xref: /linux/drivers/gpu/drm/xe/xe_device.c (revision d639d9fa162aadec1ae9980c4dcf6e50bd2f8290)
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_client.h>
14 #include <drm/drm_gem_ttm_helper.h>
15 #include <drm/drm_ioctl.h>
16 #include <drm/drm_managed.h>
17 #include <drm/drm_pagemap_util.h>
18 #include <drm/drm_print.h>
19 #include <kunit/static_stub.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_configfs.h"
29 #include "xe_debugfs.h"
30 #include "xe_defaults.h"
31 #include "xe_devcoredump.h"
32 #include "xe_device_sysfs.h"
33 #include "xe_dma_buf.h"
34 #include "xe_drm_client.h"
35 #include "xe_drv.h"
36 #include "xe_exec.h"
37 #include "xe_exec_queue.h"
38 #include "xe_force_wake.h"
39 #include "xe_ggtt.h"
40 #include "xe_gt.h"
41 #include "xe_gt_mcr.h"
42 #include "xe_gt_printk.h"
43 #include "xe_gt_sriov_vf.h"
44 #include "xe_guc.h"
45 #include "xe_guc_pc.h"
46 #include "xe_hw_engine_group.h"
47 #include "xe_hwmon.h"
48 #include "xe_i2c.h"
49 #include "xe_irq.h"
50 #include "xe_late_bind_fw.h"
51 #include "xe_mmio.h"
52 #include "xe_module.h"
53 #include "xe_nvm.h"
54 #include "xe_oa.h"
55 #include "xe_observation.h"
56 #include "xe_pagefault.h"
57 #include "xe_pat.h"
58 #include "xe_pcode.h"
59 #include "xe_pm.h"
60 #include "xe_pmu.h"
61 #include "xe_psmi.h"
62 #include "xe_pxp.h"
63 #include "xe_query.h"
64 #include "xe_shrinker.h"
65 #include "xe_soc_remapper.h"
66 #include "xe_survivability_mode.h"
67 #include "xe_sriov.h"
68 #include "xe_svm.h"
69 #include "xe_sysctrl.h"
70 #include "xe_tile.h"
71 #include "xe_ttm_stolen_mgr.h"
72 #include "xe_ttm_sys_mgr.h"
73 #include "xe_vm.h"
74 #include "xe_vm_madvise.h"
75 #include "xe_vram.h"
76 #include "xe_vram_types.h"
77 #include "xe_vsec.h"
78 #include "xe_wait_user_fence.h"
79 #include "xe_wa.h"
80 
81 #include <generated/xe_device_wa_oob.h>
82 #include <generated/xe_wa_oob.h>
83 
84 static int xe_file_open(struct drm_device *dev, struct drm_file *file)
85 {
86 	struct xe_device *xe = to_xe_device(dev);
87 	struct xe_drm_client *client;
88 	struct xe_file *xef;
89 	int ret = -ENOMEM;
90 	struct task_struct *task = NULL;
91 
92 	xef = kzalloc_obj(*xef);
93 	if (!xef)
94 		return ret;
95 
96 	client = xe_drm_client_alloc();
97 	if (!client) {
98 		kfree(xef);
99 		return ret;
100 	}
101 
102 	xef->drm = file;
103 	xef->client = client;
104 	xef->xe = xe;
105 
106 	mutex_init(&xef->vm.lock);
107 	xa_init_flags(&xef->vm.xa, XA_FLAGS_ALLOC1);
108 
109 	mutex_init(&xef->exec_queue.lock);
110 	xa_init_flags(&xef->exec_queue.xa, XA_FLAGS_ALLOC1);
111 
112 	file->driver_priv = xef;
113 	kref_init(&xef->refcount);
114 
115 	task = get_pid_task(rcu_access_pointer(file->pid), PIDTYPE_PID);
116 	if (task) {
117 		xef->process_name = kstrdup(task->comm, GFP_KERNEL);
118 		xef->pid = task->pid;
119 		put_task_struct(task);
120 	}
121 
122 	return 0;
123 }
124 
125 static void xe_file_destroy(struct kref *ref)
126 {
127 	struct xe_file *xef = container_of(ref, struct xe_file, refcount);
128 
129 	xa_destroy(&xef->exec_queue.xa);
130 	mutex_destroy(&xef->exec_queue.lock);
131 	xa_destroy(&xef->vm.xa);
132 	mutex_destroy(&xef->vm.lock);
133 
134 	xe_drm_client_put(xef->client);
135 	kfree(xef->process_name);
136 	kfree(xef);
137 }
138 
139 /**
140  * xe_file_get() - Take a reference to the xe file object
141  * @xef: Pointer to the xe file
142  *
143  * Anyone with a pointer to xef must take a reference to the xe file
144  * object using this call.
145  *
146  * Return: xe file pointer
147  */
148 struct xe_file *xe_file_get(struct xe_file *xef)
149 {
150 	kref_get(&xef->refcount);
151 	return xef;
152 }
153 
154 /**
155  * xe_file_put() - Drop a reference to the xe file object
156  * @xef: Pointer to the xe file
157  *
158  * Used to drop reference to the xef object
159  */
160 void xe_file_put(struct xe_file *xef)
161 {
162 	kref_put(&xef->refcount, xe_file_destroy);
163 }
164 
165 static void xe_file_close(struct drm_device *dev, struct drm_file *file)
166 {
167 	struct xe_device *xe = to_xe_device(dev);
168 	struct xe_file *xef = file->driver_priv;
169 	struct xe_vm *vm;
170 	struct xe_exec_queue *q;
171 	unsigned long idx;
172 
173 	guard(xe_pm_runtime)(xe);
174 
175 	/*
176 	 * No need for exec_queue.lock here as there is no contention for it
177 	 * when FD is closing as IOCTLs presumably can't be modifying the
178 	 * xarray. Taking exec_queue.lock here causes undue dependency on
179 	 * vm->lock taken during xe_exec_queue_kill().
