Loading...
1/*
2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
4 *
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 by
7 * the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 *
14 * You should have received a copy of the GNU General Public License along with
15 * this program. If not, see <http://www.gnu.org/licenses/>.
16 */
17
18#include "msm_drv.h"
19#include "msm_gpu.h"
20#include "msm_gem.h"
21
22/*
23 * Cmdstream submission:
24 */
25
26/* make sure these don't conflict w/ MSM_SUBMIT_BO_x */
27#define BO_VALID 0x8000
28#define BO_LOCKED 0x4000
29#define BO_PINNED 0x2000
30
31static inline void __user *to_user_ptr(u64 address)
32{
33 return (void __user *)(uintptr_t)address;
34}
35
36static struct msm_gem_submit *submit_create(struct drm_device *dev,
37 struct msm_gpu *gpu, int nr)
38{
39 struct msm_gem_submit *submit;
40 int sz = sizeof(*submit) + (nr * sizeof(submit->bos[0]));
41
42 submit = kmalloc(sz, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
43 if (submit) {
44 submit->dev = dev;
45 submit->gpu = gpu;
46
47 /* initially, until copy_from_user() and bo lookup succeeds: */
48 submit->nr_bos = 0;
49 submit->nr_cmds = 0;
50
51 INIT_LIST_HEAD(&submit->bo_list);
52 ww_acquire_init(&submit->ticket, &reservation_ww_class);
53 }
54
55 return submit;
56}
57
58static int submit_lookup_objects(struct msm_gem_submit *submit,
59 struct drm_msm_gem_submit *args, struct drm_file *file)
60{
61 unsigned i;
62 int ret = 0;
63
64 spin_lock(&file->table_lock);
65
66 for (i = 0; i < args->nr_bos; i++) {
67 struct drm_msm_gem_submit_bo submit_bo;
68 struct drm_gem_object *obj;
69 struct msm_gem_object *msm_obj;
70 void __user *userptr =
71 to_user_ptr(args->bos + (i * sizeof(submit_bo)));
72
73 ret = copy_from_user(&submit_bo, userptr, sizeof(submit_bo));
74 if (ret) {
75 ret = -EFAULT;
76 goto out_unlock;
77 }
78
79 if (submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) {
80 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
81 ret = -EINVAL;
82 goto out_unlock;
83 }
84
85 submit->bos[i].flags = submit_bo.flags;
86 /* in validate_objects() we figure out if this is true: */
87 submit->bos[i].iova = submit_bo.presumed;
88
89 /* normally use drm_gem_object_lookup(), but for bulk lookup
90 * all under single table_lock just hit object_idr directly:
91 */
92 obj = idr_find(&file->object_idr, submit_bo.handle);
93 if (!obj) {
94 DRM_ERROR("invalid handle %u at index %u\n", submit_bo.handle, i);
95 ret = -EINVAL;
96 goto out_unlock;
97 }
98
99 msm_obj = to_msm_bo(obj);
100
101 if (!list_empty(&msm_obj->submit_entry)) {
102 DRM_ERROR("handle %u at index %u already on submit list\n",
103 submit_bo.handle, i);
104 ret = -EINVAL;
105 goto out_unlock;
106 }
107
108 drm_gem_object_reference(obj);
109
110 submit->bos[i].obj = msm_obj;
111
112 list_add_tail(&msm_obj->submit_entry, &submit->bo_list);
113 }
114
115out_unlock:
116 submit->nr_bos = i;
117 spin_unlock(&file->table_lock);
118
119 return ret;
120}
121
122static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, int i)
123{
124 struct msm_gem_object *msm_obj = submit->bos[i].obj;
125
126 if (submit->bos[i].flags & BO_PINNED)
127 msm_gem_put_iova(&msm_obj->base, submit->gpu->id);
128
129 if (submit->bos[i].flags & BO_LOCKED)
130 ww_mutex_unlock(&msm_obj->resv->lock);
131
132 if (!(submit->bos[i].flags & BO_VALID))
133 submit->bos[i].iova = 0;
134
135 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED);
136}
137
138/* This is where we make sure all the bo's are reserved and pin'd: */
139static int submit_validate_objects(struct msm_gem_submit *submit)
140{
141 int contended, slow_locked = -1, i, ret = 0;
142
143retry:
144 submit->valid = true;
145
146 for (i = 0; i < submit->nr_bos; i++) {
147 struct msm_gem_object *msm_obj = submit->bos[i].obj;
148 uint32_t iova;
149
150 if (slow_locked == i)
151 slow_locked = -1;
152
153 contended = i;
154
155 if (!(submit->bos[i].flags & BO_LOCKED)) {
156 ret = ww_mutex_lock_interruptible(&msm_obj->resv->lock,
157 &submit->ticket);
158 if (ret)
159 goto fail;
160 submit->bos[i].flags |= BO_LOCKED;
161 }
162
163
164 /* if locking succeeded, pin bo: */
165 ret = msm_gem_get_iova_locked(&msm_obj->base,
166 submit->gpu->id, &iova);
167
168 /* this would break the logic in the fail path.. there is no
169 * reason for this to happen, but just to be on the safe side
170 * let's notice if this starts happening in the future:
171 */
