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 /* 109 * Re-enable any masked interrupts when the AFU is not 110 * activated to avoid side effects after attaching a process 111 * in dedicated mode. 112 */ 113 if (afu->current_mode == 0) { 114 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An); 115 serr &= ~CXL_PSL_SERR_An_IRQ_MASKS; 116 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr); 117 } 118 119 return rc; 120 } 121 122 static int native_afu_check_and_enable(struct cxl_afu *afu) 123 { 124 if (!cxl_ops->link_ok(afu->adapter, afu)) { 125 WARN(1, "Refusing to enable afu while link down!\n"); 126 return -EIO; 127 } 128 if (afu->enabled) 129 return 0; 130 return afu_enable(afu); 131 } 132 133 int cxl_psl_purge(struct cxl_afu *afu) 134 { 135 u64 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An); 136 u64 AFU_Cntl = cxl_p2n_read(afu, CXL_AFU_Cntl_An); 137 u64 dsisr, dar; 138 u64 start, end; 139 u64 trans_fault = 0x0ULL; 140 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 141 int rc = 0; 142 143 trace_cxl_psl_ctrl(afu, CXL_PSL_SCNTL_An_Pc); 144 145 pr_devel("PSL purge request\n"); 146 147 if (cxl_is_power8()) 148 trans_fault = CXL_PSL_DSISR_TRANS; 149 if (cxl_is_power9()) 150 trans_fault = CXL_PSL9_DSISR_An_TF; 151 152 if (!cxl_ops->link_ok(afu->adapter, afu)) { 153 dev_warn(&afu->dev, "PSL Purge called with link down, ignoring\n"); 154 rc = -EIO; 155 goto out; 156 } 157 158 if ((AFU_Cntl & CXL_AFU_Cntl_An_ES_MASK) != CXL_AFU_Cntl_An_ES_Disabled) { 159 WARN(1, "psl_purge request while AFU not disabled!\n"); 160 cxl_afu_disable(afu); 161 } 162 163 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, 164 PSL_CNTL | CXL_PSL_SCNTL_An_Pc); 165 start = local_clock(); 166 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An); 167 while ((PSL_CNTL & CXL_PSL_SCNTL_An_Ps_MASK) 168 == CXL_PSL_SCNTL_An_Ps_Pending) { 169 if (time_after_eq(jiffies, timeout)) { 170 dev_warn(&afu->dev, "WARNING: PSL Purge timed out!\n"); 171 rc = -EBUSY; 172 goto out; 173 } 174 if (!cxl_ops->link_ok(afu->adapter, afu)) { 175 rc = -EIO; 176 goto out; 177 } 178 179 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An); 180 pr_devel_ratelimited("PSL purging... PSL_CNTL: 0x%016llx PSL_DSISR: 0x%016llx\n", 181 PSL_CNTL, dsisr); 182 183 if (dsisr & trans_fault) { 184 dar = cxl_p2n_read(afu, CXL_PSL_DAR_An); 185 dev_notice(&afu->dev, "PSL purge terminating pending translation, DSISR: 0x%016llx, DAR: 0x%016llx\n", 186 dsisr, dar); 187 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE); 188 } else if (dsisr) { 189 dev_notice(&afu->dev, "PSL purge acknowledging pending non-translation fault, DSISR: 0x%016llx\n", 190 dsisr); 191 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A); 192 } else { 193 cpu_relax(); 194 } 195 PSL_CNTL = cxl_p1n_read(afu, CXL_PSL_SCNTL_An); 196 } 197 end = local_clock(); 198 pr_devel("PSL purged in %lld ns\n", end - start); 199 200 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, 201 PSL_CNTL & ~CXL_PSL_SCNTL_An_Pc); 202 out: 203 trace_cxl_psl_ctrl_done(afu, CXL_PSL_SCNTL_An_Pc, rc); 204 return rc; 205 } 206 207 static int spa_max_procs(int spa_size) 208 { 209 /* 210 * From the CAIA: 211 * end_of_SPA_area = SPA_Base + ((n+4) * 128) + (( ((n*8) + 127) >> 7) * 128) + 255 212 * Most of that junk is really just an overly-complicated way of saying 213 * the last 256 bytes are __aligned(128), so it's really: 214 * end_of_SPA_area = end_of_PSL_queue_area + __aligned(128) 255 215 * and 216 * end_of_PSL_queue_area = SPA_Base + ((n+4) * 128) + (n*8) - 1 217 * so 218 * sizeof(SPA) = ((n+4) * 128) + (n*8) + __aligned(128) 256 219 * Ignore the alignment (which is safe in this case as long as we are 220 * careful with our rounding) and solve for n: 221 */ 222 return ((spa_size / 8) - 96) / 17; 223 } 224 225 static int cxl_alloc_spa(struct cxl_afu *afu, int mode) 226 { 227 unsigned spa_size; 228 229 /* Work out how many pages to allocate */ 230 afu->native->spa_order = -1; 231 do { 232 afu->native->spa_order++; 233 spa_size = (1 << afu->native->spa_order) * PAGE_SIZE; 234 235 if (spa_size > 0x100000) { 236 dev_warn(&afu->dev, "num_of_processes too large for the SPA, limiting to %i (0x%x)\n", 237 afu->native->spa_max_procs, afu->native->spa_size); 238 if (mode != CXL_MODE_DEDICATED) 239 afu->num_procs = afu->native->spa_max_procs; 240 break; 241 } 242 243 afu->native->spa_size = spa_size; 244 afu->native->spa_max_procs = spa_max_procs(afu->native->spa_size); 245 } while (afu->native->spa_max_procs < afu->num_procs); 246 247 if (!(afu->native->spa = (struct cxl_process_element *) 248 __get_free_pages(GFP_KERNEL | __GFP_ZERO, afu->native->spa_order))) { 249 pr_err("cxl_alloc_spa: Unable to allocate scheduled process area\n"); 250 return -ENOMEM; 251 } 252 pr_devel("spa pages: %i afu->spa_max_procs: %i afu->num_procs: %i\n", 253 1<<afu->native->spa_order, afu->native->spa_max_procs, afu->num_procs); 254 255 return 0; 256 } 257 258 static void attach_spa(struct cxl_afu *afu) 259 { 260 u64 spap; 261 262 afu->native->sw_command_status = (__be64 *)((char *)afu->native->spa + 263 ((afu->native->spa_max_procs + 3) * 128)); 264 265 spap = virt_to_phys(afu->native->spa) & CXL_PSL_SPAP_Addr; 266 spap |= ((afu->native->spa_size >> (12 - CXL_PSL_SPAP_Size_Shift)) - 1) & CXL_PSL_SPAP_Size; 267 spap |= CXL_PSL_SPAP_V; 268 pr_devel("cxl: SPA allocated at 0x%p. Max processes: %i, sw_command_status: 0x%p CXL_PSL_SPAP_An=0x%016llx\n", 269 afu->native->spa, afu->native->spa_max_procs, 270 afu->native->sw_command_status, spap); 271 cxl_p1n_write(afu, CXL_PSL_SPAP_An, spap); 272 } 273 274 static inline void detach_spa(struct cxl_afu *afu) 275 { 276 cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0); 277 } 278 279 void cxl_release_spa(struct cxl_afu *afu) 280 { 281 if (afu->native->spa) { 282 free_pages((unsigned long) afu->native->spa, 283 afu->native->spa_order); 284 afu->native->spa = NULL; 285 } 286 } 287 288 /* 289 * Invalidation of all ERAT entries is no longer required by CAIA2. Use 290 * only for debug. 