1 /* 2 * Copyright 2025 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24 #include <linux/delay.h> 25 #include <linux/firmware.h> 26 #include <linux/module.h> 27 #include <linux/pci.h> 28 29 #include "amdgpu.h" 30 #include "amdgpu_ucode.h" 31 #include "amdgpu_trace.h" 32 33 #include "gc/gc_12_1_0_offset.h" 34 #include "gc/gc_12_1_0_sh_mask.h" 35 #include "ivsrcid/gfx/irqsrcs_gfx_12_1_0.h" 36 37 #include "soc15_common.h" 38 #include "soc15.h" 39 #include "sdma_v7_1_0_pkt_open.h" 40 #include "nbio_v4_3.h" 41 #include "sdma_common.h" 42 #include "sdma_v7_1.h" 43 #include "v12_structs.h" 44 #include "mes_userqueue.h" 45 46 MODULE_FIRMWARE("amdgpu/sdma_7_1_0.bin"); 47 48 #define SDMA1_REG_OFFSET 0x600 49 #define SDMA0_SDMA_IDX_0_END 0x450 50 #define SDMA1_HYP_DEC_REG_OFFSET 0x30 51 52 static const struct amdgpu_hwip_reg_entry sdma_reg_list_7_1[] = { 53 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_STATUS_REG), 54 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_STATUS1_REG), 55 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_STATUS2_REG), 56 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_STATUS3_REG), 57 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_STATUS4_REG), 58 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_STATUS5_REG), 59 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_STATUS6_REG), 60 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_UCODE_REV), 61 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_RB_RPTR_FETCH_HI), 62 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_RB_RPTR_FETCH), 63 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_UTCL1_RD_STATUS), 64 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_UTCL1_WR_STATUS), 65 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_UTCL1_RD_XNACK0), 66 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_UTCL1_RD_XNACK1), 67 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_UTCL1_WR_XNACK0), 68 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_UTCL1_WR_XNACK1), 69 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_RB_CNTL), 70 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_RB_RPTR), 71 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_RB_RPTR_HI), 72 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_RB_WPTR), 73 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_RB_WPTR_HI), 74 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_IB_OFFSET), 75 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_IB_BASE_LO), 76 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_IB_BASE_HI), 77 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_IB_CNTL), 78 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_IB_RPTR), 79 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_IB_SUB_REMAIN), 80 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE0_DUMMY_REG), 81 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE_STATUS0), 82 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_RB_CNTL), 83 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_RB_RPTR), 84 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_RB_RPTR_HI), 85 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_RB_WPTR), 86 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_RB_WPTR_HI), 87 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_IB_OFFSET), 88 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_IB_BASE_LO), 89 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_IB_BASE_HI), 90 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_IB_RPTR), 91 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_IB_SUB_REMAIN), 92 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE1_DUMMY_REG), 93 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_RB_CNTL), 94 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_RB_RPTR), 95 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_RB_RPTR_HI), 96 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_RB_WPTR), 97 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_RB_WPTR_HI), 98 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_IB_OFFSET), 99 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_IB_BASE_LO), 100 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_IB_BASE_HI), 101 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_IB_RPTR), 102 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_IB_SUB_REMAIN), 103 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_QUEUE2_DUMMY_REG), 104 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_INT_STATUS), 105 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_VM_CNTL), 106 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS2), 107 SOC15_REG_ENTRY_STR(GC, 0, regSDMA0_SDMA_CHICKEN_BITS), 108 }; 109 110 static void sdma_v7_1_set_ring_funcs(struct amdgpu_device *adev); 111 static void sdma_v7_1_set_buffer_funcs(struct amdgpu_device *adev); 112 static void sdma_v7_1_set_vm_pte_funcs(struct amdgpu_device *adev); 113 static void sdma_v7_1_set_irq_funcs(struct amdgpu_device *adev); 114 static int sdma_v7_1_inst_start(struct amdgpu_device *adev, 115 uint32_t inst_mask); 116 117 static u32 sdma_v7_1_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset) 118 { 119 u32 base; 120 u32 dev_inst = GET_INST(SDMA0, instance); 121 int xcc_id = adev->sdma.instance[instance].xcc_id; 122 int xcc_inst = dev_inst % adev->sdma.num_inst_per_xcc; 123 124 if (internal_offset >= SDMA0_SDMA_IDX_0_END) { 125 base = adev->reg_offset[GC_HWIP][xcc_id][1]; 126 if (xcc_inst != 0) 127 internal_offset += SDMA1_HYP_DEC_REG_OFFSET * xcc_inst; 128 } else { 129 base = adev->reg_offset[GC_HWIP][xcc_id][0]; 130 if (xcc_inst != 0) 131 internal_offset += SDMA1_REG_OFFSET * xcc_inst; 132 } 133 134 return base + internal_offset; 135 } 136 137 static unsigned sdma_v7_1_ring_init_cond_exec(struct amdgpu_ring *ring, 138 uint64_t addr) 139 { 140 unsigned ret; 141 142 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COND_EXE)); 143 amdgpu_ring_write(ring, lower_32_bits(addr)); 144 amdgpu_ring_write(ring, upper_32_bits(addr)); 145 amdgpu_ring_write(ring, 1); 146 /* this is the offset we need patch later */ 147 ret = ring->wptr & ring->buf_mask; 148 /* insert dummy here and patch it later */ 149 amdgpu_ring_write(ring, 0); 150 151 return ret; 152 } 153 154 /** 155 * sdma_v7_1_ring_get_rptr - get the current read pointer 156 * 157 * @ring: amdgpu ring pointer 158 * 159 * Get the current rptr from the hardware. 160 */ 161 static uint64_t sdma_v7_1_ring_get_rptr(struct amdgpu_ring *ring) 162 { 163 u64 *rptr; 164 165 /* XXX check if swapping is necessary on BE */ 166 rptr = (u64 *)ring->rptr_cpu_addr; 167 168 DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr); 169 return ((*rptr) >> 2); 170 } 171 172 /** 173 * sdma_v7_1_ring_get_wptr - get the current write pointer 174 * 175 * @ring: amdgpu ring pointer 176 * 177 * Get the current wptr from the hardware. 178 */ 179 static uint64_t sdma_v7_1_ring_get_wptr(struct amdgpu_ring *ring) 180 { 181 u64 wptr = 0; 182 183 if (ring->use_doorbell) { 184 /* XXX check if swapping is necessary on BE */ 185 wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr)); 186 DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr); 187 } 188 189 return wptr >> 2; 190 } 191 192 /** 193 * sdma_v7_1_ring_set_wptr - commit the write pointer 194 * 195 * @ring: amdgpu ring pointer 196 * 197 * Write the wptr back to the hardware. 