1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright(c) 2019 Intel Corporation. All rights rsvd. */ 3 #include <linux/init.h> 4 #include <linux/kernel.h> 5 #include <linux/module.h> 6 #include <linux/pci.h> 7 #include <linux/io-64-nonatomic-lo-hi.h> 8 #include <linux/dmaengine.h> 9 #include <linux/irq.h> 10 #include <uapi/linux/idxd.h> 11 #include "../dmaengine.h" 12 #include "idxd.h" 13 #include "registers.h" 14 15 static void idxd_cmd_exec(struct idxd_device *idxd, int cmd_code, u32 operand, 16 u32 *status); 17 static void idxd_device_wqs_clear_state(struct idxd_device *idxd); 18 static void idxd_wq_disable_cleanup(struct idxd_wq *wq); 19 20 /* Interrupt control bits */ 21 void idxd_unmask_error_interrupts(struct idxd_device *idxd) 22 { 23 union genctrl_reg genctrl; 24 25 genctrl.bits = ioread32(idxd->reg_base + IDXD_GENCTRL_OFFSET); 26 genctrl.softerr_int_en = 1; 27 genctrl.halt_int_en = 1; 28 iowrite32(genctrl.bits, idxd->reg_base + IDXD_GENCTRL_OFFSET); 29 } 30 31 void idxd_mask_error_interrupts(struct idxd_device *idxd) 32 { 33 union genctrl_reg genctrl; 34 35 genctrl.bits = ioread32(idxd->reg_base + IDXD_GENCTRL_OFFSET); 36 genctrl.softerr_int_en = 0; 37 genctrl.halt_int_en = 0; 38 iowrite32(genctrl.bits, idxd->reg_base + IDXD_GENCTRL_OFFSET); 39 } 40 41 static void free_hw_descs(struct idxd_wq *wq) 42 { 43 int i; 44 45 for (i = 0; i < wq->num_descs; i++) 46 kfree(wq->hw_descs[i]); 47 48 kfree(wq->hw_descs); 49 } 50 51 static int alloc_hw_descs(struct idxd_wq *wq, int num) 52 { 53 struct device *dev = &wq->idxd->pdev->dev; 54 int i; 55 int node = dev_to_node(dev); 56 57 wq->hw_descs = kcalloc_node(num, sizeof(struct dsa_hw_desc *), 58 GFP_KERNEL, node); 59 if (!wq->hw_descs) 60 return -ENOMEM; 61 62 for (i = 0; i < num; i++) { 63 wq->hw_descs[i] = kzalloc_node(sizeof(*wq->hw_descs[i]), 64 GFP_KERNEL, node); 65 if (!wq->hw_descs[i]) { 66 free_hw_descs(wq); 67 return -ENOMEM; 68 } 69 } 70 71 return 0; 72 } 73 74 static void free_descs(struct idxd_wq *wq) 75 { 76 int i; 77 78 for (i = 0; i < wq->num_descs; i++) 79 kfree(wq->descs[i]); 80 81 kfree(wq->descs); 82 } 83 84 static int alloc_descs(struct idxd_wq *wq, int num) 85 { 86 struct device *dev = &wq->idxd->pdev->dev; 87 int i; 88 int node = dev_to_node(dev); 89 90 wq->descs = kcalloc_node(num, sizeof(struct idxd_desc *), 91 GFP_KERNEL, node); 92 if (!wq->descs) 93 return -ENOMEM; 94 95 for (i = 0; i < num; i++) { 96 wq->descs[i] = kzalloc_node(sizeof(*wq->descs[i]), 97 GFP_KERNEL, node); 98 if (!wq->descs[i]) { 99 free_descs(wq); 100 return -ENOMEM; 101 } 102 } 103 104 return 0; 105 } 106 107 /* WQ control bits */ 108 int idxd_wq_alloc_resources(struct idxd_wq *wq) 109 { 110 struct idxd_device *idxd = wq->idxd; 111 struct device *dev = &idxd->pdev->dev; 112 int rc, num_descs, i; 113 114 if (wq->type != IDXD_WQT_KERNEL) 115 return 0; 116 117 num_descs = wq_dedicated(wq) ? wq->size : wq->threshold; 118 wq->num_descs = num_descs; 119 120 rc = alloc_hw_descs(wq, num_descs); 121 if (rc < 0) 122 return rc; 123 124 wq->compls_size = num_descs * idxd->data->compl_size; 125 wq->compls = dma_alloc_coherent(dev, wq->compls_size, &wq->compls_addr, GFP_KERNEL); 126 if (!wq->compls) { 127 rc = -ENOMEM; 128 goto fail_alloc_compls; 129 } 130 131 rc = alloc_descs(wq, num_descs); 132 if (rc < 0) 133 goto fail_alloc_descs; 134 135 rc = sbitmap_queue_init_node(&wq->sbq, num_descs, -1, false, GFP_KERNEL, 136 dev_to_node(dev)); 137 if (rc < 0) 138 goto fail_sbitmap_init; 139 140 for (i = 0; i < num_descs; i++) { 141 struct idxd_desc *desc = wq->descs[i]; 142 143 desc->hw = wq->hw_descs[i]; 144 if (idxd->data->type == IDXD_TYPE_DSA) 145 desc->completion = &wq->compls[i]; 146 else if (idxd->data->type == IDXD_TYPE_IAX) 147 desc->iax_completion = &wq->iax_compls[i]; 148 desc->compl_dma = wq->compls_addr + idxd->data->compl_size * i; 149 desc->id = i; 150 desc->wq = wq; 151 desc->cpu = -1; 152 } 153 154 return 0; 155 156 fail_sbitmap_init: 157 free_descs(wq); 158 fail_alloc_descs: 159 dma_free_coherent(dev, wq->compls_size, wq->compls, wq->compls_addr); 160 fail_alloc_compls: 161 free_hw_descs(wq); 162 return rc; 163 } 164 165 void idxd_wq_free_resources(struct idxd_wq *wq) 166 { 167 struct device *dev = &wq->idxd->pdev->dev; 168 169 if (wq->type != IDXD_WQT_KERNEL) 170 return; 171 172 free_hw_descs(wq); 173 free_descs(wq); 174 dma_free_coherent(dev, wq->compls_size, wq->compls, wq->compls_addr); 175 sbitmap_queue_free(&wq->sbq); 176 } 177 178 int idxd_wq_enable(struct idxd_wq *wq) 179 { 180 struct idxd_device *idxd = wq->idxd; 181 struct device *dev = &idxd->pdev->dev; 182 u32 status; 183 184 if (wq->state == IDXD_WQ_ENABLED) { 185 dev_dbg(dev, "WQ %d already enabled\n", wq->id); 186 return 0; 187 } 188 189 idxd_cmd_exec(idxd, IDXD_CMD_ENABLE_WQ, wq->id, &status); 190 191 if (status != IDXD_CMDSTS_SUCCESS && 192 status != IDXD_CMDSTS_ERR_WQ_ENABLED) { 193 dev_dbg(dev, "WQ enable failed: %#x\n", status); 194 return -ENXIO; 195 } 196 197 wq->state = IDXD_WQ_ENABLED; 198 set_bit(wq->id, idxd->wq_enable_map); 199 dev_dbg(dev, "WQ %d enabled\n", wq->id); 200 return 0; 201 } 202 203 int idxd_wq_disable(struct idxd_wq *wq, bool reset_config) 204 { 205 struct idxd_device *idxd = wq->idxd; 206 