1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #include "xe_device.h" 7 8 #include <linux/aperture.h> 9 #include <linux/delay.h> 10 #include <linux/fault-inject.h> 11 #include <linux/units.h> 12 13 #include <drm/drm_atomic_helper.h> 14 #include <drm/drm_client.h> 15 #include <drm/drm_gem_ttm_helper.h> 16 #include <drm/drm_ioctl.h> 17 #include <drm/drm_managed.h> 18 #include <drm/drm_pagemap_util.h> 19 #include <drm/drm_print.h> 20 #include <uapi/drm/xe_drm.h> 21 22 #include "display/xe_display.h" 23 #include "instructions/xe_gpu_commands.h" 24 #include "regs/xe_gt_regs.h" 25 #include "regs/xe_regs.h" 26 #include "xe_bo.h" 27 #include "xe_bo_evict.h" 28 #include "xe_debugfs.h" 29 #include "xe_defaults.h" 30 #include "xe_devcoredump.h" 31 #include "xe_device_sysfs.h" 32 #include "xe_dma_buf.h" 33 #include "xe_drm_client.h" 34 #include "xe_drv.h" 35 #include "xe_exec.h" 36 #include "xe_exec_queue.h" 37 #include "xe_force_wake.h" 38 #include "xe_ggtt.h" 39 #include "xe_gt.h" 40 #include "xe_gt_mcr.h" 41 #include "xe_gt_printk.h" 42 #include "xe_gt_sriov_vf.h" 43 #include "xe_guc.h" 44 #include "xe_guc_pc.h" 45 #include "xe_hw_engine_group.h" 46 #include "xe_hwmon.h" 47 #include "xe_i2c.h" 48 #include "xe_irq.h" 49 #include "xe_late_bind_fw.h" 50 #include "xe_mmio.h" 51 #include "xe_module.h" 52 #include "xe_nvm.h" 53 #include "xe_oa.h" 54 #include "xe_observation.h" 55 #include "xe_pagefault.h" 56 #include "xe_pat.h" 57 #include "xe_pcode.h" 58 #include "xe_pm.h" 59 #include "xe_pmu.h" 60 #include "xe_psmi.h" 61 #include "xe_pxp.h" 62 #include "xe_query.h" 63 #include "xe_shrinker.h" 64 #include "xe_soc_remapper.h" 65 #include "xe_survivability_mode.h" 66 #include "xe_sriov.h" 67 #include "xe_svm.h" 68 #include "xe_tile.h" 69 #include "xe_ttm_stolen_mgr.h" 70 #include "xe_ttm_sys_mgr.h" 71 #include "xe_vm.h" 72 #include "xe_vm_madvise.h" 73 #include "xe_vram.h" 74 #include "xe_vram_types.h" 75 #include "xe_vsec.h" 76 #include "xe_wait_user_fence.h" 77 #include "xe_wa.h" 78 79 #include <generated/xe_device_wa_oob.h> 80 #include <generated/xe_wa_oob.h> 81 82 static int xe_file_open(struct drm_device *dev, struct drm_file *file) 83 { 84 struct xe_device *xe = to_xe_device(dev); 85 struct xe_drm_client *client; 86 struct xe_file *xef; 87 int ret = -ENOMEM; 88 struct task_struct *task = NULL; 89 90 xef = kzalloc_obj(*xef); 91 if (!xef) 92 return ret; 93 94 client = xe_drm_client_alloc(); 95 if (!client) { 96 kfree(xef); 97 return ret; 98 } 99 100 xef->drm = file; 101 xef->client = client; 102 xef->xe = xe; 103 104 mutex_init(&xef->vm.lock); 105 xa_init_flags(&xef->vm.xa, XA_FLAGS_ALLOC1); 106 107 mutex_init(&xef->exec_queue.lock); 108 xa_init_flags(&xef->exec_queue.xa, XA_FLAGS_ALLOC1); 109 110 file->driver_priv = xef; 111 kref_init(&xef->refcount); 112 113 task = get_pid_task(rcu_access_pointer(file->pid), PIDTYPE_PID); 114 if (task) { 115 xef->process_name = kstrdup(task->comm, GFP_KERNEL); 116 xef->pid = task->pid; 117 put_task_struct(task); 118 } 119 120 return 0; 121 } 122 123 static void xe_file_destroy(struct kref *ref) 124 { 125 struct xe_file *xef = container_of(ref, struct xe_file, refcount); 126 127 xa_destroy(&xef->exec_queue.xa); 128 mutex_destroy(&xef->exec_queue.lock); 129 xa_destroy(&xef->vm.xa); 130 mutex_destroy(&xef->vm.lock); 131 132 xe_drm_client_put(xef->client); 133 kfree(xef->process_name); 134 kfree(xef); 135 } 136 137 /** 138 * xe_file_get() - Take a reference to the xe file object 139 * @xef: Pointer to the xe file 140 * 141 * Anyone with a pointer to xef must take a reference to the xe file 142 * object using this call. 143 * 144 * Return: xe file pointer 145 */ 146 struct xe_file *xe_file_get(struct xe_file *xef) 147 { 148 kref_get(&xef->refcount); 149 return xef; 150 } 151 152 /** 153 * xe_file_put() - Drop a reference to the xe file object 154 * @xef: Pointer to the xe file 155 * 156 * Used to drop reference to the xef object 157 */ 158 void xe_file_put(struct xe_file *xef) 159 { 160 kref_put(&xef->refcount, xe_file_destroy); 161 } 162 163 static void xe_file_close(struct drm_device *dev, struct drm_file *file) 164 { 165 struct xe_device *xe = to_xe_device(dev); 166 struct xe_file *xef = file->driver_priv; 167 struct xe_vm *vm; 168 struct xe_exec_queue *q; 169 unsigned long idx; 170 171 guard(xe_pm_runtime)(xe); 172 173 /* 174 * No need for exec_queue.lock here as there is no contention for it 175 * when FD is closing as IOCTLs presumably can't be modifying the 176 * xarray. Taking exec_queue.lock here causes undue dependency on 177 * vm->lock taken during xe_exec_queue_kill(). 