xref: /linux/drivers/gpu/drm/xe/xe_device.c (revision 99676aed1fec109d62822e21a06760eb098dc5f4)
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 struct xe_device *xe_device_create(struct pci_dev *pdev,
479 				   const struct pci_device_id *ent)
480 {
481 	const struct drm_driver *driver = &regular_driver;
482 	struct xe_device *xe;
483 	int err;
484 
485 #ifdef CONFIG_PCI_IOV
486 	/*
487 	 * Since XE device is not initialized yet, read from configfs
488 	 * directly to decide whether we are in admin-only PF mode or not.
489 	 */
490 	if (xe_configfs_admin_only_pf(pdev))
491 		driver = &admin_only_driver;
492 #endif
493 
494 	err = aperture_remove_conflicting_pci_devices(pdev, driver->name);
495 	if (err)
496 		return ERR_PTR(err);
497 
498 	xe = devm_drm_dev_alloc(&pdev->dev, driver, struct xe_device, drm);
499 	if (IS_ERR(xe))
500 		return xe;
501 
502 	err = ttm_device_init(&xe->ttm, &xe_ttm_funcs, xe->drm.dev,
503 			      xe->drm.anon_inode->i_mapping,
504 			      xe->drm.vma_offset_manager, 0);
505 	if (WARN_ON(err))
506 		return ERR_PTR(err);
507 
508 	xe_bo_dev_init(&xe->bo_device);
509 	err = drmm_add_action_or_reset(&xe->drm, xe_device_destroy, NULL);
510 	if (err)
511 		return ERR_PTR(err);
512 
513 	err = xe_shrinker_create(xe);
514 	if (err)
515 		return ERR_PTR(err);
516 
517 	xe->info.devid = pdev->device;
518 	xe->info.revid = pdev->revision;
519 	xe->info.force_execlist = xe_modparam.force_execlist;
520 	xe->atomic_svm_timeslice_ms = 5;
521 	xe->min_run_period_lr_ms = 5;
522 
523 	err = xe_irq_init(xe);
524 	if (err)
525 		return ERR_PTR(err);
526 
527 	xe_validation_device_init(&xe->val);
528 
529 	init_waitqueue_head(&xe->ufence_wq);
530 
531 	init_rwsem(&xe->usm.lock);
532 
533 	err = xe_pagemap_shrinker_create(xe);
534 	if (err)
535 		return ERR_PTR(err);
536 
537 	xa_init_flags(&xe->usm.asid_to_vm, XA_FLAGS_ALLOC);
538 
539 	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) {
540 		/* Trigger a large asid and an early asid wrap. */
541 		u32 asid;
542 
543 		BUILD_BUG_ON(XE_MAX_ASID < 2);
544 		err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, NULL,
545 				      XA_LIMIT(XE_MAX_ASID - 2, XE_MAX_ASID - 1),
546 				      &xe->usm.next_asid, GFP_KERNEL);
547 		drm_WARN_ON(&xe->drm, err);
548 		if (err >= 0)
549 			xa_erase(&xe->usm.asid_to_vm, asid);
550 	}
551 
552 	err = xe_bo_pinned_init(xe);
553 	if (err)
554 		return ERR_PTR(err);
555 
556 	xe->preempt_fence_wq = alloc_ordered_workqueue("xe-preempt-fence-wq",
557 						       WQ_MEM_RECLAIM);
558 	xe->ordered_wq = alloc_ordered_workqueue("xe-ordered-wq", 0);
559 	xe->unordered_wq = alloc_workqueue("xe-unordered-wq", WQ_PERCPU, 0);
560 	xe->destroy_wq = alloc_workqueue("xe-destroy-wq", WQ_PERCPU, 0);
561 	if (!xe->ordered_wq || !xe->unordered_wq ||
562 	    !xe->preempt_fence_wq || !xe->destroy_wq) {
563 		/*
564 		 * Cleanup done in xe_device_destroy via
565 		 * drmm_add_action_or_reset register above
566 		 */
567 		drm_err(&xe->drm, "Failed to allocate xe workqueues\n");
568 		return ERR_PTR(-ENOMEM);
569 	}
570 
571 	err = drmm_mutex_init(&xe->drm, &xe->pmt.lock);
572 	if (err)
573 		return ERR_PTR(err);
574 
575 	return xe;
576 }
577 ALLOW_ERROR_INJECTION(xe_device_create, ERRNO); /* See xe_pci_probe() */
578 
579 static bool xe_driver_flr_disabled(struct xe_device *xe)
580 {
581 	if (IS_SRIOV_VF(xe))
582 		return true;
583 
584 	if (xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL_PROTECTED) & DRIVERINT_FLR_DIS) {
585 		drm_info(&xe->drm, "Driver-FLR disabled by BIOS\n");
586 		return true;
587 	}
588 
589 	return false;
590 }
591 
592 /*
593  * The driver-initiated FLR is the highest level of reset that we can trigger
594  * from within the driver. It is different from the PCI FLR in that it doesn't
595  * fully reset the SGUnit and doesn't modify the PCI config space and therefore
596  * it doesn't require a re-enumeration of the PCI BARs. However, the
597  * driver-initiated FLR does still cause a reset of both GT and display and a
598  * memory wipe of local and stolen memory, so recovery would require a full HW
599  * re-init and saving/restoring (or re-populating) the wiped memory. Since we
600  * perform the FLR as the very last action before releasing access to the HW
601  * during the driver release flow, we don't attempt recovery at all, because
602  * if/when a new instance of Xe is bound to the device it will do a full
603  * re-init anyway.
