blob: a14f8dc23462229c8ba34d6aec18a14eb3fcaf7a [file] [log] [blame]
Joerg Roedelb6c02712008-06-26 21:27:53 +02001/*
Joerg Roedel5d0d7152010-10-13 11:13:21 +02002 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc.
Joerg Roedelb6c02712008-06-26 21:27:53 +02003 * 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>
Joerg Roedelcb41ed82011-04-05 11:00:53 +020021#include <linux/pci-ats.h>
Akinobu Mitaa66022c2009-12-15 16:48:28 -080022#include <linux/bitmap.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090023#include <linux/slab.h>
Joerg Roedel7f265082008-12-12 13:50:21 +010024#include <linux/debugfs.h>
Joerg Roedelb6c02712008-06-26 21:27:53 +020025#include <linux/scatterlist.h>
FUJITA Tomonori51491362009-01-05 23:47:25 +090026#include <linux/dma-mapping.h>
Joerg Roedelb6c02712008-06-26 21:27:53 +020027#include <linux/iommu-helper.h>
Joerg Roedelc156e342008-12-02 18:13:27 +010028#include <linux/iommu.h>
Joerg Roedel815b33f2011-04-06 17:26:49 +020029#include <linux/delay.h>
Joerg Roedel403f81d2011-06-14 16:44:25 +020030#include <linux/amd-iommu.h>
Joerg Roedel17f5b562011-07-06 17:14:44 +020031#include <asm/msidef.h>
Joerg Roedelb6c02712008-06-26 21:27:53 +020032#include <asm/proto.h>
FUJITA Tomonori46a7fa22008-07-11 10:23:42 +090033#include <asm/iommu.h>
Joerg Roedel1d9b16d2008-11-27 18:39:15 +010034#include <asm/gart.h>
Joerg Roedel27c21272011-05-30 15:56:24 +020035#include <asm/dma.h>
Joerg Roedel403f81d2011-06-14 16:44:25 +020036
37#include "amd_iommu_proto.h"
38#include "amd_iommu_types.h"
Joerg Roedelb6c02712008-06-26 21:27:53 +020039
40#define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28))
41
Joerg Roedel815b33f2011-04-06 17:26:49 +020042#define LOOP_TIMEOUT 100000
Joerg Roedel136f78a2008-07-11 17:14:27 +020043
Joerg Roedelb6c02712008-06-26 21:27:53 +020044static DEFINE_RWLOCK(amd_iommu_devtable_lock);
45
Joerg Roedelbd60b732008-09-11 10:24:48 +020046/* A list of preallocated protection domains */
47static LIST_HEAD(iommu_pd_list);
48static DEFINE_SPINLOCK(iommu_pd_list_lock);
49
Joerg Roedel8fa5f802011-06-09 12:24:45 +020050/* List of all available dev_data structures */
51static LIST_HEAD(dev_data_list);
52static DEFINE_SPINLOCK(dev_data_list_lock);
53
Joerg Roedel0feae532009-08-26 15:26:30 +020054/*
55 * Domain for untranslated devices - only allocated
56 * if iommu=pt passed on kernel cmd line.
57 */
58static struct protection_domain *pt_domain;
59
Joerg Roedel26961ef2008-12-03 17:00:17 +010060static struct iommu_ops amd_iommu_ops;
Joerg Roedel26961ef2008-12-03 17:00:17 +010061
Joerg Roedel431b2a22008-07-11 17:14:22 +020062/*
63 * general struct to manage commands send to an IOMMU
64 */
Joerg Roedeld6449532008-07-11 17:14:28 +020065struct iommu_cmd {
Joerg Roedelb6c02712008-06-26 21:27:53 +020066 u32 data[4];
67};
68
Joerg Roedel04bfdd82009-09-02 16:00:23 +020069static void update_domain(struct protection_domain *domain);
Chris Wrightc1eee672009-05-21 00:56:58 -070070
Joerg Roedel15898bb2009-11-24 15:39:42 +010071/****************************************************************************
72 *
73 * Helper functions
74 *
75 ****************************************************************************/
76
Joerg Roedelf62dda62011-06-09 12:55:35 +020077static struct iommu_dev_data *alloc_dev_data(u16 devid)
Joerg Roedel8fa5f802011-06-09 12:24:45 +020078{
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
Joerg Roedelf62dda62011-06-09 12:55:35 +020086 dev_data->devid = devid;
Joerg Roedel8fa5f802011-06-09 12:24:45 +020087 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
Joerg Roedel3b03bb72011-06-09 18:53:25 +0200107static 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
Joerg Roedel15898bb2009-11-24 15:39:42 +0100138static 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
Joerg Roedel657cbb62009-11-23 15:26:46 +0100145static struct iommu_dev_data *get_dev_data(struct device *dev)
146{
147 return dev->archdata.iommu;
148}
149
Joerg Roedel71c70982009-11-24 16:43:06 +0100150/*
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
Joerg Roedel98fc5a62009-11-24 17:19:23 +0100178/*
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 */
Julia Lawall339d3262010-02-06 09:42:39 +0100190 if (dev->bus != &pci_bus_type)
Joerg Roedel98fc5a62009-11-24 17:19:23 +0100191 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
Joerg Roedel657cbb62009-11-23 15:26:46 +0100205static int iommu_init_device(struct device *dev)
206{
207 struct iommu_dev_data *dev_data;
Joerg Roedel8fa5f802011-06-09 12:24:45 +0200208 u16 alias;
Joerg Roedel657cbb62009-11-23 15:26:46 +0100209
210 if (dev->archdata.iommu)
211 return 0;
212
Joerg Roedel3b03bb72011-06-09 18:53:25 +0200213 dev_data = find_dev_data(get_device_id(dev));
Joerg Roedel657cbb62009-11-23 15:26:46 +0100214 if (!dev_data)
215 return -ENOMEM;
216
Joerg Roedelf62dda62011-06-09 12:55:35 +0200217 alias = amd_iommu_alias_table[dev_data->devid];
Joerg Roedel2b02b092011-06-09 17:48:39 +0200218 if (alias != dev_data->devid) {
Joerg Roedel71f77582011-06-09 19:03:15 +0200219 struct iommu_dev_data *alias_data;
Joerg Roedelb00d3bc2009-11-26 15:35:33 +0100220
Joerg Roedel71f77582011-06-09 19:03:15 +0200221 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));
Joerg Roedel2b02b092011-06-09 17:48:39 +0200225 free_dev_data(dev_data);
226 return -ENOTSUPP;
227 }
Joerg Roedel71f77582011-06-09 19:03:15 +0200228 dev_data->alias_data = alias_data;
Joerg Roedel26018872011-06-06 16:50:14 +0200229 }
Joerg Roedel657cbb62009-11-23 15:26:46 +0100230
231 dev->archdata.iommu = dev_data;
232
Joerg Roedel657cbb62009-11-23 15:26:46 +0100233 return 0;
234}
235
Joerg Roedel26018872011-06-06 16:50:14 +0200236static 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
Joerg Roedel657cbb62009-11-23 15:26:46 +0100250static void iommu_uninit_device(struct device *dev)
251{
Joerg Roedel8fa5f802011-06-09 12:24:45 +0200252 /*
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 */
Joerg Roedel657cbb62009-11-23 15:26:46 +0100257}
Joerg Roedelb7cc9552009-12-10 11:03:39 +0100258
259void __init amd_iommu_uninit_devices(void)
260{
Joerg Roedel8fa5f802011-06-09 12:24:45 +0200261 struct iommu_dev_data *dev_data, *n;
Joerg Roedelb7cc9552009-12-10 11:03:39 +0100262 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 }
Joerg Roedel8fa5f802011-06-09 12:24:45 +0200271
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);
Joerg Roedelb7cc9552009-12-10 11:03:39 +0100275}
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);
Joerg Roedel26018872011-06-06 16:50:14 +0200288 if (ret == -ENOTSUPP)
289 iommu_ignore_device(&pdev->dev);
290 else if (ret)
Joerg Roedelb7cc9552009-12-10 11:03:39 +0100291 goto out_free;
292 }
293
294 return 0;
295
296out_free:
297
298 amd_iommu_uninit_devices();
299
300 return ret;
301}
Joerg Roedel7f265082008-12-12 13:50:21 +0100302#ifdef CONFIG_AMD_IOMMU_STATS
303
304/*
305 * Initialization code for statistics collection
306 */
307
Joerg Roedelda49f6d2008-12-12 14:59:58 +0100308DECLARE_STATS_COUNTER(compl_wait);
Joerg Roedel0f2a86f2008-12-12 15:05:16 +0100309DECLARE_STATS_COUNTER(cnt_map_single);
Joerg Roedel146a6912008-12-12 15:07:12 +0100310DECLARE_STATS_COUNTER(cnt_unmap_single);
Joerg Roedeld03f067a2008-12-12 15:09:48 +0100311DECLARE_STATS_COUNTER(cnt_map_sg);
Joerg Roedel55877a62008-12-12 15:12:14 +0100312DECLARE_STATS_COUNTER(cnt_unmap_sg);
Joerg Roedelc8f0fb32008-12-12 15:14:21 +0100313DECLARE_STATS_COUNTER(cnt_alloc_coherent);
Joerg Roedel5d31ee72008-12-12 15:16:38 +0100314DECLARE_STATS_COUNTER(cnt_free_coherent);
Joerg Roedelc1858972008-12-12 15:42:39 +0100315DECLARE_STATS_COUNTER(cross_page);
Joerg Roedelf57d98a2008-12-12 15:46:29 +0100316DECLARE_STATS_COUNTER(domain_flush_single);
Joerg Roedel18811f52008-12-12 15:48:28 +0100317DECLARE_STATS_COUNTER(domain_flush_all);
Joerg Roedel5774f7c2008-12-12 15:57:30 +0100318DECLARE_STATS_COUNTER(alloced_io_mem);
Joerg Roedel8ecaf8f2008-12-12 16:13:04 +0100319DECLARE_STATS_COUNTER(total_map_requests);
Joerg Roedelda49f6d2008-12-12 14:59:58 +0100320
Joerg Roedel7f265082008-12-12 13:50:21 +0100321static struct dentry *stats_dir;
Joerg Roedel7f265082008-12-12 13:50:21 +0100322static 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
Joerg Roedel7f265082008-12-12 13:50:21 +0100339 de_fflush = debugfs_create_bool("fullflush", 0444, stats_dir,
340 (u32 *)&amd_iommu_unmap_flush);
Joerg Roedelda49f6d2008-12-12 14:59:58 +0100341
342 amd_iommu_stats_add(&compl_wait);
Joerg Roedel0f2a86f2008-12-12 15:05:16 +0100343 amd_iommu_stats_add(&cnt_map_single);
Joerg Roedel146a6912008-12-12 15:07:12 +0100344 amd_iommu_stats_add(&cnt_unmap_single);
Joerg Roedeld03f067a2008-12-12 15:09:48 +0100345 amd_iommu_stats_add(&cnt_map_sg);
Joerg Roedel55877a62008-12-12 15:12:14 +0100346 amd_iommu_stats_add(&cnt_unmap_sg);
Joerg Roedelc8f0fb32008-12-12 15:14:21 +0100347 amd_iommu_stats_add(&cnt_alloc_coherent);
Joerg Roedel5d31ee72008-12-12 15:16:38 +0100348 amd_iommu_stats_add(&cnt_free_coherent);
Joerg Roedelc1858972008-12-12 15:42:39 +0100349 amd_iommu_stats_add(&cross_page);
Joerg Roedelf57d98a2008-12-12 15:46:29 +0100350 amd_iommu_stats_add(&domain_flush_single);
Joerg Roedel18811f52008-12-12 15:48:28 +0100351 amd_iommu_stats_add(&domain_flush_all);
Joerg Roedel5774f7c2008-12-12 15:57:30 +0100352 amd_iommu_stats_add(&alloced_io_mem);
Joerg Roedel8ecaf8f2008-12-12 16:13:04 +0100353 amd_iommu_stats_add(&total_map_requests);
Joerg Roedel7f265082008-12-12 13:50:21 +0100354}
355
356#endif
357
Joerg Roedel431b2a22008-07-11 17:14:22 +0200358/****************************************************************************
359 *
Joerg Roedela80dc3e2008-09-11 16:51:41 +0200360 * Interrupt handling functions
361 *
362 ****************************************************************************/
363
Joerg Roedele3e59872009-09-03 14:02:10 +0200364static 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
Joerg Roedel945b4ac2009-09-03 14:25:02 +0200373static 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
Joerg Roedela345b232009-09-03 15:01:43 +0200382static void iommu_print_event(struct amd_iommu *iommu, void *__evt)
Joerg Roedel90008ee2008-09-09 16:41:05 +0200383{
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
Joerg Roedel4c6f40d2009-09-01 16:43:58 +0200391 printk(KERN_ERR "AMD-Vi: Event logged [");
Joerg Roedel90008ee2008-09-09 16:41:05 +0200392
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);
Joerg Roedele3e59872009-09-03 14:02:10 +0200399 dump_dte_entry(devid);
Joerg Roedel90008ee2008-09-09 16:41:05 +0200400 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);
Joerg Roedel945b4ac2009-09-03 14:25:02 +0200421 dump_command(address);
Joerg Roedel90008ee2008-09-09 16:41:05 +0200422 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) {
Joerg Roedela345b232009-09-03 15:01:43 +0200455 iommu_print_event(iommu, iommu->evt_buf + head);
Joerg Roedel90008ee2008-09-09 16:41:05 +0200456 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
Joerg Roedel72fe00f2011-05-10 10:50:42 +0200464irqreturn_t amd_iommu_int_thread(int irq, void *data)
Joerg Roedela80dc3e2008-09-11 16:51:41 +0200465{
Joerg Roedel90008ee2008-09-09 16:41:05 +0200466 struct amd_iommu *iommu;
467
Joerg Roedel3bd22172009-05-04 15:06:20 +0200468 for_each_iommu(iommu)
Joerg Roedel90008ee2008-09-09 16:41:05 +0200469 iommu_poll_events(iommu);
470
471 return IRQ_HANDLED;
Joerg Roedela80dc3e2008-09-11 16:51:41 +0200472}
473
Joerg Roedel72fe00f2011-05-10 10:50:42 +0200474irqreturn_t amd_iommu_int_handler(int irq, void *data)
475{
476 return IRQ_WAKE_THREAD;
477}
478
Joerg Roedela80dc3e2008-09-11 16:51:41 +0200479/****************************************************************************
480 *
Joerg Roedel431b2a22008-07-11 17:14:22 +0200481 * IOMMU command queuing functions
482 *
483 ****************************************************************************/
484
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200485static int wait_on_sem(volatile u64 *sem)
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200486{
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200487 int i = 0;
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200488
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200489 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 }
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200498
499 return 0;
500}
501
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200502static void copy_cmd_to_buffer(struct amd_iommu *iommu,
503 struct iommu_cmd *cmd,
504 u32 tail)
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200505{
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200506 u8 *target;
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200507
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200508 target = iommu->cmd_buf + tail;
509 tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200510
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200511 /* 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
Joerg Roedel815b33f2011-04-06 17:26:49 +0200518static void build_completion_wait(struct iommu_cmd *cmd, u64 address)
Joerg Roedelded46732011-04-06 10:53:48 +0200519{
Joerg Roedel815b33f2011-04-06 17:26:49 +0200520 WARN_ON(address & 0x7ULL);
521
Joerg Roedelded46732011-04-06 10:53:48 +0200522 memset(cmd, 0, sizeof(*cmd));
Joerg Roedel815b33f2011-04-06 17:26:49 +0200523 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;
Joerg Roedelded46732011-04-06 10:53:48 +0200526 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT);
527}
528
Joerg Roedel94fe79e2011-04-06 11:07:21 +0200529static 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
Joerg Roedel11b64022011-04-06 11:49:28 +0200536static 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
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200567static 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
Joerg Roedel58fc7f12011-04-11 11:13:24 +0200598static void build_inv_all(struct iommu_cmd *cmd)
599{
600 memset(cmd, 0, sizeof(*cmd));
601 CMD_SET_TYPE(cmd, CMD_INV_ALL);
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200602}
603
Joerg Roedel431b2a22008-07-11 17:14:22 +0200604/*
Joerg Roedelb6c02712008-06-26 21:27:53 +0200605 * Writes the command to the IOMMUs command buffer and informs the
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200606 * hardware about the new command.
