xref: /openbmc/linux/drivers/thunderbolt/nhi.c (revision bffbf6e2)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Thunderbolt driver - NHI driver
4  *
5  * The NHI (native host interface) is the pci device that allows us to send and
6  * receive frames from the thunderbolt bus.
7  *
8  * Copyright (c) 2014 Andreas Noever <andreas.noever@gmail.com>
9  * Copyright (C) 2018, Intel Corporation
10  */
11 
12 #include <linux/pm_runtime.h>
13 #include <linux/slab.h>
14 #include <linux/errno.h>
15 #include <linux/pci.h>
16 #include <linux/interrupt.h>
17 #include <linux/module.h>
18 #include <linux/delay.h>
19 #include <linux/property.h>
20 #include <linux/platform_data/x86/apple.h>
21 
22 #include "nhi.h"
23 #include "nhi_regs.h"
24 #include "tb.h"
25 
26 #define RING_TYPE(ring) ((ring)->is_tx ? "TX ring" : "RX ring")
27 
28 #define RING_FIRST_USABLE_HOPID	1
29 
30 /*
31  * Minimal number of vectors when we use MSI-X. Two for control channel
32  * Rx/Tx and the rest four are for cross domain DMA paths.
33  */
34 #define MSIX_MIN_VECS		6
35 #define MSIX_MAX_VECS		16
36 
37 #define NHI_MAILBOX_TIMEOUT	500 /* ms */
38 
39 static int ring_interrupt_index(struct tb_ring *ring)
40 {
41 	int bit = ring->hop;
42 	if (!ring->is_tx)
43 		bit += ring->nhi->hop_count;
44 	return bit;
45 }
46 
47 /*
48  * ring_interrupt_active() - activate/deactivate interrupts for a single ring
49  *
50  * ring->nhi->lock must be held.
51  */
52 static void ring_interrupt_active(struct tb_ring *ring, bool active)
53 {
54 	int reg = REG_RING_INTERRUPT_BASE +
55 		  ring_interrupt_index(ring) / 32 * 4;
56 	int bit = ring_interrupt_index(ring) & 31;
57 	int mask = 1 << bit;
58 	u32 old, new;
59 
60 	if (ring->irq > 0) {
61 		u32 step, shift, ivr, misc;
62 		void __iomem *ivr_base;
63 		int index;
64 
65 		if (ring->is_tx)
66 			index = ring->hop;
67 		else
68 			index = ring->hop + ring->nhi->hop_count;
69 
70 		/*
71 		 * Ask the hardware to clear interrupt status bits automatically
72 		 * since we already know which interrupt was triggered.
73 		 */
74 		misc = ioread32(ring->nhi->iobase + REG_DMA_MISC);
75 		if (!(misc & REG_DMA_MISC_INT_AUTO_CLEAR)) {
76 			misc |= REG_DMA_MISC_INT_AUTO_CLEAR;
77 			iowrite32(misc, ring->nhi->iobase + REG_DMA_MISC);
78 		}
79 
80 		ivr_base = ring->nhi->iobase + REG_INT_VEC_ALLOC_BASE;
81 		step = index / REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
82 		shift = index % REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
83 		ivr = ioread32(ivr_base + step);
84 		ivr &= ~(REG_INT_VEC_ALLOC_MASK << shift);
85 		if (active)
86 			ivr |= ring->vector << shift;
87 		iowrite32(ivr, ivr_base + step);
88 	}
89 
90 	old = ioread32(ring->nhi->iobase + reg);
91 	if (active)
92 		new = old | mask;
93 	else
94 		new = old & ~mask;
95 
96 	dev_dbg(&ring->nhi->pdev->dev,
97 		"%s interrupt at register %#x bit %d (%#x -> %#x)\n",
98 		active ? "enabling" : "disabling", reg, bit, old, new);
99 
100 	if (new == old)
101 		dev_WARN(&ring->nhi->pdev->dev,
102 					 "interrupt for %s %d is already %s\n",
103 					 RING_TYPE(ring), ring->hop,
104 					 active ? "enabled" : "disabled");
105 	iowrite32(new, ring->nhi->iobase + reg);
106 }
107 
108 /*
109  * nhi_disable_interrupts() - disable interrupts for all rings
110  *
111  * Use only during init and shutdown.
112  */
113 static void nhi_disable_interrupts(struct tb_nhi *nhi)
114 {
115 	int i = 0;
116 	/* disable interrupts */
117 	for (i = 0; i < RING_INTERRUPT_REG_COUNT(nhi); i++)
118 		iowrite32(0, nhi->iobase + REG_RING_INTERRUPT_BASE + 4 * i);
119 
120 	/* clear interrupt status bits */
121 	for (i = 0; i < RING_NOTIFY_REG_COUNT(nhi); i++)
122 		ioread32(nhi->iobase + REG_RING_NOTIFY_BASE + 4 * i);
123 }
124 
125 /* ring helper methods */
126 
127 static void __iomem *ring_desc_base(struct tb_ring *ring)
128 {
129 	void __iomem *io = ring->nhi->iobase;
130 	io += ring->is_tx ? REG_TX_RING_BASE : REG_RX_RING_BASE;
131 	io += ring->hop * 16;
132 	return io;
133 }
134 
135 static void __iomem *ring_options_base(struct tb_ring *ring)
136 {
137 	void __iomem *io = ring->nhi->iobase;
138 	io += ring->is_tx ? REG_TX_OPTIONS_BASE : REG_RX_OPTIONS_BASE;
139 	io += ring->hop * 32;
140 	return io;
141 }
142 
143 static void ring_iowrite_cons(struct tb_ring *ring, u16 cons)
144 {
145 	/*
146 	 * The other 16-bits in the register is read-only and writes to it
147 	 * are ignored by the hardware so we can save one ioread32() by
148 	 * filling the read-only bits with zeroes.
149 	 */
150 	iowrite32(cons, ring_desc_base(ring) + 8);
151 }
152 
153 static void ring_iowrite_prod(struct tb_ring *ring, u16 prod)
154 {
155 	/* See ring_iowrite_cons() above for explanation */
156 	iowrite32(prod << 16, ring_desc_base(ring) + 8);
157 }
158 
159 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)
160 {
161 	iowrite32(value, ring_desc_base(ring) + offset);
162 }
163 
164 static void ring_iowrite64desc(struct tb_ring *ring, u64 value, u32 offset)
165 {
166 	iowrite32(value, ring_desc_base(ring) + offset);
167 	iowrite32(value >> 32, ring_desc_base(ring) + offset + 4);
168 }
169 
170 static void ring_iowrite32options(struct tb_ring *ring, u32 value, u32 offset)
171 {
172 	iowrite32(value, ring_options_base(ring) + offset);
173 }
174 
175 static bool ring_full(struct tb_ring *ring)
176 {
177 	return ((ring->head + 1) % ring->size) == ring->tail;
178 }
179 
180 static bool ring_empty(struct tb_ring *ring)
181 {
182 	return ring->head == ring->tail;
183 }
184 
185 /*
186  * ring_write_descriptors() - post frames from ring->queue to the controller
187  *
188  * ring->lock is held.
