xref: /openbmc/linux/drivers/thunderbolt/nhi.c (revision d7b31359)
1 /*
2  * Thunderbolt Cactus Ridge driver - NHI driver
3  *
4  * The NHI (native host interface) is the pci device that allows us to send and
5  * receive frames from the thunderbolt bus.
6  *
7  * Copyright (c) 2014 Andreas Noever <andreas.noever@gmail.com>
8  */
9 
10 #include <linux/pm_runtime.h>
11 #include <linux/slab.h>
12 #include <linux/errno.h>
13 #include <linux/pci.h>
14 #include <linux/interrupt.h>
15 #include <linux/module.h>
16 #include <linux/delay.h>
17 
18 #include "nhi.h"
19 #include "nhi_regs.h"
20 #include "tb.h"
21 
22 #define RING_TYPE(ring) ((ring)->is_tx ? "TX ring" : "RX ring")
23 
24 /*
25  * Used to enable end-to-end workaround for missing RX packets. Do not
26  * use this ring for anything else.
27  */
28 #define RING_E2E_UNUSED_HOPID	2
29 /* HopIDs 0-7 are reserved by the Thunderbolt protocol */
30 #define RING_FIRST_USABLE_HOPID	8
31 
32 /*
33  * Minimal number of vectors when we use MSI-X. Two for control channel
34  * Rx/Tx and the rest four are for cross domain DMA paths.
35  */
36 #define MSIX_MIN_VECS		6
37 #define MSIX_MAX_VECS		16
38 
39 #define NHI_MAILBOX_TIMEOUT	500 /* ms */
40 
41 static int ring_interrupt_index(struct tb_ring *ring)
42 {
43 	int bit = ring->hop;
44 	if (!ring->is_tx)
45 		bit += ring->nhi->hop_count;
46 	return bit;
47 }
48 
49 /**
50  * ring_interrupt_active() - activate/deactivate interrupts for a single ring
51  *
52  * ring->nhi->lock must be held.
53  */
54 static void ring_interrupt_active(struct tb_ring *ring, bool active)
55 {
56 	int reg = REG_RING_INTERRUPT_BASE +
57 		  ring_interrupt_index(ring) / 32 * 4;
58 	int bit = ring_interrupt_index(ring) & 31;
59 	int mask = 1 << bit;
60 	u32 old, new;
61 
62 	if (ring->irq > 0) {
63 		u32 step, shift, ivr, misc;
64 		void __iomem *ivr_base;
65 		int index;
66 
67 		if (ring->is_tx)
68 			index = ring->hop;
69 		else
70 			index = ring->hop + ring->nhi->hop_count;
71 
72 		/*
73 		 * Ask the hardware to clear interrupt status bits automatically
74 		 * since we already know which interrupt was triggered.
75 		 */
76 		misc = ioread32(ring->nhi->iobase + REG_DMA_MISC);
77 		if (!(misc & REG_DMA_MISC_INT_AUTO_CLEAR)) {
78 			misc |= REG_DMA_MISC_INT_AUTO_CLEAR;
79 			iowrite32(misc, ring->nhi->iobase + REG_DMA_MISC);
80 		}
81 
82 		ivr_base = ring->nhi->iobase + REG_INT_VEC_ALLOC_BASE;
83 		step = index / REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
84 		shift = index % REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
85 		ivr = ioread32(ivr_base + step);
86 		ivr &= ~(REG_INT_VEC_ALLOC_MASK << shift);
87 		if (active)
88 			ivr |= ring->vector << shift;
89 		iowrite32(ivr, ivr_base + step);
90 	}
91 
92 	old = ioread32(ring->nhi->iobase + reg);
93 	if (active)
94 		new = old | mask;
95 	else
96 		new = old & ~mask;
97 
98 	dev_info(&ring->nhi->pdev->dev,
99 		 "%s interrupt at register %#x bit %d (%#x -> %#x)\n",
100 		 active ? "enabling" : "disabling", reg, bit, old, new);
101 
102 	if (new == old)
103 		dev_WARN(&ring->nhi->pdev->dev,
104 					 "interrupt for %s %d is already %s\n",
105 					 RING_TYPE(ring), ring->hop,
106 					 active ? "enabled" : "disabled");
107 	iowrite32(new, ring->nhi->iobase + reg);
108 }
109 
110 /**
111  * nhi_disable_interrupts() - disable interrupts for all rings
112  *
113  * Use only during init and shutdown.
114  */
115 static void nhi_disable_interrupts(struct tb_nhi *nhi)
116 {
117 	int i = 0;
118 	/* disable interrupts */
119 	for (i = 0; i < RING_INTERRUPT_REG_COUNT(nhi); i++)
120 		iowrite32(0, nhi->iobase + REG_RING_INTERRUPT_BASE + 4 * i);
121 
122 	/* clear interrupt status bits */
123 	for (i = 0; i < RING_NOTIFY_REG_COUNT(nhi); i++)
124 		ioread32(nhi->iobase + REG_RING_NOTIFY_BASE + 4 * i);
125 }
126 
127 /* ring helper methods */
128 
129 static void __iomem *ring_desc_base(struct tb_ring *ring)
130 {
131 	void __iomem *io = ring->nhi->iobase;
132 	io += ring->is_tx ? REG_TX_RING_BASE : REG_RX_RING_BASE;
133 	io += ring->hop * 16;
134 	return io;
135 }
136 
137 static void __iomem *ring_options_base(struct tb_ring *ring)
138 {
139 	void __iomem *io = ring->nhi->iobase;
140 	io += ring->is_tx ? REG_TX_OPTIONS_BASE : REG_RX_OPTIONS_BASE;
141 	io += ring->hop * 32;
142 	return io;
143 }
144 
145 static void ring_iowrite16desc(struct tb_ring *ring, u32 value, u32 offset)
146 {
147 	iowrite16(value, ring_desc_base(ring) + offset);
148 }
149 
150 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)
151 {
152 	iowrite32(value, ring_desc_base(ring) + offset);
153 }
154 
155 static void ring_iowrite64desc(struct tb_ring *ring, u64 value, u32 offset)
156 {
157 	iowrite32(value, ring_desc_base(ring) + offset);
158 	iowrite32(value >> 32, ring_desc_base(ring) + offset + 4);
159 }
160 
161 static void ring_iowrite32options(struct tb_ring *ring, u32 value, u32 offset)
162 {
163 	iowrite32(value, ring_options_base(ring) + offset);
164 }
165 
166 static bool ring_full(struct tb_ring *ring)
167 {
168 	return ((ring->head + 1) % ring->size) == ring->tail;
169 }
170 
171 static bool ring_empty(struct tb_ring *ring)
172 {
173 	return ring->head == ring->tail;
174 }
175 
176 /**
177  * ring_write_descriptors() - post frames from ring->queue to the controller
178  *
179  * ring->lock is held.
