main.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Copyright © 2025 Intel Corporation
  4 */
  5
  6#include <linux/anon_inodes.h>
  7#include <linux/delay.h>
  8#include <linux/file.h>
  9#include <linux/module.h>
 10#include <linux/pci.h>
 11#include <linux/sizes.h>
 12#include <linux/types.h>
 13#include <linux/vfio.h>
 14#include <linux/vfio_pci_core.h>
 15
 16#include <drm/intel/xe_sriov_vfio.h>
 17#include <drm/intel/pciids.h>
 18
 19struct xe_vfio_pci_migration_file {
 20	struct file *filp;
 21	/* serializes accesses to migration data */
 22	struct mutex lock;
 23	struct xe_vfio_pci_core_device *xe_vdev;
 24	u8 disabled:1;
 25};
 26
 27struct xe_vfio_pci_core_device {
 28	struct vfio_pci_core_device core_device;
 29	struct xe_device *xe;
 30	/* PF internal control uses vfid index starting from 1 */
 31	unsigned int vfid;
 32	u8 deferred_reset:1;
 33	/* protects migration state */
 34	struct mutex state_mutex;
 35	enum vfio_device_mig_state mig_state;
 36	/* protects the reset_done flow */
 37	spinlock_t reset_lock;
 38	struct xe_vfio_pci_migration_file *migf;
 39};
 40
 41#define xe_vdev_to_dev(xe_vdev) (&(xe_vdev)->core_device.pdev->dev)
 42
 43static void xe_vfio_pci_disable_file(struct xe_vfio_pci_migration_file *migf)
 44{
 45	mutex_lock(&migf->lock);
 46	migf->disabled = true;
 47	mutex_unlock(&migf->lock);
 48}
 49
 50static void xe_vfio_pci_put_file(struct xe_vfio_pci_core_device *xe_vdev)
 51{
 52	xe_vfio_pci_disable_file(xe_vdev->migf);
 53	fput(xe_vdev->migf->filp);
 54	xe_vdev->migf = NULL;
 55}
 56
 57static void xe_vfio_pci_reset(struct xe_vfio_pci_core_device *xe_vdev)
 58{
 59	if (xe_vdev->migf)
 60		xe_vfio_pci_put_file(xe_vdev);
 61
 62	xe_vdev->mig_state = VFIO_DEVICE_STATE_RUNNING;
 63}
 64
 65static void xe_vfio_pci_state_mutex_lock(struct xe_vfio_pci_core_device *xe_vdev)
 66{
 67	mutex_lock(&xe_vdev->state_mutex);
 68}
 69
 70/*
 71 * This function is called in all state_mutex unlock cases to
 72 * handle a 'deferred_reset' if exists.
 73 */
 74static void xe_vfio_pci_state_mutex_unlock(struct xe_vfio_pci_core_device *xe_vdev)
 75{
 76again:
 77	spin_lock(&xe_vdev->reset_lock);
 78	if (xe_vdev->deferred_reset) {
 79		xe_vdev->deferred_reset = false;
 80		spin_unlock(&xe_vdev->reset_lock);
 81		xe_vfio_pci_reset(xe_vdev);
 82		goto again;
 83	}
 84	mutex_unlock(&xe_vdev->state_mutex);
 85	spin_unlock(&xe_vdev->reset_lock);
 86}
 87
 88static void xe_vfio_pci_reset_done(struct pci_dev *pdev)
 89{
 90	struct xe_vfio_pci_core_device *xe_vdev = pci_get_drvdata(pdev);
 91	int ret;
 92
 93	if (!pdev->is_virtfn)
 94		return;
 95
 96	/*
 97	 * VF FLR requires additional processing done by PF driver.
 98	 * The processing is done after FLR is already finished from PCIe
 99	 * perspective.
100	 * In order to avoid a scenario where VF is used while PF processing
101	 * is still in progress, additional synchronization point is needed.
102	 */
103	ret = xe_sriov_vfio_wait_flr_done(xe_vdev->xe, xe_vdev->vfid);
104	if (ret)
105		dev_err(&pdev->dev, "Failed to wait for FLR: %d\n", ret);
106
107	if (!xe_vdev->vfid)
108		return;
109
110	/*
111	 * As the higher VFIO layers are holding locks across reset and using
112	 * those same locks with the mm_lock we need to prevent ABBA deadlock
113	 * with the state_mutex and mm_lock.
114	 * In case the state_mutex was taken already we defer the cleanup work
115	 * to the unlock flow of the other running context.