172 WARN_ON(ret == -EDEADLK);
173
174 if (ret)
175 goto fail;
176
177 submit->bos[i].flags |= BO_PINNED;
178
179 if (iova == submit->bos[i].iova) {
180 submit->bos[i].flags |= BO_VALID;
181 } else {
182 submit->bos[i].iova = iova;
183 submit->bos[i].flags &= ~BO_VALID;
184 submit->valid = false;
185 }
186 }
187
188 ww_acquire_done(&submit->ticket);
189
190 return 0;
191
192fail:
193 for (; i >= 0; i--)
194 submit_unlock_unpin_bo(submit, i);
195
196 if (slow_locked > 0)
197 submit_unlock_unpin_bo(submit, slow_locked);
198
199 if (ret == -EDEADLK) {
200 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
201 /* we lost out in a seqno race, lock and retry.. */
202 ret = ww_mutex_lock_slow_interruptible(&msm_obj->resv->lock,
203 &submit->ticket);
204 if (!ret) {
205 submit->bos[contended].flags |= BO_LOCKED;
206 slow_locked = contended;
207 goto retry;
208 }
209 }
210
211 return ret;
212}
213
214static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
215 struct msm_gem_object **obj, uint32_t *iova, bool *valid)
216{
217 if (idx >= submit->nr_bos) {
218 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
219 idx, submit->nr_bos);
220 return -EINVAL;
221 }
222
223 if (obj)
224 *obj = submit->bos[idx].obj;
225 if (iova)
226 *iova = submit->bos[idx].iova;
227 if (valid)
228 *valid = !!(submit->bos[idx].flags & BO_VALID);
229
230 return 0;
231}
232
233/* process the reloc's and patch up the cmdstream as needed: */
234static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
235 uint32_t offset, uint32_t nr_relocs, uint64_t relocs)
236{
237 uint32_t i, last_offset = 0;
238 uint32_t *ptr;
239 int ret;
240
241 if (offset % 4) {
242 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
243 return -EINVAL;
244 }
245
246 /* For now, just map the entire thing. Eventually we probably
247 * to do it page-by-page, w/ kmap() if not vmap()d..
248 */
249 ptr = msm_gem_vaddr_locked(&obj->base);
250
251 if (IS_ERR(ptr)) {
252 ret = PTR_ERR(ptr);
253 DBG("failed to map: %d", ret);
254 return ret;
255 }
256
257 for (i = 0; i < nr_relocs; i++) {
258 struct drm_msm_gem_submit_reloc submit_reloc;
259 void __user *userptr =
260 to_user_ptr(relocs + (i * sizeof(submit_reloc)));
261 uint32_t iova, off;
262 bool valid;
263
264 ret = copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc));
265 if (ret)
266 return -EFAULT;
267
268 if (submit_reloc.submit_offset % 4) {
269 DRM_ERROR("non-aligned reloc offset: %u\n",
270 submit_reloc.submit_offset);
271 return -EINVAL;
272 }
273
274 /* offset in dwords: */
275 off = submit_reloc.submit_offset / 4;
276
277 if ((off >= (obj->base.size / 4)) ||
278 (off < last_offset)) {
279 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
280 return -EINVAL;
281 }
282
283 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
284 if (ret)
285 return ret;
286
287 if (valid)
288 continue;
289
290 iova += submit_reloc.reloc_offset;
291
292 if (submit_reloc.shift < 0)
293 iova >>= -submit_reloc.shift;
294 else
295 iova <<= submit_reloc.shift;
296
297 ptr[off] = iova | submit_reloc.or;
298
299 last_offset = off;
300 }
301
302 return 0;
303}
304
305static void submit_cleanup(struct msm_gem_submit *submit, bool fail)
306{
307 unsigned i;
308
309 for (i = 0; i < submit->nr_bos; i++) {
310 struct msm_gem_object *msm_obj = submit->bos[i].obj;
311 submit_unlock_unpin_bo(submit, i);
312 list_del_init(&msm_obj->submit_entry);
313 drm_gem_object_unreference(&msm_obj->base);
314 }
315
316 ww_acquire_fini(&submit->ticket);
317}
318
319int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
320 struct drm_file *file)
321{
322 struct msm_drm_private *priv = dev->dev_private;
323 struct drm_msm_gem_submit *args = data;
324 struct msm_file_private *ctx = file->driver_priv;
325 struct msm_gem_submit *submit;
326 struct msm_gpu *gpu = priv->gpu;
327 unsigned i;
328 int ret;
329
330 if (!gpu)
331 return -ENXIO;
332
333 /* for now, we just have 3d pipe.. eventually this would need to
334 * be more clever to dispatch to appropriate gpu module:
335 */
336 if (args->pipe != MSM_PIPE_3D0)
337 return -EINVAL;
338
339 if (args->nr_cmds > MAX_CMDS)
340 return -EINVAL;
341
342 submit = submit_create(dev, gpu, args->nr_bos);
343 if (!submit)
344 return -ENOMEM;
345
346 mutex_lock(&dev->struct_mutex);
347
348 ret = submit_lookup_objects(submit, args, file);
349 if (ret)
350 goto out;
351
352 ret = submit_validate_objects(submit);
353 if (ret)
354 goto out;