178 */ 179 xa_for_each(&xef->exec_queue.xa, idx, q) { 180 if (q->vm && q->hwe->hw_engine_group) 181 xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q); 182 xe_exec_queue_kill(q); 183 xe_exec_queue_put(q); 184 } 185 xa_for_each(&xef->vm.xa, idx, vm) 186 xe_vm_close_and_put(vm); 187 188 xe_file_put(xef); 189 } 190 191 static const struct drm_ioctl_desc xe_ioctls[] = { 192 DRM_IOCTL_DEF_DRV(XE_DEVICE_QUERY, xe_query_ioctl, DRM_RENDER_ALLOW), 193 DRM_IOCTL_DEF_DRV(XE_GEM_CREATE, xe_gem_create_ioctl, DRM_RENDER_ALLOW), 194 DRM_IOCTL_DEF_DRV(XE_GEM_MMAP_OFFSET, xe_gem_mmap_offset_ioctl, 195 DRM_RENDER_ALLOW), 196 DRM_IOCTL_DEF_DRV(XE_VM_CREATE, xe_vm_create_ioctl, DRM_RENDER_ALLOW), 197 DRM_IOCTL_DEF_DRV(XE_VM_DESTROY, xe_vm_destroy_ioctl, DRM_RENDER_ALLOW), 198 DRM_IOCTL_DEF_DRV(XE_VM_BIND, xe_vm_bind_ioctl, DRM_RENDER_ALLOW), 199 DRM_IOCTL_DEF_DRV(XE_EXEC, xe_exec_ioctl, DRM_RENDER_ALLOW), 200 DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_CREATE, xe_exec_queue_create_ioctl, 201 DRM_RENDER_ALLOW), 202 DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_DESTROY, xe_exec_queue_destroy_ioctl, 203 DRM_RENDER_ALLOW), 204 DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_GET_PROPERTY, xe_exec_queue_get_property_ioctl, 205 DRM_RENDER_ALLOW), 206 DRM_IOCTL_DEF_DRV(XE_WAIT_USER_FENCE, xe_wait_user_fence_ioctl, 207 DRM_RENDER_ALLOW), 208 DRM_IOCTL_DEF_DRV(XE_OBSERVATION, xe_observation_ioctl, DRM_RENDER_ALLOW), 209 DRM_IOCTL_DEF_DRV(XE_MADVISE, xe_vm_madvise_ioctl, DRM_RENDER_ALLOW), 210 DRM_IOCTL_DEF_DRV(XE_VM_QUERY_MEM_RANGE_ATTRS, xe_vm_query_vmas_attrs_ioctl, 211 DRM_RENDER_ALLOW), 212 DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_SET_PROPERTY, xe_exec_queue_set_property_ioctl, 213 DRM_RENDER_ALLOW), 214 DRM_IOCTL_DEF_DRV(XE_VM_GET_PROPERTY, xe_vm_get_property_ioctl, 215 DRM_RENDER_ALLOW), 216 }; 217 218 static long xe_drm_ioctl(struct file *file, unsigned int cmd, unsigned long arg) 219 { 220 struct drm_file *file_priv = file->private_data; 221 struct xe_device *xe = to_xe_device(file_priv->minor->dev); 222 long ret; 223 224 if (xe_device_wedged(xe)) 225 return -ECANCELED; 226 227 ACQUIRE(xe_pm_runtime_ioctl, pm)(xe); 228 ret = ACQUIRE_ERR(xe_pm_runtime_ioctl, &pm); 229 if (ret >= 0) 230 ret = drm_ioctl(file, cmd, arg); 231 232 return ret; 233 } 234 235 #ifdef CONFIG_COMPAT 236 static long xe_drm_compat_ioctl(struct file *file, unsigned int cmd, unsigned long arg) 237 { 238 struct drm_file *file_priv = file->private_data; 239 struct xe_device *xe = to_xe_device(file_priv->minor->dev); 240 long ret; 241 242 if (xe_device_wedged(xe)) 243 return -ECANCELED; 244 245 ACQUIRE(xe_pm_runtime_ioctl, pm)(xe); 246 ret = ACQUIRE_ERR(xe_pm_runtime_ioctl, &pm); 247 if (ret >= 0) 248 ret = drm_compat_ioctl(file, cmd, arg); 249 250 return ret; 251 } 252 #else 253 /* similarly to drm_compat_ioctl, let's it be assigned to .compat_ioct unconditionally */ 254 #define xe_drm_compat_ioctl NULL 255 #endif 256 257 static void barrier_open(struct vm_area_struct *vma) 258 { 259 drm_dev_get(vma->vm_private_data); 260 } 261 262 static void barrier_close(struct vm_area_struct *vma) 263 { 264 drm_dev_put(vma->vm_private_data); 265 } 266 267 static void barrier_release_dummy_page(struct drm_device *dev, void *res) 268 { 269 struct page *dummy_page = (struct page *)res; 270 271 __free_page(dummy_page); 272 } 273 274 static vm_fault_t barrier_fault(struct vm_fault *vmf) 275 { 276 struct drm_device *dev = vmf->vma->vm_private_data; 277 struct vm_area_struct *vma = vmf->vma; 278 vm_fault_t ret = VM_FAULT_NOPAGE; 279 pgprot_t prot; 280 int idx; 281 282 prot = vm_get_page_prot(vma->vm_flags); 283 284 if (drm_dev_enter(dev, &idx)) { 285 unsigned long pfn; 286 287 #define LAST_DB_PAGE_OFFSET 0x7ff001 288 pfn = PHYS_PFN(pci_resource_start(to_pci_dev(dev->dev), 0) + 289 LAST_DB_PAGE_OFFSET); 290 ret = vmf_insert_pfn_prot(vma, vma->vm_start, pfn, 291 pgprot_noncached(prot)); 292 drm_dev_exit(idx); 293 } else { 294 struct page *page; 295 296 /* Allocate new dummy page to map all the VA range in this VMA to it*/ 297 page = alloc_page(GFP_KERNEL | __GFP_ZERO); 298 if (!page) 299 return VM_FAULT_OOM; 300 301 /* Set the page to be freed using drmm release action */ 302 if (drmm_add_action_or_reset(dev, barrier_release_dummy_page, page)) 303 return VM_FAULT_OOM; 304 305 ret = vmf_insert_pfn_prot(vma, vma->vm_start, page_to_pfn(page), 306 prot); 307 } 308 309 return ret; 310 } 311 312 static const struct vm_operations_struct vm_ops_barrier = { 313 .open = barrier_open, 314 .close = barrier_close, 315 .fault = barrier_fault, 316 }; 317 318 static int xe_pci_barrier_mmap(struct file *filp, 319 struct vm_area_struct *vma) 320 { 321 struct drm_file *priv = filp->private_data; 322 struct drm_device *dev = priv->minor->dev; 323 struct xe_device *xe = to_xe_device(dev); 324 325 if (!IS_DGFX(xe)) 326 return -EINVAL; 327 328 if (vma->vm_end - vma->vm_start > SZ_4K) 329 return -EINVAL; 330 331 if (is_cow_mapping(vma->vm_flags)) 332 return -EINVAL; 333 334 if (vma->vm_flags & (VM_READ | VM_EXEC)) 335 return -EINVAL; 336 337 vm_flags_clear(vma, VM_MAYREAD | VM_MAYEXEC); 338 vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP | VM_IO); 339 vma->vm_ops = &vm_ops_barrier; 340 vma->vm_private_data = dev; 341 drm_dev_get(vma->vm_private_data); 342 343 return 0; 344 } 345 346 static int xe_mmap(struct file *filp, struct vm_area_struct *vma) 347 { 348 struct drm_file *priv = filp->private_data; 349 struct drm_device *dev = priv->minor->dev; 350 351 if (drm_dev_is_unplugged(dev)) 352 return -ENODEV; 353 354 switch (vma->vm_pgoff) { 355 case XE_PCI_BARRIER_MMAP_OFFSET >> XE_PTE_SHIFT: 356 return xe_pci_barrier_mmap(filp, vma); 357 } 358 359 return drm_gem_mmap(filp, vma); 360 } 361 362 static const struct file_operations xe_driver_fops = { 363 .owner = THIS_MODULE, 364 .open = drm_open, 365 .release = drm_release_noglobal, 366 .unlocked_ioctl = xe_drm_ioctl, 367 .mmap = xe_mmap, 368 .poll = drm_poll, 369 .read = drm_read, 370 .compat_ioctl = xe_drm_compat_ioctl, 371 .llseek = noop_llseek, 372 #ifdef CONFIG_PROC_FS 373 .show_fdinfo = drm_show_fdinfo, 374 #endif 375 .fop_flags = FOP_UNSIGNED_OFFSET, 376 }; 377 378 /** 379 * xe_is_xe_file() - Is the file an xe device file? 380 * @file: The file. 381 * 382 * Checks whether the file is opened against 383 * an xe device. 