?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
memcg: always create memsw files if CONFIG_CGROUP_MEM_RES_CTLR_SWAP
[linux-2.6.git] / drivers / iommu / amd_iommu.c
blobae2ec929e52f3e4ba76159e4df9448a6e315b115
1 /*
2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20 #include <linux/ratelimit.h>
21 #include <linux/pci.h>
22 #include <linux/pci-ats.h>
23 #include <linux/bitmap.h>
24 #include <linux/slab.h>
25 #include <linux/debugfs.h>
26 #include <linux/scatterlist.h>
27 #include <linux/dma-mapping.h>
28 #include <linux/iommu-helper.h>
29 #include <linux/iommu.h>
30 #include <linux/delay.h>
31 #include <linux/amd-iommu.h>
32 #include <linux/notifier.h>
33 #include <linux/export.h>
34 #include <asm/msidef.h>
35 #include <asm/proto.h>
36 #include <asm/iommu.h>
37 #include <asm/gart.h>
38 #include <asm/dma.h>
39
40 #include "amd_iommu_proto.h"
41 #include "amd_iommu_types.h"
42
43 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
44
45 #define LOOP_TIMEOUT 100000
46
47 /*
48 * This bitmap is used to advertise the page sizes our hardware support
49 * to the IOMMU core, which will then use this information to split
50 * physically contiguous memory regions it is mapping into page sizes
51 * that we support.
52 *
53 * Traditionally the IOMMU core just handed us the mappings directly,
54 * after making sure the size is an order of a 4KiB page and that the
55 * mapping has natural alignment.
56 *
57 * To retain this behavior, we currently advertise that we support
58 * all page sizes that are an order of 4KiB.
59 *
60 * If at some point we'd like to utilize the IOMMU core's new behavior,
61 * we could change this to advertise the real page sizes we support.
62 */
63 #define AMD_IOMMU_PGSIZES (~0xFFFUL)
64
65 static DEFINE_RWLOCK(amd_iommu_devtable_lock);
66
67 /* A list of preallocated protection domains */
68 static LIST_HEAD(iommu_pd_list);
69 static DEFINE_SPINLOCK(iommu_pd_list_lock);
70
71 /* List of all available dev_data structures */
72 static LIST_HEAD(dev_data_list);
73 static DEFINE_SPINLOCK(dev_data_list_lock);
74
75 /*
76 * Domain for untranslated devices - only allocated
77 * if iommu=pt passed on kernel cmd line.
78 */
79 static struct protection_domain *pt_domain;
80
81 static struct iommu_ops amd_iommu_ops;
82
83 static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
84 int amd_iommu_max_glx_val = -1;
85
86 /*
87 * general struct to manage commands send to an IOMMU
88 */
89 struct iommu_cmd {
90 u32 data[4];
91 };
92
93 static void update_domain(struct protection_domain *domain);
94 static int __init alloc_passthrough_domain(void);
95
96 /****************************************************************************
97 *
98 * Helper functions
99 *
100 ****************************************************************************/
101
102 static struct iommu_dev_data *alloc_dev_data(u16 devid)
103 {
104 struct iommu_dev_data *dev_data;
105 unsigned long flags;
106
107 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
108 if (!dev_data)
109 return NULL;
110
111 dev_data->devid = devid;
112 atomic_set(&dev_data->bind, 0);
113
114 spin_lock_irqsave(&dev_data_list_lock, flags);
115 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
116 spin_unlock_irqrestore(&dev_data_list_lock, flags);
117
118 return dev_data;
119 }
120
121 static void free_dev_data(struct iommu_dev_data *dev_data)
122 {
123 unsigned long flags;
124
125 spin_lock_irqsave(&dev_data_list_lock, flags);
126 list_del(&dev_data->dev_data_list);
127 spin_unlock_irqrestore(&dev_data_list_lock, flags);
128
129 kfree(dev_data);
130 }
131
132 static struct iommu_dev_data *search_dev_data(u16 devid)
133 {
134 struct iommu_dev_data *dev_data;
135 unsigned long flags;
136
137 spin_lock_irqsave(&dev_data_list_lock, flags);
138 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
139 if (dev_data->devid == devid)
140 goto out_unlock;
141 }
142
143 dev_data = NULL;
144
145 out_unlock:
146 spin_unlock_irqrestore(&dev_data_list_lock, flags);
147
148 return dev_data;
149 }
150
151 static struct iommu_dev_data *find_dev_data(u16 devid)
152 {
153 struct iommu_dev_data *dev_data;
154
155 dev_data = search_dev_data(devid);
156
157 if (dev_data == NULL)
158 dev_data = alloc_dev_data(devid);
159
160 return dev_data;
161 }
162
163 static inline u16 get_device_id(struct device *dev)
164 {
165 struct pci_dev *pdev = to_pci_dev(dev);
166
167 return calc_devid(pdev->bus->number, pdev->devfn);
168 }
169
170 static struct iommu_dev_data *get_dev_data(struct device *dev)
171 {
172 return dev->archdata.iommu;
173 }
174
175 static bool pci_iommuv2_capable(struct pci_dev *pdev)
176 {
177 static const int caps[] = {
178 PCI_EXT_CAP_ID_ATS,
179 PCI_EXT_CAP_ID_PRI,
180 PCI_EXT_CAP_ID_PASID,
181 };
182 int i, pos;
183
184 for (i = 0; i < 3; ++i) {
185 pos = pci_find_ext_capability(pdev, caps[i]);
186 if (pos == 0)
187 return false;
188 }
189
190 return true;
191 }
192
193 static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
194 {
195 struct iommu_dev_data *dev_data;
196
197 dev_data = get_dev_data(&pdev->dev);
198
199 return dev_data->errata & (1 << erratum) ? true : false;
200 }
201
202 /*
203 * In this function the list of preallocated protection domains is traversed to
204 * find the domain for a specific device
205 */
206 static struct dma_ops_domain *find_protection_domain(u16 devid)
207 {
208 struct dma_ops_domain *entry, *ret = NULL;
209 unsigned long flags;
210 u16 alias = amd_iommu_alias_table[devid];
211
212 if (list_empty(&iommu_pd_list))
213 return NULL;
214
215 spin_lock_irqsave(&iommu_pd_list_lock, flags);
216
217 list_for_each_entry(entry, &iommu_pd_list, list) {
218 if (entry->target_dev == devid ||
219 entry->target_dev == alias) {
220 ret = entry;
221 break;
222 }
223 }
224
225 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
226
227 return ret;
228 }
229
230 /*
231 * This function checks if the driver got a valid device from the caller to
232 * avoid dereferencing invalid pointers.
233 */
234 static bool check_device(struct device *dev)
235 {
236 u16 devid;
237
238 if (!dev || !dev->dma_mask)
239 return false;
240
241 /* No device or no PCI device */
242 if (dev->bus != &pci_bus_type)
243 return false;
244
245 devid = get_device_id(dev);
246
247 /* Out of our scope? */
248 if (devid > amd_iommu_last_bdf)
249 return false;
250
251 if (amd_iommu_rlookup_table[devid] == NULL)
252 return false;
253
254 return true;
255 }
256
257 static int iommu_init_device(struct device *dev)
258 {
259 struct pci_dev *pdev = to_pci_dev(dev);
260 struct iommu_dev_data *dev_data;
261 u16 alias;
262
263 if (dev->archdata.iommu)
264 return 0;
265
266 dev_data = find_dev_data(get_device_id(dev));
267 if (!dev_data)
268 return -ENOMEM;
269
270 alias = amd_iommu_alias_table[dev_data->devid];
271 if (alias != dev_data->devid) {
272 struct iommu_dev_data *alias_data;
273
274 alias_data = find_dev_data(alias);
275 if (alias_data == NULL) {
276 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
277 dev_name(dev));
278 free_dev_data(dev_data);
279 return -ENOTSUPP;
280 }
281 dev_data->alias_data = alias_data;
282 }
283
284 if (pci_iommuv2_capable(pdev)) {
285 struct amd_iommu *iommu;
286
287 iommu = amd_iommu_rlookup_table[dev_data->devid];
288 dev_data->iommu_v2 = iommu->is_iommu_v2;
289 }
290
291 dev->archdata.iommu = dev_data;
292
293 return 0;
294 }
295
296 static void iommu_ignore_device(struct device *dev)
297 {
298 u16 devid, alias;
299
300 devid = get_device_id(dev);
301 alias = amd_iommu_alias_table[devid];
302
303 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
304 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
305
306 amd_iommu_rlookup_table[devid] = NULL;
307 amd_iommu_rlookup_table[alias] = NULL;
308 }
309
310 static void iommu_uninit_device(struct device *dev)
311 {
312 /*
313 * Nothing to do here - we keep dev_data around for unplugged devices
314 * and reuse it when the device is re-plugged - not doing so would
315 * introduce a ton of races.
316 */
317 }
318
319 void __init amd_iommu_uninit_devices(void)
320 {
321 struct iommu_dev_data *dev_data, *n;
322 struct pci_dev *pdev = NULL;
323
324 for_each_pci_dev(pdev) {
325
326 if (!check_device(&pdev->dev))
327 continue;
328
329 iommu_uninit_device(&pdev->dev);
330 }
331
332 /* Free all of our dev_data structures */
333 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
334 free_dev_data(dev_data);
335 }
336
337 int __init amd_iommu_init_devices(void)
338 {
339 struct pci_dev *pdev = NULL;
340 int ret = 0;
341
342 for_each_pci_dev(pdev) {
343
344 if (!check_device(&pdev->dev))
345 continue;
346
347 ret = iommu_init_device(&pdev->dev);
348 if (ret == -ENOTSUPP)
349 iommu_ignore_device(&pdev->dev);
350 else if (ret)
351 goto out_free;
352 }
353
354 return 0;
355
356 out_free:
357
358 amd_iommu_uninit_devices();
359
360 return ret;
361 }
362 #ifdef CONFIG_AMD_IOMMU_STATS
363
364 /*
365 * Initialization code for statistics collection
366 */
367
368 DECLARE_STATS_COUNTER(compl_wait);
369 DECLARE_STATS_COUNTER(cnt_map_single);
370 DECLARE_STATS_COUNTER(cnt_unmap_single);
371 DECLARE_STATS_COUNTER(cnt_map_sg);
372 DECLARE_STATS_COUNTER(cnt_unmap_sg);
373 DECLARE_STATS_COUNTER(cnt_alloc_coherent);
374 DECLARE_STATS_COUNTER(cnt_free_coherent);
375 DECLARE_STATS_COUNTER(cross_page);
376 DECLARE_STATS_COUNTER(domain_flush_single);
377 DECLARE_STATS_COUNTER(domain_flush_all);
378 DECLARE_STATS_COUNTER(alloced_io_mem);
379 DECLARE_STATS_COUNTER(total_map_requests);
380 DECLARE_STATS_COUNTER(complete_ppr);
381 DECLARE_STATS_COUNTER(invalidate_iotlb);
382 DECLARE_STATS_COUNTER(invalidate_iotlb_all);
383 DECLARE_STATS_COUNTER(pri_requests);
384
385
386 static struct dentry *stats_dir;
387 static struct dentry *de_fflush;
388
389 static void amd_iommu_stats_add(struct __iommu_counter *cnt)
390 {
391 if (stats_dir == NULL)
392 return;
393
394 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
395 &cnt->value);
396 }
397
398 static void amd_iommu_stats_init(void)
399 {
400 stats_dir = debugfs_create_dir("amd-iommu", NULL);
401 if (stats_dir == NULL)
402 return;
403
404 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
405 (u32 *)&amd_iommu_unmap_flush);
406
407 amd_iommu_stats_add(&compl_wait);
408 amd_iommu_stats_add(&cnt_map_single);
409 amd_iommu_stats_add(&cnt_unmap_single);
410 amd_iommu_stats_add(&cnt_map_sg);
411 amd_iommu_stats_add(&cnt_unmap_sg);
412 amd_iommu_stats_add(&cnt_alloc_coherent);
413 amd_iommu_stats_add(&cnt_free_coherent);
414 amd_iommu_stats_add(&cross_page);
415 amd_iommu_stats_add(&domain_flush_single);
416 amd_iommu_stats_add(&domain_flush_all);
417 amd_iommu_stats_add(&alloced_io_mem);
418 amd_iommu_stats_add(&total_map_requests);
419 amd_iommu_stats_add(&complete_ppr);
420 amd_iommu_stats_add(&invalidate_iotlb);
421 amd_iommu_stats_add(&invalidate_iotlb_all);
422 amd_iommu_stats_add(&pri_requests);
423 }
424
425 #endif
426
427 /****************************************************************************
428 *
429 * Interrupt handling functions
430 *
431 ****************************************************************************/
432
433 static void dump_dte_entry(u16 devid)
434 {
435 int i;
436
437 for (i = 0; i < 4; ++i)
438 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
439 amd_iommu_dev_table[devid].data[i]);
440 }
441
442 static void dump_command(unsigned long phys_addr)
443 {
444 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
445 int i;
446
447 for (i = 0; i < 4; ++i)
448 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
449 }
450
451 static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
452 {
453 u32 *event = __evt;
454 int type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
455 int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
456 int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
457 int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
458 u64 address = (u64)(((u64)event[3]) << 32) | event[2];
459
460 printk(KERN_ERR "AMD-Vi: Event logged [");
461
462 switch (type) {
463 case EVENT_TYPE_ILL_DEV:
464 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
465 "address=0x%016llx flags=0x%04x]\n",
466 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
467 address, flags);
468 dump_dte_entry(devid);
469 break;
470 case EVENT_TYPE_IO_FAULT:
471 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
472 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
473 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
474 domid, address, flags);
475 break;
476 case EVENT_TYPE_DEV_TAB_ERR:
477 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
478 "address=0x%016llx flags=0x%04x]\n",
479 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
480 address, flags);
481 break;
482 case EVENT_TYPE_PAGE_TAB_ERR:
483 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
484 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
485 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
486 domid, address, flags);
487 break;
488 case EVENT_TYPE_ILL_CMD:
489 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
490 dump_command(address);
491 break;
492 case EVENT_TYPE_CMD_HARD_ERR:
493 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
494 "flags=0x%04x]\n", address, flags);
495 break;
496 case EVENT_TYPE_IOTLB_INV_TO:
497 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
498 "address=0x%016llx]\n",
499 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
500 address);
501 break;
502 case EVENT_TYPE_INV_DEV_REQ:
503 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
504 "address=0x%016llx flags=0x%04x]\n",
505 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
506 address, flags);
507 break;
508 default:
509 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
510 }
511 }
512
513 static void iommu_poll_events(struct amd_iommu *iommu)
514 {
515 u32 head, tail;
516 unsigned long flags;
517
518 spin_lock_irqsave(&iommu->lock, flags);
519
520 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
521 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
522
523 while (head != tail) {
524 iommu_print_event(iommu, iommu->evt_buf + head);
525 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
526 }
527
528 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
529
530 spin_unlock_irqrestore(&iommu->lock, flags);
531 }
532
533 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u32 head)
534 {
535 struct amd_iommu_fault fault;
536 volatile u64 *raw;
537 int i;
538
539 INC_STATS_COUNTER(pri_requests);
540
541 raw = (u64 *)(iommu->ppr_log + head);
542
543 /*
544 * Hardware bug: Interrupt may arrive before the entry is written to
545 * memory. If this happens we need to wait for the entry to arrive.