116	 */
117	spin_lock(&xe_vdev->reset_lock);
118	xe_vdev->deferred_reset = true;
119	if (!mutex_trylock(&xe_vdev->state_mutex)) {
120		spin_unlock(&xe_vdev->reset_lock);
121		return;
122	}
123	spin_unlock(&xe_vdev->reset_lock);
124	xe_vfio_pci_state_mutex_unlock(xe_vdev);
125
126	xe_vfio_pci_reset(xe_vdev);
127}
128
129static const struct pci_error_handlers xe_vfio_pci_err_handlers = {
130	.reset_done = xe_vfio_pci_reset_done,
131	.error_detected = vfio_pci_core_aer_err_detected,
132};
133
134static int xe_vfio_pci_open_device(struct vfio_device *core_vdev)
135{
136	struct xe_vfio_pci_core_device *xe_vdev =
137		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
138	struct vfio_pci_core_device *vdev = &xe_vdev->core_device;
139	int ret;
140
141	ret = vfio_pci_core_enable(vdev);
142	if (ret)
143		return ret;
144
145	xe_vdev->mig_state = VFIO_DEVICE_STATE_RUNNING;
146
147	vfio_pci_core_finish_enable(vdev);
148
149	return 0;
150}
151
152static void xe_vfio_pci_close_device(struct vfio_device *core_vdev)
153{
154	struct xe_vfio_pci_core_device *xe_vdev =
155		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
156
157	xe_vfio_pci_state_mutex_lock(xe_vdev);
158	xe_vfio_pci_reset(xe_vdev);
159	xe_vfio_pci_state_mutex_unlock(xe_vdev);
160	vfio_pci_core_close_device(core_vdev);
161}
162
163static int xe_vfio_pci_release_file(struct inode *inode, struct file *filp)
164{
165	struct xe_vfio_pci_migration_file *migf = filp->private_data;
166
167	mutex_destroy(&migf->lock);
168	kfree(migf);
169
170	return 0;
171}
172
173static ssize_t xe_vfio_pci_save_read(struct file *filp, char __user *buf, size_t len, loff_t *pos)
174{
175	struct xe_vfio_pci_migration_file *migf = filp->private_data;
176	ssize_t ret;
177
178	if (pos)
179		return -ESPIPE;
180
181	mutex_lock(&migf->lock);
182	if (migf->disabled) {
183		mutex_unlock(&migf->lock);
184		return -ENODEV;
185	}
186
187	ret = xe_sriov_vfio_data_read(migf->xe_vdev->xe, migf->xe_vdev->vfid, buf, len);
188	mutex_unlock(&migf->lock);
189
190	return ret;
191}
192
193static const struct file_operations xe_vfio_pci_save_fops = {
194	.owner = THIS_MODULE,
195	.read = xe_vfio_pci_save_read,
196	.release = xe_vfio_pci_release_file,
197	.llseek = noop_llseek,
198};
199
200static ssize_t xe_vfio_pci_resume_write(struct file *filp, const char __user *buf,
201					size_t len, loff_t *pos)
202{
203	struct xe_vfio_pci_migration_file *migf = filp->private_data;
204	ssize_t ret;
205
206	if (pos)
207		return -ESPIPE;
208
209	mutex_lock(&migf->lock);
210	if (migf->disabled) {
211		mutex_unlock(&migf->lock);
212		return -ENODEV;
213	}
214
215	ret = xe_sriov_vfio_data_write(migf->xe_vdev->xe, migf->xe_vdev->vfid, buf, len);
216	mutex_unlock(&migf->lock);
217
218	return ret;
219}
220
221static const struct file_operations xe_vfio_pci_resume_fops = {
222	.owner = THIS_MODULE,
223	.write = xe_vfio_pci_resume_write,
224	.release = xe_vfio_pci_release_file,
225	.llseek = noop_llseek,
226};
227
228static const char *vfio_dev_state_str(u32 state)
229{
230	switch (state) {
231	case VFIO_DEVICE_STATE_RUNNING: return "running";
232	case VFIO_DEVICE_STATE_RUNNING_P2P: return "running_p2p";
233	case VFIO_DEVICE_STATE_STOP_COPY: return "stopcopy";
234	case VFIO_DEVICE_STATE_STOP: return "stop";
235	case VFIO_DEVICE_STATE_RESUMING: return "resuming";
236	case VFIO_DEVICE_STATE_ERROR: return "error";
237	default: return "";
238	}
239}
240
241enum xe_vfio_pci_file_type {
