1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2016-2022 Advanced Micro Devices, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice shall be included in 13 * all copies or substantial portions of the Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 21 * OTHER DEALINGS IN THE SOFTWARE. 22 */ 23 24 #include "kfd_priv.h" 25 #include "kfd_events.h" 26 #include "kfd_debug.h" 27 #include "soc15_int.h" 28 #include "kfd_device_queue_manager.h" 29 #include "kfd_smi_events.h" 30 #include "amdgpu_ras.h" 31 32 /* 33 * GFX9 SQ Interrupts 34 * 35 * There are 3 encoding types of interrupts sourced from SQ sent as a 44-bit 36 * packet to the Interrupt Handler: 37 * Auto - Generated by the SQG (various cmd overflows, timestamps etc) 38 * Wave - Generated by S_SENDMSG through a shader program 39 * Error - HW generated errors (Illegal instructions, Memviols, EDC etc) 40 * 41 * The 44-bit packet is mapped as {context_id1[7:0],context_id0[31:0]} plus 42 * 4-bits for VMID (SOC15_VMID_FROM_IH_ENTRY) as such: 43 * 44 * - context_id0[27:26] 45 * Encoding type (0 = Auto, 1 = Wave, 2 = Error) 46 * 47 * - context_id0[13] 48 * PRIV bit indicates that Wave S_SEND or error occurred within trap 49 * 50 * - {context_id1[7:0],context_id0[31:28],context_id0[11:0]} 51 * 24-bit data with the following layout per encoding type: 52 * Auto - only context_id0[8:0] is used, which reports various interrupts 53 * generated by SQG. The rest is 0. 54 * Wave - user data sent from m0 via S_SENDMSG 55 * Error - Error type (context_id1[7:4]), Error Details (rest of bits) 56 * 57 * The other context_id bits show coordinates (SE/SH/CU/SIMD/WAVE) for wave 58 * S_SENDMSG and Errors. These are 0 for Auto. 59 */ 60 61 enum SQ_INTERRUPT_WORD_ENCODING { 62 SQ_INTERRUPT_WORD_ENCODING_AUTO = 0x0, 63 SQ_INTERRUPT_WORD_ENCODING_INST, 64 SQ_INTERRUPT_WORD_ENCODING_ERROR, 65 }; 66 67 enum SQ_INTERRUPT_ERROR_TYPE { 68 SQ_INTERRUPT_ERROR_TYPE_EDC_FUE = 0x0, 69 SQ_INTERRUPT_ERROR_TYPE_ILLEGAL_INST, 70 SQ_INTERRUPT_ERROR_TYPE_MEMVIOL, 71 SQ_INTERRUPT_ERROR_TYPE_EDC_FED, 72 }; 73 74 /* SQ_INTERRUPT_WORD_AUTO_CTXID */ 75 #define SQ_INTERRUPT_WORD_AUTO_CTXID__THREAD_TRACE__SHIFT 0 76 #define SQ_INTERRUPT_WORD_AUTO_CTXID__WLT__SHIFT 1 77 #define SQ_INTERRUPT_WORD_AUTO_CTXID__THREAD_TRACE_BUF_FULL__SHIFT 2 78 #define SQ_INTERRUPT_WORD_AUTO_CTXID__REG_TIMESTAMP__SHIFT 3 79 #define SQ_INTERRUPT_WORD_AUTO_CTXID__CMD_TIMESTAMP__SHIFT 4 80 #define SQ_INTERRUPT_WORD_AUTO_CTXID__HOST_CMD_OVERFLOW__SHIFT 5 81 #define SQ_INTERRUPT_WORD_AUTO_CTXID__HOST_REG_OVERFLOW__SHIFT 6 82 #define SQ_INTERRUPT_WORD_AUTO_CTXID__IMMED_OVERFLOW__SHIFT 7 83 #define SQ_INTERRUPT_WORD_AUTO_CTXID__THREAD_TRACE_UTC_ERROR__SHIFT 8 84 #define SQ_INTERRUPT_WORD_AUTO_CTXID__SE_ID__SHIFT 24 85 #define SQ_INTERRUPT_WORD_AUTO_CTXID__ENCODING__SHIFT 26 86 87 #define SQ_INTERRUPT_WORD_AUTO_CTXID__THREAD_TRACE_MASK 0x00000001 88 #define SQ_INTERRUPT_WORD_AUTO_CTXID__WLT_MASK 