1 // SPDX-License-Identifier: GPL-2.0+ 2 /* Copyright (C) 2014-2018 Broadcom */ 3 4 /** 5 * DOC: Interrupt management for the V3D engine 6 * 7 * When we take a bin, render, TFU done, or CSD done interrupt, we 8 * need to signal the fence for that job so that the scheduler can 9 * queue up the next one and unblock any waiters. 10 * 11 * When we take the binner out of memory interrupt, we need to 12 * allocate some new memory and pass it to the binner so that the 13 * current job can make progress. 14 */ 15 16 #include <linux/platform_device.h> 17 #include <linux/sched/clock.h> 18 19 #include "v3d_drv.h" 20 #include "v3d_regs.h" 21 #include "v3d_trace.h" 22 23 #define V3D_CORE_IRQS(ver) ((u32)(V3D_INT_OUTOMEM | \ 24 V3D_INT_FLDONE | \ 25 V3D_INT_FRDONE | \ 26 V3D_INT_CSDDONE(ver) | \ 27 (ver < 71 ? V3D_INT_GMPV : 0))) 28 29 #define V3D_HUB_IRQS(ver) ((u32)(V3D_HUB_INT_MMU_WRV | \ 30 V3D_HUB_INT_MMU_PTI | \ 31 V3D_HUB_INT_MMU_CAP | \ 32 V3D_HUB_INT_TFUC | \ 33 (ver >= 71 ? V3D_V7_HUB_INT_GMPV : 0))) 34 35 static irqreturn_t 36 v3d_hub_irq(int irq, void *arg); 37 38 static void 39 v3d_overflow_mem_work(struct work_struct *work) 40 { 41 struct v3d_dev *v3d = 42 container_of(work, struct v3d_dev, overflow_mem_work); 43 struct drm_device *dev = &v3d->drm; 44 struct v3d_bo *bo = v3d_bo_create(dev, NULL /* XXX: GMP */, 256 * 1024); 45 struct drm_gem_object *obj; 46 unsigned long irqflags; 47 48 if (IS_ERR(bo)) { 49 DRM_ERROR("Couldn't allocate binner overflow mem\n"); 50 return; 51 } 52 obj = &bo->base.base; 53 54 /* We lost a race, and our work task came in after the bin job 55 * completed and exited. This can happen because the HW 56 * signals OOM before it's fully OOM, so the binner might just 57 * barely complete. 58 * 59 * If we lose the race and our work task comes in after a new 60 * bin job got scheduled, that's fine. We'll just give them 61 * some binner pool anyway. 62 */ 63 spin_lock_irqsave(&v3d->job_lock, irqflags); 64 if (!v3d->bin_job) { 65 spin_unlock_irqrestore(&v3d->job_lock, irqflags); 66 goto out; 67 } 68 69 drm_gem_object_get(obj); 70 list_add_tail(&bo->unref_head, &v3d->bin_job->render->unref_list); 71 spin_unlock_irqrestore(&v3d->job_lock, irqflags); 72 73 V3D_CORE_WRITE(0, V3D_PTB_BPOA, bo->node.start << V3D_MMU_PAGE_SHIFT); 74 V3D_CORE_WRITE(0, V3D_PTB_BPOS, obj->size); 75 76 out: 77 drm_gem_object_put(obj); 78 } 79 80 static irqreturn_t 81 v3d_irq(int irq, void *arg) 82 { 83 struct v3d_dev *v3d = arg; 84 u32 intsts; 85 irqreturn_t status = IRQ_NONE; 86 87 intsts = V3D_CORE_READ(0, V3D_CTL_INT_STS); 88 89 /* Acknowledge the interrupts we're handling here. */ 90 V3D_CORE_WRITE(0, V3D_CTL_INT_CLR, intsts); 91 92 if (intsts & V3D_INT_OUTOMEM) { 93 /* Note that the OOM status is edge signaled, so the 94 * interrupt won't happen again until the we actually 95 * add more memory. Also, as of V3D 4.1, FLDONE won't 96 * be reported until any OOM state has been cleared. 