blob: 7caf2fa237f22bd7a3b0932c6387ed52e0b26424 [file] [log] [blame]
Joerg Roedele3c495c2011-11-09 12:31:15 +01001/*
2 * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
Joerg Roedel63ce3ae2015-02-04 16:12:55 +01003 * Author: Joerg Roedel <jroedel@suse.de>
Joerg Roedele3c495c2011-11-09 12:31:15 +01004 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published
7 * by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17 */
18
Joerg Roedel8736b2c2011-11-24 16:21:52 +010019#include <linux/mmu_notifier.h>
Joerg Roedeled96f222011-11-23 17:30:39 +010020#include <linux/amd-iommu.h>
21#include <linux/mm_types.h>
Joerg Roedel8736b2c2011-11-24 16:21:52 +010022#include <linux/profile.h>
Joerg Roedele3c495c2011-11-09 12:31:15 +010023#include <linux/module.h>
Joerg Roedel2d5503b2011-11-24 10:41:57 +010024#include <linux/sched.h>
Joerg Roedeled96f222011-11-23 17:30:39 +010025#include <linux/iommu.h>
Joerg Roedel028eeac2011-11-24 12:48:13 +010026#include <linux/wait.h>
Joerg Roedeled96f222011-11-23 17:30:39 +010027#include <linux/pci.h>
28#include <linux/gfp.h>
29
Joerg Roedel028eeac2011-11-24 12:48:13 +010030#include "amd_iommu_types.h"
Joerg Roedeled96f222011-11-23 17:30:39 +010031#include "amd_iommu_proto.h"
Joerg Roedele3c495c2011-11-09 12:31:15 +010032
33MODULE_LICENSE("GPL v2");
Joerg Roedel63ce3ae2015-02-04 16:12:55 +010034MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
Joerg Roedele3c495c2011-11-09 12:31:15 +010035
Joerg Roedeled96f222011-11-23 17:30:39 +010036#define MAX_DEVICES 0x10000
37#define PRI_QUEUE_SIZE 512
38
39struct pri_queue {
40 atomic_t inflight;
41 bool finish;
Joerg Roedel028eeac2011-11-24 12:48:13 +010042 int status;
Joerg Roedeled96f222011-11-23 17:30:39 +010043};
44
45struct pasid_state {
46 struct list_head list; /* For global state-list */
47 atomic_t count; /* Reference count */
Joerg Roedeld73a6d72014-06-20 16:14:22 +020048 unsigned mmu_notifier_count; /* Counting nested mmu_notifier
Joerg Roedele79df312014-05-20 23:18:26 +020049 calls */
Joerg Roedeled96f222011-11-23 17:30:39 +010050 struct mm_struct *mm; /* mm_struct for the faults */
Joerg Roedelff6d0cc2014-07-08 12:49:50 +020051 struct mmu_notifier mn; /* mmu_notifier handle */
Joerg Roedeled96f222011-11-23 17:30:39 +010052 struct pri_queue pri[PRI_QUEUE_SIZE]; /* PRI tag states */
53 struct device_state *device_state; /* Link to our device_state */
54 int pasid; /* PASID index */
Joerg Roedeld9e16112014-07-09 15:43:11 +020055 bool invalid; /* Used during setup and
56 teardown of the pasid */
Joerg Roedeld73a6d72014-06-20 16:14:22 +020057 spinlock_t lock; /* Protect pri_queues and
58 mmu_notifer_count */
Joerg Roedel028eeac2011-11-24 12:48:13 +010059 wait_queue_head_t wq; /* To wait for count == 0 */
Joerg Roedeled96f222011-11-23 17:30:39 +010060};
61
62struct device_state {
Joerg Roedel741669c2014-05-20 23:18:23 +020063 struct list_head list;
64 u16 devid;
Joerg Roedeled96f222011-11-23 17:30:39 +010065 atomic_t count;
66 struct pci_dev *pdev;
67 struct pasid_state **states;
68 struct iommu_domain *domain;
69 int pasid_levels;
70 int max_pasids;
Joerg Roedel175d6142011-11-28 14:36:36 +010071 amd_iommu_invalid_ppr_cb inv_ppr_cb;
Joerg Roedelbc216622011-12-07 12:24:42 +010072 amd_iommu_invalidate_ctx inv_ctx_cb;
Joerg Roedeled96f222011-11-23 17:30:39 +010073 spinlock_t lock;
Joerg Roedel028eeac2011-11-24 12:48:13 +010074 wait_queue_head_t wq;
75};
76
77struct fault {
78 struct work_struct work;
79 struct device_state *dev_state;
80 struct pasid_state *state;
81 struct mm_struct *mm;
82 u64 address;
83 u16 devid;
84 u16 pasid;
85 u16 tag;
86 u16 finish;
87 u16 flags;
Joerg Roedeled96f222011-11-23 17:30:39 +010088};
89
Joerg Roedel741669c2014-05-20 23:18:23 +020090static LIST_HEAD(state_list);
Joerg Roedeled96f222011-11-23 17:30:39 +010091static spinlock_t state_lock;
92
Joerg Roedel028eeac2011-11-24 12:48:13 +010093static struct workqueue_struct *iommu_wq;
94
Joerg Roedel2d5503b2011-11-24 10:41:57 +010095static void free_pasid_states(struct device_state *dev_state);
Joerg Roedeled96f222011-11-23 17:30:39 +010096
97static u16 device_id(struct pci_dev *pdev)
98{
99 u16 devid;
100
101 devid = pdev->bus->number;
102 devid = (devid << 8) | pdev->devfn;
103
104 return devid;
105}
106
Joerg Roedelb87d2d72014-05-20 23:18:22 +0200107static struct device_state *__get_device_state(u16 devid)
