1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (C) 2024 Intel Corporation
3 */
4 #define pr_fmt(fmt) "iommufd: " fmt
5
6 #include <linux/anon_inodes.h>
7 #include <linux/file.h>
8 #include <linux/fs.h>
9 #include <linux/iommufd.h>
10 #include <linux/module.h>
11 #include <linux/mutex.h>
12 #include <linux/pci.h>
13 #include <linux/pci-ats.h>
14 #include <linux/poll.h>
15 #include <uapi/linux/iommufd.h>
16
17 #include "../iommu-priv.h"
18 #include "iommufd_private.h"
19
20 /* IOMMUFD_OBJ_FAULT Functions */
21
iommufd_fault_iopf_enable(struct iommufd_device * idev)22 int iommufd_fault_iopf_enable(struct iommufd_device *idev)
23 {
24 struct device *dev = idev->dev;
25 int ret;
26
27 /*
28 * Once we turn on PCI/PRI support for VF, the response failure code
29 * should not be forwarded to the hardware due to PRI being a shared
30 * resource between PF and VFs. There is no coordination for this
31 * shared capability. This waits for a vPRI reset to recover.
32 */
33 if (dev_is_pci(dev)) {
34 struct pci_dev *pdev = to_pci_dev(dev);
35
36 if (pdev->is_virtfn && pci_pri_supported(pdev))
37 return -EINVAL;
38 }
39
40 mutex_lock(&idev->iopf_lock);
41 /* Device iopf has already been on. */
42 if (++idev->iopf_enabled > 1) {
43 mutex_unlock(&idev->iopf_lock);
44 return 0;
45 }
46
47 ret = iommu_dev_enable_feature(dev, IOMMU_DEV_FEAT_IOPF);
48 if (ret)
49 --idev->iopf_enabled;
50 mutex_unlock(&idev->iopf_lock);
51
52 return ret;
53 }
54
iommufd_fault_iopf_disable(struct iommufd_device * idev)55 void iommufd_fault_iopf_disable(struct iommufd_device *idev)
56 {
57 mutex_lock(&idev->iopf_lock);
58 if (!WARN_ON(idev->iopf_enabled == 0)) {
59 if (--idev->iopf_enabled == 0)
60 iommu_dev_disable_feature(idev->dev, IOMMU_DEV_FEAT_IOPF);
61 }
62 mutex_unlock(&idev->iopf_lock);
63 }
64
iommufd_auto_response_faults(struct iommufd_hw_pagetable * hwpt,struct iommufd_attach_handle * handle)65 void iommufd_auto_response_faults(struct iommufd_hw_pagetable *hwpt,
66 struct iommufd_attach_handle *handle)
67 {
68 struct iommufd_fault *fault = hwpt->fault;
69 struct iopf_group *group, *next;
70 struct list_head free_list;
71 unsigned long index;
72
73 if (!fault)
74 return;
75 INIT_LIST_HEAD(&free_list);
76
77 mutex_lock(&fault->mutex);
78 spin_lock(&fault->common.lock);
79 list_for_each_entry_safe(group, next, &fault->common.deliver, node) {
80 if (group->attach_handle != &handle->handle)
81 continue;
82 list_move(&group->node, &free_list);
83 }
84 spin_unlock(&fault->common.lock);
85
86 list_for_each_entry_safe(group, next, &free_list, node) {
87 list_del(&group->node);
88 iopf_group_response(group, IOMMU_PAGE_RESP_INVALID);
89 iopf_free_group(group);
90 }
91
92 xa_for_each(&fault->response, index, group) {
93 if (group->attach_handle != &handle->handle)
94 continue;
95 xa_erase(&fault->response, index);
96 iopf_group_response(group, IOMMU_PAGE_RESP_INVALID);
97 iopf_free_group(group);
98 }
99 mutex_unlock(&fault->mutex);
100 }
101
iommufd_fault_destroy(struct iommufd_object * obj)102 void iommufd_fault_destroy(struct iommufd_object *obj)
103 {
104 struct iommufd_eventq *eventq =
105 container_of(obj, struct iommufd_eventq, obj);
106 struct iommufd_fault *fault = eventq_to_fault(eventq);
107 struct iopf_group *group, *next;
108 unsigned long index;
109
110 /*
111 * The iommufd object's reference count is zero at this point.
