xref: /linux/drivers/gpu/drm/xe/xe_device.c (revision c06b6cde2a1c3bcbb561bd57bb6f34eae9030921)
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 <kunit/static_stub.h>
21 #include <uapi/drm/xe_drm.h>
22 
23 #include "display/xe_display.h"
24 #include "instructions/xe_gpu_commands.h"
25 #include "regs/xe_gt_regs.h"
26 #include "regs/xe_regs.h"
27 #include "xe_bo.h"
28 #include "xe_bo_evict.h"
29 #include "xe_configfs.h"
30 #include "xe_debugfs.h"
31 #include "xe_defaults.h"
32 #include "xe_devcoredump.h"
33 #include "xe_device_sysfs.h"
34 #include "xe_dma_buf.h"
35 #include "xe_drm_client.h"
36 #include "xe_drv.h"
37 #include "xe_exec.h"
38 #include "xe_exec_queue.h"
39 #include "xe_force_wake.h"
40 #include "xe_ggtt.h"
41 #include "xe_gt.h"
42 #include "xe_gt_mcr.h"
43 #include "xe_gt_printk.h"
44 #include "xe_gt_sriov_vf.h"
45 #include "xe_guc.h"
46 #include "xe_guc_pc.h"
47 #include "xe_hw_engine_group.h"
48 #include "xe_hwmon.h"
49 #include "xe_i2c.h"
50 #include "xe_irq.h"
51 #include "xe_late_bind_fw.h"
52 #include "xe_mmio.h"
53 #include "xe_module.h"
54 #include "xe_nvm.h"
55 #include "xe_oa.h"
56 #include "xe_observation.h"
57 #include "xe_pagefault.h"
58 #include "xe_pat.h"
59 #include "xe_pcode.h"
60 #include "xe_pm.h"
61 #include "xe_pmu.h"
62 #include "xe_psmi.h"
63 #include "xe_pxp.h"
64 #include "xe_query.h"
65 #include "xe_shrinker.h"
66 #include "xe_soc_remapper.h"
67 #include "xe_survivability_mode.h"
68 #include "xe_sriov.h"
69 #include "xe_svm.h"
70 #include "xe_sysctrl.h"
71 #include "xe_tile.h"
72 #include "xe_ttm_stolen_mgr.h"
73 #include "xe_ttm_sys_mgr.h"
74 #include "xe_vm.h"
75 #include "xe_vm_madvise.h"
76 #include "xe_vram.h"
77 #include "xe_vram_types.h"
78 #include "xe_vsec.h"
79 #include "xe_wait_user_fence.h"
80 #include "xe_wa.h"
81 
82 #include <generated/xe_device_wa_oob.h>
83 #include <generated/xe_wa_oob.h>
84 
85 static int xe_file_open(struct drm_device *dev, struct drm_file *file)
86 {
87 	struct xe_device *xe = to_xe_device(dev);
88 	struct xe_drm_client *client;
89 	struct xe_file *xef;
90 	int ret = -ENOMEM;
91 	struct task_struct *task = NULL;
92 
93 	xef = kzalloc_obj(*xef);
94 	if (!xef)
95 		return ret;
96 
97 	client = xe_drm_client_alloc();
98 	if (!client) {
99 		kfree(xef);
100 		return ret;
101 	}
102 
103 	xef->drm = file;
104 	xef->client = client;
105 	xef->xe = xe;
106 
107 	mutex_init(&xef->vm.lock);
108 	xa_init_flags(&xef->vm.xa, XA_FLAGS_ALLOC1);
109 
110 	mutex_init(&xef->exec_queue.lock);
111 	xa_init_flags(&xef->exec_queue.xa, XA_FLAGS_ALLOC1);
112 
113 	file->driver_priv = xef;
114 	kref_init(&xef->refcount);
115 
116 	task = get_pid_task(rcu_access_pointer(file->pid), PIDTYPE_PID);
117 	if (task) {
118 		xef->process_name = kstrdup(task->comm, GFP_KERNEL);
119 		xef->pid = task->pid;
120 		put_task_struct(task);
121 	}
122 
123 	return 0;
124 }
125 
126 static void xe_file_destroy(struct kref *ref)
127 {
128 	struct xe_file *xef = container_of(ref, struct xe_file, refcount);
129 
130 	xa_destroy(&xef->exec_queue.xa);
131 	mutex_destroy(&xef->exec_queue.lock);
132 	xa_destroy(&xef->vm.xa);
133 	mutex_destroy(&xef->vm.lock);
134 
135 	xe_drm_client_put(xef->client);
136 	kfree(xef->process_name);
137 	kfree(xef);
138 }
139 
140 /**
141  * xe_file_get() - Take a reference to the xe file object
142  * @xef: Pointer to the xe file
143  *
144  * Anyone with a pointer to xef must take a reference to the xe file
145  * object using this call.
146  *
147  * Return: xe file pointer
148  */
149 struct xe_file *xe_file_get(struct xe_file *xef)
150 {
151 	kref_get(&xef->refcount);
152 	return xef;
153 }
154 
155 /**
156  * xe_file_put() - Drop a reference to the xe file object
157  * @xef: Pointer to the xe file
158  *
159  * Used to drop reference to the xef object
160  */
161 void xe_file_put(struct xe_file *xef)
162 {
163 	kref_put(&xef->refcount, xe_file_destroy);
164 }
165 
166 static void xe_file_close(struct drm_device *dev, struct drm_file *file)
167 {
168 	struct xe_device *xe = to_xe_device(dev);
169 	struct xe_file *xef = file->driver_priv;
170 	struct xe_vm *vm;
171 	struct xe_exec_queue *q;
172 	unsigned long idx;
173 
174 	guard(xe_pm_runtime)(xe);
175 
176 	/*
177 	 * No need for exec_queue.lock here as there is no contention for it
178 	 * when FD is closing as IOCTLs presumably can't be modifying the
179 	 * xarray. Taking exec_queue.lock here causes undue dependency on
180 	 * vm->lock taken during xe_exec_queue_kill().