604  */
605 static void __xe_driver_flr(struct xe_device *xe)
606 {
607 	const unsigned int flr_timeout = 3 * USEC_PER_SEC; /* specs recommend a 3s wait */
608 	struct xe_mmio *mmio = xe_root_tile_mmio(xe);
609 	int ret;
610 
611 	drm_dbg(&xe->drm, "Triggering Driver-FLR\n");
612 
613 	/*
614 	 * Make sure any pending FLR requests have cleared by waiting for the
615 	 * FLR trigger bit to go to zero. Also clear GU_DEBUG's DRIVERFLR_STATUS
616 	 * to make sure it's not still set from a prior attempt (it's a write to
617 	 * clear bit).
618 	 * Note that we should never be in a situation where a previous attempt
619 	 * is still pending (unless the HW is totally dead), but better to be
620 	 * safe in case something unexpected happens
621 	 */
622 	ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false);
623 	if (ret) {
624 		drm_err(&xe->drm, "Driver-FLR-prepare wait for ready failed! %d\n", ret);
625 		return;
626 	}
627 	xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS);
628 
629 	/* Trigger the actual Driver-FLR */
630 	xe_mmio_rmw32(mmio, GU_CNTL, 0, DRIVERFLR);
631 
632 	/* Wait for hardware teardown to complete */
633 	ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false);
634 	if (ret) {
635 		drm_err(&xe->drm, "Driver-FLR-teardown wait completion failed! %d\n", ret);
636 		return;
637 	}
638 
639 	/* Wait for hardware/firmware re-init to complete */
640 	ret = xe_mmio_wait32(mmio, GU_DEBUG, DRIVERFLR_STATUS, DRIVERFLR_STATUS,
641 			     flr_timeout, NULL, false);
642 	if (ret) {
643 		drm_err(&xe->drm, "Driver-FLR-reinit wait completion failed! %d\n", ret);
644 		return;
645 	}
646 
647 	/* Clear sticky completion status */
648 	xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS);
649 }
650 
651 static void xe_driver_flr(struct xe_device *xe)
652 {
653 	if (xe_driver_flr_disabled(xe))
654 		return;
655 
656 	__xe_driver_flr(xe);
657 }
658 
659 static void xe_driver_flr_fini(void *arg)
660 {
661 	struct xe_device *xe = arg;
662 
663 	if (xe->needs_flr_on_fini)
664 		xe_driver_flr(xe);
665 }
666 
667 static void xe_device_sanitize(void *arg)
668 {
669 	struct xe_device *xe = arg;
670 	struct xe_gt *gt;
671 	u8 id;
672 
673 	for_each_gt(gt, xe, id)
674 		xe_gt_sanitize(gt);
675 }
676 
677 static int xe_set_dma_info(struct xe_device *xe)
678 {
679 	unsigned int mask_size = xe->info.dma_mask_size;
680 	int err;
681 
682 	dma_set_max_seg_size(xe->drm.dev, xe_sg_segment_size(xe->drm.dev));
683 
684 	err = dma_set_mask(xe->drm.dev, DMA_BIT_MASK(mask_size));
685 	if (err)
686 		goto mask_err;
687 
688 	err = dma_set_coherent_mask(xe->drm.dev, DMA_BIT_MASK(mask_size));
689 	if (err)
690 		goto mask_err;
691 
692 	return 0;
693 
694 mask_err:
695 	drm_err(&xe->drm, "Can't set DMA mask/consistent mask (%d)\n", err);
696 	return err;
697 }
698 
699 static void assert_lmem_ready(struct xe_device *xe)
700 {
701 	if (!IS_DGFX(xe) || IS_SRIOV_VF(xe))
702 		return;
703 
704 	xe_assert(xe, xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL) &
705 		  LMEM_INIT);
706 }
707 
708 static void vf_update_device_info(struct xe_device *xe)
709 {
710 	xe_assert(xe, IS_SRIOV_VF(xe));
711 	/* disable features that are not available/applicable to VFs */
712 	xe->info.probe_display = 0;
713 	xe->info.has_heci_cscfi = 0;
714 	xe->info.has_heci_gscfi = 0;
715 	xe->info.has_late_bind = 0;
716 	xe->info.skip_guc_pc = 1;
717 	xe->info.skip_pcode = 1;
718 }
719 
720 static int xe_device_vram_alloc(struct xe_device *xe)
721 {
722 	struct xe_vram_region *vram;
723 
724 	if (!IS_DGFX(xe))
725 		return 0;
726 
727 	vram = drmm_kzalloc(&xe->drm, sizeof(*vram), GFP_KERNEL);
728 	if (!vram)
729 		return -ENOMEM;
730 
731 	xe->mem.vram = vram;
732 	return 0;
733 }
734 
735 /**
736  * xe_device_probe_early: Device early probe
737  * @xe: xe device instance
738  *
739  * Initialize MMIO resources that don't require any
740  * knowledge about tile count. Also initialize pcode and
741  * check vram initialization on root tile.