198 */ 199 static void sdma_v7_1_ring_set_wptr(struct amdgpu_ring *ring) 200 { 201 struct amdgpu_device *adev = ring->adev; 202 203 DRM_DEBUG("Setting write pointer\n"); 204 205 if (ring->use_doorbell) { 206 DRM_DEBUG("Using doorbell -- " 207 "wptr_offs == 0x%08x " 208 "lower_32_bits(ring->wptr) << 2 == 0x%08x " 209 "upper_32_bits(ring->wptr) << 2 == 0x%08x\n", 210 ring->wptr_offs, 211 lower_32_bits(ring->wptr << 2), 212 upper_32_bits(ring->wptr << 2)); 213 /* XXX check if swapping is necessary on BE */ 214 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 215 ring->wptr << 2); 216 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n", 217 ring->doorbell_index, ring->wptr << 2); 218 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 219 } else { 220 DRM_DEBUG("Not using doorbell -- " 221 "regSDMA%i_GFX_RB_WPTR == 0x%08x " 222 "regSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n", 223 ring->me, 224 lower_32_bits(ring->wptr << 2), 225 ring->me, 226 upper_32_bits(ring->wptr << 2)); 227 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, 228 ring->me, 229 regSDMA0_SDMA_QUEUE0_RB_WPTR), 230 lower_32_bits(ring->wptr << 2)); 231 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, 232 ring->me, 233 regSDMA0_SDMA_QUEUE0_RB_WPTR_HI), 234 upper_32_bits(ring->wptr << 2)); 235 } 236 } 237 238 static void sdma_v7_1_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) 239 { 240 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 241 int i; 242 243 for (i = 0; i < count; i++) 244 if (sdma && sdma->burst_nop && (i == 0)) 245 amdgpu_ring_write(ring, ring->funcs->nop | 246 SDMA_PKT_NOP_HEADER_COUNT(count - 1)); 247 else 248 amdgpu_ring_write(ring, ring->funcs->nop); 249 } 250 251 /** 252 * sdma_v7_1_ring_emit_ib - Schedule an IB on the DMA engine 253 * 254 * @ring: amdgpu ring pointer 255 * @job: job to retrieve vmid from 256 * @ib: IB object to schedule 257 * @flags: unused 258 * 259 * Schedule an IB in the DMA ring. 260 */ 261 static void sdma_v7_1_ring_emit_ib(struct amdgpu_ring *ring, 262 struct amdgpu_job *job, 263 struct amdgpu_ib *ib, 264 uint32_t flags) 265 { 266 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 267 uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid); 268 269 /* An IB packet must end on a 8 DW boundary--the next dword 270 * must be on a 8-dword boundary. Our IB packet below is 6 271 * dwords long, thus add x number of NOPs, such that, in 272 * modular arithmetic, 273 * wptr + 6 + x = 8k, k >= 0, which in C is, 274 * (wptr + 6 + x) % 8 = 0. 275 * The expression below, is a solution of x. 276 */ 277 sdma_v7_1_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7); 278 279 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_INDIRECT) | 280 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf)); 281 /* base must be 32 byte aligned */ 282 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0); 283 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 284 amdgpu_ring_write(ring, ib->length_dw); 285 amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr)); 286 amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr)); 287 } 288 289 /** 290 * sdma_v7_1_ring_emit_mem_sync - flush the IB by graphics cache rinse 291 * 292 * @ring: amdgpu ring pointer 293 * 294 * flush the IB by graphics cache rinse. 295 */ 296 static void sdma_v7_1_ring_emit_mem_sync(struct amdgpu_ring *ring) 297 { 298 uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV | 299 SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV | 300 SDMA_GCR_GLI_INV(1); 301 302 /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */ 303 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_GCR_REQ)); 304 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0)); 305 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_56_32(0)); 306 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_0(gcr_cntl) | 307 SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_15_7(0)); 308 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_16(0)); 309 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD5_LIMIT_VA_56_48(0) | 310 SDMA_PKT_GCR_REQ_PAYLOAD5_VMID(0)); 311 } 312 313 314 /** 315 * sdma_v7_1_ring_emit_fence - emit a fence on the DMA ring 316 * 317 * @ring: amdgpu ring pointer 318 * @addr: address 319 * @seq: fence seq number 320 * @flags: fence flags 321 * 322 * Add a DMA fence packet to the ring to write 323 * the fence seq number and DMA trap packet to generate 324 * an interrupt if needed. 325 */ 326 static void sdma_v7_1_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq, 327 unsigned flags) 328 { 329 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 330 /* write the fence */ 331 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) | 332 SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */ 333 /* zero in first two bits */ 334 BUG_ON(addr & 0x3); 335 amdgpu_ring_write(ring, lower_32_bits(addr)); 336 amdgpu_ring_write(ring, upper_32_bits(addr)); 337 amdgpu_ring_write(ring, lower_32_bits(seq)); 338 339 /* optionally write high bits as well */ 340 if (write64bit) { 341 addr += 4; 342 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_FENCE) | 343 SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); 344 /* zero in first two bits */ 345 BUG_ON(addr & 0x3); 346 amdgpu_ring_write(ring, lower_32_bits(addr)); 347 amdgpu_ring_write(ring, upper_32_bits(addr)); 348 amdgpu_ring_write(ring, upper_32_bits(seq)); 349 } 350 351 if (flags & AMDGPU_FENCE_FLAG_INT) { 352 /* generate an interrupt */ 353 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_TRAP)); 354 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0)); 355 } 356 } 357 358 /** 359 * sdma_v7_1_inst_gfx_stop - stop the gfx async dma engines 360 * 361 * @adev: amdgpu_device pointer 362 * @inst_mask: mask of dma engine instances to be disabled 363 * 364 * Stop the gfx async dma ring buffers. 365 */ 366 static void sdma_v7_1_inst_gfx_stop(struct amdgpu_device *adev, 367 uint32_t inst_mask) 368 { 369 u32 rb_cntl, ib_cntl; 370 int i; 371 372 for (i = 0; i < NUM_XCC(inst_mask); i++) { 373 rb_cntl = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_CNTL)); 374 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, RB_ENABLE, 0); 375 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_CNTL), rb_cntl); 376 ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_IB_CNTL)); 377 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_SDMA_QUEUE0_IB_CNTL, IB_ENABLE, 0); 378 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_IB_CNTL), ib_cntl); 379 } 380 } 381 382 /** 383 * sdma_v7_1_inst_rlc_stop - stop the compute async dma engines 384 * 385 * @adev: amdgpu_device pointer 386 * @inst_mask: mask of dma engine instances to be disabled 387 * 388 * Stop the compute async dma queues. 389 */ 390 static void sdma_v7_1_inst_rlc_stop(struct amdgpu_device *adev, 391 uint32_t inst_mask) 392 { 393 /* XXX todo */ 394 } 395 396 /** 397 * sdma_v7_1_inst_ctx_switch_enable - stop the async dma engines context switch 398 * 399 * @adev: amdgpu_device pointer 400 * @enable: enable/disable the DMA MEs context switch. 401 * @inst_mask: mask of dma engine instances to be enabled 402 * 403 * Halt or unhalt the async dma engines context switch. 404 */ 405 static void sdma_v7_1_inst_ctx_switch_enable(struct amdgpu_device *adev, 406 bool enable, uint32_t inst_mask) 407 { 408 int i; 409 410 for_each_inst(i, inst_mask) { 411 WREG32_SOC15_IP(GC, 412 sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_UTCL1_TIMEOUT), 0x80); 413 } 414 } 415 416 /** 417 * sdma_v7_1_inst_enable - stop the async dma engines 418 * 419 * @adev: amdgpu_device pointer 420 * @enable: enable/disable the DMA MEs. 421 * @inst_mask: mask of dma engine instances to be enabled 422 * 423 * Halt or unhalt the async dma engines. 