189  */
190 static void ring_write_descriptors(struct tb_ring *ring)
191 {
192 	struct ring_frame *frame, *n;
193 	struct ring_desc *descriptor;
194 	list_for_each_entry_safe(frame, n, &ring->queue, list) {
195 		if (ring_full(ring))
196 			break;
197 		list_move_tail(&frame->list, &ring->in_flight);
198 		descriptor = &ring->descriptors[ring->head];
199 		descriptor->phys = frame->buffer_phy;
200 		descriptor->time = 0;
201 		descriptor->flags = RING_DESC_POSTED | RING_DESC_INTERRUPT;
202 		if (ring->is_tx) {
203 			descriptor->length = frame->size;
204 			descriptor->eof = frame->eof;
205 			descriptor->sof = frame->sof;
206 		}
207 		ring->head = (ring->head + 1) % ring->size;
208 		if (ring->is_tx)
209 			ring_iowrite_prod(ring, ring->head);
210 		else
211 			ring_iowrite_cons(ring, ring->head);
212 	}
213 }
214 
215 /*
216  * ring_work() - progress completed frames
217  *
218  * If the ring is shutting down then all frames are marked as canceled and
219  * their callbacks are invoked.
220  *
221  * Otherwise we collect all completed frame from the ring buffer, write new
222  * frame to the ring buffer and invoke the callbacks for the completed frames.
223  */
224 static void ring_work(struct work_struct *work)
225 {
226 	struct tb_ring *ring = container_of(work, typeof(*ring), work);
227 	struct ring_frame *frame;
228 	bool canceled = false;
229 	unsigned long flags;
230 	LIST_HEAD(done);
231 
232 	spin_lock_irqsave(&ring->lock, flags);
233 
234 	if (!ring->running) {
235 		/*  Move all frames to done and mark them as canceled. */
236 		list_splice_tail_init(&ring->in_flight, &done);
237 		list_splice_tail_init(&ring->queue, &done);
238 		canceled = true;
239 		goto invoke_callback;
240 	}
241 
242 	while (!ring_empty(ring)) {
243 		if (!(ring->descriptors[ring->tail].flags
244 				& RING_DESC_COMPLETED))
245 			break;
246 		frame = list_first_entry(&ring->in_flight, typeof(*frame),
247 					 list);
248 		list_move_tail(&frame->list, &done);
249 		if (!ring->is_tx) {
250 			frame->size = ring->descriptors[ring->tail].length;
251 			frame->eof = ring->descriptors[ring->tail].eof;
252 			frame->sof = ring->descriptors[ring->tail].sof;
253 			frame->flags = ring->descriptors[ring->tail].flags;
254 		}
255 		ring->tail = (ring->tail + 1) % ring->size;
256 	}
257 	ring_write_descriptors(ring);
258 
259 invoke_callback:
260 	/* allow callbacks to schedule new work */
261 	spin_unlock_irqrestore(&ring->lock, flags);
262 	while (!list_empty(&done)) {
263 		frame = list_first_entry(&done, typeof(*frame), list);
264 		/*
265 		 * The callback may reenqueue or delete frame.
266 		 * Do not hold on to it.
267 		 */
268 		list_del_init(&frame->list);
269 		if (frame->callback)
270 			frame->callback(ring, frame, canceled);
271 	}
272 }
273 
274 int __tb_ring_enqueue(struct tb_ring *ring, struct ring_frame *frame)
275 {
276 	unsigned long flags;
277 	int ret = 0;
278 
279 	spin_lock_irqsave(&ring->lock, flags);
280 	if (ring->running) {
281 		list_add_tail(&frame->list, &ring->queue);
282 		ring_write_descriptors(ring);
283 	} else {
284 		ret = -ESHUTDOWN;
285 	}
286 	spin_unlock_irqrestore(&ring->lock, flags);
287 	return ret;
288 }
289 EXPORT_SYMBOL_GPL(__tb_ring_enqueue);
290 
291 /**
292  * tb_ring_poll() - Poll one completed frame from the ring
293  * @ring: Ring to poll
294  *
295  * This function can be called when @start_poll callback of the @ring
296  * has been called. It will read one completed frame from the ring and
297  * return it to the caller. Returns %NULL if there is no more completed
298  * frames.
299  */
300 struct ring_frame *tb_ring_poll(struct tb_ring *ring)
301 {
302 	struct ring_frame *frame = NULL;
303 	unsigned long flags;
304 
305 	spin_lock_irqsave(&ring->lock, flags);
306 	if (!ring->running)
307 		goto unlock;
308 	if (ring_empty(ring))
309 		goto unlock;
310 
311 	if (ring->descriptors[ring->tail].flags & RING_DESC_COMPLETED) {
312 		frame = list_first_entry(&ring->in_flight, typeof(*frame),
313 					 list);
314 		list_del_init(&frame->list);
315 
316 		if (!ring->is_tx) {
317 			frame->size = ring->descriptors[ring->tail].length;
318 			frame->eof = ring->descriptors[ring->tail].eof;
319 			frame->sof = ring->descriptors[ring->tail].sof;
320 			frame->flags = ring->descriptors[ring->tail].flags;
321 		}
322 
323 		ring->tail = (ring->tail + 1) % ring->size;
324 	}
325 
326 unlock:
327 	spin_unlock_irqrestore(&ring->lock, flags);
328 	return frame;
329 }
330 EXPORT_SYMBOL_GPL(tb_ring_poll);
331 
332 static void __ring_interrupt_mask(struct tb_ring *ring, bool mask)
333 {
334 	int idx = ring_interrupt_index(ring);
335 	int reg = REG_RING_INTERRUPT_BASE + idx / 32 * 4;
336 	int bit = idx % 32;
337 	u32 val;
338 
339 	val = ioread32(ring->nhi->iobase + reg);
340 	if (mask)
341 		val &= ~BIT(bit);
342 	else
343 		val |= BIT(bit);
344 	iowrite32(val, ring->nhi->iobase + reg);
345 }
346 
347 /* Both @nhi->lock and @ring->lock should be held */
348 static void __ring_interrupt(struct tb_ring *ring)
349 {
350 	if (!ring->running)
351 		return;
352 
353 	if (ring->start_poll) {
354 		__ring_interrupt_mask(ring, true);
355 		ring->start_poll(ring->poll_data);
356 	} else {
357 		schedule_work(&ring->work);
358 	}
359 }
360 
361 /**
362  * tb_ring_poll_complete() - Re-start interrupt for the ring
363  * @ring: Ring to re-start the interrupt
364  *
365  * This will re-start (unmask) the ring interrupt once the user is done
366  * with polling.
