xref: /linux/drivers/gpu/drm/v3d/v3d_gem.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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 
15 #include "v3d_drv.h"
16 #include "v3d_regs.h"
17 #include "v3d_trace.h"
18 
19 static void
20 v3d_init_core(struct v3d_dev *v3d, int core)
21 {
22 	/* Set OVRTMUOUT, which means that the texture sampler uniform
23 	 * configuration's tmu output type field is used, instead of
24 	 * using the hardware default behavior based on the texture
25 	 * type.  If you want the default behavior, you can still put
26 	 * "2" in the indirect texture state's output_type field.
27 	 */
28 	if (v3d->ver < 40)
29 		V3D_CORE_WRITE(core, V3D_CTL_MISCCFG, V3D_MISCCFG_OVRTMUOUT);
30 
31 	/* Whenever we flush the L2T cache, we always want to flush
32 	 * the whole thing.
33 	 */
34 	V3D_CORE_WRITE(core, V3D_CTL_L2TFLSTA, 0);
35 	V3D_CORE_WRITE(core, V3D_CTL_L2TFLEND, ~0);
36 }
37 
38 /* Sets invariant state for the HW. */
39 static void
40 v3d_init_hw_state(struct v3d_dev *v3d)
41 {
42 	v3d_init_core(v3d, 0);
43 }
44 
45 static void
46 v3d_idle_axi(struct v3d_dev *v3d, int core)
47 {
48 	V3D_CORE_WRITE(core, V3D_GMP_CFG(v3d->ver), V3D_GMP_CFG_STOP_REQ);
49 
50 	if (wait_for((V3D_CORE_READ(core, V3D_GMP_STATUS(v3d->ver)) &
51 		      (V3D_GMP_STATUS_RD_COUNT_MASK |
52 		       V3D_GMP_STATUS_WR_COUNT_MASK |
53 		       V3D_GMP_STATUS_CFG_BUSY)) == 0, 100)) {
54 		DRM_ERROR("Failed to wait for safe GMP shutdown\n");
55 	}
56 }
57 
58 static void
59 v3d_idle_gca(struct v3d_dev *v3d)
60 {
61 	if (v3d->ver >= 41)
62 		return;
63 
64 	V3D_GCA_WRITE(V3D_GCA_SAFE_SHUTDOWN, V3D_GCA_SAFE_SHUTDOWN_EN);
65 
66 	if (wait_for((V3D_GCA_READ(V3D_GCA_SAFE_SHUTDOWN_ACK) &
67 		      V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED) ==
68 		     V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED, 100)) {
69 		DRM_ERROR("Failed to wait for safe GCA shutdown\n");
70 	}
71 }
72 
73 static void
74 v3d_reset_by_bridge(struct v3d_dev *v3d)
75 {
76 	int version = V3D_BRIDGE_READ(V3D_TOP_GR_BRIDGE_REVISION);
77 
78 	if (V3D_GET_FIELD(version, V3D_TOP_GR_BRIDGE_MAJOR) == 2) {
79 		V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0,
80 				 V3D_TOP_GR_BRIDGE_SW_INIT_0_V3D_CLK_108_SW_INIT);
81 		V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0, 0);
82 
83 		/* GFXH-1383: The SW_INIT may cause a stray write to address 0
84 		 * of the unit, so reset it to its power-on value here.