424 */ 425 static void sdma_v7_1_inst_enable(struct amdgpu_device *adev, 426 bool enable, uint32_t inst_mask) 427 { 428 u32 mcu_cntl; 429 int i; 430 431 if (!enable) { 432 sdma_v7_1_inst_gfx_stop(adev, inst_mask); 433 sdma_v7_1_inst_rlc_stop(adev, inst_mask); 434 } 435 436 if (amdgpu_sriov_vf(adev)) 437 return; 438 439 for (i = 0; i < NUM_XCC(inst_mask); i++) { 440 mcu_cntl = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_MCU_CNTL)); 441 mcu_cntl = REG_SET_FIELD(mcu_cntl, SDMA0_SDMA_MCU_CNTL, HALT, enable ? 0 : 1); 442 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_MCU_CNTL), mcu_cntl); 443 } 444 } 445 446 /** 447 * sdma_v7_1_gfx_resume_instance - start/restart a certain sdma engine 448 * 449 * @adev: amdgpu_device pointer 450 * @i: instance 451 * @restore: used to restore wptr when restart 452 * 453 * Set up the gfx DMA ring buffers and enable them. On restart, we will restore wptr and rptr. 454 * Return 0 for success. 455 */ 456 static int sdma_v7_1_gfx_resume_instance(struct amdgpu_device *adev, int i, bool restore) 457 { 458 struct amdgpu_ring *ring; 459 u32 rb_cntl, ib_cntl; 460 u32 rb_bufsz; 461 u32 doorbell; 462 u32 doorbell_offset; 463 u32 temp; 464 u64 wptr_gpu_addr; 465 int r; 466 467 ring = &adev->sdma.instance[i].ring; 468 469 /* Set ring buffer size in dwords */ 470 rb_bufsz = order_base_2(ring->ring_size / 4); 471 rb_cntl = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_CNTL)); 472 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, RB_SIZE, rb_bufsz); 473 #ifdef __BIG_ENDIAN 474 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, RB_SWAP_ENABLE, 1); 475 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, 476 RPTR_WRITEBACK_SWAP_ENABLE, 1); 477 #endif 478 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, RB_PRIV, 1); 479 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_CNTL), rb_cntl); 480 481 /* Initialize the ring buffer's read and write pointers */ 482 if (restore) { 483 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_RPTR), lower_32_bits(ring->wptr << 2)); 484 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_RPTR_HI), upper_32_bits(ring->wptr << 2)); 485 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_WPTR), lower_32_bits(ring->wptr << 2)); 486 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_WPTR_HI), upper_32_bits(ring->wptr << 2)); 487 } else { 488 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_RPTR), 0); 489 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_RPTR_HI), 0); 490 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_WPTR), 0); 491 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_WPTR_HI), 0); 492 } 493 /* setup the wptr shadow polling */ 494 wptr_gpu_addr = ring->wptr_gpu_addr; 495 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_WPTR_POLL_ADDR_LO), 496 lower_32_bits(wptr_gpu_addr)); 497 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_WPTR_POLL_ADDR_HI), 498 upper_32_bits(wptr_gpu_addr)); 499 500 /* set the wb address whether it's enabled or not */ 501 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_RPTR_ADDR_HI), 502 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF); 503 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_RPTR_ADDR_LO), 504 lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC); 505 506 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1); 507 if (amdgpu_sriov_vf(adev)) 508 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 1); 509 else 510 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, WPTR_POLL_ENABLE, 0); 511 512 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, MCU_WPTR_POLL_ENABLE, 1); 513 514 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_BASE), ring->gpu_addr >> 8); 515 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_BASE_HI), ring->gpu_addr >> 40); 516 517 if (!restore) 518 ring->wptr = 0; 519 520 /* before programing wptr to a less value, need set minor_ptr_update first */ 521 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_MINOR_PTR_UPDATE), 1); 522 523 if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */ 524 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_WPTR), lower_32_bits(ring->wptr) << 2); 525 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_WPTR_HI), upper_32_bits(ring->wptr) << 2); 526 } 527 528 doorbell = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_DOORBELL)); 529 doorbell_offset = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_DOORBELL_OFFSET)); 530 531 if (ring->use_doorbell) { 532 doorbell = REG_SET_FIELD(doorbell, SDMA0_SDMA_QUEUE0_DOORBELL, ENABLE, 1); 533 doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_SDMA_QUEUE0_DOORBELL_OFFSET, 534 OFFSET, ring->doorbell_index); 535 } else { 536 doorbell = REG_SET_FIELD(doorbell, SDMA0_SDMA_QUEUE0_DOORBELL, ENABLE, 0); 537 } 538 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_DOORBELL), doorbell); 539 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_DOORBELL_OFFSET), doorbell_offset); 540 541 if (i == 0) 542 adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell, 543 ring->doorbell_index, 544 adev->doorbell_index.sdma_doorbell_range * adev->sdma.num_instances); 545 546 if (amdgpu_sriov_vf(adev)) 547 sdma_v7_1_ring_set_wptr(ring); 548 549 /* set minor_ptr_update to 0 after wptr programed */ 550 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_MINOR_PTR_UPDATE), 0); 551 552 /* Set up sdma hang watchdog */ 553 temp = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_WATCHDOG_CNTL)); 554 /* 100ms per unit */ 555 temp = REG_SET_FIELD(temp, SDMA0_SDMA_WATCHDOG_CNTL, QUEUE_HANG_COUNT, 556 max(adev->usec_timeout/100000, 1)); 557 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_WATCHDOG_CNTL), temp); 558 559 /* Set up RESP_MODE to non-copy addresses */ 560 temp = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_UTCL1_CNTL)); 561 temp = REG_SET_FIELD(temp, SDMA0_SDMA_UTCL1_CNTL, RESP_MODE, 3); 562 temp = REG_SET_FIELD(temp, SDMA0_SDMA_UTCL1_CNTL, REDO_DELAY, 9); 563 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_UTCL1_CNTL), temp); 564 565 /* program default cache read and write policy */ 566 temp = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_UTCL1_PAGE)); 567 /* clean read policy and write policy bits */ 568 temp &= 0xFF0FFF; 569 temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | 570 (CACHE_WRITE_POLICY_L2__DEFAULT << 14)); 571 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_UTCL1_PAGE), temp); 572 573 if (!amdgpu_sriov_vf(adev)) { 574 /* unhalt engine */ 575 temp = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_MCU_CNTL)); 576 temp = REG_SET_FIELD(temp, SDMA0_SDMA_MCU_CNTL, HALT, 0); 577 temp = REG_SET_FIELD(temp, SDMA0_SDMA_MCU_CNTL, RESET, 0); 578 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_MCU_CNTL), temp); 579 } 580 581 /* enable DMA RB */ 582 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_SDMA_QUEUE0_RB_CNTL, RB_ENABLE, 1); 583 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_RB_CNTL), rb_cntl); 584 585 ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_IB_CNTL)); 586 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_SDMA_QUEUE0_IB_CNTL, IB_ENABLE, 1); 587 #ifdef __BIG_ENDIAN 588 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_SDMA_QUEUE0_IB_CNTL, IB_SWAP_ENABLE, 1); 589 #endif 590 /* enable DMA IBs */ 591 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_IB_CNTL), ib_cntl); 592 ring->sched.ready = true; 593 594 if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */ 595 sdma_v7_1_inst_ctx_switch_enable(adev, true, i); 596 sdma_v7_1_inst_enable(adev, true, i); 597 } 598 599 r = amdgpu_ring_test_helper(ring); 600 if (r) 601 ring->sched.ready = false; 602 603 return r; 604 } 605 606 /** 607 * sdma_v7_1_inst_gfx_resume - setup and start the async dma engines 608 * 609 * @adev: amdgpu_device pointer 610 * @inst_mask: mask of dma engine instances to be enabled 611 * 612 * Set up the gfx DMA ring buffers and enable them. 613 * Returns 0 for success, error for failure. 