1 /* 2 * Copyright 2014 IBM Corp. 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public License 6 * as published by the Free Software Foundation; either version 7 * 2 of the License, or (at your option) any later version. 8 */ 9 10 #include <linux/spinlock.h> 11 #include <linux/sched.h> 12 #include <linux/sched/clock.h> 13 #include <linux/slab.h> 14 #include <linux/mutex.h> 15 #include <linux/mm.h> 16 #include <linux/uaccess.h> 17 #include <linux/delay.h> 18 #include <asm/synch.h> 19 #include <misc/cxl-base.h> 20 21 #include "cxl.h" 22 #include "trace.h" 23 24 static int afu_control(struct cxl_afu *afu, u64 command, u64 clear, 25 u64 result, u64 mask, bool enabled) 26 { 27 u64 AFU_Cntl; 28 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 29 int rc = 0; 30 31 spin_lock(&afu->afu_cntl_lock); 32 pr_devel("AFU command starting: %llx\n", command); 33 34 trace_cxl_afu_ctrl(afu, command); 35 36 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An); 37 cxl_p2n_write(afu, CXL_AFU_Cntl_An, (AFU_Cntl & ~clear) | command); 38 39 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An); 40 while ((AFU_Cntl & mask) != result) { 41 if (time_after_eq(jiffies, timeout)) { 42 dev_warn(&afu->dev, "WARNING: AFU control timed out!\n"); 43 rc = -EBUSY; 44 goto out; 45 } 46 47 if (!cxl_ops->link_ok(afu->adapter, afu)) { 48 afu->enabled = enabled; 49 rc = -EIO; 50 goto out; 51 } 52 53 pr_devel_ratelimited("AFU control... (0x%016llx)\n", 54 AFU_Cntl | command); 55 cpu_relax(); 56 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An); 57 } 58 59 if (AFU_Cntl & CXL_AFU_Cntl_An_RA) { 60 /* 61 * Workaround for a bug in the XSL used in the Mellanox CX4 62 * that fails to clear the RA bit after an AFU reset, 63 * preventing subsequent AFU resets from working. 64 */ 65 cxl_p2n_write(afu, CXL_AFU_Cntl_An, AFU_Cntl & ~CXL_AFU_Cntl_An_RA); 66 } 67 68 pr_devel("AFU command complete: %llx\n", command); 69 afu->enabled = enabled; 70 out: 71 trace_cxl_afu_ctrl_done(afu, command, rc); 72 spin_unlock(&afu->afu_cntl_lock); 73 74 return rc; 75 } 76 77 static int afu_enable(struct cxl_afu *afu) 78 { 79 pr_devel("AFU enable request\n"); 80 81 return afu_control(afu, CXL_AFU_Cntl_An_E, 0, 82 CXL_AFU_Cntl_An_ES_Enabled, 83 CXL_AFU_Cntl_An_ES_MASK, true); 84 } 85 86 int cxl_afu_disable(struct cxl_afu *afu) 87 { 88 pr_devel("AFU disable request\n"); 89 90 return afu_control(afu, 0, CXL_AFU_Cntl_An_E, 91 CXL_AFU_Cntl_An_ES_Disabled, 92 CXL_AFU_Cntl_An_ES_MASK, false); 93 } 94 95 /* This will disable as well as reset */ 96 static int native_afu_reset(struct cxl_afu *afu) 97 { 98 int rc; 99 u64 serr; 100 101 pr_devel("AFU reset request\n"); 102 103 rc = afu_control(afu, CXL_AFU_Cntl_An_RA, 0, 104 CXL_AFU_Cntl_An_RS_Complete | CXL_AFU_Cntl_An_ES_Disabled, 105 CXL_AFU_Cntl_An_RS_MASK | CXL_AFU_Cntl_An_ES_MASK, 106 false); 107 108 /* Re-enable any masked interrupts */ 109 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An); 110 serr &= ~CXL_PSL_SERR_An_IRQ_MASKS; 111 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr); 112 113 114 return rc; 115 } 116 117 static int native_afu_check_and_enable(struct cxl_afu *afu) 118 { 119 if (!cxl_ops->link_ok(afu->adapter, afu)) { 120 WARN(1, "Refusing to enable afu while link down!\n"); 121 return -EIO; 122 } 123 if (afu->enabled) 124 return 0; 125 return afu_enable(afu); 126 } 127 128 int cxl_psl_purge(struct cxl_afu *afu) 129 { 130 u64 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An); 131 u64 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An); 132 u64 dsisr, dar; 133 u64 start, end; 134 u64 trans_fault = 0x0ULL; 135 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 136 int rc = 0; 137 138 trace_cxl_psl_ctrl(afu, CXL_PSL_SCNTL_An_Pc); 139 140 pr_devel("PSL purge request\n"); 141 142 if (cxl_is_psl8(afu)) 143 trans_fault = CXL_PSL_DSISR_TRANS; 144 if (cxl_is_psl9(afu)) 145 trans_fault = CXL_PSL9_DSISR_An_TF; 146 147 if (!cxl_ops->link_ok(afu->adapter, afu)) { 148 dev_warn(&afu->dev, "PSL Purge called with link down, ignoring\n"); 149 rc = -EIO; 150 goto out; 151 } 152 153 if ((AFU_Cntl & CXL_AFU_Cntl_An_ES_MASK) != CXL_AFU_Cntl_An_ES_Disabled) { 154 WARN(1, "psl_purge request while AFU not disabled!\n"); 155 cxl_afu_disable(afu); 156 } 157 158 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, 159 PSL_CNTL | CXL_PSL_SCNTL_An_Pc); 160 start = local_clock(); 161 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An); 162 while ((PSL_CNTL & CXL_PSL_SCNTL_An_Ps_MASK) 163 == CXL_PSL_SCNTL_An_Ps_Pending) { 164 if (time_after_eq(jiffies, timeout)) { 165 dev_warn(&afu->dev, "WARNING: PSL Purge timed out!\n"); 166 rc = -EBUSY; 167 goto out; 168 } 169 if (!cxl_ops->link_ok(afu->adapter, afu)) { 170 rc = -EIO; 171 goto out; 172 } 173 174 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An); 175 pr_devel_ratelimited("PSL purging... PSL_CNTL: 0x%016llx PSL_DSISR: 0x%016llx\n", 176 PSL_CNTL, dsisr); 177 178 if (dsisr & trans_fault) { 179 dar = cxl_p2n_read(afu, CXL_PSL_DAR_An); 180 dev_notice(&afu->dev, "PSL purge terminating pending translation, DSISR: 0x%016llx, DAR: 0x%016llx\n", 181 dsisr, dar); 182 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE); 183 } else if (dsisr) { 184 dev_notice(&afu->dev, "PSL purge acknowledging pending non-translation fault, DSISR: 0x%016llx\n", 185 dsisr); 186 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A); 187 } else { 188 cpu_relax(); 189 } 190 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An); 191 } 192 end = local_clock(); 193 pr_devel("PSL purged in %lld ns\n", end - start); 194 195 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, 196 PSL_CNTL & ~CXL_PSL_SCNTL_An_Pc); 197 out: 198 trace_cxl_psl_ctrl_done(afu, CXL_PSL_SCNTL_An_Pc, rc); 199 return rc; 200 } 201 202 static int spa_max_procs(int spa_size) 203 { 204 /* 205 * From the CAIA: 206 * end_of_SPA_area = SPA_Base + ((n+4) * 128) + (( ((n*8) + 127) >> 7) * 128) + 255 207 * Most of that junk is really just an overly-complicated way of saying 208 * the last 256 bytes are __aligned(128), so