355
356 for (i = 0; i < args->nr_cmds; i++) {
357 struct drm_msm_gem_submit_cmd submit_cmd;
358 void __user *userptr =
359 to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
360 struct msm_gem_object *msm_obj;
361 uint32_t iova;
362
363 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
364 if (ret) {
365 ret = -EFAULT;
366 goto out;
367 }
368
369 /* validate input from userspace: */
370 switch (submit_cmd.type) {
371 case MSM_SUBMIT_CMD_BUF:
372 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
373 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
374 break;
375 default:
376 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
377 ret = -EINVAL;
378 goto out;
379 }
380
381 ret = submit_bo(submit, submit_cmd.submit_idx,
382 &msm_obj, &iova, NULL);
383 if (ret)
384 goto out;
385
386 if (submit_cmd.size % 4) {
387 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
388 submit_cmd.size);
389 ret = -EINVAL;
390 goto out;
391 }
392
393 if ((submit_cmd.size + submit_cmd.submit_offset) >=
394 msm_obj->base.size) {
395 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size);
396 ret = -EINVAL;
397 goto out;
398 }
399
400 submit->cmd[i].type = submit_cmd.type;
401 submit->cmd[i].size = submit_cmd.size / 4;
402 submit->cmd[i].iova = iova + submit_cmd.submit_offset;
403 submit->cmd[i].idx = submit_cmd.submit_idx;
404
405 if (submit->valid)
406 continue;
407
408 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset,
409 submit_cmd.nr_relocs, submit_cmd.relocs);
410 if (ret)
411 goto out;
412 }
413
414 submit->nr_cmds = i;
415
416 ret = msm_gpu_submit(gpu, submit, ctx);
417
418 args->fence = submit->fence;
419
420out:
421 submit_cleanup(submit, !!ret);
422 mutex_unlock(&dev->struct_mutex);
423 return ret;
424}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
5 */
6
7#include <linux/file.h>
8#include <linux/sync_file.h>
9#include <linux/uaccess.h>
10
11#include <drm/drm_drv.h>
12#include <drm/drm_file.h>
13#include <drm/drm_syncobj.h>
14
15#include "msm_drv.h"
16#include "msm_gpu.h"
17#include "msm_gem.h"
18#include "msm_gpu_trace.h"
19
20/* For userspace errors, use DRM_UT_DRIVER.. so that userspace can enable
21 * error msgs for debugging, but we don't spam dmesg by default
22 */
23#define SUBMIT_ERROR(submit, fmt, ...) \
24 DRM_DEV_DEBUG_DRIVER((submit)->dev->dev, fmt, ##__VA_ARGS__)
25
26/*
27 * Cmdstream submission:
28 */
29
30static struct msm_gem_submit *submit_create(struct drm_device *dev,
31 struct msm_gpu *gpu,
32 struct msm_gpu_submitqueue *queue, uint32_t nr_bos,
33 uint32_t nr_cmds)
34{
35 static atomic_t ident = ATOMIC_INIT(0);
36 struct msm_gem_submit *submit;
37 uint64_t sz;
38 int ret;
39
40 sz = struct_size(submit, bos, nr_bos) +
41 ((u64)nr_cmds * sizeof(submit->cmd[0]));
42
43 if (sz > SIZE_MAX)
44 return ERR_PTR(-ENOMEM);
45
46 submit = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
47 if (!submit)
48 return ERR_PTR(-ENOMEM);
49
50 submit->hw_fence = msm_fence_alloc();
51 if (IS_ERR(submit->hw_fence)) {
52 ret = PTR_ERR(submit->hw_fence);
53 kfree(submit);
54 return ERR_PTR(ret);
55 }
56
57 ret = drm_sched_job_init(&submit->base, queue->entity, 1, queue);
58 if (ret) {
59 kfree(submit->hw_fence);
60 kfree(submit);
61 return ERR_PTR(ret);
62 }
63
64 kref_init(&submit->ref);
65 submit->dev = dev;
66 submit->aspace = queue->ctx->aspace;
67 submit->gpu = gpu;
68 submit->cmd = (void *)&submit->bos[nr_bos];
69 submit->queue = queue;
70 submit->pid = get_pid(task_pid(current));
71 submit->ring = gpu->rb[queue->ring_nr];
72 submit->fault_dumped = false;
73
74 /* Get a unique identifier for the submission for logging purposes */
75 submit->ident = atomic_inc_return(&ident) - 1;
76
77 INIT_LIST_HEAD(&submit->node);
78
79 return submit;
80}
81
82void __msm_gem_submit_destroy(struct kref *kref)
83{
84 struct msm_gem_submit *submit =
85 container_of(kref, struct msm_gem_submit, ref);
86 unsigned i;
87
88 if (submit->fence_id) {
89 spin_lock(&submit->queue->idr_lock);
90 idr_remove(&submit->queue->fence_idr, submit->fence_id);
91 spin_unlock(&submit->queue->idr_lock);
92 }
93
94 dma_fence_put(submit->user_fence);
95
96 /*
97 * If the submit is freed before msm_job_run(), then hw_fence is
98 * just some pre-allocated memory, not a reference counted fence.
99 * Once the job runs and the hw_fence is initialized, it will
100 * have a refcount of at least one, since the submit holds a ref
101 * to the hw_fence.
