xref: /openbmc/linux/drivers/gpu/drm/lima/lima_gem.c (revision 3d6af30f)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /* Copyright 2017-2019 Qiang Yu <yuq825@gmail.com> */
3 
4 #include <linux/mm.h>
5 #include <linux/iosys-map.h>
6 #include <linux/sync_file.h>
7 #include <linux/pagemap.h>
8 #include <linux/shmem_fs.h>
9 #include <linux/dma-mapping.h>
10 
11 #include <drm/drm_file.h>
12 #include <drm/drm_syncobj.h>
13 #include <drm/drm_utils.h>
14 
15 #include <drm/lima_drm.h>
16 
17 #include "lima_drv.h"
18 #include "lima_gem.h"
19 #include "lima_vm.h"
20 
21 int lima_heap_alloc(struct lima_bo *bo, struct lima_vm *vm)
22 {
23 	struct page **pages;
24 	struct address_space *mapping = bo->base.base.filp->f_mapping;
25 	struct device *dev = bo->base.base.dev->dev;
26 	size_t old_size = bo->heap_size;
27 	size_t new_size = bo->heap_size ? bo->heap_size * 2 :
28 		(lima_heap_init_nr_pages << PAGE_SHIFT);
29 	struct sg_table sgt;
30 	int i, ret;
31 
32 	if (bo->heap_size >= bo->base.base.size)
33 		return -ENOSPC;
34 
35 	new_size = min(new_size, bo->base.base.size);
36 
37 	dma_resv_lock(bo->base.base.resv, NULL);
38 
39 	if (bo->base.pages) {
40 		pages = bo->base.pages;
41 	} else {
42 		pages = kvmalloc_array(bo->base.base.size >> PAGE_SHIFT,
43 				       sizeof(*pages), GFP_KERNEL | __GFP_ZERO);
44 		if (!pages) {
45 			dma_resv_unlock(bo->base.base.resv);
46 			return -ENOMEM;
47 		}
48 
49 		bo->base.pages = pages;
50 		bo->base.pages_use_count = 1;
51 
52 		mapping_set_unevictable(mapping);
53 	}
54 
55 	for (i = old_size >> PAGE_SHIFT; i < new_size >> PAGE_SHIFT; i++) {
56 		struct page *page = shmem_read_mapping_page(mapping, i);
57 
58 		if (IS_ERR(page)) {
59 			dma_resv_unlock(bo->base.base.resv);
60 			return PTR_ERR(page);
61 		}
62 		pages[i] = page;
63 	}
64 
65 	dma_resv_unlock(bo->base.base.resv);
66 
67 	ret = sg_alloc_table_from_pages(&sgt, pages, i, 0,
68 					new_size, GFP_KERNEL);
69 	if (ret)
70 		return ret;
71 
72 	if (bo->base.sgt) {
73 		dma_unmap_sgtable(dev, bo->base.sgt, DMA_BIDIRECTIONAL, 0);
74 		sg_free_table(bo->base.sgt);
75 	} else {
76 		bo->base.sgt = kmalloc(sizeof(*bo->base.sgt), GFP_KERNEL);
77 		if (!bo->base.sgt) {
78 			ret = -ENOMEM;
79 			goto err_out0;
80 		}
81 	}
82 
83 	ret = dma_map_sgtable(dev, &sgt, DMA_BIDIRECTIONAL, 0);
84 	if (ret)
85 		goto err_out1;
86 
87 	*bo->base.sgt = sgt;
88 
89 	if (vm) {
90 		ret = lima_vm_map_bo(vm, bo, old_size >> PAGE_SHIFT);
91 		if (ret)
92 			goto err_out2;
93 	}
94 
95 	bo->heap_size = new_size;
96 	return 0;
97 
98 err_out2:
99 	dma_unmap_sgtable(dev, &sgt, DMA_BIDIRECTIONAL, 0);
100 err_out1:
101 	kfree(bo->base.sgt);
102 	bo->base.sgt = NULL;
103 err_out0:
104 	sg_free_table(&sgt);
105 	return ret;
106 }
107 
108 int lima_gem_create_handle(struct drm_device *dev, struct drm_file *file,
109 			   u32 size, u32 flags, u32 *handle)
110 {
111 	int err;
112 	gfp_t mask;
113 	struct drm_gem_shmem_object *shmem;
114 	struct drm_gem_object *obj;
115 	struct lima_bo *bo;
116 	bool is_heap = flags & LIMA_BO_FLAG_HEAP;
117 
118 	shmem = drm_gem_shmem_create(dev, size);
119 	if (IS_ERR(shmem))
120 		return PTR_ERR(shmem);
121 
122 	obj = &shmem->base;
123 
124 	/* Mali Utgard GPU can only support 32bit address space */
125 	mask = mapping_gfp_mask(obj->filp->f_mapping);
126 	mask &= ~__GFP_HIGHMEM;
127 	mask |= __GFP_DMA32;
128 	mapping_set_gfp_mask(obj->filp->f_mapping, mask);
129 
130 	if (is_heap) {
131 		bo = to_lima_bo(obj);
132 		err = lima_heap_alloc(bo, NULL);
133 		if (err)
134 			goto out;
135 	} else {
136 		struct sg_table *sgt = drm_gem_shmem_get_pages_sgt(shmem);
137 
138 		if (IS_ERR(sgt)) {
