete(domain);
2196
2197 spin_unlock_irqrestore(&domain->lock, flags);
2198
2199 return virt_addr;
2200
2201 out_free:
2202
2203 free_pages((unsigned long)virt_addr, get_order(size));
2204
2205 return NULL;
2206 }
2207
2208 /*
2209 * The exported free_coherent function for dma_ops.
2210 */
2211 static void free_coherent(struct device *dev, size_t size,
2212 void *virt_addr, dma_addr_t dma_addr)
2213 {
2214 unsigned long flags;
2215 struct protection_domain *domain;
2216
2217 INC_STATS_COUNTER(cnt_free_coherent);
2218
2219 domain = get_domain(dev);
2220 if (IS_ERR(domain))
2221 goto free_mem;
2222
2223 spin_lock_irqsave(&domain->lock, flags);
2224
2225 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2226
2227 iommu_flush_complete(domain);
2228
2229 spin_unlock_irqrestore(&domain->lock, flags);
2230
2231 free_mem:
2232 free_pages((unsigned long)virt_addr, get_order(size));
2233 }
2234
2235 /*
2236 * This function is called by the DMA layer to find out if we can handle a
2237 * particular device. It is part of the dma_ops.
2238 */
2239 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2240 {
2241 return check_device(dev);
2242 }
2243
2244 /*
2245 * The function for pre-allocating protection domains.
2246 *
2247 * If the driver core informs the DMA layer if a driver grabs a device
2248 * we don't need to preallocate the protection domains anymore.
2249 * For now we have to.
2250 */
2251 static void prealloc_protection_domains(void)
2252 {
2253 struct pci_dev *dev = NULL;
2254 struct dma_ops_domain *dma_dom;
2255 u16 devid;
2256
2257 for_each_pci_dev(dev) {
2258
2259 /* Do we handle this device? */
2260 if (!check_device(&dev->dev))
2261 continue;
2262
2263 /* Is there already any domain for it? */
2264 if (domain_for_device(&dev->dev))
2265 continue;
2266
2267 devid = get_device_id(&dev->dev);
2268
2269 dma_dom = dma_ops_domain_alloc();
2270 if (!dma_dom)
2271 continue;
2272 init_unity_mappings_for_device(dma_dom, devid);
2273 dma_dom->target_dev = devid;
2274
2275 attach_device(&dev->dev, &dma_dom->domain);
2276
2277 list_add_tail(&dma_dom->list, &iommu_pd_list);
2278 }
2279 }
2280
2281 static struct dma_map_ops amd_iommu_dma_ops = {
2282 .alloc_coherent = alloc_coherent,
2283 .free_coherent = free_coherent,
2284 .map_page = map_page,
2285 .unmap_page = unmap_page,
2286 .map_sg = map_sg,
2287 .unmap_sg = unmap_sg,
2288 .dma_supported = amd_iommu_dma_supported,
2289 };
2290
2291 /*
2292 * The function which clues the AMD IOMMU driver into dma_ops.
2293 */
2294
2295 void __init amd_iommu_init_api(void)
2296 {
2297 register_iommu(&amd_iommu_ops);
2298 }
2299
2300 int __init amd_iommu_init_dma_ops(void)
2301 {
2302 struct amd_iommu *iommu;
2303 int ret;
2304
2305 /*
2306 * first allocate a default protection domain for every IOMMU we
2307 * found in the system. Devices not assigned to any other
2308 * protection domain will be assigned to the default one.
2309 */
2310 for_each_iommu(iommu) {
2311 iommu->default_dom = dma_ops_domain_alloc();
2312 if (iommu->default_dom == NULL)
2313 return -ENOMEM;
2314 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
2315 ret = iommu_init_unity_mappings(iommu);
2316 if (ret)
2317 goto free_domains;
2318 }
2319
2320 /*
2321 * Pre-allocate the protection domains for each device.
