Loading...
1/*
2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <joerg.roedel@amd.com>
4 * Leo Duran <leo.duran@amd.com>
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20#include <linux/pci.h>
21#include <linux/pci-ats.h>
22#include <linux/bitmap.h>
23#include <linux/slab.h>
24#include <linux/debugfs.h>
25#include <linux/scatterlist.h>
26#include <linux/dma-mapping.h>
27#include <linux/iommu-helper.h>
28#include <linux/iommu.h>
29#include <linux/delay.h>
30#include <linux/amd-iommu.h>
31#include <asm/msidef.h>
32#include <asm/proto.h>
33#include <asm/iommu.h>
34#include <asm/gart.h>
35#include <asm/dma.h>
36
37#include "amd_iommu_proto.h"
38#include "amd_iommu_types.h"
39
40#define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
41
42#define LOOP_TIMEOUT 100000
43
44static DEFINE_RWLOCK(amd_iommu_devtable_lock);
45
46/* A list of preallocated protection domains */
47static LIST_HEAD(iommu_pd_list);
48static DEFINE_SPINLOCK(iommu_pd_list_lock);
49
50/* List of all available dev_data structures */
51static LIST_HEAD(dev_data_list);
52static DEFINE_SPINLOCK(dev_data_list_lock);
53
54/*
55 * Domain for untranslated devices - only allocated
56 * if iommu=pt passed on kernel cmd line.
57 */
58static struct protection_domain *pt_domain;
59
60static struct iommu_ops amd_iommu_ops;
61
62/*
63 * general struct to manage commands send to an IOMMU
64 */
65struct iommu_cmd {
66 u32 data[4];
67};
68
69static void update_domain(struct protection_domain *domain);
70
71/****************************************************************************
72 *
73 * Helper functions
74 *
75 ****************************************************************************/
76
77static struct iommu_dev_data *alloc_dev_data(u16 devid)
78{
79 struct iommu_dev_data *dev_data;
80 unsigned long flags;
81
82 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
83 if (!dev_data)
84 return NULL;
85
86 dev_data->devid = devid;
87 atomic_set(&dev_data->bind, 0);
88
89 spin_lock_irqsave(&dev_data_list_lock, flags);
90 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
91 spin_unlock_irqrestore(&dev_data_list_lock, flags);
92
93 return dev_data;
94}
95
96static void free_dev_data(struct iommu_dev_data *dev_data)
97{
98 unsigned long flags;
99
100 spin_lock_irqsave(&dev_data_list_lock, flags);
101 list_del(&dev_data->dev_data_list);
102 spin_unlock_irqrestore(&dev_data_list_lock, flags);
103
104 kfree(dev_data);
105}
106
107static struct iommu_dev_data *search_dev_data(u16 devid)
108{
109 struct iommu_dev_data *dev_data;
110 unsigned long flags;
111
112 spin_lock_irqsave(&dev_data_list_lock, flags);
113 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
114 if (dev_data->devid == devid)
115 goto out_unlock;
116 }
117
118 dev_data = NULL;
119
120out_unlock:
121 spin_unlock_irqrestore(&dev_data_list_lock, flags);
122
123 return dev_data;
124}
125
126static struct iommu_dev_data *find_dev_data(u16 devid)
127{
128 struct iommu_dev_data *dev_data;
129
130 dev_data = search_dev_data(devid);
131
132 if (dev_data == NULL)
133 dev_data = alloc_dev_data(devid);
134
135 return dev_data;
136}
137
138static inline u16 get_device_id(struct device *dev)
139{
140 struct pci_dev *pdev = to_pci_dev(dev);
141
142 return calc_devid(pdev->bus->number, pdev->devfn);
143}
144
145static struct iommu_dev_data *get_dev_data(struct device *dev)
146{
147 return dev->archdata.iommu;
148}
149
150/*
151 * In this function the list of preallocated protection domains is traversed to
152 * find the domain for a specific device
153 */
154static struct dma_ops_domain *find_protection_domain(u16 devid)
155{
156 struct dma_ops_domain *entry, *ret = NULL;
157 unsigned long flags;
158 u16 alias = amd_iommu_alias_table[devid];
159
160 if (list_empty(&iommu_pd_list))
161 return NULL;
162
163 spin_lock_irqsave(&iommu_pd_list_lock, flags);
164
165 list_for_each_entry(entry, &iommu_pd_list, list) {
166 if (entry->target_dev == devid ||
167 entry->target_dev == alias) {
168 ret = entry;
169 break;
170 }
171 }
172
173 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
174
175 return ret;
176}
177
178/*
179 * This function checks if the driver got a valid device from the caller to
180 * avoid dereferencing invalid pointers.
181 */
182static bool check_device(struct device *dev)
183{
184 u16 devid;
185
186 if (!dev || !dev->dma_mask)
187 return false;
188
189 /* No device or no PCI device */
190 if (dev->bus != &pci_bus_type)
191 return false;
192
193 devid = get_device_id(dev);
194
195 /* Out of our scope? */
196 if (devid > amd_iommu_last_bdf)
197 return false;
198
199 if (amd_iommu_rlookup_table[devid] == NULL)
200 return false;
201
202 return true;
203}
204
205static int iommu_init_device(struct device *dev)
206{
207 struct iommu_dev_data *dev_data;
208 u16 alias;
209
210 if (dev->archdata.iommu)
211 return 0;
212
213 dev_data = find_dev_data(get_device_id(dev));
214 if (!dev_data)
215 return -ENOMEM;
216
217 alias = amd_iommu_alias_table[dev_data->devid];
218 if (alias != dev_data->devid) {
219 struct iommu_dev_data *alias_data;
220
221 alias_data = find_dev_data(alias);
222 if (alias_data == NULL) {
223 pr_err("AMD-Vi: Warning: Unhandled device %s\n",
224 dev_name(dev));
225 free_dev_data(dev_data);
226 return -ENOTSUPP;
227 }
228 dev_data->alias_data = alias_data;
229 }
230
231 dev->archdata.iommu = dev_data;
232
233 return 0;
234}
235
236static void iommu_ignore_device(struct device *dev)
237{
238 u16 devid, alias;
239
240 devid = get_device_id(dev);
241 alias = amd_iommu_alias_table[devid];
242
243 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
244 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
245
246 amd_iommu_rlookup_table[devid] = NULL;
247 amd_iommu_rlookup_table[alias] = NULL;
248}
249
250static void iommu_uninit_device(struct device *dev)
251{
252 /*
253 * Nothing to do here - we keep dev_data around for unplugged devices
254 * and reuse it when the device is re-plugged - not doing so would
255 * introduce a ton of races.
256 */
257}
258
259void __init amd_iommu_uninit_devices(void)
260{
261 struct iommu_dev_data *dev_data, *n;
262 struct pci_dev *pdev = NULL;
263
264 for_each_pci_dev(pdev) {
265
266 if (!check_device(&pdev->dev))
267 continue;
268
269 iommu_uninit_device(&pdev->dev);
270 }
271
272 /* Free all of our dev_data structures */
273 list_for_each_entry_safe(dev_data, n, &dev_data_list, dev_data_list)
274 free_dev_data(dev_data);
275}
276
277int __init amd_iommu_init_devices(void)
278{
279 struct pci_dev *pdev = NULL;
280 int ret = 0;
281
282 for_each_pci_dev(pdev) {
283
284 if (!check_device(&pdev->dev))
285 continue;
286
287 ret = iommu_init_device(&pdev->dev);
288 if (ret == -ENOTSUPP)
289 iommu_ignore_device(&pdev->dev);
290 else if (ret)
291 goto out_free;
292 }
293
294 return 0;
295
296out_free:
297
298 amd_iommu_uninit_devices();
299
300 return ret;
301}
302#ifdef CONFIG_AMD_IOMMU_STATS
303
304/*
305 * Initialization code for statistics collection
306 */
307
308DECLARE_STATS_COUNTER(compl_wait);
309DECLARE_STATS_COUNTER(cnt_map_single);
310DECLARE_STATS_COUNTER(cnt_unmap_single);
311DECLARE_STATS_COUNTER(cnt_map_sg);
312DECLARE_STATS_COUNTER(cnt_unmap_sg);
313DECLARE_STATS_COUNTER(cnt_alloc_coherent);
314DECLARE_STATS_COUNTER(cnt_free_coherent);
315DECLARE_STATS_COUNTER(cross_page);
316DECLARE_STATS_COUNTER(domain_flush_single);
317DECLARE_STATS_COUNTER(domain_flush_all);
318DECLARE_STATS_COUNTER(alloced_io_mem);
319DECLARE_STATS_COUNTER(total_map_requests);
320
321static struct dentry *stats_dir;
322static struct dentry *de_fflush;
323
324static void amd_iommu_stats_add(struct __iommu_counter *cnt)
325{
326 if (stats_dir == NULL)
327 return;
328
329 cnt->dent = debugfs_create_u64(cnt->name, 0444, stats_dir,
330 &cnt->value);
331}
332
333static void amd_iommu_stats_init(void)
334{
335 stats_dir = debugfs_create_dir("amd-iommu", NULL);
336 if (stats_dir == NULL)
337 return;
338
339 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
340 (u32 *)&amd_iommu_unmap_flush);
341
342 amd_iommu_stats_add(&compl_wait);
343 amd_iommu_stats_add(&cnt_map_single);
344 amd_iommu_stats_add(&cnt_unmap_single);
345 amd_iommu_stats_add(&cnt_map_sg);
346 amd_iommu_stats_add(&cnt_unmap_sg);
347 amd_iommu_stats_add(&cnt_alloc_coherent);
348 amd_iommu_stats_add(&cnt_free_coherent);
349 amd_iommu_stats_add(&cross_page);
350 amd_iommu_stats_add(&domain_flush_single);
351 amd_iommu_stats_add(&domain_flush_all);
352 amd_iommu_stats_add(&alloced_io_mem);
353 amd_iommu_stats_add(&total_map_requests);
354}
355
356#endif
357
358/****************************************************************************
359 *
360 * Interrupt handling functions
361 *
362 ****************************************************************************/
363
364static void dump_dte_entry(u16 devid)
365{
366 int i;
367
368 for (i = 0; i < 8; ++i)
369 pr_err("AMD-Vi: DTE[%d]: %08x\n", i,
370 amd_iommu_dev_table[devid].data[i]);
371}
372
373static void dump_command(unsigned long phys_addr)
374{
375 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
376 int i;
377
378 for (i = 0; i < 4; ++i)
379 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
380}
381
382static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
383{
384 u32 *event = __evt;
385 int type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
386 int devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
387 int domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
388 int flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
389 u64 address = (u64)(((u64)event[3]) << 32) | event[2];
390
391 printk(KERN_ERR "AMD-Vi: Event logged [");
392
393 switch (type) {
394 case EVENT_TYPE_ILL_DEV:
395 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
396 "address=0x%016llx flags=0x%04x]\n",
397 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
398 address, flags);
399 dump_dte_entry(devid);
400 break;
401 case EVENT_TYPE_IO_FAULT:
402 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
403 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
404 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
405 domid, address, flags);
406 break;
407 case EVENT_TYPE_DEV_TAB_ERR:
408 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
409 "address=0x%016llx flags=0x%04x]\n",
410 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
411 address, flags);
412 break;
413 case EVENT_TYPE_PAGE_TAB_ERR:
414 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
415 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
416 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
417 domid, address, flags);
418 break;
419 case EVENT_TYPE_ILL_CMD:
420 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
421 dump_command(address);
422 break;
423 case EVENT_TYPE_CMD_HARD_ERR:
424 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
425 "flags=0x%04x]\n", address, flags);
426 break;
427 case EVENT_TYPE_IOTLB_INV_TO:
428 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
429 "address=0x%016llx]\n",
430 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
431 address);
432 break;
433 case EVENT_TYPE_INV_DEV_REQ:
434 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
435 "address=0x%016llx flags=0x%04x]\n",
436 PCI_BUS(devid), PCI_SLOT(devid), PCI_FUNC(devid),
437 address, flags);
438 break;
439 default:
440 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
441 }
442}
443
444static void iommu_poll_events(struct amd_iommu *iommu)
445{
446 u32 head, tail;
447 unsigned long flags;
448
449 spin_lock_irqsave(&iommu->lock, flags);
450
451 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
452 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
453
454 while (head != tail) {
455 iommu_print_event(iommu, iommu->evt_buf + head);
456 head = (head + EVENT_ENTRY_SIZE) % iommu->evt_buf_size;
457 }
458
459 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
460
461 spin_unlock_irqrestore(&iommu->lock, flags);
462}
463
464irqreturn_t amd_iommu_int_thread(int irq, void *data)
465{
466 struct amd_iommu *iommu;
467
468 for_each_iommu(iommu)
469 iommu_poll_events(iommu);
470
471 return IRQ_HANDLED;
472}
473
474irqreturn_t amd_iommu_int_handler(int irq, void *data)
475{
476 return IRQ_WAKE_THREAD;
477}
478
479/****************************************************************************
480 *
481 * IOMMU command queuing functions
482 *
483 ****************************************************************************/
484
485static int wait_on_sem(volatile u64 *sem)
486{
487 int i = 0;
488
489 while (*sem == 0 && i < LOOP_TIMEOUT) {
490 udelay(1);
491 i += 1;
492 }
493
494 if (i == LOOP_TIMEOUT) {
495 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
496 return -EIO;
497 }
498
499 return 0;
500}
501
502static void copy_cmd_to_buffer(struct amd_iommu *iommu,
503 struct iommu_cmd *cmd,
504 u32 tail)
505{
506 u8 *target;
507
508 target = iommu->cmd_buf + tail;
509 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
510
511 /* Copy command to buffer */
512 memcpy(target, cmd, sizeof(*cmd));
513
514 /* Tell the IOMMU about it */
515 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
516}
517
518static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
519{
520 WARN_ON(address & 0x7ULL);
521
522 memset(cmd, 0, sizeof(*cmd));
523 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
524 cmd->data[1] = upper_32_bits(__pa(address));
525 cmd->data[2] = 1;
526 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
527}
528
529static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
530{
531 memset(cmd, 0, sizeof(*cmd));
532 cmd->data[0] = devid;
533 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
534}
535
536static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
537 size_t size, u16 domid, int pde)
538{
539 u64 pages;
540 int s;
541
542 pages = iommu_num_pages(address, size, PAGE_SIZE);
543 s = 0;
544
545 if (pages > 1) {
546 /*
547 * If we have to flush more than one page, flush all
548 * TLB entries for this domain
549 */
550 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
551 s = 1;
552 }
553
554 address &= PAGE_MASK;
555
556 memset(cmd, 0, sizeof(*cmd));
557 cmd->data[1] |= domid;
558 cmd->data[2] = lower_32_bits(address);
559 cmd->data[3] = upper_32_bits(address);
560 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
561 if (s) /* size bit - we flush more than one 4kb page */
562 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
563 if (pde) /* PDE bit - we wan't flush everything not only the PTEs */
564 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
565}
566
567static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
568 u64 address, size_t size)
569{
570 u64 pages;
571 int s;
572
573 pages = iommu_num_pages(address, size, PAGE_SIZE);
574 s = 0;
575
576 if (pages > 1) {
577 /*
578 * If we have to flush more than one page, flush all
579 * TLB entries for this domain
580 */
581 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
582 s = 1;
583 }
584
585 address &= PAGE_MASK;
586
587 memset(cmd, 0, sizeof(*cmd));
588 cmd->data[0] = devid;
589 cmd->data[0] |= (qdep & 0xff) << 24;
590 cmd->data[1] = devid;
591 cmd->data[2] = lower_32_bits(address);
592 cmd->data[3] = upper_32_bits(address);
593 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
594 if (s)
595 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
596}
597
598static void build_inv_all(struct iommu_cmd *cmd)
599{
600 memset(cmd, 0, sizeof(*cmd));
601 CMD_SET_TYPE(cmd, CMD_INV_ALL);
602}
603
604/*
605 * Writes the command to the IOMMUs command buffer and informs the
606 * hardware about the new command.
607 */
608static int iommu_queue_command_sync(struct amd_iommu *iommu,
609 struct iommu_cmd *cmd,
610 bool sync)
611{
612 u32 left, tail, head, next_tail;
613 unsigned long flags;
614
615 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
616
617again:
618 spin_lock_irqsave(&iommu->lock, flags);
619
620 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
621 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
622 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
623 left = (head - next_tail) % iommu->cmd_buf_size;
624
625 if (left <= 2) {
626 struct iommu_cmd sync_cmd;
627 volatile u64 sem = 0;
628 int ret;
629
630 build_completion_wait(&sync_cmd, (u64)&sem);
631 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
632
633 spin_unlock_irqrestore(&iommu->lock, flags);
634
635 if ((ret = wait_on_sem(&sem)) != 0)
636 return ret;
637
638 goto again;
639 }
640
641 copy_cmd_to_buffer(iommu, cmd, tail);
642
643 /* We need to sync now to make sure all commands are processed */
644 iommu->need_sync = sync;
645
646 spin_unlock_irqrestore(&iommu->lock, flags);
647
648 return 0;
649}
650
651static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
652{
653 return iommu_queue_command_sync(iommu, cmd, true);
654}
655
656/*
657 * This function queues a completion wait command into the command
658 * buffer of an IOMMU
659 */
660static int iommu_completion_wait(struct amd_iommu *iommu)
661{
662 struct iommu_cmd cmd;
663 volatile u64 sem = 0;
664 int ret;
665
666 if (!iommu->need_sync)
667 return 0;
668
669 build_completion_wait(&cmd, (u64)&sem);
670
671 ret = iommu_queue_command_sync(iommu, &cmd, false);
672 if (ret)
673 return ret;
674
675 return wait_on_sem(&sem);
676}
677
678static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
679{
680 struct iommu_cmd cmd;
681
682 build_inv_dte(&cmd, devid);
683
684 return iommu_queue_command(iommu, &cmd);
685}
686
687static void iommu_flush_dte_all(struct amd_iommu *iommu)
688{
689 u32 devid;
690
691 for (devid = 0; devid <= 0xffff; ++devid)
692 iommu_flush_dte(iommu, devid);
693
694 iommu_completion_wait(iommu);
695}
696
697/*
698 * This function uses heavy locking and may disable irqs for some time. But
699 * this is no issue because it is only called during resume.
700 */
701static void iommu_flush_tlb_all(struct amd_iommu *iommu)
702{
703 u32 dom_id;
704
705 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
706 struct iommu_cmd cmd;
707 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
708 dom_id, 1);
709 iommu_queue_command(iommu, &cmd);
710 }
711
712 iommu_completion_wait(iommu);
713}
714
715static void iommu_flush_all(struct amd_iommu *iommu)
716{
717 struct iommu_cmd cmd;
718
719 build_inv_all(&cmd);
720
721 iommu_queue_command(iommu, &cmd);
722 iommu_completion_wait(iommu);
723}
724
725void iommu_flush_all_caches(struct amd_iommu *iommu)
726{
727 if (iommu_feature(iommu, FEATURE_IA)) {
728 iommu_flush_all(iommu);
729 } else {
730 iommu_flush_dte_all(iommu);
731 iommu_flush_tlb_all(iommu);
732 }
733}
734
735/*
736 * Command send function for flushing on-device TLB
737 */
738static int device_flush_iotlb(struct iommu_dev_data *dev_data,
739 u64 address, size_t size)
740{
741 struct amd_iommu *iommu;
742 struct iommu_cmd cmd;
743 int qdep;
744
745 qdep = dev_data->ats.qdep;
746 iommu = amd_iommu_rlookup_table[dev_data->devid];
747
748 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
749
750 return iommu_queue_command(iommu, &cmd);
751}
752
753/*
754 * Command send function for invalidating a device table entry
755 */
756static int device_flush_dte(struct iommu_dev_data *dev_data)
757{
758 struct amd_iommu *iommu;
759 int ret;
760
761 iommu = amd_iommu_rlookup_table[dev_data->devid];
762
763 ret = iommu_flush_dte(iommu, dev_data->devid);
764 if (ret)
765 return ret;
766
767 if (dev_data->ats.enabled)
768 ret = device_flush_iotlb(dev_data, 0, ~0UL);
769
770 return ret;
771}
772
773/*
774 * TLB invalidation function which is called from the mapping functions.
775 * It invalidates a single PTE if the range to flush is within a single
776 * page. Otherwise it flushes the whole TLB of the IOMMU.
777 */
778static void __domain_flush_pages(struct protection_domain *domain,
779 u64 address, size_t size, int pde)
780{
781 struct iommu_dev_data *dev_data;
782 struct iommu_cmd cmd;
783 int ret = 0, i;
784
785 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
786
787 for (i = 0; i < amd_iommus_present; ++i) {
788 if (!domain->dev_iommu[i])
789 continue;
790
791 /*
792 * Devices of this domain are behind this IOMMU
793 * We need a TLB flush
794 */
795 ret |= iommu_queue_command(amd_iommus[i], &cmd);
796 }
797
798 list_for_each_entry(dev_data, &domain->dev_list, list) {
799
800 if (!dev_data->ats.enabled)
801 continue;
802
803 ret |= device_flush_iotlb(dev_data, address, size);
804 }
805
806 WARN_ON(ret);
807}
808
809static void domain_flush_pages(struct protection_domain *domain,
810 u64 address, size_t size)
811{
812 __domain_flush_pages(domain, address, size, 0);
813}
814
815/* Flush the whole IO/TLB for a given protection domain */
816static void domain_flush_tlb(struct protection_domain *domain)
817{
818 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
819}
820
821/* Flush the whole IO/TLB for a given protection domain - including PDE */
822static void domain_flush_tlb_pde(struct protection_domain *domain)
823{
824 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
825}
826
827static void domain_flush_complete(struct protection_domain *domain)
828{
829 int i;
830
831 for (i = 0; i < amd_iommus_present; ++i) {
832 if (!domain->dev_iommu[i])
833 continue;
834
835 /*
836 * Devices of this domain are behind this IOMMU
837 * We need to wait for completion of all commands.
838 */
839 iommu_completion_wait(amd_iommus[i]);
840 }
841}
842
843
844/*
845 * This function flushes the DTEs for all devices in domain
846 */
847static void domain_flush_devices(struct protection_domain *domain)
848{
849 struct iommu_dev_data *dev_data;
850
851 list_for_each_entry(dev_data, &domain->dev_list, list)
852 device_flush_dte(dev_data);
853}
854
855/****************************************************************************
856 *
857 * The functions below are used the create the page table mappings for
858 * unity mapped regions.
859 *
860 ****************************************************************************/
861
862/*
863 * This function is used to add another level to an IO page table. Adding
864 * another level increases the size of the address space by 9 bits to a size up
865 * to 64 bits.
866 */
867static bool increase_address_space(struct protection_domain *domain,
868 gfp_t gfp)
869{
870 u64 *pte;
871
872 if (domain->mode == PAGE_MODE_6_LEVEL)
873 /* address space already 64 bit large */
874 return false;
875
876 pte = (void *)get_zeroed_page(gfp);
877 if (!pte)
878 return false;
879
880 *pte = PM_LEVEL_PDE(domain->mode,
881 virt_to_phys(domain->pt_root));
882 domain->pt_root = pte;
883 domain->mode += 1;
884 domain->updated = true;
885
886 return true;
887}
888
889static u64 *alloc_pte(struct protection_domain *domain,
890 unsigned long address,
891 unsigned long page_size,
892 u64 **pte_page,
893 gfp_t gfp)
894{
895 int level, end_lvl;
896 u64 *pte, *page;
897
898 BUG_ON(!is_power_of_2(page_size));
899
900 while (address > PM_LEVEL_SIZE(domain->mode))
901 increase_address_space(domain, gfp);
902
903 level = domain->mode - 1;
904 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
905 address = PAGE_SIZE_ALIGN(address, page_size);
906 end_lvl = PAGE_SIZE_LEVEL(page_size);
907
908 while (level > end_lvl) {
909 if (!IOMMU_PTE_PRESENT(*pte)) {
910 page = (u64 *)get_zeroed_page(gfp);
911 if (!page)
912 return NULL;
913 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
914 }
915
916 /* No level skipping support yet */
917 if (PM_PTE_LEVEL(*pte) != level)
918 return NULL;
919
920 level -= 1;
921
922 pte = IOMMU_PTE_PAGE(*pte);
923
924 if (pte_page && level == end_lvl)
925 *pte_page = pte;
926
927 pte = &pte[PM_LEVEL_INDEX(level, address)];
928 }
929
930 return pte;
931}
932
933/*
934 * This function checks if there is a PTE for a given dma address. If
935 * there is one, it returns the pointer to it.
936 */
937static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
938{
939 int level;
940 u64 *pte;
941
942 if (address > PM_LEVEL_SIZE(domain->mode))
943 return NULL;
944
945 level = domain->mode - 1;
946 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
947
948 while (level > 0) {
949
950 /* Not Present */
951 if (!IOMMU_PTE_PRESENT(*pte))
952 return NULL;
953
954 /* Large PTE */
955 if (PM_PTE_LEVEL(*pte) == 0x07) {
956 unsigned long pte_mask, __pte;
957
958 /*
959 * If we have a series of large PTEs, make
960 * sure to return a pointer to the first one.
961 */
962 pte_mask = PTE_PAGE_SIZE(*pte);
963 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
964 __pte = ((unsigned long)pte) & pte_mask;
965
966 return (u64 *)__pte;
967 }
968
969 /* No level skipping support yet */
970 if (PM_PTE_LEVEL(*pte) != level)
971 return NULL;
972
973 level -= 1;
974
975 /* Walk to the next level */
976 pte = IOMMU_PTE_PAGE(*pte);
977 pte = &pte[PM_LEVEL_INDEX(level, address)];
978 }
979
980 return pte;
981}
982
983/*
984 * Generic mapping functions. It maps a physical address into a DMA
985 * address space. It allocates the page table pages if necessary.
986 * In the future it can be extended to a generic mapping function
987 * supporting all features of AMD IOMMU page tables like level skipping
988 * and full 64 bit address spaces.