181 	 */
182 	xa_for_each(&xef->exec_queue.xa, idx, q) {
183 		if (q->vm && q->hwe->hw_engine_group)
184 			xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q);
185 		xe_exec_queue_kill(q);
186 		xe_exec_queue_put(q);
187 	}
188 	xa_for_each(&xef->vm.xa, idx, vm)
189 		xe_vm_close_and_put(vm);
190 
191 	xe_file_put(xef);
192 }
193 
194 static const struct drm_ioctl_desc xe_ioctls[] = {
195 	DRM_IOCTL_DEF_DRV(XE_DEVICE_QUERY, xe_query_ioctl, DRM_RENDER_ALLOW),
196 	DRM_IOCTL_DEF_DRV(XE_GEM_CREATE, xe_gem_create_ioctl, DRM_RENDER_ALLOW),
197 	DRM_IOCTL_DEF_DRV(XE_GEM_MMAP_OFFSET, xe_gem_mmap_offset_ioctl,
198 			  DRM_RENDER_ALLOW),
199 	DRM_IOCTL_DEF_DRV(XE_VM_CREATE, xe_vm_create_ioctl, DRM_RENDER_ALLOW),
200 	DRM_IOCTL_DEF_DRV(XE_VM_DESTROY, xe_vm_destroy_ioctl, DRM_RENDER_ALLOW),
201 	DRM_IOCTL_DEF_DRV(XE_VM_BIND, xe_vm_bind_ioctl, DRM_RENDER_ALLOW),
202 	DRM_IOCTL_DEF_DRV(XE_EXEC, xe_exec_ioctl, DRM_RENDER_ALLOW),
203 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_CREATE, xe_exec_queue_create_ioctl,
204 			  DRM_RENDER_ALLOW),
205 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_DESTROY, xe_exec_queue_destroy_ioctl,
206 			  DRM_RENDER_ALLOW),
207 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_GET_PROPERTY, xe_exec_queue_get_property_ioctl,
208 			  DRM_RENDER_ALLOW),
209 	DRM_IOCTL_DEF_DRV(XE_WAIT_USER_FENCE, xe_wait_user_fence_ioctl,
210 			  DRM_RENDER_ALLOW),
211 	DRM_IOCTL_DEF_DRV(XE_OBSERVATION, xe_observation_ioctl, DRM_RENDER_ALLOW),
212 	DRM_IOCTL_DEF_DRV(XE_MADVISE, xe_vm_madvise_ioctl, DRM_RENDER_ALLOW),
213 	DRM_IOCTL_DEF_DRV(XE_VM_QUERY_MEM_RANGE_ATTRS, xe_vm_query_vmas_attrs_ioctl,
214 			  DRM_RENDER_ALLOW),
215 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_SET_PROPERTY, xe_exec_queue_set_property_ioctl,
216 			  DRM_RENDER_ALLOW),
217 	DRM_IOCTL_DEF_DRV(XE_VM_GET_PROPERTY, xe_vm_get_property_ioctl,
218 			  DRM_RENDER_ALLOW),
219 };
220 
221 static long xe_drm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
222 {
223 	struct drm_file *file_priv = file->private_data;
224 	struct xe_device *xe = to_xe_device(file_priv->minor->dev);
225 	long ret;
226 
227 	if (xe_device_wedged(xe))
228 		return -ECANCELED;
229 
230 	ACQUIRE(xe_pm_runtime_ioctl, pm)(xe);
231 	ret = ACQUIRE_ERR(xe_pm_runtime_ioctl, &pm);
232 	if (ret >= 0)
233 		ret = drm_ioctl(file, cmd, arg);
234 
235 	return ret;
236 }
237 
238 #ifdef CONFIG_COMPAT
239 static long xe_drm_compat_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
240 {
241 	struct drm_file *file_priv = file->private_data;
242 	struct xe_device *xe = to_xe_device(file_priv->minor->dev);
243 	long ret;
244 
245 	if (xe_device_wedged(xe))
246 		return -ECANCELED;
247 
248 	ACQUIRE(xe_pm_runtime_ioctl, pm)(xe);
249 	ret = ACQUIRE_ERR(xe_pm_runtime_ioctl, &pm);
250 	if (ret >= 0)
251 		ret = drm_compat_ioctl(file, cmd, arg);
252 
253 	return ret;
254 }
255 #else
256 /* similarly to drm_compat_ioctl, let's it be assigned to .compat_ioct unconditionally */
257 #define xe_drm_compat_ioctl NULL
258 #endif
259 
260 static void barrier_open(struct vm_area_struct *vma)
261 {
262 	drm_dev_get(vma->vm_private_data);
263 }
264 
265 static void barrier_close(struct vm_area_struct *vma)
266 {
267 	drm_dev_put(vma->vm_private_data);
268 }
269 
270 static void barrier_release_dummy_page(struct drm_device *dev, void *res)
271 {
272 	struct page *dummy_page = (struct page *)res;
273 
274 	__free_page(dummy_page);
275 }
276 
277 static vm_fault_t barrier_fault(struct vm_fault *vmf)
278 {
279 	struct drm_device *dev = vmf->vma->vm_private_data;
280 	struct vm_area_struct *vma = vmf->vma;
281 	vm_fault_t ret = VM_FAULT_NOPAGE;
282 	pgprot_t prot;
283 	int idx;
284 
285 	prot = vm_get_page_prot(vma->vm_flags);
286 
287 	if (drm_dev_enter(dev, &idx)) {
288 		unsigned long pfn;
289 
290 #define LAST_DB_PAGE_OFFSET 0x7ff001
291 		pfn = PHYS_PFN(pci_resource_start(to_pci_dev(dev->dev), 0) +
292 				LAST_DB_PAGE_OFFSET);
293 		ret = vmf_insert_pfn_prot(vma, vma->vm_start, pfn,
294 					  pgprot_noncached(prot));
295 		drm_dev_exit(idx);
296 	} else {
297 		struct page *page;
298 
299 		/* Allocate new dummy page to map all the VA range in this VMA to it*/
300 		page = alloc_page(GFP_KERNEL | __GFP_ZERO);
301 		if (!page)
302 			return VM_FAULT_OOM;
303 
304 		/* Set the page to be freed using drmm release action */
305 		if (drmm_add_action_or_reset(dev, barrier_release_dummy_page, page))
306 			return VM_FAULT_OOM;
307 
308 		ret = vmf_insert_pfn_prot(vma, vma->vm_start, page_to_pfn(page),
309 					  prot);
310 	}
311 
312 	return ret;
313 }
314 
315 static const struct vm_operations_struct vm_ops_barrier = {
316 	.open = barrier_open,
317 	.close = barrier_close,
318 	.fault = barrier_fault,
319 };
320 
321 static int xe_pci_barrier_mmap(struct file *filp,
322 			       struct vm_area_struct *vma)
323 {
324 	struct drm_file *priv = filp->private_data;
325 	struct drm_device *dev = priv->minor->dev;
326 	struct xe_device *xe = to_xe_device(dev);
327 
328 	if (!IS_DGFX(xe))
329 		return -EINVAL;
330 
331 	if (vma->vm_end - vma->vm_start > SZ_4K)
332 		return -EINVAL;
333 
334 	if (is_cow_mapping(vma->vm_flags))
335 		return -EINVAL;
336 
337 	if (vma->vm_flags & (VM_READ | VM_EXEC))
338 		return -EINVAL;
339 
340 	vm_flags_clear(vma, VM_MAYREAD | VM_MAYEXEC);
341 	vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP | VM_IO);
342 	vma->vm_ops = &vm_ops_barrier;
343 	vma->vm_private_data = dev;
344 	drm_dev_get(vma->vm_private_data);
345 
346 	return 0;
347 }
348 
349 static int xe_mmap(struct file *filp, struct vm_area_struct *vma)
350 {
351 	struct drm_file *priv = filp->private_data;
352 	struct drm_device *dev = priv->minor->dev;
353 
354 	if (drm_dev_is_unplugged(dev))
355 		return -ENODEV;
356 
357 	switch (vma->vm_pgoff) {
358 	case XE_PCI_BARRIER_MMAP_OFFSET >> XE_PTE_SHIFT:
359 		return xe_pci_barrier_mmap(filp, vma);
360 	}
361 
362 	return drm_gem_mmap(filp, vma);
363 }
364 
365 static const struct file_operations xe_driver_fops = {
366 	.owner = THIS_MODULE,
367 	.open = drm_open,
368 	.release = drm_release_noglobal,
369 	.unlocked_ioctl = xe_drm_ioctl,
370 	.mmap = xe_mmap,
371 	.poll = drm_poll,
372 	.read = drm_read,
373 	.compat_ioctl = xe_drm_compat_ioctl,
374 	.llseek = noop_llseek,
375 #ifdef CONFIG_PROC_FS
376 	.show_fdinfo = drm_show_fdinfo,
377 #endif
378 	.fop_flags = FOP_UNSIGNED_OFFSET,
379 };
380 
381 /**
382  * xe_is_xe_file() - Is the file an xe device file?
