xref: /linux/drivers/gpu/drm/xe/xe_device.c (revision e327592cc901a3e415e28e1b6aa26381a2a93582)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include "xe_device.h"
7 
8 #include <linux/aperture.h>
9 #include <linux/delay.h>
10 #include <linux/fault-inject.h>
11 #include <linux/units.h>
12 
13 #include <drm/drm_atomic_helper.h>
14 #include <drm/drm_client.h>
15 #include <drm/drm_gem_ttm_helper.h>
16 #include <drm/drm_ioctl.h>
17 #include <drm/drm_managed.h>
18 #include <drm/drm_print.h>
19 #include <uapi/drm/xe_drm.h>
20 
21 #include "display/xe_display.h"
22 #include "instructions/xe_gpu_commands.h"
23 #include "regs/xe_gt_regs.h"
24 #include "regs/xe_regs.h"
25 #include "xe_bo.h"
26 #include "xe_bo_evict.h"
27 #include "xe_debugfs.h"
28 #include "xe_devcoredump.h"
29 #include "xe_device_sysfs.h"
30 #include "xe_dma_buf.h"
31 #include "xe_drm_client.h"
32 #include "xe_drv.h"
33 #include "xe_exec.h"
34 #include "xe_exec_queue.h"
35 #include "xe_force_wake.h"
36 #include "xe_ggtt.h"
37 #include "xe_gsc_proxy.h"
38 #include "xe_gt.h"
39 #include "xe_gt_mcr.h"
40 #include "xe_gt_printk.h"
41 #include "xe_gt_sriov_vf.h"
42 #include "xe_guc.h"
43 #include "xe_hw_engine_group.h"
44 #include "xe_hwmon.h"
45 #include "xe_irq.h"
46 #include "xe_memirq.h"
47 #include "xe_mmio.h"
48 #include "xe_module.h"
49 #include "xe_oa.h"
50 #include "xe_observation.h"
51 #include "xe_pat.h"
52 #include "xe_pcode.h"
53 #include "xe_pm.h"
54 #include "xe_pmu.h"
55 #include "xe_pxp.h"
56 #include "xe_query.h"
57 #include "xe_shrinker.h"
58 #include "xe_survivability_mode.h"
59 #include "xe_sriov.h"
60 #include "xe_tile.h"
61 #include "xe_ttm_stolen_mgr.h"
62 #include "xe_ttm_sys_mgr.h"
63 #include "xe_vm.h"
64 #include "xe_vram.h"
65 #include "xe_vsec.h"
66 #include "xe_wait_user_fence.h"
67 #include "xe_wa.h"
68 
69 #include <generated/xe_wa_oob.h>
70 
71 static int xe_file_open(struct drm_device *dev, struct drm_file *file)
72 {
73 	struct xe_device *xe = to_xe_device(dev);
74 	struct xe_drm_client *client;
75 	struct xe_file *xef;
76 	int ret = -ENOMEM;
77 	struct task_struct *task = NULL;
78 
79 	xef = kzalloc(sizeof(*xef), GFP_KERNEL);
80 	if (!xef)
81 		return ret;
82 
83 	client = xe_drm_client_alloc();
84 	if (!client) {
85 		kfree(xef);
86 		return ret;
87 	}
88 
89 	xef->drm = file;
90 	xef->client = client;
91 	xef->xe = xe;
92 
93 	mutex_init(&xef->vm.lock);
94 	xa_init_flags(&xef->vm.xa, XA_FLAGS_ALLOC1);
95 
96 	mutex_init(&xef->exec_queue.lock);
97 	xa_init_flags(&xef->exec_queue.xa, XA_FLAGS_ALLOC1);
98 
99 	file->driver_priv = xef;
100 	kref_init(&xef->refcount);
101 
102 	task = get_pid_task(rcu_access_pointer(file->pid), PIDTYPE_PID);
103 	if (task) {
104 		xef->process_name = kstrdup(task->comm, GFP_KERNEL);
105 		xef->pid = task->pid;
106 		put_task_struct(task);
107 	}
108 
109 	return 0;
110 }
111 
112 static void xe_file_destroy(struct kref *ref)
113 {
114 	struct xe_file *xef = container_of(ref, struct xe_file, refcount);
115 
116 	xa_destroy(&xef->exec_queue.xa);
117 	mutex_destroy(&xef->exec_queue.lock);
118 	xa_destroy(&xef->vm.xa);
119 	mutex_destroy(&xef->vm.lock);
120 
121 	xe_drm_client_put(xef->client);
122 	kfree(xef->process_name);
123 	kfree(xef);
124 }
125 
126 /**
127  * xe_file_get() - Take a reference to the xe file object
128  * @xef: Pointer to the xe file
129  *
130  * Anyone with a pointer to xef must take a reference to the xe file
131  * object using this call.
132  *
133  * Return: xe file pointer
134  */
135 struct xe_file *xe_file_get(struct xe_file *xef)
136 {
137 	kref_get(&xef->refcount);
138 	return xef;
139 }
140 
141 /**
142  * xe_file_put() - Drop a reference to the xe file object
143  * @xef: Pointer to the xe file
144  *
145  * Used to drop reference to the xef object
146  */
147 void xe_file_put(struct xe_file *xef)
148 {
149 	kref_put(&xef->refcount, xe_file_destroy);
150 }
151 
152 static void xe_file_close(struct drm_device *dev, struct drm_file *file)
153 {
154 	struct xe_device *xe = to_xe_device(dev);
155 	struct xe_file *xef = file->driver_priv;
156 	struct xe_vm *vm;
157 	struct xe_exec_queue *q;
158 	unsigned long idx;
159 
160 	xe_pm_runtime_get(xe);
161 
162 	/*
163 	 * No need for exec_queue.lock here as there is no contention for it
164 	 * when FD is closing as IOCTLs presumably can't be modifying the
165 	 * xarray. Taking exec_queue.lock here causes undue dependency on
166 	 * vm->lock taken during xe_exec_queue_kill().