112 * We can be confident that no other threads are currently
113 * accessing this pointer. Therefore, acquiring the mutex here
114 * is unnecessary.
115 */
116 list_for_each_entry_safe(group, next, &fault->common.deliver, node) {
117 list_del(&group->node);
118 iopf_group_response(group, IOMMU_PAGE_RESP_INVALID);
119 iopf_free_group(group);
120 }
121 xa_for_each(&fault->response, index, group) {
122 xa_erase(&fault->response, index);
123 iopf_group_response(group, IOMMU_PAGE_RESP_INVALID);
124 iopf_free_group(group);
125 }
126 xa_destroy(&fault->response);
127 mutex_destroy(&fault->mutex);
128 }
129
iommufd_compose_fault_message(struct iommu_fault * fault,struct iommu_hwpt_pgfault * hwpt_fault,struct iommufd_device * idev,u32 cookie)130 static void iommufd_compose_fault_message(struct iommu_fault *fault,
131 struct iommu_hwpt_pgfault *hwpt_fault,
132 struct iommufd_device *idev,
133 u32 cookie)
134 {
135 hwpt_fault->flags = fault->prm.flags;
136 hwpt_fault->dev_id = idev->obj.id;
137 hwpt_fault->pasid = fault->prm.pasid;
138 hwpt_fault->grpid = fault->prm.grpid;
139 hwpt_fault->perm = fault->prm.perm;
140 hwpt_fault->addr = fault->prm.addr;
141 hwpt_fault->length = 0;
142 hwpt_fault->cookie = cookie;
143 }
144
145 /* Fetch the first node out of the fault->deliver list */
146 static struct iopf_group *
iommufd_fault_deliver_fetch(struct iommufd_fault * fault)147 iommufd_fault_deliver_fetch(struct iommufd_fault *fault)
148 {
149 struct list_head *list = &fault->common.deliver;
150 struct iopf_group *group = NULL;
151
152 spin_lock(&fault->common.lock);
153 if (!list_empty(list)) {
154 group = list_first_entry(list, struct iopf_group, node);
155 list_del(&group->node);
156 }
157 spin_unlock(&fault->common.lock);
158 return group;
159 }
160
161 /* Restore a node back to the head of the fault->deliver list */
iommufd_fault_deliver_restore(struct iommufd_fault * fault,struct iopf_group * group)162 static void iommufd_fault_deliver_restore(struct iommufd_fault *fault,
163 struct iopf_group *group)
164 {
165 spin_lock(&fault->common.lock);
166 list_add(&group->node, &fault->common.deliver);
167 spin_unlock(&fault->common.lock);
168 }
169
iommufd_fault_fops_read(struct file * filep,char __user * buf,size_t count,loff_t * ppos)170 static ssize_t iommufd_fault_fops_read(struct file *filep, char __user *buf,
171 size_t count, loff_t *ppos)
172 {
173 size_t fault_size = sizeof(struct iommu_hwpt_pgfault);
174 struct iommufd_eventq *eventq = filep->private_data;
175 struct iommufd_fault *fault = eventq_to_fault(eventq);
176 struct iommu_hwpt_pgfault data = {};
177 struct iommufd_device *idev;
178 struct iopf_group *group;
179 struct iopf_fault *iopf;
180 size_t done = 0;
181 int rc = 0;
182
183 if (*ppos || count % fault_size)
184 return -ESPIPE;
185
186 mutex_lock(&fault->mutex);
187 while ((group = iommufd_fault_deliver_fetch(fault))) {
188 if (done >= count ||
189 group->fault_count * fault_size > count - done) {
190 iommufd_fault_deliver_restore(fault, group);
191 break;
192 }
193