Joerg Roedel431b2a22008-07-11 17:14:22 +0200607 */
Joerg Roedel815b33f2011-04-06 17:26:49 +0200608static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd)
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200609{
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200610 u32 left, tail, head, next_tail;
Joerg Roedel815b33f2011-04-06 17:26:49 +0200611 unsigned long flags;
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200612
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200613 WARN_ON(iommu->cmd_buf_size & CMD_BUFFER_UNINITIALIZED);
Joerg Roedelda49f6d2008-12-12 14:59:58 +0100614
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200615again:
Joerg Roedel815b33f2011-04-06 17:26:49 +0200616 spin_lock_irqsave(&iommu->lock, flags);
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200617
618 head = readl(iommu->mmio_base + MMIO_CMD_HEAD_OFFSET);
619 tail = readl(iommu->mmio_base + MMIO_CMD_TAIL_OFFSET);
620 next_tail = (tail + sizeof(*cmd)) % iommu->cmd_buf_size;
621 left = (head - next_tail) % iommu->cmd_buf_size;
622
623 if (left <= 2) {
624 struct iommu_cmd sync_cmd;
625 volatile u64 sem = 0;
626 int ret;
627
628 build_completion_wait(&sync_cmd, (u64)&sem);
629 copy_cmd_to_buffer(iommu, &sync_cmd, tail);
630
631 spin_unlock_irqrestore(&iommu->lock, flags);
632
633 if ((ret = wait_on_sem(&sem)) != 0)
634 return ret;
635
636 goto again;
Joerg Roedel136f78a2008-07-11 17:14:27 +0200637 }
638
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200639 copy_cmd_to_buffer(iommu, cmd, tail);
Joerg Roedel519c31b2008-08-14 19:55:15 +0200640
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200641 /* We need to sync now to make sure all commands are processed */
Joerg Roedel815b33f2011-04-06 17:26:49 +0200642 iommu->need_sync = true;
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200643
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200644 spin_unlock_irqrestore(&iommu->lock, flags);
645
Joerg Roedel815b33f2011-04-06 17:26:49 +0200646 return 0;
Joerg Roedel8d201962008-12-02 20:34:41 +0100647}
648
649/*
650 * This function queues a completion wait command into the command
651 * buffer of an IOMMU
652 */
Joerg Roedel8d201962008-12-02 20:34:41 +0100653static int iommu_completion_wait(struct amd_iommu *iommu)
654{
Joerg Roedel815b33f2011-04-06 17:26:49 +0200655 struct iommu_cmd cmd;
656 volatile u64 sem = 0;
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200657 int ret;
Joerg Roedel8d201962008-12-02 20:34:41 +0100658
659 if (!iommu->need_sync)
Joerg Roedel815b33f2011-04-06 17:26:49 +0200660 return 0;
Joerg Roedel8d201962008-12-02 20:34:41 +0100661
Joerg Roedel815b33f2011-04-06 17:26:49 +0200662 build_completion_wait(&cmd, (u64)&sem);
Joerg Roedel8d201962008-12-02 20:34:41 +0100663
Joerg Roedel815b33f2011-04-06 17:26:49 +0200664 ret = iommu_queue_command(iommu, &cmd);
Joerg Roedel8d201962008-12-02 20:34:41 +0100665 if (ret)
Joerg Roedel815b33f2011-04-06 17:26:49 +0200666 return ret;
Joerg Roedel8d201962008-12-02 20:34:41 +0100667
Joerg Roedelac0ea6e2011-04-06 18:38:20 +0200668 return wait_on_sem(&sem);
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200669}
670
Joerg Roedeld8c13082011-04-06 18:51:26 +0200671static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid)
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200672{
673 struct iommu_cmd cmd;
674
Joerg Roedeld8c13082011-04-06 18:51:26 +0200675 build_inv_dte(&cmd, devid);
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200676
Joerg Roedeld8c13082011-04-06 18:51:26 +0200677 return iommu_queue_command(iommu, &cmd);
678}
679
Joerg Roedel7d0c5cc2011-04-07 08:16:10 +0200680static void iommu_flush_dte_all(struct amd_iommu *iommu)
681{
682 u32 devid;
683
684 for (devid = 0; devid <= 0xffff; ++devid)
685 iommu_flush_dte(iommu, devid);
686
687 iommu_completion_wait(iommu);
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200688}
689
690/*
Joerg Roedel7d0c5cc2011-04-07 08:16:10 +0200691 * This function uses heavy locking and may disable irqs for some time. But
692 * this is no issue because it is only called during resume.
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200693 */
Joerg Roedel7d0c5cc2011-04-07 08:16:10 +0200694static void iommu_flush_tlb_all(struct amd_iommu *iommu)
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200695{
Joerg Roedel7d0c5cc2011-04-07 08:16:10 +0200696 u32 dom_id;
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200697
Joerg Roedel7d0c5cc2011-04-07 08:16:10 +0200698 for (dom_id = 0; dom_id <= 0xffff; ++dom_id) {
699 struct iommu_cmd cmd;
700 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS,
701 dom_id, 1);
702 iommu_queue_command(iommu, &cmd);
703 }
Joerg Roedel431b2a22008-07-11 17:14:22 +0200704
Joerg Roedel7d0c5cc2011-04-07 08:16:10 +0200705 iommu_completion_wait(iommu);
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200706}
707
Joerg Roedel58fc7f12011-04-11 11:13:24 +0200708static void iommu_flush_all(struct amd_iommu *iommu)
709{
710 struct iommu_cmd cmd;
711
712 build_inv_all(&cmd);
713
714 iommu_queue_command(iommu, &cmd);
715 iommu_completion_wait(iommu);
716}
717
Joerg Roedel7d0c5cc2011-04-07 08:16:10 +0200718void iommu_flush_all_caches(struct amd_iommu *iommu)
719{
Joerg Roedel58fc7f12011-04-11 11:13:24 +0200720 if (iommu_feature(iommu, FEATURE_IA)) {
721 iommu_flush_all(iommu);
722 } else {
723 iommu_flush_dte_all(iommu);
724 iommu_flush_tlb_all(iommu);
725 }
Joerg Roedel7d0c5cc2011-04-07 08:16:10 +0200726}
727
Joerg Roedel431b2a22008-07-11 17:14:22 +0200728/*
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200729 * Command send function for flushing on-device TLB
730 */
Joerg Roedel6c542042011-06-09 17:07:31 +0200731static int device_flush_iotlb(struct iommu_dev_data *dev_data,
732 u64 address, size_t size)
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200733{
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200734 struct amd_iommu *iommu;
735 struct iommu_cmd cmd;
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200736 int qdep;
737
Joerg Roedelea61cdd2011-06-09 12:56:30 +0200738 qdep = dev_data->ats.qdep;
739 iommu = amd_iommu_rlookup_table[dev_data->devid];
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200740
Joerg Roedelea61cdd2011-06-09 12:56:30 +0200741 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size);
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200742
743 return iommu_queue_command(iommu, &cmd);
744}
745
746/*
Joerg Roedel431b2a22008-07-11 17:14:22 +0200747 * Command send function for invalidating a device table entry
748 */
Joerg Roedel6c542042011-06-09 17:07:31 +0200749static int device_flush_dte(struct iommu_dev_data *dev_data)
Joerg Roedel3fa43652009-11-26 15:04:38 +0100750{
751 struct amd_iommu *iommu;
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200752 int ret;
Joerg Roedel3fa43652009-11-26 15:04:38 +0100753
Joerg Roedel6c542042011-06-09 17:07:31 +0200754 iommu = amd_iommu_rlookup_table[dev_data->devid];
Joerg Roedel3fa43652009-11-26 15:04:38 +0100755
Joerg Roedelf62dda62011-06-09 12:55:35 +0200756 ret = iommu_flush_dte(iommu, dev_data->devid);
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200757 if (ret)
758 return ret;
759
Joerg Roedelea61cdd2011-06-09 12:56:30 +0200760 if (dev_data->ats.enabled)
Joerg Roedel6c542042011-06-09 17:07:31 +0200761 ret = device_flush_iotlb(dev_data, 0, ~0UL);
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200762
763 return ret;
Joerg Roedel3fa43652009-11-26 15:04:38 +0100764}
765
Joerg Roedel431b2a22008-07-11 17:14:22 +0200766/*
767 * TLB invalidation function which is called from the mapping functions.
768 * It invalidates a single PTE if the range to flush is within a single
769 * page. Otherwise it flushes the whole TLB of the IOMMU.
770 */
Joerg Roedel17b124b2011-04-06 18:01:35 +0200771static void __domain_flush_pages(struct protection_domain *domain,
772 u64 address, size_t size, int pde)
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200773{
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200774 struct iommu_dev_data *dev_data;
Joerg Roedel11b64022011-04-06 11:49:28 +0200775 struct iommu_cmd cmd;
776 int ret = 0, i;
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200777
Joerg Roedel11b64022011-04-06 11:49:28 +0200778 build_inv_iommu_pages(&cmd, address, size, domain->id, pde);
Joerg Roedel999ba412008-07-03 19:35:08 +0200779
Joerg Roedel6de8ad92009-11-23 18:30:32 +0100780 for (i = 0; i < amd_iommus_present; ++i) {
781 if (!domain->dev_iommu[i])
782 continue;
783
784 /*
785 * Devices of this domain are behind this IOMMU
786 * We need a TLB flush
787 */
Joerg Roedel11b64022011-04-06 11:49:28 +0200788 ret |= iommu_queue_command(amd_iommus[i], &cmd);
Joerg Roedel6de8ad92009-11-23 18:30:32 +0100789 }
790
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200791 list_for_each_entry(dev_data, &domain->dev_list, list) {
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200792
Joerg Roedelea61cdd2011-06-09 12:56:30 +0200793 if (!dev_data->ats.enabled)
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200794 continue;
795
Joerg Roedel6c542042011-06-09 17:07:31 +0200796 ret |= device_flush_iotlb(dev_data, address, size);
Joerg Roedelcb41ed82011-04-05 11:00:53 +0200797 }
798
Joerg Roedel11b64022011-04-06 11:49:28 +0200799 WARN_ON(ret);
Joerg Roedel6de8ad92009-11-23 18:30:32 +0100800}
801
Joerg Roedel17b124b2011-04-06 18:01:35 +0200802static void domain_flush_pages(struct protection_domain *domain,
803 u64 address, size_t size)
Joerg Roedel6de8ad92009-11-23 18:30:32 +0100804{
Joerg Roedel17b124b2011-04-06 18:01:35 +0200805 __domain_flush_pages(domain, address, size, 0);
Joerg Roedela19ae1e2008-06-26 21:27:55 +0200806}
Joerg Roedelb6c02712008-06-26 21:27:53 +0200807
Joerg Roedel1c655772008-09-04 18:40:05 +0200808/* Flush the whole IO/TLB for a given protection domain */
Joerg Roedel17b124b2011-04-06 18:01:35 +0200809static void domain_flush_tlb(struct protection_domain *domain)
Joerg Roedel1c655772008-09-04 18:40:05 +0200810{
Joerg Roedel17b124b2011-04-06 18:01:35 +0200811 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 0);
Joerg Roedel1c655772008-09-04 18:40:05 +0200812}
813
Chris Wright42a49f92009-06-15 15:42:00 +0200814/* Flush the whole IO/TLB for a given protection domain - including PDE */
Joerg Roedel17b124b2011-04-06 18:01:35 +0200815static void domain_flush_tlb_pde(struct protection_domain *domain)
Chris Wright42a49f92009-06-15 15:42:00 +0200816{
Joerg Roedel17b124b2011-04-06 18:01:35 +0200817 __domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1);
818}
819
820static void domain_flush_complete(struct protection_domain *domain)
Joerg Roedelb6c02712008-06-26 21:27:53 +0200821{
822 int i;
823
824 for (i = 0; i < amd_iommus_present; ++i) {
825 if (!domain->dev_iommu[i])
826 continue;
827
828 /*
829 * Devices of this domain are behind this IOMMU
830 * We need to wait for completion of all commands.
831 */
832 iommu_completion_wait(amd_iommus[i]);
833 }
834}
835
Joerg Roedelb00d3bc2009-11-26 15:35:33 +0100836
Joerg Roedel43f49602008-12-02 21:01:12 +0100837/*
Joerg Roedelb00d3bc2009-11-26 15:35:33 +0100838 * This function flushes the DTEs for all devices in domain
Joerg Roedel43f49602008-12-02 21:01:12 +0100839 */
Joerg Roedel17b124b2011-04-06 18:01:35 +0200840static void domain_flush_devices(struct protection_domain *domain)
Joerg Roedelbfd1be12009-05-05 15:33:57 +0200841{
Joerg Roedelb00d3bc2009-11-26 15:35:33 +0100842 struct iommu_dev_data *dev_data;
843 unsigned long flags;
844
845 spin_lock_irqsave(&domain->lock, flags);
846
847 list_for_each_entry(dev_data, &domain->dev_list, list)
Joerg Roedel6c542042011-06-09 17:07:31 +0200848 device_flush_dte(dev_data);
Joerg Roedelb00d3bc2009-11-26 15:35:33 +0100849
850 spin_unlock_irqrestore(&domain->lock, flags);
851}
852
Joerg Roedel431b2a22008-07-11 17:14:22 +0200853/****************************************************************************
854 *
855 * The functions below are used the create the page table mappings for
856 * unity mapped regions.
857 *
858 ****************************************************************************/
859
860/*
Joerg Roedel308973d2009-11-24 17:43:32 +0100861 * This function is used to add another level to an IO page table. Adding
862 * another level increases the size of the address space by 9 bits to a size up
863 * to 64 bits.
864 */
865static bool increase_address_space(struct protection_domain *domain,
866 gfp_t gfp)
867{
868 u64 *pte;
869
870 if (domain->mode == PAGE_MODE_6_LEVEL)
871 /* address space already 64 bit large */
872 return false;
873
874 pte = (void *)get_zeroed_page(gfp);
875 if (!pte)
876 return false;
877
878 *pte = PM_LEVEL_PDE(domain->mode,
879 virt_to_phys(domain->pt_root));
880 domain->pt_root = pte;
881 domain->mode += 1;
882 domain->updated = true;
883
884 return true;
885}
886
887static u64 *alloc_pte(struct protection_domain *domain,
888 unsigned long address,
Joerg Roedelcbb9d722010-01-15 14:41:15 +0100889 unsigned long page_size,
Joerg Roedel308973d2009-11-24 17:43:32 +0100890 u64 **pte_page,
891 gfp_t gfp)
892{
Joerg Roedelcbb9d722010-01-15 14:41:15 +0100893 int level, end_lvl;
Joerg Roedel308973d2009-11-24 17:43:32 +0100894 u64 *pte, *page;
Joerg Roedelcbb9d722010-01-15 14:41:15 +0100895
896 BUG_ON(!is_power_of_2(page_size));
Joerg Roedel308973d2009-11-24 17:43:32 +0100897
898 while (address > PM_LEVEL_SIZE(domain->mode))
899 increase_address_space(domain, gfp);
900
Joerg Roedelcbb9d722010-01-15 14:41:15 +0100901 level = domain->mode - 1;
902 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
903 address = PAGE_SIZE_ALIGN(address, page_size);
904 end_lvl = PAGE_SIZE_LEVEL(page_size);
Joerg Roedel308973d2009-11-24 17:43:32 +0100905
906 while (level > end_lvl) {
907 if (!IOMMU_PTE_PRESENT(*pte)) {
908 page = (u64 *)get_zeroed_page(gfp);
909 if (!page)
910 return NULL;
911 *pte = PM_LEVEL_PDE(level, virt_to_phys(page));
912 }
913
Joerg Roedelcbb9d722010-01-15 14:41:15 +0100914 /* No level skipping support yet */
915 if (PM_PTE_LEVEL(*pte) != level)
916 return NULL;
917
Joerg Roedel308973d2009-11-24 17:43:32 +0100918 level -= 1;
919
920 pte = IOMMU_PTE_PAGE(*pte);
921
922 if (pte_page && level == end_lvl)
923 *pte_page = pte;
924
925 pte = &pte[PM_LEVEL_INDEX(level, address)];
926 }
927
928 return pte;
929}
930
931/*
932 * This function checks if there is a PTE for a given dma address. If
933 * there is one, it returns the pointer to it.
934 */
Joerg Roedel24cd7722010-01-19 17:27:39 +0100935static u64 *fetch_pte(struct protection_domain *domain, unsigned long address)
Joerg Roedel308973d2009-11-24 17:43:32 +0100936{
937 int level;
938 u64 *pte;
939
Joerg Roedel24cd7722010-01-19 17:27:39 +0100940 if (address > PM_LEVEL_SIZE(domain->mode))
941 return NULL;
Joerg Roedel308973d2009-11-24 17:43:32 +0100942
Joerg Roedel24cd7722010-01-19 17:27:39 +0100943 level = domain->mode - 1;
944 pte = &domain->pt_root[PM_LEVEL_INDEX(level, address)];
945
946 while (level > 0) {
947
948 /* Not Present */
Joerg Roedel308973d2009-11-24 17:43:32 +0100949 if (!IOMMU_PTE_PRESENT(*pte))
950 return NULL;
951
Joerg Roedel24cd7722010-01-19 17:27:39 +0100952 /* Large PTE */
953 if (PM_PTE_LEVEL(*pte) == 0x07) {
954 unsigned long pte_mask, __pte;
955
956 /*
957 * If we have a series of large PTEs, make
958 * sure to return a pointer to the first one.
