1 /* 2 * Copyright 2019 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_10_3_0_offset.h" 34 #include "gc/gc_10_3_0_sh_mask.h" 35 #include "ivsrcid/sdma0/irqsrcs_sdma0_5_0.h" 36 #include "ivsrcid/sdma1/irqsrcs_sdma1_5_0.h" 37 #include "ivsrcid/sdma2/irqsrcs_sdma2_5_0.h" 38 #include "ivsrcid/sdma3/irqsrcs_sdma3_5_0.h" 39 40 #include "soc15_common.h" 41 #include "soc15.h" 42 #include "navi10_sdma_pkt_open.h" 43 #include "nbio_v2_3.h" 44 #include "sdma_common.h" 45 #include "sdma_v5_2.h" 46 47 MODULE_FIRMWARE("amdgpu/sienna_cichlid_sdma.bin"); 48 MODULE_FIRMWARE("amdgpu/navy_flounder_sdma.bin"); 49 MODULE_FIRMWARE("amdgpu/dimgrey_cavefish_sdma.bin"); 50 MODULE_FIRMWARE("amdgpu/beige_goby_sdma.bin"); 51 52 MODULE_FIRMWARE("amdgpu/vangogh_sdma.bin"); 53 MODULE_FIRMWARE("amdgpu/yellow_carp_sdma.bin"); 54 MODULE_FIRMWARE("amdgpu/sdma_5_2_6.bin"); 55 MODULE_FIRMWARE("amdgpu/sdma_5_2_7.bin"); 56 57 #define SDMA1_REG_OFFSET 0x600 58 #define SDMA3_REG_OFFSET 0x400 59 #define SDMA0_HYP_DEC_REG_START 0x5880 60 #define SDMA0_HYP_DEC_REG_END 0x5893 61 #define SDMA1_HYP_DEC_REG_OFFSET 0x20 62 63 static const struct amdgpu_hwip_reg_entry sdma_reg_list_5_2[] = { 64 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS_REG), 65 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS1_REG), 66 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS2_REG), 67 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_STATUS3_REG), 68 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UCODE_CHECKSUM), 69 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RB_RPTR_FETCH_HI), 70 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RB_RPTR_FETCH), 71 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_STATUS), 72 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_STATUS), 73 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_XNACK0), 74 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_RD_XNACK1), 75 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_XNACK0), 76 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_UTCL1_WR_XNACK1), 77 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_CNTL), 78 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_RPTR), 79 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_RPTR_HI), 80 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_WPTR), 81 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_RB_WPTR_HI), 82 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_OFFSET), 83 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_BASE_LO), 84 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_BASE_HI), 85 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_CNTL), 86 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_RPTR), 87 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_IB_SUB_REMAIN), 88 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_GFX_DUMMY_REG), 89 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_CNTL), 90 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_RPTR), 91 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_RPTR_HI), 92 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_WPTR), 93 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_RB_WPTR_HI), 94 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_OFFSET), 95 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_BASE_LO), 96 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_IB_BASE_HI), 97 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_PAGE_DUMMY_REG), 98 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_CNTL), 99 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_RPTR), 100 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_RPTR_HI), 101 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_WPTR), 102 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_RB_WPTR_HI), 103 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_OFFSET), 104 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_BASE_LO), 105 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_IB_BASE_HI), 106 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_RLC0_DUMMY_REG), 107 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_INT_STATUS), 108 SOC15_REG_ENTRY_STR(GC, 0, mmSDMA0_VM_CNTL), 109 SOC15_REG_ENTRY_STR(GC, 0, mmGRBM_STATUS2) 110 }; 111 112 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev); 113 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev); 114 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev); 115 static int sdma_v5_2_stop_queue(struct amdgpu_ring *ring); 116 static int sdma_v5_2_restore_queue(struct amdgpu_ring *ring); 117 118 static u32 sdma_v5_2_get_reg_offset(struct amdgpu_device *adev, u32 instance, u32 internal_offset) 119 { 120 u32 base; 121 122 if (internal_offset >= SDMA0_HYP_DEC_REG_START && 123 internal_offset <= SDMA0_HYP_DEC_REG_END) { 124 base = adev->reg_offset[GC_HWIP][0][1]; 125 if (instance != 0) 126 internal_offset += SDMA1_HYP_DEC_REG_OFFSET * instance; 127 } else { 128 if (instance < 2) { 129 base = adev->reg_offset[GC_HWIP][0][0]; 130 if (instance == 1) 131 internal_offset += SDMA1_REG_OFFSET; 132 } else { 133 base = adev->reg_offset[GC_HWIP][0][2]; 134 if (instance == 3) 135 internal_offset += SDMA3_REG_OFFSET; 136 } 137 } 138 139 return base + internal_offset; 140 } 141 142 static unsigned sdma_v5_2_ring_init_cond_exec(struct amdgpu_ring *ring, 143 uint64_t addr) 144 { 145 unsigned ret; 146 147 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_COND_EXE)); 148 amdgpu_ring_write(ring, lower_32_bits(addr)); 149 amdgpu_ring_write(ring, upper_32_bits(addr)); 150 amdgpu_ring_write(ring, 1); 151 /* this is the offset we need patch later */ 152 ret = ring->wptr & ring->buf_mask; 153 /* insert dummy here and patch it later */ 154 amdgpu_ring_write(ring, 0); 155 156 return ret; 157 } 158 159 /** 160 * sdma_v5_2_ring_get_rptr - get the current read pointer 161 * 162 * @ring: amdgpu ring pointer 163 * 164 * Get the current rptr from the hardware (NAVI10+). 165 */ 166 static uint64_t sdma_v5_2_ring_get_rptr(struct amdgpu_ring *ring) 167 { 168 u64 *rptr; 169 170 /* XXX check if swapping is necessary on BE */ 171 rptr = (u64 *)ring->rptr_cpu_addr; 172 173 DRM_DEBUG("rptr before shift == 0x%016llx\n", *rptr); 174 return ((*rptr) >> 2); 175 } 176 177 /** 178 * sdma_v5_2_ring_get_wptr - get the current write pointer 179 * 180 * @ring: amdgpu ring pointer 181 * 182 * Get the current wptr from the hardware (NAVI10+). 183 */ 184 static uint64_t sdma_v5_2_ring_get_wptr(struct amdgpu_ring *ring) 185 { 186 struct amdgpu_device *adev = ring->adev; 187 u64 wptr; 188 189 if (ring->use_doorbell) { 190 /* XXX check if swapping is necessary on BE */ 191 wptr = READ_ONCE(*((u64 *)ring->wptr_cpu_addr)); 192 DRM_DEBUG("wptr/doorbell before shift == 0x%016llx\n", wptr); 193 } else { 194 wptr = RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI)); 195 wptr = wptr << 32; 196 wptr |= RREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR)); 197 DRM_DEBUG("wptr before shift [%i] wptr == 0x%016llx\n", ring->me, wptr); 198 } 199 200 return wptr >> 2; 201 } 202 203 /** 204 * sdma_v5_2_ring_set_wptr - commit the write pointer 205 * 206 * @ring: amdgpu ring pointer 207 * 208 * Write the wptr back to the hardware (NAVI10+). 209 */ 210 static void sdma_v5_2_ring_set_wptr(struct amdgpu_ring *ring) 211 { 212 struct amdgpu_device *adev = ring->adev; 213 214 DRM_DEBUG("Setting write pointer\n"); 215 if (ring->use_doorbell) { 216 DRM_DEBUG("Using doorbell -- " 217 "wptr_offs == 0x%08x " 218 "lower_32_bits(ring->wptr << 2) == 0x%08x " 219 "upper_32_bits(ring->wptr << 2) == 0x%08x\n", 220 ring->wptr_offs, 221 lower_32_bits(ring->wptr << 2), 222 upper_32_bits(ring->wptr << 2)); 223 /* XXX check if swapping is necessary on BE */ 224 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 225 ring->wptr << 2); 226 DRM_DEBUG("calling WDOORBELL64(0x%08x, 0x%016llx)\n", 227 ring->doorbell_index, ring->wptr << 2); 228 WDOORBELL64(ring->doorbell_index, ring->wptr << 2); 229 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) == IP_VERSION(5, 2, 1)) { 230 /* SDMA seems to miss doorbells sometimes when powergating kicks in. 231 * Updating the wptr directly will wake it. This is only safe because 232 * we disallow gfxoff in begin_use() and then allow it again in end_use(). 