xref: /openbmc/linux/drivers/gpu/drm/lima/lima_gem.c (revision ffcdf473)
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 	mutex_lock(&bo->base.pages_lock);
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 			mutex_unlock(&bo->base.pages_lock);
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 			mutex_unlock(&bo->base.pages_lock);
60 			return PTR_ERR(page);
61 		}
62 		pages[i] = page;
63 	}
64 
65 	mutex_unlock(&bo->base.pages_lock);
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 			sg_free_table(&sgt);
79 			return -ENOMEM;
80 		}
81 	}
82 
83 	ret = dma_map_sgtable(dev, &sgt, DMA_BIDIRECTIONAL, 0);
84 	if (ret) {
85 		sg_free_table(&sgt);
86 		kfree(bo->base.sgt);
87 		bo->base.sgt = NULL;
88 		return ret;
89 	}
90 
91 	*bo->base.sgt = sgt;
92 
93 	if (vm) {
94 		ret = lima_vm_map_bo(vm, bo, old_size >> PAGE_SHIFT);
95 		if (ret)
96 			return ret;
97 	}
98 
99 	bo->heap_size = new_size;
100 	return 0;
101 }
102 
103 int lima_gem_create_handle(struct drm_device *dev, struct drm_file *file,
104 			   u32 size, u32 flags, u32 *handle)
105 {
106 	int err;
107 	gfp_t mask;
108 	struct drm_gem_shmem_object *shmem;
109 	struct drm_gem_object *obj;
110 	struct lima_bo *bo;
111 	bool is_heap = flags & LIMA_BO_FLAG_HEAP;
112 
113 	shmem = drm_gem_shmem_create(dev, size);
114 	if (IS_ERR(shmem))
115 		return PTR_ERR(shmem);
116 
117 	obj = &shmem->base;
118 
119 	/* Mali Utgard GPU can only support 32bit address space */
120 	mask = mapping_gfp_mask(obj->filp->f_mapping);
121 	mask &= ~__GFP_HIGHMEM;
122 	mask |= __GFP_DMA32;
123 	mapping_set_gfp_mask(obj->filp->f_mapping, mask);
124 
125 	if (is_heap) {
126 		bo = to_lima_bo(obj);
127 		err = lima_heap_alloc(bo, NULL);
128 		if (err)
129 			goto out;
130 	} else {
131 		struct sg_table *sgt = drm_gem_shmem_get_pages_sgt(shmem);
132 
133 		if (IS_ERR(sgt)) {
134 			err = PTR_ERR(sgt);
135 			goto out;
136 		}
137 	}
138 
139 	err = drm_gem_handle_create(file, obj, handle);
140 
141 out:
142 	/* drop reference from allocate - handle holds it now */
143 	drm_gem_object_put(obj);
144 
145 	return err;
146 }
147 
148 static void lima_gem_free_object(struct drm_gem_object *obj)
149 {
150 	struct lima_bo *bo = to_lima_bo(obj);
151 
152 	if (!list_empty(&bo->va))
153 		dev_err(obj->dev->dev, "lima gem free bo still has va\n");
154 
155 	drm_gem_shmem_free(&bo->base);
156 }
157 
158 static int lima_gem_object_open(struct drm_gem_object *obj, struct drm_file *file)
159 {
160 	struct lima_bo *bo = to_lima_bo(obj);
161 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
162 	struct lima_vm *vm = priv->vm;
163 
164 	return lima_vm_bo_add(vm, bo, true);
165 }
166 
167 static void lima_gem_object_close(struct drm_gem_object *obj, struct drm_file *file)
168 {
169 	struct lima_bo *bo = to_lima_bo(obj);
170 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
171 	struct lima_vm *vm = priv->vm;
172 
173 	lima_vm_bo_del(vm, bo);
174 }
175 
176 static int lima_gem_pin(struct drm_gem_object *obj)
177 {
178 	struct lima_bo *bo = to_lima_bo(obj);