struct device *dev = &idxd->pdev->dev; 207 u32 status, operand; 208 209 dev_dbg(dev, "Disabling WQ %d\n", wq->id); 210 211 if (wq->state != IDXD_WQ_ENABLED) { 212 dev_dbg(dev, "WQ %d in wrong state: %d\n", wq->id, wq->state); 213 return 0; 214 } 215 216 operand = BIT(wq->id % 16) | ((wq->id / 16) << 16); 217 idxd_cmd_exec(idxd, IDXD_CMD_DISABLE_WQ, operand, &status); 218 219 if (status != IDXD_CMDSTS_SUCCESS) { 220 dev_dbg(dev, "WQ disable failed: %#x\n", status); 221 return -ENXIO; 222 } 223 224 if (reset_config) 225 idxd_wq_disable_cleanup(wq); 226 clear_bit(wq->id, idxd->wq_enable_map); 227 wq->state = IDXD_WQ_DISABLED; 228 dev_dbg(dev, "WQ %d disabled\n", wq->id); 229 return 0; 230 } 231 232 void idxd_wq_drain(struct idxd_wq *wq) 233 { 234 struct idxd_device *idxd = wq->idxd; 235 struct device *dev = &idxd->pdev->dev; 236 u32 operand; 237 238 if (wq->state != IDXD_WQ_ENABLED) { 239 dev_dbg(dev, "WQ %d in wrong state: %d\n", wq->id, wq->state); 240 return; 241 } 242 243 dev_dbg(dev, "Draining WQ %d\n", wq->id); 244 operand = BIT(wq->id % 16) | ((wq->id / 16) << 16); 245 idxd_cmd_exec(idxd, IDXD_CMD_DRAIN_WQ, operand, NULL); 246 } 247 248 void idxd_wq_reset(struct idxd_wq *wq) 249 { 250 struct idxd_device *idxd = wq->idxd; 251 struct device *dev = &idxd->pdev->dev; 252 u32 operand; 253 254 if (wq->state != IDXD_WQ_ENABLED) { 255 dev_dbg(dev, "WQ %d in wrong state: %d\n", wq->id, wq->state); 256 return; 257 } 258 259 operand = BIT(wq->id % 16) | ((wq->id / 16) << 16); 260 idxd_cmd_exec(idxd, IDXD_CMD_RESET_WQ, operand, NULL); 261 idxd_wq_disable_cleanup(wq); 262 } 263 264 int idxd_wq_map_portal(struct idxd_wq *wq) 265 { 266 struct idxd_device *idxd = wq->idxd; 267 struct pci_dev *pdev = idxd->pdev; 268 struct device *dev = &pdev->dev; 269 resource_size_t start; 270 271 start = pci_resource_start(pdev, IDXD_WQ_BAR); 272 start += idxd_get_wq_portal_full_offset(wq->id, IDXD_PORTAL_LIMITED); 273 274 wq->portal = devm_ioremap(dev, start, IDXD_PORTAL_SIZE); 275 if (!wq->portal) 276 return -ENOMEM; 277 278 return 0; 279 } 280 281 void idxd_wq_unmap_portal(struct idxd_wq *wq) 282 { 283 struct device *dev = &wq->idxd->pdev->dev; 284 285 devm_iounmap(dev, wq->portal); 286 wq->portal = NULL; 287 wq->portal_offset = 0; 288 } 289 290 void idxd_wqs_unmap_portal(struct idxd_device *idxd) 291 { 292 int i; 293 294 for (i = 0; i < idxd->max_wqs; i++) { 295 struct idxd_wq *wq = idxd->wqs[i]; 296 297 if (wq->portal) 298 idxd_wq_unmap_portal(wq); 299 } 300 } 301 302 static void __idxd_wq_set_priv_locked(struct idxd_wq *wq, int priv) 303 { 304 struct idxd_device *idxd = wq->idxd; 305 union wqcfg wqcfg; 306 unsigned int offset; 307 308 offset = WQCFG_OFFSET(idxd, wq->id, WQCFG_PRIVL_IDX); 309 spin_lock(&idxd->dev_lock); 310 wqcfg.bits[WQCFG_PRIVL_IDX] = ioread32(idxd->reg_base + offset); 311 wqcfg.priv = priv; 312 wq->wqcfg->bits[WQCFG_PRIVL_IDX] = wqcfg.bits[WQCFG_PRIVL_IDX]; 313 iowrite32(wqcfg.bits[WQCFG_PRIVL_IDX], idxd->reg_base + offset); 314 spin_unlock(&idxd->dev_lock); 315 } 316 317 static void __idxd_wq_set_pasid_locked(struct idxd_wq *wq, int pasid) 318 { 319 struct idxd_device *idxd = wq->idxd; 320 union wqcfg wqcfg; 321 unsigned int offset; 322 323 offset = WQCFG_OFFSET(idxd, wq->id, WQCFG_PASID_IDX); 324 spin_lock(&idxd->dev_lock); 325 wqcfg.bits[WQCFG_PASID_IDX] = ioread32(idxd->reg_base + offset); 326 wqcfg.pasid_en = 1; 327 wqcfg.pasid = pasid; 328 wq->wqcfg->bits[WQCFG_PASID_IDX] = wqcfg.bits[WQCFG_PASID_IDX]; 329 iowrite32(wqcfg.bits[WQCFG_PASID_IDX], idxd->reg_base + offset); 330 spin_unlock(&idxd->dev_lock); 331 } 332 333 int idxd_wq_set_pasid(struct idxd_wq *wq, int pasid) 334 { 335 int rc; 336 337 rc = idxd_wq_disable(wq, false); 338 if (rc < 0) 339 return rc; 340 341 __idxd_wq_set_pasid_locked(wq, pasid); 342 343 rc = idxd_wq_enable(wq); 344 if (rc < 0) 345 return rc; 346 347 return 0; 348 } 349 350 int idxd_wq_disable_pasid(struct idxd_wq *wq) 351 { 352 struct idxd_device *idxd = wq->idxd; 353 int rc; 354 union wqcfg wqcfg; 355 unsigned int offset; 356 357 rc = idxd_wq_disable(wq, false); 358 if (rc < 0) 359 return rc; 360 361 offset = WQCFG_OFFSET(idxd, wq->id, WQCFG_PASID_IDX); 362 spin_lock(&idxd->dev_lock); 363 wqcfg.bits[WQCFG_PASID_IDX] = ioread32(idxd->reg_base + offset); 364 wqcfg.pasid_en = 0; 365 wqcfg.pasid = 0; 366 iowrite32(wqcfg.bits[WQCFG_PASID_IDX], idxd->reg_base + offset); 367 spin_unlock(&idxd->dev_lock); 368 369 rc = idxd_wq_enable(wq); 370 if (rc < 0) 371 return rc; 372 373 return 0; 374 } 375 376 static void idxd_wq_disable_cleanup(struct idxd_wq *wq) 377 { 378 struct idxd_device *idxd = wq->idxd; 379 380 lockdep_assert_held(&wq->wq_lock); 381 wq->state = IDXD_WQ_DISABLED; 382 memset(wq->wqcfg, 0, idxd->wqcfg_size); 383 wq->type = IDXD_WQT_NONE; 384 wq->threshold = 0; 385 wq->priority = 0; 386 wq->enqcmds_retries = IDXD_ENQCMDS_RETRIES; 387 clear_bit(WQ_FLAG_DEDICATED, &wq->flags); 388 clear_bit(WQ_FLAG_BLOCK_ON_FAULT, &wq->flags); 389 clear_bit(WQ_FLAG_ATS_DISABLE, &wq->flags); 390 memset(wq->name, 0, WQ_NAME_SIZE); 391 wq->max_xfer_bytes = WQ_DEFAULT_MAX_XFER; 392 idxd_wq_set_max_batch_size(idxd->data->type, wq, WQ_DEFAULT_MAX_BATCH); 393 if (wq->opcap_bmap) 394 bitmap_copy(wq->opcap_bmap, idxd->opcap_bmap, IDXD_MAX_OPCAP_BITS); 