97 */ 98 schedule_work(&v3d->overflow_mem_work); 99 status = IRQ_HANDLED; 100 } 101 102 if (intsts & V3D_INT_FLDONE) { 103 struct v3d_fence *fence = 104 to_v3d_fence(v3d->bin_job->base.irq_fence); 105 struct v3d_file_priv *file = v3d->bin_job->base.file->driver_priv; 106 u64 runtime = local_clock() - file->start_ns[V3D_BIN]; 107 108 file->enabled_ns[V3D_BIN] += local_clock() - file->start_ns[V3D_BIN]; 109 file->jobs_sent[V3D_BIN]++; 110 v3d->queue[V3D_BIN].jobs_sent++; 111 112 file->start_ns[V3D_BIN] = 0; 113 v3d->queue[V3D_BIN].start_ns = 0; 114 115 file->enabled_ns[V3D_BIN] += runtime; 116 v3d->queue[V3D_BIN].enabled_ns += runtime; 117 118 trace_v3d_bcl_irq(&v3d->drm, fence->seqno); 119 dma_fence_signal(&fence->base); 120 status = IRQ_HANDLED; 121 } 122 123 if (intsts & V3D_INT_FRDONE) { 124 struct v3d_fence *fence = 125 to_v3d_fence(v3d->render_job->base.irq_fence); 126 struct v3d_file_priv *file = v3d->render_job->base.file->driver_priv; 127 u64 runtime = local_clock() - file->start_ns[V3D_RENDER]; 128 129 file->enabled_ns[V3D_RENDER] += local_clock() - file->start_ns[V3D_RENDER]; 130 file->jobs_sent[V3D_RENDER]++; 131 v3d->queue[V3D_RENDER].jobs_sent++; 132 133 file->start_ns[V3D_RENDER] = 0; 134 v3d->queue[V3D_RENDER].start_ns = 0; 135 136 file->enabled_ns[V3D_RENDER] += runtime; 137 v3d->queue[V3D_RENDER].enabled_ns += runtime; 138 139 trace_v3d_rcl_irq(&v3d->drm, fence->seqno); 140 dma_fence_signal(&fence->base); 141 status = IRQ_HANDLED; 142 } 143 144 if (intsts & V3D_INT_CSDDONE(v3d->ver)) { 145 struct v3d_fence *fence = 146 to_v3d_fence(v3d->csd_job->base.irq_fence); 147 struct v3d_file_priv *file = v3d->csd_job->base.file->driver_priv; 148 u64 runtime = local_clock() - file->start_ns[V3D_CSD]; 149 150 file->enabled_ns[V3D_CSD] += local_clock() - file->start_ns[V3D_CSD]; 151 file->jobs_sent[V3D_CSD]++; 152 v3d->queue[V3D_CSD].jobs_sent++; 153 154 file->start_ns[V3D_CSD] = 0; 155 v3d->queue[V3D_CSD].start_ns = 0; 156 157 file->enabled_ns[V3D_CSD] += runtime; 158 v3d->queue[V3D_CSD].enabled_ns += runtime; 159 160 trace_v3d_csd_irq(&v3d->drm, fence->seqno); 161 dma_fence_signal(&fence->base); 162 status = IRQ_HANDLED; 163 } 164 165 /* We shouldn't be triggering these if we have GMP in 166 * always-allowed mode. 167 */ 168 if (v3d->ver < 71 && (intsts & V3D_INT_GMPV)) 169 dev_err(v3d->drm.dev, "GMP violation\n"); 170 171 /* V3D 4.2 wires the hub and core IRQs together, so if we & 172 * didn't see the common one then check hub for MMU IRQs. 173 */ 174 if (v3d->single_irq_line && status == IRQ_NONE) 175 return v3d_hub_irq(irq, arg); 176 177 return status; 178 } 179 180 static irqreturn_t 181 v3d_hub_irq(int irq, void *arg) 182 { 183 struct v3d_dev *v3d = arg; 184 u32 intsts; 185 irqreturn_t status = IRQ_NONE; 186 187 intsts = V3D_READ(V3D_HUB_INT_STS); 188 189 /* Acknowledge the interrupts we're handling here. */ 190 V3D_WRITE(V3D_HUB_INT_CLR, intsts); 191 192 if (intsts & V3D_HUB_INT_TFUC) { 193 struct v3d_fence *fence = 194 to_v3d_fence(v3d->tfu_job->base.irq_fence); 195 struct v3d_file_priv *file = v3d->tfu_job->base.file->driver_priv; 196 u64 runtime = local_clock() - file->start_ns[V3D_TFU]; 197 198 file->enabled_ns[V3D_TFU] += local_clock() - file->start_ns[V3D_TFU]; 199 file->jobs_sent[V3D_TFU]++; 200 v3d->queue[V3D_TFU].jobs_sent++; 201 202 file->start_ns[V3D_TFU] = 0; 203 v3d->queue[V3D_TFU].start_ns = 0; 204 205 file->enabled_ns[V3D_TFU] += runtime; 206 v3d->queue[V3D_TFU].enabled_ns += runtime; 207 208 trace_v3d_tfu_irq(&v3d->drm, fence->seqno); 209 dma_fence_signal(&fence->base); 210 status = IRQ_HANDLED; 211 } 212 213 if (intsts & (V3D_HUB_INT_MMU_WRV | 214 V3D_HUB_INT_MMU_PTI | 215 V3D_HUB_INT_MMU_CAP)) { 216 u32 axi_id = V3D_READ(V3D_MMU_VIO_ID); 217 u64 vio_addr = ((u64)V3D_READ(V3D_MMU_VIO_ADDR) << 