1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (c) 2021-2022, NVIDIA CORPORATION & AFFILIATES
4 */
5 #include <linux/interval_tree.h>
6 #include <linux/iommufd.h>
7 #include <linux/iommu.h>
8 #include <uapi/linux/iommufd.h>
9
10 #include "io_pagetable.h"
11
iommufd_ioas_destroy(struct iommufd_object * obj)12 void iommufd_ioas_destroy(struct iommufd_object *obj)
13 {
14 struct iommufd_ioas *ioas = container_of(obj, struct iommufd_ioas, obj);
15 int rc;
16
17 rc = iopt_unmap_all(&ioas->iopt, NULL);
18 WARN_ON(rc && rc != -ENOENT);
19 iopt_destroy_table(&ioas->iopt);
20 mutex_destroy(&ioas->mutex);
21 }
22
iommufd_ioas_alloc(struct iommufd_ctx * ictx)23 struct iommufd_ioas *iommufd_ioas_alloc(struct iommufd_ctx *ictx)
24 {
25 struct iommufd_ioas *ioas;
26
27 ioas = iommufd_object_alloc(ictx, ioas, IOMMUFD_OBJ_IOAS);
28 if (IS_ERR(ioas))
29 return ioas;
30
31 iopt_init_table(&ioas->iopt);
32 INIT_LIST_HEAD(&ioas->hwpt_list);
33 mutex_init(&ioas->mutex);
34 return ioas;
35 }
36
iommufd_ioas_alloc_ioctl(struct iommufd_ucmd * ucmd)37 int iommufd_ioas_alloc_ioctl(struct iommufd_ucmd *ucmd)
38 {
39 struct iommu_ioas_alloc *cmd = ucmd->cmd;
40 struct iommufd_ioas *ioas;
41 int rc;
42
43 if (cmd->flags)
44 return -EOPNOTSUPP;
45
46 ioas = iommufd_ioas_alloc(ucmd->ictx);
47 if (IS_ERR(ioas))
48 return PTR_ERR(ioas);
49
50 cmd->out_ioas_id = ioas->obj.id;
51 rc = iommufd_ucmd_respond(ucmd, sizeof(*cmd));
52 if (rc)
53 goto out_table;
54 iommufd_object_finalize(ucmd->ictx, &ioas->obj);
55 return 0;
56
57 out_table:
58 iommufd_object_abort_and_destroy(ucmd->ictx, &ioas->obj);
59 return rc;
60 }
61
iommufd_ioas_iova_ranges(struct iommufd_ucmd * ucmd)62 int iommufd_ioas_iova_ranges(struct iommufd_ucmd *ucmd)
63 {
64 struct iommu_iova_range __user *ranges;
65 struct iommu_ioas_iova_ranges *cmd = ucmd->cmd;
66 struct iommufd_ioas *ioas;
67 struct interval_tree_span_iter span;
68 u32 max_iovas;
69 int rc;
70
71 if (cmd->__reserved)
72 return -EOPNOTSUPP;
73
74 ioas = iommufd_get_ioas(ucmd->ictx, cmd->ioas_id);
75 if (IS_ERR(ioas))
76 return PTR_ERR(ioas);
77
78 down_read(&ioas->iopt.iova_rwsem);
79 max_iovas = cmd->num_iovas;
80 ranges = u64_to_user_ptr(cmd->allowed_iovas);
81 cmd->num_iovas = 0;
82 cmd->out_iova_alignment = ioas->iopt.iova_alignment;
83 interval_tree_for_each_span(&span, &ioas->iopt.reserved_itree, 0,
84 ULONG_MAX) {
85 if (!span.is_hole)
86 continue;
87 if (cmd->num_iovas < max_iovas) {
88 struct iommu_iova_range elm = {
89 .start = span.start_hole,
90 .last = span.last_hole,
91 };
92
93 if (copy_to_user(&ranges[cmd->num_iovas], &elm,
94 sizeof(elm))) {
95 rc = -EFAULT;
96 goto out_put;
97 }
98 }
99 cmd->num_iovas++;
100 }
101 rc = iommufd_ucmd_respond(ucmd, sizeof(*cmd));
102 if (rc)
103 goto out_put;
