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_mmu_flush_all(v3d); 74 75 V3D_CORE_WRITE(0, V3D_PTB_BPOA, bo->node.start << V3D_MMU_PAGE_SHIFT); 76 V3D_CORE_WRITE(0, V3D_PTB_BPOS, obj->size); 77 78 out: 79 drm_gem_object_put(obj); 80 } 81 82 static irqreturn_t 83 v3d_irq(int irq, void *arg) 84 { 85 struct v3d_dev *v3d = arg; 86 u32 intsts; 87 irqreturn_t status = IRQ_NONE; 88 89 intsts = V3D_CORE_READ(0, V3D_CTL_INT_STS); 90 91 /* Acknowledge the interrupts we're handling here. */ 92 V3D_CORE_WRITE(0, V3D_CTL_INT_CLR, intsts); 93 94 if (intsts & V3D_INT_OUTOMEM) { 95 /* Note that the OOM status is edge signaled, so the 96 * interrupt won't happen again until the we actually 97 * add more memory. Also, as of V3D 4.1, FLDONE won't 98 * be reported until any OOM state has been cleared. 99 */ 100 schedule_work(&v3d->overflow_mem_work); 101 status = IRQ_HANDLED; 102 } 103 104 if (intsts & V3D_INT_FLDONE) { 105 struct v3d_fence *fence = 106 to_v3d_fence(v3d->bin_job->base.irq_fence); 107 108 v3d_job_update_stats(&v3d->bin_job->base, V3D_BIN); 109 trace_v3d_bcl_irq(&v3d->drm, fence->seqno); 110 111 v3d->bin_job = NULL; 112 dma_fence_signal(&fence->base); 113 114 status = IRQ_HANDLED; 115 } 116 117 if (intsts & V3D_INT_FRDONE) { 118 struct v3d_fence *fence = 119 to_v3d_fence(v3d->render_job->base.irq_fence); 120 121 v3d_job_update_stats(&v3d->render_job->base, V3D_RENDER); 122 trace_v3d_rcl_irq(&v3d->drm, fence->seqno); 123 124 v3d->render_job = NULL; 125 dma_fence_signal(&fence->base); 126 127 status = IRQ_HANDLED; 128 } 129 130 if (intsts & V3D_INT_CSDDONE(v3d->ver)) { 131 struct v3d_fence *fence = 132 to_v3d_fence(v3d->csd_job->base.irq_fence); 133 134 v3d_job_update_stats(&v3d->csd_job->base, V3D_CSD); 135 trace_v3d_csd_irq(&v3d->drm, fence->seqno); 136 137 v3d->csd_job = NULL; 138 dma_fence_signal(&fence->base); 139 140 status = IRQ_HANDLED; 141 } 142 143 /* We shouldn't be triggering these if we have GMP in 144 * always-allowed mode. 145 */ 146 if (v3d->ver < 71 && (intsts & V3D_INT_GMPV)) 147 dev_err(v3d->drm.dev, "GMP violation\n"); 148 149 /* V3D 4.2 wires the hub and core IRQs together, so if we & 150 * didn't see the common one then check hub for MMU IRQs. 151 */ 152 if (v3d->single_irq_line && status == IRQ_NONE) 153 return v3d_hub_irq(irq, arg); 154 155 return status; 156 } 157 158 static irqreturn_t 159 v3d_hub_irq(int irq, void *arg) 160 { 161 struct v3d_dev *v3d = arg; 162 u32 intsts; 163 irqreturn_t status = IRQ_NONE; 164 165 intsts = V3D_READ(V3D_HUB_INT_STS); 166 167 /* Acknowledge the interrupts we're handling here. */ 168 V3D_WRITE(V3D_HUB_INT_CLR, intsts); 169 170 if (intsts & V3D_HUB_INT_TFUC) { 171 struct v3d_fence *fence = 172 to_v3d_fence(v3d->tfu_job->base.irq_fence); 173 174 v3d_job_update_stats(&v3d->tfu_job->base, V3D_TFU); 175 trace_v3d_tfu_irq(&v3d->drm, fence->seqno); 176 177 v3d->tfu_job = NULL; 178 dma_fence_signal(&fence->base); 179 180 status = IRQ_HANDLED; 181 } 182 183 if (intsts & (V3D_HUB_INT_MMU_WRV | 184 V3D_HUB_INT_MMU_PTI | 185 V3D_HUB_INT_MMU_CAP)) { 186 u32 axi_id = V3D_READ(V3D_MMU_VIO_ID); 187 u64 vio_addr = ((u64)V3D_READ(V3D_MMU_VIO_ADDR) << 188 (v3d->va_width - 32)); 189 static const char *const v3d41_axi_ids[] = { 190 "L2T", 191 "PTB", 192 "PSE", 193 "TLB", 194 "CLE", 195 "TFU", 196 "MMU", 197 "GMP", 198 }; 199 const char *client = "?"; 200 201 V3D_WRITE(V3D_MMU_CTL, V3D_READ(V3D_MMU_CTL)); 202 203 if (v3d->ver >= 41) { 204 axi_id = axi_id >> 5; 205 if (axi_id < ARRAY_SIZE(v3d41_axi_ids)) 206 client = v3d41_axi_ids[axi_id]; 207 } 208 209 dev_err(v3d->drm.dev, "MMU error from client %s (%d) at 0x%llx%s%s%s\n", 210 client, axi_id, (long long)vio_addr, 211 ((intsts & V3D_HUB_INT_MMU_WRV) ? 