180  */
181 static void ring_write_descriptors(struct tb_ring *ring)
182 {
183 	struct ring_frame *frame, *n;
184 	struct ring_desc *descriptor;
185 	list_for_each_entry_safe(frame, n, &ring->queue, list) {
186 		if (ring_full(ring))
187 			break;
188 		list_move_tail(&frame->list, &ring->in_flight);
189 		descriptor = &ring->descriptors[ring->head];
190 		descriptor->phys = frame->buffer_phy;
191 		descriptor->time = 0;
192 		descriptor->flags = RING_DESC_POSTED | RING_DESC_INTERRUPT;
193 		if (ring->is_tx) {
194 			descriptor->length = frame->size;
195 			descriptor->eof = frame->eof;
196 			descriptor->sof = frame->sof;
197 		}
198 		ring->head = (ring->head + 1) % ring->size;
199 		ring_iowrite16desc(ring, ring->head, ring->is_tx ? 10 : 8);
200 	}
201 }
202 
203 /**
204  * ring_work() - progress completed frames
205  *
206  * If the ring is shutting down then all frames are marked as canceled and
207  * their callbacks are invoked.
208  *
209  * Otherwise we collect all completed frame from the ring buffer, write new
210  * frame to the ring buffer and invoke the callbacks for the completed frames.
211  */
212 static void ring_work(struct work_struct *work)
213 {
214 	struct tb_ring *ring = container_of(work, typeof(*ring), work);
215 	struct ring_frame *frame;
216 	bool canceled = false;
217 	unsigned long flags;
218 	LIST_HEAD(done);
219 
220 	spin_lock_irqsave(&ring->lock, flags);
221 
222 	if (!ring->running) {
223 		/*  Move all frames to done and mark them as canceled. */
224 		list_splice_tail_init(&ring->in_flight, &done);
225 		list_splice_tail_init(&ring->queue, &done);
226 		canceled = true;
227 		goto invoke_callback;
228 	}
229 
230 	while (!ring_empty(ring)) {
231 		if (!(ring->descriptors[ring->tail].flags
232 				& RING_DESC_COMPLETED))
233 			break;
234 		frame = list_first_entry(&ring->in_flight, typeof(*frame),
235 					 list);
236 		list_move_tail(&frame->list, &done);
237 		if (!ring->is_tx) {
238 			frame->size = ring->descriptors[ring->tail].length;
239 			frame->eof = ring->descriptors[ring->tail].eof;
240 			frame->sof = ring->descriptors[ring->tail].sof;
241 			frame->flags = ring->descriptors[ring->tail].flags;
242 		}
243 		ring->tail = (ring->tail + 1) % ring->size;
244 	}
245 	ring_write_descriptors(ring);
246 
247 invoke_callback:
248 	/* allow callbacks to schedule new work */
249 	spin_unlock_irqrestore(&ring->lock, flags);
250 	while (!list_empty(&done)) {
251 		frame = list_first_entry(&done, typeof(*frame), list);
252 		/*
253 		 * The callback may reenqueue or delete frame.
254 		 * Do not hold on to it.
255 		 */
256 		list_del_init(&frame->list);
257 		if (frame->callback)
258 			frame->callback(ring, frame, canceled);
259 	}
260 }
261 
262 int __tb_ring_enqueue(struct tb_ring *ring, struct ring_frame *frame)
263 {
264 	unsigned long flags;
265 	int ret = 0;
266 
267 	spin_lock_irqsave(&ring->lock, flags);
268 	if (ring->running) {
269 		list_add_tail(&frame->list, &ring->queue);
270 		ring_write_descriptors(ring);
271 	} else {
272 		ret = -ESHUTDOWN;
273 	}
274 	spin_unlock_irqrestore(&ring->lock, flags);
275 	return ret;
276 }
277 EXPORT_SYMBOL_GPL(__tb_ring_enqueue);
278 
279 /**
280  * tb_ring_poll() - Poll one completed frame from the ring
281  * @ring: Ring to poll
282  *
283  * This function can be called when @start_poll callback of the @ring
284  * has been called. It will read one completed frame from the ring and
285  * return it to the caller. Returns %NULL if there is no more completed
286  * frames.