989 */
990static int iommu_map_page(struct protection_domain *dom,
991 unsigned long bus_addr,
992 unsigned long phys_addr,
993 int prot,
994 unsigned long page_size)
995{
996 u64 __pte, *pte;
997 int i, count;
998
999 if (!(prot & IOMMU_PROT_MASK))
1000 return -EINVAL;
1001
1002 bus_addr = PAGE_ALIGN(bus_addr);
1003 phys_addr = PAGE_ALIGN(phys_addr);
1004 count = PAGE_SIZE_PTE_COUNT(page_size);
1005 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
1006
1007 for (i = 0; i < count; ++i)
1008 if (IOMMU_PTE_PRESENT(pte[i]))
1009 return -EBUSY;
1010
1011 if (page_size > PAGE_SIZE) {
1012 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1013 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1014 } else
1015 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1016
1017 if (prot & IOMMU_PROT_IR)
1018 __pte |= IOMMU_PTE_IR;
1019 if (prot & IOMMU_PROT_IW)
1020 __pte |= IOMMU_PTE_IW;
1021
1022 for (i = 0; i < count; ++i)
1023 pte[i] = __pte;
1024
1025 update_domain(dom);
1026
1027 return 0;
1028}
1029
1030static unsigned long iommu_unmap_page(struct protection_domain *dom,
1031 unsigned long bus_addr,
1032 unsigned long page_size)
1033{
1034 unsigned long long unmap_size, unmapped;
1035 u64 *pte;
1036
1037 BUG_ON(!is_power_of_2(page_size));
1038
1039 unmapped = 0;
1040
1041 while (unmapped < page_size) {
1042
1043 pte = fetch_pte(dom, bus_addr);
1044
1045 if (!pte) {
1046 /*
1047 * No PTE for this address
1048 * move forward in 4kb steps
1049 */
1050 unmap_size = PAGE_SIZE;
1051 } else if (PM_PTE_LEVEL(*pte) == 0) {
1052 /* 4kb PTE found for this address */
1053 unmap_size = PAGE_SIZE;
1054 *pte = 0ULL;
1055 } else {
1056 int count, i;
1057
1058 /* Large PTE found which maps this address */
1059 unmap_size = PTE_PAGE_SIZE(*pte);
1060 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1061 for (i = 0; i < count; i++)
1062 pte[i] = 0ULL;
1063 }
1064
1065 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1066 unmapped += unmap_size;
1067 }
1068
1069 BUG_ON(!is_power_of_2(unmapped));
1070
1071 return unmapped;
1072}
1073
1074/*
1075 * This function checks if a specific unity mapping entry is needed for
1076 * this specific IOMMU.
1077 */
1078static int iommu_for_unity_map(struct amd_iommu *iommu,
1079 struct unity_map_entry *entry)
1080{
1081 u16 bdf, i;
1082
1083 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1084 bdf = amd_iommu_alias_table[i];
1085 if (amd_iommu_rlookup_table[bdf] == iommu)
1086 return 1;
1087 }
1088
1089 return 0;
1090}
1091
1092/*
1093 * This function actually applies the mapping to the page table of the
1094 * dma_ops domain.
1095 */
1096static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1097 struct unity_map_entry *e)
1098{
1099 u64 addr;
1100 int ret;
1101
1102 for (addr = e->address_start; addr < e->address_end;
1103 addr += PAGE_SIZE) {
1104 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
1105 PAGE_SIZE);
1106 if (ret)
1107 return ret;
1108 /*
1109 * if unity mapping is in aperture range mark the page
1110 * as allocated in the aperture
1111 */
1112 if (addr < dma_dom->aperture_size)
1113 __set_bit(addr >> PAGE_SHIFT,
1114 dma_dom->aperture[0]->bitmap);
1115 }
1116
1117 return 0;
1118}
1119
1120/*
1121 * Init the unity mappings for a specific IOMMU in the system
1122 *
1123 * Basically iterates over all unity mapping entries and applies them to
1124 * the default domain DMA of that IOMMU if necessary.
1125 */
1126static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1127{
1128 struct unity_map_entry *entry;
1129 int ret;
1130
1131 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1132 if (!iommu_for_unity_map(iommu, entry))
1133 continue;
1134 ret = dma_ops_unity_map(iommu->default_dom, entry);
1135 if (ret)
1136 return ret;
1137 }
1138
1139 return 0;
1140}
1141
1142/*
1143 * Inits the unity mappings required for a specific device
1144 */
1145static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1146 u16 devid)
1147{
1148 struct unity_map_entry *e;
1149 int ret;
1150
1151 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1152 if (!(devid >= e->devid_start && devid <= e->devid_end))
1153 continue;
1154 ret = dma_ops_unity_map(dma_dom, e);
1155 if (ret)
1156 return ret;
1157 }
1158
1159 return 0;
1160}
1161
1162/****************************************************************************
1163 *
1164 * The next functions belong to the address allocator for the dma_ops
1165 * interface functions. They work like the allocators in the other IOMMU
1166 * drivers. Its basically a bitmap which marks the allocated pages in
1167 * the aperture. Maybe it could be enhanced in the future to a more
1168 * efficient allocator.
1169 *
1170 ****************************************************************************/
1171
1172/*
1173 * The address allocator core functions.
1174 *
1175 * called with domain->lock held
1176 */
1177
1178/*
1179 * Used to reserve address ranges in the aperture (e.g. for exclusion
1180 * ranges.
1181 */
1182static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1183 unsigned long start_page,
1184 unsigned int pages)
1185{
1186 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1187
1188 if (start_page + pages > last_page)
1189 pages = last_page - start_page;
1190
1191 for (i = start_page; i < start_page + pages; ++i) {
1192 int index = i / APERTURE_RANGE_PAGES;
1193 int page = i % APERTURE_RANGE_PAGES;
1194 __set_bit(page, dom->aperture[index]->bitmap);
1195 }
1196}
1197
1198/*
1199 * This function is used to add a new aperture range to an existing
1200 * aperture in case of dma_ops domain allocation or address allocation
1201 * failure.
1202 */
1203static int alloc_new_range(struct dma_ops_domain *dma_dom,
1204 bool populate, gfp_t gfp)
1205{
1206 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
1207 struct amd_iommu *iommu;
1208 unsigned long i, old_size;
1209
1210#ifdef CONFIG_IOMMU_STRESS
1211 populate = false;
1212#endif
1213
1214 if (index >= APERTURE_MAX_RANGES)
1215 return -ENOMEM;
1216
1217 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1218 if (!dma_dom->aperture[index])
1219 return -ENOMEM;
1220
1221 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1222 if (!dma_dom->aperture[index]->bitmap)
1223 goto out_free;
1224
1225 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1226
1227 if (populate) {
1228 unsigned long address = dma_dom->aperture_size;
1229 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1230 u64 *pte, *pte_page;
1231
1232 for (i = 0; i < num_ptes; ++i) {
1233 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
1234 &pte_page, gfp);
1235 if (!pte)
1236 goto out_free;
1237
1238 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1239
1240 address += APERTURE_RANGE_SIZE / 64;
1241 }
1242 }
1243
1244 old_size = dma_dom->aperture_size;
1245 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1246
1247 /* Reserve address range used for MSI messages */
1248 if (old_size < MSI_ADDR_BASE_LO &&
1249 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1250 unsigned long spage;
1251 int pages;
1252
1253 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1254 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1255
1256 dma_ops_reserve_addresses(dma_dom, spage, pages);
1257 }
1258
1259 /* Initialize the exclusion range if necessary */
1260 for_each_iommu(iommu) {
1261 if (iommu->exclusion_start &&
1262 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1263 && iommu->exclusion_start < dma_dom->aperture_size) {
1264 unsigned long startpage;
1265 int pages = iommu_num_pages(iommu->exclusion_start,
1266 iommu->exclusion_length,
1267 PAGE_SIZE);
1268 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1269 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1270 }
1271 }
1272
1273 /*
1274 * Check for areas already mapped as present in the new aperture
1275 * range and mark those pages as reserved in the allocator. Such
1276 * mappings may already exist as a result of requested unity
1277 * mappings for devices.
1278 */
1279 for (i = dma_dom->aperture[index]->offset;
1280 i < dma_dom->aperture_size;
1281 i += PAGE_SIZE) {
1282 u64 *pte = fetch_pte(&dma_dom->domain, i);
1283 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1284 continue;
1285
1286 dma_ops_reserve_addresses(dma_dom, i << PAGE_SHIFT, 1);
1287 }
1288
1289 update_domain(&dma_dom->domain);
1290
1291 return 0;
1292
1293out_free:
1294 update_domain(&dma_dom->domain);
1295
1296 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1297
1298 kfree(dma_dom->aperture[index]);
1299 dma_dom->aperture[index] = NULL;
1300
1301 return -ENOMEM;
1302}
1303
1304static unsigned long dma_ops_area_alloc(struct device *dev,
1305 struct dma_ops_domain *dom,
1306 unsigned int pages,
1307 unsigned long align_mask,
1308 u64 dma_mask,
1309 unsigned long start)
1310{
1311 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
1312 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1313 int i = start >> APERTURE_RANGE_SHIFT;
1314 unsigned long boundary_size;
1315 unsigned long address = -1;
1316 unsigned long limit;
1317
1318 next_bit >>= PAGE_SHIFT;
1319
1320 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1321 PAGE_SIZE) >> PAGE_SHIFT;
1322
1323 for (;i < max_index; ++i) {
1324 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1325
1326 if (dom->aperture[i]->offset >= dma_mask)
1327 break;
1328
1329 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1330 dma_mask >> PAGE_SHIFT);
1331
1332 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1333 limit, next_bit, pages, 0,
1334 boundary_size, align_mask);
1335 if (address != -1) {
1336 address = dom->aperture[i]->offset +
1337 (address << PAGE_SHIFT);
1338 dom->next_address = address + (pages << PAGE_SHIFT);
1339 break;
1340 }
1341
1342 next_bit = 0;
1343 }
1344
1345 return address;
1346}
1347
1348static unsigned long dma_ops_alloc_addresses(struct device *dev,
1349 struct dma_ops_domain *dom,
1350 unsigned int pages,
1351 unsigned long align_mask,
1352 u64 dma_mask)
1353{
1354 unsigned long address;
1355
1356#ifdef CONFIG_IOMMU_STRESS
1357 dom->next_address = 0;
1358 dom->need_flush = true;
1359#endif
1360
1361 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1362 dma_mask, dom->next_address);
1363
1364 if (address == -1) {
1365 dom->next_address = 0;
1366 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1367 dma_mask, 0);
1368 dom->need_flush = true;
1369 }
1370
1371 if (unlikely(address == -1))
1372 address = DMA_ERROR_CODE;
1373
1374 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1375
1376 return address;
1377}
1378
1379/*
1380 * The address free function.
1381 *
1382 * called with domain->lock held
1383 */
1384static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1385 unsigned long address,
1386 unsigned int pages)
1387{
1388 unsigned i = address >> APERTURE_RANGE_SHIFT;
1389 struct aperture_range *range = dom->aperture[i];
1390
1391 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1392
1393#ifdef CONFIG_IOMMU_STRESS
1394 if (i < 4)
1395 return;
1396#endif
1397
1398 if (address >= dom->next_address)
1399 dom->need_flush = true;
1400
1401 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
1402
1403 bitmap_clear(range->bitmap, address, pages);
1404
1405}
1406
1407/****************************************************************************
1408 *
1409 * The next functions belong to the domain allocation. A domain is
1410 * allocated for every IOMMU as the default domain. If device isolation
1411 * is enabled, every device get its own domain. The most important thing
1412 * about domains is the page table mapping the DMA address space they
1413 * contain.
1414 *
1415 ****************************************************************************/
1416
1417/*
1418 * This function adds a protection domain to the global protection domain list
1419 */
1420static void add_domain_to_list(struct protection_domain *domain)
1421{
1422 unsigned long flags;
1423
1424 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1425 list_add(&domain->list, &amd_iommu_pd_list);
1426 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1427}
1428
1429/*
1430 * This function removes a protection domain to the global
1431 * protection domain list
1432 */
1433static void del_domain_from_list(struct protection_domain *domain)
1434{
1435 unsigned long flags;
1436
1437 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1438 list_del(&domain->list);
1439 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1440}
1441
1442static u16 domain_id_alloc(void)
1443{
1444 unsigned long flags;
1445 int id;
1446
1447 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1448 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1449 BUG_ON(id == 0);
1450 if (id > 0 && id < MAX_DOMAIN_ID)
1451 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1452 else
1453 id = 0;
1454 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1455
1456 return id;
1457}
1458
1459static void domain_id_free(int id)
1460{
1461 unsigned long flags;
1462
1463 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1464 if (id > 0 && id < MAX_DOMAIN_ID)
1465 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1466 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1467}
1468
1469static void free_pagetable(struct protection_domain *domain)
1470{
1471 int i, j;
1472 u64 *p1, *p2, *p3;
1473
1474 p1 = domain->pt_root;
1475
1476 if (!p1)
1477 return;
1478
1479 for (i = 0; i < 512; ++i) {
1480 if (!IOMMU_PTE_PRESENT(p1[i]))
1481 continue;
1482
1483 p2 = IOMMU_PTE_PAGE(p1[i]);
1484 for (j = 0; j < 512; ++j) {
1485 if (!IOMMU_PTE_PRESENT(p2[j]))
1486 continue;
1487 p3 = IOMMU_PTE_PAGE(p2[j]);
1488 free_page((unsigned long)p3);
1489 }
1490
1491 free_page((unsigned long)p2);
1492 }
1493
1494 free_page((unsigned long)p1);
1495
1496 domain->pt_root = NULL;
1497}
1498
1499/*
1500 * Free a domain, only used if something went wrong in the
1501 * allocation path and we need to free an already allocated page table
1502 */
1503static void dma_ops_domain_free(struct dma_ops_domain *dom)
1504{
1505 int i;
1506
1507 if (!dom)
1508 return;
1509
1510 del_domain_from_list(&dom->domain);
1511
1512 free_pagetable(&dom->domain);
1513
1514 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1515 if (!dom->aperture[i])
1516 continue;
1517 free_page((unsigned long)dom->aperture[i]->bitmap);
1518 kfree(dom->aperture[i]);
1519 }
1520
1521 kfree(dom);
1522}
1523
1524/*
1525 * Allocates a new protection domain usable for the dma_ops functions.
1526 * It also initializes the page table and the address allocator data
1527 * structures required for the dma_ops interface
1528 */
1529static struct dma_ops_domain *dma_ops_domain_alloc(void)
1530{
1531 struct dma_ops_domain *dma_dom;
1532
1533 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1534 if (!dma_dom)
1535 return NULL;
1536
1537 spin_lock_init(&dma_dom->domain.lock);
1538
1539 dma_dom->domain.id = domain_id_alloc();
1540 if (dma_dom->domain.id == 0)
1541 goto free_dma_dom;
1542 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
1543 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
1544 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1545 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1546 dma_dom->domain.priv = dma_dom;
1547 if (!dma_dom->domain.pt_root)
1548 goto free_dma_dom;
1549
1550 dma_dom->need_flush = false;
1551 dma_dom->target_dev = 0xffff;
1552
1553 add_domain_to_list(&dma_dom->domain);
1554
1555 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
1556 goto free_dma_dom;
1557
1558 /*
1559 * mark the first page as allocated so we never return 0 as
1560 * a valid dma-address. So we can use 0 as error value
1561 */
1562 dma_dom->aperture[0]->bitmap[0] = 1;
1563 dma_dom->next_address = 0;
1564
1565
1566 return dma_dom;
1567
1568free_dma_dom:
1569 dma_ops_domain_free(dma_dom);
1570
1571 return NULL;
1572}
1573
1574/*
1575 * little helper function to check whether a given protection domain is a
1576 * dma_ops domain
1577 */
1578static bool dma_ops_domain(struct protection_domain *domain)
1579{
1580 return domain->flags & PD_DMA_OPS_MASK;
1581}
1582
1583static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1584{
1585 u64 pte_root = virt_to_phys(domain->pt_root);
1586 u32 flags = 0;
1587
1588 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1589 << DEV_ENTRY_MODE_SHIFT;
1590 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1591
1592 if (ats)
1593 flags |= DTE_FLAG_IOTLB;
1594
1595 amd_iommu_dev_table[devid].data[3] |= flags;
1596 amd_iommu_dev_table[devid].data[2] = domain->id;
1597 amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
1598 amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
1599}
1600
1601static void clear_dte_entry(u16 devid)
1602{
1603 /* remove entry from the device table seen by the hardware */
1604 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1605 amd_iommu_dev_table[devid].data[1] = 0;
1606 amd_iommu_dev_table[devid].data[2] = 0;
1607
1608 amd_iommu_apply_erratum_63(devid);
1609}
1610
1611static void do_attach(struct iommu_dev_data *dev_data,
1612 struct protection_domain *domain)
1613{
1614 struct amd_iommu *iommu;
1615 bool ats;
1616
1617 iommu = amd_iommu_rlookup_table[dev_data->devid];
1618 ats = dev_data->ats.enabled;
1619
1620 /* Update data structures */
1621 dev_data->domain = domain;
1622 list_add(&dev_data->list, &domain->dev_list);
1623 set_dte_entry(dev_data->devid, domain, ats);
1624
1625 /* Do reference counting */
1626 domain->dev_iommu[iommu->index] += 1;
1627 domain->dev_cnt += 1;
1628
1629 /* Flush the DTE entry */
1630 device_flush_dte(dev_data);
1631}
1632
1633static void do_detach(struct iommu_dev_data *dev_data)
1634{
1635 struct amd_iommu *iommu;
1636
1637 iommu = amd_iommu_rlookup_table[dev_data->devid];
1638
1639 /* decrease reference counters */
1640 dev_data->domain->dev_iommu[iommu->index] -= 1;
1641 dev_data->domain->dev_cnt -= 1;
1642
1643 /* Update data structures */
1644 dev_data->domain = NULL;
1645 list_del(&dev_data->list);
1646 clear_dte_entry(dev_data->devid);
1647
1648 /* Flush the DTE entry */
1649 device_flush_dte(dev_data);
1650}
1651
1652/*
1653 * If a device is not yet associated with a domain, this function does
1654 * assigns it visible for the hardware
1655 */
1656static int __attach_device(struct iommu_dev_data *dev_data,
1657 struct protection_domain *domain)
1658{
1659 int ret;
1660
1661 /* lock domain */
1662 spin_lock(&domain->lock);
1663
1664 if (dev_data->alias_data != NULL) {
1665 struct iommu_dev_data *alias_data = dev_data->alias_data;
1666
1667 /* Some sanity checks */
1668 ret = -EBUSY;
1669 if (alias_data->domain != NULL &&
1670 alias_data->domain != domain)
1671 goto out_unlock;
1672
1673 if (dev_data->domain != NULL &&
1674 dev_data->domain != domain)
1675 goto out_unlock;
1676
1677 /* Do real assignment */
1678 if (alias_data->domain == NULL)
1679 do_attach(alias_data, domain);
1680
1681 atomic_inc(&alias_data->bind);
1682 }
1683
1684 if (dev_data->domain == NULL)
1685 do_attach(dev_data, domain);
1686
1687 atomic_inc(&dev_data->bind);
1688
1689 ret = 0;
1690
1691out_unlock:
1692
1693 /* ready */
1694 spin_unlock(&domain->lock);
1695
1696 return ret;
1697}
1698
1699/*
1700 * If a device is not yet associated with a domain, this function does
1701 * assigns it visible for the hardware
1702 */
1703static int attach_device(struct device *dev,
1704 struct protection_domain *domain)
1705{
1706 struct pci_dev *pdev = to_pci_dev(dev);
1707 struct iommu_dev_data *dev_data;
1708 unsigned long flags;
1709 int ret;
1710
1711 dev_data = get_dev_data(dev);
1712
1713 if (amd_iommu_iotlb_sup && pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1714 dev_data->ats.enabled = true;
1715 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1716 }
1717
1718 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1719 ret = __attach_device(dev_data, domain);
1720 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1721
1722 /*
1723 * We might boot into a crash-kernel here. The crashed kernel
1724 * left the caches in the IOMMU dirty. So we have to flush
1725 * here to evict all dirty stuff.
1726 */
1727 domain_flush_tlb_pde(domain);
1728
1729 return ret;
1730}
1731
1732/*
1733 * Removes a device from a protection domain (unlocked)
1734 */
1735static void __detach_device(struct iommu_dev_data *dev_data)
1736{
1737 struct protection_domain *domain;
1738 unsigned long flags;
1739
1740 BUG_ON(!dev_data->domain);
1741
1742 domain = dev_data->domain;
1743
1744 spin_lock_irqsave(&domain->lock, flags);
1745
1746 if (dev_data->alias_data != NULL) {
1747 struct iommu_dev_data *alias_data = dev_data->alias_data;
1748
1749 if (atomic_dec_and_test(&alias_data->bind))
1750 do_detach(alias_data);
1751 }
1752
1753 if (atomic_dec_and_test(&dev_data->bind))
1754 do_detach(dev_data);
1755
1756 spin_unlock_irqrestore(&domain->lock, flags);
1757
1758 /*
1759 * If we run in passthrough mode the device must be assigned to the
1760 * passthrough domain if it is detached from any other domain.
1761 * Make sure we can deassign from the pt_domain itself.
1762 */
1763 if (iommu_pass_through &&
1764 (dev_data->domain == NULL && domain != pt_domain))
1765 __attach_device(dev_data, pt_domain);
1766}
1767
1768/*
1769 * Removes a device from a protection domain (with devtable_lock held)
1770 */
1771static void detach_device(struct device *dev)
1772{
1773 struct iommu_dev_data *dev_data;
1774 unsigned long flags;
1775
1776 dev_data = get_dev_data(dev);
1777
1778 /* lock device table */
1779 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1780 __detach_device(dev_data);
1781 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1782
1783 if (dev_data->ats.enabled) {
1784 pci_disable_ats(to_pci_dev(dev));
1785 dev_data->ats.enabled = false;
1786 }
1787}
1788
1789/*
1790 * Find out the protection domain structure for a given PCI device. This
1791 * will give us the pointer to the page table root for example.
1792 */
1793static struct protection_domain *domain_for_device(struct device *dev)
1794{
1795 struct iommu_dev_data *dev_data;
1796 struct protection_domain *dom = NULL;
1797 unsigned long flags;
1798
1799 dev_data = get_dev_data(dev);
1800
1801 if (dev_data->domain)
1802 return dev_data->domain;
1803
1804 if (dev_data->alias_data != NULL) {
1805 struct iommu_dev_data *alias_data = dev_data->alias_data;
1806
1807 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
1808 if (alias_data->domain != NULL) {
1809 __attach_device(dev_data, alias_data->domain);
1810 dom = alias_data->domain;
1811 }
1812 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1813 }
1814
1815 return dom;
1816}
1817
1818static int device_change_notifier(struct notifier_block *nb,
1819 unsigned long action, void *data)
1820{
1821 struct device *dev = data;
1822 u16 devid;
1823 struct protection_domain *domain;
1824 struct dma_ops_domain *dma_domain;
1825 struct amd_iommu *iommu;
1826 unsigned long flags;
1827
1828 if (!check_device(dev))
1829 return 0;
1830
1831 devid = get_device_id(dev);
1832 iommu = amd_iommu_rlookup_table[devid];
1833
1834 switch (action) {
1835 case BUS_NOTIFY_UNBOUND_DRIVER:
1836
1837 domain = domain_for_device(dev);
1838
1839 if (!domain)
1840 goto out;
1841 if (iommu_pass_through)
1842 break;
1843 detach_device(dev);
1844 break;
1845 case BUS_NOTIFY_ADD_DEVICE:
1846
1847 iommu_init_device(dev);
1848
1849 domain = domain_for_device(dev);
1850
1851 /* allocate a protection domain if a device is added */
1852 dma_domain = find_protection_domain(devid);
1853 if (dma_domain)
1854 goto out;
1855 dma_domain = dma_ops_domain_alloc();
1856 if (!dma_domain)
1857 goto out;
1858 dma_domain->target_dev = devid;
1859
1860 spin_lock_irqsave(&iommu_pd_list_lock, flags);
1861 list_add_tail(&dma_domain->list, &iommu_pd_list);
1862 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
1863
1864 break;
1865 case BUS_NOTIFY_DEL_DEVICE:
1866
1867 iommu_uninit_device(dev);
1868
1869 default:
1870 goto out;
1871 }
1872
1873 iommu_completion_wait(iommu);
1874
1875out:
1876 return 0;
1877}
1878
1879static struct notifier_block device_nb = {
1880 .notifier_call = device_change_notifier,
1881};
1882
1883void amd_iommu_init_notifier(void)
1884{
1885 bus_register_notifier(&pci_bus_type, &device_nb);
1886}
1887
1888/*****************************************************************************
1889 *
1890 * The next functions belong to the dma_ops mapping/unmapping code.
1891 *
1892 *****************************************************************************/
1893
1894/*
1895 * In the dma_ops path we only have the struct device. This function
1896 * finds the corresponding IOMMU, the protection domain and the
1897 * requestor id for a given device.
1898 * If the device is not yet associated with a domain this is also done
1899 * in this function.
1900 */
1901static struct protection_domain *get_domain(struct device *dev)
1902{
1903 struct protection_domain *domain;
1904 struct dma_ops_domain *dma_dom;
1905 u16 devid = get_device_id(dev);
1906
1907 if (!check_device(dev))
1908 return ERR_PTR(-EINVAL);
1909
1910 domain = domain_for_device(dev);
1911 if (domain != NULL && !dma_ops_domain(domain))
1912 return ERR_PTR(-EBUSY);
1913
1914 if (domain != NULL)
1915 return domain;
1916
1917 /* Device not bount yet - bind it */
1918 dma_dom = find_protection_domain(devid);
1919 if (!dma_dom)
1920 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
1921 attach_device(dev, &dma_dom->domain);
1922 DUMP_printk("Using protection domain %d for device %s\n",
1923 dma_dom->domain.id, dev_name(dev));
1924
1925 return &dma_dom->domain;
1926}
1927
1928static void update_device_table(struct protection_domain *domain)
1929{
1930 struct iommu_dev_data *dev_data;
1931
1932 list_for_each_entry(dev_data, &domain->dev_list, list)
1933 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
1934}
1935
1936static void update_domain(struct protection_domain *domain)
1937{
1938 if (!domain->updated)
1939 return;
1940
1941 update_device_table(domain);
1942
1943 domain_flush_devices(domain);
1944 domain_flush_tlb_pde(domain);
1945
1946 domain->updated = false;
1947}
1948
1949/*
1950 * This function fetches the PTE for a given address in the aperture
1951 */
1952static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
1953 unsigned long address)
1954{
1955 struct aperture_range *aperture;
1956 u64 *pte, *pte_page;
1957
1958 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
1959 if (!aperture)
1960 return NULL;
1961
1962 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
1963 if (!pte) {
1964 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
1965 GFP_ATOMIC);
1966 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
1967 } else
1968 pte += PM_LEVEL_INDEX(0, address);
1969
1970 update_domain(&dom->domain);
1971
1972 return pte;
1973}
1974
1975/*
1976 * This is the generic map function. It maps one 4kb page at paddr to
1977 * the given address in the DMA address space for the domain.