108{
Joerg Roedel741669c2014-05-20 23:18:23 +0200109 struct device_state *dev_state;
110
111 list_for_each_entry(dev_state, &state_list, list) {
112 if (dev_state->devid == devid)
113 return dev_state;
114 }
115
116 return NULL;
Joerg Roedelb87d2d72014-05-20 23:18:22 +0200117}
118
Joerg Roedeled96f222011-11-23 17:30:39 +0100119static struct device_state *get_device_state(u16 devid)
120{
121 struct device_state *dev_state;
122 unsigned long flags;
123
124 spin_lock_irqsave(&state_lock, flags);
Joerg Roedelb87d2d72014-05-20 23:18:22 +0200125 dev_state = __get_device_state(devid);
Joerg Roedeled96f222011-11-23 17:30:39 +0100126 if (dev_state != NULL)
127 atomic_inc(&dev_state->count);
128 spin_unlock_irqrestore(&state_lock, flags);
129
130 return dev_state;
131}
132
133static void free_device_state(struct device_state *dev_state)
134{
Joerg Roedel55c99a42015-07-28 16:58:47 +0200135 struct iommu_group *group;
136
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100137 /*
138 * First detach device from domain - No more PRI requests will arrive
139 * from that device after it is unbound from the IOMMUv2 domain.
140 */
Joerg Roedel55c99a42015-07-28 16:58:47 +0200141 group = iommu_group_get(&dev_state->pdev->dev);
142 if (WARN_ON(!group))
143 return;
144
145 iommu_detach_group(dev_state->domain, group);
146
147 iommu_group_put(group);
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100148
149 /* Everything is down now, free the IOMMUv2 domain */
Joerg Roedeled96f222011-11-23 17:30:39 +0100150 iommu_domain_free(dev_state->domain);
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100151
152 /* Finally get rid of the device-state */
Joerg Roedeled96f222011-11-23 17:30:39 +0100153 kfree(dev_state);
154}
155
156static void put_device_state(struct device_state *dev_state)
157{
158 if (atomic_dec_and_test(&dev_state->count))
Joerg Roedel028eeac2011-11-24 12:48:13 +0100159 wake_up(&dev_state->wq);
Joerg Roedeled96f222011-11-23 17:30:39 +0100160}
161
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100162/* Must be called under dev_state->lock */
163static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
164 int pasid, bool alloc)
165{
166 struct pasid_state **root, **ptr;
167 int level, index;
168
169 level = dev_state->pasid_levels;
170 root = dev_state->states;
171
172 while (true) {
173
174 index = (pasid >> (9 * level)) & 0x1ff;
175 ptr = &root[index];
176
177 if (level == 0)
178 break;
179
180 if (*ptr == NULL) {
181 if (!alloc)
182 return NULL;
183
184 *ptr = (void *)get_zeroed_page(GFP_ATOMIC);
185 if (*ptr == NULL)
186 return NULL;
187 }
188
189 root = (struct pasid_state **)*ptr;
190 level -= 1;
191 }
192
193 return ptr;
194}
195
196static int set_pasid_state(struct device_state *dev_state,
197 struct pasid_state *pasid_state,
198 int pasid)
199{
200 struct pasid_state **ptr;
201 unsigned long flags;
202 int ret;
203
204 spin_lock_irqsave(&dev_state->lock, flags);
205 ptr = __get_pasid_state_ptr(dev_state, pasid, true);
206
207 ret = -ENOMEM;
208 if (ptr == NULL)
209 goto out_unlock;
210
211 ret = -ENOMEM;
212 if (*ptr != NULL)
213 goto out_unlock;
214
215 *ptr = pasid_state;
216
217 ret = 0;
218
219out_unlock:
220 spin_unlock_irqrestore(&dev_state->lock, flags);
221
222 return ret;
223}
224
225static void clear_pasid_state(struct device_state *dev_state, int pasid)
226{
227 struct pasid_state **ptr;
228 unsigned long flags;
229
230 spin_lock_irqsave(&dev_state->lock, flags);
231 ptr = __get_pasid_state_ptr(dev_state, pasid, true);
232
233 if (ptr == NULL)
234 goto out_unlock;
235
236 *ptr = NULL;
237
238out_unlock:
239 spin_unlock_irqrestore(&dev_state->lock, flags);
240}
241
242static struct pasid_state *get_pasid_state(struct device_state *dev_state,
243 int pasid)
244{
245 struct pasid_state **ptr, *ret = NULL;
246 unsigned long flags;
247
248 spin_lock_irqsave(&dev_state->lock, flags);
249 ptr = __get_pasid_state_ptr(dev_state, pasid, false);
250
251 if (ptr == NULL)
252 goto out_unlock;
253
254 ret = *ptr;
255 if (ret)
256 atomic_inc(&ret->count);
257
258out_unlock:
259 spin_unlock_irqrestore(&dev_state->lock, flags);
260
261 return ret;
262}
263
264static void free_pasid_state(struct pasid_state *pasid_state)
265{
266 kfree(pasid_state);
267}
268
269static void put_pasid_state(struct pasid_state *pasid_state)