242	XE_VFIO_FILE_SAVE = 0,
243	XE_VFIO_FILE_RESUME,
244};
245
246static struct xe_vfio_pci_migration_file *
247xe_vfio_pci_alloc_file(struct xe_vfio_pci_core_device *xe_vdev,
248		       enum xe_vfio_pci_file_type type)
249{
250	struct xe_vfio_pci_migration_file *migf;
251	const struct file_operations *fops;
252	int flags;
253	int ret;
254
255	migf = kzalloc(sizeof(*migf), GFP_KERNEL_ACCOUNT);
256	if (!migf)
257		return ERR_PTR(-ENOMEM);
258
259	fops = type == XE_VFIO_FILE_SAVE ? &xe_vfio_pci_save_fops : &xe_vfio_pci_resume_fops;
260	flags = type == XE_VFIO_FILE_SAVE ? O_RDONLY : O_WRONLY;
261	migf->filp = anon_inode_getfile("xe_vfio_mig", fops, migf, flags);
262	if (IS_ERR(migf->filp)) {
263		ret = PTR_ERR(migf->filp);
264		kfree(migf);
265		return ERR_PTR(ret);
266	}
267
268	mutex_init(&migf->lock);
269	migf->xe_vdev = xe_vdev;
270	xe_vdev->migf = migf;
271
272	stream_open(migf->filp->f_inode, migf->filp);
273
274	return migf;
275}
276
277static struct file *
278xe_vfio_set_state(struct xe_vfio_pci_core_device *xe_vdev, u32 new)
279{
280	u32 cur = xe_vdev->mig_state;
281	int ret;
282
283	dev_dbg(xe_vdev_to_dev(xe_vdev),
284		"state: %s->%s\n", vfio_dev_state_str(cur), vfio_dev_state_str(new));
285
286	/*
287	 * "STOP" handling is reused for "RUNNING_P2P", as the device doesn't
288	 * have the capability to selectively block outgoing p2p DMA transfers.
289	 * While the device is allowing BAR accesses when the VF is stopped, it
290	 * is not processing any new workload requests, effectively stopping
291	 * any outgoing DMA transfers (not just p2p).
292	 * Any VRAM / MMIO accesses occurring during "RUNNING_P2P" are kept and
293	 * will be migrated to target VF during stop-copy.
294	 */
295	if (cur == VFIO_DEVICE_STATE_RUNNING && new == VFIO_DEVICE_STATE_RUNNING_P2P) {
296		ret = xe_sriov_vfio_suspend_device(xe_vdev->xe, xe_vdev->vfid);
297		if (ret)
298			goto err;
299
300		return NULL;
301	}
302
303	if ((cur == VFIO_DEVICE_STATE_RUNNING_P2P && new == VFIO_DEVICE_STATE_STOP) ||
304	    (cur == VFIO_DEVICE_STATE_STOP && new == VFIO_DEVICE_STATE_RUNNING_P2P))
305		return NULL;
306
307	if (cur == VFIO_DEVICE_STATE_RUNNING_P2P && new == VFIO_DEVICE_STATE_RUNNING) {
308		ret = xe_sriov_vfio_resume_device(xe_vdev->xe, xe_vdev->vfid);
309		if (ret)
310			goto err;
311
312		return NULL;
313	}
314
315	if (cur == VFIO_DEVICE_STATE_STOP && new == VFIO_DEVICE_STATE_STOP_COPY) {
316		struct xe_vfio_pci_migration_file *migf;
317
318		migf = xe_vfio_pci_alloc_file(xe_vdev, XE_VFIO_FILE_SAVE);
319		if (IS_ERR(migf)) {
320			ret = PTR_ERR(migf);
321			goto err;
322		}
323		get_file(migf->filp);
324
325		ret = xe_sriov_vfio_stop_copy_enter(xe_vdev->xe, xe_vdev->vfid);
326		if (ret) {
327			fput(migf->filp);
328			goto err;
329		}
330
331		return migf->filp;
332	}
333
334	if (cur == VFIO_DEVICE_STATE_STOP_COPY && new == VFIO_DEVICE_STATE_STOP) {
335		if (xe_vdev->migf)
336			xe_vfio_pci_put_file(xe_vdev);
337
338		ret = xe_sriov_vfio_stop_copy_exit(xe_vdev->xe, xe_vdev->vfid);
339		if (ret)
340			goto err;
341
342		return NULL;
343	}
344
345	if (cur == VFIO_DEVICE_STATE_STOP && new == VFIO_DEVICE_STATE_RESUMING) {
346		struct xe_vfio_pci_migration_file *migf;
347
348		migf = xe_vfio_pci_alloc_file(xe_vdev, XE_VFIO_FILE_RESUME);
349		if (IS_ERR(migf)) {
350			ret = PTR_ERR(migf);
351			goto err;
352		}
353		get_file(migf->filp);
354
355		ret = xe_sriov_vfio_resume_data_enter(xe_vdev->xe, xe_vdev->vfid);