180 	 */
181 	xa_for_each(&xef->exec_queue.xa, idx, q) {
182 		if (q->vm && q->hwe->hw_engine_group)
183 			xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q);
184 		xe_exec_queue_kill(q);
185 		xe_exec_queue_put(q);
186 	}
187 	xa_for_each(&xef->vm.xa, idx, vm)
188 		xe_vm_close_and_put(vm);
189 
190 	xe_file_put(xef);
191 }
192 
193 static const struct drm_ioctl_desc xe_ioctls[] = {
194 	DRM_IOCTL_DEF_DRV(XE_DEVICE_QUERY, xe_query_ioctl, DRM_RENDER_ALLOW),
195 	DRM_IOCTL_DEF_DRV(XE_GEM_CREATE, xe_gem_create_ioctl, DRM_RENDER_ALLOW),
196 	DRM_IOCTL_DEF_DRV(XE_GEM_MMAP_OFFSET, xe_gem_mmap_offset_ioctl,
197 			  DRM_RENDER_ALLOW),
198 	DRM_IOCTL_DEF_DRV(XE_VM_CREATE, xe_vm_create_ioctl, DRM_RENDER_ALLOW),
199 	DRM_IOCTL_DEF_DRV(XE_VM_DESTROY, xe_vm_destroy_ioctl, DRM_RENDER_ALLOW),
200 	DRM_IOCTL_DEF_DRV(XE_VM_BIND, xe_vm_bind_ioctl, DRM_RENDER_ALLOW),
201 	DRM_IOCTL_DEF_DRV(XE_EXEC, xe_exec_ioctl, DRM_RENDER_ALLOW),
202 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_CREATE, xe_exec_queue_create_ioctl,
203 			  DRM_RENDER_ALLOW),
204 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_DESTROY, xe_exec_queue_destroy_ioctl,
205 			  DRM_RENDER_ALLOW),
206 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_GET_PROPERTY, xe_exec_queue_get_property_ioctl,
207 			  DRM_RENDER_ALLOW),
208 	DRM_IOCTL_DEF_DRV(XE_WAIT_USER_FENCE, xe_wait_user_fence_ioctl,
209 			  DRM_RENDER_ALLOW),
210 	DRM_IOCTL_DEF_DRV(XE_OBSERVATION, xe_observation_ioctl, DRM_RENDER_ALLOW),
211 	DRM_IOCTL_DEF_DRV(XE_MADVISE, xe_vm_madvise_ioctl, DRM_RENDER_ALLOW),
212 	DRM_IOCTL_DEF_DRV(XE_VM_QUERY_MEM_RANGE_ATTRS, xe_vm_query_vmas_attrs_ioctl,
213 			  DRM_RENDER_ALLOW),
214 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_SET_PROPERTY, xe_exec_queue_set_property_ioctl,
215 			  DRM_RENDER_ALLOW),
216 	DRM_IOCTL_DEF_DRV(XE_VM_GET_PROPERTY, xe_vm_get_property_ioctl,
217 			  DRM_RENDER_ALLOW),
218 };
219 
220 static long xe_drm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
221 {
222 	struct drm_file *file_priv = file->private_data;
223 	struct xe_device *xe = to_xe_device(file_priv->minor->dev);
224 	long ret;
225 
226 	if (xe_device_wedged(xe))
227 		return -ECANCELED;
228 
229 	ACQUIRE(xe_pm_runtime_ioctl, pm)(xe);
230 	ret = ACQUIRE_ERR(xe_pm_runtime_ioctl, &pm);
231 	if (ret >= 0)
232 		ret = drm_ioctl(file, cmd, arg);
233 
234 	return ret;
235 }
236 
237 #ifdef CONFIG_COMPAT
238 static long xe_drm_compat_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
239 {
240 	struct drm_file *file_priv = file->private_data;
241 	struct xe_device *xe = to_xe_device(file_priv->minor->dev);
242 	long ret;
243 
244 	if (xe_device_wedged(xe))
245 		return -ECANCELED;
246 
247 	ACQUIRE(xe_pm_runtime_ioctl, pm)(xe);
248 	ret = ACQUIRE_ERR(xe_pm_runtime_ioctl, &pm);
249 	if (ret >= 0)
250 		ret = drm_compat_ioctl(file, cmd, arg);
251 
252 	return ret;
253 }
254 #else
255 /* similarly to drm_compat_ioctl, let's it be assigned to .compat_ioct unconditionally */
256 #define xe_drm_compat_ioctl NULL
257 #endif
258 
259 static void barrier_open(struct vm_area_struct *vma)
260 {
261 	drm_dev_get(vma->vm_private_data);
262 }
263 
264 static void barrier_close(struct vm_area_struct *vma)
265 {
266 	drm_dev_put(vma->vm_private_data);
267 }
268 
269 static void barrier_release_dummy_page(struct drm_device *dev, void *res)
270 {
271 	struct page *dummy_page = (struct page *)res;
272 
273 	__free_page(dummy_page);
274 }
275 
276 static vm_fault_t barrier_fault(struct vm_fault *vmf)
277 {
278 	struct drm_device *dev = vmf->vma->vm_private_data;
279 	struct vm_area_struct *vma = vmf->vma;
280 	vm_fault_t ret = VM_FAULT_NOPAGE;
281 	pgprot_t prot;
282 	int idx;
283 
284 	prot = vm_get_page_prot(vma->vm_flags);
285 
286 	if (drm_dev_enter(dev, &idx)) {
287 		unsigned long pfn;
288 
289 #define LAST_DB_PAGE_OFFSET 0x7ff001
290 		pfn = PHYS_PFN(pci_resource_start(to_pci_dev(dev->dev), 0) +
291 				LAST_DB_PAGE_OFFSET);
292 		ret = vmf_insert_pfn_prot(vma, vma->vm_start, pfn,
293 					  pgprot_noncached(prot));
294 		drm_dev_exit(idx);
295 	} else {
296 		struct page *page;
297 
298 		/* Allocate new dummy page to map all the VA range in this VMA to it*/
299 		page = alloc_page(GFP_KERNEL | __GFP_ZERO);
300 		if (!page)
301 			return VM_FAULT_OOM;
302 
303 		/* Set the page to be freed using drmm release action */
304 		if (drmm_add_action_or_reset(dev, barrier_release_dummy_page, page))
305 			return VM_FAULT_OOM;
306 
307 		ret = vmf_insert_pfn_prot(vma, vma->vm_start, page_to_pfn(page),
308 					  prot);
309 	}
310 
311 	return ret;
312 }
313 
314 static const struct vm_operations_struct vm_ops_barrier = {
315 	.open = barrier_open,
316 	.close = barrier_close,
317 	.fault = barrier_fault,
318 };
319 
320 static int xe_pci_barrier_mmap(struct file *filp,
321 			       struct vm_area_struct *vma)
322 {
323 	struct drm_file *priv = filp->private_data;
324 	struct drm_device *dev = priv->minor->dev;
325 	struct xe_device *xe = to_xe_device(dev);
326 
327 	if (!IS_DGFX(xe))
328 		return -EINVAL;
329 
330 	if (vma->vm_end - vma->vm_start > SZ_4K)
331 		return -EINVAL;
332 
333 	if (is_cow_mapping(vma->vm_flags))
334 		return -EINVAL;
335 
336 	if (vma->vm_flags & (VM_READ | VM_EXEC))
337 		return -EINVAL;
338 
339 	vm_flags_clear(vma, VM_MAYREAD | VM_MAYEXEC);
340 	vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP | VM_IO);
341 	vma->vm_ops = &vm_ops_barrier;
342 	vma->vm_private_data = dev;
343 	drm_dev_get(vma->vm_private_data);
344 
345 	return 0;
346 }
347 
348 static int xe_mmap(struct file *filp, struct vm_area_struct *vma)
349 {
350 	struct drm_file *priv = filp->private_data;
351 	struct drm_device *dev = priv->minor->dev;
352 
353 	if (drm_dev_is_unplugged(dev))
354 		return -ENODEV;
355 
356 	switch (vma->vm_pgoff) {
357 	case XE_PCI_BARRIER_MMAP_OFFSET >> XE_PTE_SHIFT:
358 		return xe_pci_barrier_mmap(filp, vma);
359 	}
360 
361 	return drm_gem_mmap(filp, vma);
362 }
363 
364 static const struct file_operations xe_driver_fops = {
365 	.owner = THIS_MODULE,
366 	.open = drm_open,
367 	.release = drm_release_noglobal,
368 	.unlocked_ioctl = xe_drm_ioctl,
369 	.mmap = xe_mmap,
370 	.poll = drm_poll,
371 	.read = drm_read,
372 	.compat_ioctl = xe_drm_compat_ioctl,
373 	.llseek = noop_llseek,
374 #ifdef CONFIG_PROC_FS
375 	.show_fdinfo = drm_show_fdinfo,
376 #endif
377 	.fop_flags = FOP_UNSIGNED_OFFSET,
378 };
379 
380 /**
381  * xe_is_xe_file() - Is the file an xe device file?