614 */ 615 static int sdma_v7_1_inst_gfx_resume(struct amdgpu_device *adev, 616 uint32_t inst_mask) 617 { 618 int i, r; 619 620 for (i = 0; i < NUM_XCC(inst_mask); i++) { 621 r = sdma_v7_1_gfx_resume_instance(adev, i, false); 622 if (r) 623 return r; 624 } 625 626 return 0; 627 628 } 629 630 /** 631 * sdma_v7_1_inst_rlc_resume - setup and start the async dma engines 632 * 633 * @adev: amdgpu_device pointer 634 * @inst_mask: mask of dma engine instances to be enabled 635 * 636 * Set up the compute DMA queues and enable them. 637 * Returns 0 for success, error for failure. 638 */ 639 static int sdma_v7_1_inst_rlc_resume(struct amdgpu_device *adev, 640 uint32_t inst_mask) 641 { 642 return 0; 643 } 644 645 static void sdma_v7_1_inst_free_ucode_buffer(struct amdgpu_device *adev, 646 uint32_t inst_mask) 647 { 648 int i; 649 650 for (i = 0; i < NUM_XCC(inst_mask); i++) { 651 amdgpu_bo_free_kernel(&adev->sdma.instance[i].sdma_fw_obj, 652 &adev->sdma.instance[i].sdma_fw_gpu_addr, 653 (void **)&adev->sdma.instance[i].sdma_fw_ptr); 654 } 655 } 656 657 /** 658 * sdma_v7_1_inst_load_microcode - load the sDMA ME ucode 659 * 660 * @adev: amdgpu_device pointer 661 * @inst_mask: mask of dma engine instances to be enabled 662 * 663 * Loads the sDMA0/1 ucode. 664 * Returns 0 for success, -EINVAL if the ucode is not available. 665 */ 666 static int sdma_v7_1_inst_load_microcode(struct amdgpu_device *adev, 667 uint32_t inst_mask) 668 { 669 const struct sdma_firmware_header_v3_0 *hdr; 670 const __le32 *fw_data; 671 u32 fw_size; 672 uint32_t tmp, sdma_status, ic_op_cntl; 673 int i, r, j; 674 675 /* halt the MEs */ 676 sdma_v7_1_inst_enable(adev, false, inst_mask); 677 678 if (!adev->sdma.instance[0].fw) 679 return -EINVAL; 680 681 hdr = (const struct sdma_firmware_header_v3_0 *) 682 adev->sdma.instance[0].fw->data; 683 amdgpu_ucode_print_sdma_hdr(&hdr->header); 684 685 fw_data = (const __le32 *)(adev->sdma.instance[0].fw->data + 686 le32_to_cpu(hdr->ucode_offset_bytes)); 687 fw_size = le32_to_cpu(hdr->ucode_size_bytes); 688 689 for (i = 0; i < NUM_XCC(inst_mask); i++) { 690 r = amdgpu_bo_create_reserved(adev, fw_size, 691 PAGE_SIZE, 692 AMDGPU_GEM_DOMAIN_VRAM, 693 &adev->sdma.instance[i].sdma_fw_obj, 694 &adev->sdma.instance[i].sdma_fw_gpu_addr, 695 (void **)&adev->sdma.instance[i].sdma_fw_ptr); 696 if (r) { 697 dev_err(adev->dev, "(%d) failed to create sdma ucode bo\n", r); 698 return r; 699 } 700 701 memcpy(adev->sdma.instance[i].sdma_fw_ptr, fw_data, fw_size); 702 703 amdgpu_bo_kunmap(adev->sdma.instance[i].sdma_fw_obj); 704 amdgpu_bo_unreserve(adev->sdma.instance[i].sdma_fw_obj); 705 706 tmp = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_IC_CNTL)); 707 tmp = REG_SET_FIELD(tmp, SDMA0_SDMA_IC_CNTL, GPA, 0); 708 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_IC_CNTL), tmp); 709 710 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_IC_BASE_LO), 711 lower_32_bits(adev->sdma.instance[i].sdma_fw_gpu_addr)); 712 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_IC_BASE_HI), 713 upper_32_bits(adev->sdma.instance[i].sdma_fw_gpu_addr)); 714 715 tmp = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_IC_OP_CNTL)); 716 tmp = REG_SET_FIELD(tmp, SDMA0_SDMA_IC_OP_CNTL, PRIME_ICACHE, 1); 717 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_IC_OP_CNTL), tmp); 718 719 /* Wait for sdma ucode init complete */ 720 for (j = 0; j < adev->usec_timeout; j++) { 721 ic_op_cntl = RREG32_SOC15_IP(GC, 722 sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_IC_OP_CNTL)); 723 sdma_status = RREG32_SOC15_IP(GC, 724 sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_STATUS_REG)); 725 if ((REG_GET_FIELD(ic_op_cntl, SDMA0_SDMA_IC_OP_CNTL, ICACHE_PRIMED) == 1) && 726 (REG_GET_FIELD(sdma_status, SDMA0_SDMA_STATUS_REG, UCODE_INIT_DONE) == 1)) 727 break; 728 udelay(1); 729 } 730 731 if (j >= adev->usec_timeout) { 732 dev_err(adev->dev, "failed to init sdma ucode\n"); 733 return -EINVAL; 734 } 735 } 736 737 return 0; 738 } 739 740 static int sdma_v7_1_soft_reset(struct amdgpu_ip_block *ip_block) 741 { 742 struct amdgpu_device *adev = ip_block->adev; 743 uint32_t inst_mask; 744 u32 tmp; 745 int i; 746 747 inst_mask = adev->sdma.sdma_mask; 748 sdma_v7_1_inst_gfx_stop(adev, inst_mask); 749 750 for (i = 0; i < NUM_XCC(inst_mask); i++) { 751 //tmp = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_FREEZE)); 752 //tmp |= SDMA0_SDMA_FREEZE__FREEZE_MASK; 753 //WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_FREEZE), tmp); 754 tmp = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_MCU_CNTL)); 755 tmp |= SDMA0_SDMA_MCU_CNTL__HALT_MASK; 756 tmp |= SDMA0_SDMA_MCU_CNTL__RESET_MASK; 757 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_MCU_CNTL), tmp); 758 759 WREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_QUEUE0_PREEMPT), 0); 760 761 udelay(100); 762 763 tmp = GRBM_SOFT_RESET__SOFT_RESET_SDMA0_MASK << i; 764 WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, tmp); 765 tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET); 766 767 udelay(100); 768 769 WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, 0); 770 tmp = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET); 771 772 udelay(100); 773 } 774 775 return sdma_v7_1_inst_start(adev, inst_mask); 776 } 777 778 static bool sdma_v7_1_check_soft_reset(struct amdgpu_ip_block *ip_block) 779 { 780 struct amdgpu_device *adev = ip_block->adev; 781 struct amdgpu_ring *ring; 782 int i, r; 783 long tmo = msecs_to_jiffies(1000); 784 785 for (i = 0; i < adev->sdma.num_instances; i++) { 786 ring = &adev->sdma.instance[i].ring; 787 r = amdgpu_ring_test_ib(ring, tmo); 788 if (r) 789 return true; 790 } 791 792 return false; 793 } 794 795 static int sdma_v7_1_reset_queue(struct amdgpu_ring *ring, 796 unsigned int vmid, 797 struct amdgpu_fence *timedout_fence) 798 { 799 struct amdgpu_device *adev = ring->adev; 800 int r; 801 802 if (ring->me >= adev->sdma.num_instances) { 803 dev_err(adev->dev, "sdma instance not found\n"); 804 return -EINVAL; 805 } 806 807 amdgpu_ring_reset_helper_begin(ring, timedout_fence); 808 809 r = amdgpu_mes_reset_legacy_queue(adev, ring, vmid, true, 0); 810 if (r) 811 return r; 812 813 r = sdma_v7_1_gfx_resume_instance(adev, ring->me, true); 814 if (r) 815 return r; 816 817 return amdgpu_ring_reset_helper_end(ring, timedout_fence); 818 } 819 820 /** 821 * sdma_v7_1_inst_start - setup and start the async dma engines 822 * 823 * @adev: amdgpu_device pointer 824 * @inst_mask: mask of dma engine instances to be enabled 825 * 826 * Set up the DMA engines and enable them. 827 * Returns 0 for success, error for failure. 