959 */
960 pte_mask = PTE_PAGE_SIZE(*pte);
961 pte_mask = ~((PAGE_SIZE_PTE_COUNT(pte_mask) << 3) - 1);
962 __pte = ((unsigned long)pte) & pte_mask;
963
964 return (u64 *)__pte;
965 }
966
967 /* No level skipping support yet */
968 if (PM_PTE_LEVEL(*pte) != level)
969 return NULL;
970
Joerg Roedel308973d2009-11-24 17:43:32 +0100971 level -= 1;
972
Joerg Roedel24cd7722010-01-19 17:27:39 +0100973 /* Walk to the next level */
Joerg Roedel308973d2009-11-24 17:43:32 +0100974 pte = IOMMU_PTE_PAGE(*pte);
975 pte = &pte[PM_LEVEL_INDEX(level, address)];
Joerg Roedel308973d2009-11-24 17:43:32 +0100976 }
977
978 return pte;
979}
980
981/*
Joerg Roedel431b2a22008-07-11 17:14:22 +0200982 * Generic mapping functions. It maps a physical address into a DMA
983 * address space. It allocates the page table pages if necessary.
984 * In the future it can be extended to a generic mapping function
985 * supporting all features of AMD IOMMU page tables like level skipping
986 * and full 64 bit address spaces.
987 */
Joerg Roedel38e817f2008-12-02 17:27:52 +0100988static int iommu_map_page(struct protection_domain *dom,
989 unsigned long bus_addr,
990 unsigned long phys_addr,
Joerg Roedelabdc5eb2009-09-03 11:33:51 +0200991 int prot,
Joerg Roedelcbb9d722010-01-15 14:41:15 +0100992 unsigned long page_size)
Joerg Roedelbd0e5212008-06-26 21:27:56 +0200993{
Joerg Roedel8bda3092009-05-12 12:02:46 +0200994 u64 __pte, *pte;
Joerg Roedelcbb9d722010-01-15 14:41:15 +0100995 int i, count;
Joerg Roedelabdc5eb2009-09-03 11:33:51 +0200996
Joerg Roedelbad1cac2009-09-02 16:52:23 +0200997 if (!(prot & IOMMU_PROT_MASK))
Joerg Roedelbd0e5212008-06-26 21:27:56 +0200998 return -EINVAL;
999
Joerg Roedelcbb9d722010-01-15 14:41:15 +01001000 bus_addr = PAGE_ALIGN(bus_addr);
1001 phys_addr = PAGE_ALIGN(phys_addr);
1002 count = PAGE_SIZE_PTE_COUNT(page_size);
1003 pte = alloc_pte(dom, bus_addr, page_size, NULL, GFP_KERNEL);
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001004
Joerg Roedelcbb9d722010-01-15 14:41:15 +01001005 for (i = 0; i < count; ++i)
1006 if (IOMMU_PTE_PRESENT(pte[i]))
1007 return -EBUSY;
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001008
Joerg Roedelcbb9d722010-01-15 14:41:15 +01001009 if (page_size > PAGE_SIZE) {
1010 __pte = PAGE_SIZE_PTE(phys_addr, page_size);
1011 __pte |= PM_LEVEL_ENC(7) | IOMMU_PTE_P | IOMMU_PTE_FC;
1012 } else
1013 __pte = phys_addr | IOMMU_PTE_P | IOMMU_PTE_FC;
1014
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001015 if (prot & IOMMU_PROT_IR)
1016 __pte |= IOMMU_PTE_IR;
1017 if (prot & IOMMU_PROT_IW)
1018 __pte |= IOMMU_PTE_IW;
1019
Joerg Roedelcbb9d722010-01-15 14:41:15 +01001020 for (i = 0; i < count; ++i)
1021 pte[i] = __pte;
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001022
Joerg Roedel04bfdd82009-09-02 16:00:23 +02001023 update_domain(dom);
1024
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001025 return 0;
1026}
1027
Joerg Roedel24cd7722010-01-19 17:27:39 +01001028static unsigned long iommu_unmap_page(struct protection_domain *dom,
1029 unsigned long bus_addr,
1030 unsigned long page_size)
Joerg Roedeleb74ff62008-12-02 19:59:10 +01001031{
Joerg Roedel24cd7722010-01-19 17:27:39 +01001032 unsigned long long unmap_size, unmapped;
1033 u64 *pte;
Joerg Roedeleb74ff62008-12-02 19:59:10 +01001034
Joerg Roedel24cd7722010-01-19 17:27:39 +01001035 BUG_ON(!is_power_of_2(page_size));
1036
1037 unmapped = 0;
1038
1039 while (unmapped < page_size) {
1040
1041 pte = fetch_pte(dom, bus_addr);
1042
1043 if (!pte) {
1044 /*
1045 * No PTE for this address
1046 * move forward in 4kb steps
1047 */
1048 unmap_size = PAGE_SIZE;
1049 } else if (PM_PTE_LEVEL(*pte) == 0) {
1050 /* 4kb PTE found for this address */
1051 unmap_size = PAGE_SIZE;
1052 *pte = 0ULL;
1053 } else {
1054 int count, i;
1055
1056 /* Large PTE found which maps this address */
1057 unmap_size = PTE_PAGE_SIZE(*pte);
1058 count = PAGE_SIZE_PTE_COUNT(unmap_size);
1059 for (i = 0; i < count; i++)
1060 pte[i] = 0ULL;
1061 }
1062
1063 bus_addr = (bus_addr & ~(unmap_size - 1)) + unmap_size;
1064 unmapped += unmap_size;
1065 }
1066
1067 BUG_ON(!is_power_of_2(unmapped));
1068
1069 return unmapped;
Joerg Roedeleb74ff62008-12-02 19:59:10 +01001070}
Joerg Roedeleb74ff62008-12-02 19:59:10 +01001071
Joerg Roedel431b2a22008-07-11 17:14:22 +02001072/*
1073 * This function checks if a specific unity mapping entry is needed for
1074 * this specific IOMMU.
1075 */
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001076static int iommu_for_unity_map(struct amd_iommu *iommu,
1077 struct unity_map_entry *entry)
1078{
1079 u16 bdf, i;
1080
1081 for (i = entry->devid_start; i <= entry->devid_end; ++i) {
1082 bdf = amd_iommu_alias_table[i];
1083 if (amd_iommu_rlookup_table[bdf] == iommu)
1084 return 1;
1085 }
1086
1087 return 0;
1088}
1089
Joerg Roedel431b2a22008-07-11 17:14:22 +02001090/*
Joerg Roedel431b2a22008-07-11 17:14:22 +02001091 * This function actually applies the mapping to the page table of the
1092 * dma_ops domain.
1093 */
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001094static int dma_ops_unity_map(struct dma_ops_domain *dma_dom,
1095 struct unity_map_entry *e)
1096{
1097 u64 addr;
1098 int ret;
1099
1100 for (addr = e->address_start; addr < e->address_end;
1101 addr += PAGE_SIZE) {
Joerg Roedelabdc5eb2009-09-03 11:33:51 +02001102 ret = iommu_map_page(&dma_dom->domain, addr, addr, e->prot,
Joerg Roedelcbb9d722010-01-15 14:41:15 +01001103 PAGE_SIZE);
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001104 if (ret)
1105 return ret;
1106 /*
1107 * if unity mapping is in aperture range mark the page
1108 * as allocated in the aperture
1109 */
1110 if (addr < dma_dom->aperture_size)
Joerg Roedelc3239562009-05-12 10:56:44 +02001111 __set_bit(addr >> PAGE_SHIFT,
Joerg Roedel384de722009-05-15 12:30:05 +02001112 dma_dom->aperture[0]->bitmap);
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001113 }
1114
1115 return 0;
1116}
1117
Joerg Roedel431b2a22008-07-11 17:14:22 +02001118/*
Joerg Roedel171e7b32009-11-24 17:47:56 +01001119 * Init the unity mappings for a specific IOMMU in the system
1120 *
1121 * Basically iterates over all unity mapping entries and applies them to
1122 * the default domain DMA of that IOMMU if necessary.
1123 */
1124static int iommu_init_unity_mappings(struct amd_iommu *iommu)
1125{
1126 struct unity_map_entry *entry;
1127 int ret;
1128
1129 list_for_each_entry(entry, &amd_iommu_unity_map, list) {
1130 if (!iommu_for_unity_map(iommu, entry))
1131 continue;
1132 ret = dma_ops_unity_map(iommu->default_dom, entry);
1133 if (ret)
1134 return ret;
1135 }
1136
1137 return 0;
1138}
1139
1140/*
Joerg Roedel431b2a22008-07-11 17:14:22 +02001141 * Inits the unity mappings required for a specific device
1142 */
Joerg Roedelbd0e5212008-06-26 21:27:56 +02001143static int init_unity_mappings_for_device(struct dma_ops_domain *dma_dom,
1144 u16 devid)
1145{
1146 struct unity_map_entry *e;
1147 int ret;
1148
1149 list_for_each_entry(e, &amd_iommu_unity_map, list) {
1150 if (!(devid >= e->devid_start && devid <= e->devid_end))
1151 continue;
1152 ret = dma_ops_unity_map(dma_dom, e);
1153 if (ret)
1154 return ret;
1155 }
1156
1157 return 0;
1158}
1159
Joerg Roedel431b2a22008-07-11 17:14:22 +02001160/****************************************************************************
1161 *
1162 * The next functions belong to the address allocator for the dma_ops
1163 * interface functions. They work like the allocators in the other IOMMU
1164 * drivers. Its basically a bitmap which marks the allocated pages in
1165 * the aperture. Maybe it could be enhanced in the future to a more
1166 * efficient allocator.
1167 *
1168 ****************************************************************************/
Joerg Roedeld3086442008-06-26 21:27:57 +02001169
Joerg Roedel431b2a22008-07-11 17:14:22 +02001170/*
Joerg Roedel384de722009-05-15 12:30:05 +02001171 * The address allocator core functions.
Joerg Roedel431b2a22008-07-11 17:14:22 +02001172 *
1173 * called with domain->lock held
1174 */
Joerg Roedel384de722009-05-15 12:30:05 +02001175
Joerg Roedel9cabe892009-05-18 16:38:55 +02001176/*
Joerg Roedel171e7b32009-11-24 17:47:56 +01001177 * Used to reserve address ranges in the aperture (e.g. for exclusion
1178 * ranges.
1179 */
1180static void dma_ops_reserve_addresses(struct dma_ops_domain *dom,
1181 unsigned long start_page,
1182 unsigned int pages)
1183{
1184 unsigned int i, last_page = dom->aperture_size >> PAGE_SHIFT;
1185
1186 if (start_page + pages > last_page)
1187 pages = last_page - start_page;
1188
1189 for (i = start_page; i < start_page + pages; ++i) {
1190 int index = i / APERTURE_RANGE_PAGES;
1191 int page = i % APERTURE_RANGE_PAGES;
1192 __set_bit(page, dom->aperture[index]->bitmap);
1193 }
1194}
1195
1196/*
Joerg Roedel9cabe892009-05-18 16:38:55 +02001197 * This function is used to add a new aperture range to an existing
1198 * aperture in case of dma_ops domain allocation or address allocation
1199 * failure.
1200 */
Joerg Roedel576175c2009-11-23 19:08:46 +01001201static int alloc_new_range(struct dma_ops_domain *dma_dom,
Joerg Roedel9cabe892009-05-18 16:38:55 +02001202 bool populate, gfp_t gfp)
1203{
1204 int index = dma_dom->aperture_size >> APERTURE_RANGE_SHIFT;
Joerg Roedel576175c2009-11-23 19:08:46 +01001205 struct amd_iommu *iommu;
Joerg Roedel17f5b562011-07-06 17:14:44 +02001206 unsigned long i, old_size;
Joerg Roedel9cabe892009-05-18 16:38:55 +02001207
Joerg Roedelf5e97052009-05-22 12:31:53 +02001208#ifdef CONFIG_IOMMU_STRESS
1209 populate = false;
1210#endif
1211
Joerg Roedel9cabe892009-05-18 16:38:55 +02001212 if (index >= APERTURE_MAX_RANGES)
1213 return -ENOMEM;
1214
1215 dma_dom->aperture[index] = kzalloc(sizeof(struct aperture_range), gfp);
1216 if (!dma_dom->aperture[index])
1217 return -ENOMEM;
1218
1219 dma_dom->aperture[index]->bitmap = (void *)get_zeroed_page(gfp);
1220 if (!dma_dom->aperture[index]->bitmap)
1221 goto out_free;
1222
1223 dma_dom->aperture[index]->offset = dma_dom->aperture_size;
1224
1225 if (populate) {
1226 unsigned long address = dma_dom->aperture_size;
1227 int i, num_ptes = APERTURE_RANGE_PAGES / 512;
1228 u64 *pte, *pte_page;
1229
1230 for (i = 0; i < num_ptes; ++i) {
Joerg Roedelcbb9d722010-01-15 14:41:15 +01001231 pte = alloc_pte(&dma_dom->domain, address, PAGE_SIZE,
Joerg Roedel9cabe892009-05-18 16:38:55 +02001232 &pte_page, gfp);
1233 if (!pte)
1234 goto out_free;
1235
1236 dma_dom->aperture[index]->pte_pages[i] = pte_page;
1237
1238 address += APERTURE_RANGE_SIZE / 64;
1239 }
1240 }
1241
Joerg Roedel17f5b562011-07-06 17:14:44 +02001242 old_size = dma_dom->aperture_size;
Joerg Roedel9cabe892009-05-18 16:38:55 +02001243 dma_dom->aperture_size += APERTURE_RANGE_SIZE;
1244
Joerg Roedel17f5b562011-07-06 17:14:44 +02001245 /* Reserve address range used for MSI messages */
1246 if (old_size < MSI_ADDR_BASE_LO &&
1247 dma_dom->aperture_size > MSI_ADDR_BASE_LO) {
1248 unsigned long spage;
1249 int pages;
1250
1251 pages = iommu_num_pages(MSI_ADDR_BASE_LO, 0x10000, PAGE_SIZE);
1252 spage = MSI_ADDR_BASE_LO >> PAGE_SHIFT;
1253
1254 dma_ops_reserve_addresses(dma_dom, spage, pages);
1255 }
1256
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001257 /* Initialize the exclusion range if necessary */
Joerg Roedel576175c2009-11-23 19:08:46 +01001258 for_each_iommu(iommu) {
1259 if (iommu->exclusion_start &&
1260 iommu->exclusion_start >= dma_dom->aperture[index]->offset
1261 && iommu->exclusion_start < dma_dom->aperture_size) {
1262 unsigned long startpage;
1263 int pages = iommu_num_pages(iommu->exclusion_start,
1264 iommu->exclusion_length,
1265 PAGE_SIZE);
1266 startpage = iommu->exclusion_start >> PAGE_SHIFT;
1267 dma_ops_reserve_addresses(dma_dom, startpage, pages);
1268 }
Joerg Roedel00cd1222009-05-19 09:52:40 +02001269 }
1270
1271 /*
1272 * Check for areas already mapped as present in the new aperture
1273 * range and mark those pages as reserved in the allocator. Such
1274 * mappings may already exist as a result of requested unity
1275 * mappings for devices.