287  */
288 struct ring_frame *tb_ring_poll(struct tb_ring *ring)
289 {
290 	struct ring_frame *frame = NULL;
291 	unsigned long flags;
292 
293 	spin_lock_irqsave(&ring->lock, flags);
294 	if (!ring->running)
295 		goto unlock;
296 	if (ring_empty(ring))
297 		goto unlock;
298 
299 	if (ring->descriptors[ring->tail].flags & RING_DESC_COMPLETED) {
300 		frame = list_first_entry(&ring->in_flight, typeof(*frame),
301 					 list);
302 		list_del_init(&frame->list);
303 
304 		if (!ring->is_tx) {
305 			frame->size = ring->descriptors[ring->tail].length;
306 			frame->eof = ring->descriptors[ring->tail].eof;
307 			frame->sof = ring->descriptors[ring->tail].sof;
308 			frame->flags = ring->descriptors[ring->tail].flags;
309 		}
310 
311 		ring->tail = (ring->tail + 1) % ring->size;
312 	}
313 
314 unlock:
315 	spin_unlock_irqrestore(&ring->lock, flags);
316 	return frame;
317 }
318 EXPORT_SYMBOL_GPL(tb_ring_poll);
319 
320 static void __ring_interrupt_mask(struct tb_ring *ring, bool mask)
321 {
322 	int idx = ring_interrupt_index(ring);
323 	int reg = REG_RING_INTERRUPT_BASE + idx / 32 * 4;
324 	int bit = idx % 32;
325 	u32 val;
326 
327 	val = ioread32(ring->nhi->iobase + reg);
328 	if (mask)
329 		val &= ~BIT(bit);
330 	else
331 		val |= BIT(bit);
332 	iowrite32(val, ring->nhi->iobase + reg);
333 }
334 
335 /* Both @nhi->lock and @ring->lock should be held */
336 static void __ring_interrupt(struct tb_ring *ring)
337 {
338 	if (!ring->running)
339 		return;
340 
341 	if (ring->start_poll) {
342 		__ring_interrupt_mask(ring, true);
343 		ring->start_poll(ring->poll_data);
344 	} else {
345 		schedule_work(&ring->work);
346 	}
347 }
348 
349 /**
350  * tb_ring_poll_complete() - Re-start interrupt for the ring
351  * @ring: Ring to re-start the interrupt
352  *
353  * This will re-start (unmask) the ring interrupt once the user is done
354  * with polling.
355  */
356 void tb_ring_poll_complete(struct tb_ring *ring)
357 {
358 	unsigned long flags;
359 
360 	spin_lock_irqsave(&ring->nhi->lock, flags);
361 	spin_lock(&ring->lock);
362 	if (ring->start_poll)
363 		__ring_interrupt_mask(ring, false);
364 	spin_unlock(&ring->lock);
365 	spin_unlock_irqrestore(&ring->nhi->lock, flags);
366 }
367 EXPORT_SYMBOL_GPL(tb_ring_poll_complete);
368 
369 static irqreturn_t ring_msix(int irq, void *data)
370 {
371 	struct tb_ring *ring = data;
372 
373 	spin_lock(&ring->nhi->lock);
374 	spin_lock(&ring->lock);
375 	__ring_interrupt(ring);
376 	spin_unlock(&ring->lock);
377 	spin_unlock(&ring->nhi->lock);
378 
379 	return IRQ_HANDLED;
380 }
381 
382 static int ring_request_msix(struct tb_ring *ring, bool no_suspend)
383 {
384 	struct tb_nhi *nhi = ring->nhi;
385 	unsigned long irqflags;
386 	int ret;
387 
388 	if (!nhi->pdev->msix_enabled)
389 		return 0;
390 
391 	ret = ida_simple_get(&nhi->msix_ida, 0, MSIX_MAX_VECS, GFP_KERNEL);
392 	if (ret < 0)
393 		return ret;
394 
395 	ring->vector = ret;
396 
397 	ring->irq = pci_irq_vector(ring->nhi->pdev, ring->vector);
398 	if (ring->irq < 0)
399 		return ring->irq;
400 
401 	irqflags = no_suspend ? IRQF_NO_SUSPEND : 0;
402 	return request_irq(ring->irq, ring_msix, irqflags, "thunderbolt", ring);
403 }
404 
405 static void ring_release_msix(struct tb_ring *ring)
406 {
407 	if (ring->irq <= 0)
408 		return;
409 
410 	free_irq(ring->irq, ring);
411 	ida_simple_remove(&ring->nhi->msix_ida, ring->vector);
412 	ring->vector = 0;
413 	ring->irq = 0;
414 }
415 
416 static int nhi_alloc_hop(struct tb_nhi *nhi, struct tb_ring *ring)
417 {
418 	int ret = 0;
419 
420 	spin_lock_irq(&nhi->lock);
421 
422 	if (ring->hop < 0) {
423 		unsigned int i;
424 
425 		/*
426 		 * Automatically allocate HopID from the non-reserved
427 		 * range 8 .. hop_count - 1.