0x00000002 89 #define SQ_INTERRUPT_WORD_AUTO_CTXID__THREAD_TRACE_BUF_FULL_MASK 0x00000004 90 #define SQ_INTERRUPT_WORD_AUTO_CTXID__REG_TIMESTAMP_MASK 0x00000008 91 #define SQ_INTERRUPT_WORD_AUTO_CTXID__CMD_TIMESTAMP_MASK 0x00000010 92 #define SQ_INTERRUPT_WORD_AUTO_CTXID__HOST_CMD_OVERFLOW_MASK 0x00000020 93 #define SQ_INTERRUPT_WORD_AUTO_CTXID__HOST_REG_OVERFLOW_MASK 0x00000040 94 #define SQ_INTERRUPT_WORD_AUTO_CTXID__IMMED_OVERFLOW_MASK 0x00000080 95 #define SQ_INTERRUPT_WORD_AUTO_CTXID__THREAD_TRACE_UTC_ERROR_MASK 0x00000100 96 #define SQ_INTERRUPT_WORD_AUTO_CTXID__SE_ID_MASK 0x03000000 97 #define SQ_INTERRUPT_WORD_AUTO_CTXID__ENCODING_MASK 0x0c000000 98 99 /* SQ_INTERRUPT_WORD_WAVE_CTXID */ 100 #define SQ_INTERRUPT_WORD_WAVE_CTXID__DATA__SHIFT 0 101 #define SQ_INTERRUPT_WORD_WAVE_CTXID__SH_ID__SHIFT 12 102 #define SQ_INTERRUPT_WORD_WAVE_CTXID__PRIV__SHIFT 13 103 #define SQ_INTERRUPT_WORD_WAVE_CTXID__WAVE_ID__SHIFT 14 104 #define SQ_INTERRUPT_WORD_WAVE_CTXID__SIMD_ID__SHIFT 18 105 #define SQ_INTERRUPT_WORD_WAVE_CTXID__CU_ID__SHIFT 20 106 #define SQ_INTERRUPT_WORD_WAVE_CTXID__SE_ID__SHIFT 24 107 #define SQ_INTERRUPT_WORD_WAVE_CTXID__ENCODING__SHIFT 26 108 109 #define SQ_INTERRUPT_WORD_WAVE_CTXID__DATA_MASK 0x00000fff 110 #define SQ_INTERRUPT_WORD_WAVE_CTXID__SH_ID_MASK 0x00001000 111 #define SQ_INTERRUPT_WORD_WAVE_CTXID__PRIV_MASK 0x00002000 112 #define SQ_INTERRUPT_WORD_WAVE_CTXID__WAVE_ID_MASK 0x0003c000 113 #define SQ_INTERRUPT_WORD_WAVE_CTXID__SIMD_ID_MASK 0x000c0000 114 #define SQ_INTERRUPT_WORD_WAVE_CTXID__CU_ID_MASK 0x00f00000 115 #define SQ_INTERRUPT_WORD_WAVE_CTXID__SE_ID_MASK 0x03000000 116 #define SQ_INTERRUPT_WORD_WAVE_CTXID__ENCODING_MASK 0x0c000000 117 118 /* GFX9 SQ interrupt 24-bit data from context_id<0,1> */ 119 #define KFD_CONTEXT_ID_GET_SQ_INT_DATA(ctx0, ctx1) \ 120 ((ctx0 & 0xfff) | ((ctx0 >> 16) & 0xf000) | ((ctx1 << 16) & 0xff0000)) 121 122 #define KFD_SQ_INT_DATA__ERR_TYPE_MASK 0xF00000 123 #define KFD_SQ_INT_DATA__ERR_TYPE__SHIFT 20 124 125 /* 126 * The debugger will send user data(m0) with PRIV=1 to indicate it requires 127 * notification from the KFD with the following queue id (DOORBELL_ID) and 128 * trap code (TRAP_CODE). 129 */ 130 #define KFD_INT_DATA_DEBUG_DOORBELL_MASK 0x0003ff 131 #define KFD_INT_DATA_DEBUG_TRAP_CODE_SHIFT 10 132 #define KFD_INT_DATA_DEBUG_TRAP_CODE_MASK 0x07fc00 133 #define KFD_DEBUG_DOORBELL_ID(sq_int_data) ((sq_int_data) & \ 134 KFD_INT_DATA_DEBUG_DOORBELL_MASK) 135 #define KFD_DEBUG_TRAP_CODE(sq_int_data) (((sq_int_data) & \ 136 KFD_INT_DATA_DEBUG_TRAP_CODE_MASK) \ 137 >> KFD_INT_DATA_DEBUG_TRAP_CODE_SHIFT) 138 #define KFD_DEBUG_CP_BAD_OP_ECODE_MASK 0x3fffc00 139 #define KFD_DEBUG_CP_BAD_OP_ECODE_SHIFT 10 140 #define KFD_DEBUG_CP_BAD_OP_ECODE(ctxid0) (((ctxid0) & \ 141 KFD_DEBUG_CP_BAD_OP_ECODE_MASK) \ 142 >> KFD_DEBUG_CP_BAD_OP_ECODE_SHIFT) 143 144 static void event_interrupt_poison_consumption_v9(struct kfd_node *dev, 145 uint16_t pasid, uint16_t client_id) 146 { 147 enum amdgpu_ras_block block = 0; 148 uint32_t reset = 0; 149 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); 150 enum ras_event_type type = RAS_EVENT_TYPE_POISON_CONSUMPTION; 151 u64 event_id; 152 int old_poison, ret; 153 154 if (!p) 155 return; 156 157 /* all queues of a process will be unmapped in one time */ 158 old_poison = atomic_cmpxchg(&p->poison, 0, 1); 159 kfd_unref_process(p); 160 if (old_poison) 161 return; 162 163 switch (client_id) { 164 case SOC15_IH_CLIENTID_SE0SH: 165 case SOC15_IH_CLIENTID_SE1SH: 166 case SOC15_IH_CLIENTID_SE2SH: 167 case SOC15_IH_CLIENTID_SE3SH: 168 case SOC15_IH_CLIENTID_UTCL2: 169 block = AMDGPU_RAS_BLOCK__GFX; 170 if (amdgpu_ip_version(dev->adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) || 171 amdgpu_ip_version(dev->adev, GC_HWIP, 0) == IP_VERSION(9, 4, 4)) 172 reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET; 173 else 174 reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET; 175 break; 176 case SOC15_IH_CLIENTID_VMC: 177 case SOC15_IH_CLIENTID_VMC1: 178 block = AMDGPU_RAS_BLOCK__MMHUB; 179 reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET; 180 break; 181 case SOC15_IH_CLIENTID_SDMA0: 182 case SOC15_IH_CLIENTID_SDMA1: 183 case SOC15_IH_CLIENTID_SDMA2: 184 case SOC15_IH_CLIENTID_SDMA3: 185 case SOC15_IH_CLIENTID_SDMA4: 186 block = AMDGPU_RAS_BLOCK__SDMA; 187 if (amdgpu_ip_version(dev->adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) || 188 amdgpu_ip_version(dev->adev, GC_HWIP, 0) == IP_VERSION(9, 4, 4)) 189 reset = AMDGPU_RAS_GPU_RESET_MODE1_RESET; 190 else 191 reset = AMDGPU_RAS_GPU_RESET_MODE2_RESET; 192 break; 193 default: 194 dev_warn(dev->adev->dev, 195 "client %d does not support poison consumption\n", client_id); 196 return; 197 } 198 199 ret = amdgpu_ras_mark_ras_event(dev->adev, type); 200 if (ret) 201 return; 202 203 kfd_signal_poison_consumed_event(dev, pasid); 204 205 event_id = amdgpu_ras_acquire_event_id(dev->adev, type); 206 207 RAS_EVENT_LOG(dev->adev, event_id, 208 "poison is consumed by client %d, kick off gpu reset flow\n", client_id); 209 210 amdgpu_amdkfd_ras_pasid_poison_consumption_handler(dev->adev, 211 block, pasid, NULL, NULL, reset); 212 } 213 214 static bool context_id_expected(struct kfd_dev *dev) 215 { 216 switch (KFD_GC_VERSION(dev)) { 217 case IP_VERSION(9, 0, 1): 218 return dev->mec_fw_version >= 0x817a; 219 case IP_VERSION(9, 1, 0): 220 case IP_VERSION(9, 2, 1): 221 case IP_VERSION(9, 2, 2): 222 case IP_VERSION(9, 3, 0): 223 case IP_VERSION(9, 4, 0): 224 return dev->mec_fw_version >= 0x17a; 225 default: 226 /* Other GFXv9 and later GPUs always sent valid context IDs 227 * on legitimate events 228 */ 229 return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 1); 230 } 231 } 232 233 static bool event_interrupt_isr_v9(struct kfd_node *dev, 234 const uint32_t *ih_ring_entry, 235 uint32_t *patched_ihre, 