it's really: 209 * end_of_SPA_area = end_of_PSL_queue_area + __aligned(128) 255 210 * and 211 * end_of_PSL_queue_area = SPA_Base + ((n+4) * 128) + (n*8) - 1 212 * so 213 * sizeof(SPA) = ((n+4) * 128) + (n*8) + __aligned(128) 256 214 * Ignore the alignment (which is safe in this case as long as we are 215 * careful with our rounding) and solve for n: 216 */ 217 return ((spa_size / 8) - 96) / 17; 218 } 219 220 static int cxl_alloc_spa(struct cxl_afu *afu, int mode) 221 { 222 unsigned spa_size; 223 224 /* Work out how many pages to allocate */ 225 afu->native->spa_order = -1; 226 do { 227 afu->native->spa_order++; 228 spa_size = (1 << afu->native->spa_order) * PAGE_SIZE; 229 230 if (spa_size > 0x100000) { 231 dev_warn(&afu->dev, "num_of_processes too large for the SPA, limiting to %i (0x%x)\n", 232 afu->native->spa_max_procs, afu->native->spa_size); 233 if (mode != CXL_MODE_DEDICATED) 234 afu->num_procs = afu->native->spa_max_procs; 235 break; 236 } 237 238 afu->native->spa_size = spa_size; 239 afu->native->spa_max_procs = spa_max_procs(afu->native->spa_size); 240 } while (afu->native->spa_max_procs < afu->num_procs); 241 242 if (!(afu->native->spa = (struct cxl_process_element *) 243 __get_free_pages(GFP_KERNEL | __GFP_ZERO, afu->native->spa_order))) { 244 pr_err("cxl_alloc_spa: Unable to allocate scheduled process area\n"); 245 return -ENOMEM; 246 } 247 pr_devel("spa pages: %i afu->spa_max_procs: %i afu->num_procs: %i\n", 248 1<<afu->native->spa_order, afu->native->spa_max_procs, afu->num_procs); 249 250 return 0; 251 } 252 253 static void attach_spa(struct cxl_afu *afu) 254 { 255 u64 spap; 256 257 afu->native->sw_command_status = (__be64 *)((char *)afu->native->spa + 258 ((afu->native->spa_max_procs + 3) * 128)); 259 260 spap = virt_to_phys(afu->native->spa) & CXL_PSL_SPAP_Addr; 261 spap |= ((afu->native->spa_size >> (12 - CXL_PSL_SPAP_Size_Shift)) - 1) & CXL_PSL_SPAP_Size; 262 spap |= CXL_PSL_SPAP_V; 263 pr_devel("cxl: SPA allocated at 0x%p. Max processes: %i, sw_command_status: 0x%p CXL_PSL_SPAP_An=0x%016llx\n", 264 afu->native->spa, afu->native->spa_max_procs, 265 afu->native->sw_command_status, spap); 266 cxl_p1n_write(afu, CXL_PSL_SPAP_An, spap); 267 } 268 269 static inline void detach_spa(struct cxl_afu *afu) 270 { 271 cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0); 272 } 273 274 void cxl_release_spa(struct cxl_afu *afu) 275 { 276 if (afu->native->spa) { 277 free_pages((unsigned long) afu->native->spa, 278 afu->native->spa_order); 279 afu->native->spa = NULL; 280 } 281 } 282 283 /* 284 * Invalidation of all ERAT entries is no longer required by CAIA2. Use 285 * only for debug. 286 */ 287 int cxl_invalidate_all_psl9(struct cxl *adapter) 288 { 289 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 290 u64 ierat; 291 292 pr_devel("CXL adapter - invalidation of all ERAT entries\n"); 293 294 /* Invalidates all ERAT entries for Radix or HPT */ 295 ierat = CXL_XSL9_IERAT_IALL; 296 if (radix_enabled()) 297 ierat |= CXL_XSL9_IERAT_INVR; 298 cxl_p1_write(adapter, CXL_XSL9_IERAT, ierat); 299 300 while (cxl_p1_read(adapter, CXL_XSL9_IERAT) & CXL_XSL9_IERAT_IINPROG) { 301 if (time_after_eq(jiffies, timeout)) { 302 dev_warn(&adapter->dev, 303 "WARNING: CXL adapter invalidation of all ERAT entries timed out!\n"); 304 return -EBUSY; 305 } 306 if (!cxl_ops->link_ok(adapter, NULL)) 307 return -EIO; 308 cpu_relax(); 309 } 310 return 0; 311 } 312 313 int cxl_invalidate_all_psl8(struct cxl *adapter) 314 { 315 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 316 317 pr_devel("CXL adapter wide TLBIA & SLBIA\n"); 318 319 cxl_p1_write(adapter, CXL_PSL_AFUSEL, CXL_PSL_AFUSEL_A); 320 321 cxl_p1_write(adapter, CXL_PSL_TLBIA, CXL_TLB_SLB_IQ_ALL); 322 while (cxl_p1_read(adapter, CXL_PSL_TLBIA) & CXL_TLB_SLB_P) { 323 if (time_after_eq(jiffies, timeout)) { 324 dev_warn(&adapter->dev, "WARNING: CXL adapter wide TLBIA timed out!\n"); 325 return -EBUSY; 326 } 327 if (!cxl_ops->link_ok(adapter, NULL)) 328 return -EIO; 329 cpu_relax(); 330 } 331 332 cxl_p1_write(adapter, CXL_PSL_SLBIA, CXL_TLB_SLB_IQ_ALL); 333 while (cxl_p1_read(adapter, CXL_PSL_SLBIA) & CXL_TLB_SLB_P) { 334 if (time_after_eq(jiffies, timeout)) { 335 dev_warn(&adapter->dev, "WARNING: CXL adapter wide SLBIA timed out!\n"); 336 return -EBUSY; 337 } 338 if (!cxl_ops->link_ok(adapter, NULL)) 339 return -EIO; 340 cpu_relax(); 341 } 342 return 0; 343 } 344 345 int cxl_data_cache_flush(struct cxl *adapter) 346 { 347 u64 reg; 348 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 349 350 pr_devel("Flushing data cache\n"); 351 352 reg = cxl_p1_read(adapter, CXL_PSL_Control); 353 reg |= CXL_PSL_Control_Fr; 354 cxl_p1_write(adapter, CXL_PSL_Control, reg); 355 356 reg = cxl_p1_read(adapter, CXL_PSL_Control); 357 while ((reg & CXL_PSL_Control_Fs_MASK) != CXL_PSL_Control_Fs_Complete) { 358 if (time_after_eq(jiffies, timeout)) { 359 dev_warn(&adapter->dev, "WARNING: cache flush timed out!\n"); 360 return -EBUSY; 361 } 362 363 if (!cxl_ops->link_ok(adapter, NULL)) { 364 dev_warn(&adapter->dev, "WARNING: link down when flushing cache\n"); 365 return -EIO; 366 } 367 cpu_relax(); 368 reg = cxl_p1_read(adapter, CXL_PSL_Control); 369 } 370 371 reg &= ~CXL_PSL_Control_Fr; 372 cxl_p1_write(adapter, CXL_PSL_Control, reg); 373 return 0; 374 } 375 376 static int cxl_write_sstp(struct cxl_afu *afu, u64 sstp0, u64 sstp1) 377 { 378 int rc; 379 380 /* 1. Disable SSTP by writing 0 to SSTP1[V] */ 381 cxl_p2n_write(afu, CXL_SSTP1_An, 0); 382 383 /* 2. Invalidate all SLB entries */ 384 if ((rc = cxl_afu_slbia(afu))) 385 return rc; 386 387 /* 3. Set SSTP0_An */ 388 cxl_p2n_write(afu, CXL_SSTP0_An, sstp0); 389 390 /* 4. Set SSTP1_An */ 391 cxl_p2n_write(afu, CXL_SSTP1_An, sstp1); 392 393 return 0; 394 } 395 396 /* Using per slice version may improve performance here. (ie. SLBIA_An) */ 397 static void slb_invalid(struct cxl_context *ctx) 398 { 399 struct cxl *adapter = ctx->afu->adapter; 400 u64 slbia; 401 402 WARN_ON(!mutex_is_locked(&ctx->afu->native->spa_mutex)); 403 404 cxl_p1_write(adapter, CXL_PSL_LBISEL, 405 ((u64)be32_to_cpu(ctx->elem->common.pid) << 32) | 406 be32_to_cpu(ctx->elem->lpid)); 407 cxl_p1_write(adapter, CXL_PSL_SLBIA, CXL_TLB_SLB_IQ_LPIDPID); 408 409 while (1) { 410 if (!cxl_ops->link_ok(adapter, NULL)) 411 break; 412 slbia = cxl_p1_read(adapter, CXL_PSL_SLBIA); 413 if (!