546 */
547 for (i = 0; i < LOOP_TIMEOUT; ++i) {
548 if (PPR_REQ_TYPE(raw[0]) != 0)
549 break;
550 udelay(1);
551 }
552
553 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
554 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
555 return;
556 }
557
558 fault.address = raw[1];
559 fault.pasid = PPR_PASID(raw[0]);
560 fault.device_id = PPR_DEVID(raw[0]);
561 fault.tag = PPR_TAG(raw[0]);
562 fault.flags = PPR_FLAGS(raw[0]);
563
564 /*
565 * To detect the hardware bug we need to clear the entry
566 * to back to zero.
567 */
568 raw[0] = raw[1] = 0;
569
570 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
571 }
572
573 static void iommu_poll_ppr_log(struct amd_iommu *iommu)
574 {
575 unsigned long flags;
576 u32 head, tail;
577
578 if (iommu->ppr_log == NULL)
579 return;
580
581 spin_lock_irqsave(&iommu->lock, flags);
582
583 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
584 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
585
586 while (head != tail) {
587
588 /* Handle PPR entry */
589 iommu_handle_ppr_entry(iommu, head);
590
591 /* Update and refresh ring-buffer state*/
592 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
593 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
594 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
595 }
596
597 /* enable ppr interrupts again */
598 writel(MMIO_STATUS_PPR_INT_MASK, iommu->mmio_base + MMIO_STATUS_OFFSET);
599
600 spin_unlock_irqrestore(&iommu->lock, flags);
601 }
602
603 irqreturn_t amd_iommu_int_thread(int irq, void *data)
604 {
605 struct amd_iommu *iommu;
606
607 for_each_iommu(iommu) {
608 iommu_poll_events(iommu);
609 iommu_poll_ppr_log(iommu);
610 }
611
612 return IRQ_HANDLED;
613 }
614
615 irqreturn_t amd_iommu_int_handler(int irq, void *data)
616 {
617 return IRQ_WAKE_THREAD;
618 }
619
620 /****************************************************************************
621 *
622 * IOMMU command queuing functions
623 *
624 ****************************************************************************/
625
626 static int wait_on_sem(volatile u64 *sem)
627 {
628 int i = 0;
629
630 while (*sem == 0 && i < LOOP_TIMEOUT) {
631 udelay(1);
632 i += 1;
633 }
634
635 if (i == LOOP_TIMEOUT) {
636 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
637 return -EIO;
638 }
639
640 return 0;
641 }
642
643 static void copy_cmd_to_buffer(struct amd_iommu *iommu,
644 struct iommu_cmd *cmd,
645 u32 tail)
646 {
647 u8 *target;
648
649 target = iommu->cmd_buf + tail;
650 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
651
652 /* Copy command to buffer */
653 memcpy(target, cmd, sizeof(*cmd));
654
655 /* Tell the IOMMU about it */
656 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
657 }
658
659 static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
660 {
661 WARN_ON(address & 0x7ULL);
662
663 memset(cmd, 0, sizeof(*cmd));
664 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
665 cmd->data[1] = upper_32_bits(__pa(address));
666 cmd->data[2] = 1;
667 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
668 }
669
670 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
671 {
672 memset(cmd, 0, sizeof(*cmd));
673 cmd->data[0] = devid;
674 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
675 }
676
677 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
678 size_t size, u16 domid, int pde)
679 {
680 u64 pages;
681 int s;
682
683 pages = iommu_num_pages(address, size, PAGE_SIZE);
684 s = 0;
685
686 if (pages > 1) {
687 /*
688 * If we have to flush more than one page, flush all
689 * TLB entries for this domain
690 */
691 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
692 s = 1;
693 }
694
695 address &= PAGE_MASK;
696
697 memset(cmd, 0, sizeof(*cmd));
698 cmd->data[1] |= domid;
699 cmd->data[2] = lower_32_bits(address);
700 cmd->data[3] = upper_32_bits(address);
701 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
702 if (s) /* size bit - we flush more than one 4kb page */
703 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
704 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
705 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
706 }
707
708 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
709 u64 address, size_t size)
710 {
711 u64 pages;
712 int s;
713
714 pages = iommu_num_pages(address, size, PAGE_SIZE);
715 s = 0;
716
717 if (pages > 1) {
718 /*
719 * If we have to flush more than one page, flush all
720 * TLB entries for this domain
721 */
722 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
723 s = 1;
724 }
725
726 address &= PAGE_MASK;
727
728 memset(cmd, 0, sizeof(*cmd));
729 cmd->data[0] = devid;
730 cmd->data[0] |= (qdep & 0xff) << 24;
731 cmd->data[1] = devid;
732 cmd->data[2] = lower_32_bits(address);
733 cmd->data[3] = upper_32_bits(address);
734 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
735 if (s)
736 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
737 }
738
739 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
740 u64 address, bool size)
741 {
742 memset(cmd, 0, sizeof(*cmd));
743
744 address &= ~(0xfffULL);
745
746 cmd->data[0] = pasid & PASID_MASK;
747 cmd->data[1] = domid;
748 cmd->data[2] = lower_32_bits(address);
749 cmd->data[3] = upper_32_bits(address);
750 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
751 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
752 if (size)
753 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
754 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
755 }
756
757 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
758 int qdep, u64 address, bool size)
759 {
760 memset(cmd, 0, sizeof(*cmd));
761
762 address &= ~(0xfffULL);
763
764 cmd->data[0] = devid;
765 cmd->data[0] |= (pasid & 0xff) << 16;
766 cmd->data[0] |= (qdep & 0xff) << 24;
767 cmd->data[1] = devid;
768 cmd->data[1] |= ((pasid >> 8) & 0xfff) << 16;
769 cmd->data[2] = lower_32_bits(address);
770 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
771 cmd->data[3] = upper_32_bits(address);
772 if (size)
773 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
774 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
775 }
776
777 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
778 int status, int tag, bool gn)
779 {
780 memset(cmd, 0, sizeof(*cmd));
781
782 cmd->data[0] = devid;
783 if (gn) {
784 cmd->data[1] = pasid & PASID_MASK;
785 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
786 }
787 cmd->data[3] = tag & 0x1ff;
788 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
789
790 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
791 }
792
793 static void build_inv_all(struct iommu_cmd *cmd)
794 {
795 memset(cmd, 0, sizeof(*cmd));
796 CMD_SET_TYPE(cmd, CMD_INV_ALL);
797 }
798
799 /*
800 * Writes the command to the IOMMUs command buffer and informs the
801 * hardware about the new command.
802 */
803 static int iommu_queue_command_sync(struct amd_iommu *iommu,
804 struct iommu_cmd *cmd,
805 bool sync)
806 {
807 u32 left, tail, head, next_tail;
808 unsigned long flags;
809
810 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
811
812 again:
813 spin_lock_irqsave(&iommu->lock, flags);
814
815 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
816 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
817 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
818 left = (head - next_tail) % iommu->cmd_buf_size;
819
820 if (left <= 2) {
821 struct iommu_cmd sync_cmd;
822 volatile u64 sem = 0;
823 int ret;
824
825 build_completion_wait(&sync_cmd, (u64)&sem);
826 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
827
828 spin_unlock_irqrestore(&iommu->lock, flags);
829
830 if ((ret = wait_on_sem(&sem)) != 0)
831 return ret;
832
833 goto again;
834 }
835
836 copy_cmd_to_buffer(iommu, cmd, tail);
837
838 /* We need to sync now to make sure all commands are processed */
839 iommu->need_sync = sync;
840
841 spin_unlock_irqrestore(&iommu->lock, flags);
842
843 return 0;
844 }
845
846 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
847 {
848 return iommu_queue_command_sync(iommu, cmd, true);
849 }
850
851 /*
852 * This function queues a completion wait command into the command
853 * buffer of an IOMMU
854 */
855 static int iommu_completion_wait(struct amd_iommu *iommu)
856 {
857 struct iommu_cmd cmd;
858 volatile u64 sem = 0;
859 int ret;
860
861 if (!iommu->need_sync)
862 return 0;
863
864 build_completion_wait(&cmd, (u64)&sem);
865
866 ret = iommu_queue_command_sync(iommu, &cmd, false);
867 if (ret)
868 return ret;
869
870 return wait_on_sem(&sem);
871 }
872
873 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
874 {
875 struct iommu_cmd cmd;
876
877 build_inv_dte(&cmd, devid);
878
879 return iommu_queue_command(iommu, &cmd);
880 }
881
882 static void iommu_flush_dte_all(struct amd_iommu *iommu)
883 {
884 u32 devid;
885
886 for (devid = 0; devid <= 0xffff; ++devid)
887 iommu_flush_dte(iommu, devid);
888
889 iommu_completion_wait(iommu);
890 }
891
892 /*
893 * This function uses heavy locking and may disable irqs for some time. But
894 * this is no issue because it is only called during resume.
895 */
896 static void iommu_flush_tlb_all(struct amd_iommu *iommu)
897 {
898 u32 dom_id;
899
900 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
901 struct iommu_cmd cmd;
902 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
903 dom_id, 1);
904 iommu_queue_command(iommu, &cmd);
905 }
906
907 iommu_completion_wait(iommu);
908 }
909
910 static void iommu_flush_all(struct amd_iommu *iommu)
911 {
912 struct iommu_cmd cmd;
913
914 build_inv_all(&cmd);
915
916 iommu_queue_command(iommu, &cmd);
917 iommu_completion_wait(iommu);
918 }
919
920 void iommu_flush_all_caches(struct amd_iommu *iommu)
921 {
922 if (iommu_feature(iommu, FEATURE_IA)) {
923 iommu_flush_all(iommu);
924 } else {
925 iommu_flush_dte_all(iommu);
926 iommu_flush_tlb_all(iommu);
927 }
928 }
929
930 /*
931 * Command send function for flushing on-device TLB
932 */
933 static int device_flush_iotlb(struct iommu_dev_data *dev_data,
934 u64 address, size_t size)
935 {
936 struct amd_iommu *iommu;
937 struct iommu_cmd cmd;
938 int qdep;
939
940 qdep = dev_data->ats.qdep;
941 iommu = amd_iommu_rlookup_table[dev_data->devid];
942
943 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
944
945 return iommu_queue_command(iommu, &cmd);
946 }
947
948 /*
949 * Command send function for invalidating a device table entry
950 */
951 static int device_flush_dte(struct iommu_dev_data *dev_data)
952 {
953 struct amd_iommu *iommu;
954 int ret;
955
956 iommu = amd_iommu_rlookup_table[dev_data->devid];
957
958 ret = iommu_flush_dte(iommu, dev_data->devid);
959 if (ret)
960 return ret;
961
962 if (dev_data->ats.enabled)
963 ret = device_flush_iotlb(dev_data, 0, ~0UL);
964
965 return ret;
966 }
967
968 /*
969 * TLB invalidation function which is called from the mapping functions.