270{
Oded Gabbay1c510992014-11-10 12:21:39 +0200271 if (atomic_dec_and_test(&pasid_state->count))
Joerg Roedel028eeac2011-11-24 12:48:13 +0100272 wake_up(&pasid_state->wq);
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100273}
274
Joerg Roedel028eeac2011-11-24 12:48:13 +0100275static void put_pasid_state_wait(struct pasid_state *pasid_state)
276{
Oded Gabbay1bf1b432015-04-16 17:08:44 +0300277 atomic_dec(&pasid_state->count);
Joerg Roedela1bec062015-02-04 15:50:38 +0100278 wait_event(pasid_state->wq, !atomic_read(&pasid_state->count));
Joerg Roedel028eeac2011-11-24 12:48:13 +0100279 free_pasid_state(pasid_state);
280}
281
Joerg Roedel61feb432014-07-08 14:19:35 +0200282static void unbind_pasid(struct pasid_state *pasid_state)
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100283{
284 struct iommu_domain *domain;
285
286 domain = pasid_state->device_state->domain;
287
Joerg Roedel53d340e2014-07-08 15:01:43 +0200288 /*
289 * Mark pasid_state as invalid, no more faults will we added to the
290 * work queue after this is visible everywhere.
291 */
292 pasid_state->invalid = true;
293
294 /* Make sure this is visible */
295 smp_wmb();
296
297 /* After this the device/pasid can't access the mm anymore */
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100298 amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100299
300 /* Make sure no more pending faults are in the queue */
301 flush_workqueue(iommu_wq);
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100302}
303
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100304static void free_pasid_states_level1(struct pasid_state **tbl)
305{
306 int i;
307
308 for (i = 0; i < 512; ++i) {
309 if (tbl[i] == NULL)
310 continue;
311
312 free_page((unsigned long)tbl[i]);
313 }
314}
315
316static void free_pasid_states_level2(struct pasid_state **tbl)
317{
318 struct pasid_state **ptr;
319 int i;
320
321 for (i = 0; i < 512; ++i) {
322 if (tbl[i] == NULL)
323 continue;
324
325 ptr = (struct pasid_state **)tbl[i];
326 free_pasid_states_level1(ptr);
327 }
328}
329
330static void free_pasid_states(struct device_state *dev_state)
331{
332 struct pasid_state *pasid_state;
333 int i;
334
335 for (i = 0; i < dev_state->max_pasids; ++i) {
336 pasid_state = get_pasid_state(dev_state, i);
337 if (pasid_state == NULL)
338 continue;
339
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100340 put_pasid_state(pasid_state);
Joerg Roedela40d4c62014-05-20 23:18:24 +0200341
342 /*
343 * This will call the mn_release function and
344 * unbind the PASID
345 */
346 mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
Joerg Roedelc5db16a2014-07-08 14:15:45 +0200347
348 put_pasid_state_wait(pasid_state); /* Reference taken in
Joerg Roedeldaff2f92014-07-30 16:04:40 +0200349 amd_iommu_bind_pasid */
Joerg Roedel75058a32014-07-30 16:04:39 +0200350
351 /* Drop reference taken in amd_iommu_bind_pasid */
352 put_device_state(dev_state);
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100353 }
354
355 if (dev_state->pasid_levels == 2)
356 free_pasid_states_level2(dev_state->states);
357 else if (dev_state->pasid_levels == 1)
358 free_pasid_states_level1(dev_state->states);
Joerg Roedel23d3a982015-08-13 11:15:13 +0200359 else
360 BUG_ON(dev_state->pasid_levels != 0);
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100361
362 free_page((unsigned long)dev_state->states);
363}
364
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100365static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
366{
367 return container_of(mn, struct pasid_state, mn);
368}
369
370static void __mn_flush_page(struct mmu_notifier *mn,
371 unsigned long address)
372{
373 struct pasid_state *pasid_state;
374 struct device_state *dev_state;
375
376 pasid_state = mn_to_state(mn);
377 dev_state = pasid_state->device_state;
378
379 amd_iommu_flush_page(dev_state->domain, pasid_state->pasid, address);
380}
381
382static int mn_clear_flush_young(struct mmu_notifier *mn,
383 struct mm_struct *mm,
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700384 unsigned long start,
385 unsigned long end)
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100386{
Andres Lagar-Cavilla57128462014-09-22 14:54:42 -0700387 for (; start < end; start += PAGE_SIZE)
388 __mn_flush_page(mn, start);