104 if (cmd->num_iovas > max_iovas)
105 rc = -EMSGSIZE;
106 out_put:
107 up_read(&ioas->iopt.iova_rwsem);
108 iommufd_put_object(&ioas->obj);
109 return rc;
110 }
111
iommufd_ioas_load_iovas(struct rb_root_cached * itree,struct iommu_iova_range __user * ranges,u32 num)112 static int iommufd_ioas_load_iovas(struct rb_root_cached *itree,
113 struct iommu_iova_range __user *ranges,
114 u32 num)
115 {
116 u32 i;
117
118 for (i = 0; i != num; i++) {
119 struct iommu_iova_range range;
120 struct iopt_allowed *allowed;
121
122 if (copy_from_user(&range, ranges + i, sizeof(range)))
123 return -EFAULT;
124
125 if (range.start >= range.last)
126 return -EINVAL;
127
128 if (interval_tree_iter_first(itree, range.start, range.last))
129 return -EINVAL;
130
131 allowed = kzalloc(sizeof(*allowed), GFP_KERNEL_ACCOUNT);
132 if (!allowed)
133 return -ENOMEM;
134 allowed->node.start = range.start;
135 allowed->node.last = range.last;
136
137 interval_tree_insert(&allowed->node, itree);
138 }
139 return 0;
140 }
141
iommufd_ioas_allow_iovas(struct iommufd_ucmd * ucmd)142 int iommufd_ioas_allow_iovas(struct iommufd_ucmd *ucmd)
143 {
144 struct iommu_ioas_allow_iovas *cmd = ucmd->cmd;
145 struct rb_root_cached allowed_iova = RB_ROOT_CACHED;
146 struct interval_tree_node *node;
147 struct iommufd_ioas *ioas;
148 struct io_pagetable *iopt;
149 int rc = 0;
150
151 if (cmd->__reserved)
152 return -EOPNOTSUPP;
153
154 ioas = iommufd_get_ioas(ucmd->ictx, cmd->ioas_id);
155 if (IS_ERR(ioas))
156 return PTR_ERR(ioas);
157 iopt = &ioas->iopt;
158
159 rc = iommufd_ioas_load_iovas(&allowed_iova,
160 u64_to_user_ptr(cmd->allowed_iovas),
161 cmd->num_iovas);
162 if (rc)
163 goto out_free;
164
165 /*
166 * We want the allowed tree update to be atomic, so we have to keep the
167 * original nodes around, and keep track of the new nodes as we allocate
168 * memory for them. The simplest solution is to have a new/old tree and
169 * then swap new for old. On success we free the old tree, on failure we
170 * free the new tree.
171 */
172 rc = iopt_set_allow_iova(iopt, &allowed_iova);
173 out_free:
174 while ((node = interval_tree_iter_first(&allowed_iova, 0, ULONG_MAX))) {
175 interval_tree_remove(node, &allowed_iova);
176 kfree(container_of(node, struct iopt_allowed, node));
177 }
178 iommufd_put_object(&ioas->obj);
179 return rc;
180 }
181
conv_iommu_prot(u32 map_flags)182 static int conv_iommu_prot(u32 map_flags)
183 {
184 /*
185 * We provide no manual cache coherency ioctls to userspace and most
186 * architectures make the CPU ops for cache flushing privileged.
187 * Therefore we require the underlying IOMMU to support CPU coherent
188 * operation. Support for IOMMU_CACHE is enforced by the
189 * IOMMU_CAP_CACHE_COHERENCY test during bind.