(slbia & CXL_TLB_SLB_P)) 414 break; 415 cpu_relax(); 416 } 417 } 418 419 static int do_process_element_cmd(struct cxl_context *ctx, 420 u64 cmd, u64 pe_state) 421 { 422 u64 state; 423 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 424 int rc = 0; 425 426 trace_cxl_llcmd(ctx, cmd); 427 428 WARN_ON(!ctx->afu->enabled); 429 430 ctx->elem->software_state = cpu_to_be32(pe_state); 431 smp_wmb(); 432 *(ctx->afu->native->sw_command_status) = cpu_to_be64(cmd | 0 | ctx->pe); 433 smp_mb(); 434 cxl_p1n_write(ctx->afu, CXL_PSL_LLCMD_An, cmd | ctx->pe); 435 while (1) { 436 if (time_after_eq(jiffies, timeout)) { 437 dev_warn(&ctx->afu->dev, "WARNING: Process Element Command timed out!\n"); 438 rc = -EBUSY; 439 goto out; 440 } 441 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) { 442 dev_warn(&ctx->afu->dev, "WARNING: Device link down, aborting Process Element Command!\n"); 443 rc = -EIO; 444 goto out; 445 } 446 state = be64_to_cpup(ctx->afu->native->sw_command_status); 447 if (state == ~0ULL) { 448 pr_err("cxl: Error adding process element to AFU\n"); 449 rc = -1; 450 goto out; 451 } 452 if ((state & (CXL_SPA_SW_CMD_MASK | CXL_SPA_SW_STATE_MASK | CXL_SPA_SW_LINK_MASK)) == 453 (cmd | (cmd >> 16) | ctx->pe)) 454 break; 455 /* 456 * The command won't finish in the PSL if there are 457 * outstanding DSIs. Hence we need to yield here in 458 * case there are outstanding DSIs that we need to 459 * service. Tuning possiblity: we could wait for a 460 * while before sched 461 */ 462 schedule(); 463 464 } 465 out: 466 trace_cxl_llcmd_done(ctx, cmd, rc); 467 return rc; 468 } 469 470 static int add_process_element(struct cxl_context *ctx) 471 { 472 int rc = 0; 473 474 mutex_lock(&ctx->afu->native->spa_mutex); 475 pr_devel("%s Adding pe: %i started\n", __func__, ctx->pe); 476 if (!(rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_ADD, CXL_PE_SOFTWARE_STATE_V))) 477 ctx->pe_inserted = true; 478 pr_devel("%s Adding pe: %i finished\n", __func__, ctx->pe); 479 mutex_unlock(&ctx->afu->native->spa_mutex); 480 return rc; 481 } 482 483 static int terminate_process_element(struct cxl_context *ctx) 484 { 485 int rc = 0; 486 487 /* fast path terminate if it's already invalid */ 488 if (!(ctx->elem->software_state & cpu_to_be32(CXL_PE_SOFTWARE_STATE_V))) 489 return rc; 490 491 mutex_lock(&ctx->afu->native->spa_mutex); 492 pr_devel("%s Terminate pe: %i started\n", __func__, ctx->pe); 493 /* We could be asked to terminate when the hw is down. That 494 * should always succeed: it's not running if the hw has gone 495 * away and is being reset. 496 */ 497 if (cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 498 rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_TERMINATE, 499 CXL_PE_SOFTWARE_STATE_V | CXL_PE_SOFTWARE_STATE_T); 500 ctx->elem->software_state = 0; /* Remove Valid bit */ 501 pr_devel("%s Terminate pe: %i finished\n", __func__, ctx->pe); 502 mutex_unlock(&ctx->afu->native->spa_mutex); 503 return rc; 504 } 505 506 static int remove_process_element(struct cxl_context *ctx) 507 { 508 int rc = 0; 509 510 mutex_lock(&ctx->afu->native->spa_mutex); 511 pr_devel("%s Remove pe: %i started\n", __func__, ctx->pe); 512 513 /* We could be asked to remove when the hw is down. Again, if 514 * the hw is down, the PE is gone, so we succeed. 515 */ 516 if (cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 517 rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_REMOVE, 0); 518 519 if (!rc) 520 ctx->pe_inserted = false; 521 if (cxl_is_power8()) 522 slb_invalid(ctx); 523 pr_devel("%s Remove pe: %i finished\n", __func__, ctx->pe); 524 mutex_unlock(&ctx->afu->native->spa_mutex); 525 526 return rc; 527 } 528 529 void cxl_assign_psn_space(struct cxl_context *ctx) 530 { 531 if (!ctx->afu->pp_size || ctx->master) { 532 ctx->psn_phys = ctx->afu->psn_phys; 533 ctx->psn_size = ctx->afu->adapter->ps_size; 534 } else { 535 ctx->psn_phys = ctx->afu->psn_phys + 536 (ctx->afu->native->pp_offset + ctx->afu->pp_size * ctx->pe); 537 ctx->psn_size = ctx->afu->pp_size; 538 } 539 } 540 541 static int activate_afu_directed(struct cxl_afu *afu) 542 { 543 int rc; 544 545 dev_info(&afu->dev, "Activating AFU directed mode\n"); 546 547 afu->num_procs = afu->max_procs_virtualised; 548 if (afu->native->spa == NULL) { 549 if (cxl_alloc_spa(afu, CXL_MODE_DIRECTED)) 550 return -ENOMEM; 551 } 552 attach_spa(afu); 553 554 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_AFU); 555 if (cxl_is_power8()) 556 cxl_p1n_write(afu, CXL_PSL_AMOR_An, 0xFFFFFFFFFFFFFFFFULL); 557 cxl_p1n_write(afu, CXL_PSL_ID_An, CXL_PSL_ID_An_F | CXL_PSL_ID_An_L); 558 559 afu->current_mode = CXL_MODE_DIRECTED; 560 561 if ((rc = cxl_chardev_m_afu_add(afu))) 562 return rc; 563 564 if ((rc = cxl_sysfs_afu_m_add(afu))) 565 goto err; 566 567 if ((rc = cxl_chardev_s_afu_add(afu))) 568 goto err1; 569 570 return 0; 571 err1: 572 cxl_sysfs_afu_m_remove(afu); 573 err: 574 cxl_chardev_afu_remove(afu); 575 return rc; 576 } 577 578 #ifdef CONFIG_CPU_LITTLE_ENDIAN 579 #define set_endian(sr) ((sr) |= CXL_PSL_SR_An_LE) 580 #else 581 #define set_endian(sr) ((sr) &= ~(CXL_PSL_SR_An_LE)) 582 #endif 583 584 static u64 calculate_sr(struct cxl_context *ctx) 585 { 586 u64 sr = 0; 587 588 set_endian(sr); 589 if (ctx->master) 590 sr |= CXL_PSL_SR_An_MP; 591 if (mfspr(SPRN_LPCR) & LPCR_TC) 592 sr |= CXL_PSL_SR_An_TC; 593 if (ctx->kernel) { 594 if (!ctx->real_mode) 595 sr |= CXL_PSL_SR_An_R; 596 sr |= (mfmsr() & MSR_SF) | CXL_PSL_SR_An_HV; 597 } else { 598 sr |= CXL_PSL_SR_An_PR | CXL_PSL_SR_An_R; 599 if (radix_enabled()) 600 sr |= CXL_PSL_SR_An_HV; 601 else 602 sr &= ~(CXL_PSL_SR_An_HV); 603 if (!test_tsk_thread_flag(current, TIF_32BIT)) 604 sr |= CXL_PSL_SR_An_SF; 605 } 606 if (cxl_is_psl9(ctx->afu)) { 607 if (radix_enabled()) 608 sr |= CXL_PSL_SR_An_XLAT_ror; 609 else 610 sr |= CXL_PSL_SR_An_XLAT_hpt; 611 } 612 return sr; 613 } 614 615 static void update_ivtes_directed(struct cxl_context *ctx) 616 { 617 bool need_update = (ctx->status == STARTED); 618 int r; 619 620 if (need_update) { 621 WARN_ON(terminate_process_element(ctx)); 622 WARN_ON(remove_process_element(ctx)); 623 } 624 625 for (r = 0; r < CXL_IRQ_RANGES; r++) { 626 ctx->elem->ivte_offsets[r] = cpu_to_be16(ctx->irqs.offset[r]); 627 ctx->elem->ivte_ranges[r] = cpu_to_be16(ctx->irqs.range[r]); 628 } 629 630 /* 631 * Theoretically we could use the update llcmd, instead of a 632 * terminate/remove/add (or if an atomic update was required we could 633 * do a suspend/update/resume), however it seems there might be issues 634 * with the update llcmd on some cards (including those using an XSL on 635 * an ASIC) so for now it's safest to go with the commands that are 636 * known to work. In the future if we come across a situation where the 637 * card may be performing transactions using the same PE while we are 638 * doing this update we might need to revisit this. 639 */ 640 if (need_update) 641 WARN_ON(add_process_element(ctx)); 642 } 643 644 static int process_element_entry_psl9(struct cxl_context *ctx, u64 wed, u64 amr) 645 { 646 u32 pid; 647 648 cxl_assign_psn_space(ctx); 649 650 ctx->elem->ctxtime = 0; /* disable */ 651 ctx->elem->lpid = cpu_to_be32(mfspr(SPRN_LPID)); 652 ctx->elem->haurp = 0; /* disable */ 653 654 if (ctx->kernel) 655 pid = 0; 656 else { 657 if (ctx->mm == NULL) { 658 pr_devel("%s: unable to get mm for pe=%d pid=%i\n", 659 __func__, ctx->pe, pid_nr(ctx->pid)); 660 return -EINVAL; 661 } 662 pid = ctx->mm->context.id; 663 } 664 665 ctx->elem->common.tid = 0; 666 ctx->elem->common.pid = cpu_to_be32(pid); 667 668 ctx->elem->sr = cpu_to_be64(calculate_sr(ctx)); 669 670 ctx->elem->common.csrp = 0; /* disable */ 671 672 cxl_prefault(ctx, wed); 673 674 /* 675 * Ensure we have the multiplexed PSL interrupt set up to take faults 676 * for kernel contexts that may not have allocated any AFU IRQs at all: 677 */ 678 if (ctx->irqs.range[0] == 0) { 679 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq; 680 ctx->irqs.range[0] = 1; 681 } 682 683 ctx->elem->common.amr = cpu_to_be64(amr); 684 ctx->elem->common.wed = cpu_to_be64(wed); 685 686 return 0; 687 } 688 689 int cxl_attach_afu_directed_psl9(struct cxl_context *ctx, u64 wed, u64 amr) 690 { 691 int result; 692 693 /* fill the process element entry */ 694 result = process_element_entry_psl9(ctx, wed, amr); 695 if (result) 696 return result; 697 698 update_ivtes_directed(ctx); 699 700 /* first guy needs to enable */ 701 result = cxl_ops->afu_check_and_enable(ctx->afu); 702 if (result) 703 return result; 704 705 return add_process_element(ctx); 706 } 707 708 int cxl_attach_afu_directed_psl8(struct cxl_context *ctx, u64 wed, u64 amr) 709 { 710 u32 pid; 711 int result; 712 713 cxl_assign_psn_space(ctx); 714 715 ctx->elem->ctxtime = 0; /* disable */ 716 ctx->elem->lpid = cpu_to_be32(mfspr(SPRN_LPID)); 717 ctx->elem->haurp = 0; /* disable */ 718 ctx->elem->u.sdr = cpu_to_be64(mfspr(SPRN_SDR1)); 719 720 pid = current->pid; 721 if (ctx->kernel) 722 pid = 0; 723 ctx->elem->common.tid = 0; 724 ctx->elem->common.pid = cpu_to_be32(pid); 725 726 ctx->elem->sr = cpu_to_be64(calculate_sr(ctx)); 727 728 ctx->elem->common.csrp = 0; /* disable */ 729 ctx->elem->common.u.psl8.aurp0 = 0; /* disable */ 730 ctx->elem->common.u.psl8.aurp1 = 0; /* disable */ 731 732 cxl_prefault(ctx, wed); 733 734 ctx->elem->common.u.psl8.sstp0 = cpu_to_be64(ctx->sstp0); 735 ctx->elem->common.u.psl8.sstp1 = cpu_to_be64(ctx->sstp1); 736 737 /* 738 * Ensure we have the multiplexed PSL interrupt set up to take faults 739 * for kernel contexts that may not have allocated any AFU IRQs at all: 740 */ 741 if (ctx->irqs.range[0] == 0) { 742 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq; 743 ctx->irqs.range[0] = 1; 744 } 745 746 update_ivtes_directed(ctx); 747 748 ctx->elem->common.amr = cpu_to_be64(amr); 749 ctx->elem->common.wed = cpu_to_be64(wed); 750 751 /* first guy needs to enable */ 752 if ((result = cxl_ops->afu_check_and_enable(ctx->afu))) 753 return result; 754 755 return add_process_element(ctx); 756 } 757 758 static int deactivate_afu_directed(struct cxl_afu *afu) 759 { 760 dev_info(&afu->dev, "Deactivating AFU directed mode\n"); 761 762 afu->current_mode = 0; 763 afu->num_procs = 0; 764 765 cxl_sysfs_afu_m_remove(afu); 766 cxl_chardev_afu_remove(afu); 767 768 /* 769 * The CAIA section 2.2.1 indicates that the procedure for starting and 770 * stopping an AFU in AFU directed mode is AFU specific, which is not 771 * ideal since this code is generic and with one exception has no 772 * knowledge of the AFU. This is in contrast to the procedure for 773 * disabling a dedicated process AFU, which is documented to just 774 * require a reset. The architecture does indicate that both an AFU 775 * reset and an AFU disable should result in the AFU being disabled and 776 * we do both followed by a PSL purge for safety. 777 * 778 * Notably we used to have some issues with the disable sequence on PSL 779 * cards, which is why we ended up using this heavy weight procedure in 780 * the first place, however a bug was discovered that had rendered the 781 * disable operation ineffective, so it is conceivable that was the 782 * sole explanation for those difficulties. Careful regression testing 783 * is recommended if anyone attempts to remove or reorder these 784 * operations. 785 * 786 * The XSL on the Mellanox CX4 behaves a little differently from the 787 * PSL based cards and will time out an AFU reset if the AFU is still 788 * enabled. That card is special in that we do have a means to identify 789 * it from this code, so in that case we skip the reset and just use a 790 * disable/purge to avoid the timeout and corresponding noise in the 791 * kernel log. 792 */ 793 if (afu->adapter->native->sl_ops->needs_reset_before_disable) 794 cxl_ops->afu_reset(afu); 795 cxl_afu_disable(afu); 796 cxl_psl_purge(afu); 797 798 return 0; 799 } 800 801 int cxl_activate_dedicated_process_psl9(struct cxl_afu *afu) 802 { 803 dev_info(&afu->dev, "Activating dedicated process mode\n"); 804 805 /* 806 * If XSL is set to dedicated mode (Set in PSL_SCNTL reg), the 807 * XSL and AFU are programmed to work with a single context. 