xref: /linux/drivers/accel/qaic/qaic_drv.c (revision 82f78acd5a9270370ef4aa3f032ede25f3dc91ee)
1 // SPDX-License-Identifier: GPL-2.0-only
2 
3 /* Copyright (c) 2019-2021, The Linux Foundation. All rights reserved. */
4 /* Copyright (c) 2021-2023 Qualcomm Innovation Center, Inc. All rights reserved. */
5 
6 #include <linux/delay.h>
7 #include <linux/dma-mapping.h>
8 #include <linux/idr.h>
9 #include <linux/interrupt.h>
10 #include <linux/list.h>
11 #include <linux/kobject.h>
12 #include <linux/kref.h>
13 #include <linux/mhi.h>
14 #include <linux/module.h>
15 #include <linux/msi.h>
16 #include <linux/mutex.h>
17 #include <linux/pci.h>
18 #include <linux/spinlock.h>
19 #include <linux/workqueue.h>
20 #include <linux/wait.h>
21 #include <drm/drm_accel.h>
22 #include <drm/drm_drv.h>
23 #include <drm/drm_file.h>
24 #include <drm/drm_gem.h>
25 #include <drm/drm_ioctl.h>
26 #include <drm/drm_managed.h>
27 #include <uapi/drm/qaic_accel.h>
28 
29 #include "mhi_controller.h"
30 #include "qaic.h"
31 #include "qaic_debugfs.h"
32 #include "qaic_ras.h"
33 #include "qaic_ssr.h"
34 #include "qaic_timesync.h"
35 #include "sahara.h"
36 
37 MODULE_IMPORT_NS("DMA_BUF");
38 
39 #define PCI_DEVICE_ID_QCOM_AIC080	0xa080
40 #define PCI_DEVICE_ID_QCOM_AIC100	0xa100
41 #define PCI_DEVICE_ID_QCOM_AIC200	0xa110
42 #define QAIC_NAME			"qaic"
43 #define QAIC_DESC			"Qualcomm Cloud AI Accelerators"
44 #define CNTL_MAJOR			5
45 #define CNTL_MINOR			0
46 
47 struct qaic_device_config {
48 	/* Indicates the AIC family the device belongs to */
49 	int family;
50 	/* A bitmask representing the available BARs */
51 	int bar_mask;
52 	/* An index value used to identify the MHI controller BAR */
53 	unsigned int mhi_bar_idx;
54 	/* An index value used to identify the DBCs BAR */
55 	unsigned int dbc_bar_idx;
56 };
57 
58 static const struct qaic_device_config aic080_config = {
59 	.family = FAMILY_AIC100,
60 	.bar_mask = BIT(0) | BIT(2) | BIT(4),
61 	.mhi_bar_idx = 0,
62 	.dbc_bar_idx = 2,
63 };
64 
65 static const struct qaic_device_config aic100_config = {
66 	.family = FAMILY_AIC100,
67 	.bar_mask = BIT(0) | BIT(2) | BIT(4),
68 	.mhi_bar_idx = 0,
69 	.dbc_bar_idx = 2,
70 };
71 
72 static const struct qaic_device_config aic200_config = {
73 	.family = FAMILY_AIC200,
74 	.bar_mask = BIT(0) | BIT(1) | BIT(2) | BIT(4),
75 	.mhi_bar_idx = 1,
76 	.dbc_bar_idx = 2,
77 };
78 
79 bool datapath_polling;
80 module_param(datapath_polling, bool, 0400);
81 MODULE_PARM_DESC(datapath_polling, "Operate the datapath in polling mode");
82 static bool link_up;
83 static DEFINE_IDA(qaic_usrs);
84 
qaicm_wq_release(struct drm_device * dev,void * res)85 static void qaicm_wq_release(struct drm_device *dev, void *res)
86 {
87 	struct workqueue_struct *wq = res;
88 
89 	destroy_workqueue(wq);
90 }
91 
qaicm_wq_init(struct drm_device * dev,const char * name)92 static struct workqueue_struct *qaicm_wq_init(struct drm_device *dev, const char *name)
93 {
94 	struct workqueue_struct *wq;
95 	int ret;
96 
97 	wq = alloc_workqueue("%s", WQ_UNBOUND, 0, name);
98 	if (!wq)
99 		return ERR_PTR(-ENOMEM);
100 	ret = drmm_add_action_or_reset(dev, qaicm_wq_release, wq);
101 	if (ret)
102 		return ERR_PTR(ret);
103 
104 	return wq;
105 }
106 