383  * @file: The file.
384  *
385  * Checks whether the file is opened against
386  * an xe device.
387  *
388  * Return: %true if an xe file, %false if not.
389  */
390 bool xe_is_xe_file(const struct file *file)
391 {
392 	return file->f_op == &xe_driver_fops;
393 }
394 
395 static struct drm_driver regular_driver = {
396 	.driver_features =
397 	    DRIVER_GEM |
398 	    DRIVER_RENDER | DRIVER_SYNCOBJ |
399 	    DRIVER_SYNCOBJ_TIMELINE | DRIVER_GEM_GPUVA,
400 	.open = xe_file_open,
401 	.postclose = xe_file_close,
402 
403 	.gem_prime_import = xe_gem_prime_import,
404 
405 	.dumb_create = xe_bo_dumb_create,
406 	.dumb_map_offset = drm_gem_ttm_dumb_map_offset,
407 #ifdef CONFIG_PROC_FS
408 	.show_fdinfo = xe_drm_client_fdinfo,
409 #endif
410 	.ioctls = xe_ioctls,
411 	.num_ioctls = ARRAY_SIZE(xe_ioctls),
412 	.fops = &xe_driver_fops,
413 	.name = DRIVER_NAME,
414 	.desc = DRIVER_DESC,
415 	.major = DRIVER_MAJOR,
416 	.minor = DRIVER_MINOR,
417 	.patchlevel = DRIVER_PATCHLEVEL,
418 };
419 
420 #ifdef CONFIG_PCI_IOV
421 static const struct drm_ioctl_desc xe_ioctls_admin_only[] = {
422 	DRM_IOCTL_DEF_DRV(XE_DEVICE_QUERY, xe_query_ioctl, DRM_RENDER_ALLOW),
423 	DRM_IOCTL_DEF_DRV(XE_OBSERVATION, xe_observation_ioctl, DRM_RENDER_ALLOW),
424 };
425 
426 static struct drm_driver admin_only_driver = {
427 	.driver_features =
428 	    DRIVER_GEM | DRIVER_RENDER | DRIVER_GEM_GPUVA,
429 	.open = xe_file_open,
430 	.postclose = xe_file_close,
431 	.ioctls = xe_ioctls_admin_only,
432 	.num_ioctls = ARRAY_SIZE(xe_ioctls_admin_only),
433 	.fops = &xe_driver_fops,
434 	.name = DRIVER_NAME,
435 	.desc = DRIVER_DESC,
436 	.major = DRIVER_MAJOR,
437 	.minor = DRIVER_MINOR,
438 	.patchlevel = DRIVER_PATCHLEVEL,
439 };
440 
441 /**
442  * xe_device_is_admin_only() - Check whether device is admin only or not.
443  * @xe: the &xe_device to check
444  *
445  * Return: true if the device is admin only, false otherwise.
446  */
447 bool xe_device_is_admin_only(const struct xe_device *xe)
448 {
449 	KUNIT_STATIC_STUB_REDIRECT(xe_device_is_admin_only, xe);
450 	return xe->drm.driver == &admin_only_driver;
451 }
452 #endif
453 
454 static void xe_device_destroy(struct drm_device *dev, void *dummy)
455 {
456 	struct xe_device *xe = to_xe_device(dev);
457 
458 	xe_bo_dev_fini(&xe->bo_device);
459 
460 	if (xe->preempt_fence_wq)
461 		destroy_workqueue(xe->preempt_fence_wq);
462 
463 	if (xe->ordered_wq)
464 		destroy_workqueue(xe->ordered_wq);
465 
466 	if (xe->unordered_wq)
467 		destroy_workqueue(xe->unordered_wq);
468 
469 	if (xe->destroy_wq)
470 		destroy_workqueue(xe->destroy_wq);
471 
472 	ttm_device_fini(&xe->ttm);
473 }
474 
475 struct xe_device *xe_device_create(struct pci_dev *pdev,
476 				   const struct pci_device_id *ent)
477 {
478 	struct drm_driver *driver = &regular_driver;
479 	struct xe_device *xe;
480 	int err;
481 
482 #ifdef CONFIG_PCI_IOV
483 	/*
484 	 * Since XE device is not initialized yet, read from configfs
485 	 * directly to decide whether we are in admin-only PF mode or not.
486 	 */
487 	if (xe_configfs_admin_only_pf(pdev))
488 		driver = &admin_only_driver;
489 #endif
490 
491 	xe_display_driver_set_hooks(driver);
492 
493 	err = aperture_remove_conflicting_pci_devices(pdev, driver->name);
494 	if (err)
495 		return ERR_PTR(err);
496 
497 	xe = devm_drm_dev_alloc(&pdev->dev, driver, struct xe_device, drm);
498 	if (IS_ERR(xe))
499 		return xe;
500 
501 	err = ttm_device_init(&xe->ttm, &xe_ttm_funcs, xe->drm.dev,
502 			      xe->drm.anon_inode->i_mapping,
503 			      xe->drm.vma_offset_manager, 0);
504 	if (WARN_ON(err))
505 		return ERR_PTR(err);
506 
507 	xe_bo_dev_init(&xe->bo_device);
508 	err = drmm_add_action_or_reset(&xe->drm, xe_device_destroy, NULL);
509 	if (err)
510 		return ERR_PTR(err);
511 
512 	err = xe_shrinker_create(xe);
513 	if (err)
514 		return ERR_PTR(err);
515 
516 	xe->info.devid = pdev->device;
517 	xe->info.revid = pdev->revision;
518 	xe->info.force_execlist = xe_modparam.force_execlist;
519 	xe->atomic_svm_timeslice_ms = 5;
520 	xe->min_run_period_lr_ms = 5;
521 
522 	err = xe_irq_init(xe);
523 	if (err)
524 		return ERR_PTR(err);
525 
526 	xe_validation_device_init(&xe->val);
527 
528 	init_waitqueue_head(&xe->ufence_wq);
529 
530 	init_rwsem(&xe->usm.lock);
531 
532 	err = xe_pagemap_shrinker_create(xe);
533 	if (err)
534 		return ERR_PTR(err);
535 
536 	xa_init_flags(&xe->usm.asid_to_vm, XA_FLAGS_ALLOC);
537 
538 	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) {
539 		/* Trigger a large asid and an early asid wrap. */
540 		u32 asid;
541 
542 		BUILD_BUG_ON(XE_MAX_ASID < 2);
543 		err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, NULL,
544 				      XA_LIMIT(XE_MAX_ASID - 2, XE_MAX_ASID - 1),
545 				      &xe->usm.next_asid, GFP_KERNEL);
546 		drm_WARN_ON(&xe->drm, err);
547 		if (err >= 0)
548 			xa_erase(&xe->usm.asid_to_vm, asid);
549 	}
550 
551 	err = xe_bo_pinned_init(xe);
552 	if (err)
553 		return ERR_PTR(err);
554 
555 	xe->preempt_fence_wq = alloc_ordered_workqueue("xe-preempt-fence-wq",
556 						       WQ_MEM_RECLAIM);
557 	xe->ordered_wq = alloc_ordered_workqueue("xe-ordered-wq", 0);
558 	xe->unordered_wq = alloc_workqueue("xe-unordered-wq", WQ_PERCPU, 0);
559 	xe->destroy_wq = alloc_workqueue("xe-destroy-wq", WQ_PERCPU, 0);
560 	if (!xe->ordered_wq || !xe->unordered_wq ||
561 	    !xe->preempt_fence_wq || !xe->destroy_wq) {
562 		/*
563 		 * Cleanup done in xe_device_destroy via
564 		 * drmm_add_action_or_reset register above
