| 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 | /* Sets invariant state for the HW. */ |
| 40 | static void |
| 41 | v3d_init_hw_state(struct v3d_dev *v3d) |
| 42 | { |
| 43 | v3d_init_core(v3d, core: 0); |
| 44 | } |
| 45 | |
| 46 | static void |
| 47 | v3d_idle_axi(struct v3d_dev *v3d, int core) |
| 48 | { |
| 49 | V3D_CORE_WRITE(core, V3D_GMP_CFG(v3d->ver), V3D_GMP_CFG_STOP_REQ); |
| 50 | |
| 51 | if (wait_for((V3D_CORE_READ(core, V3D_GMP_STATUS(v3d->ver)) & |
| 52 | (V3D_GMP_STATUS_RD_COUNT_MASK | |
| 53 | V3D_GMP_STATUS_WR_COUNT_MASK | |
| 54 | V3D_GMP_STATUS_CFG_BUSY)) == 0, 100)) { |
| 55 | DRM_ERROR("Failed to wait for safe GMP shutdown\n" ); |
| 56 | } |
| 57 | } |
| 58 | |
| 59 | static void |
| 60 | v3d_idle_gca(struct v3d_dev *v3d) |
| 61 | { |
| 62 | if (v3d->ver >= V3D_GEN_41) |
| 63 | return; |
| 64 | |
| 65 | V3D_GCA_WRITE(V3D_GCA_SAFE_SHUTDOWN, V3D_GCA_SAFE_SHUTDOWN_EN); |
| 66 | |
| 67 | if (wait_for((V3D_GCA_READ(V3D_GCA_SAFE_SHUTDOWN_ACK) & |
| 68 | V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED) == |
| 69 | V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED, 100)) { |
| 70 | DRM_ERROR("Failed to wait for safe GCA shutdown\n" ); |
| 71 | } |
| 72 | } |
| 73 | |
| 74 | static void |
| 75 | v3d_reset_by_bridge(struct v3d_dev *v3d) |
| 76 | { |
| 77 | int version = V3D_BRIDGE_READ(V3D_TOP_GR_BRIDGE_REVISION); |
| 78 | |
| 79 | if (V3D_GET_FIELD(version, V3D_TOP_GR_BRIDGE_MAJOR) == 2) { |
| 80 | V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0, |
| 81 | V3D_TOP_GR_BRIDGE_SW_INIT_0_V3D_CLK_108_SW_INIT); |
| 82 | V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0, 0); |
| 83 | |
| 84 | /* GFXH-1383: The SW_INIT may cause a stray write to address 0 |
| 85 | * of the unit, so reset it to its power-on value here. |
| 86 | */ |
| 87 | V3D_WRITE(V3D_HUB_AXICFG, V3D_HUB_AXICFG_MAX_LEN_MASK); |
| 88 | } else { |
| 89 | WARN_ON_ONCE(V3D_GET_FIELD(version, |
| 90 | V3D_TOP_GR_BRIDGE_MAJOR) != 7); |
| 91 | V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1, |
| 92 | V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT); |
| 93 | V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1, 0); |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | static void |
| 98 | v3d_reset_v3d(struct v3d_dev *v3d) |
| 99 | { |
| 100 | if (v3d->reset) |
| 101 | reset_control_reset(rstc: v3d->reset); |
| 102 | else |
| 103 | v3d_reset_by_bridge(v3d); |
| 104 | |
| 105 | v3d_init_hw_state(v3d); |
| 106 | } |
| 107 | |
| 108 | void |
| 109 | v3d_reset_sms(struct v3d_dev *v3d) |
| 110 | { |
| 111 | if (v3d->ver < V3D_GEN_71) |
| 112 | return; |
| 113 | |
| 114 | V3D_SMS_WRITE(V3D_SMS_REE_CS, V3D_SET_FIELD(0x4, V3D_SMS_STATE)); |
| 115 | |
| 116 | if (wait_for(!(V3D_GET_FIELD(V3D_SMS_READ(V3D_SMS_REE_CS), |
| 117 | V3D_SMS_STATE) == V3D_SMS_ISOLATING_FOR_RESET) && |
| 118 | !(V3D_GET_FIELD(V3D_SMS_READ(V3D_SMS_REE_CS), |
| 119 | V3D_SMS_STATE) == V3D_SMS_RESETTING), 100)) { |
| 120 | DRM_ERROR("Failed to wait for SMS reset\n" ); |
| 121 | } |
| 122 | } |
| 123 | |
| 124 | void |
| 125 | v3d_reset(struct v3d_dev *v3d) |
| 126 | { |
| 127 | struct drm_device *dev = &v3d->drm; |
| 128 | |
| 129 | DRM_DEV_ERROR(dev->dev, "Resetting GPU for hang.\n" ); |
| 130 | DRM_DEV_ERROR(dev->dev, "V3D_ERR_STAT: 0x%08x\n" , |
| 131 | V3D_CORE_READ(0, V3D_ERR_STAT)); |
| 132 | trace_v3d_reset_begin(dev); |
| 133 | |
| 134 | /* XXX: only needed for safe powerdown, not reset. */ |
| 135 | if (false) |
| 136 | v3d_idle_axi(v3d, core: 0); |
| 137 | |
| 138 | v3d_irq_disable(v3d); |
| 139 | |
| 140 | v3d_idle_gca(v3d); |
| 141 | v3d_reset_sms(v3d); |
| 142 | v3d_reset_v3d(v3d); |