382  * @file: The file.
383  *
384  * Checks whether the file is opened against
385  * an xe device.
386  *
387  * Return: %true if an xe file, %false if not.
388  */
389 bool xe_is_xe_file(const struct file *file)
390 {
391 	return file->f_op == &xe_driver_fops;
392 }
393 
394 static const struct drm_driver regular_driver = {
395 	.driver_features =
396 	    XE_DISPLAY_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 	XE_DISPLAY_DRIVER_OPS,
419 };
420 
421 #ifdef CONFIG_PCI_IOV
422 static const struct drm_ioctl_desc xe_ioctls_admin_only[] = {
423 	DRM_IOCTL_DEF_DRV(XE_DEVICE_QUERY, xe_query_ioctl, DRM_RENDER_ALLOW),
424 	DRM_IOCTL_DEF_DRV(XE_OBSERVATION, xe_observation_ioctl, DRM_RENDER_ALLOW),
425 };
426 
427 static const struct drm_driver admin_only_driver = {
428 	.driver_features =
429 	    XE_DISPLAY_DRIVER_FEATURES |
430 	    DRIVER_GEM | DRIVER_RENDER | DRIVER_GEM_GPUVA,
431 	.open = xe_file_open,
432 	.postclose = xe_file_close,
433 	.ioctls = xe_ioctls_admin_only,
434 	.num_ioctls = ARRAY_SIZE(xe_ioctls_admin_only),
435 	.fops = &xe_driver_fops,
436 	.name = DRIVER_NAME,
437 	.desc = DRIVER_DESC,
438 	.major = DRIVER_MAJOR,
439 	.minor = DRIVER_MINOR,
440 	.patchlevel = DRIVER_PATCHLEVEL,
441 	XE_DISPLAY_DRIVER_OPS,
442 };
443 
444 /**
445  * xe_device_is_admin_only() - Check whether device is admin only or not.
446  * @xe: the &xe_device to check
447  *
448  * Return: true if the device is admin only, false otherwise.
449  */
450 bool xe_device_is_admin_only(const struct xe_device *xe)
451 {
452 	KUNIT_STATIC_STUB_REDIRECT(xe_device_is_admin_only, xe);
453 	return xe->drm.driver == &admin_only_driver;
454 }
455 #endif
456 
457 static void xe_device_destroy(struct drm_device *dev, void *dummy)
458 {
459 	struct xe_device *xe = to_xe_device(dev);
460 
461 	xe_bo_dev_fini(&xe->bo_device);
462 
463 	if (xe->preempt_fence_wq)
464 		destroy_workqueue(xe->preempt_fence_wq);
465 
466 	if (xe->ordered_wq)
467 		destroy_workqueue(xe->ordered_wq);
468 
469 	if (xe->unordered_wq)
470 		destroy_workqueue(xe->unordered_wq);
471 
472 	if (xe->destroy_wq)
473 		destroy_workqueue(xe->destroy_wq);
474 
475 	ttm_device_fini(&xe->ttm);
476 }
477 
478 /**
479  * xe_device_create() - Create a new &xe_device instance
480  * @pdev: the parent &pci_dev
481  *
482  * Allocate and initialize a device managed Xe device structure.
483  *
484  * Return: pointer to new &xe_device on success, or ERR_PTR on failure.
485  */
486 struct xe_device *xe_device_create(struct pci_dev *pdev)
487 {
488 	const struct drm_driver *driver = &regular_driver;
489 	struct xe_device *xe;
490 	int err;
491 
492 #ifdef CONFIG_PCI_IOV
493 	/*
494 	 * Since XE device is not initialized yet, read from configfs
495 	 * directly to decide whether we are in admin-only PF mode or not.
496 	 */
497 	if (xe_configfs_admin_only_pf(pdev))
498 		driver = &admin_only_driver;
499 #endif
500 
501 	err = aperture_remove_conflicting_pci_devices(pdev, driver->name);
502 	if (err)
503 		return ERR_PTR(err);
504 
505 	xe = devm_drm_dev_alloc(&pdev->dev, driver, struct xe_device, drm);
506 	if (IS_ERR(xe))
507 		return xe;
508 
509 	err = xe_device_init_early(xe);
510 	if (err)
511 		return ERR_PTR(err);
512 
513 	return xe;
514 }
515 ALLOW_ERROR_INJECTION(xe_device_create, ERRNO); /* See xe_pci_probe() */
516 
517 /**
518  * xe_device_init_early() - Initialize a new &xe_device instance
519  * @xe: the &xe_device to initialize
520  *
521  * Return: 0 on success or a negative error code on failure.
