xref: /openbmc/linux/drivers/gpu/drm/lima/lima_gem.c (revision e953aeaa)
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/sync_file.h>
6 #include <linux/pagemap.h>
7 #include <linux/shmem_fs.h>
8 #include <linux/dma-mapping.h>
9 
10 #include <drm/drm_file.h>
11 #include <drm/drm_syncobj.h>
12 #include <drm/drm_utils.h>
13 
14 #include <drm/lima_drm.h>
15 
16 #include "lima_drv.h"
17 #include "lima_gem.h"
18 #include "lima_vm.h"
19 
20 int lima_heap_alloc(struct lima_bo *bo, struct lima_vm *vm)
21 {
22 	struct page **pages;
23 	struct address_space *mapping = bo->base.base.filp->f_mapping;
24 	struct device *dev = bo->base.base.dev->dev;
25 	size_t old_size = bo->heap_size;
26 	size_t new_size = bo->heap_size ? bo->heap_size * 2 :
27 		(lima_heap_init_nr_pages << PAGE_SHIFT);
28 	struct sg_table sgt;
29 	int i, ret;
30 
31 	if (bo->heap_size >= bo->base.base.size)
32 		return -ENOSPC;
33 
34 	new_size = min(new_size, bo->base.base.size);
35 
36 	mutex_lock(&bo->base.pages_lock);
37 
38 	if (bo->base.pages) {
39 		pages = bo->base.pages;
40 	} else {
41 		pages = kvmalloc_array(bo->base.base.size >> PAGE_SHIFT,
42 				       sizeof(*pages), GFP_KERNEL | __GFP_ZERO);
43 		if (!pages) {
44 			mutex_unlock(&bo->base.pages_lock);
45 			return -ENOMEM;
46 		}
47 
48 		bo->base.pages = pages;
49 		bo->base.pages_use_count = 1;
50 
51 		mapping_set_unevictable(mapping);
52 	}
53 
54 	for (i = old_size >> PAGE_SHIFT; i < new_size >> PAGE_SHIFT; i++) {
55 		struct page *page = shmem_read_mapping_page(mapping, i);
56 
57 		if (IS_ERR(page)) {
58 			mutex_unlock(&bo->base.pages_lock);
59 			return PTR_ERR(page);
60 		}
61 		pages[i] = page;
62 	}
63 
64 	mutex_unlock(&bo->base.pages_lock);
65 
66 	ret = sg_alloc_table_from_pages(&sgt, pages, i, 0,
67 					new_size, GFP_KERNEL);
68 	if (ret)
69 		return ret;
70 
71 	if (bo->base.sgt) {
72 		dma_unmap_sg(dev, bo->base.sgt->sgl,
73 			     bo->base.sgt->nents, DMA_BIDIRECTIONAL);
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 	dma_map_sg(dev, sgt.sgl, sgt.nents, DMA_BIDIRECTIONAL);
84 
85 	*bo->base.sgt = sgt;
86 
87 	if (vm) {
88 		ret = lima_vm_map_bo(vm, bo, old_size >> PAGE_SHIFT);
89 		if (ret)
90 			return ret;
91 	}
92 
93 	bo->heap_size = new_size;
94 	return 0;
95 }
96 
97 int lima_gem_create_handle(struct drm_device *dev, struct drm_file *file,
98 			   u32 size, u32 flags, u32 *handle)
99 {
100 	int err;
101 	gfp_t mask;
102 	struct drm_gem_shmem_object *shmem;
103 	struct drm_gem_object *obj;
104 	struct lima_bo *bo;
105 	bool is_heap = flags & LIMA_BO_FLAG_HEAP;
106 
107 	shmem = drm_gem_shmem_create(dev, size);
108 	if (IS_ERR(shmem))
109 		return PTR_ERR(shmem);
110 
111 	obj = &shmem->base;
112 
113 	/* Mali Utgard GPU can only support 32bit address space */
114 	mask = mapping_gfp_mask(obj->filp->f_mapping);
115 	mask &= ~__GFP_HIGHMEM;
116 	mask |= __GFP_DMA32;
117 	mapping_set_gfp_mask(obj->filp->f_mapping, mask);
118 
119 	if (is_heap) {
120 		bo = to_lima_bo(obj);
121 		err = lima_heap_alloc(bo, NULL);
122 		if (err)
123 			goto out;
124 	} else {
125 		struct sg_table *sgt = drm_gem_shmem_get_pages_sgt(obj);
126 
127 		if (IS_ERR(sgt)) {
128 			err = PTR_ERR(sgt);
129 			goto out;
130 		}
131 	}
132 
133 	err = drm_gem_handle_create(file, obj, handle);
134 
135 out:
136 	/* drop reference from allocate - handle holds it now */
137 	drm_gem_object_put(obj);
138 
139 	return err;
140 }
141 
142 static void lima_gem_free_object(struct drm_gem_object *obj)
143 {
144 	struct lima_bo *bo = to_lima_bo(obj);