742  *
743  * Return: 0 on success, error code on failure
744  */
745 int xe_device_probe_early(struct xe_device *xe)
746 {
747 	int err;
748 
749 	xe_wa_device_init(xe);
750 	xe_wa_process_device_oob(xe);
751 
752 	err = xe_mmio_probe_early(xe);
753 	if (err)
754 		return err;
755 
756 	xe_sriov_probe_early(xe);
757 
758 	if (xe_device_is_admin_only(xe) && !IS_SRIOV_PF(xe)) {
759 		xe_err(xe, "Can't run Admin-only mode without SR-IOV PF mode!\n");
760 		return -ENODEV;
761 	}
762 
763 	if (IS_SRIOV_VF(xe))
764 		vf_update_device_info(xe);
765 
766 	/*
767 	 * Check for pcode uncore_init status to confirm if the SoC
768 	 * initialization is complete. Until done, any MMIO or lmem access from
769 	 * the driver will be blocked
770 	 */
771 	err = xe_pcode_probe_early(xe);
772 	if (err || xe_survivability_mode_is_requested(xe)) {
773 		int save_err = err;
774 
775 		/*
776 		 * Try to leave device in survivability mode if device is
777 		 * possible, but still return the previous error for error
778 		 * propagation
779 		 */
780 		err = xe_survivability_mode_boot_enable(xe);
781 		if (err)
782 			return err;
783 
784 		return save_err;
785 	}
786 
787 	/*
788 	 * Make sure the lmem is initialized and ready to use. xe_pcode_ready()
789 	 * is flagged after full initialization is complete. Assert if lmem is
790 	 * not initialized.
791 	 */
792 	assert_lmem_ready(xe);
793 
794 	xe->wedged.mode = xe_device_validate_wedged_mode(xe, xe_modparam.wedged_mode) ?
795 			  XE_DEFAULT_WEDGED_MODE : xe_modparam.wedged_mode;
796 	drm_dbg(&xe->drm, "wedged_mode: setting mode (%u) %s\n",
797 		xe->wedged.mode, xe_wedged_mode_to_string(xe->wedged.mode));
798 
799 	err = xe_device_vram_alloc(xe);
800 	if (err)
801 		return err;
802 
803 	return 0;
804 }
805 ALLOW_ERROR_INJECTION(xe_device_probe_early, ERRNO); /* See xe_pci_probe() */
806 
807 static int probe_has_flat_ccs(struct xe_device *xe)
808 {
809 	struct xe_gt *gt;
810 	u32 reg;
811 
812 	/* Always enabled/disabled, no runtime check to do */
813 	if (GRAPHICS_VER(xe) < 20 || !xe->info.has_flat_ccs || IS_SRIOV_VF(xe))
814 		return 0;
815 
816 	gt = xe_root_mmio_gt(xe);
817 	if (!gt)
818 		return 0;
819 
820 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
821 	if (!fw_ref.domains)
822 		return -ETIMEDOUT;
823 
824 	reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_LOWER);
825 	xe->info.has_flat_ccs = (reg & XE2_FLAT_CCS_ENABLE);
826 
827 	if (!xe->info.has_flat_ccs)
828 		drm_dbg(&xe->drm,
829 			"Flat CCS has been disabled in bios, May lead to performance impact");
830 
831 	return 0;
832 }
833 
834 /*
835  * Detect if the driver is being run on pre-production hardware.  We don't
836  * keep workarounds for pre-production hardware long term, so print an
837  * error and add taint if we're being loaded on a pre-production platform
838  * for which the pre-prod workarounds have already been removed.
839  *
840  * The general policy is that we'll remove any workarounds that only apply to
841  * pre-production hardware around the time force_probe restrictions are lifted
842  * for a platform of the next major IP generation (for example, Xe2 pre-prod
843  * workarounds should be removed around the time the first Xe3 platforms have
844  * force_probe lifted).
845  */
846 static void detect_preproduction_hw(struct xe_device *xe)
847 {
848 	struct xe_gt *gt;
849 	int id;
850 
851 	/*
852 	 * SR-IOV VFs don't have access to the FUSE2 register, so we can't
853 	 * check pre-production status there.  But the host OS will notice
854 	 * and report the pre-production status, which should be enough to
855 	 * help us catch mistaken use of pre-production hardware.