233 */ 234 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR), 235 lower_32_bits(ring->wptr << 2)); 236 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI), 237 upper_32_bits(ring->wptr << 2)); 238 } 239 } else { 240 DRM_DEBUG("Not using doorbell -- " 241 "mmSDMA%i_GFX_RB_WPTR == 0x%08x " 242 "mmSDMA%i_GFX_RB_WPTR_HI == 0x%08x\n", 243 ring->me, 244 lower_32_bits(ring->wptr << 2), 245 ring->me, 246 upper_32_bits(ring->wptr << 2)); 247 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR), 248 lower_32_bits(ring->wptr << 2)); 249 WREG32(sdma_v5_2_get_reg_offset(adev, ring->me, mmSDMA0_GFX_RB_WPTR_HI), 250 upper_32_bits(ring->wptr << 2)); 251 } 252 } 253 254 static void sdma_v5_2_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count) 255 { 256 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 257 int i; 258 259 for (i = 0; i < count; i++) 260 if (sdma && sdma->burst_nop && (i == 0)) 261 amdgpu_ring_write(ring, ring->funcs->nop | 262 SDMA_PKT_NOP_HEADER_COUNT(count - 1)); 263 else 264 amdgpu_ring_write(ring, ring->funcs->nop); 265 } 266 267 /** 268 * sdma_v5_2_ring_emit_ib - Schedule an IB on the DMA engine 269 * 270 * @ring: amdgpu ring pointer 271 * @job: job to retrieve vmid from 272 * @ib: IB object to schedule 273 * @flags: unused 274 * 275 * Schedule an IB in the DMA ring. 276 */ 277 static void sdma_v5_2_ring_emit_ib(struct amdgpu_ring *ring, 278 struct amdgpu_job *job, 279 struct amdgpu_ib *ib, 280 uint32_t flags) 281 { 282 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 283 uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid); 284 285 /* An IB packet must end on a 8 DW boundary--the next dword 286 * must be on a 8-dword boundary. Our IB packet below is 6 287 * dwords long, thus add x number of NOPs, such that, in 288 * modular arithmetic, 289 * wptr + 6 + x = 8k, k >= 0, which in C is, 290 * (wptr + 6 + x) % 8 = 0. 291 * The expression below, is a solution of x. 292 */ 293 sdma_v5_2_ring_insert_nop(ring, (2 - lower_32_bits(ring->wptr)) & 7); 294 295 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_INDIRECT) | 296 SDMA_PKT_INDIRECT_HEADER_VMID(vmid & 0xf)); 297 /* base must be 32 byte aligned */ 298 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr) & 0xffffffe0); 299 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 300 amdgpu_ring_write(ring, ib->length_dw); 301 amdgpu_ring_write(ring, lower_32_bits(csa_mc_addr)); 302 amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr)); 303 } 304 305 /** 306 * sdma_v5_2_ring_emit_mem_sync - flush the IB by graphics cache rinse 307 * 308 * @ring: amdgpu ring pointer 309 * 310 * flush the IB by graphics cache rinse. 311 */ 312 static void sdma_v5_2_ring_emit_mem_sync(struct amdgpu_ring *ring) 313 { 314 uint32_t gcr_cntl = SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | 315 SDMA_GCR_GLM_INV | SDMA_GCR_GL1_INV | 316 SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV | 317 SDMA_GCR_GLI_INV(1); 318 319 /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */ 320 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ)); 321 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0)); 322 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) | 323 SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0)); 324 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) | 325 SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16)); 326 amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) | 327 SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0)); 328 } 329 330 /** 331 * sdma_v5_2_ring_emit_hdp_flush - emit an hdp flush on the DMA ring 332 * 333 * @ring: amdgpu ring pointer 334 * 335 * Emit an hdp flush packet on the requested DMA ring. 336 */ 337 static void sdma_v5_2_ring_emit_hdp_flush(struct amdgpu_ring *ring) 338 { 339 struct amdgpu_device *adev = ring->adev; 340 u32 ref_and_mask = 0; 341 const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg; 342 343 if (ring->me > 1) { 344 amdgpu_hdp_flush(adev, ring); 345 } else { 346 ref_and_mask = nbio_hf_reg->ref_and_mask_sdma0 << ring->me; 347 348 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 349 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(1) | 350 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3)); /* == */ 351 amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_done_offset(adev)) << 2); 352 amdgpu_ring_write(ring, (adev->nbio.funcs->get_hdp_flush_req_offset(adev)) << 2); 353 amdgpu_ring_write(ring, ref_and_mask); /* reference */ 354 amdgpu_ring_write(ring, ref_and_mask); /* mask */ 355 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 356 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(10)); /* retry count, poll interval */ 357 } 358 } 359 360 /** 361 * sdma_v5_2_ring_emit_fence - emit a fence on the DMA ring 362 * 363 * @ring: amdgpu ring pointer 364 * @addr: address 365 * @seq: sequence number 366 * @flags: fence related flags 367 * 368 * Add a DMA fence packet to the ring to write 369 * the fence seq number and DMA trap packet to generate 370 * an interrupt if needed. 371 */ 372 static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 seq, 373 unsigned flags) 374 { 375 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 376 /* write the fence */ 377 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) | 378 SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); /* Ucached(UC) */ 379 /* zero in first two bits */ 380 BUG_ON(addr & 0x3); 381 amdgpu_ring_write(ring, lower_32_bits(addr)); 382 amdgpu_ring_write(ring, upper_32_bits(addr)); 383 amdgpu_ring_write(ring, lower_32_bits(seq)); 384 385 /* optionally write high bits as well */ 386 if (write64bit) { 387 addr += 4; 388 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_FENCE) | 389 SDMA_PKT_FENCE_HEADER_MTYPE(0x3)); 390 /* zero in first two bits */ 391 BUG_ON(addr & 0x3); 392 amdgpu_ring_write(ring, lower_32_bits(addr)); 393 amdgpu_ring_write(ring, upper_32_bits(addr)); 394 amdgpu_ring_write(ring, upper_32_bits(seq)); 395 } 396 397 if ((flags & AMDGPU_FENCE_FLAG_INT)) { 398 /* generate an interrupt */ 399 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_TRAP)); 400 amdgpu_ring_write(ring, SDMA_PKT_TRAP_INT_CONTEXT_INT_CONTEXT(0)); 401 } 402 } 403 404 405 /** 406 * sdma_v5_2_gfx_stop - stop the gfx async dma engines 407 * 408 * @adev: amdgpu_device pointer 409 * @inst_mask: mask of dma engine instances to be disabled 410 * Stop the gfx async dma ring buffers. 411 */ 412 static void sdma_v5_2_gfx_stop(struct amdgpu_device *adev, uint32_t inst_mask) 413 { 414 u32 rb_cntl, ib_cntl; 415 int i; 416 417 for_each_inst(i, inst_mask) { 418 rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL)); 419 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 0); 420 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl); 421 ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL)); 422 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 0); 423 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl); 424 } 425 } 426 427 /** 428 * sdma_v5_2_rlc_stop - stop the compute async dma engines 429 * 430 * @adev: amdgpu_device pointer 431 * 432 * Stop the compute async dma queues. 433 */ 434 static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev) 435 { 436 /* XXX todo */ 437 } 438 439 /** 440 * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch 441 * 442 * @adev: amdgpu_device pointer 443 * @enable: enable/disable the DMA MEs context switch. 444 * 445 * Halt or unhalt the async dma engines context switch. 446 */ 447 static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable) 448 { 449 u32 f32_cntl, phase_quantum = 0; 450 int i; 451 452 if (amdgpu_sdma_phase_quantum) { 453 unsigned value = amdgpu_sdma_phase_quantum; 454 unsigned unit = 0; 455 456 while (value > (SDMA0_PHASE0_QUANTUM__VALUE_MASK >> 457 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT)) { 458 value = (value + 1) >> 1; 459 unit++; 460 } 461 if (unit > (SDMA0_PHASE0_QUANTUM__UNIT_MASK >> 462 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT)) { 463 value = (SDMA0_PHASE0_QUANTUM__VALUE_MASK >> 464 SDMA0_PHASE0_QUANTUM__VALUE__SHIFT); 465 unit = (SDMA0_PHASE0_QUANTUM__UNIT_MASK >> 466 SDMA0_PHASE0_QUANTUM__UNIT__SHIFT); 467 WARN_ONCE(1, 468 "clamping sdma_phase_quantum to %uK clock cycles\n", 469 value << unit); 470 } 471 phase_quantum = 472 value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT | 473 unit << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT; 474 } 475 476 for (i = 0; i < adev->sdma.num_instances; i++) { 477 if (enable && amdgpu_sdma_phase_quantum) { 478 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM), 479 phase_quantum); 480 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM), 481 phase_quantum); 482 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM), 483 phase_quantum); 484 } 485 486 if (!amdgpu_sriov_vf(adev)) { 487 f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL)); 488 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL, 489 AUTO_CTXSW_ENABLE, enable ? 