102 */
103 if (kref_read(&submit->hw_fence->refcount) == 0) {
104 kfree(submit->hw_fence);
105 } else {
106 dma_fence_put(submit->hw_fence);
107 }
108
109 put_pid(submit->pid);
110 msm_submitqueue_put(submit->queue);
111
112 for (i = 0; i < submit->nr_cmds; i++)
113 kfree(submit->cmd[i].relocs);
114
115 kfree(submit);
116}
117
118static int submit_lookup_objects(struct msm_gem_submit *submit,
119 struct drm_msm_gem_submit *args, struct drm_file *file)
120{
121 unsigned i;
122 int ret = 0;
123
124 for (i = 0; i < args->nr_bos; i++) {
125 struct drm_msm_gem_submit_bo submit_bo;
126 void __user *userptr =
127 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
128
129 /* make sure we don't have garbage flags, in case we hit
130 * error path before flags is initialized:
131 */
132 submit->bos[i].flags = 0;
133
134 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
135 ret = -EFAULT;
136 i = 0;
137 goto out;
138 }
139
140/* at least one of READ and/or WRITE flags should be set: */
141#define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
142
143 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
144 !(submit_bo.flags & MANDATORY_FLAGS)) {
145 SUBMIT_ERROR(submit, "invalid flags: %x\n", submit_bo.flags);
146 ret = -EINVAL;
147 i = 0;
148 goto out;
149 }
150
151 submit->bos[i].handle = submit_bo.handle;
152 submit->bos[i].flags = submit_bo.flags;
153 }
154
155 spin_lock(&file->table_lock);
156
157 for (i = 0; i < args->nr_bos; i++) {
158 struct drm_gem_object *obj;
159
160 /* normally use drm_gem_object_lookup(), but for bulk lookup
161 * all under single table_lock just hit object_idr directly:
162 */
163 obj = idr_find(&file->object_idr, submit->bos[i].handle);
164 if (!obj) {
165 SUBMIT_ERROR(submit, "invalid handle %u at index %u\n", submit->bos[i].handle, i);
166 ret = -EINVAL;
167 goto out_unlock;
168 }
169
170 drm_gem_object_get(obj);
171
172 submit->bos[i].obj = obj;
173 }
174
175out_unlock:
176 spin_unlock(&file->table_lock);
177
178out:
179 submit->nr_bos = i;
180
181 return ret;
182}
183
184static int submit_lookup_cmds(struct msm_gem_submit *submit,
185 struct drm_msm_gem_submit *args, struct drm_file *file)
186{
187 unsigned i;
188 size_t sz;
189 int ret = 0;
190
191 for (i = 0; i < args->nr_cmds; i++) {
192 struct drm_msm_gem_submit_cmd submit_cmd;
193 void __user *userptr =
194 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
195
196 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
197 if (ret) {
198 ret = -EFAULT;
199 goto out;
200 }
201
202 /* validate input from userspace: */
203 switch (submit_cmd.type) {
204 case MSM_SUBMIT_CMD_BUF:
205 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
206 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
207 break;
208 default:
209 SUBMIT_ERROR(submit, "invalid type: %08x\n", submit_cmd.type);
210 return -EINVAL;
211 }
212
213 if (submit_cmd.size % 4) {
214 SUBMIT_ERROR(submit, "non-aligned cmdstream buffer size: %u\n",
215 submit_cmd.size);
216 ret = -EINVAL;
217 goto out;
218 }
219
220 submit->cmd[i].type = submit_cmd.type;
221 submit->cmd[i].size = submit_cmd.size / 4;
222 submit->cmd[i].offset = submit_cmd.submit_offset / 4;
223 submit->cmd[i].idx = submit_cmd.submit_idx;
224 submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
225
226 userptr = u64_to_user_ptr(submit_cmd.relocs);
227
228 sz = array_size(submit_cmd.nr_relocs,
229 sizeof(struct drm_msm_gem_submit_reloc));
230 /* check for overflow: */
231 if (sz == SIZE_MAX) {
232 ret = -ENOMEM;
233 goto out;
234 }
235 submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN);
236 if (!submit->cmd[i].relocs) {
237 ret = -ENOMEM;
238 goto out;
239 }
240 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
241 if (ret) {
242 ret = -EFAULT;
243 goto out;
244 }
245 }
246
247out:
248 return ret;
249}
250
251/* This is where we make sure all the bo's are reserved and pin'd: */
252static int submit_lock_objects(struct msm_gem_submit *submit)
253{
254 int ret;
255
256 drm_exec_init(&submit->exec, DRM_EXEC_INTERRUPTIBLE_WAIT, submit->nr_bos);
257
258 drm_exec_until_all_locked (&submit->exec) {
259 for (unsigned i = 0; i < submit->nr_bos; i++) {
260 struct drm_gem_object *obj = submit->bos[i].obj;
261 ret = drm_exec_prepare_obj(&submit->exec, obj, 1);
262 drm_exec_retry_on_contention(&submit->exec);
263 if (ret)
264 goto error;
265 }
266 }
267
268 return 0;
269
270error:
271 return ret;
272}
273
274static int submit_fence_sync(struct msm_gem_submit *submit)
275{
276 int i, ret = 0;
277
278 for (i = 0; i < submit->nr_bos; i++) {
279 struct drm_gem_object *obj = submit->bos[i].obj;
280 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
281
282 /* Otherwise userspace can ask for implicit sync to be
283 * disabled on specific buffers. This is useful for internal
284 * usermode driver managed buffers, suballocation, etc.