| 143 | |
| 144 | v3d_mmu_set_page_table(v3d); |
| 145 | v3d_irq_reset(v3d); |
| 146 | |
| 147 | v3d_perfmon_stop(v3d, perfmon: v3d->active_perfmon, capture: false); |
| 148 | |
| 149 | trace_v3d_reset_end(dev); |
| 150 | } |
| 151 | |
| 152 | static void |
| 153 | v3d_flush_l3(struct v3d_dev *v3d) |
| 154 | { |
| 155 | if (v3d->ver < V3D_GEN_41) { |
| 156 | u32 gca_ctrl = V3D_GCA_READ(V3D_GCA_CACHE_CTRL); |
| 157 | |
| 158 | V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL, |
| 159 | gca_ctrl | V3D_GCA_CACHE_CTRL_FLUSH); |
| 160 | |
| 161 | if (v3d->ver < V3D_GEN_33) { |
| 162 | V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL, |
| 163 | gca_ctrl & ~V3D_GCA_CACHE_CTRL_FLUSH); |
| 164 | } |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | /* Invalidates the (read-only) L2C cache. This was the L2 cache for |
| 169 | * uniforms and instructions on V3D 3.2. |
| 170 | */ |
| 171 | static void |
| 172 | v3d_invalidate_l2c(struct v3d_dev *v3d, int core) |
| 173 | { |
| 174 | if (v3d->ver >= V3D_GEN_33) |
| 175 | return; |
| 176 | |
| 177 | V3D_CORE_WRITE(core, V3D_CTL_L2CACTL, |
| 178 | V3D_L2CACTL_L2CCLR | |
| 179 | V3D_L2CACTL_L2CENA); |
| 180 | } |
| 181 | |
| 182 | /* Invalidates texture L2 cachelines */ |
| 183 | static void |
| 184 | v3d_flush_l2t(struct v3d_dev *v3d, int core) |
| 185 | { |
| 186 | /* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't |
| 187 | * need to wait for completion before dispatching the job -- |
| 188 | * L2T accesses will be stalled until the flush has completed. |
| 189 | * However, we do need to make sure we don't try to trigger a |
| 190 | * new flush while the L2_CLEAN queue is trying to |
| 191 | * synchronously clean after a job. |
| 192 | */ |
| 193 | mutex_lock(&v3d->cache_clean_lock); |
| 194 | V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, |
| 195 | V3D_L2TCACTL_L2TFLS | |
| 196 | V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM)); |
| 197 | mutex_unlock(lock: &v3d->cache_clean_lock); |
| 198 | } |
| 199 | |
| 200 | /* Cleans texture L1 and L2 cachelines (writing back dirty data). |
| 201 | * |
| 202 | * For cleaning, which happens from the CACHE_CLEAN queue after CSD has |
| 203 | * executed, we need to make sure that the clean is done before |
| 204 | * signaling job completion. So, we synchronously wait before |
| 205 | * returning, and we make sure that L2 invalidates don't happen in the |
| 206 | * meantime to confuse our are-we-done checks. |
| 207 | */ |
| 208 | void |
| 209 | v3d_clean_caches(struct v3d_dev *v3d) |
| 210 | { |
| 211 | struct drm_device *dev = &v3d->drm; |
| 212 | int core = 0; |
| 213 | |
| 214 | trace_v3d_cache_clean_begin(dev); |
| 215 | |
| 216 | V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF); |
| 217 | if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) & |
| 218 | V3D_L2TCACTL_TMUWCF), 100)) { |
| 219 | DRM_ERROR("Timeout waiting for TMU write combiner flush\n" ); |
| 220 | } |
| 221 | |
| 222 | mutex_lock(&v3d->cache_clean_lock); |
| 223 | V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, |
| 224 | V3D_L2TCACTL_L2TFLS | |
| 225 | V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM)); |
| 226 | |
| 227 | if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) & |
| 228 | V3D_L2TCACTL_L2TFLS), 100)) { |
| 229 | DRM_ERROR("Timeout waiting for L2T clean\n" ); |
| 230 | } |
| 231 | |
| 232 | mutex_unlock(lock: &v3d->cache_clean_lock); |
| 233 | |
| 234 | trace_v3d_cache_clean_end(dev); |
| 235 | } |
| 236 | |
| 237 | /* Invalidates the slice caches. These are read-only caches. */ |
| 238 | static void |
| 239 | v3d_invalidate_slices(struct v3d_dev *v3d, int core) |