167 	 */
168 	xa_for_each(&xef->exec_queue.xa, idx, q) {
169 		if (q->vm && q->hwe->hw_engine_group)
170 			xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q);
171 		xe_exec_queue_kill(q);
172 		xe_exec_queue_put(q);
173 	}
174 	xa_for_each(&xef->vm.xa, idx, vm)
175 		xe_vm_close_and_put(vm);
176 
177 	xe_file_put(xef);
178 
179 	xe_pm_runtime_put(xe);
180 }
181 
182 static const struct drm_ioctl_desc xe_ioctls[] = {
183 	DRM_IOCTL_DEF_DRV(XE_DEVICE_QUERY, xe_query_ioctl, DRM_RENDER_ALLOW),
184 	DRM_IOCTL_DEF_DRV(XE_GEM_CREATE, xe_gem_create_ioctl, DRM_RENDER_ALLOW),
185 	DRM_IOCTL_DEF_DRV(XE_GEM_MMAP_OFFSET, xe_gem_mmap_offset_ioctl,
186 			  DRM_RENDER_ALLOW),
187 	DRM_IOCTL_DEF_DRV(XE_VM_CREATE, xe_vm_create_ioctl, DRM_RENDER_ALLOW),
188 	DRM_IOCTL_DEF_DRV(XE_VM_DESTROY, xe_vm_destroy_ioctl, DRM_RENDER_ALLOW),
189 	DRM_IOCTL_DEF_DRV(XE_VM_BIND, xe_vm_bind_ioctl, DRM_RENDER_ALLOW),
190 	DRM_IOCTL_DEF_DRV(XE_EXEC, xe_exec_ioctl, DRM_RENDER_ALLOW),
191 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_CREATE, xe_exec_queue_create_ioctl,
192 			  DRM_RENDER_ALLOW),
193 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_DESTROY, xe_exec_queue_destroy_ioctl,
194 			  DRM_RENDER_ALLOW),
195 	DRM_IOCTL_DEF_DRV(XE_EXEC_QUEUE_GET_PROPERTY, xe_exec_queue_get_property_ioctl,
196 			  DRM_RENDER_ALLOW),
197 	DRM_IOCTL_DEF_DRV(XE_WAIT_USER_FENCE, xe_wait_user_fence_ioctl,
198 			  DRM_RENDER_ALLOW),
199 	DRM_IOCTL_DEF_DRV(XE_OBSERVATION, xe_observation_ioctl, DRM_RENDER_ALLOW),
200 };
201 
202 static long xe_drm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
203 {
204 	struct drm_file *file_priv = file->private_data;
205 	struct xe_device *xe = to_xe_device(file_priv->minor->dev);
206 	long ret;
207 
208 	if (xe_device_wedged(xe))
209 		return -ECANCELED;
210 
211 	ret = xe_pm_runtime_get_ioctl(xe);
212 	if (ret >= 0)
213 		ret = drm_ioctl(file, cmd, arg);
214 	xe_pm_runtime_put(xe);
215 
216 	return ret;
217 }
218 
219 #ifdef CONFIG_COMPAT
220 static long xe_drm_compat_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 	ret = xe_pm_runtime_get_ioctl(xe);
230 	if (ret >= 0)
231 		ret = drm_compat_ioctl(file, cmd, arg);
232 	xe_pm_runtime_put(xe);
233 
234 	return ret;
235 }
236 #else
237 /* similarly to drm_compat_ioctl, let's it be assigned to .compat_ioct unconditionally */
238 #define xe_drm_compat_ioctl NULL
239 #endif
240 
241 static void barrier_open(struct vm_area_struct *vma)
242 {
243 	drm_dev_get(vma->vm_private_data);
244 }
245 
246 static void barrier_close(struct vm_area_struct *vma)
247 {
248 	drm_dev_put(vma->vm_private_data);
249 }
250 
251 static void barrier_release_dummy_page(struct drm_device *dev, void *res)
252 {
253 	struct page *dummy_page = (struct page *)res;
254 
255 	__free_page(dummy_page);
256 }
257 
258 static vm_fault_t barrier_fault(struct vm_fault *vmf)
259 {
260 	struct drm_device *dev = vmf->vma->vm_private_data;
261 	struct vm_area_struct *vma = vmf->vma;
262 	vm_fault_t ret = VM_FAULT_NOPAGE;
263 	pgprot_t prot;
264 	int idx;
265 
266 	prot = vm_get_page_prot(vma->vm_flags);
267 
268 	if (drm_dev_enter(dev, &idx)) {
269 		unsigned long pfn;
270 
271 #define LAST_DB_PAGE_OFFSET 0x7ff001
272 		pfn = PHYS_PFN(pci_resource_start(to_pci_dev(dev->dev), 0) +
273 				LAST_DB_PAGE_OFFSET);
274 		ret = vmf_insert_pfn_prot(vma, vma->vm_start, pfn,
275 					  pgprot_noncached(prot));
276 		drm_dev_exit(idx);
277 	} else {
278 		struct page *page;
279 
280 		/* Allocate new dummy page to map all the VA range in this VMA to it*/
281 		page = alloc_page(GFP_KERNEL | __GFP_ZERO);
282 		if (!page)
283 			return VM_FAULT_OOM;
284 
285 		/* Set the page to be freed using drmm release action */
286 		if (drmm_add_action_or_reset(dev, barrier_release_dummy_page, page))
287 			return VM_FAULT_OOM;
288 