856 	 */
857 	if (IS_SRIOV_VF(xe))
858 		return;
859 
860 	/*
861 	 * The "SW_CAP" fuse contains a bit indicating whether the device is a
862 	 * production or pre-production device.  This fuse is reflected through
863 	 * the GT "FUSE2" register, even though the contents of the fuse are
864 	 * not GT-specific.  Every GT's reflection of this fuse should show the
865 	 * same value, so we'll just use the first available GT for lookup.
866 	 */
867 	for_each_gt(gt, xe, id)
868 		break;
869 
870 	if (!gt)
871 		return;
872 
873 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
874 	if (!xe_force_wake_ref_has_domain(fw_ref.domains, XE_FW_GT)) {
875 		xe_gt_err(gt, "Forcewake failure; cannot determine production/pre-production hw status.\n");
876 		return;
877 	}
878 
879 	if (xe_mmio_read32(&gt->mmio, FUSE2) & PRODUCTION_HW)
880 		return;
881 
882 	xe_info(xe, "Pre-production hardware detected.\n");
883 	if (!xe->info.has_pre_prod_wa) {
884 		xe_err(xe, "Pre-production workarounds for this platform have already been removed.\n");
885 		add_taint(TAINT_MACHINE_CHECK, LOCKDEP_STILL_OK);
886 	}
887 }
888 
889 static void xe_device_wedged_fini(struct drm_device *drm, void *arg)
890 {
891 	struct xe_device *xe = arg;
892 
893 	if (atomic_read(&xe->wedged.flag))
894 		xe_pm_runtime_put(xe);
895 }
896 
897 int xe_device_probe(struct xe_device *xe)
898 {
899 	struct xe_tile *tile;
900 	struct xe_gt *gt;
901 	int err;
902 	u8 id;
903 
904 	xe_pat_init_early(xe);
905 
906 	err = xe_sriov_init(xe);
907 	if (err)
908 		return err;
909 
910 	xe->info.mem_region_mask = 1;
911 
912 	err = xe_set_dma_info(xe);
913 	if (err)
914 		return err;
915 
916 	err = xe_mmio_probe_tiles(xe);
917 	if (err)
918 		return err;
919 
920 	for_each_gt(gt, xe, id) {
921 		err = xe_gt_init_early(gt);
922 		if (err)
923 			return err;
924 	}
925 
926 	for_each_tile(tile, xe, id) {
927 		err = xe_ggtt_init_early(tile->mem.ggtt);
928 		if (err)
929 			return err;
930 	}
931 
932 	/*
933 	 * From here on, if a step fails, make sure a Driver-FLR is triggereed
934 	 */
935 	err = devm_add_action_or_reset(xe->drm.dev, xe_driver_flr_fini, xe);
936 	if (err)
937 		return err;
938 
939 	err = probe_has_flat_ccs(xe);
940 	if (err)
941 		return err;
942 
943 	err = xe_vram_probe(xe);
944 	if (err)
945 		return err;
946 
947 	for_each_tile(tile, xe, id) {
948 		err = xe_tile_init_noalloc(tile);
949 		if (err)
950 			return err;
951 	}
952 
953 	/*
954 	 * Allow allocations only now to ensure xe_display_init_early()
955 	 * is the first to allocate, always.
956 	 */
957 	err = xe_ttm_sys_mgr_init(xe);
958 	if (err)
959 		return err;
960 
961 	/* Allocate and map stolen after potential VRAM resize */
962 	err = xe_ttm_stolen_mgr_init(xe);
963 	if (err)
964 		return err;
965 
966 	/*
967 	 * Now that GT is initialized (TTM in particular),
968 	 * we can try to init display, and inherit the initial fb.
969 	 * This is the reason the first allocation needs to be done
970 	 * inside display.