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100389
390 return 0;
391}
392
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100393static void mn_invalidate_page(struct mmu_notifier *mn,
394 struct mm_struct *mm,
395 unsigned long address)
396{
397 __mn_flush_page(mn, address);
398}
399
Joerg Roedele7cc3dd2014-11-13 13:46:09 +1100400static void mn_invalidate_range(struct mmu_notifier *mn,
401 struct mm_struct *mm,
402 unsigned long start, unsigned long end)
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100403{
404 struct pasid_state *pasid_state;
405 struct device_state *dev_state;
406
407 pasid_state = mn_to_state(mn);
408 dev_state = pasid_state->device_state;
409
Joerg Roedele7cc3dd2014-11-13 13:46:09 +1100410 if ((start ^ (end - 1)) < PAGE_SIZE)
411 amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
412 start);
413 else
414 amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100415}
416
Joerg Roedela40d4c62014-05-20 23:18:24 +0200417static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
418{
419 struct pasid_state *pasid_state;
420 struct device_state *dev_state;
Joerg Roedeld9e16112014-07-09 15:43:11 +0200421 bool run_inv_ctx_cb;
Joerg Roedela40d4c62014-05-20 23:18:24 +0200422
423 might_sleep();
424
Joerg Roedeld9e16112014-07-09 15:43:11 +0200425 pasid_state = mn_to_state(mn);
426 dev_state = pasid_state->device_state;
427 run_inv_ctx_cb = !pasid_state->invalid;
Joerg Roedela40d4c62014-05-20 23:18:24 +0200428
Dan Carpenter940f7002015-02-20 13:52:01 +0300429 if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
Joerg Roedela40d4c62014-05-20 23:18:24 +0200430 dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
431
Joerg Roedel61feb432014-07-08 14:19:35 +0200432 unbind_pasid(pasid_state);
Joerg Roedela40d4c62014-05-20 23:18:24 +0200433}
434
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100435static struct mmu_notifier_ops iommu_mn = {
Joerg Roedela40d4c62014-05-20 23:18:24 +0200436 .release = mn_release,
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100437 .clear_flush_young = mn_clear_flush_young,
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100438 .invalidate_page = mn_invalidate_page,
Joerg Roedele7cc3dd2014-11-13 13:46:09 +1100439 .invalidate_range = mn_invalidate_range,
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100440};
441
Joerg Roedel028eeac2011-11-24 12:48:13 +0100442static void set_pri_tag_status(struct pasid_state *pasid_state,
443 u16 tag, int status)
444{
445 unsigned long flags;
446
447 spin_lock_irqsave(&pasid_state->lock, flags);
448 pasid_state->pri[tag].status = status;
449 spin_unlock_irqrestore(&pasid_state->lock, flags);
450}
451
452static void finish_pri_tag(struct device_state *dev_state,
453 struct pasid_state *pasid_state,
454 u16 tag)
455{
456 unsigned long flags;
457
458 spin_lock_irqsave(&pasid_state->lock, flags);
459 if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
460 pasid_state->pri[tag].finish) {
461 amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
462 pasid_state->pri[tag].status, tag);
463 pasid_state->pri[tag].finish = false;
464 pasid_state->pri[tag].status = PPR_SUCCESS;
465 }
466 spin_unlock_irqrestore(&pasid_state->lock, flags);
467}
468
Jesse Barnes9dc00f42014-12-12 16:55:30 -0800469static void handle_fault_error(struct fault *fault)
470{
471 int status;
472
473 if (!fault->dev_state->inv_ppr_cb) {
474 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
475 return;
476 }
477
478 status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
479 fault->pasid,
480 fault->address,
481 fault->flags);
482 switch (status) {
483 case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
484 set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
485 break;
486 case AMD_IOMMU_INV_PRI_RSP_INVALID:
487 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
488 break;
489 case AMD_IOMMU_INV_PRI_RSP_FAIL:
490 set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
491 break;
492 default:
493 BUG();
494 }
495}
496
Joerg Roedel7b5cc1a2015-11-17 16:11:36 +0100497static bool access_error(struct vm_area_struct *vma, struct fault *fault)
498{
499 unsigned long requested = 0;
500
501 if (fault->flags & PPR_FAULT_EXEC)
502 requested |= VM_EXEC;
503
504 if (fault->flags & PPR_FAULT_READ)
505 requested |= VM_READ;
506
507 if (fault->flags & PPR_FAULT_WRITE)