808 * The context information should be configured in the SPA area 809 * index 0 (so PSL_SPAP must be configured before enabling the 810 * AFU). 811 */ 812 afu->num_procs = 1; 813 if (afu->native->spa == NULL) { 814 if (cxl_alloc_spa(afu, CXL_MODE_DEDICATED)) 815 return -ENOMEM; 816 } 817 attach_spa(afu); 818 819 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_Process); 820 cxl_p1n_write(afu, CXL_PSL_ID_An, CXL_PSL_ID_An_F | CXL_PSL_ID_An_L); 821 822 afu->current_mode = CXL_MODE_DEDICATED; 823 824 return cxl_chardev_d_afu_add(afu); 825 } 826 827 int cxl_activate_dedicated_process_psl8(struct cxl_afu *afu) 828 { 829 dev_info(&afu->dev, "Activating dedicated process mode\n"); 830 831 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_Process); 832 833 cxl_p1n_write(afu, CXL_PSL_CtxTime_An, 0); /* disable */ 834 cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0); /* disable */ 835 cxl_p1n_write(afu, CXL_PSL_AMOR_An, 0xFFFFFFFFFFFFFFFFULL); 836 cxl_p1n_write(afu, CXL_PSL_LPID_An, mfspr(SPRN_LPID)); 837 cxl_p1n_write(afu, CXL_HAURP_An, 0); /* disable */ 838 cxl_p1n_write(afu, CXL_PSL_SDR_An, mfspr(SPRN_SDR1)); 839 840 cxl_p2n_write(afu, CXL_CSRP_An, 0); /* disable */ 841 cxl_p2n_write(afu, CXL_AURP0_An, 0); /* disable */ 842 cxl_p2n_write(afu, CXL_AURP1_An, 0); /* disable */ 843 844 afu->current_mode = CXL_MODE_DEDICATED; 845 afu->num_procs = 1; 846 847 return cxl_chardev_d_afu_add(afu); 848 } 849 850 void cxl_update_dedicated_ivtes_psl9(struct cxl_context *ctx) 851 { 852 int r; 853 854 for (r = 0; r < CXL_IRQ_RANGES; r++) { 855 ctx->elem->ivte_offsets[r] = cpu_to_be16(ctx->irqs.offset[r]); 856 ctx->elem->ivte_ranges[r] = cpu_to_be16(ctx->irqs.range[r]); 857 } 858 } 859 860 void cxl_update_dedicated_ivtes_psl8(struct cxl_context *ctx) 861 { 862 struct cxl_afu *afu = ctx->afu; 863 864 cxl_p1n_write(afu, CXL_PSL_IVTE_Offset_An, 865 (((u64)ctx->irqs.offset[0] & 0xffff) << 48) | 866 (((u64)ctx->irqs.offset[1] & 0xffff) << 32) | 867 (((u64)ctx->irqs.offset[2] & 0xffff) << 16) | 868 ((u64)ctx->irqs.offset[3] & 0xffff)); 869 cxl_p1n_write(afu, CXL_PSL_IVTE_Limit_An, (u64) 870 (((u64)ctx->irqs.range[0] & 0xffff) << 48) | 871 (((u64)ctx->irqs.range[1] & 0xffff) << 32) | 872 (((u64)ctx->irqs.range[2] & 0xffff) << 16) | 873 ((u64)ctx->irqs.range[3] & 0xffff)); 874 } 875 876 int cxl_attach_dedicated_process_psl9(struct cxl_context *ctx, u64 wed, u64 amr) 877 { 878 struct cxl_afu *afu = ctx->afu; 879 int result; 880 881 /* fill the process element entry */ 882 result = process_element_entry_psl9(ctx, wed, amr); 883 if (result) 884 return result; 885 886 if (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes) 887 afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx); 888 889 result = cxl_ops->afu_reset(afu); 890 if (result) 891 return result; 892 893 return afu_enable(afu); 894 } 895 896 int cxl_attach_dedicated_process_psl8(struct cxl_context *ctx, u64 wed, u64 amr) 897 { 898 struct cxl_afu *afu = ctx->afu; 899 u64 pid; 900 int rc; 901 902 pid = (u64)current->pid << 32; 903 if (ctx->kernel) 904 pid = 0; 905 cxl_p2n_write(afu, CXL_PSL_PID_TID_An, pid); 906 907 cxl_p1n_write(afu, CXL_PSL_SR_An, calculate_sr(ctx)); 908 909 if ((rc = cxl_write_sstp(afu, ctx->sstp0, ctx->sstp1))) 910 return rc; 911 912 cxl_prefault(ctx, wed); 913 914 if (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes) 915 afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx); 916 917 cxl_p2n_write(afu, CXL_PSL_AMR_An, amr); 918 919 /* master only context for dedicated */ 920 cxl_assign_psn_space(ctx); 921 922 if ((rc = cxl_ops->afu_reset(afu))) 923 return rc; 924 925 cxl_p2n_write(afu, CXL_PSL_WED_An, wed); 926 927 return afu_enable(afu); 928 } 929 930 static int deactivate_dedicated_process(struct cxl_afu *afu) 931 { 932 dev_info(&afu->dev, "Deactivating dedicated process mode\n"); 933 934 afu->current_mode = 0; 935 afu->num_procs = 0; 936 937 cxl_chardev_afu_remove(afu); 938 939 return 0; 940 } 941 942 static int native_afu_deactivate_mode(struct cxl_afu *afu, int mode) 943 { 944 if (mode == CXL_MODE_DIRECTED) 945 return deactivate_afu_directed(afu); 946 if (mode == CXL_MODE_DEDICATED) 947 return deactivate_dedicated_process(afu); 948 return 0; 949 } 950 951 static int native_afu_activate_mode(struct cxl_afu *afu, int mode) 952 { 953 if (!mode) 954 return 0; 955 if (!(mode & afu->modes_supported)) 956 return -EINVAL; 957 958 if (!cxl_ops->link_ok(afu->adapter, afu)) { 959 WARN(1, "Device link is down, refusing to activate!\n"); 960 return -EIO; 961 } 962 963 if (mode == CXL_MODE_DIRECTED) 964 return activate_afu_directed(afu); 965 if ((mode == CXL_MODE_DEDICATED) && 966 (afu->adapter->native->sl_ops->activate_dedicated_process)) 967 return afu->adapter->native->sl_ops->activate_dedicated_process(afu); 968 969 return -EINVAL; 970 } 971 972 static int native_attach_process(struct cxl_context *ctx, bool kernel, 973 u64 wed, u64 amr) 974 { 975 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) { 976 WARN(1, "Device link is down, refusing to attach process!