971 	 */
972 	err = xe_display_init_early(xe);
973 	if (err)
974 		return err;
975 
976 	for_each_tile(tile, xe, id) {
977 		err = xe_tile_init(tile);
978 		if (err)
979 			return err;
980 	}
981 
982 	err = xe_irq_install(xe);
983 	if (err)
984 		return err;
985 
986 	for_each_gt(gt, xe, id) {
987 		err = xe_gt_init(gt);
988 		if (err)
989 			return err;
990 	}
991 
992 	err = xe_pagefault_init(xe);
993 	if (err)
994 		return err;
995 
996 	if (xe->tiles->media_gt &&
997 	    XE_GT_WA(xe->tiles->media_gt, 15015404425_disable))
998 		XE_DEVICE_WA_DISABLE(xe, 15015404425);
999 
1000 	err = xe_devcoredump_init(xe);
1001 	if (err)
1002 		return err;
1003 
1004 	xe_nvm_init(xe);
1005 
1006 	err = xe_soc_remapper_init(xe);
1007 	if (err)
1008 		return err;
1009 
1010 	err = xe_heci_gsc_init(xe);
1011 	if (err)
1012 		return err;
1013 
1014 	err = xe_late_bind_init(&xe->late_bind);
1015 	if (err)
1016 		return err;
1017 
1018 	err = xe_oa_init(xe);
1019 	if (err)
1020 		return err;
1021 
1022 	err = xe_display_init(xe);
1023 	if (err)
1024 		return err;
1025 
1026 	err = xe_pxp_init(xe);
1027 	if (err)
1028 		return err;
1029 
1030 	err = xe_psmi_init(xe);
1031 	if (err)
1032 		return err;
1033 
1034 	err = drm_dev_register(&xe->drm, 0);
1035 	if (err)
1036 		return err;
1037 
1038 	xe_display_register(xe);
1039 
1040 	err = xe_oa_register(xe);
1041 	if (err)
1042 		goto err_unregister_display;
1043 
1044 	err = xe_pmu_register(&xe->pmu);
1045 	if (err)
1046 		goto err_unregister_display;
1047 
1048 	err = xe_sysctrl_init(xe);
1049 	if (err)
1050 		goto err_unregister_display;
1051 
1052 	err = xe_device_sysfs_init(xe);
1053 	if (err)
1054 		goto err_unregister_display;
1055 
1056 	xe_debugfs_register(xe);
1057 
1058 	err = xe_hwmon_register(xe);
1059 	if (err)
1060 		goto err_unregister_display;
1061 
1062 	err = xe_i2c_probe(xe);
1063 	if (err)
1064 		goto err_unregister_display;
1065 
1066 	for_each_gt(gt, xe, id)
1067 		xe_gt_sanitize_freq(gt);
1068 
1069 	xe_vsec_init(xe);
1070 
1071 	err = xe_sriov_init_late(xe);
1072 	if (err)
1073 		goto err_unregister_display;
1074 
1075 	detect_preproduction_hw(xe);
1076 
1077 	err = drmm_add_action_or_reset(&xe->drm, xe_device_wedged_fini, xe);
1078 	if (err)
1079 		goto err_unregister_display;
1080 
1081 	return devm_add_action_or_reset(xe->drm.dev, xe_device_sanitize, xe);
1082 
1083 err_unregister_display:
1084 	xe_display_unregister(xe);
1085 	drm_dev_unregister(&xe->drm);
1086 
1087 	return err;
1088 }
1089 
1090 void xe_device_remove(struct xe_device *xe)
1091 {
1092 	xe_display_unregister(xe);
1093 
1094 	drm_dev_unplug(&xe->drm);
1095 
1096 	xe_bo_pci_dev_remove_all(xe);
1097 }
1098 
1099 void xe_device_shutdown(struct xe_device *xe)
1100 {
1101 	struct xe_gt *gt;
1102 	u8 id;
1103 
1104 	drm_dbg(&xe->drm, "Shutting down device\n");
1105 
1106 	xe_display_pm_shutdown(xe);
1107 
1108 	xe_irq_suspend(xe);
1109 
1110 	for_each_gt(gt, xe, id)
1111 		xe_gt_shutdown(gt);
1112 
1113 	xe_display_pm_shutdown_late(xe);
1114 
1115 	if (!xe_driver_flr_disabled(xe)) {
1116 		/* BOOM! */
1117 		__xe_driver_flr(xe);
1118 	}
1119 }
1120 
1121 /**
1122  * xe_device_wmb() - Device specific write memory barrier
1123  * @xe: the &xe_device
1124  *
1125  * While wmb() is sufficient for a barrier if we use system memory, on discrete
1126  * platforms with device memory we additionally need to issue a register write.
1127  * Since it doesn't matter which register we write to, use the read-only VF_CAP
1128  * register that is also marked as accessible by the VFs.
1129  */
1130 void xe_device_wmb(struct xe_device *xe)
1131 {
1132 	wmb();
1133 	if (IS_DGFX(xe))
1134 		xe_mmio_write32(xe_root_tile_mmio(xe), VF_CAP_REG, 0);
1135 }
1136 
1137 /*
1138  * Issue a TRANSIENT_FLUSH_REQUEST and wait for completion on each gt.
1139  */
1140 static void tdf_request_sync(struct xe_device *xe)
1141 {
1142 	struct xe_gt *gt;
1143 	u8 id;
1144 
1145 	for_each_gt_with_type(gt, xe, id, BIT(XE_GT_TYPE_MAIN)) {
1146 		CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
1147 		if (!fw_ref.domains)
1148 			return;
1149 
1150 		xe_mmio_write32(&gt->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST);
1151 
1152 		/*
1153 		 * FIXME: We can likely do better here with our choice of
1154 		 * timeout. Currently we just assume the worst case, i.e. 150us,
1155 		 * which is believed to be sufficient to cover the worst case
1156 		 * scenario on current platforms if all cache entries are
1157 		 * transient and need to be flushed..
1158 		 */
1159 		if (xe_mmio_wait32(&gt->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST, 0,
1160 				   300, NULL, false))
1161 			xe_gt_err_once(gt, "TD flush timeout\n");
1162 	}
1163 }
1164 
1165 /**
1166  * xe_device_is_l2_flush_optimized - if L2 flush is optimized by HW
1167  * @xe: The device to check.