508 requested |= VM_WRITE;
509
510 return (requested & ~vma->vm_flags) != 0;
511}
512
Joerg Roedel028eeac2011-11-24 12:48:13 +0100513static void do_fault(struct work_struct *work)
514{
515 struct fault *fault = container_of(work, struct fault, work);
Jesse Barnes9dc00f42014-12-12 16:55:30 -0800516 struct mm_struct *mm;
517 struct vm_area_struct *vma;
518 u64 address;
519 int ret, write;
Joerg Roedel028eeac2011-11-24 12:48:13 +0100520
521 write = !!(fault->flags & PPR_FAULT_WRITE);
522
Jesse Barnes9dc00f42014-12-12 16:55:30 -0800523 mm = fault->state->mm;
524 address = fault->address;
Joerg Roedel028eeac2011-11-24 12:48:13 +0100525
Jesse Barnes9dc00f42014-12-12 16:55:30 -0800526 down_read(&mm->mmap_sem);
527 vma = find_extend_vma(mm, address);
528 if (!vma || address < vma->vm_start) {
529 /* failed to get a vma in the right range */
530 up_read(&mm->mmap_sem);
531 handle_fault_error(fault);
532 goto out;
Joerg Roedel175d6142011-11-28 14:36:36 +0100533 }
Joerg Roedel028eeac2011-11-24 12:48:13 +0100534
Joerg Roedel7b5cc1a2015-11-17 16:11:36 +0100535 /* Check if we have the right permissions on the vma */
536 if (access_error(vma, fault)) {
Jay Cornwalld14f6fc2015-09-16 14:10:03 -0500537 up_read(&mm->mmap_sem);
538 handle_fault_error(fault);
539 goto out;
540 }
541
Jesse Barnes9dc00f42014-12-12 16:55:30 -0800542 ret = handle_mm_fault(mm, vma, address, write);
543 if (ret & VM_FAULT_ERROR) {
544 /* failed to service fault */
545 up_read(&mm->mmap_sem);
546 handle_fault_error(fault);
547 goto out;
548 }
549
550 up_read(&mm->mmap_sem);
551
552out:
Joerg Roedel028eeac2011-11-24 12:48:13 +0100553 finish_pri_tag(fault->dev_state, fault->state, fault->tag);
554
555 put_pasid_state(fault->state);
556
557 kfree(fault);
558}
559
560static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
561{
562 struct amd_iommu_fault *iommu_fault;
563 struct pasid_state *pasid_state;
564 struct device_state *dev_state;
565 unsigned long flags;
566 struct fault *fault;
567 bool finish;
568 u16 tag;
569 int ret;
570
571 iommu_fault = data;
572 tag = iommu_fault->tag & 0x1ff;
573 finish = (iommu_fault->tag >> 9) & 1;
574
575 ret = NOTIFY_DONE;
576 dev_state = get_device_state(iommu_fault->device_id);
577 if (dev_state == NULL)
578 goto out;
579
580 pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
Joerg Roedel53d340e2014-07-08 15:01:43 +0200581 if (pasid_state == NULL || pasid_state->invalid) {
Joerg Roedel028eeac2011-11-24 12:48:13 +0100582 /* We know the device but not the PASID -> send INVALID */
583 amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
584 PPR_INVALID, tag);
585 goto out_drop_state;
586 }
587
588 spin_lock_irqsave(&pasid_state->lock, flags);
589 atomic_inc(&pasid_state->pri[tag].inflight);
590 if (finish)
591 pasid_state->pri[tag].finish = true;
592 spin_unlock_irqrestore(&pasid_state->lock, flags);
593
594 fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
595 if (fault == NULL) {
596 /* We are OOM - send success and let the device re-fault */
597 finish_pri_tag(dev_state, pasid_state, tag);
598 goto out_drop_state;
599 }
600
601 fault->dev_state = dev_state;
602 fault->address = iommu_fault->address;
603 fault->state = pasid_state;
604 fault->tag = tag;
605 fault->finish = finish;
Alexey Skidanovb00675b2014-07-08 17:30:16 +0300606 fault->pasid = iommu_fault->pasid;
Joerg Roedel028eeac2011-11-24 12:48:13 +0100607 fault->flags = iommu_fault->flags;
608 INIT_WORK(&fault->work, do_fault);
609
610 queue_work(iommu_wq, &fault->work);
611
612 ret = NOTIFY_OK;
613
614out_drop_state:
Joerg Roedeldc88db72014-07-08 14:55:10 +0200615
616 if (ret != NOTIFY_OK && pasid_state)
617 put_pasid_state(pasid_state);
618
Joerg Roedel028eeac2011-11-24 12:48:13 +0100619 put_device_state(dev_state);
620
621out:
622 return ret;
623}
624
625static struct notifier_block ppr_nb = {
626 .notifier_call = ppr_notifier,
627};
628
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100629int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
630 struct task_struct *task)
631{
632 struct pasid_state *pasid_state;
633 struct device_state *dev_state;
Joerg Roedelf0aac632014-07-08 15:15:07 +0200634 struct mm_struct *mm;
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100635 u16 devid;
636 int ret;
637
638 might_sleep();
639
640 if (!amd_iommu_v2_supported())
641 return -ENODEV;