291 */ 292 int cxl_invalidate_all_psl9(struct cxl *adapter) 293 { 294 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 295 u64 ierat; 296 297 pr_devel("CXL adapter - invalidation of all ERAT entries\n"); 298 299 /* Invalidates all ERAT entries for Radix or HPT */ 300 ierat = CXL_XSL9_IERAT_IALL; 301 if (radix_enabled()) 302 ierat |= CXL_XSL9_IERAT_INVR; 303 cxl_p1_write(adapter, CXL_XSL9_IERAT, ierat); 304 305 while (cxl_p1_read(adapter, CXL_XSL9_IERAT) & CXL_XSL9_IERAT_IINPROG) { 306 if (time_after_eq(jiffies, timeout)) { 307 dev_warn(&adapter->dev, 308 "WARNING: CXL adapter invalidation of all ERAT entries timed out!\n"); 309 return -EBUSY; 310 } 311 if (!cxl_ops->link_ok(adapter, NULL)) 312 return -EIO; 313 cpu_relax(); 314 } 315 return 0; 316 } 317 318 int cxl_invalidate_all_psl8(struct cxl *adapter) 319 { 320 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 321 322 pr_devel("CXL adapter wide TLBIA & SLBIA\n"); 323 324 cxl_p1_write(adapter, CXL_PSL_AFUSEL, CXL_PSL_AFUSEL_A); 325 326 cxl_p1_write(adapter, CXL_PSL_TLBIA, CXL_TLB_SLB_IQ_ALL); 327 while (cxl_p1_read(adapter, CXL_PSL_TLBIA) & CXL_TLB_SLB_P) { 328 if (time_after_eq(jiffies, timeout)) { 329 dev_warn(&adapter->dev, "WARNING: CXL adapter wide TLBIA timed out!\n"); 330 return -EBUSY; 331 } 332 if (!cxl_ops->link_ok(adapter, NULL)) 333 return -EIO; 334 cpu_relax(); 335 } 336 337 cxl_p1_write(adapter, CXL_PSL_SLBIA, CXL_TLB_SLB_IQ_ALL); 338 while (cxl_p1_read(adapter, CXL_PSL_SLBIA) & CXL_TLB_SLB_P) { 339 if (time_after_eq(jiffies, timeout)) { 340 dev_warn(&adapter->dev, "WARNING: CXL adapter wide SLBIA timed out!\n"); 341 return -EBUSY; 342 } 343 if (!cxl_ops->link_ok(adapter, NULL)) 344 return -EIO; 345 cpu_relax(); 346 } 347 return 0; 348 } 349 350 int cxl_data_cache_flush(struct cxl *adapter) 351 { 352 u64 reg; 353 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 354 355 pr_devel("Flushing data cache\n"); 356 357 reg = cxl_p1_read(adapter, CXL_PSL_Control); 358 reg |= CXL_PSL_Control_Fr; 359 cxl_p1_write(adapter, CXL_PSL_Control, reg); 360 361 reg = cxl_p1_read(adapter, CXL_PSL_Control); 362 while ((reg & CXL_PSL_Control_Fs_MASK) != CXL_PSL_Control_Fs_Complete) { 363 if (time_after_eq(jiffies, timeout)) { 364 dev_warn(&adapter->dev, "WARNING: cache flush timed out!\n"); 365 return -EBUSY; 366 } 367 368 if (!cxl_ops->link_ok(adapter, NULL)) { 369 dev_warn(&adapter->dev, "WARNING: link down when flushing cache\n"); 370 return -EIO; 371 } 372 cpu_relax(); 373 reg = cxl_p1_read(adapter, CXL_PSL_Control); 374 } 375 376 reg &= ~CXL_PSL_Control_Fr; 377 cxl_p1_write(adapter, CXL_PSL_Control, reg); 378 return 0; 379 } 380 381 static int cxl_write_sstp(struct cxl_afu *afu, u64 sstp0, u64 sstp1) 382 { 383 int rc; 384 385 /* 1. Disable SSTP by writing 0 to SSTP1[V] */ 386 cxl_p2n_write(afu, CXL_SSTP1_An, 0); 387 388 /* 2. Invalidate all SLB entries */ 389 if ((rc = cxl_afu_slbia(afu))) 390 return rc; 391 392 /* 3. Set SSTP0_An */ 393 cxl_p2n_write(afu, CXL_SSTP0_An, sstp0); 394 395 /* 4. Set SSTP1_An */ 396 cxl_p2n_write(afu, CXL_SSTP1_An, sstp1); 397 398 return 0; 399 } 400 401 /* Using per slice version may improve performance here. (ie. SLBIA_An) */ 402 static void slb_invalid(struct cxl_context *ctx) 403 { 404 struct cxl *adapter = ctx->afu->adapter; 405 u64 slbia; 406 407 WARN_ON(!mutex_is_locked(&ctx->afu->native->spa_mutex)); 408 409 cxl_p1_write(adapter, CXL_PSL_LBISEL, 410 ((u64)be32_to_cpu(ctx->elem->common.pid) << 32) | 411 be32_to_cpu(ctx->elem->lpid)); 412 cxl_p1_write(adapter, CXL_PSL_SLBIA, CXL_TLB_SLB_IQ_LPIDPID); 413 414 while (1) { 415 if (!cxl_ops->link_ok(adapter, NULL)) 416 break; 417 slbia = cxl_p1_read(adapter, CXL_PSL_SLBIA); 418 if (!(slbia & CXL_TLB_SLB_P)) 419 break; 420 cpu_relax(); 421 } 422 } 423 424 static int do_process_element_cmd(struct cxl_context *ctx, 425 u64 cmd, u64 pe_state) 426 { 427 u64 state; 428 unsigned long timeout = jiffies + (HZ * CXL_TIMEOUT); 429 int rc = 0; 430 431 trace_cxl_llcmd(ctx, cmd); 432 433 WARN_ON(!ctx->afu->enabled); 434 435 ctx->elem->software_state = cpu_to_be32(pe_state); 436 smp_wmb(); 437 *(ctx->afu->native->sw_command_status) = cpu_to_be64(cmd | 0 | ctx->pe); 438 smp_mb(); 439 cxl_p1n_write(ctx->afu, CXL_PSL_LLCMD_An, cmd | ctx->pe); 440 while (1) { 441 if (time_after_eq(jiffies, timeout)) { 442 dev_warn(&ctx->afu->dev, "WARNING: Process Element Command timed out!\n"); 443 rc = -EBUSY; 444 goto out; 445 } 446 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) { 447 dev_warn(&ctx->afu->dev, "WARNING: Device link down, aborting Process Element Command!\n"); 448 rc = -EIO; 449 goto out; 450 } 451 state = be64_to_cpup(ctx->afu->native->sw_command_status); 452 if (state == ~0ULL) { 453 pr_err("cxl: Error adding process element to AFU\n"); 454 rc = -1; 455 goto out; 456 } 457 if ((state & (CXL_SPA_SW_CMD_MASK | CXL_SPA_SW_STATE_MASK | CXL_SPA_SW_LINK_MASK)) == 458 (cmd | (cmd >> 16) | ctx->pe)) 459 break; 460 /* 461 * The command won't finish in the PSL if there are 462 * outstanding DSIs. Hence we need to yield here in 463 * case there are outstanding DSIs that we need to 464 * service. Tuning possiblity: we could wait for a 465 * while before sched 466 */ 467 schedule(); 468 469 } 470 out: 471 trace_cxl_llcmd_done(ctx, cmd, rc); 472 return rc; 473 } 474 475 static int add_process_element(struct cxl_context *ctx) 476 { 477 int rc = 0; 478 479 mutex_lock(&ctx->afu->native->spa_mutex); 480 pr_devel("%s Adding pe: %i started\n", __func__, ctx->pe); 481 if (!(rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_ADD, CXL_PE_SOFTWARE_STATE_V))) 482 ctx->pe_inserted = true; 483 pr_devel("%s Adding pe: %i finished\n", __func__, ctx->pe); 484 mutex_unlock(&ctx->afu->native->spa_mutex); 485 return rc; 486 } 487 488 static int terminate_process_element(struct cxl_context *ctx) 489 { 490 int rc = 0; 491 492 /* fast path terminate if it's already invalid */ 493 if (!(ctx->elem->software_state & cpu_to_be32(CXL_PE_SOFTWARE_STATE_V))) 494 return rc; 495 496 mutex_lock(&ctx->afu->native->spa_mutex); 497 pr_devel("%s Terminate pe: %i started\n", __func__, ctx->pe); 498 /* We could be asked to terminate when the hw is down. That 499 * should always succeed: it's not running if the hw has gone 500 * away and is being reset. 501 */ 502 if (cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 503 rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_TERMINATE, 504 CXL_PE_SOFTWARE_STATE_V | CXL_PE_SOFTWARE_STATE_T); 505 ctx->elem->software_state = 0; /* Remove Valid bit */ 506 pr_devel("%s Terminate pe: %i finished\n", __func__, ctx->pe); 507 mutex_unlock(&ctx->afu->native->spa_mutex); 508 return rc; 509 } 510 511 static int remove_process_element(struct cxl_context *ctx) 512 { 513 int rc = 0; 514 515 mutex_lock(&ctx->afu->native->spa_mutex); 516 pr_devel("%s Remove pe: %i started\n", __func__, ctx->pe); 517 518 /* We could be asked to remove when the hw is down. Again, if 519 * the hw is down, the PE is gone, so we succeed. 520 */ 521 if (cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) 522 rc = do_process_element_cmd(ctx, CXL_SPA_SW_CMD_REMOVE, 0); 523 524 if (!rc) 525 ctx->pe_inserted = false; 526 if (cxl_is_power8()) 527 slb_invalid(ctx); 528 pr_devel("%s Remove pe: %i finished\n", __func__, ctx->pe); 529 mutex_unlock(&ctx->afu->native->spa_mutex); 530 531 return rc; 532 } 533 534 void cxl_assign_psn_space(struct cxl_context *ctx) 535 { 536 if (!ctx->afu->pp_size || ctx->master) { 537 ctx->psn_phys = ctx->afu->psn_phys; 538 ctx->psn_size = ctx->afu->adapter->ps_size; 539 } else { 540 ctx->psn_phys = ctx->afu->psn_phys + 541 (ctx->afu->native->pp_offset + ctx->afu->pp_size * ctx->pe); 542 ctx->psn_size = ctx->afu->pp_size; 543 } 544 } 545 546 static int activate_afu_directed(struct cxl_afu *afu) 547 { 548 int rc; 549 550 dev_info(&afu->dev, "Activating AFU directed mode\n"); 551 552 afu->num_procs = afu->max_procs_virtualised; 553 if (afu->native->spa == NULL) { 554 if (cxl_alloc_spa(afu, CXL_MODE_DIRECTED)) 555 return -ENOMEM; 556 } 557 attach_spa(afu); 558 559 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_AFU); 560 if (cxl_is_power8()) 561 cxl_p1n_write(afu, CXL_PSL_AMOR_An, 0xFFFFFFFFFFFFFFFFULL); 562 cxl_p1n_write(afu, CXL_PSL_ID_An, CXL_PSL_ID_An_F | CXL_PSL_ID_An_L); 563 564 afu->current_mode = CXL_MODE_DIRECTED; 565 566 if ((rc = cxl_chardev_m_afu_add(afu))) 567 return rc; 568 569 if ((rc = cxl_sysfs_afu_m_add(afu))) 570 goto err; 571 572 if ((rc = cxl_chardev_s_afu_add(afu))) 573 goto err1; 574 575 return 0; 576 err1: 577 cxl_sysfs_afu_m_remove(afu); 578 err: 579 cxl_chardev_afu_remove(afu); 580 return rc; 581 } 582 583 #ifdef CONFIG_CPU_LITTLE_ENDIAN 584 #define set_endian(sr) ((sr) |= CXL_PSL_SR_An_LE) 585 #else 586 #define set_endian(sr) ((sr) &= ~(CXL_PSL_SR_An_LE)) 587 #endif 588 589 u64 cxl_calculate_sr(bool master, bool kernel, bool real_mode, bool p9) 590 { 591 u64 sr = 0; 592 593 set_endian(sr); 594 if (master) 595 sr |= CXL_PSL_SR_An_MP; 596 if (mfspr(SPRN_LPCR) & LPCR_TC) 597 sr |= CXL_PSL_SR_An_TC; 598 if (kernel) { 599 if (!real_mode) 600 sr |= CXL_PSL_SR_An_R; 601 sr |= (mfmsr() & MSR_SF) | CXL_PSL_SR_An_HV; 602 } else { 603 sr |= CXL_PSL_SR_An_PR | CXL_PSL_SR_An_R; 604 if (radix_enabled()) 605 sr |= CXL_PSL_SR_An_HV; 606 else 607 sr &= ~(CXL_PSL_SR_An_HV); 608 if (!test_tsk_thread_flag(current, TIF_32BIT)) 609 sr |= CXL_PSL_SR_An_SF; 610 } 611 if (p9) { 612 if (radix_enabled()) 613 sr |= CXL_PSL_SR_An_XLAT_ror; 614 else 615 sr |= CXL_PSL_SR_An_XLAT_hpt; 616 } 617 return sr; 618 } 619 620 static u64 calculate_sr(struct cxl_context *ctx) 621 { 622 return cxl_calculate_sr(ctx->master, ctx->kernel, ctx->real_mode, 623 cxl_is_power9()); 624 } 625 626 static void update_ivtes_directed(struct cxl_context *ctx) 627 { 628 bool need_update = (ctx->status == STARTED); 629 int r; 630 631 if (need_update) { 632 WARN_ON(terminate_process_element(ctx)); 633 WARN_ON(remove_process_element(ctx)); 634 } 635 636 for (r = 0; r < CXL_IRQ_RANGES; r++) { 637 ctx->elem->ivte_offsets[r] = cpu_to_be16(ctx->irqs.offset[r]); 638 ctx->elem->ivte_ranges[r] = cpu_to_be16(ctx->irqs.range[r]); 639 } 640 641 /* 642 * Theoretically we could use the update llcmd, instead of a 643 * terminate/remove/add (or if an atomic update was required we could 644 * do a suspend/update/resume), however it seems there might be issues 645 * with the update llcmd on some cards (including those using an XSL on 646 * an ASIC) so for now it's safest to go with the commands that are 647 * known to work. In the future if we come across a situation where the 648 * card may be performing transactions using the same PE while we are 649 * doing this update we might need to revisit this. 650 */ 651 if (need_update) 652 WARN_ON(add_process_element(ctx)); 653 } 654 655 static int process_element_entry_psl9(struct cxl_context *ctx, u64 wed, u64 amr) 656 { 657 u32 pid; 658 659 cxl_assign_psn_space(ctx); 660 661 ctx->elem->ctxtime = 0; /* disable */ 662 ctx->elem->lpid = cpu_to_be32(mfspr(SPRN_LPID)); 663 ctx->elem->haurp = 0; /* disable */ 664 665 if (ctx->kernel) 666 pid = 0; 667 else { 668 if (ctx->mm == NULL) { 669 pr_devel("%s: unable to get mm for pe=%d pid=%i\n", 670 __func__, ctx->pe, pid_nr(ctx->pid)); 671 return -EINVAL; 672 } 673 pid = ctx->mm->context.id; 674 } 675 676 ctx->elem->common.tid = 0; 