395 } 396 397 static void idxd_wq_device_reset_cleanup(struct idxd_wq *wq) 398 { 399 lockdep_assert_held(&wq->wq_lock); 400 401 wq->size = 0; 402 wq->group = NULL; 403 } 404 405 static void idxd_wq_ref_release(struct percpu_ref *ref) 406 { 407 struct idxd_wq *wq = container_of(ref, struct idxd_wq, wq_active); 408 409 complete(&wq->wq_dead); 410 } 411 412 int idxd_wq_init_percpu_ref(struct idxd_wq *wq) 413 { 414 int rc; 415 416 memset(&wq->wq_active, 0, sizeof(wq->wq_active)); 417 rc = percpu_ref_init(&wq->wq_active, idxd_wq_ref_release, 418 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL); 419 if (rc < 0) 420 return rc; 421 reinit_completion(&wq->wq_dead); 422 reinit_completion(&wq->wq_resurrect); 423 return 0; 424 } 425 426 void __idxd_wq_quiesce(struct idxd_wq *wq) 427 { 428 lockdep_assert_held(&wq->wq_lock); 429 reinit_completion(&wq->wq_resurrect); 430 percpu_ref_kill(&wq->wq_active); 431 complete_all(&wq->wq_resurrect); 432 wait_for_completion(&wq->wq_dead); 433 } 434 435 void idxd_wq_quiesce(struct idxd_wq *wq) 436 { 437 mutex_lock(&wq->wq_lock); 438 __idxd_wq_quiesce(wq); 439 mutex_unlock(&wq->wq_lock); 440 } 441 442 /* Device control bits */ 443 static inline bool idxd_is_enabled(struct idxd_device *idxd) 444 { 445 union gensts_reg gensts; 446 447 gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET); 448 449 if (gensts.state == IDXD_DEVICE_STATE_ENABLED) 450 return true; 451 return false; 452 } 453 454 static inline bool idxd_device_is_halted(struct idxd_device *idxd) 455 { 456 union gensts_reg gensts; 457 458 gensts.bits = ioread32(idxd->reg_base + IDXD_GENSTATS_OFFSET); 459 460 return (gensts.state == IDXD_DEVICE_STATE_HALT); 461 } 462 463 /* 464 * This is function is only used for reset during probe and will 465 * poll for completion. Once the device is setup with interrupts, 466 * all commands will be done via interrupt completion. 467 */ 468 int idxd_device_init_reset(struct idxd_device *idxd) 469 { 470 struct device *dev = &idxd->pdev->dev; 471 union idxd_command_reg cmd; 472 473 if (idxd_device_is_halted(idxd)) { 474 dev_warn(&idxd->pdev->dev, "Device is HALTED!\n"); 475 return -ENXIO; 476 } 477 478 memset(&cmd, 0, sizeof(cmd)); 479 cmd.cmd = IDXD_CMD_RESET_DEVICE; 480 dev_dbg(dev, "%s: sending reset for init.\n", __func__); 481 spin_lock(&idxd->cmd_lock); 482 iowrite32(cmd.bits, idxd->reg_base + IDXD_CMD_OFFSET); 483 484 while (ioread32(idxd->reg_base + IDXD_CMDSTS_OFFSET) & 485 IDXD_CMDSTS_ACTIVE) 486 cpu_relax(); 487 spin_unlock(&idxd->cmd_lock); 488 return 0; 489 } 490 491 static void idxd_cmd_exec(struct idxd_device *idxd, int cmd_code, u32 operand, 492 u32 *status) 493 { 494 union idxd_command_reg cmd; 495 DECLARE_COMPLETION_ONSTACK(done); 496 u32 stat; 497 498 if (idxd_device_is_halted(idxd)) { 499 dev_warn(&idxd->pdev->dev, "Device is HALTED!\n"); 500 if (status) 501 *status = IDXD_CMDSTS_HW_ERR; 502 return; 503 } 504 505 memset(&cmd, 0, sizeof(cmd)); 506 cmd.cmd = cmd_code; 507 cmd.operand = operand; 508 cmd.int_req = 1; 509 510 spin_lock(&idxd->cmd_lock); 511 wait_event_lock_irq(idxd->cmd_waitq, 512 !test_bit(IDXD_FLAG_CMD_RUNNING, &idxd->flags), 513 idxd->cmd_lock); 514 515 dev_dbg(&idxd->pdev->dev, "%s: sending cmd: %#x op: %#x\n", 516 __func__, cmd_code, operand); 517 518 idxd->cmd_status = 0; 519 __set_bit(IDXD_FLAG_CMD_RUNNING, &idxd->flags); 520 idxd->cmd_done = &done; 521 iowrite32(cmd.bits, idxd->reg_base + IDXD_CMD_OFFSET); 522 523 /* 524 * After command submitted, release lock and go to sleep until 525 * the command completes via interrupt. 526 */ 527 spin_unlock(&idxd->cmd_lock); 528 wait_for_completion(&done); 529 stat = ioread32(idxd->reg_base + IDXD_CMDSTS_OFFSET); 530 spin_lock(&idxd->cmd_lock); 531 if (status) 532 *status = stat; 533 idxd->cmd_status = stat & GENMASK(7, 0); 534 535 __clear_bit(IDXD_FLAG_CMD_RUNNING, &idxd->flags); 536 /* Wake up other pending commands */ 537 wake_up(&idxd->cmd_waitq); 538 spin_unlock(&idxd->cmd_lock); 539 } 540 541 int idxd_device_enable(struct idxd_device *idxd) 542 { 543 struct device *dev = &idxd->pdev->dev; 544 u32 status; 545 546 if (idxd_is_enabled(idxd)) { 547 dev_dbg(dev, "Device already enabled\n"); 548 return -ENXIO; 549 } 550 551 idxd_cmd_exec(idxd, IDXD_CMD_ENABLE_DEVICE, 0, &status); 552 553 /* If the command is successful or if the device was enabled */ 554 if (status != IDXD_CMDSTS_SUCCESS && 555 status != IDXD_CMDSTS_ERR_DEV_ENABLED) { 556 dev_dbg(dev, "%s: err_code: %#x\n", __func__, status); 557 return -ENXIO; 558 } 559 560 idxd->state = IDXD_DEV_ENABLED; 561 return 0; 562 } 563 564 int idxd_device_disable(struct idxd_device *idxd) 565 { 566 struct device *dev = &idxd->pdev->dev; 567 u32 status; 568 569 if (!idxd_is_enabled(idxd)) { 570 dev_dbg(dev, "Device is not enabled\n"); 571 return 0; 572 } 573 574 idxd_cmd_exec(idxd, IDXD_CMD_DISABLE_DEVICE, 0, &status); 575 576 /* If the command is successful or if the device was disabled */ 577 if (status != IDXD_CMDSTS_SUCCESS && 578 !