970 * It invalidates a single PTE if the range to flush is within a single
971 * page. Otherwise it flushes the whole TLB of the IOMMU.
972 */
973 static void __domain_flush_pages(struct protection_domain *domain,
974 u64 address, size_t size, int pde)
975 {
976 struct iommu_dev_data *dev_data;
977 struct iommu_cmd cmd;
978 int ret = 0, i;
979
980 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
981
982 for (i = 0; i < amd_iommus_present; ++i) {
983 if (!domain->dev_iommu[i])
984 continue;
985
986 /*
987 * Devices of this domain are behind this IOMMU
988 * We need a TLB flush
989 */
990 ret |= iommu_queue_command(amd_iommus[i], &cmd);
991 }
992
993 list_for_each_entry(dev_data, &domain->dev_list, list) {
994
995 if (!dev_data->ats.enabled)
996 continue;
997
998 ret |= device_flush_iotlb(dev_data, address, size);
999 }
1000
1001 WARN_ON(ret);
1002 }
1003
1004 static void domain_flush_pages(struct protection_domain *domain,
1005 u64 address, size_t size)
1006 {
1007 __domain_flush_pages(domain, address, size, 0);
1008 }
1009
1010 /* Flush the whole IO/TLB for a given protection domain */
1011 static void domain_flush_tlb(struct protection_domain *domain)
1012 {
1013 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1014 }
1015
1016 /* Flush the whole IO/TLB for a given protection domain - including PDE */
1017 static void domain_flush_tlb_pde(struct protection_domain *domain)
1018 {
1019 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1020 }
1021
1022 static void domain_flush_complete(struct protection_domain *domain)
1023 {
1024 int i;
1025
1026 for (i = 0; i < amd_iommus_present; ++i) {
1027 if (!domain->dev_iommu[i])
1028 continue;
1029
1030 /*
1031 * Devices of this domain are behind this IOMMU
1032 * We need to wait for completion of all commands.
1033 */
1034 iommu_completion_wait(amd_iommus[i]);
1035 }
1036 }
1037
1038
1039 /*
1040 * This function flushes the DTEs for all devices in domain
1041 */
1042 static void domain_flush_devices(struct protection_domain *domain)
1043 {
1044 struct iommu_dev_data *dev_data;
1045
1046 list_for_each_entry(dev_data, &domain->dev_list, list)
1047 device_flush_dte(dev_data);
1048 }
1049
1050 /****************************************************************************
1051 *
1052 * The functions below are used the create the page table mappings for
1053 * unity mapped regions.
1054 *
1055 ****************************************************************************/
1056
1057 /*
1058 * This function is used to add another level to an IO page table. Adding
1059 * another level increases the size of the address space by 9 bits to a size up
1060 * to 64 bits.
1061 */
1062 static bool increase_address_space(struct protection_domain *domain,
1063 gfp_t gfp)
1064 {
1065 u64 *pte;
1066
1067 if (domain->mode == PAGE_MODE_6_LEVEL)
1068 /* address space already 64 bit large */
1069 return false;
1070
1071 pte = (void *)get_zeroed_page(gfp);
1072 if (!pte)
1073 return false;
1074
1075 *pte = PM_LEVEL_PDE(domain->mode,
1076 virt_to_phys(domain->pt_root));
1077 domain->pt_root = pte;
1078 domain->mode += 1;
1079 domain->updated = true;
1080
1081 return true;
1082 }
1083
1084 static u64 *alloc_pte(struct protection_domain *domain,
1085 unsigned long address,
1086 unsigned long page_size,
1087 u64 **pte_page,
1088 gfp_t gfp)
1089 {
1090 int level, end_lvl;
1091 u64 *pte, *page;
1092
1093 BUG_ON(!is_power_of_2(page_size));
1094
1095 while (address > PM_LEVEL_SIZE(domain->mode))
1096 increase_address_space(domain, gfp);
1097
1098 level = domain->mode - 1;
1099 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1100 address = PAGE_SIZE_ALIGN(address, page_size);
1101 end_lvl = PAGE_SIZE_LEVEL(page_size);
1102
1103 while (level > end_lvl) {
1104 if (!IOMMU_PTE_PRESENT(*pte)) {
1105 page = (u64 *)get_zeroed_page(gfp);
1106 if (!page)
1107 return NULL;
1108 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
1109 }
1110
1111 /* No level skipping support yet */
1112 if (PM_PTE_LEVEL(*pte) != level)
1113 return NULL;
1114
1115 level -= 1;
1116
1117 pte = IOMMU_PTE_PAGE(*pte);
1118
1119 if (pte_page && level == end_lvl)
1120 *pte_page = pte;
1121
1122 pte = &pte[PM_LEVEL_INDEX(level, address)];
1123 }
1124
1125 return pte;
1126 }
1127
1128 /*
1129 * This function checks if there is a PTE for a given dma address. If
1130 * there is one, it returns the pointer to it.
1131 */
1132 static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
1133 {
1134 int level;
1135 u64 *pte;
1136
1137 if (address > PM_LEVEL_SIZE(domain->mode))
1138 return NULL;
1139
1140 level = domain->mode - 1;
1141 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1142
1143 while (level > 0) {
1144
1145 /* Not Present */
1146 if (!IOMMU_PTE_PRESENT(*pte))
1147 return NULL;
1148
1149 /* Large PTE */
1150 if (PM_PTE_LEVEL(*pte) == 0x07) {
1151 unsigned long pte_mask, __pte;
1152
1153 /*
1154 * If we have a series of large PTEs, make
1155 * sure to return a pointer to the first one.
1156 */
1157 pte_mask = PTE_PAGE_SIZE(*pte);
1158 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1159 __pte = ((unsigned long)pte) & pte_mask;
1160
1161 return (u64 *)__pte;
1162 }
1163
1164 /* No level skipping support yet */
1165 if (PM_PTE_LEVEL(*pte) != level)
1166 return NULL;
1167
1168 level -= 1;
1169
1170 /* Walk to the next level */
1171 pte = IOMMU_PTE_PAGE(*pte);
1172 pte = &pte[PM_LEVEL_INDEX(level, address)];
1173 }
1174
1175 return pte;
1176 }
1177
1178 /*
1179 * Generic mapping functions. It maps a physical address into a DMA
1180 * address space. It allocates the page table pages if necessary.
1181 * In the future it can be extended to a generic mapping function
1182 * supporting all features of AMD IOMMU page tables like level skipping
1183 * and full 64 bit address spaces.
1184 */
1185 static int iommu_map_page(struct protection_domain *dom,
1186 unsigned long bus_addr,
1187 unsigned long phys_addr,
1188 int prot,
1189 unsigned long page_size)
1190 {
1191 u64 __pte, *pte;
1192 int i, count;
1193
1194 if (!(prot & IOMMU_PROT_MASK))
1195 return -EINVAL;
1196
1197 bus_addr = PAGE_ALIGN(bus_addr);
1198 phys_addr = PAGE_ALIGN(phys_addr);
1199 count = PAGE_SIZE_PTE_COUNT(page_size);
1200 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1201
1202 for (i = 0; i < count; ++i)
1203 if (IOMMU_PTE_PRESENT(pte[i]))
1204 return -EBUSY;
1205
1206 if (page_size > PAGE_SIZE) {
1207 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1208 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1209 } else
1210 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1211
1212 if (prot & IOMMU_PROT_IR)
1213 __pte |= IOMMU_PTE_IR;
1214 if (prot & IOMMU_PROT_IW)
1215 __pte |= IOMMU_PTE_IW;
1216
1217 for (i = 0; i < count; ++i)
1218 pte[i] = __pte;
1219
1220 update_domain(dom);
1221
1222 return 0;
1223 }
1224
1225 static unsigned long iommu_unmap_page(struct protection_domain *dom,
1226 unsigned long bus_addr,
1227 unsigned long page_size)
1228 {
1229 unsigned long long unmap_size, unmapped;
1230 u64 *pte;
1231
1232 BUG_ON(!is_power_of_2(page_size));
1233
1234 unmapped = 0;
1235
1236 while (unmapped < page_size) {
1237
1238 pte = fetch_pte(dom, bus_addr);
1239
1240 if (!pte) {
1241 /*
1242 * No PTE for this address
1243 * move forward in 4kb steps
1244 */
1245 unmap_size = PAGE_SIZE;
1246 } else if (PM_PTE_LEVEL(*pte) == 0) {
1247 /* 4kb PTE found for this address */
1248 unmap_size = PAGE_SIZE;
1249 *pte = 0ULL;
1250 } else {
1251 int count, i;
1252
1253 /* Large PTE found which maps this address */
1254 unmap_size = PTE_PAGE_SIZE(*pte);
1255 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1256 for (i = 0; i < count; i++)
1257 pte[i] = 0ULL;
1258 }
1259
1260 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1261 unmapped += unmap_size;
1262 }
1263
1264 BUG_ON(!is_power_of_2(unmapped));
1265
1266 return unmapped;
1267 }
1268
1269 /*
1270 * This function checks if a specific unity mapping entry is needed for
1271 * this specific IOMMU.
1272 */
1273 static int iommu_for_unity_map(struct amd_iommu *iommu,
1274 struct unity_map_entry *entry)
1275 {
1276 u16 bdf, i;
1277
1278 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1279 bdf = amd_iommu_alias_table[i];
1280 if (amd_iommu_rlookup_table[bdf] == iommu)
1281 return 1;
1282 }
1283
1284 return 0;
1285 }
1286
1287 /*
1288 * This function actually applies the mapping to the page table of the
1289 * dma_ops domain.
1290 */
1291 static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1292 struct unity_map_entry *e)
1293 {
1294 u64 addr;
1295 int ret;
1296
1297 for (addr = e->address_start; addr < e->address_end;
1298 addr += PAGE_SIZE) {
1299 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1300 PAGE_SIZE);
1301 if (ret)
1302 return ret;
1303 /*
1304 * if unity mapping is in aperture range mark the page
1305 * as allocated in the aperture
1306 */
1307 if (addr < dma_dom->aperture_size)
1308 __set_bit(addr >> PAGE_SHIFT,
1309 dma_dom->aperture[0]->bitmap);
1310 }
1311
1312 return 0;
1313 }
1314
1315 /*
1316 * Init the unity mappings for a specific IOMMU in the system
1317 *
1318 * Basically iterates over all unity mapping entries and applies them to
1319 * the default domain DMA of that IOMMU if necessary.
1320 */
1321 static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1322 {
1323 struct unity_map_entry *entry;
1324 int ret;
1325
1326 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1327 if (!iommu_for_unity_map(iommu, entry))
1328 continue;
1329 ret = dma_ops_unity_map(iommu->default_dom, entry);
1330 if (ret)
1331 return ret;
1332 }
1333
1334 return 0;
1335 }
1336
1337 /*
1338 * Inits the unity mappings required for a specific device
1339 */
1340 static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1341 u16 devid)
1342 {
1343 struct unity_map_entry *e;
1344 int ret;
1345
1346 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1347 if (!(devid >= e->devid_start && devid <= e->devid_end))
1348 continue;
1349 ret = dma_ops_unity_map(dma_dom, e);
1350 if (ret)
1351 return ret;
1352 }
1353
1354 return 0;
1355 }
1356
1357 /****************************************************************************
1358 *
1359 * The next functions belong to the address allocator for the dma_ops
1360 * interface functions. They work like the allocators in the other IOMMU
1361 * drivers. Its basically a bitmap which marks the allocated pages in
1362 * the aperture. Maybe it could be enhanced in the future to a more
1363 * efficient allocator.
1364 *
1365 ****************************************************************************/
1366
1367 /*
1368 * The address allocator core functions.
1369 *
1370 * called with domain->lock held
1371 */
1372
1373 /*
1374 * Used to reserve address ranges in the aperture (e.g. for exclusion
1375 * ranges.