1276 */
1277 for (i = dma_dom->aperture[index]->offset;
1278 i < dma_dom->aperture_size;
1279 i += PAGE_SIZE) {
Joerg Roedel24cd7722010-01-19 17:27:39 +01001280 u64 *pte = fetch_pte(&dma_dom->domain, i);
Joerg Roedel00cd1222009-05-19 09:52:40 +02001281 if (!pte || !IOMMU_PTE_PRESENT(*pte))
1282 continue;
1283
1284 dma_ops_reserve_addresses(dma_dom, i << PAGE_SHIFT, 1);
1285 }
1286
Joerg Roedel04bfdd82009-09-02 16:00:23 +02001287 update_domain(&dma_dom->domain);
1288
Joerg Roedel9cabe892009-05-18 16:38:55 +02001289 return 0;
1290
1291out_free:
Joerg Roedel04bfdd82009-09-02 16:00:23 +02001292 update_domain(&dma_dom->domain);
1293
Joerg Roedel9cabe892009-05-18 16:38:55 +02001294 free_page((unsigned long)dma_dom->aperture[index]->bitmap);
1295
1296 kfree(dma_dom->aperture[index]);
1297 dma_dom->aperture[index] = NULL;
1298
1299 return -ENOMEM;
1300}
1301
Joerg Roedel384de722009-05-15 12:30:05 +02001302static unsigned long dma_ops_area_alloc(struct device *dev,
1303 struct dma_ops_domain *dom,
1304 unsigned int pages,
1305 unsigned long align_mask,
1306 u64 dma_mask,
1307 unsigned long start)
1308{
Joerg Roedel803b8cb2009-05-18 15:32:48 +02001309 unsigned long next_bit = dom->next_address % APERTURE_RANGE_SIZE;
Joerg Roedel384de722009-05-15 12:30:05 +02001310 int max_index = dom->aperture_size >> APERTURE_RANGE_SHIFT;
1311 int i = start >> APERTURE_RANGE_SHIFT;
1312 unsigned long boundary_size;
1313 unsigned long address = -1;
1314 unsigned long limit;
1315
Joerg Roedel803b8cb2009-05-18 15:32:48 +02001316 next_bit >>= PAGE_SHIFT;
1317
Joerg Roedel384de722009-05-15 12:30:05 +02001318 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
1319 PAGE_SIZE) >> PAGE_SHIFT;
1320
1321 for (;i < max_index; ++i) {
1322 unsigned long offset = dom->aperture[i]->offset >> PAGE_SHIFT;
1323
1324 if (dom->aperture[i]->offset >= dma_mask)
1325 break;
1326
1327 limit = iommu_device_max_index(APERTURE_RANGE_PAGES, offset,
1328 dma_mask >> PAGE_SHIFT);
1329
1330 address = iommu_area_alloc(dom->aperture[i]->bitmap,
1331 limit, next_bit, pages, 0,
1332 boundary_size, align_mask);
1333 if (address != -1) {
1334 address = dom->aperture[i]->offset +
1335 (address << PAGE_SHIFT);
Joerg Roedel803b8cb2009-05-18 15:32:48 +02001336 dom->next_address = address + (pages << PAGE_SHIFT);
Joerg Roedel384de722009-05-15 12:30:05 +02001337 break;
1338 }
1339
1340 next_bit = 0;
1341 }
1342
1343 return address;
1344}
1345
Joerg Roedeld3086442008-06-26 21:27:57 +02001346static unsigned long dma_ops_alloc_addresses(struct device *dev,
1347 struct dma_ops_domain *dom,
Joerg Roedel6d4f3432008-09-04 19:18:02 +02001348 unsigned int pages,
Joerg Roedel832a90c2008-09-18 15:54:23 +02001349 unsigned long align_mask,
1350 u64 dma_mask)
Joerg Roedeld3086442008-06-26 21:27:57 +02001351{
Joerg Roedeld3086442008-06-26 21:27:57 +02001352 unsigned long address;
Joerg Roedeld3086442008-06-26 21:27:57 +02001353
Joerg Roedelfe16f082009-05-22 12:27:53 +02001354#ifdef CONFIG_IOMMU_STRESS
1355 dom->next_address = 0;
1356 dom->need_flush = true;
1357#endif
Joerg Roedeld3086442008-06-26 21:27:57 +02001358
Joerg Roedel384de722009-05-15 12:30:05 +02001359 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
Joerg Roedel803b8cb2009-05-18 15:32:48 +02001360 dma_mask, dom->next_address);
Joerg Roedeld3086442008-06-26 21:27:57 +02001361
Joerg Roedel1c655772008-09-04 18:40:05 +02001362 if (address == -1) {
Joerg Roedel803b8cb2009-05-18 15:32:48 +02001363 dom->next_address = 0;
Joerg Roedel384de722009-05-15 12:30:05 +02001364 address = dma_ops_area_alloc(dev, dom, pages, align_mask,
1365 dma_mask, 0);
Joerg Roedel1c655772008-09-04 18:40:05 +02001366 dom->need_flush = true;
1367 }
Joerg Roedeld3086442008-06-26 21:27:57 +02001368
Joerg Roedel384de722009-05-15 12:30:05 +02001369 if (unlikely(address == -1))
FUJITA Tomonori8fd524b2009-11-15 21:19:53 +09001370 address = DMA_ERROR_CODE;
Joerg Roedeld3086442008-06-26 21:27:57 +02001371
1372 WARN_ON((address + (PAGE_SIZE*pages)) > dom->aperture_size);
1373
1374 return address;
1375}
1376
Joerg Roedel431b2a22008-07-11 17:14:22 +02001377/*
1378 * The address free function.
1379 *
1380 * called with domain->lock held
1381 */
Joerg Roedeld3086442008-06-26 21:27:57 +02001382static void dma_ops_free_addresses(struct dma_ops_domain *dom,
1383 unsigned long address,
1384 unsigned int pages)
1385{
Joerg Roedel384de722009-05-15 12:30:05 +02001386 unsigned i = address >> APERTURE_RANGE_SHIFT;
1387 struct aperture_range *range = dom->aperture[i];
Joerg Roedel80be3082008-11-06 14:59:05 +01001388
Joerg Roedel384de722009-05-15 12:30:05 +02001389 BUG_ON(i >= APERTURE_MAX_RANGES || range == NULL);
1390
Joerg Roedel47bccd62009-05-22 12:40:54 +02001391#ifdef CONFIG_IOMMU_STRESS
1392 if (i < 4)
1393 return;
1394#endif
1395
Joerg Roedel803b8cb2009-05-18 15:32:48 +02001396 if (address >= dom->next_address)
Joerg Roedel80be3082008-11-06 14:59:05 +01001397 dom->need_flush = true;
Joerg Roedel384de722009-05-15 12:30:05 +02001398
1399 address = (address % APERTURE_RANGE_SIZE) >> PAGE_SHIFT;
Joerg Roedel803b8cb2009-05-18 15:32:48 +02001400
Akinobu Mitaa66022c2009-12-15 16:48:28 -08001401 bitmap_clear(range->bitmap, address, pages);
Joerg Roedel384de722009-05-15 12:30:05 +02001402
Joerg Roedeld3086442008-06-26 21:27:57 +02001403}
1404
Joerg Roedel431b2a22008-07-11 17:14:22 +02001405/****************************************************************************
1406 *
1407 * The next functions belong to the domain allocation. A domain is
1408 * allocated for every IOMMU as the default domain. If device isolation
1409 * is enabled, every device get its own domain. The most important thing
1410 * about domains is the page table mapping the DMA address space they
1411 * contain.
1412 *
1413 ****************************************************************************/
1414
Joerg Roedelaeb26f52009-11-20 16:44:01 +01001415/*
1416 * This function adds a protection domain to the global protection domain list
1417 */
1418static void add_domain_to_list(struct protection_domain *domain)
1419{
1420 unsigned long flags;
1421
1422 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1423 list_add(&domain->list, &amd_iommu_pd_list);
1424 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1425}
1426
1427/*
1428 * This function removes a protection domain to the global
1429 * protection domain list
1430 */
1431static void del_domain_from_list(struct protection_domain *domain)
1432{
1433 unsigned long flags;
1434
1435 spin_lock_irqsave(&amd_iommu_pd_lock, flags);
1436 list_del(&domain->list);
1437 spin_unlock_irqrestore(&amd_iommu_pd_lock, flags);
1438}
1439
Joerg Roedelec487d12008-06-26 21:27:58 +02001440static u16 domain_id_alloc(void)
1441{
1442 unsigned long flags;
1443 int id;
1444
1445 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1446 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID);
1447 BUG_ON(id == 0);
1448 if (id > 0 && id < MAX_DOMAIN_ID)
1449 __set_bit(id, amd_iommu_pd_alloc_bitmap);
1450 else
1451 id = 0;
1452 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1453
1454 return id;
1455}
1456
Joerg Roedela2acfb72008-12-02 18:28:53 +01001457static void domain_id_free(int id)
1458{
1459 unsigned long flags;
1460
1461 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
1462 if (id > 0 && id < MAX_DOMAIN_ID)
1463 __clear_bit(id, amd_iommu_pd_alloc_bitmap);
1464 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1465}
Joerg Roedela2acfb72008-12-02 18:28:53 +01001466
Joerg Roedel86db2e52008-12-02 18:20:21 +01001467static void free_pagetable(struct protection_domain *domain)
Joerg Roedelec487d12008-06-26 21:27:58 +02001468{
1469 int i, j;
1470 u64 *p1, *p2, *p3;
1471
Joerg Roedel86db2e52008-12-02 18:20:21 +01001472 p1 = domain->pt_root;
Joerg Roedelec487d12008-06-26 21:27:58 +02001473
1474 if (!p1)
1475 return;
1476
1477 for (i = 0; i < 512; ++i) {
1478 if (!IOMMU_PTE_PRESENT(p1[i]))
1479 continue;
1480
1481 p2 = IOMMU_PTE_PAGE(p1[i]);
Joerg Roedel3cc3d842008-12-04 16:44:31 +01001482 for (j = 0; j < 512; ++j) {
Joerg Roedelec487d12008-06-26 21:27:58 +02001483 if (!IOMMU_PTE_PRESENT(p2[j]))
1484 continue;
1485 p3 = IOMMU_PTE_PAGE(p2[j]);
1486 free_page((unsigned long)p3);
1487 }
1488
1489 free_page((unsigned long)p2);
1490 }
1491
1492 free_page((unsigned long)p1);
Joerg Roedel86db2e52008-12-02 18:20:21 +01001493
1494 domain->pt_root = NULL;
Joerg Roedelec487d12008-06-26 21:27:58 +02001495}
1496
Joerg Roedel431b2a22008-07-11 17:14:22 +02001497/*
1498 * Free a domain, only used if something went wrong in the
1499 * allocation path and we need to free an already allocated page table
1500 */
Joerg Roedelec487d12008-06-26 21:27:58 +02001501static void dma_ops_domain_free(struct dma_ops_domain *dom)
1502{
Joerg Roedel384de722009-05-15 12:30:05 +02001503 int i;
1504
Joerg Roedelec487d12008-06-26 21:27:58 +02001505 if (!dom)
1506 return;
1507
Joerg Roedelaeb26f52009-11-20 16:44:01 +01001508 del_domain_from_list(&dom->domain);
1509
Joerg Roedel86db2e52008-12-02 18:20:21 +01001510 free_pagetable(&dom->domain);
Joerg Roedelec487d12008-06-26 21:27:58 +02001511
Joerg Roedel384de722009-05-15 12:30:05 +02001512 for (i = 0; i < APERTURE_MAX_RANGES; ++i) {
1513 if (!dom->aperture[i])
1514 continue;
1515 free_page((unsigned long)dom->aperture[i]->bitmap);
1516 kfree(dom->aperture[i]);
1517 }
Joerg Roedelec487d12008-06-26 21:27:58 +02001518
1519 kfree(dom);
1520}
1521
Joerg Roedel431b2a22008-07-11 17:14:22 +02001522/*
1523 * Allocates a new protection domain usable for the dma_ops functions.
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001524 * It also initializes the page table and the address allocator data
Joerg Roedel431b2a22008-07-11 17:14:22 +02001525 * structures required for the dma_ops interface
1526 */
Joerg Roedel87a64d52009-11-24 17:26:43 +01001527static struct dma_ops_domain *dma_ops_domain_alloc(void)
Joerg Roedelec487d12008-06-26 21:27:58 +02001528{
1529 struct dma_ops_domain *dma_dom;
Joerg Roedelec487d12008-06-26 21:27:58 +02001530
1531 dma_dom = kzalloc(sizeof(struct dma_ops_domain), GFP_KERNEL);
1532 if (!dma_dom)
1533 return NULL;
1534
1535 spin_lock_init(&dma_dom->domain.lock);
1536
1537 dma_dom->domain.id = domain_id_alloc();
1538 if (dma_dom->domain.id == 0)
1539 goto free_dma_dom;
Joerg Roedel7c392cb2009-11-26 11:13:32 +01001540 INIT_LIST_HEAD(&dma_dom->domain.dev_list);
Joerg Roedel8f7a0172009-09-02 16:55:24 +02001541 dma_dom->domain.mode = PAGE_MODE_2_LEVEL;
Joerg Roedelec487d12008-06-26 21:27:58 +02001542 dma_dom->domain.pt_root = (void *)get_zeroed_page(GFP_KERNEL);
Joerg Roedel9fdb19d2008-12-02 17:46:25 +01001543 dma_dom->domain.flags = PD_DMA_OPS_MASK;
Joerg Roedelec487d12008-06-26 21:27:58 +02001544 dma_dom->domain.priv = dma_dom;
1545 if (!dma_dom->domain.pt_root)
1546 goto free_dma_dom;
Joerg Roedelec487d12008-06-26 21:27:58 +02001547
Joerg Roedel1c655772008-09-04 18:40:05 +02001548 dma_dom->need_flush = false;
Joerg Roedelbd60b732008-09-11 10:24:48 +02001549 dma_dom->target_dev = 0xffff;
Joerg Roedel1c655772008-09-04 18:40:05 +02001550
Joerg Roedelaeb26f52009-11-20 16:44:01 +01001551 add_domain_to_list(&dma_dom->domain);
1552
Joerg Roedel576175c2009-11-23 19:08:46 +01001553 if (alloc_new_range(dma_dom, true, GFP_KERNEL))
Joerg Roedelec487d12008-06-26 21:27:58 +02001554 goto free_dma_dom;
Joerg Roedelec487d12008-06-26 21:27:58 +02001555
Joerg Roedel431b2a22008-07-11 17:14:22 +02001556 /*
Joerg Roedelec487d12008-06-26 21:27:58 +02001557 * mark the first page as allocated so we never return 0 as
1558 * a valid dma-address. So we can use 0 as error value
Joerg Roedel431b2a22008-07-11 17:14:22 +02001559 */
Joerg Roedel384de722009-05-15 12:30:05 +02001560 dma_dom->aperture[0]->bitmap[0] = 1;
Joerg Roedel803b8cb2009-05-18 15:32:48 +02001561 dma_dom->next_address = 0;
Joerg Roedelec487d12008-06-26 21:27:58 +02001562
Joerg Roedelec487d12008-06-26 21:27:58 +02001563
1564 return dma_dom;
1565
1566free_dma_dom:
1567 dma_ops_domain_free(dma_dom);
1568
1569 return NULL;
1570}
1571
Joerg Roedel431b2a22008-07-11 17:14:22 +02001572/*
Joerg Roedel5b28df62008-12-02 17:49:42 +01001573 * little helper function to check whether a given protection domain is a
1574 * dma_ops domain
1575 */
1576static bool dma_ops_domain(struct protection_domain *domain)
1577{
1578 return domain->flags & PD_DMA_OPS_MASK;
1579}
1580
Joerg Roedelfd7b5532011-04-05 15:31:08 +02001581static void set_dte_entry(u16 devid, struct protection_domain *domain, bool ats)
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001582{
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001583 u64 pte_root = virt_to_phys(domain->pt_root);
Joerg Roedelfd7b5532011-04-05 15:31:08 +02001584 u32 flags = 0;
Joerg Roedel863c74e2008-12-02 17:56:36 +01001585
Joerg Roedel38ddf412008-09-11 10:38:32 +02001586 pte_root |= (domain->mode & DEV_ENTRY_MODE_MASK)
1587 << DEV_ENTRY_MODE_SHIFT;
1588 pte_root |= IOMMU_PTE_IR | IOMMU_PTE_IW | IOMMU_PTE_P | IOMMU_PTE_TV;
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001589
Joerg Roedelfd7b5532011-04-05 15:31:08 +02001590 if (ats)
1591 flags |= DTE_FLAG_IOTLB;
1592
1593 amd_iommu_dev_table[devid].data[3] |= flags;
1594 amd_iommu_dev_table[devid].data[2] = domain->id;
1595 amd_iommu_dev_table[devid].data[1] = upper_32_bits(pte_root);
1596 amd_iommu_dev_table[devid].data[0] = lower_32_bits(pte_root);
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001597}
1598
Joerg Roedel15898bb2009-11-24 15:39:42 +01001599static void clear_dte_entry(u16 devid)
Joerg Roedel355bf552008-12-08 12:02:41 +01001600{
Joerg Roedel355bf552008-12-08 12:02:41 +01001601 /* remove entry from the device table seen by the hardware */
1602 amd_iommu_dev_table[devid].data[0] = IOMMU_PTE_P | IOMMU_PTE_TV;
1603 amd_iommu_dev_table[devid].data[1] = 0;
1604 amd_iommu_dev_table[devid].data[2] = 0;
1605
Joerg Roedelc5cca142009-10-09 18:31:20 +02001606 amd_iommu_apply_erratum_63(devid);
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001607}
1608
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001609static void do_attach(struct iommu_dev_data *dev_data,
1610 struct protection_domain *domain)
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001611{
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001612 struct amd_iommu *iommu;
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001613 bool ats;
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001614
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001615 iommu = amd_iommu_rlookup_table[dev_data->devid];
1616 ats = dev_data->ats.enabled;
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001617
1618 /* Update data structures */
1619 dev_data->domain = domain;
1620 list_add(&dev_data->list, &domain->dev_list);
Joerg Roedelf62dda62011-06-09 12:55:35 +02001621 set_dte_entry(dev_data->devid, domain, ats);
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001622
1623 /* Do reference counting */
1624 domain->dev_iommu[iommu->index] += 1;
1625 domain->dev_cnt += 1;
1626
1627 /* Flush the DTE entry */
Joerg Roedel6c542042011-06-09 17:07:31 +02001628 device_flush_dte(dev_data);
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001629}
1630
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001631static void do_detach(struct iommu_dev_data *dev_data)
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001632{
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001633 struct amd_iommu *iommu;
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001634
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001635 iommu = amd_iommu_rlookup_table[dev_data->devid];
Joerg Roedelc5cca142009-10-09 18:31:20 +02001636
Joerg Roedelc4596112009-11-20 14:57:32 +01001637 /* decrease reference counters */
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001638 dev_data->domain->dev_iommu[iommu->index] -= 1;
1639 dev_data->domain->dev_cnt -= 1;
Joerg Roedel355bf552008-12-08 12:02:41 +01001640
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001641 /* Update data structures */
1642 dev_data->domain = NULL;
1643 list_del(&dev_data->list);
Joerg Roedelf62dda62011-06-09 12:55:35 +02001644 clear_dte_entry(dev_data->devid);
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001645
1646 /* Flush the DTE entry */
Joerg Roedel6c542042011-06-09 17:07:31 +02001647 device_flush_dte(dev_data);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001648}
1649
1650/*
1651 * If a device is not yet associated with a domain, this function does
1652 * assigns it visible for the hardware
1653 */
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001654static int __attach_device(struct iommu_dev_data *dev_data,
Joerg Roedel15898bb2009-11-24 15:39:42 +01001655 struct protection_domain *domain)
1656{
Julia Lawall84fe6c12010-05-27 12:31:51 +02001657 int ret;
Joerg Roedel657cbb62009-11-23 15:26:46 +01001658
Joerg Roedel15898bb2009-11-24 15:39:42 +01001659 /* lock domain */
1660 spin_lock(&domain->lock);
1661
Joerg Roedel71f77582011-06-09 19:03:15 +02001662 if (dev_data->alias_data != NULL) {
1663 struct iommu_dev_data *alias_data = dev_data->alias_data;
Joerg Roedel15898bb2009-11-24 15:39:42 +01001664
Joerg Roedel2b02b092011-06-09 17:48:39 +02001665 /* Some sanity checks */
1666 ret = -EBUSY;
1667 if (alias_data->domain != NULL &&
1668 alias_data->domain != domain)
1669 goto out_unlock;
Joerg Roedel15898bb2009-11-24 15:39:42 +01001670
Joerg Roedel2b02b092011-06-09 17:48:39 +02001671 if (dev_data->domain != NULL &&
1672 dev_data->domain != domain)
1673 goto out_unlock;
1674
1675 /* Do real assignment */
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001676 if (alias_data->domain == NULL)
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001677 do_attach(alias_data, domain);
Joerg Roedel24100052009-11-25 15:59:57 +01001678
1679 atomic_inc(&alias_data->bind);
Joerg Roedel657cbb62009-11-23 15:26:46 +01001680 }
Joerg Roedel15898bb2009-11-24 15:39:42 +01001681
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001682 if (dev_data->domain == NULL)
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001683 do_attach(dev_data, domain);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001684
Joerg Roedel24100052009-11-25 15:59:57 +01001685 atomic_inc(&dev_data->bind);
1686
Julia Lawall84fe6c12010-05-27 12:31:51 +02001687 ret = 0;
1688
1689out_unlock:
1690
Joerg Roedel355bf552008-12-08 12:02:41 +01001691 /* ready */
1692 spin_unlock(&domain->lock);
Joerg Roedel21129f72009-09-01 11:59:42 +02001693
Julia Lawall84fe6c12010-05-27 12:31:51 +02001694 return ret;
Joerg Roedel15898bb2009-11-24 15:39:42 +01001695}
1696
1697/*
1698 * If a device is not yet associated with a domain, this function does
1699 * assigns it visible for the hardware
1700 */
1701static int attach_device(struct device *dev,
1702 struct protection_domain *domain)
1703{
Joerg Roedelfd7b5532011-04-05 15:31:08 +02001704 struct pci_dev *pdev = to_pci_dev(dev);
Joerg Roedelea61cdd2011-06-09 12:56:30 +02001705 struct iommu_dev_data *dev_data;
Joerg Roedel15898bb2009-11-24 15:39:42 +01001706 unsigned long flags;
1707 int ret;
1708
Joerg Roedelea61cdd2011-06-09 12:56:30 +02001709 dev_data = get_dev_data(dev);
1710
1711 if (amd_iommu_iotlb_sup && pci_enable_ats(pdev, PAGE_SHIFT) == 0) {
1712 dev_data->ats.enabled = true;
1713 dev_data->ats.qdep = pci_ats_queue_depth(pdev);
1714 }
Joerg Roedelfd7b5532011-04-05 15:31:08 +02001715
Joerg Roedel15898bb2009-11-24 15:39:42 +01001716 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001717 ret = __attach_device(dev_data, domain);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001718 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
1719
1720 /*
1721 * We might boot into a crash-kernel here. The crashed kernel
1722 * left the caches in the IOMMU dirty. So we have to flush
1723 * here to evict all dirty stuff.