285 */
286 if (submit->bos[i].flags & MSM_SUBMIT_BO_NO_IMPLICIT)
287 continue;
288
289 ret = drm_sched_job_add_implicit_dependencies(&submit->base,
290 obj,
291 write);
292 if (ret)
293 break;
294 }
295
296 return ret;
297}
298
299static int submit_pin_objects(struct msm_gem_submit *submit)
300{
301 struct msm_drm_private *priv = submit->dev->dev_private;
302 int i, ret = 0;
303
304 for (i = 0; i < submit->nr_bos; i++) {
305 struct drm_gem_object *obj = submit->bos[i].obj;
306 struct msm_gem_vma *vma;
307
308 /* if locking succeeded, pin bo: */
309 vma = msm_gem_get_vma_locked(obj, submit->aspace);
310 if (IS_ERR(vma)) {
311 ret = PTR_ERR(vma);
312 break;
313 }
314
315 ret = msm_gem_pin_vma_locked(obj, vma);
316 if (ret)
317 break;
318
319 submit->bos[i].iova = vma->iova;
320 }
321
322 /*
323 * A second loop while holding the LRU lock (a) avoids acquiring/dropping
324 * the LRU lock for each individual bo, while (b) avoiding holding the
325 * LRU lock while calling msm_gem_pin_vma_locked() (which could trigger
326 * get_pages() which could trigger reclaim.. and if we held the LRU lock
327 * could trigger deadlock with the shrinker).
328 */
329 mutex_lock(&priv->lru.lock);
330 for (i = 0; i < submit->nr_bos; i++) {
331 msm_gem_pin_obj_locked(submit->bos[i].obj);
332 }
333 mutex_unlock(&priv->lru.lock);
334
335 submit->bos_pinned = true;
336
337 return ret;
338}
339
340static void submit_unpin_objects(struct msm_gem_submit *submit)
341{
342 if (!submit->bos_pinned)
343 return;
344
345 for (int i = 0; i < submit->nr_bos; i++) {
346 struct drm_gem_object *obj = submit->bos[i].obj;
347
348 msm_gem_unpin_locked(obj);
349 }
350
351 submit->bos_pinned = false;
352}
353
354static void submit_attach_object_fences(struct msm_gem_submit *submit)
355{
356 int i;
357
358 for (i = 0; i < submit->nr_bos; i++) {
359 struct drm_gem_object *obj = submit->bos[i].obj;
360
361 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
362 dma_resv_add_fence(obj->resv, submit->user_fence,
363 DMA_RESV_USAGE_WRITE);
364 else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
365 dma_resv_add_fence(obj->resv, submit->user_fence,
366 DMA_RESV_USAGE_READ);
367 }
368}
369
370static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
371 struct drm_gem_object **obj, uint64_t *iova)
372{
373 if (idx >= submit->nr_bos) {
374 SUBMIT_ERROR(submit, "invalid buffer index: %u (out of %u)\n",
375 idx, submit->nr_bos);
376 return -EINVAL;
377 }
378
379 if (obj)
380 *obj = submit->bos[idx].obj;
381 if (iova)
382 *iova = submit->bos[idx].iova;
383
384 return 0;
385}
386
387/* process the reloc's and patch up the cmdstream as needed: */
388static int submit_reloc(struct msm_gem_submit *submit, struct drm_gem_object *obj,
389 uint32_t offset, uint32_t nr_relocs, struct drm_msm_gem_submit_reloc *relocs)
390{
391 uint32_t i, last_offset = 0;
392 uint32_t *ptr;
393 int ret = 0;
394
395 if (offset % 4) {
396 SUBMIT_ERROR(submit, "non-aligned cmdstream buffer: %u\n", offset);
397 return -EINVAL;
398 }
399
400 /* For now, just map the entire thing. Eventually we probably
401 * to do it page-by-page, w/ kmap() if not vmap()d..
402 */
403 ptr = msm_gem_get_vaddr_locked(obj);
404
405 if (IS_ERR(ptr)) {
406 ret = PTR_ERR(ptr);
407 DBG("failed to map: %d", ret);
408 return ret;
409 }
410
411 for (i = 0; i < nr_relocs; i++) {
412 struct drm_msm_gem_submit_reloc submit_reloc = relocs[i];
413 uint32_t off;
414 uint64_t iova;
415
416 if (submit_reloc.submit_offset % 4) {
417 SUBMIT_ERROR(submit, "non-aligned reloc offset: %u\n",
418 submit_reloc.submit_offset);
419 ret = -EINVAL;
420 goto out;
421 }
422
423 /* offset in dwords: */
424 off = submit_reloc.submit_offset / 4;
425
426 if ((off >= (obj->size / 4)) ||
427 (off < last_offset)) {
428 SUBMIT_ERROR(submit, "invalid offset %u at reloc %u\n", off, i);
429 ret = -EINVAL;
430 goto out;
431 }
432
433 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova);
434 if (ret)
435 goto out;
436
437 iova += submit_reloc.reloc_offset;
438
439 if (submit_reloc.shift < 0)
440 iova >>= -submit_reloc.shift;
441 else
442 iova <<= submit_reloc.shift;
443
444 ptr[off] = iova | submit_reloc.or;
445
446 last_offset = off;
447 }
448
449out:
450 msm_gem_put_vaddr_locked(obj);
451
452 return ret;
453}
454
455/* Cleanup submit at end of ioctl. In the error case, this also drops
456 * references, unpins, and drops active refcnt. In the non-error case,
457 * this is done when the submit is retired.