384 * 385 * Return: %true if an xe file, %false if not. 386 */ 387 bool xe_is_xe_file(const struct file *file) 388 { 389 return file->f_op == &xe_driver_fops; 390 } 391 392 static struct drm_driver driver = { 393 .driver_features = 394 DRIVER_GEM | 395 DRIVER_RENDER | DRIVER_SYNCOBJ | 396 DRIVER_SYNCOBJ_TIMELINE | DRIVER_GEM_GPUVA, 397 .open = xe_file_open, 398 .postclose = xe_file_close, 399 400 .gem_prime_import = xe_gem_prime_import, 401 402 .dumb_create = xe_bo_dumb_create, 403 .dumb_map_offset = drm_gem_ttm_dumb_map_offset, 404 #ifdef CONFIG_PROC_FS 405 .show_fdinfo = xe_drm_client_fdinfo, 406 #endif 407 .ioctls = xe_ioctls, 408 .num_ioctls = ARRAY_SIZE(xe_ioctls), 409 .fops = &xe_driver_fops, 410 .name = DRIVER_NAME, 411 .desc = DRIVER_DESC, 412 .major = DRIVER_MAJOR, 413 .minor = DRIVER_MINOR, 414 .patchlevel = DRIVER_PATCHLEVEL, 415 }; 416 417 static void xe_device_destroy(struct drm_device *dev, void *dummy) 418 { 419 struct xe_device *xe = to_xe_device(dev); 420 421 xe_bo_dev_fini(&xe->bo_device); 422 423 if (xe->preempt_fence_wq) 424 destroy_workqueue(xe->preempt_fence_wq); 425 426 if (xe->ordered_wq) 427 destroy_workqueue(xe->ordered_wq); 428 429 if (xe->unordered_wq) 430 destroy_workqueue(xe->unordered_wq); 431 432 if (xe->destroy_wq) 433 destroy_workqueue(xe->destroy_wq); 434 435 ttm_device_fini(&xe->ttm); 436 } 437 438 struct xe_device *xe_device_create(struct pci_dev *pdev, 439 const struct pci_device_id *ent) 440 { 441 struct xe_device *xe; 442 int err; 443 444 xe_display_driver_set_hooks(&driver); 445 446 err = aperture_remove_conflicting_pci_devices(pdev, driver.name); 447 if (err) 448 return ERR_PTR(err); 449 450 xe = devm_drm_dev_alloc(&pdev->dev, &driver, struct xe_device, drm); 451 if (IS_ERR(xe)) 452 return xe; 453 454 err = ttm_device_init(&xe->ttm, &xe_ttm_funcs, xe->drm.dev, 455 xe->drm.anon_inode->i_mapping, 456 xe->drm.vma_offset_manager, 0); 457 if (WARN_ON(err)) 458 return ERR_PTR(err); 459 460 xe_bo_dev_init(&xe->bo_device); 461 err = drmm_add_action_or_reset(&xe->drm, xe_device_destroy, NULL); 462 if (err) 463 return ERR_PTR(err); 464 465 err = xe_shrinker_create(xe); 466 if (err) 467 return ERR_PTR(err); 468 469 xe->info.devid = pdev->device; 470 xe->info.revid = pdev->revision; 471 xe->info.force_execlist = xe_modparam.force_execlist; 472 xe->atomic_svm_timeslice_ms = 5; 473 xe->min_run_period_lr_ms = 5; 474 475 err = xe_irq_init(xe); 476 if (err) 477 return ERR_PTR(err); 478 479 xe_validation_device_init(&xe->val); 480 481 init_waitqueue_head(&xe->ufence_wq); 482 483 init_rwsem(&xe->usm.lock); 484 485 err = xe_pagemap_shrinker_create(xe); 486 if (err) 487 return ERR_PTR(err); 488 489 xa_init_flags(&xe->usm.asid_to_vm, XA_FLAGS_ALLOC); 490 491 if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) { 492 /* Trigger a large asid and an early asid wrap. */ 493 u32 asid; 494 495 BUILD_BUG_ON(XE_MAX_ASID < 2); 496 err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, NULL, 497 XA_LIMIT(XE_MAX_ASID - 2, XE_MAX_ASID - 1), 498 &xe->usm.next_asid, GFP_KERNEL); 499 drm_WARN_ON(&xe->drm, err); 500 if (err >= 0) 501 xa_erase(&xe->usm.asid_to_vm, asid); 502 } 503 504 err = xe_bo_pinned_init(xe); 505 if (err) 506 return ERR_PTR(err); 507 508 xe->preempt_fence_wq = alloc_ordered_workqueue("xe-preempt-fence-wq", 509 WQ_MEM_RECLAIM); 510 xe->ordered_wq = alloc_ordered_workqueue("xe-ordered-wq", 0); 511 xe->unordered_wq = alloc_workqueue("xe-unordered-wq", WQ_PERCPU, 0); 512 xe->destroy_wq = alloc_workqueue("xe-destroy-wq", WQ_PERCPU, 0); 513 if (!xe->ordered_wq || !xe->unordered_wq || 514 !xe->preempt_fence_wq || !xe->destroy_wq) { 515 /* 516 * Cleanup done in xe_device_destroy via 517 * drmm_add_action_or_reset register above 518 */ 519 drm_err(&xe->drm, "Failed to allocate xe workqueues\n"); 520 return ERR_PTR(-ENOMEM); 521 } 522 523 err = drmm_mutex_init(&xe->drm, &xe->pmt.lock); 524 if (err) 525 return ERR_PTR(err); 526 527 return xe; 528 } 529 ALLOW_ERROR_INJECTION(xe_device_create, ERRNO); /* See xe_pci_probe() */ 530 531 static bool xe_driver_flr_disabled(struct xe_device *xe) 532 { 533 if (IS_SRIOV_VF(xe)) 534 return true; 535 536 if (xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL_PROTECTED) & DRIVERINT_FLR_DIS) { 537 drm_info(&xe->drm, "Driver-FLR disabled by BIOS\n"); 538 return true; 539 } 540 541 return false; 542 } 543 544 /* 545 * The driver-initiated FLR is the highest level of reset that we can trigger 546 * from within the driver. It is different from the PCI FLR in that it doesn't 547 * fully reset the SGUnit and doesn't modify the PCI config space and therefore 548 * it doesn't require a re-enumeration of the PCI BARs. However, the 549 * driver-initiated FLR does still cause a reset of both GT and display and a 550 * memory wipe of local and stolen memory, so recovery would require a full HW 551 * re-init and saving/restoring (or re-populating) the wiped memory. Since we 552 * perform the FLR as the very last action before releasing access to the HW 553 * during the driver release flow, we don't attempt recovery at all, because 554 * if/when a new instance of Xe is bound to the device it will do a full 555 * re-init anyway. 