289 		ret = vmf_insert_pfn_prot(vma, vma->vm_start, page_to_pfn(page),
290 					  prot);
291 	}
292 
293 	return ret;
294 }
295 
296 static const struct vm_operations_struct vm_ops_barrier = {
297 	.open = barrier_open,
298 	.close = barrier_close,
299 	.fault = barrier_fault,
300 };
301 
302 static int xe_pci_barrier_mmap(struct file *filp,
303 			       struct vm_area_struct *vma)
304 {
305 	struct drm_file *priv = filp->private_data;
306 	struct drm_device *dev = priv->minor->dev;
307 	struct xe_device *xe = to_xe_device(dev);
308 
309 	if (!IS_DGFX(xe))
310 		return -EINVAL;
311 
312 	if (vma->vm_end - vma->vm_start > SZ_4K)
313 		return -EINVAL;
314 
315 	if (is_cow_mapping(vma->vm_flags))
316 		return -EINVAL;
317 
318 	if (vma->vm_flags & (VM_READ | VM_EXEC))
319 		return -EINVAL;
320 
321 	vm_flags_clear(vma, VM_MAYREAD | VM_MAYEXEC);
322 	vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP | VM_IO);
323 	vma->vm_ops = &vm_ops_barrier;
324 	vma->vm_private_data = dev;
325 	drm_dev_get(vma->vm_private_data);
326 
327 	return 0;
328 }
329 
330 static int xe_mmap(struct file *filp, struct vm_area_struct *vma)
331 {
332 	struct drm_file *priv = filp->private_data;
333 	struct drm_device *dev = priv->minor->dev;
334 
335 	if (drm_dev_is_unplugged(dev))
336 		return -ENODEV;
337 
338 	switch (vma->vm_pgoff) {
339 	case XE_PCI_BARRIER_MMAP_OFFSET >> XE_PTE_SHIFT:
340 		return xe_pci_barrier_mmap(filp, vma);
341 	}
342 
343 	return drm_gem_mmap(filp, vma);
344 }
345 
346 static const struct file_operations xe_driver_fops = {
347 	.owner = THIS_MODULE,
348 	.open = drm_open,
349 	.release = drm_release_noglobal,
350 	.unlocked_ioctl = xe_drm_ioctl,
351 	.mmap = xe_mmap,
352 	.poll = drm_poll,
353 	.read = drm_read,
354 	.compat_ioctl = xe_drm_compat_ioctl,
355 	.llseek = noop_llseek,
356 #ifdef CONFIG_PROC_FS
357 	.show_fdinfo = drm_show_fdinfo,
358 #endif
359 	.fop_flags = FOP_UNSIGNED_OFFSET,
360 };
361 
362 static struct drm_driver driver = {
363 	/* Don't use MTRRs here; the Xserver or userspace app should
364 	 * deal with them for Intel hardware.
365 	 */
366 	.driver_features =
367 	    DRIVER_GEM |
368 	    DRIVER_RENDER | DRIVER_SYNCOBJ |
369 	    DRIVER_SYNCOBJ_TIMELINE | DRIVER_GEM_GPUVA,
370 	.open = xe_file_open,
371 	.postclose = xe_file_close,
372 
373 	.gem_prime_import = xe_gem_prime_import,
374 
375 	.dumb_create = xe_bo_dumb_create,
376 	.dumb_map_offset = drm_gem_ttm_dumb_map_offset,
377 #ifdef CONFIG_PROC_FS
378 	.show_fdinfo = xe_drm_client_fdinfo,
379 #endif
380 	.ioctls = xe_ioctls,
381 	.num_ioctls = ARRAY_SIZE(xe_ioctls),
382 	.fops = &xe_driver_fops,
383 	.name = DRIVER_NAME,
384 	.desc = DRIVER_DESC,
385 	.major = DRIVER_MAJOR,
386 	.minor = DRIVER_MINOR,
387 	.patchlevel = DRIVER_PATCHLEVEL,
388 };
389 
390 static void xe_device_destroy(struct drm_device *dev, void *dummy)
391 {
392 	struct xe_device *xe = to_xe_device(dev);
393 
394 	xe_bo_dev_fini(&xe->bo_device);
395 
396 	if (xe->preempt_fence_wq)
397 		destroy_workqueue(xe->preempt_fence_wq);
398 
399 	if (xe->ordered_wq)
400 		destroy_workqueue(xe->ordered_wq);
401 
402 	if (xe->unordered_wq)
403 		destroy_workqueue(xe->unordered_wq);
404 
405 	if (xe->destroy_wq)
406 		destroy_workqueue(xe->destroy_wq);
407 
408 	ttm_device_fini(&xe->ttm);
409 }
410 
411 struct xe_device *xe_device_create(struct pci_dev *pdev,
412 				   const struct pci_device_id *ent)
413 {
414 	struct xe_device *xe;
415 	int err;
416 
417 	xe_display_driver_set_hooks(&driver);
418 
419 	err = aperture_remove_conflicting_pci_devices(pdev, driver.name);
420 	if (err)
421 		return ERR_PTR(err);
422 
423 	xe = devm_drm_dev_alloc(&pdev->dev, &driver, struct xe_device, drm);
424 	if (IS_ERR(xe))
425 		return xe;
426 
427 	err = ttm_device_init(&xe->ttm, &xe_ttm_funcs, xe->drm.dev,
428 			      xe->drm.anon_inode->i_mapping,