367  */
368 void tb_ring_poll_complete(struct tb_ring *ring)
369 {
370 	unsigned long flags;
371 
372 	spin_lock_irqsave(&ring->nhi->lock, flags);
373 	spin_lock(&ring->lock);
374 	if (ring->start_poll)
375 		__ring_interrupt_mask(ring, false);
376 	spin_unlock(&ring->lock);
377 	spin_unlock_irqrestore(&ring->nhi->lock, flags);
378 }
379 EXPORT_SYMBOL_GPL(tb_ring_poll_complete);
380 
381 static irqreturn_t ring_msix(int irq, void *data)
382 {
383 	struct tb_ring *ring = data;
384 
385 	spin_lock(&ring->nhi->lock);
386 	spin_lock(&ring->lock);
387 	__ring_interrupt(ring);
388 	spin_unlock(&ring->lock);
389 	spin_unlock(&ring->nhi->lock);
390 
391 	return IRQ_HANDLED;
392 }
393 
394 static int ring_request_msix(struct tb_ring *ring, bool no_suspend)
395 {
396 	struct tb_nhi *nhi = ring->nhi;
397 	unsigned long irqflags;
398 	int ret;
399 
400 	if (!nhi->pdev->msix_enabled)
401 		return 0;
402 
403 	ret = ida_simple_get(&nhi->msix_ida, 0, MSIX_MAX_VECS, GFP_KERNEL);
404 	if (ret < 0)
405 		return ret;
406 
407 	ring->vector = ret;
408 
409 	ret = pci_irq_vector(ring->nhi->pdev, ring->vector);
410 	if (ret < 0)
411 		goto err_ida_remove;
412 
413 	ring->irq = ret;
414 
415 	irqflags = no_suspend ? IRQF_NO_SUSPEND : 0;
416 	ret = request_irq(ring->irq, ring_msix, irqflags, "thunderbolt", ring);
417 	if (ret)
418 		goto err_ida_remove;
419 
420 	return 0;
421 
422 err_ida_remove:
423 	ida_simple_remove(&nhi->msix_ida, ring->vector);
424 
425 	return ret;
426 }
427 
428 static void ring_release_msix(struct tb_ring *ring)
429 {
430 	if (ring->irq <= 0)
431 		return;
432 
433 	free_irq(ring->irq, ring);
434 	ida_simple_remove(&ring->nhi->msix_ida, ring->vector);
435 	ring->vector = 0;
436 	ring->irq = 0;
437 }
438 
439 static int nhi_alloc_hop(struct tb_nhi *nhi, struct tb_ring *ring)
440 {
441 	int ret = 0;
442 
443 	spin_lock_irq(&nhi->lock);
444 
445 	if (ring->hop < 0) {
446 		unsigned int i;
447 
448 		/*
449 		 * Automatically allocate HopID from the non-reserved
450 		 * range 1 .. hop_count - 1.
451 		 */
452 		for (i = RING_FIRST_USABLE_HOPID; i < nhi->hop_count; i++) {
453 			if (ring->is_tx) {
454 				if (!nhi->tx_rings[i]) {
455 					ring->hop = i;
456 					break;
457 				}
458 			} else {
459 				if (!nhi->rx_rings[i]) {
460 					ring->hop = i;
461 					break;
462 				}
463 			}
464 		}
465 	}
466 
467 	if (ring->hop < 0 || ring->hop >= nhi->hop_count) {
468 		dev_warn(&nhi->pdev->dev, "invalid hop: %d\n", ring->hop);
469 		ret = -EINVAL;
470 		goto err_unlock;
471 	}
472 	if (ring->is_tx && nhi->tx_rings[ring->hop]) {
473 		dev_warn(&nhi->pdev->dev, "TX hop %d already allocated\n",
474 			 ring->hop);
475 		ret = -EBUSY;
476 		goto err_unlock;
477 	} else if (!ring->is_tx && nhi->rx_rings[ring->hop]) {
478 		dev_warn(&nhi->pdev->dev, "RX hop %d already allocated\n",
479 			 ring->hop);
480 		ret = -EBUSY;
481 		goto err_unlock;
482 	}
483 
484 	if (ring->is_tx)
485 		nhi->tx_rings[ring->hop] = ring;
486 	else
487 		nhi->rx_rings[ring->hop] = ring;
488 
489 err_unlock:
490 	spin_unlock_irq(&nhi->lock);
491 
492 	return ret;
493 }
494 
495 static struct tb_ring *tb_ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
496 				     bool transmit, unsigned int flags,
497 				     int e2e_tx_hop, u16 sof_mask, u16 eof_mask,
498 				     void (*start_poll)(void *),
499 				     void *poll_data)
500 {
501 	struct tb_ring *ring = NULL;
502 
503 	dev_dbg(&nhi->pdev->dev, "allocating %s ring %d of size %d\n",
504 		transmit ? "TX" : "RX", hop, size);
505 
506 	ring = kzalloc(sizeof(*ring), GFP_KERNEL);
507 	if (!ring)
508 		return NULL;
509 
510 	spin_lock_init(&ring->lock);
511 	INIT_LIST_HEAD(&ring->queue);
512 	INIT_LIST_HEAD(&ring->in_flight);
513 	INIT_WORK(&ring->work, ring_work);
514 
515 	ring->nhi = nhi;
516 	ring->hop = hop;
517 	ring->is_tx = transmit;
518 	ring->size = size;
519 	ring->flags = flags;
520 	ring->e2e_tx_hop = e2e_tx_hop;
521 	ring->sof_mask = sof_mask;
522 	ring->eof_mask = eof_mask;
523 	ring->head = 0;
524 	ring->tail = 0;
525 	ring->running = false;
526 	ring->start_poll = start_poll;
527 	ring->poll_data = poll_data;
528 
529 	ring->descriptors = dma_alloc_coherent(&ring->nhi->pdev->dev,
530 			size * sizeof(*ring->descriptors),
531 			&ring->descriptors_dma, GFP_KERNEL | __GFP_ZERO);
532 	if (!ring->descriptors)
533 		goto err_free_ring;
534 
535 	if (ring_request_msix(ring, flags & RING_FLAG_NO_SUSPEND))
536 		goto err_free_descs;
537 
538 	if (nhi_alloc_hop(nhi, ring))
539 		goto err_release_msix;
540 
541 	return ring;
542 
543 err_release_msix:
544 	ring_release_msix(ring);
545 err_free_descs:
546 	dma_free_coherent(&ring->nhi->pdev->dev,
547 			  ring->size * sizeof(*ring->descriptors),
548 			  ring->descriptors, ring->descriptors_dma);
549 err_free_ring:
550 	kfree(ring);
551 
552 	return NULL;
553 }
554 
555 /**
556  * tb_ring_alloc_tx() - Allocate DMA ring for transmit
557  * @nhi: Pointer to the NHI the ring is to be allocated
558  * @hop: HopID (ring) to allocate
559  * @size: Number of entries in the ring
560  * @flags: Flags for the ring
561  */
562 struct tb_ring *tb_ring_alloc_tx(struct tb_nhi *nhi, int hop, int size,
563 				 unsigned int flags)
564 {
565 	return tb_ring_alloc(nhi, hop, size, true, flags, 0, 0, 0, NULL, NULL);
566 }
567 EXPORT_SYMBOL_GPL(tb_ring_alloc_tx);
568 
569 /**
570  * tb_ring_alloc_rx() - Allocate DMA ring for receive
571  * @nhi: Pointer to the NHI the ring is to be allocated
572  * @hop: HopID (ring) to allocate. Pass %-1 for automatic allocation.