1724 */
Joerg Roedel17b124b2011-04-06 18:01:35 +02001725 domain_flush_tlb_pde(domain);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001726
1727 return ret;
1728}
1729
1730/*
1731 * Removes a device from a protection domain (unlocked)
1732 */
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001733static void __detach_device(struct iommu_dev_data *dev_data)
Joerg Roedel15898bb2009-11-24 15:39:42 +01001734{
Joerg Roedel2ca76272010-01-22 16:45:31 +01001735 struct protection_domain *domain;
Joerg Roedel7c392cb2009-11-26 11:13:32 +01001736 unsigned long flags;
Joerg Roedel15898bb2009-11-24 15:39:42 +01001737
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001738 BUG_ON(!dev_data->domain);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001739
Joerg Roedel2ca76272010-01-22 16:45:31 +01001740 domain = dev_data->domain;
1741
1742 spin_lock_irqsave(&domain->lock, flags);
Joerg Roedel24100052009-11-25 15:59:57 +01001743
Joerg Roedel71f77582011-06-09 19:03:15 +02001744 if (dev_data->alias_data != NULL) {
1745 struct iommu_dev_data *alias_data = dev_data->alias_data;
1746
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001747 if (atomic_dec_and_test(&alias_data->bind))
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001748 do_detach(alias_data);
Joerg Roedel24100052009-11-25 15:59:57 +01001749 }
1750
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001751 if (atomic_dec_and_test(&dev_data->bind))
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001752 do_detach(dev_data);
Joerg Roedel7f760dd2009-11-26 14:49:59 +01001753
Joerg Roedel2ca76272010-01-22 16:45:31 +01001754 spin_unlock_irqrestore(&domain->lock, flags);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001755
Joerg Roedel21129f72009-09-01 11:59:42 +02001756 /*
1757 * If we run in passthrough mode the device must be assigned to the
Joerg Roedeld3ad9372010-01-22 17:55:27 +01001758 * passthrough domain if it is detached from any other domain.
1759 * Make sure we can deassign from the pt_domain itself.
Joerg Roedel21129f72009-09-01 11:59:42 +02001760 */
Joerg Roedeld3ad9372010-01-22 17:55:27 +01001761 if (iommu_pass_through &&
1762 (dev_data->domain == NULL && domain != pt_domain))
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001763 __attach_device(dev_data, pt_domain);
Joerg Roedel355bf552008-12-08 12:02:41 +01001764}
1765
1766/*
1767 * Removes a device from a protection domain (with devtable_lock held)
1768 */
Joerg Roedel15898bb2009-11-24 15:39:42 +01001769static void detach_device(struct device *dev)
Joerg Roedel355bf552008-12-08 12:02:41 +01001770{
Joerg Roedelea61cdd2011-06-09 12:56:30 +02001771 struct iommu_dev_data *dev_data;
Joerg Roedel355bf552008-12-08 12:02:41 +01001772 unsigned long flags;
1773
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001774 dev_data = get_dev_data(dev);
1775
Joerg Roedel355bf552008-12-08 12:02:41 +01001776 /* lock device table */
1777 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
Joerg Roedelec9e79e2011-06-09 17:25:50 +02001778 __detach_device(dev_data);
Joerg Roedel355bf552008-12-08 12:02:41 +01001779 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
Joerg Roedelfd7b5532011-04-05 15:31:08 +02001780
Joerg Roedelea61cdd2011-06-09 12:56:30 +02001781 if (dev_data->ats.enabled) {
1782 pci_disable_ats(to_pci_dev(dev));
1783 dev_data->ats.enabled = false;
1784 }
Joerg Roedel355bf552008-12-08 12:02:41 +01001785}
Joerg Roedele275a2a2008-12-10 18:27:25 +01001786
Joerg Roedel15898bb2009-11-24 15:39:42 +01001787/*
1788 * Find out the protection domain structure for a given PCI device. This
1789 * will give us the pointer to the page table root for example.
1790 */
1791static struct protection_domain *domain_for_device(struct device *dev)
1792{
Joerg Roedel71f77582011-06-09 19:03:15 +02001793 struct iommu_dev_data *dev_data;
Joerg Roedel2b02b092011-06-09 17:48:39 +02001794 struct protection_domain *dom = NULL;
Joerg Roedel15898bb2009-11-24 15:39:42 +01001795 unsigned long flags;
Joerg Roedel15898bb2009-11-24 15:39:42 +01001796
Joerg Roedel657cbb62009-11-23 15:26:46 +01001797 dev_data = get_dev_data(dev);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001798
Joerg Roedel2b02b092011-06-09 17:48:39 +02001799 if (dev_data->domain)
1800 return dev_data->domain;
1801
Joerg Roedel71f77582011-06-09 19:03:15 +02001802 if (dev_data->alias_data != NULL) {
1803 struct iommu_dev_data *alias_data = dev_data->alias_data;
Joerg Roedel2b02b092011-06-09 17:48:39 +02001804
1805 read_lock_irqsave(&amd_iommu_devtable_lock, flags);
1806 if (alias_data->domain != NULL) {
1807 __attach_device(dev_data, alias_data->domain);
1808 dom = alias_data->domain;
1809 }
1810 read_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001811 }
1812
Joerg Roedel15898bb2009-11-24 15:39:42 +01001813 return dom;
1814}
1815
Joerg Roedele275a2a2008-12-10 18:27:25 +01001816static int device_change_notifier(struct notifier_block *nb,
1817 unsigned long action, void *data)
1818{
1819 struct device *dev = data;
Joerg Roedel98fc5a62009-11-24 17:19:23 +01001820 u16 devid;
Joerg Roedele275a2a2008-12-10 18:27:25 +01001821 struct protection_domain *domain;
1822 struct dma_ops_domain *dma_domain;
1823 struct amd_iommu *iommu;
Joerg Roedel1ac4cbb2008-12-10 19:33:26 +01001824 unsigned long flags;
Joerg Roedele275a2a2008-12-10 18:27:25 +01001825
Joerg Roedel98fc5a62009-11-24 17:19:23 +01001826 if (!check_device(dev))
1827 return 0;
Joerg Roedele275a2a2008-12-10 18:27:25 +01001828
Joerg Roedel98fc5a62009-11-24 17:19:23 +01001829 devid = get_device_id(dev);
1830 iommu = amd_iommu_rlookup_table[devid];
Joerg Roedele275a2a2008-12-10 18:27:25 +01001831
1832 switch (action) {
Chris Wrightc1eee672009-05-21 00:56:58 -07001833 case BUS_NOTIFY_UNBOUND_DRIVER:
Joerg Roedel657cbb62009-11-23 15:26:46 +01001834
1835 domain = domain_for_device(dev);
1836
Joerg Roedele275a2a2008-12-10 18:27:25 +01001837 if (!domain)
1838 goto out;
Joerg Roedela1ca3312009-09-01 12:22:22 +02001839 if (iommu_pass_through)
1840 break;
Joerg Roedel15898bb2009-11-24 15:39:42 +01001841 detach_device(dev);
Joerg Roedele275a2a2008-12-10 18:27:25 +01001842 break;
Joerg Roedel1ac4cbb2008-12-10 19:33:26 +01001843 case BUS_NOTIFY_ADD_DEVICE:
Joerg Roedel657cbb62009-11-23 15:26:46 +01001844
1845 iommu_init_device(dev);
1846
1847 domain = domain_for_device(dev);
1848
Joerg Roedel1ac4cbb2008-12-10 19:33:26 +01001849 /* allocate a protection domain if a device is added */
1850 dma_domain = find_protection_domain(devid);
1851 if (dma_domain)
1852 goto out;
Joerg Roedel87a64d52009-11-24 17:26:43 +01001853 dma_domain = dma_ops_domain_alloc();
Joerg Roedel1ac4cbb2008-12-10 19:33:26 +01001854 if (!dma_domain)
1855 goto out;
1856 dma_domain->target_dev = devid;
1857
1858 spin_lock_irqsave(&iommu_pd_list_lock, flags);
1859 list_add_tail(&dma_domain->list, &iommu_pd_list);
1860 spin_unlock_irqrestore(&iommu_pd_list_lock, flags);
1861
1862 break;
Joerg Roedel657cbb62009-11-23 15:26:46 +01001863 case BUS_NOTIFY_DEL_DEVICE:
1864
1865 iommu_uninit_device(dev);
1866
Joerg Roedele275a2a2008-12-10 18:27:25 +01001867 default:
1868 goto out;
1869 }
1870
Joerg Roedele275a2a2008-12-10 18:27:25 +01001871 iommu_completion_wait(iommu);
1872
1873out:
1874 return 0;
1875}
1876
Jaswinder Singh Rajputb25ae672009-07-01 19:53:14 +05301877static struct notifier_block device_nb = {
Joerg Roedele275a2a2008-12-10 18:27:25 +01001878 .notifier_call = device_change_notifier,
1879};
Joerg Roedel355bf552008-12-08 12:02:41 +01001880
Joerg Roedel8638c492009-12-10 11:12:25 +01001881void amd_iommu_init_notifier(void)
1882{
1883 bus_register_notifier(&pci_bus_type, &device_nb);
1884}
1885
Joerg Roedel431b2a22008-07-11 17:14:22 +02001886/*****************************************************************************
1887 *
1888 * The next functions belong to the dma_ops mapping/unmapping code.
1889 *
1890 *****************************************************************************/
1891
1892/*
1893 * In the dma_ops path we only have the struct device. This function
1894 * finds the corresponding IOMMU, the protection domain and the
1895 * requestor id for a given device.
1896 * If the device is not yet associated with a domain this is also done
1897 * in this function.
1898 */
Joerg Roedel94f6d192009-11-24 16:40:02 +01001899static struct protection_domain *get_domain(struct device *dev)
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001900{
Joerg Roedel94f6d192009-11-24 16:40:02 +01001901 struct protection_domain *domain;
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001902 struct dma_ops_domain *dma_dom;
Joerg Roedel94f6d192009-11-24 16:40:02 +01001903 u16 devid = get_device_id(dev);
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001904
Joerg Roedelf99c0f12009-11-23 16:52:56 +01001905 if (!check_device(dev))
Joerg Roedel94f6d192009-11-24 16:40:02 +01001906 return ERR_PTR(-EINVAL);
Joerg Roedeldbcc1122008-09-04 15:04:26 +02001907
Joerg Roedel94f6d192009-11-24 16:40:02 +01001908 domain = domain_for_device(dev);
1909 if (domain != NULL && !dma_ops_domain(domain))
1910 return ERR_PTR(-EBUSY);
Joerg Roedelf99c0f12009-11-23 16:52:56 +01001911
Joerg Roedel94f6d192009-11-24 16:40:02 +01001912 if (domain != NULL)
1913 return domain;
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001914
Joerg Roedel15898bb2009-11-24 15:39:42 +01001915 /* Device not bount yet - bind it */
Joerg Roedel94f6d192009-11-24 16:40:02 +01001916 dma_dom = find_protection_domain(devid);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001917 if (!dma_dom)
Joerg Roedel94f6d192009-11-24 16:40:02 +01001918 dma_dom = amd_iommu_rlookup_table[devid]->default_dom;
1919 attach_device(dev, &dma_dom->domain);
Joerg Roedel15898bb2009-11-24 15:39:42 +01001920 DUMP_printk("Using protection domain %d for device %s\n",
Joerg Roedel94f6d192009-11-24 16:40:02 +01001921 dma_dom->domain.id, dev_name(dev));
Joerg Roedelf91ba192008-11-25 12:56:12 +01001922
Joerg Roedel94f6d192009-11-24 16:40:02 +01001923 return &dma_dom->domain;
Joerg Roedelb20ac0d2008-06-26 21:27:59 +02001924}
1925
Joerg Roedel04bfdd82009-09-02 16:00:23 +02001926static void update_device_table(struct protection_domain *domain)
1927{
Joerg Roedel492667d2009-11-27 13:25:47 +01001928 struct iommu_dev_data *dev_data;
Joerg Roedel04bfdd82009-09-02 16:00:23 +02001929
Joerg Roedelea61cdd2011-06-09 12:56:30 +02001930 list_for_each_entry(dev_data, &domain->dev_list, list)
1931 set_dte_entry(dev_data->devid, domain, dev_data->ats.enabled);
Joerg Roedel04bfdd82009-09-02 16:00:23 +02001932}
1933
1934static void update_domain(struct protection_domain *domain)
1935{
1936 if (!domain->updated)
1937 return;
1938
1939 update_device_table(domain);
Joerg Roedel17b124b2011-04-06 18:01:35 +02001940
1941 domain_flush_devices(domain);
1942 domain_flush_tlb_pde(domain);
Joerg Roedel04bfdd82009-09-02 16:00:23 +02001943
1944 domain->updated = false;
1945}
1946
Joerg Roedel431b2a22008-07-11 17:14:22 +02001947/*
Joerg Roedel8bda3092009-05-12 12:02:46 +02001948 * This function fetches the PTE for a given address in the aperture
1949 */
1950static u64* dma_ops_get_pte(struct dma_ops_domain *dom,
1951 unsigned long address)
1952{
Joerg Roedel384de722009-05-15 12:30:05 +02001953 struct aperture_range *aperture;
Joerg Roedel8bda3092009-05-12 12:02:46 +02001954 u64 *pte, *pte_page;
1955
Joerg Roedel384de722009-05-15 12:30:05 +02001956 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
1957 if (!aperture)
1958 return NULL;
1959
1960 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
Joerg Roedel8bda3092009-05-12 12:02:46 +02001961 if (!pte) {
Joerg Roedelcbb9d722010-01-15 14:41:15 +01001962 pte = alloc_pte(&dom->domain, address, PAGE_SIZE, &pte_page,
Joerg Roedelabdc5eb2009-09-03 11:33:51 +02001963 GFP_ATOMIC);
Joerg Roedel384de722009-05-15 12:30:05 +02001964 aperture->pte_pages[APERTURE_PAGE_INDEX(address)] = pte_page;
1965 } else
Joerg Roedel8c8c1432009-09-02 17:30:00 +02001966 pte += PM_LEVEL_INDEX(0, address);
Joerg Roedel8bda3092009-05-12 12:02:46 +02001967
Joerg Roedel04bfdd82009-09-02 16:00:23 +02001968 update_domain(&dom->domain);
Joerg Roedel8bda3092009-05-12 12:02:46 +02001969
1970 return pte;
1971}
1972
1973/*
Joerg Roedel431b2a22008-07-11 17:14:22 +02001974 * This is the generic map function. It maps one 4kb page at paddr to
1975 * the given address in the DMA address space for the domain.