1376 */
1377 static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1378 unsigned long start_page,
1379 unsigned int pages)
1380 {
1381 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1382
1383 if (start_page + pages > last_page)
1384 pages = last_page - start_page;
1385
1386 for (i = start_page; i < start_page + pages; ++i) {
1387 int index = i / APERTURE_RANGE_PAGES;
1388 int page = i % APERTURE_RANGE_PAGES;
1389 __set_bit(page, dom->aperture[index]->bitmap);
1390 }
1391 }
1392
1393 /*
1394 * This function is used to add a new aperture range to an existing
1395 * aperture in case of dma_ops domain allocation or address allocation
1396 * failure.
1397 */
1398 static int alloc_new_range(struct dma_ops_domain *dma_dom,
1399 bool populate, gfp_t gfp)
1400 {
1401 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1402 struct amd_iommu *iommu;
1403 unsigned long i, old_size;
1404
1405 #ifdef CONFIG_IOMMU_STRESS
1406 populate = false;
1407 #endif
1408
1409 if (index >= APERTURE_MAX_RANGES)
1410 return -ENOMEM;
1411
1412 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1413 if (!dma_dom->aperture[index])
1414 return -ENOMEM;
1415
1416 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1417 if (!dma_dom->aperture[index]->bitmap)
1418 goto out_free;
1419
1420 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1421
1422 if (populate) {
1423 unsigned long address = dma_dom->aperture_size;
1424 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1425 u64 *pte, *pte_page;
1426
1427 for (i = 0; i < num_ptes; ++i) {
1428 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1429 &pte_page, gfp);
1430 if (!pte)
1431 goto out_free;
1432
1433 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1434
1435 address += APERTURE_RANGE_SIZE / 64;
1436 }
1437 }
1438
1439 old_size = dma_dom->aperture_size;
1440 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1441
1442 /* Reserve address range used for MSI messages */
1443 if (old_size < MSI_ADDR_BASE_LO &&
1444 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1445 unsigned long spage;
1446 int pages;
1447
1448 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1449 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1450
1451 dma_ops_reserve_addresses(dma_dom, spage, pages);
1452 }
1453
1454 /* Initialize the exclusion range if necessary */
1455 for_each_iommu(iommu) {
1456 if (iommu->exclusion_start &&
1457 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1458 && iommu->exclusion_start < dma_dom->aperture_size) {
1459 unsigned long startpage;
1460 int pages = iommu_num_pages(iommu->exclusion_start,
1461 iommu->exclusion_length,
1462 PAGE_SIZE);
1463 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1464 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1465 }
1466 }
1467
1468 /*
1469 * Check for areas already mapped as present in the new aperture
1470 * range and mark those pages as reserved in the allocator. Such
1471 * mappings may already exist as a result of requested unity
1472 * mappings for devices.
1473 */
1474 for (i = dma_dom->aperture[index]->offset;
1475 i < dma_dom->aperture_size;
1476 i += PAGE_SIZE) {
1477 u64 *pte = fetch_pte(&dma_dom->domain, i);
1478 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1479 continue;
1480
1481 dma_ops_reserve_addresses(dma_dom, i >> PAGE_SHIFT, 1);
1482 }
1483
1484 update_domain(&dma_dom->domain);
1485
1486 return 0;
1487
1488 out_free:
1489 update_domain(&dma_dom->domain);
1490
1491 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1492
1493 kfree(dma_dom->aperture[index]);
1494 dma_dom->aperture[index] = NULL;
1495
1496 return -ENOMEM;
1497 }
1498
1499 static unsigned long dma_ops_area_alloc(struct device *dev,
1500 struct dma_ops_domain *dom,
1501 unsigned int pages,
1502 unsigned long align_mask,
1503 u64 dma_mask,
1504 unsigned long start)
1505 {
1506 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1507 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1508 int i = start >> APERTURE_RANGE_SHIFT;
1509 unsigned long boundary_size;
1510 unsigned long address = -1;
1511 unsigned long limit;
1512
1513 next_bit >>= PAGE_SHIFT;
1514
1515 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1516 PAGE_SIZE) >> PAGE_SHIFT;
1517
1518 for (;i < max_index; ++i) {
1519 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1520
1521 if (dom->aperture[i]->offset >= dma_mask)
1522 break;
1523
1524 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1525 dma_mask >> PAGE_SHIFT);
1526
1527 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1528 limit, next_bit, pages, 0,
1529 boundary_size, align_mask);
1530 if (address != -1) {
1531 address = dom->aperture[i]->offset +
1532 (address << PAGE_SHIFT);
1533 dom->next_address = address + (pages << PAGE_SHIFT);
1534 break;
1535 }
1536
1537 next_bit = 0;
1538 }
1539
1540 return address;
1541 }
1542
1543 static unsigned long dma_ops_alloc_addresses(struct device *dev,
1544 struct dma_ops_domain *dom,
1545 unsigned int pages,
1546 unsigned long align_mask,
1547 u64 dma_mask)
1548 {
1549 unsigned long address;
1550
1551 #ifdef CONFIG_IOMMU_STRESS
1552 dom->next_address = 0;
1553 dom->need_flush = true;
1554 #endif
1555
1556 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1557 dma_mask, dom->next_address);
1558
1559 if (address == -1) {
1560 dom->next_address = 0;
1561 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1562 dma_mask, 0);
1563 dom->need_flush = true;
1564 }
1565
1566 if (unlikely(address == -1))
1567 address = DMA_ERROR_CODE;
1568
1569 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1570
1571 return address;
1572 }
1573
1574 /*
1575 * The address free function.
1576 *
1577 * called with domain->lock held
1578 */
1579 static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1580 unsigned long address,
1581 unsigned int pages)
1582 {
1583 unsigned i = address >> APERTURE_RANGE_SHIFT;
1584 struct aperture_range *range = dom->aperture[i];
1585
1586 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1587
1588 #ifdef CONFIG_IOMMU_STRESS
1589 if (i < 4)
1590 return;
1591 #endif
1592
1593 if (address >= dom->next_address)
1594 dom->need_flush = true;
1595
1596 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1597
1598 bitmap_clear(range->bitmap, address, pages);
1599
1600 }
1601
1602 /****************************************************************************
1603 *
1604 * The next functions belong to the domain allocation. A domain is
1605 * allocated for every IOMMU as the default domain. If device isolation
1606 * is enabled, every device get its own domain. The most important thing
1607 * about domains is the page table mapping the DMA address space they
1608 * contain.
1609 *
1610 ****************************************************************************/
1611
1612 /*
1613 * This function adds a protection domain to the global protection domain list
1614 */
1615 static void add_domain_to_list(struct protection_domain *domain)
1616 {
1617 unsigned long flags;
1618
1619 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1620 list_add(&domain->list, &amd_iommu_pd_list);
1621 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1622 }
1623
1624 /*
1625 * This function removes a protection domain to the global
1626 * protection domain list
1627 */
1628 static void del_domain_from_list(struct protection_domain *domain)
1629 {
1630 unsigned long flags;
1631
1632 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1633 list_del(&domain->list);
1634 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1635 }
1636
1637 static u16 domain_id_alloc(void)
1638 {
1639 unsigned long flags;
1640 int id;
1641
1642 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1643 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1644 BUG_ON(id == 0);
1645 if (id > 0 && id < MAX_DOMAIN_ID)
1646 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1647 else
1648 id = 0;
1649 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1650
1651 return id;
1652 }
1653
1654 static void domain_id_free(int id)
1655 {
1656 unsigned long flags;
1657
1658 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1659 if (id > 0 && id < MAX_DOMAIN_ID)
1660 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1661 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1662 }
1663
1664 static void free_pagetable(struct protection_domain *domain)
1665 {
1666 int i, j;
1667 u64 *p1, *p2, *p3;
1668
1669 p1 = domain->pt_root;
1670
1671 if (!p1)
1672 return;
1673
1674 for (i = 0; i < 512; ++i) {
1675 if (!IOMMU_PTE_PRESENT(p1[i]))
1676 continue;
1677
1678 p2 = IOMMU_PTE_PAGE(p1[i]);
1679 for (j = 0; j < 512; ++j) {
1680 if (!IOMMU_PTE_PRESENT(p2[j]))
1681 continue;
1682 p3 = IOMMU_PTE_PAGE(p2[j]);
1683 free_page((unsigned long)p3);
1684 }
1685
1686 free_page((unsigned long)p2);
1687 }
1688
1689 free_page((unsigned long)p1);
1690
1691 domain->pt_root = NULL;
1692 }
1693
1694 static void free_gcr3_tbl_level1(u64 *tbl)
1695 {
1696 u64 *ptr;
1697 int i;
1698
1699 for (i = 0; i < 512; ++i) {
1700 if (!(tbl[i] & GCR3_VALID))
1701 continue;
1702
1703 ptr = __va(tbl[i] & PAGE_MASK);
1704
1705 free_page((unsigned long)ptr);
1706 }
1707 }
1708
1709 static void free_gcr3_tbl_level2(u64 *tbl)
1710 {
1711 u64 *ptr;
1712 int i;
1713
1714 for (i = 0; i < 512; ++i) {
1715 if (!(tbl[i] & GCR3_VALID))
1716 continue;
1717
1718 ptr = __va(tbl[i] & PAGE_MASK);
1719
1720 free_gcr3_tbl_level1(ptr);
1721 }
1722 }
1723
1724 static void free_gcr3_table(struct protection_domain *domain)
1725 {
1726 if (domain->glx == 2)
1727 free_gcr3_tbl_level2(domain->gcr3_tbl);
1728 else if (domain->glx == 1)
1729 free_gcr3_tbl_level1(domain->gcr3_tbl);
1730 else if (domain->glx != 0)
1731 BUG();
1732
1733 free_page((unsigned long)domain->gcr3_tbl);
1734 }
1735
1736 /*
1737 * Free a domain, only used if something went wrong in the
1738 * allocation path and we need to free an already allocated page table
1739 */
1740 static void dma_ops_domain_free(struct dma_ops_domain *dom)
1741 {
1742 int i;
1743
1744 if (!dom)
1745 return;
1746
1747 del_domain_from_list(&dom->domain);
1748
1749 free_pagetable(&dom->domain);
1750
1751 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1752 if (!dom->aperture[i])
1753 continue;
1754 free_page((unsigned long)dom->aperture[i]->bitmap);
1755 kfree(dom->aperture[i]);
1756 }
1757
1758 kfree(dom);
1759 }
1760
1761 /*
1762 * Allocates a new protection domain usable for the dma_ops functions.