565 		 */
566 		drm_err(&xe->drm, "Failed to allocate xe workqueues\n");
567 		return ERR_PTR(-ENOMEM);
568 	}
569 
570 	err = drmm_mutex_init(&xe->drm, &xe->pmt.lock);
571 	if (err)
572 		return ERR_PTR(err);
573 
574 	return xe;
575 }
576 ALLOW_ERROR_INJECTION(xe_device_create, ERRNO); /* See xe_pci_probe() */
577 
578 static bool xe_driver_flr_disabled(struct xe_device *xe)
579 {
580 	if (IS_SRIOV_VF(xe))
581 		return true;
582 
583 	if (xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL_PROTECTED) & DRIVERINT_FLR_DIS) {
584 		drm_info(&xe->drm, "Driver-FLR disabled by BIOS\n");
585 		return true;
586 	}
587 
588 	return false;
589 }
590 
591 /*
592  * The driver-initiated FLR is the highest level of reset that we can trigger
593  * from within the driver. It is different from the PCI FLR in that it doesn't
594  * fully reset the SGUnit and doesn't modify the PCI config space and therefore
595  * it doesn't require a re-enumeration of the PCI BARs. However, the
596  * driver-initiated FLR does still cause a reset of both GT and display and a
597  * memory wipe of local and stolen memory, so recovery would require a full HW
598  * re-init and saving/restoring (or re-populating) the wiped memory. Since we
599  * perform the FLR as the very last action before releasing access to the HW
600  * during the driver release flow, we don't attempt recovery at all, because
601  * if/when a new instance of Xe is bound to the device it will do a full
602  * re-init anyway.
603  */
604 static void __xe_driver_flr(struct xe_device *xe)
605 {
606 	const unsigned int flr_timeout = 3 * USEC_PER_SEC; /* specs recommend a 3s wait */
607 	struct xe_mmio *mmio = xe_root_tile_mmio(xe);
608 	int ret;
609 
610 	drm_dbg(&xe->drm, "Triggering Driver-FLR\n");
611 
612 	/*
613 	 * Make sure any pending FLR requests have cleared by waiting for the
614 	 * FLR trigger bit to go to zero. Also clear GU_DEBUG's DRIVERFLR_STATUS
615 	 * to make sure it's not still set from a prior attempt (it's a write to
616 	 * clear bit).
617 	 * Note that we should never be in a situation where a previous attempt
618 	 * is still pending (unless the HW is totally dead), but better to be
619 	 * safe in case something unexpected happens
620 	 */
621 	ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false);
622 	if (ret) {
623 		drm_err(&xe->drm, "Driver-FLR-prepare wait for ready failed! %d\n", ret);
624 		return;
625 	}
626 	xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS);
627 
628 	/* Trigger the actual Driver-FLR */
629 	xe_mmio_rmw32(mmio, GU_CNTL, 0, DRIVERFLR);
630 
631 	/* Wait for hardware teardown to complete */
632 	ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false);
633 	if (ret) {
634 		drm_err(&xe->drm, "Driver-FLR-teardown wait completion failed! %d\n", ret);
635 		return;
636 	}
637 
638 	/* Wait for hardware/firmware re-init to complete */
639 	ret = xe_mmio_wait32(mmio, GU_DEBUG, DRIVERFLR_STATUS, DRIVERFLR_STATUS,
640 			     flr_timeout, NULL, false);
641 	if (ret) {
642 		drm_err(&xe->drm, "Driver-FLR-reinit wait completion failed! %d\n", ret);
643 		return;
644 	}
645 
646 	/* Clear sticky completion status */
647 	xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS);
648 }
649 
650 static void xe_driver_flr(struct xe_device *xe)
651 {
652 	if (xe_driver_flr_disabled(xe))
653 		return;
654 
655 	__xe_driver_flr(xe);
656 }
657 
658 static void xe_driver_flr_fini(void *arg)
659 {
660 	struct xe_device *xe = arg;
661 
662 	if (xe->needs_flr_on_fini)
663 		xe_driver_flr(xe);
664 }
665 
666 static void xe_device_sanitize(void *arg)
667 {
668 	struct xe_device *xe = arg;
669 	struct xe_gt *gt;
670 	u8 id;
671 
672 	for_each_gt(gt, xe, id)
673 		xe_gt_sanitize(gt);
674 }
675 
676 static int xe_set_dma_info(struct xe_device *xe)
677 {
678 	unsigned int mask_size = xe->info.dma_mask_size;
679 	int err;
680 
681 	dma_set_max_seg_size(xe->drm.dev, xe_sg_segment_size(xe->drm.dev));
682 
683 	err = dma_set_mask(xe->drm.dev, DMA_BIT_MASK(mask_size));
684 	if (err)
685 		goto mask_err;
686 
687 	err = dma_set_coherent_mask(xe->drm.dev, DMA_BIT_MASK(mask_size));
688 	if (err)
689 		goto mask_err;
690 
691 	return 0;
692 
693 mask_err:
694 	drm_err(&xe->drm, "Can't set DMA mask/consistent mask (%d)\n", err);
695 	return err;
696 }
697 
698 static void assert_lmem_ready(struct xe_device *xe)
699 {
700 	if (!IS_DGFX(xe) || IS_SRIOV_VF(xe))
701 		return;
702 
703 	xe_assert(xe, xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL) &
704 		  LMEM_INIT);
705 }
706 
707 static void vf_update_device_info(struct xe_device *xe)
708 {
709 	xe_assert(xe, IS_SRIOV_VF(xe));
710 	/* disable features that are not available/applicable to VFs */
711 	xe->info.probe_display = 0;
712 	xe->info.has_heci_cscfi = 0;
713 	xe->info.has_heci_gscfi = 0;
714 	xe->info.has_late_bind = 0;
715 	xe->info.skip_guc_pc = 1;
716 	xe->info.skip_pcode = 1;
717 }
718 
719 static int xe_device_vram_alloc(struct xe_device *xe)
720 {
721 	struct xe_vram_region *vram;
722 
723 	if (!IS_DGFX(xe))
724 		return 0;
725 
726 	vram = drmm_kzalloc(&xe->drm, sizeof(*vram), GFP_KERNEL);
727 	if (!vram)
728 		return -ENOMEM;
729 
730 	xe->mem.vram = vram;
731 	return 0;
732 }
733 
734 /**
735  * xe_device_probe_early: Device early probe
736  * @xe: xe device instance
737  *
738  * Initialize MMIO resources that don't require any
739  * knowledge about tile count. Also initialize pcode and
740  * check vram initialization on root tile.