522  */
523 int xe_device_init_early(struct xe_device *xe)
524 {
525 	int err;
526 
527 	err = ttm_device_init(&xe->ttm, &xe_ttm_funcs, xe->drm.dev,
528 			      xe->drm.anon_inode->i_mapping,
529 			      xe->drm.vma_offset_manager, 0);
530 	if (err)
531 		return err;
532 
533 	xe_bo_dev_init(&xe->bo_device);
534 	err = drmm_add_action_or_reset(&xe->drm, xe_device_destroy, NULL);
535 	if (err)
536 		return err;
537 
538 	err = xe_shrinker_create(xe);
539 	if (err)
540 		return err;
541 
542 	xe->atomic_svm_timeslice_ms = 5;
543 	xe->min_run_period_lr_ms = 5;
544 
545 	err = xe_irq_init(xe);
546 	if (err)
547 		return err;
548 
549 	xe_validation_device_init(&xe->val);
550 
551 	init_waitqueue_head(&xe->ufence_wq);
552 
553 	init_rwsem(&xe->usm.lock);
554 
555 	err = xe_pagemap_shrinker_create(xe);
556 	if (err)
557 		return err;
558 
559 	xa_init_flags(&xe->usm.asid_to_vm, XA_FLAGS_ALLOC);
560 
561 	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) {
562 		/* Trigger a large asid and an early asid wrap. */
563 		u32 asid;
564 
565 		BUILD_BUG_ON(XE_MAX_ASID < 2);
566 		err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, NULL,
567 				      XA_LIMIT(XE_MAX_ASID - 2, XE_MAX_ASID - 1),
568 				      &xe->usm.next_asid, GFP_KERNEL);
569 		drm_WARN_ON(&xe->drm, err);
570 		if (err >= 0)
571 			xa_erase(&xe->usm.asid_to_vm, asid);
572 	}
573 
574 	err = xe_bo_pinned_init(xe);
575 	if (err)
576 		return err;
577 
578 	xe->preempt_fence_wq = alloc_ordered_workqueue("xe-preempt-fence-wq",
579 						       WQ_MEM_RECLAIM);
580 	xe->ordered_wq = alloc_ordered_workqueue("xe-ordered-wq", 0);
581 	xe->unordered_wq = alloc_workqueue("xe-unordered-wq", WQ_PERCPU, 0);
582 	xe->destroy_wq = alloc_workqueue("xe-destroy-wq", WQ_PERCPU, 0);
583 	if (!xe->ordered_wq || !xe->unordered_wq ||
584 	    !xe->preempt_fence_wq || !xe->destroy_wq) {
585 		/*
586 		 * Cleanup done in xe_device_destroy via
587 		 * drmm_add_action_or_reset register above
588 		 */
589 		drm_err(&xe->drm, "Failed to allocate xe workqueues\n");
590 		return -ENOMEM;
591 	}
592 
593 	err = drmm_mutex_init(&xe->drm, &xe->pmt.lock);
594 	if (err)
595 		return err;
596 
597 	err = xe_pm_init_early(xe);
598 	if (err)
599 		return err;
600 
601 	return 0;
602 }
603 
604 static bool xe_driver_flr_disabled(struct xe_device *xe)
605 {
606 	if (IS_SRIOV_VF(xe))
607 		return true;
608 
609 	if (xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL_PROTECTED) & DRIVERINT_FLR_DIS) {
610 		drm_info(&xe->drm, "Driver-FLR disabled by BIOS\n");
611 		return true;
612 	}
613 
614 	return false;
615 }
616 
617 /*
618  * The driver-initiated FLR is the highest level of reset that we can trigger
619  * from within the driver. It is different from the PCI FLR in that it doesn't
620  * fully reset the SGUnit and doesn't modify the PCI config space and therefore
621  * it doesn't require a re-enumeration of the PCI BARs. However, the
622  * driver-initiated FLR does still cause a reset of both GT and display and a
623  * memory wipe of local and stolen memory, so recovery would require a full HW
624  * re-init and saving/restoring (or re-populating) the wiped memory. Since we
625  * perform the FLR as the very last action before releasing access to the HW
626  * during the driver release flow, we don't attempt recovery at all, because
627  * if/when a new instance of Xe is bound to the device it will do a full
628  * re-init anyway.
629  */
630 static void __xe_driver_flr(struct xe_device *xe)
631 {
632 	const unsigned int flr_timeout = 3 * USEC_PER_SEC; /* specs recommend a 3s wait */
633 	struct xe_mmio *mmio = xe_root_tile_mmio(xe);
634 	int ret;
635 
636 	drm_dbg(&xe->drm, "Triggering Driver-FLR\n");
637 
638 	/*
639 	 * Make sure any pending FLR requests have cleared by waiting for the
640 	 * FLR trigger bit to go to zero. Also clear GU_DEBUG's DRIVERFLR_STATUS
641 	 * to make sure it's not still set from a prior attempt (it's a write to
642 	 * clear bit).
643 	 * Note that we should never be in a situation where a previous attempt
644 	 * is still pending (unless the HW is totally dead), but better to be
645 	 * safe in case something unexpected happens
646 	 */
647 	ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false);
648 	if (ret) {
649 		drm_err(&xe->drm, "Driver-FLR-prepare wait for ready failed! %d\n", ret);
650 		return;
651 	}
652 	xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS);
653 
654 	/* Trigger the actual Driver-FLR */
655 	xe_mmio_rmw32(mmio, GU_CNTL, 0, DRIVERFLR);
656 
657 	/* Wait for hardware teardown to complete */
658 	ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false);
659 	if (ret) {
660 		drm_err(&xe->drm, "Driver-FLR-teardown wait completion failed! %d\n", ret);
661 		return;
662 	}
663 
664 	/* Wait for hardware/firmware re-init to complete */
665 	ret = xe_mmio_wait32(mmio, GU_DEBUG, DRIVERFLR_STATUS, DRIVERFLR_STATUS,
666 			     flr_timeout, NULL, false);
667 	if (ret) {
668 		drm_err(&xe->drm, "Driver-FLR-reinit wait completion failed! %d\n", ret);
669 		return;
670 	}
671 
672 	/* Clear sticky completion status */
673 	xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS);
674 }
675 
676 static void xe_driver_flr(struct xe_device *xe)
677 {
678 	if (xe_driver_flr_disabled(xe))
679 		return;
680 
681 	__xe_driver_flr(xe);
682 }
683 
684 static void xe_driver_flr_fini(void *arg)
685 {
686 	struct xe_device *xe = arg;
687 
688 	if (xe->needs_flr_on_fini)
689 		xe_driver_flr(xe);
690 }
691 
692 static void xe_device_sanitize(void *arg)
693 {
694 	struct xe_device *xe = arg;
695 	struct xe_gt *gt;
696 	u8 id;
697 
698 	for_each_gt(gt, xe, id)
699 		xe_gt_sanitize(gt);
700 }
701 
702 static int xe_set_dma_info(struct xe_device *xe)
703 {
704 	unsigned int mask_size = xe->info.dma_mask_size;
705 	int err;
706 
707 	dma_set_max_seg_size(xe->drm.dev, xe_sg_segment_size(xe->drm.dev));
708 
709 	err = dma_set_mask(xe->drm.dev, DMA_BIT_MASK(mask_size));
710 	if (err)
711 		goto mask_err;
712 
713 	err = dma_set_coherent_mask(xe->drm.dev, DMA_BIT_MASK(mask_size));
714 	if (err)
715 		goto mask_err;
716 
717 	return 0;
718 
719 mask_err:
720 	drm_err(&xe->drm, "Can't set DMA mask/consistent mask (%d)\n", err);
721 	return err;
722 }
723 
724 static void assert_lmem_ready(struct xe_device *xe)
725 {
726 	if (!IS_DGFX(xe) || IS_SRIOV_VF(xe))
727 		return;
728 
729 	xe_assert(xe, xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL) &
730 		  LMEM_INIT);
731 }
732 
733 static void vf_update_device_info(struct xe_device *xe)
734 {
735 	xe_assert(xe, IS_SRIOV_VF(xe));
736 	/* disable features that are not available/applicable to VFs */
737 	xe->info.probe_display = 0;
738 	xe->info.has_heci_cscfi = 0;
739 	xe->info.has_heci_gscfi = 0;
740 	xe->info.has_late_bind = 0;
741 	xe->info.skip_guc_pc = 1;
742 	xe->info.skip_pcode = 1;
743 }
744 
745 static int xe_device_vram_alloc(struct xe_device *xe)
746 {
747 	struct xe_vram_region *vram;
748 
749 	if (!IS_DGFX(xe))
750 		return 0;
751 
752 	vram = drmm_kzalloc(&xe->drm, sizeof(*vram), GFP_KERNEL);
753 	if (!vram)
754 		return -ENOMEM;
755 
756 	xe->mem.vram = vram;
757 	return 0;
758 }
759 
760 /**
761  * xe_device_probe_early: Device early probe
762  * @xe: xe device instance
763  *
764  * Initialize MMIO resources that don't require any
765  * knowledge about tile count. Also initialize pcode and
766  * check vram initialization on root tile.