1 : 0); 490 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl); 491 } 492 } 493 494 } 495 496 /** 497 * sdma_v5_2_enable - stop the async dma engines 498 * 499 * @adev: amdgpu_device pointer 500 * @enable: enable/disable the DMA MEs. 501 * 502 * Halt or unhalt the async dma engines. 503 */ 504 static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable) 505 { 506 u32 f32_cntl; 507 int i; 508 uint32_t inst_mask; 509 510 inst_mask = GENMASK(adev->sdma.num_instances - 1, 0); 511 if (!enable) { 512 sdma_v5_2_gfx_stop(adev, inst_mask); 513 sdma_v5_2_rlc_stop(adev); 514 } 515 516 if (!amdgpu_sriov_vf(adev)) { 517 for (i = 0; i < adev->sdma.num_instances; i++) { 518 f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL)); 519 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1); 520 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl); 521 } 522 } 523 } 524 525 /** 526 * sdma_v5_2_gfx_resume_instance - start/restart a certain sdma engine 527 * 528 * @adev: amdgpu_device pointer 529 * @i: instance 530 * @restore: used to restore wptr when restart 531 * 532 * Set up the gfx DMA ring buffers and enable them. On restart, we will restore wptr and rptr. 533 * Return 0 for success. 534 */ 535 536 static int sdma_v5_2_gfx_resume_instance(struct amdgpu_device *adev, int i, bool restore) 537 { 538 struct amdgpu_ring *ring; 539 u32 rb_cntl, ib_cntl; 540 u32 rb_bufsz; 541 u32 doorbell; 542 u32 doorbell_offset; 543 u32 temp; 544 u32 wptr_poll_cntl; 545 u64 wptr_gpu_addr; 546 547 ring = &adev->sdma.instance[i].ring; 548 549 if (!amdgpu_sriov_vf(adev)) 550 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_SEM_WAIT_FAIL_TIMER_CNTL), 0); 551 552 /* Set ring buffer size in dwords */ 553 rb_bufsz = order_base_2(ring->ring_size / 4); 554 rb_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL)); 555 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SIZE, rb_bufsz); 556 #ifdef __BIG_ENDIAN 557 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_SWAP_ENABLE, 1); 558 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, 559 RPTR_WRITEBACK_SWAP_ENABLE, 1); 560 #endif 561 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl); 562 563 /* Initialize the ring buffer's read and write pointers */ 564 if (restore) { 565 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), lower_32_bits(ring->wptr << 2)); 566 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), upper_32_bits(ring->wptr << 2)); 567 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2)); 568 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2)); 569 } else { 570 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR), 0); 571 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_HI), 0); 572 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), 0); 573 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), 0); 574 } 575 576 /* setup the wptr shadow polling */ 577 wptr_gpu_addr = ring->wptr_gpu_addr; 578 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_LO), 579 lower_32_bits(wptr_gpu_addr)); 580 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_ADDR_HI), 581 upper_32_bits(wptr_gpu_addr)); 582 wptr_poll_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, 583 mmSDMA0_GFX_RB_WPTR_POLL_CNTL)); 584 wptr_poll_cntl = REG_SET_FIELD(wptr_poll_cntl, 585 SDMA0_GFX_RB_WPTR_POLL_CNTL, 586 F32_POLL_ENABLE, 1); 587 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_POLL_CNTL), 588 wptr_poll_cntl); 589 590 /* set the wb address whether it's enabled or not */ 591 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_HI), 592 upper_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFF); 593 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_RPTR_ADDR_LO), 594 lower_32_bits(ring->rptr_gpu_addr) & 0xFFFFFFFC); 595 596 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RPTR_WRITEBACK_ENABLE, 1); 597 598 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE), ring->gpu_addr >> 8); 599 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_BASE_HI), ring->gpu_addr >> 40); 600 601 if (!restore) 602 ring->wptr = 0; 603 604 /* before programing wptr to a less value, need set minor_ptr_update first */ 605 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 1); 606 607 if (!amdgpu_sriov_vf(adev)) { /* only bare-metal use register write for wptr */ 608 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR), lower_32_bits(ring->wptr << 2)); 609 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_WPTR_HI), upper_32_bits(ring->wptr << 2)); 610 } 611 612 doorbell = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL)); 613 doorbell_offset = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET)); 614 615 if (ring->use_doorbell) { 616 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 1); 617 doorbell_offset = REG_SET_FIELD(doorbell_offset, SDMA0_GFX_DOORBELL_OFFSET, 618 OFFSET, ring->doorbell_index); 619 } else { 620 doorbell = REG_SET_FIELD(doorbell, SDMA0_GFX_DOORBELL, ENABLE, 0); 621 } 622 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL), doorbell); 623 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_DOORBELL_OFFSET), doorbell_offset); 624 625 adev->nbio.funcs->sdma_doorbell_range(adev, i, ring->use_doorbell, 626 ring->doorbell_index, 627 adev->doorbell_index.sdma_doorbell_range); 628 629 if (amdgpu_sriov_vf(adev)) 630 sdma_v5_2_ring_set_wptr(ring); 631 632 /* set minor_ptr_update to 0 after wptr programed */ 633 634 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_MINOR_PTR_UPDATE), 0); 635 636 /* SRIOV VF has no control of any of registers below */ 637 if (!amdgpu_sriov_vf(adev)) { 638 /* set utc l1 enable flag always to 1 */ 639 temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL)); 640 temp = REG_SET_FIELD(temp, SDMA0_CNTL, UTC_L1_ENABLE, 1); 641 642 /* enable MCBP */ 643 temp = REG_SET_FIELD(temp, SDMA0_CNTL, MIDCMD_PREEMPT_ENABLE, 1); 644 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), temp); 645 646 /* Set up RESP_MODE to non-copy addresses */ 647 temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL)); 648 temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, RESP_MODE, 3); 649 temp = REG_SET_FIELD(temp, SDMA0_UTCL1_CNTL, REDO_DELAY, 9); 650 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_CNTL), temp); 651 652 /* program default cache read and write policy */ 653 temp = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE)); 654 /* clean read policy and write policy bits */ 655 temp &= 0xFF0FFF; 656 temp |= ((CACHE_READ_POLICY_L2__DEFAULT << 12) | 657 (CACHE_WRITE_POLICY_L2__DEFAULT << 14) | 658 SDMA0_UTCL1_PAGE__LLC_NOALLOC_MASK); 659 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UTCL1_PAGE), temp); 660 661 /* unhalt engine */ 662 temp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL)); 663 temp = REG_SET_FIELD(temp, SDMA0_F32_CNTL, HALT, 0); 664 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), temp); 665 } 666 667 /* enable DMA RB */ 668 rb_cntl = REG_SET_FIELD(rb_cntl, SDMA0_GFX_RB_CNTL, RB_ENABLE, 1); 669 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_RB_CNTL), rb_cntl); 670 671 ib_cntl = RREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL)); 672 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_ENABLE, 1); 673 #ifdef __BIG_ENDIAN 674 ib_cntl = REG_SET_FIELD(ib_cntl, SDMA0_GFX_IB_CNTL, IB_SWAP_ENABLE, 1); 675 #endif 676 /* enable DMA IBs */ 677 WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_GFX_IB_CNTL), ib_cntl); 678 679 if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */ 680 sdma_v5_2_ctx_switch_enable(adev, true); 681 sdma_v5_2_enable(adev, true); 682 } 683 684 return amdgpu_ring_test_helper(ring); 685 } 686 687 /** 688 * sdma_v5_2_gfx_resume - setup and start the async dma engines 689 * 690 * @adev: amdgpu_device pointer 691 * 692 * Set up the gfx DMA ring buffers and enable them. 693 * Returns 0 for success, error for failure. 694 */ 695 static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev) 696 { 697 int i, r; 698 699 for (i = 0; i < adev->sdma.num_instances; i++) { 700 r = sdma_v5_2_gfx_resume_instance(adev, i, false); 701 if (r) 702 return r; 703 } 704 705 return 0; 706 } 707 708 /** 709 * sdma_v5_2_rlc_resume - setup and start the async dma engines 710 * 711 * @adev: amdgpu_device pointer 712 * 713 * Set up the compute DMA queues and enable them. 714 * Returns 0 for success, error for failure. 