429 			      xe->drm.vma_offset_manager, false, false);
430 	if (WARN_ON(err))
431 		goto err;
432 
433 	xe_bo_dev_init(&xe->bo_device);
434 	err = drmm_add_action_or_reset(&xe->drm, xe_device_destroy, NULL);
435 	if (err)
436 		goto err;
437 
438 	err = xe_shrinker_create(xe);
439 	if (err)
440 		goto err;
441 
442 	xe->info.devid = pdev->device;
443 	xe->info.revid = pdev->revision;
444 	xe->info.force_execlist = xe_modparam.force_execlist;
445 
446 	err = xe_irq_init(xe);
447 	if (err)
448 		goto err;
449 
450 	init_waitqueue_head(&xe->ufence_wq);
451 
452 	init_rwsem(&xe->usm.lock);
453 
454 	xa_init_flags(&xe->usm.asid_to_vm, XA_FLAGS_ALLOC);
455 
456 	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) {
457 		/* Trigger a large asid and an early asid wrap. */
458 		u32 asid;
459 
460 		BUILD_BUG_ON(XE_MAX_ASID < 2);
461 		err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, NULL,
462 				      XA_LIMIT(XE_MAX_ASID - 2, XE_MAX_ASID - 1),
463 				      &xe->usm.next_asid, GFP_KERNEL);
464 		drm_WARN_ON(&xe->drm, err);
465 		if (err >= 0)
466 			xa_erase(&xe->usm.asid_to_vm, asid);
467 	}
468 
469 	err = xe_bo_pinned_init(xe);
470 	if (err)
471 		goto err;
472 
473 	xe->preempt_fence_wq = alloc_ordered_workqueue("xe-preempt-fence-wq",
474 						       WQ_MEM_RECLAIM);
475 	xe->ordered_wq = alloc_ordered_workqueue("xe-ordered-wq", 0);
476 	xe->unordered_wq = alloc_workqueue("xe-unordered-wq", 0, 0);
477 	xe->destroy_wq = alloc_workqueue("xe-destroy-wq", 0, 0);
478 	if (!xe->ordered_wq || !xe->unordered_wq ||
479 	    !xe->preempt_fence_wq || !xe->destroy_wq) {
480 		/*
481 		 * Cleanup done in xe_device_destroy via
482 		 * drmm_add_action_or_reset register above
483 		 */
484 		drm_err(&xe->drm, "Failed to allocate xe workqueues\n");
485 		err = -ENOMEM;
486 		goto err;
487 	}
488 
489 	err = drmm_mutex_init(&xe->drm, &xe->pmt.lock);
490 	if (err)
491 		goto err;
492 
493 	err = xe_display_create(xe);
494 	if (WARN_ON(err))
495 		goto err;
496 
497 	return xe;
498 
499 err:
500 	return ERR_PTR(err);
501 }
502 ALLOW_ERROR_INJECTION(xe_device_create, ERRNO); /* See xe_pci_probe() */
503 
504 static bool xe_driver_flr_disabled(struct xe_device *xe)
505 {
506 	if (IS_SRIOV_VF(xe))
507 		return true;
508 
509 	if (xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL_PROTECTED) & DRIVERINT_FLR_DIS) {
510 		drm_info(&xe->drm, "Driver-FLR disabled by BIOS\n");
511 		return true;
512 	}
513 
514 	return false;
515 }
516 
517 /*
518  * The driver-initiated FLR is the highest level of reset that we can trigger
519  * from within the driver. It is different from the PCI FLR in that it doesn't
520  * fully reset the SGUnit and doesn't modify the PCI config space and therefore
521  * it doesn't require a re-enumeration of the PCI BARs. However, the
522  * driver-initiated FLR does still cause a reset of both GT and display and a
523  * memory wipe of local and stolen memory, so recovery would require a full HW
524  * re-init and saving/restoring (or re-populating) the wiped memory. Since we
525  * perform the FLR as the very last action before releasing access to the HW
526  * during the driver release flow, we don't attempt recovery at all, because
527  * if/when a new instance of i915 is bound to the device it will do a full
528  * re-init anyway.
529  */
530 static void __xe_driver_flr(struct xe_device *xe)
531 {
532 	const unsigned int flr_timeout = 3 * USEC_PER_SEC; /* specs recommend a 3s wait */
533 	struct xe_mmio *mmio = xe_root_tile_mmio(xe);
534 	int ret;
535 
536 	drm_dbg(&xe->drm, "Triggering Driver-FLR\n");
537 
538 	/*
539 	 * Make sure any pending FLR requests have cleared by waiting for the
540 	 * FLR trigger bit to go to zero. Also clear GU_DEBUG's DRIVERFLR_STATUS
541 	 * to make sure it's not still set from a prior attempt (it's a write to
542 	 * clear bit).