677 ctx->elem->common.pid = cpu_to_be32(pid); 678 679 ctx->elem->sr = cpu_to_be64(calculate_sr(ctx)); 680 681 ctx->elem->common.csrp = 0; /* disable */ 682 683 cxl_prefault(ctx, wed); 684 685 /* 686 * Ensure we have the multiplexed PSL interrupt set up to take faults 687 * for kernel contexts that may not have allocated any AFU IRQs at all: 688 */ 689 if (ctx->irqs.range[0] == 0) { 690 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq; 691 ctx->irqs.range[0] = 1; 692 } 693 694 ctx->elem->common.amr = cpu_to_be64(amr); 695 ctx->elem->common.wed = cpu_to_be64(wed); 696 697 return 0; 698 } 699 700 int cxl_attach_afu_directed_psl9(struct cxl_context *ctx, u64 wed, u64 amr) 701 { 702 int result; 703 704 /* fill the process element entry */ 705 result = process_element_entry_psl9(ctx, wed, amr); 706 if (result) 707 return result; 708 709 update_ivtes_directed(ctx); 710 711 /* first guy needs to enable */ 712 result = cxl_ops->afu_check_and_enable(ctx->afu); 713 if (result) 714 return result; 715 716 return add_process_element(ctx); 717 } 718 719 int cxl_attach_afu_directed_psl8(struct cxl_context *ctx, u64 wed, u64 amr) 720 { 721 u32 pid; 722 int result; 723 724 cxl_assign_psn_space(ctx); 725 726 ctx->elem->ctxtime = 0; /* disable */ 727 ctx->elem->lpid = cpu_to_be32(mfspr(SPRN_LPID)); 728 ctx->elem->haurp = 0; /* disable */ 729 ctx->elem->u.sdr = cpu_to_be64(mfspr(SPRN_SDR1)); 730 731 pid = current->pid; 732 if (ctx->kernel) 733 pid = 0; 734 ctx->elem->common.tid = 0; 735 ctx->elem->common.pid = cpu_to_be32(pid); 736 737 ctx->elem->sr = cpu_to_be64(calculate_sr(ctx)); 738 739 ctx->elem->common.csrp = 0; /* disable */ 740 ctx->elem->common.u.psl8.aurp0 = 0; /* disable */ 741 ctx->elem->common.u.psl8.aurp1 = 0; /* disable */ 742 743 cxl_prefault(ctx, wed); 744 745 ctx->elem->common.u.psl8.sstp0 = cpu_to_be64(ctx->sstp0); 746 ctx->elem->common.u.psl8.sstp1 = cpu_to_be64(ctx->sstp1); 747 748 /* 749 * Ensure we have the multiplexed PSL interrupt set up to take faults 750 * for kernel contexts that may not have allocated any AFU IRQs at all: 751 */ 752 if (ctx->irqs.range[0] == 0) { 753 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq; 754 ctx->irqs.range[0] = 1; 755 } 756 757 update_ivtes_directed(ctx); 758 759 ctx->elem->common.amr = cpu_to_be64(amr); 760 ctx->elem->common.wed = cpu_to_be64(wed); 761 762 /* first guy needs to enable */ 763 if ((result = cxl_ops->afu_check_and_enable(ctx->afu))) 764 return result; 765 766 return add_process_element(ctx); 767 } 768 769 static int deactivate_afu_directed(struct cxl_afu *afu) 770 { 771 dev_info(&afu->dev, "Deactivating AFU directed mode\n"); 772 773 afu->current_mode = 0; 774 afu->num_procs = 0; 775 776 cxl_sysfs_afu_m_remove(afu); 777 cxl_chardev_afu_remove(afu); 778 779 /* 780 * The CAIA section 2.2.1 indicates that the procedure for starting and 781 * stopping an AFU in AFU directed mode is AFU specific, which is not 782 * ideal since this code is generic and with one exception has no 783 * knowledge of the AFU. This is in contrast to the procedure for 784 * disabling a dedicated process AFU, which is documented to just 785 * require a reset. The architecture does indicate that both an AFU 786 * reset and an AFU disable should result in the AFU being disabled and 787 * we do both followed by a PSL purge for safety. 788 * 789 * Notably we used to have some issues with the disable sequence on PSL 790 * cards, which is why we ended up using this heavy weight procedure in 791 * the first place, however a bug was discovered that had rendered the 792 * disable operation ineffective, so it is conceivable that was the 793 * sole explanation for those difficulties. Careful regression testing 794 * is recommended if anyone attempts to remove or reorder these 795 * operations. 796 * 797 * The XSL on the Mellanox CX4 behaves a little differently from the 798 * PSL based cards and will time out an AFU reset if the AFU is still 799 * enabled. That card is special in that we do have a means to identify 800 * it from this code, so in that case we skip the reset and just use a 801 * disable/purge to avoid the timeout and corresponding noise in the 802 * kernel log. 803 */ 804 if (afu->adapter->native->sl_ops->needs_reset_before_disable) 805 cxl_ops->afu_reset(afu); 806 cxl_afu_disable(afu); 807 cxl_psl_purge(afu); 808 809 return 0; 810 } 811 812 int cxl_activate_dedicated_process_psl9(struct cxl_afu *afu) 813 { 814 dev_info(&afu->dev, "Activating dedicated process mode\n"); 815 816 /* 817 * If XSL is set to dedicated mode (Set in PSL_SCNTL reg), the 818 * XSL and AFU are programmed to work with a single context. 819 * The context information should be configured in the SPA area 820 * index 0 (so PSL_SPAP must be configured before enabling the 821 * AFU). 822 */ 823 afu->num_procs = 1; 824 if (afu->native->spa == NULL) { 825 if (cxl_alloc_spa(afu, CXL_MODE_DEDICATED)) 826 return -ENOMEM; 827 } 828 attach_spa(afu); 829 830 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_Process); 831 cxl_p1n_write(afu, CXL_PSL_ID_An, CXL_PSL_ID_An_F | CXL_PSL_ID_An_L); 832 833 afu->current_mode = CXL_MODE_DEDICATED; 834 835 return cxl_chardev_d_afu_add(afu); 836 } 837 838 int cxl_activate_dedicated_process_psl8(struct cxl_afu *afu) 839 { 840 dev_info(&afu->dev, "Activating dedicated process mode\n"); 841 842 cxl_p1n_write(afu, CXL_PSL_SCNTL_An, CXL_PSL_SCNTL_An_PM_Process); 843 844 cxl_p1n_write(afu, CXL_PSL_CtxTime_An, 0); /* disable */ 845 cxl_p1n_write(afu, CXL_PSL_SPAP_An, 0); /* disable */ 846 cxl_p1n_write(afu, CXL_PSL_AMOR_An, 0xFFFFFFFFFFFFFFFFULL); 847 cxl_p1n_write(afu, CXL_PSL_LPID_An, mfspr(SPRN_LPID)); 848 cxl_p1n_write(afu, CXL_HAURP_An, 0); /* disable */ 849 cxl_p1n_write(afu, CXL_PSL_SDR_An, mfspr(SPRN_SDR1)); 850 851 