qaicm_srcu_release(struct drm_device * dev,void * res)107 static void qaicm_srcu_release(struct drm_device *dev, void *res)
108 {
109 	struct srcu_struct *lock = res;
110 
111 	cleanup_srcu_struct(lock);
112 }
113 
qaicm_srcu_init(struct drm_device * dev,struct srcu_struct * lock)114 static int qaicm_srcu_init(struct drm_device *dev, struct srcu_struct *lock)
115 {
116 	int ret;
117 
118 	ret = init_srcu_struct(lock);
119 	if (ret)
120 		return ret;
121 
122 	return drmm_add_action_or_reset(dev, qaicm_srcu_release, lock);
123 }
124 
qaicm_pci_release(struct drm_device * dev,void * res)125 static void qaicm_pci_release(struct drm_device *dev, void *res)
126 {
127 	struct qaic_device *qdev = to_qaic_device(dev);
128 
129 	pci_set_drvdata(qdev->pdev, NULL);
130 }
131 
free_usr(struct kref * kref)132 static void free_usr(struct kref *kref)
133 {
134 	struct qaic_user *usr = container_of(kref, struct qaic_user, ref_count);
135 
136 	cleanup_srcu_struct(&usr->qddev_lock);
137 	ida_free(&qaic_usrs, usr->handle);
138 	kfree(usr);
139 }
140 
qaic_open(struct drm_device * dev,struct drm_file * file)141 static int qaic_open(struct drm_device *dev, struct drm_file *file)
142 {
143 	struct qaic_drm_device *qddev = to_qaic_drm_device(dev);
144 	struct qaic_device *qdev = qddev->qdev;
145 	struct qaic_user *usr;
146 	int rcu_id;
147 	int ret;
148 
149 	rcu_id = srcu_read_lock(&qdev->dev_lock);
150 	if (qdev->dev_state != QAIC_ONLINE) {
151 		ret = -ENODEV;
152 		goto dev_unlock;
153 	}
154 
155 	usr = kmalloc(sizeof(*usr), GFP_KERNEL);
156 	if (!usr) {
157 		ret = -ENOMEM;
158 		goto dev_unlock;
159 	}
160 
161 	usr->handle = ida_alloc(&qaic_usrs, GFP_KERNEL);
162 	if (usr->handle < 0) {
163 		ret = usr->handle;
164 		goto free_usr;
165 	}
166 	usr->qddev = qddev;
167 	atomic_set(&usr->chunk_id, 0);
168 	init_srcu_struct(&usr->qddev_lock);
169 	kref_init(&usr->ref_count);
170 
171 	ret = mutex_lock_interruptible(&qddev->users_mutex);
172 	if (ret)
173 		goto cleanup_usr;
174 
175 	list_add(&usr->node, &qddev->users);
176 	mutex_unlock(&qddev->users_mutex);
177 
178 	file->driver_priv = usr;
179 
180 	srcu_read_unlock(&qdev->dev_lock, rcu_id);
181 	return 0;
182 
183 cleanup_usr:
184 	cleanup_srcu_struct(&usr->qddev_lock);
185 	ida_free(&qaic_usrs, usr->handle);
186 free_usr:
187 	kfree(usr);
188 dev_unlock:
189 	srcu_read_unlock(&qdev->dev_lock, rcu_id);
190 	return ret;
191 }
192 
qaic_postclose(struct drm_device * dev,struct drm_file * file)193 static void qaic_postclose(struct drm_device *dev, struct drm_file *file)
194 {
195 	struct qaic_user *usr = file->driver_priv;
196 	struct qaic_drm_device *qddev;
197 	struct qaic_device *qdev;
198 	int qdev_rcu_id;
199 	int usr_rcu_id;
200 	int i;
201 
202 	qddev = usr->qddev;
203 	usr_rcu_id = srcu_read_lock(&usr->qddev_lock);
204 	if (qddev) {
205 		qdev = qddev->qdev;
206 		qdev_rcu_id = srcu_read_lock(&qdev->dev_lock);
207 		if (qdev->dev_state == QAIC_ONLINE) {
208 			qaic_release_usr(qdev, usr);
209 			for (i = 0; i < qdev->num_dbc; ++i)
210 				if (qdev->dbc[i].usr && qdev->dbc[i].usr->handle == usr->handle)
211 					release_dbc(qdev, i);
212 		}
213 		srcu_read_unlock(&qdev->dev_lock, qdev_rcu_id);
214 
215 		mutex_lock(&qddev->users_mutex);
216 		if (!list_empty(&usr->node))
217 			list_del_init(&usr->node);
218 		mutex_unlock(&qddev->users_mutex);