428 		 */
429 		for (i = RING_FIRST_USABLE_HOPID; i < nhi->hop_count; i++) {
430 			if (ring->is_tx) {
431 				if (!nhi->tx_rings[i]) {
432 					ring->hop = i;
433 					break;
434 				}
435 			} else {
436 				if (!nhi->rx_rings[i]) {
437 					ring->hop = i;
438 					break;
439 				}
440 			}
441 		}
442 	}
443 
444 	if (ring->hop < 0 || ring->hop >= nhi->hop_count) {
445 		dev_warn(&nhi->pdev->dev, "invalid hop: %d\n", ring->hop);
446 		ret = -EINVAL;
447 		goto err_unlock;
448 	}
449 	if (ring->is_tx && nhi->tx_rings[ring->hop]) {
450 		dev_warn(&nhi->pdev->dev, "TX hop %d already allocated\n",
451 			 ring->hop);
452 		ret = -EBUSY;
453 		goto err_unlock;
454 	} else if (!ring->is_tx && nhi->rx_rings[ring->hop]) {
455 		dev_warn(&nhi->pdev->dev, "RX hop %d already allocated\n",
456 			 ring->hop);
457 		ret = -EBUSY;
458 		goto err_unlock;
459 	}
460 
461 	if (ring->is_tx)
462 		nhi->tx_rings[ring->hop] = ring;
463 	else
464 		nhi->rx_rings[ring->hop] = ring;
465 
466 err_unlock:
467 	spin_unlock_irq(&nhi->lock);
468 
469 	return ret;
470 }
471 
472 static struct tb_ring *tb_ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
473 				     bool transmit, unsigned int flags,
474 				     u16 sof_mask, u16 eof_mask,
475 				     void (*start_poll)(void *),
476 				     void *poll_data)
477 {
478 	struct tb_ring *ring = NULL;
479 	dev_info(&nhi->pdev->dev, "allocating %s ring %d of size %d\n",
480 		 transmit ? "TX" : "RX", hop, size);
481 
482 	/* Tx Ring 2 is reserved for E2E workaround */
483 	if (transmit && hop == RING_E2E_UNUSED_HOPID)
484 		return NULL;
485 
486 	ring = kzalloc(sizeof(*ring), GFP_KERNEL);
487 	if (!ring)
488 		return NULL;
489 
490 	spin_lock_init(&ring->lock);
491 	INIT_LIST_HEAD(&ring->queue);
492 	INIT_LIST_HEAD(&ring->in_flight);
493 	INIT_WORK(&ring->work, ring_work);
494 
495 	ring->nhi = nhi;
496 	ring->hop = hop;
497 	ring->is_tx = transmit;
498 	ring->size = size;
499 	ring->flags = flags;
500 	ring->sof_mask = sof_mask;
501 	ring->eof_mask = eof_mask;
502 	ring->head = 0;
503 	ring->tail = 0;
504 	ring->running = false;
505 	ring->start_poll = start_poll;
506 	ring->poll_data = poll_data;
507 
508 	ring->descriptors = dma_alloc_coherent(&ring->nhi->pdev->dev,
509 			size * sizeof(*ring->descriptors),
510 			&ring->descriptors_dma, GFP_KERNEL | __GFP_ZERO);
511 	if (!ring->descriptors)
512 		goto err_free_ring;
513 
514 	if (ring_request_msix(ring, flags & RING_FLAG_NO_SUSPEND))
515 		goto err_free_descs;
516 
517 	if (nhi_alloc_hop(nhi, ring))
518 		goto err_release_msix;
519 
520 	return ring;
521 
522 err_release_msix:
523 	ring_release_msix(ring);
524 err_free_descs:
525 	dma_free_coherent(&ring->nhi->pdev->dev,
526 			  ring->size * sizeof(*ring->descriptors),
527 			  ring->descriptors, ring->descriptors_dma);
528 err_free_ring:
529 	kfree(ring);
530 
531 	return NULL;
532 }
533 
534 /**
535  * tb_ring_alloc_tx() - Allocate DMA ring for transmit
536  * @nhi: Pointer to the NHI the ring is to be allocated
537  * @hop: HopID (ring) to allocate
538  * @size: Number of entries in the ring
539  * @flags: Flags for the ring
540  */
541 struct tb_ring *tb_ring_alloc_tx(struct tb_nhi *nhi, int hop, int size,
542 				 unsigned int flags)
543 {
544 	return tb_ring_alloc(nhi, hop, size, true, flags, 0, 0, NULL, NULL);
545 }
546 EXPORT_SYMBOL_GPL(tb_ring_alloc_tx);
547 
548 /**
549  * tb_ring_alloc_rx() - Allocate DMA ring for receive
550  * @nhi: Pointer to the NHI the ring is to be allocated
551  * @hop: HopID (ring) to allocate. Pass %-1 for automatic allocation.
552  * @size: Number of entries in the ring
553  * @flags: Flags for the ring
554  * @sof_mask: Mask of PDF values that start a frame
555  * @eof_mask: Mask of PDF values that end a frame
556  * @start_poll: If not %NULL the ring will call this function when an
557  *		interrupt is triggered and masked, instead of callback
558  *		in each Rx frame.
559  * @poll_data: Optional data passed to @start_poll
560  */
561 struct tb_ring *tb_ring_alloc_rx(struct tb_nhi *nhi, int hop, int size,
562 				 unsigned int flags, u16 sof_mask, u16 eof_mask,
563 				 void (*start_poll)(void *), void *poll_data)
564 {
565 	return tb_ring_alloc(nhi, hop, size, false, flags, sof_mask, eof_mask,
566 			     start_poll, poll_data);
567 }
568 EXPORT_SYMBOL_GPL(tb_ring_alloc_rx);
569 
570 /**
571  * tb_ring_start() - enable a ring
572  *
573  * Must not be invoked in parallel with tb_ring_stop().
574  */
575 void tb_ring_start(struct tb_ring *ring)
576 {
577 	u16 frame_size;
578 	u32 flags;
579 
580 	spin_lock_irq(&ring->nhi->lock);
581 	spin_lock(&ring->lock);
582 	if (ring->nhi->going_away)
583 		goto err;
584 	if (ring->running) {
585 		dev_WARN(&ring->nhi->pdev->dev, "ring already started\n");
586 		goto err;
587 	}
588 	dev_info(&ring->nhi->pdev->dev, "starting %s %d\n",
589 		 RING_TYPE(ring), ring->hop);
590 
591 	if (ring->flags & RING_FLAG_FRAME) {
592 		/* Means 4096 */
593 		frame_size = 0;
594 		flags = RING_FLAG_ENABLE;
595 	} else {
596 		frame_size = TB_FRAME_SIZE;
597 		flags = RING_FLAG_ENABLE | RING_FLAG_RAW;
598 	}
599 
600 	if (ring->flags & RING_FLAG_E2E && !ring->is_tx) {
601 		u32 hop;
602 
603 		/*
604 		 * In order not to lose Rx packets we enable end-to-end
605 		 * workaround which transfers Rx credits to an unused Tx
606 		 * HopID.