741  *
742  * Return: 0 on success, error code on failure
743  */
744 int xe_device_probe_early(struct xe_device *xe)
745 {
746 	int err;
747 
748 	xe_wa_device_init(xe);
749 	xe_wa_process_device_oob(xe);
750 
751 	err = xe_mmio_probe_early(xe);
752 	if (err)
753 		return err;
754 
755 	xe_sriov_probe_early(xe);
756 
757 	if (xe_device_is_admin_only(xe) && !IS_SRIOV_PF(xe)) {
758 		xe_err(xe, "Can't run Admin-only mode without SR-IOV PF mode!\n");
759 		return -ENODEV;
760 	}
761 
762 	if (IS_SRIOV_VF(xe))
763 		vf_update_device_info(xe);
764 
765 	/*
766 	 * Check for pcode uncore_init status to confirm if the SoC
767 	 * initialization is complete. Until done, any MMIO or lmem access from
768 	 * the driver will be blocked
769 	 */
770 	err = xe_pcode_probe_early(xe);
771 	if (err || xe_survivability_mode_is_requested(xe)) {
772 		int save_err = err;
773 
774 		/*
775 		 * Try to leave device in survivability mode if device is
776 		 * possible, but still return the previous error for error
777 		 * propagation
778 		 */
779 		err = xe_survivability_mode_boot_enable(xe);
780 		if (err)
781 			return err;
782 
783 		return save_err;
784 	}
785 
786 	/*
787 	 * Make sure the lmem is initialized and ready to use. xe_pcode_ready()
788 	 * is flagged after full initialization is complete. Assert if lmem is
789 	 * not initialized.
790 	 */
791 	assert_lmem_ready(xe);
792 
793 	xe->wedged.mode = xe_device_validate_wedged_mode(xe, xe_modparam.wedged_mode) ?
794 			  XE_DEFAULT_WEDGED_MODE : xe_modparam.wedged_mode;
795 	drm_dbg(&xe->drm, "wedged_mode: setting mode (%u) %s\n",
796 		xe->wedged.mode, xe_wedged_mode_to_string(xe->wedged.mode));
797 
798 	err = xe_device_vram_alloc(xe);
799 	if (err)
800 		return err;
801 
802 	return 0;
803 }
804 ALLOW_ERROR_INJECTION(xe_device_probe_early, ERRNO); /* See xe_pci_probe() */
805 
806 static int probe_has_flat_ccs(struct xe_device *xe)
807 {
808 	struct xe_gt *gt;
809 	u32 reg;
810 
811 	/* Always enabled/disabled, no runtime check to do */
812 	if (GRAPHICS_VER(xe) < 20 || !xe->info.has_flat_ccs || IS_SRIOV_VF(xe))
813 		return 0;
814 
815 	gt = xe_root_mmio_gt(xe);
816 	if (!gt)
817 		return 0;
818 
819 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
820 	if (!fw_ref.domains)
821 		return -ETIMEDOUT;
822 
823 	reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_LOWER);
824 	xe->info.has_flat_ccs = (reg & XE2_FLAT_CCS_ENABLE);
825 
826 	if (!xe->info.has_flat_ccs)
827 		drm_dbg(&xe->drm,
828 			"Flat CCS has been disabled in bios, May lead to performance impact");
829 
830 	return 0;
831 }
832 
833 /*
834  * Detect if the driver is being run on pre-production hardware.  We don't
835  * keep workarounds for pre-production hardware long term, so print an
836  * error and add taint if we're being loaded on a pre-production platform
837  * for which the pre-prod workarounds have already been removed.
838  *
839  * The general policy is that we'll remove any workarounds that only apply to
840  * pre-production hardware around the time force_probe restrictions are lifted
841  * for a platform of the next major IP generation (for example, Xe2 pre-prod
842  * workarounds should be removed around the time the first Xe3 platforms have
843  * force_probe lifted).
844  */
845 static void detect_preproduction_hw(struct xe_device *xe)
846 {
847 	struct xe_gt *gt;
848 	int id;
849 
850 	/*
851 	 * SR-IOV VFs don't have access to the FUSE2 register, so we can't
852 	 * check pre-production status there.  But the host OS will notice
853 	 * and report the pre-production status, which should be enough to
854 	 * help us catch mistaken use of pre-production hardware.
855 	 */
856 	if (IS_SRIOV_VF(xe))
857 		return;
858 
859 	/*
860 	 * The "SW_CAP" fuse contains a bit indicating whether the device is a
861 	 * production or pre-production device.  This fuse is reflected through
862 	 * the GT "FUSE2" register, even though the contents of the fuse are
863 	 * not GT-specific.  Every GT's reflection of this fuse should show the
864 	 * same value, so we'll just use the first available GT for lookup.
865 	 */
866 	for_each_gt(gt, xe, id)
867 		break;
868 
869 	if (!gt)
870 		return;
871 
872 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
873 	if (!xe_force_wake_ref_has_domain(fw_ref.domains, XE_FW_GT)) {
874 		xe_gt_err(gt, "Forcewake failure; cannot determine production/pre-production hw status.\n");
875 		return;
876 	}
877 
878 	if (xe_mmio_read32(&gt->mmio, FUSE2) & PRODUCTION_HW)
879 		return;
880 
881 	xe_info(xe, "Pre-production hardware detected.\n");
882 	if (!xe->info.has_pre_prod_wa) {
883 		xe_err(xe, "Pre-production workarounds for this platform have already been removed.\n");
884 		add_taint(TAINT_MACHINE_CHECK, LOCKDEP_STILL_OK);
885 	}
886 }
887 
888 static void xe_device_wedged_fini(struct drm_device *drm, void *arg)
889 {
890 	struct xe_device *xe = arg;
891 
892 	if (atomic_read(&xe->wedged.flag))
893 		xe_pm_runtime_put(xe);
894 }
895 
896 int xe_device_probe(struct xe_device *xe)
897 {
898 	struct xe_tile *tile;
899 	struct xe_gt *gt;
900 	int err;
901 	u8 id;
902 
903 	xe_pat_init_early(xe);
904 
905 	err = xe_sriov_init(xe);
906 	if (err)
907 		return err;
908 
909 	xe->info.mem_region_mask = 1;
910 
911 	err = xe_set_dma_info(xe);
912 	if (err)
913 		return err;
914 
915 	err = xe_mmio_probe_tiles(xe);
916 	if (err)
917 		return err;
918 
919 	for_each_gt(gt, xe, id) {
920 		err = xe_gt_init_early(gt);
921 		if (err)
922 			return err;
923 	}
924 
925 	for_each_tile(tile, xe, id) {
926 		err = xe_ggtt_init_early(tile->mem.ggtt);
927 		if (err)
928 			return err;
929 	}
930 
931 	/*
932 	 * From here on, if a step fails, make sure a Driver-FLR is triggereed
933 	 */
934 	err = devm_add_action_or_reset(xe->drm.dev, xe_driver_flr_fini, xe);
935 	if (err)
936 		return err;
937 
938 	err = probe_has_flat_ccs(xe);
939 	if (err)
940 		return err;
941 
942 	err = xe_vram_probe(xe);
943 	if (err)
944 		return err;
945 
946 	for_each_tile(tile, xe, id) {
947 		err = xe_tile_init_noalloc(tile);
948 		if (err)
949 			return err;
950 	}
951 
952 	/*
953 	 * Allow allocations only now to ensure xe_display_init_early()
954 	 * is the first to allocate, always.