356		if (ret) {
357			fput(migf->filp);
358			goto err;
359		}
360
361		return migf->filp;
362	}
363
364	if (cur == VFIO_DEVICE_STATE_RESUMING && new == VFIO_DEVICE_STATE_STOP) {
365		if (xe_vdev->migf)
366			xe_vfio_pci_put_file(xe_vdev);
367
368		ret = xe_sriov_vfio_resume_data_exit(xe_vdev->xe, xe_vdev->vfid);
369		if (ret)
370			goto err;
371
372		return NULL;
373	}
374
375	WARN(true, "Unknown state transition %d->%d", cur, new);
376	return ERR_PTR(-EINVAL);
377
378err:
379	dev_dbg(xe_vdev_to_dev(xe_vdev),
380		"Failed to transition state: %s->%s err=%d\n",
381		vfio_dev_state_str(cur), vfio_dev_state_str(new), ret);
382	return ERR_PTR(ret);
383}
384
385static struct file *
386xe_vfio_pci_set_device_state(struct vfio_device *core_vdev,
387			     enum vfio_device_mig_state new_state)
388{
389	struct xe_vfio_pci_core_device *xe_vdev =
390		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
391	enum vfio_device_mig_state next_state;
392	struct file *f = NULL;
393	int ret;
394
395	xe_vfio_pci_state_mutex_lock(xe_vdev);
396	while (new_state != xe_vdev->mig_state) {
397		ret = vfio_mig_get_next_state(core_vdev, xe_vdev->mig_state,
398					      new_state, &next_state);
399		if (ret) {
400			xe_sriov_vfio_error(xe_vdev->xe, xe_vdev->vfid);
401			f = ERR_PTR(ret);
402			break;
403		}
404		f = xe_vfio_set_state(xe_vdev, next_state);
405		if (IS_ERR(f))
406			break;
407
408		xe_vdev->mig_state = next_state;
409
410		/* Multiple state transitions with non-NULL file in the middle */
411		if (f && new_state != xe_vdev->mig_state) {
412			fput(f);
413			f = ERR_PTR(-EINVAL);
414			break;
415		}
416	}
417	xe_vfio_pci_state_mutex_unlock(xe_vdev);
418
419	return f;
420}
421
422static int xe_vfio_pci_get_device_state(struct vfio_device *core_vdev,
423					enum vfio_device_mig_state *curr_state)
424{
425	struct xe_vfio_pci_core_device *xe_vdev =
426		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
427
428	xe_vfio_pci_state_mutex_lock(xe_vdev);
429	*curr_state = xe_vdev->mig_state;
430	xe_vfio_pci_state_mutex_unlock(xe_vdev);
431
432	return 0;
433}
434
435static int xe_vfio_pci_get_data_size(struct vfio_device *vdev,
436				     unsigned long *stop_copy_length)
437{
438	struct xe_vfio_pci_core_device *xe_vdev =
439		container_of(vdev, struct xe_vfio_pci_core_device, core_device.vdev);
440
441	xe_vfio_pci_state_mutex_lock(xe_vdev);
442	*stop_copy_length = xe_sriov_vfio_stop_copy_size(xe_vdev->xe, xe_vdev->vfid);
443	xe_vfio_pci_state_mutex_unlock(xe_vdev);
444
445	return 0;
446}
447
448static const struct vfio_migration_ops xe_vfio_pci_migration_ops = {
449	.migration_set_state = xe_vfio_pci_set_device_state,
450	.migration_get_state = xe_vfio_pci_get_device_state,
451	.migration_get_data_size = xe_vfio_pci_get_data_size,
452};
453
454static void xe_vfio_pci_migration_init(struct xe_vfio_pci_core_device *xe_vdev)
455{
456	struct vfio_device *core_vdev = &xe_vdev->core_device.vdev;
457
458	if (!xe_sriov_vfio_migration_supported(xe_vdev->xe))
459		return;
460
461	core_vdev->migration_flags = VFIO_MIGRATION_STOP_COPY | VFIO_MIGRATION_P2P;
462	core_vdev->mig_ops = &xe_vfio_pci_migration_ops;
463}
464
465static int xe_vfio_pci_vf_init(struct xe_vfio_pci_core_device *xe_vdev)
466{
467	struct vfio_device *core_vdev = &xe_vdev->core_device.vdev;
468	struct pci_dev *pdev = to_pci_dev(core_vdev->dev);
469	struct xe_device *xe = xe_sriov_vfio_get_pf(pdev);