1763 * It also initializes the page table and the address allocator data
1764 * structures required for the dma_ops interface
1765 */
1766 static struct dma_ops_domain *dma_ops_domain_alloc(void)
1767 {
1768 struct dma_ops_domain *dma_dom;
1769
1770 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1771 if (!dma_dom)
1772 return NULL;
1773
1774 spin_lock_init(&dma_dom->domain.lock);
1775
1776 dma_dom->domain.id = domain_id_alloc();
1777 if (dma_dom->domain.id == 0)
1778 goto free_dma_dom;
1779 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1780 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1781 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1782 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1783 dma_dom->domain.priv = dma_dom;
1784 if (!dma_dom->domain.pt_root)
1785 goto free_dma_dom;
1786
1787 dma_dom->need_flush = false;
1788 dma_dom->target_dev = 0xffff;
1789
1790 add_domain_to_list(&dma_dom->domain);
1791
1792 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1793 goto free_dma_dom;
1794
1795 /*
1796 * mark the first page as allocated so we never return 0 as
1797 * a valid dma-address. So we can use 0 as error value
1798 */
1799 dma_dom->aperture[0]->bitmap[0] = 1;
1800 dma_dom->next_address = 0;
1801
1802
1803 return dma_dom;
1804
1805 free_dma_dom:
1806 dma_ops_domain_free(dma_dom);
1807
1808 return NULL;
1809 }
1810
1811 /*
1812 * little helper function to check whether a given protection domain is a
1813 * dma_ops domain
1814 */
1815 static bool dma_ops_domain(struct protection_domain *domain)
1816 {
1817 return domain->flags & PD_DMA_OPS_MASK;
1818 }
1819
1820 static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1821 {
1822 u64 pte_root = 0;
1823 u64 flags = 0;
1824
1825 if (domain->mode != PAGE_MODE_NONE)
1826 pte_root = virt_to_phys(domain->pt_root);
1827
1828 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1829 << DEV_ENTRY_MODE_SHIFT;
1830 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1831
1832 flags = amd_iommu_dev_table[devid].data[1];
1833
1834 if (ats)
1835 flags |= DTE_FLAG_IOTLB;
1836
1837 if (domain->flags & PD_IOMMUV2_MASK) {
1838 u64 gcr3 = __pa(domain->gcr3_tbl);
1839 u64 glx = domain->glx;
1840 u64 tmp;
1841
1842 pte_root |= DTE_FLAG_GV;
1843 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1844
1845 /* First mask out possible old values for GCR3 table */
1846 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1847 flags &= ~tmp;
1848
1849 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1850 flags &= ~tmp;
1851
1852 /* Encode GCR3 table into DTE */
1853 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1854 pte_root |= tmp;
1855
1856 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1857 flags |= tmp;
1858
1859 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1860 flags |= tmp;
1861 }
1862
1863 flags &= ~(0xffffUL);
1864 flags |= domain->id;
1865
1866 amd_iommu_dev_table[devid].data[1] = flags;
1867 amd_iommu_dev_table[devid].data[0] = pte_root;
1868 }
1869
1870 static void clear_dte_entry(u16 devid)
1871 {
1872 /* remove entry from the device table seen by the hardware */
1873 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1874 amd_iommu_dev_table[devid].data[1] = 0;
1875
1876 amd_iommu_apply_erratum_63(devid);
1877 }
1878
1879 static void do_attach(struct iommu_dev_data *dev_data,
1880 struct protection_domain *domain)
1881 {
1882 struct amd_iommu *iommu;
1883 bool ats;
1884
1885 iommu = amd_iommu_rlookup_table[dev_data->devid];
1886 ats = dev_data->ats.enabled;
1887
1888 /* Update data structures */
1889 dev_data->domain = domain;
1890 list_add(&dev_data->list, &domain->dev_list);
1891 set_dte_entry(dev_data->devid, domain, ats);
1892
1893 /* Do reference counting */
1894 domain->dev_iommu[iommu->index] += 1;
1895 domain->dev_cnt += 1;
1896
1897 /* Flush the DTE entry */
1898 device_flush_dte(dev_data);
1899 }
1900
1901 static void do_detach(struct iommu_dev_data *dev_data)
1902 {
1903 struct amd_iommu *iommu;
1904
1905 iommu = amd_iommu_rlookup_table[dev_data->devid];
1906
1907 /* decrease reference counters */
1908 dev_data->domain->dev_iommu[iommu->index] -= 1;
1909 dev_data->domain->dev_cnt -= 1;
1910
1911 /* Update data structures */
1912 dev_data->domain = NULL;
1913 list_del(&dev_data->list);
1914 clear_dte_entry(dev_data->devid);
1915
1916 /* Flush the DTE entry */
1917 device_flush_dte(dev_data);
1918 }
1919
1920 /*
1921 * If a device is not yet associated with a domain, this function does
1922 * assigns it visible for the hardware
1923 */
1924 static int __attach_device(struct iommu_dev_data *dev_data,
1925 struct protection_domain *domain)
1926 {
1927 int ret;
1928
1929 /* lock domain */
1930 spin_lock(&domain->lock);
1931
1932 if (dev_data->alias_data != NULL) {
1933 struct iommu_dev_data *alias_data = dev_data->alias_data;
1934
1935 /* Some sanity checks */
1936 ret = -EBUSY;
1937 if (alias_data->domain != NULL &&
1938 alias_data->domain != domain)
1939 goto out_unlock;
1940
1941 if (dev_data->domain != NULL &&
1942 dev_data->domain != domain)
1943 goto out_unlock;
1944
1945 /* Do real assignment */
1946 if (alias_data->domain == NULL)
1947 do_attach(alias_data, domain);
1948
1949 atomic_inc(&alias_data->bind);
1950 }
1951
1952 if (dev_data->domain == NULL)
1953 do_attach(dev_data, domain);
1954
1955 atomic_inc(&dev_data->bind);
1956
1957 ret = 0;
1958
1959 out_unlock:
1960
1961 /* ready */
1962 spin_unlock(&domain->lock);
1963
1964 return ret;
1965 }
1966
1967
1968 static void pdev_iommuv2_disable(struct pci_dev *pdev)
1969 {
1970 pci_disable_ats(pdev);
1971 pci_disable_pri(pdev);
1972 pci_disable_pasid(pdev);
1973 }
1974
1975 /* FIXME: Change generic reset-function to do the same */
1976 static int pri_reset_while_enabled(struct pci_dev *pdev)
1977 {
1978 u16 control;
1979 int pos;
1980
1981 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1982 if (!pos)
1983 return -EINVAL;
1984
1985 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1986 control |= PCI_PRI_CTRL_RESET;
1987 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1988
1989 return 0;
1990 }
1991
1992 static int pdev_iommuv2_enable(struct pci_dev *pdev)
1993 {
1994 bool reset_enable;
1995 int reqs, ret;
1996
1997 /* FIXME: Hardcode number of outstanding requests for now */
1998 reqs = 32;
1999 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
2000 reqs = 1;
2001 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
2002
2003 /* Only allow access to user-accessible pages */
2004 ret = pci_enable_pasid(pdev, 0);
2005 if (ret)
2006 goto out_err;
2007
2008 /* First reset the PRI state of the device */
2009 ret = pci_reset_pri(pdev);
2010 if (ret)
2011 goto out_err;
2012
2013 /* Enable PRI */
2014 ret = pci_enable_pri(pdev, reqs);
2015 if (ret)
2016 goto out_err;
2017
2018 if (reset_enable) {
2019 ret = pri_reset_while_enabled(pdev);
2020 if (ret)
2021 goto out_err;
2022 }
2023
2024 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2025 if (ret)
2026 goto out_err;
2027
2028 return 0;
2029
2030 out_err:
2031 pci_disable_pri(pdev);
2032 pci_disable_pasid(pdev);
2033
2034 return ret;
2035 }
2036
2037 /* FIXME: Move this to PCI code */
2038 #define PCI_PRI_TLP_OFF (1 << 2)
2039
2040 bool pci_pri_tlp_required(struct pci_dev *pdev)
2041 {
2042 u16 control;
2043 int pos;
2044
2045 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2046 if (!pos)
2047 return false;
2048
2049 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
2050
2051 return (control & PCI_PRI_TLP_OFF) ? true : false;
2052 }
2053
2054 /*
2055 * If a device is not yet associated with a domain, this function does
2056 * assigns it visible for the hardware
2057 */
2058 static int attach_device(struct device *dev,
2059 struct protection_domain *domain)
2060 {
2061 struct pci_dev *pdev = to_pci_dev(dev);
2062 struct iommu_dev_data *dev_data;
2063 unsigned long flags;
2064 int ret;
2065
2066 dev_data = get_dev_data(dev);
2067
2068 if (domain->flags & PD_IOMMUV2_MASK) {
2069 if (!dev_data->iommu_v2 || !dev_data->passthrough)
2070 return -EINVAL;
2071
2072 if (pdev_iommuv2_enable(pdev) != 0)
2073 return -EINVAL;
2074
2075 dev_data->ats.enabled = true;
2076 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2077 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2078 } else if (amd_iommu_iotlb_sup &&
2079 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2080 dev_data->ats.enabled = true;
2081 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2082 }
2083
2084 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2085 ret = __attach_device(dev_data, domain);
2086 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2087
2088 /*
2089 * We might boot into a crash-kernel here. The crashed kernel
2090 * left the caches in the IOMMU dirty. So we have to flush
2091 * here to evict all dirty stuff.
2092 */
2093 domain_flush_tlb_pde(domain);
2094
2095 return ret;
2096 }
2097
2098 /*
2099 * Removes a device from a protection domain (unlocked)
2100 */
2101 static void __detach_device(struct iommu_dev_data *dev_data)
2102 {
2103 struct protection_domain *domain;
2104 unsigned long flags;
2105
2106 BUG_ON(!dev_data->domain);
2107
2108 domain = dev_data->domain;
2109
2110 spin_lock_irqsave(&domain->lock, flags);
2111
2112 if (dev_data->alias_data != NULL) {
2113 struct iommu_dev_data *alias_data = dev_data->alias_data;
2114
2115 if (atomic_dec_and_test(&alias_data->bind))
2116 do_detach(alias_data);
2117 }
2118
2119 if (atomic_dec_and_test(&dev_data->bind))
2120 do_detach(dev_data);
2121
2122 spin_unlock_irqrestore(&domain->lock, flags);
2123
2124 /*
2125 * If we run in passthrough mode the device must be assigned to the
2126 * passthrough domain if it is detached from any other domain.
2127 * Make sure we can deassign from the pt_domain itself.
2128 */
2129 if (dev_data->passthrough &&
2130 (dev_data->domain == NULL && domain != pt_domain))
2131 __attach_device(dev_data, pt_domain);
2132 }
2133
2134 /*
2135 * Removes a device from a protection domain (with devtable_lock held)
2136 */
2137 static void detach_device(struct device *dev)
2138 {
2139 struct protection_domain *domain;
2140 struct iommu_dev_data *dev_data;
2141 unsigned long flags;
2142
2143 dev_data = get_dev_data(dev);
2144 domain = dev_data->domain;
2145
2146 /* lock device table */
2147 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2148 __detach_device(dev_data);
2149 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2150
2151 if (domain->flags & PD_IOMMUV2_MASK)
2152 pdev_iommuv2_disable(to_pci_dev(dev));
2153 else if (dev_data->ats.enabled)
2154 pci_disable_ats(to_pci_dev(dev));
2155
2156 dev_data->ats.enabled = false;
2157 }
2158
2159 /*
2160 * Find out the protection domain structure for a given PCI device. This
2161 * will give us the pointer to the page table root for example.
2162 */
2163 static struct protection_domain *domain_for_device(struct device *dev)
2164 {
2165 struct iommu_dev_data *dev_data;
2166 struct protection_domain *dom = NULL;
2167 unsigned long flags;
2168
2169 dev_data = get_dev_data(dev);
2170
2171 if (dev_data->domain)
2172 return dev_data->domain;
2173
2174 if (dev_data->alias_data != NULL) {
2175 struct iommu_dev_data *alias_data = dev_data->alias_data;
2176
2177 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
2178 if (alias_data->domain != NULL) {
2179 __attach_device(dev_data, alias_data->domain);
2180 dom = alias_data->domain;
2181 }
2182 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2183 }
2184
2185 return dom;
2186 }
2187
2188 static int device_change_notifier(struct notifier_block *nb,
2189 unsigned long action, void *data)
2190 {
2191 struct dma_ops_domain *dma_domain;
2192 struct protection_domain *domain;
2193 struct iommu_dev_data *dev_data;
2194 struct device *dev = data;
2195 struct amd_iommu *iommu;
2196 unsigned long flags;
2197 u16 devid;
2198
2199 if (!check_device(dev))
2200 return 0;
2201
2202 devid = get_device_id(dev);
2203 iommu = amd_iommu_rlookup_table[devid];
2204 dev_data = get_dev_data(dev);
2205
2206 switch (action) {
2207 case BUS_NOTIFY_UNBOUND_DRIVER:
2208
2209 domain = domain_for_device(dev);
2210
2211 if (!domain)
2212 goto out;
2213 if (dev_data->passthrough)
2214 break;
2215 detach_device(dev);
2216 break;
2217 case BUS_NOTIFY_ADD_DEVICE:
2218
2219 iommu_init_device(dev);
2220
2221 domain = domain_for_device(dev);
2222
2223 /* allocate a protection domain if a device is added */
2224 dma_domain = find_protection_domain(devid);
2225 if (dma_domain)
2226 goto out;
2227 dma_domain = dma_ops_domain_alloc();
2228 if (!dma_domain)
2229 goto out;
2230 dma_domain->target_dev = devid;
2231
2232 spin_lock_irqsave(&iommu_pd_list_lock, flags);
2233 list_add_tail(&dma_domain->list, &iommu_pd_list);
2234 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
2235
2236 break;
2237 case BUS_NOTIFY_DEL_DEVICE:
2238
2239 iommu_uninit_device(dev);
2240
2241 default:
2242 goto out;
2243 }
2244
2245 iommu_completion_wait(iommu);
2246
2247 out:
2248 return 0;
2249 }
2250
2251 static struct notifier_block device_nb = {
2252 .notifier_call = device_change_notifier,
2253 };
2254
2255 void amd_iommu_init_notifier(void)
2256 {
2257 bus_register_notifier(&pci_bus_type, &device_nb);
2258 }
2259
2260 /*****************************************************************************
2261 *
2262 * The next functions belong to the dma_ops mapping/unmapping code.
2263 *
2264 *****************************************************************************/
2265
2266 /*
2267 * In the dma_ops path we only have the struct device. This function
2268 * finds the corresponding IOMMU, the protection domain and the
2269 * requestor id for a given device.
2270 * If the device is not yet associated with a domain this is also done
2271 * in this function.