219 	}
220 
221 	srcu_read_unlock(&usr->qddev_lock, usr_rcu_id);
222 	kref_put(&usr->ref_count, free_usr);
223 
224 	file->driver_priv = NULL;
225 }
226 
227 DEFINE_DRM_ACCEL_FOPS(qaic_accel_fops);
228 
229 static const struct drm_ioctl_desc qaic_drm_ioctls[] = {
230 	DRM_IOCTL_DEF_DRV(QAIC_MANAGE, qaic_manage_ioctl, 0),
231 	DRM_IOCTL_DEF_DRV(QAIC_CREATE_BO, qaic_create_bo_ioctl, 0),
232 	DRM_IOCTL_DEF_DRV(QAIC_MMAP_BO, qaic_mmap_bo_ioctl, 0),
233 	DRM_IOCTL_DEF_DRV(QAIC_ATTACH_SLICE_BO, qaic_attach_slice_bo_ioctl, 0),
234 	DRM_IOCTL_DEF_DRV(QAIC_EXECUTE_BO, qaic_execute_bo_ioctl, 0),
235 	DRM_IOCTL_DEF_DRV(QAIC_PARTIAL_EXECUTE_BO, qaic_partial_execute_bo_ioctl, 0),
236 	DRM_IOCTL_DEF_DRV(QAIC_WAIT_BO, qaic_wait_bo_ioctl, 0),
237 	DRM_IOCTL_DEF_DRV(QAIC_PERF_STATS_BO, qaic_perf_stats_bo_ioctl, 0),
238 	DRM_IOCTL_DEF_DRV(QAIC_DETACH_SLICE_BO, qaic_detach_slice_bo_ioctl, 0),
239 };
240 
241 static const struct drm_driver qaic_accel_driver = {
242 	.driver_features	= DRIVER_GEM | DRIVER_COMPUTE_ACCEL,
243 
244 	.name			= QAIC_NAME,
245 	.desc			= QAIC_DESC,
246 
247 	.fops			= &qaic_accel_fops,
248 	.open			= qaic_open,
249 	.postclose		= qaic_postclose,
250 
251 	.ioctls			= qaic_drm_ioctls,
252 	.num_ioctls		= ARRAY_SIZE(qaic_drm_ioctls),
253 	.gem_prime_import	= qaic_gem_prime_import,
254 };
255 
qaic_create_drm_device(struct qaic_device * qdev,s32 partition_id)256 static int qaic_create_drm_device(struct qaic_device *qdev, s32 partition_id)
257 {
258 	struct qaic_drm_device *qddev = qdev->qddev;
259 	struct drm_device *drm = to_drm(qddev);
260 	int ret;
261 
262 	/* Hold off implementing partitions until the uapi is determined */
263 	if (partition_id != QAIC_NO_PARTITION)
264 		return -EINVAL;
265 
266 	qddev->partition_id = partition_id;
267 
268 	ret = drm_dev_register(drm, 0);
269 	if (ret) {
270 		pci_dbg(qdev->pdev, "drm_dev_register failed %d\n", ret);
271 		return ret;
272 	}
273 
274 	ret = qaic_sysfs_init(qddev);
275 	if (ret) {
276 		drm_dev_unregister(drm);
277 		pci_dbg(qdev->pdev, "qaic_sysfs_init failed %d\n", ret);
278 		return ret;
279 	}
280 
281 	qaic_debugfs_init(qddev);
282 
283 	return ret;
284 }
285 
qaic_destroy_drm_device(struct qaic_device * qdev,s32 partition_id)286 static void qaic_destroy_drm_device(struct qaic_device *qdev, s32 partition_id)
287 {
288 	struct qaic_drm_device *qddev = qdev->qddev;
289 	struct drm_device *drm = to_drm(qddev);
290 	struct qaic_user *usr;
291 
292 	qaic_sysfs_remove(qddev);
293 	drm_dev_unregister(drm);
294 	qddev->partition_id = 0;
295 	/*
296 	 * Existing users get unresolvable errors till they close FDs.
297 	 * Need to sync carefully with users calling close(). The
298 	 * list of users can be modified elsewhere when the lock isn't
299 	 * held here, but the sync'ing the srcu with the mutex held
300 	 * could deadlock. Grab the mutex so that the list will be
301 	 * unmodified. The user we get will exist as long as the
302 	 * lock is held. Signal that the qcdev is going away, and
303 	 * grab a reference to the user so they don't go away for
304 	 * synchronize_srcu(). Then release the mutex to avoid
305 	 * deadlock and make sure the user has observed the signal.
306 	 * With the lock released, we cannot maintain any state of the
307 	 * user list.