179 
180 	if (bo->heap_size)
181 		return -EINVAL;
182 
183 	return drm_gem_shmem_pin(&bo->base);
184 }
185 
186 static int lima_gem_vmap(struct drm_gem_object *obj, struct iosys_map *map)
187 {
188 	struct lima_bo *bo = to_lima_bo(obj);
189 
190 	if (bo->heap_size)
191 		return -EINVAL;
192 
193 	return drm_gem_shmem_vmap(&bo->base, map);
194 }
195 
196 static int lima_gem_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
197 {
198 	struct lima_bo *bo = to_lima_bo(obj);
199 
200 	if (bo->heap_size)
201 		return -EINVAL;
202 
203 	return drm_gem_shmem_mmap(&bo->base, vma);
204 }
205 
206 static const struct drm_gem_object_funcs lima_gem_funcs = {
207 	.free = lima_gem_free_object,
208 	.open = lima_gem_object_open,
209 	.close = lima_gem_object_close,
210 	.print_info = drm_gem_shmem_object_print_info,
211 	.pin = lima_gem_pin,
212 	.unpin = drm_gem_shmem_object_unpin,
213 	.get_sg_table = drm_gem_shmem_object_get_sg_table,
214 	.vmap = lima_gem_vmap,
215 	.vunmap = drm_gem_shmem_object_vunmap,
216 	.mmap = lima_gem_mmap,
217 	.vm_ops = &drm_gem_shmem_vm_ops,
218 };
219 
220 struct drm_gem_object *lima_gem_create_object(struct drm_device *dev, size_t size)
221 {
222 	struct lima_bo *bo;
223 
224 	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
225 	if (!bo)
226 		return ERR_PTR(-ENOMEM);
227 
228 	mutex_init(&bo->lock);
229 	INIT_LIST_HEAD(&bo->va);
230 	bo->base.map_wc = true;
231 	bo->base.base.funcs = &lima_gem_funcs;
232 
233 	return &bo->base.base;
234 }
235 
236 int lima_gem_get_info(struct drm_file *file, u32 handle, u32 *va, u64 *offset)
237 {
238 	struct drm_gem_object *obj;
239 	struct lima_bo *bo;
240 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
241 	struct lima_vm *vm = priv->vm;
242 
243 	obj = drm_gem_object_lookup(file, handle);
244 	if (!obj)
245 		return -ENOENT;
246 
247 	bo = to_lima_bo(obj);
248 
249 	*va = lima_vm_get_va(vm, bo);
250 
251 	*offset = drm_vma_node_offset_addr(&obj->vma_node);
252 
253 	drm_gem_object_put(obj);
254 	return 0;
255 }
256 
257 static int lima_gem_sync_bo(struct lima_sched_task *task, struct lima_bo *bo,
258 			    bool write, bool explicit)
259 {
260 	int err;
261 
262 	err = dma_resv_reserve_fences(lima_bo_resv(bo), 1);
263 	if (err)
264 		return err;
265 
266 	/* explicit sync use user passed dep fence */
267 	if (explicit)
268 		return 0;
269 
270 	return drm_sched_job_add_implicit_dependencies(&task->base,
271 						       &bo->base.base,
272 						       write);
273 }
274 
275 static int lima_gem_add_deps(struct drm_file *file, struct lima_submit *submit)
276 {
277 	int i, err;
278 
279 	for (i = 0; i < ARRAY_SIZE(submit->in_sync); i++) {
280 		if (!submit->in_sync[i])
281 			continue;
282 
283 		err = drm_sched_job_add_syncobj_dependency(&submit->task->base, file,
284 							   submit->in_sync[i], 0);
285 		if (err)
286 			return err;
287 	}
288 
289 	return 0;
290 }
291 
292 int lima_gem_submit(struct drm_file *file, struct lima_submit *submit)
293 {
294 	int i, err = 0;
295 	struct ww_acquire_ctx ctx;