2272 */
2273 static struct protection_domain *get_domain(struct device *dev)
2274 {
2275 struct protection_domain *domain;
2276 struct dma_ops_domain *dma_dom;
2277 u16 devid = get_device_id(dev);
2278
2279 if (!check_device(dev))
2280 return ERR_PTR(-EINVAL);
2281
2282 domain = domain_for_device(dev);
2283 if (domain != NULL && !dma_ops_domain(domain))
2284 return ERR_PTR(-EBUSY);
2285
2286 if (domain != NULL)
2287 return domain;
2288
2289 /* Device not bount yet - bind it */
2290 dma_dom = find_protection_domain(devid);
2291 if (!dma_dom)
2292 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
2293 attach_device(dev, &dma_dom->domain);
2294 DUMP_printk("Using protection domain %d for device %s\n",
2295 dma_dom->domain.id, dev_name(dev));
2296
2297 return &dma_dom->domain;
2298 }
2299
2300 static void update_device_table(struct protection_domain *domain)
2301 {
2302 struct iommu_dev_data *dev_data;
2303
2304 list_for_each_entry(dev_data, &domain->dev_list, list)
2305 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2306 }
2307
2308 static void update_domain(struct protection_domain *domain)
2309 {
2310 if (!domain->updated)
2311 return;
2312
2313 update_device_table(domain);
2314
2315 domain_flush_devices(domain);
2316 domain_flush_tlb_pde(domain);
2317
2318 domain->updated = false;
2319 }
2320
2321 /*
2322 * This function fetches the PTE for a given address in the aperture
2323 */
2324 static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
2325 unsigned long address)
2326 {
2327 struct aperture_range *aperture;
2328 u64 *pte, *pte_page;
2329
2330 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2331 if (!aperture)
2332 return NULL;
2333
2334 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2335 if (!pte) {
2336 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
2337 GFP_ATOMIC);
2338 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
2339 } else
2340 pte += PM_LEVEL_INDEX(0, address);
2341
2342 update_domain(&dom->domain);
2343
2344 return pte;
2345 }
2346
2347 /*
2348 * This is the generic map function. It maps one 4kb page at paddr to
2349 * the given address in the DMA address space for the domain.
2350 */
2351 static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
2352 unsigned long address,
2353 phys_addr_t paddr,
2354 int direction)
2355 {
2356 u64 *pte, __pte;
2357
2358 WARN_ON(address > dom->aperture_size);
2359
2360 paddr &= PAGE_MASK;
2361
2362 pte = dma_ops_get_pte(dom, address);
2363 if (!pte)
2364 return DMA_ERROR_CODE;
2365
2366 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
2367
2368 if (direction == DMA_TO_DEVICE)
2369 __pte |= IOMMU_PTE_IR;
2370 else if (direction == DMA_FROM_DEVICE)
2371 __pte |= IOMMU_PTE_IW;
2372 else if (direction == DMA_BIDIRECTIONAL)
2373 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2374
2375 WARN_ON(*pte);
2376
2377 *pte = __pte;
2378
2379 return (dma_addr_t)address;
2380 }
2381
2382 /*
2383 * The generic unmapping function for on page in the DMA address space.
2384 */
2385 static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2386 unsigned long address)
2387 {
2388 struct aperture_range *aperture;
2389 u64 *pte;
2390
2391 if (address >= dom->aperture_size)
2392 return;
2393
2394 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2395 if (!aperture)
2396 return;
2397
2398 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2399 if (!pte)
2400 return;
2401
2402 pte += PM_LEVEL_INDEX(0, address);
2403
2404 WARN_ON(!*pte);
2405
2406 *pte = 0ULL;
2407 }
2408
2409 /*
2410 * This function contains common code for mapping of a physically
2411 * contiguous memory region into DMA address space. It is used by all
2412 * mapping functions provided with this IOMMU driver.
2413 * Must be called with the domain lock held.
2414 */
2415 static dma_addr_t __map_single(struct device *dev,
2416 struct dma_ops_domain *dma_dom,
2417 phys_addr_t paddr,
2418 size_t size,
2419 int dir,
2420 bool align,
2421 u64 dma_mask)
2422 {
2423 dma_addr_t offset = paddr & ~PAGE_MASK;
2424 dma_addr_t address, start, ret;
2425 unsigned int pages;
2426 unsigned long align_mask = 0;
2427 int i;
2428
2429 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2430 paddr &= PAGE_MASK;
2431
2432 INC_STATS_COUNTER(total_map_requests);
2433
2434 if (pages > 1)
2435 INC_STATS_COUNTER(cross_page);
2436
2437 if (align)
2438 align_mask = (1UL << get_order(size)) - 1;
2439
2440 retry:
2441 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2442 dma_mask);
2443 if (unlikely(address == DMA_ERROR_CODE)) {
2444 /*
2445 * setting next_address here will let the address
2446 * allocator only scan the new allocated range in the
2447 * first run. This is a small optimization.
2448 */
2449 dma_dom->next_address = dma_dom->aperture_size;
2450
2451 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2452 goto out;
2453
2454 /*
2455 * aperture was successfully enlarged by 128 MB, try
2456 * allocation again
2457 */
2458 goto retry;
2459 }
2460
2461 start = address;
2462 for (i = 0; i < pages; ++i) {
2463 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2464 if (ret == DMA_ERROR_CODE)
2465 goto out_unmap;
2466
2467 paddr += PAGE_SIZE;
2468 start += PAGE_SIZE;
2469 }
2470 address += offset;
2471
2472 ADD_STATS_COUNTER(alloced_io_mem, size);
2473
2474 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2475 domain_flush_tlb(&dma_dom->domain);
2476 dma_dom->need_flush = false;
2477 } else if (unlikely(amd_iommu_np_cache))
2478 domain_flush_pages(&dma_dom->domain, address, size);
2479
2480 out:
2481 return address;
2482
2483 out_unmap:
2484
2485 for (--i; i >= 0; --i) {
2486 start -= PAGE_SIZE;
2487 dma_ops_domain_unmap(dma_dom, start);
2488 }
2489
2490 dma_ops_free_addresses(dma_dom, address, pages);
2491
2492 return DMA_ERROR_CODE;
2493 }
2494
2495 /*
2496 * Does the reverse of the __map_single function. Must be called with
2497 * the domain lock held too
2498 */
2499 static void __unmap_single(struct dma_ops_domain *dma_dom,
2500 dma_addr_t dma_addr,
2501 size_t size,
2502 int dir)
2503 {
2504 dma_addr_t flush_addr;
2505 dma_addr_t i, start;
2506 unsigned int pages;
2507
2508 if ((dma_addr == DMA_ERROR_CODE) ||
2509 (dma_addr + size > dma_dom->aperture_size))
2510 return;
2511
2512 flush_addr = dma_addr;
2513 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2514 dma_addr &= PAGE_MASK;
2515 start = dma_addr;
2516
2517 for (i = 0; i < pages; ++i) {
2518 dma_ops_domain_unmap(dma_dom, start);
2519 start += PAGE_SIZE;
2520 }
2521
2522 SUB_STATS_COUNTER(alloced_io_mem, size);
2523
2524 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2525
2526 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2527 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2528 dma_dom->need_flush = false;
2529 }
2530 }
2531
2532 /*
2533 * The exported map_single function for dma_ops.
2534 */
2535 static dma_addr_t map_page(struct device *dev, struct page *page,
2536 unsigned long offset, size_t size,
2537 enum dma_data_direction dir,
2538 struct dma_attrs *attrs)
2539 {
2540 unsigned long flags;
2541 struct protection_domain *domain;
2542 dma_addr_t addr;
2543 u64 dma_mask;
2544 phys_addr_t paddr = page_to_phys(page) + offset;
2545
2546 INC_STATS_COUNTER(cnt_map_single);
2547
2548 domain = get_domain(dev);
2549 if (PTR_ERR(domain) == -EINVAL)
2550 return (dma_addr_t)paddr;
2551 else if (IS_ERR(domain))
2552 return DMA_ERROR_CODE;
2553
2554 dma_mask = *dev->dma_mask;
2555
2556 spin_lock_irqsave(&domain->lock, flags);
2557
2558 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2559 dma_mask);
2560 if (addr == DMA_ERROR_CODE)
2561 goto out;
2562
2563 domain_flush_complete(domain);
2564
2565 out:
2566 spin_unlock_irqrestore(&domain->lock, flags);
2567
2568 return addr;
2569 }
2570
2571 /*
2572 * The exported unmap_single function for dma_ops.
2573 */
2574 static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2575 enum dma_data_direction dir, struct dma_attrs *attrs)
2576 {
2577 unsigned long flags;
2578 struct protection_domain *domain;
2579
2580 INC_STATS_COUNTER(cnt_unmap_single);
2581
2582 domain = get_domain(dev);
2583 if (IS_ERR(domain))
2584 return;
2585
2586 spin_lock_irqsave(&domain->lock, flags);
2587
2588 __unmap_single(domain->priv, dma_addr, size, dir);
2589
2590 domain_flush_complete(domain);
2591
2592 spin_unlock_irqrestore(&domain->lock, flags);
2593 }
2594
2595 /*
2596 * This is a special map_sg function which is used if we should map a
2597 * device which is not handled by an AMD IOMMU in the system.
2598 */
2599 static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2600 int nelems, int dir)
2601 {
2602 struct scatterlist *s;
2603 int i;
2604
2605 for_each_sg(sglist, s, nelems, i) {
2606 s->dma_address = (dma_addr_t)sg_phys(s);
2607 s->dma_length = s->length;
2608 }
2609
2610 return nelems;
2611 }
2612
2613 /*
2614 * The exported map_sg function for dma_ops (handles scatter-gather
2615 * lists).
2616 */
2617 static int map_sg(struct device *dev, struct scatterlist *sglist,
2618 int nelems, enum dma_data_direction dir,
2619 struct dma_attrs *attrs)
2620 {
2621 unsigned long flags;
2622 struct protection_domain *domain;
2623 int i;
2624 struct scatterlist *s;
2625 phys_addr_t paddr;
2626 int mapped_elems = 0;
2627 u64 dma_mask;
2628
2629 INC_STATS_COUNTER(cnt_map_sg);
2630
2631 domain = get_domain(dev);
2632 if (PTR_ERR(domain) == -EINVAL)
2633 return map_sg_no_iommu(dev, sglist, nelems, dir);
2634 else if (IS_ERR(domain))
2635 return 0;
2636
2637 dma_mask = *dev->dma_mask;
2638
2639 spin_lock_irqsave(&domain->lock, flags);
2640
2641 for_each_sg(sglist, s, nelems, i) {
2642 paddr = sg_phys(s);
2643
2644 s->dma_address = __map_single(dev, domain->priv,
2645 paddr, s->length, dir, false,
2646 dma_mask);
2647
2648 if (s->dma_address) {
2649 s->dma_length = s->length;
2650 mapped_elems++;
2651 } else
2652 goto unmap;
2653 }
2654
2655 domain_flush_complete(domain);
2656
2657 out:
2658 spin_unlock_irqrestore(&domain->lock, flags);
2659
2660 return mapped_elems;
2661 unmap:
2662 for_each_sg(sglist, s, mapped_elems, i) {
2663 if (s->dma_address)
2664 __unmap_single(domain->priv, s->dma_address,
2665 s->dma_length, dir);
2666 s->dma_address = s->dma_length = 0;
2667 }
2668
2669 mapped_elems = 0;
2670
2671 goto out;
2672 }
2673
2674 /*
2675 * The exported map_sg function for dma_ops (handles scatter-gather
2676 * lists).
2677 */
2678 static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2679 int nelems, enum dma_data_direction dir,
2680 struct dma_attrs *attrs)
2681 {
2682 unsigned long flags;
2683 struct protection_domain *domain;
2684 struct scatterlist *s;
2685 int i;
2686
2687 INC_STATS_COUNTER(cnt_unmap_sg);
2688
2689 domain = get_domain(dev);
2690 if (IS_ERR(domain))
2691 return;
2692
2693 spin_lock_irqsave(&domain->lock, flags);
2694
2695 for_each_sg(sglist, s, nelems, i) {
2696 __unmap_single(domain->priv, s->dma_address,
2697 s->dma_length, dir);
2698 s->dma_address = s->dma_length = 0;
2699 }
2700
2701 domain_flush_complete(domain);
2702
2703 spin_unlock_irqrestore(&domain->lock, flags);
2704 }
2705
2706 /*
2707 * The exported alloc_coherent function for dma_ops.