190 */
191 int iommu_prot = IOMMU_CACHE;
192
193 if (map_flags & IOMMU_IOAS_MAP_WRITEABLE)
194 iommu_prot |= IOMMU_WRITE;
195 if (map_flags & IOMMU_IOAS_MAP_READABLE)
196 iommu_prot |= IOMMU_READ;
197 return iommu_prot;
198 }
199
iommufd_ioas_map(struct iommufd_ucmd * ucmd)200 int iommufd_ioas_map(struct iommufd_ucmd *ucmd)
201 {
202 struct iommu_ioas_map *cmd = ucmd->cmd;
203 unsigned long iova = cmd->iova;
204 struct iommufd_ioas *ioas;
205 unsigned int flags = 0;
206 int rc;
207
208 if ((cmd->flags &
209 ~(IOMMU_IOAS_MAP_FIXED_IOVA | IOMMU_IOAS_MAP_WRITEABLE |
210 IOMMU_IOAS_MAP_READABLE)) ||
211 cmd->__reserved)
212 return -EOPNOTSUPP;
213 if (cmd->iova >= ULONG_MAX || cmd->length >= ULONG_MAX)
214 return -EOVERFLOW;
215
216 ioas = iommufd_get_ioas(ucmd->ictx, cmd->ioas_id);
217 if (IS_ERR(ioas))
218 return PTR_ERR(ioas);
219
220 if (!(cmd->flags & IOMMU_IOAS_MAP_FIXED_IOVA))
221 flags = IOPT_ALLOC_IOVA;
222 rc = iopt_map_user_pages(ucmd->ictx, &ioas->iopt, &iova,
223 u64_to_user_ptr(cmd->user_va), cmd->length,
224 conv_iommu_prot(cmd->flags), flags);
225 if (rc)
226 goto out_put;
227
228 cmd->iova = iova;
229 rc = iommufd_ucmd_respond(ucmd, sizeof(*cmd));
230 out_put:
231 iommufd_put_object(&ioas->obj);
232 return rc;
233 }
234
iommufd_ioas_copy(struct iommufd_ucmd * ucmd)235 int iommufd_ioas_copy(struct iommufd_ucmd *ucmd)
236 {
237 struct iommu_ioas_copy *cmd = ucmd->cmd;
238 struct iommufd_ioas *src_ioas;
239 struct iommufd_ioas *dst_ioas;
240 unsigned int flags = 0;
241 LIST_HEAD(pages_list);
242 unsigned long iova;
243 int rc;
244
245 iommufd_test_syz_conv_iova_id(ucmd, cmd->src_ioas_id, &cmd->src_iova,
246 &cmd->flags);
247
248 if ((cmd->flags &
249 ~(IOMMU_IOAS_MAP_FIXED_IOVA | IOMMU_IOAS_MAP_WRITEABLE |
250 IOMMU_IOAS_MAP_READABLE)))
251 return -EOPNOTSUPP;
252 if (cmd->length >= ULONG_MAX || cmd->src_iova >= ULONG_MAX ||
253 cmd->dst_iova >= ULONG_MAX)
254 return -EOVERFLOW;
255
256 src_ioas = iommufd_get_ioas(ucmd->ictx, cmd->src_ioas_id);
257 if (IS_ERR(src_ioas))
258 return PTR_ERR(src_ioas);
259 rc = iopt_get_pages(&src_ioas->iopt, cmd->src_iova, cmd->length,
260 &pages_list);
261 iommufd_put_object(&src_ioas->obj);
262 if (rc)
263 return rc;
264
265 dst_ioas = iommufd_get_ioas(ucmd->ictx, cmd->dst_ioas_id);
266 if (IS_ERR(dst_ioas)) {
267 rc = PTR_ERR(dst_ioas);
268 goto out_pages;
269 }
270
271 if (!(cmd->flags & IOMMU_IOAS_MAP_FIXED_IOVA))
272 flags = IOPT_ALLOC_IOVA;
273 iova = cmd->dst_iova;
274 rc = iopt_map_pages(&dst_ioas->iopt, &pages_list, cmd->length, &iova,
275 conv_iommu_prot(cmd->flags), flags);
276 if (rc)
277 goto out_put_dst;
278
279 cmd->dst_iova = iova;
280 rc = iommufd_ucmd_respond(ucmd, sizeof(*cmd));
281 out_put_dst:
282 iommufd_put_object(&dst_ioas->obj);
283 out_pages:
284 iopt_free_pages_list(&pages_list);
285 return rc;
286 }
287