194 rc = xa_alloc(&fault->response, &group->cookie, group,
195 xa_limit_32b, GFP_KERNEL);
196 if (rc) {
197 iommufd_fault_deliver_restore(fault, group);
198 break;
199 }
200
201 idev = to_iommufd_handle(group->attach_handle)->idev;
202 list_for_each_entry(iopf, &group->faults, list) {
203 iommufd_compose_fault_message(&iopf->fault,
204 &data, idev,
205 group->cookie);
206 if (copy_to_user(buf + done, &data, fault_size)) {
207 xa_erase(&fault->response, group->cookie);
208 iommufd_fault_deliver_restore(fault, group);
209 rc = -EFAULT;
210 break;
211 }
212 done += fault_size;
213 }
214 }
215 mutex_unlock(&fault->mutex);
216
217 return done == 0 ? rc : done;
218 }
219
iommufd_fault_fops_write(struct file * filep,const char __user * buf,size_t count,loff_t * ppos)220 static ssize_t iommufd_fault_fops_write(struct file *filep, const char __user *buf,
221 size_t count, loff_t *ppos)
222 {
223 size_t response_size = sizeof(struct iommu_hwpt_page_response);
224 struct iommufd_eventq *eventq = filep->private_data;
225 struct iommufd_fault *fault = eventq_to_fault(eventq);
226 struct iommu_hwpt_page_response response;
227 struct iopf_group *group;
228 size_t done = 0;
229 int rc = 0;
230
231 if (*ppos || count % response_size)
232 return -ESPIPE;
233
234 mutex_lock(&fault->mutex);
235 while (count > done) {
236 rc = copy_from_user(&response, buf + done, response_size);
237 if (rc)
238 break;
239
240 static_assert((int)IOMMUFD_PAGE_RESP_SUCCESS ==
241 (int)IOMMU_PAGE_RESP_SUCCESS);
242 static_assert((int)IOMMUFD_PAGE_RESP_INVALID ==
243 (int)IOMMU_PAGE_RESP_INVALID);
244 if (response.code != IOMMUFD_PAGE_RESP_SUCCESS &&
245 response.code != IOMMUFD_PAGE_RESP_INVALID) {
246 rc = -EINVAL;
247 break;
248 }
249
250 group = xa_erase(&fault->response, response.cookie);
251 if (!group) {
252 rc = -EINVAL;
253 break;
254 }
255
256 iopf_group_response(group, response.code);
257 iopf_free_group(group);
258 done += response_size;
259 }
260 mutex_unlock(&fault->mutex);
261
262 return done == 0 ? rc : done;
263 }
264
265 /* IOMMUFD_OBJ_VEVENTQ Functions */
266
iommufd_veventq_abort(struct iommufd_object * obj)267 void iommufd_veventq_abort(struct iommufd_object *obj)
268 {
269 struct iommufd_eventq *eventq =
270 container_of(obj, struct iommufd_eventq, obj);
271 struct iommufd_veventq *veventq = eventq_to_veventq(eventq);
272 struct iommufd_viommu *viommu = veventq->viommu;
273 struct iommufd_vevent *cur, *next;
274
275 lockdep_assert_held_write(&viommu->veventqs_rwsem);
276
277 list_for_each_entry_safe(cur, next, &eventq->deliver, node) {
278 list_del(&cur->node);
279 if (cur != &veventq->lost_events_header)
280 kfree(cur);
281 }
282
283 refcount_dec(&viommu->obj.users);
284 list_del(&veventq->node);
285 }
286
iommufd_veventq_destroy(struct iommufd_object * obj)287 void iommufd_veventq_destroy(struct iommufd_object *obj)
288 {
289 struct iommufd_veventq *veventq = eventq_to_veventq(
290 container_of(obj, struct iommufd_eventq, obj));
291
292 down_write(&veventq->viommu->veventqs_rwsem);
293 iommufd_veventq_abort(obj);