2708 */
2709 static void *alloc_coherent(struct device *dev, size_t size,
2710 dma_addr_t *dma_addr, gfp_t flag)
2711 {
2712 unsigned long flags;
2713 void *virt_addr;
2714 struct protection_domain *domain;
2715 phys_addr_t paddr;
2716 u64 dma_mask = dev->coherent_dma_mask;
2717
2718 INC_STATS_COUNTER(cnt_alloc_coherent);
2719
2720 domain = get_domain(dev);
2721 if (PTR_ERR(domain) == -EINVAL) {
2722 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2723 *dma_addr = __pa(virt_addr);
2724 return virt_addr;
2725 } else if (IS_ERR(domain))
2726 return NULL;
2727
2728 dma_mask = dev->coherent_dma_mask;
2729 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2730 flag |= __GFP_ZERO;
2731
2732 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2733 if (!virt_addr)
2734 return NULL;
2735
2736 paddr = virt_to_phys(virt_addr);
2737
2738 if (!dma_mask)
2739 dma_mask = *dev->dma_mask;
2740
2741 spin_lock_irqsave(&domain->lock, flags);
2742
2743 *dma_addr = __map_single(dev, domain->priv, paddr,
2744 size, DMA_BIDIRECTIONAL, true, dma_mask);
2745
2746 if (*dma_addr == DMA_ERROR_CODE) {
2747 spin_unlock_irqrestore(&domain->lock, flags);
2748 goto out_free;
2749 }
2750
2751 domain_flush_complete(domain);
2752
2753 spin_unlock_irqrestore(&domain->lock, flags);
2754
2755 return virt_addr;
2756
2757 out_free:
2758
2759 free_pages((unsigned long)virt_addr, get_order(size));
2760
2761 return NULL;
2762 }
2763
2764 /*
2765 * The exported free_coherent function for dma_ops.
2766 */
2767 static void free_coherent(struct device *dev, size_t size,
2768 void *virt_addr, dma_addr_t dma_addr)
2769 {
2770 unsigned long flags;
2771 struct protection_domain *domain;
2772
2773 INC_STATS_COUNTER(cnt_free_coherent);
2774
2775 domain = get_domain(dev);
2776 if (IS_ERR(domain))
2777 goto free_mem;
2778
2779 spin_lock_irqsave(&domain->lock, flags);
2780
2781 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2782
2783 domain_flush_complete(domain);
2784
2785 spin_unlock_irqrestore(&domain->lock, flags);
2786
2787 free_mem:
2788 free_pages((unsigned long)virt_addr, get_order(size));
2789 }
2790
2791 /*
2792 * This function is called by the DMA layer to find out if we can handle a
2793 * particular device. It is part of the dma_ops.
2794 */
2795 static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2796 {
2797 return check_device(dev);
2798 }
2799
2800 /*
2801 * The function for pre-allocating protection domains.
2802 *
2803 * If the driver core informs the DMA layer if a driver grabs a device
2804 * we don't need to preallocate the protection domains anymore.
2805 * For now we have to.
2806 */
2807 static void __init prealloc_protection_domains(void)
2808 {
2809 struct iommu_dev_data *dev_data;
2810 struct dma_ops_domain *dma_dom;
2811 struct pci_dev *dev = NULL;
2812 u16 devid;
2813
2814 for_each_pci_dev(dev) {
2815
2816 /* Do we handle this device? */
2817 if (!check_device(&dev->dev))
2818 continue;
2819
2820 dev_data = get_dev_data(&dev->dev);
2821 if (!amd_iommu_force_isolation && dev_data->iommu_v2) {
2822 /* Make sure passthrough domain is allocated */
2823 alloc_passthrough_domain();
2824 dev_data->passthrough = true;
2825 attach_device(&dev->dev, pt_domain);
2826 pr_info("AMD-Vi: Using passthough domain for device %s\n",
2827 dev_name(&dev->dev));
2828 }
2829
2830 /* Is there already any domain for it? */
2831 if (domain_for_device(&dev->dev))
2832 continue;
2833
2834 devid = get_device_id(&dev->dev);
2835
2836 dma_dom = dma_ops_domain_alloc();
2837 if (!dma_dom)
2838 continue;
2839 init_unity_mappings_for_device(dma_dom, devid);
2840 dma_dom->target_dev = devid;
2841
2842 attach_device(&dev->dev, &dma_dom->domain);
2843
2844 list_add_tail(&dma_dom->list, &iommu_pd_list);
2845 }
2846 }
2847
2848 static struct dma_map_ops amd_iommu_dma_ops = {
2849 .alloc_coherent = alloc_coherent,
2850 .free_coherent = free_coherent,
2851 .map_page = map_page,
2852 .unmap_page = unmap_page,
2853 .map_sg = map_sg,
2854 .unmap_sg = unmap_sg,
2855 .dma_supported = amd_iommu_dma_supported,
2856 };
2857
2858 static unsigned device_dma_ops_init(void)
2859 {
2860 struct iommu_dev_data *dev_data;
2861 struct pci_dev *pdev = NULL;
2862 unsigned unhandled = 0;
2863
2864 for_each_pci_dev(pdev) {
2865 if (!check_device(&pdev->dev)) {
2866
2867 iommu_ignore_device(&pdev->dev);
2868
2869 unhandled += 1;
2870 continue;
2871 }
2872
2873 dev_data = get_dev_data(&pdev->dev);
2874
2875 if (!dev_data->passthrough)
2876 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
2877 else
2878 pdev->dev.archdata.dma_ops = &nommu_dma_ops;
2879 }
2880
2881 return unhandled;
2882 }
2883
2884 /*
2885 * The function which clues the AMD IOMMU driver into dma_ops.
2886 */
2887
2888 void __init amd_iommu_init_api(void)
2889 {
2890 bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2891 }
2892
2893 int __init amd_iommu_init_dma_ops(void)
2894 {
2895 struct amd_iommu *iommu;
2896 int ret, unhandled;
2897
2898 /*
2899 * first allocate a default protection domain for every IOMMU we
2900 * found in the system. Devices not assigned to any other
2901 * protection domain will be assigned to the default one.
2902 */
2903 for_each_iommu(iommu) {
2904 iommu->default_dom = dma_ops_domain_alloc();
2905 if (iommu->default_dom == NULL)
2906 return -ENOMEM;
2907 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
2908 ret = iommu_init_unity_mappings(iommu);
2909 if (ret)
2910 goto free_domains;
2911 }
2912
2913 /*
2914 * Pre-allocate the protection domains for each device.
2915 */
2916 prealloc_protection_domains();
2917
2918 iommu_detected = 1;
2919 swiotlb = 0;
2920
2921 /* Make the driver finally visible to the drivers */
2922 unhandled = device_dma_ops_init();
2923 if (unhandled && max_pfn > MAX_DMA32_PFN) {
2924 /* There are unhandled devices - initialize swiotlb for them */
2925 swiotlb = 1;
2926 }
2927
2928 amd_iommu_stats_init();
2929
2930 return 0;
2931
2932 free_domains:
2933
2934 for_each_iommu(iommu) {
2935 if (iommu->default_dom)
2936 dma_ops_domain_free(iommu->default_dom);
2937 }
2938
2939 return ret;
2940 }
2941
2942 /*****************************************************************************
2943 *
2944 * The following functions belong to the exported interface of AMD IOMMU
2945 *
2946 * This interface allows access to lower level functions of the IOMMU
2947 * like protection domain handling and assignement of devices to domains
2948 * which is not possible with the dma_ops interface.
2949 *
2950 *****************************************************************************/
2951
2952 static void cleanup_domain(struct protection_domain *domain)
2953 {
2954 struct iommu_dev_data *dev_data, *next;
2955 unsigned long flags;
2956
2957 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2958
2959 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
2960 __detach_device(dev_data);
2961 atomic_set(&dev_data->bind, 0);
2962 }
2963
2964 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2965 }
2966
2967 static void protection_domain_free(struct protection_domain *domain)
2968 {
2969 if (!domain)
2970 return;
2971
2972 del_domain_from_list(domain);
2973
2974 if (domain->id)
2975 domain_id_free(domain->id);
2976
2977 kfree(domain);
2978 }
2979
2980 static struct protection_domain *protection_domain_alloc(void)
2981 {
2982 struct protection_domain *domain;
2983
2984 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2985 if (!domain)
2986 return NULL;
2987
2988 spin_lock_init(&domain->lock);
2989 mutex_init(&domain->api_lock);
2990 domain->id = domain_id_alloc();
2991 if (!domain->id)
2992 goto out_err;
2993 INIT_LIST_HEAD(&domain->dev_list);
2994
2995 add_domain_to_list(domain);
2996
2997 return domain;
2998
2999 out_err:
3000 kfree(domain);
3001
3002 return NULL;
3003 }
3004
3005 static int __init alloc_passthrough_domain(void)
3006 {
3007 if (pt_domain != NULL)
3008 return 0;
3009
3010 /* allocate passthrough domain */
3011 pt_domain = protection_domain_alloc();
3012 if (!pt_domain)
3013 return -ENOMEM;
3014
3015 pt_domain->mode = PAGE_MODE_NONE;
3016
3017 return 0;
3018 }
3019 static int amd_iommu_domain_init(struct iommu_domain *dom)
3020 {
3021 struct protection_domain *domain;
3022
3023 domain = protection_domain_alloc();
3024 if (!domain)
3025 goto out_free;
3026
3027 domain->mode = PAGE_MODE_3_LEVEL;
3028 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
3029 if (!domain->pt_root)
3030 goto out_free;
3031
3032 domain->iommu_domain = dom;
3033
3034 dom->priv = domain;
3035
3036 return 0;
3037
3038 out_free:
3039 protection_domain_free(domain);
3040
3041 return -ENOMEM;
3042 }
3043
3044 static void amd_iommu_domain_destroy(struct iommu_domain *dom)
3045 {
3046 struct protection_domain *domain = dom->priv;
3047
3048 if (!domain)
3049 return;
3050
3051 if (domain->dev_cnt > 0)
3052 cleanup_domain(domain);
3053
3054 BUG_ON(domain->dev_cnt != 0);
3055
3056 if (domain->mode != PAGE_MODE_NONE)
3057 free_pagetable(domain);
3058
3059 if (domain->flags & PD_IOMMUV2_MASK)
3060 free_gcr3_table(domain);
3061
3062 protection_domain_free(domain);
3063
3064 dom->priv = NULL;
3065 }
3066
3067 static void amd_iommu_detach_device(struct iommu_domain *dom,
3068 struct device *dev)
3069 {
3070 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3071 struct amd_iommu *iommu;
3072 u16 devid;
3073
3074 if (!check_device(dev))
3075 return;
3076
3077 devid = get_device_id(dev);
3078
3079 if (dev_data->domain != NULL)
3080 detach_device(dev);
3081
3082 iommu = amd_iommu_rlookup_table[devid];
3083 if (!iommu)
3084 return;
3085
3086 iommu_completion_wait(iommu);
3087 }
3088
3089 static int amd_iommu_attach_device(struct iommu_domain *dom,
3090 struct device *dev)
3091 {
3092 struct protection_domain *domain = dom->priv;
3093 struct iommu_dev_data *dev_data;
3094 struct amd_iommu *iommu;
3095 int ret;
3096
3097 if (!check_device(dev))
3098 return -EINVAL;
3099
3100 dev_data = dev->archdata.iommu;
3101
3102 iommu = amd_iommu_rlookup_table[dev_data->devid];
3103 if (!iommu)
3104 return -EINVAL;
3105
3106 if (dev_data->domain)
3107 detach_device(dev);
3108
3109 ret = attach_device(dev, domain);
3110
3111 iommu_completion_wait(iommu);
3112
3113 return ret;
3114 }
3115
3116 static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3117 phys_addr_t paddr, size_t page_size, int iommu_prot)
3118 {
3119 struct protection_domain *domain = dom->priv;
3120 int prot = 0;
3121 int ret;
3122
3123 if (domain->mode == PAGE_MODE_NONE)
3124 return -EINVAL;
3125
3126 if (iommu_prot & IOMMU_READ)
3127 prot |= IOMMU_PROT_IR;
3128 if (iommu_prot & IOMMU_WRITE)
3129 prot |= IOMMU_PROT_IW;
3130
3131 mutex_lock(&domain->api_lock);
3132 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
3133 mutex_unlock(&domain->api_lock);
3134
3135 return ret;
3136 }
3137
3138 static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3139 size_t page_size)
3140 {
3141 struct protection_domain *domain = dom->priv;
3142 size_t unmap_size;
3143
3144 if (domain->mode == PAGE_MODE_NONE)
3145 return -EINVAL;
3146
3147 mutex_lock(&domain->api_lock);
3148 unmap_size = iommu_unmap_page(domain, iova, page_size);
3149 mutex_unlock(&domain->api_lock);
3150
3151 domain_flush_tlb_pde(domain);
3152
3153 return unmap_size;
3154 }
3155
3156 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3157 unsigned long iova)
3158 {
3159 struct protection_domain *domain = dom->priv;
3160 unsigned long offset_mask;
3161 phys_addr_t paddr;
3162 u64 *pte, __pte;
3163
3164 if (domain->mode == PAGE_MODE_NONE)
3165 return iova;
3166
3167 pte = fetch_pte(domain, iova);
3168
3169 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3170 return 0;
3171
3172 if (PM_PTE_LEVEL(*pte) == 0)
3173 offset_mask = PAGE_SIZE - 1;
3174 else
3175 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
3176
3177 __pte = *pte & PM_ADDR_MASK;
3178 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
3179
3180 return paddr;
3181 }
3182
3183 static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
3184 unsigned long cap)
3185 {
3186 switch (cap) {
3187 case IOMMU_CAP_CACHE_COHERENCY:
3188 return 1;
3189 }
3190
3191 return 0;
3192 }
3193
3194 static int amd_iommu_device_group(struct device *dev, unsigned int *groupid)
3195 {
3196 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3197 struct pci_dev *pdev = to_pci_dev(dev);
3198 u16 devid;
3199
3200 if (!dev_data)
3201 return -ENODEV;
3202
3203 if (pdev->is_virtfn || !iommu_group_mf)
3204 devid = dev_data->devid;
3205 else
3206 devid = calc_devid(pdev->bus->number,
3207 PCI_DEVFN(PCI_SLOT(pdev->devfn), 0));
3208
3209 *groupid = amd_iommu_alias_table[devid];
3210
3211 return 0;
3212 }
3213
3214 static struct iommu_ops amd_iommu_ops = {
3215 .domain_init = amd_iommu_domain_init,
3216 .domain_destroy = amd_iommu_domain_destroy,
3217 .attach_dev = amd_iommu_attach_device,
3218 .detach_dev = amd_iommu_detach_device,
3219 .map = amd_iommu_map,
3220 .unmap = amd_iommu_unmap,
3221 .iova_to_phys = amd_iommu_iova_to_phys,
3222 .domain_has_cap = amd_iommu_domain_has_cap,
3223 .device_group = amd_iommu_device_group,
3224 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3225 };
3226
3227 /*****************************************************************************
3228 *
3229 * The next functions do a basic initialization of IOMMU for pass through
3230 * mode
3231 *
3232 * In passthrough mode the IOMMU is initialized and enabled but not used for
3233 * DMA-API translation.