145 
146 	if (!list_empty(&bo->va))
147 		dev_err(obj->dev->dev, "lima gem free bo still has va\n");
148 
149 	drm_gem_shmem_free_object(obj);
150 }
151 
152 static int lima_gem_object_open(struct drm_gem_object *obj, struct drm_file *file)
153 {
154 	struct lima_bo *bo = to_lima_bo(obj);
155 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
156 	struct lima_vm *vm = priv->vm;
157 
158 	return lima_vm_bo_add(vm, bo, true);
159 }
160 
161 static void lima_gem_object_close(struct drm_gem_object *obj, struct drm_file *file)
162 {
163 	struct lima_bo *bo = to_lima_bo(obj);
164 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
165 	struct lima_vm *vm = priv->vm;
166 
167 	lima_vm_bo_del(vm, bo);
168 }
169 
170 static int lima_gem_pin(struct drm_gem_object *obj)
171 {
172 	struct lima_bo *bo = to_lima_bo(obj);
173 
174 	if (bo->heap_size)
175 		return -EINVAL;
176 
177 	return drm_gem_shmem_pin(obj);
178 }
179 
180 static void *lima_gem_vmap(struct drm_gem_object *obj)
181 {
182 	struct lima_bo *bo = to_lima_bo(obj);
183 
184 	if (bo->heap_size)
185 		return ERR_PTR(-EINVAL);
186 
187 	return drm_gem_shmem_vmap(obj);
188 }
189 
190 static int lima_gem_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
191 {
192 	struct lima_bo *bo = to_lima_bo(obj);
193 
194 	if (bo->heap_size)
195 		return -EINVAL;
196 
197 	return drm_gem_shmem_mmap(obj, vma);
198 }
199 
200 static const struct drm_gem_object_funcs lima_gem_funcs = {
201 	.free = lima_gem_free_object,
202 	.open = lima_gem_object_open,
203 	.close = lima_gem_object_close,
204 	.print_info = drm_gem_shmem_print_info,
205 	.pin = lima_gem_pin,
206 	.unpin = drm_gem_shmem_unpin,
207 	.get_sg_table = drm_gem_shmem_get_sg_table,
208 	.vmap = lima_gem_vmap,
209 	.vunmap = drm_gem_shmem_vunmap,
210 	.mmap = lima_gem_mmap,
211 };
212 
213 struct drm_gem_object *lima_gem_create_object(struct drm_device *dev, size_t size)
214 {
215 	struct lima_bo *bo;
216 
217 	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
218 	if (!bo)
219 		return NULL;
220 
221 	mutex_init(&bo->lock);
222 	INIT_LIST_HEAD(&bo->va);
223 
224 	bo->base.base.funcs = &lima_gem_funcs;
225 
226 	return &bo->base.base;
227 }
228 
229 int lima_gem_get_info(struct drm_file *file, u32 handle, u32 *va, u64 *offset)
230 {
231 	struct drm_gem_object *obj;
232 	struct lima_bo *bo;
233 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
234 	struct lima_vm *vm = priv->vm;
235 
236 	obj = drm_gem_object_lookup(file, handle);
237 	if (!obj)
238 		return -ENOENT;
239 
240 	bo = to_lima_bo(obj);
241 
242 	*va = lima_vm_get_va(vm, bo);
243 
244 	*offset = drm_vma_node_offset_addr(&obj->vma_node);
245 
246 	drm_gem_object_put(obj);
247 	return 0;
248 }
249 
250 static int lima_gem_sync_bo(struct lima_sched_task *task, struct lima_bo *bo,
251 			    bool write, bool explicit)
252 {
253 	int err = 0;
254 
255 	if (!write) {
256 		err = dma_resv_reserve_shared(lima_bo_resv(bo), 1);
257 		if (err)
258 			return err;
259 	}
260 
261 	/* explicit sync use user passed dep fence */
262 	if (explicit)
263 		return 0;
264 
265 	return drm_gem_fence_array_add_implicit(&task->deps, &bo->base.base, write);
266 }
267 
268 static int lima_gem_add_deps(struct drm_file *file, struct lima_submit *submit)
269 {
270 	int i, err;
271 
272 	for (i = 0; i < ARRAY_SIZE(submit->in_sync); i++) {
273 		struct dma_fence *fence = NULL;
274 
275 		if (!submit->in_sync[i])
276 			continue;
277 
278 		err = drm_syncobj_find_fence(file, submit->in_sync[i],