1976 */
Joerg Roedel680525e2009-11-23 18:44:42 +01001977static dma_addr_t dma_ops_domain_map(struct dma_ops_domain *dom,
Joerg Roedelcb76c322008-06-26 21:28:00 +02001978 unsigned long address,
1979 phys_addr_t paddr,
1980 int direction)
1981{
1982 u64 *pte, __pte;
1983
1984 WARN_ON(address > dom->aperture_size);
1985
1986 paddr &= PAGE_MASK;
1987
Joerg Roedel8bda3092009-05-12 12:02:46 +02001988 pte = dma_ops_get_pte(dom, address);
Joerg Roedel53812c12009-05-12 12:17:38 +02001989 if (!pte)
FUJITA Tomonori8fd524b2009-11-15 21:19:53 +09001990 return DMA_ERROR_CODE;
Joerg Roedelcb76c322008-06-26 21:28:00 +02001991
1992 __pte = paddr | IOMMU_PTE_P | IOMMU_PTE_FC;
1993
1994 if (direction == DMA_TO_DEVICE)
1995 __pte |= IOMMU_PTE_IR;
1996 else if (direction == DMA_FROM_DEVICE)
1997 __pte |= IOMMU_PTE_IW;
1998 else if (direction == DMA_BIDIRECTIONAL)
1999 __pte |= IOMMU_PTE_IR | IOMMU_PTE_IW;
2000
2001 WARN_ON(*pte);
2002
2003 *pte = __pte;
2004
2005 return (dma_addr_t)address;
2006}
2007
Joerg Roedel431b2a22008-07-11 17:14:22 +02002008/*
2009 * The generic unmapping function for on page in the DMA address space.
2010 */
Joerg Roedel680525e2009-11-23 18:44:42 +01002011static void dma_ops_domain_unmap(struct dma_ops_domain *dom,
Joerg Roedelcb76c322008-06-26 21:28:00 +02002012 unsigned long address)
2013{
Joerg Roedel384de722009-05-15 12:30:05 +02002014 struct aperture_range *aperture;
Joerg Roedelcb76c322008-06-26 21:28:00 +02002015 u64 *pte;
2016
2017 if (address >= dom->aperture_size)
2018 return;
2019
Joerg Roedel384de722009-05-15 12:30:05 +02002020 aperture = dom->aperture[APERTURE_RANGE_INDEX(address)];
2021 if (!aperture)
2022 return;
Joerg Roedelcb76c322008-06-26 21:28:00 +02002023
Joerg Roedel384de722009-05-15 12:30:05 +02002024 pte = aperture->pte_pages[APERTURE_PAGE_INDEX(address)];
2025 if (!pte)
2026 return;
2027
Joerg Roedel8c8c1432009-09-02 17:30:00 +02002028 pte += PM_LEVEL_INDEX(0, address);
Joerg Roedelcb76c322008-06-26 21:28:00 +02002029
2030 WARN_ON(!*pte);
2031
2032 *pte = 0ULL;
2033}
2034
Joerg Roedel431b2a22008-07-11 17:14:22 +02002035/*
2036 * This function contains common code for mapping of a physically
Joerg Roedel24f81162008-12-08 14:25:39 +01002037 * contiguous memory region into DMA address space. It is used by all
2038 * mapping functions provided with this IOMMU driver.
Joerg Roedel431b2a22008-07-11 17:14:22 +02002039 * Must be called with the domain lock held.
2040 */
Joerg Roedelcb76c322008-06-26 21:28:00 +02002041static dma_addr_t __map_single(struct device *dev,
Joerg Roedelcb76c322008-06-26 21:28:00 +02002042 struct dma_ops_domain *dma_dom,
2043 phys_addr_t paddr,
2044 size_t size,
Joerg Roedel6d4f3432008-09-04 19:18:02 +02002045 int dir,
Joerg Roedel832a90c2008-09-18 15:54:23 +02002046 bool align,
2047 u64 dma_mask)
Joerg Roedelcb76c322008-06-26 21:28:00 +02002048{
2049 dma_addr_t offset = paddr & ~PAGE_MASK;
Joerg Roedel53812c12009-05-12 12:17:38 +02002050 dma_addr_t address, start, ret;
Joerg Roedelcb76c322008-06-26 21:28:00 +02002051 unsigned int pages;
Joerg Roedel6d4f3432008-09-04 19:18:02 +02002052 unsigned long align_mask = 0;
Joerg Roedelcb76c322008-06-26 21:28:00 +02002053 int i;
2054
Joerg Roedele3c449f2008-10-15 22:02:11 -07002055 pages = iommu_num_pages(paddr, size, PAGE_SIZE);
Joerg Roedelcb76c322008-06-26 21:28:00 +02002056 paddr &= PAGE_MASK;
2057
Joerg Roedel8ecaf8f2008-12-12 16:13:04 +01002058 INC_STATS_COUNTER(total_map_requests);
2059
Joerg Roedelc1858972008-12-12 15:42:39 +01002060 if (pages > 1)
2061 INC_STATS_COUNTER(cross_page);
2062
Joerg Roedel6d4f3432008-09-04 19:18:02 +02002063 if (align)
2064 align_mask = (1UL << get_order(size)) - 1;
2065
Joerg Roedel11b83882009-05-19 10:23:15 +02002066retry:
Joerg Roedel832a90c2008-09-18 15:54:23 +02002067 address = dma_ops_alloc_addresses(dev, dma_dom, pages, align_mask,
2068 dma_mask);
FUJITA Tomonori8fd524b2009-11-15 21:19:53 +09002069 if (unlikely(address == DMA_ERROR_CODE)) {
Joerg Roedel11b83882009-05-19 10:23:15 +02002070 /*
2071 * setting next_address here will let the address
2072 * allocator only scan the new allocated range in the
2073 * first run. This is a small optimization.
2074 */
2075 dma_dom->next_address = dma_dom->aperture_size;
2076
Joerg Roedel576175c2009-11-23 19:08:46 +01002077 if (alloc_new_range(dma_dom, false, GFP_ATOMIC))
Joerg Roedel11b83882009-05-19 10:23:15 +02002078 goto out;
2079
2080 /*
André Goddard Rosaaf901ca2009-11-14 13:09:05 -02002081 * aperture was successfully enlarged by 128 MB, try
Joerg Roedel11b83882009-05-19 10:23:15 +02002082 * allocation again
2083 */
2084 goto retry;
2085 }
Joerg Roedelcb76c322008-06-26 21:28:00 +02002086
2087 start = address;
2088 for (i = 0; i < pages; ++i) {
Joerg Roedel680525e2009-11-23 18:44:42 +01002089 ret = dma_ops_domain_map(dma_dom, start, paddr, dir);
FUJITA Tomonori8fd524b2009-11-15 21:19:53 +09002090 if (ret == DMA_ERROR_CODE)
Joerg Roedel53812c12009-05-12 12:17:38 +02002091 goto out_unmap;
2092
Joerg Roedelcb76c322008-06-26 21:28:00 +02002093 paddr += PAGE_SIZE;
2094 start += PAGE_SIZE;
2095 }
2096 address += offset;
2097
Joerg Roedel5774f7c2008-12-12 15:57:30 +01002098 ADD_STATS_COUNTER(alloced_io_mem, size);
2099
FUJITA Tomonoriafa9fdc2008-09-20 01:23:30 +09002100 if (unlikely(dma_dom->need_flush && !amd_iommu_unmap_flush)) {
Joerg Roedel17b124b2011-04-06 18:01:35 +02002101 domain_flush_tlb(&dma_dom->domain);
Joerg Roedel1c655772008-09-04 18:40:05 +02002102 dma_dom->need_flush = false;
Joerg Roedel318afd42009-11-23 18:32:38 +01002103 } else if (unlikely(amd_iommu_np_cache))
Joerg Roedel17b124b2011-04-06 18:01:35 +02002104 domain_flush_pages(&dma_dom->domain, address, size);
Joerg Roedel270cab242008-09-04 15:49:46 +02002105
Joerg Roedelcb76c322008-06-26 21:28:00 +02002106out:
2107 return address;
Joerg Roedel53812c12009-05-12 12:17:38 +02002108
2109out_unmap:
2110
2111 for (--i; i >= 0; --i) {
2112 start -= PAGE_SIZE;
Joerg Roedel680525e2009-11-23 18:44:42 +01002113 dma_ops_domain_unmap(dma_dom, start);
Joerg Roedel53812c12009-05-12 12:17:38 +02002114 }
2115
2116 dma_ops_free_addresses(dma_dom, address, pages);
2117
FUJITA Tomonori8fd524b2009-11-15 21:19:53 +09002118 return DMA_ERROR_CODE;
Joerg Roedelcb76c322008-06-26 21:28:00 +02002119}
2120
Joerg Roedel431b2a22008-07-11 17:14:22 +02002121/*
2122 * Does the reverse of the __map_single function. Must be called with
2123 * the domain lock held too
2124 */
Joerg Roedelcd8c82e2009-11-23 19:33:56 +01002125static void __unmap_single(struct dma_ops_domain *dma_dom,
Joerg Roedelcb76c322008-06-26 21:28:00 +02002126 dma_addr_t dma_addr,
2127 size_t size,
2128 int dir)
2129{
Joerg Roedel04e04632010-09-23 16:12:48 +02002130 dma_addr_t flush_addr;
Joerg Roedelcb76c322008-06-26 21:28:00 +02002131 dma_addr_t i, start;
2132 unsigned int pages;
2133
FUJITA Tomonori8fd524b2009-11-15 21:19:53 +09002134 if ((dma_addr == DMA_ERROR_CODE) ||
Joerg Roedelb8d99052008-12-08 14:40:26 +01002135 (dma_addr + size > dma_dom->aperture_size))
Joerg Roedelcb76c322008-06-26 21:28:00 +02002136 return;
2137
Joerg Roedel04e04632010-09-23 16:12:48 +02002138 flush_addr = dma_addr;
Joerg Roedele3c449f2008-10-15 22:02:11 -07002139 pages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
Joerg Roedelcb76c322008-06-26 21:28:00 +02002140 dma_addr &= PAGE_MASK;
2141 start = dma_addr;
2142
2143 for (i = 0; i < pages; ++i) {
Joerg Roedel680525e2009-11-23 18:44:42 +01002144 dma_ops_domain_unmap(dma_dom, start);
Joerg Roedelcb76c322008-06-26 21:28:00 +02002145 start += PAGE_SIZE;
2146 }
2147
Joerg Roedel5774f7c2008-12-12 15:57:30 +01002148 SUB_STATS_COUNTER(alloced_io_mem, size);
2149
Joerg Roedelcb76c322008-06-26 21:28:00 +02002150 dma_ops_free_addresses(dma_dom, dma_addr, pages);
Joerg Roedel270cab242008-09-04 15:49:46 +02002151
Joerg Roedel80be3082008-11-06 14:59:05 +01002152 if (amd_iommu_unmap_flush || dma_dom->need_flush) {
Joerg Roedel17b124b2011-04-06 18:01:35 +02002153 domain_flush_pages(&dma_dom->domain, flush_addr, size);
Joerg Roedel80be3082008-11-06 14:59:05 +01002154 dma_dom->need_flush = false;
2155 }
Joerg Roedelcb76c322008-06-26 21:28:00 +02002156}
2157
Joerg Roedel431b2a22008-07-11 17:14:22 +02002158/*
2159 * The exported map_single function for dma_ops.
2160 */
FUJITA Tomonori51491362009-01-05 23:47:25 +09002161static dma_addr_t map_page(struct device *dev, struct page *page,
2162 unsigned long offset, size_t size,
2163 enum dma_data_direction dir,
2164 struct dma_attrs *attrs)
Joerg Roedel4da70b92008-06-26 21:28:01 +02002165{
2166 unsigned long flags;
Joerg Roedel4da70b92008-06-26 21:28:01 +02002167 struct protection_domain *domain;
Joerg Roedel4da70b92008-06-26 21:28:01 +02002168 dma_addr_t addr;
Joerg Roedel832a90c2008-09-18 15:54:23 +02002169 u64 dma_mask;
FUJITA Tomonori51491362009-01-05 23:47:25 +09002170 phys_addr_t paddr = page_to_phys(page) + offset;
Joerg Roedel4da70b92008-06-26 21:28:01 +02002171
Joerg Roedel0f2a86f2008-12-12 15:05:16 +01002172 INC_STATS_COUNTER(cnt_map_single);
2173
Joerg Roedel94f6d192009-11-24 16:40:02 +01002174 domain = get_domain(dev);
2175 if (PTR_ERR(domain) == -EINVAL)
Joerg Roedel4da70b92008-06-26 21:28:01 +02002176 return (dma_addr_t)paddr;
Joerg Roedel94f6d192009-11-24 16:40:02 +01002177 else if (IS_ERR(domain))
2178 return DMA_ERROR_CODE;
Joerg Roedel4da70b92008-06-26 21:28:01 +02002179
Joerg Roedelf99c0f12009-11-23 16:52:56 +01002180 dma_mask = *dev->dma_mask;
2181
Joerg Roedel4da70b92008-06-26 21:28:01 +02002182 spin_lock_irqsave(&domain->lock, flags);
Joerg Roedel94f6d192009-11-24 16:40:02 +01002183
Joerg Roedelcd8c82e2009-11-23 19:33:56 +01002184 addr = __map_single(dev, domain->priv, paddr, size, dir, false,
Joerg Roedel832a90c2008-09-18 15:54:23 +02002185 dma_mask);
FUJITA Tomonori8fd524b2009-11-15 21:19:53 +09002186 if (addr == DMA_ERROR_CODE)
Joerg Roedel4da70b92008-06-26 21:28:01 +02002187 goto out;
2188
Joerg Roedel17b124b2011-04-06 18:01:35 +02002189 domain_flush_complete(domain);
Joerg Roedel4da70b92008-06-26 21:28:01 +02002190
2191out:
2192 spin_unlock_irqrestore(&domain->lock, flags);
2193
2194 return addr;
2195}
2196
Joerg Roedel431b2a22008-07-11 17:14:22 +02002197/*
2198 * The exported unmap_single function for dma_ops.
2199 */
FUJITA Tomonori51491362009-01-05 23:47:25 +09002200static void unmap_page(struct device *dev, dma_addr_t dma_addr, size_t size,
2201 enum dma_data_direction dir, struct dma_attrs *attrs)
Joerg Roedel4da70b92008-06-26 21:28:01 +02002202{
2203 unsigned long flags;
Joerg Roedel4da70b92008-06-26 21:28:01 +02002204 struct protection_domain *domain;
Joerg Roedel4da70b92008-06-26 21:28:01 +02002205
Joerg Roedel146a6912008-12-12 15:07:12 +01002206 INC_STATS_COUNTER(cnt_unmap_single);
2207
Joerg Roedel94f6d192009-11-24 16:40:02 +01002208 domain = get_domain(dev);
2209 if (IS_ERR(domain))
Joerg Roedel5b28df62008-12-02 17:49:42 +01002210 return;
2211
Joerg Roedel4da70b92008-06-26 21:28:01 +02002212 spin_lock_irqsave(&domain->lock, flags);
2213
Joerg Roedelcd8c82e2009-11-23 19:33:56 +01002214 __unmap_single(domain->priv, dma_addr, size, dir);
Joerg Roedel4da70b92008-06-26 21:28:01 +02002215
Joerg Roedel17b124b2011-04-06 18:01:35 +02002216 domain_flush_complete(domain);
Joerg Roedel4da70b92008-06-26 21:28:01 +02002217
2218 spin_unlock_irqrestore(&domain->lock, flags);
2219}
2220
Joerg Roedel431b2a22008-07-11 17:14:22 +02002221/*
2222 * This is a special map_sg function which is used if we should map a
2223 * device which is not handled by an AMD IOMMU in the system.