543 	 * Note that we should never be in a situation where a previous attempt
544 	 * is still pending (unless the HW is totally dead), but better to be
545 	 * safe in case something unexpected happens
546 	 */
547 	ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false);
548 	if (ret) {
549 		drm_err(&xe->drm, "Driver-FLR-prepare wait for ready failed! %d\n", ret);
550 		return;
551 	}
552 	xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS);
553 
554 	/* Trigger the actual Driver-FLR */
555 	xe_mmio_rmw32(mmio, GU_CNTL, 0, DRIVERFLR);
556 
557 	/* Wait for hardware teardown to complete */
558 	ret = xe_mmio_wait32(mmio, GU_CNTL, DRIVERFLR, 0, flr_timeout, NULL, false);
559 	if (ret) {
560 		drm_err(&xe->drm, "Driver-FLR-teardown wait completion failed! %d\n", ret);
561 		return;
562 	}
563 
564 	/* Wait for hardware/firmware re-init to complete */
565 	ret = xe_mmio_wait32(mmio, GU_DEBUG, DRIVERFLR_STATUS, DRIVERFLR_STATUS,
566 			     flr_timeout, NULL, false);
567 	if (ret) {
568 		drm_err(&xe->drm, "Driver-FLR-reinit wait completion failed! %d\n", ret);
569 		return;
570 	}
571 
572 	/* Clear sticky completion status */
573 	xe_mmio_write32(mmio, GU_DEBUG, DRIVERFLR_STATUS);
574 }
575 
576 static void xe_driver_flr(struct xe_device *xe)
577 {
578 	if (xe_driver_flr_disabled(xe))
579 		return;
580 
581 	__xe_driver_flr(xe);
582 }
583 
584 static void xe_driver_flr_fini(void *arg)
585 {
586 	struct xe_device *xe = arg;
587 
588 	if (xe->needs_flr_on_fini)
589 		xe_driver_flr(xe);
590 }
591 
592 static void xe_device_sanitize(void *arg)
593 {
594 	struct xe_device *xe = arg;
595 	struct xe_gt *gt;
596 	u8 id;
597 
598 	for_each_gt(gt, xe, id)
599 		xe_gt_sanitize(gt);
600 }
601 
602 static int xe_set_dma_info(struct xe_device *xe)
603 {
604 	unsigned int mask_size = xe->info.dma_mask_size;
605 	int err;
606 
607 	dma_set_max_seg_size(xe->drm.dev, xe_sg_segment_size(xe->drm.dev));
608 
609 	err = dma_set_mask(xe->drm.dev, DMA_BIT_MASK(mask_size));
610 	if (err)
611 		goto mask_err;
612 
613 	err = dma_set_coherent_mask(xe->drm.dev, DMA_BIT_MASK(mask_size));
614 	if (err)
615 		goto mask_err;
616 
617 	return 0;
618 
619 mask_err:
620 	drm_err(&xe->drm, "Can't set DMA mask/consistent mask (%d)\n", err);
621 	return err;
622 }
623 
624 static bool verify_lmem_ready(struct xe_device *xe)
625 {
626 	u32 val = xe_mmio_read32(xe_root_tile_mmio(xe), GU_CNTL) & LMEM_INIT;
627 
628 	return !!val;
629 }
630 
631 static int wait_for_lmem_ready(struct xe_device *xe)
632 {
633 	unsigned long timeout, start;
634 
635 	if (!IS_DGFX(xe))
636 		return 0;
637 
638 	if (IS_SRIOV_VF(xe))
639 		return 0;
640 
641 	if (verify_lmem_ready(xe))
642 		return 0;
643 
644 	drm_dbg(&xe->drm, "Waiting for lmem initialization\n");
645 
646 	start = jiffies;
647 	timeout = start + secs_to_jiffies(60); /* 60 sec! */
648 
649 	do {
650 		if (signal_pending(current))
651 			return -EINTR;
652 
653 		/*
654 		 * The boot firmware initializes local memory and
655 		 * assesses its health. If memory training fails,
656 		 * the punit will have been instructed to keep the GT powered
657 		 * down.we won't be able to communicate with it
658 		 *
659 		 * If the status check is done before punit updates the register,
660 		 * it can lead to the system being unusable.
661 		 * use a timeout and defer the probe to prevent this.
662 		 */
663 		if (time_after(jiffies, timeout)) {
664 			drm_dbg(&xe->drm, "lmem not initialized by firmware\n");
665 			return -EPROBE_DEFER;
666 		}
667 
668 		msleep(20);
669 
670 	} while (!verify_lmem_ready(xe));
671 
672 	drm_dbg(&xe->drm, "lmem ready after %ums",
673 		jiffies_to_msecs(jiffies - start));
674 
675 	return 0;
676 }
677 ALLOW_ERROR_INJECTION(wait_for_lmem_ready, ERRNO); /* See xe_pci_probe() */
678 
679 static void sriov_update_device_info(struct xe_device *xe)
680 {
681 	/* disable features that are not available/applicable to VFs */
682 	if (IS_SRIOV_VF(xe)) {
683 		xe->info.probe_display = 0;
684 		xe->info.has_heci_gscfi = 0;
685 		xe->info.skip_guc_pc = 1;
686 		xe->info.skip_pcode = 1;
687 	}
688 }
689 
690 /**
691  * xe_device_probe_early: Device early probe
692  * @xe: xe device instance
693  *
694  * Initialize MMIO resources that don't require any
695  * knowledge about tile count. Also initialize pcode and
696  * check vram initialization on root tile.