308 	 */
309 	mutex_lock(&qddev->users_mutex);
310 	while (!list_empty(&qddev->users)) {
311 		usr = list_first_entry(&qddev->users, struct qaic_user, node);
312 		list_del_init(&usr->node);
313 		kref_get(&usr->ref_count);
314 		usr->qddev = NULL;
315 		mutex_unlock(&qddev->users_mutex);
316 		synchronize_srcu(&usr->qddev_lock);
317 		kref_put(&usr->ref_count, free_usr);
318 		mutex_lock(&qddev->users_mutex);
319 	}
320 	mutex_unlock(&qddev->users_mutex);
321 }
322 
qaic_mhi_probe(struct mhi_device * mhi_dev,const struct mhi_device_id * id)323 static int qaic_mhi_probe(struct mhi_device *mhi_dev, const struct mhi_device_id *id)
324 {
325 	u16 major = -1, minor = -1;
326 	struct qaic_device *qdev;
327 	int ret;
328 
329 	/*
330 	 * Invoking this function indicates that the control channel to the
331 	 * device is available. We use that as a signal to indicate that
332 	 * the device side firmware has booted. The device side firmware
333 	 * manages the device resources, so we need to communicate with it
334 	 * via the control channel in order to utilize the device. Therefore
335 	 * we wait until this signal to create the drm dev that userspace will
336 	 * use to control the device, because without the device side firmware,
337 	 * userspace can't do anything useful.
338 	 */
339 
340 	qdev = pci_get_drvdata(to_pci_dev(mhi_dev->mhi_cntrl->cntrl_dev));
341 
342 	dev_set_drvdata(&mhi_dev->dev, qdev);
343 	qdev->cntl_ch = mhi_dev;
344 
345 	ret = qaic_control_open(qdev);
346 	if (ret) {
347 		pci_dbg(qdev->pdev, "%s: control_open failed %d\n", __func__, ret);
348 		return ret;
349 	}
350 
351 	qdev->dev_state = QAIC_BOOT;
352 	ret = get_cntl_version(qdev, NULL, &major, &minor);
353 	if (ret || major != CNTL_MAJOR || minor > CNTL_MINOR) {
354 		pci_err(qdev->pdev, "%s: Control protocol version (%d.%d) not supported. Supported version is (%d.%d). Ret: %d\n",
355 			__func__, major, minor, CNTL_MAJOR, CNTL_MINOR, ret);
356 		ret = -EINVAL;
357 		goto close_control;
358 	}
359 	qdev->dev_state = QAIC_ONLINE;
360 	kobject_uevent(&(to_accel_kdev(qdev->qddev))->kobj, KOBJ_ONLINE);
361 
362 	return ret;
363 
364 close_control:
365 	qaic_control_close(qdev);
366 	return ret;
367 }
368 
qaic_mhi_remove(struct mhi_device * mhi_dev)369 static void qaic_mhi_remove(struct mhi_device *mhi_dev)
370 {
371 /* This is redundant since we have already observed the device crash */
372 }
373 
qaic_notify_reset(struct qaic_device * qdev)374 static void qaic_notify_reset(struct qaic_device *qdev)
375 {
376 	int i;
377 
378 	kobject_uevent(&(to_accel_kdev(qdev->qddev))->kobj, KOBJ_OFFLINE);
379 	qdev->dev_state = QAIC_OFFLINE;
380 	/* wake up any waiters to avoid waiting for timeouts at sync */
381 	wake_all_cntl(qdev);
382 	for (i = 0; i < qdev->num_dbc; ++i)
383 		wakeup_dbc(qdev, i);
384 	synchronize_srcu(&qdev->dev_lock);
385 }
386 
qaic_dev_reset_clean_local_state(struct qaic_device * qdev)387 void qaic_dev_reset_clean_local_state(struct qaic_device *qdev)
388 {
389 	int i;
390 
391 	qaic_notify_reset(qdev);
392 
393 	/* start tearing things down */
394 	qaic_clean_up_ssr(qdev);
395 	for (i = 0; i < qdev->num_dbc; ++i)
396 		release_dbc(qdev, i);
397 }
398 
create_qdev(struct pci_dev * pdev,const struct qaic_device_config * config)399 static struct qaic_device *create_qdev(struct pci_dev *pdev,
400 				       const struct qaic_device_config *config)
401 {
402 	struct device *dev = &pdev->dev;
403 	struct qaic_drm_device *qddev;
404 	struct qaic_device *qdev;
405 	struct drm_device *drm;
406 	int i, ret;
407 
408 	qdev = devm_kzalloc(dev, sizeof(*qdev), GFP_KERNEL);
409 	if (!qdev)
410 		return NULL;
411 
412 	qdev->dev_state = QAIC_OFFLINE;
413 	qdev->num_dbc = 16;
414 	qdev->dbc = devm_kcalloc(dev, qdev->num_dbc, sizeof(*qdev->dbc), GFP_KERNEL);
415 	if (!qdev->dbc)
416 		return NULL;
417 
418 	qddev = devm_drm_dev_alloc(&pdev->dev, &qaic_accel_driver, struct qaic_drm_device, drm);