\n"); 977 return -EIO; 978 } 979 980 ctx->kernel = kernel; 981 if ((ctx->afu->current_mode == CXL_MODE_DIRECTED) && 982 (ctx->afu->adapter->native->sl_ops->attach_afu_directed)) 983 return ctx->afu->adapter->native->sl_ops->attach_afu_directed(ctx, wed, amr); 984 985 if ((ctx->afu->current_mode == CXL_MODE_DEDICATED) && 986 (ctx->afu->adapter->native->sl_ops->attach_dedicated_process)) 987 return ctx->afu->adapter->native->sl_ops->attach_dedicated_process(ctx, wed, amr); 988 989 return -EINVAL; 990 } 991 992 static inline int detach_process_native_dedicated(struct cxl_context *ctx) 993 { 994 /* 995 * The CAIA section 2.1.1 indicates that we need to do an AFU reset to 996 * stop the AFU in dedicated mode (we therefore do not make that 997 * optional like we do in the afu directed path). It does not indicate 998 * that we need to do an explicit disable (which should occur 999 * implicitly as part of the reset) or purge, but we do these as well 1000 * to be on the safe side. 1001 * 1002 * Notably we used to have some issues with the disable sequence 1003 * (before the sequence was spelled out in the architecture) which is 1004 * why we were so heavy weight in the first place, however a bug was 1005 * discovered that had rendered the disable operation ineffective, so 1006 * it is conceivable that was the sole explanation for those 1007 * difficulties. Point is, we should be careful and do some regression 1008 * testing if we ever attempt to remove any part of this procedure. 1009 */ 1010 cxl_ops->afu_reset(ctx->afu); 1011 cxl_afu_disable(ctx->afu); 1012 cxl_psl_purge(ctx->afu); 1013 return 0; 1014 } 1015 1016 static void native_update_ivtes(struct cxl_context *ctx) 1017 { 1018 if (ctx->afu->current_mode == CXL_MODE_DIRECTED) 1019 return update_ivtes_directed(ctx); 1020 if ((ctx->afu->current_mode == CXL_MODE_DEDICATED) && 1021 (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes)) 1022 return ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx); 1023 WARN(1, "native_update_ivtes: Bad mode\n"); 1024 } 1025 1026 static inline int detach_process_native_afu_directed(struct cxl_context *ctx) 1027 { 1028 if (!ctx->pe_inserted) 1029 return 0; 1030 if (terminate_process_element(ctx)) 1031 return -1; 1032 if (remove_process_element(ctx)) 1033 return -1; 1034 1035 return 0; 1036 } 1037 1038 static int native_detach_process(struct cxl_context *ctx) 1039 { 1040 trace_cxl_detach(ctx); 1041 1042 if (ctx->afu->current_mode == CXL_MODE_DEDICATED) 1043 return detach_process_native_dedicated(ctx); 1044 1045 return detach_process_native_afu_directed(ctx); 1046 } 1047 1048 static int native_get_irq_info(struct cxl_afu *afu, struct cxl_irq_info *info) 1049 { 1050 /* If the adapter has gone away, we can't get any meaningful 1051 * information. 1052 */ 1053 if (!cxl_ops->link_ok(afu->adapter, afu)) 1054 return -EIO; 1055 1056 info->dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An); 1057 info->dar = cxl_p2n_read(afu, CXL_PSL_DAR_An); 1058 if (cxl_is_power8()) 1059 info->dsr = cxl_p2n_read(afu, CXL_PSL_DSR_An); 1060 info->afu_err = cxl_p2n_read(afu, CXL_AFU_ERR_An); 1061 info->errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An); 1062 info->proc_handle = 0; 1063 1064 return 0; 1065 } 1066 1067 void cxl_native_irq_dump_regs_psl9(struct cxl_context *ctx) 1068 { 1069 u64 fir1, fir2, serr; 1070 1071 fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL9_FIR1); 1072 fir2 = cxl_p1_read(ctx->afu->adapter, CXL_PSL9_FIR2); 1073 1074 dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1); 1075 dev_crit(&ctx->afu->dev, "PSL_FIR2: 0x%016llx\n", fir2); 1076 if (ctx->afu->adapter->native->sl_ops->register_serr_irq) { 1077 serr = cxl_p1n_read(ctx->afu, CXL_PSL_SERR_An); 1078 cxl_afu_decode_psl_serr(ctx->afu, serr); 1079 } 1080 } 1081 1082 void cxl_native_irq_dump_regs_psl8(struct cxl_context *ctx) 1083 { 1084 u64 fir1, fir2, fir_slice, serr, afu_debug; 1085 1086 fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR1); 1087 fir2 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR2); 1088 fir_slice = cxl_p1n_read(ctx->afu, CXL_PSL_FIR_SLICE_An); 1089 afu_debug = cxl_p1n_read(ctx->afu, CXL_AFU_DEBUG_An); 1090 1091 dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1); 1092 dev_crit(&ctx->afu->dev, "PSL_FIR2: 0x%016llx\n", fir2); 1093 if (ctx->afu->adapter->native->sl_ops->register_serr_irq) { 1094 serr = cxl_p1n_read(ctx->afu, CXL_PSL_SERR_An); 1095 cxl_afu_decode_psl_serr(ctx->afu, serr); 1096 } 1097 dev_crit(&ctx->afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice); 1098 dev_crit(&ctx->afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug); 1099 } 1100 1101 static irqreturn_t native_handle_psl_slice_error(struct cxl_context *ctx, 1102 u64 dsisr, u64 errstat) 1103 { 1104 1105 dev_crit(&ctx->afu->dev, "PSL ERROR STATUS: 0x%016llx\n", errstat); 1106 1107 if (ctx->afu->adapter->native->sl_ops->psl_irq_dump_registers) 1108 ctx->afu->adapter->native->sl_ops->psl_irq_dump_registers(ctx); 1109 1110 if (ctx->afu->adapter->native->sl_ops->debugfs_stop_trace) { 1111 dev_crit(&ctx->afu->dev, "STOPPING CXL TRACE\n"); 1112 ctx->afu->adapter->native->sl_ops->debugfs_stop_trace(ctx->afu->adapter); 1113 } 1114 1115 return cxl_ops->ack_irq(ctx, 0, errstat); 1116 } 1117 1118 static bool cxl_is_translation_fault(struct cxl_afu *afu, u64 dsisr) 1119 { 1120 if ((cxl_is_psl8(afu)) && (dsisr & CXL_PSL_DSISR_TRANS)) 1121 return true; 1122 1123 if ((cxl_is_psl9(afu)) && (dsisr & CXL_PSL9_DSISR_An_TF)) 1124 return true; 1125 1126 return false; 1127 } 1128 1129 irqreturn_t cxl_fail_irq_psl(struct cxl_afu *afu, struct cxl_irq_info *irq_info) 1130 { 1131 if (cxl_is_translation_fault(afu, irq_info->dsisr)) 1132 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE); 1133 else 1134 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A); 1135 1136 return IRQ_HANDLED; 1137 } 1138 1139 static irqreturn_t native_irq_multiplexed(int irq, void *data) 1140 { 1141 struct cxl_afu *afu = data; 1142 struct cxl_context *ctx; 1143 struct cxl_irq_info irq_info; 1144 u64 phreg = cxl_p2n_read(afu, CXL_PSL_PEHandle_An); 1145 int ph, ret = IRQ_HANDLED, res; 1146 1147 /* check if eeh kicked in while the interrupt was in flight */ 1148 if (unlikely(phreg == ~0ULL)) { 1149 dev_warn(&afu->dev, 1150 "Ignoring slice interrupt(%d) due to fenced card", 1151 irq); 1152 return IRQ_HANDLED; 1153 } 1154 /* Mask the pe-handle from register value */ 1155 ph = phreg & 0xffff; 1156 if ((res = native_get_irq_info(afu, &irq_info))) { 1157 WARN(1, "Unable to get CXL IRQ Info: %i\n", res); 1158 if (afu->adapter->native->sl_ops->fail_irq) 1159 return afu->adapter->native->sl_ops->fail_irq(afu, &irq_info); 1160 return ret; 1161 } 1162 1163 rcu_read_lock(); 1164 ctx = idr_find(&afu->contexts_idr, ph); 1165 if (ctx) { 1166 if (afu->adapter->native->sl_ops->handle_interrupt) 1167 ret = afu->adapter->native->sl_ops->handle_interrupt(irq, ctx, &irq_info); 1168 rcu_read_unlock(); 1169 return ret; 1170 } 1171 rcu_read_unlock(); 1172 1173 WARN(1, "Unable to demultiplex CXL PSL IRQ for PE %i DSISR %016llx DAR" 1174 " %016llx\n(Possible AFU HW issue - was a term/remove acked" 1175 " with outstanding transactions?)