294 up_write(&veventq->viommu->veventqs_rwsem);
295 }
296
297 static struct iommufd_vevent *
iommufd_veventq_deliver_fetch(struct iommufd_veventq * veventq)298 iommufd_veventq_deliver_fetch(struct iommufd_veventq *veventq)
299 {
300 struct iommufd_eventq *eventq = &veventq->common;
301 struct list_head *list = &eventq->deliver;
302 struct iommufd_vevent *vevent = NULL;
303
304 spin_lock(&eventq->lock);
305 if (!list_empty(list)) {
306 struct iommufd_vevent *next;
307
308 next = list_first_entry(list, struct iommufd_vevent, node);
309 /* Make a copy of the lost_events_header for copy_to_user */
310 if (next == &veventq->lost_events_header) {
311 vevent = kzalloc(sizeof(*vevent), GFP_ATOMIC);
312 if (!vevent)
313 goto out_unlock;
314 }
315 list_del(&next->node);
316 if (vevent)
317 memcpy(vevent, next, sizeof(*vevent));
318 else
319 vevent = next;
320 }
321 out_unlock:
322 spin_unlock(&eventq->lock);
323 return vevent;
324 }
325
iommufd_veventq_deliver_restore(struct iommufd_veventq * veventq,struct iommufd_vevent * vevent)326 static void iommufd_veventq_deliver_restore(struct iommufd_veventq *veventq,
327 struct iommufd_vevent *vevent)
328 {
329 struct iommufd_eventq *eventq = &veventq->common;
330 struct list_head *list = &eventq->deliver;
331
332 spin_lock(&eventq->lock);
333 if (vevent_for_lost_events_header(vevent)) {
334 /* Remove the copy of the lost_events_header */
335 kfree(vevent);
336 vevent = NULL;
337 /* An empty list needs the lost_events_header back */
338 if (list_empty(list))
339 vevent = &veventq->lost_events_header;
340 }
341 if (vevent)
342 list_add(&vevent->node, list);
343 spin_unlock(&eventq->lock);
344 }
345
iommufd_veventq_fops_read(struct file * filep,char __user * buf,size_t count,loff_t * ppos)346 static ssize_t iommufd_veventq_fops_read(struct file *filep, char __user *buf,
347 size_t count, loff_t *ppos)
348 {
349 struct iommufd_eventq *eventq = filep->private_data;
350 struct iommufd_veventq *veventq = eventq_to_veventq(eventq);
351 struct iommufd_vevent_header *hdr;
352 struct iommufd_vevent *cur;
353 size_t done = 0;
354 int rc = 0;
355
356 if (*ppos)
357 return -ESPIPE;
358
359 while ((cur = iommufd_veventq_deliver_fetch(veventq))) {
360 /* Validate the remaining bytes against the header size */
361 if (done >= count || sizeof(*hdr) > count - done) {
362 iommufd_veventq_deliver_restore(veventq, cur);
363 break;
364 }
365 hdr = &cur->header;
366
367 /* If being a normal vEVENT, validate against the full size */
368 if (!vevent_for_lost_events_header(cur) &&
369 sizeof(hdr) + cur->data_len > count - done) {
370 iommufd_veventq_deliver_restore(veventq, cur);
371 break;
372 }
373
374 if (copy_to_user(buf + done, hdr, sizeof(*hdr))) {
375 iommufd_veventq_deliver_restore(veventq, cur);
376 rc = -EFAULT;
377 break;
378 }
379 done += sizeof(*hdr);
380
381 if (cur->data_len &&
382 copy_to_user(buf + done, cur->event_data, cur->data_len)) {
383 iommufd_veventq_deliver_restore(veventq, cur);
384 rc = -EFAULT;
385 break;
386 }
387 spin_lock(&eventq->lock);