1978 */
1979static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
1980 unsigned long address,
1981 phys_addr_t paddr,
1982 int direction)
1983{
1984 u64 *pte, __pte;
1985
1986 WARN_ON(address > dom->aperture_size);
1987
1988 paddr &= PAGE_MASK;
1989
1990 pte = dma_ops_get_pte(dom, address);
1991 if (!pte)
1992 return DMA_ERROR_CODE;
1993
1994 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
1995
1996 if (direction == DMA_TO_DEVICE)
1997 __pte |= IOMMU_PTE_IR;
1998 else if (direction == DMA_FROM_DEVICE)
1999 __pte |= IOMMU_PTE_IW;
2000 else if (direction == DMA_BIDIRECTIONAL)
2001 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2002
2003 WARN_ON(*pte);
2004
2005 *pte = __pte;
2006
2007 return (dma_addr_t)address;
2008}
2009
2010/*
2011 * The generic unmapping function for on page in the DMA address space.
2012 */
2013static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
2014 unsigned long address)
2015{
2016 struct aperture_range *aperture;
2017 u64 *pte;
2018
2019 if (address >= dom->aperture_size)
2020 return;
2021
2022 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2023 if (!aperture)
2024 return;
2025
2026 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2027 if (!pte)
2028 return;
2029
2030 pte += PM_LEVEL_INDEX(0, address);
2031
2032 WARN_ON(!*pte);
2033
2034 *pte = 0ULL;
2035}
2036
2037/*
2038 * This function contains common code for mapping of a physically
2039 * contiguous memory region into DMA address space. It is used by all
2040 * mapping functions provided with this IOMMU driver.
2041 * Must be called with the domain lock held.
2042 */
2043static dma_addr_t __map_single(struct device *dev,
2044 struct dma_ops_domain *dma_dom,
2045 phys_addr_t paddr,
2046 size_t size,
2047 int dir,
2048 bool align,
2049 u64 dma_mask)
2050{
2051 dma_addr_t offset = paddr & ~PAGE_MASK;
2052 dma_addr_t address, start, ret;
2053 unsigned int pages;
2054 unsigned long align_mask = 0;
2055 int i;
2056
2057 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2058 paddr &= PAGE_MASK;
2059
2060 INC_STATS_COUNTER(total_map_requests);
2061
2062 if (pages > 1)
2063 INC_STATS_COUNTER(cross_page);
2064
2065 if (align)
2066 align_mask = (1UL << get_order(size)) - 1;
2067
2068retry:
2069 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2070 dma_mask);
2071 if (unlikely(address == DMA_ERROR_CODE)) {
2072 /*
2073 * setting next_address here will let the address
2074 * allocator only scan the new allocated range in the
2075 * first run. This is a small optimization.
2076 */
2077 dma_dom->next_address = dma_dom->aperture_size;
2078
2079 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
2080 goto out;
2081
2082 /*
2083 * aperture was successfully enlarged by 128 MB, try
2084 * allocation again
2085 */
2086 goto retry;
2087 }
2088
2089 start = address;
2090 for (i = 0; i < pages; ++i) {
2091 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
2092 if (ret == DMA_ERROR_CODE)
2093 goto out_unmap;
2094
2095 paddr += PAGE_SIZE;
2096 start += PAGE_SIZE;
2097 }
2098 address += offset;
2099
2100 ADD_STATS_COUNTER(alloced_io_mem, size);
2101
2102 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
2103 domain_flush_tlb(&dma_dom->domain);
2104 dma_dom->need_flush = false;
2105 } else if (unlikely(amd_iommu_np_cache))
2106 domain_flush_pages(&dma_dom->domain, address, size);
2107
2108out:
2109 return address;
2110
2111out_unmap:
2112
2113 for (--i; i >= 0; --i) {
2114 start -= PAGE_SIZE;
2115 dma_ops_domain_unmap(dma_dom, start);
2116 }
2117
2118 dma_ops_free_addresses(dma_dom, address, pages);
2119
2120 return DMA_ERROR_CODE;
2121}
2122
2123/*
2124 * Does the reverse of the __map_single function. Must be called with
2125 * the domain lock held too
2126 */
2127static void __unmap_single(struct dma_ops_domain *dma_dom,
2128 dma_addr_t dma_addr,
2129 size_t size,
2130 int dir)
2131{
2132 dma_addr_t flush_addr;
2133 dma_addr_t i, start;
2134 unsigned int pages;
2135
2136 if ((dma_addr == DMA_ERROR_CODE) ||
2137 (dma_addr + size > dma_dom->aperture_size))
2138 return;
2139
2140 flush_addr = dma_addr;
2141 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2142 dma_addr &= PAGE_MASK;
2143 start = dma_addr;
2144
2145 for (i = 0; i < pages; ++i) {
2146 dma_ops_domain_unmap(dma_dom, start);
2147 start += PAGE_SIZE;
2148 }
2149
2150 SUB_STATS_COUNTER(alloced_io_mem, size);
2151
2152 dma_ops_free_addresses(dma_dom, dma_addr, pages);
2153
2154 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
2155 domain_flush_pages(&dma_dom->domain, flush_addr, size);
2156 dma_dom->need_flush = false;
2157 }
2158}
2159
2160/*
2161 * The exported map_single function for dma_ops.
2162 */
2163static dma_addr_t map_page(struct device *dev, struct page *page,
2164 unsigned long offset, size_t size,
2165 enum dma_data_direction dir,
2166 struct dma_attrs *attrs)
2167{
2168 unsigned long flags;
2169 struct protection_domain *domain;
2170 dma_addr_t addr;
2171 u64 dma_mask;
2172 phys_addr_t paddr = page_to_phys(page) + offset;
2173
2174 INC_STATS_COUNTER(cnt_map_single);
2175
2176 domain = get_domain(dev);
2177 if (PTR_ERR(domain) == -EINVAL)
2178 return (dma_addr_t)paddr;
2179 else if (IS_ERR(domain))
2180 return DMA_ERROR_CODE;
2181
2182 dma_mask = *dev->dma_mask;
2183
2184 spin_lock_irqsave(&domain->lock, flags);
2185
2186 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
2187 dma_mask);
2188 if (addr == DMA_ERROR_CODE)
2189 goto out;
2190
2191 domain_flush_complete(domain);
2192
2193out:
2194 spin_unlock_irqrestore(&domain->lock, flags);
2195
2196 return addr;
2197}
2198
2199/*
2200 * The exported unmap_single function for dma_ops.
2201 */
2202static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2203 enum dma_data_direction dir, struct dma_attrs *attrs)
2204{
2205 unsigned long flags;
2206 struct protection_domain *domain;
2207
2208 INC_STATS_COUNTER(cnt_unmap_single);
2209
2210 domain = get_domain(dev);
2211 if (IS_ERR(domain))
2212 return;
2213
2214 spin_lock_irqsave(&domain->lock, flags);
2215
2216 __unmap_single(domain->priv, dma_addr, size, dir);
2217
2218 domain_flush_complete(domain);
2219
2220 spin_unlock_irqrestore(&domain->lock, flags);
2221}
2222
2223/*
2224 * This is a special map_sg function which is used if we should map a
2225 * device which is not handled by an AMD IOMMU in the system.
2226 */
2227static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2228 int nelems, int dir)
2229{
2230 struct scatterlist *s;
2231 int i;
2232
2233 for_each_sg(sglist, s, nelems, i) {
2234 s->dma_address = (dma_addr_t)sg_phys(s);
2235 s->dma_length = s->length;
2236 }
2237
2238 return nelems;
2239}
2240
2241/*
2242 * The exported map_sg function for dma_ops (handles scatter-gather
2243 * lists).
2244 */
2245static int map_sg(struct device *dev, struct scatterlist *sglist,
2246 int nelems, enum dma_data_direction dir,
2247 struct dma_attrs *attrs)
2248{
2249 unsigned long flags;
2250 struct protection_domain *domain;
2251 int i;
2252 struct scatterlist *s;
2253 phys_addr_t paddr;
2254 int mapped_elems = 0;
2255 u64 dma_mask;
2256
2257 INC_STATS_COUNTER(cnt_map_sg);
2258
2259 domain = get_domain(dev);
2260 if (PTR_ERR(domain) == -EINVAL)
2261 return map_sg_no_iommu(dev, sglist, nelems, dir);
2262 else if (IS_ERR(domain))
2263 return 0;
2264
2265 dma_mask = *dev->dma_mask;
2266
2267 spin_lock_irqsave(&domain->lock, flags);
2268
2269 for_each_sg(sglist, s, nelems, i) {
2270 paddr = sg_phys(s);
2271
2272 s->dma_address = __map_single(dev, domain->priv,
2273 paddr, s->length, dir, false,
2274 dma_mask);
2275
2276 if (s->dma_address) {
2277 s->dma_length = s->length;
2278 mapped_elems++;
2279 } else
2280 goto unmap;
2281 }
2282
2283 domain_flush_complete(domain);
2284
2285out:
2286 spin_unlock_irqrestore(&domain->lock, flags);
2287
2288 return mapped_elems;
2289unmap:
2290 for_each_sg(sglist, s, mapped_elems, i) {
2291 if (s->dma_address)
2292 __unmap_single(domain->priv, s->dma_address,
2293 s->dma_length, dir);
2294 s->dma_address = s->dma_length = 0;
2295 }
2296
2297 mapped_elems = 0;
2298
2299 goto out;
2300}
2301
2302/*
2303 * The exported map_sg function for dma_ops (handles scatter-gather
2304 * lists).
2305 */
2306static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2307 int nelems, enum dma_data_direction dir,
2308 struct dma_attrs *attrs)
2309{
2310 unsigned long flags;
2311 struct protection_domain *domain;
2312 struct scatterlist *s;
2313 int i;
2314
2315 INC_STATS_COUNTER(cnt_unmap_sg);
2316
2317 domain = get_domain(dev);
2318 if (IS_ERR(domain))
2319 return;
2320
2321 spin_lock_irqsave(&domain->lock, flags);
2322
2323 for_each_sg(sglist, s, nelems, i) {
2324 __unmap_single(domain->priv, s->dma_address,
2325 s->dma_length, dir);
2326 s->dma_address = s->dma_length = 0;
2327 }
2328
2329 domain_flush_complete(domain);
2330
2331 spin_unlock_irqrestore(&domain->lock, flags);
2332}
2333
2334/*
2335 * The exported alloc_coherent function for dma_ops.
2336 */
2337static void *alloc_coherent(struct device *dev, size_t size,
2338 dma_addr_t *dma_addr, gfp_t flag)
2339{
2340 unsigned long flags;
2341 void *virt_addr;
2342 struct protection_domain *domain;
2343 phys_addr_t paddr;
2344 u64 dma_mask = dev->coherent_dma_mask;
2345
2346 INC_STATS_COUNTER(cnt_alloc_coherent);
2347
2348 domain = get_domain(dev);
2349 if (PTR_ERR(domain) == -EINVAL) {
2350 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2351 *dma_addr = __pa(virt_addr);
2352 return virt_addr;
2353 } else if (IS_ERR(domain))
2354 return NULL;
2355
2356 dma_mask = dev->coherent_dma_mask;
2357 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2358 flag |= __GFP_ZERO;
2359
2360 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2361 if (!virt_addr)
2362 return NULL;
2363
2364 paddr = virt_to_phys(virt_addr);
2365
2366 if (!dma_mask)
2367 dma_mask = *dev->dma_mask;
2368
2369 spin_lock_irqsave(&domain->lock, flags);
2370
2371 *dma_addr = __map_single(dev, domain->priv, paddr,
2372 size, DMA_BIDIRECTIONAL, true, dma_mask);
2373
2374 if (*dma_addr == DMA_ERROR_CODE) {
2375 spin_unlock_irqrestore(&domain->lock, flags);
2376 goto out_free;
2377 }
2378
2379 domain_flush_complete(domain);
2380
2381 spin_unlock_irqrestore(&domain->lock, flags);
2382
2383 return virt_addr;
2384
2385out_free:
2386
2387 free_pages((unsigned long)virt_addr, get_order(size));
2388
2389 return NULL;
2390}
2391
2392/*
2393 * The exported free_coherent function for dma_ops.
2394 */
2395static void free_coherent(struct device *dev, size_t size,
2396 void *virt_addr, dma_addr_t dma_addr)
2397{
2398 unsigned long flags;
2399 struct protection_domain *domain;
2400
2401 INC_STATS_COUNTER(cnt_free_coherent);
2402
2403 domain = get_domain(dev);
2404 if (IS_ERR(domain))
2405 goto free_mem;
2406
2407 spin_lock_irqsave(&domain->lock, flags);
2408
2409 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
2410
2411 domain_flush_complete(domain);
2412
2413 spin_unlock_irqrestore(&domain->lock, flags);
2414
2415free_mem:
2416 free_pages((unsigned long)virt_addr, get_order(size));
2417}
2418
2419/*
2420 * This function is called by the DMA layer to find out if we can handle a
2421 * particular device. It is part of the dma_ops.
2422 */
2423static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2424{
2425 return check_device(dev);
2426}
2427
2428/*
2429 * The function for pre-allocating protection domains.
2430 *
2431 * If the driver core informs the DMA layer if a driver grabs a device
2432 * we don't need to preallocate the protection domains anymore.
2433 * For now we have to.
2434 */
2435static void prealloc_protection_domains(void)
2436{
2437 struct pci_dev *dev = NULL;
2438 struct dma_ops_domain *dma_dom;
2439 u16 devid;
2440
2441 for_each_pci_dev(dev) {
2442
2443 /* Do we handle this device? */
2444 if (!check_device(&dev->dev))
2445 continue;
2446
2447 /* Is there already any domain for it? */
2448 if (domain_for_device(&dev->dev))
2449 continue;
2450
2451 devid = get_device_id(&dev->dev);
2452
2453 dma_dom = dma_ops_domain_alloc();
2454 if (!dma_dom)
2455 continue;
2456 init_unity_mappings_for_device(dma_dom, devid);
2457 dma_dom->target_dev = devid;
2458
2459 attach_device(&dev->dev, &dma_dom->domain);
2460
2461 list_add_tail(&dma_dom->list, &iommu_pd_list);
2462 }
2463}
2464
2465static struct dma_map_ops amd_iommu_dma_ops = {
2466 .alloc_coherent = alloc_coherent,
2467 .free_coherent = free_coherent,
2468 .map_page = map_page,
2469 .unmap_page = unmap_page,
2470 .map_sg = map_sg,
2471 .unmap_sg = unmap_sg,
2472 .dma_supported = amd_iommu_dma_supported,
2473};
2474
2475static unsigned device_dma_ops_init(void)
2476{
2477 struct pci_dev *pdev = NULL;
2478 unsigned unhandled = 0;
2479
2480 for_each_pci_dev(pdev) {
2481 if (!check_device(&pdev->dev)) {
2482 unhandled += 1;
2483 continue;
2484 }
2485
2486 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
2487 }
2488
2489 return unhandled;
2490}
2491
2492/*
2493 * The function which clues the AMD IOMMU driver into dma_ops.
2494 */
2495
2496void __init amd_iommu_init_api(void)
2497{
2498 register_iommu(&amd_iommu_ops);
2499}
2500
2501int __init amd_iommu_init_dma_ops(void)
2502{
2503 struct amd_iommu *iommu;
2504 int ret, unhandled;
2505
2506 /*
2507 * first allocate a default protection domain for every IOMMU we
2508 * found in the system. Devices not assigned to any other
2509 * protection domain will be assigned to the default one.
2510 */
2511 for_each_iommu(iommu) {
2512 iommu->default_dom = dma_ops_domain_alloc();
2513 if (iommu->default_dom == NULL)
2514 return -ENOMEM;
2515 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
2516 ret = iommu_init_unity_mappings(iommu);
2517 if (ret)
2518 goto free_domains;
2519 }
2520
2521 /*
2522 * Pre-allocate the protection domains for each device.
2523 */
2524 prealloc_protection_domains();
2525
2526 iommu_detected = 1;
2527 swiotlb = 0;
2528
2529 /* Make the driver finally visible to the drivers */
2530 unhandled = device_dma_ops_init();
2531 if (unhandled && max_pfn > MAX_DMA32_PFN) {
2532 /* There are unhandled devices - initialize swiotlb for them */
2533 swiotlb = 1;
2534 }
2535
2536 amd_iommu_stats_init();
2537
2538 return 0;
2539
2540free_domains:
2541
2542 for_each_iommu(iommu) {
2543 if (iommu->default_dom)
2544 dma_ops_domain_free(iommu->default_dom);
2545 }
2546
2547 return ret;
2548}
2549
2550/*****************************************************************************
2551 *
2552 * The following functions belong to the exported interface of AMD IOMMU
2553 *
2554 * This interface allows access to lower level functions of the IOMMU
2555 * like protection domain handling and assignement of devices to domains
2556 * which is not possible with the dma_ops interface.
2557 *
2558 *****************************************************************************/
2559
2560static void cleanup_domain(struct protection_domain *domain)
2561{
2562 struct iommu_dev_data *dev_data, *next;
2563 unsigned long flags;
2564
2565 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2566
2567 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
2568 __detach_device(dev_data);
2569 atomic_set(&dev_data->bind, 0);
2570 }
2571
2572 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2573}
2574
2575static void protection_domain_free(struct protection_domain *domain)
2576{
2577 if (!domain)
2578 return;
2579
2580 del_domain_from_list(domain);
2581
2582 if (domain->id)
2583 domain_id_free(domain->id);
2584
2585 kfree(domain);
2586}
2587
2588static struct protection_domain *protection_domain_alloc(void)
2589{
2590 struct protection_domain *domain;
2591
2592 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2593 if (!domain)
2594 return NULL;
2595
2596 spin_lock_init(&domain->lock);
2597 mutex_init(&domain->api_lock);
2598 domain->id = domain_id_alloc();
2599 if (!domain->id)
2600 goto out_err;
2601 INIT_LIST_HEAD(&domain->dev_list);
2602
2603 add_domain_to_list(domain);
2604
2605 return domain;
2606
2607out_err:
2608 kfree(domain);
2609
2610 return NULL;
2611}
2612
2613static int amd_iommu_domain_init(struct iommu_domain *dom)
2614{
2615 struct protection_domain *domain;
2616
2617 domain = protection_domain_alloc();
2618 if (!domain)
2619 goto out_free;
2620
2621 domain->mode = PAGE_MODE_3_LEVEL;
2622 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2623 if (!domain->pt_root)
2624 goto out_free;
2625
2626 dom->priv = domain;
2627
2628 return 0;
2629
2630out_free:
2631 protection_domain_free(domain);
2632
2633 return -ENOMEM;
2634}
2635
2636static void amd_iommu_domain_destroy(struct iommu_domain *dom)
2637{
2638 struct protection_domain *domain = dom->priv;
2639
2640 if (!domain)
2641 return;
2642
2643 if (domain->dev_cnt > 0)
2644 cleanup_domain(domain);
2645
2646 BUG_ON(domain->dev_cnt != 0);
2647
2648 free_pagetable(domain);
2649
2650 protection_domain_free(domain);
2651
2652 dom->priv = NULL;
2653}
2654
2655static void amd_iommu_detach_device(struct iommu_domain *dom,
2656 struct device *dev)
2657{
2658 struct iommu_dev_data *dev_data = dev->archdata.iommu;
2659 struct amd_iommu *iommu;
2660 u16 devid;
2661
2662 if (!check_device(dev))
2663 return;
2664
2665 devid = get_device_id(dev);
2666
2667 if (dev_data->domain != NULL)
2668 detach_device(dev);
2669
2670 iommu = amd_iommu_rlookup_table[devid];
2671 if (!iommu)
2672 return;
2673
2674 iommu_completion_wait(iommu);
2675}
2676
2677static int amd_iommu_attach_device(struct iommu_domain *dom,
2678 struct device *dev)
2679{
2680 struct protection_domain *domain = dom->priv;
2681 struct iommu_dev_data *dev_data;
2682 struct amd_iommu *iommu;
2683 int ret;
2684
2685 if (!check_device(dev))
2686 return -EINVAL;
2687
2688 dev_data = dev->archdata.iommu;
2689
2690 iommu = amd_iommu_rlookup_table[dev_data->devid];
2691 if (!iommu)
2692 return -EINVAL;
2693
2694 if (dev_data->domain)
2695 detach_device(dev);
2696
2697 ret = attach_device(dev, domain);
2698
2699 iommu_completion_wait(iommu);
2700
2701 return ret;
2702}
2703
2704static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2705 phys_addr_t paddr, int gfp_order, int iommu_prot)
2706{
2707 unsigned long page_size = 0x1000UL << gfp_order;
2708 struct protection_domain *domain = dom->priv;
2709 int prot = 0;
2710 int ret;
2711
2712 if (iommu_prot & IOMMU_READ)
2713 prot |= IOMMU_PROT_IR;
2714 if (iommu_prot & IOMMU_WRITE)
2715 prot |= IOMMU_PROT_IW;
2716
2717 mutex_lock(&domain->api_lock);
2718 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
2719 mutex_unlock(&domain->api_lock);
2720
2721 return ret;
2722}
2723
2724static int amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2725 int gfp_order)
2726{
2727 struct protection_domain *domain = dom->priv;
2728 unsigned long page_size, unmap_size;
2729
2730 page_size = 0x1000UL << gfp_order;
2731
2732 mutex_lock(&domain->api_lock);
2733 unmap_size = iommu_unmap_page(domain, iova, page_size);
2734 mutex_unlock(&domain->api_lock);
2735
2736 domain_flush_tlb_pde(domain);
2737
2738 return get_order(unmap_size);
2739}
2740
2741static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2742 unsigned long iova)
2743{
2744 struct protection_domain *domain = dom->priv;
2745 unsigned long offset_mask;
2746 phys_addr_t paddr;
2747 u64 *pte, __pte;
2748
2749 pte = fetch_pte(domain, iova);
2750
2751 if (!pte || !IOMMU_PTE_PRESENT(*pte))
2752 return 0;
2753
2754 if (PM_PTE_LEVEL(*pte) == 0)
2755 offset_mask = PAGE_SIZE - 1;
2756 else
2757 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
2758
2759 __pte = *pte & PM_ADDR_MASK;
2760 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
2761
2762 return paddr;
2763}
2764
2765static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
2766 unsigned long cap)
2767{
2768 switch (cap) {
2769 case IOMMU_CAP_CACHE_COHERENCY:
2770 return 1;
2771 }
2772
2773 return 0;
2774}
2775
2776static struct iommu_ops amd_iommu_ops = {
2777 .domain_init = amd_iommu_domain_init,
2778 .domain_destroy = amd_iommu_domain_destroy,
2779 .attach_dev = amd_iommu_attach_device,
2780 .detach_dev = amd_iommu_detach_device,
2781 .map = amd_iommu_map,
2782 .unmap = amd_iommu_unmap,
2783 .iova_to_phys = amd_iommu_iova_to_phys,
2784 .domain_has_cap = amd_iommu_domain_has_cap,
2785};
2786
2787/*****************************************************************************
2788 *
2789 * The next functions do a basic initialization of IOMMU for pass through
2790 * mode
2791 *
2792 * In passthrough mode the IOMMU is initialized and enabled but not used for
2793 * DMA-API translation.
2794 *
2795 *****************************************************************************/
2796
2797int __init amd_iommu_init_passthrough(void)
2798{
2799 struct amd_iommu *iommu;
2800 struct pci_dev *dev = NULL;
2801 u16 devid;
2802
2803 /* allocate passthrough domain */
2804 pt_domain = protection_domain_alloc();
2805 if (!pt_domain)
2806 return -ENOMEM;
2807
2808 pt_domain->mode |= PAGE_MODE_NONE;
2809
2810 for_each_pci_dev(dev) {
2811 if (!check_device(&dev->dev))
2812 continue;
2813
2814 devid = get_device_id(&dev->dev);
2815
2816 iommu = amd_iommu_rlookup_table[devid];
2817 if (!iommu)
2818 continue;
2819
2820 attach_device(&dev->dev, pt_domain);
2821 }
2822
2823 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
2824
2825 return 0;
2826}
1/*
2 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
3 * Author: Joerg Roedel <jroedel@suse.de>
4 * Leo Duran <leo.duran@amd.com>
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
18 */
19
20#include <linux/ratelimit.h>
21#include <linux/pci.h>
22#include <linux/acpi.h>
23#include <linux/amba/bus.h>
24#include <linux/platform_device.h>
25#include <linux/pci-ats.h>
26#include <linux/bitmap.h>
27#include <linux/slab.h>
28#include <linux/debugfs.h>
29#include <linux/scatterlist.h>
30#include <linux/dma-mapping.h>
31#include <linux/iommu-helper.h>
32#include <linux/iommu.h>
33#include <linux/delay.h>
34#include <linux/amd-iommu.h>
35#include <linux/notifier.h>
36#include <linux/export.h>
37#include <linux/irq.h>
38#include <linux/msi.h>
39#include <linux/dma-contiguous.h>
40#include <linux/irqdomain.h>
41#include <linux/percpu.h>
42#include <linux/iova.h>
43#include <asm/irq_remapping.h>
44#include <asm/io_apic.h>
45#include <asm/apic.h>
46#include <asm/hw_irq.h>
47#include <asm/msidef.h>
48#include <asm/proto.h>
49#include <asm/iommu.h>
50#include <asm/gart.h>
51#include <asm/dma.h>
52
53#include "amd_iommu_proto.h"
54#include "amd_iommu_types.h"
55#include "irq_remapping.h"
56
57#define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
58
59#define LOOP_TIMEOUT 100000
60
61/* IO virtual address start page frame number */
62#define IOVA_START_PFN (1)
63#define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT)
64#define DMA_32BIT_PFN IOVA_PFN(DMA_BIT_MASK(32))
65
66/* Reserved IOVA ranges */
67#define MSI_RANGE_START (0xfee00000)
68#define MSI_RANGE_END (0xfeefffff)
69#define HT_RANGE_START (0xfd00000000ULL)
70#define HT_RANGE_END (0xffffffffffULL)
71
72/*
73 * This bitmap is used to advertise the page sizes our hardware support
74 * to the IOMMU core, which will then use this information to split
75 * physically contiguous memory regions it is mapping into page sizes
76 * that we support.
77 *
78 * 512GB Pages are not supported due to a hardware bug
79 */
80#define AMD_IOMMU_PGSIZES ((~0xFFFUL) & ~(2ULL << 38))
81
82static DEFINE_RWLOCK(amd_iommu_devtable_lock);
83
84/* List of all available dev_data structures */
85static LIST_HEAD(dev_data_list);
86static DEFINE_SPINLOCK(dev_data_list_lock);
87
88LIST_HEAD(ioapic_map);
89LIST_HEAD(hpet_map);
90LIST_HEAD(acpihid_map);
91
92#define FLUSH_QUEUE_SIZE 256
93
94struct flush_queue_entry {
95 unsigned long iova_pfn;
96 unsigned long pages;
97 struct dma_ops_domain *dma_dom;
98};
99
100struct flush_queue {
101 spinlock_t lock;
102 unsigned next;
103 struct flush_queue_entry *entries;
104};
105
106static DEFINE_PER_CPU(struct flush_queue, flush_queue);
107
108static atomic_t queue_timer_on;
109static struct timer_list queue_timer;
110
111/*
112 * Domain for untranslated devices - only allocated
113 * if iommu=pt passed on kernel cmd line.