828 */ 829 static int sdma_v7_1_inst_start(struct amdgpu_device *adev, 830 uint32_t inst_mask) 831 { 832 int r = 0; 833 834 if (amdgpu_sriov_vf(adev)) { 835 sdma_v7_1_inst_ctx_switch_enable(adev, false, inst_mask); 836 sdma_v7_1_inst_enable(adev, false, inst_mask); 837 838 /* set RB registers */ 839 r = sdma_v7_1_inst_gfx_resume(adev, inst_mask); 840 return r; 841 } 842 843 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 844 r = sdma_v7_1_inst_load_microcode(adev, inst_mask); 845 if (r) { 846 sdma_v7_1_inst_free_ucode_buffer(adev, inst_mask); 847 return r; 848 } 849 850 if (amdgpu_emu_mode == 1) 851 msleep(1000); 852 } 853 854 /* unhalt the MEs */ 855 sdma_v7_1_inst_enable(adev, true, inst_mask); 856 /* enable sdma ring preemption */ 857 sdma_v7_1_inst_ctx_switch_enable(adev, true, inst_mask); 858 859 /* start the gfx rings and rlc compute queues */ 860 r = sdma_v7_1_inst_gfx_resume(adev, inst_mask); 861 if (r) 862 return r; 863 r = sdma_v7_1_inst_rlc_resume(adev, inst_mask); 864 865 return r; 866 } 867 868 static int sdma_v7_1_mqd_init(struct amdgpu_device *adev, void *mqd, 869 struct amdgpu_mqd_prop *prop) 870 { 871 struct v12_sdma_mqd *m = mqd; 872 uint64_t wb_gpu_addr; 873 874 m->sdmax_rlcx_rb_cntl = 875 order_base_2(prop->queue_size / 4) << SDMA0_SDMA_QUEUE0_RB_CNTL__RB_SIZE__SHIFT | 876 1 << SDMA0_SDMA_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | 877 4 << SDMA0_SDMA_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT | 878 1 << SDMA0_SDMA_QUEUE0_RB_CNTL__MCU_WPTR_POLL_ENABLE__SHIFT; 879 880 m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8); 881 m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8); 882 883 wb_gpu_addr = prop->wptr_gpu_addr; 884 m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr); 885 m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr); 886 887 wb_gpu_addr = prop->rptr_gpu_addr; 888 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr); 889 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr); 890 891 m->sdmax_rlcx_ib_cntl = RREG32_SOC15_IP(GC, sdma_v7_1_get_reg_offset(adev, 0, 892 regSDMA0_SDMA_QUEUE0_IB_CNTL)); 893 894 m->sdmax_rlcx_doorbell_offset = 895 prop->doorbell_index << SDMA0_SDMA_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT; 896 897 m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_SDMA_QUEUE0_DOORBELL, ENABLE, 1); 898 899 m->sdmax_rlcx_doorbell_log = 0; 900 m->sdmax_rlcx_rb_aql_cntl = 0x4000; //regSDMA0_SDMA_QUEUE0_RB_AQL_CNTL_DEFAULT; 901 m->sdmax_rlcx_dummy_reg = 0xf; //regSDMA0_SDMA_QUEUE0_DUMMY_REG_DEFAULT; 902 903 m->sdmax_rlcx_csa_addr_lo = lower_32_bits(prop->csa_addr); 904 m->sdmax_rlcx_csa_addr_hi = upper_32_bits(prop->csa_addr); 905 906 return 0; 907 } 908 909 static void sdma_v7_1_set_mqd_funcs(struct amdgpu_device *adev) 910 { 911 adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v12_sdma_mqd); 912 adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v7_1_mqd_init; 913 } 914 915 /** 916 * sdma_v7_1_ring_test_ring - simple async dma engine test 917 * 918 * @ring: amdgpu_ring structure holding ring information 919 * 920 * Test the DMA engine by writing using it to write an 921 * value to memory. 922 * Returns 0 for success, error for failure. 923 */ 924 static int sdma_v7_1_ring_test_ring(struct amdgpu_ring *ring) 925 { 926 struct amdgpu_device *adev = ring->adev; 927 unsigned i; 928 unsigned index; 929 int r; 930 u32 tmp; 931 u64 gpu_addr; 932 933 tmp = 0xCAFEDEAD; 934 935 r = amdgpu_device_wb_get(adev, &index); 936 if (r) { 937 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r); 938 return r; 939 } 940 941 gpu_addr = adev->wb.gpu_addr + (index * 4); 942 adev->wb.wb[index] = cpu_to_le32(tmp); 943 944 r = amdgpu_ring_alloc(ring, 5); 945 if (r) { 946 DRM_ERROR("amdgpu: dma failed to lock ring %d (%d).\n", ring->idx, r); 947 amdgpu_device_wb_free(adev, index); 948 return r; 949 } 950 951 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) | 952 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR)); 953 amdgpu_ring_write(ring, lower_32_bits(gpu_addr)); 954 amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); 955 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0)); 956 amdgpu_ring_write(ring, 0xDEADBEEF); 957 amdgpu_ring_commit(ring); 958 959 for (i = 0; i < adev->usec_timeout; i++) { 960 tmp = le32_to_cpu(adev->wb.wb[index]); 961 if (tmp == 0xDEADBEEF) 962 break; 963 if (amdgpu_emu_mode == 1) 964 msleep(1); 965 else 966 udelay(1); 967 } 968 969 if (i >= adev->usec_timeout) 970 r = -ETIMEDOUT; 971 972 amdgpu_device_wb_free(adev, index); 973 974 return r; 975 } 976 977 /** 978 * sdma_v7_1_ring_test_ib - test an IB on the DMA engine 979 * 980 * @ring: amdgpu_ring structure holding ring information 981 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT 982 * 983 * Test a simple IB in the DMA ring. 984 * Returns 0 on success, error on failure. 985 */ 986 static int sdma_v7_1_ring_test_ib(struct amdgpu_ring *ring, long timeout) 987 { 988 struct amdgpu_device *adev = ring->adev; 989 struct amdgpu_ib ib; 990 struct dma_fence *f = NULL; 991 unsigned index; 992 long r; 993 u32 tmp = 0; 994 u64 gpu_addr; 995 996 tmp = 0xCAFEDEAD; 997 memset(&ib, 0, sizeof(ib)); 998 999 r = amdgpu_device_wb_get(adev, &index); 1000 if (r) { 1001 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r); 1002 return r; 1003 } 1004 1005 gpu_addr = adev->wb.gpu_addr + (index * 4); 1006 adev->wb.wb[index] = cpu_to_le32(tmp); 1007 1008 r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib); 1009 if (r) { 1010 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r); 1011 goto err0; 1012 } 1013 1014 ib.ptr[0] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) | 1015 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1016 ib.ptr[1] = lower_32_bits(gpu_addr); 1017 ib.ptr[2] = upper_32_bits(gpu_addr); 1018 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0); 1019 ib.ptr[4] = 0xDEADBEEF; 1020 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1021 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1022 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1023 ib.length_dw = 8; 1024 1025 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 1026 if (r) 1027 goto err1; 1028 1029 r = dma_fence_wait_timeout(f, false, timeout); 1030 if (r == 0) { 1031 DRM_ERROR("amdgpu: IB test timed out\n"); 1032 r = -ETIMEDOUT; 1033 goto err1; 1034 } else if (r < 0) { 1035 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r); 1036 goto err1; 1037 } 1038 1039 tmp = le32_to_cpu(adev->wb.wb[index]); 1040 1041 if (tmp == 0xDEADBEEF) 1042 r = 0; 1043 else 1044 r = -EINVAL; 1045 1046 err1: 1047 amdgpu_ib_free(&ib, NULL); 1048 dma_fence_put(f); 1049 err0: 1050 amdgpu_device_wb_free(adev, index); 1051 return r; 1052 } 1053 1054 1055 /** 1056 * sdma_v7_1_vm_copy_pte - update PTEs by copying them from the GART 1057 * 1058 * @ib: indirect buffer to fill with commands 1059 * @pe: addr of the page entry 1060 * @src: src addr to copy from 1061 * @count: number of page entries to update 1062 * 1063 * Update PTEs by copying them from the GART using sDMA. 1064 */ 1065 static void sdma_v7_1_vm_copy_pte(struct amdgpu_ib *ib, 1066 uint64_t pe, uint64_t src, 1067 unsigned count) 1068 { 1069 unsigned bytes = count * 8; 1070 1071 ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) | 1072 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); 1073 1074 ib->ptr[ib->length_dw++] = bytes - 1; 1075 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 1076 ib->ptr[ib->length_dw++] = lower_32_bits(src); 1077 ib->ptr[ib->length_dw++] = upper_32_bits(src); 1078 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1079 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1080 1081 } 1082 1083 /** 1084 * sdma_v7_1_vm_write_pte - update PTEs by writing them manually 1085 * 1086 * @ib: indirect buffer to fill with commands 1087 * @pe: addr of the page entry 1088 * @value: dst addr to write into pe 1089 * @count: number of page entries to update 1090 * @incr: increase next addr by incr bytes 1091 * 1092 * Update PTEs by writing them manually using sDMA. 1093 */ 1094 static void sdma_v7_1_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe, 1095 uint64_t value, unsigned count, 1096 uint32_t incr) 1097 { 1098 unsigned ndw = count * 2; 1099 1100 ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_WRITE) | 1101 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1102 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1103 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1104 ib->ptr[ib->length_dw++] = ndw - 1; 1105 for (; ndw > 0; ndw -= 2) { 1106 ib->ptr[ib->length_dw++] = lower_32_bits(value); 1107 ib->ptr[ib->length_dw++] = upper_32_bits(value); 1108 value += incr; 1109 } 1110 } 1111 