573  * @size: Number of entries in the ring
574  * @flags: Flags for the ring
575  * @e2e_tx_hop: Transmit HopID when E2E is enabled in @flags
576  * @sof_mask: Mask of PDF values that start a frame
577  * @eof_mask: Mask of PDF values that end a frame
578  * @start_poll: If not %NULL the ring will call this function when an
579  *		interrupt is triggered and masked, instead of callback
580  *		in each Rx frame.
581  * @poll_data: Optional data passed to @start_poll
582  */
583 struct tb_ring *tb_ring_alloc_rx(struct tb_nhi *nhi, int hop, int size,
584 				 unsigned int flags, int e2e_tx_hop,
585 				 u16 sof_mask, u16 eof_mask,
586 				 void (*start_poll)(void *), void *poll_data)
587 {
588 	return tb_ring_alloc(nhi, hop, size, false, flags, e2e_tx_hop, sof_mask, eof_mask,
589 			     start_poll, poll_data);
590 }
591 EXPORT_SYMBOL_GPL(tb_ring_alloc_rx);
592 
593 /**
594  * tb_ring_start() - enable a ring
595  * @ring: Ring to start
596  *
597  * Must not be invoked in parallel with tb_ring_stop().
598  */
599 void tb_ring_start(struct tb_ring *ring)
600 {
601 	u16 frame_size;
602 	u32 flags;
603 
604 	spin_lock_irq(&ring->nhi->lock);
605 	spin_lock(&ring->lock);
606 	if (ring->nhi->going_away)
607 		goto err;
608 	if (ring->running) {
609 		dev_WARN(&ring->nhi->pdev->dev, "ring already started\n");
610 		goto err;
611 	}
612 	dev_dbg(&ring->nhi->pdev->dev, "starting %s %d\n",
613 		RING_TYPE(ring), ring->hop);
614 
615 	if (ring->flags & RING_FLAG_FRAME) {
616 		/* Means 4096 */
617 		frame_size = 0;
618 		flags = RING_FLAG_ENABLE;
619 	} else {
620 		frame_size = TB_FRAME_SIZE;
621 		flags = RING_FLAG_ENABLE | RING_FLAG_RAW;
622 	}
623 
624 	ring_iowrite64desc(ring, ring->descriptors_dma, 0);
625 	if (ring->is_tx) {
626 		ring_iowrite32desc(ring, ring->size, 12);
627 		ring_iowrite32options(ring, 0, 4); /* time releated ? */
628 		ring_iowrite32options(ring, flags, 0);
629 	} else {
630 		u32 sof_eof_mask = ring->sof_mask << 16 | ring->eof_mask;
631 
632 		ring_iowrite32desc(ring, (frame_size << 16) | ring->size, 12);
633 		ring_iowrite32options(ring, sof_eof_mask, 4);
634 		ring_iowrite32options(ring, flags, 0);
635 	}
636 
637 	/*
638 	 * Now that the ring valid bit is set we can configure E2E if
639 	 * enabled for the ring.
640 	 */
641 	if (ring->flags & RING_FLAG_E2E) {
642 		if (!ring->is_tx) {
643 			u32 hop;
644 
645 			hop = ring->e2e_tx_hop << REG_RX_OPTIONS_E2E_HOP_SHIFT;
646 			hop &= REG_RX_OPTIONS_E2E_HOP_MASK;
647 			flags |= hop;
648 
649 			dev_dbg(&ring->nhi->pdev->dev,
650 				"enabling E2E for %s %d with TX HopID %d\n",
651 				RING_TYPE(ring), ring->hop, ring->e2e_tx_hop);
652 		} else {
653 			dev_dbg(&ring->nhi->pdev->dev, "enabling E2E for %s %d\n",
654 				RING_TYPE(ring), ring->hop);
655 		}
656 
657 		flags |= RING_FLAG_E2E_FLOW_CONTROL;
658 		ring_iowrite32options(ring, flags, 0);
659 	}
660 
661 	ring_interrupt_active(ring, true);
662 	ring->running = true;
663 err:
664 	spin_unlock(&ring->lock);
665 	spin_unlock_irq(&ring->nhi->lock);
666 }
667 EXPORT_SYMBOL_GPL(tb_ring_start);
668 
669 /**
670  * tb_ring_stop() - shutdown a ring
671  * @ring: Ring to stop
672  *
673  * Must not be invoked from a callback.
674  *
675  * This method will disable the ring. Further calls to
676  * tb_ring_tx/tb_ring_rx will return -ESHUTDOWN until ring_stop has been
677  * called.
678  *
679  * All enqueued frames will be canceled and their callbacks will be executed
680  * with frame->canceled set to true (on the callback thread). This method
681  * returns only after all callback invocations have finished.
682  */
683 void tb_ring_stop(struct tb_ring *ring)
684 {
685 	spin_lock_irq(&ring->nhi->lock);
686 	spin_lock(&ring->lock);
687 	dev_dbg(&ring->nhi->pdev->dev, "stopping %s %d\n",
688 		RING_TYPE(ring), ring->hop);
689 	if (ring->nhi->going_away)
690 		goto err;
691 	if (!ring->running) {
692 		dev_WARN(&ring->nhi->pdev->dev, "%s %d already stopped\n",
693 			 RING_TYPE(ring), ring->hop);
694 		goto err;
695 	}
696 	ring_interrupt_active(ring, false);
697 
698 	ring_iowrite32options(ring, 0, 0);
699 	ring_iowrite64desc(ring, 0, 0);
700 	ring_iowrite32desc(ring, 0, 8);
701 	ring_iowrite32desc(ring, 0, 12);
702 	ring->head = 0;
703 	ring->tail = 0;
704 	ring->running = false;
705 
706 err:
707 	spin_unlock(&ring->lock);
708 	spin_unlock_irq(&ring->nhi->lock);
709 
710 	/*
711 	 * schedule ring->work to invoke callbacks on all remaining frames.