715 */ 716 static int sdma_v5_2_rlc_resume(struct amdgpu_device *adev) 717 { 718 return 0; 719 } 720 721 /** 722 * sdma_v5_2_load_microcode - load the sDMA ME ucode 723 * 724 * @adev: amdgpu_device pointer 725 * 726 * Loads the sDMA0/1/2/3 ucode. 727 * Returns 0 for success, -EINVAL if the ucode is not available. 728 */ 729 static int sdma_v5_2_load_microcode(struct amdgpu_device *adev) 730 { 731 const struct sdma_firmware_header_v1_0 *hdr; 732 const __le32 *fw_data; 733 u32 fw_size; 734 int i, j; 735 736 /* halt the MEs */ 737 sdma_v5_2_enable(adev, false); 738 739 for (i = 0; i < adev->sdma.num_instances; i++) { 740 if (!adev->sdma.instance[i].fw) 741 return -EINVAL; 742 743 hdr = (const struct sdma_firmware_header_v1_0 *)adev->sdma.instance[i].fw->data; 744 amdgpu_ucode_print_sdma_hdr(&hdr->header); 745 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 746 747 fw_data = (const __le32 *) 748 (adev->sdma.instance[i].fw->data + 749 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 750 751 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), 0); 752 753 for (j = 0; j < fw_size; j++) { 754 if (amdgpu_emu_mode == 1 && j % 500 == 0) 755 msleep(1); 756 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_DATA), le32_to_cpup(fw_data++)); 757 } 758 759 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_UCODE_ADDR), adev->sdma.instance[i].fw_version); 760 } 761 762 return 0; 763 } 764 765 static int sdma_v5_2_soft_reset_engine(struct amdgpu_device *adev, u32 instance_id) 766 { 767 u32 grbm_soft_reset; 768 u32 tmp; 769 770 grbm_soft_reset = REG_SET_FIELD(0, 771 GRBM_SOFT_RESET, SOFT_RESET_SDMA0, 772 1); 773 grbm_soft_reset <<= instance_id; 774 775 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET); 776 tmp |= grbm_soft_reset; 777 DRM_DEBUG("GRBM_SOFT_RESET=0x%08X\n", tmp); 778 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp); 779 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET); 780 781 udelay(50); 782 783 tmp &= ~grbm_soft_reset; 784 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp); 785 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET); 786 return 0; 787 } 788 789 static int sdma_v5_2_soft_reset(struct amdgpu_ip_block *ip_block) 790 { 791 struct amdgpu_device *adev = ip_block->adev; 792 int i; 793 794 for (i = 0; i < adev->sdma.num_instances; i++) { 795 sdma_v5_2_soft_reset_engine(adev, i); 796 udelay(50); 797 } 798 799 return 0; 800 } 801 802 static const struct amdgpu_sdma_funcs sdma_v5_2_sdma_funcs = { 803 .stop_kernel_queue = &sdma_v5_2_stop_queue, 804 .start_kernel_queue = &sdma_v5_2_restore_queue, 805 .soft_reset_kernel_queue = &sdma_v5_2_soft_reset_engine, 806 }; 807 808 /** 809 * sdma_v5_2_start - setup and start the async dma engines 810 * 811 * @adev: amdgpu_device pointer 812 * 813 * Set up the DMA engines and enable them. 814 * Returns 0 for success, error for failure. 815 */ 816 static int sdma_v5_2_start(struct amdgpu_device *adev) 817 { 818 int r = 0; 819 struct amdgpu_ip_block *ip_block; 820 821 if (amdgpu_sriov_vf(adev)) { 822 sdma_v5_2_ctx_switch_enable(adev, false); 823 sdma_v5_2_enable(adev, false); 824 825 /* set RB registers */ 826 r = sdma_v5_2_gfx_resume(adev); 827 return r; 828 } 829 830 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 831 r = sdma_v5_2_load_microcode(adev); 832 if (r) 833 return r; 834 835 /* The value of mmSDMA_F32_CNTL is invalid the moment after loading fw */ 836 if (amdgpu_emu_mode == 1) 837 msleep(1000); 838 } 839 840 ip_block = amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_SDMA); 841 if (!ip_block) 842 return -EINVAL; 843 844 sdma_v5_2_soft_reset(ip_block); 845 /* unhalt the MEs */ 846 sdma_v5_2_enable(adev, true); 847 /* enable sdma ring preemption */ 848 sdma_v5_2_ctx_switch_enable(adev, true); 849 850 /* start the gfx rings and rlc compute queues */ 851 r = sdma_v5_2_gfx_resume(adev); 852 if (r) 853 return r; 854 r = sdma_v5_2_rlc_resume(adev); 855 856 return r; 857 } 858 859 static int sdma_v5_2_mqd_init(struct amdgpu_device *adev, void *mqd, 860 struct amdgpu_mqd_prop *prop) 861 { 862 struct v10_sdma_mqd *m = mqd; 863 uint64_t wb_gpu_addr; 864 865 m->sdmax_rlcx_rb_cntl = 866 order_base_2(prop->queue_size / 4) << SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT | 867 1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT | 868 6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT | 869 1 << SDMA0_RLC0_RB_CNTL__RB_PRIV__SHIFT; 870 871 m->sdmax_rlcx_rb_base = lower_32_bits(prop->hqd_base_gpu_addr >> 8); 872 m->sdmax_rlcx_rb_base_hi = upper_32_bits(prop->hqd_base_gpu_addr >> 8); 873 874 m->sdmax_rlcx_rb_wptr_poll_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0, 875 mmSDMA0_GFX_RB_WPTR_POLL_CNTL)); 876 877 wb_gpu_addr = prop->wptr_gpu_addr; 878 m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits(wb_gpu_addr); 879 m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr); 880 881 wb_gpu_addr = prop->rptr_gpu_addr; 882 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits(wb_gpu_addr); 883 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits(wb_gpu_addr); 884 885 m->sdmax_rlcx_ib_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, 0, 886 mmSDMA0_GFX_IB_CNTL)); 887 888 m->sdmax_rlcx_doorbell_offset = 889 prop->doorbell_index << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT; 890 891 m->sdmax_rlcx_doorbell = REG_SET_FIELD(0, SDMA0_RLC0_DOORBELL, ENABLE, 1); 892 893 return 0; 894 } 895 896 static void sdma_v5_2_set_mqd_funcs(struct amdgpu_device *adev) 897 { 898 adev->mqds[AMDGPU_HW_IP_DMA].mqd_size = sizeof(struct v10_sdma_mqd); 899 adev->mqds[AMDGPU_HW_IP_DMA].init_mqd = sdma_v5_2_mqd_init; 900 } 901 902 /** 903 * sdma_v5_2_ring_test_ring - simple async dma engine test 904 * 905 * @ring: amdgpu_ring structure holding ring information 906 * 907 * Test the DMA engine by writing using it to write an 908 * value to memory. 909 * Returns 0 for success, error for failure. 910 */ 911 static int sdma_v5_2_ring_test_ring(struct amdgpu_ring *ring) 912 { 913 struct amdgpu_device *adev = ring->adev; 914 unsigned i; 915 unsigned index; 916 int r; 917 u32 tmp; 918 u64 gpu_addr; 919 920 tmp = 0xCAFEDEAD; 921 922 r = amdgpu_device_wb_get(adev, &index); 923 if (r) { 924 dev_err(adev->dev, "(%d) failed to allocate wb slot\n", r); 925 return r; 926 } 927 928 gpu_addr = adev->wb.gpu_addr + (index * 4); 929 adev->wb.wb[index] = cpu_to_le32(tmp); 930 931 r = amdgpu_ring_alloc(ring, 20); 932 if (r) { 933 drm_err(adev_to_drm(adev), "dma failed to lock ring %d (%d).\n", ring->idx, r); 934 amdgpu_device_wb_free(adev, index); 935 return r; 936 } 937 938 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 939 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR)); 940 amdgpu_ring_write(ring, lower_32_bits(gpu_addr)); 941 amdgpu_ring_write(ring, upper_32_bits(gpu_addr)); 942 amdgpu_ring_write(ring, SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0)); 943 amdgpu_ring_write(ring, 0xDEADBEEF); 944 amdgpu_ring_commit(ring); 945 946 for (i = 0; i < adev->usec_timeout; i++) { 947 tmp = le32_to_cpu(adev->wb.wb[index]); 948 if (tmp == 0xDEADBEEF) 949 break; 950 if (amdgpu_emu_mode == 1) 951 msleep(1); 952 else 953 udelay(1); 954 } 955 956 if (i >= adev->usec_timeout) 957 r = -ETIMEDOUT; 958 959 amdgpu_device_wb_free(adev, index); 960 961 return r; 962 } 963 964 /** 965 * sdma_v5_2_ring_test_ib - test an IB on the DMA engine 966 * 967 * @ring: amdgpu_ring structure holding ring information 968 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT 969 * 970 * Test a simple IB in the DMA ring. 971 * Returns 0 on success, error on failure. 972 */ 973 static int sdma_v5_2_ring_test_ib(struct amdgpu_ring *ring, long timeout) 974 { 975 struct amdgpu_device *adev = ring->adev; 976 struct amdgpu_ib ib; 977 struct dma_fence *f = NULL; 978 unsigned index; 979 long r; 980 u32 tmp = 0; 981 u64 gpu_addr; 982 983 tmp = 0xCAFEDEAD; 984 memset(&ib, 0, sizeof(ib)); 985 986 r = amdgpu_device_wb_get(adev, &index); 987 if (r) { 988 dev_err(adev->dev, "(%ld) failed to allocate wb slot\n", r); 989 return r; 990 } 991 992 gpu_addr = adev->wb.gpu_addr + (index * 4); 993 adev->wb.wb[index] = cpu_to_le32(tmp); 994 995 r = amdgpu_ib_get(adev, NULL, 256, AMDGPU_IB_POOL_DIRECT, &ib); 996 if (r) { 997 drm_err(adev_to_drm(adev), "failed to get ib (%ld).