767  *
768  * Return: 0 on success, error code on failure
769  */
770 int xe_device_probe_early(struct xe_device *xe)
771 {
772 	int err;
773 
774 	xe_wa_device_init(xe);
775 	xe_wa_process_device_oob(xe);
776 
777 	err = xe_mmio_probe_early(xe);
778 	if (err)
779 		return err;
780 
781 	xe_sriov_probe_early(xe);
782 
783 	if (xe_device_is_admin_only(xe) && !IS_SRIOV_PF(xe)) {
784 		xe_err(xe, "Can't run Admin-only mode without SR-IOV PF mode!\n");
785 		return -ENODEV;
786 	}
787 
788 	if (IS_SRIOV_VF(xe))
789 		vf_update_device_info(xe);
790 
791 	/*
792 	 * Check for pcode uncore_init status to confirm if the SoC
793 	 * initialization is complete. Until done, any MMIO or lmem access from
794 	 * the driver will be blocked
795 	 */
796 	err = xe_pcode_probe_early(xe);
797 	if (err || xe_survivability_mode_is_requested(xe)) {
798 		int save_err = err;
799 
800 		/*
801 		 * Try to leave device in survivability mode if device is
802 		 * possible, but still return the previous error for error
803 		 * propagation
804 		 */
805 		err = xe_survivability_mode_boot_enable(xe);
806 		if (err)
807 			return err;
808 
809 		return save_err;
810 	}
811 
812 	/*
813 	 * Make sure the lmem is initialized and ready to use. xe_pcode_ready()
814 	 * is flagged after full initialization is complete. Assert if lmem is
815 	 * not initialized.
816 	 */
817 	assert_lmem_ready(xe);
818 
819 	xe->wedged.mode = xe_device_validate_wedged_mode(xe, xe_modparam.wedged_mode) ?
820 			  XE_DEFAULT_WEDGED_MODE : xe_modparam.wedged_mode;
821 	drm_dbg(&xe->drm, "wedged_mode: setting mode (%u) %s\n",
822 		xe->wedged.mode, xe_wedged_mode_to_string(xe->wedged.mode));
823 
824 	err = xe_device_vram_alloc(xe);
825 	if (err)
826 		return err;
827 
828 	return 0;
829 }
830 ALLOW_ERROR_INJECTION(xe_device_probe_early, ERRNO); /* See xe_pci_probe() */
831 
832 static int probe_has_flat_ccs(struct xe_device *xe)
833 {
834 	struct xe_gt *gt;
835 	u32 reg;
836 
837 	/* Always enabled/disabled, no runtime check to do */
838 	if (GRAPHICS_VER(xe) < 20 || !xe->info.has_flat_ccs || IS_SRIOV_VF(xe))
839 		return 0;
840 
841 	gt = xe_root_mmio_gt(xe);
842 	if (!gt)
843 		return 0;
844 
845 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
846 	if (!fw_ref.domains)
847 		return -ETIMEDOUT;
848 
849 	reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_LOWER);
850 	xe->info.has_flat_ccs = (reg & XE2_FLAT_CCS_ENABLE);
851 
852 	if (!xe->info.has_flat_ccs)
853 		drm_dbg(&xe->drm,
854 			"Flat CCS has been disabled in bios, May lead to performance impact");
855 
856 	return 0;
857 }
858 
859 /*
860  * Detect if the driver is being run on pre-production hardware.  We don't
861  * keep workarounds for pre-production hardware long term, so print an
862  * error and add taint if we're being loaded on a pre-production platform
863  * for which the pre-prod workarounds have already been removed.
864  *
865  * The general policy is that we'll remove any workarounds that only apply to
866  * pre-production hardware around the time force_probe restrictions are lifted
867  * for a platform of the next major IP generation (for example, Xe2 pre-prod
868  * workarounds should be removed around the time the first Xe3 platforms have
869  * force_probe lifted).
870  */
871 static void detect_preproduction_hw(struct xe_device *xe)
872 {
873 	struct xe_gt *gt;
874 	int id;
875 
876 	/*
877 	 * SR-IOV VFs don't have access to the FUSE2 register, so we can't
878 	 * check pre-production status there.  But the host OS will notice
879 	 * and report the pre-production status, which should be enough to
880 	 * help us catch mistaken use of pre-production hardware.
881 	 */
882 	if (IS_SRIOV_VF(xe))
883 		return;
884 
885 	/*
886 	 * The "SW_CAP" fuse contains a bit indicating whether the device is a
887 	 * production or pre-production device.  This fuse is reflected through
888 	 * the GT "FUSE2" register, even though the contents of the fuse are
889 	 * not GT-specific.  Every GT's reflection of this fuse should show the
890 	 * same value, so we'll just use the first available GT for lookup.
891 	 */
892 	for_each_gt(gt, xe, id)
893 		break;
894 
895 	if (!gt)
896 		return;
897 
898 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
899 	if (!xe_force_wake_ref_has_domain(fw_ref.domains, XE_FW_GT)) {
900 		xe_gt_err(gt, "Forcewake failure; cannot determine production/pre-production hw status.\n");
901 		return;
902 	}
903 
904 	if (xe_mmio_read32(&gt->mmio, FUSE2) & PRODUCTION_HW)
905 		return;
906 
907 	xe_info(xe, "Pre-production hardware detected.\n");
908 	if (!xe->info.has_pre_prod_wa) {
909 		xe_err(xe, "Pre-production workarounds for this platform have already been removed.\n");
910 		add_taint(TAINT_MACHINE_CHECK, LOCKDEP_STILL_OK);
911 	}
912 }
913 
914 static void xe_device_wedged_fini(struct drm_device *drm, void *arg)
915 {
916 	struct xe_device *xe = arg;
917 
918 	if (atomic_read(&xe->wedged.flag))
919 		xe_pm_runtime_put(xe);
920 }
921 
922 int xe_device_probe(struct xe_device *xe)
923 {
924 	struct xe_tile *tile;
925 	struct xe_gt *gt;
926 	int err;
927 	u8 id;
928 
929 	xe_pat_init_early(xe);
930 
931 	err = xe_sriov_init(xe);
932 	if (err)
933 		return err;
934 
935 	xe->info.mem_region_mask = 1;
936 
937 	err = xe_set_dma_info(xe);
938 	if (err)
939 		return err;
940 
941 	err = xe_mmio_probe_tiles(xe);
942 	if (err)
943 		return err;
944 
945 	for_each_gt(gt, xe, id) {
946 		err = xe_gt_init_early(gt);
947 		if (err)
948 			return err;
949 	}
950 
951 	for_each_tile(tile, xe, id) {
952 		err = xe_ggtt_init_early(tile->mem.ggtt);
953 		if (err)
954 			return err;
955 	}
956 
957 	/*
958 	 * From here on, if a step fails, make sure a Driver-FLR is triggereed
959 	 */
960 	err = devm_add_action_or_reset(xe->drm.dev, xe_driver_flr_fini, xe);
961 	if (err)
962 		return err;
963 
964 	err = probe_has_flat_ccs(xe);
965 	if (err)
966 		return err;
967 
968 	err = xe_vram_probe(xe);
969 	if (err)
970 		return err;
971 
972 	for_each_tile(tile, xe, id) {
973 		err = xe_tile_init_noalloc(tile);
974 		if (err)
975 			return err;
976 	}
977 
978 	/*
979 	 * Allow allocations only now to ensure xe_display_init_early()
980 	 * is the first to allocate, always.
