xref: /linux/drivers/gpu/drm/v3d/v3d_gem.c (revision 6bfca938471b1804c4a63774ca2fd30cb3596215)
1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2014-2018 Broadcom */
3 
4 #include <linux/device.h>
5 #include <linux/dma-mapping.h>
6 #include <linux/io.h>
7 #include <linux/module.h>
8 #include <linux/platform_device.h>
9 #include <linux/reset.h>
10 #include <linux/sched/signal.h>
11 #include <linux/uaccess.h>
12 
13 #include <drm/drm_managed.h>
14 #include <drm/drm_print.h>
15 
16 #include "v3d_drv.h"
17 #include "v3d_regs.h"
18 #include "v3d_trace.h"
19 
20 static void
21 v3d_init_core(struct v3d_dev *v3d, int core)
22 {
23 	/* Set OVRTMUOUT, which means that the texture sampler uniform
24 	 * configuration's tmu output type field is used, instead of
25 	 * using the hardware default behavior based on the texture
26 	 * type.  If you want the default behavior, you can still put
27 	 * "2" in the indirect texture state's output_type field.
28 	 */
29 	if (v3d->ver < V3D_GEN_41)
30 		V3D_CORE_WRITE(core, V3D_CTL_MISCCFG, V3D_MISCCFG_OVRTMUOUT);
31 
32 	/* Whenever we flush the L2T cache, we always want to flush
33 	 * the whole thing.
34 	 */
35 	V3D_CORE_WRITE(core, V3D_CTL_L2TFLSTA, 0);
36 	V3D_CORE_WRITE(core, V3D_CTL_L2TFLEND, ~0);
37 }
38 
39 static void
40 v3d_idle_axi(struct v3d_dev *v3d, int core)
41 {
42 	V3D_CORE_WRITE(core, V3D_GMP_CFG(v3d->ver), V3D_GMP_CFG_STOP_REQ);
43 
44 	if (wait_for((V3D_CORE_READ(core, V3D_GMP_STATUS(v3d->ver)) &
45 		      (V3D_GMP_STATUS_RD_COUNT_MASK |
46 		       V3D_GMP_STATUS_WR_COUNT_MASK |
47 		       V3D_GMP_STATUS_CFG_BUSY)) == 0, 100)) {
48 		drm_err(&v3d->drm, "Failed to wait for safe GMP shutdown\n");
49 	}
50 }
51 
52 static void
53 v3d_idle_gca(struct v3d_dev *v3d)
54 {
55 	if (v3d->ver >= V3D_GEN_41)
56 		return;
57 
58 	V3D_GCA_WRITE(V3D_GCA_SAFE_SHUTDOWN, V3D_GCA_SAFE_SHUTDOWN_EN);
59 
60 	if (wait_for((V3D_GCA_READ(V3D_GCA_SAFE_SHUTDOWN_ACK) &
61 		      V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED) ==
62 		     V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED, 100)) {
63 		drm_err(&v3d->drm, "Failed to wait for safe GCA shutdown\n");
64 	}
65 }
66 
67 static void
68 v3d_reset_by_bridge(struct v3d_dev *v3d)
69 {
70 	int version = V3D_BRIDGE_READ(V3D_TOP_GR_BRIDGE_REVISION);
71 
72 	if (V3D_GET_FIELD(version, V3D_TOP_GR_BRIDGE_MAJOR) == 2) {
73 		V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0,
74 				 V3D_TOP_GR_BRIDGE_SW_INIT_0_V3D_CLK_108_SW_INIT);
75 		V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0, 0);
76 
77 		/* GFXH-1383: The SW_INIT may cause a stray write to address 0
78 		 * of the unit, so reset it to its power-on value here.