955 	 */
956 	err = xe_ttm_sys_mgr_init(xe);
957 	if (err)
958 		return err;
959 
960 	/* Allocate and map stolen after potential VRAM resize */
961 	err = xe_ttm_stolen_mgr_init(xe);
962 	if (err)
963 		return err;
964 
965 	/*
966 	 * Now that GT is initialized (TTM in particular),
967 	 * we can try to init display, and inherit the initial fb.
968 	 * This is the reason the first allocation needs to be done
969 	 * inside display.
970 	 */
971 	err = xe_display_init_early(xe);
972 	if (err)
973 		return err;
974 
975 	for_each_tile(tile, xe, id) {
976 		err = xe_tile_init(tile);
977 		if (err)
978 			return err;
979 	}
980 
981 	err = xe_irq_install(xe);
982 	if (err)
983 		return err;
984 
985 	for_each_gt(gt, xe, id) {
986 		err = xe_gt_init(gt);
987 		if (err)
988 			return err;
989 	}
990 
991 	err = xe_pagefault_init(xe);
992 	if (err)
993 		return err;
994 
995 	if (xe->tiles->media_gt &&
996 	    XE_GT_WA(xe->tiles->media_gt, 15015404425_disable))
997 		XE_DEVICE_WA_DISABLE(xe, 15015404425);
998 
999 	err = xe_devcoredump_init(xe);
1000 	if (err)
1001 		return err;
1002 
1003 	xe_nvm_init(xe);
1004 
1005 	err = xe_soc_remapper_init(xe);
1006 	if (err)
1007 		return err;
1008 
1009 	err = xe_heci_gsc_init(xe);
1010 	if (err)
1011 		return err;
1012 
1013 	err = xe_late_bind_init(&xe->late_bind);
1014 	if (err)
1015 		return err;
1016 
1017 	err = xe_oa_init(xe);
1018 	if (err)
1019 		return err;
1020 
1021 	err = xe_display_init(xe);
1022 	if (err)
1023 		return err;
1024 
1025 	err = xe_pxp_init(xe);
1026 	if (err)
1027 		return err;
1028 
1029 	err = xe_psmi_init(xe);
1030 	if (err)
1031 		return err;
1032 
1033 	err = drm_dev_register(&xe->drm, 0);
1034 	if (err)
1035 		return err;
1036 
1037 	xe_display_register(xe);
1038 
1039 	err = xe_oa_register(xe);
1040 	if (err)
1041 		goto err_unregister_display;
1042 
1043 	err = xe_pmu_register(&xe->pmu);
1044 	if (err)
1045 		goto err_unregister_display;
1046 
1047 	err = xe_sysctrl_init(xe);
1048 	if (err)
1049 		goto err_unregister_display;
1050 
1051 	err = xe_device_sysfs_init(xe);
1052 	if (err)
1053 		goto err_unregister_display;
1054 
1055 	xe_debugfs_register(xe);
1056 
1057 	err = xe_hwmon_register(xe);
1058 	if (err)
1059 		goto err_unregister_display;
1060 
1061 	err = xe_i2c_probe(xe);
1062 	if (err)
1063 		goto err_unregister_display;
1064 
1065 	for_each_gt(gt, xe, id)
1066 		xe_gt_sanitize_freq(gt);
1067 
1068 	xe_vsec_init(xe);
1069 
1070 	err = xe_sriov_init_late(xe);
1071 	if (err)
1072 		goto err_unregister_display;
1073 
1074 	detect_preproduction_hw(xe);
1075 
1076 	err = drmm_add_action_or_reset(&xe->drm, xe_device_wedged_fini, xe);
1077 	if (err)
1078 		goto err_unregister_display;
1079 
1080 	return devm_add_action_or_reset(xe->drm.dev, xe_device_sanitize, xe);
1081 
1082 err_unregister_display:
1083 	xe_display_unregister(xe);
1084 	drm_dev_unregister(&xe->drm);
1085 
1086 	return err;
1087 }
1088 
1089 void xe_device_remove(struct xe_device *xe)
1090 {
1091 	xe_display_unregister(xe);
1092 
1093 	drm_dev_unplug(&xe->drm);
1094 
1095 	xe_bo_pci_dev_remove_all(xe);
1096 }
1097 
1098 void xe_device_shutdown(struct xe_device *xe)
1099 {
1100 	struct xe_gt *gt;
1101 	u8 id;
1102 
1103 	drm_dbg(&xe->drm, "Shutting down device\n");
1104 
1105 	xe_display_pm_shutdown(xe);
1106 
1107 	xe_irq_suspend(xe);
1108 
1109 	for_each_gt(gt, xe, id)
1110 		xe_gt_shutdown(gt);
1111 
1112 	xe_display_pm_shutdown_late(xe);
1113 
1114 	if (!xe_driver_flr_disabled(xe)) {
1115 		/* BOOM! */
1116 		__xe_driver_flr(xe);
1117 	}
1118 }
1119 
1120 /**
1121  * xe_device_wmb() - Device specific write memory barrier
1122  * @xe: the &xe_device
1123  *
1124  * While wmb() is sufficient for a barrier if we use system memory, on discrete
1125  * platforms with device memory we additionally need to issue a register write.
1126  * Since it doesn't matter which register we write to, use the read-only VF_CAP
1127  * register that is also marked as accessible by the VFs.
1128  */
1129 void xe_device_wmb(struct xe_device *xe)
1130 {
1131 	wmb();
1132 	if (IS_DGFX(xe))
1133 		xe_mmio_write32(xe_root_tile_mmio(xe), VF_CAP_REG, 0);
1134 }
1135 
1136 /*
1137  * Issue a TRANSIENT_FLUSH_REQUEST and wait for completion on each gt.
1138  */
1139 static void tdf_request_sync(struct xe_device *xe)
1140 {
1141 	struct xe_gt *gt;
1142 	u8 id;
1143 
1144 	for_each_gt_with_type(gt, xe, id, BIT(XE_GT_TYPE_MAIN)) {
1145 		CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
1146 		if (!fw_ref.domains)
1147 			return;
1148 
1149 		xe_mmio_write32(&gt->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST);
1150 
1151 		/*
1152 		 * FIXME: We can likely do better here with our choice of
1153 		 * timeout. Currently we just assume the worst case, i.e. 150us,
1154 		 * which is believed to be sufficient to cover the worst case
1155 		 * scenario on current platforms if all cache entries are
1156 		 * transient and need to be flushed..
1157 		 */
1158 		if (xe_mmio_wait32(&gt->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST, 0,
1159 				   300, NULL, false))
1160 			xe_gt_err_once(gt, "TD flush timeout\n");
1161 	}
1162 }
1163 
1164 /**
1165  * xe_device_is_l2_flush_optimized - if L2 flush is optimized by HW
1166  * @xe: The device to check.