212 ", write violation" : ""), 213 ((intsts & V3D_HUB_INT_MMU_PTI) ? 214 ", pte invalid" : ""), 215 ((intsts & V3D_HUB_INT_MMU_CAP) ? 216 ", cap exceeded" : "")); 217 status = IRQ_HANDLED; 218 } 219 220 if (v3d->ver >= 71 && (intsts & V3D_V7_HUB_INT_GMPV)) { 221 dev_err(v3d->drm.dev, "GMP Violation\n"); 222 status = IRQ_HANDLED; 223 } 224 225 return status; 226 } 227 228 int 229 v3d_irq_init(struct v3d_dev *v3d) 230 { 231 int irq1, ret, core; 232 233 INIT_WORK(&v3d->overflow_mem_work, v3d_overflow_mem_work); 234 235 /* Clear any pending interrupts someone might have left around 236 * for us. 237 */ 238 for (core = 0; core < v3d->cores; core++) 239 V3D_CORE_WRITE(core, V3D_CTL_INT_CLR, V3D_CORE_IRQS(v3d->ver)); 240 V3D_WRITE(V3D_HUB_INT_CLR, V3D_HUB_IRQS(v3d->ver)); 241 242 irq1 = platform_get_irq_optional(v3d_to_pdev(v3d), 1); 243 if (irq1 == -EPROBE_DEFER) 244 return irq1; 245 if (irq1 > 0) { 246 ret = devm_request_irq(v3d->drm.dev, irq1, 247 v3d_irq, IRQF_SHARED, 248 "v3d_core0", v3d); 249 if (ret) 250 goto fail; 251 ret = devm_request_irq(v3d->drm.dev, 252 platform_get_irq(v3d_to_pdev(v3d), 0), 253 v3d_hub_irq, IRQF_SHARED, 254 "v3d_hub", v3d); 255 if (ret) 256 goto fail; 257 } else { 258 v3d->single_irq_line = true; 259 260 ret = devm_request_irq(v3d->drm.dev, 261 platform_get_irq(v3d_to_pdev(v3d), 0), 262 v3d_irq, IRQF_SHARED, 263 "v3d", v3d); 264 if (ret) 265 goto fail; 266 } 267 268 v3d_irq_enable(v3d); 269 return 0; 270 271 fail: 272 if (ret != -EPROBE_DEFER) 273 dev_err(v3d->drm.dev, "IRQ setup failed: %d\n", ret); 274 return ret; 275 } 276 277 void 278 v3d_irq_enable(struct v3d_dev *v3d) 279 { 280 int core; 281 282 /* Enable our set of interrupts, masking out any others. */ 283 for (core = 0; core < v3d->cores; core++) { 284 V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_SET, ~V3D_CORE_IRQS(v3d->ver)); 285 V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_CLR, V3D_CORE_IRQS(v3d->ver)); 286 } 287 288 V3D_WRITE(V3D_HUB_INT_MSK_SET, ~V3D_HUB_IRQS(v3d->ver)); 289 V3D_WRITE(V3D_HUB_INT_MSK_CLR, V3D_HUB_IRQS(v3d->ver)); 290 } 291 292 void 293 v3d_irq_disable(struct v3d_dev *v3d) 294 { 295 int core; 296 297 /* Disable all interrupts. */ 298 for (core = 0; core < v3d->cores; core++) 299 V3D_CORE_WRITE(core, V3D_CTL_INT_MSK_SET, ~0); 300 V3D_WRITE(V3D_HUB_INT_MSK_SET, ~0); 301 302 /* Clear any pending interrupts we might have left. */ 303 for (core = 0; core < v3d->cores; core++) 304 V3D_CORE_WRITE(core, V3D_CTL_INT_CLR, V3D_CORE_IRQS(v3d->ver)); 305 V3D_WRITE(V3D_HUB_INT_CLR, V3D_HUB_IRQS(v3d->ver)); 306 307 cancel_work_sync(&v3d->overflow_mem_work); 308 } 309 310 /** Reinitializes interrupt registers when a GPU reset is performed. */ 311 void v3d_irq_reset(struct v3d_dev *v3d) 312 { 313 v3d_irq_enable(v3d); 314 } 315