cxl_p2n_write(afu, CXL_CSRP_An, 0); /* disable */ 852 cxl_p2n_write(afu, CXL_AURP0_An, 0); /* disable */ 853 cxl_p2n_write(afu, CXL_AURP1_An, 0); /* disable */ 854 855 afu->current_mode = CXL_MODE_DEDICATED; 856 afu->num_procs = 1; 857 858 return cxl_chardev_d_afu_add(afu); 859 } 860 861 void cxl_update_dedicated_ivtes_psl9(struct cxl_context *ctx) 862 { 863 int r; 864 865 for (r = 0; r < CXL_IRQ_RANGES; r++) { 866 ctx->elem->ivte_offsets[r] = cpu_to_be16(ctx->irqs.offset[r]); 867 ctx->elem->ivte_ranges[r] = cpu_to_be16(ctx->irqs.range[r]); 868 } 869 } 870 871 void cxl_update_dedicated_ivtes_psl8(struct cxl_context *ctx) 872 { 873 struct cxl_afu *afu = ctx->afu; 874 875 cxl_p1n_write(afu, CXL_PSL_IVTE_Offset_An, 876 (((u64)ctx->irqs.offset[0] & 0xffff) << 48) | 877 (((u64)ctx->irqs.offset[1] & 0xffff) << 32) | 878 (((u64)ctx->irqs.offset[2] & 0xffff) << 16) | 879 ((u64)ctx->irqs.offset[3] & 0xffff)); 880 cxl_p1n_write(afu, CXL_PSL_IVTE_Limit_An, (u64) 881 (((u64)ctx->irqs.range[0] & 0xffff) << 48) | 882 (((u64)ctx->irqs.range[1] & 0xffff) << 32) | 883 (((u64)ctx->irqs.range[2] & 0xffff) << 16) | 884 ((u64)ctx->irqs.range[3] & 0xffff)); 885 } 886 887 int cxl_attach_dedicated_process_psl9(struct cxl_context *ctx, u64 wed, u64 amr) 888 { 889 struct cxl_afu *afu = ctx->afu; 890 int result; 891 892 /* fill the process element entry */ 893 result = process_element_entry_psl9(ctx, wed, amr); 894 if (result) 895 return result; 896 897 if (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes) 898 afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx); 899 900 ctx->elem->software_state = cpu_to_be32(CXL_PE_SOFTWARE_STATE_V); 901 /* 902 * Ideally we should do a wmb() here to make sure the changes to the 903 * PE are visible to the card before we call afu_enable. 904 * On ppc64 though all mmios are preceded by a 'sync' instruction hence 905 * we dont dont need one here. 906 */ 907 908 result = cxl_ops->afu_reset(afu); 909 if (result) 910 return result; 911 912 return afu_enable(afu); 913 } 914 915 int cxl_attach_dedicated_process_psl8(struct cxl_context *ctx, u64 wed, u64 amr) 916 { 917 struct cxl_afu *afu = ctx->afu; 918 u64 pid; 919 int rc; 920 921 pid = (u64)current->pid << 32; 922 if (ctx->kernel) 923 pid = 0; 924 cxl_p2n_write(afu, CXL_PSL_PID_TID_An, pid); 925 926 cxl_p1n_write(afu, CXL_PSL_SR_An, calculate_sr(ctx)); 927 928 if ((rc = cxl_write_sstp(afu, ctx->sstp0, ctx->sstp1))) 929 return rc; 930 931 cxl_prefault(ctx, wed); 932 933 if (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes) 934 afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx); 935 936 cxl_p2n_write(afu, CXL_PSL_AMR_An, amr); 937 938 /* master only context for dedicated */ 939 cxl_assign_psn_space(ctx); 940 941 if ((rc = cxl_ops->afu_reset(afu))) 942 return rc; 943 944 cxl_p2n_write(afu, CXL_PSL_WED_An, wed); 945 946 return afu_enable(afu); 947 } 948 949 static int deactivate_dedicated_process(struct cxl_afu *afu) 950 { 951 dev_info(&afu->dev, "Deactivating dedicated process mode\n"); 952 953 afu->current_mode = 0; 954 afu->num_procs = 0; 955 956 cxl_chardev_afu_remove(afu); 957 958 return 0; 959 } 960 961 static int native_afu_deactivate_mode(struct cxl_afu *afu, int mode) 962 { 963 if (mode == CXL_MODE_DIRECTED) 964 return deactivate_afu_directed(afu); 965 if (mode == CXL_MODE_DEDICATED) 966 return deactivate_dedicated_process(afu); 967 return 0; 968 } 969 970 static int native_afu_activate_mode(struct cxl_afu *afu, int mode) 971 { 972 if (!mode) 973 return 0; 974 if (!(mode & afu->modes_supported)) 975 return -EINVAL; 976 977 if (!cxl_ops->link_ok(afu->adapter, afu)) { 978 WARN(1, "Device link is down, refusing to activate!\n"); 979 return -EIO; 980 } 981 982 if (mode == CXL_MODE_DIRECTED) 983 return activate_afu_directed(afu); 984 if ((mode == CXL_MODE_DEDICATED) && 985 (afu->adapter->native->sl_ops->activate_dedicated_process)) 986 return afu->adapter->native->sl_ops->activate_dedicated_process(afu); 987 988 return -EINVAL; 989 } 990 991 static int native_attach_process(struct cxl_context *ctx, bool kernel, 992 u64 wed, u64 amr) 993 { 994 if (!cxl_ops->link_ok(ctx->afu->adapter, ctx->afu)) { 995 WARN(1, "Device link is down, refusing to attach process!\n"); 996 return -EIO; 997 } 998 999 ctx->kernel = kernel; 1000 if ((ctx->afu->current_mode == CXL_MODE_DIRECTED) && 1001 (ctx->afu->adapter->native->sl_ops->attach_afu_directed)) 1002 return ctx->afu->adapter->native->sl_ops->attach_afu_directed(ctx, wed, amr); 1003 1004 if ((ctx->afu->current_mode == CXL_MODE_DEDICATED) && 1005 (ctx->afu->adapter->native->sl_ops->attach_dedicated_process)) 1006 return ctx->afu->adapter->native->sl_ops->attach_dedicated_process(ctx, wed, amr); 1007 1008 return -EINVAL; 1009 } 1010 1011 static inline int detach_process_native_dedicated(struct cxl_context *ctx) 1012 { 1013 /* 1014 * The CAIA section 2.1.1 indicates that we need to do an AFU reset to 1015 * stop the AFU in dedicated mode (we therefore do not make that 1016 * optional like we do in the afu directed path). It does not indicate 1017 * that we need to do an explicit disable (which should occur 1018 * implicitly as part of the reset) or purge, but we do these as well 1019 * to be on the safe side. 1020 * 1021 * Notably we used to have some issues with the disable sequence 1022 * (before the sequence was spelled out in the architecture) which is 1023 * why we were so heavy weight in the first place, however a bug was 1024 * discovered that had rendered the disable operation ineffective, so 1025 * it is conceivable that was the sole explanation for those 1026 * difficulties. Point is, we should be careful and do some regression 1027 * testing if we ever attempt to remove any part of this procedure. 