419 	if (IS_ERR(qddev))
420 		return NULL;
421 
422 	drm = to_drm(qddev);
423 	pci_set_drvdata(pdev, qdev);
424 
425 	ret = drmm_mutex_init(drm, &qddev->users_mutex);
426 	if (ret)
427 		return NULL;
428 	ret = drmm_add_action_or_reset(drm, qaicm_pci_release, NULL);
429 	if (ret)
430 		return NULL;
431 	ret = drmm_mutex_init(drm, &qdev->cntl_mutex);
432 	if (ret)
433 		return NULL;
434 	ret = drmm_mutex_init(drm, &qdev->bootlog_mutex);
435 	if (ret)
436 		return NULL;
437 
438 	qdev->cntl_wq = qaicm_wq_init(drm, "qaic_cntl");
439 	if (IS_ERR(qdev->cntl_wq))
440 		return NULL;
441 	qdev->qts_wq = qaicm_wq_init(drm, "qaic_ts");
442 	if (IS_ERR(qdev->qts_wq))
443 		return NULL;
444 	qdev->ssr_wq = qaicm_wq_init(drm, "qaic_ssr");
445 	if (IS_ERR(qdev->ssr_wq))
446 		return NULL;
447 
448 	ret = qaicm_srcu_init(drm, &qdev->dev_lock);
449 	if (ret)
450 		return NULL;
451 
452 	ret = qaic_ssr_init(qdev, drm);
453 	if (ret)
454 		pci_info(pdev, "QAIC SSR crashdump collection not supported.\n");
455 
456 	qdev->qddev = qddev;
457 	qdev->pdev = pdev;
458 	qddev->qdev = qdev;
459 
460 	INIT_LIST_HEAD(&qdev->cntl_xfer_list);
461 	INIT_LIST_HEAD(&qdev->bootlog);
462 	INIT_LIST_HEAD(&qddev->users);
463 
464 	for (i = 0; i < qdev->num_dbc; ++i) {
465 		spin_lock_init(&qdev->dbc[i].xfer_lock);
466 		qdev->dbc[i].qdev = qdev;
467 		qdev->dbc[i].id = i;
468 		INIT_LIST_HEAD(&qdev->dbc[i].xfer_list);
469 		ret = qaicm_srcu_init(drm, &qdev->dbc[i].ch_lock);
470 		if (ret)
471 			return NULL;
472 		init_waitqueue_head(&qdev->dbc[i].dbc_release);
473 		INIT_LIST_HEAD(&qdev->dbc[i].bo_lists);
474 		ret = drmm_mutex_init(drm, &qdev->dbc[i].req_lock);
475 		if (ret)
476 			return NULL;
477 	}
478 
479 	return qdev;
480 }
481 
init_pci(struct qaic_device * qdev,struct pci_dev * pdev,const struct qaic_device_config * config)482 static int init_pci(struct qaic_device *qdev, struct pci_dev *pdev,
483 		    const struct qaic_device_config *config)
484 {
485 	int bars;
486 	int ret;
487 
488 	bars = pci_select_bars(pdev, IORESOURCE_MEM) & 0x3f;
489 
490 	/* make sure the device has the expected BARs */
491 	if (bars != config->bar_mask) {
492 		pci_dbg(pdev, "%s: expected BARs %#x not found in device. Found %#x\n",
493 			__func__, config->bar_mask, bars);
494 		return -EINVAL;
495 	}
496 
497 	ret = pcim_enable_device(pdev);
498 	if (ret)
499 		return ret;
500 
501 	ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
502 	if (ret)
503 		return ret;
504 	dma_set_max_seg_size(&pdev->dev, UINT_MAX);
505 
506 	qdev->bar_mhi = devm_ioremap_resource(&pdev->dev, &pdev->resource[config->mhi_bar_idx]);
507 	if (IS_ERR(qdev->bar_mhi))
508 		return PTR_ERR(qdev->bar_mhi);
509 
510 	qdev->bar_dbc = devm_ioremap_resource(&pdev->dev, &pdev->resource[config->dbc_bar_idx]);
511 	if (IS_ERR(qdev->bar_dbc))
512 		return PTR_ERR(qdev->bar_dbc);
513 
514 	/* Managed release since we use pcim_enable_device above */
515 	pci_set_master(pdev);
516 
517 	return 0;
518 }
519 
init_msi(struct qaic_device * qdev,struct pci_dev * pdev)520 static int init_msi(struct qaic_device *qdev, struct pci_dev *pdev)
521 {
522 	int irq_count = qdev->num_dbc + 1;
523 	int mhi_irq;
524 	int ret;
525 	int i;
526 
527 	/* Managed release since we use pcim_enable_device */
528 	ret = pci_alloc_irq_vectors(pdev, irq_count, irq_count, PCI_IRQ_MSI | PCI_IRQ_MSIX);
529 	if (ret == -ENOSPC) {
530 		ret = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI | PCI_IRQ_MSIX);
531 		if (ret < 0)
532 			return ret;
533 
534 		/*
535 		 * Operate in one MSI mode. All interrupts will be directed to
536 		 * MSI0; every interrupt will wake up all the interrupt handlers
537 		 * (MHI and DBC[0-15]). Since the interrupt is now shared, it is
538 		 * not disabled during DBC threaded handler, but only one thread
539 		 * will be allowed to run per DBC, so while it can be
540 		 * interrupted, it shouldn't race with itself.
