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
v3d_init_core(struct v3d_dev * v3d,int core)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 < V3D_GEN_41)
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
v3d_init_hw_state(struct v3d_dev * v3d)40 v3d_init_hw_state(struct v3d_dev *v3d)
41 {
42 v3d_init_core(v3d, 0);
43 }
44
45 static void
v3d_idle_axi(struct v3d_dev * v3d,int core)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
v3d_idle_gca(struct v3d_dev * v3d)59 v3d_idle_gca(struct v3d_dev *v3d)
60 {
61 if (v3d->ver >= V3D_GEN_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
v3d_reset_by_bridge(struct v3d_dev * v3d)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
v3d_reset_v3d(struct v3d_dev * v3d)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
v3d_reset_sms(struct v3d_dev * v3d)108 v3d_reset_sms(struct v3d_dev *v3d)
109 {
110 if (v3d->ver < V3D_GEN_71)
111 return;
112
113 V3D_SMS_WRITE(V3D_SMS_REE_CS, V3D_SET_FIELD(0x4, V3D_SMS_STATE));
114
115 if (wait_for(!(V3D_GET_FIELD(V3D_SMS_READ(V3D_SMS_REE_CS),
116 V3D_SMS_STATE) == V3D_SMS_ISOLATING_FOR_RESET) &&
117 !(V3D_GET_FIELD(V3D_SMS_READ(V3D_SMS_REE_CS),
118 V3D_SMS_STATE) == V3D_SMS_RESETTING), 100)) {
119 DRM_ERROR("Failed to wait for SMS reset\n");
120 }
121 }
122
123 void
v3d_reset(struct v3d_dev * v3d)124 v3d_reset(struct v3d_dev *v3d)
125 {
126 struct drm_device *dev = &v3d->drm;
127
128 DRM_DEV_ERROR(dev->dev, "Resetting GPU for hang.\n");
129 DRM_DEV_ERROR(dev->dev, "V3D_ERR_STAT: 0x%08x\n",
130 V3D_CORE_READ(0, V3D_ERR_STAT));
131 trace_v3d_reset_begin(dev);
132
133 /* XXX: only needed for safe powerdown, not reset. */
134 if (false)
135 v3d_idle_axi(v3d, 0);
136
137 v3d_irq_disable(v3d);
138
139 v3d_idle_gca(v3d);
140 v3d_reset_sms(v3d);
141 v3d_reset_v3d(v3d);
142
143 v3d_mmu_set_page_table(v3d);
144 v3d_irq_reset(v3d);
145
146 v3d_perfmon_stop(v3d, v3d->active_perfmon, false);
147
148 trace_v3d_reset_end(dev);
149 }
150
151 static void
v3d_flush_l3(struct v3d_dev * v3d)152 v3d_flush_l3(struct v3d_dev *v3d)
153 {
154 if (v3d->ver < V3D_GEN_41) {
155 u32 gca_ctrl = V3D_GCA_READ(V3D_GCA_CACHE_CTRL);
156
157 V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
158 gca_ctrl | V3D_GCA_CACHE_CTRL_FLUSH);
159
160 if (v3d->ver < V3D_GEN_33) {
161 V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
162 gca_ctrl & ~V3D_GCA_CACHE_CTRL_FLUSH);
163 }
164 }
165 }
166
167 /* Invalidates the (read-only) L2C cache. This was the L2 cache for
168 * uniforms and instructions on V3D 3.2.
169 */
170 static void
v3d_invalidate_l2c(struct v3d_dev * v3d,int core)171 v3d_invalidate_l2c(struct v3d_dev *v3d, int core)
172 {
173 if (v3d->ver >= V3D_GEN_33)
174 return;
175
176 V3D_CORE_WRITE(core, V3D_CTL_L2CACTL,
177 V3D_L2CACTL_L2CCLR |
178 V3D_L2CACTL_L2CENA);
179 }
180
181 /* Invalidates texture L2 cachelines */
182 static void
v3d_flush_l2t(struct v3d_dev * v3d,int core)183 v3d_flush_l2t(struct v3d_dev *v3d, int core)
184 {
185 /* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't
186 * need to wait for completion before dispatching the job --
187 * L2T accesses will be stalled until the flush has completed.
188 * However, we do need to make sure we don't try to trigger a
189 * new flush while the L2_CLEAN queue is trying to
190 * synchronously clean after a job.
191 */
192 mutex_lock(&v3d->cache_clean_lock);
193 V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
194 V3D_L2TCACTL_L2TFLS |
195 V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
196 mutex_unlock(&v3d->cache_clean_lock);
197 }
198
199 /* Cleans texture L1 and L2 cachelines (writing back dirty data).
200 *
201 * For cleaning, which happens from the CACHE_CLEAN queue after CSD has
202 * executed, we need to make sure that the clean is done before
203 * signaling job completion. So, we synchronously wait before
204 * returning, and we make sure that L2 invalidates don't happen in the
205 * meantime to confuse our are-we-done checks.