642
643 devid = device_id(pdev);
644 dev_state = get_device_state(devid);
645
646 if (dev_state == NULL)
647 return -EINVAL;
648
649 ret = -EINVAL;
650 if (pasid < 0 || pasid >= dev_state->max_pasids)
651 goto out;
652
653 ret = -ENOMEM;
654 pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
655 if (pasid_state == NULL)
656 goto out;
657
Joerg Roedelf0aac632014-07-08 15:15:07 +0200658
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100659 atomic_set(&pasid_state->count, 1);
Joerg Roedel028eeac2011-11-24 12:48:13 +0100660 init_waitqueue_head(&pasid_state->wq);
Joerg Roedel2c13d472012-07-19 10:56:10 +0200661 spin_lock_init(&pasid_state->lock);
662
Joerg Roedelf0aac632014-07-08 15:15:07 +0200663 mm = get_task_mm(task);
Joerg Roedelf0aac632014-07-08 15:15:07 +0200664 pasid_state->mm = mm;
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100665 pasid_state->device_state = dev_state;
666 pasid_state->pasid = pasid;
Joerg Roedeld9e16112014-07-09 15:43:11 +0200667 pasid_state->invalid = true; /* Mark as valid only if we are
668 done with setting up the pasid */
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100669 pasid_state->mn.ops = &iommu_mn;
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100670
671 if (pasid_state->mm == NULL)
672 goto out_free;
673
Joerg Roedelf0aac632014-07-08 15:15:07 +0200674 mmu_notifier_register(&pasid_state->mn, mm);
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100675
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100676 ret = set_pasid_state(dev_state, pasid_state, pasid);
677 if (ret)
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100678 goto out_unregister;
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100679
680 ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
681 __pa(pasid_state->mm->pgd));
682 if (ret)
683 goto out_clear_state;
684
Joerg Roedeld9e16112014-07-09 15:43:11 +0200685 /* Now we are ready to handle faults */
686 pasid_state->invalid = false;
687
Joerg Roedelf0aac632014-07-08 15:15:07 +0200688 /*
689 * Drop the reference to the mm_struct here. We rely on the
690 * mmu_notifier release call-back to inform us when the mm
691 * is going away.
692 */
693 mmput(mm);
694
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100695 return 0;
696
697out_clear_state:
698 clear_pasid_state(dev_state, pasid);
699
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100700out_unregister:
Joerg Roedelf0aac632014-07-08 15:15:07 +0200701 mmu_notifier_unregister(&pasid_state->mn, mm);
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100702
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100703out_free:
Joerg Roedelf0aac632014-07-08 15:15:07 +0200704 mmput(mm);
Joerg Roedel028eeac2011-11-24 12:48:13 +0100705 free_pasid_state(pasid_state);
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100706
707out:
708 put_device_state(dev_state);
709
710 return ret;
711}
712EXPORT_SYMBOL(amd_iommu_bind_pasid);
713
714void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
715{
Joerg Roedela40d4c62014-05-20 23:18:24 +0200716 struct pasid_state *pasid_state;
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100717 struct device_state *dev_state;
718 u16 devid;
719
720 might_sleep();
721
722 if (!amd_iommu_v2_supported())
723 return;
724
725 devid = device_id(pdev);
726 dev_state = get_device_state(devid);
727 if (dev_state == NULL)
728 return;
729
730 if (pasid < 0 || pasid >= dev_state->max_pasids)
731 goto out;
732
Joerg Roedela40d4c62014-05-20 23:18:24 +0200733 pasid_state = get_pasid_state(dev_state, pasid);
734 if (pasid_state == NULL)
735 goto out;
736 /*
737 * Drop reference taken here. We are safe because we still hold
738 * the reference taken in the amd_iommu_bind_pasid function.
739 */
740 put_pasid_state(pasid_state);
741
Joerg Roedel53d340e2014-07-08 15:01:43 +0200742 /* Clear the pasid state so that the pasid can be re-used */
743 clear_pasid_state(dev_state, pasid_state->pasid);
744
Joerg Roedelf0aac632014-07-08 15:15:07 +0200745 /*
Joerg Roedelfcaa9602014-07-30 16:04:37 +0200746 * Call mmu_notifier_unregister to drop our reference
747 * to pasid_state->mm
Joerg Roedelf0aac632014-07-08 15:15:07 +0200748 */
Joerg Roedelfcaa9602014-07-30 16:04:37 +0200749 mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100750