712 	 */
713 	schedule_work(&ring->work);
714 	flush_work(&ring->work);
715 }
716 EXPORT_SYMBOL_GPL(tb_ring_stop);
717 
718 /*
719  * tb_ring_free() - free ring
720  *
721  * When this method returns all invocations of ring->callback will have
722  * finished.
723  *
724  * Ring must be stopped.
725  *
726  * Must NOT be called from ring_frame->callback!
727  */
728 void tb_ring_free(struct tb_ring *ring)
729 {
730 	spin_lock_irq(&ring->nhi->lock);
731 	/*
732 	 * Dissociate the ring from the NHI. This also ensures that
733 	 * nhi_interrupt_work cannot reschedule ring->work.
734 	 */
735 	if (ring->is_tx)
736 		ring->nhi->tx_rings[ring->hop] = NULL;
737 	else
738 		ring->nhi->rx_rings[ring->hop] = NULL;
739 
740 	if (ring->running) {
741 		dev_WARN(&ring->nhi->pdev->dev, "%s %d still running\n",
742 			 RING_TYPE(ring), ring->hop);
743 	}
744 	spin_unlock_irq(&ring->nhi->lock);
745 
746 	ring_release_msix(ring);
747 
748 	dma_free_coherent(&ring->nhi->pdev->dev,
749 			  ring->size * sizeof(*ring->descriptors),
750 			  ring->descriptors, ring->descriptors_dma);
751 
752 	ring->descriptors = NULL;
753 	ring->descriptors_dma = 0;
754 
755 
756 	dev_dbg(&ring->nhi->pdev->dev, "freeing %s %d\n", RING_TYPE(ring),
757 		ring->hop);
758 
759 	/*
760 	 * ring->work can no longer be scheduled (it is scheduled only
761 	 * by nhi_interrupt_work, ring_stop and ring_msix). Wait for it
762 	 * to finish before freeing the ring.
763 	 */
764 	flush_work(&ring->work);
765 	kfree(ring);
766 }
767 EXPORT_SYMBOL_GPL(tb_ring_free);
768 
769 /**
770  * nhi_mailbox_cmd() - Send a command through NHI mailbox
771  * @nhi: Pointer to the NHI structure
772  * @cmd: Command to send
773  * @data: Data to be send with the command
774  *
775  * Sends mailbox command to the firmware running on NHI. Returns %0 in
776  * case of success and negative errno in case of failure.
777  */
778 int nhi_mailbox_cmd(struct tb_nhi *nhi, enum nhi_mailbox_cmd cmd, u32 data)
779 {
780 	ktime_t timeout;
781 	u32 val;
782 
783 	iowrite32(data, nhi->iobase + REG_INMAIL_DATA);
784 
785 	val = ioread32(nhi->iobase + REG_INMAIL_CMD);
786 	val &= ~(REG_INMAIL_CMD_MASK | REG_INMAIL_ERROR);
787 	val |= REG_INMAIL_OP_REQUEST | cmd;
788 	iowrite32(val, nhi->iobase + REG_INMAIL_CMD);
789 
790 	timeout = ktime_add_ms(ktime_get(), NHI_MAILBOX_TIMEOUT);
791 	do {
792 		val = ioread32(nhi->iobase + REG_INMAIL_CMD);
793 		if (!(val & REG_INMAIL_OP_REQUEST))
794 			break;
795 		usleep_range(10, 20);
796 	} while (ktime_before(ktime_get(), timeout));
797 
798 	if (val & REG_INMAIL_OP_REQUEST)
799 		return -ETIMEDOUT;
800 	if (val & REG_INMAIL_ERROR)
801 		return -EIO;
802 
803 	return 0;
804 }
805 
806 /**
807  * nhi_mailbox_mode() - Return current firmware operation mode
808  * @nhi: Pointer to the NHI structure
809  *
810  * The function reads current firmware operation mode using NHI mailbox
811  * registers and returns it to the caller.
812  */
813 enum nhi_fw_mode nhi_mailbox_mode(struct tb_nhi *nhi)
814 {
815 	u32 val;
816 
817 	val = ioread32(nhi->iobase + REG_OUTMAIL_CMD);
818 	val &= REG_OUTMAIL_CMD_OPMODE_MASK;
819 	val >>= REG_OUTMAIL_CMD_OPMODE_SHIFT;
820 
821 	return (enum nhi_fw_mode)val;
822 }
823 
824 static void nhi_interrupt_work(struct work_struct *work)
825 {
826 	struct tb_nhi *nhi = container_of(work, typeof(*nhi), interrupt_work);
827 	int value = 0; /* Suppress uninitialized usage warning. */
828 	int bit;
829 	int hop = -1;
830 	int type = 0; /* current interrupt type 0: TX, 1: RX, 2: RX overflow */
831 	struct tb_ring *ring;
832 
833 	spin_lock_irq(&nhi->lock);
834 
835 	/*
836 	 * Starting at REG_RING_NOTIFY_BASE there are three status bitfields
837 	 * (TX, RX, RX overflow). We iterate over the bits and read a new
838 	 * dwords as required. The registers are cleared on read.
839 	 */
840 	for (bit = 0; bit < 3 * nhi->hop_count; bit++) {
841 		if (bit % 32 == 0)
842 			value = ioread32(nhi->iobase
843 					 + REG_RING_NOTIFY_BASE
844 					 + 4 * (bit / 32));
845 		if (++hop == nhi->hop_count) {
846 			hop = 0;
847 			type++;
848 		}
849 		if ((value & (1 << (bit % 32))) == 0)
850 			continue;
851 		if (type == 2) {
852 			dev_warn(&nhi->pdev->dev,
853 				 "RX overflow for ring %d\n",
854 				 hop);
855 			continue;
856 		}
857 		if (type == 0)
858 			ring = nhi->tx_rings[hop];
859 		else
860 			ring = nhi->rx_rings[hop];
861 		if (ring == NULL) {
862 			dev_warn(&nhi->pdev->dev,
863 				 "got interrupt for inactive %s ring %d\n",
864 				 type ? "RX" : "TX",
865 				 hop);
866 			continue;
867 		}
868 
869 		spin_lock(&ring->lock);
870 		__ring_interrupt(ring);
871 		spin_unlock(&ring->lock);
872 	}
873 	spin_unlock_irq(&nhi->lock);
874 }
875 
876 static irqreturn_t nhi_msi(int irq, void *data)
877 {
878 	struct tb_nhi *nhi = data;
879 	schedule_work(&nhi->interrupt_work);
880 	return IRQ_HANDLED;
881 }
882 
883 static int __nhi_suspend_noirq(struct device *dev, bool wakeup)
884 {
885 	struct pci_dev *pdev = to_pci_dev(dev);
886 	struct tb *tb = pci_get_drvdata(pdev);
887 	struct tb_nhi *nhi = tb->nhi;
888 	int ret;
889 
890 	ret = tb_domain_suspend_noirq(tb);
891 	if (ret)
892 		return ret;
893 
894 	if (nhi->ops && nhi->ops->suspend_noirq) {
895 		ret = nhi->ops->suspend_noirq(tb->nhi, wakeup);
896 		if (ret)
897 			return ret;
898 	}
899 
900 	return 0;
901 }
902 
903 static int nhi_suspend_noirq(struct device *dev)
904 {
905 	return __nhi_suspend_noirq(dev, device_may_wakeup(dev));
906 }
907 
908 static int nhi_freeze_noirq(struct device *dev)
909 {
910 	struct pci_dev *pdev = to_pci_dev(dev);
911 	struct tb *tb = pci_get_drvdata(pdev);
912 
913 	return tb_domain_freeze_noirq(tb);
914 }
915 
916 static int nhi_thaw_noirq(struct device *dev)
917 {
918 	struct pci_dev *pdev = to_pci_dev(dev);
919 	struct tb *tb = pci_get_drvdata(pdev);
920 
921 	return tb_domain_thaw_noirq(tb);
922 }
923 
924 static bool nhi_wake_supported(struct pci_dev *pdev)
925 {
926 	u8 val;
927 
928 	/*
929 	 * If power rails are sustainable for wakeup from S4 this
930 	 * property is set by the BIOS.