607 		 */
608 		hop = RING_E2E_UNUSED_HOPID << REG_RX_OPTIONS_E2E_HOP_SHIFT;
609 		hop &= REG_RX_OPTIONS_E2E_HOP_MASK;
610 		flags |= hop | RING_FLAG_E2E_FLOW_CONTROL;
611 	}
612 
613 	ring_iowrite64desc(ring, ring->descriptors_dma, 0);
614 	if (ring->is_tx) {
615 		ring_iowrite32desc(ring, ring->size, 12);
616 		ring_iowrite32options(ring, 0, 4); /* time releated ? */
617 		ring_iowrite32options(ring, flags, 0);
618 	} else {
619 		u32 sof_eof_mask = ring->sof_mask << 16 | ring->eof_mask;
620 
621 		ring_iowrite32desc(ring, (frame_size << 16) | ring->size, 12);
622 		ring_iowrite32options(ring, sof_eof_mask, 4);
623 		ring_iowrite32options(ring, flags, 0);
624 	}
625 	ring_interrupt_active(ring, true);
626 	ring->running = true;
627 err:
628 	spin_unlock(&ring->lock);
629 	spin_unlock_irq(&ring->nhi->lock);
630 }
631 EXPORT_SYMBOL_GPL(tb_ring_start);
632 
633 /**
634  * tb_ring_stop() - shutdown a ring
635  *
636  * Must not be invoked from a callback.
637  *
638  * This method will disable the ring. Further calls to
639  * tb_ring_tx/tb_ring_rx will return -ESHUTDOWN until ring_stop has been
640  * called.
641  *
642  * All enqueued frames will be canceled and their callbacks will be executed
643  * with frame->canceled set to true (on the callback thread). This method
644  * returns only after all callback invocations have finished.
645  */
646 void tb_ring_stop(struct tb_ring *ring)
647 {
648 	spin_lock_irq(&ring->nhi->lock);
649 	spin_lock(&ring->lock);
650 	dev_info(&ring->nhi->pdev->dev, "stopping %s %d\n",
651 		 RING_TYPE(ring), ring->hop);
652 	if (ring->nhi->going_away)
653 		goto err;
654 	if (!ring->running) {
655 		dev_WARN(&ring->nhi->pdev->dev, "%s %d already stopped\n",
656 			 RING_TYPE(ring), ring->hop);
657 		goto err;
658 	}
659 	ring_interrupt_active(ring, false);
660 
661 	ring_iowrite32options(ring, 0, 0);
662 	ring_iowrite64desc(ring, 0, 0);
663 	ring_iowrite16desc(ring, 0, ring->is_tx ? 10 : 8);
664 	ring_iowrite32desc(ring, 0, 12);
665 	ring->head = 0;
666 	ring->tail = 0;
667 	ring->running = false;
668 
669 err:
670 	spin_unlock(&ring->lock);
671 	spin_unlock_irq(&ring->nhi->lock);
672 
673 	/*
674 	 * schedule ring->work to invoke callbacks on all remaining frames.
675 	 */
676 	schedule_work(&ring->work);
677 	flush_work(&ring->work);
678 }
679 EXPORT_SYMBOL_GPL(tb_ring_stop);
680 
681 /*
682  * tb_ring_free() - free ring
683  *
684  * When this method returns all invocations of ring->callback will have
685  * finished.
686  *
687  * Ring must be stopped.
688  *
689  * Must NOT be called from ring_frame->callback!
690  */
691 void tb_ring_free(struct tb_ring *ring)
692 {
693 	spin_lock_irq(&ring->nhi->lock);
694 	/*
695 	 * Dissociate the ring from the NHI. This also ensures that
696 	 * nhi_interrupt_work cannot reschedule ring->work.
697 	 */
698 	if (ring->is_tx)
699 		ring->nhi->tx_rings[ring->hop] = NULL;
700 	else
701 		ring->nhi->rx_rings[ring->hop] = NULL;
702 
703 	if (ring->running) {
704 		dev_WARN(&ring->nhi->pdev->dev, "%s %d still running\n",
705 			 RING_TYPE(ring), ring->hop);
706 	}
707 	spin_unlock_irq(&ring->nhi->lock);
708 
709 	ring_release_msix(ring);
710 
711 	dma_free_coherent(&ring->nhi->pdev->dev,
712 			  ring->size * sizeof(*ring->descriptors),
713 			  ring->descriptors, ring->descriptors_dma);
714 
715 	ring->descriptors = NULL;
716 	ring->descriptors_dma = 0;
717 
718 
719 	dev_info(&ring->nhi->pdev->dev,
720 		 "freeing %s %d\n",
721 		 RING_TYPE(ring),
722 		 ring->hop);
723 
724 	/**
725 	 * ring->work can no longer be scheduled (it is scheduled only
726 	 * by nhi_interrupt_work, ring_stop and ring_msix). Wait for it
727 	 * to finish before freeing the ring.
728 	 */
729 	flush_work(&ring->work);
730 	kfree(ring);
731 }
732 EXPORT_SYMBOL_GPL(tb_ring_free);
733 
734 /**
735  * nhi_mailbox_cmd() - Send a command through NHI mailbox
736  * @nhi: Pointer to the NHI structure
737  * @cmd: Command to send
738  * @data: Data to be send with the command
739  *
740  * Sends mailbox command to the firmware running on NHI. Returns %0 in
741  * case of success and negative errno in case of failure.