Joerg Roedelc5db16a2014-07-08 14:15:45 +0200751 put_pasid_state_wait(pasid_state); /* Reference taken in
Joerg Roedeldaff2f92014-07-30 16:04:40 +0200752 amd_iommu_bind_pasid */
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100753out:
Joerg Roedel75058a32014-07-30 16:04:39 +0200754 /* Drop reference taken in this function */
755 put_device_state(dev_state);
756
757 /* Drop reference taken in amd_iommu_bind_pasid */
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100758 put_device_state(dev_state);
759}
760EXPORT_SYMBOL(amd_iommu_unbind_pasid);
761
Joerg Roedeled96f222011-11-23 17:30:39 +0100762int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
763{
764 struct device_state *dev_state;
Joerg Roedel55c99a42015-07-28 16:58:47 +0200765 struct iommu_group *group;
Joerg Roedeled96f222011-11-23 17:30:39 +0100766 unsigned long flags;
767 int ret, tmp;
768 u16 devid;
769
770 might_sleep();
771
772 if (!amd_iommu_v2_supported())
773 return -ENODEV;
774
775 if (pasids <= 0 || pasids > (PASID_MASK + 1))
776 return -EINVAL;
777
778 devid = device_id(pdev);
779
780 dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
781 if (dev_state == NULL)
782 return -ENOMEM;
783
784 spin_lock_init(&dev_state->lock);
Joerg Roedel028eeac2011-11-24 12:48:13 +0100785 init_waitqueue_head(&dev_state->wq);
Joerg Roedel741669c2014-05-20 23:18:23 +0200786 dev_state->pdev = pdev;
787 dev_state->devid = devid;
Joerg Roedeled96f222011-11-23 17:30:39 +0100788
789 tmp = pasids;
790 for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
791 dev_state->pasid_levels += 1;
792
793 atomic_set(&dev_state->count, 1);
794 dev_state->max_pasids = pasids;
795
796 ret = -ENOMEM;
797 dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
798 if (dev_state->states == NULL)
799 goto out_free_dev_state;
800
801 dev_state->domain = iommu_domain_alloc(&pci_bus_type);
802 if (dev_state->domain == NULL)
803 goto out_free_states;
804
805 amd_iommu_domain_direct_map(dev_state->domain);
806
807 ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
808 if (ret)
809 goto out_free_domain;
810
Joerg Roedel55c99a42015-07-28 16:58:47 +0200811 group = iommu_group_get(&pdev->dev);
812 if (!group)
Joerg Roedeled96f222011-11-23 17:30:39 +0100813 goto out_free_domain;
814
Joerg Roedel55c99a42015-07-28 16:58:47 +0200815 ret = iommu_attach_group(dev_state->domain, group);
816 if (ret != 0)
817 goto out_drop_group;
818
819 iommu_group_put(group);
820
Joerg Roedeled96f222011-11-23 17:30:39 +0100821 spin_lock_irqsave(&state_lock, flags);
822
Joerg Roedel741669c2014-05-20 23:18:23 +0200823 if (__get_device_state(devid) != NULL) {
Joerg Roedeled96f222011-11-23 17:30:39 +0100824 spin_unlock_irqrestore(&state_lock, flags);
825 ret = -EBUSY;
826 goto out_free_domain;
827 }
828
Joerg Roedel741669c2014-05-20 23:18:23 +0200829 list_add_tail(&dev_state->list, &state_list);
Joerg Roedeled96f222011-11-23 17:30:39 +0100830
831 spin_unlock_irqrestore(&state_lock, flags);
832
833 return 0;
834
Joerg Roedel55c99a42015-07-28 16:58:47 +0200835out_drop_group:
836 iommu_group_put(group);
837
Joerg Roedeled96f222011-11-23 17:30:39 +0100838out_free_domain:
839 iommu_domain_free(dev_state->domain);
840
841out_free_states:
842 free_page((unsigned long)dev_state->states);
843
844out_free_dev_state:
845 kfree(dev_state);
846
847 return ret;
848}
849EXPORT_SYMBOL(amd_iommu_init_device);
850
851void amd_iommu_free_device(struct pci_dev *pdev)
852{
853 struct device_state *dev_state;
854 unsigned long flags;
855 u16 devid;
856
857 if (!amd_iommu_v2_supported())
858 return;
859
860 devid = device_id(pdev);
861
862 spin_lock_irqsave(&state_lock, flags);
863
Joerg Roedelb87d2d72014-05-20 23:18:22 +0200864 dev_state = __get_device_state(devid);
Joerg Roedeled96f222011-11-23 17:30:39 +0100865 if (dev_state == NULL) {
866 spin_unlock_irqrestore(&state_lock, flags);
867 return;
868 }
869
Joerg Roedel741669c2014-05-20 23:18:23 +0200870 list_del(&dev_state->list);
Joerg Roedeled96f222011-11-23 17:30:39 +0100871
872 spin_unlock_irqrestore(&state_lock, flags);
873
Joerg Roedel2d5503b2011-11-24 10:41:57 +0100874 /* Get rid of any remaining pasid states */
875 free_pasid_states(dev_state);
876
Peter Zijlstra91f65fa2015-02-03 13:25:51 +0100877 put_device_state(dev_state);
878 /*
879 * Wait until the last reference is dropped before freeing
880 * the device state.