931 	 */
932 	if (device_property_read_u8(&pdev->dev, "WAKE_SUPPORTED", &val))
933 		return !!val;
934 
935 	return true;
936 }
937 
938 static int nhi_poweroff_noirq(struct device *dev)
939 {
940 	struct pci_dev *pdev = to_pci_dev(dev);
941 	bool wakeup;
942 
943 	wakeup = device_may_wakeup(dev) && nhi_wake_supported(pdev);
944 	return __nhi_suspend_noirq(dev, wakeup);
945 }
946 
947 static void nhi_enable_int_throttling(struct tb_nhi *nhi)
948 {
949 	/* Throttling is specified in 256ns increments */
950 	u32 throttle = DIV_ROUND_UP(128 * NSEC_PER_USEC, 256);
951 	unsigned int i;
952 
953 	/*
954 	 * Configure interrupt throttling for all vectors even if we
955 	 * only use few.
956 	 */
957 	for (i = 0; i < MSIX_MAX_VECS; i++) {
958 		u32 reg = REG_INT_THROTTLING_RATE + i * 4;
959 		iowrite32(throttle, nhi->iobase + reg);
960 	}
961 }
962 
963 static int nhi_resume_noirq(struct device *dev)
964 {
965 	struct pci_dev *pdev = to_pci_dev(dev);
966 	struct tb *tb = pci_get_drvdata(pdev);
967 	struct tb_nhi *nhi = tb->nhi;
968 	int ret;
969 
970 	/*
971 	 * Check that the device is still there. It may be that the user
972 	 * unplugged last device which causes the host controller to go
973 	 * away on PCs.
974 	 */
975 	if (!pci_device_is_present(pdev)) {
976 		nhi->going_away = true;
977 	} else {
978 		if (nhi->ops && nhi->ops->resume_noirq) {
979 			ret = nhi->ops->resume_noirq(nhi);
980 			if (ret)
981 				return ret;
982 		}
983 		nhi_enable_int_throttling(tb->nhi);
984 	}
985 
986 	return tb_domain_resume_noirq(tb);
987 }
988 
989 static int nhi_suspend(struct device *dev)
990 {
991 	struct pci_dev *pdev = to_pci_dev(dev);
992 	struct tb *tb = pci_get_drvdata(pdev);
993 
994 	return tb_domain_suspend(tb);
995 }
996 
997 static void nhi_complete(struct device *dev)
998 {
999 	struct pci_dev *pdev = to_pci_dev(dev);
1000 	struct tb *tb = pci_get_drvdata(pdev);
1001 
1002 	/*
1003 	 * If we were runtime suspended when system suspend started,
1004 	 * schedule runtime resume now. It should bring the domain back
1005 	 * to functional state.
1006 	 */
1007 	if (pm_runtime_suspended(&pdev->dev))
1008 		pm_runtime_resume(&pdev->dev);
1009 	else
1010 		tb_domain_complete(tb);
1011 }
1012 
1013 static int nhi_runtime_suspend(struct device *dev)
1014 {
1015 	struct pci_dev *pdev = to_pci_dev(dev);
1016 	struct tb *tb = pci_get_drvdata(pdev);
1017 	struct tb_nhi *nhi = tb->nhi;
1018 	int ret;
1019 
1020 	ret = tb_domain_runtime_suspend(tb);
1021 	if (ret)
1022 		return ret;
1023 
1024 	if (nhi->ops && nhi->ops->runtime_suspend) {
1025 		ret = nhi->ops->runtime_suspend(tb->nhi);
1026 		if (ret)
1027 			return ret;
1028 	}
1029 	return 0;
1030 }
1031 
1032 static int nhi_runtime_resume(struct device *dev)
1033 {
1034 	struct pci_dev *pdev = to_pci_dev(dev);
1035 	struct tb *tb = pci_get_drvdata(pdev);
1036 	struct tb_nhi *nhi = tb->nhi;
1037 	int ret;
1038 
1039 	if (nhi->ops && nhi->ops->runtime_resume) {
1040 		ret = nhi->ops->runtime_resume(nhi);
1041 		if (ret)
1042 			return ret;
1043 	}
1044 
1045 	nhi_enable_int_throttling(nhi);
1046 	return tb_domain_runtime_resume(tb);
1047 }
1048 
1049 static void nhi_shutdown(struct tb_nhi *nhi)
1050 {
1051 	int i;
1052 
1053 	dev_dbg(&nhi->pdev->dev, "shutdown\n");
1054 
1055 	for (i = 0; i < nhi->hop_count; i++) {
1056 		if (nhi->tx_rings[i])
1057 			dev_WARN(&nhi->pdev->dev,
1058 				 "TX ring %d is still active\n", i);
1059 		if (nhi->rx_rings[i])
1060 			dev_WARN(&nhi->pdev->dev,
1061 				 "RX ring %d is still active\n", i);
1062 	}
1063 	nhi_disable_interrupts(nhi);
1064 	/*
1065 	 * We have to release the irq before calling flush_work. Otherwise an
1066 	 * already executing IRQ handler could call schedule_work again.
