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