\n", ph, irq_info.dsisr, 1176 irq_info.dar); 1177 if (afu->adapter->native->sl_ops->fail_irq) 1178 ret = afu->adapter->native->sl_ops->fail_irq(afu, &irq_info); 1179 return ret; 1180 } 1181 1182 static void native_irq_wait(struct cxl_context *ctx) 1183 { 1184 u64 dsisr; 1185 int timeout = 1000; 1186 int ph; 1187 1188 /* 1189 * Wait until no further interrupts are presented by the PSL 1190 * for this context. 1191 */ 1192 while (timeout--) { 1193 ph = cxl_p2n_read(ctx->afu, CXL_PSL_PEHandle_An) & 0xffff; 1194 if (ph != ctx->pe) 1195 return; 1196 dsisr = cxl_p2n_read(ctx->afu, CXL_PSL_DSISR_An); 1197 if (cxl_is_psl8(ctx->afu) && 1198 ((dsisr & CXL_PSL_DSISR_PENDING) == 0)) 1199 return; 1200 if (cxl_is_psl9(ctx->afu) && 1201 ((dsisr & CXL_PSL9_DSISR_PENDING) == 0)) 1202 return; 1203 /* 1204 * We are waiting for the workqueue to process our 1205 * irq, so need to let that run here. 1206 */ 1207 msleep(1); 1208 } 1209 1210 dev_warn(&ctx->afu->dev, "WARNING: waiting on DSI for PE %i" 1211 " DSISR %016llx!\n", ph, dsisr); 1212 return; 1213 } 1214 1215 static irqreturn_t native_slice_irq_err(int irq, void *data) 1216 { 1217 struct cxl_afu *afu = data; 1218 u64 errstat, serr, afu_error, dsisr; 1219 u64 fir_slice, afu_debug, irq_mask; 1220 1221 /* 1222 * slice err interrupt is only used with full PSL (no XSL) 1223 */ 1224 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An); 1225 errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An); 1226 afu_error = cxl_p2n_read(afu, CXL_AFU_ERR_An); 1227 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An); 1228 cxl_afu_decode_psl_serr(afu, serr); 1229 1230 if (cxl_is_power8()) { 1231 fir_slice = cxl_p1n_read(afu, CXL_PSL_FIR_SLICE_An); 1232 afu_debug = cxl_p1n_read(afu, CXL_AFU_DEBUG_An); 1233 dev_crit(&afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice); 1234 dev_crit(&afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug); 1235 } 1236 dev_crit(&afu->dev, "CXL_PSL_ErrStat_An: 0x%016llx\n", errstat); 1237 dev_crit(&afu->dev, "AFU_ERR_An: 0x%.16llx\n", afu_error); 1238 dev_crit(&afu->dev, "PSL_DSISR_An: 0x%.16llx\n", dsisr); 1239 1240 /* mask off the IRQ so it won't retrigger until the AFU is reset */ 1241 irq_mask = (serr & CXL_PSL_SERR_An_IRQS) >> 32; 1242 serr |= irq_mask; 1243 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr); 1244 dev_info(&afu->dev, "Further such interrupts will be masked until the AFU is reset\n"); 1245 1246 return IRQ_HANDLED; 1247 } 1248 1249 void cxl_native_err_irq_dump_regs(struct cxl *adapter) 1250 { 1251 u64 fir1, fir2; 1252 1253 fir1 = cxl_p1_read(adapter, CXL_PSL_FIR1); 1254 fir2 = cxl_p1_read(adapter, CXL_PSL_FIR2); 1255 1256 dev_crit(&adapter->dev, "PSL_FIR1: 0x%016llx\nPSL_FIR2: 0x%016llx\n", fir1, fir2); 1257 } 1258 1259 static irqreturn_t native_irq_err(int irq, void *data) 1260 { 1261 struct cxl *adapter = data; 1262 u64 err_ivte; 1263 1264 WARN(1, "CXL ERROR interrupt %i\n", irq); 1265 1266 err_ivte = cxl_p1_read(adapter, CXL_PSL_ErrIVTE); 1267 dev_crit(&adapter->dev, "PSL_ErrIVTE: 0x%016llx\n", err_ivte); 1268 1269 if (adapter->native->sl_ops->debugfs_stop_trace) { 1270 dev_crit(&adapter->dev, "STOPPING CXL TRACE\n"); 1271 adapter->native->sl_ops->debugfs_stop_trace(adapter); 1272 } 1273 1274 if (adapter->native->sl_ops->err_irq_dump_registers) 1275 adapter->native->sl_ops->err_irq_dump_registers(adapter); 1276 1277 return IRQ_HANDLED; 1278 } 1279 1280 int cxl_native_register_psl_err_irq(struct cxl *adapter) 1281 { 1282 int rc; 1283 1284 adapter->irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err", 1285 dev_name(&adapter->dev)); 1286 if (!adapter->irq_name) 1287 return -ENOMEM; 1288 1289 if ((rc = cxl_register_one_irq(adapter, native_irq_err, adapter, 1290 &adapter->native->err_hwirq, 1291 &adapter->native->err_virq, 1292 adapter->irq_name))) { 1293 kfree(adapter->irq_name); 1294 adapter->irq_name = NULL; 1295 return rc; 1296 } 1297 1298 cxl_p1_write(adapter, CXL_PSL_ErrIVTE, adapter->native->err_hwirq & 0xffff); 1299 1300 return 0; 1301 } 1302 1303 void cxl_native_release_psl_err_irq(struct cxl *adapter) 1304 { 1305 if (adapter->native->err_virq != irq_find_mapping(NULL, adapter->native->err_hwirq)) 1306 return; 1307 1308 cxl_p1_write(adapter, CXL_PSL_ErrIVTE, 0x0000000000000000); 1309 cxl_unmap_irq(adapter->native->err_virq, adapter); 1310 cxl_ops->release_one_irq(adapter, adapter->native->err_hwirq); 1311 kfree(adapter->irq_name); 1312 } 1313 1314 int cxl_native_register_serr_irq(struct cxl_afu *afu) 1315 { 1316 u64 serr; 1317 int rc; 1318 1319 afu->err_irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err", 1320 dev_name(&afu->dev)); 1321 if (!afu->err_irq_name) 1322 return -ENOMEM; 1323 1324 if ((rc = cxl_register_one_irq(afu->adapter, native_slice_irq_err, afu, 1325 &afu->serr_hwirq, 1326 &afu->serr_virq, afu->err_irq_name))) { 1327 kfree(afu->err_irq_name); 1328 afu->err_irq_name = NULL; 1329 return rc; 1330 } 1331 1332 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An); 1333 if (cxl_is_power8()) 1334 serr = (serr & 0x00ffffffffff0000ULL) | (afu->serr_hwirq & 0xffff); 1335 if (cxl_is_power9()) { 1336 /* 1337 * By default, all errors are masked. So don't set all masks. 1338 * Slice errors will be transfered. 