(status & IDXD_CMDSTS_ERR_DIS_DEV_EN)) { 579 dev_dbg(dev, "%s: err_code: %#x\n", __func__, status); 580 return -ENXIO; 581 } 582 583 idxd_device_clear_state(idxd); 584 return 0; 585 } 586 587 void idxd_device_reset(struct idxd_device *idxd) 588 { 589 idxd_cmd_exec(idxd, IDXD_CMD_RESET_DEVICE, 0, NULL); 590 idxd_device_clear_state(idxd); 591 spin_lock(&idxd->dev_lock); 592 idxd_unmask_error_interrupts(idxd); 593 spin_unlock(&idxd->dev_lock); 594 } 595 596 void idxd_device_drain_pasid(struct idxd_device *idxd, int pasid) 597 { 598 struct device *dev = &idxd->pdev->dev; 599 u32 operand; 600 601 operand = pasid; 602 dev_dbg(dev, "cmd: %u operand: %#x\n", IDXD_CMD_DRAIN_PASID, operand); 603 idxd_cmd_exec(idxd, IDXD_CMD_DRAIN_PASID, operand, NULL); 604 dev_dbg(dev, "pasid %d drained\n", pasid); 605 } 606 607 int idxd_device_request_int_handle(struct idxd_device *idxd, int idx, int *handle, 608 enum idxd_interrupt_type irq_type) 609 { 610 struct device *dev = &idxd->pdev->dev; 611 u32 operand, status; 612 613 if (!(idxd->hw.cmd_cap & BIT(IDXD_CMD_REQUEST_INT_HANDLE))) 614 return -EOPNOTSUPP; 615 616 dev_dbg(dev, "get int handle, idx %d\n", idx); 617 618 operand = idx & GENMASK(15, 0); 619 if (irq_type == IDXD_IRQ_IMS) 620 operand |= CMD_INT_HANDLE_IMS; 621 622 dev_dbg(dev, "cmd: %u operand: %#x\n", IDXD_CMD_REQUEST_INT_HANDLE, operand); 623 624 idxd_cmd_exec(idxd, IDXD_CMD_REQUEST_INT_HANDLE, operand, &status); 625 626 if ((status & IDXD_CMDSTS_ERR_MASK) != IDXD_CMDSTS_SUCCESS) { 627 dev_dbg(dev, "request int handle failed: %#x\n", status); 628 return -ENXIO; 629 } 630 631 *handle = (status >> IDXD_CMDSTS_RES_SHIFT) & GENMASK(15, 0); 632 633 dev_dbg(dev, "int handle acquired: %u\n", *handle); 634 return 0; 635 } 636 637 int idxd_device_release_int_handle(struct idxd_device *idxd, int handle, 638 enum idxd_interrupt_type irq_type) 639 { 640 struct device *dev = &idxd->pdev->dev; 641 u32 operand, status; 642 union idxd_command_reg cmd; 643 644 if (!(idxd->hw.cmd_cap & BIT(IDXD_CMD_RELEASE_INT_HANDLE))) 645 return -EOPNOTSUPP; 646 647 dev_dbg(dev, "release int handle, handle %d\n", handle); 648 649 memset(&cmd, 0, sizeof(cmd)); 650 operand = handle & GENMASK(15, 0); 651 652 if (irq_type == IDXD_IRQ_IMS) 653 operand |= CMD_INT_HANDLE_IMS; 654 655 cmd.cmd = IDXD_CMD_RELEASE_INT_HANDLE; 656 cmd.operand = operand; 657 658 dev_dbg(dev, "cmd: %u operand: %#x\n", IDXD_CMD_RELEASE_INT_HANDLE, operand); 659 660 spin_lock(&idxd->cmd_lock); 661 iowrite32(cmd.bits, idxd->reg_base + IDXD_CMD_OFFSET); 662 663 while (ioread32(idxd->reg_base + IDXD_CMDSTS_OFFSET) & IDXD_CMDSTS_ACTIVE) 664 cpu_relax(); 665 status = ioread32(idxd->reg_base + IDXD_CMDSTS_OFFSET); 666 spin_unlock(&idxd->cmd_lock); 667 668 if ((status & IDXD_CMDSTS_ERR_MASK) != IDXD_CMDSTS_SUCCESS) { 669 dev_dbg(dev, "release int handle failed: %#x\n", status); 670 return -ENXIO; 671 } 672 673 dev_dbg(dev, "int handle released.\n"); 674 return 0; 675 } 676 677 /* Device configuration bits */ 678 static void idxd_engines_clear_state(struct idxd_device *idxd) 679 { 680 struct idxd_engine *engine; 681 int i; 682 683 lockdep_assert_held(&idxd->dev_lock); 684 for (i = 0; i < idxd->max_engines; i++) { 685 engine = idxd->engines[i]; 686 engine->group = NULL; 687 } 688 } 689 690 static void idxd_groups_clear_state(struct idxd_device *idxd) 691 { 692 struct idxd_group *group; 693 int i; 694 695 lockdep_assert_held(&idxd->dev_lock); 696 for (i = 0; i < idxd->max_groups; i++) { 697 group = idxd->groups[i]; 698 memset(&group->grpcfg, 0, sizeof(group->grpcfg)); 699 group->num_engines = 0; 700 group->num_wqs = 0; 701 group->use_rdbuf_limit = false; 702 group->rdbufs_allowed = 0; 703 group->rdbufs_reserved = 0; 704 if (idxd->hw.version < DEVICE_VERSION_2 && !tc_override) { 705 group->tc_a = 1; 706 group->tc_b = 1; 707 } else { 708 group->tc_a = -1; 709 group->tc_b = -1; 710 } 711 group->desc_progress_limit = 0; 712 group->batch_progress_limit = 0; 713 } 714 } 715 716 static void idxd_device_wqs_clear_state(struct idxd_device *idxd) 717 { 718 int i; 719 720 for (i = 0; i < idxd->max_wqs; i++) { 721 struct idxd_wq *wq = idxd->wqs[i]; 722 723 mutex_lock(&wq->wq_lock); 724 idxd_wq_disable_cleanup(wq); 725 idxd_wq_device_reset_cleanup(wq); 726 mutex_unlock(&wq->wq_lock); 727 } 728 } 729 730 void idxd_device_clear_state(struct idxd_device *idxd) 731 { 732 /* IDXD is always disabled. Other states are cleared only when IDXD is configurable. */ 733 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) { 734 /* 735 * Clearing wq state is protected by wq lock. 736 * So no need to be protected by device lock. 737 */ 738 idxd_device_wqs_clear_state(idxd); 739 740 spin_lock(&idxd->dev_lock); 741 idxd_groups_clear_state(idxd); 742 idxd_engines_clear_state(idxd); 743 } else { 744 spin_lock(&idxd->dev_lock); 745 } 746 747 idxd->state = IDXD_DEV_DISABLED; 748 spin_unlock(&idxd->dev_lock); 749 } 750 751 static void idxd_group_config_write(struct idxd_group *group) 752 { 753 struct idxd_device *idxd = group->idxd; 754 struct device *dev = &idxd->pdev->dev; 755 int i; 756 u32 grpcfg_offset; 757 758 dev_dbg(dev, "Writing group %d cfg registers\n", group->id); 759 760 /* setup GRPWQCFG */ 761 for (i = 0; i < GRPWQCFG_STRIDES; i++) { 762 grpcfg_offset = GRPWQCFG_OFFSET(idxd, group->id, i); 763 iowrite64(group->grpcfg.wqs[i], idxd->reg_base + grpcfg_offset); 764 dev_dbg(dev, "GRPCFG wq[%d:%d: %#x]: %#llx\n", 765 group->id, i, grpcfg_offset, 766 