\n", r); 998 goto err0; 999 } 1000 1001 ib.ptr[0] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1002 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1003 ib.ptr[1] = lower_32_bits(gpu_addr); 1004 ib.ptr[2] = upper_32_bits(gpu_addr); 1005 ib.ptr[3] = SDMA_PKT_WRITE_UNTILED_DW_3_COUNT(0); 1006 ib.ptr[4] = 0xDEADBEEF; 1007 ib.ptr[5] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1008 ib.ptr[6] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1009 ib.ptr[7] = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP); 1010 ib.length_dw = 8; 1011 1012 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 1013 if (r) 1014 goto err1; 1015 1016 r = dma_fence_wait_timeout(f, false, timeout); 1017 if (r == 0) { 1018 drm_err(adev_to_drm(adev), "IB test timed out\n"); 1019 r = -ETIMEDOUT; 1020 goto err1; 1021 } else if (r < 0) { 1022 drm_err(adev_to_drm(adev), "fence wait failed (%ld).\n", r); 1023 goto err1; 1024 } 1025 1026 tmp = le32_to_cpu(adev->wb.wb[index]); 1027 1028 if (tmp == 0xDEADBEEF) 1029 r = 0; 1030 else 1031 r = -EINVAL; 1032 1033 err1: 1034 amdgpu_ib_free(&ib, NULL); 1035 dma_fence_put(f); 1036 err0: 1037 amdgpu_device_wb_free(adev, index); 1038 return r; 1039 } 1040 1041 1042 /** 1043 * sdma_v5_2_vm_copy_pte - update PTEs by copying them from the GART 1044 * 1045 * @ib: indirect buffer to fill with commands 1046 * @pe: addr of the page entry 1047 * @src: src addr to copy from 1048 * @count: number of page entries to update 1049 * 1050 * Update PTEs by copying them from the GART using sDMA. 1051 */ 1052 static void sdma_v5_2_vm_copy_pte(struct amdgpu_ib *ib, 1053 uint64_t pe, uint64_t src, 1054 unsigned count) 1055 { 1056 unsigned bytes = count * 8; 1057 1058 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 1059 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR); 1060 ib->ptr[ib->length_dw++] = bytes - 1; 1061 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 1062 ib->ptr[ib->length_dw++] = lower_32_bits(src); 1063 ib->ptr[ib->length_dw++] = upper_32_bits(src); 1064 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1065 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1066 1067 } 1068 1069 /** 1070 * sdma_v5_2_vm_write_pte - update PTEs by writing them manually 1071 * 1072 * @ib: indirect buffer to fill with commands 1073 * @pe: addr of the page entry 1074 * @value: dst addr to write into pe 1075 * @count: number of page entries to update 1076 * @incr: increase next addr by incr bytes 1077 * 1078 * Update PTEs by writing them manually using sDMA. 1079 */ 1080 static void sdma_v5_2_vm_write_pte(struct amdgpu_ib *ib, uint64_t pe, 1081 uint64_t value, unsigned count, 1082 uint32_t incr) 1083 { 1084 unsigned ndw = count * 2; 1085 1086 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_WRITE) | 1087 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_WRITE_LINEAR); 1088 ib->ptr[ib->length_dw++] = lower_32_bits(pe); 1089 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1090 ib->ptr[ib->length_dw++] = ndw - 1; 1091 for (; ndw > 0; ndw -= 2) { 1092 ib->ptr[ib->length_dw++] = lower_32_bits(value); 1093 ib->ptr[ib->length_dw++] = upper_32_bits(value); 1094 value += incr; 1095 } 1096 } 1097 1098 /** 1099 * sdma_v5_2_vm_set_pte_pde - update the page tables using sDMA 1100 * 1101 * @ib: indirect buffer to fill with commands 1102 * @pe: addr of the page entry 1103 * @addr: dst addr to write into pe 1104 * @count: number of page entries to update 1105 * @incr: increase next addr by incr bytes 1106 * @flags: access flags 1107 * 1108 * Update the page tables using sDMA. 1109 */ 1110 static void sdma_v5_2_vm_set_pte_pde(struct amdgpu_ib *ib, 1111 uint64_t pe, 1112 uint64_t addr, unsigned count, 1113 uint32_t incr, uint64_t flags) 1114 { 1115 /* for physically contiguous pages (vram) */ 1116 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_PTEPDE); 1117 ib->ptr[ib->length_dw++] = lower_32_bits(pe); /* dst addr */ 1118 ib->ptr[ib->length_dw++] = upper_32_bits(pe); 1119 ib->ptr[ib->length_dw++] = lower_32_bits(flags); /* mask */ 1120 ib->ptr[ib->length_dw++] = upper_32_bits(flags); 1121 ib->ptr[ib->length_dw++] = lower_32_bits(addr); /* value */ 1122 ib->ptr[ib->length_dw++] = upper_32_bits(addr); 1123 ib->ptr[ib->length_dw++] = incr; /* increment size */ 1124 ib->ptr[ib->length_dw++] = 0; 1125 ib->ptr[ib->length_dw++] = count - 1; /* number of entries */ 1126 } 1127 1128 /** 1129 * sdma_v5_2_ring_pad_ib - pad the IB 1130 * 1131 * @ib: indirect buffer to fill with padding 1132 * @ring: amdgpu_ring structure holding ring information 1133 * 1134 * Pad the IB with NOPs to a boundary multiple of 8. 1135 */ 1136 static void sdma_v5_2_ring_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib) 1137 { 1138 struct amdgpu_sdma_instance *sdma = amdgpu_sdma_get_instance_from_ring(ring); 1139 u32 pad_count; 1140 int i; 1141 1142 pad_count = (-ib->length_dw) & 0x7; 1143 for (i = 0; i < pad_count; i++) 1144 if (sdma && sdma->burst_nop && (i == 0)) 1145 ib->ptr[ib->length_dw++] = 1146 SDMA_PKT_HEADER_OP(SDMA_OP_NOP) | 1147 SDMA_PKT_NOP_HEADER_COUNT(pad_count - 1); 1148 else 1149 ib->ptr[ib->length_dw++] = 1150 SDMA_PKT_HEADER_OP(SDMA_OP_NOP); 1151 } 1152 1153 1154 /** 1155 * sdma_v5_2_ring_emit_pipeline_sync - sync the pipeline 1156 * 1157 * @ring: amdgpu_ring pointer 1158 * 1159 * Make sure all previous operations are completed (CIK). 1160 */ 1161 static void sdma_v5_2_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 1162 { 1163 uint32_t seq = ring->fence_drv.sync_seq; 1164 uint64_t addr = ring->fence_drv.gpu_addr; 1165 1166 /* wait for idle */ 1167 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 1168 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) | 1169 SDMA_PKT_POLL_REGMEM_HEADER_FUNC(3) | /* equal */ 1170 SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)); 1171 amdgpu_ring_write(ring, addr & 0xfffffffc); 1172 amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff); 1173 amdgpu_ring_write(ring, seq); /* reference */ 1174 amdgpu_ring_write(ring, 0xffffffff); /* mask */ 1175 amdgpu_ring_write(ring, SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff) | 1176 SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(4)); /* retry count, poll interval */ 1177 } 1178 1179 1180 /** 1181 * sdma_v5_2_ring_emit_vm_flush - vm flush using sDMA 1182 * 1183 * @ring: amdgpu_ring pointer 1184 * @vmid: vmid number to use 1185 * @pd_addr: address 1186 * 1187 * Update the page table base and flush the VM TLB 1188 * using sDMA. 1189 */ 1190 static void sdma_v5_2_ring_emit_vm_flush(struct amdgpu_ring *ring, 1191 unsigned vmid, uint64_t pd_addr) 1192 { 1193 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub]; 1194 uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0); 1195 1196 /* Update the PD address for this VMID. */ 1197 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 1198 (hub->ctx_addr_distance * vmid), 1199 lower_32_bits(pd_addr)); 1200 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 1201 (hub->ctx_addr_distance * vmid), 1202 upper_32_bits(pd_addr)); 1203 1204 /* Trigger invalidation. */ 1205 amdgpu_ring_write(ring, 1206 SDMA_PKT_VM_INVALIDATION_HEADER_OP(SDMA_OP_POLL_REGMEM) | 1207 SDMA_PKT_VM_INVALIDATION_HEADER_SUB_OP(SDMA_SUBOP_VM_INVALIDATION) | 1208 SDMA_PKT_VM_INVALIDATION_HEADER_GFX_ENG_ID(ring->vm_inv_eng) | 1209 SDMA_PKT_VM_INVALIDATION_HEADER_MM_ENG_ID(0x1f)); 1210 amdgpu_ring_write(ring, req); 1211 amdgpu_ring_write(ring, 0xFFFFFFFF); 1212 amdgpu_ring_write(ring, 1213 SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_INVALIDATEACK(1 << vmid) | 1214 SDMA_PKT_VM_INVALIDATION_ADDRESSRANGEHI_ADDRESSRANGEHI(0x1F)); 1215 } 1216 1217 static void sdma_v5_2_ring_emit_wreg(struct amdgpu_ring *ring, 1218 uint32_t reg, uint32_t val) 1219 { 1220 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_SRBM_WRITE) | 1221 SDMA_PKT_SRBM_WRITE_HEADER_BYTE_EN(0xf)); 1222 amdgpu_ring_write(ring, reg); 1223 amdgpu_ring_write(ring, val); 1224 } 1225 1226 static void sdma_v5_2_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg, 1227 uint32_t val, uint32_t mask) 1228 { 1229 amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) | 1230 SDMA_PKT_POLL_REGMEM_HEADER_HDP_FLUSH(0) | 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_v5_2_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 const struct amdgpu_vm_pte_funcs sdma_v5_2_vm_pte_funcs = { 1251 .copy_pte_num_dw = 7, 1252 .copy_pte = sdma_v5_2_vm_copy_pte, 1253 .write_pte = sdma_v5_2_vm_write_pte, 1254 .set_pte_pde = sdma_v5_2_vm_set_pte_pde, 1255 }; 1256 1257 static int sdma_v5_2_early_init(struct amdgpu_ip_block *ip_block) 1258 { 1259 struct amdgpu_device *adev = ip_block->adev; 1260 int r; 1261 1262 r = amdgpu_sdma_init_microcode(adev, 0, true); 1263 if (r) 1264 return r; 1265 1266 sdma_v5_2_set_ring_funcs(adev); 1267 sdma_v5_2_set_buffer_funcs(adev); 1268 amdgpu_sdma_set_vm_pte_scheds(adev, &sdma_v5_2_vm_pte_funcs); 1269 sdma_v5_2_set_irq_funcs(adev); 1270 sdma_v5_2_set_mqd_funcs(adev); 1271 1272 return 0; 1273 } 1274 1275 static unsigned sdma_v5_2_seq_to_irq_id(int seq_num) 1276 { 1277 switch (seq_num) { 1278 case 0: 1279 return SOC15_IH_CLIENTID_SDMA0; 1280 case 1: 1281 return SOC15_IH_CLIENTID_SDMA1; 1282 case 2: 1283 return SOC15_IH_CLIENTID_SDMA2; 1284 case 3: 1285 return SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid; 1286 default: 1287 break; 1288 } 1289 return -EINVAL; 1290 } 1291 1292 static unsigned sdma_v5_2_seq_to_trap_id(int seq_num) 1293 { 1294 switch (seq_num) { 1295 case 0: 1296 return SDMA0_5_0__SRCID__SDMA_TRAP; 1297 case 1: 1298 return SDMA1_5_0__SRCID__SDMA_TRAP; 1299 case 2: 1300 return SDMA2_5_0__SRCID__SDMA_TRAP; 1301 case 3: 1302 return SDMA3_5_0__SRCID__SDMA_TRAP; 1303 default: 1304 break; 1305 } 1306 return -EINVAL; 1307 } 1308 1309 static int sdma_v5_2_sw_init(struct amdgpu_ip_block *ip_block) 1310 { 1311 struct amdgpu_ring *ring; 1312 int r, i; 1313 struct amdgpu_device *adev = ip_block->adev; 1314 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2); 1315 uint32_t *ptr; 1316 1317 /* SDMA trap event */ 1318 for (i = 0; i < adev->sdma.num_instances; i++) { 1319 r = amdgpu_irq_add_id(adev, sdma_v5_2_seq_to_irq_id(i), 1320 sdma_v5_2_seq_to_trap_id(i), 1321 &adev->sdma.trap_irq); 1322 if (r) 1323 return r; 1324 } 1325 1326 for (i = 0; i < adev->sdma.num_instances; i++) { 1327 mutex_init(&adev->sdma.instance[i].engine_reset_mutex); 1328 adev->sdma.instance[i].funcs = &sdma_v5_2_sdma_funcs; 1329 ring = &adev->sdma.instance[i].ring; 1330 ring->ring_obj = NULL; 1331 ring->use_doorbell = true; 1332 ring->me = i; 1333 1334 drm_info(adev_to_drm(adev), "use_doorbell being set to: [%s]\n", 1335 ring->use_doorbell?"true":"false"); 1336 1337 ring->doorbell_index = 1338 (adev->doorbell_index.sdma_engine[i] << 1); //get DWORD offset 1339 1340 ring->vm_hub = AMDGPU_GFXHUB(0); 1341 sprintf(ring->name, "sdma%d", i); 1342 r = amdgpu_ring_init(adev, ring, 1024, &adev->sdma.trap_irq, 1343 AMDGPU_SDMA_IRQ_INSTANCE0 + i, 1344 AMDGPU_RING_PRIO_DEFAULT, NULL); 1345 if (r) 1346 return r; 1347 } 1348 1349 adev->sdma.supported_reset = 1350 amdgpu_get_soft_full_reset_mask(&adev->sdma.instance[0].ring); 1351 if (!amdgpu_sriov_vf(adev) && 1352 !adev->debug_disable_gpu_ring_reset) 1353 adev->sdma.supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE; 1354 1355 /* Allocate memory for SDMA IP Dump buffer */ 1356 ptr = kcalloc(adev->sdma.num_instances * reg_count, sizeof(uint32_t), GFP_KERNEL); 1357 if (ptr) 1358 adev->sdma.ip_dump = ptr; 1359 else 1360 DRM_ERROR("Failed to allocated memory for SDMA IP Dump\n"); 1361 1362 r = amdgpu_sdma_sysfs_reset_mask_init(adev); 1363 if (r) 1364 return r; 1365 1366 return r; 1367 } 1368 1369 static int sdma_v5_2_sw_fini(struct amdgpu_ip_block *ip_block) 1370 { 1371 struct amdgpu_device *adev = ip_block->adev; 1372 int i; 1373 1374 for (i = 0; i < adev->sdma.num_instances; i++) 1375 amdgpu_ring_fini(&adev->sdma.instance[i].ring); 1376 1377 amdgpu_sdma_sysfs_reset_mask_fini(adev); 1378 amdgpu_sdma_destroy_inst_ctx(adev, true); 1379 1380 kfree(adev->sdma.ip_dump); 1381 1382 return 0; 1383 } 1384 1385 static int sdma_v5_2_hw_init(struct amdgpu_ip_block *ip_block) 1386 { 1387 struct amdgpu_device *adev = ip_block->adev; 1388 1389 return sdma_v5_2_start(adev); 1390 } 1391 1392 static int sdma_v5_2_hw_fini(struct amdgpu_ip_block *ip_block) 1393 { 1394 struct amdgpu_device *adev = ip_block->adev; 1395 1396 if (amdgpu_sriov_vf(adev)) 1397 return 0; 1398 1399 sdma_v5_2_ctx_switch_enable(adev, false); 1400 sdma_v5_2_enable(adev, false); 1401 1402 return 0; 1403 } 1404 1405 static int sdma_v5_2_suspend(struct amdgpu_ip_block *ip_block) 1406 { 1407 return sdma_v5_2_hw_fini(ip_block); 1408 } 1409 1410 static int sdma_v5_2_resume(struct amdgpu_ip_block *ip_block) 1411 { 1412 return sdma_v5_2_hw_init(ip_block); 1413 } 1414 1415 static bool sdma_v5_2_is_idle(struct amdgpu_ip_block *ip_block) 1416 { 1417 struct amdgpu_device *adev = ip_block->adev; 1418 u32 i; 1419 1420 for (i = 0; i < adev->sdma.num_instances; i++) { 1421 u32 tmp = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS_REG)); 1422 1423 if (!(tmp & SDMA0_STATUS_REG__IDLE_MASK)) 1424 return false; 1425 } 1426 1427 return true; 1428 } 1429 1430 static int sdma_v5_2_wait_for_idle(struct amdgpu_ip_block *ip_block) 1431 { 1432 unsigned i; 1433 u32 sdma0, sdma1, sdma2, sdma3; 1434 struct amdgpu_device *adev = ip_block->adev; 1435 1436 for (i = 0; i < adev->usec_timeout; i++) { 1437 sdma0 = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_STATUS_REG)); 1438 sdma1 = RREG32(sdma_v5_2_get_reg_offset(adev, 1, mmSDMA0_STATUS_REG)); 1439 sdma2 = RREG32(sdma_v5_2_get_reg_offset(adev, 2, mmSDMA0_STATUS_REG)); 1440 sdma3 = RREG32(sdma_v5_2_get_reg_offset(adev, 3, mmSDMA0_STATUS_REG)); 1441 1442 if (sdma0 & sdma1 & sdma2 & sdma3 & SDMA0_STATUS_REG__IDLE_MASK) 1443 return 0; 1444 udelay(1); 1445 } 1446 return -ETIMEDOUT; 1447 } 1448 1449 static int sdma_v5_2_reset_queue(struct amdgpu_ring *ring, 1450 unsigned int vmid, 1451 struct amdgpu_fence *timedout_fence) 1452 { 1453 struct amdgpu_device *adev = ring->adev; 1454 int r; 1455 1456 if (ring->me >= adev->sdma.num_instances) { 1457 dev_err(adev->dev, "sdma instance not found\n"); 1458 return -EINVAL; 1459 } 1460 1461 amdgpu_ring_reset_helper_begin(ring, timedout_fence); 1462 1463 amdgpu_amdkfd_suspend(adev, true); 1464 r = amdgpu_sdma_reset_engine(adev, ring->me, true); 1465 amdgpu_amdkfd_resume(adev, true); 1466 if (r) 1467 return r; 1468 1469 return amdgpu_ring_reset_helper_end(ring, timedout_fence); 1470 } 1471 1472 static int sdma_v5_2_stop_queue(struct amdgpu_ring *ring) 1473 { 1474 u32 f32_cntl, freeze, cntl, stat1_reg; 1475 struct amdgpu_device *adev = ring->adev; 1476 int i, j, r = 0; 1477 1478 if (amdgpu_sriov_vf(adev)) 1479 return -EINVAL; 1480 1481 i = ring->me; 1482 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 1483 1484 /* stop queue */ 1485 sdma_v5_2_gfx_stop(adev, 1 << i); 1486 1487 /*engine stop SDMA1_F32_CNTL.HALT to 1 and SDMAx_FREEZE freeze bit to 1 */ 1488 freeze = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE)); 1489 freeze = REG_SET_FIELD(freeze, SDMA0_FREEZE, FREEZE, 1); 1490 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE), freeze); 1491 1492 for (j = 0; j < adev->usec_timeout; j++) { 1493 freeze = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_FREEZE)); 1494 1495 if (REG_GET_FIELD(freeze, SDMA0_FREEZE, FROZEN) & 1) 1496 break; 1497 udelay(1); 1498 } 1499 1500 1501 if (j == adev->usec_timeout) { 1502 stat1_reg = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_STATUS1_REG)); 1503 if ((stat1_reg & 0x3FF) != 0x3FF) { 1504 DRM_ERROR("cannot soft reset as sdma not idle\n"); 1505 r = -ETIMEDOUT; 1506 goto err0; 1507 } 1508 } 1509 1510 f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL)); 1511 f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, 1); 1512 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl); 1513 1514 cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL)); 1515 cntl = REG_SET_FIELD(cntl, SDMA0_CNTL, UTC_L1_ENABLE, 0); 1516 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), cntl); 1517 1518 err0: 1519 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 1520 return r; 1521 } 1522 1523 static int sdma_v5_2_restore_queue(struct amdgpu_ring *ring) 1524 { 1525 struct amdgpu_device *adev = ring->adev; 1526 u32 inst_id = ring->me; 1527 u32 freeze; 1528 int r; 1529 1530 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 1531 /* unfreeze and unhalt */ 1532 freeze = RREG32(sdma_v5_2_get_reg_offset(adev, inst_id, mmSDMA0_FREEZE)); 1533 freeze = REG_SET_FIELD(freeze, SDMA0_FREEZE, FREEZE, 0); 1534 WREG32(sdma_v5_2_get_reg_offset(adev, inst_id, mmSDMA0_FREEZE), freeze); 1535 1536 r = sdma_v5_2_gfx_resume_instance(adev, inst_id, true); 1537 1538 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 1539 1540 return r; 1541 } 1542 1543 static int sdma_v5_2_ring_preempt_ib(struct