697  *
698  * Return: 0 on success, error code on failure
699  */
700 int xe_device_probe_early(struct xe_device *xe)
701 {
702 	int err;
703 
704 	err = xe_mmio_probe_early(xe);
705 	if (err)
706 		return err;
707 
708 	xe_sriov_probe_early(xe);
709 
710 	sriov_update_device_info(xe);
711 
712 	err = xe_pcode_probe_early(xe);
713 	if (err || xe_survivability_mode_is_requested(xe)) {
714 		int save_err = err;
715 
716 		/*
717 		 * Try to leave device in survivability mode if device is
718 		 * possible, but still return the previous error for error
719 		 * propagation
720 		 */
721 		err = xe_survivability_mode_enable(xe);
722 		if (err)
723 			return err;
724 
725 		return save_err;
726 	}
727 
728 	err = wait_for_lmem_ready(xe);
729 	if (err)
730 		return err;
731 
732 	xe->wedged.mode = xe_modparam.wedged_mode;
733 
734 	return 0;
735 }
736 ALLOW_ERROR_INJECTION(xe_device_probe_early, ERRNO); /* See xe_pci_probe() */
737 
738 static int probe_has_flat_ccs(struct xe_device *xe)
739 {
740 	struct xe_gt *gt;
741 	unsigned int fw_ref;
742 	u32 reg;
743 
744 	/* Always enabled/disabled, no runtime check to do */
745 	if (GRAPHICS_VER(xe) < 20 || !xe->info.has_flat_ccs || IS_SRIOV_VF(xe))
746 		return 0;
747 
748 	gt = xe_root_mmio_gt(xe);
749 
750 	fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
751 	if (!fw_ref)
752 		return -ETIMEDOUT;
753 
754 	reg = xe_gt_mcr_unicast_read_any(gt, XE2_FLAT_CCS_BASE_RANGE_LOWER);
755 	xe->info.has_flat_ccs = (reg & XE2_FLAT_CCS_ENABLE);
756 
757 	if (!xe->info.has_flat_ccs)
758 		drm_dbg(&xe->drm,
759 			"Flat CCS has been disabled in bios, May lead to performance impact");
760 
761 	xe_force_wake_put(gt_to_fw(gt), fw_ref);
762 
763 	return 0;
764 }
765 
766 int xe_device_probe(struct xe_device *xe)
767 {
768 	struct xe_tile *tile;
769 	struct xe_gt *gt;
770 	int err;
771 	u8 id;
772 
773 	xe_pat_init_early(xe);
774 
775 	err = xe_sriov_init(xe);
776 	if (err)
777 		return err;
778 
779 	xe->info.mem_region_mask = 1;
780 
781 	err = xe_set_dma_info(xe);
782 	if (err)
783 		return err;
784 
785 	err = xe_mmio_probe_tiles(xe);
786 	if (err)
787 		return err;
788 
789 	err = xe_ttm_sys_mgr_init(xe);
790 	if (err)
791 		return err;
792 
793 	for_each_gt(gt, xe, id) {
794 		err = xe_gt_init_early(gt);
795 		if (err)
796 			return err;
797 
798 		/*
799 		 * Only after this point can GT-specific MMIO operations
800 		 * (including things like communication with the GuC)
801 		 * be performed.
802 		 */
803 		xe_gt_mmio_init(gt);
804 	}
805 
806 	for_each_tile(tile, xe, id) {
807 		if (IS_SRIOV_VF(xe)) {
808 			xe_guc_comm_init_early(&tile->primary_gt->uc.guc);
809 			err = xe_gt_sriov_vf_bootstrap(tile->primary_gt);
810 			if (err)
811 				return err;
812 			err = xe_gt_sriov_vf_query_config(tile->primary_gt);
813 			if (err)
814 				return err;
815 		}
816 		err = xe_ggtt_init_early(tile->mem.ggtt);
817 		if (err)
818 			return err;
819 		err = xe_memirq_init(&tile->memirq);
820 		if (err)
821 			return err;
822 	}
823 
824 	for_each_gt(gt, xe, id) {
825 		err = xe_gt_init_hwconfig(gt);
826 		if (err)
827 			return err;
828 	}
829 
830 	err = xe_devcoredump_init(xe);
831 	if (err)
832 		return err;
833 
834 	/*
835 	 * From here on, if a step fails, make sure a Driver-FLR is triggereed
836 	 */
837 	err = devm_add_action_or_reset(xe->drm.dev, xe_driver_flr_fini, xe);
838 	if (err)
839 		return err;
840 
841 	err = probe_has_flat_ccs(xe);
842 	if (err)
843 		return err;
844 
845 	err = xe_vram_probe(xe);
846 	if (err)
847 		return err;
848 
849 	for_each_tile(tile, xe, id) {
850 		err = xe_tile_init_noalloc(tile);
851 		if (err)
852 			return err;
853 	}
854 
855 	/* Allocate and map stolen after potential VRAM resize */
856 	err = xe_ttm_stolen_mgr_init(xe);
857 	if (err)
858 		return err;
859 
860 	/*
861 	 * Now that GT is initialized (TTM in particular),
862 	 * we can try to init display, and inherit the initial fb.
863 	 * This is the reason the first allocation needs to be done
864 	 * inside display.