1028 */ 1029 cxl_ops->afu_reset(ctx->afu); 1030 cxl_afu_disable(ctx->afu); 1031 cxl_psl_purge(ctx->afu); 1032 return 0; 1033 } 1034 1035 static void native_update_ivtes(struct cxl_context *ctx) 1036 { 1037 if (ctx->afu->current_mode == CXL_MODE_DIRECTED) 1038 return update_ivtes_directed(ctx); 1039 if ((ctx->afu->current_mode == CXL_MODE_DEDICATED) && 1040 (ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes)) 1041 return ctx->afu->adapter->native->sl_ops->update_dedicated_ivtes(ctx); 1042 WARN(1, "native_update_ivtes: Bad mode\n"); 1043 } 1044 1045 static inline int detach_process_native_afu_directed(struct cxl_context *ctx) 1046 { 1047 if (!ctx->pe_inserted) 1048 return 0; 1049 if (terminate_process_element(ctx)) 1050 return -1; 1051 if (remove_process_element(ctx)) 1052 return -1; 1053 1054 return 0; 1055 } 1056 1057 static int native_detach_process(struct cxl_context *ctx) 1058 { 1059 trace_cxl_detach(ctx); 1060 1061 if (ctx->afu->current_mode == CXL_MODE_DEDICATED) 1062 return detach_process_native_dedicated(ctx); 1063 1064 return detach_process_native_afu_directed(ctx); 1065 } 1066 1067 static int native_get_irq_info(struct cxl_afu *afu, struct cxl_irq_info *info) 1068 { 1069 /* If the adapter has gone away, we can't get any meaningful 1070 * information. 1071 */ 1072 if (!cxl_ops->link_ok(afu->adapter, afu)) 1073 return -EIO; 1074 1075 info->dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An); 1076 info->dar = cxl_p2n_read(afu, CXL_PSL_DAR_An); 1077 if (cxl_is_power8()) 1078 info->dsr = cxl_p2n_read(afu, CXL_PSL_DSR_An); 1079 info->afu_err = cxl_p2n_read(afu, CXL_AFU_ERR_An); 1080 info->errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An); 1081 info->proc_handle = 0; 1082 1083 return 0; 1084 } 1085 1086 void cxl_native_irq_dump_regs_psl9(struct cxl_context *ctx) 1087 { 1088 u64 fir1, serr; 1089 1090 fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL9_FIR1); 1091 1092 dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1); 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 } 1098 1099 void cxl_native_irq_dump_regs_psl8(struct cxl_context *ctx) 1100 { 1101 u64 fir1, fir2, fir_slice, serr, afu_debug; 1102 1103 fir1 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR1); 1104 fir2 = cxl_p1_read(ctx->afu->adapter, CXL_PSL_FIR2); 1105 fir_slice = cxl_p1n_read(ctx->afu, CXL_PSL_FIR_SLICE_An); 1106 afu_debug = cxl_p1n_read(ctx->afu, CXL_AFU_DEBUG_An); 1107 1108 dev_crit(&ctx->afu->dev, "PSL_FIR1: 0x%016llx\n", fir1); 1109 dev_crit(&ctx->afu->dev, "PSL_FIR2: 0x%016llx\n", fir2); 1110 if (ctx->afu->adapter->native->sl_ops->register_serr_irq) { 1111 serr = cxl_p1n_read(ctx->afu, CXL_PSL_SERR_An); 1112 cxl_afu_decode_psl_serr(ctx->afu, serr); 1113 } 1114 dev_crit(&ctx->afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice); 1115 dev_crit(&ctx->afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug); 1116 } 1117 1118 static irqreturn_t native_handle_psl_slice_error(struct cxl_context *ctx, 1119 u64 dsisr, u64 errstat) 1120 { 1121 1122 dev_crit(&ctx->afu->dev, "PSL ERROR STATUS: 0x%016llx\n", errstat); 1123 1124 if (ctx->afu->adapter->native->sl_ops->psl_irq_dump_registers) 1125 ctx->afu->adapter->native->sl_ops->psl_irq_dump_registers(ctx); 1126 1127 if (ctx->afu->adapter->native->sl_ops->debugfs_stop_trace) { 1128 dev_crit(&ctx->afu->dev, "STOPPING CXL TRACE\n"); 1129 ctx->afu->adapter->native->sl_ops->debugfs_stop_trace(ctx->afu->adapter); 1130 } 1131 1132 return cxl_ops->ack_irq(ctx, 0, errstat); 1133 } 1134 1135 static bool cxl_is_translation_fault(struct cxl_afu *afu, u64 dsisr) 1136 { 1137 if ((cxl_is_power8()) && (dsisr & CXL_PSL_DSISR_TRANS)) 1138 return true; 1139 1140 if ((cxl_is_power9()) && (dsisr & CXL_PSL9_DSISR_An_TF)) 1141 return true; 1142 1143 return false; 1144 } 1145 1146 irqreturn_t cxl_fail_irq_psl(struct cxl_afu *afu, struct cxl_irq_info *irq_info) 1147 { 1148 if (cxl_is_translation_fault(afu, irq_info->dsisr)) 1149 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_AE); 1150 else 1151 cxl_p2n_write(afu, CXL_PSL_TFC_An, CXL_PSL_TFC_An_A); 1152 1153 return IRQ_HANDLED; 1154 } 1155 1156 static irqreturn_t native_irq_multiplexed(int irq, void *data) 1157 { 1158 struct cxl_afu *afu = data; 1159 struct cxl_context *ctx; 1160 struct cxl_irq_info irq_info; 1161 u64 phreg = cxl_p2n_read(afu, CXL_PSL_PEHandle_An); 1162 int ph, ret = IRQ_HANDLED, res; 1163 1164 /* check if eeh kicked in while the interrupt was in flight */ 1165 if (unlikely(phreg == ~0ULL)) { 1166 dev_warn(&afu->dev, 1167 "Ignoring slice interrupt(%d) due to fenced card", 1168 irq); 1169 return IRQ_HANDLED; 1170 } 1171 /* Mask the pe-handle from register value */ 1172 ph = phreg & 0xffff; 1173 if ((res = native_get_irq_info(afu, &irq_info))) { 1174 WARN(1, "Unable to get CXL IRQ Info: %i\n", res); 1175 if (afu->adapter->native->sl_ops->fail_irq) 1176 return afu->adapter->native->sl_ops->fail_irq(afu, &irq_info); 1177 return ret; 1178 } 1179 1180 rcu_read_lock(); 1181 ctx = idr_find(&afu->contexts_idr, ph); 1182 if (ctx) { 1183 if (afu->adapter->native->sl_ops->handle_interrupt) 1184 ret = afu->adapter->native->sl_ops->handle_interrupt(irq, ctx, &irq_info); 1185 rcu_read_unlock(); 1186 return ret; 1187 } 1188 rcu_read_unlock(); 1189 1190 WARN(1, "Unable to demultiplex CXL PSL IRQ for PE %i DSISR %016llx DAR" 1191 " %016llx\n(Possible AFU HW issue - was a term/remove acked" 1192 " with outstanding transactions?)\n", ph, irq_info.dsisr, 1193 irq_info.dar); 1194 if (afu->adapter->native->sl_ops->fail_irq) 1195 ret = afu->adapter->native->sl_ops->fail_irq(afu, &irq_info); 1196 return ret; 1197 } 1198 1199 static void native_irq_wait(struct cxl_context *ctx) 1200 { 1201 u64 dsisr; 1202 int timeout = 1000; 1203 int ph; 1204 1205 /* 1206 * Wait until no further interrupts are presented by the PSL 1207 * for this context. 