| 240 | { |
| 241 | V3D_CORE_WRITE(core, V3D_CTL_SLCACTL, |
| 242 | V3D_SET_FIELD(0xf, V3D_SLCACTL_TVCCS) | |
| 243 | V3D_SET_FIELD(0xf, V3D_SLCACTL_TDCCS) | |
| 244 | V3D_SET_FIELD(0xf, V3D_SLCACTL_UCC) | |
| 245 | V3D_SET_FIELD(0xf, V3D_SLCACTL_ICC)); |
| 246 | } |
| 247 | |
| 248 | void |
| 249 | v3d_invalidate_caches(struct v3d_dev *v3d) |
| 250 | { |
| 251 | /* Invalidate the caches from the outside in. That way if |
| 252 | * another CL's concurrent use of nearby memory were to pull |
| 253 | * an invalidated cacheline back in, we wouldn't leave stale |
| 254 | * data in the inner cache. |
| 255 | */ |
| 256 | v3d_flush_l3(v3d); |
| 257 | v3d_invalidate_l2c(v3d, core: 0); |
| 258 | v3d_flush_l2t(v3d, core: 0); |
| 259 | v3d_invalidate_slices(v3d, core: 0); |
| 260 | } |
| 261 | |
| 262 | int |
| 263 | v3d_gem_init(struct drm_device *dev) |
| 264 | { |
| 265 | struct v3d_dev *v3d = to_v3d_dev(dev); |
| 266 | u32 pt_size = 4096 * 1024; |
| 267 | int ret, i; |
| 268 | |
| 269 | for (i = 0; i < V3D_MAX_QUEUES; i++) { |
| 270 | struct v3d_queue_state *queue = &v3d->queue[i]; |
| 271 | |
| 272 | queue->fence_context = dma_fence_context_alloc(num: 1); |
| 273 | memset(&queue->stats, 0, sizeof(queue->stats)); |
| 274 | seqcount_init(&queue->stats.lock); |
| 275 | |
| 276 | spin_lock_init(&queue->queue_lock); |
| 277 | spin_lock_init(&queue->fence_lock); |
| 278 | } |
| 279 | |
| 280 | spin_lock_init(&v3d->mm_lock); |
| 281 | ret = drmm_mutex_init(dev, &v3d->bo_lock); |
| 282 | if (ret) |
| 283 | return ret; |
| 284 | ret = drmm_mutex_init(dev, &v3d->reset_lock); |
| 285 | if (ret) |
| 286 | return ret; |
| 287 | ret = drmm_mutex_init(dev, &v3d->sched_lock); |
| 288 | if (ret) |
| 289 | return ret; |
| 290 | ret = drmm_mutex_init(dev, &v3d->cache_clean_lock); |
| 291 | if (ret) |
| 292 | return ret; |
| 293 | |
| 294 | /* Note: We don't allocate address 0. Various bits of HW |
| 295 | * treat 0 as special, such as the occlusion query counters |
| 296 | * where 0 means "disabled". |
| 297 | */ |
| 298 | drm_mm_init(mm: &v3d->mm, start: 1, size: pt_size / sizeof(u32) - 1); |
| 299 | |
| 300 | v3d->pt = dma_alloc_wc(dev: v3d->drm.dev, size: pt_size, |
| 301 | dma_addr: &v3d->pt_paddr, |
| 302 | GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO); |
| 303 | if (!v3d->pt) { |
| 304 | drm_mm_takedown(mm: &v3d->mm); |
| 305 | dev_err(v3d->drm.dev, |
| 306 | "Failed to allocate page tables. Please ensure you have DMA enabled.\n" ); |
| 307 | return -ENOMEM; |
| 308 | } |
| 309 | |
| 310 | v3d_init_hw_state(v3d); |
| 311 | v3d_mmu_set_page_table(v3d); |
| 312 | |
| 313 | v3d_gemfs_init(v3d); |
| 314 | |
| 315 | ret = v3d_sched_init(v3d); |
| 316 | if (ret) { |
| 317 | drm_mm_takedown(mm: &v3d->mm); |
| 318 | dma_free_coherent(dev: v3d->drm.dev, size: pt_size, cpu_addr: (void *)v3d->pt, |
| 319 | dma_handle: v3d->pt_paddr); |
| 320 | return ret; |
| 321 | } |
| 322 | |
| 323 | return 0; |
| 324 | } |
| 325 | |
| 326 | void |
| 327 | v3d_gem_destroy(struct drm_device *dev) |
| 328 | { |
| 329 | struct v3d_dev *v3d = to_v3d_dev(dev); |
| 330 | enum v3d_queue q; |
| 331 | |
| 332 | v3d_sched_fini(v3d); |
| 333 | v3d_gemfs_fini(v3d); |
| 334 | |
| 335 | /* Waiting for jobs to finish would need to be done before |
| 336 | * unregistering V3D. |
| 337 | */ |
| 338 | for (q = 0; q < V3D_MAX_QUEUES; q++) |
| 339 | WARN_ON(v3d->queue[q].active_job); |
| 340 | |
| 341 | drm_mm_takedown(mm: &v3d->mm); |
| 342 | |
| 343 | dma_free_coherent(dev: v3d->drm.dev, size: 4096 * 1024, cpu_addr: (void *)v3d->pt, |
| 344 | dma_handle: v3d->pt_paddr); |
| 345 | } |
| 346 | |