1067 	 */
1068 	if (!nhi->pdev->msix_enabled) {
1069 		devm_free_irq(&nhi->pdev->dev, nhi->pdev->irq, nhi);
1070 		flush_work(&nhi->interrupt_work);
1071 	}
1072 	ida_destroy(&nhi->msix_ida);
1073 
1074 	if (nhi->ops && nhi->ops->shutdown)
1075 		nhi->ops->shutdown(nhi);
1076 }
1077 
1078 static int nhi_init_msi(struct tb_nhi *nhi)
1079 {
1080 	struct pci_dev *pdev = nhi->pdev;
1081 	int res, irq, nvec;
1082 
1083 	/* In case someone left them on. */
1084 	nhi_disable_interrupts(nhi);
1085 
1086 	nhi_enable_int_throttling(nhi);
1087 
1088 	ida_init(&nhi->msix_ida);
1089 
1090 	/*
1091 	 * The NHI has 16 MSI-X vectors or a single MSI. We first try to
1092 	 * get all MSI-X vectors and if we succeed, each ring will have
1093 	 * one MSI-X. If for some reason that does not work out, we
1094 	 * fallback to a single MSI.
1095 	 */
1096 	nvec = pci_alloc_irq_vectors(pdev, MSIX_MIN_VECS, MSIX_MAX_VECS,
1097 				     PCI_IRQ_MSIX);
1098 	if (nvec < 0) {
1099 		nvec = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI);
1100 		if (nvec < 0)
1101 			return nvec;
1102 
1103 		INIT_WORK(&nhi->interrupt_work, nhi_interrupt_work);
1104 
1105 		irq = pci_irq_vector(nhi->pdev, 0);
1106 		if (irq < 0)
1107 			return irq;
1108 
1109 		res = devm_request_irq(&pdev->dev, irq, nhi_msi,
1110 				       IRQF_NO_SUSPEND, "thunderbolt", nhi);
1111 		if (res) {
1112 			dev_err(&pdev->dev, "request_irq failed, aborting\n");
1113 			return res;
1114 		}
1115 	}
1116 
1117 	return 0;
1118 }
1119 
1120 static bool nhi_imr_valid(struct pci_dev *pdev)
1121 {
1122 	u8 val;
1123 
1124 	if (!device_property_read_u8(&pdev->dev, "IMR_VALID", &val))
1125 		return !!val;
1126 
1127 	return true;
1128 }
1129 
1130 /*
1131  * During suspend the Thunderbolt controller is reset and all PCIe
1132  * tunnels are lost. The NHI driver will try to reestablish all tunnels
1133  * during resume. This adds device links between the tunneled PCIe
1134  * downstream ports and the NHI so that the device core will make sure
1135  * NHI is resumed first before the rest.
1136  */
1137 static void tb_apple_add_links(struct tb_nhi *nhi)
1138 {
1139 	struct pci_dev *upstream, *pdev;
1140 
1141 	if (!x86_apple_machine)
1142 		return;
1143 
1144 	switch (nhi->pdev->device) {
1145 	case PCI_DEVICE_ID_INTEL_LIGHT_RIDGE:
1146 	case PCI_DEVICE_ID_INTEL_CACTUS_RIDGE_4C:
1147 	case PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI:
1148 	case PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI:
1149 		break;
1150 	default:
1151 		return;
1152 	}
1153 
1154 	upstream = pci_upstream_bridge(nhi->pdev);
1155 	while (upstream) {
1156 		if (!pci_is_pcie(upstream))
1157 			return;
1158 		if (pci_pcie_type(upstream) == PCI_EXP_TYPE_UPSTREAM)
1159 			break;
1160 		upstream = pci_upstream_bridge(upstream);
1161 	}
1162 
1163 	if (!upstream)
1164 		return;
1165 
1166 	/*
1167 	 * For each hotplug downstream port, create add device link
1168 	 * back to NHI so that PCIe tunnels can be re-established after
1169 	 * sleep.
1170 	 */
1171 	for_each_pci_bridge(pdev, upstream->subordinate) {
1172 		const struct device_link *link;
1173 
1174 		if (!pci_is_pcie(pdev))
1175 			continue;
1176 		if (pci_pcie_type(pdev) != PCI_EXP_TYPE_DOWNSTREAM ||
1177 		    !pdev->is_hotplug_bridge)
1178 			continue;
1179 
1180 		link = device_link_add(&pdev->dev, &nhi->pdev->dev,
1181 				       DL_FLAG_AUTOREMOVE_SUPPLIER |
1182 				       DL_FLAG_PM_RUNTIME);
1183 		if (link) {
1184 			dev_dbg(&nhi->pdev->dev, "created link from %s\n",
1185 				dev_name(&pdev->dev));
1186 		} else {
1187 			dev_warn(&nhi->pdev->dev, "device link creation from %s failed\n",
1188 				 dev_name(&pdev->dev));
1189 		}
1190 	}
1191 }
1192 
1193 static struct tb *nhi_select_cm(struct tb_nhi *nhi)
1194 {
1195 	struct tb *tb;
1196 
1197 	/*
1198 	 * USB4 case is simple. If we got control of any of the
1199 	 * capabilities, we use software CM.
1200 	 */
1201 	if (tb_acpi_is_native())
1202 		return tb_probe(nhi);
1203 
1204 	/*
1205 	 * Either firmware based CM is running (we did not get control
1206 	 * from the firmware) or this is pre-USB4 PC so try first
1207 	 * firmware CM and then fallback to software CM.
1208 	 */
1209 	tb = icm_probe(nhi);
1210 	if (!tb)
1211 		tb = tb_probe(nhi);
1212 
1213 	return tb;
1214 }
1215 
1216 static int nhi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1217 {
1218 	struct tb_nhi *nhi;
1219 	struct tb *tb;
1220 	int res;
1221 
1222 	if (!nhi_imr_valid(pdev)) {
1223 		dev_warn(&pdev->dev, "firmware image not valid, aborting\n");
1224 		return -ENODEV;
1225 	}
1226 
1227 	res = pcim_enable_device(pdev);
1228 	if (res) {
1229 		dev_err(&pdev->dev, "cannot enable PCI device, aborting\n");
1230 		return res;
1231 	}
1232 
1233 	res = pcim_iomap_regions(pdev, 1 << 0, "thunderbolt");
1234 	if (res) {
1235 		dev_err(&pdev->dev, "cannot obtain PCI resources, aborting\n");
1236 		return res;
1237 	}
1238 
1239 	nhi = devm_kzalloc(&pdev->dev, sizeof(*nhi), GFP_KERNEL);
1240 	if (!nhi)
1241 		return -ENOMEM;
1242 
1243 	nhi->pdev = pdev;
1244 	nhi->ops = (const struct tb_nhi_ops *)id->driver_data;
1245 	/* cannot fail - table is allocated bin pcim_iomap_regions */
1246 	nhi->iobase = pcim_iomap_table(pdev)[0];
1247 	nhi->hop_count = ioread32(nhi->iobase + REG_HOP_COUNT) & 0x3ff;
1248 	dev_dbg(&pdev->dev, "total paths: %d\n", nhi->hop_count);
1249 
1250 	nhi->tx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1251 				     sizeof(*nhi->tx_rings), GFP_KERNEL);
1252 	nhi->rx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1253 				     sizeof(*nhi->rx_rings), GFP_KERNEL);
1254 	if (!nhi->tx_rings || !nhi->rx_rings)
1255 		return -ENOMEM;
1256 
1257 	res = nhi_init_msi(nhi);
1258 	if (res) {
1259 		dev_err(&pdev->dev, "cannot enable MSI, aborting\n");
1260 		return res;
1261 	}
1262 
1263 	spin_lock_init(&nhi->lock);
1264 
1265 	res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1266 	if (res)
1267 		res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
1268 	if (res) {
1269 		dev_err(&pdev->dev, "failed to set DMA mask\n");
1270 		return res;
1271 	}
1272 
1273 	pci_set_master(pdev);
1274 
1275 	if (nhi->ops && nhi->ops->init) {
1276 		res = nhi->ops->init(nhi);
1277 		if (res)
1278 			return res;
1279 	}
1280 
1281 	tb_apple_add_links(nhi);
1282 	tb_acpi_add_links(nhi);
1283 
1284 	tb = nhi_select_cm(nhi);
1285 	if (!tb) {
1286 		dev_err(&nhi->pdev->dev,
1287 			"failed to determine connection manager, aborting\n");
1288 		return -ENODEV;
1289 	}
1290 
1291 	dev_dbg(&nhi->pdev->dev, "NHI initialized, starting thunderbolt\n");
1292 
1293 	res = tb_domain_add(tb);
1294 	if (res) {
1295 		/*
1296 		 * At this point the RX/TX rings might already have been
1297 		 * activated. Do a proper shutdown.