981 	 */
982 	err = xe_ttm_sys_mgr_init(xe);
983 	if (err)
984 		return err;
985 
986 	/* Allocate and map stolen after potential VRAM resize */
987 	err = xe_ttm_stolen_mgr_init(xe);
988 	if (err)
989 		return err;
990 
991 	/*
992 	 * Now that GT is initialized (TTM in particular),
993 	 * we can try to init display, and inherit the initial fb.
994 	 * This is the reason the first allocation needs to be done
995 	 * inside display.
996 	 */
997 	err = xe_display_init_early(xe);
998 	if (err)
999 		return err;
1000 
1001 	for_each_tile(tile, xe, id) {
1002 		err = xe_tile_init(tile);
1003 		if (err)
1004 			return err;
1005 	}
1006 
1007 	err = xe_irq_install(xe);
1008 	if (err)
1009 		return err;
1010 
1011 	for_each_gt(gt, xe, id) {
1012 		err = xe_gt_init(gt);
1013 		if (err)
1014 			return err;
1015 	}
1016 
1017 	err = xe_pagefault_init(xe);
1018 	if (err)
1019 		return err;
1020 
1021 	if (xe->tiles->media_gt &&
1022 	    XE_GT_WA(xe->tiles->media_gt, 15015404425_disable))
1023 		XE_DEVICE_WA_DISABLE(xe, 15015404425);
1024 
1025 	err = xe_devcoredump_init(xe);
1026 	if (err)
1027 		return err;
1028 
1029 	xe_nvm_init(xe);
1030 
1031 	err = xe_soc_remapper_init(xe);
1032 	if (err)
1033 		return err;
1034 
1035 	err = xe_heci_gsc_init(xe);
1036 	if (err)
1037 		return err;
1038 
1039 	err = xe_late_bind_init(&xe->late_bind);
1040 	if (err)
1041 		return err;
1042 
1043 	err = xe_oa_init(xe);
1044 	if (err)
1045 		return err;
1046 
1047 	err = xe_display_init(xe);
1048 	if (err)
1049 		return err;
1050 
1051 	err = xe_pxp_init(xe);
1052 	if (err)
1053 		return err;
1054 
1055 	err = xe_psmi_init(xe);
1056 	if (err)
1057 		return err;
1058 
1059 	err = drm_dev_register(&xe->drm, 0);
1060 	if (err)
1061 		return err;
1062 
1063 	xe_display_register(xe);
1064 
1065 	err = xe_oa_register(xe);
1066 	if (err)
1067 		goto err_unregister_display;
1068 
1069 	err = xe_pmu_register(&xe->pmu);
1070 	if (err)
1071 		goto err_unregister_display;
1072 
1073 	err = xe_sysctrl_init(xe);
1074 	if (err)
1075 		goto err_unregister_display;
1076 
1077 	err = xe_device_sysfs_init(xe);
1078 	if (err)
1079 		goto err_unregister_display;
1080 
1081 	xe_debugfs_register(xe);
1082 
1083 	err = xe_hwmon_register(xe);
1084 	if (err)
1085 		goto err_unregister_display;
1086 
1087 	err = xe_i2c_probe(xe);
1088 	if (err)
1089 		goto err_unregister_display;
1090 
1091 	for_each_gt(gt, xe, id)
1092 		xe_gt_sanitize_freq(gt);
1093 
1094 	xe_vsec_init(xe);
1095 
1096 	err = xe_sriov_init_late(xe);
1097 	if (err)
1098 		goto err_unregister_display;
1099 
1100 	detect_preproduction_hw(xe);
1101 
1102 	err = drmm_add_action_or_reset(&xe->drm, xe_device_wedged_fini, xe);
1103 	if (err)
1104 		goto err_unregister_display;
1105 
1106 	return devm_add_action_or_reset(xe->drm.dev, xe_device_sanitize, xe);
1107 
1108 err_unregister_display:
1109 	xe_display_unregister(xe);
1110 	drm_dev_unregister(&xe->drm);
1111 
1112 	return err;
1113 }
1114 
1115 void xe_device_remove(struct xe_device *xe)
1116 {
1117 	xe_display_unregister(xe);
1118 
1119 	drm_dev_unplug(&xe->drm);
1120 
1121 	xe_bo_pci_dev_remove_all(xe);
1122 }
1123 
1124 void xe_device_shutdown(struct xe_device *xe)
1125 {
1126 	struct xe_gt *gt;
1127 	u8 id;
1128 
1129 	drm_dbg(&xe->drm, "Shutting down device\n");
1130 
1131 	xe_display_pm_shutdown(xe);
1132 
1133 	xe_irq_suspend(xe);
1134 
1135 	for_each_gt(gt, xe, id)
1136 		xe_gt_shutdown(gt);
1137 
1138 	xe_display_pm_shutdown_late(xe);
1139 
1140 	if (!xe_driver_flr_disabled(xe)) {
1141 		/* BOOM! */
1142 		__xe_driver_flr(xe);
1143 	}
1144 }
1145 
1146 /**
1147  * xe_device_wmb() - Device specific write memory barrier
1148  * @xe: the &xe_device
1149  *
1150  * While wmb() is sufficient for a barrier if we use system memory, on discrete
1151  * platforms with device memory we additionally need to issue a register write.
1152  * Since it doesn't matter which register we write to, use the read-only VF_CAP
1153  * register that is also marked as accessible by the VFs.
1154  */
1155 void xe_device_wmb(struct xe_device *xe)
1156 {
1157 	wmb();
1158 	if (IS_DGFX(xe))
1159 		xe_mmio_write32(xe_root_tile_mmio(xe), VF_CAP_REG, 0);
1160 }
1161 
1162 /*
1163  * Issue a TRANSIENT_FLUSH_REQUEST and wait for completion on each gt.
1164  */
1165 static void tdf_request_sync(struct xe_device *xe)
1166 {
1167 	struct xe_gt *gt;
1168 	u8 id;
1169 
1170 	for_each_gt_with_type(gt, xe, id, BIT(XE_GT_TYPE_MAIN)) {
1171 		CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
1172 		if (!fw_ref.domains)
1173 			return;
1174 
1175 		xe_mmio_write32(&gt->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST);
1176 
1177 		/*
1178 		 * FIXME: We can likely do better here with our choice of
1179 		 * timeout. Currently we just assume the worst case, i.e. 150us,
1180 		 * which is believed to be sufficient to cover the worst case
1181 		 * scenario on current platforms if all cache entries are
1182 		 * transient and need to be flushed..
1183 		 */
1184 		if (xe_mmio_wait32(&gt->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST, 0,
1185 				   300, NULL, false))
1186 			xe_gt_err_once(gt, "TD flush timeout\n");
1187 	}
1188 }
1189 
1190 /**
1191  * xe_device_is_l2_flush_optimized - if L2 flush is optimized by HW
1192  * @xe: The device to check.