1168  *
1169  * Return: true if the HW device optimizing L2 flush, false otherwise.
1170  */
1171 bool xe_device_is_l2_flush_optimized(struct xe_device *xe)
1172 {
1173 	/* XA is *always* flushed, like at the end-of-submssion (and maybe other
1174 	 * places), just that internally as an optimisation hw doesn't need to make
1175 	 * that a full flush (which will also include XA) when Media is
1176 	 * off/powergated, since it doesn't need to worry about GT caches vs Media
1177 	 * coherency, and only CPU vs GPU coherency, so can make that flush a
1178 	 * targeted XA flush, since stuff tagged with XA now means it's shared with
1179 	 * the CPU. The main implication is that we now need to somehow flush non-XA before
1180 	 * freeing system memory pages, otherwise dirty cachelines could be flushed after the free
1181 	 * (like if Media suddenly turns on and does a full flush)
1182 	 */
1183 	if (GRAPHICS_VER(xe) >= 35 && !IS_DGFX(xe))
1184 		return true;
1185 	return false;
1186 }
1187 
1188 void xe_device_l2_flush(struct xe_device *xe)
1189 {
1190 	struct xe_gt *gt;
1191 
1192 	gt = xe_root_mmio_gt(xe);
1193 	if (!gt)
1194 		return;
1195 
1196 	if (!XE_GT_WA(gt, 16023588340))
1197 		return;
1198 
1199 	CLASS(xe_force_wake, fw_ref)(gt_to_fw(gt), XE_FW_GT);
1200 	if (!fw_ref.domains)
1201 		return;
1202 
1203 	spin_lock(&gt->global_invl_lock);
1204 
1205 	xe_mmio_write32(&gt->mmio, XE2_GLOBAL_INVAL, 0x1);
1206 	if (xe_mmio_wait32(&gt->mmio, XE2_GLOBAL_INVAL, 0x1, 0x0, 1000, NULL, true))
1207 		xe_gt_err_once(gt, "Global invalidation timeout\n");
1208 
1209 	spin_unlock(&gt->global_invl_lock);
1210 }
1211 
1212 /**
1213  * xe_device_td_flush() - Flush transient L3 cache entries
1214  * @xe: The device
1215  *
1216  * Display engine has direct access to memory and is never coherent with L3/L4
1217  * caches (or CPU caches), however KMD is responsible for specifically flushing
1218  * transient L3 GPU cache entries prior to the flip sequence to ensure scanout
1219  * can happen from such a surface without seeing corruption.
1220  *
1221  * Display surfaces can be tagged as transient by mapping it using one of the
1222  * various L3:XD PAT index modes on Xe2.
1223  *
1224  * Note: On non-discrete xe2 platforms, like LNL, the entire L3 cache is flushed
1225  * at the end of each submission via PIPE_CONTROL for compute/render, since SA
1226  * Media is not coherent with L3 and we want to support render-vs-media
1227  * usescases. For other engines like copy/blt the HW internally forces uncached
1228  * behaviour, hence why we can skip the TDF on such platforms.
1229  */
1230 void xe_device_td_flush(struct xe_device *xe)
1231 {
1232 	struct xe_gt *root_gt;
1233 
1234 	/*
1235 	 * From Xe3p onward the HW takes care of flush of TD entries also along
1236 	 * with flushing XA entries, which will be at the usual sync points,
1237 	 * like at the end of submission, so no manual flush is needed here.
1238 	 */
1239 	if (GRAPHICS_VER(xe) >= 35)
1240 		return;
1241 
1242 	if (!IS_DGFX(xe) || GRAPHICS_VER(xe) < 20)
1243 		return;
1244 
1245 	root_gt = xe_root_mmio_gt(xe);
1246 	if (!root_gt)
1247 		return;
1248 
1249 	if (XE_GT_WA(root_gt, 16023588340)) {
1250 		/* A transient flush is not sufficient: flush the L2 */
1251 		xe_device_l2_flush(xe);
1252 	} else {
1253 		xe_guc_pc_apply_flush_freq_limit(&root_gt->uc.guc.pc);
1254 		tdf_request_sync(xe);
1255 		xe_guc_pc_remove_flush_freq_limit(&root_gt->uc.guc.pc);
1256 	}
1257 }
1258 
1259 u32 xe_device_ccs_bytes(struct xe_device *xe, u64 size)
1260 {
1261 	return xe_device_has_flat_ccs(xe) ?
1262 		DIV_ROUND_UP_ULL(size, NUM_BYTES_PER_CCS_BYTE(xe)) : 0;
1263 }
1264 
1265 /**
1266  * xe_device_assert_mem_access - Inspect the current runtime_pm state.