206 */
207 void
v3d_clean_caches(struct v3d_dev * v3d)208 v3d_clean_caches(struct v3d_dev *v3d)
209 {
210 struct drm_device *dev = &v3d->drm;
211 int core = 0;
212
213 trace_v3d_cache_clean_begin(dev);
214
215 V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
216 if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
217 V3D_L2TCACTL_TMUWCF), 100)) {
218 DRM_ERROR("Timeout waiting for TMU write combiner flush\n");
219 }
220
221 mutex_lock(&v3d->cache_clean_lock);
222 V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
223 V3D_L2TCACTL_L2TFLS |
224 V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM));
225
226 if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
227 V3D_L2TCACTL_L2TFLS), 100)) {
228 DRM_ERROR("Timeout waiting for L2T clean\n");
229 }
230
231 mutex_unlock(&v3d->cache_clean_lock);
232
233 trace_v3d_cache_clean_end(dev);
234 }
235
236 /* Invalidates the slice caches. These are read-only caches. */
237 static void
v3d_invalidate_slices(struct v3d_dev * v3d,int core)238 v3d_invalidate_slices(struct v3d_dev *v3d, int core)
239 {
240 V3D_CORE_WRITE(core, V3D_CTL_SLCACTL,
241 V3D_SET_FIELD(0xf, V3D_SLCACTL_TVCCS) |
242 V3D_SET_FIELD(0xf, V3D_SLCACTL_TDCCS) |
243 V3D_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
244 V3D_SET_FIELD(0xf, V3D_SLCACTL_ICC));
245 }
246
247 void
v3d_invalidate_caches(struct v3d_dev * v3d)248 v3d_invalidate_caches(struct v3d_dev *v3d)
249 {
250 /* Invalidate the caches from the outside in. That way if
251 * another CL's concurrent use of nearby memory were to pull
252 * an invalidated cacheline back in, we wouldn't leave stale
253 * data in the inner cache.
254 */
255 v3d_flush_l3(v3d);
256 v3d_invalidate_l2c(v3d, 0);
257 v3d_flush_l2t(v3d, 0);
258 v3d_invalidate_slices(v3d, 0);
259 }
260
261 int
v3d_gem_init(struct drm_device * dev)262 v3d_gem_init(struct drm_device *dev)
263 {
264 struct v3d_dev *v3d = to_v3d_dev(dev);
265 u32 pt_size = 4096 * 1024;
266 int ret, i;
267
268 for (i = 0; i < V3D_MAX_QUEUES; i++) {
269 struct v3d_queue_state *queue = &v3d->queue[i];
270
271 queue->fence_context = dma_fence_context_alloc(1);
272 memset(&queue->stats, 0, sizeof(queue->stats));
273 seqcount_init(&queue->stats.lock);
274 }
275
276 spin_lock_init(&v3d->mm_lock);
277 spin_lock_init(&v3d->job_lock);
278 ret = drmm_mutex_init(dev, &v3d->bo_lock);
279 if (ret)
280 return ret;
281 ret = drmm_mutex_init(dev, &v3d->reset_lock);
282 if (ret)
283 return ret;
284 ret = drmm_mutex_init(dev, &v3d->sched_lock);
285 if (ret)
286 return ret;
287 ret = drmm_mutex_init(dev, &v3d->cache_clean_lock);
288 if (ret)
289 return ret;
290
291 /* Note: We don't allocate address 0. Various bits of HW
292 * treat 0 as special, such as the occlusion query counters
293 * where 0 means "disabled".
294 */
295 drm_mm_init(&v3d->mm, 1, pt_size / sizeof(u32) - 1);
296
297 v3d->pt = dma_alloc_wc(v3d->drm.dev, pt_size,
298 &v3d->pt_paddr,
299 GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
300 if (!v3d->pt) {
301 drm_mm_takedown(&v3d->mm);
302 dev_err(v3d->drm.dev,
303 "Failed to allocate page tables. Please ensure you have DMA enabled.\n");
304 return -ENOMEM;
305 }
306
307 v3d_init_hw_state(v3d);
308 v3d_mmu_set_page_table(v3d);
309
310 v3d_gemfs_init(v3d);
311
312 ret = v3d_sched_init(v3d);
313 if (ret) {
314 drm_mm_takedown(&v3d->mm);
315 dma_free_coherent(v3d->drm.dev, pt_size, (void *)v3d->pt,
316 v3d->pt_paddr);
317 return ret;
318 }
319
320 return 0;
321 }
322
323 void
v3d_gem_destroy(struct drm_device * dev)324 v3d_gem_destroy(struct drm_device *dev)
325 {
326 struct v3d_dev *v3d = to_v3d_dev(dev);
327
328 v3d_sched_fini(v3d);
329 v3d_gemfs_fini(v3d);
330
331 /* Waiting for jobs to finish would need to be done before
332 * unregistering V3D.
333 */
334 WARN_ON(v3d->bin_job);
335 WARN_ON(v3d->render_job);
336 WARN_ON(v3d->tfu_job);
337 WARN_ON(v3d->csd_job);
338
339 drm_mm_takedown(&v3d->mm);
340
341 dma_free_coherent(v3d->drm.dev, 4096 * 1024, (void *)v3d->pt,
342 v3d->pt_paddr);
343 }
344