556 */ 557 static void __xe_driver_flr(struct xe_device *xe) 558 { 559 const unsigned int flr_timeout = 3 * USEC_PER_SEC; /* specs recommend a 3s wait */ 560 struct xe_mmio *mmio = xe_root_tile_mmio(xe); 561 int ret; 562 563 drm_dbg(&xe->drm, "Triggering Driver-FLR\n"); 564 565 /* 566 * Make sure any pending FLR requests have cleared by waiting for the 567 * FLR trigger bit to go to zero. Also clear GU_DEBUG's DRIVERFLR_STATUS 568 * to make sure it's not still set from a prior attempt (it's a write to 569 * clear bit). 570 * Note that we should never be in a situation where a previous attempt 571 * is still pending (unless the HW is totally dead), but better to be 572 * safe in case something unexpected happens 573 */ 574 ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false); 575 if (ret) { 576 drm_err(&xe->drm, "Driver-FLR-prepare wait for ready failed! %d\n", ret); 577 return; 578 } 579 xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS); 580 581 /* Trigger the actual Driver-FLR */ 582 xe_mmio_rmw32(mmio, GU_CNTL, 0, DRIVERFLR); 583 584 /* Wait for hardware teardown to complete */ 585 ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false); 586 if (ret) { 587 drm_err(&xe->drm, "Driver-FLR-teardown wait completion failed! %d\n", ret); 588 return; 589 } 590 591 /* Wait for hardware/firmware re-init to complete */ 592 ret = xe_mmio_wait32(mmio, GU_DEBUG, DRIVERFLR_STATUS, DRIVERFLR_STATUS, 593 flr_timeout, NULL, false); 594 if (ret) { 595 drm_err(&xe->drm, "Driver-FLR-reinit wait completion failed! %d\n", ret); 596 return; 597 } 598 599 /* Clear sticky completion status */ 600 xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS); 601 } 602 603 static void xe_driver_flr(struct xe_device *xe) 604 { 605 if (xe_driver_flr_disabled(xe)) 606 return; 607 608 __xe_driver_flr(xe); 609 } 610 611 static void xe_driver_flr_fini(void *arg) 612 { 613 struct xe_device *xe = arg; 614 615 if (xe->needs_flr_on_fini) 616 xe_driver_flr(xe); 617 } 618 619 static void xe_device_sanitize(void *arg) 620 { 621 struct xe_device *xe = arg; 622 struct xe_gt *gt; 623 u8 id; 624 625 for_each_gt(gt, xe, id) 626 xe_gt_sanitize(gt); 627 } 628 629 static int xe_set_dma_info(struct xe_device *xe) 630 { 631 unsigned int mask_size = xe->info.dma_mask_size; 632 int err; 633 634 dma_set_max_seg_size(xe->drm.dev, xe_sg_segment_size(xe->drm.dev)); 635 636 err = dma_set_mask(xe->drm.dev, DMA_BIT_MASK(mask_size)); 637 if (err) 638 goto mask_err; 639 640 err = dma_set_coherent_mask(xe->drm.dev, DMA_BIT_MASK(mask_size)); 641 if (err) 642 goto mask_err; 643 644 return 0; 645 646 mask_err: 647 drm_err(&xe->drm, "Can't set DMA mask/consistent mask (%d)\n", err); 648 return err; 649 } 650 651 static void assert_lmem_ready(struct xe_device *xe) 652 { 653 if (!IS_DGFX(xe) || IS_SRIOV_VF(xe)) 654 return; 655 656 xe_assert(xe, xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL) & 657 LMEM_INIT); 658 } 659 660 static void vf_update_device_info(struct xe_device *xe) 661 { 662 xe_assert(xe, IS_SRIOV_VF(xe)); 663 /* disable features that are not available/applicable to VFs */ 664 xe->info.probe_display = 0; 665 xe->info.has_heci_cscfi = 0; 666 xe->info.has_heci_gscfi = 0; 667 xe->info.has_late_bind = 0; 668 xe->info.skip_guc_pc = 1; 669 xe->info.skip_pcode = 1; 670 } 671 672 static int xe_device_vram_alloc(struct xe_device *xe) 673 { 674 struct xe_vram_region *vram; 675 676 if (!IS_DGFX(xe)) 677 return 0; 678 679 vram = drmm_kzalloc(&xe->drm, sizeof(*vram), GFP_KERNEL); 680 if (!vram) 681 return -ENOMEM; 682 683 xe->mem.vram = vram; 684 return 0; 685 } 686 687 /** 688 * xe_device_probe_early: Device early probe 689 * @xe: xe device instance 690 * 691 * Initialize MMIO resources that don't require any 692 * knowledge about tile count. Also initialize pcode and 693 * check vram initialization on root tile. 694 * 695 * Return: 0 on success, error code on failure 696 */ 697 int xe_device_probe_early(struct xe_device *xe) 698 { 699 int err; 700 701 xe_wa_device_init(xe); 702 xe_wa_process_device_oob(xe); 703 704 err = xe_mmio_probe_early(xe); 705 if (err) 706 return err; 707 708 xe_sriov_probe_early(xe); 709 710 if (IS_SRIOV_VF(xe)) 711 vf_update_device_info(xe); 712 713 /* 714 * Check for pcode uncore_init status to confirm if the SoC 715 * initialization is complete. Until done, any MMIO or lmem access from 716 * the driver will be blocked 717 */ 718 err = xe_pcode_probe_early(xe); 719 if (err || xe_survivability_mode_is_requested(xe)) { 720 int save_err = err; 721 722 /* 723 * Try to leave device in survivability mode if device is 724 * possible, but still return the previous error for error 725 * propagation 726 */ 727 err = xe_survivability_mode_boot_enable(xe); 728 if (err) 729 return err; 730 731 return save_err; 732 } 733 734 /* 735 * Make sure the lmem is initialized and ready to use. xe_pcode_ready() 736 * is flagged after full initialization is complete. Assert if lmem is 737 * not initialized. 738 */ 739 assert_lmem_ready(xe); 740 741 xe->wedged.mode = xe_device_validate_wedged_mode(xe, xe_modparam.wedged_mode) ? 