139 			err = PTR_ERR(sgt);
140 			goto out;
141 		}
142 	}
143 
144 	err = drm_gem_handle_create(file, obj, handle);
145 
146 out:
147 	/* drop reference from allocate - handle holds it now */
148 	drm_gem_object_put(obj);
149 
150 	return err;
151 }
152 
153 static void lima_gem_free_object(struct drm_gem_object *obj)
154 {
155 	struct lima_bo *bo = to_lima_bo(obj);
156 
157 	if (!list_empty(&bo->va))
158 		dev_err(obj->dev->dev, "lima gem free bo still has va\n");
159 
160 	drm_gem_shmem_free(&bo->base);
161 }
162 
163 static int lima_gem_object_open(struct drm_gem_object *obj, struct drm_file *file)
164 {
165 	struct lima_bo *bo = to_lima_bo(obj);
166 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
167 	struct lima_vm *vm = priv->vm;
168 
169 	return lima_vm_bo_add(vm, bo, true);
170 }
171 
172 static void lima_gem_object_close(struct drm_gem_object *obj, struct drm_file *file)
173 {
174 	struct lima_bo *bo = to_lima_bo(obj);
175 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
176 	struct lima_vm *vm = priv->vm;
177 
178 	lima_vm_bo_del(vm, bo);
179 }
180 
181 static int lima_gem_pin(struct drm_gem_object *obj)
182 {
183 	struct lima_bo *bo = to_lima_bo(obj);
184 
185 	if (bo->heap_size)
186 		return -EINVAL;
187 
188 	return drm_gem_shmem_pin(&bo->base);
189 }
190 
191 static int lima_gem_vmap(struct drm_gem_object *obj, struct iosys_map *map)
192 {
193 	struct lima_bo *bo = to_lima_bo(obj);
194 
195 	if (bo->heap_size)
196 		return -EINVAL;
197 
198 	return drm_gem_shmem_vmap(&bo->base, map);
199 }
200 
201 static int lima_gem_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
202 {
203 	struct lima_bo *bo = to_lima_bo(obj);
204 
205 	if (bo->heap_size)
206 		return -EINVAL;
207 
208 	return drm_gem_shmem_mmap(&bo->base, vma);
209 }
210 
211 static const struct drm_gem_object_funcs lima_gem_funcs = {
212 	.free = lima_gem_free_object,
213 	.open = lima_gem_object_open,
214 	.close = lima_gem_object_close,
215 	.print_info = drm_gem_shmem_object_print_info,
216 	.pin = lima_gem_pin,
217 	.unpin = drm_gem_shmem_object_unpin,
218 	.get_sg_table = drm_gem_shmem_object_get_sg_table,
219 	.vmap = lima_gem_vmap,
220 	.vunmap = drm_gem_shmem_object_vunmap,
221 	.mmap = lima_gem_mmap,
222 	.vm_ops = &drm_gem_shmem_vm_ops,
223 };
224 
225 struct drm_gem_object *lima_gem_create_object(struct drm_device *dev, size_t size)
226 {
227 	struct lima_bo *bo;
228 
229 	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
230 	if (!bo)
231 		return ERR_PTR(-ENOMEM);
232 
233 	mutex_init(&bo->lock);
234 	INIT_LIST_HEAD(&bo->va);
235 	bo->base.map_wc = true;
236 	bo->base.base.funcs = &lima_gem_funcs;
237 
238 	return &bo->base.base;
239 }
240 
241 int lima_gem_get_info(struct drm_file *file, u32 handle, u32 *va, u64 *offset)
242 {
243 	struct drm_gem_object *obj;
244 	struct lima_bo *bo;
245 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
246 	struct lima_vm *vm = priv->vm;
247 
248 	obj = drm_gem_object_lookup(file, handle);
249 	if (!obj)
250 		return -ENOENT;
251 
252 	bo = to_lima_bo(obj);
253 
254 	*va = lima_vm_get_va(vm, bo);
255 
256 	*offset = drm_vma_node_offset_addr(&obj->vma_node);
257 
258 	drm_gem_object_put(obj);
259 	return 0;
260 }
261 
262 static int lima_gem_sync_bo(struct lima_sched_task *task, struct lima_bo *bo,
263 			    bool write, bool explicit)
264 {
265 	int err;
266 
267 	err = dma_resv_reserve_fences(lima_bo_resv(bo), 1);
268 	if (err)
269 		return err;
270 
271 	/* explicit sync use user passed dep fence */
272 	if (explicit)
273 		return 0;
274 
275 	return drm_sched_job_add_implicit_dependencies(&task->base,
276 						       &bo->base.base,