2224 */
Joerg Roedel65b050a2008-06-26 21:28:02 +02002225static int map_sg_no_iommu(struct device *dev, struct scatterlist *sglist,
2226 int nelems, int dir)
2227{
2228 struct scatterlist *s;
2229 int i;
2230
2231 for_each_sg(sglist, s, nelems, i) {
2232 s->dma_address = (dma_addr_t)sg_phys(s);
2233 s->dma_length = s->length;
2234 }
2235
2236 return nelems;
2237}
2238
Joerg Roedel431b2a22008-07-11 17:14:22 +02002239/*
2240 * The exported map_sg function for dma_ops (handles scatter-gather
2241 * lists).
2242 */
Joerg Roedel65b050a2008-06-26 21:28:02 +02002243static int map_sg(struct device *dev, struct scatterlist *sglist,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +09002244 int nelems, enum dma_data_direction dir,
2245 struct dma_attrs *attrs)
Joerg Roedel65b050a2008-06-26 21:28:02 +02002246{
2247 unsigned long flags;
Joerg Roedel65b050a2008-06-26 21:28:02 +02002248 struct protection_domain *domain;
Joerg Roedel65b050a2008-06-26 21:28:02 +02002249 int i;
2250 struct scatterlist *s;
2251 phys_addr_t paddr;
2252 int mapped_elems = 0;
Joerg Roedel832a90c2008-09-18 15:54:23 +02002253 u64 dma_mask;
Joerg Roedel65b050a2008-06-26 21:28:02 +02002254
Joerg Roedeld03f067a2008-12-12 15:09:48 +01002255 INC_STATS_COUNTER(cnt_map_sg);
2256
Joerg Roedel94f6d192009-11-24 16:40:02 +01002257 domain = get_domain(dev);
2258 if (PTR_ERR(domain) == -EINVAL)
Joerg Roedelf99c0f12009-11-23 16:52:56 +01002259 return map_sg_no_iommu(dev, sglist, nelems, dir);
Joerg Roedel94f6d192009-11-24 16:40:02 +01002260 else if (IS_ERR(domain))
2261 return 0;
Joerg Roedeldbcc1122008-09-04 15:04:26 +02002262
Joerg Roedel832a90c2008-09-18 15:54:23 +02002263 dma_mask = *dev->dma_mask;
Joerg Roedel65b050a2008-06-26 21:28:02 +02002264
Joerg Roedel65b050a2008-06-26 21:28:02 +02002265 spin_lock_irqsave(&domain->lock, flags);
2266
2267 for_each_sg(sglist, s, nelems, i) {
2268 paddr = sg_phys(s);
2269
Joerg Roedelcd8c82e2009-11-23 19:33:56 +01002270 s->dma_address = __map_single(dev, domain->priv,
Joerg Roedel832a90c2008-09-18 15:54:23 +02002271 paddr, s->length, dir, false,
2272 dma_mask);
Joerg Roedel65b050a2008-06-26 21:28:02 +02002273
2274 if (s->dma_address) {
2275 s->dma_length = s->length;
2276 mapped_elems++;
2277 } else
2278 goto unmap;
Joerg Roedel65b050a2008-06-26 21:28:02 +02002279 }
2280
Joerg Roedel17b124b2011-04-06 18:01:35 +02002281 domain_flush_complete(domain);
Joerg Roedel65b050a2008-06-26 21:28:02 +02002282
2283out:
2284 spin_unlock_irqrestore(&domain->lock, flags);
2285
2286 return mapped_elems;
2287unmap:
2288 for_each_sg(sglist, s, mapped_elems, i) {
2289 if (s->dma_address)
Joerg Roedelcd8c82e2009-11-23 19:33:56 +01002290 __unmap_single(domain->priv, s->dma_address,
Joerg Roedel65b050a2008-06-26 21:28:02 +02002291 s->dma_length, dir);
2292 s->dma_address = s->dma_length = 0;
2293 }
2294
2295 mapped_elems = 0;
2296
2297 goto out;
2298}
2299
Joerg Roedel431b2a22008-07-11 17:14:22 +02002300/*
2301 * The exported map_sg function for dma_ops (handles scatter-gather
2302 * lists).
2303 */
Joerg Roedel65b050a2008-06-26 21:28:02 +02002304static void unmap_sg(struct device *dev, struct scatterlist *sglist,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +09002305 int nelems, enum dma_data_direction dir,
2306 struct dma_attrs *attrs)
Joerg Roedel65b050a2008-06-26 21:28:02 +02002307{
2308 unsigned long flags;
Joerg Roedel65b050a2008-06-26 21:28:02 +02002309 struct protection_domain *domain;
2310 struct scatterlist *s;
Joerg Roedel65b050a2008-06-26 21:28:02 +02002311 int i;
2312
Joerg Roedel55877a62008-12-12 15:12:14 +01002313 INC_STATS_COUNTER(cnt_unmap_sg);
2314
Joerg Roedel94f6d192009-11-24 16:40:02 +01002315 domain = get_domain(dev);
2316 if (IS_ERR(domain))
Joerg Roedel5b28df62008-12-02 17:49:42 +01002317 return;
2318
Joerg Roedel65b050a2008-06-26 21:28:02 +02002319 spin_lock_irqsave(&domain->lock, flags);
2320
2321 for_each_sg(sglist, s, nelems, i) {
Joerg Roedelcd8c82e2009-11-23 19:33:56 +01002322 __unmap_single(domain->priv, s->dma_address,
Joerg Roedel65b050a2008-06-26 21:28:02 +02002323 s->dma_length, dir);
Joerg Roedel65b050a2008-06-26 21:28:02 +02002324 s->dma_address = s->dma_length = 0;
2325 }
2326
Joerg Roedel17b124b2011-04-06 18:01:35 +02002327 domain_flush_complete(domain);
Joerg Roedel65b050a2008-06-26 21:28:02 +02002328
2329 spin_unlock_irqrestore(&domain->lock, flags);
2330}
2331
Joerg Roedel431b2a22008-07-11 17:14:22 +02002332/*
2333 * The exported alloc_coherent function for dma_ops.
2334 */
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002335static void *alloc_coherent(struct device *dev, size_t size,
2336 dma_addr_t *dma_addr, gfp_t flag)
2337{
2338 unsigned long flags;
2339 void *virt_addr;
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002340 struct protection_domain *domain;
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002341 phys_addr_t paddr;
Joerg Roedel832a90c2008-09-18 15:54:23 +02002342 u64 dma_mask = dev->coherent_dma_mask;
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002343
Joerg Roedelc8f0fb32008-12-12 15:14:21 +01002344 INC_STATS_COUNTER(cnt_alloc_coherent);
2345
Joerg Roedel94f6d192009-11-24 16:40:02 +01002346 domain = get_domain(dev);
2347 if (PTR_ERR(domain) == -EINVAL) {
Joerg Roedelf99c0f12009-11-23 16:52:56 +01002348 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2349 *dma_addr = __pa(virt_addr);
2350 return virt_addr;
Joerg Roedel94f6d192009-11-24 16:40:02 +01002351 } else if (IS_ERR(domain))
2352 return NULL;
Joerg Roedeldbcc1122008-09-04 15:04:26 +02002353
Joerg Roedelf99c0f12009-11-23 16:52:56 +01002354 dma_mask = dev->coherent_dma_mask;
2355 flag &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
2356 flag |= __GFP_ZERO;
FUJITA Tomonori13d9fea2008-09-10 20:19:40 +09002357
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002358 virt_addr = (void *)__get_free_pages(flag, get_order(size));
2359 if (!virt_addr)
Jaswinder Singh Rajputb25ae672009-07-01 19:53:14 +05302360 return NULL;
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002361
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002362 paddr = virt_to_phys(virt_addr);
2363
Joerg Roedel832a90c2008-09-18 15:54:23 +02002364 if (!dma_mask)
2365 dma_mask = *dev->dma_mask;
2366
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002367 spin_lock_irqsave(&domain->lock, flags);
2368
Joerg Roedelcd8c82e2009-11-23 19:33:56 +01002369 *dma_addr = __map_single(dev, domain->priv, paddr,
Joerg Roedel832a90c2008-09-18 15:54:23 +02002370 size, DMA_BIDIRECTIONAL, true, dma_mask);
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002371
FUJITA Tomonori8fd524b2009-11-15 21:19:53 +09002372 if (*dma_addr == DMA_ERROR_CODE) {
Jiri Slaby367d04c2009-05-28 09:54:48 +02002373 spin_unlock_irqrestore(&domain->lock, flags);
Joerg Roedel5b28df62008-12-02 17:49:42 +01002374 goto out_free;
Jiri Slaby367d04c2009-05-28 09:54:48 +02002375 }
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002376
Joerg Roedel17b124b2011-04-06 18:01:35 +02002377 domain_flush_complete(domain);
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002378
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002379 spin_unlock_irqrestore(&domain->lock, flags);
2380
2381 return virt_addr;
Joerg Roedel5b28df62008-12-02 17:49:42 +01002382
2383out_free:
2384
2385 free_pages((unsigned long)virt_addr, get_order(size));
2386
2387 return NULL;
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002388}
2389
Joerg Roedel431b2a22008-07-11 17:14:22 +02002390/*
2391 * The exported free_coherent function for dma_ops.
Joerg Roedel431b2a22008-07-11 17:14:22 +02002392 */
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002393static void free_coherent(struct device *dev, size_t size,
2394 void *virt_addr, dma_addr_t dma_addr)
2395{
2396 unsigned long flags;
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002397 struct protection_domain *domain;
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002398
Joerg Roedel5d31ee72008-12-12 15:16:38 +01002399 INC_STATS_COUNTER(cnt_free_coherent);
2400
Joerg Roedel94f6d192009-11-24 16:40:02 +01002401 domain = get_domain(dev);
2402 if (IS_ERR(domain))
Joerg Roedel5b28df62008-12-02 17:49:42 +01002403 goto free_mem;
2404
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002405 spin_lock_irqsave(&domain->lock, flags);
2406
Joerg Roedelcd8c82e2009-11-23 19:33:56 +01002407 __unmap_single(domain->priv, dma_addr, size, DMA_BIDIRECTIONAL);
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002408
Joerg Roedel17b124b2011-04-06 18:01:35 +02002409 domain_flush_complete(domain);
Joerg Roedel5d8b53c2008-06-26 21:28:03 +02002410
2411 spin_unlock_irqrestore(&domain->lock, flags);
2412
2413free_mem:
2414 free_pages((unsigned long)virt_addr, get_order(size));
2415}
2416
Joerg Roedelc432f3d2008-06-26 21:28:04 +02002417/*
Joerg Roedelb39ba6a2008-09-09 18:40:46 +02002418 * This function is called by the DMA layer to find out if we can handle a
2419 * particular device. It is part of the dma_ops.
2420 */
2421static int amd_iommu_dma_supported(struct device *dev, u64 mask)
2422{
Joerg Roedel420aef82009-11-23 16:14:57 +01002423 return check_device(dev);
Joerg Roedelb39ba6a2008-09-09 18:40:46 +02002424}
2425
2426/*
Joerg Roedel431b2a22008-07-11 17:14:22 +02002427 * The function for pre-allocating protection domains.
2428 *
Joerg Roedelc432f3d2008-06-26 21:28:04 +02002429 * If the driver core informs the DMA layer if a driver grabs a device
2430 * we don't need to preallocate the protection domains anymore.
2431 * For now we have to.
2432 */
Jaswinder Singh Rajput0e93dd82008-12-29 21:45:22 +05302433static void prealloc_protection_domains(void)
Joerg Roedelc432f3d2008-06-26 21:28:04 +02002434{
2435 struct pci_dev *dev = NULL;
2436 struct dma_ops_domain *dma_dom;
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002437 u16 devid;
Joerg Roedelc432f3d2008-06-26 21:28:04 +02002438
Chris Wrightd18c69d2010-04-02 18:27:55 -07002439 for_each_pci_dev(dev) {
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002440
2441 /* Do we handle this device? */
2442 if (!check_device(&dev->dev))
Joerg Roedelc432f3d2008-06-26 21:28:04 +02002443 continue;
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002444
2445 /* Is there already any domain for it? */
Joerg Roedel15898bb2009-11-24 15:39:42 +01002446 if (domain_for_device(&dev->dev))
Joerg Roedelc432f3d2008-06-26 21:28:04 +02002447 continue;
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002448
2449 devid = get_device_id(&dev->dev);
2450
Joerg Roedel87a64d52009-11-24 17:26:43 +01002451 dma_dom = dma_ops_domain_alloc();
Joerg Roedelc432f3d2008-06-26 21:28:04 +02002452 if (!dma_dom)
2453 continue;
2454 init_unity_mappings_for_device(dma_dom, devid);
Joerg Roedelbd60b732008-09-11 10:24:48 +02002455 dma_dom->target_dev = devid;
2456
Joerg Roedel15898bb2009-11-24 15:39:42 +01002457 attach_device(&dev->dev, &dma_dom->domain);
Joerg Roedelbe831292009-11-23 12:50:00 +01002458
Joerg Roedelbd60b732008-09-11 10:24:48 +02002459 list_add_tail(&dma_dom->list, &iommu_pd_list);
Joerg Roedelc432f3d2008-06-26 21:28:04 +02002460 }
2461}
2462
FUJITA Tomonori160c1d82009-01-05 23:59:02 +09002463static struct dma_map_ops amd_iommu_dma_ops = {
Joerg Roedel6631ee92008-06-26 21:28:05 +02002464 .alloc_coherent = alloc_coherent,
2465 .free_coherent = free_coherent,
FUJITA Tomonori51491362009-01-05 23:47:25 +09002466 .map_page = map_page,
2467 .unmap_page = unmap_page,
Joerg Roedel6631ee92008-06-26 21:28:05 +02002468 .map_sg = map_sg,
2469 .unmap_sg = unmap_sg,
Joerg Roedelb39ba6a2008-09-09 18:40:46 +02002470 .dma_supported = amd_iommu_dma_supported,
Joerg Roedel6631ee92008-06-26 21:28:05 +02002471};
2472
Joerg Roedel27c21272011-05-30 15:56:24 +02002473static unsigned device_dma_ops_init(void)
2474{
2475 struct pci_dev *pdev = NULL;
2476 unsigned unhandled = 0;
2477
2478 for_each_pci_dev(pdev) {
2479 if (!check_device(&pdev->dev)) {
2480 unhandled += 1;
2481 continue;
2482 }
2483
2484 pdev->dev.archdata.dma_ops = &amd_iommu_dma_ops;
2485 }
2486
2487 return unhandled;
2488}
2489
Joerg Roedel431b2a22008-07-11 17:14:22 +02002490/*
2491 * The function which clues the AMD IOMMU driver into dma_ops.
2492 */
Joerg Roedelf5325092010-01-22 17:44:35 +01002493
2494void __init amd_iommu_init_api(void)
2495{
2496 register_iommu(&amd_iommu_ops);
2497}
2498
Joerg Roedel6631ee92008-06-26 21:28:05 +02002499int __init amd_iommu_init_dma_ops(void)
2500{
2501 struct amd_iommu *iommu;
Joerg Roedel27c21272011-05-30 15:56:24 +02002502 int ret, unhandled;
Joerg Roedel6631ee92008-06-26 21:28:05 +02002503
Joerg Roedel431b2a22008-07-11 17:14:22 +02002504 /*
2505 * first allocate a default protection domain for every IOMMU we
2506 * found in the system. Devices not assigned to any other
2507 * protection domain will be assigned to the default one.
2508 */
Joerg Roedel3bd22172009-05-04 15:06:20 +02002509 for_each_iommu(iommu) {
Joerg Roedel87a64d52009-11-24 17:26:43 +01002510 iommu->default_dom = dma_ops_domain_alloc();
Joerg Roedel6631ee92008-06-26 21:28:05 +02002511 if (iommu->default_dom == NULL)
2512 return -ENOMEM;
Joerg Roedele2dc14a2008-12-10 18:48:59 +01002513 iommu->default_dom->domain.flags |= PD_DEFAULT_MASK;
Joerg Roedel6631ee92008-06-26 21:28:05 +02002514 ret = iommu_init_unity_mappings(iommu);
2515 if (ret)
2516 goto free_domains;
2517 }
2518
Joerg Roedel431b2a22008-07-11 17:14:22 +02002519 /*
Joerg Roedel8793abe2009-11-27 11:40:33 +01002520 * Pre-allocate the protection domains for each device.