236 bool *patched_flag) 237 { 238 uint16_t source_id, client_id, pasid, vmid; 239 const uint32_t *data = ih_ring_entry; 240 241 source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry); 242 client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); 243 244 /* Only handle interrupts from KFD VMIDs */ 245 vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); 246 if (!KFD_IRQ_IS_FENCE(client_id, source_id) && 247 (vmid < dev->vm_info.first_vmid_kfd || 248 vmid > dev->vm_info.last_vmid_kfd)) 249 return false; 250 251 pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry); 252 253 /* Only handle clients we care about */ 254 if (client_id != SOC15_IH_CLIENTID_GRBM_CP && 255 client_id != SOC15_IH_CLIENTID_SDMA0 && 256 client_id != SOC15_IH_CLIENTID_SDMA1 && 257 client_id != SOC15_IH_CLIENTID_SDMA2 && 258 client_id != SOC15_IH_CLIENTID_SDMA3 && 259 client_id != SOC15_IH_CLIENTID_SDMA4 && 260 client_id != SOC15_IH_CLIENTID_SDMA5 && 261 client_id != SOC15_IH_CLIENTID_SDMA6 && 262 client_id != SOC15_IH_CLIENTID_SDMA7 && 263 client_id != SOC15_IH_CLIENTID_VMC && 264 client_id != SOC15_IH_CLIENTID_VMC1 && 265 client_id != SOC15_IH_CLIENTID_UTCL2 && 266 client_id != SOC15_IH_CLIENTID_SE0SH && 267 client_id != SOC15_IH_CLIENTID_SE1SH && 268 client_id != SOC15_IH_CLIENTID_SE2SH && 269 client_id != SOC15_IH_CLIENTID_SE3SH && 270 !KFD_IRQ_IS_FENCE(client_id, source_id)) 271 return false; 272 273 /* This is a known issue for gfx9. Under non HWS, pasid is not set 274 * in the interrupt payload, so we need to find out the pasid on our 275 * own. 276 */ 277 if (!pasid && dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { 278 const uint32_t pasid_mask = 0xffff; 279 280 *patched_flag = true; 281 memcpy(patched_ihre, ih_ring_entry, 282 dev->kfd->device_info.ih_ring_entry_size); 283 284 pasid = dev->dqm->vmid_pasid[vmid]; 285 286 /* Patch the pasid field */ 287 patched_ihre[3] = cpu_to_le32((le32_to_cpu(patched_ihre[3]) 288 & ~pasid_mask) | pasid); 289 } 290 291 pr_debug("client id 0x%x, source id %d, vmid %d, pasid 0x%x. raw data:\n", 292 client_id, source_id, vmid, pasid); 293 pr_debug("%8X, %8X, %8X, %8X, %8X, %8X, %8X, %8X.\n", 294 data[0], data[1], data[2], data[3], 295 data[4], data[5], data[6], data[7]); 296 297 /* If there is no valid PASID, it's likely a bug */ 298 if (WARN_ONCE(pasid == 0, "Bug: No PASID in KFD interrupt")) 299 return false; 300 301 /* Workaround CP firmware sending bogus signals with 0 context_id. 302 * Those can be safely ignored on hardware and firmware versions that 303 * include a valid context_id on legitimate signals. This avoids the 304 * slow path in kfd_signal_event_interrupt that scans all event slots 305 * for signaled events. 306 */ 307 if (source_id == SOC15_INTSRC_CP_END_OF_PIPE) { 308 uint32_t context_id = 309 SOC15_CONTEXT_ID0_FROM_IH_ENTRY(ih_ring_entry); 310 311 if (context_id == 0 && context_id_expected(dev->kfd)) 312 return false; 313 } 314 315 /* Interrupt types we care about: various signals and faults. 