ioread64(idxd->reg_base + grpcfg_offset)); 767 } 768 769 /* setup GRPENGCFG */ 770 grpcfg_offset = GRPENGCFG_OFFSET(idxd, group->id); 771 iowrite64(group->grpcfg.engines, idxd->reg_base + grpcfg_offset); 772 dev_dbg(dev, "GRPCFG engs[%d: %#x]: %#llx\n", group->id, 773 grpcfg_offset, ioread64(idxd->reg_base + grpcfg_offset)); 774 775 /* setup GRPFLAGS */ 776 grpcfg_offset = GRPFLGCFG_OFFSET(idxd, group->id); 777 iowrite64(group->grpcfg.flags.bits, idxd->reg_base + grpcfg_offset); 778 dev_dbg(dev, "GRPFLAGS flags[%d: %#x]: %#llx\n", 779 group->id, grpcfg_offset, 780 ioread64(idxd->reg_base + grpcfg_offset)); 781 } 782 783 static int idxd_groups_config_write(struct idxd_device *idxd) 784 785 { 786 union gencfg_reg reg; 787 int i; 788 struct device *dev = &idxd->pdev->dev; 789 790 /* Setup bandwidth rdbuf limit */ 791 if (idxd->hw.gen_cap.config_en && idxd->rdbuf_limit) { 792 reg.bits = ioread32(idxd->reg_base + IDXD_GENCFG_OFFSET); 793 reg.rdbuf_limit = idxd->rdbuf_limit; 794 iowrite32(reg.bits, idxd->reg_base + IDXD_GENCFG_OFFSET); 795 } 796 797 dev_dbg(dev, "GENCFG(%#x): %#x\n", IDXD_GENCFG_OFFSET, 798 ioread32(idxd->reg_base + IDXD_GENCFG_OFFSET)); 799 800 for (i = 0; i < idxd->max_groups; i++) { 801 struct idxd_group *group = idxd->groups[i]; 802 803 idxd_group_config_write(group); 804 } 805 806 return 0; 807 } 808 809 static bool idxd_device_pasid_priv_enabled(struct idxd_device *idxd) 810 { 811 struct pci_dev *pdev = idxd->pdev; 812 813 if (pdev->pasid_enabled && (pdev->pasid_features & PCI_PASID_CAP_PRIV)) 814 return true; 815 return false; 816 } 817 818 static int idxd_wq_config_write(struct idxd_wq *wq) 819 { 820 struct idxd_device *idxd = wq->idxd; 821 struct device *dev = &idxd->pdev->dev; 822 u32 wq_offset; 823 int i, n; 824 825 if (!wq->group) 826 return 0; 827 828 /* 829 * Instead of memset the entire shadow copy of WQCFG, copy from the hardware after 830 * wq reset. This will copy back the sticky values that are present on some devices. 831 */ 832 for (i = 0; i < WQCFG_STRIDES(idxd); i++) { 833 wq_offset = WQCFG_OFFSET(idxd, wq->id, i); 834 wq->wqcfg->bits[i] |= ioread32(idxd->reg_base + wq_offset); 835 } 836 837 if (wq->size == 0 && wq->type != IDXD_WQT_NONE) 838 wq->size = WQ_DEFAULT_QUEUE_DEPTH; 839 840 /* byte 0-3 */ 841 wq->wqcfg->wq_size = wq->size; 842 843 /* bytes 4-7 */ 844 wq->wqcfg->wq_thresh = wq->threshold; 845 846 /* byte 8-11 */ 847 if (wq_dedicated(wq)) 848 wq->wqcfg->mode = 1; 849 850 /* 851 * The WQ priv bit is set depending on the WQ type. priv = 1 if the 852 * WQ type is kernel to indicate privileged access. This setting only 853 * matters for dedicated WQ. According to the DSA spec: 854 * If the WQ is in dedicated mode, WQ PASID Enable is 1, and the 855 * Privileged Mode Enable field of the PCI Express PASID capability 856 * is 0, this field must be 0. 857 * 858 * In the case of a dedicated kernel WQ that is not able to support 859 * the PASID cap, then the configuration will be rejected. 860 */ 861 if (wq_dedicated(wq) && wq->wqcfg->pasid_en && 862 !idxd_device_pasid_priv_enabled(idxd) && 863 wq->type == IDXD_WQT_KERNEL) { 864 idxd->cmd_status = IDXD_SCMD_WQ_NO_PRIV; 865 return -EOPNOTSUPP; 866 } 867 868 wq->wqcfg->priority = wq->priority; 869 870 if (idxd->hw.gen_cap.block_on_fault && 871 test_bit(WQ_FLAG_BLOCK_ON_FAULT, &wq->flags)) 872 wq->wqcfg->bof = 1; 873 874 if (idxd->hw.wq_cap.wq_ats_support) 875 wq->wqcfg->wq_ats_disable = test_bit(WQ_FLAG_ATS_DISABLE, &wq->flags); 876 877 /* bytes 12-15 */ 878 wq->wqcfg->max_xfer_shift = ilog2(wq->max_xfer_bytes); 879 idxd_wqcfg_set_max_batch_shift(idxd->data->type, wq->wqcfg, ilog2(wq->max_batch_size)); 880 881 /* bytes 32-63 */ 882 if (idxd->hw.wq_cap.op_config && wq->opcap_bmap) { 883 memset(wq->wqcfg->op_config, 0, IDXD_MAX_OPCAP_BITS / 8); 884 for_each_set_bit(n, wq->opcap_bmap, IDXD_MAX_OPCAP_BITS) { 885 int pos = n % BITS_PER_LONG_LONG; 886 int idx = n / BITS_PER_LONG_LONG; 887 888 wq->wqcfg->op_config[idx] |= BIT(pos); 889 } 890 } 891 892 dev_dbg(dev, "WQ %d CFGs\n", wq->id); 893 for (i = 0; i < WQCFG_STRIDES(idxd); i++) { 894 wq_offset = WQCFG_OFFSET(idxd, wq->id, i); 895 iowrite32(wq->wqcfg->bits[i], idxd->reg_base + wq_offset); 896 dev_dbg(dev, "WQ[%d][%d][%#x]: %#x\n", 897 wq->id, i, wq_offset, 898 ioread32(idxd->reg_base + wq_offset)); 899 } 900 901 return 0; 902 } 903 904 static int idxd_wqs_config_write(struct idxd_device *idxd) 905 { 906 int i, rc; 907 908 for (i = 0; i < idxd->max_wqs; i++) { 909 struct idxd_wq *wq = idxd->wqs[i]; 910 911 rc = idxd_wq_config_write(wq); 912 if (rc < 0) 913 return rc; 914 } 915 916 return 0; 917 } 918 919 static void idxd_group_flags_setup(struct idxd_device *idxd) 920 { 921 int i; 922 923 /* TC-A 0 and TC-B 1 should be defaults */ 924 for (i = 0; i < idxd->max_groups; i++) { 925 struct idxd_group *group = idxd->groups[i]; 926 927 if (group->tc_a == -1) 928 group->tc_a = group->grpcfg.flags.tc_a = 0; 929 else 930 group->grpcfg.flags.tc_a = group->tc_a; 931 if (group->tc_b == -1) 932 group->tc_b = group->grpcfg.flags.tc_b = 1; 933 else 934 group->grpcfg.flags.tc_b = group->tc_b; 935 group->grpcfg.flags.use_rdbuf_limit = group->use_rdbuf_limit; 936 group->grpcfg.flags.rdbufs_reserved = group->rdbufs_reserved; 