541 		 */
542 		qdev->single_msi = true;
543 		pci_info(pdev, "Allocating %d MSIs failed, operating in 1 MSI mode. Performance may be impacted.\n",
544 			 irq_count);
545 	} else if (ret < 0) {
546 		return ret;
547 	}
548 
549 	mhi_irq = pci_irq_vector(pdev, 0);
550 	if (mhi_irq < 0)
551 		return mhi_irq;
552 
553 	for (i = 0; i < qdev->num_dbc; ++i) {
554 		ret = devm_request_threaded_irq(&pdev->dev,
555 						pci_irq_vector(pdev, qdev->single_msi ? 0 : i + 1),
556 						dbc_irq_handler, dbc_irq_threaded_fn, IRQF_SHARED,
557 						"qaic_dbc", &qdev->dbc[i]);
558 		if (ret)
559 			return ret;
560 
561 		if (datapath_polling) {
562 			qdev->dbc[i].irq = pci_irq_vector(pdev, qdev->single_msi ? 0 : i + 1);
563 			if (!qdev->single_msi)
564 				disable_irq_nosync(qdev->dbc[i].irq);
565 			INIT_WORK(&qdev->dbc[i].poll_work, qaic_irq_polling_work);
566 		}
567 	}
568 
569 	return mhi_irq;
570 }
571 
qaic_pci_probe(struct pci_dev * pdev,const struct pci_device_id * id)572 static int qaic_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
573 {
574 	struct qaic_device_config *config = (struct qaic_device_config *)id->driver_data;
575 	struct qaic_device *qdev;
576 	int mhi_irq;
577 	int ret;
578 	int i;
579 
580 	qdev = create_qdev(pdev, config);
581 	if (!qdev)
582 		return -ENOMEM;
583 
584 	ret = init_pci(qdev, pdev, config);
585 	if (ret)
586 		return ret;
587 
588 	for (i = 0; i < qdev->num_dbc; ++i)
589 		qdev->dbc[i].dbc_base = qdev->bar_dbc + QAIC_DBC_OFF(i);
590 
591 	mhi_irq = init_msi(qdev, pdev);
592 	if (mhi_irq < 0)
593 		return mhi_irq;
594 
595 	ret = qaic_create_drm_device(qdev, QAIC_NO_PARTITION);
596 	if (ret)
597 		return ret;
598 
599 	qdev->mhi_cntrl = qaic_mhi_register_controller(pdev, qdev->bar_mhi, mhi_irq,
600 						       qdev->single_msi, config->family);
601 	if (IS_ERR(qdev->mhi_cntrl)) {
602 		ret = PTR_ERR(qdev->mhi_cntrl);
603 		qaic_destroy_drm_device(qdev, QAIC_NO_PARTITION);
604 		return ret;
605 	}
606 
607 	return 0;
608 }
609 
qaic_pci_remove(struct pci_dev * pdev)610 static void qaic_pci_remove(struct pci_dev *pdev)
611 {
612 	struct qaic_device *qdev = pci_get_drvdata(pdev);
613 
614 	if (!qdev)
615 		return;
616 
617 	qaic_dev_reset_clean_local_state(qdev);
618 	qaic_mhi_free_controller(qdev->mhi_cntrl, link_up);
619 	qaic_destroy_drm_device(qdev, QAIC_NO_PARTITION);
620 }
621 
qaic_pci_shutdown(struct pci_dev * pdev)622 static void qaic_pci_shutdown(struct pci_dev *pdev)
623 {
624 	/* see qaic_exit for what link_up is doing */
625 	link_up = true;
626 	qaic_pci_remove(pdev);
627 }
628 
qaic_pci_error_detected(struct pci_dev * pdev,pci_channel_state_t error)629 static pci_ers_result_t qaic_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t error)
630 {
631 	return PCI_ERS_RESULT_NEED_RESET;
632 }
633 
qaic_pci_reset_prepare(struct pci_dev * pdev)634 static void qaic_pci_reset_prepare(struct pci_dev *pdev)
635 {
636 	struct qaic_device *qdev = pci_get_drvdata(pdev);
637 
638 	qaic_notify_reset(qdev);
639 	qaic_mhi_start_reset(qdev->mhi_cntrl);
640 	qaic_dev_reset_clean_local_state(qdev);
641 }
642 
qaic_pci_reset_done(struct pci_dev * pdev)643 static void qaic_pci_reset_done(struct pci_dev *pdev)
644 {
645 	struct qaic_device *qdev = pci_get_drvdata(pdev);
646 