742  */
743 int nhi_mailbox_cmd(struct tb_nhi *nhi, enum nhi_mailbox_cmd cmd, u32 data)
744 {
745 	ktime_t timeout;
746 	u32 val;
747 
748 	iowrite32(data, nhi->iobase + REG_INMAIL_DATA);
749 
750 	val = ioread32(nhi->iobase + REG_INMAIL_CMD);
751 	val &= ~(REG_INMAIL_CMD_MASK | REG_INMAIL_ERROR);
752 	val |= REG_INMAIL_OP_REQUEST | cmd;
753 	iowrite32(val, nhi->iobase + REG_INMAIL_CMD);
754 
755 	timeout = ktime_add_ms(ktime_get(), NHI_MAILBOX_TIMEOUT);
756 	do {
757 		val = ioread32(nhi->iobase + REG_INMAIL_CMD);
758 		if (!(val & REG_INMAIL_OP_REQUEST))
759 			break;
760 		usleep_range(10, 20);
761 	} while (ktime_before(ktime_get(), timeout));
762 
763 	if (val & REG_INMAIL_OP_REQUEST)
764 		return -ETIMEDOUT;
765 	if (val & REG_INMAIL_ERROR)
766 		return -EIO;
767 
768 	return 0;
769 }
770 
771 /**
772  * nhi_mailbox_mode() - Return current firmware operation mode
773  * @nhi: Pointer to the NHI structure
774  *
775  * The function reads current firmware operation mode using NHI mailbox
776  * registers and returns it to the caller.
777  */
778 enum nhi_fw_mode nhi_mailbox_mode(struct tb_nhi *nhi)
779 {
780 	u32 val;
781 
782 	val = ioread32(nhi->iobase + REG_OUTMAIL_CMD);
783 	val &= REG_OUTMAIL_CMD_OPMODE_MASK;
784 	val >>= REG_OUTMAIL_CMD_OPMODE_SHIFT;
785 
786 	return (enum nhi_fw_mode)val;
787 }
788 
789 static void nhi_interrupt_work(struct work_struct *work)
790 {
791 	struct tb_nhi *nhi = container_of(work, typeof(*nhi), interrupt_work);
792 	int value = 0; /* Suppress uninitialized usage warning. */
793 	int bit;
794 	int hop = -1;
795 	int type = 0; /* current interrupt type 0: TX, 1: RX, 2: RX overflow */
796 	struct tb_ring *ring;
797 
798 	spin_lock_irq(&nhi->lock);
799 
800 	/*
801 	 * Starting at REG_RING_NOTIFY_BASE there are three status bitfields
802 	 * (TX, RX, RX overflow). We iterate over the bits and read a new
803 	 * dwords as required. The registers are cleared on read.
804 	 */
805 	for (bit = 0; bit < 3 * nhi->hop_count; bit++) {
806 		if (bit % 32 == 0)
807 			value = ioread32(nhi->iobase
808 					 + REG_RING_NOTIFY_BASE
809 					 + 4 * (bit / 32));
810 		if (++hop == nhi->hop_count) {
811 			hop = 0;
812 			type++;
813 		}
814 		if ((value & (1 << (bit % 32))) == 0)
815 			continue;
816 		if (type == 2) {
817 			dev_warn(&nhi->pdev->dev,
818 				 "RX overflow for ring %d\n",
819 				 hop);
820 			continue;
821 		}
822 		if (type == 0)
823 			ring = nhi->tx_rings[hop];
824 		else
825 			ring = nhi->rx_rings[hop];
826 		if (ring == NULL) {
827 			dev_warn(&nhi->pdev->dev,
828 				 "got interrupt for inactive %s ring %d\n",
829 				 type ? "RX" : "TX",
830 				 hop);
831 			continue;
832 		}
833 
834 		spin_lock(&ring->lock);
835 		__ring_interrupt(ring);
836 		spin_unlock(&ring->lock);
837 	}
838 	spin_unlock_irq(&nhi->lock);
839 }
840 
841 static irqreturn_t nhi_msi(int irq, void *data)
842 {
843 	struct tb_nhi *nhi = data;
844 	schedule_work(&nhi->interrupt_work);
845 	return IRQ_HANDLED;
846 }
847 
848 static int nhi_suspend_noirq(struct device *dev)
849 {
850 	struct pci_dev *pdev = to_pci_dev(dev);
851 	struct tb *tb = pci_get_drvdata(pdev);
852 
853 	return tb_domain_suspend_noirq(tb);
854 }
855 
856 static void nhi_enable_int_throttling(struct tb_nhi *nhi)
857 {
858 	/* Throttling is specified in 256ns increments */
859 	u32 throttle = DIV_ROUND_UP(128 * NSEC_PER_USEC, 256);
860 	unsigned int i;
861 
862 	/*
863 	 * Configure interrupt throttling for all vectors even if we
864 	 * only use few.
865 	 */
866 	for (i = 0; i < MSIX_MAX_VECS; i++) {
867 		u32 reg = REG_INT_THROTTLING_RATE + i * 4;
868 		iowrite32(throttle, nhi->iobase + reg);
869 	}
870 }
871 
872 static int nhi_resume_noirq(struct device *dev)
873 {
874 	struct pci_dev *pdev = to_pci_dev(dev);
875 	struct tb *tb = pci_get_drvdata(pdev);
876 
877 	/*
878 	 * Check that the device is still there. It may be that the user
879 	 * unplugged last device which causes the host controller to go
880 	 * away on PCs.