3234 *
3235 *****************************************************************************/
3236
3237 int __init amd_iommu_init_passthrough(void)
3238 {
3239 struct iommu_dev_data *dev_data;
3240 struct pci_dev *dev = NULL;
3241 struct amd_iommu *iommu;
3242 u16 devid;
3243 int ret;
3244
3245 ret = alloc_passthrough_domain();
3246 if (ret)
3247 return ret;
3248
3249 for_each_pci_dev(dev) {
3250 if (!check_device(&dev->dev))
3251 continue;
3252
3253 dev_data = get_dev_data(&dev->dev);
3254 dev_data->passthrough = true;
3255
3256 devid = get_device_id(&dev->dev);
3257
3258 iommu = amd_iommu_rlookup_table[devid];
3259 if (!iommu)
3260 continue;
3261
3262 attach_device(&dev->dev, pt_domain);
3263 }
3264
3265 amd_iommu_stats_init();
3266
3267 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
3268
3269 return 0;
3270 }
3271
3272 /* IOMMUv2 specific functions */
3273 int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3274 {
3275 return atomic_notifier_chain_register(&ppr_notifier, nb);
3276 }
3277 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3278
3279 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3280 {
3281 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3282 }
3283 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3284
3285 void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3286 {
3287 struct protection_domain *domain = dom->priv;
3288 unsigned long flags;
3289
3290 spin_lock_irqsave(&domain->lock, flags);
3291
3292 /* Update data structure */
3293 domain->mode = PAGE_MODE_NONE;
3294 domain->updated = true;
3295
3296 /* Make changes visible to IOMMUs */
3297 update_domain(domain);
3298
3299 /* Page-table is not visible to IOMMU anymore, so free it */
3300 free_pagetable(domain);
3301
3302 spin_unlock_irqrestore(&domain->lock, flags);
3303 }
3304 EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3305
3306 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3307 {
3308 struct protection_domain *domain = dom->priv;
3309 unsigned long flags;
3310 int levels, ret;
3311
3312 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3313 return -EINVAL;
3314
3315 /* Number of GCR3 table levels required */
3316 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3317 levels += 1;
3318
3319 if (levels > amd_iommu_max_glx_val)
3320 return -EINVAL;
3321
3322 spin_lock_irqsave(&domain->lock, flags);
3323
3324 /*
3325 * Save us all sanity checks whether devices already in the
3326 * domain support IOMMUv2. Just force that the domain has no
3327 * devices attached when it is switched into IOMMUv2 mode.
3328 */
3329 ret = -EBUSY;
3330 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3331 goto out;
3332
3333 ret = -ENOMEM;
3334 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3335 if (domain->gcr3_tbl == NULL)
3336 goto out;
3337
3338 domain->glx = levels;
3339 domain->flags |= PD_IOMMUV2_MASK;
3340 domain->updated = true;
3341
3342 update_domain(domain);
3343
3344 ret = 0;
3345
3346 out:
3347 spin_unlock_irqrestore(&domain->lock, flags);
3348
3349 return ret;
3350 }
3351 EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3352
3353 static int __flush_pasid(struct protection_domain *domain, int pasid,
3354 u64 address, bool size)
3355 {
3356 struct iommu_dev_data *dev_data;
3357 struct iommu_cmd cmd;
3358 int i, ret;
3359
3360 if (!(domain->flags & PD_IOMMUV2_MASK))
3361 return -EINVAL;
3362
3363 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3364
3365 /*
3366 * IOMMU TLB needs to be flushed before Device TLB to
3367 * prevent device TLB refill from IOMMU TLB
3368 */
3369 for (i = 0; i < amd_iommus_present; ++i) {
3370 if (domain->dev_iommu[i] == 0)
3371 continue;
3372
3373 ret = iommu_queue_command(amd_iommus[i], &cmd);
3374 if (ret != 0)
3375 goto out;
3376 }
3377
3378 /* Wait until IOMMU TLB flushes are complete */
3379 domain_flush_complete(domain);
3380
3381 /* Now flush device TLBs */
3382 list_for_each_entry(dev_data, &domain->dev_list, list) {
3383 struct amd_iommu *iommu;
3384 int qdep;
3385
3386 BUG_ON(!dev_data->ats.enabled);
3387
3388 qdep = dev_data->ats.qdep;
3389 iommu = amd_iommu_rlookup_table[dev_data->devid];
3390
3391 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3392 qdep, address, size);
3393
3394 ret = iommu_queue_command(iommu, &cmd);
3395 if (ret != 0)
3396 goto out;
3397 }
3398
3399 /* Wait until all device TLBs are flushed */
3400 domain_flush_complete(domain);
3401
3402 ret = 0;
3403
3404 out:
3405
3406 return ret;
3407 }
3408
3409 static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3410 u64 address)
3411 {
3412 INC_STATS_COUNTER(invalidate_iotlb);
3413
3414 return __flush_pasid(domain, pasid, address, false);
3415 }
3416
3417 int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3418 u64 address)
3419 {
3420 struct protection_domain *domain = dom->priv;
3421 unsigned long flags;
3422 int ret;
3423
3424 spin_lock_irqsave(&domain->lock, flags);
3425 ret = __amd_iommu_flush_page(domain, pasid, address);
3426 spin_unlock_irqrestore(&domain->lock, flags);
3427
3428 return ret;
3429 }
3430 EXPORT_SYMBOL(amd_iommu_flush_page);
3431
3432 static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3433 {
3434 INC_STATS_COUNTER(invalidate_iotlb_all);
3435
3436 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3437 true);
3438 }
3439
3440 int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3441 {
3442 struct protection_domain *domain = dom->priv;
3443 unsigned long flags;
3444 int ret;
3445
3446 spin_lock_irqsave(&domain->lock, flags);
3447 ret = __amd_iommu_flush_tlb(domain, pasid);
3448 spin_unlock_irqrestore(&domain->lock, flags);
3449
3450 return ret;
3451 }
3452 EXPORT_SYMBOL(amd_iommu_flush_tlb);
3453
3454 static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3455 {
3456 int index;
3457 u64 *pte;
3458
3459 while (true) {
3460
3461 index = (pasid >> (9 * level)) & 0x1ff;
3462 pte = &root[index];
3463
3464 if (level == 0)
3465 break;
3466
3467 if (!(*pte & GCR3_VALID)) {
3468 if (!alloc)
3469 return NULL;
3470
3471 root = (void *)get_zeroed_page(GFP_ATOMIC);
3472 if (root == NULL)
3473 return NULL;
3474
3475 *pte = __pa(root) | GCR3_VALID;
3476 }
3477
3478 root = __va(*pte & PAGE_MASK);
3479
3480 level -= 1;
3481 }
3482
3483 return pte;
3484 }
3485
3486 static int __set_gcr3(struct protection_domain *domain, int pasid,
3487 unsigned long cr3)
3488 {
3489 u64 *pte;
3490
3491 if (domain->mode != PAGE_MODE_NONE)
3492 return -EINVAL;
3493
3494 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3495 if (pte == NULL)
3496 return -ENOMEM;
3497
3498 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3499
3500 return __amd_iommu_flush_tlb(domain, pasid);
3501 }
3502
3503 static int __clear_gcr3(struct protection_domain *domain, int pasid)
3504 {
3505 u64 *pte;
3506
3507 if (domain->mode != PAGE_MODE_NONE)
3508 return -EINVAL;
3509
3510 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3511 if (pte == NULL)
3512 return 0;
3513
3514 *pte = 0;
3515
3516 return __amd_iommu_flush_tlb(domain, pasid);
3517 }
3518
3519 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3520 unsigned long cr3)
3521 {
3522 struct protection_domain *domain = dom->priv;
3523 unsigned long flags;
3524 int ret;
3525
3526 spin_lock_irqsave(&domain->lock, flags);
3527 ret = __set_gcr3(domain, pasid, cr3);
3528 spin_unlock_irqrestore(&domain->lock, flags);
3529
3530 return ret;
3531 }
3532 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3533
3534 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3535 {
3536 struct protection_domain *domain = dom->priv;
3537 unsigned long flags;
3538 int ret;
3539
3540 spin_lock_irqsave(&domain->lock, flags);
3541 ret = __clear_gcr3(domain, pasid);
3542 spin_unlock_irqrestore(&domain->lock, flags);
3543
3544 return ret;
3545 }
3546 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3547
3548 int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3549 int status, int tag)
3550 {
3551 struct iommu_dev_data *dev_data;
3552 struct amd_iommu *iommu;
3553 struct iommu_cmd cmd;
3554
3555 INC_STATS_COUNTER(complete_ppr);
3556
3557 dev_data = get_dev_data(&pdev->dev);
3558 iommu = amd_iommu_rlookup_table[dev_data->devid];
3559
3560 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3561 tag, dev_data->pri_tlp);
3562
3563 return iommu_queue_command(iommu, &cmd);
3564 }
3565 EXPORT_SYMBOL(amd_iommu_complete_ppr);
3566
3567 struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3568 {
3569 struct protection_domain *domain;
3570
3571 domain = get_domain(&pdev->dev);
3572 if (IS_ERR(domain))
3573 return NULL;
3574
3575 /* Only return IOMMUv2 domains */
3576 if (!(domain->flags & PD_IOMMUV2_MASK))
3577 return NULL;
3578
3579 return domain->iommu_domain;
3580 }
3581 EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3582
3583 void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3584 {
3585 struct iommu_dev_data *dev_data;
3586
3587 if (!amd_iommu_v2_supported())
3588 return;
3589
3590 dev_data = get_dev_data(&pdev->dev);
3591 dev_data->errata |= (1 << erratum);
3592 }
3593 EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3594
3595 int amd_iommu_device_info(struct pci_dev *pdev,
3596 struct amd_iommu_device_info *info)
3597 {
3598 int max_pasids;
3599 int pos;
3600
3601 if (pdev == NULL || info == NULL)
3602 return -EINVAL;
3603
3604 if (!amd_iommu_v2_supported())
3605 return -EINVAL;
3606
3607 memset(info, 0, sizeof(*info));
3608
3609 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3610 if (pos)
3611 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3612
3613 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3614 if (pos)
3615 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3616
3617 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3618 if (pos) {
3619 int features;
3620
3621 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3622 max_pasids = min(max_pasids, (1 << 20));
3623
3624 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3625 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3626
3627 features = pci_pasid_features(pdev);
3628 if (features & PCI_PASID_CAP_EXEC)
3629 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3630 if (features & PCI_PASID_CAP_PRIV)
3631 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3632 }
3633
3634 return 0;
3635 }
3636 EXPORT_SYMBOL(amd_iommu_device_info);
Linus' kernel tree