79 		 */
80 		V3D_WRITE(V3D_HUB_AXICFG, V3D_HUB_AXICFG_MAX_LEN_MASK);
81 	} else {
82 		WARN_ON_ONCE(V3D_GET_FIELD(version,
83 					   V3D_TOP_GR_BRIDGE_MAJOR) != 7);
84 		V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1,
85 				 V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT);
86 		V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1, 0);
87 	}
88 }
89 
90 static void
91 v3d_reset_v3d(struct v3d_dev *v3d)
92 {
93 	if (v3d->reset)
94 		reset_control_reset(v3d->reset);
95 	else
96 		v3d_reset_by_bridge(v3d);
97 
98 	v3d_init_hw_state(v3d);
99 }
100 
101 void
102 v3d_reset_sms(struct v3d_dev *v3d)
103 {
104 	if (v3d->ver < V3D_GEN_71)
105 		return;
106 
107 	V3D_SMS_WRITE(V3D_SMS_REE_CS, V3D_SET_FIELD(0x4, V3D_SMS_STATE));
108 
109 	if (wait_for(!(V3D_GET_FIELD(V3D_SMS_READ(V3D_SMS_REE_CS),
110 				     V3D_SMS_STATE) == V3D_SMS_ISOLATING_FOR_RESET) &&
111 		     !(V3D_GET_FIELD(V3D_SMS_READ(V3D_SMS_REE_CS),
112 				     V3D_SMS_STATE) == V3D_SMS_RESETTING), 100)) {
113 		drm_err(&v3d->drm, "Failed to wait for SMS reset\n");
114 	}
115 }
116 
117 void
118 v3d_reset(struct v3d_dev *v3d)
119 {
120 	struct drm_device *dev = &v3d->drm;
121 
122 	drm_err(dev, "Resetting GPU for hang.\n");
123 	drm_err(dev, "V3D_ERR_STAT: 0x%08x\n", V3D_CORE_READ(0, V3D_ERR_STAT));
124 
125 	trace_v3d_reset_begin(dev);
126 
127 	/* XXX: only needed for safe powerdown, not reset. */
128 	if (false)
129 		v3d_idle_axi(v3d, 0);
130 
131 	v3d_irq_disable(v3d);
132 
133 	v3d_idle_gca(v3d);
134 	v3d_reset_sms(v3d);
135 	v3d_reset_v3d(v3d);
136 
137 	v3d_mmu_set_page_table(v3d);
138 	v3d_irq_reset(v3d);
139 
140 	/* Re-arm the global perfmon HW counters that the reset zeroed. */
141 	v3d_perfmon_resume(v3d);
142 
143 	trace_v3d_reset_end(dev);
144 }
145 
146 static void
147 v3d_flush_l3(struct v3d_dev *v3d)
148 {
149 	if (v3d->ver < V3D_GEN_41) {
150 		u32 gca_ctrl = V3D_GCA_READ(V3D_GCA_CACHE_CTRL);
151 
152 		V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
153 			      gca_ctrl | V3D_GCA_CACHE_CTRL_FLUSH);
154 
155 		if (v3d->ver < V3D_GEN_33) {
156 			V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
157 				      gca_ctrl & ~V3D_GCA_CACHE_CTRL_FLUSH);
158 		}
159 	}
160 }
161 
162 /* Invalidates the (read-only) L2C cache.  This was the L2 cache for
163  * uniforms and instructions on V3D 3.2.
164  */
165 static void
166 v3d_invalidate_l2c(struct v3d_dev *v3d, int core)
167 {
168 	if (v3d->ver >= V3D_GEN_33)
169 		return;
170 
171 	V3D_CORE_WRITE(core, V3D_CTL_L2CACTL,
172 		       V3D_L2CACTL_L2CCLR |
173 		       V3D_L2CACTL_L2CENA);
174 }
175 
176 /* Invalidates texture L2 cachelines */
177 static void
178 v3d_flush_l2t(struct v3d_dev *v3d, int core)
179 {
180 	/* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't
181 	 * need to wait for completion before dispatching the job --
182 	 * L2T accesses will be stalled until the flush has completed.