1208 */ 1209 while (timeout--) { 1210 ph = cxl_p2n_read(ctx->afu, CXL_PSL_PEHandle_An) & 0xffff; 1211 if (ph != ctx->pe) 1212 return; 1213 dsisr = cxl_p2n_read(ctx->afu, CXL_PSL_DSISR_An); 1214 if (cxl_is_power8() && 1215 ((dsisr & CXL_PSL_DSISR_PENDING) == 0)) 1216 return; 1217 if (cxl_is_power9() && 1218 ((dsisr & CXL_PSL9_DSISR_PENDING) == 0)) 1219 return; 1220 /* 1221 * We are waiting for the workqueue to process our 1222 * irq, so need to let that run here. 1223 */ 1224 msleep(1); 1225 } 1226 1227 dev_warn(&ctx->afu->dev, "WARNING: waiting on DSI for PE %i" 1228 " DSISR %016llx!\n", ph, dsisr); 1229 return; 1230 } 1231 1232 static irqreturn_t native_slice_irq_err(int irq, void *data) 1233 { 1234 struct cxl_afu *afu = data; 1235 u64 errstat, serr, afu_error, dsisr; 1236 u64 fir_slice, afu_debug, irq_mask; 1237 1238 /* 1239 * slice err interrupt is only used with full PSL (no XSL) 1240 */ 1241 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An); 1242 errstat = cxl_p2n_read(afu, CXL_PSL_ErrStat_An); 1243 afu_error = cxl_p2n_read(afu, CXL_AFU_ERR_An); 1244 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An); 1245 cxl_afu_decode_psl_serr(afu, serr); 1246 1247 if (cxl_is_power8()) { 1248 fir_slice = cxl_p1n_read(afu, CXL_PSL_FIR_SLICE_An); 1249 afu_debug = cxl_p1n_read(afu, CXL_AFU_DEBUG_An); 1250 dev_crit(&afu->dev, "PSL_FIR_SLICE_An: 0x%016llx\n", fir_slice); 1251 dev_crit(&afu->dev, "CXL_PSL_AFU_DEBUG_An: 0x%016llx\n", afu_debug); 1252 } 1253 dev_crit(&afu->dev, "CXL_PSL_ErrStat_An: 0x%016llx\n", errstat); 1254 dev_crit(&afu->dev, "AFU_ERR_An: 0x%.16llx\n", afu_error); 1255 dev_crit(&afu->dev, "PSL_DSISR_An: 0x%.16llx\n", dsisr); 1256 1257 /* mask off the IRQ so it won't retrigger until the AFU is reset */ 1258 irq_mask = (serr & CXL_PSL_SERR_An_IRQS) >> 32; 1259 serr |= irq_mask; 1260 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr); 1261 dev_info(&afu->dev, "Further such interrupts will be masked until the AFU is reset\n"); 1262 1263 return IRQ_HANDLED; 1264 } 1265 1266 void cxl_native_err_irq_dump_regs_psl9(struct cxl *adapter) 1267 { 1268 u64 fir1; 1269 1270 fir1 = cxl_p1_read(adapter, CXL_PSL9_FIR1); 1271 dev_crit(&adapter->dev, "PSL_FIR: 0x%016llx\n", fir1); 1272 } 1273 1274 void cxl_native_err_irq_dump_regs_psl8(struct cxl *adapter) 1275 { 1276 u64 fir1, fir2; 1277 1278 fir1 = cxl_p1_read(adapter, CXL_PSL_FIR1); 1279 fir2 = cxl_p1_read(adapter, CXL_PSL_FIR2); 1280 dev_crit(&adapter->dev, 1281 "PSL_FIR1: 0x%016llx\nPSL_FIR2: 0x%016llx\n", 1282 fir1, fir2); 1283 } 1284 1285 static irqreturn_t native_irq_err(int irq, void *data) 1286 { 1287 struct cxl *adapter = data; 1288 u64 err_ivte; 1289 1290 WARN(1, "CXL ERROR interrupt %i\n", irq); 1291 1292 err_ivte = cxl_p1_read(adapter, CXL_PSL_ErrIVTE); 1293 dev_crit(&adapter->dev, "PSL_ErrIVTE: 0x%016llx\n", err_ivte); 1294 1295 if (adapter->native->sl_ops->debugfs_stop_trace) { 1296 dev_crit(&adapter->dev, "STOPPING CXL TRACE\n"); 1297 adapter->native->sl_ops->debugfs_stop_trace(adapter); 1298 } 1299 1300 if (adapter->native->sl_ops->err_irq_dump_registers) 1301 adapter->native->sl_ops->err_irq_dump_registers(adapter); 1302 1303 return IRQ_HANDLED; 1304 } 1305 1306 int cxl_native_register_psl_err_irq(struct cxl *adapter) 1307 { 1308 int rc; 1309 1310 adapter->irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err", 1311 dev_name(&adapter->dev)); 1312 if (!adapter->irq_name) 1313 return -ENOMEM; 1314 1315 if ((rc = cxl_register_one_irq(adapter, native_irq_err, adapter, 1316 &adapter->native->err_hwirq, 1317 &adapter->native->err_virq, 1318 adapter->irq_name))) { 1319 kfree(adapter->irq_name); 1320 adapter->irq_name = NULL; 1321 return rc; 1322 } 1323 1324 cxl_p1_write(adapter, CXL_PSL_ErrIVTE, adapter->native->err_hwirq & 0xffff); 1325 1326 return 0; 1327 } 1328 1329 void cxl_native_release_psl_err_irq(struct cxl *adapter) 1330 { 1331 if (adapter->native->err_virq == 0 || 1332 adapter->native->err_virq != 1333 irq_find_mapping(NULL, adapter->native->err_hwirq)) 1334 return; 1335 1336 cxl_p1_write(adapter, CXL_PSL_ErrIVTE, 0x0000000000000000); 1337 cxl_unmap_irq(adapter->native->err_virq, adapter); 1338 cxl_ops->release_one_irq(adapter, adapter->native->err_hwirq); 1339 kfree(adapter->irq_name); 1340 adapter->native->err_virq = 0; 1341 } 1342 1343 int cxl_native_register_serr_irq(struct cxl_afu *afu) 1344 { 1345 u64 serr; 1346 int rc; 1347 1348 afu->err_irq_name = kasprintf(GFP_KERNEL, "cxl-%s-err", 1349 dev_name(&afu->dev)); 1350 if (!afu->err_irq_name) 1351 return -ENOMEM; 1352 1353 if ((rc = cxl_register_one_irq(afu->adapter, native_slice_irq_err, afu, 1354 &afu->serr_hwirq, 1355 &afu->serr_virq, afu->err_irq_name))) { 1356 kfree(afu->err_irq_name); 1357 afu->err_irq_name = NULL; 1358 return rc; 1359 } 1360 1361 serr = cxl_p1n_read(afu, CXL_PSL_SERR_An); 1362 if (cxl_is_power8()) 1363 serr = (serr & 0x00ffffffffff0000ULL) | (afu->serr_hwirq & 0xffff); 1364 if (cxl_is_power9()) { 1365 /* 1366 * By default, all errors are masked. So don't set all masks. 1367 * Slice errors will be transfered. 