277 						       write);
278 }
279 
280 static int lima_gem_add_deps(struct drm_file *file, struct lima_submit *submit)
281 {
282 	int i, err;
283 
284 	for (i = 0; i < ARRAY_SIZE(submit->in_sync); i++) {
285 		if (!submit->in_sync[i])
286 			continue;
287 
288 		err = drm_sched_job_add_syncobj_dependency(&submit->task->base, file,
289 							   submit->in_sync[i], 0);
290 		if (err)
291 			return err;
292 	}
293 
294 	return 0;
295 }
296 
297 int lima_gem_submit(struct drm_file *file, struct lima_submit *submit)
298 {
299 	int i, err = 0;
300 	struct ww_acquire_ctx ctx;
301 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
302 	struct lima_vm *vm = priv->vm;
303 	struct drm_syncobj *out_sync = NULL;
304 	struct dma_fence *fence;
305 	struct lima_bo **bos = submit->lbos;
306 
307 	if (submit->out_sync) {
308 		out_sync = drm_syncobj_find(file, submit->out_sync);
309 		if (!out_sync)
310 			return -ENOENT;
311 	}
312 
313 	for (i = 0; i < submit->nr_bos; i++) {
314 		struct drm_gem_object *obj;
315 		struct lima_bo *bo;
316 
317 		obj = drm_gem_object_lookup(file, submit->bos[i].handle);
318 		if (!obj) {
319 			err = -ENOENT;
320 			goto err_out0;
321 		}
322 
323 		bo = to_lima_bo(obj);
324 
325 		/* increase refcnt of gpu va map to prevent unmapped when executing,
326 		 * will be decreased when task done
327 		 */
328 		err = lima_vm_bo_add(vm, bo, false);
329 		if (err) {
330 			drm_gem_object_put(obj);
331 			goto err_out0;
332 		}
333 
334 		bos[i] = bo;
335 	}
336 
337 	err = drm_gem_lock_reservations((struct drm_gem_object **)bos,
338 					submit->nr_bos, &ctx);
339 	if (err)
340 		goto err_out0;
341 
342 	err = lima_sched_task_init(
343 		submit->task, submit->ctx->context + submit->pipe,
344 		bos, submit->nr_bos, vm);
345 	if (err)
346 		goto err_out1;
347 
348 	err = lima_gem_add_deps(file, submit);
349 	if (err)
350 		goto err_out2;
351 
352 	for (i = 0; i < submit->nr_bos; i++) {
353 		err = lima_gem_sync_bo(
354 			submit->task, bos[i],
355 			submit->bos[i].flags & LIMA_SUBMIT_BO_WRITE,
356 			submit->flags & LIMA_SUBMIT_FLAG_EXPLICIT_FENCE);
357 		if (err)
358 			goto err_out2;
359 	}
360 
361 	fence = lima_sched_context_queue_task(submit->task);
362 
363 	for (i = 0; i < submit->nr_bos; i++) {
364 		dma_resv_add_fence(lima_bo_resv(bos[i]), fence,
365 				   submit->bos[i].flags & LIMA_SUBMIT_BO_WRITE ?
366 				   DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_READ);
367 	}
368 
369 	drm_gem_unlock_reservations((struct drm_gem_object **)bos,
370 				    submit->nr_bos, &ctx);
371 
372 	for (i = 0; i < submit->nr_bos; i++)
373 		drm_gem_object_put(&bos[i]->base.base);
374 
375 	if (out_sync) {
376 		drm_syncobj_replace_fence(out_sync, fence);
377 		drm_syncobj_put(out_sync);
378 	}
379 
380 	dma_fence_put(fence);
381 
382 	return 0;
383 
384 err_out2:
385 	lima_sched_task_fini(submit->task);
386 err_out1:
387 	drm_gem_unlock_reservations((struct drm_gem_object **)bos,
388 				    submit->nr_bos, &ctx);
389 err_out0:
390 	for (i = 0; i < submit->nr_bos; i++) {
391 		if (!bos[i])
392 			break;
393 		lima_vm_bo_del(vm, bos[i]);
394 		drm_gem_object_put(&bos[i]->base.base);
395 	}
396 	if (out_sync)
397 		drm_syncobj_put(out_sync);
398 	return err;
399 }
400 
401 int lima_gem_wait(struct drm_file *file, u32 handle, u32 op, s64 timeout_ns)
402 {
403 	bool write = op & LIMA_GEM_WAIT_WRITE;
404 	long ret, timeout;
405 
406 	if (!op)
407 		return 0;
408 
409 	timeout = drm_timeout_abs_to_jiffies(timeout_ns);
410 
411 	ret = drm_gem_dma_resv_wait(file, handle, write, timeout);
412 	if (ret == -ETIME)
413 		ret = timeout ? -ETIMEDOUT : -EBUSY;
414 
415 	return ret;
416 }
417