865 	 */
866 	err = xe_display_init_early(xe);
867 	if (err)
868 		return err;
869 
870 	for_each_tile(tile, xe, id) {
871 		err = xe_tile_init(tile);
872 		if (err)
873 			return err;
874 	}
875 
876 	err = xe_irq_install(xe);
877 	if (err)
878 		return err;
879 
880 	for_each_gt(gt, xe, id) {
881 		err = xe_gt_init(gt);
882 		if (err)
883 			return err;
884 	}
885 
886 	err = xe_heci_gsc_init(xe);
887 	if (err)
888 		return err;
889 
890 	err = xe_oa_init(xe);
891 	if (err)
892 		return err;
893 
894 	err = xe_display_init(xe);
895 	if (err)
896 		return err;
897 
898 	err = xe_pxp_init(xe);
899 	if (err)
900 		return err;
901 
902 	err = drm_dev_register(&xe->drm, 0);
903 	if (err)
904 		return err;
905 
906 	xe_display_register(xe);
907 
908 	err = xe_oa_register(xe);
909 	if (err)
910 		goto err_unregister_display;
911 
912 	err = xe_pmu_register(&xe->pmu);
913 	if (err)
914 		goto err_unregister_display;
915 
916 	err = xe_device_sysfs_init(xe);
917 	if (err)
918 		goto err_unregister_display;
919 
920 	xe_debugfs_register(xe);
921 
922 	err = xe_hwmon_register(xe);
923 	if (err)
924 		goto err_unregister_display;
925 
926 	for_each_gt(gt, xe, id)
927 		xe_gt_sanitize_freq(gt);
928 
929 	xe_vsec_init(xe);
930 
931 	return devm_add_action_or_reset(xe->drm.dev, xe_device_sanitize, xe);
932 
933 err_unregister_display:
934 	xe_display_unregister(xe);
935 
936 	return err;
937 }
938 
939 void xe_device_remove(struct xe_device *xe)
940 {
941 	xe_display_unregister(xe);
942 
943 	drm_dev_unplug(&xe->drm);
944 
945 	xe_bo_pci_dev_remove_all(xe);
946 }
947 
948 void xe_device_shutdown(struct xe_device *xe)
949 {
950 	struct xe_gt *gt;
951 	u8 id;
952 
953 	drm_dbg(&xe->drm, "Shutting down device\n");
954 
955 	if (xe_driver_flr_disabled(xe)) {
956 		xe_display_pm_shutdown(xe);
957 
958 		xe_irq_suspend(xe);
959 
960 		for_each_gt(gt, xe, id)
961 			xe_gt_shutdown(gt);
962 
963 		xe_display_pm_shutdown_late(xe);
964 	} else {
965 		/* BOOM! */
966 		__xe_driver_flr(xe);
967 	}
968 }
969 
970 /**
971  * xe_device_wmb() - Device specific write memory barrier
972  * @xe: the &xe_device
973  *
974  * While wmb() is sufficient for a barrier if we use system memory, on discrete
975  * platforms with device memory we additionally need to issue a register write.
976  * Since it doesn't matter which register we write to, use the read-only VF_CAP
977  * register that is also marked as accessible by the VFs.
978  */
979 void xe_device_wmb(struct xe_device *xe)
980 {
981 	wmb();
982 	if (IS_DGFX(xe))
983 		xe_mmio_write32(xe_root_tile_mmio(xe), VF_CAP_REG, 0);
984 }
985 
986 /**
987  * xe_device_td_flush() - Flush transient L3 cache entries
988  * @xe: The device
989  *
990  * Display engine has direct access to memory and is never coherent with L3/L4
991  * caches (or CPU caches), however KMD is responsible for specifically flushing
992  * transient L3 GPU cache entries prior to the flip sequence to ensure scanout
993  * can happen from such a surface without seeing corruption.
994  *
995  * Display surfaces can be tagged as transient by mapping it using one of the
996  * various L3:XD PAT index modes on Xe2.
997  *
998  * Note: On non-discrete xe2 platforms, like LNL, the entire L3 cache is flushed
999  * at the end of each submission via PIPE_CONTROL for compute/render, since SA
1000  * Media is not coherent with L3 and we want to support render-vs-media
1001  * usescases. For other engines like copy/blt the HW internally forces uncached
1002  * behaviour, hence why we can skip the TDF on such platforms.
1003  */
1004 void xe_device_td_flush(struct xe_device *xe)
1005 {
1006 	struct xe_gt *gt;
1007 	unsigned int fw_ref;
1008 	u8 id;
1009 
1010 	if (!IS_DGFX(xe) || GRAPHICS_VER(xe) < 20)
1011 		return;
1012 
1013 	if (XE_WA(xe_root_mmio_gt(xe), 16023588340)) {
1014 		xe_device_l2_flush(xe);
1015 		return;
1016 	}
1017 
1018 	for_each_gt(gt, xe, id) {
1019 		if (xe_gt_is_media_type(gt))
1020 			continue;
1021 
1022 		fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
1023 		if (!fw_ref)
1024 			return;
1025 
1026 		xe_mmio_write32(&gt->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST);
1027 		/*
1028 		 * FIXME: We can likely do better here with our choice of
1029 		 * timeout. Currently we just assume the worst case, i.e. 150us,
1030 		 * which is believed to be sufficient to cover the worst case
1031 		 * scenario on current platforms if all cache entries are
1032 		 * transient and need to be flushed..
1033 		 */
1034 		if (xe_mmio_wait32(&gt->mmio, XE2_TDF_CTRL, TRANSIENT_FLUSH_REQUEST, 0,
1035 				   150, NULL, false))
1036 			xe_gt_err_once(gt, "TD flush timeout\n");
1037 
1038 		xe_force_wake_put(gt_to_fw(gt), fw_ref);
1039 	}
1040 }
1041 
1042 void xe_device_l2_flush(struct xe_device *xe)
1043 {
1044 	struct xe_gt *gt;
1045 	unsigned int fw_ref;
1046 
1047 	gt = xe_root_mmio_gt(xe);
1048 
1049 	if (!XE_WA(gt, 16023588340))
1050 		return;
1051 
1052 	fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
1053 	if (!fw_ref)
1054 		return;
1055 
1056 	spin_lock(&gt->global_invl_lock);
1057 	xe_mmio_write32(&gt->mmio, XE2_GLOBAL_INVAL, 0x1);
1058 
1059 	if (xe_mmio_wait32(&gt->mmio, XE2_GLOBAL_INVAL, 0x1, 0x0, 500, NULL, true))
1060 		xe_gt_err_once(gt, "Global invalidation timeout\n");
1061 	spin_unlock(&gt->global_invl_lock);
1062 
1063 	xe_force_wake_put(gt_to_fw(gt), fw_ref);
1064 }
1065 
1066 u32 xe_device_ccs_bytes(struct xe_device *xe, u64 size)
1067 {
1068 	return xe_device_has_flat_ccs(xe) ?