114 */
115static const struct iommu_ops amd_iommu_ops;
116
117static ATOMIC_NOTIFIER_HEAD(ppr_notifier);
118int amd_iommu_max_glx_val = -1;
119
120static struct dma_map_ops amd_iommu_dma_ops;
121
122/*
123 * This struct contains device specific data for the IOMMU
124 */
125struct iommu_dev_data {
126 struct list_head list; /* For domain->dev_list */
127 struct list_head dev_data_list; /* For global dev_data_list */
128 struct protection_domain *domain; /* Domain the device is bound to */
129 u16 devid; /* PCI Device ID */
130 u16 alias; /* Alias Device ID */
131 bool iommu_v2; /* Device can make use of IOMMUv2 */
132 bool passthrough; /* Device is identity mapped */
133 struct {
134 bool enabled;
135 int qdep;
136 } ats; /* ATS state */
137 bool pri_tlp; /* PASID TLB required for
138 PPR completions */
139 u32 errata; /* Bitmap for errata to apply */
140 bool use_vapic; /* Enable device to use vapic mode */
141};
142
143/*
144 * general struct to manage commands send to an IOMMU
145 */
146struct iommu_cmd {
147 u32 data[4];
148};
149
150struct kmem_cache *amd_iommu_irq_cache;
151
152static void update_domain(struct protection_domain *domain);
153static int protection_domain_init(struct protection_domain *domain);
154static void detach_device(struct device *dev);
155
156/*
157 * Data container for a dma_ops specific protection domain
158 */
159struct dma_ops_domain {
160 /* generic protection domain information */
161 struct protection_domain domain;
162
163 /* IOVA RB-Tree */
164 struct iova_domain iovad;
165};
166
167static struct iova_domain reserved_iova_ranges;
168static struct lock_class_key reserved_rbtree_key;
169
170/****************************************************************************
171 *
172 * Helper functions
173 *
174 ****************************************************************************/
175
176static inline int match_hid_uid(struct device *dev,
177 struct acpihid_map_entry *entry)
178{
179 const char *hid, *uid;
180
181 hid = acpi_device_hid(ACPI_COMPANION(dev));
182 uid = acpi_device_uid(ACPI_COMPANION(dev));
183
184 if (!hid || !(*hid))
185 return -ENODEV;
186
187 if (!uid || !(*uid))
188 return strcmp(hid, entry->hid);
189
190 if (!(*entry->uid))
191 return strcmp(hid, entry->hid);
192
193 return (strcmp(hid, entry->hid) || strcmp(uid, entry->uid));
194}
195
196static inline u16 get_pci_device_id(struct device *dev)
197{
198 struct pci_dev *pdev = to_pci_dev(dev);
199
200 return PCI_DEVID(pdev->bus->number, pdev->devfn);
201}
202
203static inline int get_acpihid_device_id(struct device *dev,
204 struct acpihid_map_entry **entry)
205{
206 struct acpihid_map_entry *p;
207
208 list_for_each_entry(p, &acpihid_map, list) {
209 if (!match_hid_uid(dev, p)) {
210 if (entry)
211 *entry = p;
212 return p->devid;
213 }
214 }
215 return -EINVAL;
216}
217
218static inline int get_device_id(struct device *dev)
219{
220 int devid;
221
222 if (dev_is_pci(dev))
223 devid = get_pci_device_id(dev);
224 else
225 devid = get_acpihid_device_id(dev, NULL);
226
227 return devid;
228}
229
230static struct protection_domain *to_pdomain(struct iommu_domain *dom)
231{
232 return container_of(dom, struct protection_domain, domain);
233}
234
235static struct dma_ops_domain* to_dma_ops_domain(struct protection_domain *domain)
236{
237 BUG_ON(domain->flags != PD_DMA_OPS_MASK);
238 return container_of(domain, struct dma_ops_domain, domain);
239}
240
241static struct iommu_dev_data *alloc_dev_data(u16 devid)
242{
243 struct iommu_dev_data *dev_data;
244 unsigned long flags;
245
246 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL);
247 if (!dev_data)
248 return NULL;
249
250 dev_data->devid = devid;
251
252 spin_lock_irqsave(&dev_data_list_lock, flags);
253 list_add_tail(&dev_data->dev_data_list, &dev_data_list);
254 spin_unlock_irqrestore(&dev_data_list_lock, flags);
255
256 return dev_data;
257}
258
259static struct iommu_dev_data *search_dev_data(u16 devid)
260{
261 struct iommu_dev_data *dev_data;
262 unsigned long flags;
263
264 spin_lock_irqsave(&dev_data_list_lock, flags);
265 list_for_each_entry(dev_data, &dev_data_list, dev_data_list) {
266 if (dev_data->devid == devid)
267 goto out_unlock;
268 }
269
270 dev_data = NULL;
271
272out_unlock:
273 spin_unlock_irqrestore(&dev_data_list_lock, flags);
274
275 return dev_data;
276}
277
278static int __last_alias(struct pci_dev *pdev, u16 alias, void *data)
279{
280 *(u16 *)data = alias;
281 return 0;
282}
283
284static u16 get_alias(struct device *dev)
285{
286 struct pci_dev *pdev = to_pci_dev(dev);
287 u16 devid, ivrs_alias, pci_alias;
288
289 /* The callers make sure that get_device_id() does not fail here */
290 devid = get_device_id(dev);
291 ivrs_alias = amd_iommu_alias_table[devid];
292 pci_for_each_dma_alias(pdev, __last_alias, &pci_alias);
293
294 if (ivrs_alias == pci_alias)
295 return ivrs_alias;
296
297 /*
298 * DMA alias showdown
299 *
300 * The IVRS is fairly reliable in telling us about aliases, but it
301 * can't know about every screwy device. If we don't have an IVRS
302 * reported alias, use the PCI reported alias. In that case we may
303 * still need to initialize the rlookup and dev_table entries if the
304 * alias is to a non-existent device.
305 */
306 if (ivrs_alias == devid) {
307 if (!amd_iommu_rlookup_table[pci_alias]) {
308 amd_iommu_rlookup_table[pci_alias] =
309 amd_iommu_rlookup_table[devid];
310 memcpy(amd_iommu_dev_table[pci_alias].data,
311 amd_iommu_dev_table[devid].data,
312 sizeof(amd_iommu_dev_table[pci_alias].data));
313 }
314
315 return pci_alias;
316 }
317
318 pr_info("AMD-Vi: Using IVRS reported alias %02x:%02x.%d "
319 "for device %s[%04x:%04x], kernel reported alias "
320 "%02x:%02x.%d\n", PCI_BUS_NUM(ivrs_alias), PCI_SLOT(ivrs_alias),
321 PCI_FUNC(ivrs_alias), dev_name(dev), pdev->vendor, pdev->device,
322 PCI_BUS_NUM(pci_alias), PCI_SLOT(pci_alias),
323 PCI_FUNC(pci_alias));
324
325 /*
326 * If we don't have a PCI DMA alias and the IVRS alias is on the same
327 * bus, then the IVRS table may know about a quirk that we don't.
328 */
329 if (pci_alias == devid &&
330 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) {
331 pci_add_dma_alias(pdev, ivrs_alias & 0xff);
332 pr_info("AMD-Vi: Added PCI DMA alias %02x.%d for %s\n",
333 PCI_SLOT(ivrs_alias), PCI_FUNC(ivrs_alias),
334 dev_name(dev));
335 }
336
337 return ivrs_alias;
338}
339
340static struct iommu_dev_data *find_dev_data(u16 devid)
341{
342 struct iommu_dev_data *dev_data;
343
344 dev_data = search_dev_data(devid);
345
346 if (dev_data == NULL)
347 dev_data = alloc_dev_data(devid);
348
349 return dev_data;
350}
351
352static struct iommu_dev_data *get_dev_data(struct device *dev)
353{
354 return dev->archdata.iommu;
355}
356
357/*
358* Find or create an IOMMU group for a acpihid device.
359*/
360static struct iommu_group *acpihid_device_group(struct device *dev)
361{
362 struct acpihid_map_entry *p, *entry = NULL;
363 int devid;
364
365 devid = get_acpihid_device_id(dev, &entry);
366 if (devid < 0)
367 return ERR_PTR(devid);
368
369 list_for_each_entry(p, &acpihid_map, list) {
370 if ((devid == p->devid) && p->group)
371 entry->group = p->group;
372 }
373
374 if (!entry->group)
375 entry->group = generic_device_group(dev);
376 else
377 iommu_group_ref_get(entry->group);
378
379 return entry->group;
380}
381
382static bool pci_iommuv2_capable(struct pci_dev *pdev)
383{
384 static const int caps[] = {
385 PCI_EXT_CAP_ID_ATS,
386 PCI_EXT_CAP_ID_PRI,
387 PCI_EXT_CAP_ID_PASID,
388 };
389 int i, pos;
390
391 for (i = 0; i < 3; ++i) {
392 pos = pci_find_ext_capability(pdev, caps[i]);
393 if (pos == 0)
394 return false;
395 }
396
397 return true;
398}
399
400static bool pdev_pri_erratum(struct pci_dev *pdev, u32 erratum)
401{
402 struct iommu_dev_data *dev_data;
403
404 dev_data = get_dev_data(&pdev->dev);
405
406 return dev_data->errata & (1 << erratum) ? true : false;
407}
408
409/*
410 * This function checks if the driver got a valid device from the caller to
411 * avoid dereferencing invalid pointers.
412 */
413static bool check_device(struct device *dev)
414{
415 int devid;
416
417 if (!dev || !dev->dma_mask)
418 return false;
419
420 devid = get_device_id(dev);
421 if (devid < 0)
422 return false;
423
424 /* Out of our scope? */
425 if (devid > amd_iommu_last_bdf)
426 return false;
427
428 if (amd_iommu_rlookup_table[devid] == NULL)
429 return false;
430
431 return true;
432}
433
434static void init_iommu_group(struct device *dev)
435{
436 struct iommu_group *group;
437
438 group = iommu_group_get_for_dev(dev);
439 if (IS_ERR(group))
440 return;
441
442 iommu_group_put(group);
443}
444
445static int iommu_init_device(struct device *dev)
446{
447 struct iommu_dev_data *dev_data;
448 int devid;
449
450 if (dev->archdata.iommu)
451 return 0;
452
453 devid = get_device_id(dev);
454 if (devid < 0)
455 return devid;
456
457 dev_data = find_dev_data(devid);
458 if (!dev_data)
459 return -ENOMEM;
460
461 dev_data->alias = get_alias(dev);
462
463 if (dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) {
464 struct amd_iommu *iommu;
465
466 iommu = amd_iommu_rlookup_table[dev_data->devid];
467 dev_data->iommu_v2 = iommu->is_iommu_v2;
468 }
469
470 dev->archdata.iommu = dev_data;
471
472 iommu_device_link(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
473 dev);
474
475 return 0;
476}
477
478static void iommu_ignore_device(struct device *dev)
479{
480 u16 alias;
481 int devid;
482
483 devid = get_device_id(dev);
484 if (devid < 0)
485 return;
486
487 alias = get_alias(dev);
488
489 memset(&amd_iommu_dev_table[devid], 0, sizeof(struct dev_table_entry));
490 memset(&amd_iommu_dev_table[alias], 0, sizeof(struct dev_table_entry));
491
492 amd_iommu_rlookup_table[devid] = NULL;
493 amd_iommu_rlookup_table[alias] = NULL;
494}
495
496static void iommu_uninit_device(struct device *dev)
497{
498 int devid;
499 struct iommu_dev_data *dev_data;
500
501 devid = get_device_id(dev);
502 if (devid < 0)
503 return;
504
505 dev_data = search_dev_data(devid);
506 if (!dev_data)
507 return;
508
509 if (dev_data->domain)
510 detach_device(dev);
511
512 iommu_device_unlink(amd_iommu_rlookup_table[dev_data->devid]->iommu_dev,
513 dev);
514
515 iommu_group_remove_device(dev);
516
517 /* Remove dma-ops */
518 dev->archdata.dma_ops = NULL;
519
520 /*
521 * We keep dev_data around for unplugged devices and reuse it when the
522 * device is re-plugged - not doing so would introduce a ton of races.
523 */
524}
525
526/****************************************************************************
527 *
528 * Interrupt handling functions
529 *
530 ****************************************************************************/
531
532static void dump_dte_entry(u16 devid)
533{
534 int i;
535
536 for (i = 0; i < 4; ++i)
537 pr_err("AMD-Vi: DTE[%d]: %016llx\n", i,
538 amd_iommu_dev_table[devid].data[i]);
539}
540
541static void dump_command(unsigned long phys_addr)
542{
543 struct iommu_cmd *cmd = phys_to_virt(phys_addr);
544 int i;
545
546 for (i = 0; i < 4; ++i)
547 pr_err("AMD-Vi: CMD[%d]: %08x\n", i, cmd->data[i]);
548}
549
550static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
551{
552 int type, devid, domid, flags;
553 volatile u32 *event = __evt;
554 int count = 0;
555 u64 address;
556
557retry:
558 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK;
559 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK;
560 domid = (event[1] >> EVENT_DOMID_SHIFT) & EVENT_DOMID_MASK;
561 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK;
562 address = (u64)(((u64)event[3]) << 32) | event[2];
563
564 if (type == 0) {
565 /* Did we hit the erratum? */
566 if (++count == LOOP_TIMEOUT) {
567 pr_err("AMD-Vi: No event written to event log\n");
568 return;
569 }
570 udelay(1);
571 goto retry;
572 }
573
574 printk(KERN_ERR "AMD-Vi: Event logged [");
575
576 switch (type) {
577 case EVENT_TYPE_ILL_DEV:
578 printk("ILLEGAL_DEV_TABLE_ENTRY device=%02x:%02x.%x "
579 "address=0x%016llx flags=0x%04x]\n",
580 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
581 address, flags);
582 dump_dte_entry(devid);
583 break;
584 case EVENT_TYPE_IO_FAULT:
585 printk("IO_PAGE_FAULT device=%02x:%02x.%x "
586 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
587 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
588 domid, address, flags);
589 break;
590 case EVENT_TYPE_DEV_TAB_ERR:
591 printk("DEV_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
592 "address=0x%016llx flags=0x%04x]\n",
593 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
594 address, flags);
595 break;
596 case EVENT_TYPE_PAGE_TAB_ERR:
597 printk("PAGE_TAB_HARDWARE_ERROR device=%02x:%02x.%x "
598 "domain=0x%04x address=0x%016llx flags=0x%04x]\n",
599 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
600 domid, address, flags);
601 break;
602 case EVENT_TYPE_ILL_CMD:
603 printk("ILLEGAL_COMMAND_ERROR address=0x%016llx]\n", address);
604 dump_command(address);
605 break;
606 case EVENT_TYPE_CMD_HARD_ERR:
607 printk("COMMAND_HARDWARE_ERROR address=0x%016llx "
608 "flags=0x%04x]\n", address, flags);
609 break;
610 case EVENT_TYPE_IOTLB_INV_TO:
611 printk("IOTLB_INV_TIMEOUT device=%02x:%02x.%x "
612 "address=0x%016llx]\n",
613 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
614 address);
615 break;
616 case EVENT_TYPE_INV_DEV_REQ:
617 printk("INVALID_DEVICE_REQUEST device=%02x:%02x.%x "
618 "address=0x%016llx flags=0x%04x]\n",
619 PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid),
620 address, flags);
621 break;
622 default:
623 printk(KERN_ERR "UNKNOWN type=0x%02x]\n", type);
624 }
625
626 memset(__evt, 0, 4 * sizeof(u32));
627}
628
629static void iommu_poll_events(struct amd_iommu *iommu)
630{
631 u32 head, tail;
632
633 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
634 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET);
635
636 while (head != tail) {
637 iommu_print_event(iommu, iommu->evt_buf + head);
638 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE;
639 }
640
641 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET);
642}
643
644static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw)
645{
646 struct amd_iommu_fault fault;
647
648 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) {
649 pr_err_ratelimited("AMD-Vi: Unknown PPR request received\n");
650 return;
651 }
652
653 fault.address = raw[1];
654 fault.pasid = PPR_PASID(raw[0]);
655 fault.device_id = PPR_DEVID(raw[0]);
656 fault.tag = PPR_TAG(raw[0]);
657 fault.flags = PPR_FLAGS(raw[0]);
658
659 atomic_notifier_call_chain(&ppr_notifier, 0, &fault);
660}
661
662static void iommu_poll_ppr_log(struct amd_iommu *iommu)
663{
664 u32 head, tail;
665
666 if (iommu->ppr_log == NULL)
667 return;
668
669 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
670 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
671
672 while (head != tail) {
673 volatile u64 *raw;
674 u64 entry[2];
675 int i;
676
677 raw = (u64 *)(iommu->ppr_log + head);
678
679 /*
680 * Hardware bug: Interrupt may arrive before the entry is
681 * written to memory. If this happens we need to wait for the
682 * entry to arrive.
683 */
684 for (i = 0; i < LOOP_TIMEOUT; ++i) {
685 if (PPR_REQ_TYPE(raw[0]) != 0)
686 break;
687 udelay(1);
688 }
689
690 /* Avoid memcpy function-call overhead */
691 entry[0] = raw[0];
692 entry[1] = raw[1];
693
694 /*
695 * To detect the hardware bug we need to clear the entry
696 * back to zero.
697 */
698 raw[0] = raw[1] = 0UL;
699
700 /* Update head pointer of hardware ring-buffer */
701 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE;
702 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
703
704 /* Handle PPR entry */
705 iommu_handle_ppr_entry(iommu, entry);
706
707 /* Refresh ring-buffer information */
708 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET);
709 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET);
710 }
711}
712
713#ifdef CONFIG_IRQ_REMAP
714static int (*iommu_ga_log_notifier)(u32);
715
716int amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
717{
718 iommu_ga_log_notifier = notifier;
719
720 return 0;
721}
722EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier);
723
724static void iommu_poll_ga_log(struct amd_iommu *iommu)
725{
726 u32 head, tail, cnt = 0;
727
728 if (iommu->ga_log == NULL)
729 return;
730
731 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
732 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET);
733
734 while (head != tail) {
735 volatile u64 *raw;
736 u64 log_entry;
737
738 raw = (u64 *)(iommu->ga_log + head);
739 cnt++;
740
741 /* Avoid memcpy function-call overhead */
742 log_entry = *raw;
743
744 /* Update head pointer of hardware ring-buffer */
745 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE;
746 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET);
747
748 /* Handle GA entry */
749 switch (GA_REQ_TYPE(log_entry)) {
750 case GA_GUEST_NR:
751 if (!iommu_ga_log_notifier)
752 break;
753
754 pr_debug("AMD-Vi: %s: devid=%#x, ga_tag=%#x\n",
755 __func__, GA_DEVID(log_entry),
756 GA_TAG(log_entry));
757
758 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0)
759 pr_err("AMD-Vi: GA log notifier failed.\n");
760 break;
761 default:
762 break;
763 }
764 }
765}
766#endif /* CONFIG_IRQ_REMAP */
767
768#define AMD_IOMMU_INT_MASK \
769 (MMIO_STATUS_EVT_INT_MASK | \
770 MMIO_STATUS_PPR_INT_MASK | \
771 MMIO_STATUS_GALOG_INT_MASK)
772
773irqreturn_t amd_iommu_int_thread(int irq, void *data)
774{
775 struct amd_iommu *iommu = (struct amd_iommu *) data;
776 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
777
778 while (status & AMD_IOMMU_INT_MASK) {
779 /* Enable EVT and PPR and GA interrupts again */
780 writel(AMD_IOMMU_INT_MASK,
781 iommu->mmio_base + MMIO_STATUS_OFFSET);
782
783 if (status & MMIO_STATUS_EVT_INT_MASK) {
784 pr_devel("AMD-Vi: Processing IOMMU Event Log\n");
785 iommu_poll_events(iommu);
786 }
787
788 if (status & MMIO_STATUS_PPR_INT_MASK) {
789 pr_devel("AMD-Vi: Processing IOMMU PPR Log\n");
790 iommu_poll_ppr_log(iommu);
791 }
792
793#ifdef CONFIG_IRQ_REMAP
794 if (status & MMIO_STATUS_GALOG_INT_MASK) {
795 pr_devel("AMD-Vi: Processing IOMMU GA Log\n");
796 iommu_poll_ga_log(iommu);
797 }
798#endif
799
800 /*
801 * Hardware bug: ERBT1312
802 * When re-enabling interrupt (by writing 1
803 * to clear the bit), the hardware might also try to set
804 * the interrupt bit in the event status register.
805 * In this scenario, the bit will be set, and disable
806 * subsequent interrupts.
807 *
808 * Workaround: The IOMMU driver should read back the
809 * status register and check if the interrupt bits are cleared.
810 * If not, driver will need to go through the interrupt handler
811 * again and re-clear the bits
812 */
813 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET);
814 }
815 return IRQ_HANDLED;
816}
817
818irqreturn_t amd_iommu_int_handler(int irq, void *data)
819{
820 return IRQ_WAKE_THREAD;
821}
822
823/****************************************************************************
824 *
825 * IOMMU command queuing functions
826 *
827 ****************************************************************************/
828
829static int wait_on_sem(volatile u64 *sem)
830{
831 int i = 0;
832
833 while (*sem == 0 && i < LOOP_TIMEOUT) {
834 udelay(1);
835 i += 1;
836 }
837
838 if (i == LOOP_TIMEOUT) {
839 pr_alert("AMD-Vi: Completion-Wait loop timed out\n");
840 return -EIO;
841 }
842
843 return 0;
844}
845
846static void copy_cmd_to_buffer(struct amd_iommu *iommu,
847 struct iommu_cmd *cmd,
848 u32 tail)
849{
850 u8 *target;
851
852 target = iommu->cmd_buf + tail;
853 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
854
855 /* Copy command to buffer */
856 memcpy(target, cmd, sizeof(*cmd));
857
858 /* Tell the IOMMU about it */
859 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
860}
861
862static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
863{
864 WARN_ON(address & 0x7ULL);
865
866 memset(cmd, 0, sizeof(*cmd));
867 cmd->data[0] = lower_32_bits(__pa(address)) | CMD_COMPL_WAIT_STORE_MASK;
868 cmd->data[1] = upper_32_bits(__pa(address));
869 cmd->data[2] = 1;
870 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
871}
872
873static void build_inv_dte(struct iommu_cmd *cmd, u16 devid)
874{
875 memset(cmd, 0, sizeof(*cmd));
876 cmd->data[0] = devid;
877 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY);
878}
879
880static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address,
881 size_t size, u16 domid, int pde)
882{
883 u64 pages;
884 bool s;
885
886 pages = iommu_num_pages(address, size, PAGE_SIZE);
887 s = false;
888
889 if (pages > 1) {
890 /*
891 * If we have to flush more than one page, flush all
892 * TLB entries for this domain
893 */
894 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
895 s = true;
896 }
897
898 address &= PAGE_MASK;
899
900 memset(cmd, 0, sizeof(*cmd));
901 cmd->data[1] |= domid;
902 cmd->data[2] = lower_32_bits(address);
903 cmd->data[3] = upper_32_bits(address);
904 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
905 if (s) /* size bit - we flush more than one 4kb page */
906 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
907 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */
908 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
909}
910
911static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep,
912 u64 address, size_t size)
913{
914 u64 pages;
915 bool s;
916
917 pages = iommu_num_pages(address, size, PAGE_SIZE);
918 s = false;
919
920 if (pages > 1) {
921 /*
922 * If we have to flush more than one page, flush all
923 * TLB entries for this domain
924 */
925 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS;
926 s = true;
927 }
928
929 address &= PAGE_MASK;
930
931 memset(cmd, 0, sizeof(*cmd));
932 cmd->data[0] = devid;
933 cmd->data[0] |= (qdep & 0xff) << 24;
934 cmd->data[1] = devid;
935 cmd->data[2] = lower_32_bits(address);
936 cmd->data[3] = upper_32_bits(address);
937 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
938 if (s)
939 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
940}
941
942static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, int pasid,
943 u64 address, bool size)
944{
945 memset(cmd, 0, sizeof(*cmd));
946
947 address &= ~(0xfffULL);
948
949 cmd->data[0] = pasid;
950 cmd->data[1] = domid;
951 cmd->data[2] = lower_32_bits(address);
952 cmd->data[3] = upper_32_bits(address);
953 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK;
954 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
955 if (size)
956 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
957 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES);
958}
959
960static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, int pasid,
961 int qdep, u64 address, bool size)
962{
963 memset(cmd, 0, sizeof(*cmd));
964
965 address &= ~(0xfffULL);
966
967 cmd->data[0] = devid;
968 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16;
969 cmd->data[0] |= (qdep & 0xff) << 24;
970 cmd->data[1] = devid;
971 cmd->data[1] |= (pasid & 0xff) << 16;
972 cmd->data[2] = lower_32_bits(address);
973 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK;
974 cmd->data[3] = upper_32_bits(address);
975 if (size)
976 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK;
977 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES);
978}
979
980static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, int pasid,
981 int status, int tag, bool gn)
982{
983 memset(cmd, 0, sizeof(*cmd));
984
985 cmd->data[0] = devid;
986 if (gn) {
987 cmd->data[1] = pasid;
988 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK;
989 }
990 cmd->data[3] = tag & 0x1ff;
991 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT;
992
993 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR);
994}
995
996static void build_inv_all(struct iommu_cmd *cmd)
997{
998 memset(cmd, 0, sizeof(*cmd));
999 CMD_SET_TYPE(cmd, CMD_INV_ALL);
1000}
1001
1002static void build_inv_irt(struct iommu_cmd *cmd, u16 devid)
1003{
1004 memset(cmd, 0, sizeof(*cmd));
1005 cmd->data[0] = devid;
1006 CMD_SET_TYPE(cmd, CMD_INV_IRT);
1007}
1008
1009/*
1010 * Writes the command to the IOMMUs command buffer and informs the
1011 * hardware about the new command.