316 * They will be forwarded to a work queue (see below). 317 */ 318 return source_id == SOC15_INTSRC_CP_END_OF_PIPE || 319 source_id == SOC15_INTSRC_SDMA_TRAP || 320 source_id == SOC15_INTSRC_SDMA_ECC || 321 source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG || 322 source_id == SOC15_INTSRC_CP_BAD_OPCODE || 323 KFD_IRQ_IS_FENCE(client_id, source_id) || 324 ((client_id == SOC15_IH_CLIENTID_VMC || 325 client_id == SOC15_IH_CLIENTID_VMC1 || 326 client_id == SOC15_IH_CLIENTID_UTCL2) && 327 !amdgpu_no_queue_eviction_on_vm_fault); 328 } 329 330 static void event_interrupt_wq_v9(struct kfd_node *dev, 331 const uint32_t *ih_ring_entry) 332 { 333 uint16_t source_id, client_id, pasid, vmid; 334 uint32_t context_id0, context_id1; 335 uint32_t sq_intr_err, sq_int_data, encoding; 336 337 source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry); 338 client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry); 339 pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry); 340 vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); 341 context_id0 = SOC15_CONTEXT_ID0_FROM_IH_ENTRY(ih_ring_entry); 342 context_id1 = SOC15_CONTEXT_ID1_FROM_IH_ENTRY(ih_ring_entry); 343 344 if (client_id == SOC15_IH_CLIENTID_GRBM_CP || 345 client_id == SOC15_IH_CLIENTID_SE0SH || 346 client_id == SOC15_IH_CLIENTID_SE1SH || 347 client_id == SOC15_IH_CLIENTID_SE2SH || 348 client_id == SOC15_IH_CLIENTID_SE3SH) { 349 if (source_id == SOC15_INTSRC_CP_END_OF_PIPE) 350 kfd_signal_event_interrupt(pasid, context_id0, 32); 351 else if (source_id == SOC15_INTSRC_SQ_INTERRUPT_MSG) { 352 sq_int_data = KFD_CONTEXT_ID_GET_SQ_INT_DATA(context_id0, context_id1); 353 encoding = REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, ENCODING); 354 switch (encoding) { 355 case SQ_INTERRUPT_WORD_ENCODING_AUTO: 356 pr_debug_ratelimited( 357 "sq_intr: auto, se %d, ttrace %d, wlt %d, ttrac_buf_full %d, reg_tms %d, cmd_tms %d, host_cmd_ovf %d, host_reg_ovf %d, immed_ovf %d, ttrace_utc_err %d\n", 358 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, SE_ID), 359 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, THREAD_TRACE), 360 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, WLT), 361 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, THREAD_TRACE_BUF_FULL), 362 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, REG_TIMESTAMP), 363 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, CMD_TIMESTAMP), 364 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, HOST_CMD_OVERFLOW), 365 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, HOST_REG_OVERFLOW), 366 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, IMMED_OVERFLOW), 367 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_AUTO_CTXID, THREAD_TRACE_UTC_ERROR)); 368 break; 369 case SQ_INTERRUPT_WORD_ENCODING_INST: 370 pr_debug_ratelimited("sq_intr: inst, se %d, data 0x%x, sh %d, priv %d, wave_id %d, simd_id %d, cu_id %d, intr_data 0x%x\n", 371 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, SE_ID), 372 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, DATA), 373 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, SH_ID), 374 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, PRIV), 375 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, WAVE_ID), 376 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, SIMD_ID), 377 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, CU_ID), 378 sq_int_data); 379 if (context_id0 & SQ_INTERRUPT_WORD_WAVE_CTXID__PRIV_MASK) { 380 if (kfd_set_dbg_ev_from_interrupt(dev, pasid, 381 KFD_DEBUG_DOORBELL_ID(sq_int_data), 382 KFD_DEBUG_TRAP_CODE(sq_int_data), 383 NULL, 0)) 384 return; 385 } 386 break; 387 case SQ_INTERRUPT_WORD_ENCODING_ERROR: 388 sq_intr_err = REG_GET_FIELD(sq_int_data, KFD_SQ_INT_DATA, ERR_TYPE); 389 pr_warn_ratelimited("sq_intr: error, se %d, data 0x%x, sh %d, priv %d, wave_id %d, simd_id %d, cu_id %d, err_type %d\n", 390 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, SE_ID), 391 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, DATA), 392 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, SH_ID), 393 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, PRIV), 394 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, WAVE_ID), 395 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, SIMD_ID), 396 REG_GET_FIELD(context_id0, SQ_INTERRUPT_WORD_WAVE_CTXID, CU_ID), 397 sq_intr_err); 398 if (sq_intr_err != SQ_INTERRUPT_ERROR_TYPE_ILLEGAL_INST && 399 sq_intr_err != SQ_INTERRUPT_ERROR_TYPE_MEMVIOL) { 400 event_interrupt_poison_consumption_v9(dev, pasid, client_id); 401 return; 402 } 403 break; 404 default: 405 break; 406 } 407 kfd_signal_event_interrupt(pasid, sq_int_data, 24); 408 } else if (source_id == SOC15_INTSRC_CP_BAD_OPCODE && 409 KFD_DBG_EC_TYPE_IS_PACKET(KFD_DEBUG_CP_BAD_OP_ECODE(context_id0))) { 410 kfd_set_dbg_ev_from_interrupt(dev, pasid, 411 KFD_DEBUG_DOORBELL_ID(context_id0), 412 KFD_EC_MASK(KFD_DEBUG_CP_BAD_OP_ECODE(context_id0)), 413 NULL, 0); 414 } 415 } else if (client_id == SOC15_IH_CLIENTID_SDMA0 || 416 client_id == SOC15_IH_CLIENTID_SDMA1 || 417 client_id == SOC15_IH_CLIENTID_SDMA2 || 418 client_id == SOC15_IH_CLIENTID_SDMA3 || 419 client_id == SOC15_IH_CLIENTID_SDMA4 || 420 client_id == SOC15_IH_CLIENTID_SDMA5 || 421 client_id == SOC15_IH_CLIENTID_SDMA6 || 422 client_id == SOC15_IH_CLIENTID_SDMA7) { 423 if (source_id == SOC15_INTSRC_SDMA_TRAP) { 424 kfd_signal_event_interrupt(pasid, context_id0 & 0xfffffff, 28); 425 } else if (source_id == SOC15_INTSRC_SDMA_ECC) { 426 event_interrupt_poison_consumption_v9(dev, pasid, client_id); 427 return; 428 } 429 } else if (client_id == SOC15_IH_CLIENTID_VMC || 430 client_id == SOC15_IH_CLIENTID_VMC1 || 431 client_id == SOC15_IH_CLIENTID_UTCL2) { 432 struct kfd_vm_fault_info info = {0}; 433 uint16_t ring_id = SOC15_RING_ID_FROM_IH_ENTRY(ih_ring_entry); 434 uint32_t node_id = SOC15_NODEID_FROM_IH_ENTRY(ih_ring_entry); 435 uint32_t vmid_type = SOC15_VMID_TYPE_FROM_IH_ENTRY(ih_ring_entry); 436 int hub_inst = 0; 437 struct kfd_hsa_memory_exception_data exception_data; 438 439 /* gfxhub */ 440 if (!vmid_type && dev->adev->gfx.funcs->ih_node_to_logical_xcc) { 441 hub_inst = dev->adev->gfx.funcs->ih_node_to_logical_xcc(dev->adev, 442 node_id); 443 if (hub_inst < 0) 444 hub_inst = 0; 445 } 446 447 /* mmhub */ 448 if (vmid_type && client_id == SOC15_IH_CLIENTID_VMC) 449 hub_inst = node_id / 4; 450 451 if (amdgpu_amdkfd_ras_query_utcl2_poison_status(dev->adev, 452 hub_inst, vmid_type)) { 453 event_interrupt_poison_consumption_v9(dev, pasid, client_id); 454 return; 455 } 456 457 info.vmid = vmid; 458 info.mc_id = client_id; 459 info.page_addr = ih_ring_entry[4] | 460 (uint64_t)(ih_ring_entry[5] & 0xf) << 32; 461 info.prot_valid = ring_id & 0x08; 462 info.prot_read = ring_id & 0x10; 463 info.prot_write = ring_id & 0x20; 464 465 memset(&exception_data, 0, sizeof(exception_data)); 466 exception_data.gpu_id = dev->id; 467 exception_data.va = (info.page_addr) << PAGE_SHIFT; 468 exception_data.failure.NotPresent = info.prot_valid ? 1 : 0; 469 exception_data.failure.NoExecute = info.prot_exec ? 1 : 0; 470 exception_data.failure.ReadOnly = info.prot_write ? 1 : 0; 471 exception_data.failure.imprecise = 0; 472 473 kfd_set_dbg_ev_from_interrupt(dev, 474 pasid, 475 -1, 476 KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION), 477 &exception_data, 478 sizeof(exception_data)); 479 kfd_smi_event_update_vmfault(dev, pasid); 480 } else if (KFD_IRQ_IS_FENCE(client_id, source_id)) { 481 kfd_process_close_interrupt_drain(pasid); 482 } 483 } 484 485 static bool event_interrupt_isr_v9_4_3(struct kfd_node *node, 486 const uint32_t *ih_ring_entry, 487 uint32_t *patched_ihre, 488 bool *patched_flag) 489 { 490 uint16_t node_id, vmid; 491 492 /* 493 * For GFX 9.4.3, process the interrupt if: 494 * - NodeID field in IH entry matches the corresponding bit 495 * set in interrupt_bitmap Bits 0-15. 496 * OR 497 * - If partition mode is CPX and interrupt came from 498 * Node_id 0,4,8,12, then check if the Bit (16 + client id) 499 * is set in interrupt bitmap Bits 16-31. 500 */ 501 node_id = SOC15_NODEID_FROM_IH_ENTRY(ih_ring_entry); 502 vmid = SOC15_VMID_FROM_IH_ENTRY(ih_ring_entry); 503 if (kfd_irq_is_from_node(node, node_id, vmid)) 504 return event_interrupt_isr_v9(node, ih_ring_entry, 505 patched_ihre, patched_flag); 506 return false; 507 } 508 509 const struct kfd_event_interrupt_class event_interrupt_class_v9 = { 510 .interrupt_isr = event_interrupt_isr_v9, 511 .interrupt_wq = event_interrupt_wq_v9, 512 }; 513 514 const struct kfd_event_interrupt_class event_interrupt_class_v9_4_3 = { 515 .interrupt_isr = event_interrupt_isr_v9_4_3, 516 .interrupt_wq = event_interrupt_wq_v9, 517 }; 518