742 XE_DEFAULT_WEDGED_MODE : xe_modparam.wedged_mode; 743 drm_dbg(&xe->drm, "wedged_mode: setting mode (%u) %s\n", 744 xe->wedged.mode, xe_wedged_mode_to_string(xe->wedged.mode)); 745 746 err = xe_device_vram_alloc(xe); 747 if (err) 748 return err; 749 750 return 0; 751 } 752 ALLOW_ERROR_INJECTION(xe_device_probe_early, ERRNO); /* See xe_pci_probe() */ 753 754 static int probe_has_flat_ccs(struct xe_device *xe) 755 { 756 struct xe_gt *gt; 757 u32 reg; 758 759 /* Always enabled/disabled, no runtime check to do */ 760 if (GRAPHICS_VER(xe) < 20 || !xe->info.has_flat_ccs || IS_SRIOV_VF(xe)) 761 return 0; 762 763 gt = xe_root_mmio_gt(xe); 764 if (!gt) 765 return 0; 766 767 CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT); 768 if (!fw_ref.domains) 769 return -ETIMEDOUT; 770 771 reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_LOWER); 772 xe->info.has_flat_ccs = (reg & XE2_FLAT_CCS_ENABLE); 773 774 if (!xe->info.has_flat_ccs) 775 drm_dbg(&xe->drm, 776 "Flat CCS has been disabled in bios, May lead to performance impact"); 777 778 return 0; 779 } 780 781 /* 782 * Detect if the driver is being run on pre-production hardware. We don't 783 * keep workarounds for pre-production hardware long term, so print an 784 * error and add taint if we're being loaded on a pre-production platform 785 * for which the pre-prod workarounds have already been removed. 786 * 787 * The general policy is that we'll remove any workarounds that only apply to 788 * pre-production hardware around the time force_probe restrictions are lifted 789 * for a platform of the next major IP generation (for example, Xe2 pre-prod 790 * workarounds should be removed around the time the first Xe3 platforms have 791 * force_probe lifted). 792 */ 793 static void detect_preproduction_hw(struct xe_device *xe) 794 { 795 struct xe_gt *gt; 796 int id; 797 798 /* 799 * SR-IOV VFs don't have access to the FUSE2 register, so we can't 800 * check pre-production status there. But the host OS will notice 801 * and report the pre-production status, which should be enough to 802 * help us catch mistaken use of pre-production hardware. 803 */ 804 if (IS_SRIOV_VF(xe)) 805 return; 806 807 /* 808 * The "SW_CAP" fuse contains a bit indicating whether the device is a 809 * production or pre-production device. This fuse is reflected through 810 * the GT "FUSE2" register, even though the contents of the fuse are 811 * not GT-specific. Every GT's reflection of this fuse should show the 812 * same value, so we'll just use the first available GT for lookup. 813 */ 814 for_each_gt(gt, xe, id) 815 break; 816 817 if (!gt) 818 return; 819 820 CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT); 821 if (!xe_force_wake_ref_has_domain(fw_ref.domains, XE_FW_GT)) { 822 xe_gt_err(gt, "Forcewake failure; cannot determine production/pre-production hw status.\n"); 823 return; 824 } 825 826 if (xe_mmio_read32(>->mmio, FUSE2) & PRODUCTION_HW) 827 return; 828 829 xe_info(xe, "Pre-production hardware detected.\n"); 830 if (!xe->info.has_pre_prod_wa) { 831 xe_err(xe, "Pre-production workarounds for this platform have already been removed.\n"); 832 add_taint(TAINT_MACHINE_CHECK, LOCKDEP_STILL_OK); 833 } 834 } 835 836 static void xe_device_wedged_fini(struct drm_device *drm, void *arg) 837 { 838 struct xe_device *xe = arg; 839 840 if (atomic_read(&xe->wedged.flag)) 841 xe_pm_runtime_put(xe); 842 } 843 844 int xe_device_probe(struct xe_device *xe) 845 { 846 struct xe_tile *tile; 847 struct xe_gt *gt; 848 int err; 849 u8 id; 850 851 xe_pat_init_early(xe); 852 853 err = xe_sriov_init(xe); 854 if (err) 855 return err; 856 857 xe->info.mem_region_mask = 1; 858 859 err = xe_set_dma_info(xe); 860 if (err) 861 return err; 862 863 err = xe_mmio_probe_tiles(xe); 864 if (err) 865 return err; 866 867 for_each_gt(gt, xe, id) { 868 err = xe_gt_init_early(gt); 869 if (err) 870 return err; 871 } 872 873 for_each_tile(tile, xe, id) { 874 err = xe_ggtt_init_early(tile->mem.ggtt); 875 if (err) 876 return err; 877 } 878 879 /* 880 * From here on, if a step fails, make sure a Driver-FLR is triggereed 881 */ 882 err = devm_add_action_or_reset(xe->drm.dev, xe_driver_flr_fini, xe); 883 if (err) 884 return err; 885 886 err = probe_has_flat_ccs(xe); 887 if (err) 888 return err; 889 890 err = xe_vram_probe(xe); 891 if (err) 892 return err; 893 894 for_each_tile(tile, xe, id) { 895 err = xe_tile_init_noalloc(tile); 896 if (err) 897 return err; 898 } 899 900 /* 901 * Allow allocations only now to ensure xe_display_init_early() 902 * is the first to allocate, always. 903 */ 904 err = xe_ttm_sys_mgr_init(xe); 905 if (err) 906 return err; 907 908 /* Allocate and map stolen after potential VRAM resize */ 909 err = xe_ttm_stolen_mgr_init(xe); 910 if (err) 911 return err; 912 913 /* 914 * Now that GT is initialized (TTM in particular), 915 * we can try to init display, and inherit the initial fb. 916 * This is the reason the first allocation needs to be done 917 * inside display. 