388 if (!vevent_for_lost_events_header(cur))
389 veventq->num_events--;
390 spin_unlock(&eventq->lock);
391 done += cur->data_len;
392 kfree(cur);
393 }
394
395 return done == 0 ? rc : done;
396 }
397
398 /* Common Event Queue Functions */
399
iommufd_eventq_fops_poll(struct file * filep,struct poll_table_struct * wait)400 static __poll_t iommufd_eventq_fops_poll(struct file *filep,
401 struct poll_table_struct *wait)
402 {
403 struct iommufd_eventq *eventq = filep->private_data;
404 __poll_t pollflags = 0;
405
406 if (eventq->obj.type == IOMMUFD_OBJ_FAULT)
407 pollflags |= EPOLLOUT;
408
409 poll_wait(filep, &eventq->wait_queue, wait);
410 spin_lock(&eventq->lock);
411 if (!list_empty(&eventq->deliver))
412 pollflags |= EPOLLIN | EPOLLRDNORM;
413 spin_unlock(&eventq->lock);
414
415 return pollflags;
416 }
417
iommufd_eventq_fops_release(struct inode * inode,struct file * filep)418 static int iommufd_eventq_fops_release(struct inode *inode, struct file *filep)
419 {
420 struct iommufd_eventq *eventq = filep->private_data;
421
422 refcount_dec(&eventq->obj.users);
423 iommufd_ctx_put(eventq->ictx);
424 return 0;
425 }
426
427 #define INIT_EVENTQ_FOPS(read_op, write_op) \
428 ((const struct file_operations){ \
429 .owner = THIS_MODULE, \
430 .open = nonseekable_open, \
431 .read = read_op, \
432 .write = write_op, \
433 .poll = iommufd_eventq_fops_poll, \
434 .release = iommufd_eventq_fops_release, \
435 })
436
iommufd_eventq_init(struct iommufd_eventq * eventq,char * name,struct iommufd_ctx * ictx,const struct file_operations * fops)437 static int iommufd_eventq_init(struct iommufd_eventq *eventq, char *name,
438 struct iommufd_ctx *ictx,
439 const struct file_operations *fops)
440 {
441 struct file *filep;
442 int fdno;
443
444 spin_lock_init(&eventq->lock);
445 INIT_LIST_HEAD(&eventq->deliver);
446 init_waitqueue_head(&eventq->wait_queue);
447
448 filep = anon_inode_getfile(name, fops, eventq, O_RDWR);
449 if (IS_ERR(filep))
450 return PTR_ERR(filep);
451
452 eventq->ictx = ictx;
453 iommufd_ctx_get(eventq->ictx);
454 eventq->filep = filep;
455 refcount_inc(&eventq->obj.users);
456
457 fdno = get_unused_fd_flags(O_CLOEXEC);
458 if (fdno < 0)
459 fput(filep);
460 return fdno;
461 }
462
463 static const struct file_operations iommufd_fault_fops =
464 INIT_EVENTQ_FOPS(iommufd_fault_fops_read, iommufd_fault_fops_write);
465
iommufd_fault_alloc(struct iommufd_ucmd * ucmd)466 int iommufd_fault_alloc(struct iommufd_ucmd *ucmd)
467 {
468 struct iommu_fault_alloc *cmd = ucmd->cmd;
469 struct iommufd_fault *fault;
470 int fdno;
471 int rc;
472
473 if (cmd->flags)
474 return -EOPNOTSUPP;
475
476 fault = __iommufd_object_alloc(ucmd->ictx, fault, IOMMUFD_OBJ_FAULT,
477 common.obj);
478 if (IS_ERR(fault))
479 return PTR_ERR(fault);
480
481 xa_init_flags(&fault->response, XA_FLAGS_ALLOC1);
482 mutex_init(&fault->mutex);
483
484 fdno = iommufd_eventq_init(&fault->common, "[iommufd-pgfault]",
485 ucmd->ictx, &iommufd_fault_fops);
486 if (fdno < 0) {
487 rc = fdno;
488 goto out_abort;
489 }
490
491 cmd->out_fault_id = fault->common.obj.id;