Joerg Roedel431b2a22008-07-11 17:14:22 +02002521 */
Joerg Roedel8793abe2009-11-27 11:40:33 +01002522 prealloc_protection_domains();
Joerg Roedel6631ee92008-06-26 21:28:05 +02002523
2524 iommu_detected = 1;
FUJITA Tomonori75f1cdf2009-11-10 19:46:20 +09002525 swiotlb = 0;
Joerg Roedel6631ee92008-06-26 21:28:05 +02002526
Joerg Roedel431b2a22008-07-11 17:14:22 +02002527 /* Make the driver finally visible to the drivers */
Joerg Roedel27c21272011-05-30 15:56:24 +02002528 unhandled = device_dma_ops_init();
2529 if (unhandled && max_pfn > MAX_DMA32_PFN) {
2530 /* There are unhandled devices - initialize swiotlb for them */
2531 swiotlb = 1;
2532 }
Joerg Roedel6631ee92008-06-26 21:28:05 +02002533
Joerg Roedel7f265082008-12-12 13:50:21 +01002534 amd_iommu_stats_init();
2535
Joerg Roedel6631ee92008-06-26 21:28:05 +02002536 return 0;
2537
2538free_domains:
2539
Joerg Roedel3bd22172009-05-04 15:06:20 +02002540 for_each_iommu(iommu) {
Joerg Roedel6631ee92008-06-26 21:28:05 +02002541 if (iommu->default_dom)
2542 dma_ops_domain_free(iommu->default_dom);
2543 }
2544
2545 return ret;
2546}
Joerg Roedel6d98cd82008-12-08 12:05:55 +01002547
2548/*****************************************************************************
2549 *
2550 * The following functions belong to the exported interface of AMD IOMMU
2551 *
2552 * This interface allows access to lower level functions of the IOMMU
2553 * like protection domain handling and assignement of devices to domains
2554 * which is not possible with the dma_ops interface.
2555 *
2556 *****************************************************************************/
2557
Joerg Roedel6d98cd82008-12-08 12:05:55 +01002558static void cleanup_domain(struct protection_domain *domain)
2559{
Joerg Roedel492667d2009-11-27 13:25:47 +01002560 struct iommu_dev_data *dev_data, *next;
Joerg Roedel6d98cd82008-12-08 12:05:55 +01002561 unsigned long flags;
Joerg Roedel6d98cd82008-12-08 12:05:55 +01002562
2563 write_lock_irqsave(&amd_iommu_devtable_lock, flags);
2564
Joerg Roedel492667d2009-11-27 13:25:47 +01002565 list_for_each_entry_safe(dev_data, next, &domain->dev_list, list) {
Joerg Roedelec9e79e2011-06-09 17:25:50 +02002566 __detach_device(dev_data);
Joerg Roedel492667d2009-11-27 13:25:47 +01002567 atomic_set(&dev_data->bind, 0);
2568 }
Joerg Roedel6d98cd82008-12-08 12:05:55 +01002569
2570 write_unlock_irqrestore(&amd_iommu_devtable_lock, flags);
2571}
2572
Joerg Roedel26508152009-08-26 16:52:40 +02002573static void protection_domain_free(struct protection_domain *domain)
2574{
2575 if (!domain)
2576 return;
2577
Joerg Roedelaeb26f52009-11-20 16:44:01 +01002578 del_domain_from_list(domain);
2579
Joerg Roedel26508152009-08-26 16:52:40 +02002580 if (domain->id)
2581 domain_id_free(domain->id);
2582
2583 kfree(domain);
2584}
2585
2586static struct protection_domain *protection_domain_alloc(void)
Joerg Roedelc156e342008-12-02 18:13:27 +01002587{
2588 struct protection_domain *domain;
2589
2590 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2591 if (!domain)
Joerg Roedel26508152009-08-26 16:52:40 +02002592 return NULL;
Joerg Roedelc156e342008-12-02 18:13:27 +01002593
2594 spin_lock_init(&domain->lock);
Joerg Roedel5d214fe2010-02-08 14:44:49 +01002595 mutex_init(&domain->api_lock);
Joerg Roedelc156e342008-12-02 18:13:27 +01002596 domain->id = domain_id_alloc();
2597 if (!domain->id)
Joerg Roedel26508152009-08-26 16:52:40 +02002598 goto out_err;
Joerg Roedel7c392cb2009-11-26 11:13:32 +01002599 INIT_LIST_HEAD(&domain->dev_list);
Joerg Roedel26508152009-08-26 16:52:40 +02002600
Joerg Roedelaeb26f52009-11-20 16:44:01 +01002601 add_domain_to_list(domain);
2602
Joerg Roedel26508152009-08-26 16:52:40 +02002603 return domain;
2604
2605out_err:
2606 kfree(domain);
2607
2608 return NULL;
2609}
2610
2611static int amd_iommu_domain_init(struct iommu_domain *dom)
2612{
2613 struct protection_domain *domain;
2614
2615 domain = protection_domain_alloc();
2616 if (!domain)
Joerg Roedelc156e342008-12-02 18:13:27 +01002617 goto out_free;
Joerg Roedel26508152009-08-26 16:52:40 +02002618
2619 domain->mode = PAGE_MODE_3_LEVEL;
Joerg Roedelc156e342008-12-02 18:13:27 +01002620 domain->pt_root = (void *)get_zeroed_page(GFP_KERNEL);
2621 if (!domain->pt_root)
2622 goto out_free;
2623
2624 dom->priv = domain;
2625
2626 return 0;
2627
2628out_free:
Joerg Roedel26508152009-08-26 16:52:40 +02002629 protection_domain_free(domain);
Joerg Roedelc156e342008-12-02 18:13:27 +01002630
2631 return -ENOMEM;
2632}
2633
Joerg Roedel98383fc2008-12-02 18:34:12 +01002634static void amd_iommu_domain_destroy(struct iommu_domain *dom)
2635{
2636 struct protection_domain *domain = dom->priv;
2637
2638 if (!domain)
2639 return;
2640
2641 if (domain->dev_cnt > 0)
2642 cleanup_domain(domain);
2643
2644 BUG_ON(domain->dev_cnt != 0);
2645
2646 free_pagetable(domain);
2647
Joerg Roedel8b408fe2010-03-08 14:20:07 +01002648 protection_domain_free(domain);
Joerg Roedel98383fc2008-12-02 18:34:12 +01002649
2650 dom->priv = NULL;
2651}
2652
Joerg Roedel684f2882008-12-08 12:07:44 +01002653static void amd_iommu_detach_device(struct iommu_domain *dom,
2654 struct device *dev)
2655{
Joerg Roedel657cbb62009-11-23 15:26:46 +01002656 struct iommu_dev_data *dev_data = dev->archdata.iommu;
Joerg Roedel684f2882008-12-08 12:07:44 +01002657 struct amd_iommu *iommu;
Joerg Roedel684f2882008-12-08 12:07:44 +01002658 u16 devid;
2659
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002660 if (!check_device(dev))
Joerg Roedel684f2882008-12-08 12:07:44 +01002661 return;
2662
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002663 devid = get_device_id(dev);
Joerg Roedel684f2882008-12-08 12:07:44 +01002664
Joerg Roedel657cbb62009-11-23 15:26:46 +01002665 if (dev_data->domain != NULL)
Joerg Roedel15898bb2009-11-24 15:39:42 +01002666 detach_device(dev);
Joerg Roedel684f2882008-12-08 12:07:44 +01002667
2668 iommu = amd_iommu_rlookup_table[devid];
2669 if (!iommu)
2670 return;
2671
Joerg Roedel684f2882008-12-08 12:07:44 +01002672 iommu_completion_wait(iommu);
2673}
2674
Joerg Roedel01106062008-12-02 19:34:11 +01002675static int amd_iommu_attach_device(struct iommu_domain *dom,
2676 struct device *dev)
2677{
2678 struct protection_domain *domain = dom->priv;
Joerg Roedel657cbb62009-11-23 15:26:46 +01002679 struct iommu_dev_data *dev_data;
Joerg Roedel01106062008-12-02 19:34:11 +01002680 struct amd_iommu *iommu;
Joerg Roedel15898bb2009-11-24 15:39:42 +01002681 int ret;
Joerg Roedel01106062008-12-02 19:34:11 +01002682
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002683 if (!check_device(dev))
Joerg Roedel01106062008-12-02 19:34:11 +01002684 return -EINVAL;
2685
Joerg Roedel657cbb62009-11-23 15:26:46 +01002686 dev_data = dev->archdata.iommu;
2687
Joerg Roedelf62dda62011-06-09 12:55:35 +02002688 iommu = amd_iommu_rlookup_table[dev_data->devid];
Joerg Roedel01106062008-12-02 19:34:11 +01002689 if (!iommu)
2690 return -EINVAL;
2691
Joerg Roedel657cbb62009-11-23 15:26:46 +01002692 if (dev_data->domain)
Joerg Roedel15898bb2009-11-24 15:39:42 +01002693 detach_device(dev);
Joerg Roedel01106062008-12-02 19:34:11 +01002694
Joerg Roedel15898bb2009-11-24 15:39:42 +01002695 ret = attach_device(dev, domain);
Joerg Roedel01106062008-12-02 19:34:11 +01002696
2697 iommu_completion_wait(iommu);
2698
Joerg Roedel15898bb2009-11-24 15:39:42 +01002699 return ret;
Joerg Roedel01106062008-12-02 19:34:11 +01002700}
2701
Joerg Roedel468e2362010-01-21 16:37:36 +01002702static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
2703 phys_addr_t paddr, int gfp_order, int iommu_prot)
Joerg Roedelc6229ca2008-12-02 19:48:43 +01002704{
Joerg Roedel468e2362010-01-21 16:37:36 +01002705 unsigned long page_size = 0x1000UL << gfp_order;
Joerg Roedelc6229ca2008-12-02 19:48:43 +01002706 struct protection_domain *domain = dom->priv;
Joerg Roedelc6229ca2008-12-02 19:48:43 +01002707 int prot = 0;
2708 int ret;
2709
2710 if (iommu_prot & IOMMU_READ)
2711 prot |= IOMMU_PROT_IR;
2712 if (iommu_prot & IOMMU_WRITE)
2713 prot |= IOMMU_PROT_IW;
2714
Joerg Roedel5d214fe2010-02-08 14:44:49 +01002715 mutex_lock(&domain->api_lock);
Joerg Roedel795e74f2010-05-11 17:40:57 +02002716 ret = iommu_map_page(domain, iova, paddr, prot, page_size);
Joerg Roedel5d214fe2010-02-08 14:44:49 +01002717 mutex_unlock(&domain->api_lock);
2718
Joerg Roedel795e74f2010-05-11 17:40:57 +02002719 return ret;
Joerg Roedelc6229ca2008-12-02 19:48:43 +01002720}
2721
Joerg Roedel468e2362010-01-21 16:37:36 +01002722static int amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
2723 int gfp_order)
Joerg Roedeleb74ff62008-12-02 19:59:10 +01002724{
Joerg Roedeleb74ff62008-12-02 19:59:10 +01002725 struct protection_domain *domain = dom->priv;
Joerg Roedel468e2362010-01-21 16:37:36 +01002726 unsigned long page_size, unmap_size;
Joerg Roedeleb74ff62008-12-02 19:59:10 +01002727
Joerg Roedel468e2362010-01-21 16:37:36 +01002728 page_size = 0x1000UL << gfp_order;
Joerg Roedeleb74ff62008-12-02 19:59:10 +01002729
Joerg Roedel5d214fe2010-02-08 14:44:49 +01002730 mutex_lock(&domain->api_lock);
Joerg Roedel468e2362010-01-21 16:37:36 +01002731 unmap_size = iommu_unmap_page(domain, iova, page_size);
Joerg Roedel795e74f2010-05-11 17:40:57 +02002732 mutex_unlock(&domain->api_lock);
Joerg Roedeleb74ff62008-12-02 19:59:10 +01002733
Joerg Roedel17b124b2011-04-06 18:01:35 +02002734 domain_flush_tlb_pde(domain);
Joerg Roedel5d214fe2010-02-08 14:44:49 +01002735
Joerg Roedel468e2362010-01-21 16:37:36 +01002736 return get_order(unmap_size);
Joerg Roedeleb74ff62008-12-02 19:59:10 +01002737}
2738
Joerg Roedel645c4c82008-12-02 20:05:50 +01002739static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom,
2740 unsigned long iova)
2741{
2742 struct protection_domain *domain = dom->priv;
Joerg Roedelf03152b2010-01-21 16:15:24 +01002743 unsigned long offset_mask;
Joerg Roedel645c4c82008-12-02 20:05:50 +01002744 phys_addr_t paddr;
Joerg Roedelf03152b2010-01-21 16:15:24 +01002745 u64 *pte, __pte;
Joerg Roedel645c4c82008-12-02 20:05:50 +01002746
Joerg Roedel24cd7722010-01-19 17:27:39 +01002747 pte = fetch_pte(domain, iova);
Joerg Roedel645c4c82008-12-02 20:05:50 +01002748
Joerg Roedela6d41a42009-09-02 17:08:55 +02002749 if (!pte || !IOMMU_PTE_PRESENT(*pte))
Joerg Roedel645c4c82008-12-02 20:05:50 +01002750 return 0;
2751
Joerg Roedelf03152b2010-01-21 16:15:24 +01002752 if (PM_PTE_LEVEL(*pte) == 0)
2753 offset_mask = PAGE_SIZE - 1;
2754 else
2755 offset_mask = PTE_PAGE_SIZE(*pte) - 1;
2756
2757 __pte = *pte & PM_ADDR_MASK;
2758 paddr = (__pte & ~offset_mask) | (iova & offset_mask);
Joerg Roedel645c4c82008-12-02 20:05:50 +01002759
2760 return paddr;
2761}
2762
Sheng Yangdbb9fd82009-03-18 15:33:06 +08002763static int amd_iommu_domain_has_cap(struct iommu_domain *domain,
2764 unsigned long cap)
2765{
Joerg Roedel80a506b2010-07-27 17:14:24 +02002766 switch (cap) {
2767 case IOMMU_CAP_CACHE_COHERENCY:
2768 return 1;
2769 }
2770
Sheng Yangdbb9fd82009-03-18 15:33:06 +08002771 return 0;
2772}
2773
Joerg Roedel26961ef2008-12-03 17:00:17 +01002774static struct iommu_ops amd_iommu_ops = {
2775 .domain_init = amd_iommu_domain_init,
2776 .domain_destroy = amd_iommu_domain_destroy,
2777 .attach_dev = amd_iommu_attach_device,
2778 .detach_dev = amd_iommu_detach_device,
Joerg Roedel468e2362010-01-21 16:37:36 +01002779 .map = amd_iommu_map,
2780 .unmap = amd_iommu_unmap,
Joerg Roedel26961ef2008-12-03 17:00:17 +01002781 .iova_to_phys = amd_iommu_iova_to_phys,
Sheng Yangdbb9fd82009-03-18 15:33:06 +08002782 .domain_has_cap = amd_iommu_domain_has_cap,
Joerg Roedel26961ef2008-12-03 17:00:17 +01002783};
2784
Joerg Roedel0feae532009-08-26 15:26:30 +02002785/*****************************************************************************
2786 *
2787 * The next functions do a basic initialization of IOMMU for pass through
2788 * mode
2789 *
2790 * In passthrough mode the IOMMU is initialized and enabled but not used for
2791 * DMA-API translation.
2792 *
2793 *****************************************************************************/
2794
2795int __init amd_iommu_init_passthrough(void)
2796{
Joerg Roedel15898bb2009-11-24 15:39:42 +01002797 struct amd_iommu *iommu;
Joerg Roedel0feae532009-08-26 15:26:30 +02002798 struct pci_dev *dev = NULL;
Joerg Roedel15898bb2009-11-24 15:39:42 +01002799 u16 devid;
Joerg Roedel0feae532009-08-26 15:26:30 +02002800
André Goddard Rosaaf901ca2009-11-14 13:09:05 -02002801 /* allocate passthrough domain */
Joerg Roedel0feae532009-08-26 15:26:30 +02002802 pt_domain = protection_domain_alloc();
2803 if (!pt_domain)
2804 return -ENOMEM;
2805
2806 pt_domain->mode |= PAGE_MODE_NONE;
2807
Kulikov Vasiliy6c54aab2010-07-03 12:03:51 -04002808 for_each_pci_dev(dev) {
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002809 if (!check_device(&dev->dev))
Joerg Roedel0feae532009-08-26 15:26:30 +02002810 continue;
2811
Joerg Roedel98fc5a62009-11-24 17:19:23 +01002812 devid = get_device_id(&dev->dev);
2813
Joerg Roedel15898bb2009-11-24 15:39:42 +01002814 iommu = amd_iommu_rlookup_table[devid];
Joerg Roedel0feae532009-08-26 15:26:30 +02002815 if (!iommu)
2816 continue;
2817
Joerg Roedel15898bb2009-11-24 15:39:42 +01002818 attach_device(&dev->dev, pt_domain);
Joerg Roedel0feae532009-08-26 15:26:30 +02002819 }
2820
2821 pr_info("AMD-Vi: Initialized for Passthrough Mode\n");
2822
2823 return 0;
2824}