1112 /** 1113 * sdma_v7_1_vm_set_pte_pde - update the page tables using sDMA 1114 * 1115 * @ib: indirect buffer to fill with commands 1116 * @pe: addr of the page entry 1117 * @addr: dst addr to write into pe 1118 * @count: number of page entries to update 1119 * @incr: increase next addr by incr bytes 1120 * @flags: access flags 1121 * 1122 * Update the page tables using sDMA. 1123 */ 1124 static void sdma_v7_1_vm_set_pte_pde(struct amdgpu_ib *ib, 1125 uint64_t pe, 1126 uint64_t addr, unsigned count, 1127 uint32_t incr, uint64_t flags) 1128 { 1129 /* for physically contiguous pages (vram) */ 1130 u32 header = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_PTEPDE); 1131 1132 if (amdgpu_mtype_local) 1133 header |= SDMA_PKT_PTEPDE_COPY_HEADER_MTYPE(0x3); 1134 else 1135 header |= (SDMA_PKT_PTEPDE_COPY_HEADER_MTYPE(0x2) | 1136 SDMA_PKT_PTEPDE_COPY_HEADER_SNOOP(0x1) | 1137 SDMA_PKT_PTEPDE_COPY_HEADER_SCOPE(0x3)); 1138 1139 ib->ptr[ib->length_dw++] = header; 1140 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */ 1141 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1142 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */ 1143 ib->ptr[ib->length_dw++] = upper_32_bits(flags); 1144 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */ 1145 ib->ptr[ib->length_dw++] = upper_32_bits(addr); 1146 ib->ptr[ib->length_dw++] = incr; /* increment size */ 1147 ib->ptr[ib->length_dw++] = 0; 1148 ib->ptr[ib->length_dw++] = count - 1; /* number of entries */ 1149 } 1150 1151 /** 1152 * sdma_v7_1_ring_pad_ib - pad the IB 1153 * 1154 * @ring: amdgpu ring pointer 1155 * @ib: indirect buffer to fill with padding 1156 * 1157 * Pad the IB with NOPs to a boundary multiple of 8. 1158 */ 1159 static void sdma_v7_1_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib) 1160 { 1161 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 1162 u32 pad_count; 1163 int i; 1164 1165 pad_count = (-ib->length_dw) & 0x7; 1166 for (i = 0; i < pad_count; i++) 1167 if (sdma && sdma->burst_nop && (i == 0)) 1168 ib->ptr[ib->length_dw++] = 1169 SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP) | 1170 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1); 1171 else 1172 ib->ptr[ib->length_dw++] = 1173 SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_NOP); 1174 } 1175 1176 /** 1177 * sdma_v7_1_ring_emit_pipeline_sync - sync the pipeline 1178 * 1179 * @ring: amdgpu_ring pointer 1180 * 1181 * Make sure all previous operations are completed (CIK). 1182 */ 1183 static void sdma_v7_1_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 1184 { 1185 uint32_t seq = ring->fence_drv.sync_seq; 1186 uint64_t addr = ring->fence_drv.gpu_addr; 1187 1188 /* wait for idle */ 1189 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) | 1190 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */ 1191 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)); 1192 amdgpu_ring_write(ring, addr & 0xfffffffc); 1193 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff); 1194 amdgpu_ring_write(ring, seq); /* reference */ 1195 amdgpu_ring_write(ring, 0xffffffff); /* mask */ 1196 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 1197 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */ 1198 } 1199 1200 /** 1201 * sdma_v7_1_ring_emit_vm_flush - vm flush using sDMA 1202 * 1203 * @ring: amdgpu_ring pointer 1204 * @vmid: vmid number to use 1205 * @pd_addr: address 1206 * 1207 * Update the page table base and flush the VM TLB 1208 * using sDMA. 1209 */ 1210 static void sdma_v7_1_ring_emit_vm_flush(struct amdgpu_ring *ring, 1211 unsigned vmid, uint64_t pd_addr) 1212 { 1213 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr); 1214 } 1215 1216 static void sdma_v7_1_ring_emit_wreg(struct amdgpu_ring *ring, 1217 uint32_t reg, uint32_t val) 1218 { 1219 /* SRBM WRITE command will not support on sdma v7. 1220 * Use Register WRITE command instead, which OPCODE is same as SRBM WRITE 1221 */ 1222 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_SRBM_WRITE)); 1223 amdgpu_ring_write(ring, reg << 2); 1224 amdgpu_ring_write(ring, val); 1225 } 1226 1227 static void sdma_v7_1_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg, 1228 uint32_t val, uint32_t mask) 1229 { 1230 amdgpu_ring_write(ring, SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_POLL_REGMEM) | 1231 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* equal */ 1232 amdgpu_ring_write(ring, reg << 2); 1233 amdgpu_ring_write(ring, 0); 1234 amdgpu_ring_write(ring, val); /* reference */ 1235 amdgpu_ring_write(ring, mask); /* mask */ 1236 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 1237 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); 1238 } 1239 1240 static void sdma_v7_1_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring, 1241 uint32_t reg0, uint32_t reg1, 1242 uint32_t ref, uint32_t mask) 1243 { 1244 amdgpu_ring_emit_wreg(ring, reg0, ref); 1245 /* wait for a cycle to reset vm_inv_eng*_ack */ 1246 amdgpu_ring_emit_reg_wait(ring, reg0, 0, 0); 1247 amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask); 1248 } 1249 1250 static int sdma_v7_1_early_init(struct amdgpu_ip_block *ip_block) 1251 { 1252 struct amdgpu_device *adev = ip_block->adev; 1253 int r; 1254 1255 r = amdgpu_sdma_init_microcode(adev, 0, true); 1256 if (r) { 1257 DRM_ERROR("Failed to init sdma firmware!\n"); 1258 return r; 1259 } 1260 1261 sdma_v7_1_set_ring_funcs(adev); 1262 sdma_v7_1_set_buffer_funcs(adev); 1263 sdma_v7_1_set_vm_pte_funcs(adev); 1264 sdma_v7_1_set_irq_funcs(adev); 1265 sdma_v7_1_set_mqd_funcs(adev); 1266 1267 return 0; 1268 } 1269 1270 static int sdma_v7_1_sw_init(struct amdgpu_ip_block *ip_block) 1271 { 1272 struct amdgpu_ring *ring; 1273 int r, i; 1274 struct amdgpu_device *adev = ip_block->adev; 1275 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_1); 1276 uint32_t *ptr; 1277 u32 xcc_id; 1278 1279 /* SDMA trap event */ 1280 r = amdgpu_irq_add_id(adev, SOC_V1_0_IH_CLIENTID_GFX, 1281 GFX_12_1_0__SRCID__SDMA_TRAP, 1282 &adev->sdma.trap_irq); 1283 if (r) 1284 return r; 1285 1286 for (i = 0; i < adev->sdma.num_instances; i++) { 1287 ring = &adev->sdma.instance[i].ring; 1288 ring->ring_obj = NULL; 1289 ring->use_doorbell = true; 1290 ring->me = i; 1291 1292 for (xcc_id = 0; xcc_id < fls(adev->gfx.xcc_mask); xcc_id++) { 1293 if (adev->sdma.instance[i].xcc_id == GET_INST(GC, xcc_id)) 1294 break; 1295 } 1296 1297 DRM_DEBUG("SDMA%d.%d use_doorbell being set to: [%s]\n", 1298 xcc_id, GET_INST(SDMA0, i) % adev->sdma.num_inst_per_xcc, 1299 ring->use_doorbell?"true":"false"); 1300 1301 ring->doorbell_index = 1302 (adev->doorbell_index.sdma_engine[i] << 1); // get DWORD offset 1303 1304 ring->vm_hub = AMDGPU_GFXHUB(xcc_id); 1305 sprintf(ring->name, "sdma%d.