2322 */
2323 prealloc_protection_domains();
2324
2325 iommu_detected = 1;
2326 swiotlb = 0;
2327 #ifdef CONFIG_GART_IOMMU
2328 gart_iommu_aperture_disabled = 1;
2329 gart_iommu_aperture = 0;
2330 #endif
2331
2332 /* Make the driver finally visible to the drivers */
2333 dma_ops = &amd_iommu_dma_ops;
2334
2335 amd_iommu_stats_init();
2336
2337 return 0;
2338
2339 free_domains:
2340
2341 for_each_iommu(iommu) {
2342 if (iommu->default_dom)
2343 dma_ops_domain_free(iommu->default_dom);
2344 }
2345
2346 return ret;
2347 }
2348
2349 /*****************************************************************************
2350 *
2351 * The following functions belong to the exported interface of AMD IOMMU
2352 *
2353 * This interface allows access to lower level functions of the IOMMU
2354 * like protection domain handling and assignement of devices to domains
2355 * which is not possible with the dma_ops interface.
2356 *
2357 *****************************************************************************/
2358
2359 static void cleanup_domain(struct protection_domain *domain)
2360 {
2361 struct iommu_dev_data *dev_data, *next;
2362 unsigned long flags;
2363
2364 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2365
2366 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
2367 struct device *dev = dev_data->dev;
2368
2369 __detach_device(dev);
2370 atomic_set(&dev_data->bind, 0);
2371 }
2372
2373 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2374 }
2375
2376 static void protection_domain_free(struct protection_domain *domain)
2377 {
2378 if (!domain)
2379 return;
2380
2381 del_domain_from_list(domain);
2382
2383 if (domain->id)
2384 domain_id_free(domain->id);
2385
2386 kfree(domain);
2387 }
2388
2389 static struct protection_domain *protection_domain_alloc(void)
2390 {
2391 struct protection_domain *domain;
2392
2393 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2394 if (!domain)
2395 return NULL;
2396
2397 spin_lock_init(&domain->lock);
2398 mutex_init(&domain->api_lock);
2399 domain->id = domain_id_alloc();
2400 if (!domain->id)
2401 goto out_err;
2402 INIT_LIST_HEAD(&domain->dev_list);
2403
2404 add_domain_to_list(domain);
2405
2406 return domain;
2407
2408 out_err:
2409 kfree(domain);
2410
2411 return NULL;
2412 }
2413
2414 static int amd_iommu_domain_init(struct iommu_domain *dom)
2415 {
2416 struct protection_domain *domain;
2417
2418 domain = protection_domain_alloc();
2419 if (!domain)
2420 goto out_free;
2421
2422 domain->mode = PAGE_MODE_3_LEVEL;
2423 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2424 if (!domain->pt_root)
2425 goto out_free;
2426
2427 dom->priv = domain;
2428
2429 return 0;
2430
2431 out_free:
2432 protection_domain_free(domain);
2433
2434 return -ENOMEM;
2435 }
2436
2437 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
2438 {
2439 struct protection_domain *domain = dom->priv;
2440
2441 if (!domain)
2442 return;
2443
2444 if (domain->dev_cnt > 0)
2445 cleanup_domain(domain);
2446
2447 BUG_ON(domain->dev_cnt != 0);
2448
2449 free_pagetable(domain);
2450
2451 protection_domain_free(domain);
2452
2453 dom->priv = NULL;
2454 }
2455
2456 static void amd_iommu_detach_device(struct iommu_domain *dom,
2457 struct device *dev)
2458 {
2459 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2460 struct amd_iommu *iommu;
2461 u16 devid;
2462
2463 if (!check_device(dev))
2464 return;
2465
2466 devid = get_device_id(dev);
2467
2468 if (dev_data->domain != NULL)
2469 detach_device(dev);
2470
2471 iommu = amd_iommu_rlookup_table[devid];
2472 if (!iommu)
2473 return;
2474
2475 iommu_flush_device(dev);
2476 iommu_completion_wait(iommu);
2477 }
2478
2479 static int amd_iommu_attach_device(struct iommu_domain *dom,
2480 struct device *dev)
2481 {
2482 struct protection_domain *domain = dom->priv;