918 */ 919 err = xe_display_init_early(xe); 920 if (err) 921 return err; 922 923 for_each_tile(tile, xe, id) { 924 err = xe_tile_init(tile); 925 if (err) 926 return err; 927 } 928 929 err = xe_irq_install(xe); 930 if (err) 931 return err; 932 933 for_each_gt(gt, xe, id) { 934 err = xe_gt_init(gt); 935 if (err) 936 return err; 937 } 938 939 err = xe_pagefault_init(xe); 940 if (err) 941 return err; 942 943 if (xe->tiles->media_gt && 944 XE_GT_WA(xe->tiles->media_gt, 15015404425_disable)) 945 XE_DEVICE_WA_DISABLE(xe, 15015404425); 946 947 err = xe_devcoredump_init(xe); 948 if (err) 949 return err; 950 951 xe_nvm_init(xe); 952 953 err = xe_soc_remapper_init(xe); 954 if (err) 955 return err; 956 957 err = xe_heci_gsc_init(xe); 958 if (err) 959 return err; 960 961 err = xe_late_bind_init(&xe->late_bind); 962 if (err) 963 return err; 964 965 err = xe_oa_init(xe); 966 if (err) 967 return err; 968 969 err = xe_display_init(xe); 970 if (err) 971 return err; 972 973 err = xe_pxp_init(xe); 974 if (err) 975 return err; 976 977 err = xe_psmi_init(xe); 978 if (err) 979 return err; 980 981 err = drm_dev_register(&xe->drm, 0); 982 if (err) 983 return err; 984 985 xe_display_register(xe); 986 987 err = xe_oa_register(xe); 988 if (err) 989 goto err_unregister_display; 990 991 err = xe_pmu_register(&xe->pmu); 992 if (err) 993 goto err_unregister_display; 994 995 err = xe_device_sysfs_init(xe); 996 if (err) 997 goto err_unregister_display; 998 999 xe_debugfs_register(xe); 1000 1001 err = xe_hwmon_register(xe); 1002 if (err) 1003 goto err_unregister_display; 1004 1005 err = xe_i2c_probe(xe); 1006 if (err) 1007 goto err_unregister_display; 1008 1009 for_each_gt(gt, xe, id) 1010 xe_gt_sanitize_freq(gt); 1011 1012 xe_vsec_init(xe); 1013 1014 err = xe_sriov_init_late(xe); 1015 if (err) 1016 goto err_unregister_display; 1017 1018 detect_preproduction_hw(xe); 1019 1020 err = drmm_add_action_or_reset(&xe->drm, xe_device_wedged_fini, xe); 1021 if (err) 1022 goto err_unregister_display; 1023 1024 return devm_add_action_or_reset(xe->drm.dev, xe_device_sanitize, xe); 1025 1026 err_unregister_display: 1027 xe_display_unregister(xe); 1028 drm_dev_unregister(&xe->drm); 1029 1030 return err; 1031 } 1032 1033 void xe_device_remove(struct xe_device *xe) 1034 { 1035 xe_display_unregister(xe); 1036 1037 drm_dev_unplug(&xe->drm); 1038 1039 xe_bo_pci_dev_remove_all(xe); 1040 } 1041 1042 void xe_device_shutdown(struct xe_device *xe) 1043 { 1044 struct xe_gt *gt; 1045 u8 id; 1046 1047 drm_dbg(&xe->drm, "Shutting down device\n"); 1048 1049 xe_display_pm_shutdown(xe); 1050 1051 xe_irq_suspend(xe); 1052 1053 for_each_gt(gt, xe, id) 1054 xe_gt_shutdown(gt); 1055 1056 xe_display_pm_shutdown_late(xe); 1057 1058 if (!xe_driver_flr_disabled(xe)) { 1059 /* BOOM! */ 1060 __xe_driver_flr(xe); 1061 } 1062 } 1063 1064 /** 1065 * xe_device_wmb() - Device specific write memory barrier 1066 * @xe: the &xe_device 1067 * 1068 * While wmb() is sufficient for a barrier if we use system memory, on discrete 1069 * platforms with device memory we additionally need to issue a register write. 1070 * Since it doesn't matter which register we write to, use the read-only VF_CAP 1071 * register that is also marked as accessible by the VFs. 1072 */ 1073 void xe_device_wmb(struct xe_device *xe) 1074 { 1075 wmb(); 1076 if (IS_DGFX(xe)) 1077 xe_mmio_write32(xe_root_tile_mmio(xe), VF_CAP_REG, 0); 1078 } 1079 1080 /* 1081 * Issue a TRANSIENT_FLUSH_REQUEST and wait for completion on each gt. 1082 */ 1083 static void tdf_request_sync(struct xe_device *xe) 1084 { 1085 struct xe_gt *gt; 1086 u8 id; 1087 1088 for_each_gt_with_type(gt, xe, id, BIT(XE_GT_TYPE_MAIN)) { 1089 CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT); 1090 if (!fw_ref.domains) 1091 return; 1092 1093 xe_mmio_write32(>->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST); 1094 1095 /* 1096 * FIXME: We can likely do better here with our choice of 1097 * timeout. Currently we just assume the worst case, i.e. 150us, 1098 * which is believed to be sufficient to cover the worst case 1099 * scenario on current platforms if all cache entries are 1100 * transient and need to be flushed.. 1101 */ 1102 if (xe_mmio_wait32(>->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST, 0, 1103 300, NULL, false)) 1104 xe_gt_err_once(gt, "TD flush timeout\n"); 1105 } 1106 } 1107 1108 /** 1109 * xe_device_is_l2_flush_optimized - if L2 flush is optimized by HW 1110 * @xe: The device to check. 1111 * 1112 * Return: true if the HW device optimizing L2 flush, false otherwise. 1113 */ 1114 bool xe_device_is_l2_flush_optimized(struct xe_device *xe) 1115 { 1116 /* XA is *always* flushed, like at the end-of-submssion (and maybe other 1117 * places), just that internally as an optimisation hw doesn't need to make 1118 * that a full flush (which will also include XA) when Media is 1119 * off/powergated, since it doesn't need to worry about GT caches vs Media 1120 * coherency, and only CPU vs GPU coherency, so can make that flush a 1121 * targeted XA flush, since stuff tagged with XA now means it's shared with 1122 * the CPU. The main implication is that we now need to somehow flush non-XA before 1123 * freeing system memory pages, otherwise dirty cachelines could be flushed after the free 1124 * (like if Media suddenly turns on and does a full flush) 1125 */ 1126 if (GRAPHICS_VER(xe) >= 35 && !IS_DGFX(xe)) 1127 return true; 1128 return false; 1129 } 1130 1131 void xe_device_l2_flush(struct xe_device *xe) 1132 { 1133 struct xe_gt *gt; 1134 1135 gt = xe_root_mmio_gt(xe); 1136 if (!gt) 1137 return; 1138 1139 if (!XE_GT_WA(gt, 16023588340)) 1140 return; 1141 