492 cmd->out_fault_fd = fdno;
493
494 rc = iommufd_ucmd_respond(ucmd, sizeof(*cmd));
495 if (rc)
496 goto out_put_fdno;
497 iommufd_object_finalize(ucmd->ictx, &fault->common.obj);
498
499 fd_install(fdno, fault->common.filep);
500
501 return 0;
502 out_put_fdno:
503 put_unused_fd(fdno);
504 fput(fault->common.filep);
505 out_abort:
506 iommufd_object_abort_and_destroy(ucmd->ictx, &fault->common.obj);
507
508 return rc;
509 }
510
iommufd_fault_iopf_handler(struct iopf_group * group)511 int iommufd_fault_iopf_handler(struct iopf_group *group)
512 {
513 struct iommufd_hw_pagetable *hwpt;
514 struct iommufd_fault *fault;
515
516 hwpt = group->attach_handle->domain->iommufd_hwpt;
517 fault = hwpt->fault;
518
519 spin_lock(&fault->common.lock);
520 list_add_tail(&group->node, &fault->common.deliver);
521 spin_unlock(&fault->common.lock);
522
523 wake_up_interruptible(&fault->common.wait_queue);
524
525 return 0;
526 }
527
528 static const struct file_operations iommufd_veventq_fops =
529 INIT_EVENTQ_FOPS(iommufd_veventq_fops_read, NULL);
530
iommufd_veventq_alloc(struct iommufd_ucmd * ucmd)531 int iommufd_veventq_alloc(struct iommufd_ucmd *ucmd)
532 {
533 struct iommu_veventq_alloc *cmd = ucmd->cmd;
534 struct iommufd_veventq *veventq;
535 struct iommufd_viommu *viommu;
536 int fdno;
537 int rc;
538
539 if (cmd->flags || cmd->__reserved ||
540 cmd->type == IOMMU_VEVENTQ_TYPE_DEFAULT)
541 return -EOPNOTSUPP;
542 if (!cmd->veventq_depth)
543 return -EINVAL;
544
545 viommu = iommufd_get_viommu(ucmd, cmd->viommu_id);
546 if (IS_ERR(viommu))
547 return PTR_ERR(viommu);
548
549 down_write(&viommu->veventqs_rwsem);
550
551 if (iommufd_viommu_find_veventq(viommu, cmd->type)) {
552 rc = -EEXIST;
553 goto out_unlock_veventqs;
554 }
555
556 veventq = __iommufd_object_alloc(ucmd->ictx, veventq,
557 IOMMUFD_OBJ_VEVENTQ, common.obj);
558 if (IS_ERR(veventq)) {
559 rc = PTR_ERR(veventq);
560 goto out_unlock_veventqs;
561 }
562
563 veventq->type = cmd->type;
564 veventq->viommu = viommu;
565 refcount_inc(&viommu->obj.users);
566 veventq->depth = cmd->veventq_depth;
567 list_add_tail(&veventq->node, &viommu->veventqs);
568 veventq->lost_events_header.header.flags =
569 IOMMU_VEVENTQ_FLAG_LOST_EVENTS;
570
571 fdno = iommufd_eventq_init(&veventq->common, "[iommufd-viommu-event]",
572 ucmd->ictx, &iommufd_veventq_fops);
573 if (fdno < 0) {
574 rc = fdno;
575 goto out_abort;
576 }
577
578 cmd->out_veventq_id = veventq->common.obj.id;
579 cmd->out_veventq_fd = fdno;
580
581 rc = iommufd_ucmd_respond(ucmd, sizeof(*cmd));
582 if (rc)
583 goto out_put_fdno;
584
585 iommufd_object_finalize(ucmd->ictx, &veventq->common.obj);
586 fd_install(fdno, veventq->common.filep);
587 goto out_unlock_veventqs;
588
589 out_put_fdno:
590 put_unused_fd(fdno);
591 fput(veventq->common.filep);
592 out_abort:
593 iommufd_object_abort_and_destroy(ucmd->ictx, &veventq->common.obj);
594 out_unlock_veventqs:
595 up_write(&viommu->veventqs_rwsem);
596 iommufd_put_object(ucmd->ictx, &viommu->obj);
597 return rc;
598 }
599