183 	 * However, we do need to make sure we don't try to trigger a
184 	 * new flush while the L2_CLEAN queue is trying to
185 	 * synchronously clean after a job.
186 	 */
187 	mutex_lock(&v3d->cache_clean_lock);
188 	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
189 		       V3D_L2TCACTL_L2TFLS |
190 		       V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
191 	mutex_unlock(&v3d->cache_clean_lock);
192 }
193 
194 /* Cleans texture L1 and L2 cachelines (writing back dirty data).
195  *
196  * For cleaning, which happens from the CACHE_CLEAN queue after CSD has
197  * executed, we need to make sure that the clean is done before
198  * signaling job completion.  So, we synchronously wait before
199  * returning, and we make sure that L2 invalidates don't happen in the
200  * meantime to confuse our are-we-done checks.
201  */
202 void
203 v3d_clean_caches(struct v3d_dev *v3d)
204 {
205 	struct drm_device *dev = &v3d->drm;
206 	int core = 0;
207 
208 	trace_v3d_cache_clean_begin(dev);
209 
210 	/* GFXH-1897: Ensure pending flushes complete before writing L2TCACTL */
211 	if (v3d->ver < V3D_GEN_71) {
212 		if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
213 			       V3D_L2TCACTL_L2TFLS), 100)) {
214 			drm_err(dev, "Timeout waiting for L2T clean\n");
215 		}
216 	}
217 
218 	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
219 	if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
220 		       V3D_L2TCACTL_TMUWCF), 100)) {
221 		drm_err(dev, "Timeout waiting for TMU write combiner flush\n");
222 	}
223 
224 	mutex_lock(&v3d->cache_clean_lock);
225 	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
226 		       V3D_L2TCACTL_L2TFLS |
227 		       V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM));
228 
229 	if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
230 		       V3D_L2TCACTL_L2TFLS), 100)) {
231 		drm_err(dev, "Timeout waiting for L2T clean\n");
232 	}
233 
234 	mutex_unlock(&v3d->cache_clean_lock);
235 
236 	trace_v3d_cache_clean_end(dev);
237 }
238 
239 /* Invalidates the slice caches.  These are read-only caches. */
240 static void
241 v3d_invalidate_slices(struct v3d_dev *v3d, int core)
242 {
243 	V3D_CORE_WRITE(core, V3D_CTL_SLCACTL,
244 		       V3D_SET_FIELD(0xf, V3D_SLCACTL_TVCCS) |
245 		       V3D_SET_FIELD(0xf, V3D_SLCACTL_TDCCS) |
246 		       V3D_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
247 		       V3D_SET_FIELD(0xf, V3D_SLCACTL_ICC));
248 }
249 
250 void
251 v3d_invalidate_caches(struct v3d_dev *v3d)
252 {
253 	/* Invalidate the caches from the outside in.  That way if
254 	 * another CL's concurrent use of nearby memory were to pull
255 	 * an invalidated cacheline back in, we wouldn't leave stale
256 	 * data in the inner cache.