881 	 */
882 	if (!pci_device_is_present(pdev))
883 		tb->nhi->going_away = true;
884 	else
885 		nhi_enable_int_throttling(tb->nhi);
886 
887 	return tb_domain_resume_noirq(tb);
888 }
889 
890 static int nhi_suspend(struct device *dev)
891 {
892 	struct pci_dev *pdev = to_pci_dev(dev);
893 	struct tb *tb = pci_get_drvdata(pdev);
894 
895 	return tb_domain_suspend(tb);
896 }
897 
898 static void nhi_complete(struct device *dev)
899 {
900 	struct pci_dev *pdev = to_pci_dev(dev);
901 	struct tb *tb = pci_get_drvdata(pdev);
902 
903 	/*
904 	 * If we were runtime suspended when system suspend started,
905 	 * schedule runtime resume now. It should bring the domain back
906 	 * to functional state.
907 	 */
908 	if (pm_runtime_suspended(&pdev->dev))
909 		pm_runtime_resume(&pdev->dev);
910 	else
911 		tb_domain_complete(tb);
912 }
913 
914 static int nhi_runtime_suspend(struct device *dev)
915 {
916 	struct pci_dev *pdev = to_pci_dev(dev);
917 	struct tb *tb = pci_get_drvdata(pdev);
918 
919 	return tb_domain_runtime_suspend(tb);
920 }
921 
922 static int nhi_runtime_resume(struct device *dev)
923 {
924 	struct pci_dev *pdev = to_pci_dev(dev);
925 	struct tb *tb = pci_get_drvdata(pdev);
926 
927 	nhi_enable_int_throttling(tb->nhi);
928 	return tb_domain_runtime_resume(tb);
929 }
930 
931 static void nhi_shutdown(struct tb_nhi *nhi)
932 {
933 	int i;
934 	dev_info(&nhi->pdev->dev, "shutdown\n");
935 
936 	for (i = 0; i < nhi->hop_count; i++) {
937 		if (nhi->tx_rings[i])
938 			dev_WARN(&nhi->pdev->dev,
939 				 "TX ring %d is still active\n", i);
940 		if (nhi->rx_rings[i])
941 			dev_WARN(&nhi->pdev->dev,
942 				 "RX ring %d is still active\n", i);
943 	}
944 	nhi_disable_interrupts(nhi);
945 	/*
946 	 * We have to release the irq before calling flush_work. Otherwise an
947 	 * already executing IRQ handler could call schedule_work again.
948 	 */
949 	if (!nhi->pdev->msix_enabled) {
950 		devm_free_irq(&nhi->pdev->dev, nhi->pdev->irq, nhi);
951 		flush_work(&nhi->interrupt_work);
952 	}
953 	ida_destroy(&nhi->msix_ida);
954 }
955 
956 static int nhi_init_msi(struct tb_nhi *nhi)
957 {
958 	struct pci_dev *pdev = nhi->pdev;
959 	int res, irq, nvec;
960 
961 	/* In case someone left them on. */
962 	nhi_disable_interrupts(nhi);
963 
964 	nhi_enable_int_throttling(nhi);
965 
966 	ida_init(&nhi->msix_ida);
967 
968 	/*
969 	 * The NHI has 16 MSI-X vectors or a single MSI. We first try to
970 	 * get all MSI-X vectors and if we succeed, each ring will have
971 	 * one MSI-X. If for some reason that does not work out, we
972 	 * fallback to a single MSI.
973 	 */
974 	nvec = pci_alloc_irq_vectors(pdev, MSIX_MIN_VECS, MSIX_MAX_VECS,
975 				     PCI_IRQ_MSIX);
976 	if (nvec < 0) {
977 		nvec = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI);
978 		if (nvec < 0)
979 			return nvec;
980 
981 		INIT_WORK(&nhi->interrupt_work, nhi_interrupt_work);
982 
983 		irq = pci_irq_vector(nhi->pdev, 0);
984 		if (irq < 0)
985 			return irq;
986 
987 		res = devm_request_irq(&pdev->dev, irq, nhi_msi,
988 				       IRQF_NO_SUSPEND, "thunderbolt", nhi);
989 		if (res) {
990 			dev_err(&pdev->dev, "request_irq failed, aborting\n");
991 			return res;
992 		}
993 	}
994 
995 	return 0;
996 }
997 
998 static int nhi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
999 {
1000 	struct tb_nhi *nhi;
1001 	struct tb *tb;
1002 	int res;
1003 
1004 	res = pcim_enable_device(pdev);
1005 	if (res) {
1006 		dev_err(&pdev->dev, "cannot enable PCI device, aborting\n");
1007 		return res;
1008 	}
1009 
1010 	res = pcim_iomap_regions(pdev, 1 << 0, "thunderbolt");
1011 	if (res) {
1012 		dev_err(&pdev->dev, "cannot obtain PCI resources, aborting\n");
1013 		return res;
1014 	}
1015 
1016 	nhi = devm_kzalloc(&pdev->dev, sizeof(*nhi), GFP_KERNEL);
1017 	if (!nhi)
1018 		return -ENOMEM;
1019 
1020 	nhi->pdev = pdev;
1021 	/* cannot fail - table is allocated bin pcim_iomap_regions */
1022 	nhi->iobase = pcim_iomap_table(pdev)[0];
1023 	nhi->hop_count = ioread32(nhi->iobase + REG_HOP_COUNT) & 0x3ff;
1024 	if (nhi->hop_count != 12 && nhi->hop_count != 32)
1025 		dev_warn(&pdev->dev, "unexpected hop count: %d\n",
1026 			 nhi->hop_count);
1027 
1028 	nhi->tx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1029 				     sizeof(*nhi->tx_rings), GFP_KERNEL);
1030 	nhi->rx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1031 				     sizeof(*nhi->rx_rings), GFP_KERNEL);
1032 	if (!nhi->tx_rings || !nhi->rx_rings)
1033 		return -ENOMEM;
1034 
1035 	res = nhi_init_msi(nhi);
1036 	if (res) {
1037 		dev_err(&pdev->dev, "cannot enable MSI, aborting\n");
1038 		return res;
1039 	}
1040 
1041 	spin_lock_init(&nhi->lock);
1042 
1043 	res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1044 	if (res)
1045 		res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
1046 	if (res) {
1047 		dev_err(&pdev->dev, "failed to set DMA mask\n");
1048 		return res;
1049 	}
1050 
1051 	pci_set_master(pdev);
1052 
1053 	tb = icm_probe(nhi);
1054 	if (!tb)
1055 		tb = tb_probe(nhi);
1056 	if (!tb) {
1057 		dev_err(&nhi->pdev->dev,
1058 			"failed to determine connection manager, aborting\n");
1059 		return -ENODEV;
1060 	}
1061 
1062 	dev_info(&nhi->pdev->dev, "NHI initialized, starting thunderbolt\n");
1063 
1064 	res = tb_domain_add(tb);
1065 	if (res) {
1066 		/*
1067 		 * At this point the RX/TX rings might already have been
1068 		 * activated. Do a proper shutdown.