296 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
297 	struct lima_vm *vm = priv->vm;
298 	struct drm_syncobj *out_sync = NULL;
299 	struct dma_fence *fence;
300 	struct lima_bo **bos = submit->lbos;
301 
302 	if (submit->out_sync) {
303 		out_sync = drm_syncobj_find(file, submit->out_sync);
304 		if (!out_sync)
305 			return -ENOENT;
306 	}
307 
308 	for (i = 0; i < submit->nr_bos; i++) {
309 		struct drm_gem_object *obj;
310 		struct lima_bo *bo;
311 
312 		obj = drm_gem_object_lookup(file, submit->bos[i].handle);
313 		if (!obj) {
314 			err = -ENOENT;
315 			goto err_out0;
316 		}
317 
318 		bo = to_lima_bo(obj);
319 
320 		/* increase refcnt of gpu va map to prevent unmapped when executing,
321 		 * will be decreased when task done
322 		 */
323 		err = lima_vm_bo_add(vm, bo, false);
324 		if (err) {
325 			drm_gem_object_put(obj);
326 			goto err_out0;
327 		}
328 
329 		bos[i] = bo;
330 	}
331 
332 	err = drm_gem_lock_reservations((struct drm_gem_object **)bos,
333 					submit->nr_bos, &ctx);
334 	if (err)
335 		goto err_out0;
336 
337 	err = lima_sched_task_init(
338 		submit->task, submit->ctx->context + submit->pipe,
339 		bos, submit->nr_bos, vm);
340 	if (err)
341 		goto err_out1;
342 
343 	err = lima_gem_add_deps(file, submit);
344 	if (err)
345 		goto err_out2;
346 
347 	for (i = 0; i < submit->nr_bos; i++) {
348 		err = lima_gem_sync_bo(
349 			submit->task, bos[i],
350 			submit->bos[i].flags & LIMA_SUBMIT_BO_WRITE,
351 			submit->flags & LIMA_SUBMIT_FLAG_EXPLICIT_FENCE);
352 		if (err)
353 			goto err_out2;
354 	}
355 
356 	fence = lima_sched_context_queue_task(submit->task);
357 
358 	for (i = 0; i < submit->nr_bos; i++) {
359 		dma_resv_add_fence(lima_bo_resv(bos[i]), fence,
360 				   submit->bos[i].flags & LIMA_SUBMIT_BO_WRITE ?
361 				   DMA_RESV_USAGE_WRITE : DMA_RESV_USAGE_READ);
362 	}
363 
364 	drm_gem_unlock_reservations((struct drm_gem_object **)bos,
365 				    submit->nr_bos, &ctx);
366 
367 	for (i = 0; i < submit->nr_bos; i++)
368 		drm_gem_object_put(&bos[i]->base.base);
369 
370 	if (out_sync) {
371 		drm_syncobj_replace_fence(out_sync, fence);
372 		drm_syncobj_put(out_sync);
373 	}
374 
375 	dma_fence_put(fence);
376 
377 	return 0;
378 
379 err_out2:
380 	lima_sched_task_fini(submit->task);
381 err_out1:
382 	drm_gem_unlock_reservations((struct drm_gem_object **)bos,
383 				    submit->nr_bos, &ctx);
384 err_out0:
385 	for (i = 0; i < submit->nr_bos; i++) {
386 		if (!bos[i])
387 			break;
388 		lima_vm_bo_del(vm, bos[i]);
389 		drm_gem_object_put(&bos[i]->base.base);
390 	}
391 	if (out_sync)
392 		drm_syncobj_put(out_sync);
393 	return err;
394 }
395 
396 int lima_gem_wait(struct drm_file *file, u32 handle, u32 op, s64 timeout_ns)
397 {
398 	bool write = op & LIMA_GEM_WAIT_WRITE;
399 	long ret, timeout;
400 
401 	if (!op)
402 		return 0;
403 
404 	timeout = drm_timeout_abs_to_jiffies(timeout_ns);
405 
406 	ret = drm_gem_dma_resv_wait(file, handle, write, timeout);
407 	if (ret == -ETIME)
408 		ret = timeout ? -ETIMEDOUT : -EBUSY;
409 
410 	return ret;
411 }
412