458 */
459static void submit_cleanup(struct msm_gem_submit *submit, bool error)
460{
461 if (error) {
462 submit_unpin_objects(submit);
463 /* job wasn't enqueued to scheduler, so early retirement: */
464 msm_submit_retire(submit);
465 }
466
467 if (submit->exec.objects)
468 drm_exec_fini(&submit->exec);
469}
470
471void msm_submit_retire(struct msm_gem_submit *submit)
472{
473 int i;
474
475 for (i = 0; i < submit->nr_bos; i++) {
476 struct drm_gem_object *obj = submit->bos[i].obj;
477
478 drm_gem_object_put(obj);
479 }
480}
481
482struct msm_submit_post_dep {
483 struct drm_syncobj *syncobj;
484 uint64_t point;
485 struct dma_fence_chain *chain;
486};
487
488static struct drm_syncobj **msm_parse_deps(struct msm_gem_submit *submit,
489 struct drm_file *file,
490 uint64_t in_syncobjs_addr,
491 uint32_t nr_in_syncobjs,
492 size_t syncobj_stride)
493{
494 struct drm_syncobj **syncobjs = NULL;
495 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
496 int ret = 0;
497 uint32_t i, j;
498
499 syncobjs = kcalloc(nr_in_syncobjs, sizeof(*syncobjs),
500 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
501 if (!syncobjs)
502 return ERR_PTR(-ENOMEM);
503
504 for (i = 0; i < nr_in_syncobjs; ++i) {
505 uint64_t address = in_syncobjs_addr + i * syncobj_stride;
506
507 if (copy_from_user(&syncobj_desc,
508 u64_to_user_ptr(address),
509 min(syncobj_stride, sizeof(syncobj_desc)))) {
510 ret = -EFAULT;
511 break;
512 }
513
514 if (syncobj_desc.point &&
515 !drm_core_check_feature(submit->dev, DRIVER_SYNCOBJ_TIMELINE)) {
516 ret = -EOPNOTSUPP;
517 break;
518 }
519
520 if (syncobj_desc.flags & ~MSM_SUBMIT_SYNCOBJ_FLAGS) {
521 ret = -EINVAL;
522 break;
523 }
524
525 ret = drm_sched_job_add_syncobj_dependency(&submit->base, file,
526 syncobj_desc.handle, syncobj_desc.point);
527 if (ret)
528 break;
529
530 if (syncobj_desc.flags & MSM_SUBMIT_SYNCOBJ_RESET) {
531 syncobjs[i] =
532 drm_syncobj_find(file, syncobj_desc.handle);
533 if (!syncobjs[i]) {
534 ret = -EINVAL;
535 break;
536 }
537 }
538 }
539
540 if (ret) {
541 for (j = 0; j <= i; ++j) {
542 if (syncobjs[j])
543 drm_syncobj_put(syncobjs[j]);
544 }
545 kfree(syncobjs);
546 return ERR_PTR(ret);
547 }
548 return syncobjs;
549}
550
551static void msm_reset_syncobjs(struct drm_syncobj **syncobjs,
552 uint32_t nr_syncobjs)
553{
554 uint32_t i;
555
556 for (i = 0; syncobjs && i < nr_syncobjs; ++i) {
557 if (syncobjs[i])
558 drm_syncobj_replace_fence(syncobjs[i], NULL);
559 }
560}
561
562static struct msm_submit_post_dep *msm_parse_post_deps(struct drm_device *dev,
563 struct drm_file *file,
564 uint64_t syncobjs_addr,
565 uint32_t nr_syncobjs,
566 size_t syncobj_stride)
567{
568 struct msm_submit_post_dep *post_deps;
569 struct drm_msm_gem_submit_syncobj syncobj_desc = {0};
570 int ret = 0;
571 uint32_t i, j;
572
573 post_deps = kcalloc(nr_syncobjs, sizeof(*post_deps),
574 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
575 if (!post_deps)
576 return ERR_PTR(-ENOMEM);
577
578 for (i = 0; i < nr_syncobjs; ++i) {
579 uint64_t address = syncobjs_addr + i * syncobj_stride;
580
581 if (copy_from_user(&syncobj_desc,
582 u64_to_user_ptr(address),
583 min(syncobj_stride, sizeof(syncobj_desc)))) {
584 ret = -EFAULT;
585 break;
586 }
587
588 post_deps[i].point = syncobj_desc.point;
589
590 if (syncobj_desc.flags) {
591 ret = -EINVAL;
592 break;
593 }
594
595 if (syncobj_desc.point) {
596 if (!drm_core_check_feature(dev,
597 DRIVER_SYNCOBJ_TIMELINE)) {
598 ret = -EOPNOTSUPP;
599 break;
600 }
601
602 post_deps[i].chain = dma_fence_chain_alloc();
603 if (!post_deps[i].chain) {
604 ret = -ENOMEM;
605 break;
606 }
607 }
608
609 post_deps[i].syncobj =
610 drm_syncobj_find(file, syncobj_desc.handle);
611 if (!post_deps[i].syncobj) {
612 ret = -EINVAL;
613 break;
614 }
615 }
616
617 if (ret) {
618 for (j = 0; j <= i; ++j) {
619 dma_fence_chain_free(post_deps[j].chain);
620 if (post_deps[j].syncobj)
621 drm_syncobj_put(post_deps[j].syncobj);
622 }
623
624 kfree(post_deps);
625 return ERR_PTR(ret);