1142 CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT); 1143 if (!fw_ref.domains) 1144 return; 1145 1146 spin_lock(>->global_invl_lock); 1147 1148 xe_mmio_write32(>->mmio, XE2_GLOBAL_INVAL, 0x1); 1149 if (xe_mmio_wait32(>->mmio, XE2_GLOBAL_INVAL, 0x1, 0x0, 1000, NULL, true)) 1150 xe_gt_err_once(gt, "Global invalidation timeout\n"); 1151 1152 spin_unlock(>->global_invl_lock); 1153 } 1154 1155 /** 1156 * xe_device_td_flush() - Flush transient L3 cache entries 1157 * @xe: The device 1158 * 1159 * Display engine has direct access to memory and is never coherent with L3/L4 1160 * caches (or CPU caches), however KMD is responsible for specifically flushing 1161 * transient L3 GPU cache entries prior to the flip sequence to ensure scanout 1162 * can happen from such a surface without seeing corruption. 1163 * 1164 * Display surfaces can be tagged as transient by mapping it using one of the 1165 * various L3:XD PAT index modes on Xe2. 1166 * 1167 * Note: On non-discrete xe2 platforms, like LNL, the entire L3 cache is flushed 1168 * at the end of each submission via PIPE_CONTROL for compute/render, since SA 1169 * Media is not coherent with L3 and we want to support render-vs-media 1170 * usescases. For other engines like copy/blt the HW internally forces uncached 1171 * behaviour, hence why we can skip the TDF on such platforms. 1172 */ 1173 void xe_device_td_flush(struct xe_device *xe) 1174 { 1175 struct xe_gt *root_gt; 1176 1177 /* 1178 * From Xe3p onward the HW takes care of flush of TD entries also along 1179 * with flushing XA entries, which will be at the usual sync points, 1180 * like at the end of submission, so no manual flush is needed here. 1181 */ 1182 if (GRAPHICS_VER(xe) >= 35) 1183 return; 1184 1185 if (!IS_DGFX(xe) || GRAPHICS_VER(xe) < 20) 1186 return; 1187 1188 root_gt = xe_root_mmio_gt(xe); 1189 if (!root_gt) 1190 return; 1191 1192 if (XE_GT_WA(root_gt, 16023588340)) { 1193 /* A transient flush is not sufficient: flush the L2 */ 1194 xe_device_l2_flush(xe); 1195 } else { 1196 xe_guc_pc_apply_flush_freq_limit(&root_gt->uc.guc.pc); 1197 tdf_request_sync(xe); 1198 xe_guc_pc_remove_flush_freq_limit(&root_gt->uc.guc.pc); 1199 } 1200 } 1201 1202 u32 xe_device_ccs_bytes(struct xe_device *xe, u64 size) 1203 { 1204 return xe_device_has_flat_ccs(xe) ? 1205 DIV_ROUND_UP_ULL(size, NUM_BYTES_PER_CCS_BYTE(xe)) : 0; 1206 } 1207 1208 /** 1209 * xe_device_assert_mem_access - Inspect the current runtime_pm state. 1210 * @xe: xe device instance 1211 * 1212 * To be used before any kind of memory access. It will splat a debug warning 1213 * if the device is currently sleeping. But it doesn't guarantee in any way 1214 * that the device is going to remain awake. Xe PM runtime get and put 1215 * functions might be added to the outer bound of the memory access, while 1216 * this check is intended for inner usage to splat some warning if the worst 1217 * case has just happened. 1218 */ 1219 void xe_device_assert_mem_access(struct xe_device *xe) 1220 { 1221 xe_assert(xe, !xe_pm_runtime_suspended(xe)); 1222 } 1223 1224 void xe_device_snapshot_print(struct xe_device *xe, struct drm_printer *p) 1225 { 1226 struct xe_gt *gt; 1227 u8 id; 1228 1229 drm_printf(p, "PCI ID: 0x%04x\n", xe->info.devid); 1230 drm_printf(p, "PCI revision: 0x%02x\n", xe->info.revid); 1231 1232 for_each_gt(gt, xe, id) { 1233 drm_printf(p, "GT id: %u\n", id); 1234 drm_printf(p, "\tTile: %u\n", gt->tile->id); 1235 drm_printf(p, "\tType: %s\n", 1236 gt->info.type == XE_GT_TYPE_MAIN ? "main" : "media"); 1237 drm_printf(p, "\tIP ver: %u.%u.%u\n", 1238 REG_FIELD_GET(GMD_ID_ARCH_MASK, gt->info.gmdid), 1239 REG_FIELD_GET(GMD_ID_RELEASE_MASK, gt->info.gmdid), 1240 REG_FIELD_GET(GMD_ID_REVID, gt->info.gmdid)); 1241 drm_printf(p, "\tCS reference clock: %u\n", gt->info.reference_clock); 1242 } 1243 } 1244 1245 u64 xe_device_canonicalize_addr(struct xe_device *xe, u64 address) 1246 { 1247 return sign_extend64(address, xe->info.va_bits - 1); 1248 } 1249 1250 u64 xe_device_uncanonicalize_addr(struct xe_device *xe, u64 address) 1251 { 1252 return address & GENMASK_ULL(xe->info.va_bits - 1, 0); 1253 } 1254 1255 /** 1256 * DOC: Xe Device Wedging 1257 * 1258 * Xe driver uses drm device wedged uevent as documented in Documentation/gpu/drm-uapi.rst. 1259 * When device is in wedged state, every IOCTL will be blocked and GT cannot 1260 * be used. The conditions under which the driver declares the device wedged 1261 * depend on the wedged mode configuration (see &enum xe_wedged_mode). The 1262 * default recovery method for a wedged state is rebind/bus-reset. 1263 * 1264 * Another recovery method is vendor-specific. Below are the cases that send 1265 * ``WEDGED=vendor-specific`` recovery method in drm device wedged uevent. 1266 * 1267 * Case: Firmware Flash 1268 * -------------------- 1269 * 1270 * Identification Hint 1271 * +++++++++++++++++++ 1272 * 1273 * ``WEDGED=vendor-specific`` drm device wedged uevent with 1274 * :ref:`Runtime Survivability mode <xe-survivability-mode>` is used to notify 1275 * admin/userspace consumer about the need for a firmware flash. 1276 * 1277 * Recovery Procedure 1278 * ++++++++++++++++++ 1279 * 1280 * Once ``WEDGED=vendor-specific`` drm device wedged uevent is received, follow 1281 * the below steps 1282 * 1283 * - Check Runtime Survivability mode sysfs. 