1167  *
1168  * Return: true if the HW device optimizing L2 flush, false otherwise.
1169  */
1170 bool xe_device_is_l2_flush_optimized(struct xe_device *xe)
1171 {
1172 	/* XA is *always* flushed, like at the end-of-submssion (and maybe other
1173 	 * places), just that internally as an optimisation hw doesn't need to make
1174 	 * that a full flush (which will also include XA) when Media is
1175 	 * off/powergated, since it doesn't need to worry about GT caches vs Media
1176 	 * coherency, and only CPU vs GPU coherency, so can make that flush a
1177 	 * targeted XA flush, since stuff tagged with XA now means it's shared with
1178 	 * the CPU. The main implication is that we now need to somehow flush non-XA before
1179 	 * freeing system memory pages, otherwise dirty cachelines could be flushed after the free
1180 	 * (like if Media suddenly turns on and does a full flush)
1181 	 */
1182 	if (GRAPHICS_VER(xe) >= 35 && !IS_DGFX(xe))
1183 		return true;
1184 	return false;
1185 }
1186 
1187 void xe_device_l2_flush(struct xe_device *xe)
1188 {
1189 	struct xe_gt *gt;
1190 
1191 	gt = xe_root_mmio_gt(xe);
1192 	if (!gt)
1193 		return;
1194 
1195 	if (!XE_GT_WA(gt, 16023588340))
1196 		return;
1197 
1198 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
1199 	if (!fw_ref.domains)
1200 		return;
1201 
1202 	spin_lock(&gt->global_invl_lock);
1203 
1204 	xe_mmio_write32(&gt->mmio, XE2_GLOBAL_INVAL, 0x1);
1205 	if (xe_mmio_wait32(&gt->mmio, XE2_GLOBAL_INVAL, 0x1, 0x0, 1000, NULL, true))
1206 		xe_gt_err_once(gt, "Global invalidation timeout\n");
1207 
1208 	spin_unlock(&gt->global_invl_lock);
1209 }
1210 
1211 /**
1212  * xe_device_td_flush() - Flush transient L3 cache entries
1213  * @xe: The device
1214  *
1215  * Display engine has direct access to memory and is never coherent with L3/L4
1216  * caches (or CPU caches), however KMD is responsible for specifically flushing
1217  * transient L3 GPU cache entries prior to the flip sequence to ensure scanout
1218  * can happen from such a surface without seeing corruption.
1219  *
1220  * Display surfaces can be tagged as transient by mapping it using one of the
1221  * various L3:XD PAT index modes on Xe2.
1222  *
1223  * Note: On non-discrete xe2 platforms, like LNL, the entire L3 cache is flushed
1224  * at the end of each submission via PIPE_CONTROL for compute/render, since SA
1225  * Media is not coherent with L3 and we want to support render-vs-media
1226  * usescases. For other engines like copy/blt the HW internally forces uncached
1227  * behaviour, hence why we can skip the TDF on such platforms.
1228  */
1229 void xe_device_td_flush(struct xe_device *xe)
1230 {
1231 	struct xe_gt *root_gt;
1232 
1233 	/*
1234 	 * From Xe3p onward the HW takes care of flush of TD entries also along
1235 	 * with flushing XA entries, which will be at the usual sync points,
1236 	 * like at the end of submission, so no manual flush is needed here.
1237 	 */
1238 	if (GRAPHICS_VER(xe) >= 35)
1239 		return;
1240 
1241 	if (!IS_DGFX(xe) || GRAPHICS_VER(xe) < 20)
1242 		return;
1243 
1244 	root_gt = xe_root_mmio_gt(xe);
1245 	if (!root_gt)
1246 		return;
1247 
1248 	if (XE_GT_WA(root_gt, 16023588340)) {
1249 		/* A transient flush is not sufficient: flush the L2 */
1250 		xe_device_l2_flush(xe);
1251 	} else {
1252 		xe_guc_pc_apply_flush_freq_limit(&root_gt->uc.guc.pc);
1253 		tdf_request_sync(xe);
1254 		xe_guc_pc_remove_flush_freq_limit(&root_gt->uc.guc.pc);
1255 	}
1256 }
1257 
1258 u32 xe_device_ccs_bytes(struct xe_device *xe, u64 size)
1259 {
1260 	return xe_device_has_flat_ccs(xe) ?
1261 		DIV_ROUND_UP_ULL(size, NUM_BYTES_PER_CCS_BYTE(xe)) : 0;
1262 }
1263 
1264 /**
1265  * xe_device_assert_mem_access - Inspect the current runtime_pm state.
1266  * @xe: xe device instance
1267  *
1268  * To be used before any kind of memory access. It will splat a debug warning
1269  * if the device is currently sleeping. But it doesn't guarantee in any way
1270  * that the device is going to remain awake. Xe PM runtime get and put
1271  * functions might be added to the outer bound of the memory access, while
1272  * this check is intended for inner usage to splat some warning if the worst
1273  * case has just happened.
1274  */
1275 void xe_device_assert_mem_access(struct xe_device *xe)
1276 {
1277 	xe_assert(xe, !xe_pm_runtime_suspended(xe));
1278 }
1279 
1280 void xe_device_snapshot_print(struct xe_device *xe, struct drm_printer *p)
1281 {
1282 	struct xe_gt *gt;
1283 	u8 id;
1284 
1285 	drm_printf(p, "PCI ID: 0x%04x\n", xe->info.devid);
1286 	drm_printf(p, "PCI revision: 0x%02x\n", xe->info.revid);
1287 
1288 	for_each_gt(gt, xe, id) {
1289 		drm_printf(p, "GT id: %u\n", id);
1290 		drm_printf(p, "\tTile: %u\n", gt->tile->id);
1291 		drm_printf(p, "\tType: %s\n",
1292 			   gt->info.type == XE_GT_TYPE_MAIN ? "main" : "media");
1293 		drm_printf(p, "\tIP ver: %u.%u.%u\n",
1294 			   REG_FIELD_GET(GMD_ID_ARCH_MASK, gt->info.gmdid),
1295 			   REG_FIELD_GET(GMD_ID_RELEASE_MASK, gt->info.gmdid),
1296 			   REG_FIELD_GET(GMD_ID_REVID, gt->info.gmdid));
1297 		drm_printf(p, "\tCS reference clock: %u\n", gt->info.reference_clock);
1298 	}
1299 }
1300 
1301 u64 xe_device_canonicalize_addr(struct xe_device *xe, u64 address)
1302 {
1303 	return sign_extend64(address, xe->info.va_bits - 1);
1304 }
1305 
1306 u64 xe_device_uncanonicalize_addr(struct xe_device *xe, u64 address)
1307 {
1308 	return address & GENMASK_ULL(xe->info.va_bits - 1, 0);
1309 }
1310 
1311 /**
1312  * DOC: Xe Device Wedging
1313  *
1314  * Xe driver uses drm device wedged uevent as documented in Documentation/gpu/drm-uapi.rst.