626 }
627
628 return post_deps;
629}
630
631static void msm_process_post_deps(struct msm_submit_post_dep *post_deps,
632 uint32_t count, struct dma_fence *fence)
633{
634 uint32_t i;
635
636 for (i = 0; post_deps && i < count; ++i) {
637 if (post_deps[i].chain) {
638 drm_syncobj_add_point(post_deps[i].syncobj,
639 post_deps[i].chain,
640 fence, post_deps[i].point);
641 post_deps[i].chain = NULL;
642 } else {
643 drm_syncobj_replace_fence(post_deps[i].syncobj,
644 fence);
645 }
646 }
647}
648
649int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
650 struct drm_file *file)
651{
652 struct msm_drm_private *priv = dev->dev_private;
653 struct drm_msm_gem_submit *args = data;
654 struct msm_file_private *ctx = file->driver_priv;
655 struct msm_gem_submit *submit = NULL;
656 struct msm_gpu *gpu = priv->gpu;
657 struct msm_gpu_submitqueue *queue;
658 struct msm_ringbuffer *ring;
659 struct msm_submit_post_dep *post_deps = NULL;
660 struct drm_syncobj **syncobjs_to_reset = NULL;
661 int out_fence_fd = -1;
662 unsigned i;
663 int ret;
664
665 if (!gpu)
666 return -ENXIO;
667
668 if (args->pad)
669 return -EINVAL;
670
671 if (unlikely(!ctx->aspace) && !capable(CAP_SYS_RAWIO)) {
672 DRM_ERROR_RATELIMITED("IOMMU support or CAP_SYS_RAWIO required!\n");
673 return -EPERM;
674 }
675
676 /* for now, we just have 3d pipe.. eventually this would need to
677 * be more clever to dispatch to appropriate gpu module:
678 */
679 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
680 return -EINVAL;
681
682 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
683 return -EINVAL;
684
685 if (args->flags & MSM_SUBMIT_SUDO) {
686 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
687 !capable(CAP_SYS_RAWIO))
688 return -EINVAL;
689 }
690
691 queue = msm_submitqueue_get(ctx, args->queueid);
692 if (!queue)
693 return -ENOENT;
694
695 ring = gpu->rb[queue->ring_nr];
696
697 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
698 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
699 if (out_fence_fd < 0) {
700 ret = out_fence_fd;
701 goto out_post_unlock;
702 }
703 }
704
705 submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds);
706 if (IS_ERR(submit)) {
707 ret = PTR_ERR(submit);
708 goto out_post_unlock;
709 }
710
711 trace_msm_gpu_submit(pid_nr(submit->pid), ring->id, submit->ident,
712 args->nr_bos, args->nr_cmds);
713
714 ret = mutex_lock_interruptible(&queue->lock);
715 if (ret)
716 goto out_post_unlock;
717
718 if (args->flags & MSM_SUBMIT_SUDO)
719 submit->in_rb = true;
720
721 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
722 struct dma_fence *in_fence;
723
724 in_fence = sync_file_get_fence(args->fence_fd);
725
726 if (!in_fence) {
727 ret = -EINVAL;
728 goto out_unlock;
729 }
730
731 ret = drm_sched_job_add_dependency(&submit->base, in_fence);
732 if (ret)
733 goto out_unlock;
734 }
735
736 if (args->flags & MSM_SUBMIT_SYNCOBJ_IN) {
737 syncobjs_to_reset = msm_parse_deps(submit, file,
738 args->in_syncobjs,
739 args->nr_in_syncobjs,
740 args->syncobj_stride);
741 if (IS_ERR(syncobjs_to_reset)) {
742 ret = PTR_ERR(syncobjs_to_reset);
743 goto out_unlock;
744 }
745 }
746
747 if (args->flags & MSM_SUBMIT_SYNCOBJ_OUT) {
748 post_deps = msm_parse_post_deps(dev, file,
749 args->out_syncobjs,
750 args->nr_out_syncobjs,
751 args->syncobj_stride);
752 if (IS_ERR(post_deps)) {
753 ret = PTR_ERR(post_deps);
754 goto out_unlock;
755 }
756 }
757
758 ret = submit_lookup_objects(submit, args, file);
759 if (ret)
760 goto out;
761
762 ret = submit_lookup_cmds(submit, args, file);
763 if (ret)
764 goto out;
765
766 /* copy_*_user while holding a ww ticket upsets lockdep */
767 ret = submit_lock_objects(submit);
768 if (ret)
769 goto out;
770
771 if (!(args->flags & MSM_SUBMIT_NO_IMPLICIT)) {
772 ret = submit_fence_sync(submit);
773 if (ret)
774 goto out;
775 }
776
777 ret = submit_pin_objects(submit);
778 if (ret)
779 goto out;
780
781 for (i = 0; i < args->nr_cmds; i++) {
782 struct drm_gem_object *obj;
783 uint64_t iova;