1012 */
1013static int __iommu_queue_command_sync(struct amd_iommu *iommu,
1014 struct iommu_cmd *cmd,
1015 bool sync)
1016{
1017 u32 left, tail, head, next_tail;
1018
1019again:
1020
1021 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
1022 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
1023 next_tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE;
1024 left = (head - next_tail) % CMD_BUFFER_SIZE;
1025
1026 if (left <= 0x20) {
1027 struct iommu_cmd sync_cmd;
1028 int ret;
1029
1030 iommu->cmd_sem = 0;
1031
1032 build_completion_wait(&sync_cmd, (u64)&iommu->cmd_sem);
1033 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
1034
1035 if ((ret = wait_on_sem(&iommu->cmd_sem)) != 0)
1036 return ret;
1037
1038 goto again;
1039 }
1040
1041 copy_cmd_to_buffer(iommu, cmd, tail);
1042
1043 /* We need to sync now to make sure all commands are processed */
1044 iommu->need_sync = sync;
1045
1046 return 0;
1047}
1048
1049static int iommu_queue_command_sync(struct amd_iommu *iommu,
1050 struct iommu_cmd *cmd,
1051 bool sync)
1052{
1053 unsigned long flags;
1054 int ret;
1055
1056 spin_lock_irqsave(&iommu->lock, flags);
1057 ret = __iommu_queue_command_sync(iommu, cmd, sync);
1058 spin_unlock_irqrestore(&iommu->lock, flags);
1059
1060 return ret;
1061}
1062
1063static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
1064{
1065 return iommu_queue_command_sync(iommu, cmd, true);
1066}
1067
1068/*
1069 * This function queues a completion wait command into the command
1070 * buffer of an IOMMU
1071 */
1072static int iommu_completion_wait(struct amd_iommu *iommu)
1073{
1074 struct iommu_cmd cmd;
1075 unsigned long flags;
1076 int ret;
1077
1078 if (!iommu->need_sync)
1079 return 0;
1080
1081
1082 build_completion_wait(&cmd, (u64)&iommu->cmd_sem);
1083
1084 spin_lock_irqsave(&iommu->lock, flags);
1085
1086 iommu->cmd_sem = 0;
1087
1088 ret = __iommu_queue_command_sync(iommu, &cmd, false);
1089 if (ret)
1090 goto out_unlock;
1091
1092 ret = wait_on_sem(&iommu->cmd_sem);
1093
1094out_unlock:
1095 spin_unlock_irqrestore(&iommu->lock, flags);
1096
1097 return ret;
1098}
1099
1100static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
1101{
1102 struct iommu_cmd cmd;
1103
1104 build_inv_dte(&cmd, devid);
1105
1106 return iommu_queue_command(iommu, &cmd);
1107}
1108
1109static void iommu_flush_dte_all(struct amd_iommu *iommu)
1110{
1111 u32 devid;
1112
1113 for (devid = 0; devid <= 0xffff; ++devid)
1114 iommu_flush_dte(iommu, devid);
1115
1116 iommu_completion_wait(iommu);
1117}
1118
1119/*
1120 * This function uses heavy locking and may disable irqs for some time. But
1121 * this is no issue because it is only called during resume.
1122 */
1123static void iommu_flush_tlb_all(struct amd_iommu *iommu)
1124{
1125 u32 dom_id;
1126
1127 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
1128 struct iommu_cmd cmd;
1129 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
1130 dom_id, 1);
1131 iommu_queue_command(iommu, &cmd);
1132 }
1133
1134 iommu_completion_wait(iommu);
1135}
1136
1137static void iommu_flush_all(struct amd_iommu *iommu)
1138{
1139 struct iommu_cmd cmd;
1140
1141 build_inv_all(&cmd);
1142
1143 iommu_queue_command(iommu, &cmd);
1144 iommu_completion_wait(iommu);
1145}
1146
1147static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid)
1148{
1149 struct iommu_cmd cmd;
1150
1151 build_inv_irt(&cmd, devid);
1152
1153 iommu_queue_command(iommu, &cmd);
1154}
1155
1156static void iommu_flush_irt_all(struct amd_iommu *iommu)
1157{
1158 u32 devid;
1159
1160 for (devid = 0; devid <= MAX_DEV_TABLE_ENTRIES; devid++)
1161 iommu_flush_irt(iommu, devid);
1162
1163 iommu_completion_wait(iommu);
1164}
1165
1166void iommu_flush_all_caches(struct amd_iommu *iommu)
1167{
1168 if (iommu_feature(iommu, FEATURE_IA)) {
1169 iommu_flush_all(iommu);
1170 } else {
1171 iommu_flush_dte_all(iommu);
1172 iommu_flush_irt_all(iommu);
1173 iommu_flush_tlb_all(iommu);
1174 }
1175}
1176
1177/*
1178 * Command send function for flushing on-device TLB
1179 */
1180static int device_flush_iotlb(struct iommu_dev_data *dev_data,
1181 u64 address, size_t size)
1182{
1183 struct amd_iommu *iommu;
1184 struct iommu_cmd cmd;
1185 int qdep;
1186
1187 qdep = dev_data->ats.qdep;
1188 iommu = amd_iommu_rlookup_table[dev_data->devid];
1189
1190 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
1191
1192 return iommu_queue_command(iommu, &cmd);
1193}
1194
1195/*
1196 * Command send function for invalidating a device table entry
1197 */
1198static int device_flush_dte(struct iommu_dev_data *dev_data)
1199{
1200 struct amd_iommu *iommu;
1201 u16 alias;
1202 int ret;
1203
1204 iommu = amd_iommu_rlookup_table[dev_data->devid];
1205 alias = dev_data->alias;
1206
1207 ret = iommu_flush_dte(iommu, dev_data->devid);
1208 if (!ret && alias != dev_data->devid)
1209 ret = iommu_flush_dte(iommu, alias);
1210 if (ret)
1211 return ret;
1212
1213 if (dev_data->ats.enabled)
1214 ret = device_flush_iotlb(dev_data, 0, ~0UL);
1215
1216 return ret;
1217}
1218
1219/*
1220 * TLB invalidation function which is called from the mapping functions.
1221 * It invalidates a single PTE if the range to flush is within a single
1222 * page. Otherwise it flushes the whole TLB of the IOMMU.
1223 */
1224static void __domain_flush_pages(struct protection_domain *domain,
1225 u64 address, size_t size, int pde)
1226{
1227 struct iommu_dev_data *dev_data;
1228 struct iommu_cmd cmd;
1229 int ret = 0, i;
1230
1231 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
1232
1233 for (i = 0; i < amd_iommus_present; ++i) {
1234 if (!domain->dev_iommu[i])
1235 continue;
1236
1237 /*
1238 * Devices of this domain are behind this IOMMU
1239 * We need a TLB flush
1240 */
1241 ret |= iommu_queue_command(amd_iommus[i], &cmd);
1242 }
1243
1244 list_for_each_entry(dev_data, &domain->dev_list, list) {
1245
1246 if (!dev_data->ats.enabled)
1247 continue;
1248
1249 ret |= device_flush_iotlb(dev_data, address, size);
1250 }
1251
1252 WARN_ON(ret);
1253}
1254
1255static void domain_flush_pages(struct protection_domain *domain,
1256 u64 address, size_t size)
1257{
1258 __domain_flush_pages(domain, address, size, 0);
1259}
1260
1261/* Flush the whole IO/TLB for a given protection domain */
1262static void domain_flush_tlb(struct protection_domain *domain)
1263{
1264 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
1265}
1266
1267/* Flush the whole IO/TLB for a given protection domain - including PDE */
1268static void domain_flush_tlb_pde(struct protection_domain *domain)
1269{
1270 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
1271}
1272
1273static void domain_flush_complete(struct protection_domain *domain)
1274{
1275 int i;
1276
1277 for (i = 0; i < amd_iommus_present; ++i) {
1278 if (domain && !domain->dev_iommu[i])
1279 continue;
1280
1281 /*
1282 * Devices of this domain are behind this IOMMU
1283 * We need to wait for completion of all commands.
1284 */
1285 iommu_completion_wait(amd_iommus[i]);
1286 }
1287}
1288
1289
1290/*
1291 * This function flushes the DTEs for all devices in domain
1292 */
1293static void domain_flush_devices(struct protection_domain *domain)
1294{
1295 struct iommu_dev_data *dev_data;
1296
1297 list_for_each_entry(dev_data, &domain->dev_list, list)
1298 device_flush_dte(dev_data);
1299}
1300
1301/****************************************************************************
1302 *
1303 * The functions below are used the create the page table mappings for
1304 * unity mapped regions.
1305 *
1306 ****************************************************************************/
1307
1308/*
1309 * This function is used to add another level to an IO page table. Adding
1310 * another level increases the size of the address space by 9 bits to a size up
1311 * to 64 bits.
1312 */
1313static bool increase_address_space(struct protection_domain *domain,
1314 gfp_t gfp)
1315{
1316 u64 *pte;
1317
1318 if (domain->mode == PAGE_MODE_6_LEVEL)
1319 /* address space already 64 bit large */
1320 return false;
1321
1322 pte = (void *)get_zeroed_page(gfp);
1323 if (!pte)
1324 return false;
1325
1326 *pte = PM_LEVEL_PDE(domain->mode,
1327 virt_to_phys(domain->pt_root));
1328 domain->pt_root = pte;
1329 domain->mode += 1;
1330 domain->updated = true;
1331
1332 return true;
1333}
1334
1335static u64 *alloc_pte(struct protection_domain *domain,
1336 unsigned long address,
1337 unsigned long page_size,
1338 u64 **pte_page,
1339 gfp_t gfp)
1340{
1341 int level, end_lvl;
1342 u64 *pte, *page;
1343
1344 BUG_ON(!is_power_of_2(page_size));
1345
1346 while (address > PM_LEVEL_SIZE(domain->mode))
1347 increase_address_space(domain, gfp);
1348
1349 level = domain->mode - 1;
1350 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1351 address = PAGE_SIZE_ALIGN(address, page_size);
1352 end_lvl = PAGE_SIZE_LEVEL(page_size);
1353
1354 while (level > end_lvl) {
1355 u64 __pte, __npte;
1356
1357 __pte = *pte;
1358
1359 if (!IOMMU_PTE_PRESENT(__pte)) {
1360 page = (u64 *)get_zeroed_page(gfp);
1361 if (!page)
1362 return NULL;
1363
1364 __npte = PM_LEVEL_PDE(level, virt_to_phys(page));
1365
1366 /* pte could have been changed somewhere. */
1367 if (cmpxchg64(pte, __pte, __npte) != __pte) {
1368 free_page((unsigned long)page);
1369 continue;
1370 }
1371 }
1372
1373 /* No level skipping support yet */
1374 if (PM_PTE_LEVEL(*pte) != level)
1375 return NULL;
1376
1377 level -= 1;
1378
1379 pte = IOMMU_PTE_PAGE(*pte);
1380
1381 if (pte_page && level == end_lvl)
1382 *pte_page = pte;
1383
1384 pte = &pte[PM_LEVEL_INDEX(level, address)];
1385 }
1386
1387 return pte;
1388}
1389
1390/*
1391 * This function checks if there is a PTE for a given dma address. If
1392 * there is one, it returns the pointer to it.
1393 */
1394static u64 *fetch_pte(struct protection_domain *domain,
1395 unsigned long address,
1396 unsigned long *page_size)
1397{
1398 int level;
1399 u64 *pte;
1400
1401 if (address > PM_LEVEL_SIZE(domain->mode))
1402 return NULL;
1403
1404 level = domain->mode - 1;
1405 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
1406 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1407
1408 while (level > 0) {
1409
1410 /* Not Present */
1411 if (!IOMMU_PTE_PRESENT(*pte))
1412 return NULL;
1413
1414 /* Large PTE */
1415 if (PM_PTE_LEVEL(*pte) == 7 ||
1416 PM_PTE_LEVEL(*pte) == 0)
1417 break;
1418
1419 /* No level skipping support yet */
1420 if (PM_PTE_LEVEL(*pte) != level)
1421 return NULL;
1422
1423 level -= 1;
1424
1425 /* Walk to the next level */
1426 pte = IOMMU_PTE_PAGE(*pte);
1427 pte = &pte[PM_LEVEL_INDEX(level, address)];
1428 *page_size = PTE_LEVEL_PAGE_SIZE(level);
1429 }
1430
1431 if (PM_PTE_LEVEL(*pte) == 0x07) {
1432 unsigned long pte_mask;
1433
1434 /*
1435 * If we have a series of large PTEs, make
1436 * sure to return a pointer to the first one.
1437 */
1438 *page_size = pte_mask = PTE_PAGE_SIZE(*pte);
1439 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
1440 pte = (u64 *)(((unsigned long)pte) & pte_mask);
1441 }
1442
1443 return pte;
1444}
1445
1446/*
1447 * Generic mapping functions. It maps a physical address into a DMA
1448 * address space. It allocates the page table pages if necessary.
1449 * In the future it can be extended to a generic mapping function
1450 * supporting all features of AMD IOMMU page tables like level skipping
1451 * and full 64 bit address spaces.
1452 */
1453static int iommu_map_page(struct protection_domain *dom,
1454 unsigned long bus_addr,
1455 unsigned long phys_addr,
1456 unsigned long page_size,
1457 int prot,
1458 gfp_t gfp)
1459{
1460 u64 __pte, *pte;
1461 int i, count;
1462
1463 BUG_ON(!IS_ALIGNED(bus_addr, page_size));
1464 BUG_ON(!IS_ALIGNED(phys_addr, page_size));
1465
1466 if (!(prot & IOMMU_PROT_MASK))
1467 return -EINVAL;
1468
1469 count = PAGE_SIZE_PTE_COUNT(page_size);
1470 pte = alloc_pte(dom, bus_addr, page_size, NULL, gfp);
1471
1472 if (!pte)
1473 return -ENOMEM;
1474
1475 for (i = 0; i < count; ++i)
1476 if (IOMMU_PTE_PRESENT(pte[i]))
1477 return -EBUSY;
1478
1479 if (count > 1) {
1480 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1481 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1482 } else
1483 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1484
1485 if (prot & IOMMU_PROT_IR)
1486 __pte |= IOMMU_PTE_IR;
1487 if (prot & IOMMU_PROT_IW)
1488 __pte |= IOMMU_PTE_IW;
1489
1490 for (i = 0; i < count; ++i)
1491 pte[i] = __pte;
1492
1493 update_domain(dom);
1494
1495 return 0;
1496}
1497
1498static unsigned long iommu_unmap_page(struct protection_domain *dom,
1499 unsigned long bus_addr,
1500 unsigned long page_size)
1501{
1502 unsigned long long unmapped;
1503 unsigned long unmap_size;
1504 u64 *pte;
1505
1506 BUG_ON(!is_power_of_2(page_size));
1507
1508 unmapped = 0;
1509
1510 while (unmapped < page_size) {
1511
1512 pte = fetch_pte(dom, bus_addr, &unmap_size);
1513
1514 if (pte) {
1515 int i, count;
1516
1517 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1518 for (i = 0; i < count; i++)
1519 pte[i] = 0ULL;
1520 }
1521
1522 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1523 unmapped += unmap_size;
1524 }
1525
1526 BUG_ON(unmapped && !is_power_of_2(unmapped));
1527
1528 return unmapped;
1529}
1530
1531/****************************************************************************
1532 *
1533 * The next functions belong to the address allocator for the dma_ops
1534 * interface functions.
1535 *
1536 ****************************************************************************/
1537
1538
1539static unsigned long dma_ops_alloc_iova(struct device *dev,
1540 struct dma_ops_domain *dma_dom,
1541 unsigned int pages, u64 dma_mask)
1542{
1543 unsigned long pfn = 0;
1544
1545 pages = __roundup_pow_of_two(pages);
1546
1547 if (dma_mask > DMA_BIT_MASK(32))
1548 pfn = alloc_iova_fast(&dma_dom->iovad, pages,
1549 IOVA_PFN(DMA_BIT_MASK(32)));
1550
1551 if (!pfn)
1552 pfn = alloc_iova_fast(&dma_dom->iovad, pages, IOVA_PFN(dma_mask));
1553
1554 return (pfn << PAGE_SHIFT);
1555}
1556
1557static void dma_ops_free_iova(struct dma_ops_domain *dma_dom,
1558 unsigned long address,
1559 unsigned int pages)
1560{
1561 pages = __roundup_pow_of_two(pages);
1562 address >>= PAGE_SHIFT;
1563
1564 free_iova_fast(&dma_dom->iovad, address, pages);
1565}
1566
1567/****************************************************************************
1568 *
1569 * The next functions belong to the domain allocation. A domain is
1570 * allocated for every IOMMU as the default domain. If device isolation
1571 * is enabled, every device get its own domain. The most important thing
1572 * about domains is the page table mapping the DMA address space they
1573 * contain.
1574 *
1575 ****************************************************************************/
1576
1577/*
1578 * This function adds a protection domain to the global protection domain list
1579 */
1580static void add_domain_to_list(struct protection_domain *domain)
1581{
1582 unsigned long flags;
1583
1584 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1585 list_add(&domain->list, &amd_iommu_pd_list);
1586 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1587}
1588
1589/*
1590 * This function removes a protection domain to the global
1591 * protection domain list
1592 */
1593static void del_domain_from_list(struct protection_domain *domain)
1594{
1595 unsigned long flags;
1596
1597 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1598 list_del(&domain->list);
1599 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1600}
1601
1602static u16 domain_id_alloc(void)
1603{
1604 unsigned long flags;
1605 int id;
1606
1607 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1608 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1609 BUG_ON(id == 0);
1610 if (id > 0 && id < MAX_DOMAIN_ID)
1611 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1612 else
1613 id = 0;
1614 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1615
1616 return id;
1617}
1618
1619static void domain_id_free(int id)
1620{
1621 unsigned long flags;
1622
1623 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1624 if (id > 0 && id < MAX_DOMAIN_ID)
1625 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1626 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1627}
1628
1629#define DEFINE_FREE_PT_FN(LVL, FN) \
1630static void free_pt_##LVL (unsigned long __pt) \
1631{ \
1632 unsigned long p; \
1633 u64 *pt; \
1634 int i; \
1635 \
1636 pt = (u64 *)__pt; \
1637 \
1638 for (i = 0; i < 512; ++i) { \
1639 /* PTE present? */ \
1640 if (!IOMMU_PTE_PRESENT(pt[i])) \
1641 continue; \
1642 \
1643 /* Large PTE? */ \
1644 if (PM_PTE_LEVEL(pt[i]) == 0 || \
1645 PM_PTE_LEVEL(pt[i]) == 7) \
1646 continue; \
1647 \
1648 p = (unsigned long)IOMMU_PTE_PAGE(pt[i]); \
1649 FN(p); \
1650 } \
1651 free_page((unsigned long)pt); \
1652}
1653
1654DEFINE_FREE_PT_FN(l2, free_page)
1655DEFINE_FREE_PT_FN(l3, free_pt_l2)
1656DEFINE_FREE_PT_FN(l4, free_pt_l3)
1657DEFINE_FREE_PT_FN(l5, free_pt_l4)
1658DEFINE_FREE_PT_FN(l6, free_pt_l5)
1659
1660static void free_pagetable(struct protection_domain *domain)
1661{
1662 unsigned long root = (unsigned long)domain->pt_root;
1663
1664 switch (domain->mode) {
1665 case PAGE_MODE_NONE:
1666 break;
1667 case PAGE_MODE_1_LEVEL:
1668 free_page(root);
1669 break;
1670 case PAGE_MODE_2_LEVEL:
1671 free_pt_l2(root);
1672 break;
1673 case PAGE_MODE_3_LEVEL:
1674 free_pt_l3(root);
1675 break;
1676 case PAGE_MODE_4_LEVEL:
1677 free_pt_l4(root);
1678 break;
1679 case PAGE_MODE_5_LEVEL:
1680 free_pt_l5(root);
1681 break;
1682 case PAGE_MODE_6_LEVEL:
1683 free_pt_l6(root);
1684 break;
1685 default:
1686 BUG();
1687 }
1688}
1689
1690static void free_gcr3_tbl_level1(u64 *tbl)
1691{
1692 u64 *ptr;
1693 int i;
1694
1695 for (i = 0; i < 512; ++i) {
1696 if (!(tbl[i] & GCR3_VALID))
1697 continue;
1698
1699 ptr = __va(tbl[i] & PAGE_MASK);
1700
1701 free_page((unsigned long)ptr);
1702 }
1703}
1704
1705static void free_gcr3_tbl_level2(u64 *tbl)
1706{
1707 u64 *ptr;
1708 int i;
1709
1710 for (i = 0; i < 512; ++i) {
1711 if (!(tbl[i] & GCR3_VALID))
1712 continue;
1713
1714 ptr = __va(tbl[i] & PAGE_MASK);
1715
1716 free_gcr3_tbl_level1(ptr);
1717 }
1718}
1719
1720static void free_gcr3_table(struct protection_domain *domain)
1721{
1722 if (domain->glx == 2)
1723 free_gcr3_tbl_level2(domain->gcr3_tbl);
1724 else if (domain->glx == 1)
1725 free_gcr3_tbl_level1(domain->gcr3_tbl);
1726 else
1727 BUG_ON(domain->glx != 0);
1728
1729 free_page((unsigned long)domain->gcr3_tbl);
1730}
1731
1732/*
1733 * Free a domain, only used if something went wrong in the
1734 * allocation path and we need to free an already allocated page table
1735 */
1736static void dma_ops_domain_free(struct dma_ops_domain *dom)
1737{
1738 if (!dom)
1739 return;
1740
1741 del_domain_from_list(&dom->domain);
1742
1743 put_iova_domain(&dom->iovad);
1744
1745 free_pagetable(&dom->domain);
1746
1747 if (dom->domain.id)
1748 domain_id_free(dom->domain.id);
1749
1750 kfree(dom);
1751}
1752
1753/*
1754 * Allocates a new protection domain usable for the dma_ops functions.
1755 * It also initializes the page table and the address allocator data
1756 * structures required for the dma_ops interface
1757 */
1758static struct dma_ops_domain *dma_ops_domain_alloc(void)
1759{
1760 struct dma_ops_domain *dma_dom;
1761
1762 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1763 if (!dma_dom)
1764 return NULL;
1765
1766 if (protection_domain_init(&dma_dom->domain))
1767 goto free_dma_dom;
1768
1769 dma_dom->domain.mode = PAGE_MODE_3_LEVEL;
1770 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
1771 dma_dom->domain.flags = PD_DMA_OPS_MASK;
1772 if (!dma_dom->domain.pt_root)
1773 goto free_dma_dom;
1774
1775 init_iova_domain(&dma_dom->iovad, PAGE_SIZE,
1776 IOVA_START_PFN, DMA_32BIT_PFN);
1777
1778 /* Initialize reserved ranges */
1779 copy_reserved_iova(&reserved_iova_ranges, &dma_dom->iovad);
1780
1781 add_domain_to_list(&dma_dom->domain);
1782
1783 return dma_dom;
1784
1785free_dma_dom:
1786 dma_ops_domain_free(dma_dom);
1787
1788 return NULL;
1789}
1790
1791/*
1792 * little helper function to check whether a given protection domain is a
1793 * dma_ops domain
1794 */
1795static bool dma_ops_domain(struct protection_domain *domain)
1796{
1797 return domain->flags & PD_DMA_OPS_MASK;
1798}
1799
1800static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
1801{
1802 u64 pte_root = 0;
1803 u64 flags = 0;
1804
1805 if (domain->mode != PAGE_MODE_NONE)
1806 pte_root = virt_to_phys(domain->pt_root);
1807
1808 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1809 << DEV_ENTRY_MODE_SHIFT;
1810 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
1811
1812 flags = amd_iommu_dev_table[devid].data[1];
1813
1814 if (ats)
1815 flags |= DTE_FLAG_IOTLB;
1816
1817 if (domain->flags & PD_IOMMUV2_MASK) {
1818 u64 gcr3 = __pa(domain->gcr3_tbl);
1819 u64 glx = domain->glx;
1820 u64 tmp;
1821
1822 pte_root |= DTE_FLAG_GV;
1823 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT;
1824
1825 /* First mask out possible old values for GCR3 table */
1826 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B;
1827 flags &= ~tmp;
1828
1829 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C;
1830 flags &= ~tmp;
1831
1832 /* Encode GCR3 table into DTE */
1833 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A;
1834 pte_root |= tmp;
1835
1836 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B;
1837 flags |= tmp;
1838
1839 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C;
1840 flags |= tmp;
1841 }
1842
1843 flags &= ~(0xffffUL);
1844 flags |= domain->id;
1845
1846 amd_iommu_dev_table[devid].data[1] = flags;
1847 amd_iommu_dev_table[devid].data[0] = pte_root;
1848}
1849
1850static void clear_dte_entry(u16 devid)
1851{
1852 /* remove entry from the device table seen by the hardware */
1853 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1854 amd_iommu_dev_table[devid].data[1] &= DTE_FLAG_MASK;
1855
1856 amd_iommu_apply_erratum_63(devid);
1857}
1858
1859static void do_attach(struct iommu_dev_data *dev_data,
1860 struct protection_domain *domain)
1861{
1862 struct amd_iommu *iommu;
1863 u16 alias;
1864 bool ats;
1865
1866 iommu = amd_iommu_rlookup_table[dev_data->devid];
1867 alias = dev_data->alias;
1868 ats = dev_data->ats.enabled;
1869
1870 /* Update data structures */
1871 dev_data->domain = domain;
1872 list_add(&dev_data->list, &domain->dev_list);
1873
1874 /* Do reference counting */
1875 domain->dev_iommu[iommu->index] += 1;
1876 domain->dev_cnt += 1;
1877
1878 /* Update device table */
1879 set_dte_entry(dev_data->devid, domain, ats);
1880 if (alias != dev_data->devid)
1881 set_dte_entry(alias, domain, ats);
1882
1883 device_flush_dte(dev_data);
1884}
1885
1886static void do_detach(struct iommu_dev_data *dev_data)
1887{
1888 struct amd_iommu *iommu;
1889 u16 alias;
1890
1891 /*
1892 * First check if the device is still attached. It might already
1893 * be detached from its domain because the generic
1894 * iommu_detach_group code detached it and we try again here in
1895 * our alias handling.
1896 */
1897 if (!dev_data->domain)
1898 return;
1899
1900 iommu = amd_iommu_rlookup_table[dev_data->devid];
1901 alias = dev_data->alias;
1902
1903 /* decrease reference counters */
1904 dev_data->domain->dev_iommu[iommu->index] -= 1;
1905 dev_data->domain->dev_cnt -= 1;
1906
1907 /* Update data structures */
1908 dev_data->domain = NULL;
1909 list_del(&dev_data->list);
1910 clear_dte_entry(dev_data->devid);
1911 if (alias != dev_data->devid)
1912 clear_dte_entry(alias);
1913
1914 /* Flush the DTE entry */
1915 device_flush_dte(dev_data);
1916}
1917
1918/*
1919 * If a device is not yet associated with a domain, this function does
1920 * assigns it visible for the hardware
1921 */
1922static int __attach_device(struct iommu_dev_data *dev_data,
1923 struct protection_domain *domain)
1924{
1925 int ret;
1926
1927 /*
1928 * Must be called with IRQs disabled. Warn here to detect early
1929 * when its not.