647 	qaic_mhi_reset_done(qdev->mhi_cntrl);
648 }
649 
650 static const struct mhi_device_id qaic_mhi_match_table[] = {
651 	{ .chan = "QAIC_CONTROL", },
652 	{},
653 };
654 
655 static struct mhi_driver qaic_mhi_driver = {
656 	.id_table = qaic_mhi_match_table,
657 	.remove = qaic_mhi_remove,
658 	.probe = qaic_mhi_probe,
659 	.ul_xfer_cb = qaic_mhi_ul_xfer_cb,
660 	.dl_xfer_cb = qaic_mhi_dl_xfer_cb,
661 	.driver = {
662 		.name = "qaic_mhi",
663 	},
664 };
665 
666 static const struct pci_device_id qaic_ids[] = {
667 	{ PCI_DEVICE_DATA(QCOM, AIC080, (kernel_ulong_t)&aic080_config), },
668 	{ PCI_DEVICE_DATA(QCOM, AIC100, (kernel_ulong_t)&aic100_config), },
669 	{ PCI_DEVICE_DATA(QCOM, AIC200, (kernel_ulong_t)&aic200_config), },
670 	{ }
671 };
672 MODULE_DEVICE_TABLE(pci, qaic_ids);
673 
674 static const struct pci_error_handlers qaic_pci_err_handler = {
675 	.error_detected = qaic_pci_error_detected,
676 	.reset_prepare = qaic_pci_reset_prepare,
677 	.reset_done = qaic_pci_reset_done,
678 };
679 
qaic_is_under_reset(struct qaic_device * qdev)680 static bool qaic_is_under_reset(struct qaic_device *qdev)
681 {
682 	int rcu_id;
683 	bool ret;
684 
685 	rcu_id = srcu_read_lock(&qdev->dev_lock);
686 	ret = qdev->dev_state != QAIC_ONLINE;
687 	srcu_read_unlock(&qdev->dev_lock, rcu_id);
688 	return ret;
689 }
690 
qaic_data_path_busy(struct qaic_device * qdev)691 static bool qaic_data_path_busy(struct qaic_device *qdev)
692 {
693 	bool ret = false;
694 	int dev_rcu_id;
695 	int i;
696 
697 	dev_rcu_id = srcu_read_lock(&qdev->dev_lock);
698 	if (qdev->dev_state != QAIC_ONLINE) {
699 		srcu_read_unlock(&qdev->dev_lock, dev_rcu_id);
700 		return false;
701 	}
702 	for (i = 0; i < qdev->num_dbc; i++) {
703 		struct dma_bridge_chan *dbc = &qdev->dbc[i];
704 		unsigned long flags;
705 		int ch_rcu_id;
706 
707 		ch_rcu_id = srcu_read_lock(&dbc->ch_lock);
708 		if (!dbc->usr || !dbc->in_use) {
709 			srcu_read_unlock(&dbc->ch_lock, ch_rcu_id);
710 			continue;
711 		}
712 		spin_lock_irqsave(&dbc->xfer_lock, flags);
713 		ret = !list_empty(&dbc->xfer_list);
714 		spin_unlock_irqrestore(&dbc->xfer_lock, flags);
715 		srcu_read_unlock(&dbc->ch_lock, ch_rcu_id);
716 		if (ret)
717 			break;
718 	}
719 	srcu_read_unlock(&qdev->dev_lock, dev_rcu_id);
720 	return ret;
721 }
722 
qaic_pm_suspend(struct device * dev)723 static int qaic_pm_suspend(struct device *dev)
724 {
725 	struct qaic_device *qdev = pci_get_drvdata(to_pci_dev(dev));
726 
727 	dev_dbg(dev, "Suspending..\n");
728 	if (qaic_data_path_busy(qdev)) {
729 		dev_dbg(dev, "Device's datapath is busy. Aborting suspend..\n");
730 		return -EBUSY;
731 	}
732 	if (qaic_is_under_reset(qdev)) {
733 		dev_dbg(dev, "Device is under reset. Aborting suspend..\n");
734 		return -EBUSY;
735 	}
736 	qaic_mqts_ch_stop_timer(qdev->mqts_ch);
737 	qaic_pci_reset_prepare(qdev->pdev);
738 	pci_save_state(qdev->pdev);
739 	pci_disable_device(qdev->pdev);
740 	pci_set_power_state(qdev->pdev, PCI_D3hot);
741 	return 0;
742 }
743 
qaic_pm_resume(struct device * dev)744 static int qaic_pm_resume(struct device *dev)
745 {
746 	struct qaic_device *qdev = pci_get_drvdata(to_pci_dev(dev));
747 	int ret;
748 
749 	dev_dbg(dev, "Resuming..\n");
750 	pci_set_power_state(qdev->pdev, PCI_D0);
751 	pci_restore_state(qdev->pdev);