784
785 ret = submit_bo(submit, submit->cmd[i].idx, &obj, &iova);
786 if (ret)
787 goto out;
788
789 if (!submit->cmd[i].size ||
790 ((submit->cmd[i].size + submit->cmd[i].offset) >
791 obj->size / 4)) {
792 SUBMIT_ERROR(submit, "invalid cmdstream size: %u\n", submit->cmd[i].size * 4);
793 ret = -EINVAL;
794 goto out;
795 }
796
797 submit->cmd[i].iova = iova + (submit->cmd[i].offset * 4);
798
799 if (likely(!submit->cmd[i].nr_relocs))
800 continue;
801
802 if (!gpu->allow_relocs) {
803 SUBMIT_ERROR(submit, "relocs not allowed\n");
804 ret = -EINVAL;
805 goto out;
806 }
807
808 ret = submit_reloc(submit, obj, submit->cmd[i].offset * 4,
809 submit->cmd[i].nr_relocs, submit->cmd[i].relocs);
810 if (ret)
811 goto out;
812 }
813
814 submit->nr_cmds = i;
815
816 idr_preload(GFP_KERNEL);
817
818 spin_lock(&queue->idr_lock);
819
820 /*
821 * If using userspace provided seqno fence, validate that the id
822 * is available before arming sched job. Since access to fence_idr
823 * is serialized on the queue lock, the slot should be still avail
824 * after the job is armed
825 */
826 if ((args->flags & MSM_SUBMIT_FENCE_SN_IN) &&
827 (!args->fence || idr_find(&queue->fence_idr, args->fence))) {
828 spin_unlock(&queue->idr_lock);
829 idr_preload_end();
830 ret = -EINVAL;
831 goto out;
832 }
833
834 drm_sched_job_arm(&submit->base);
835
836 submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
837
838 if (args->flags & MSM_SUBMIT_FENCE_SN_IN) {
839 /*
840 * Userspace has assigned the seqno fence that it wants
841 * us to use. It is an error to pick a fence sequence
842 * number that is not available.
843 */
844 submit->fence_id = args->fence;
845 ret = idr_alloc_u32(&queue->fence_idr, submit->user_fence,
846 &submit->fence_id, submit->fence_id,
847 GFP_NOWAIT);
848 /*
849 * We've already validated that the fence_id slot is valid,
850 * so if idr_alloc_u32 failed, it is a kernel bug
851 */
852 WARN_ON(ret);
853 } else {
854 /*
855 * Allocate an id which can be used by WAIT_FENCE ioctl to map
856 * back to the underlying fence.
857 */
858 submit->fence_id = idr_alloc_cyclic(&queue->fence_idr,
859 submit->user_fence, 1,
860 INT_MAX, GFP_NOWAIT);
861 }
862
863 spin_unlock(&queue->idr_lock);
864 idr_preload_end();
865
866 if (submit->fence_id < 0) {
867 ret = submit->fence_id;
868 submit->fence_id = 0;
869 }
870
871 if (ret == 0 && args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
872 struct sync_file *sync_file = sync_file_create(submit->user_fence);
873 if (!sync_file) {
874 ret = -ENOMEM;
875 } else {
876 fd_install(out_fence_fd, sync_file->file);
877 args->fence_fd = out_fence_fd;
878 }
879 }
880
881 if (ret)
882 goto out;
883
884 submit_attach_object_fences(submit);
885
886 /* The scheduler owns a ref now: */
887 msm_gem_submit_get(submit);
888
889 msm_rd_dump_submit(priv->rd, submit, NULL);
890
891 drm_sched_entity_push_job(&submit->base);
892
893 args->fence = submit->fence_id;
894 queue->last_fence = submit->fence_id;
895
896 msm_reset_syncobjs(syncobjs_to_reset, args->nr_in_syncobjs);
897 msm_process_post_deps(post_deps, args->nr_out_syncobjs,
898 submit->user_fence);
899
900
901out:
902 submit_cleanup(submit, !!ret);
903out_unlock:
904 mutex_unlock(&queue->lock);
905out_post_unlock:
906 if (ret && (out_fence_fd >= 0))
907 put_unused_fd(out_fence_fd);
908
909 if (!IS_ERR_OR_NULL(submit)) {
910 msm_gem_submit_put(submit);
911 } else {
912 /*
913 * If the submit hasn't yet taken ownership of the queue
914 * then we need to drop the reference ourself:
915 */
916 msm_submitqueue_put(queue);
917 }
918 if (!IS_ERR_OR_NULL(post_deps)) {
919 for (i = 0; i < args->nr_out_syncobjs; ++i) {
920 kfree(post_deps[i].chain);
921 drm_syncobj_put(post_deps[i].syncobj);
922 }
923 kfree(post_deps);
924 }
925
926 if (!IS_ERR_OR_NULL(syncobjs_to_reset)) {
927 for (i = 0; i < args->nr_in_syncobjs; ++i) {
928 if (syncobjs_to_reset[i])
929 drm_syncobj_put(syncobjs_to_reset[i]);
930 }
931 kfree(syncobjs_to_reset);
932 }
933
934 return ret;
935}