1368 */ 1369 serr = (serr & ~0xff0000007fffffffULL) | (afu->serr_hwirq & 0xffff); 1370 } 1371 cxl_p1n_write(afu, CXL_PSL_SERR_An, serr); 1372 1373 return 0; 1374 } 1375 1376 void cxl_native_release_serr_irq(struct cxl_afu *afu) 1377 { 1378 if (afu->serr_virq == 0 || 1379 afu->serr_virq != irq_find_mapping(NULL, afu->serr_hwirq)) 1380 return; 1381 1382 cxl_p1n_write(afu, CXL_PSL_SERR_An, 0x0000000000000000); 1383 cxl_unmap_irq(afu->serr_virq, afu); 1384 cxl_ops->release_one_irq(afu->adapter, afu->serr_hwirq); 1385 kfree(afu->err_irq_name); 1386 afu->serr_virq = 0; 1387 } 1388 1389 int cxl_native_register_psl_irq(struct cxl_afu *afu) 1390 { 1391 int rc; 1392 1393 afu->psl_irq_name = kasprintf(GFP_KERNEL, "cxl-%s", 1394 dev_name(&afu->dev)); 1395 if (!afu->psl_irq_name) 1396 return -ENOMEM; 1397 1398 if ((rc = cxl_register_one_irq(afu->adapter, native_irq_multiplexed, 1399 afu, &afu->native->psl_hwirq, &afu->native->psl_virq, 1400 afu->psl_irq_name))) { 1401 kfree(afu->psl_irq_name); 1402 afu->psl_irq_name = NULL; 1403 } 1404 return rc; 1405 } 1406 1407 void cxl_native_release_psl_irq(struct cxl_afu *afu) 1408 { 1409 if (afu->native->psl_virq == 0 || 1410 afu->native->psl_virq != 1411 irq_find_mapping(NULL, afu->native->psl_hwirq)) 1412 return; 1413 1414 cxl_unmap_irq(afu->native->psl_virq, afu); 1415 cxl_ops->release_one_irq(afu->adapter, afu->native->psl_hwirq); 1416 kfree(afu->psl_irq_name); 1417 afu->native->psl_virq = 0; 1418 } 1419 1420 static void recover_psl_err(struct cxl_afu *afu, u64 errstat) 1421 { 1422 u64 dsisr; 1423 1424 pr_devel("RECOVERING FROM PSL ERROR... (0x%016llx)\n", errstat); 1425 1426 /* Clear PSL_DSISR[PE] */ 1427 dsisr = cxl_p2n_read(afu, CXL_PSL_DSISR_An); 1428 cxl_p2n_write(afu, CXL_PSL_DSISR_An, dsisr & ~CXL_PSL_DSISR_An_PE); 1429 1430 /* Write 1s to clear error status bits */ 1431 cxl_p2n_write(afu, CXL_PSL_ErrStat_An, errstat); 1432 } 1433 1434 static int native_ack_irq(struct cxl_context *ctx, u64 tfc, u64 psl_reset_mask) 1435 { 1436 trace_cxl_psl_irq_ack(ctx, tfc); 1437 if (tfc) 1438 cxl_p2n_write(ctx->afu, CXL_PSL_TFC_An, tfc); 1439 if (psl_reset_mask) 1440 recover_psl_err(ctx->afu, psl_reset_mask); 1441 1442 return 0; 1443 } 1444 1445 int cxl_check_error(struct cxl_afu *afu) 1446 { 1447 return (cxl_p1n_read(afu, CXL_PSL_SCNTL_An) == ~0ULL); 1448 } 1449 1450 static bool native_support_attributes(const char *attr_name, 1451 enum cxl_attrs type) 1452 { 1453 return true; 1454 } 1455 1456 static int native_afu_cr_read64(struct cxl_afu *afu, int cr, u64 off, u64 *out) 1457 { 1458 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu))) 1459 return -EIO; 1460 if (unlikely(off >= afu->crs_len)) 1461 return -ERANGE; 1462 *out = in_le64(afu->native->afu_desc_mmio + afu->crs_offset + 1463 (cr * afu->crs_len) + off); 1464 return 0; 1465 } 1466 1467 static int native_afu_cr_read32(struct cxl_afu *afu, int cr, u64 off, u32 *out) 1468 { 1469 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu))) 1470 return -EIO; 1471 if (unlikely(off >= afu->crs_len)) 1472 return -ERANGE; 1473 *out = in_le32(afu->native->afu_desc_mmio + afu->crs_offset + 1474 (cr * afu->crs_len) + off); 1475 return 0; 1476 } 1477 1478 static int native_afu_cr_read16(struct cxl_afu *afu, int cr, u64 off, u16 *out) 1479 { 1480 u64 aligned_off = off & ~0x3L; 1481 u32 val; 1482 int rc; 1483 1484 rc = native_afu_cr_read32(afu, cr, aligned_off, &val); 1485 if (!rc) 1486 *out = (val >> ((off & 0x3) * 8)) & 0xffff; 1487 return rc; 1488 } 1489 1490 static int native_afu_cr_read8(struct cxl_afu *afu, int cr, u64 off, u8 *out) 1491 { 1492 u64 aligned_off = off & ~0x3L; 1493 u32 val; 1494 int rc; 1495 1496 rc = native_afu_cr_read32(afu, cr, aligned_off, &val); 1497 if (!rc) 1498 *out = (val >> ((off & 0x3) * 8)) & 0xff; 1499 return rc; 1500 } 1501 1502 static int native_afu_cr_write32(struct cxl_afu *afu, int cr, u64 off, u32 in) 1503 { 1504 if (unlikely(!cxl_ops->link_ok(afu->adapter, afu))) 1505 return -EIO; 1506 if (unlikely(off >= afu->crs_len)) 1507 return -ERANGE; 1508 out_le32(afu->native->afu_desc_mmio + afu->crs_offset + 1509 (cr * afu->crs_len) + off, in); 1510 return 0; 1511 } 1512 1513 static int native_afu_cr_write16(struct cxl_afu *afu, int cr, u64 off, u16 in) 1514 { 1515 u64 aligned_off = off & ~0x3L; 1516 u32 val32, mask, shift; 1517 int rc; 1518 1519 rc = native_afu_cr_read32(afu, cr, aligned_off, &val32); 1520 if (rc) 1521 return rc; 1522 shift = (off & 0x3) * 8; 1523 WARN_ON(shift == 24); 1524 mask = 0xffff << shift; 1525 val32 = (val32 & ~mask) | (in << shift); 1526 1527 rc = native_afu_cr_write32(afu, cr, aligned_off, val32); 1528 return rc; 1529 } 1530 1531 static int native_afu_cr_write8(struct cxl_afu *afu, int cr, u64 off, u8 in) 1532 { 1533 u64 aligned_off = off & ~0x3L; 1534 u32 val32, mask, shift; 1535 int rc; 1536 1537 rc = native_afu_cr_read32(afu, cr, aligned_off, &val32); 1538 if (rc) 1539 return rc; 1540 shift = (off & 0x3) * 8; 1541 mask = 0xff << shift; 1542 val32 = (val32 & ~mask) | (in << shift); 1543 1544 rc = native_afu_cr_write32(afu, cr, aligned_off, val32); 1545 return rc; 1546 } 1547 1548 const struct cxl_backend_ops cxl_native_ops = { 1549 .module = THIS_MODULE, 1550 .adapter_reset = cxl_pci_reset, 1551 .alloc_one_irq = cxl_pci_alloc_one_irq, 1552 .release_one_irq = cxl_pci_release_one_irq, 1553 .alloc_irq_ranges = cxl_pci_alloc_irq_ranges, 1554 .release_irq_ranges = cxl_pci_release_irq_ranges, 1555 .setup_irq = cxl_pci_setup_irq, 1556 .handle_psl_slice_error = native_handle_psl_slice_error, 1557 .psl_interrupt = NULL, 1558 .ack_irq = native_ack_irq, 1559 .irq_wait = native_irq_wait, 1560 .attach_process = native_attach_process, 1561 .detach_process = native_detach_process, 1562 .update_ivtes = native_update_ivtes, 1563 .support_attributes = native_support_attributes, 1564 .link_ok = cxl_adapter_link_ok, 1565 .release_afu = cxl_pci_release_afu, 1566 .afu_read_err_buffer = cxl_pci_afu_read_err_buffer, 1567 .afu_check_and_enable = native_afu_check_and_enable, 1568 .afu_activate_mode = native_afu_activate_mode, 1569 .afu_deactivate_mode = native_afu_deactivate_mode, 1570 .afu_reset = native_afu_reset, 1571 .afu_cr_read8 = native_afu_cr_read8, 1572 .afu_cr_read16 = native_afu_cr_read16, 1573 .afu_cr_read32 = native_afu_cr_read32, 1574 .afu_cr_read64 = native_afu_cr_read64, 1575 .afu_cr_write8 = native_afu_cr_write8, 1576 .afu_cr_write16 = native_afu_cr_write16, 1577 .afu_cr_write32 = native_afu_cr_write32, 1578 .read_adapter_vpd = cxl_pci_read_adapter_vpd, 1579 }; 1580