470
471	if (!pdev->is_virtfn)
472		return 0;
473	if (!xe)
474		return -ENODEV;
475	xe_vdev->xe = xe;
476
477	/* PF internal control uses vfid index starting from 1 */
478	xe_vdev->vfid = pci_iov_vf_id(pdev) + 1;
479
480	xe_vfio_pci_migration_init(xe_vdev);
481
482	return 0;
483}
484
485static int xe_vfio_pci_init_dev(struct vfio_device *core_vdev)
486{
487	struct xe_vfio_pci_core_device *xe_vdev =
488		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
489	int ret;
490
491	mutex_init(&xe_vdev->state_mutex);
492	spin_lock_init(&xe_vdev->reset_lock);
493
494	ret = xe_vfio_pci_vf_init(xe_vdev);
495	if (ret)
496		return ret;
497
498	return vfio_pci_core_init_dev(core_vdev);
499}
500
501static void xe_vfio_pci_release_dev(struct vfio_device *core_vdev)
502{
503	struct xe_vfio_pci_core_device *xe_vdev =
504		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
505
506	mutex_destroy(&xe_vdev->state_mutex);
507}
508
509static const struct vfio_device_ops xe_vfio_pci_ops = {
510	.name = "xe-vfio-pci",
511	.init = xe_vfio_pci_init_dev,
512	.release = xe_vfio_pci_release_dev,
513	.open_device = xe_vfio_pci_open_device,
514	.close_device = xe_vfio_pci_close_device,
515	.ioctl = vfio_pci_core_ioctl,
516	.get_region_info_caps = vfio_pci_ioctl_get_region_info,
517	.device_feature = vfio_pci_core_ioctl_feature,
518	.read = vfio_pci_core_read,
519	.write = vfio_pci_core_write,
520	.mmap = vfio_pci_core_mmap,
521	.request = vfio_pci_core_request,
522	.match = vfio_pci_core_match,
523	.match_token_uuid = vfio_pci_core_match_token_uuid,
524	.bind_iommufd = vfio_iommufd_physical_bind,
525	.unbind_iommufd = vfio_iommufd_physical_unbind,
526	.attach_ioas = vfio_iommufd_physical_attach_ioas,
527	.detach_ioas = vfio_iommufd_physical_detach_ioas,
528};
529
530static int xe_vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
531{
532	struct xe_vfio_pci_core_device *xe_vdev;
533	int ret;
534
535	xe_vdev = vfio_alloc_device(xe_vfio_pci_core_device, core_device.vdev, &pdev->dev,
536				    &xe_vfio_pci_ops);
537	if (IS_ERR(xe_vdev))
538		return PTR_ERR(xe_vdev);
539
540	dev_set_drvdata(&pdev->dev, &xe_vdev->core_device);
541
542	ret = vfio_pci_core_register_device(&xe_vdev->core_device);
543	if (ret) {
544		vfio_put_device(&xe_vdev->core_device.vdev);
545		return ret;
546	}
547
548	return 0;
549}
550
551static void xe_vfio_pci_remove(struct pci_dev *pdev)
552{
553	struct xe_vfio_pci_core_device *xe_vdev = pci_get_drvdata(pdev);
554
555	vfio_pci_core_unregister_device(&xe_vdev->core_device);
556	vfio_put_device(&xe_vdev->core_device.vdev);
557}
558
559#define INTEL_PCI_VFIO_DEVICE(_id) { \
560	PCI_DRIVER_OVERRIDE_DEVICE_VFIO(PCI_VENDOR_ID_INTEL, (_id)) \
561}
562
563static const struct pci_device_id xe_vfio_pci_table[] = {
564	INTEL_PTL_IDS(INTEL_PCI_VFIO_DEVICE),
565	INTEL_WCL_IDS(INTEL_PCI_VFIO_DEVICE),
566	INTEL_BMG_IDS(INTEL_PCI_VFIO_DEVICE),
567	{}
568};
569MODULE_DEVICE_TABLE(pci, xe_vfio_pci_table);
570
571static struct pci_driver xe_vfio_pci_driver = {
572	.name = "xe-vfio-pci",
573	.id_table = xe_vfio_pci_table,
574	.probe = xe_vfio_pci_probe,
575	.remove = xe_vfio_pci_remove,
576	.err_handler = &xe_vfio_pci_err_handlers,
577	.driver_managed_dma = true,
578};
579module_pci_driver(xe_vfio_pci_driver);
580
581MODULE_LICENSE("GPL");
582MODULE_AUTHOR("Michał Winiarski <michal.winiarski@intel.com>");
583MODULE_DESCRIPTION("VFIO PCI driver with migration support for Intel Graphics");