85 		 */
86 		V3D_WRITE(V3D_HUB_AXICFG, V3D_HUB_AXICFG_MAX_LEN_MASK);
87 	} else {
88 		WARN_ON_ONCE(V3D_GET_FIELD(version,
89 					   V3D_TOP_GR_BRIDGE_MAJOR) != 7);
90 		V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1,
91 				 V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT);
92 		V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1, 0);
93 	}
94 }
95 
96 static void
97 v3d_reset_v3d(struct v3d_dev *v3d)
98 {
99 	if (v3d->reset)
100 		reset_control_reset(v3d->reset);
101 	else
102 		v3d_reset_by_bridge(v3d);
103 
104 	v3d_init_hw_state(v3d);
105 }
106 
107 void
108 v3d_reset(struct v3d_dev *v3d)
109 {
110 	struct drm_device *dev = &v3d->drm;
111 
112 	DRM_DEV_ERROR(dev->dev, "Resetting GPU for hang.\n");
113 	DRM_DEV_ERROR(dev->dev, "V3D_ERR_STAT: 0x%08x\n",
114 		      V3D_CORE_READ(0, V3D_ERR_STAT));
115 	trace_v3d_reset_begin(dev);
116 
117 	/* XXX: only needed for safe powerdown, not reset. */
118 	if (false)
119 		v3d_idle_axi(v3d, 0);
120 
121 	v3d_idle_gca(v3d);
122 	v3d_reset_v3d(v3d);
123 
124 	v3d_mmu_set_page_table(v3d);
125 	v3d_irq_reset(v3d);
126 
127 	v3d_perfmon_stop(v3d, v3d->active_perfmon, false);
128 
129 	trace_v3d_reset_end(dev);
130 }
131 
132 static void
133 v3d_flush_l3(struct v3d_dev *v3d)
134 {
135 	if (v3d->ver < 41) {
136 		u32 gca_ctrl = V3D_GCA_READ(V3D_GCA_CACHE_CTRL);
137 
138 		V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
139 			      gca_ctrl | V3D_GCA_CACHE_CTRL_FLUSH);
140 
141 		if (v3d->ver < 33) {
142 			V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
143 				      gca_ctrl & ~V3D_GCA_CACHE_CTRL_FLUSH);
144 		}
145 	}
146 }
147 
148 /* Invalidates the (read-only) L2C cache.  This was the L2 cache for
149  * uniforms and instructions on V3D 3.2.
150  */
151 static void
152 v3d_invalidate_l2c(struct v3d_dev *v3d, int core)
153 {
154 	if (v3d->ver > 32)
155 		return;
156 
157 	V3D_CORE_WRITE(core, V3D_CTL_L2CACTL,
158 		       V3D_L2CACTL_L2CCLR |
159 		       V3D_L2CACTL_L2CENA);
160 }
161 
162 /* Invalidates texture L2 cachelines */
163 static void
164 v3d_flush_l2t(struct v3d_dev *v3d, int core)
165 {
166 	/* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't
167 	 * need to wait for completion before dispatching the job --
168 	 * L2T accesses will be stalled until the flush has completed.
169 	 * However, we do need to make sure we don't try to trigger a
170 	 * new flush while the L2_CLEAN queue is trying to
171 	 * synchronously clean after a job.
172 	 */
173 	mutex_lock(&v3d->cache_clean_lock);
174 	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
175 		       V3D_L2TCACTL_L2TFLS |
176 		       V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
177 	mutex_unlock(&v3d->cache_clean_lock);
178 }
179 
180 /* Cleans texture L1 and L2 cachelines (writing back dirty data).
181  *
182  * For cleaning, which happens from the CACHE_CLEAN queue after CSD has
183  * executed, we need to make sure that the clean is done before
184  * signaling job completion.  So, we synchronously wait before
185  * returning, and we make sure that L2 invalidates don't happen in the
186  * meantime to confuse our are-we-done checks.
187  */
188 void
189 v3d_clean_caches(struct v3d_dev *v3d)
190 {
191 	struct drm_device *dev = &v3d->drm;
192 	int core = 0;
193 
194 	trace_v3d_cache_clean_begin(dev);
195 
196 	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
197 	if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
198 		       V3D_L2TCACTL_TMUWCF), 100)) {
199 		DRM_ERROR("Timeout waiting for TMU write combiner flush\n");
200 	}
201 
202 	mutex_lock(&v3d->cache_clean_lock);
203 	V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
204 		       V3D_L2TCACTL_L2TFLS |
205 		       V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM));
206 
207 	if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
208 		       V3D_L2TCACTL_L2TFLS), 100)) {
209 		DRM_ERROR("Timeout waiting for L2T clean\n");
210 	}
211 
212 	mutex_unlock(&v3d->cache_clean_lock);
213 
214 	trace_v3d_cache_clean_end(dev);
215 }
216 
217 /* Invalidates the slice caches.  These are read-only caches. */
218 static void
219 v3d_invalidate_slices(struct v3d_dev *v3d, int core)
220 {
221 	V3D_CORE_WRITE(core, V3D_CTL_SLCACTL,
222 		       V3D_SET_FIELD(0xf, V3D_SLCACTL_TVCCS) |
223 		       V3D_SET_FIELD(0xf, V3D_SLCACTL_TDCCS) |
224 		       V3D_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
225 		       V3D_SET_FIELD(0xf, V3D_SLCACTL_ICC));
226 }
227 
228 void
229 v3d_invalidate_caches(struct v3d_dev *v3d)
230 {
231 	/* Invalidate the caches from the outside in.  That way if
232 	 * another CL's concurrent use of nearby memory were to pull
233 	 * an invalidated cacheline back in, we wouldn't leave stale
234 	 * data in the inner cache.