iommufd_ioas_unmap(struct iommufd_ucmd * ucmd)288 int iommufd_ioas_unmap(struct iommufd_ucmd *ucmd)
289 {
290 struct iommu_ioas_unmap *cmd = ucmd->cmd;
291 struct iommufd_ioas *ioas;
292 unsigned long unmapped = 0;
293 int rc;
294
295 ioas = iommufd_get_ioas(ucmd->ictx, cmd->ioas_id);
296 if (IS_ERR(ioas))
297 return PTR_ERR(ioas);
298
299 if (cmd->iova == 0 && cmd->length == U64_MAX) {
300 rc = iopt_unmap_all(&ioas->iopt, &unmapped);
301 if (rc)
302 goto out_put;
303 } else {
304 if (cmd->iova >= ULONG_MAX || cmd->length >= ULONG_MAX) {
305 rc = -EOVERFLOW;
306 goto out_put;
307 }
308 rc = iopt_unmap_iova(&ioas->iopt, cmd->iova, cmd->length,
309 &unmapped);
310 if (rc)
311 goto out_put;
312 }
313
314 cmd->length = unmapped;
315 rc = iommufd_ucmd_respond(ucmd, sizeof(*cmd));
316
317 out_put:
318 iommufd_put_object(&ioas->obj);
319 return rc;
320 }
321
iommufd_option_rlimit_mode(struct iommu_option * cmd,struct iommufd_ctx * ictx)322 int iommufd_option_rlimit_mode(struct iommu_option *cmd,
323 struct iommufd_ctx *ictx)
324 {
325 if (cmd->object_id)
326 return -EOPNOTSUPP;
327
328 if (cmd->op == IOMMU_OPTION_OP_GET) {
329 cmd->val64 = ictx->account_mode == IOPT_PAGES_ACCOUNT_MM;
330 return 0;
331 }
332 if (cmd->op == IOMMU_OPTION_OP_SET) {
333 int rc = 0;
334
335 if (!capable(CAP_SYS_RESOURCE))
336 return -EPERM;
337
338 xa_lock(&ictx->objects);
339 if (!xa_empty(&ictx->objects)) {
340 rc = -EBUSY;
341 } else {
342 if (cmd->val64 == 0)
343 ictx->account_mode = IOPT_PAGES_ACCOUNT_USER;
344 else if (cmd->val64 == 1)
345 ictx->account_mode = IOPT_PAGES_ACCOUNT_MM;
346 else
347 rc = -EINVAL;
348 }
349 xa_unlock(&ictx->objects);
350
351 return rc;
352 }
353 return -EOPNOTSUPP;
354 }
355
iommufd_ioas_option_huge_pages(struct iommu_option * cmd,struct iommufd_ioas * ioas)356 static int iommufd_ioas_option_huge_pages(struct iommu_option *cmd,
357 struct iommufd_ioas *ioas)
358 {
359 if (cmd->op == IOMMU_OPTION_OP_GET) {
360 cmd->val64 = !ioas->iopt.disable_large_pages;
361 return 0;
362 }
363 if (cmd->op == IOMMU_OPTION_OP_SET) {
364 if (cmd->val64 == 0)
365 return iopt_disable_large_pages(&ioas->iopt);
366 if (cmd->val64 == 1) {
367 iopt_enable_large_pages(&ioas->iopt);
368 return 0;
369 }
370 return -EINVAL;
371 }
372 return -EOPNOTSUPP;
373 }
374
iommufd_ioas_option(struct iommufd_ucmd * ucmd)375 int iommufd_ioas_option(struct iommufd_ucmd *ucmd)
376 {
377 struct iommu_option *cmd = ucmd->cmd;
378 struct iommufd_ioas *ioas;
379 int rc = 0;
380
381 if (cmd->__reserved)
382 return -EOPNOTSUPP;
383
384 ioas = iommufd_get_ioas(ucmd->ictx, cmd->object_id);
385 if (IS_ERR(ioas))
386 return PTR_ERR(ioas);
387
388 switch (cmd->option_id) {
389 case IOMMU_OPTION_HUGE_PAGES:
390 rc = iommufd_ioas_option_huge_pages(cmd, ioas);
391 break;
392 default:
393 rc = -EOPNOTSUPP;
394 }
395
396 iommufd_put_object(&ioas->obj);
397 return rc;
398 }
399