1267  * @xe: xe device instance
1268  *
1269  * To be used before any kind of memory access. It will splat a debug warning
1270  * if the device is currently sleeping. But it doesn't guarantee in any way
1271  * that the device is going to remain awake. Xe PM runtime get and put
1272  * functions might be added to the outer bound of the memory access, while
1273  * this check is intended for inner usage to splat some warning if the worst
1274  * case has just happened.
1275  */
1276 void xe_device_assert_mem_access(struct xe_device *xe)
1277 {
1278 	xe_assert(xe, !xe_pm_runtime_suspended(xe));
1279 }
1280 
1281 void xe_device_snapshot_print(struct xe_device *xe, struct drm_printer *p)
1282 {
1283 	struct xe_gt *gt;
1284 	u8 id;
1285 
1286 	drm_printf(p, "PCI ID: 0x%04x\n", xe->info.devid);
1287 	drm_printf(p, "PCI revision: 0x%02x\n", xe->info.revid);
1288 
1289 	for_each_gt(gt, xe, id) {
1290 		drm_printf(p, "GT id: %u\n", id);
1291 		drm_printf(p, "\tTile: %u\n", gt->tile->id);
1292 		drm_printf(p, "\tType: %s\n",
1293 			   gt->info.type == XE_GT_TYPE_MAIN ? "main" : "media");
1294 		drm_printf(p, "\tIP ver: %u.%u.%u\n",
1295 			   REG_FIELD_GET(GMD_ID_ARCH_MASK, gt->info.gmdid),
1296 			   REG_FIELD_GET(GMD_ID_RELEASE_MASK, gt->info.gmdid),
1297 			   REG_FIELD_GET(GMD_ID_REVID, gt->info.gmdid));
1298 		drm_printf(p, "\tCS reference clock: %u\n", gt->info.reference_clock);
1299 	}
1300 }
1301 
1302 u64 xe_device_canonicalize_addr(struct xe_device *xe, u64 address)
1303 {
1304 	return sign_extend64(address, xe->info.va_bits - 1);
1305 }
1306 
1307 u64 xe_device_uncanonicalize_addr(struct xe_device *xe, u64 address)
1308 {
1309 	return address & GENMASK_ULL(xe->info.va_bits - 1, 0);
1310 }
1311 
1312 /**
1313  * DOC: Xe Device Wedging
1314  *
1315  * Xe driver uses drm device wedged uevent as documented in Documentation/gpu/drm-uapi.rst.
1316  * When device is in wedged state, every IOCTL will be blocked and GT cannot
1317  * be used. The conditions under which the driver declares the device wedged
1318  * depend on the wedged mode configuration (see &enum xe_wedged_mode). The
1319  * default recovery method for a wedged state is rebind/bus-reset.
1320  *
1321  * Another recovery method is vendor-specific. Below are the cases that send
1322  * ``WEDGED=vendor-specific`` recovery method in drm device wedged uevent.
1323  *
1324  * Case: Firmware Flash
1325  * --------------------
1326  *
1327  * Identification Hint
1328  * +++++++++++++++++++
1329  *
1330  * ``WEDGED=vendor-specific`` drm device wedged uevent with
1331  * :ref:`Runtime Survivability mode <xe-survivability-mode>` is used to notify
1332  * admin/userspace consumer about the need for a firmware flash.
1333  *
1334  * Recovery Procedure
1335  * ++++++++++++++++++
1336  *
1337  * Once ``WEDGED=vendor-specific`` drm device wedged uevent is received, follow
1338  * the below steps
1339  *
1340  * - Check Runtime Survivability mode sysfs.
1341  *   If enabled, firmware flash is required to recover the device.
1342  *
1343  *   /sys/bus/pci/devices/<device>/survivability_mode
1344  *
1345  * - Admin/userspace consumer can use firmware flashing tools like fwupd to flash
1346  *   firmware and restore device to normal operation.
1347  */
1348 
1349 /**
1350  * xe_device_set_wedged_method - Set wedged recovery method
1351  * @xe: xe device instance
1352  * @method: recovery method to set
1353  *
1354  * Set wedged recovery method to be sent in drm wedged uevent.
1355  */
1356 void xe_device_set_wedged_method(struct xe_device *xe, unsigned long method)
1357 {
1358 	xe->wedged.method = method;
1359 }
1360 
1361 /**
1362  * xe_device_declare_wedged - Declare device wedged
1363  * @xe: xe device instance
1364  *
1365  * This is a final state that can only be cleared with the recovery method
1366  * specified in the drm wedged uevent. The method can be set using
1367  * xe_device_set_wedged_method before declaring the device as wedged. If no method
1368  * is set, reprobe (unbind/re-bind) will be sent by default.
1369  *
1370  * In this state every IOCTL will be blocked so the GT cannot be used.
1371  * In general it will be called upon any critical error such as gt reset
1372  * failure or guc loading failure. Userspace will be notified of this state
1373  * through device wedged uevent.