1069 		 */
1070 		tb_domain_put(tb);
1071 		nhi_shutdown(nhi);
1072 		return res;
1073 	}
1074 	pci_set_drvdata(pdev, tb);
1075 
1076 	pm_runtime_allow(&pdev->dev);
1077 	pm_runtime_set_autosuspend_delay(&pdev->dev, TB_AUTOSUSPEND_DELAY);
1078 	pm_runtime_use_autosuspend(&pdev->dev);
1079 	pm_runtime_put_autosuspend(&pdev->dev);
1080 
1081 	return 0;
1082 }
1083 
1084 static void nhi_remove(struct pci_dev *pdev)
1085 {
1086 	struct tb *tb = pci_get_drvdata(pdev);
1087 	struct tb_nhi *nhi = tb->nhi;
1088 
1089 	pm_runtime_get_sync(&pdev->dev);
1090 	pm_runtime_dont_use_autosuspend(&pdev->dev);
1091 	pm_runtime_forbid(&pdev->dev);
1092 
1093 	tb_domain_remove(tb);
1094 	nhi_shutdown(nhi);
1095 }
1096 
1097 /*
1098  * The tunneled pci bridges are siblings of us. Use resume_noirq to reenable
1099  * the tunnels asap. A corresponding pci quirk blocks the downstream bridges
1100  * resume_noirq until we are done.
1101  */
1102 static const struct dev_pm_ops nhi_pm_ops = {
1103 	.suspend_noirq = nhi_suspend_noirq,
1104 	.resume_noirq = nhi_resume_noirq,
1105 	.freeze_noirq = nhi_suspend_noirq, /*
1106 					    * we just disable hotplug, the
1107 					    * pci-tunnels stay alive.
1108 					    */
1109 	.thaw_noirq = nhi_resume_noirq,
1110 	.restore_noirq = nhi_resume_noirq,
1111 	.suspend = nhi_suspend,
1112 	.freeze = nhi_suspend,
1113 	.poweroff = nhi_suspend,
1114 	.complete = nhi_complete,
1115 	.runtime_suspend = nhi_runtime_suspend,
1116 	.runtime_resume = nhi_runtime_resume,
1117 };
1118 
1119 static struct pci_device_id nhi_ids[] = {
1120 	/*
1121 	 * We have to specify class, the TB bridges use the same device and
1122 	 * vendor (sub)id on gen 1 and gen 2 controllers.
1123 	 */
1124 	{
1125 		.class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1126 		.vendor = PCI_VENDOR_ID_INTEL,
1127 		.device = PCI_DEVICE_ID_INTEL_LIGHT_RIDGE,
1128 		.subvendor = 0x2222, .subdevice = 0x1111,
1129 	},
1130 	{
1131 		.class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1132 		.vendor = PCI_VENDOR_ID_INTEL,
1133 		.device = PCI_DEVICE_ID_INTEL_CACTUS_RIDGE_4C,
1134 		.subvendor = 0x2222, .subdevice = 0x1111,
1135 	},
1136 	{
1137 		.class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1138 		.vendor = PCI_VENDOR_ID_INTEL,
1139 		.device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI,
1140 		.subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1141 	},
1142 	{
1143 		.class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1144 		.vendor = PCI_VENDOR_ID_INTEL,
1145 		.device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI,
1146 		.subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1147 	},
1148 
1149 	/* Thunderbolt 3 */
1150 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_2C_NHI) },
1151 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_4C_NHI) },
1152 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_USBONLY_NHI) },
1153 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_NHI) },
1154 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_USBONLY_NHI) },
1155 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_2C_NHI) },
1156 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_4C_NHI) },
1157 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_USBONLY_NHI) },
1158 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI) },
1159 	{ PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI) },
1160 
1161 	{ 0,}
1162 };
1163 
1164 MODULE_DEVICE_TABLE(pci, nhi_ids);
1165 MODULE_LICENSE("GPL");
1166 
1167 static struct pci_driver nhi_driver = {
1168 	.name = "thunderbolt",
1169 	.id_table = nhi_ids,
1170 	.probe = nhi_probe,
1171 	.remove = nhi_remove,
1172 	.driver.pm = &nhi_pm_ops,
1173 };
1174 
1175 static int __init nhi_init(void)
1176 {
1177 	int ret;
1178 
1179 	ret = tb_domain_init();
1180 	if (ret)
1181 		return ret;
1182 	ret = pci_register_driver(&nhi_driver);
1183 	if (ret)
1184 		tb_domain_exit();
1185 	return ret;
1186 }
1187 
1188 static void __exit nhi_unload(void)
1189 {
1190 	pci_unregister_driver(&nhi_driver);
1191 	tb_domain_exit();
1192 }
1193 
1194 rootfs_initcall(nhi_init);
1195 module_exit(nhi_unload);
1196