279 					     0, 0, &fence);
280 		if (err)
281 			return err;
282 
283 		err = drm_gem_fence_array_add(&submit->task->deps, fence);
284 		if (err) {
285 			dma_fence_put(fence);
286 			return err;
287 		}
288 	}
289 
290 	return 0;
291 }
292 
293 int lima_gem_submit(struct drm_file *file, struct lima_submit *submit)
294 {
295 	int i, err = 0;
296 	struct ww_acquire_ctx ctx;
297 	struct lima_drm_priv *priv = to_lima_drm_priv(file);
298 	struct lima_vm *vm = priv->vm;
299 	struct drm_syncobj *out_sync = NULL;
300 	struct dma_fence *fence;
301 	struct lima_bo **bos = submit->lbos;
302 
303 	if (submit->out_sync) {
304 		out_sync = drm_syncobj_find(file, submit->out_sync);
305 		if (!out_sync)
306 			return -ENOENT;
307 	}
308 
309 	for (i = 0; i < submit->nr_bos; i++) {
310 		struct drm_gem_object *obj;
311 		struct lima_bo *bo;
312 
313 		obj = drm_gem_object_lookup(file, submit->bos[i].handle);
314 		if (!obj) {
315 			err = -ENOENT;
316 			goto err_out0;
317 		}
318 
319 		bo = to_lima_bo(obj);
320 
321 		/* increase refcnt of gpu va map to prevent unmapped when executing,
322 		 * will be decreased when task done
323 		 */
324 		err = lima_vm_bo_add(vm, bo, false);
325 		if (err) {
326 			drm_gem_object_put(obj);
327 			goto err_out0;
328 		}
329 
330 		bos[i] = bo;
331 	}
332 
333 	err = drm_gem_lock_reservations((struct drm_gem_object **)bos,
334 					submit->nr_bos, &ctx);
335 	if (err)
336 		goto err_out0;
337 
338 	err = lima_sched_task_init(
339 		submit->task, submit->ctx->context + submit->pipe,
340 		bos, submit->nr_bos, vm);
341 	if (err)
342 		goto err_out1;
343 
344 	err = lima_gem_add_deps(file, submit);
345 	if (err)
346 		goto err_out2;
347 
348 	for (i = 0; i < submit->nr_bos; i++) {
349 		err = lima_gem_sync_bo(
350 			submit->task, bos[i],
351 			submit->bos[i].flags & LIMA_SUBMIT_BO_WRITE,
352 			submit->flags & LIMA_SUBMIT_FLAG_EXPLICIT_FENCE);
353 		if (err)
354 			goto err_out2;
355 	}
356 
357 	fence = lima_sched_context_queue_task(
358 		submit->ctx->context + submit->pipe, submit->task);
359 
360 	for (i = 0; i < submit->nr_bos; i++) {
361 		if (submit->bos[i].flags & LIMA_SUBMIT_BO_WRITE)
362 			dma_resv_add_excl_fence(lima_bo_resv(bos[i]), fence);
363 		else
364 			dma_resv_add_shared_fence(lima_bo_resv(bos[i]), fence);
365 	}
366 
367 	drm_gem_unlock_reservations((struct drm_gem_object **)bos,
368 				    submit->nr_bos, &ctx);
369 
370 	for (i = 0; i < submit->nr_bos; i++)
371 		drm_gem_object_put(&bos[i]->base.base);
372 
373 	if (out_sync) {
374 		drm_syncobj_replace_fence(out_sync, fence);
375 		drm_syncobj_put(out_sync);
376 	}
377 
378 	dma_fence_put(fence);
379 
380 	return 0;
381 
382 err_out2:
383 	lima_sched_task_fini(submit->task);
384 err_out1:
385 	drm_gem_unlock_reservations((struct drm_gem_object **)bos,
386 				    submit->nr_bos, &ctx);
387 err_out0:
388 	for (i = 0; i < submit->nr_bos; i++) {
389 		if (!bos[i])
390 			break;
391 		lima_vm_bo_del(vm, bos[i]);
392 		drm_gem_object_put(&bos[i]->base.base);
393 	}
394 	if (out_sync)
395 		drm_syncobj_put(out_sync);
396 	return err;
397 }
398 
399 int lima_gem_wait(struct drm_file *file, u32 handle, u32 op, s64 timeout_ns)
400 {
401 	bool write = op & LIMA_GEM_WAIT_WRITE;
402 	long ret, timeout;
403 
404 	if (!op)
405 		return 0;
406 
407 	timeout = drm_timeout_abs_to_jiffies(timeout_ns);
408 
409 	ret = drm_gem_dma_resv_wait(file, handle, write, timeout);
410 	if (ret == -ETIME)
411 		ret = timeout ? -ETIMEDOUT : -EBUSY;
412 
413 	return ret;
414 }
415