937 if (group->rdbufs_allowed) 938 group->grpcfg.flags.rdbufs_allowed = group->rdbufs_allowed; 939 else 940 group->grpcfg.flags.rdbufs_allowed = idxd->max_rdbufs; 941 942 group->grpcfg.flags.desc_progress_limit = group->desc_progress_limit; 943 group->grpcfg.flags.batch_progress_limit = group->batch_progress_limit; 944 } 945 } 946 947 static int idxd_engines_setup(struct idxd_device *idxd) 948 { 949 int i, engines = 0; 950 struct idxd_engine *eng; 951 struct idxd_group *group; 952 953 for (i = 0; i < idxd->max_groups; i++) { 954 group = idxd->groups[i]; 955 group->grpcfg.engines = 0; 956 } 957 958 for (i = 0; i < idxd->max_engines; i++) { 959 eng = idxd->engines[i]; 960 group = eng->group; 961 962 if (!group) 963 continue; 964 965 group->grpcfg.engines |= BIT(eng->id); 966 engines++; 967 } 968 969 if (!engines) 970 return -EINVAL; 971 972 return 0; 973 } 974 975 static int idxd_wqs_setup(struct idxd_device *idxd) 976 { 977 struct idxd_wq *wq; 978 struct idxd_group *group; 979 int i, j, configured = 0; 980 struct device *dev = &idxd->pdev->dev; 981 982 for (i = 0; i < idxd->max_groups; i++) { 983 group = idxd->groups[i]; 984 for (j = 0; j < 4; j++) 985 group->grpcfg.wqs[j] = 0; 986 } 987 988 for (i = 0; i < idxd->max_wqs; i++) { 989 wq = idxd->wqs[i]; 990 group = wq->group; 991 992 if (!wq->group) 993 continue; 994 995 if (wq_shared(wq) && !wq_shared_supported(wq)) { 996 idxd->cmd_status = IDXD_SCMD_WQ_NO_SWQ_SUPPORT; 997 dev_warn(dev, "No shared wq support but configured.\n"); 998 return -EINVAL; 999 } 1000 1001 group->grpcfg.wqs[wq->id / 64] |= BIT(wq->id % 64); 1002 configured++; 1003 } 1004 1005 if (configured == 0) { 1006 idxd->cmd_status = IDXD_SCMD_WQ_NONE_CONFIGURED; 1007 return -EINVAL; 1008 } 1009 1010 return 0; 1011 } 1012 1013 int idxd_device_config(struct idxd_device *idxd) 1014 { 1015 int rc; 1016 1017 lockdep_assert_held(&idxd->dev_lock); 1018 rc = idxd_wqs_setup(idxd); 1019 if (rc < 0) 1020 return rc; 1021 1022 rc = idxd_engines_setup(idxd); 1023 if (rc < 0) 1024 return rc; 1025 1026 idxd_group_flags_setup(idxd); 1027 1028 rc = idxd_wqs_config_write(idxd); 1029 if (rc < 0) 1030 return rc; 1031 1032 rc = idxd_groups_config_write(idxd); 1033 if (rc < 0) 1034 return rc; 1035 1036 return 0; 1037 } 1038 1039 static int idxd_wq_load_config(struct idxd_wq *wq) 1040 { 1041 struct idxd_device *idxd = wq->idxd; 1042 struct device *dev = &idxd->pdev->dev; 1043 int wqcfg_offset; 1044 int i; 1045 1046 wqcfg_offset = WQCFG_OFFSET(idxd, wq->id, 0); 1047 memcpy_fromio(wq->wqcfg, idxd->reg_base + wqcfg_offset, idxd->wqcfg_size); 1048 1049 wq->size = wq->wqcfg->wq_size; 1050 wq->threshold = wq->wqcfg->wq_thresh; 1051 1052 /* The driver does not support shared WQ mode in read-only config yet */ 1053 if (wq->wqcfg->mode == 0 || wq->wqcfg->pasid_en) 1054 return -EOPNOTSUPP; 1055 1056 set_bit(WQ_FLAG_DEDICATED, &wq->flags); 1057 1058 wq->priority = wq->wqcfg->priority; 1059 1060 wq->max_xfer_bytes = 1ULL << wq->wqcfg->max_xfer_shift; 1061 idxd_wq_set_max_batch_size(idxd->data->type, wq, 1U << wq->wqcfg->max_batch_shift); 1062 1063 for (i = 0; i < WQCFG_STRIDES(idxd); i++) { 1064 wqcfg_offset = WQCFG_OFFSET(idxd, wq->id, i); 1065 dev_dbg(dev, "WQ[%d][%d][%#x]: %#x\n", wq->id, i, wqcfg_offset, wq->wqcfg->bits[i]); 1066 } 1067 1068 return 0; 1069 } 1070 1071 static void idxd_group_load_config(struct idxd_group *group) 1072 { 1073 struct idxd_device *idxd = group->idxd; 1074 struct device *dev = &idxd->pdev->dev; 1075 int i, j, grpcfg_offset; 1076 1077 /* 1078 * Load WQS bit fields 1079 * Iterate through all 256 bits 64 bits at a time 1080 */ 1081 for (i = 0; i < GRPWQCFG_STRIDES; i++) { 1082 struct idxd_wq *wq; 1083 1084 grpcfg_offset = GRPWQCFG_OFFSET(idxd, group->id, i); 1085 group->grpcfg.wqs[i] = ioread64(idxd->reg_base + grpcfg_offset); 1086 dev_dbg(dev, "GRPCFG wq[%d:%d: %#x]: %#llx\n", 1087 group->id, i, grpcfg_offset, group->grpcfg.wqs[i]); 1088 1089 if (i * 64 >= idxd->max_wqs) 1090 break; 1091 1092 /* Iterate through all 64 bits and check for wq set */ 1093 for (j = 0; j < 64; j++) { 1094 int id = i * 64 + j; 1095 1096 /* No need to check beyond max wqs */ 1097 if (id >= idxd->max_wqs) 1098 break; 1099 1100 /* Set group assignment for wq if wq bit is set */ 1101 if (group->grpcfg.wqs[i] & BIT(j)) { 1102 wq = idxd->wqs[id]; 1103 wq->group = group; 1104 } 1105 } 1106 } 1107 1108 grpcfg_offset = GRPENGCFG_OFFSET(idxd, group->id); 1109 group->grpcfg.engines = ioread64(idxd->reg_base + grpcfg_offset); 1110 dev_dbg(dev, "GRPCFG engs[%d: %#x]: %#llx\n", group->id, 1111 grpcfg_offset, group->grpcfg.engines); 1112 1113 /* Iterate through all 64 bits to check engines set */ 1114 for (i = 0; i < 64; i++) { 1115 if (i >= idxd->max_engines) 1116 break; 1117 1118 if (group->grpcfg.engines & BIT(i)) { 1119 struct idxd_engine *engine = idxd->engines[i]; 1120 1121 engine->group = group; 1122 } 1123 } 1124 1125 grpcfg_offset = GRPFLGCFG_OFFSET(idxd, group->id); 1126 group->grpcfg.flags.bits = ioread64(idxd->reg_base + grpcfg_offset); 1127 dev_dbg(dev, "GRPFLAGS flags[%d: %#x]: %#llx\n", 1128 group->id, grpcfg_offset, group->grpcfg.flags.bits); 1129 } 1130 1131 int idxd_device_load_config(struct idxd_device *idxd) 1132 { 1133 union gencfg_reg