881 */
882 wait_event(dev_state->wq, !atomic_read(&dev_state->count));
883 free_device_state(dev_state);
Joerg Roedeled96f222011-11-23 17:30:39 +0100884}
885EXPORT_SYMBOL(amd_iommu_free_device);
886
Joerg Roedel175d6142011-11-28 14:36:36 +0100887int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
888 amd_iommu_invalid_ppr_cb cb)
889{
890 struct device_state *dev_state;
891 unsigned long flags;
892 u16 devid;
893 int ret;
894
895 if (!amd_iommu_v2_supported())
896 return -ENODEV;
897
898 devid = device_id(pdev);
899
900 spin_lock_irqsave(&state_lock, flags);
901
902 ret = -EINVAL;
Joerg Roedelb87d2d72014-05-20 23:18:22 +0200903 dev_state = __get_device_state(devid);
Joerg Roedel175d6142011-11-28 14:36:36 +0100904 if (dev_state == NULL)
905 goto out_unlock;
906
907 dev_state->inv_ppr_cb = cb;
908
909 ret = 0;
910
911out_unlock:
912 spin_unlock_irqrestore(&state_lock, flags);
913
914 return ret;
915}
916EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
917
Joerg Roedelbc216622011-12-07 12:24:42 +0100918int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
919 amd_iommu_invalidate_ctx cb)
920{
921 struct device_state *dev_state;
922 unsigned long flags;
923 u16 devid;
924 int ret;
925
926 if (!amd_iommu_v2_supported())
927 return -ENODEV;
928
929 devid = device_id(pdev);
930
931 spin_lock_irqsave(&state_lock, flags);
932
933 ret = -EINVAL;
Joerg Roedelb87d2d72014-05-20 23:18:22 +0200934 dev_state = __get_device_state(devid);
Joerg Roedelbc216622011-12-07 12:24:42 +0100935 if (dev_state == NULL)
936 goto out_unlock;
937
938 dev_state->inv_ctx_cb = cb;
939
940 ret = 0;
941
942out_unlock:
943 spin_unlock_irqrestore(&state_lock, flags);
944
945 return ret;
946}
947EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
948
Joerg Roedele3c495c2011-11-09 12:31:15 +0100949static int __init amd_iommu_v2_init(void)
950{
Joerg Roedel028eeac2011-11-24 12:48:13 +0100951 int ret;
Joerg Roedeled96f222011-11-23 17:30:39 +0100952
Joerg Roedel63ce3ae2015-02-04 16:12:55 +0100953 pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
Joerg Roedel474d567d2012-03-15 12:46:40 +0100954
955 if (!amd_iommu_v2_supported()) {
Masanari Iida07db0402012-07-22 02:21:32 +0900956 pr_info("AMD IOMMUv2 functionality not available on this system\n");
Joerg Roedel474d567d2012-03-15 12:46:40 +0100957 /*
958 * Load anyway to provide the symbols to other modules
959 * which may use AMD IOMMUv2 optionally.
960 */
961 return 0;
962 }
Joerg Roedele3c495c2011-11-09 12:31:15 +0100963
Joerg Roedeled96f222011-11-23 17:30:39 +0100964 spin_lock_init(&state_lock);
965
Joerg Roedel028eeac2011-11-24 12:48:13 +0100966 ret = -ENOMEM;
967 iommu_wq = create_workqueue("amd_iommu_v2");
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100968 if (iommu_wq == NULL)
Joerg Roedel741669c2014-05-20 23:18:23 +0200969 goto out;
Joerg Roedel8736b2c2011-11-24 16:21:52 +0100970
Joerg Roedel028eeac2011-11-24 12:48:13 +0100971 amd_iommu_register_ppr_notifier(&ppr_nb);
972
Joerg Roedele3c495c2011-11-09 12:31:15 +0100973 return 0;
Joerg Roedel028eeac2011-11-24 12:48:13 +0100974
Joerg Roedel741669c2014-05-20 23:18:23 +0200975out:
Joerg Roedel028eeac2011-11-24 12:48:13 +0100976 return ret;
Joerg Roedele3c495c2011-11-09 12:31:15 +0100977}
978
979static void __exit amd_iommu_v2_exit(void)
980{
Joerg Roedeled96f222011-11-23 17:30:39 +0100981 struct device_state *dev_state;
Joerg Roedeled96f222011-11-23 17:30:39 +0100982 int i;
983
Joerg Roedel474d567d2012-03-15 12:46:40 +0100984 if (!amd_iommu_v2_supported())
985 return;
986
Joerg Roedel028eeac2011-11-24 12:48:13 +0100987 amd_iommu_unregister_ppr_notifier(&ppr_nb);
988
989 flush_workqueue(iommu_wq);
990
991 /*
992 * The loop below might call flush_workqueue(), so call
993 * destroy_workqueue() after it
994 */
Joerg Roedeled96f222011-11-23 17:30:39 +0100995 for (i = 0; i < MAX_DEVICES; ++i) {
996 dev_state = get_device_state(i);
997
998 if (dev_state == NULL)
999 continue;
1000
1001 WARN_ON_ONCE(1);
1002
Joerg Roedeled96f222011-11-23 17:30:39 +01001003 put_device_state(dev_state);
Joerg Roedel028eeac2011-11-24 12:48:13 +01001004 amd_iommu_free_device(dev_state->pdev);
Joerg Roedeled96f222011-11-23 17:30:39 +01001005 }
1006
Joerg Roedel028eeac2011-11-24 12:48:13 +01001007 destroy_workqueue(iommu_wq);
Joerg Roedele3c495c2011-11-09 12:31:15 +01001008}
1009
1010module_init(amd_iommu_v2_init);
1011module_exit(amd_iommu_v2_exit);