1339 */ 1340 serr = (serr & ~0xff0000007fffffffULL) | (afu->serr_hwirq & 0xffff); 1341 } 1342 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr); 1343 1344 return 0; 1345 } 1346 1347 void cxl_native_release_serr_irq(struct cxl_afu *afu) 1348 { 1349 if (afu->serr_virq != irq_find_mapping(NULL, afu->serr_hwirq)) 1350 return; 1351 1352 cxl_p1n_write(afu, CXL_PSL_SERR_An, 0x0000000000000000); 1353 cxl_unmap_irq(afu->serr_virq, afu); 1354 cxl_ops->release_one_irq(afu->adapter, afu->serr_hwirq); 1355 kfree(afu->err_irq_name); 1356 } 1357 1358 int cxl_native_register_psl_irq(struct cxl_afu *afu) 1359 { 1360 int rc; 1361 1362 afu->psl_irq_name = kasprintf(GFP_KERNEL, "cxl-%s", 1363 dev_name(&afu->dev)); 1364 if (!afu->psl_irq_name) 1365 return -ENOMEM; 1366 1367 if ((rc = cxl_register_one_irq(afu->adapter, native_irq_multiplexed, 1368 afu, &afu->native->psl_hwirq, &afu->native->psl_virq, 1369 afu->psl_irq_name))) { 1370 kfree(afu->psl_irq_name); 1371 afu->psl_irq_name = NULL; 1372 } 1373 return rc; 1374 } 1375 1376 void cxl_native_release_psl_irq(struct cxl_afu *afu) 1377 { 1378 if (afu->native->psl_virq != irq_find_mapping(NULL, afu->native->psl_hwirq)) 1379 return; 1380 1381 cxl_unmap_irq(afu->native->psl_virq, afu); 1382 cxl_ops->release_one_irq(afu->adapter, afu->native->psl_hwirq); 1383 kfree(afu->psl_irq_name); 1384 } 1385 1386 static void recover_psl_err(struct cxl_afu *afu, u64 errstat) 1387 { 1388 u64 dsisr; 1389 1390 pr_devel("RECOVERING FROM PSL ERROR... (0x%016llx)\n", errstat); 1391 1392 /* Clear PSL_DSISR[PE] */ 1393 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An); 1394 cxl_p2n_write(afu, CXL_PSL_DSISR_An, dsisr & ~CXL_PSL_DSISR_An_PE); 1395 1396 /* Write 1s to clear error status bits */ 1397 cxl_p2n_write(afu, CXL_PSL_ErrStat_An, errstat); 1398 } 1399 1400 static int native_ack_irq(struct cxl_context *ctx, u64 tfc, u64 psl_reset_mask) 1401 { 1402 trace_cxl_psl_irq_ack(ctx, tfc); 1403 if (tfc) 1404 cxl_p2n_write(ctx->afu, CXL_PSL_TFC_An, tfc); 1405 if (psl_reset_mask) 1406 recover_psl_err(ctx->afu, psl_reset_mask); 1407 1408 return 0; 1409 } 1410 1411 int cxl_check_error(struct cxl_afu *afu) 1412 { 1413 return (cxl_p1n_read(afu, CXL_PSL_SCNTL_An) == ~0ULL); 1414 } 1415 1416 static bool native_support_attributes(const char *attr_name, 1417 enum cxl_attrs type) 1418 { 1419 return true; 1420 } 1421 1422 static int native_afu_cr_read64(struct cxl_afu *afu, int cr, u64 off, u64 *out) 1423 { 1424 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu))) 1425 return -EIO; 1426 if (unlikely(off >= afu->crs_len)) 1427 return -ERANGE; 1428 *out = in_le64(afu->native->afu_desc_mmio + afu->crs_offset + 1429 (cr * afu->crs_len) + off); 1430 return 0; 1431 } 1432 1433 static int native_afu_cr_read32(struct cxl_afu *afu, int cr, u64 off, u32 *out) 1434 { 1435 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu))) 1436 return -EIO; 1437 if (unlikely(off >= afu->crs_len)) 1438 return -ERANGE; 1439 *out = in_le32(afu->native->afu_desc_mmio + afu->crs_offset + 1440 (cr * afu->crs_len) + off); 1441 return 0; 1442 } 1443 1444 static int native_afu_cr_read16(struct cxl_afu *afu, int cr, u64 off, u16 *out) 1445 { 1446 u64 aligned_off = off & ~0x3L; 1447 u32 val; 1448 int rc; 1449 1450 rc = native_afu_cr_read32(afu, cr, aligned_off, &val); 1451 if (!rc) 1452 *out = (val >> ((off & 0x3) * 8)) & 0xffff; 1453 return rc; 1454 } 1455 1456 static int native_afu_cr_read8(struct cxl_afu *afu, int cr, u64 off, u8 *out) 1457 { 1458 u64 aligned_off = off & ~0x3L; 1459 u32 val; 1460 int rc; 1461 1462 rc = native_afu_cr_read32(afu, cr, aligned_off, &val); 1463 if (!rc) 1464 *out = (val >> ((off & 0x3) * 8)) & 0xff; 1465 return rc; 1466 } 1467 1468 static int native_afu_cr_write32(struct cxl_afu *afu, int cr, u64 off, u32 in) 1469 { 1470 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu))) 1471 return -EIO; 1472 if (unlikely(off >= afu->crs_len)) 1473 return -ERANGE; 1474 out_le32(afu->native->afu_desc_mmio + afu->crs_offset + 1475 (cr * afu->crs_len) + off, in); 1476 return 0; 1477 } 1478 1479 static int native_afu_cr_write16(struct cxl_afu *afu, int cr, u64 off, u16 in) 1480 { 1481 u64 aligned_off = off & ~0x3L; 1482 u32 val32, mask, shift; 1483 int rc; 1484 1485 rc = native_afu_cr_read32(afu, cr, aligned_off, &val32); 1486 if (rc) 1487 return rc; 1488 shift = (off & 0x3) * 8; 1489 WARN_ON(shift == 24); 1490 mask = 0xffff << shift; 1491 val32 = (val32 & ~mask) | (in << shift); 1492 1493 rc = native_afu_cr_write32(afu, cr, aligned_off, val32); 1494 return rc; 1495 } 1496 1497 static int native_afu_cr_write8(struct cxl_afu *afu, int cr, u64 off, u8 in) 1498 { 1499 u64 aligned_off = off & ~0x3L; 1500 u32 val32, mask, shift; 1501 int rc; 1502 1503 rc = native_afu_cr_read32(afu, cr, aligned_off, &val32); 1504 if (rc) 1505 return rc; 1506 shift = (off & 0x3) * 8; 1507 mask = 0xff << shift; 1508 val32 = (val32 & ~mask) | (in << shift); 1509 1510 rc = native_afu_cr_write32(afu, cr, aligned_off, val32); 1511 return rc; 1512 } 1513 1514 const struct cxl_backend_ops cxl_native_ops = { 1515 .module = THIS_MODULE, 1516 .adapter_reset = cxl_pci_reset, 1517 .alloc_one_irq = cxl_pci_alloc_one_irq, 1518 .release_one_irq = cxl_pci_release_one_irq, 1519 .alloc_irq_ranges = cxl_pci_alloc_irq_ranges, 1520 .release_irq_ranges = cxl_pci_release_irq_ranges, 1521 .setup_irq = cxl_pci_setup_irq, 1522 .handle_psl_slice_error = native_handle_psl_slice_error, 1523 .psl_interrupt = NULL, 1524 .ack_irq = native_ack_irq, 1525 .irq_wait = native_irq_wait, 1526 .attach_process = native_attach_process, 1527 .detach_process = native_detach_process, 1528 .update_ivtes = native_update_ivtes, 1529 .support_attributes = native_support_attributes, 1530 .link_ok = cxl_adapter_link_ok, 1531 .release_afu = cxl_pci_release_afu, 1532 .afu_read_err_buffer = cxl_pci_afu_read_err_buffer, 1533 .afu_check_and_enable = native_afu_check_and_enable, 1534 .afu_activate_mode = native_afu_activate_mode, 1535 .afu_deactivate_mode = native_afu_deactivate_mode, 1536 .afu_reset = native_afu_reset, 1537 .afu_cr_read8 = native_afu_cr_read8, 1538 .afu_cr_read16 = native_afu_cr_read16, 1539 .afu_cr_read32 = native_afu_cr_read32, 1540 .afu_cr_read64 = native_afu_cr_read64, 1541 .afu_cr_write8 = native_afu_cr_write8, 1542 .afu_cr_write16 = native_afu_cr_write16, 1543 .afu_cr_write32 = native_afu_cr_write32, 1544 .read_adapter_vpd = cxl_pci_read_adapter_vpd, 1545 }; 1546