1374  * If xe.wedged module parameter is set to 2, this function will be called
1375  * on every single execution timeout (a.k.a. GPU hang) right after devcoredump
1376  * snapshot capture. In this mode, GT reset won't be attempted so the state of
1377  * the issue is preserved for further debugging.
1378  */
1379 void xe_device_declare_wedged(struct xe_device *xe)
1380 {
1381 	struct xe_gt *gt;
1382 	u8 id;
1383 
1384 	if (xe->wedged.mode == XE_WEDGED_MODE_NEVER) {
1385 		drm_dbg(&xe->drm, "Wedged mode is forcibly disabled\n");
1386 		return;
1387 	}
1388 
1389 	if (!atomic_xchg(&xe->wedged.flag, 1)) {
1390 		xe->needs_flr_on_fini = true;
1391 		xe_pm_runtime_get_noresume(xe);
1392 		drm_err(&xe->drm,
1393 			"CRITICAL: Xe has declared device %s as wedged.\n"
1394 			"IOCTLs and executions are blocked.\n"
1395 			"For recovery procedure, refer to https://docs.kernel.org/gpu/drm-uapi.html#device-wedging\n"
1396 			"Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/xe/kernel/issues/new\n",
1397 			dev_name(xe->drm.dev));
1398 	}
1399 
1400 	for_each_gt(gt, xe, id)
1401 		xe_gt_declare_wedged(gt);
1402 
1403 	if (xe_device_wedged(xe)) {
1404 		/*
1405 		 * XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET is intended for debugging
1406 		 * hangs, so wedge the device with 'none' recovery method and have
1407 		 * it available to the user for debugging.
1408 		 */
1409 		if (xe->wedged.mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET)
1410 			xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_NONE);
1411 		/* If no wedge recovery method is set, use default */
1412 		else if (!xe->wedged.method)
1413 			xe_device_set_wedged_method(xe, DRM_WEDGE_RECOVERY_REBIND |
1414 						    DRM_WEDGE_RECOVERY_BUS_RESET);
1415 
1416 		/* Notify userspace of wedged device */
1417 		drm_dev_wedged_event(&xe->drm, xe->wedged.method, NULL);
1418 	}
1419 }
1420 
1421 /**
1422  * xe_device_validate_wedged_mode - Check if given mode is supported
1423  * @xe: the &xe_device
1424  * @mode: requested mode to validate
1425  *
1426  * Check whether the provided wedged mode is supported.
1427  *
1428  * Return: 0 if mode is supported, error code otherwise.
1429  */
1430 int xe_device_validate_wedged_mode(struct xe_device *xe, unsigned int mode)
1431 {
1432 	if (mode > XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET) {
1433 		drm_dbg(&xe->drm, "wedged_mode: invalid value (%u)\n", mode);
1434 		return -EINVAL;
1435 	} else if (mode == XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET && (IS_SRIOV_VF(xe) ||
1436 		   (IS_SRIOV_PF(xe) && !IS_ENABLED(CONFIG_DRM_XE_DEBUG)))) {
1437 		drm_dbg(&xe->drm, "wedged_mode: (%u) %s mode is not supported for %s\n",
1438 			mode, xe_wedged_mode_to_string(mode),
1439 			xe_sriov_mode_to_string(xe_device_sriov_mode(xe)));
1440 		return -EPERM;
1441 	}
1442 
1443 	return 0;
1444 }
1445 
1446 /**
1447  * xe_wedged_mode_to_string - Convert enum value to string.
1448  * @mode: the &xe_wedged_mode to convert
1449  *
1450  * Returns: wedged mode as a user friendly string.
1451  */
1452 const char *xe_wedged_mode_to_string(enum xe_wedged_mode mode)
1453 {
1454 	switch (mode) {
1455 	case XE_WEDGED_MODE_NEVER:
1456 		return "never";
1457 	case XE_WEDGED_MODE_UPON_CRITICAL_ERROR:
1458 		return "upon-critical-error";
1459 	case XE_WEDGED_MODE_UPON_ANY_HANG_NO_RESET:
1460 		return "upon-any-hang-no-reset";
1461 	default:
1462 		return "<invalid>";
1463 	}
1464 }
1465 
1466 /**
1467  * xe_device_asid_to_vm() - Find VM from ASID
1468  * @xe: the &xe_device
1469  * @asid: Address space ID
1470  *
1471  * Find a VM from ASID and take a reference to VM which caller must drop.
1472  * Reclaim safe.
1473  *
1474  * Return: VM on success, ERR_PTR on failure
1475  */
1476 struct xe_vm *xe_device_asid_to_vm(struct xe_device *xe, u32 asid)
1477 {
1478 	struct xe_vm *vm;
1479 
1480 	down_read(&xe->usm.lock);
1481 	vm = xa_load(&xe->usm.asid_to_vm, asid);
1482 	if (vm)
1483 		xe_vm_get(vm);
1484 	else
1485 		vm = ERR_PTR(-EINVAL);
1486 	up_read(&xe->usm.lock);
1487 
1488 	return vm;
1489 }
1490