752 	ret = pci_enable_device(qdev->pdev);
753 	if (ret) {
754 		dev_err(dev, "pci_enable_device failed on resume %d\n", ret);
755 		return ret;
756 	}
757 	pci_set_master(qdev->pdev);
758 	qaic_pci_reset_done(qdev->pdev);
759 	return 0;
760 }
761 
762 static const struct dev_pm_ops qaic_pm_ops = {
763 	SYSTEM_SLEEP_PM_OPS(qaic_pm_suspend, qaic_pm_resume)
764 };
765 
766 static struct pci_driver qaic_pci_driver = {
767 	.name = QAIC_NAME,
768 	.id_table = qaic_ids,
769 	.probe = qaic_pci_probe,
770 	.remove = qaic_pci_remove,
771 	.shutdown = qaic_pci_shutdown,
772 	.err_handler = &qaic_pci_err_handler,
773 	.driver = {
774 		.pm = pm_sleep_ptr(&qaic_pm_ops),
775 	},
776 };
777 
qaic_init(void)778 static int __init qaic_init(void)
779 {
780 	int ret;
781 
782 	ret = pci_register_driver(&qaic_pci_driver);
783 	if (ret) {
784 		pr_debug("qaic: pci_register_driver failed %d\n", ret);
785 		return ret;
786 	}
787 
788 	ret = mhi_driver_register(&qaic_mhi_driver);
789 	if (ret) {
790 		pr_debug("qaic: mhi_driver_register failed %d\n", ret);
791 		goto free_pci;
792 	}
793 
794 	ret = sahara_register();
795 	if (ret) {
796 		pr_debug("qaic: sahara_register failed %d\n", ret);
797 		goto free_mhi;
798 	}
799 
800 	ret = qaic_timesync_init();
801 	if (ret)
802 		pr_debug("qaic: qaic_timesync_init failed %d\n", ret);
803 
804 	ret = qaic_bootlog_register();
805 	if (ret)
806 		pr_debug("qaic: qaic_bootlog_register failed %d\n", ret);
807 
808 	ret = qaic_ras_register();
809 	if (ret)
810 		pr_debug("qaic: qaic_ras_register failed %d\n", ret);
811 	ret = qaic_ssr_register();
812 	if (ret) {
813 		pr_debug("qaic: qaic_ssr_register failed %d\n", ret);
814 		goto free_bootlog;
815 	}
816 
817 	return 0;
818 
819 free_bootlog:
820 	qaic_bootlog_unregister();
821 free_mhi:
822 	mhi_driver_unregister(&qaic_mhi_driver);
823 free_pci:
824 	pci_unregister_driver(&qaic_pci_driver);
825 	return ret;
826 }
827 
qaic_exit(void)828 static void __exit qaic_exit(void)
829 {
830 	/*
831 	 * We assume that qaic_pci_remove() is called due to a hotplug event
832 	 * which would mean that the link is down, and thus
833 	 * qaic_mhi_free_controller() should not try to access the device during
834 	 * cleanup.
835 	 * We call pci_unregister_driver() below, which also triggers
836 	 * qaic_pci_remove(), but since this is module exit, we expect the link
837 	 * to the device to be up, in which case qaic_mhi_free_controller()
838 	 * should try to access the device during cleanup to put the device in
839 	 * a sane state.
840 	 * For that reason, we set link_up here to let qaic_mhi_free_controller
841 	 * know the expected link state. Since the module is going to be
842 	 * removed at the end of this, we don't need to worry about
843 	 * reinitializing the link_up state after the cleanup is done.
844 	 */
845 	link_up = true;
846 	qaic_ssr_unregister();
847 	qaic_ras_unregister();
848 	qaic_bootlog_unregister();
849 	qaic_timesync_deinit();
850 	sahara_unregister();
851 	mhi_driver_unregister(&qaic_mhi_driver);
852 	pci_unregister_driver(&qaic_pci_driver);
853 }
854 
855 module_init(qaic_init);
856 module_exit(qaic_exit);
857 
858 MODULE_AUTHOR(QAIC_DESC " Kernel Driver Team");
859 MODULE_DESCRIPTION(QAIC_DESC " Accel Driver");
860 MODULE_LICENSE("GPL");
861