1930 */
1931 WARN_ON(!irqs_disabled());
1932
1933 /* lock domain */
1934 spin_lock(&domain->lock);
1935
1936 ret = -EBUSY;
1937 if (dev_data->domain != NULL)
1938 goto out_unlock;
1939
1940 /* Attach alias group root */
1941 do_attach(dev_data, domain);
1942
1943 ret = 0;
1944
1945out_unlock:
1946
1947 /* ready */
1948 spin_unlock(&domain->lock);
1949
1950 return ret;
1951}
1952
1953
1954static void pdev_iommuv2_disable(struct pci_dev *pdev)
1955{
1956 pci_disable_ats(pdev);
1957 pci_disable_pri(pdev);
1958 pci_disable_pasid(pdev);
1959}
1960
1961/* FIXME: Change generic reset-function to do the same */
1962static int pri_reset_while_enabled(struct pci_dev *pdev)
1963{
1964 u16 control;
1965 int pos;
1966
1967 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
1968 if (!pos)
1969 return -EINVAL;
1970
1971 pci_read_config_word(pdev, pos + PCI_PRI_CTRL, &control);
1972 control |= PCI_PRI_CTRL_RESET;
1973 pci_write_config_word(pdev, pos + PCI_PRI_CTRL, control);
1974
1975 return 0;
1976}
1977
1978static int pdev_iommuv2_enable(struct pci_dev *pdev)
1979{
1980 bool reset_enable;
1981 int reqs, ret;
1982
1983 /* FIXME: Hardcode number of outstanding requests for now */
1984 reqs = 32;
1985 if (pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_LIMIT_REQ_ONE))
1986 reqs = 1;
1987 reset_enable = pdev_pri_erratum(pdev, AMD_PRI_DEV_ERRATUM_ENABLE_RESET);
1988
1989 /* Only allow access to user-accessible pages */
1990 ret = pci_enable_pasid(pdev, 0);
1991 if (ret)
1992 goto out_err;
1993
1994 /* First reset the PRI state of the device */
1995 ret = pci_reset_pri(pdev);
1996 if (ret)
1997 goto out_err;
1998
1999 /* Enable PRI */
2000 ret = pci_enable_pri(pdev, reqs);
2001 if (ret)
2002 goto out_err;
2003
2004 if (reset_enable) {
2005 ret = pri_reset_while_enabled(pdev);
2006 if (ret)
2007 goto out_err;
2008 }
2009
2010 ret = pci_enable_ats(pdev, PAGE_SHIFT);
2011 if (ret)
2012 goto out_err;
2013
2014 return 0;
2015
2016out_err:
2017 pci_disable_pri(pdev);
2018 pci_disable_pasid(pdev);
2019
2020 return ret;
2021}
2022
2023/* FIXME: Move this to PCI code */
2024#define PCI_PRI_TLP_OFF (1 << 15)
2025
2026static bool pci_pri_tlp_required(struct pci_dev *pdev)
2027{
2028 u16 status;
2029 int pos;
2030
2031 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
2032 if (!pos)
2033 return false;
2034
2035 pci_read_config_word(pdev, pos + PCI_PRI_STATUS, &status);
2036
2037 return (status & PCI_PRI_TLP_OFF) ? true : false;
2038}
2039
2040/*
2041 * If a device is not yet associated with a domain, this function
2042 * assigns it visible for the hardware
2043 */
2044static int attach_device(struct device *dev,
2045 struct protection_domain *domain)
2046{
2047 struct pci_dev *pdev;
2048 struct iommu_dev_data *dev_data;
2049 unsigned long flags;
2050 int ret;
2051
2052 dev_data = get_dev_data(dev);
2053
2054 if (!dev_is_pci(dev))
2055 goto skip_ats_check;
2056
2057 pdev = to_pci_dev(dev);
2058 if (domain->flags & PD_IOMMUV2_MASK) {
2059 if (!dev_data->passthrough)
2060 return -EINVAL;
2061
2062 if (dev_data->iommu_v2) {
2063 if (pdev_iommuv2_enable(pdev) != 0)
2064 return -EINVAL;
2065
2066 dev_data->ats.enabled = true;
2067 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2068 dev_data->pri_tlp = pci_pri_tlp_required(pdev);
2069 }
2070 } else if (amd_iommu_iotlb_sup &&
2071 pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
2072 dev_data->ats.enabled = true;
2073 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
2074 }
2075
2076skip_ats_check:
2077 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2078 ret = __attach_device(dev_data, domain);
2079 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2080
2081 /*
2082 * We might boot into a crash-kernel here. The crashed kernel
2083 * left the caches in the IOMMU dirty. So we have to flush
2084 * here to evict all dirty stuff.
2085 */
2086 domain_flush_tlb_pde(domain);
2087
2088 return ret;
2089}
2090
2091/*
2092 * Removes a device from a protection domain (unlocked)
2093 */
2094static void __detach_device(struct iommu_dev_data *dev_data)
2095{
2096 struct protection_domain *domain;
2097
2098 /*
2099 * Must be called with IRQs disabled. Warn here to detect early
2100 * when its not.
2101 */
2102 WARN_ON(!irqs_disabled());
2103
2104 if (WARN_ON(!dev_data->domain))
2105 return;
2106
2107 domain = dev_data->domain;
2108
2109 spin_lock(&domain->lock);
2110
2111 do_detach(dev_data);
2112
2113 spin_unlock(&domain->lock);
2114}
2115
2116/*
2117 * Removes a device from a protection domain (with devtable_lock held)
2118 */
2119static void detach_device(struct device *dev)
2120{
2121 struct protection_domain *domain;
2122 struct iommu_dev_data *dev_data;
2123 unsigned long flags;
2124
2125 dev_data = get_dev_data(dev);
2126 domain = dev_data->domain;
2127
2128 /* lock device table */
2129 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2130 __detach_device(dev_data);
2131 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2132
2133 if (!dev_is_pci(dev))
2134 return;
2135
2136 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2)
2137 pdev_iommuv2_disable(to_pci_dev(dev));
2138 else if (dev_data->ats.enabled)
2139 pci_disable_ats(to_pci_dev(dev));
2140
2141 dev_data->ats.enabled = false;
2142}
2143
2144static int amd_iommu_add_device(struct device *dev)
2145{
2146 struct iommu_dev_data *dev_data;
2147 struct iommu_domain *domain;
2148 struct amd_iommu *iommu;
2149 int ret, devid;
2150
2151 if (!check_device(dev) || get_dev_data(dev))
2152 return 0;
2153
2154 devid = get_device_id(dev);
2155 if (devid < 0)
2156 return devid;
2157
2158 iommu = amd_iommu_rlookup_table[devid];
2159
2160 ret = iommu_init_device(dev);
2161 if (ret) {
2162 if (ret != -ENOTSUPP)
2163 pr_err("Failed to initialize device %s - trying to proceed anyway\n",
2164 dev_name(dev));
2165
2166 iommu_ignore_device(dev);
2167 dev->archdata.dma_ops = &nommu_dma_ops;
2168 goto out;
2169 }
2170 init_iommu_group(dev);
2171
2172 dev_data = get_dev_data(dev);
2173
2174 BUG_ON(!dev_data);
2175
2176 if (iommu_pass_through || dev_data->iommu_v2)
2177 iommu_request_dm_for_dev(dev);
2178
2179 /* Domains are initialized for this device - have a look what we ended up with */
2180 domain = iommu_get_domain_for_dev(dev);
2181 if (domain->type == IOMMU_DOMAIN_IDENTITY)
2182 dev_data->passthrough = true;
2183 else
2184 dev->archdata.dma_ops = &amd_iommu_dma_ops;
2185
2186out:
2187 iommu_completion_wait(iommu);
2188
2189 return 0;
2190}
2191
2192static void amd_iommu_remove_device(struct device *dev)
2193{
2194 struct amd_iommu *iommu;
2195 int devid;
2196
2197 if (!check_device(dev))
2198 return;
2199
2200 devid = get_device_id(dev);
2201 if (devid < 0)
2202 return;
2203
2204 iommu = amd_iommu_rlookup_table[devid];
2205
2206 iommu_uninit_device(dev);
2207 iommu_completion_wait(iommu);
2208}
2209
2210static struct iommu_group *amd_iommu_device_group(struct device *dev)
2211{
2212 if (dev_is_pci(dev))
2213 return pci_device_group(dev);
2214
2215 return acpihid_device_group(dev);
2216}
2217
2218/*****************************************************************************
2219 *
2220 * The next functions belong to the dma_ops mapping/unmapping code.
2221 *
2222 *****************************************************************************/
2223
2224static void __queue_flush(struct flush_queue *queue)
2225{
2226 struct protection_domain *domain;
2227 unsigned long flags;
2228 int idx;
2229
2230 /* First flush TLB of all known domains */
2231 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
2232 list_for_each_entry(domain, &amd_iommu_pd_list, list)
2233 domain_flush_tlb(domain);
2234 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
2235
2236 /* Wait until flushes have completed */
2237 domain_flush_complete(NULL);
2238
2239 for (idx = 0; idx < queue->next; ++idx) {
2240 struct flush_queue_entry *entry;
2241
2242 entry = queue->entries + idx;
2243
2244 free_iova_fast(&entry->dma_dom->iovad,
2245 entry->iova_pfn,
2246 entry->pages);
2247
2248 /* Not really necessary, just to make sure we catch any bugs */
2249 entry->dma_dom = NULL;
2250 }
2251
2252 queue->next = 0;
2253}
2254
2255static void queue_flush_all(void)
2256{
2257 int cpu;
2258
2259 for_each_possible_cpu(cpu) {
2260 struct flush_queue *queue;
2261 unsigned long flags;
2262
2263 queue = per_cpu_ptr(&flush_queue, cpu);
2264 spin_lock_irqsave(&queue->lock, flags);
2265 if (queue->next > 0)
2266 __queue_flush(queue);
2267 spin_unlock_irqrestore(&queue->lock, flags);
2268 }
2269}
2270
2271static void queue_flush_timeout(unsigned long unsused)
2272{
2273 atomic_set(&queue_timer_on, 0);
2274 queue_flush_all();
2275}
2276
2277static void queue_add(struct dma_ops_domain *dma_dom,
2278 unsigned long address, unsigned long pages)
2279{
2280 struct flush_queue_entry *entry;
2281 struct flush_queue *queue;
2282 unsigned long flags;
2283 int idx;
2284
2285 pages = __roundup_pow_of_two(pages);
2286 address >>= PAGE_SHIFT;
2287
2288 queue = get_cpu_ptr(&flush_queue);
2289 spin_lock_irqsave(&queue->lock, flags);
2290
2291 if (queue->next == FLUSH_QUEUE_SIZE)
2292 __queue_flush(queue);
2293
2294 idx = queue->next++;
2295 entry = queue->entries + idx;
2296
2297 entry->iova_pfn = address;
2298 entry->pages = pages;
2299 entry->dma_dom = dma_dom;
2300
2301 spin_unlock_irqrestore(&queue->lock, flags);
2302
2303 if (atomic_cmpxchg(&queue_timer_on, 0, 1) == 0)
2304 mod_timer(&queue_timer, jiffies + msecs_to_jiffies(10));
2305
2306 put_cpu_ptr(&flush_queue);
2307}
2308
2309
2310/*
2311 * In the dma_ops path we only have the struct device. This function
2312 * finds the corresponding IOMMU, the protection domain and the
2313 * requestor id for a given device.
2314 * If the device is not yet associated with a domain this is also done
2315 * in this function.
2316 */
2317static struct protection_domain *get_domain(struct device *dev)
2318{
2319 struct protection_domain *domain;
2320
2321 if (!check_device(dev))
2322 return ERR_PTR(-EINVAL);
2323
2324 domain = get_dev_data(dev)->domain;
2325 if (!dma_ops_domain(domain))
2326 return ERR_PTR(-EBUSY);
2327
2328 return domain;
2329}
2330
2331static void update_device_table(struct protection_domain *domain)
2332{
2333 struct iommu_dev_data *dev_data;
2334
2335 list_for_each_entry(dev_data, &domain->dev_list, list) {
2336 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
2337
2338 if (dev_data->devid == dev_data->alias)
2339 continue;
2340
2341 /* There is an alias, update device table entry for it */
2342 set_dte_entry(dev_data->alias, domain, dev_data->ats.enabled);
2343 }
2344}
2345
2346static void update_domain(struct protection_domain *domain)
2347{
2348 if (!domain->updated)
2349 return;
2350
2351 update_device_table(domain);
2352
2353 domain_flush_devices(domain);
2354 domain_flush_tlb_pde(domain);
2355
2356 domain->updated = false;
2357}
2358
2359static int dir2prot(enum dma_data_direction direction)
2360{
2361 if (direction == DMA_TO_DEVICE)
2362 return IOMMU_PROT_IR;
2363 else if (direction == DMA_FROM_DEVICE)
2364 return IOMMU_PROT_IW;
2365 else if (direction == DMA_BIDIRECTIONAL)
2366 return IOMMU_PROT_IW | IOMMU_PROT_IR;
2367 else
2368 return 0;
2369}
2370/*
2371 * This function contains common code for mapping of a physically
2372 * contiguous memory region into DMA address space. It is used by all
2373 * mapping functions provided with this IOMMU driver.
2374 * Must be called with the domain lock held.
2375 */
2376static dma_addr_t __map_single(struct device *dev,
2377 struct dma_ops_domain *dma_dom,
2378 phys_addr_t paddr,
2379 size_t size,
2380 enum dma_data_direction direction,
2381 u64 dma_mask)
2382{
2383 dma_addr_t offset = paddr & ~PAGE_MASK;
2384 dma_addr_t address, start, ret;
2385 unsigned int pages;
2386 int prot = 0;
2387 int i;
2388
2389 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
2390 paddr &= PAGE_MASK;
2391
2392 address = dma_ops_alloc_iova(dev, dma_dom, pages, dma_mask);
2393 if (address == DMA_ERROR_CODE)
2394 goto out;
2395
2396 prot = dir2prot(direction);
2397
2398 start = address;
2399 for (i = 0; i < pages; ++i) {
2400 ret = iommu_map_page(&dma_dom->domain, start, paddr,
2401 PAGE_SIZE, prot, GFP_ATOMIC);
2402 if (ret)
2403 goto out_unmap;
2404
2405 paddr += PAGE_SIZE;
2406 start += PAGE_SIZE;
2407 }
2408 address += offset;
2409
2410 if (unlikely(amd_iommu_np_cache)) {
2411 domain_flush_pages(&dma_dom->domain, address, size);
2412 domain_flush_complete(&dma_dom->domain);
2413 }
2414
2415out:
2416 return address;
2417
2418out_unmap:
2419
2420 for (--i; i >= 0; --i) {
2421 start -= PAGE_SIZE;
2422 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2423 }
2424
2425 domain_flush_tlb(&dma_dom->domain);
2426 domain_flush_complete(&dma_dom->domain);
2427
2428 dma_ops_free_iova(dma_dom, address, pages);
2429
2430 return DMA_ERROR_CODE;
2431}
2432
2433/*
2434 * Does the reverse of the __map_single function. Must be called with
2435 * the domain lock held too
2436 */
2437static void __unmap_single(struct dma_ops_domain *dma_dom,
2438 dma_addr_t dma_addr,
2439 size_t size,
2440 int dir)
2441{
2442 dma_addr_t flush_addr;
2443 dma_addr_t i, start;
2444 unsigned int pages;
2445
2446 flush_addr = dma_addr;
2447 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
2448 dma_addr &= PAGE_MASK;
2449 start = dma_addr;
2450
2451 for (i = 0; i < pages; ++i) {
2452 iommu_unmap_page(&dma_dom->domain, start, PAGE_SIZE);
2453 start += PAGE_SIZE;
2454 }
2455
2456 if (amd_iommu_unmap_flush) {
2457 dma_ops_free_iova(dma_dom, dma_addr, pages);
2458 domain_flush_tlb(&dma_dom->domain);
2459 domain_flush_complete(&dma_dom->domain);
2460 } else {
2461 queue_add(dma_dom, dma_addr, pages);
2462 }
2463}
2464
2465/*
2466 * The exported map_single function for dma_ops.
2467 */
2468static dma_addr_t map_page(struct device *dev, struct page *page,
2469 unsigned long offset, size_t size,
2470 enum dma_data_direction dir,
2471 unsigned long attrs)
2472{
2473 phys_addr_t paddr = page_to_phys(page) + offset;
2474 struct protection_domain *domain;
2475 struct dma_ops_domain *dma_dom;
2476 u64 dma_mask;
2477
2478 domain = get_domain(dev);
2479 if (PTR_ERR(domain) == -EINVAL)
2480 return (dma_addr_t)paddr;
2481 else if (IS_ERR(domain))
2482 return DMA_ERROR_CODE;
2483
2484 dma_mask = *dev->dma_mask;
2485 dma_dom = to_dma_ops_domain(domain);
2486
2487 return __map_single(dev, dma_dom, paddr, size, dir, dma_mask);
2488}
2489
2490/*
2491 * The exported unmap_single function for dma_ops.
2492 */
2493static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2494 enum dma_data_direction dir, unsigned long attrs)
2495{
2496 struct protection_domain *domain;
2497 struct dma_ops_domain *dma_dom;
2498
2499 domain = get_domain(dev);
2500 if (IS_ERR(domain))
2501 return;
2502
2503 dma_dom = to_dma_ops_domain(domain);
2504
2505 __unmap_single(dma_dom, dma_addr, size, dir);
2506}
2507
2508static int sg_num_pages(struct device *dev,
2509 struct scatterlist *sglist,
2510 int nelems)
2511{
2512 unsigned long mask, boundary_size;
2513 struct scatterlist *s;
2514 int i, npages = 0;
2515
2516 mask = dma_get_seg_boundary(dev);
2517 boundary_size = mask + 1 ? ALIGN(mask + 1, PAGE_SIZE) >> PAGE_SHIFT :
2518 1UL << (BITS_PER_LONG - PAGE_SHIFT);
2519
2520 for_each_sg(sglist, s, nelems, i) {
2521 int p, n;
2522
2523 s->dma_address = npages << PAGE_SHIFT;
2524 p = npages % boundary_size;
2525 n = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2526 if (p + n > boundary_size)
2527 npages += boundary_size - p;
2528 npages += n;
2529 }
2530
2531 return npages;
2532}
2533
2534/*
2535 * The exported map_sg function for dma_ops (handles scatter-gather
2536 * lists).
2537 */
2538static int map_sg(struct device *dev, struct scatterlist *sglist,
2539 int nelems, enum dma_data_direction direction,
2540 unsigned long attrs)
2541{
2542 int mapped_pages = 0, npages = 0, prot = 0, i;
2543 struct protection_domain *domain;
2544 struct dma_ops_domain *dma_dom;
2545 struct scatterlist *s;
2546 unsigned long address;
2547 u64 dma_mask;
2548
2549 domain = get_domain(dev);
2550 if (IS_ERR(domain))
2551 return 0;
2552
2553 dma_dom = to_dma_ops_domain(domain);
2554 dma_mask = *dev->dma_mask;
2555
2556 npages = sg_num_pages(dev, sglist, nelems);
2557
2558 address = dma_ops_alloc_iova(dev, dma_dom, npages, dma_mask);
2559 if (address == DMA_ERROR_CODE)
2560 goto out_err;
2561
2562 prot = dir2prot(direction);
2563
2564 /* Map all sg entries */
2565 for_each_sg(sglist, s, nelems, i) {
2566 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2567
2568 for (j = 0; j < pages; ++j) {
2569 unsigned long bus_addr, phys_addr;
2570 int ret;
2571
2572 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2573 phys_addr = (sg_phys(s) & PAGE_MASK) + (j << PAGE_SHIFT);
2574 ret = iommu_map_page(domain, bus_addr, phys_addr, PAGE_SIZE, prot, GFP_ATOMIC);
2575 if (ret)
2576 goto out_unmap;
2577
2578 mapped_pages += 1;
2579 }
2580 }
2581
2582 /* Everything is mapped - write the right values into s->dma_address */
2583 for_each_sg(sglist, s, nelems, i) {
2584 s->dma_address += address + s->offset;
2585 s->dma_length = s->length;
2586 }
2587
2588 return nelems;
2589
2590out_unmap:
2591 pr_err("%s: IOMMU mapping error in map_sg (io-pages: %d)\n",
2592 dev_name(dev), npages);
2593
2594 for_each_sg(sglist, s, nelems, i) {
2595 int j, pages = iommu_num_pages(sg_phys(s), s->length, PAGE_SIZE);
2596
2597 for (j = 0; j < pages; ++j) {
2598 unsigned long bus_addr;
2599
2600 bus_addr = address + s->dma_address + (j << PAGE_SHIFT);
2601 iommu_unmap_page(domain, bus_addr, PAGE_SIZE);
2602
2603 if (--mapped_pages)
2604 goto out_free_iova;
2605 }
2606 }
2607
2608out_free_iova:
2609 free_iova_fast(&dma_dom->iovad, address, npages);
2610
2611out_err:
2612 return 0;
2613}
2614
2615/*
2616 * The exported map_sg function for dma_ops (handles scatter-gather
2617 * lists).
2618 */
2619static void unmap_sg(struct device *dev, struct scatterlist *sglist,
2620 int nelems, enum dma_data_direction dir,
2621 unsigned long attrs)
2622{
2623 struct protection_domain *domain;
2624 struct dma_ops_domain *dma_dom;
2625 unsigned long startaddr;
2626 int npages = 2;
2627
2628 domain = get_domain(dev);
2629 if (IS_ERR(domain))
2630 return;
2631
2632 startaddr = sg_dma_address(sglist) & PAGE_MASK;
2633 dma_dom = to_dma_ops_domain(domain);
2634 npages = sg_num_pages(dev, sglist, nelems);
2635
2636 __unmap_single(dma_dom, startaddr, npages << PAGE_SHIFT, dir);
2637}
2638
2639/*
2640 * The exported alloc_coherent function for dma_ops.
2641 */
2642static void *alloc_coherent(struct device *dev, size_t size,
2643 dma_addr_t *dma_addr, gfp_t flag,
2644 unsigned long attrs)
2645{
2646 u64 dma_mask = dev->coherent_dma_mask;
2647 struct protection_domain *domain;
2648 struct dma_ops_domain *dma_dom;
2649 struct page *page;
2650
2651 domain = get_domain(dev);
2652 if (PTR_ERR(domain) == -EINVAL) {
2653 page = alloc_pages(flag, get_order(size));
2654 *dma_addr = page_to_phys(page);
2655 return page_address(page);
2656 } else if (IS_ERR(domain))
2657 return NULL;
2658
2659 dma_dom = to_dma_ops_domain(domain);
2660 size = PAGE_ALIGN(size);
2661 dma_mask = dev->coherent_dma_mask;
2662 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2663 flag |= __GFP_ZERO;
2664
2665 page = alloc_pages(flag | __GFP_NOWARN, get_order(size));
2666 if (!page) {
2667 if (!gfpflags_allow_blocking(flag))
2668 return NULL;
2669
2670 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
2671 get_order(size));
2672 if (!page)
2673 return NULL;
2674 }
2675
2676 if (!dma_mask)
2677 dma_mask = *dev->dma_mask;
2678
2679 *dma_addr = __map_single(dev, dma_dom, page_to_phys(page),
2680 size, DMA_BIDIRECTIONAL, dma_mask);
2681
2682 if (*dma_addr == DMA_ERROR_CODE)
2683 goto out_free;
2684
2685 return page_address(page);
2686
2687out_free:
2688
2689 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2690 __free_pages(page, get_order(size));
2691
2692 return NULL;
2693}
2694
2695/*
2696 * The exported free_coherent function for dma_ops.
2697 */
2698static void free_coherent(struct device *dev, size_t size,
2699 void *virt_addr, dma_addr_t dma_addr,
2700 unsigned long attrs)
2701{
2702 struct protection_domain *domain;
2703 struct dma_ops_domain *dma_dom;
2704 struct page *page;
2705
2706 page = virt_to_page(virt_addr);
2707 size = PAGE_ALIGN(size);
2708
2709 domain = get_domain(dev);
2710 if (IS_ERR(domain))
2711 goto free_mem;
2712
2713 dma_dom = to_dma_ops_domain(domain);
2714
2715 __unmap_single(dma_dom, dma_addr, size, DMA_BIDIRECTIONAL);
2716
2717free_mem:
2718 if (!dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT))
2719 __free_pages(page, get_order(size));
2720}
2721
2722/*
2723 * This function is called by the DMA layer to find out if we can handle a
2724 * particular device. It is part of the dma_ops.