reg; 1134 int i, rc; 1135 1136 reg.bits = ioread32(idxd->reg_base + IDXD_GENCFG_OFFSET); 1137 idxd->rdbuf_limit = reg.rdbuf_limit; 1138 1139 for (i = 0; i < idxd->max_groups; i++) { 1140 struct idxd_group *group = idxd->groups[i]; 1141 1142 idxd_group_load_config(group); 1143 } 1144 1145 for (i = 0; i < idxd->max_wqs; i++) { 1146 struct idxd_wq *wq = idxd->wqs[i]; 1147 1148 rc = idxd_wq_load_config(wq); 1149 if (rc < 0) 1150 return rc; 1151 } 1152 1153 return 0; 1154 } 1155 1156 static void idxd_flush_pending_descs(struct idxd_irq_entry *ie) 1157 { 1158 struct idxd_desc *desc, *itr; 1159 struct llist_node *head; 1160 LIST_HEAD(flist); 1161 enum idxd_complete_type ctype; 1162 1163 spin_lock(&ie->list_lock); 1164 head = llist_del_all(&ie->pending_llist); 1165 if (head) { 1166 llist_for_each_entry_safe(desc, itr, head, llnode) 1167 list_add_tail(&desc->list, &ie->work_list); 1168 } 1169 1170 list_for_each_entry_safe(desc, itr, &ie->work_list, list) 1171 list_move_tail(&desc->list, &flist); 1172 spin_unlock(&ie->list_lock); 1173 1174 list_for_each_entry_safe(desc, itr, &flist, list) { 1175 struct dma_async_tx_descriptor *tx; 1176 1177 list_del(&desc->list); 1178 ctype = desc->completion->status ? IDXD_COMPLETE_NORMAL : IDXD_COMPLETE_ABORT; 1179 /* 1180 * wq is being disabled. Any remaining descriptors are 1181 * likely to be stuck and can be dropped. callback could 1182 * point to code that is no longer accessible, for example 1183 * if dmatest module has been unloaded. 1184 */ 1185 tx = &desc->txd; 1186 tx->callback = NULL; 1187 tx->callback_result = NULL; 1188 idxd_dma_complete_txd(desc, ctype, true); 1189 } 1190 } 1191 1192 static void idxd_device_set_perm_entry(struct idxd_device *idxd, 1193 struct idxd_irq_entry *ie) 1194 { 1195 union msix_perm mperm; 1196 1197 if (ie->pasid == INVALID_IOASID) 1198 return; 1199 1200 mperm.bits = 0; 1201 mperm.pasid = ie->pasid; 1202 mperm.pasid_en = 1; 1203 iowrite32(mperm.bits, idxd->reg_base + idxd->msix_perm_offset + ie->id * 8); 1204 } 1205 1206 static void idxd_device_clear_perm_entry(struct idxd_device *idxd, 1207 struct idxd_irq_entry *ie) 1208 { 1209 iowrite32(0, idxd->reg_base + idxd->msix_perm_offset + ie->id * 8); 1210 } 1211 1212 void idxd_wq_free_irq(struct idxd_wq *wq) 1213 { 1214 struct idxd_device *idxd = wq->idxd; 1215 struct idxd_irq_entry *ie = &wq->ie; 1216 1217 if (wq->type != IDXD_WQT_KERNEL) 1218 return; 1219 1220 free_irq(ie->vector, ie); 1221 idxd_flush_pending_descs(ie); 1222 if (idxd->request_int_handles) 1223 idxd_device_release_int_handle(idxd, ie->int_handle, IDXD_IRQ_MSIX); 1224 idxd_device_clear_perm_entry(idxd, ie); 1225 ie->vector = -1; 1226 ie->int_handle = INVALID_INT_HANDLE; 1227 ie->pasid = INVALID_IOASID; 1228 } 1229 1230 int idxd_wq_request_irq(struct idxd_wq *wq) 1231 { 1232 struct idxd_device *idxd = wq->idxd; 1233 struct pci_dev *pdev = idxd->pdev; 1234 struct device *dev = &pdev->dev; 1235 struct idxd_irq_entry *ie; 1236 int rc; 1237 1238 if (wq->type != IDXD_WQT_KERNEL) 1239 return 0; 1240 1241 ie = &wq->ie; 1242 ie->vector = pci_irq_vector(pdev, ie->id); 1243 ie->pasid = device_pasid_enabled(idxd) ? idxd->pasid : INVALID_IOASID; 1244 idxd_device_set_perm_entry(idxd, ie); 1245 1246 rc = request_threaded_irq(ie->vector, NULL, idxd_wq_thread, 0, "idxd-portal", ie); 1247 if (rc < 0) { 1248 dev_err(dev, "Failed to request irq %d.\n", ie->vector); 1249 goto err_irq; 1250 } 1251 1252 if (idxd->request_int_handles) { 1253 rc = idxd_device_request_int_handle(idxd, ie->id, &ie->int_handle, 1254 IDXD_IRQ_MSIX); 1255 if (rc < 0) 1256 goto err_int_handle; 1257 } else { 1258 ie->int_handle = ie->id; 1259 } 1260 1261 return 0; 1262 1263 err_int_handle: 1264 ie->int_handle = INVALID_INT_HANDLE; 1265 free_irq(ie->vector, ie); 1266 err_irq: 1267 idxd_device_clear_perm_entry(idxd, ie); 1268 ie->pasid = INVALID_IOASID; 1269 return rc; 1270 } 1271 1272 int drv_enable_wq(struct idxd_wq *wq) 1273 { 1274 struct idxd_device *idxd = wq->idxd; 1275 struct device *dev = &idxd->pdev->dev; 1276 int rc = -ENXIO; 1277 1278 lockdep_assert_held(&wq->wq_lock); 1279 1280 if (idxd->state != IDXD_DEV_ENABLED) { 1281 idxd->cmd_status = IDXD_SCMD_DEV_NOT_ENABLED; 1282 goto err; 1283 } 1284 1285 if (wq->state != IDXD_WQ_DISABLED) { 1286 dev_dbg(dev, "wq %d already enabled.\n", wq->id); 1287 idxd->cmd_status = IDXD_SCMD_WQ_ENABLED; 1288 rc = -EBUSY; 1289 goto err; 1290 } 1291 1292 if (!wq->group) { 1293 dev_dbg(dev, "wq %d not attached to group.\n", wq->id); 1294 idxd->cmd_status = IDXD_SCMD_WQ_NO_GRP; 1295 goto err; 1296 } 1297 1298 if (strlen(wq->name) == 0) { 1299 idxd->cmd_status = IDXD_SCMD_WQ_NO_NAME; 1300 dev_dbg(dev, "wq %d name not set.\n", wq->id); 1301 goto err; 1302 } 1303 1304 /* Shared WQ checks */ 1305 if (wq_shared(wq)) { 1306 if (!wq_shared_supported(wq)) { 1307 idxd->cmd_status = IDXD_SCMD_WQ_NO_SVM; 1308 dev_dbg(dev, "PASID not enabled and shared wq.\n"); 1309 goto err; 1310 } 1311 /* 1312 * Shared wq with the threshold set to 0 means the user 1313 * did not set the threshold or transitioned from a 1314 * dedicated wq but did not set threshold. A value 1315 * of 0 would effectively disable the shared wq. The 1316 * driver does not allow a value of 0 to be set for 1317 * threshold via sysfs. 