235 	 */
236 	v3d_flush_l3(v3d);
237 	v3d_invalidate_l2c(v3d, 0);
238 	v3d_flush_l2t(v3d, 0);
239 	v3d_invalidate_slices(v3d, 0);
240 }
241 
242 int
243 v3d_gem_init(struct drm_device *dev)
244 {
245 	struct v3d_dev *v3d = to_v3d_dev(dev);
246 	u32 pt_size = 4096 * 1024;
247 	int ret, i;
248 
249 	for (i = 0; i < V3D_MAX_QUEUES; i++) {
250 		struct v3d_queue_state *queue = &v3d->queue[i];
251 
252 		queue->fence_context = dma_fence_context_alloc(1);
253 		memset(&queue->stats, 0, sizeof(queue->stats));
254 		seqcount_init(&queue->stats.lock);
255 	}
256 
257 	spin_lock_init(&v3d->mm_lock);
258 	spin_lock_init(&v3d->job_lock);
259 	ret = drmm_mutex_init(dev, &v3d->bo_lock);
260 	if (ret)
261 		return ret;
262 	ret = drmm_mutex_init(dev, &v3d->reset_lock);
263 	if (ret)
264 		return ret;
265 	ret = drmm_mutex_init(dev, &v3d->sched_lock);
266 	if (ret)
267 		return ret;
268 	ret = drmm_mutex_init(dev, &v3d->cache_clean_lock);
269 	if (ret)
270 		return ret;
271 
272 	/* Note: We don't allocate address 0.  Various bits of HW
273 	 * treat 0 as special, such as the occlusion query counters
274 	 * where 0 means "disabled".
275 	 */
276 	drm_mm_init(&v3d->mm, 1, pt_size / sizeof(u32) - 1);
277 
278 	v3d->pt = dma_alloc_wc(v3d->drm.dev, pt_size,
279 			       &v3d->pt_paddr,
280 			       GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
281 	if (!v3d->pt) {
282 		drm_mm_takedown(&v3d->mm);
283 		dev_err(v3d->drm.dev,
284 			"Failed to allocate page tables. Please ensure you have DMA enabled.\n");
285 		return -ENOMEM;
286 	}
287 
288 	v3d_init_hw_state(v3d);
289 	v3d_mmu_set_page_table(v3d);
290 
291 	v3d_gemfs_init(v3d);
292 
293 	ret = v3d_sched_init(v3d);
294 	if (ret) {
295 		drm_mm_takedown(&v3d->mm);
296 		dma_free_coherent(v3d->drm.dev, pt_size, (void *)v3d->pt,
297 				  v3d->pt_paddr);
298 		return ret;
299 	}
300 
301 	return 0;
302 }
303 
304 void
305 v3d_gem_destroy(struct drm_device *dev)
306 {
307 	struct v3d_dev *v3d = to_v3d_dev(dev);
308 
309 	v3d_sched_fini(v3d);
310 	v3d_gemfs_fini(v3d);
311 
312 	/* Waiting for jobs to finish would need to be done before
313 	 * unregistering V3D.
314 	 */
315 	WARN_ON(v3d->bin_job);
316 	WARN_ON(v3d->render_job);
317 	WARN_ON(v3d->tfu_job);
318 	WARN_ON(v3d->csd_job);
319 
320 	drm_mm_takedown(&v3d->mm);
321 
322 	dma_free_coherent(v3d->drm.dev, 4096 * 1024, (void *)v3d->pt,
323 			  v3d->pt_paddr);
324 }
325