1284 * If enabled, firmware flash is required to recover the device. 1285 * 1286 * /sys/bus/pci/devices/<device>/survivability_mode 1287 * 1288 * - Admin/userspace consumer can use firmware flashing tools like fwupd to flash 1289 * firmware and restore device to normal operation. 1290 */ 1291 1292 /** 1293 * xe_device_set_wedged_method - Set wedged recovery method 1294 * @xe: xe device instance 1295 * @method: recovery method to set 1296 * 1297 * Set wedged recovery method to be sent in drm wedged uevent. 1298 */ 1299 void xe_device_set_wedged_method(struct xe_device *xe, unsigned long method) 1300 { 1301 xe->wedged.method = method; 1302 } 1303 1304 /** 1305 * xe_device_declare_wedged - Declare device wedged 1306 * @xe: xe device instance 1307 * 1308 * This is a final state that can only be cleared with the recovery method 1309 * specified in the drm wedged uevent. The method can be set using 1310 * xe_device_set_wedged_method before declaring the device as wedged. If no method 1311 * is set, reprobe (unbind/re-bind) will be sent by default. 1312 * 1313 * In this state every IOCTL will be blocked so the GT cannot be used. 1314 * In general it will be called upon any critical error such as gt reset 1315 * failure or guc loading failure. Userspace will be notified of this state 1316 * through device wedged uevent. 1317 * If xe.wedged module parameter is set to 2, this function will be called 1318 * on every single execution timeout (a.k.a. GPU hang) right after devcoredump 1319 * snapshot capture. In this mode, GT reset won't be attempted so the state of 1320 * the issue is preserved for further debugging. 1321 */ 1322 void xe_device_declare_wedged(struct xe_device *xe) 1323 { 1324 struct xe_gt *gt; 1325 u8 id; 1326 1327 if (xe->wedged.mode == XE_WEDGED_MODE_NEVER) { 1328 drm_dbg(&xe->drm, "Wedged mode is forcibly disabled\n"); 1329 return; 1330 } 1331 1332 if (!atomic_xchg(&xe->wedged.flag, 1)) { 1333 xe->needs_flr_on_fini = true; 1334 xe_pm_runtime_get_noresume(xe); 1335 drm_err(&xe->drm, 1336 "CRITICAL: Xe has declared device %s as wedged.\n" 1337 "IOCTLs and executions are blocked.\n" 1338 "For recovery procedure, refer to https://docs.kernel.org/gpu/drm-uapi.html#device-wedging\n" 1339 "Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/xe/kernel/issues/new\n", 1340 dev_name(xe->drm.dev)); 1341 } 1342 1343 for_each_gt(gt, xe, id) 1344 xe_gt_declare_wedged(gt); 1345 1346 if (xe_device_wedged(xe)) { 1347 /* 1348 * XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET is intended for debugging 1349 * hangs, so wedge the device with 'none' recovery method and have 1350 * it available to the user for debugging. 1351 */ 1352 if (xe->wedged.mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET) 1353 xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_NONE); 1354 /* If no wedge recovery method is set, use default */ 1355 else if (!xe->wedged.method) 1356 xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_REBIND | 1357 DRM_WEDGE_RECOVERY_BUS_RESET); 1358 1359 /* Notify userspace of wedged device */ 1360 drm_dev_wedged_event(&xe->drm, xe->wedged.method, NULL); 1361 } 1362 } 1363 1364 /** 1365 * xe_device_validate_wedged_mode - Check if given mode is supported 1366 * @xe: the &xe_device 1367 * @mode: requested mode to validate 1368 * 1369 * Check whether the provided wedged mode is supported. 1370 * 1371 * Return: 0 if mode is supported, error code otherwise. 1372 */ 1373 int xe_device_validate_wedged_mode(struct xe_device *xe, unsigned int mode) 1374 { 1375 if (mode > XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET) { 1376 drm_dbg(&xe->drm, "wedged_mode: invalid value (%u)\n", mode); 1377 return -EINVAL; 1378 } else if (mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET && (IS_SRIOV_VF(xe) || 1379 (IS_SRIOV_PF(xe) && !IS_ENABLED(CONFIG_DRM_XE_DEBUG)))) { 1380 drm_dbg(&xe->drm, "wedged_mode: (%u) %s mode is not supported for %s\n", 1381 mode, xe_wedged_mode_to_string(mode), 1382 xe_sriov_mode_to_string(xe_device_sriov_mode(xe))); 1383 return -EPERM; 1384 } 1385 1386 return 0; 1387 } 1388 1389 /** 1390 * xe_wedged_mode_to_string - Convert enum value to string. 1391 * @mode: the &xe_wedged_mode to convert 1392 * 1393 * Returns: wedged mode as a user friendly string. 1394 */ 1395 const char *xe_wedged_mode_to_string(enum xe_wedged_mode mode) 1396 { 1397 switch (mode) { 1398 case XE_WEDGED_MODE_NEVER: 1399 return "never"; 1400 case XE_WEDGED_MODE_UPON_CRITICAL_ERROR: 1401 return "upon-critical-error"; 1402 case XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET: 1403 return "upon-any-hang-no-reset"; 1404 default: 1405 return "<invalid>"; 1406 } 1407 } 1408 1409 /** 1410 * xe_device_asid_to_vm() - Find VM from ASID 1411 * @xe: the &xe_device 1412 * @asid: Address space ID 1413 * 1414 * Find a VM from ASID and take a reference to VM which caller must drop. 1415 * Reclaim safe. 1416 * 1417 * Return: VM on success, ERR_PTR on failure 1418 */ 1419 struct xe_vm *xe_device_asid_to_vm(struct xe_device *xe, u32 asid) 1420 { 1421 struct xe_vm *vm; 1422 1423 down_read(&xe->usm.lock); 1424 vm = xa_load(&xe->usm.asid_to_vm, asid); 1425 if (vm) 1426 xe_vm_get(vm); 1427 else 1428 vm = ERR_PTR(-EINVAL); 1429 up_read(&xe->usm.lock); 1430 1431 return vm; 1432 } 1433