218 (v3d->va_width - 32)); 219 static const char *const v3d41_axi_ids[] = { 220 "L2T", 221 "PTB", 222 "PSE", 223 "TLB", 224 "CLE", 225 "TFU", 226 "MMU", 227 "GMP", 228 }; 229 const char *client = "?"; 230 231 V3D_WRITE(V3D_MMU_CTL, V3D_READ(V3D_MMU_CTL)); 232 233 if (v3d->ver >= 41) { 234 axi_id = axi_id >> 5; 235 if (axi_id < ARRAY_SIZE(v3d41_axi_ids)) 236 client = v3d41_axi_ids[axi_id]; 237 } 238 239 dev_err(v3d->drm.dev, "MMU error from client %s (%d) at 0x%llx%s%s%s\n", 240 client, axi_id, (long long)vio_addr, 241 ((intsts & V3D_HUB_INT_MMU_WRV) ? 242 ", write violation" : ""), 243 ((intsts & V3D_HUB_INT_MMU_PTI) ? 244 ", pte invalid" : ""), 245 ((intsts & V3D_HUB_INT_MMU_CAP) ? 246 ", cap exceeded" : "")); 247 status = IRQ_HANDLED; 248 } 249 250 if (v3d->ver >= 71 && (intsts & V3D_V7_HUB_INT_GMPV)) { 251 dev_err(v3d->drm.dev, "GMP Violation\n"); 252 status = IRQ_HANDLED; 253 } 254 255 return status; 256 } 257 258 int 259 v3d_irq_init(struct v3d_dev *v3d) 260 { 261 int irq1, ret, core; 262 263 INIT_WORK(&v3d->overflow_mem_work, v3d_overflow_mem_work); 264 265 /* Clear any pending interrupts someone might have left around 266 * for us. 267 */ 268 for (core = 0; core < v3d->cores; core++) 269 V3D_CORE_WRITE(core, V3D_CTL_INT_CLR, V3D_CORE_IRQS(v3d->ver)); 270 V3D_WRITE(V3D_HUB_INT_CLR, V3D_HUB_IRQS(v3d->ver)); 271 272 irq1 = platform_get_irq_optional(v3d_to_pdev(v3d), 1); 273 if (irq1 == -EPROBE_DEFER) 274 return irq1; 275 if (irq1 > 0) { 276 ret = devm_request_irq(v3d->drm.dev, irq1, 277 v3d_irq, IRQF_SHARED, 278 "v3d_core0", v3d); 279 if (ret) 280 goto fail; 281 ret = devm_request_irq(v3d->drm.dev, 282 platform_get_irq(v3d_to_pdev(v3d), 0), 283 v3d_hub_irq, IRQF_SHARED, 284 "v3d_hub", v3d); 285 if (ret) 286 goto fail; 287 } else { 288 v3d->single_irq_line = true; 289 290 ret = devm_request_irq(v3d->drm.dev, 291 platform_get_irq(v3d_to_pdev(v3d), 0), 292 v3d_irq, IRQF_SHARED, 293 "v3d", v3d); 294 if (ret) 295 goto fail; 296 } 297 298 v3d_irq_enable(v3d); 299 return 0; 300 301 fail: 302 if (ret != -EPROBE_DEFER) 303 dev_err(v3d->drm.dev, "IRQ setup failed: %d\n", ret); 304 return ret; 305 } 306 307 void 308 v3d_irq_enable(struct v3d_dev *v3d) 309 { 310 int core; 311 312 /* Enable our set of interrupts, masking out any others. */ 313 for (core = 0; core < v3d->cores; core++) { 314 V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_SET, ~V3D_CORE_IRQS(v3d->ver)); 315 V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_CLR, V3D_CORE_IRQS(v3d->ver)); 316 } 317 318 V3D_WRITE(V3D_HUB_INT_MSK_SET, ~V3D_HUB_IRQS(v3d->ver)); 319 V3D_WRITE(V3D_HUB_INT_MSK_CLR, V3D_HUB_IRQS(v3d->ver)); 320 } 321 322 void 323 v3d_irq_disable(struct v3d_dev *v3d) 324 { 325 int core; 326 327 /* Disable all interrupts. */ 328 for (core = 0; core < v3d->cores; core++) 329 V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_SET, ~0); 330 V3D_WRITE(V3D_HUB_INT_MSK_SET, ~0); 331 332 /* Clear any pending interrupts we might have left. */ 333 for (core = 0; core < v3d->cores; core++) 334 V3D_CORE_WRITE(core, V3D_CTL_INT_CLR, V3D_CORE_IRQS(v3d->ver)); 335 V3D_WRITE(V3D_HUB_INT_CLR, V3D_HUB_IRQS(v3d->ver)); 336 337 cancel_work_sync(&v3d->overflow_mem_work); 338 } 339 340 /** Reinitializes interrupt registers when a GPU reset is performed. */ 341 void v3d_irq_reset(struct v3d_dev *v3d) 342 { 343 v3d_irq_enable(v3d); 344 } 345