257 	 */
258 	v3d_flush_l3(v3d);
259 	v3d_invalidate_l2c(v3d, 0);
260 	v3d_flush_l2t(v3d, 0);
261 	v3d_invalidate_slices(v3d, 0);
262 }
263 
264 /* Sets invariant state for the HW. */
265 void
266 v3d_init_hw_state(struct v3d_dev *v3d)
267 {
268 	v3d_init_core(v3d, 0);
269 }
270 
271 static void
272 v3d_huge_mnt_init(struct v3d_dev *v3d)
273 {
274 	int err = 0;
275 
276 	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && super_pages)
277 		err = drm_gem_huge_mnt_create(&v3d->drm, "within_size");
278 
279 	if (drm_gem_get_huge_mnt(&v3d->drm))
280 		drm_info(&v3d->drm, "Using Transparent Hugepages\n");
281 	else if (err)
282 		drm_warn(&v3d->drm, "Can't use Transparent Hugepages (%d)\n",
283 			 err);
284 	else
285 		drm_notice(&v3d->drm,
286 			   "Transparent Hugepage support is recommended for optimal performance on this platform!\n");
287 }
288 
289 int
290 v3d_gem_init(struct drm_device *dev)
291 {
292 	struct v3d_dev *v3d = to_v3d_dev(dev);
293 	u32 pt_size = 4096 * 1024;
294 	int ret, i;
295 
296 	for (i = 0; i < V3D_MAX_QUEUES; i++) {
297 		struct v3d_queue_state *queue = &v3d->queue[i];
298 
299 		queue->stats = v3d_stats_alloc();
300 		if (!queue->stats) {
301 			ret = -ENOMEM;
302 			goto err_stats;
303 		}
304 
305 		queue->fence_context = dma_fence_context_alloc(1);
306 
307 		spin_lock_init(&queue->queue_lock);
308 	}
309 
310 	spin_lock_init(&v3d->mm_lock);
311 	spin_lock_init(&v3d->perfmon_state.lock);
312 	ret = drmm_mutex_init(dev, &v3d->bo_lock);
313 	if (ret)
314 		goto err_stats;
315 	ret = drmm_mutex_init(dev, &v3d->reset_lock);
316 	if (ret)
317 		goto err_stats;
318 	ret = drmm_mutex_init(dev, &v3d->sched_lock);
319 	if (ret)
320 		goto err_stats;
321 	ret = drmm_mutex_init(dev, &v3d->cache_clean_lock);
322 	if (ret)
323 		goto err_stats;
324 
325 	/* Note: We don't allocate address 0.  Various bits of HW
326 	 * treat 0 as special, such as the occlusion query counters
327 	 * where 0 means "disabled".
328 	 */
329 	drm_mm_init(&v3d->mm, 1, pt_size / sizeof(u32) - 1);
330 
331 	v3d->pt = dma_alloc_wc(v3d->drm.dev, pt_size,
332 			       &v3d->pt_paddr,
333 			       GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
334 	if (!v3d->pt) {
335 		dev_err(v3d->drm.dev,
336 			"Failed to allocate page tables. Please ensure you have DMA enabled.\n");
337 		ret = -ENOMEM;
338 		goto err_dma_alloc;
339 	}
340 
341 	v3d_huge_mnt_init(v3d);
342 
343 	ret = v3d_sched_init(v3d);
344 	if (ret)
345 		goto err_sched;
346 
347 	return 0;
348 
349 err_sched:
350 	dma_free_coherent(v3d->drm.dev, pt_size, (void *)v3d->pt, v3d->pt_paddr);
351 err_dma_alloc:
352 	drm_mm_takedown(&v3d->mm);
353 err_stats:
354 	for (i--; i >= 0; i--)
355 		v3d_stats_put(v3d->queue[i].stats);
356 
357 	return ret;
358 }
359 
360 void
361 v3d_gem_destroy(struct drm_device *dev)
362 {
363 	struct v3d_dev *v3d = to_v3d_dev(dev);
364 	enum v3d_queue q;
365 
366 	v3d_sched_fini(v3d);
367 
368 	/* Waiting for jobs to finish would need to be done before
369 	 * unregistering V3D.
370 	 */
371 	for (q = 0; q < V3D_MAX_QUEUES; q++) {
372 		WARN_ON(v3d->queue[q].active_job);
373 		v3d_stats_put(v3d->queue[q].stats);
374 		dma_fence_put(v3d->perfmon_state.last_hw_fence[q]);
375 	}
376 
377 	dma_fence_put(v3d->perfmon_state.fence);
378 
379 	drm_mm_takedown(&v3d->mm);
380 
381 	dma_free_coherent(v3d->drm.dev, 4096 * 1024, (void *)v3d->pt,
382 			  v3d->pt_paddr);
383 }
384