1069 		DIV_ROUND_UP_ULL(size, NUM_BYTES_PER_CCS_BYTE(xe)) : 0;
1070 }
1071 
1072 /**
1073  * xe_device_assert_mem_access - Inspect the current runtime_pm state.
1074  * @xe: xe device instance
1075  *
1076  * To be used before any kind of memory access. It will splat a debug warning
1077  * if the device is currently sleeping. But it doesn't guarantee in any way
1078  * that the device is going to remain awake. Xe PM runtime get and put
1079  * functions might be added to the outer bound of the memory access, while
1080  * this check is intended for inner usage to splat some warning if the worst
1081  * case has just happened.
1082  */
1083 void xe_device_assert_mem_access(struct xe_device *xe)
1084 {
1085 	xe_assert(xe, !xe_pm_runtime_suspended(xe));
1086 }
1087 
1088 void xe_device_snapshot_print(struct xe_device *xe, struct drm_printer *p)
1089 {
1090 	struct xe_gt *gt;
1091 	u8 id;
1092 
1093 	drm_printf(p, "PCI ID: 0x%04x\n", xe->info.devid);
1094 	drm_printf(p, "PCI revision: 0x%02x\n", xe->info.revid);
1095 
1096 	for_each_gt(gt, xe, id) {
1097 		drm_printf(p, "GT id: %u\n", id);
1098 		drm_printf(p, "\tTile: %u\n", gt->tile->id);
1099 		drm_printf(p, "\tType: %s\n",
1100 			   gt->info.type == XE_GT_TYPE_MAIN ? "main" : "media");
1101 		drm_printf(p, "\tIP ver: %u.%u.%u\n",
1102 			   REG_FIELD_GET(GMD_ID_ARCH_MASK, gt->info.gmdid),
1103 			   REG_FIELD_GET(GMD_ID_RELEASE_MASK, gt->info.gmdid),
1104 			   REG_FIELD_GET(GMD_ID_REVID, gt->info.gmdid));
1105 		drm_printf(p, "\tCS reference clock: %u\n", gt->info.reference_clock);
1106 	}
1107 }
1108 
1109 u64 xe_device_canonicalize_addr(struct xe_device *xe, u64 address)
1110 {
1111 	return sign_extend64(address, xe->info.va_bits - 1);
1112 }
1113 
1114 u64 xe_device_uncanonicalize_addr(struct xe_device *xe, u64 address)
1115 {
1116 	return address & GENMASK_ULL(xe->info.va_bits - 1, 0);
1117 }
1118 
1119 static void xe_device_wedged_fini(struct drm_device *drm, void *arg)
1120 {
1121 	struct xe_device *xe = arg;
1122 
1123 	xe_pm_runtime_put(xe);
1124 }
1125 
1126 /**
1127  * xe_device_declare_wedged - Declare device wedged
1128  * @xe: xe device instance
1129  *
1130  * This is a final state that can only be cleared with a module
1131  * re-probe (unbind + bind).
1132  * In this state every IOCTL will be blocked so the GT cannot be used.
1133  * In general it will be called upon any critical error such as gt reset
1134  * failure or guc loading failure. Userspace will be notified of this state
1135  * through device wedged uevent.
1136  * If xe.wedged module parameter is set to 2, this function will be called
1137  * on every single execution timeout (a.k.a. GPU hang) right after devcoredump
1138  * snapshot capture. In this mode, GT reset won't be attempted so the state of
1139  * the issue is preserved for further debugging.
1140  */
1141 void xe_device_declare_wedged(struct xe_device *xe)
1142 {
1143 	struct xe_gt *gt;
1144 	u8 id;
1145 
1146 	if (xe->wedged.mode == 0) {
1147 		drm_dbg(&xe->drm, "Wedged mode is forcibly disabled\n");
1148 		return;
1149 	}
1150 
1151 	xe_pm_runtime_get_noresume(xe);
1152 
1153 	if (drmm_add_action_or_reset(&xe->drm, xe_device_wedged_fini, xe)) {
1154 		drm_err(&xe->drm, "Failed to register xe_device_wedged_fini clean-up. Although device is wedged.\n");
1155 		return;
1156 	}
1157 
1158 	if (!atomic_xchg(&xe->wedged.flag, 1)) {
1159 		xe->needs_flr_on_fini = true;
1160 		drm_err(&xe->drm,
1161 			"CRITICAL: Xe has declared device %s as wedged.\n"
1162 			"IOCTLs and executions are blocked. Only a rebind may clear the failure\n"
1163 			"Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/xe/kernel/issues/new\n",
1164 			dev_name(xe->drm.dev));
1165 
1166 		/* Notify userspace of wedged device */
1167 		drm_dev_wedged_event(&xe->drm,
1168 				     DRM_WEDGE_RECOVERY_REBIND | DRM_WEDGE_RECOVERY_BUS_RESET);
1169 	}
1170 
1171 	for_each_gt(gt, xe, id)
1172 		xe_gt_declare_wedged(gt);
1173 }
1174