1193  *
1194  * Return: true if the HW device optimizing L2 flush, false otherwise.
1195  */
1196 bool xe_device_is_l2_flush_optimized(struct xe_device *xe)
1197 {
1198 	/* XA is *always* flushed, like at the end-of-submssion (and maybe other
1199 	 * places), just that internally as an optimisation hw doesn't need to make
1200 	 * that a full flush (which will also include XA) when Media is
1201 	 * off/powergated, since it doesn't need to worry about GT caches vs Media
1202 	 * coherency, and only CPU vs GPU coherency, so can make that flush a
1203 	 * targeted XA flush, since stuff tagged with XA now means it's shared with
1204 	 * the CPU. The main implication is that we now need to somehow flush non-XA before
1205 	 * freeing system memory pages, otherwise dirty cachelines could be flushed after the free
1206 	 * (like if Media suddenly turns on and does a full flush)
1207 	 */
1208 	if (GRAPHICS_VER(xe) >= 35 && !IS_DGFX(xe))
1209 		return true;
1210 	return false;
1211 }
1212 
1213 void xe_device_l2_flush(struct xe_device *xe)
1214 {
1215 	struct xe_gt *gt;
1216 
1217 	gt = xe_root_mmio_gt(xe);
1218 	if (!gt)
1219 		return;
1220 
1221 	if (!XE_GT_WA(gt, 16023588340))
1222 		return;
1223 
1224 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
1225 	if (!fw_ref.domains)
1226 		return;
1227 
1228 	spin_lock(&gt->global_invl_lock);
1229 
1230 	xe_mmio_write32(&gt->mmio, XE2_GLOBAL_INVAL, 0x1);
1231 	if (xe_mmio_wait32(&gt->mmio, XE2_GLOBAL_INVAL, 0x1, 0x0, 1000, NULL, true))
1232 		xe_gt_err_once(gt, "Global invalidation timeout\n");
1233 
1234 	spin_unlock(&gt->global_invl_lock);
1235 }
1236 
1237 /**
1238  * xe_device_td_flush() - Flush transient L3 cache entries
1239  * @xe: The device
1240  *
1241  * Display engine has direct access to memory and is never coherent with L3/L4
1242  * caches (or CPU caches), however KMD is responsible for specifically flushing
1243  * transient L3 GPU cache entries prior to the flip sequence to ensure scanout
1244  * can happen from such a surface without seeing corruption.
1245  *
1246  * Display surfaces can be tagged as transient by mapping it using one of the
1247  * various L3:XD PAT index modes on Xe2.
1248  *
1249  * Note: On non-discrete xe2 platforms, like LNL, the entire L3 cache is flushed
1250  * at the end of each submission via PIPE_CONTROL for compute/render, since SA
1251  * Media is not coherent with L3 and we want to support render-vs-media
1252  * usescases. For other engines like copy/blt the HW internally forces uncached
1253  * behaviour, hence why we can skip the TDF on such platforms.
1254  */
1255 void xe_device_td_flush(struct xe_device *xe)
1256 {
1257 	struct xe_gt *root_gt;
1258 
1259 	/*
1260 	 * From Xe3p onward the HW takes care of flush of TD entries also along
1261 	 * with flushing XA entries, which will be at the usual sync points,
1262 	 * like at the end of submission, so no manual flush is needed here.
1263 	 */
1264 	if (GRAPHICS_VER(xe) >= 35)
1265 		return;
1266 
1267 	if (!IS_DGFX(xe) || GRAPHICS_VER(xe) < 20)
1268 		return;
1269 
1270 	root_gt = xe_root_mmio_gt(xe);
1271 	if (!root_gt)
1272 		return;
1273 
1274 	if (XE_GT_WA(root_gt, 16023588340)) {
1275 		/* A transient flush is not sufficient: flush the L2 */
1276 		xe_device_l2_flush(xe);
1277 	} else {
1278 		xe_guc_pc_apply_flush_freq_limit(&root_gt->uc.guc.pc);
1279 		tdf_request_sync(xe);
1280 		xe_guc_pc_remove_flush_freq_limit(&root_gt->uc.guc.pc);
1281 	}
1282 }
1283 
1284 u32 xe_device_ccs_bytes(struct xe_device *xe, u64 size)
1285 {
1286 	return xe_device_has_flat_ccs(xe) ?
1287 		DIV_ROUND_UP_ULL(size, NUM_BYTES_PER_CCS_BYTE(xe)) : 0;
1288 }
1289 
1290 /**
1291  * xe_device_assert_mem_access - Inspect the current runtime_pm state.
1292  * @xe: xe device instance
1293  *
1294  * To be used before any kind of memory access. It will splat a debug warning
1295  * if the device is currently sleeping. But it doesn't guarantee in any way
1296  * that the device is going to remain awake. Xe PM runtime get and put
1297  * functions might be added to the outer bound of the memory access, while
1298  * this check is intended for inner usage to splat some warning if the worst
1299  * case has just happened.
1300  */
1301 void xe_device_assert_mem_access(struct xe_device *xe)
1302 {
1303 	xe_assert(xe, !xe_pm_runtime_suspended(xe));
1304 }
1305 
1306 void xe_device_snapshot_print(struct xe_device *xe, struct drm_printer *p)
1307 {
1308 	struct xe_gt *gt;
1309 	u8 id;
1310 
1311 	drm_printf(p, "PCI ID: 0x%04x\n", xe->info.devid);
1312 	drm_printf(p, "PCI revision: 0x%02x\n", xe->info.revid);
1313 
1314 	for_each_gt(gt, xe, id) {
1315 		drm_printf(p, "GT id: %u\n", id);
1316 		drm_printf(p, "\tTile: %u\n", gt->tile->id);
1317 		drm_printf(p, "\tType: %s\n",
1318 			   gt->info.type == XE_GT_TYPE_MAIN ? "main" : "media");
1319 		drm_printf(p, "\tIP ver: %u.%u.%u\n",
1320 			   REG_FIELD_GET(GMD_ID_ARCH_MASK, gt->info.gmdid),
1321 			   REG_FIELD_GET(GMD_ID_RELEASE_MASK, gt->info.gmdid),
1322 			   REG_FIELD_GET(GMD_ID_REVID, gt->info.gmdid));
1323 		drm_printf(p, "\tCS reference clock: %u\n", gt->info.reference_clock);
1324 	}
1325 }
1326 
1327 u64 xe_device_canonicalize_addr(struct xe_device *xe, u64 address)
1328 {
1329 	return sign_extend64(address, xe->info.va_bits - 1);
1330 }
1331 
1332 u64 xe_device_uncanonicalize_addr(struct xe_device *xe, u64 address)
1333 {
1334 	return address & GENMASK_ULL(xe->info.va_bits - 1, 0);
1335 }
1336 
1337 /**
1338  * DOC: Xe Device Wedging
1339  *
1340  * Xe driver uses drm device wedged uevent as documented in Documentation/gpu/drm-uapi.rst.
1341  * When device is in wedged state, every IOCTL will be blocked and GT cannot
1342  * be used. The conditions under which the driver declares the device wedged
1343  * depend on the wedged mode configuration (see &enum xe_wedged_mode). The
1344  * default recovery method for a wedged state is rebind/bus-reset.