1318 */ 1319 if (wq->threshold == 0) { 1320 idxd->cmd_status = IDXD_SCMD_WQ_NO_THRESH; 1321 dev_dbg(dev, "Shared wq and threshold 0.\n"); 1322 goto err; 1323 } 1324 } 1325 1326 /* 1327 * In the event that the WQ is configurable for pasid and priv bits. 1328 * For kernel wq, the driver should setup the pasid, pasid_en, and priv bit. 1329 * However, for non-kernel wq, the driver should only set the pasid_en bit for 1330 * shared wq. A dedicated wq that is not 'kernel' type will configure pasid and 1331 * pasid_en later on so there is no need to setup. 1332 */ 1333 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) { 1334 int priv = 0; 1335 1336 if (wq_pasid_enabled(wq)) { 1337 if (is_idxd_wq_kernel(wq) || wq_shared(wq)) { 1338 u32 pasid = wq_dedicated(wq) ? idxd->pasid : 0; 1339 1340 __idxd_wq_set_pasid_locked(wq, pasid); 1341 } 1342 } 1343 1344 if (is_idxd_wq_kernel(wq)) 1345 priv = 1; 1346 __idxd_wq_set_priv_locked(wq, priv); 1347 } 1348 1349 rc = 0; 1350 spin_lock(&idxd->dev_lock); 1351 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) 1352 rc = idxd_device_config(idxd); 1353 spin_unlock(&idxd->dev_lock); 1354 if (rc < 0) { 1355 dev_dbg(dev, "Writing wq %d config failed: %d\n", wq->id, rc); 1356 goto err; 1357 } 1358 1359 rc = idxd_wq_enable(wq); 1360 if (rc < 0) { 1361 dev_dbg(dev, "wq %d enabling failed: %d\n", wq->id, rc); 1362 goto err; 1363 } 1364 1365 rc = idxd_wq_map_portal(wq); 1366 if (rc < 0) { 1367 idxd->cmd_status = IDXD_SCMD_WQ_PORTAL_ERR; 1368 dev_dbg(dev, "wq %d portal mapping failed: %d\n", wq->id, rc); 1369 goto err_map_portal; 1370 } 1371 1372 wq->client_count = 0; 1373 1374 rc = idxd_wq_request_irq(wq); 1375 if (rc < 0) { 1376 idxd->cmd_status = IDXD_SCMD_WQ_IRQ_ERR; 1377 dev_dbg(dev, "WQ %d irq setup failed: %d\n", wq->id, rc); 1378 goto err_irq; 1379 } 1380 1381 rc = idxd_wq_alloc_resources(wq); 1382 if (rc < 0) { 1383 idxd->cmd_status = IDXD_SCMD_WQ_RES_ALLOC_ERR; 1384 dev_dbg(dev, "WQ resource alloc failed\n"); 1385 goto err_res_alloc; 1386 } 1387 1388 rc = idxd_wq_init_percpu_ref(wq); 1389 if (rc < 0) { 1390 idxd->cmd_status = IDXD_SCMD_PERCPU_ERR; 1391 dev_dbg(dev, "percpu_ref setup failed\n"); 1392 goto err_ref; 1393 } 1394 1395 return 0; 1396 1397 err_ref: 1398 idxd_wq_free_resources(wq); 1399 err_res_alloc: 1400 idxd_wq_free_irq(wq); 1401 err_irq: 1402 idxd_wq_unmap_portal(wq); 1403 err_map_portal: 1404 if (idxd_wq_disable(wq, false)) 1405 dev_dbg(dev, "wq %s disable failed\n", dev_name(wq_confdev(wq))); 1406 err: 1407 return rc; 1408 } 1409 1410 void drv_disable_wq(struct idxd_wq *wq) 1411 { 1412 struct idxd_device *idxd = wq->idxd; 1413 struct device *dev = &idxd->pdev->dev; 1414 1415 lockdep_assert_held(&wq->wq_lock); 1416 1417 if (idxd_wq_refcount(wq)) 1418 dev_warn(dev, "Clients has claim on wq %d: %d\n", 1419 wq->id, idxd_wq_refcount(wq)); 1420 1421 idxd_wq_unmap_portal(wq); 1422 idxd_wq_drain(wq); 1423 idxd_wq_free_irq(wq); 1424 idxd_wq_reset(wq); 1425 idxd_wq_free_resources(wq); 1426 percpu_ref_exit(&wq->wq_active); 1427 wq->type = IDXD_WQT_NONE; 1428 wq->client_count = 0; 1429 } 1430 1431 int idxd_device_drv_probe(struct idxd_dev *idxd_dev) 1432 { 1433 struct idxd_device *idxd = idxd_dev_to_idxd(idxd_dev); 1434 int rc = 0; 1435 1436 /* 1437 * Device should be in disabled state for the idxd_drv to load. If it's in 1438 * enabled state, then the device was altered outside of driver's control. 1439 * If the state is in halted state, then we don't want to proceed. 1440 */ 1441 if (idxd->state != IDXD_DEV_DISABLED) { 1442 idxd->cmd_status = IDXD_SCMD_DEV_ENABLED; 1443 return -ENXIO; 1444 } 1445 1446 /* Device configuration */ 1447 spin_lock(&idxd->dev_lock); 1448 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) 1449 rc = idxd_device_config(idxd); 1450 spin_unlock(&idxd->dev_lock); 1451 if (rc < 0) 1452 return -ENXIO; 1453 1454 /* Start device */ 1455 rc = idxd_device_enable(idxd); 1456 if (rc < 0) 1457 return rc; 1458 1459 /* Setup DMA device without channels */ 1460 rc = idxd_register_dma_device(idxd); 1461 if (rc < 0) { 1462 idxd_device_disable(idxd); 1463 idxd->cmd_status = IDXD_SCMD_DEV_DMA_ERR; 1464 return rc; 1465 } 1466 1467 idxd->cmd_status = 0; 1468 return 0; 1469 } 1470 1471 void idxd_device_drv_remove(struct idxd_dev *idxd_dev) 1472 { 1473 struct device *dev = &idxd_dev->conf_dev; 1474 struct idxd_device *idxd = idxd_dev_to_idxd(idxd_dev); 1475 int i; 1476 1477 for (i = 0; i < idxd->max_wqs; i++) { 1478 struct idxd_wq *wq = idxd->wqs[i]; 1479 struct device *wq_dev = wq_confdev(wq); 1480 1481 if (wq->state == IDXD_WQ_DISABLED) 1482 continue; 1483 dev_warn(dev, "Active wq %d on disable %s.\n", i, dev_name(wq_dev)); 1484 device_release_driver(wq_dev); 1485 } 1486 1487 idxd_unregister_dma_device(idxd); 1488 idxd_device_disable(idxd); 1489 if (test_bit(IDXD_FLAG_CONFIGURABLE, &idxd->flags)) 1490 idxd_device_reset(idxd); 1491 } 1492 1493 static enum idxd_dev_type dev_types[] = { 1494 IDXD_DEV_DSA, 1495 IDXD_DEV_IAX, 1496 IDXD_DEV_NONE, 1497 }; 1498 1499 struct idxd_device_driver idxd_drv = { 1500 .type = dev_types, 1501 .probe = idxd_device_drv_probe, 1502 .remove = idxd_device_drv_remove, 1503 .name = "idxd", 1504 }; 1505 EXPORT_SYMBOL_GPL(idxd_drv); 1506