2725 */
2726static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2727{
2728 return check_device(dev);
2729}
2730
2731static struct dma_map_ops amd_iommu_dma_ops = {
2732 .alloc = alloc_coherent,
2733 .free = free_coherent,
2734 .map_page = map_page,
2735 .unmap_page = unmap_page,
2736 .map_sg = map_sg,
2737 .unmap_sg = unmap_sg,
2738 .dma_supported = amd_iommu_dma_supported,
2739};
2740
2741static int init_reserved_iova_ranges(void)
2742{
2743 struct pci_dev *pdev = NULL;
2744 struct iova *val;
2745
2746 init_iova_domain(&reserved_iova_ranges, PAGE_SIZE,
2747 IOVA_START_PFN, DMA_32BIT_PFN);
2748
2749 lockdep_set_class(&reserved_iova_ranges.iova_rbtree_lock,
2750 &reserved_rbtree_key);
2751
2752 /* MSI memory range */
2753 val = reserve_iova(&reserved_iova_ranges,
2754 IOVA_PFN(MSI_RANGE_START), IOVA_PFN(MSI_RANGE_END));
2755 if (!val) {
2756 pr_err("Reserving MSI range failed\n");
2757 return -ENOMEM;
2758 }
2759
2760 /* HT memory range */
2761 val = reserve_iova(&reserved_iova_ranges,
2762 IOVA_PFN(HT_RANGE_START), IOVA_PFN(HT_RANGE_END));
2763 if (!val) {
2764 pr_err("Reserving HT range failed\n");
2765 return -ENOMEM;
2766 }
2767
2768 /*
2769 * Memory used for PCI resources
2770 * FIXME: Check whether we can reserve the PCI-hole completly
2771 */
2772 for_each_pci_dev(pdev) {
2773 int i;
2774
2775 for (i = 0; i < PCI_NUM_RESOURCES; ++i) {
2776 struct resource *r = &pdev->resource[i];
2777
2778 if (!(r->flags & IORESOURCE_MEM))
2779 continue;
2780
2781 val = reserve_iova(&reserved_iova_ranges,
2782 IOVA_PFN(r->start),
2783 IOVA_PFN(r->end));
2784 if (!val) {
2785 pr_err("Reserve pci-resource range failed\n");
2786 return -ENOMEM;
2787 }
2788 }
2789 }
2790
2791 return 0;
2792}
2793
2794int __init amd_iommu_init_api(void)
2795{
2796 int ret, cpu, err = 0;
2797
2798 ret = iova_cache_get();
2799 if (ret)
2800 return ret;
2801
2802 ret = init_reserved_iova_ranges();
2803 if (ret)
2804 return ret;
2805
2806 for_each_possible_cpu(cpu) {
2807 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2808
2809 queue->entries = kzalloc(FLUSH_QUEUE_SIZE *
2810 sizeof(*queue->entries),
2811 GFP_KERNEL);
2812 if (!queue->entries)
2813 goto out_put_iova;
2814
2815 spin_lock_init(&queue->lock);
2816 }
2817
2818 err = bus_set_iommu(&pci_bus_type, &amd_iommu_ops);
2819 if (err)
2820 return err;
2821#ifdef CONFIG_ARM_AMBA
2822 err = bus_set_iommu(&amba_bustype, &amd_iommu_ops);
2823 if (err)
2824 return err;
2825#endif
2826 err = bus_set_iommu(&platform_bus_type, &amd_iommu_ops);
2827 if (err)
2828 return err;
2829 return 0;
2830
2831out_put_iova:
2832 for_each_possible_cpu(cpu) {
2833 struct flush_queue *queue = per_cpu_ptr(&flush_queue, cpu);
2834
2835 kfree(queue->entries);
2836 }
2837
2838 return -ENOMEM;
2839}
2840
2841int __init amd_iommu_init_dma_ops(void)
2842{
2843 setup_timer(&queue_timer, queue_flush_timeout, 0);
2844 atomic_set(&queue_timer_on, 0);
2845
2846 swiotlb = iommu_pass_through ? 1 : 0;
2847 iommu_detected = 1;
2848
2849 /*
2850 * In case we don't initialize SWIOTLB (actually the common case
2851 * when AMD IOMMU is enabled), make sure there are global
2852 * dma_ops set as a fall-back for devices not handled by this
2853 * driver (for example non-PCI devices).
2854 */
2855 if (!swiotlb)
2856 dma_ops = &nommu_dma_ops;
2857
2858 if (amd_iommu_unmap_flush)
2859 pr_info("AMD-Vi: IO/TLB flush on unmap enabled\n");
2860 else
2861 pr_info("AMD-Vi: Lazy IO/TLB flushing enabled\n");
2862
2863 return 0;
2864
2865}
2866
2867/*****************************************************************************
2868 *
2869 * The following functions belong to the exported interface of AMD IOMMU
2870 *
2871 * This interface allows access to lower level functions of the IOMMU
2872 * like protection domain handling and assignement of devices to domains
2873 * which is not possible with the dma_ops interface.
2874 *
2875 *****************************************************************************/
2876
2877static void cleanup_domain(struct protection_domain *domain)
2878{
2879 struct iommu_dev_data *entry;
2880 unsigned long flags;
2881
2882 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2883
2884 while (!list_empty(&domain->dev_list)) {
2885 entry = list_first_entry(&domain->dev_list,
2886 struct iommu_dev_data, list);
2887 __detach_device(entry);
2888 }
2889
2890 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2891}
2892
2893static void protection_domain_free(struct protection_domain *domain)
2894{
2895 if (!domain)
2896 return;
2897
2898 del_domain_from_list(domain);
2899
2900 if (domain->id)
2901 domain_id_free(domain->id);
2902
2903 kfree(domain);
2904}
2905
2906static int protection_domain_init(struct protection_domain *domain)
2907{
2908 spin_lock_init(&domain->lock);
2909 mutex_init(&domain->api_lock);
2910 domain->id = domain_id_alloc();
2911 if (!domain->id)
2912 return -ENOMEM;
2913 INIT_LIST_HEAD(&domain->dev_list);
2914
2915 return 0;
2916}
2917
2918static struct protection_domain *protection_domain_alloc(void)
2919{
2920 struct protection_domain *domain;
2921
2922 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2923 if (!domain)
2924 return NULL;
2925
2926 if (protection_domain_init(domain))
2927 goto out_err;
2928
2929 add_domain_to_list(domain);
2930
2931 return domain;
2932
2933out_err:
2934 kfree(domain);
2935
2936 return NULL;
2937}
2938
2939static struct iommu_domain *amd_iommu_domain_alloc(unsigned type)
2940{
2941 struct protection_domain *pdomain;
2942 struct dma_ops_domain *dma_domain;
2943
2944 switch (type) {
2945 case IOMMU_DOMAIN_UNMANAGED:
2946 pdomain = protection_domain_alloc();
2947 if (!pdomain)
2948 return NULL;
2949
2950 pdomain->mode = PAGE_MODE_3_LEVEL;
2951 pdomain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2952 if (!pdomain->pt_root) {
2953 protection_domain_free(pdomain);
2954 return NULL;
2955 }
2956
2957 pdomain->domain.geometry.aperture_start = 0;
2958 pdomain->domain.geometry.aperture_end = ~0ULL;
2959 pdomain->domain.geometry.force_aperture = true;
2960
2961 break;
2962 case IOMMU_DOMAIN_DMA:
2963 dma_domain = dma_ops_domain_alloc();
2964 if (!dma_domain) {
2965 pr_err("AMD-Vi: Failed to allocate\n");
2966 return NULL;
2967 }
2968 pdomain = &dma_domain->domain;
2969 break;
2970 case IOMMU_DOMAIN_IDENTITY:
2971 pdomain = protection_domain_alloc();
2972 if (!pdomain)
2973 return NULL;
2974
2975 pdomain->mode = PAGE_MODE_NONE;
2976 break;
2977 default:
2978 return NULL;
2979 }
2980
2981 return &pdomain->domain;
2982}
2983
2984static void amd_iommu_domain_free(struct iommu_domain *dom)
2985{
2986 struct protection_domain *domain;
2987 struct dma_ops_domain *dma_dom;
2988
2989 domain = to_pdomain(dom);
2990
2991 if (domain->dev_cnt > 0)
2992 cleanup_domain(domain);
2993
2994 BUG_ON(domain->dev_cnt != 0);
2995
2996 if (!dom)
2997 return;
2998
2999 switch (dom->type) {
3000 case IOMMU_DOMAIN_DMA:
3001 /*
3002 * First make sure the domain is no longer referenced from the
3003 * flush queue
3004 */
3005 queue_flush_all();
3006
3007 /* Now release the domain */
3008 dma_dom = to_dma_ops_domain(domain);
3009 dma_ops_domain_free(dma_dom);
3010 break;
3011 default:
3012 if (domain->mode != PAGE_MODE_NONE)
3013 free_pagetable(domain);
3014
3015 if (domain->flags & PD_IOMMUV2_MASK)
3016 free_gcr3_table(domain);
3017
3018 protection_domain_free(domain);
3019 break;
3020 }
3021}
3022
3023static void amd_iommu_detach_device(struct iommu_domain *dom,
3024 struct device *dev)
3025{
3026 struct iommu_dev_data *dev_data = dev->archdata.iommu;
3027 struct amd_iommu *iommu;
3028 int devid;
3029
3030 if (!check_device(dev))
3031 return;
3032
3033 devid = get_device_id(dev);
3034 if (devid < 0)
3035 return;
3036
3037 if (dev_data->domain != NULL)
3038 detach_device(dev);
3039
3040 iommu = amd_iommu_rlookup_table[devid];
3041 if (!iommu)
3042 return;
3043
3044#ifdef CONFIG_IRQ_REMAP
3045 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) &&
3046 (dom->type == IOMMU_DOMAIN_UNMANAGED))
3047 dev_data->use_vapic = 0;
3048#endif
3049
3050 iommu_completion_wait(iommu);
3051}
3052
3053static int amd_iommu_attach_device(struct iommu_domain *dom,
3054 struct device *dev)
3055{
3056 struct protection_domain *domain = to_pdomain(dom);
3057 struct iommu_dev_data *dev_data;
3058 struct amd_iommu *iommu;
3059 int ret;
3060
3061 if (!check_device(dev))
3062 return -EINVAL;
3063
3064 dev_data = dev->archdata.iommu;
3065
3066 iommu = amd_iommu_rlookup_table[dev_data->devid];
3067 if (!iommu)
3068 return -EINVAL;
3069
3070 if (dev_data->domain)
3071 detach_device(dev);
3072
3073 ret = attach_device(dev, domain);
3074
3075#ifdef CONFIG_IRQ_REMAP
3076 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
3077 if (dom->type == IOMMU_DOMAIN_UNMANAGED)
3078 dev_data->use_vapic = 1;
3079 else
3080 dev_data->use_vapic = 0;
3081 }
3082#endif
3083
3084 iommu_completion_wait(iommu);
3085
3086 return ret;
3087}
3088
3089static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
3090 phys_addr_t paddr, size_t page_size, int iommu_prot)
3091{
3092 struct protection_domain *domain = to_pdomain(dom);
3093 int prot = 0;
3094 int ret;
3095
3096 if (domain->mode == PAGE_MODE_NONE)
3097 return -EINVAL;
3098
3099 if (iommu_prot & IOMMU_READ)
3100 prot |= IOMMU_PROT_IR;
3101 if (iommu_prot & IOMMU_WRITE)
3102 prot |= IOMMU_PROT_IW;
3103
3104 mutex_lock(&domain->api_lock);
3105 ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
3106 mutex_unlock(&domain->api_lock);
3107
3108 return ret;
3109}
3110
3111static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
3112 size_t page_size)
3113{
3114 struct protection_domain *domain = to_pdomain(dom);
3115 size_t unmap_size;
3116
3117 if (domain->mode == PAGE_MODE_NONE)
3118 return -EINVAL;
3119
3120 mutex_lock(&domain->api_lock);
3121 unmap_size = iommu_unmap_page(domain, iova, page_size);
3122 mutex_unlock(&domain->api_lock);
3123
3124 domain_flush_tlb_pde(domain);
3125
3126 return unmap_size;
3127}
3128
3129static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
3130 dma_addr_t iova)
3131{
3132 struct protection_domain *domain = to_pdomain(dom);
3133 unsigned long offset_mask, pte_pgsize;
3134 u64 *pte, __pte;
3135
3136 if (domain->mode == PAGE_MODE_NONE)
3137 return iova;
3138
3139 pte = fetch_pte(domain, iova, &pte_pgsize);
3140
3141 if (!pte || !IOMMU_PTE_PRESENT(*pte))
3142 return 0;
3143
3144 offset_mask = pte_pgsize - 1;
3145 __pte = *pte & PM_ADDR_MASK;
3146
3147 return (__pte & ~offset_mask) | (iova & offset_mask);
3148}
3149
3150static bool amd_iommu_capable(enum iommu_cap cap)
3151{
3152 switch (cap) {
3153 case IOMMU_CAP_CACHE_COHERENCY:
3154 return true;
3155 case IOMMU_CAP_INTR_REMAP:
3156 return (irq_remapping_enabled == 1);
3157 case IOMMU_CAP_NOEXEC:
3158 return false;
3159 }
3160
3161 return false;
3162}
3163
3164static void amd_iommu_get_dm_regions(struct device *dev,
3165 struct list_head *head)
3166{
3167 struct unity_map_entry *entry;
3168 int devid;
3169
3170 devid = get_device_id(dev);
3171 if (devid < 0)
3172 return;
3173
3174 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
3175 struct iommu_dm_region *region;
3176
3177 if (devid < entry->devid_start || devid > entry->devid_end)
3178 continue;
3179
3180 region = kzalloc(sizeof(*region), GFP_KERNEL);
3181 if (!region) {
3182 pr_err("Out of memory allocating dm-regions for %s\n",
3183 dev_name(dev));
3184 return;
3185 }
3186
3187 region->start = entry->address_start;
3188 region->length = entry->address_end - entry->address_start;
3189 if (entry->prot & IOMMU_PROT_IR)
3190 region->prot |= IOMMU_READ;
3191 if (entry->prot & IOMMU_PROT_IW)
3192 region->prot |= IOMMU_WRITE;
3193
3194 list_add_tail(®ion->list, head);
3195 }
3196}
3197
3198static void amd_iommu_put_dm_regions(struct device *dev,
3199 struct list_head *head)
3200{
3201 struct iommu_dm_region *entry, *next;
3202
3203 list_for_each_entry_safe(entry, next, head, list)
3204 kfree(entry);
3205}
3206
3207static void amd_iommu_apply_dm_region(struct device *dev,
3208 struct iommu_domain *domain,
3209 struct iommu_dm_region *region)
3210{
3211 struct dma_ops_domain *dma_dom = to_dma_ops_domain(to_pdomain(domain));
3212 unsigned long start, end;
3213
3214 start = IOVA_PFN(region->start);
3215 end = IOVA_PFN(region->start + region->length);
3216
3217 WARN_ON_ONCE(reserve_iova(&dma_dom->iovad, start, end) == NULL);
3218}
3219
3220static const struct iommu_ops amd_iommu_ops = {
3221 .capable = amd_iommu_capable,
3222 .domain_alloc = amd_iommu_domain_alloc,
3223 .domain_free = amd_iommu_domain_free,
3224 .attach_dev = amd_iommu_attach_device,
3225 .detach_dev = amd_iommu_detach_device,
3226 .map = amd_iommu_map,
3227 .unmap = amd_iommu_unmap,
3228 .map_sg = default_iommu_map_sg,
3229 .iova_to_phys = amd_iommu_iova_to_phys,
3230 .add_device = amd_iommu_add_device,
3231 .remove_device = amd_iommu_remove_device,
3232 .device_group = amd_iommu_device_group,
3233 .get_dm_regions = amd_iommu_get_dm_regions,
3234 .put_dm_regions = amd_iommu_put_dm_regions,
3235 .apply_dm_region = amd_iommu_apply_dm_region,
3236 .pgsize_bitmap = AMD_IOMMU_PGSIZES,
3237};
3238
3239/*****************************************************************************
3240 *
3241 * The next functions do a basic initialization of IOMMU for pass through
3242 * mode
3243 *
3244 * In passthrough mode the IOMMU is initialized and enabled but not used for
3245 * DMA-API translation.
3246 *
3247 *****************************************************************************/
3248
3249/* IOMMUv2 specific functions */
3250int amd_iommu_register_ppr_notifier(struct notifier_block *nb)
3251{
3252 return atomic_notifier_chain_register(&ppr_notifier, nb);
3253}
3254EXPORT_SYMBOL(amd_iommu_register_ppr_notifier);
3255
3256int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb)
3257{
3258 return atomic_notifier_chain_unregister(&ppr_notifier, nb);
3259}
3260EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier);
3261
3262void amd_iommu_domain_direct_map(struct iommu_domain *dom)
3263{
3264 struct protection_domain *domain = to_pdomain(dom);
3265 unsigned long flags;
3266
3267 spin_lock_irqsave(&domain->lock, flags);
3268
3269 /* Update data structure */
3270 domain->mode = PAGE_MODE_NONE;
3271 domain->updated = true;
3272
3273 /* Make changes visible to IOMMUs */
3274 update_domain(domain);
3275
3276 /* Page-table is not visible to IOMMU anymore, so free it */
3277 free_pagetable(domain);
3278
3279 spin_unlock_irqrestore(&domain->lock, flags);
3280}
3281EXPORT_SYMBOL(amd_iommu_domain_direct_map);
3282
3283int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids)
3284{
3285 struct protection_domain *domain = to_pdomain(dom);
3286 unsigned long flags;
3287 int levels, ret;
3288
3289 if (pasids <= 0 || pasids > (PASID_MASK + 1))
3290 return -EINVAL;
3291
3292 /* Number of GCR3 table levels required */
3293 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9)
3294 levels += 1;
3295
3296 if (levels > amd_iommu_max_glx_val)
3297 return -EINVAL;
3298
3299 spin_lock_irqsave(&domain->lock, flags);
3300
3301 /*
3302 * Save us all sanity checks whether devices already in the
3303 * domain support IOMMUv2. Just force that the domain has no
3304 * devices attached when it is switched into IOMMUv2 mode.
3305 */
3306 ret = -EBUSY;
3307 if (domain->dev_cnt > 0 || domain->flags & PD_IOMMUV2_MASK)
3308 goto out;
3309
3310 ret = -ENOMEM;
3311 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC);
3312 if (domain->gcr3_tbl == NULL)
3313 goto out;
3314
3315 domain->glx = levels;
3316 domain->flags |= PD_IOMMUV2_MASK;
3317 domain->updated = true;
3318
3319 update_domain(domain);
3320
3321 ret = 0;
3322
3323out:
3324 spin_unlock_irqrestore(&domain->lock, flags);
3325
3326 return ret;
3327}
3328EXPORT_SYMBOL(amd_iommu_domain_enable_v2);
3329
3330static int __flush_pasid(struct protection_domain *domain, int pasid,
3331 u64 address, bool size)
3332{
3333 struct iommu_dev_data *dev_data;
3334 struct iommu_cmd cmd;
3335 int i, ret;
3336
3337 if (!(domain->flags & PD_IOMMUV2_MASK))
3338 return -EINVAL;
3339
3340 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size);
3341
3342 /*
3343 * IOMMU TLB needs to be flushed before Device TLB to
3344 * prevent device TLB refill from IOMMU TLB
3345 */
3346 for (i = 0; i < amd_iommus_present; ++i) {
3347 if (domain->dev_iommu[i] == 0)
3348 continue;
3349
3350 ret = iommu_queue_command(amd_iommus[i], &cmd);
3351 if (ret != 0)
3352 goto out;
3353 }
3354
3355 /* Wait until IOMMU TLB flushes are complete */
3356 domain_flush_complete(domain);
3357
3358 /* Now flush device TLBs */
3359 list_for_each_entry(dev_data, &domain->dev_list, list) {
3360 struct amd_iommu *iommu;
3361 int qdep;
3362
3363 /*
3364 There might be non-IOMMUv2 capable devices in an IOMMUv2
3365 * domain.
3366 */
3367 if (!dev_data->ats.enabled)
3368 continue;
3369
3370 qdep = dev_data->ats.qdep;
3371 iommu = amd_iommu_rlookup_table[dev_data->devid];
3372
3373 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid,
3374 qdep, address, size);
3375
3376 ret = iommu_queue_command(iommu, &cmd);
3377 if (ret != 0)
3378 goto out;
3379 }
3380
3381 /* Wait until all device TLBs are flushed */
3382 domain_flush_complete(domain);
3383
3384 ret = 0;
3385
3386out:
3387
3388 return ret;
3389}
3390
3391static int __amd_iommu_flush_page(struct protection_domain *domain, int pasid,
3392 u64 address)
3393{
3394 return __flush_pasid(domain, pasid, address, false);
3395}
3396
3397int amd_iommu_flush_page(struct iommu_domain *dom, int pasid,
3398 u64 address)
3399{
3400 struct protection_domain *domain = to_pdomain(dom);
3401 unsigned long flags;
3402 int ret;
3403
3404 spin_lock_irqsave(&domain->lock, flags);
3405 ret = __amd_iommu_flush_page(domain, pasid, address);
3406 spin_unlock_irqrestore(&domain->lock, flags);
3407
3408 return ret;
3409}
3410EXPORT_SYMBOL(amd_iommu_flush_page);
3411
3412static int __amd_iommu_flush_tlb(struct protection_domain *domain, int pasid)
3413{
3414 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
3415 true);
3416}
3417
3418int amd_iommu_flush_tlb(struct iommu_domain *dom, int pasid)
3419{
3420 struct protection_domain *domain = to_pdomain(dom);
3421 unsigned long flags;
3422 int ret;
3423
3424 spin_lock_irqsave(&domain->lock, flags);
3425 ret = __amd_iommu_flush_tlb(domain, pasid);
3426 spin_unlock_irqrestore(&domain->lock, flags);
3427
3428 return ret;
3429}
3430EXPORT_SYMBOL(amd_iommu_flush_tlb);
3431
3432static u64 *__get_gcr3_pte(u64 *root, int level, int pasid, bool alloc)
3433{
3434 int index;
3435 u64 *pte;
3436
3437 while (true) {
3438
3439 index = (pasid >> (9 * level)) & 0x1ff;
3440 pte = &root[index];
3441
3442 if (level == 0)
3443 break;
3444
3445 if (!(*pte & GCR3_VALID)) {
3446 if (!alloc)
3447 return NULL;
3448
3449 root = (void *)get_zeroed_page(GFP_ATOMIC);
3450 if (root == NULL)
3451 return NULL;
3452
3453 *pte = __pa(root) | GCR3_VALID;
3454 }
3455
3456 root = __va(*pte & PAGE_MASK);
3457
3458 level -= 1;
3459 }
3460
3461 return pte;
3462}
3463
3464static int __set_gcr3(struct protection_domain *domain, int pasid,
3465 unsigned long cr3)
3466{
3467 u64 *pte;
3468
3469 if (domain->mode != PAGE_MODE_NONE)
3470 return -EINVAL;
3471
3472 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true);
3473 if (pte == NULL)
3474 return -ENOMEM;
3475
3476 *pte = (cr3 & PAGE_MASK) | GCR3_VALID;
3477
3478 return __amd_iommu_flush_tlb(domain, pasid);
3479}
3480
3481static int __clear_gcr3(struct protection_domain *domain, int pasid)
3482{
3483 u64 *pte;
3484
3485 if (domain->mode != PAGE_MODE_NONE)
3486 return -EINVAL;
3487
3488 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false);
3489 if (pte == NULL)
3490 return 0;
3491
3492 *pte = 0;
3493
3494 return __amd_iommu_flush_tlb(domain, pasid);
3495}
3496
3497int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, int pasid,
3498 unsigned long cr3)
3499{
3500 struct protection_domain *domain = to_pdomain(dom);
3501 unsigned long flags;
3502 int ret;
3503
3504 spin_lock_irqsave(&domain->lock, flags);
3505 ret = __set_gcr3(domain, pasid, cr3);
3506 spin_unlock_irqrestore(&domain->lock, flags);
3507
3508 return ret;
3509}
3510EXPORT_SYMBOL(amd_iommu_domain_set_gcr3);
3511
3512int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, int pasid)
3513{
3514 struct protection_domain *domain = to_pdomain(dom);
3515 unsigned long flags;
3516 int ret;
3517
3518 spin_lock_irqsave(&domain->lock, flags);
3519 ret = __clear_gcr3(domain, pasid);
3520 spin_unlock_irqrestore(&domain->lock, flags);
3521
3522 return ret;
3523}
3524EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3);
3525
3526int amd_iommu_complete_ppr(struct pci_dev *pdev, int pasid,
3527 int status, int tag)
3528{
3529 struct iommu_dev_data *dev_data;
3530 struct amd_iommu *iommu;
3531 struct iommu_cmd cmd;
3532
3533 dev_data = get_dev_data(&pdev->dev);
3534 iommu = amd_iommu_rlookup_table[dev_data->devid];
3535
3536 build_complete_ppr(&cmd, dev_data->devid, pasid, status,
3537 tag, dev_data->pri_tlp);
3538
3539 return iommu_queue_command(iommu, &cmd);
3540}
3541EXPORT_SYMBOL(amd_iommu_complete_ppr);
3542
3543struct iommu_domain *amd_iommu_get_v2_domain(struct pci_dev *pdev)
3544{
3545 struct protection_domain *pdomain;
3546
3547 pdomain = get_domain(&pdev->dev);
3548 if (IS_ERR(pdomain))
3549 return NULL;
3550
3551 /* Only return IOMMUv2 domains */
3552 if (!(pdomain->flags & PD_IOMMUV2_MASK))
3553 return NULL;
3554
3555 return &pdomain->domain;
3556}
3557EXPORT_SYMBOL(amd_iommu_get_v2_domain);
3558
3559void amd_iommu_enable_device_erratum(struct pci_dev *pdev, u32 erratum)
3560{
3561 struct iommu_dev_data *dev_data;
3562
3563 if (!amd_iommu_v2_supported())
3564 return;
3565
3566 dev_data = get_dev_data(&pdev->dev);
3567 dev_data->errata |= (1 << erratum);
3568}
3569EXPORT_SYMBOL(amd_iommu_enable_device_erratum);
3570
3571int amd_iommu_device_info(struct pci_dev *pdev,
3572 struct amd_iommu_device_info *info)
3573{
3574 int max_pasids;
3575 int pos;
3576
3577 if (pdev == NULL || info == NULL)
3578 return -EINVAL;
3579
3580 if (!amd_iommu_v2_supported())
3581 return -EINVAL;
3582
3583 memset(info, 0, sizeof(*info));
3584
3585 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_ATS);
3586 if (pos)
3587 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP;
3588
3589 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI);
3590 if (pos)
3591 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP;
3592
3593 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID);
3594 if (pos) {
3595 int features;
3596
3597 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1));
3598 max_pasids = min(max_pasids, (1 << 20));
3599
3600 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
3601 info->max_pasids = min(pci_max_pasids(pdev), max_pasids);
3602
3603 features = pci_pasid_features(pdev);
3604 if (features & PCI_PASID_CAP_EXEC)
3605 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP;
3606 if (features & PCI_PASID_CAP_PRIV)