1315  * When device is in wedged state, every IOCTL will be blocked and GT cannot
1316  * be used. The conditions under which the driver declares the device wedged
1317  * depend on the wedged mode configuration (see &enum xe_wedged_mode). The
1318  * default recovery method for a wedged state is rebind/bus-reset.
1319  *
1320  * Another recovery method is vendor-specific. Below are the cases that send
1321  * ``WEDGED=vendor-specific`` recovery method in drm device wedged uevent.
1322  *
1323  * Case: Firmware Flash
1324  * --------------------
1325  *
1326  * Identification Hint
1327  * +++++++++++++++++++
1328  *
1329  * ``WEDGED=vendor-specific`` drm device wedged uevent with
1330  * :ref:`Runtime Survivability mode <xe-survivability-mode>` is used to notify
1331  * admin/userspace consumer about the need for a firmware flash.
1332  *
1333  * Recovery Procedure
1334  * ++++++++++++++++++
1335  *
1336  * Once ``WEDGED=vendor-specific`` drm device wedged uevent is received, follow
1337  * the below steps
1338  *
1339  * - Check Runtime Survivability mode sysfs.
1340  *   If enabled, firmware flash is required to recover the device.
1341  *
1342  *   /sys/bus/pci/devices/<device>/survivability_mode
1343  *
1344  * - Admin/userspace consumer can use firmware flashing tools like fwupd to flash
1345  *   firmware and restore device to normal operation.
1346  */
1347 
1348 /**
1349  * xe_device_set_wedged_method - Set wedged recovery method
1350  * @xe: xe device instance
1351  * @method: recovery method to set
1352  *
1353  * Set wedged recovery method to be sent in drm wedged uevent.
1354  */
1355 void xe_device_set_wedged_method(struct xe_device *xe, unsigned long method)
1356 {
1357 	xe->wedged.method = method;
1358 }
1359 
1360 /**
1361  * xe_device_declare_wedged - Declare device wedged
1362  * @xe: xe device instance
1363  *
1364  * This is a final state that can only be cleared with the recovery method
1365  * specified in the drm wedged uevent. The method can be set using
1366  * xe_device_set_wedged_method before declaring the device as wedged. If no method
1367  * is set, reprobe (unbind/re-bind) will be sent by default.
1368  *
1369  * In this state every IOCTL will be blocked so the GT cannot be used.
1370  * In general it will be called upon any critical error such as gt reset
1371  * failure or guc loading failure. Userspace will be notified of this state
1372  * through device wedged uevent.
1373  * If xe.wedged module parameter is set to 2, this function will be called
1374  * on every single execution timeout (a.k.a. GPU hang) right after devcoredump
1375  * snapshot capture. In this mode, GT reset won't be attempted so the state of
1376  * the issue is preserved for further debugging.
1377  */
1378 void xe_device_declare_wedged(struct xe_device *xe)
1379 {
1380 	struct xe_gt *gt;
1381 	u8 id;
1382 
1383 	if (xe->wedged.mode == XE_WEDGED_MODE_NEVER) {
1384 		drm_dbg(&xe->drm, "Wedged mode is forcibly disabled\n");
1385 		return;
1386 	}
1387 
1388 	if (!atomic_xchg(&xe->wedged.flag, 1)) {
1389 		xe->needs_flr_on_fini = true;
1390 		xe_pm_runtime_get_noresume(xe);
1391 		drm_err(&xe->drm,
1392 			"CRITICAL: Xe has declared device %s as wedged.\n"
1393 			"IOCTLs and executions are blocked.\n"
1394 			"For recovery procedure, refer to https://docs.kernel.org/gpu/drm-uapi.html#device-wedging\n"
1395 			"Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/xe/kernel/issues/new\n",
1396 			dev_name(xe->drm.dev));
1397 	}
1398 
1399 	for_each_gt(gt, xe, id)
1400 		xe_gt_declare_wedged(gt);
1401 
1402 	if (xe_device_wedged(xe)) {
1403 		/*
1404 		 * XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET is intended for debugging
1405 		 * hangs, so wedge the device with 'none' recovery method and have
1406 		 * it available to the user for debugging.
1407 		 */
1408 		if (xe->wedged.mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET)
1409 			xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_NONE);
1410 		/* If no wedge recovery method is set, use default */
1411 		else if (!xe->wedged.method)
1412 			xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_REBIND |
1413 						    DRM_WEDGE_RECOVERY_BUS_RESET);
1414 
1415 		/* Notify userspace of wedged device */
1416 		drm_dev_wedged_event(&xe->drm, xe->wedged.method, NULL);
1417 	}
1418 }
1419 
1420 /**
1421  * xe_device_validate_wedged_mode - Check if given mode is supported
1422  * @xe: the &xe_device
1423  * @mode: requested mode to validate
1424  *
1425  * Check whether the provided wedged mode is supported.
1426  *
1427  * Return: 0 if mode is supported, error code otherwise.
1428  */
1429 int xe_device_validate_wedged_mode(struct xe_device *xe, unsigned int mode)
1430 {
1431 	if (mode > XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET) {
1432 		drm_dbg(&xe->drm, "wedged_mode: invalid value (%u)\n", mode);
1433 		return -EINVAL;
1434 	} else if (mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET && (IS_SRIOV_VF(xe) ||
1435 		   (IS_SRIOV_PF(xe) && !IS_ENABLED(CONFIG_DRM_XE_DEBUG)))) {
1436 		drm_dbg(&xe->drm, "wedged_mode: (%u) %s mode is not supported for %s\n",
1437 			mode, xe_wedged_mode_to_string(mode),
1438 			xe_sriov_mode_to_string(xe_device_sriov_mode(xe)));
1439 		return -EPERM;
1440 	}
1441 
1442 	return 0;
1443 }
1444 
1445 /**
1446  * xe_wedged_mode_to_string - Convert enum value to string.
1447  * @mode: the &xe_wedged_mode to convert
1448  *
1449  * Returns: wedged mode as a user friendly string.
1450  */
1451 const char *xe_wedged_mode_to_string(enum xe_wedged_mode mode)
1452 {
1453 	switch (mode) {
1454 	case XE_WEDGED_MODE_NEVER:
1455 		return "never";
1456 	case XE_WEDGED_MODE_UPON_CRITICAL_ERROR:
1457 		return "upon-critical-error";
1458 	case XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET:
1459 		return "upon-any-hang-no-reset";
1460 	default:
1461 		return "<invalid>";
1462 	}
1463 }
1464 
1465 /**
1466  * xe_device_asid_to_vm() - Find VM from ASID
1467  * @xe: the &xe_device
1468  * @asid: Address space ID
1469  *
1470  * Find a VM from ASID and take a reference to VM which caller must drop.
1471  * Reclaim safe.
1472  *
1473  * Return: VM on success, ERR_PTR on failure
1474  */
1475 struct xe_vm *xe_device_asid_to_vm(struct xe_device *xe, u32 asid)
1476 {
1477 	struct xe_vm *vm;
1478 
1479 	down_read(&xe->usm.lock);
1480 	vm = xa_load(&xe->usm.asid_to_vm, asid);
1481 	if (vm)
1482 		xe_vm_get(vm);
1483 	else
1484 		vm = ERR_PTR(-EINVAL);
1485 	up_read(&xe->usm.lock);
1486 
1487 	return vm;
1488 }
1489