amdgpu_ring *ring) 1544 { 1545 int i, r = 0; 1546 struct amdgpu_device *adev = ring->adev; 1547 u32 index = 0; 1548 u64 sdma_gfx_preempt; 1549 1550 amdgpu_sdma_get_index_from_ring(ring, &index); 1551 sdma_gfx_preempt = 1552 sdma_v5_2_get_reg_offset(adev, index, mmSDMA0_GFX_PREEMPT); 1553 1554 /* assert preemption condition */ 1555 amdgpu_ring_set_preempt_cond_exec(ring, false); 1556 1557 /* emit the trailing fence */ 1558 ring->trail_seq += 1; 1559 amdgpu_ring_alloc(ring, 10); 1560 sdma_v5_2_ring_emit_fence(ring, ring->trail_fence_gpu_addr, 1561 ring->trail_seq, 0); 1562 amdgpu_ring_commit(ring); 1563 1564 /* assert IB preemption */ 1565 WREG32(sdma_gfx_preempt, 1); 1566 1567 /* poll the trailing fence */ 1568 for (i = 0; i < adev->usec_timeout; i++) { 1569 if (ring->trail_seq == 1570 le32_to_cpu(*(ring->trail_fence_cpu_addr))) 1571 break; 1572 udelay(1); 1573 } 1574 1575 if (i >= adev->usec_timeout) { 1576 r = -EINVAL; 1577 DRM_ERROR("ring %d failed to be preempted\n", ring->idx); 1578 } 1579 1580 /* deassert IB preemption */ 1581 WREG32(sdma_gfx_preempt, 0); 1582 1583 /* deassert the preemption condition */ 1584 amdgpu_ring_set_preempt_cond_exec(ring, true); 1585 return r; 1586 } 1587 1588 static int sdma_v5_2_set_trap_irq_state(struct amdgpu_device *adev, 1589 struct amdgpu_irq_src *source, 1590 unsigned type, 1591 enum amdgpu_interrupt_state state) 1592 { 1593 u32 sdma_cntl; 1594 u32 reg_offset = sdma_v5_2_get_reg_offset(adev, type, mmSDMA0_CNTL); 1595 1596 if (!amdgpu_sriov_vf(adev)) { 1597 sdma_cntl = RREG32(reg_offset); 1598 sdma_cntl = REG_SET_FIELD(sdma_cntl, SDMA0_CNTL, TRAP_ENABLE, 1599 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 1600 WREG32(reg_offset, sdma_cntl); 1601 } 1602 1603 return 0; 1604 } 1605 1606 static int sdma_v5_2_process_trap_irq(struct amdgpu_device *adev, 1607 struct amdgpu_irq_src *source, 1608 struct amdgpu_iv_entry *entry) 1609 { 1610 DRM_DEBUG("IH: SDMA trap\n"); 1611 1612 if (drm_WARN_ON_ONCE(&adev->ddev, 1613 adev->enable_mes && 1614 (entry->src_data[0] & AMDGPU_FENCE_MES_QUEUE_FLAG))) 1615 return 0; 1616 1617 switch (entry->client_id) { 1618 case SOC15_IH_CLIENTID_SDMA0: 1619 switch (entry->ring_id) { 1620 case 0: 1621 amdgpu_fence_process(&adev->sdma.instance[0].ring); 1622 break; 1623 case 1: 1624 /* XXX compute */ 1625 break; 1626 case 2: 1627 /* XXX compute */ 1628 break; 1629 case 3: 1630 /* XXX page queue*/ 1631 break; 1632 } 1633 break; 1634 case SOC15_IH_CLIENTID_SDMA1: 1635 switch (entry->ring_id) { 1636 case 0: 1637 amdgpu_fence_process(&adev->sdma.instance[1].ring); 1638 break; 1639 case 1: 1640 /* XXX compute */ 1641 break; 1642 case 2: 1643 /* XXX compute */ 1644 break; 1645 case 3: 1646 /* XXX page queue*/ 1647 break; 1648 } 1649 break; 1650 case SOC15_IH_CLIENTID_SDMA2: 1651 switch (entry->ring_id) { 1652 case 0: 1653 amdgpu_fence_process(&adev->sdma.instance[2].ring); 1654 break; 1655 case 1: 1656 /* XXX compute */ 1657 break; 1658 case 2: 1659 /* XXX compute */ 1660 break; 1661 case 3: 1662 /* XXX page queue*/ 1663 break; 1664 } 1665 break; 1666 case SOC15_IH_CLIENTID_SDMA3_Sienna_Cichlid: 1667 switch (entry->ring_id) { 1668 case 0: 1669 amdgpu_fence_process(&adev->sdma.instance[3].ring); 1670 break; 1671 case 1: 1672 /* XXX compute */ 1673 break; 1674 case 2: 1675 /* XXX compute */ 1676 break; 1677 case 3: 1678 /* XXX page queue*/ 1679 break; 1680 } 1681 break; 1682 } 1683 return 0; 1684 } 1685 1686 static int sdma_v5_2_process_illegal_inst_irq(struct amdgpu_device *adev, 1687 struct amdgpu_irq_src *source, 1688 struct amdgpu_iv_entry *entry) 1689 { 1690 return 0; 1691 } 1692 1693 static bool sdma_v5_2_firmware_mgcg_support(struct amdgpu_device *adev, 1694 int i) 1695 { 1696 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 1697 case IP_VERSION(5, 2, 1): 1698 if (adev->sdma.instance[i].fw_version < 70) 1699 return false; 1700 break; 1701 case IP_VERSION(5, 2, 3): 1702 if (adev->sdma.instance[i].fw_version < 47) 1703 return false; 1704 break; 1705 case IP_VERSION(5, 2, 7): 1706 if (adev->sdma.instance[i].fw_version < 9) 1707 return false; 1708 break; 1709 default: 1710 return true; 1711 } 1712 1713 return true; 1714 1715 } 1716 1717 static void sdma_v5_2_update_medium_grain_clock_gating(struct amdgpu_device *adev, 1718 bool enable) 1719 { 1720 uint32_t data, def; 1721 int i; 1722 1723 for (i = 0; i < adev->sdma.num_instances; i++) { 1724 1725 if (!sdma_v5_2_firmware_mgcg_support(adev, i)) 1726 adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_MGCG; 1727 1728 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_MGCG)) { 1729 /* Enable sdma clock gating */ 1730 def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL)); 1731 data &= ~(SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK | 1732 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK | 1733 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK | 1734 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK | 1735 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK | 1736 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK); 1737 if (def != data) 1738 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data); 1739 } else { 1740 /* Disable sdma clock gating */ 1741 def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL)); 1742 data |= (SDMA0_CLK_CTRL__SOFT_OVERRIDE4_MASK | 1743 SDMA0_CLK_CTRL__SOFT_OVERRIDE3_MASK | 1744 SDMA0_CLK_CTRL__SOFT_OVERRIDE2_MASK | 1745 SDMA0_CLK_CTRL__SOFT_OVERRIDE1_MASK | 1746 SDMA0_CLK_CTRL__SOFT_OVERRIDE0_MASK | 1747 SDMA0_CLK_CTRL__SOFT_OVERRIDER_REG_MASK); 1748 if (def != data) 1749 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CLK_CTRL), data); 1750 } 1751 } 1752 } 1753 1754 static void sdma_v5_2_update_medium_grain_light_sleep(struct amdgpu_device *adev, 1755 bool enable) 1756 { 1757 uint32_t data, def; 1758 int i; 1759 1760 for (i = 0; i < adev->sdma.num_instances; i++) { 1761 if (adev->sdma.instance[i].fw_version < 70 && 1762 amdgpu_ip_version(adev, SDMA0_HWIP, 0) == 1763 IP_VERSION(5, 2, 1)) 1764 adev->cg_flags &= ~AMD_CG_SUPPORT_SDMA_LS; 1765 1766 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_SDMA_LS)) { 1767 /* Enable sdma mem light sleep */ 1768 def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL)); 1769 data |= SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK; 1770 if (def != data) 1771 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data); 1772 1773 } else { 1774 /* Disable sdma mem light sleep */ 1775 def = data = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL)); 1776 data &= ~SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK; 1777 if (def != data) 1778 WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_POWER_CNTL), data); 1779 1780 } 1781 } 1782 } 1783 1784 static int sdma_v5_2_set_clockgating_state(struct amdgpu_ip_block *ip_block, 1785 enum amd_clockgating_state state) 1786 { 1787 struct amdgpu_device *adev = ip_block->adev; 1788 1789 if (amdgpu_sriov_vf(adev)) 1790 return 0; 1791 1792 switch (amdgpu_ip_version(adev, SDMA0_HWIP, 0)) { 1793 case IP_VERSION(5, 2, 0): 1794 case IP_VERSION(5, 2, 2): 1795 case IP_VERSION(5, 2, 1): 1796 case IP_VERSION(5, 2, 4): 1797 case IP_VERSION(5, 2, 5): 1798 case IP_VERSION(5, 2, 6): 1799 case IP_VERSION(5, 2, 3): 1800 case IP_VERSION(5, 2, 7): 1801 sdma_v5_2_update_medium_grain_clock_gating(adev, 1802 state == AMD_CG_STATE_GATE); 1803 sdma_v5_2_update_medium_grain_light_sleep(adev, 1804 state == AMD_CG_STATE_GATE); 1805 break; 1806 default: 1807 break; 1808 } 1809 1810 return 0; 1811 } 1812 1813 static int sdma_v5_2_set_powergating_state(struct amdgpu_ip_block *ip_block, 1814 enum amd_powergating_state state) 1815 { 1816 return 0; 1817 } 1818 1819 static void sdma_v5_2_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags) 1820 { 1821 struct amdgpu_device *adev = ip_block->adev; 1822 int data; 1823 1824 if (amdgpu_sriov_vf(adev)) 1825 *flags = 0; 1826 1827 /* AMD_CG_SUPPORT_SDMA_MGCG */ 1828 data = RREG32(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_CLK_CTRL)); 1829 if (!(data & SDMA0_CLK_CTRL__CGCG_EN_OVERRIDE_MASK)) 1830 *flags |= AMD_CG_SUPPORT_SDMA_MGCG; 1831 1832 /* AMD_CG_SUPPORT_SDMA_LS */ 1833 data = RREG32_KIQ(sdma_v5_2_get_reg_offset(adev, 0, mmSDMA0_POWER_CNTL)); 1834 if (data & SDMA0_POWER_CNTL__MEM_POWER_OVERRIDE_MASK) 1835 *flags |= AMD_CG_SUPPORT_SDMA_LS; 1836 } 1837 1838 static void sdma_v5_2_ring_begin_use(struct amdgpu_ring *ring) 1839 { 1840 struct amdgpu_device *adev = ring->adev; 1841 1842 /* SDMA 5.2.3 (RMB) FW doesn't seem to properly 1843 * disallow GFXOFF in some cases leading to 1844 * hangs in SDMA. Disallow GFXOFF while SDMA is active. 