%d", xcc_id, 1306 GET_INST(SDMA0, i) % adev->sdma.num_inst_per_xcc); 1307 r = amdgpu_ring_init(adev, ring, 1024, 1308 &adev->sdma.trap_irq, 1309 AMDGPU_SDMA_IRQ_INSTANCE0 + i, 1310 AMDGPU_RING_PRIO_DEFAULT, NULL); 1311 if (r) 1312 return r; 1313 } 1314 1315 adev->sdma.supported_reset = 1316 amdgpu_get_soft_full_reset_mask(&adev->sdma.instance[0].ring); 1317 if (!amdgpu_sriov_vf(adev) && 1318 !adev->debug_disable_gpu_ring_reset) 1319 adev->sdma.supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE; 1320 1321 r = amdgpu_sdma_sysfs_reset_mask_init(adev); 1322 if (r) 1323 return r; 1324 /* Allocate memory for SDMA IP Dump buffer */ 1325 ptr = kcalloc(adev->sdma.num_instances * reg_count, sizeof(uint32_t), GFP_KERNEL); 1326 if (ptr) 1327 adev->sdma.ip_dump = ptr; 1328 else 1329 DRM_ERROR("Failed to allocated memory for SDMA IP Dump\n"); 1330 1331 #ifdef CONFIG_DRM_AMDGPU_NAVI3X_USERQ 1332 adev->userq_funcs[AMDGPU_HW_IP_DMA] = &userq_mes_funcs; 1333 #endif 1334 1335 return r; 1336 } 1337 1338 static int sdma_v7_1_sw_fini(struct amdgpu_ip_block *ip_block) 1339 { 1340 struct amdgpu_device *adev = ip_block->adev; 1341 int i; 1342 1343 for (i = 0; i < adev->sdma.num_instances; i++) 1344 amdgpu_ring_fini(&adev->sdma.instance[i].ring); 1345 1346 amdgpu_sdma_sysfs_reset_mask_fini(adev); 1347 amdgpu_sdma_destroy_inst_ctx(adev, true); 1348 1349 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) 1350 sdma_v7_1_inst_free_ucode_buffer(adev, adev->sdma.sdma_mask); 1351 1352 kfree(adev->sdma.ip_dump); 1353 1354 return 0; 1355 } 1356 1357 static int sdma_v7_1_hw_init(struct amdgpu_ip_block *ip_block) 1358 { 1359 struct amdgpu_device *adev = ip_block->adev; 1360 1361 return sdma_v7_1_inst_start(adev, adev->sdma.sdma_mask); 1362 } 1363 1364 static int sdma_v7_1_hw_fini(struct amdgpu_ip_block *ip_block) 1365 { 1366 struct amdgpu_device *adev = ip_block->adev; 1367 1368 if (amdgpu_sriov_vf(adev)) 1369 return 0; 1370 1371 sdma_v7_1_inst_ctx_switch_enable(adev, false, adev->sdma.sdma_mask); 1372 sdma_v7_1_inst_enable(adev, false, adev->sdma.sdma_mask); 1373 1374 return 0; 1375 } 1376 1377 static int sdma_v7_1_suspend(struct amdgpu_ip_block *ip_block) 1378 { 1379 return sdma_v7_1_hw_fini(ip_block); 1380 } 1381 1382 static int sdma_v7_1_resume(struct amdgpu_ip_block *ip_block) 1383 { 1384 return sdma_v7_1_hw_init(ip_block); 1385 } 1386 1387 static bool sdma_v7_1_is_idle(struct amdgpu_ip_block *ip_block) 1388 { 1389 struct amdgpu_device *adev = ip_block->adev; 1390 u32 i; 1391 1392 for (i = 0; i < adev->sdma.num_instances; i++) { 1393 u32 tmp = RREG32(sdma_v7_1_get_reg_offset(adev, i, regSDMA0_SDMA_STATUS_REG)); 1394 1395 if (!(tmp & SDMA0_SDMA_STATUS_REG__IDLE_MASK)) 1396 return false; 1397 } 1398 1399 return true; 1400 } 1401 1402 static int sdma_v7_1_wait_for_idle(struct amdgpu_ip_block *ip_block) 1403 { 1404 unsigned i, j; 1405 u32 sdma[AMDGPU_MAX_SDMA_INSTANCES]; 1406 struct amdgpu_device *adev = ip_block->adev; 1407 1408 for (i = 0; i < adev->usec_timeout; i++) { 1409 for (j = 0; j < adev->sdma.num_instances; j++) { 1410 sdma[j] = RREG32(sdma_v7_1_get_reg_offset(adev, 1411 j, regSDMA0_SDMA_STATUS_REG)); 1412 if (!(sdma[j] & SDMA0_SDMA_STATUS_REG__IDLE_MASK)) 1413 break; 1414 } 1415 if (j == adev->sdma.num_instances) 1416 return 0; 1417 udelay(1); 1418 } 1419 return -ETIMEDOUT; 1420 } 1421 1422 static int sdma_v7_1_ring_preempt_ib(struct amdgpu_ring *ring) 1423 { 1424 int i, r = 0; 1425 struct amdgpu_device *adev = ring->adev; 1426 u32 index = 0; 1427 u64 sdma_gfx_preempt; 1428 1429 amdgpu_sdma_get_index_from_ring(ring, &index); 1430 sdma_gfx_preempt = 1431 sdma_v7_1_get_reg_offset(adev, index, regSDMA0_SDMA_QUEUE0_PREEMPT); 1432 1433 /* assert preemption condition */ 1434 amdgpu_ring_set_preempt_cond_exec(ring, false); 1435 1436 /* emit the trailing fence */ 1437 ring->trail_seq += 1; 1438 r = amdgpu_ring_alloc(ring, 10); 1439 if (r) { 1440 DRM_ERROR("ring %d failed to be allocated \n", ring->idx); 1441 return r; 1442 } 1443 sdma_v7_1_ring_emit_fence(ring, ring->trail_fence_gpu_addr, 1444 ring->trail_seq, 0); 1445 amdgpu_ring_commit(ring); 1446 1447 /* assert IB preemption */ 1448 WREG32(sdma_gfx_preempt, 1); 1449 1450 /* poll the trailing fence */ 1451 for (i = 0; i < adev->usec_timeout; i++) { 1452 if (ring->trail_seq == 1453 le32_to_cpu(*(ring->trail_fence_cpu_addr))) 1454 break; 1455 udelay(1); 1456 } 1457 1458 if (i >= adev->usec_timeout) { 1459 r = -EINVAL; 1460 DRM_ERROR("ring %d failed to be preempted\n", ring->idx); 1461 } 1462 1463 /* deassert IB preemption */ 1464 WREG32(sdma_gfx_preempt, 0); 1465 1466 /* deassert the preemption condition */ 1467 amdgpu_ring_set_preempt_cond_exec(ring, true); 1468 return r; 1469 } 1470 1471 static int sdma_v7_1_set_trap_irq_state(struct amdgpu_device *adev, 1472 struct amdgpu_irq_src *source, 1473 unsigned type, 1474 enum amdgpu_interrupt_state state) 1475 { 1476 u32 sdma_cntl; 1477 1478 u32 reg_offset = sdma_v7_1_get_reg_offset(adev, type, regSDMA0_SDMA_CNTL); 1479 1480 sdma_cntl = RREG32(reg_offset); 1481 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_SDMA_CNTL, TRAP_ENABLE, 1482 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 1483 WREG32(reg_offset, sdma_cntl); 1484 1485 return 0; 1486 } 1487 1488 static int sdma_v7_1_process_trap_irq(struct amdgpu_device *adev, 1489 struct amdgpu_irq_src *source, 1490 struct amdgpu_iv_entry *entry) 1491 { 1492 int inst, instances, queue, xcc_id = 0; 1493 uint32_t mes_queue_id = entry->src_data[0]; 1494 1495 DRM_DEBUG("IH: SDMA trap\n"); 1496 1497 if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) { 1498 struct amdgpu_mes_queue *queue; 1499 1500 mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK; 1501 1502 spin_lock(&adev->mes.queue_id_lock); 1503 queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id); 1504 if (queue) { 1505 DRM_DEBUG("process smda queue id = %d\n", mes_queue_id); 1506 amdgpu_fence_process(queue->ring); 1507 } 1508 spin_unlock(&adev->mes.queue_id_lock); 1509 return 0; 1510 } 1511 1512 queue = entry->ring_id & 0xf; 1513 if (adev->gfx.funcs && adev->gfx.funcs->ih_node_to_logical_xcc) 1514 xcc_id = adev->gfx.funcs->ih_node_to_logical_xcc(adev, entry->node_id); 1515 else 1516 dev_warn(adev->dev, "IH: SDMA may get wrong xcc id as gfx function not available\n"); 1517 inst = ((entry->ring_id & 0xf0) >> 4) + 1518 GET_INST(GC, xcc_id) * adev->sdma.num_inst_per_xcc; 1519 for (instances = 0; instances < adev->sdma.num_instances; instances++) { 1520 if (inst == GET_INST(SDMA0, instances)) 1521 break; 1522 } 1523 if (instances > adev->sdma.num_instances - 1) { 1524 DRM_ERROR("IH: wrong ring_ID detected, as wrong sdma instance\n"); 1525 return -EINVAL; 1526 } 1527 1528 switch (entry->client_id) { 1529 case SOC_V1_0_IH_CLIENTID_GFX: 1530 switch (queue) { 1531 case 0: 1532 amdgpu_fence_process(&adev->sdma.instance[instances].ring); 1533 break; 1534 default: 1535 break; 1536 } 1537 break; 1538 } 1539 return 0; 1540 } 1541 1542 static int sdma_v7_1_process_illegal_inst_irq(struct amdgpu_device *adev, 1543 struct amdgpu_irq_src *source, 1544 struct amdgpu_iv_entry *entry) 1545 { 1546 return 0; 1547 } 1548 1549 static int sdma_v7_1_set_clockgating_state(struct amdgpu_ip_block *ip_block, 1550 enum amd_clockgating_state state) 1551 { 1552 return 0; 1553 } 1554 1555 static int sdma_v7_1_set_powergating_state(struct amdgpu_ip_block *ip_block, 1556 enum amd_powergating_state state) 1557 { 1558 return 0; 1559 } 1560 1561 static void sdma_v7_1_get_clockgating_state(struct amdgpu_ip_block *ip_block, 1562 u64 *flags) 1563 { 1564 } 1565 1566 static void sdma_v7_1_print_ip_state(struct amdgpu_ip_block *ip_block, struct drm_printer *p) 1567 { 1568 struct amdgpu_device *adev = ip_block->adev; 1569 int i, j; 1570 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_1); 1571 uint32_t instance_offset; 1572 1573 if (!adev->sdma.ip_dump) 1574 return; 1575 1576 drm_printf(p, "num_instances:%d\n", adev->sdma.num_instances); 1577 for (i = 0; i < adev->sdma.num_instances; i++) { 1578 instance_offset = i * reg_count; 1579 drm_printf(p, "\nInstance:%d\n", i); 1580 1581 for (j = 0; j < reg_count; j++) 1582 drm_printf(p, "%-50s \t 0x%08x\n", sdma_reg_list_7_1[j].reg_name, 1583 adev->sdma.ip_dump[instance_offset + j]); 1584 } 1585 } 1586 1587 static void sdma_v7_1_dump_ip_state(struct amdgpu_ip_block *ip_block) 1588 { 1589 struct amdgpu_device *adev = ip_block->adev; 1590 int i, j; 1591 uint32_t instance_offset; 1592 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_7_1); 1593 1594 if (!adev->sdma.ip_dump) 1595 return; 1596 1597 amdgpu_gfx_off_ctrl(adev, false); 1598 for (i = 0; i < adev->sdma.num_instances; i++) { 1599 instance_offset = i * reg_count; 1600 for (j = 0; j < reg_count; j++) 1601 adev->sdma.ip_dump[instance_offset + j] = 1602 RREG32(sdma_v7_1_get_reg_offset(adev, i, 1603 sdma_reg_list_7_1[j].reg_offset)); 1604 } 1605 amdgpu_gfx_off_ctrl(adev, true); 1606 } 1607 1608 const struct amd_ip_funcs sdma_v7_1_ip_funcs = { 1609 .name = "sdma_v7_1", 1610 .early_init = sdma_v7_1_early_init, 1611 .late_init = NULL, 1612 .sw_init = sdma_v7_1_sw_init, 1613 .sw_fini = sdma_v7_1_sw_fini, 1614 .hw_init = sdma_v7_1_hw_init, 1615 .hw_fini = sdma_v7_1_hw_fini, 1616 .suspend = sdma_v7_1_suspend, 1617 .resume = sdma_v7_1_resume, 1618 .is_idle = sdma_v7_1_is_idle, 1619 .wait_for_idle = sdma_v7_1_wait_for_idle, 1620 .soft_reset = sdma_v7_1_soft_reset, 1621 .check_soft_reset = sdma_v7_1_check_soft_reset, 1622 .set_clockgating_state = sdma_v7_1_set_clockgating_state, 1623 .set_powergating_state = sdma_v7_1_set_powergating_state, 1624 .get_clockgating_state = sdma_v7_1_get_clockgating_state, 1625 .dump_ip_state = sdma_v7_1_dump_ip_state, 1626 .print_ip_state = sdma_v7_1_print_ip_state, 1627 }; 1628 1629 static const struct amdgpu_ring_funcs sdma_v7_1_ring_funcs = { 1630 .type = AMDGPU_RING_TYPE_SDMA, 1631 .align_mask = 0xf, 1632 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 1633 .support_64bit_ptrs = true, 1634 .secure_submission_supported = true, 1635 .get_rptr = sdma_v7_1_ring_get_rptr, 1636 .get_wptr = sdma_v7_1_ring_get_wptr, 1637 .set_wptr = sdma_v7_1_ring_set_wptr, 1638 .emit_frame_size = 1639 5 + /* sdma_v7_1_ring_init_cond_exec */ 1640 6 + /* sdma_v7_1_ring_emit_pipeline_sync */ 1641 /* sdma_v7_1_ring_emit_vm_flush */ 1642 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 1643 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 + 1644 10 + 10 + 10, /* sdma_v7_1_ring_emit_fence x3 for user fence, vm fence */ 1645 .emit_ib_size = 5 + 7 + 6, /* sdma_v7_1_ring_emit_ib */ 1646 .emit_ib = sdma_v7_1_ring_emit_ib, 1647 .emit_mem_sync = sdma_v7_1_ring_emit_mem_sync, 1648 .emit_fence = sdma_v7_1_ring_emit_fence, 1649 .emit_pipeline_sync = sdma_v7_1_ring_emit_pipeline_sync, 1650 .emit_vm_flush = sdma_v7_1_ring_emit_vm_flush, 1651 .test_ring = sdma_v7_1_ring_test_ring, 1652 .test_ib = sdma_v7_1_ring_test_ib, 1653 .insert_nop = sdma_v7_1_ring_insert_nop, 1654 .pad_ib = sdma_v7_1_ring_pad_ib, 1655 .emit_wreg = sdma_v7_1_ring_emit_wreg, 1656 .emit_reg_wait = sdma_v7_1_ring_emit_reg_wait, 1657 .emit_reg_write_reg_wait = sdma_v7_1_ring_emit_reg_write_reg_wait, 1658 .init_cond_exec = sdma_v7_1_ring_init_cond_exec, 1659 .preempt_ib = sdma_v7_1_ring_preempt_ib, 1660 .reset = sdma_v7_1_reset_queue, 1661 }; 1662 1663 static void sdma_v7_1_set_ring_funcs(struct amdgpu_device *adev) 1664 { 1665 int i, dev_inst; 1666 1667 for (i = 0; i < adev->sdma.num_instances; i++) { 1668 adev->sdma.instance[i].ring.funcs = &sdma_v7_1_ring_funcs; 1669 adev->sdma.instance[i].ring.me = i; 1670 1671 dev_inst = GET_INST(SDMA0, i); 1672 /* XCC to which SDMA belongs depends on physical instance */ 1673 adev->sdma.instance[i].xcc_id = 1674 dev_inst / adev->sdma.num_inst_per_xcc; 1675 } 1676 } 1677 1678 static const struct amdgpu_irq_src_funcs sdma_v7_1_trap_irq_funcs = { 1679 .set = sdma_v7_1_set_trap_irq_state, 1680 .process = sdma_v7_1_process_trap_irq, 1681 }; 1682 1683 static const struct amdgpu_irq_src_funcs sdma_v7_1_illegal_inst_irq_funcs = { 1684 .process = sdma_v7_1_process_illegal_inst_irq, 1685 }; 1686 1687 static void sdma_v7_1_set_irq_funcs(struct amdgpu_device *adev) 1688 { 1689 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 + 1690 adev->sdma.num_instances; 1691 adev->sdma.trap_irq.funcs = &sdma_v7_1_trap_irq_funcs; 1692 adev->sdma.illegal_inst_irq.funcs = &sdma_v7_1_illegal_inst_irq_funcs; 1693 } 1694 1695 /** 1696 * sdma_v7_1_emit_copy_buffer - copy buffer using the sDMA engine 1697 * 1698 * @ib: indirect buffer to fill with commands 1699 * @src_offset: src GPU address 1700 * @dst_offset: dst GPU address 1701 * @byte_count: number of bytes to xfer 1702 * @copy_flags: copy flags for the buffers 1703 * 1704 * Copy GPU buffers using the DMA engine. 1705 * Used by the amdgpu ttm implementation to move pages if 1706 * registered as the asic copy callback. 1707 */ 1708 static void sdma_v7_1_emit_copy_buffer(struct amdgpu_ib *ib, 1709 uint64_t src_offset, 1710 uint64_t dst_offset, 1711 uint32_t byte_count, 1712 uint32_t copy_flags) 1713 { 1714 ib->ptr[ib->length_dw++] = SDMA_PKT_COPY_LINEAR_HEADER_OP(SDMA_OP_COPY) | 1715 SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) | 1716 SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0); 1717 1718 ib->ptr[ib->length_dw++] = byte_count - 1; 1719 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 1720 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); 1721 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); 1722 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 1723 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 1724 } 1725 1726 /** 1727 * sdma_v7_1_emit_fill_buffer - fill buffer using the sDMA engine 1728 * 1729 * @ib: indirect buffer to fill 1730 * @src_data: value to write to buffer 1731 * @dst_offset: dst GPU address 1732 * @byte_count: number of bytes to xfer 1733 * 1734 * Fill GPU buffers using the DMA engine. 1735 */ 1736 static void sdma_v7_1_emit_fill_buffer(struct amdgpu_ib *ib, 1737 uint32_t src_data, 1738 uint64_t dst_offset, 1739 uint32_t byte_count) 1740 { 1741 ib->ptr[ib->length_dw++] = SDMA_PKT_CONSTANT_FILL_HEADER_OP(SDMA_OP_CONST_FILL); 1742 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 1743 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 1744 ib->ptr[ib->length_dw++] = src_data; 1745 ib->ptr[ib->length_dw++] = byte_count - 1; 1746 } 1747 1748 static const struct amdgpu_buffer_funcs sdma_v7_1_buffer_funcs = { 1749 .copy_max_bytes = 0x400000, 1750 .copy_num_dw = 8, 1751 .emit_copy_buffer = sdma_v7_1_emit_copy_buffer, 1752 .fill_max_bytes = 0x400000, 1753 .fill_num_dw = 5, 1754 .emit_fill_buffer = sdma_v7_1_emit_fill_buffer, 1755 }; 1756 1757 static void sdma_v7_1_set_buffer_funcs(struct amdgpu_device *adev) 1758 { 1759 adev->mman.buffer_funcs = &sdma_v7_1_buffer_funcs; 1760 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring; 1761 } 1762 1763 static const struct amdgpu_vm_pte_funcs sdma_v7_1_vm_pte_funcs = { 1764 .copy_pte_num_dw = 8, 1765 .copy_pte = sdma_v7_1_vm_copy_pte, 1766 .write_pte = sdma_v7_1_vm_write_pte, 1767 .set_pte_pde = sdma_v7_1_vm_set_pte_pde, 1768 }; 1769 1770 static void sdma_v7_1_set_vm_pte_funcs(struct amdgpu_device *adev) 1771 { 1772 unsigned i; 1773 1774 adev->vm_manager.vm_pte_funcs = &sdma_v7_1_vm_pte_funcs; 1775 for (i = 0; i < adev->sdma.num_instances; i++) { 1776 adev->vm_manager.vm_pte_scheds[i] = 1777 &adev->sdma.instance[i].ring.sched; 1778 } 1779 adev->vm_manager.vm_pte_num_scheds = adev->sdma.num_instances; 1780 } 1781 1782 const struct amdgpu_ip_block_version sdma_v7_1_ip_block = { 1783 .type = AMD_IP_BLOCK_TYPE_SDMA, 1784 .major = 7, 1785 .minor = 1, 1786 .rev = 0, 1787 .funcs = &sdma_v7_1_ip_funcs, 1788 }; 1789 1790 static int sdma_v7_1_xcp_resume(void *handle, uint32_t inst_mask) 1791 { 1792 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1793 int r; 1794 1795 r = sdma_v7_1_inst_start(adev, inst_mask); 1796 1797 return r; 1798 } 1799 1800 static int sdma_v7_1_xcp_suspend(void *handle, uint32_t inst_mask) 1801 { 1802 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1803 1804 sdma_v7_1_inst_ctx_switch_enable(adev, false, inst_mask); 1805 sdma_v7_1_inst_enable(adev, false, inst_mask); 1806 1807 return 0; 1808 } 1809 1810 struct amdgpu_xcp_ip_funcs sdma_v7_1_xcp_funcs = { 1811 .suspend = &sdma_v7_1_xcp_suspend, 1812 .resume = &sdma_v7_1_xcp_resume 1813 }; 1814