1298 		 */
1299 		tb_domain_put(tb);
1300 		nhi_shutdown(nhi);
1301 		return res;
1302 	}
1303 	pci_set_drvdata(pdev, tb);
1304 
1305 	device_wakeup_enable(&pdev->dev);
1306 
1307 	pm_runtime_allow(&pdev->dev);
1308 	pm_runtime_set_autosuspend_delay(&pdev->dev, TB_AUTOSUSPEND_DELAY);
1309 	pm_runtime_use_autosuspend(&pdev->dev);
1310 	pm_runtime_put_autosuspend(&pdev->dev);
1311 
1312 	return 0;
1313 }
1314 
1315 static void nhi_remove(struct pci_dev *pdev)
1316 {
1317 	struct tb *tb = pci_get_drvdata(pdev);
1318 	struct tb_nhi *nhi = tb->nhi;
1319 
1320 	pm_runtime_get_sync(&pdev->dev);
1321 	pm_runtime_dont_use_autosuspend(&pdev->dev);
1322 	pm_runtime_forbid(&pdev->dev);
1323 
1324 	tb_domain_remove(tb);
1325 	nhi_shutdown(nhi);
1326 }
1327 
1328 /*
1329  * The tunneled pci bridges are siblings of us. Use resume_noirq to reenable
1330  * the tunnels asap. A corresponding pci quirk blocks the downstream bridges
1331  * resume_noirq until we are done.
1332  */
1333 static const struct dev_pm_ops nhi_pm_ops = {
1334 	.suspend_noirq = nhi_suspend_noirq,
1335 	.resume_noirq = nhi_resume_noirq,
1336 	.freeze_noirq = nhi_freeze_noirq,  /*
1337 					    * we just disable hotplug, the
1338 					    * pci-tunnels stay alive.
1339 					    */
1340 	.thaw_noirq = nhi_thaw_noirq,
1341 	.restore_noirq = nhi_resume_noirq,
1342 	.suspend = nhi_suspend,
1343 	.poweroff_noirq = nhi_poweroff_noirq,
1344 	.poweroff = nhi_suspend,
1345 	.complete = nhi_complete,
1346 	.runtime_suspend = nhi_runtime_suspend,
1347 	.runtime_resume = nhi_runtime_resume,
1348 };
1349 
1350 static struct pci_device_id nhi_ids[] = {
1351 	/*
1352 	 * We have to specify class, the TB bridges use the same device and
1353 	 * vendor (sub)id on gen 1 and gen 2 controllers.
1354 	 */
1355 	{
1356 		.class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1357 		.vendor = PCI_VENDOR_ID_INTEL,
1358 		.device = PCI_DEVICE_ID_INTEL_LIGHT_RIDGE,
1359 		.subvendor = 0x2222, .subdevice = 0x1111,
1360 	},
1361 	{
1362 		.class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1363 		.vendor = PCI_VENDOR_ID_INTEL,
1364 		.device = PCI_DEVICE_ID_INTEL_CACTUS_RIDGE_4C,
1365 		.subvendor = 0x2222, .subdevice = 0x1111,
1366 	},
1367 	{
1368 		.class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1369 		.vendor = PCI_VENDOR_ID_INTEL,
1370 		.device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI,
1371 		.subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1372 	},
1373 	{
1374 		.class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1375 		.vendor = PCI_VENDOR_ID_INTEL,
1376 		.device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI,
1377 		.subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1378 	},
1379 
1380 	/* Thunderbolt 3 */
1381 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_2C_NHI) },
1382 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_4C_NHI) },
1383 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_USBONLY_NHI) },
1384 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_NHI) },
1385 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_USBONLY_NHI) },
1386 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_2C_NHI) },
1387 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_4C_NHI) },
1388 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_USBONLY_NHI) },
1389 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI) },
1390 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI) },
1391 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ICL_NHI0),
1392 	  .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1393 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ICL_NHI1),
1394 	  .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1395 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_NHI0),
1396 	  .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1397 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_NHI1),
1398 	  .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1399 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_H_NHI0),
1400 	  .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1401 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_H_NHI1),
1402 	  .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1403 
1404 	/* Any USB4 compliant host */
1405 	{ PCI_DEVICE_CLASS(PCI_CLASS_SERIAL_USB_USB4, ~0) },
1406 
1407 	{ 0,}
1408 };
1409 
1410 MODULE_DEVICE_TABLE(pci, nhi_ids);
1411 MODULE_LICENSE("GPL");
1412 
1413 static struct pci_driver nhi_driver = {
1414 	.name = "thunderbolt",
1415 	.id_table = nhi_ids,
1416 	.probe = nhi_probe,
1417 	.remove = nhi_remove,
1418 	.shutdown = nhi_remove,
1419 	.driver.pm = &nhi_pm_ops,
1420 };
1421 
1422 static int __init nhi_init(void)
1423 {
1424 	int ret;
1425 
1426 	ret = tb_domain_init();
1427 	if (ret)
1428 		return ret;
1429 	ret = pci_register_driver(&nhi_driver);
1430 	if (ret)
1431 		tb_domain_exit();
1432 	return ret;
1433 }
1434 
1435 static void __exit nhi_unload(void)
1436 {
1437 	pci_unregister_driver(&nhi_driver);
1438 	tb_domain_exit();
1439 }
1440 
1441 rootfs_initcall(nhi_init);
1442 module_exit(nhi_unload);
1443