1345  *
1346  * Another recovery method is vendor-specific. Below are the cases that send
1347  * ``WEDGED=vendor-specific`` recovery method in drm device wedged uevent.
1348  *
1349  * Case: Firmware Flash
1350  * --------------------
1351  *
1352  * Identification Hint
1353  * +++++++++++++++++++
1354  *
1355  * ``WEDGED=vendor-specific`` drm device wedged uevent with
1356  * :ref:`Runtime Survivability mode <xe-survivability-mode>` is used to notify
1357  * admin/userspace consumer about the need for a firmware flash.
1358  *
1359  * Recovery Procedure
1360  * ++++++++++++++++++
1361  *
1362  * Once ``WEDGED=vendor-specific`` drm device wedged uevent is received, follow
1363  * the below steps
1364  *
1365  * - Check Runtime Survivability mode sysfs.
1366  *   If enabled, firmware flash is required to recover the device.
1367  *
1368  *   /sys/bus/pci/devices/<device>/survivability_mode
1369  *
1370  * - Admin/userspace consumer can use firmware flashing tools like fwupd to flash
1371  *   firmware and restore device to normal operation.
1372  */
1373 
1374 /**
1375  * xe_device_set_wedged_method - Set wedged recovery method
1376  * @xe: xe device instance
1377  * @method: recovery method to set
1378  *
1379  * Set wedged recovery method to be sent in drm wedged uevent.
1380  */
1381 void xe_device_set_wedged_method(struct xe_device *xe, unsigned long method)
1382 {
1383 	xe->wedged.method = method;
1384 }
1385 
1386 /**
1387  * xe_device_declare_wedged - Declare device wedged
1388  * @xe: xe device instance
1389  *
1390  * This is a final state that can only be cleared with the recovery method
1391  * specified in the drm wedged uevent. The method can be set using
1392  * xe_device_set_wedged_method before declaring the device as wedged. If no method
1393  * is set, reprobe (unbind/re-bind) will be sent by default.
1394  *
1395  * In this state every IOCTL will be blocked so the GT cannot be used.
1396  * In general it will be called upon any critical error such as gt reset
1397  * failure or guc loading failure. Userspace will be notified of this state
1398  * through device wedged uevent.
1399  * If xe.wedged module parameter is set to 2, this function will be called
1400  * on every single execution timeout (a.k.a. GPU hang) right after devcoredump
1401  * snapshot capture. In this mode, GT reset won't be attempted so the state of
1402  * the issue is preserved for further debugging.
1403  */
1404 void xe_device_declare_wedged(struct xe_device *xe)
1405 {
1406 	struct xe_gt *gt;
1407 	u8 id;
1408 
1409 	if (xe->wedged.mode == XE_WEDGED_MODE_NEVER) {
1410 		drm_dbg(&xe->drm, "Wedged mode is forcibly disabled\n");
1411 		return;
1412 	}
1413 
1414 	if (!atomic_xchg(&xe->wedged.flag, 1)) {
1415 		xe->needs_flr_on_fini = true;
1416 		xe_pm_runtime_get_noresume(xe);
1417 		drm_err(&xe->drm,
1418 			"CRITICAL: Xe has declared device %s as wedged.\n"
1419 			"IOCTLs and executions are blocked.\n"
1420 			"For recovery procedure, refer to https://docs.kernel.org/gpu/drm-uapi.html#device-wedging\n"
1421 			"Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/xe/kernel/issues/new\n",
1422 			dev_name(xe->drm.dev));
1423 	}
1424 
1425 	for_each_gt(gt, xe, id)
1426 		xe_gt_declare_wedged(gt);
1427 
1428 	if (xe_device_wedged(xe)) {
1429 		/*
1430 		 * XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET is intended for debugging
1431 		 * hangs, so wedge the device with 'none' recovery method and have
1432 		 * it available to the user for debugging.
1433 		 */
1434 		if (xe->wedged.mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET)
1435 			xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_NONE);
1436 		/* If no wedge recovery method is set, use default */
1437 		else if (!xe->wedged.method)
1438 			xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_REBIND |
1439 						    DRM_WEDGE_RECOVERY_BUS_RESET);
1440 
1441 		/* Notify userspace of wedged device */
1442 		drm_dev_wedged_event(&xe->drm, xe->wedged.method, NULL);
1443 	}
1444 }
1445 
1446 /**
1447  * xe_device_validate_wedged_mode - Check if given mode is supported
1448  * @xe: the &xe_device
1449  * @mode: requested mode to validate
1450  *
1451  * Check whether the provided wedged mode is supported.
1452  *
1453  * Return: 0 if mode is supported, error code otherwise.
1454  */
1455 int xe_device_validate_wedged_mode(struct xe_device *xe, unsigned int mode)
1456 {
1457 	if (mode > XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET) {
1458 		drm_dbg(&xe->drm, "wedged_mode: invalid value (%u)\n", mode);
1459 		return -EINVAL;
1460 	} else if (mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET && (IS_SRIOV_VF(xe) ||
1461 		   (IS_SRIOV_PF(xe) && !IS_ENABLED(CONFIG_DRM_XE_DEBUG)))) {
1462 		drm_dbg(&xe->drm, "wedged_mode: (%u) %s mode is not supported for %s\n",
1463 			mode, xe_wedged_mode_to_string(mode),
1464 			xe_sriov_mode_to_string(xe_device_sriov_mode(xe)));
1465 		return -EPERM;
1466 	}
1467 
1468 	return 0;
1469 }
1470 
1471 /**
1472  * xe_wedged_mode_to_string - Convert enum value to string.
1473  * @mode: the &xe_wedged_mode to convert
1474  *
1475  * Returns: wedged mode as a user friendly string.
1476  */
1477 const char *xe_wedged_mode_to_string(enum xe_wedged_mode mode)
1478 {
1479 	switch (mode) {
1480 	case XE_WEDGED_MODE_NEVER:
1481 		return "never";
1482 	case XE_WEDGED_MODE_UPON_CRITICAL_ERROR:
1483 		return "upon-critical-error";
1484 	case XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET:
1485 		return "upon-any-hang-no-reset";
1486 	default:
1487 		return "<invalid>";
1488 	}
1489 }
1490 
1491 /**
1492  * xe_device_asid_to_vm() - Find VM from ASID
1493  * @xe: the &xe_device
1494  * @asid: Address space ID
1495  *
1496  * Find a VM from ASID and take a reference to VM which caller must drop.
1497  * Reclaim safe.
1498  *
1499  * Return: VM on success, ERR_PTR on failure
1500  */
1501 struct xe_vm *xe_device_asid_to_vm(struct xe_device *xe, u32 asid)
1502 {
1503 	struct xe_vm *vm;
1504 
1505 	down_read(&xe->usm.lock);
1506 	vm = xa_load(&xe->usm.asid_to_vm, asid);
1507 	if (vm)
1508 		xe_vm_get(vm);
1509 	else
1510 		vm = ERR_PTR(-EINVAL);
1511 	up_read(&xe->usm.lock);
1512 
1513 	return vm;
1514 }
1515