2483 struct iommu_dev_data *dev_data;
2484 struct amd_iommu *iommu;
2485 int ret;
2486 u16 devid;
2487
2488 if (!check_device(dev))
2489 return -EINVAL;
2490
2491 dev_data = dev->archdata.iommu;
2492
2493 devid = get_device_id(dev);
2494
2495 iommu = amd_iommu_rlookup_table[devid];
2496 if (!iommu)
2497 return -EINVAL;
2498
2499 if (dev_data->domain)
2500 detach_device(dev);
2501
2502 ret = attach_device(dev, domain);
2503
2504 iommu_completion_wait(iommu);
2505
2506 return ret;
2507 }
2508
2509 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2510 phys_addr_t paddr, int gfp_order, int iommu_prot)
2511 {
2512 unsigned long page_size = 0x1000UL << gfp_order;
2513 struct protection_domain *domain = dom->priv;
2514 int prot = 0;
2515 int ret;
2516
2517 if (iommu_prot & IOMMU_READ)
2518 prot |= IOMMU_PROT_IR;
2519 if (iommu_prot & IOMMU_WRITE)
2520 prot |= IOMMU_PROT_IW;
2521
2522 mutex_lock(&domain->api_lock);
2523 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
2524 mutex_unlock(&domain->api_lock);
2525
2526 return ret;
2527 }
2528
2529 static int amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2530 int gfp_order)
2531 {
2532 struct protection_domain *domain = dom->priv;
2533 unsigned long page_size, unmap_size;
2534
2535 page_size = 0x1000UL << gfp_order;
2536
2537 mutex_lock(&domain->api_lock);
2538 unmap_size = iommu_unmap_page(domain, iova, page_size);
2539 mutex_unlock(&domain->api_lock);
2540
2541 iommu_flush_tlb_pde(domain);
2542
2543 return get_order(unmap_size);
2544 }
2545
2546 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2547 unsigned long iova)
2548 {
2549 struct protection_domain *domain = dom->priv;
2550 unsigned long offset_mask;
2551 phys_addr_t paddr;
2552 u64 *pte, __pte;
2553
2554 pte = fetch_pte(domain, iova);
2555
2556 if (!pte || !IOMMU_PTE_PRESENT(*pte))
2557 return 0;
2558
2559 if (PM_PTE_LEVEL(*pte) == 0)
2560 offset_mask = PAGE_SIZE - 1;
2561 else
2562 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
2563
2564 __pte = *pte & PM_ADDR_MASK;
2565 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
2566
2567 return paddr;
2568 }
2569
2570 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
2571 unsigned long cap)
2572 {
2573 return 0;
2574 }
2575
2576 static struct iommu_ops amd_iommu_ops = {
2577 .domain_init = amd_iommu_domain_init,
2578 .domain_destroy = amd_iommu_domain_destroy,
2579 .attach_dev = amd_iommu_attach_device,
2580 .detach_dev = amd_iommu_detach_device,
2581 .map = amd_iommu_map,
2582 .unmap = amd_iommu_unmap,
2583 .iova_to_phys = amd_iommu_iova_to_phys,
2584 .domain_has_cap = amd_iommu_domain_has_cap,
2585 };
2586
2587 /*****************************************************************************
2588 *
2589 * The next functions do a basic initialization of IOMMU for pass through
2590 * mode
2591 *
2592 * In passthrough mode the IOMMU is initialized and enabled but not used for
2593 * DMA-API translation.
2594 *
2595 *****************************************************************************/
2596
2597 int __init amd_iommu_init_passthrough(void)
2598 {
2599 struct amd_iommu *iommu;
2600 struct pci_dev *dev = NULL;
2601 u16 devid;
2602
2603 /* allocate passthrough domain */
2604 pt_domain = protection_domain_alloc();
2605 if (!pt_domain)
2606 return -ENOMEM;
2607
2608 pt_domain->mode |= PAGE_MODE_NONE;
2609
2610 while ((dev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, dev)) != NULL) {
2611
2612 if (!check_device(&dev->dev))
2613 continue;
2614
2615 devid = get_device_id(&dev->dev);
2616
2617 iommu = amd_iommu_rlookup_table[devid];
2618 if (!iommu)
2619 continue;
2620
2621 attach_device(&dev->dev, pt_domain);
2622 }
2623
2624 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
2625
2626 return 0;
2627 }
CRIS linux kernel tree