1845 * We can probably just limit this to 5.2.3, 1846 * but it shouldn't hurt for other parts since 1847 * this GFXOFF will be disallowed anyway when SDMA is 1848 * active, this just makes it explicit. 1849 * sdma_v5_2_ring_set_wptr() takes advantage of this 1850 * to update the wptr because sometimes SDMA seems to miss 1851 * doorbells when entering PG. If you remove this, update 1852 * sdma_v5_2_ring_set_wptr() as well! 1853 */ 1854 amdgpu_gfx_off_ctrl(adev, false); 1855 } 1856 1857 static void sdma_v5_2_ring_end_use(struct amdgpu_ring *ring) 1858 { 1859 struct amdgpu_device *adev = ring->adev; 1860 1861 /* SDMA 5.2.3 (RMB) FW doesn't seem to properly 1862 * disallow GFXOFF in some cases leading to 1863 * hangs in SDMA. Allow GFXOFF when SDMA is complete. 1864 */ 1865 amdgpu_gfx_off_ctrl(adev, true); 1866 } 1867 1868 static void sdma_v5_2_print_ip_state(struct amdgpu_ip_block *ip_block, struct drm_printer *p) 1869 { 1870 struct amdgpu_device *adev = ip_block->adev; 1871 int i, j; 1872 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2); 1873 uint32_t instance_offset; 1874 1875 if (!adev->sdma.ip_dump) 1876 return; 1877 1878 drm_printf(p, "num_instances:%d\n", adev->sdma.num_instances); 1879 for (i = 0; i < adev->sdma.num_instances; i++) { 1880 instance_offset = i * reg_count; 1881 drm_printf(p, "\nInstance:%d\n", i); 1882 1883 for (j = 0; j < reg_count; j++) 1884 drm_printf(p, "%-50s \t 0x%08x\n", sdma_reg_list_5_2[j].reg_name, 1885 adev->sdma.ip_dump[instance_offset + j]); 1886 } 1887 } 1888 1889 static void sdma_v5_2_dump_ip_state(struct amdgpu_ip_block *ip_block) 1890 { 1891 struct amdgpu_device *adev = ip_block->adev; 1892 int i, j; 1893 uint32_t instance_offset; 1894 uint32_t reg_count = ARRAY_SIZE(sdma_reg_list_5_2); 1895 1896 if (!adev->sdma.ip_dump) 1897 return; 1898 1899 amdgpu_gfx_off_ctrl(adev, false); 1900 for (i = 0; i < adev->sdma.num_instances; i++) { 1901 instance_offset = i * reg_count; 1902 for (j = 0; j < reg_count; j++) 1903 adev->sdma.ip_dump[instance_offset + j] = 1904 RREG32(sdma_v5_2_get_reg_offset(adev, i, 1905 sdma_reg_list_5_2[j].reg_offset)); 1906 } 1907 amdgpu_gfx_off_ctrl(adev, true); 1908 } 1909 1910 static const struct amd_ip_funcs sdma_v5_2_ip_funcs = { 1911 .name = "sdma_v5_2", 1912 .early_init = sdma_v5_2_early_init, 1913 .sw_init = sdma_v5_2_sw_init, 1914 .sw_fini = sdma_v5_2_sw_fini, 1915 .hw_init = sdma_v5_2_hw_init, 1916 .hw_fini = sdma_v5_2_hw_fini, 1917 .suspend = sdma_v5_2_suspend, 1918 .resume = sdma_v5_2_resume, 1919 .is_idle = sdma_v5_2_is_idle, 1920 .wait_for_idle = sdma_v5_2_wait_for_idle, 1921 .soft_reset = sdma_v5_2_soft_reset, 1922 .set_clockgating_state = sdma_v5_2_set_clockgating_state, 1923 .set_powergating_state = sdma_v5_2_set_powergating_state, 1924 .get_clockgating_state = sdma_v5_2_get_clockgating_state, 1925 .dump_ip_state = sdma_v5_2_dump_ip_state, 1926 .print_ip_state = sdma_v5_2_print_ip_state, 1927 }; 1928 1929 static const struct amdgpu_ring_funcs sdma_v5_2_ring_funcs = { 1930 .type = AMDGPU_RING_TYPE_SDMA, 1931 .align_mask = 0xf, 1932 .nop = SDMA_PKT_NOP_HEADER_OP(SDMA_OP_NOP), 1933 .support_64bit_ptrs = true, 1934 .secure_submission_supported = true, 1935 .get_rptr = sdma_v5_2_ring_get_rptr, 1936 .get_wptr = sdma_v5_2_ring_get_wptr, 1937 .set_wptr = sdma_v5_2_ring_set_wptr, 1938 .emit_frame_size = 1939 5 + /* sdma_v5_2_ring_init_cond_exec */ 1940 6 + /* sdma_v5_2_ring_emit_hdp_flush */ 1941 3 + /* hdp_invalidate */ 1942 6 + /* sdma_v5_2_ring_emit_pipeline_sync */ 1943 /* sdma_v5_2_ring_emit_vm_flush */ 1944 SOC15_FLUSH_GPU_TLB_NUM_WREG * 3 + 1945 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 6 + 1946 10 + 10 + 10, /* sdma_v5_2_ring_emit_fence x3 for user fence, vm fence */ 1947 .emit_ib_size = 7 + 6, /* sdma_v5_2_ring_emit_ib */ 1948 .emit_ib = sdma_v5_2_ring_emit_ib, 1949 .emit_mem_sync = sdma_v5_2_ring_emit_mem_sync, 1950 .emit_fence = sdma_v5_2_ring_emit_fence, 1951 .emit_pipeline_sync = sdma_v5_2_ring_emit_pipeline_sync, 1952 .emit_vm_flush = sdma_v5_2_ring_emit_vm_flush, 1953 .emit_hdp_flush = sdma_v5_2_ring_emit_hdp_flush, 1954 .test_ring = sdma_v5_2_ring_test_ring, 1955 .test_ib = sdma_v5_2_ring_test_ib, 1956 .insert_nop = sdma_v5_2_ring_insert_nop, 1957 .pad_ib = sdma_v5_2_ring_pad_ib, 1958 .begin_use = sdma_v5_2_ring_begin_use, 1959 .end_use = sdma_v5_2_ring_end_use, 1960 .emit_wreg = sdma_v5_2_ring_emit_wreg, 1961 .emit_reg_wait = sdma_v5_2_ring_emit_reg_wait, 1962 .emit_reg_write_reg_wait = sdma_v5_2_ring_emit_reg_write_reg_wait, 1963 .init_cond_exec = sdma_v5_2_ring_init_cond_exec, 1964 .preempt_ib = sdma_v5_2_ring_preempt_ib, 1965 .reset = sdma_v5_2_reset_queue, 1966 }; 1967 1968 static void sdma_v5_2_set_ring_funcs(struct amdgpu_device *adev) 1969 { 1970 int i; 1971 1972 for (i = 0; i < adev->sdma.num_instances; i++) { 1973 adev->sdma.instance[i].ring.funcs = &sdma_v5_2_ring_funcs; 1974 adev->sdma.instance[i].ring.me = i; 1975 } 1976 } 1977 1978 static const struct amdgpu_irq_src_funcs sdma_v5_2_trap_irq_funcs = { 1979 .set = sdma_v5_2_set_trap_irq_state, 1980 .process = sdma_v5_2_process_trap_irq, 1981 }; 1982 1983 static const struct amdgpu_irq_src_funcs sdma_v5_2_illegal_inst_irq_funcs = { 1984 .process = sdma_v5_2_process_illegal_inst_irq, 1985 }; 1986 1987 static void sdma_v5_2_set_irq_funcs(struct amdgpu_device *adev) 1988 { 1989 adev->sdma.trap_irq.num_types = AMDGPU_SDMA_IRQ_INSTANCE0 + 1990 adev->sdma.num_instances; 1991 adev->sdma.trap_irq.funcs = &sdma_v5_2_trap_irq_funcs; 1992 adev->sdma.illegal_inst_irq.funcs = &sdma_v5_2_illegal_inst_irq_funcs; 1993 } 1994 1995 /** 1996 * sdma_v5_2_emit_copy_buffer - copy buffer using the sDMA engine 1997 * 1998 * @ib: indirect buffer to copy to 1999 * @src_offset: src GPU address 2000 * @dst_offset: dst GPU address 2001 * @byte_count: number of bytes to xfer 2002 * @copy_flags: copy flags for the buffers 2003 * 2004 * Copy GPU buffers using the DMA engine. 2005 * Used by the amdgpu ttm implementation to move pages if 2006 * registered as the asic copy callback. 2007 */ 2008 static void sdma_v5_2_emit_copy_buffer(struct amdgpu_ib *ib, 2009 uint64_t src_offset, 2010 uint64_t dst_offset, 2011 uint32_t byte_count, 2012 uint32_t copy_flags) 2013 { 2014 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_COPY) | 2015 SDMA_PKT_HEADER_SUB_OP(SDMA_SUBOP_COPY_LINEAR) | 2016 SDMA_PKT_COPY_LINEAR_HEADER_TMZ((copy_flags & AMDGPU_COPY_FLAGS_TMZ) ? 1 : 0); 2017 ib->ptr[ib->length_dw++] = byte_count - 1; 2018 ib->ptr[ib->length_dw++] = 0; /* src/dst endian swap */ 2019 ib->ptr[ib->length_dw++] = lower_32_bits(src_offset); 2020 ib->ptr[ib->length_dw++] = upper_32_bits(src_offset); 2021 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 2022 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 2023 } 2024 2025 /** 2026 * sdma_v5_2_emit_fill_buffer - fill buffer using the sDMA engine 2027 * 2028 * @ib: indirect buffer to fill 2029 * @src_data: value to write to buffer 2030 * @dst_offset: dst GPU address 2031 * @byte_count: number of bytes to xfer 2032 * 2033 * Fill GPU buffers using the DMA engine. 2034 */ 2035 static void sdma_v5_2_emit_fill_buffer(struct amdgpu_ib *ib, 2036 uint32_t src_data, 2037 uint64_t dst_offset, 2038 uint32_t byte_count) 2039 { 2040 ib->ptr[ib->length_dw++] = SDMA_PKT_HEADER_OP(SDMA_OP_CONST_FILL); 2041 ib->ptr[ib->length_dw++] = lower_32_bits(dst_offset); 2042 ib->ptr[ib->length_dw++] = upper_32_bits(dst_offset); 2043 ib->ptr[ib->length_dw++] = src_data; 2044 ib->ptr[ib->length_dw++] = byte_count - 1; 2045 } 2046 2047 static const struct amdgpu_buffer_funcs sdma_v5_2_buffer_funcs = { 2048 .copy_max_bytes = 1 << 30, 2049 .copy_num_dw = 7, 2050 .emit_copy_buffer = sdma_v5_2_emit_copy_buffer, 2051 2052 .fill_max_bytes = 1 << 30, /* HW supports 1 << 30, but PAL uses 1 << 22 */ 2053 .fill_num_dw = 5, 2054 .emit_fill_buffer = sdma_v5_2_emit_fill_buffer, 2055 }; 2056 2057 static void sdma_v5_2_set_buffer_funcs(struct amdgpu_device *adev) 2058 { 2059 if (adev->mman.buffer_funcs == NULL) { 2060 adev->mman.buffer_funcs = &sdma_v5_2_buffer_funcs; 2061 adev->mman.buffer_funcs_ring = &adev->sdma.instance[0].ring; 2062 } 2063 } 2064 2065 const struct amdgpu_ip_block_version sdma_v5_2_ip_block = { 2066 .type = AMD_IP_BLOCK_TYPE_SDMA, 2067 .major = 5, 2068 .minor = 2, 2069 .rev = 0, 2070 .funcs = &sdma_v5_2_ip_funcs, 2071 }; 2072