3607 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP;
3608 }
3609
3610 return 0;
3611}
3612EXPORT_SYMBOL(amd_iommu_device_info);
3613
3614#ifdef CONFIG_IRQ_REMAP
3615
3616/*****************************************************************************
3617 *
3618 * Interrupt Remapping Implementation
3619 *
3620 *****************************************************************************/
3621
3622static struct irq_chip amd_ir_chip;
3623
3624#define DTE_IRQ_PHYS_ADDR_MASK (((1ULL << 45)-1) << 6)
3625#define DTE_IRQ_REMAP_INTCTL (2ULL << 60)
3626#define DTE_IRQ_TABLE_LEN (8ULL << 1)
3627#define DTE_IRQ_REMAP_ENABLE 1ULL
3628
3629static void set_dte_irq_entry(u16 devid, struct irq_remap_table *table)
3630{
3631 u64 dte;
3632
3633 dte = amd_iommu_dev_table[devid].data[2];
3634 dte &= ~DTE_IRQ_PHYS_ADDR_MASK;
3635 dte |= virt_to_phys(table->table);
3636 dte |= DTE_IRQ_REMAP_INTCTL;
3637 dte |= DTE_IRQ_TABLE_LEN;
3638 dte |= DTE_IRQ_REMAP_ENABLE;
3639
3640 amd_iommu_dev_table[devid].data[2] = dte;
3641}
3642
3643static struct irq_remap_table *get_irq_table(u16 devid, bool ioapic)
3644{
3645 struct irq_remap_table *table = NULL;
3646 struct amd_iommu *iommu;
3647 unsigned long flags;
3648 u16 alias;
3649
3650 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
3651
3652 iommu = amd_iommu_rlookup_table[devid];
3653 if (!iommu)
3654 goto out_unlock;
3655
3656 table = irq_lookup_table[devid];
3657 if (table)
3658 goto out_unlock;
3659
3660 alias = amd_iommu_alias_table[devid];
3661 table = irq_lookup_table[alias];
3662 if (table) {
3663 irq_lookup_table[devid] = table;
3664 set_dte_irq_entry(devid, table);
3665 iommu_flush_dte(iommu, devid);
3666 goto out;
3667 }
3668
3669 /* Nothing there yet, allocate new irq remapping table */
3670 table = kzalloc(sizeof(*table), GFP_ATOMIC);
3671 if (!table)
3672 goto out_unlock;
3673
3674 /* Initialize table spin-lock */
3675 spin_lock_init(&table->lock);
3676
3677 if (ioapic)
3678 /* Keep the first 32 indexes free for IOAPIC interrupts */
3679 table->min_index = 32;
3680
3681 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_ATOMIC);
3682 if (!table->table) {
3683 kfree(table);
3684 table = NULL;
3685 goto out_unlock;
3686 }
3687
3688 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
3689 memset(table->table, 0,
3690 MAX_IRQS_PER_TABLE * sizeof(u32));
3691 else
3692 memset(table->table, 0,
3693 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2)));
3694
3695 if (ioapic) {
3696 int i;
3697
3698 for (i = 0; i < 32; ++i)
3699 iommu->irte_ops->set_allocated(table, i);
3700 }
3701
3702 irq_lookup_table[devid] = table;
3703 set_dte_irq_entry(devid, table);
3704 iommu_flush_dte(iommu, devid);
3705 if (devid != alias) {
3706 irq_lookup_table[alias] = table;
3707 set_dte_irq_entry(alias, table);
3708 iommu_flush_dte(iommu, alias);
3709 }
3710
3711out:
3712 iommu_completion_wait(iommu);
3713
3714out_unlock:
3715 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
3716
3717 return table;
3718}
3719
3720static int alloc_irq_index(u16 devid, int count)
3721{
3722 struct irq_remap_table *table;
3723 unsigned long flags;
3724 int index, c;
3725 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
3726
3727 if (!iommu)
3728 return -ENODEV;
3729
3730 table = get_irq_table(devid, false);
3731 if (!table)
3732 return -ENODEV;
3733
3734 spin_lock_irqsave(&table->lock, flags);
3735
3736 /* Scan table for free entries */
3737 for (c = 0, index = table->min_index;
3738 index < MAX_IRQS_PER_TABLE;
3739 ++index) {
3740 if (!iommu->irte_ops->is_allocated(table, index))
3741 c += 1;
3742 else
3743 c = 0;
3744
3745 if (c == count) {
3746 for (; c != 0; --c)
3747 iommu->irte_ops->set_allocated(table, index - c + 1);
3748
3749 index -= count - 1;
3750 goto out;
3751 }
3752 }
3753
3754 index = -ENOSPC;
3755
3756out:
3757 spin_unlock_irqrestore(&table->lock, flags);
3758
3759 return index;
3760}
3761
3762static int modify_irte_ga(u16 devid, int index, struct irte_ga *irte,
3763 struct amd_ir_data *data)
3764{
3765 struct irq_remap_table *table;
3766 struct amd_iommu *iommu;
3767 unsigned long flags;
3768 struct irte_ga *entry;
3769
3770 iommu = amd_iommu_rlookup_table[devid];
3771 if (iommu == NULL)
3772 return -EINVAL;
3773
3774 table = get_irq_table(devid, false);
3775 if (!table)
3776 return -ENOMEM;
3777
3778 spin_lock_irqsave(&table->lock, flags);
3779
3780 entry = (struct irte_ga *)table->table;
3781 entry = &entry[index];
3782 entry->lo.fields_remap.valid = 0;
3783 entry->hi.val = irte->hi.val;
3784 entry->lo.val = irte->lo.val;
3785 entry->lo.fields_remap.valid = 1;
3786 if (data)
3787 data->ref = entry;
3788
3789 spin_unlock_irqrestore(&table->lock, flags);
3790
3791 iommu_flush_irt(iommu, devid);
3792 iommu_completion_wait(iommu);
3793
3794 return 0;
3795}
3796
3797static int modify_irte(u16 devid, int index, union irte *irte)
3798{
3799 struct irq_remap_table *table;
3800 struct amd_iommu *iommu;
3801 unsigned long flags;
3802
3803 iommu = amd_iommu_rlookup_table[devid];
3804 if (iommu == NULL)
3805 return -EINVAL;
3806
3807 table = get_irq_table(devid, false);
3808 if (!table)
3809 return -ENOMEM;
3810
3811 spin_lock_irqsave(&table->lock, flags);
3812 table->table[index] = irte->val;
3813 spin_unlock_irqrestore(&table->lock, flags);
3814
3815 iommu_flush_irt(iommu, devid);
3816 iommu_completion_wait(iommu);
3817
3818 return 0;
3819}
3820
3821static void free_irte(u16 devid, int index)
3822{
3823 struct irq_remap_table *table;
3824 struct amd_iommu *iommu;
3825 unsigned long flags;
3826
3827 iommu = amd_iommu_rlookup_table[devid];
3828 if (iommu == NULL)
3829 return;
3830
3831 table = get_irq_table(devid, false);
3832 if (!table)
3833 return;
3834
3835 spin_lock_irqsave(&table->lock, flags);
3836 iommu->irte_ops->clear_allocated(table, index);
3837 spin_unlock_irqrestore(&table->lock, flags);
3838
3839 iommu_flush_irt(iommu, devid);
3840 iommu_completion_wait(iommu);
3841}
3842
3843static void irte_prepare(void *entry,
3844 u32 delivery_mode, u32 dest_mode,
3845 u8 vector, u32 dest_apicid, int devid)
3846{
3847 union irte *irte = (union irte *) entry;
3848
3849 irte->val = 0;
3850 irte->fields.vector = vector;
3851 irte->fields.int_type = delivery_mode;
3852 irte->fields.destination = dest_apicid;
3853 irte->fields.dm = dest_mode;
3854 irte->fields.valid = 1;
3855}
3856
3857static void irte_ga_prepare(void *entry,
3858 u32 delivery_mode, u32 dest_mode,
3859 u8 vector, u32 dest_apicid, int devid)
3860{
3861 struct irte_ga *irte = (struct irte_ga *) entry;
3862 struct iommu_dev_data *dev_data = search_dev_data(devid);
3863
3864 irte->lo.val = 0;
3865 irte->hi.val = 0;
3866 irte->lo.fields_remap.guest_mode = dev_data ? dev_data->use_vapic : 0;
3867 irte->lo.fields_remap.int_type = delivery_mode;
3868 irte->lo.fields_remap.dm = dest_mode;
3869 irte->hi.fields.vector = vector;
3870 irte->lo.fields_remap.destination = dest_apicid;
3871 irte->lo.fields_remap.valid = 1;
3872}
3873
3874static void irte_activate(void *entry, u16 devid, u16 index)
3875{
3876 union irte *irte = (union irte *) entry;
3877
3878 irte->fields.valid = 1;
3879 modify_irte(devid, index, irte);
3880}
3881
3882static void irte_ga_activate(void *entry, u16 devid, u16 index)
3883{
3884 struct irte_ga *irte = (struct irte_ga *) entry;
3885
3886 irte->lo.fields_remap.valid = 1;
3887 modify_irte_ga(devid, index, irte, NULL);
3888}
3889
3890static void irte_deactivate(void *entry, u16 devid, u16 index)
3891{
3892 union irte *irte = (union irte *) entry;
3893
3894 irte->fields.valid = 0;
3895 modify_irte(devid, index, irte);
3896}
3897
3898static void irte_ga_deactivate(void *entry, u16 devid, u16 index)
3899{
3900 struct irte_ga *irte = (struct irte_ga *) entry;
3901
3902 irte->lo.fields_remap.valid = 0;
3903 modify_irte_ga(devid, index, irte, NULL);
3904}
3905
3906static void irte_set_affinity(void *entry, u16 devid, u16 index,
3907 u8 vector, u32 dest_apicid)
3908{
3909 union irte *irte = (union irte *) entry;
3910
3911 irte->fields.vector = vector;
3912 irte->fields.destination = dest_apicid;
3913 modify_irte(devid, index, irte);
3914}
3915
3916static void irte_ga_set_affinity(void *entry, u16 devid, u16 index,
3917 u8 vector, u32 dest_apicid)
3918{
3919 struct irte_ga *irte = (struct irte_ga *) entry;
3920 struct iommu_dev_data *dev_data = search_dev_data(devid);
3921
3922 if (!dev_data || !dev_data->use_vapic) {
3923 irte->hi.fields.vector = vector;
3924 irte->lo.fields_remap.destination = dest_apicid;
3925 irte->lo.fields_remap.guest_mode = 0;
3926 modify_irte_ga(devid, index, irte, NULL);
3927 }
3928}
3929
3930#define IRTE_ALLOCATED (~1U)
3931static void irte_set_allocated(struct irq_remap_table *table, int index)
3932{
3933 table->table[index] = IRTE_ALLOCATED;
3934}
3935
3936static void irte_ga_set_allocated(struct irq_remap_table *table, int index)
3937{
3938 struct irte_ga *ptr = (struct irte_ga *)table->table;
3939 struct irte_ga *irte = &ptr[index];
3940
3941 memset(&irte->lo.val, 0, sizeof(u64));
3942 memset(&irte->hi.val, 0, sizeof(u64));
3943 irte->hi.fields.vector = 0xff;
3944}
3945
3946static bool irte_is_allocated(struct irq_remap_table *table, int index)
3947{
3948 union irte *ptr = (union irte *)table->table;
3949 union irte *irte = &ptr[index];
3950
3951 return irte->val != 0;
3952}
3953
3954static bool irte_ga_is_allocated(struct irq_remap_table *table, int index)
3955{
3956 struct irte_ga *ptr = (struct irte_ga *)table->table;
3957 struct irte_ga *irte = &ptr[index];
3958
3959 return irte->hi.fields.vector != 0;
3960}
3961
3962static void irte_clear_allocated(struct irq_remap_table *table, int index)
3963{
3964 table->table[index] = 0;
3965}
3966
3967static void irte_ga_clear_allocated(struct irq_remap_table *table, int index)
3968{
3969 struct irte_ga *ptr = (struct irte_ga *)table->table;
3970 struct irte_ga *irte = &ptr[index];
3971
3972 memset(&irte->lo.val, 0, sizeof(u64));
3973 memset(&irte->hi.val, 0, sizeof(u64));
3974}
3975
3976static int get_devid(struct irq_alloc_info *info)
3977{
3978 int devid = -1;
3979
3980 switch (info->type) {
3981 case X86_IRQ_ALLOC_TYPE_IOAPIC:
3982 devid = get_ioapic_devid(info->ioapic_id);
3983 break;
3984 case X86_IRQ_ALLOC_TYPE_HPET:
3985 devid = get_hpet_devid(info->hpet_id);
3986 break;
3987 case X86_IRQ_ALLOC_TYPE_MSI:
3988 case X86_IRQ_ALLOC_TYPE_MSIX:
3989 devid = get_device_id(&info->msi_dev->dev);
3990 break;
3991 default:
3992 BUG_ON(1);
3993 break;
3994 }
3995
3996 return devid;
3997}
3998
3999static struct irq_domain *get_ir_irq_domain(struct irq_alloc_info *info)
4000{
4001 struct amd_iommu *iommu;
4002 int devid;
4003
4004 if (!info)
4005 return NULL;
4006
4007 devid = get_devid(info);
4008 if (devid >= 0) {
4009 iommu = amd_iommu_rlookup_table[devid];
4010 if (iommu)
4011 return iommu->ir_domain;
4012 }
4013
4014 return NULL;
4015}
4016
4017static struct irq_domain *get_irq_domain(struct irq_alloc_info *info)
4018{
4019 struct amd_iommu *iommu;
4020 int devid;
4021
4022 if (!info)
4023 return NULL;
4024
4025 switch (info->type) {
4026 case X86_IRQ_ALLOC_TYPE_MSI:
4027 case X86_IRQ_ALLOC_TYPE_MSIX:
4028 devid = get_device_id(&info->msi_dev->dev);
4029 if (devid < 0)
4030 return NULL;
4031
4032 iommu = amd_iommu_rlookup_table[devid];
4033 if (iommu)
4034 return iommu->msi_domain;
4035 break;
4036 default:
4037 break;
4038 }
4039
4040 return NULL;
4041}
4042
4043struct irq_remap_ops amd_iommu_irq_ops = {
4044 .prepare = amd_iommu_prepare,
4045 .enable = amd_iommu_enable,
4046 .disable = amd_iommu_disable,
4047 .reenable = amd_iommu_reenable,
4048 .enable_faulting = amd_iommu_enable_faulting,
4049 .get_ir_irq_domain = get_ir_irq_domain,
4050 .get_irq_domain = get_irq_domain,
4051};
4052
4053static void irq_remapping_prepare_irte(struct amd_ir_data *data,
4054 struct irq_cfg *irq_cfg,
4055 struct irq_alloc_info *info,
4056 int devid, int index, int sub_handle)
4057{
4058 struct irq_2_irte *irte_info = &data->irq_2_irte;
4059 struct msi_msg *msg = &data->msi_entry;
4060 struct IO_APIC_route_entry *entry;
4061 struct amd_iommu *iommu = amd_iommu_rlookup_table[devid];
4062
4063 if (!iommu)
4064 return;
4065
4066 data->irq_2_irte.devid = devid;
4067 data->irq_2_irte.index = index + sub_handle;
4068 iommu->irte_ops->prepare(data->entry, apic->irq_delivery_mode,
4069 apic->irq_dest_mode, irq_cfg->vector,
4070 irq_cfg->dest_apicid, devid);
4071
4072 switch (info->type) {
4073 case X86_IRQ_ALLOC_TYPE_IOAPIC:
4074 /* Setup IOAPIC entry */
4075 entry = info->ioapic_entry;
4076 info->ioapic_entry = NULL;
4077 memset(entry, 0, sizeof(*entry));
4078 entry->vector = index;
4079 entry->mask = 0;
4080 entry->trigger = info->ioapic_trigger;
4081 entry->polarity = info->ioapic_polarity;
4082 /* Mask level triggered irqs. */
4083 if (info->ioapic_trigger)
4084 entry->mask = 1;
4085 break;
4086
4087 case X86_IRQ_ALLOC_TYPE_HPET:
4088 case X86_IRQ_ALLOC_TYPE_MSI:
4089 case X86_IRQ_ALLOC_TYPE_MSIX:
4090 msg->address_hi = MSI_ADDR_BASE_HI;
4091 msg->address_lo = MSI_ADDR_BASE_LO;
4092 msg->data = irte_info->index;
4093 break;
4094
4095 default:
4096 BUG_ON(1);
4097 break;
4098 }
4099}
4100
4101struct amd_irte_ops irte_32_ops = {
4102 .prepare = irte_prepare,
4103 .activate = irte_activate,
4104 .deactivate = irte_deactivate,
4105 .set_affinity = irte_set_affinity,
4106 .set_allocated = irte_set_allocated,
4107 .is_allocated = irte_is_allocated,
4108 .clear_allocated = irte_clear_allocated,
4109};
4110
4111struct amd_irte_ops irte_128_ops = {
4112 .prepare = irte_ga_prepare,
4113 .activate = irte_ga_activate,
4114 .deactivate = irte_ga_deactivate,
4115 .set_affinity = irte_ga_set_affinity,
4116 .set_allocated = irte_ga_set_allocated,
4117 .is_allocated = irte_ga_is_allocated,
4118 .clear_allocated = irte_ga_clear_allocated,
4119};
4120
4121static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq,
4122 unsigned int nr_irqs, void *arg)
4123{
4124 struct irq_alloc_info *info = arg;
4125 struct irq_data *irq_data;
4126 struct amd_ir_data *data = NULL;
4127 struct irq_cfg *cfg;
4128 int i, ret, devid;
4129 int index = -1;
4130
4131 if (!info)
4132 return -EINVAL;
4133 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_MSI &&
4134 info->type != X86_IRQ_ALLOC_TYPE_MSIX)
4135 return -EINVAL;
4136
4137 /*
4138 * With IRQ remapping enabled, don't need contiguous CPU vectors
4139 * to support multiple MSI interrupts.
4140 */
4141 if (info->type == X86_IRQ_ALLOC_TYPE_MSI)
4142 info->flags &= ~X86_IRQ_ALLOC_CONTIGUOUS_VECTORS;
4143
4144 devid = get_devid(info);
4145 if (devid < 0)
4146 return -EINVAL;
4147
4148 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg);
4149 if (ret < 0)
4150 return ret;
4151
4152 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) {
4153 if (get_irq_table(devid, true))
4154 index = info->ioapic_pin;
4155 else
4156 ret = -ENOMEM;
4157 } else {
4158 index = alloc_irq_index(devid, nr_irqs);
4159 }
4160 if (index < 0) {
4161 pr_warn("Failed to allocate IRTE\n");
4162 ret = index;
4163 goto out_free_parent;
4164 }
4165
4166 for (i = 0; i < nr_irqs; i++) {
4167 irq_data = irq_domain_get_irq_data(domain, virq + i);
4168 cfg = irqd_cfg(irq_data);
4169 if (!irq_data || !cfg) {
4170 ret = -EINVAL;
4171 goto out_free_data;
4172 }
4173
4174 ret = -ENOMEM;
4175 data = kzalloc(sizeof(*data), GFP_KERNEL);
4176 if (!data)
4177 goto out_free_data;
4178
4179 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir))
4180 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL);
4181 else
4182 data->entry = kzalloc(sizeof(struct irte_ga),
4183 GFP_KERNEL);
4184 if (!data->entry) {
4185 kfree(data);
4186 goto out_free_data;
4187 }
4188
4189 irq_data->hwirq = (devid << 16) + i;
4190 irq_data->chip_data = data;
4191 irq_data->chip = &amd_ir_chip;
4192 irq_remapping_prepare_irte(data, cfg, info, devid, index, i);
4193 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT);
4194 }
4195
4196 return 0;
4197
4198out_free_data:
4199 for (i--; i >= 0; i--) {
4200 irq_data = irq_domain_get_irq_data(domain, virq + i);
4201 if (irq_data)
4202 kfree(irq_data->chip_data);
4203 }
4204 for (i = 0; i < nr_irqs; i++)
4205 free_irte(devid, index + i);
4206out_free_parent:
4207 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4208 return ret;
4209}
4210
4211static void irq_remapping_free(struct irq_domain *domain, unsigned int virq,
4212 unsigned int nr_irqs)
4213{
4214 struct irq_2_irte *irte_info;
4215 struct irq_data *irq_data;
4216 struct amd_ir_data *data;
4217 int i;
4218
4219 for (i = 0; i < nr_irqs; i++) {
4220 irq_data = irq_domain_get_irq_data(domain, virq + i);
4221 if (irq_data && irq_data->chip_data) {
4222 data = irq_data->chip_data;
4223 irte_info = &data->irq_2_irte;
4224 free_irte(irte_info->devid, irte_info->index);
4225 kfree(data->entry);
4226 kfree(data);
4227 }
4228 }
4229 irq_domain_free_irqs_common(domain, virq, nr_irqs);
4230}
4231
4232static void irq_remapping_activate(struct irq_domain *domain,
4233 struct irq_data *irq_data)
4234{
4235 struct amd_ir_data *data = irq_data->chip_data;
4236 struct irq_2_irte *irte_info = &data->irq_2_irte;
4237 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4238
4239 if (iommu)
4240 iommu->irte_ops->activate(data->entry, irte_info->devid,
4241 irte_info->index);
4242}
4243
4244static void irq_remapping_deactivate(struct irq_domain *domain,
4245 struct irq_data *irq_data)
4246{
4247 struct amd_ir_data *data = irq_data->chip_data;
4248 struct irq_2_irte *irte_info = &data->irq_2_irte;
4249 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4250
4251 if (iommu)
4252 iommu->irte_ops->deactivate(data->entry, irte_info->devid,
4253 irte_info->index);
4254}
4255
4256static struct irq_domain_ops amd_ir_domain_ops = {
4257 .alloc = irq_remapping_alloc,
4258 .free = irq_remapping_free,
4259 .activate = irq_remapping_activate,
4260 .deactivate = irq_remapping_deactivate,
4261};
4262
4263static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
4264{
4265 struct amd_iommu *iommu;
4266 struct amd_iommu_pi_data *pi_data = vcpu_info;
4267 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
4268 struct amd_ir_data *ir_data = data->chip_data;
4269 struct irte_ga *irte = (struct irte_ga *) ir_data->entry;
4270 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4271 struct iommu_dev_data *dev_data = search_dev_data(irte_info->devid);
4272
4273 /* Note:
4274 * This device has never been set up for guest mode.
4275 * we should not modify the IRTE
4276 */
4277 if (!dev_data || !dev_data->use_vapic)
4278 return 0;
4279
4280 pi_data->ir_data = ir_data;
4281
4282 /* Note:
4283 * SVM tries to set up for VAPIC mode, but we are in
4284 * legacy mode. So, we force legacy mode instead.
4285 */
4286 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) {
4287 pr_debug("AMD-Vi: %s: Fall back to using intr legacy remap\n",
4288 __func__);
4289 pi_data->is_guest_mode = false;
4290 }
4291
4292 iommu = amd_iommu_rlookup_table[irte_info->devid];
4293 if (iommu == NULL)
4294 return -EINVAL;
4295
4296 pi_data->prev_ga_tag = ir_data->cached_ga_tag;
4297 if (pi_data->is_guest_mode) {
4298 /* Setting */
4299 irte->hi.fields.ga_root_ptr = (pi_data->base >> 12);
4300 irte->hi.fields.vector = vcpu_pi_info->vector;
4301 irte->lo.fields_vapic.guest_mode = 1;
4302 irte->lo.fields_vapic.ga_tag = pi_data->ga_tag;
4303
4304 ir_data->cached_ga_tag = pi_data->ga_tag;
4305 } else {
4306 /* Un-Setting */
4307 struct irq_cfg *cfg = irqd_cfg(data);
4308
4309 irte->hi.val = 0;
4310 irte->lo.val = 0;
4311 irte->hi.fields.vector = cfg->vector;
4312 irte->lo.fields_remap.guest_mode = 0;
4313 irte->lo.fields_remap.destination = cfg->dest_apicid;
4314 irte->lo.fields_remap.int_type = apic->irq_delivery_mode;
4315 irte->lo.fields_remap.dm = apic->irq_dest_mode;
4316
4317 /*
4318 * This communicates the ga_tag back to the caller
4319 * so that it can do all the necessary clean up.
4320 */
4321 ir_data->cached_ga_tag = 0;
4322 }
4323
4324 return modify_irte_ga(irte_info->devid, irte_info->index, irte, ir_data);
4325}
4326
4327static int amd_ir_set_affinity(struct irq_data *data,
4328 const struct cpumask *mask, bool force)
4329{
4330 struct amd_ir_data *ir_data = data->chip_data;
4331 struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
4332 struct irq_cfg *cfg = irqd_cfg(data);
4333 struct irq_data *parent = data->parent_data;
4334 struct amd_iommu *iommu = amd_iommu_rlookup_table[irte_info->devid];
4335 int ret;
4336
4337 if (!iommu)
4338 return -ENODEV;
4339
4340 ret = parent->chip->irq_set_affinity(parent, mask, force);
4341 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE)
4342 return ret;
4343
4344 /*
4345 * Atomically updates the IRTE with the new destination, vector
4346 * and flushes the interrupt entry cache.
4347 */
4348 iommu->irte_ops->set_affinity(ir_data->entry, irte_info->devid,
4349 irte_info->index, cfg->vector, cfg->dest_apicid);
4350
4351 /*
4352 * After this point, all the interrupts will start arriving
4353 * at the new destination. So, time to cleanup the previous
4354 * vector allocation.
4355 */
4356 send_cleanup_vector(cfg);
4357
4358 return IRQ_SET_MASK_OK_DONE;
4359}
4360
4361static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg)
4362{
4363 struct amd_ir_data *ir_data = irq_data->chip_data;
4364
4365 *msg = ir_data->msi_entry;
4366}
4367
4368static struct irq_chip amd_ir_chip = {
4369 .irq_ack = ir_ack_apic_edge,
4370 .irq_set_affinity = amd_ir_set_affinity,
4371 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity,
4372 .irq_compose_msi_msg = ir_compose_msi_msg,
4373};
4374
4375int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
4376{
4377 iommu->ir_domain = irq_domain_add_tree(NULL, &amd_ir_domain_ops, iommu);
4378 if (!iommu->ir_domain)
4379 return -ENOMEM;
4380
4381 iommu->ir_domain->parent = arch_get_ir_parent_domain();
4382 iommu->msi_domain = arch_create_msi_irq_domain(iommu->ir_domain);
4383
4384 return 0;
4385}
4386
4387int amd_iommu_update_ga(int cpu, bool is_run, void *data)
4388{
4389 unsigned long flags;
4390 struct amd_iommu *iommu;
4391 struct irq_remap_table *irt;
4392 struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
4393 int devid = ir_data->irq_2_irte.devid;
4394 struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
4395 struct irte_ga *ref = (struct irte_ga *) ir_data->ref;
4396
4397 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
4398 !ref || !entry || !entry->lo.fields_vapic.guest_mode)
4399 return 0;
4400
4401 iommu = amd_iommu_rlookup_table[devid];
4402 if (!iommu)
4403 return -ENODEV;
4404
4405 irt = get_irq_table(devid, false);
4406 if (!irt)
4407 return -ENODEV;
4408
4409 spin_lock_irqsave(&irt->lock, flags);
4410
4411 if (ref->lo.fields_vapic.guest_mode) {
4412 if (cpu >= 0)
4413 ref->lo.fields_vapic.destination = cpu;
4414 ref->lo.fields_vapic.is_run = is_run;
4415 barrier();
4416 }
4417
4418 spin_unlock_irqrestore(&irt->lock, flags);
4419
4420 iommu_flush_irt(iommu, devid);
4421 iommu_completion_wait(iommu);
4422 return 0;
4423}
4424EXPORT_SYMBOL(amd_iommu_update_ga);
4425#endif