Loading...
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
21#include "nhi.h"
22#include "nhi_regs.h"
23#include "tb.h"
24
25#define RING_TYPE(ring) ((ring)->is_tx ? "TX ring" : "RX ring")
26
27#define RING_FIRST_USABLE_HOPID 1
28
29/*
30 * Minimal number of vectors when we use MSI-X. Two for control channel
31 * Rx/Tx and the rest four are for cross domain DMA paths.
32 */
33#define MSIX_MIN_VECS 6
34#define MSIX_MAX_VECS 16
35
36#define NHI_MAILBOX_TIMEOUT 500 /* ms */
37
38static int ring_interrupt_index(struct tb_ring *ring)
39{
40 int bit = ring->hop;
41 if (!ring->is_tx)
42 bit += ring->nhi->hop_count;
43 return bit;
44}
45
46/**
47 * ring_interrupt_active() - activate/deactivate interrupts for a single ring
48 *
49 * ring->nhi->lock must be held.
50 */
51static void ring_interrupt_active(struct tb_ring *ring, bool active)
52{
53 int reg = REG_RING_INTERRUPT_BASE +
54 ring_interrupt_index(ring) / 32 * 4;
55 int bit = ring_interrupt_index(ring) & 31;
56 int mask = 1 << bit;
57 u32 old, new;
58
59 if (ring->irq > 0) {
60 u32 step, shift, ivr, misc;
61 void __iomem *ivr_base;
62 int index;
63
64 if (ring->is_tx)
65 index = ring->hop;
66 else
67 index = ring->hop + ring->nhi->hop_count;
68
69 /*
70 * Ask the hardware to clear interrupt status bits automatically
71 * since we already know which interrupt was triggered.
72 */
73 misc = ioread32(ring->nhi->iobase + REG_DMA_MISC);
74 if (!(misc & REG_DMA_MISC_INT_AUTO_CLEAR)) {
75 misc |= REG_DMA_MISC_INT_AUTO_CLEAR;
76 iowrite32(misc, ring->nhi->iobase + REG_DMA_MISC);
77 }
78
79 ivr_base = ring->nhi->iobase + REG_INT_VEC_ALLOC_BASE;
80 step = index / REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
81 shift = index % REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
82 ivr = ioread32(ivr_base + step);
83 ivr &= ~(REG_INT_VEC_ALLOC_MASK << shift);
84 if (active)
85 ivr |= ring->vector << shift;
86 iowrite32(ivr, ivr_base + step);
87 }
88
89 old = ioread32(ring->nhi->iobase + reg);
90 if (active)
91 new = old | mask;
92 else
93 new = old & ~mask;
94
95 dev_dbg(&ring->nhi->pdev->dev,
96 "%s interrupt at register %#x bit %d (%#x -> %#x)\n",
97 active ? "enabling" : "disabling", reg, bit, old, new);
98
99 if (new == old)
100 dev_WARN(&ring->nhi->pdev->dev,
101 "interrupt for %s %d is already %s\n",
102 RING_TYPE(ring), ring->hop,
103 active ? "enabled" : "disabled");
104 iowrite32(new, ring->nhi->iobase + reg);
105}
106
107/**
108 * nhi_disable_interrupts() - disable interrupts for all rings
109 *
110 * Use only during init and shutdown.
111 */
112static void nhi_disable_interrupts(struct tb_nhi *nhi)
113{
114 int i = 0;
115 /* disable interrupts */
116 for (i = 0; i < RING_INTERRUPT_REG_COUNT(nhi); i++)
117 iowrite32(0, nhi->iobase + REG_RING_INTERRUPT_BASE + 4 * i);
118
119 /* clear interrupt status bits */
120 for (i = 0; i < RING_NOTIFY_REG_COUNT(nhi); i++)
121 ioread32(nhi->iobase + REG_RING_NOTIFY_BASE + 4 * i);
122}
123
124/* ring helper methods */
125
126static void __iomem *ring_desc_base(struct tb_ring *ring)
127{
128 void __iomem *io = ring->nhi->iobase;
129 io += ring->is_tx ? REG_TX_RING_BASE : REG_RX_RING_BASE;
130 io += ring->hop * 16;
131 return io;
132}
133
134static void __iomem *ring_options_base(struct tb_ring *ring)
135{
136 void __iomem *io = ring->nhi->iobase;
137 io += ring->is_tx ? REG_TX_OPTIONS_BASE : REG_RX_OPTIONS_BASE;
138 io += ring->hop * 32;
139 return io;
140}
141
142static void ring_iowrite_cons(struct tb_ring *ring, u16 cons)
143{
144 /*
145 * The other 16-bits in the register is read-only and writes to it
146 * are ignored by the hardware so we can save one ioread32() by
147 * filling the read-only bits with zeroes.
148 */
149 iowrite32(cons, ring_desc_base(ring) + 8);
150}
151
152static void ring_iowrite_prod(struct tb_ring *ring, u16 prod)
153{
154 /* See ring_iowrite_cons() above for explanation */
155 iowrite32(prod << 16, ring_desc_base(ring) + 8);
156}
157
158static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)
159{
160 iowrite32(value, ring_desc_base(ring) + offset);
161}
162
163static void ring_iowrite64desc(struct tb_ring *ring, u64 value, u32 offset)
164{
165 iowrite32(value, ring_desc_base(ring) + offset);
166 iowrite32(value >> 32, ring_desc_base(ring) + offset + 4);
167}
168
169static void ring_iowrite32options(struct tb_ring *ring, u32 value, u32 offset)
170{
171 iowrite32(value, ring_options_base(ring) + offset);
172}
173
174static bool ring_full(struct tb_ring *ring)
175{
176 return ((ring->head + 1) % ring->size) == ring->tail;
177}
178
179static bool ring_empty(struct tb_ring *ring)
180{
181 return ring->head == ring->tail;
182}
183
184/**
185 * ring_write_descriptors() - post frames from ring->queue to the controller
186 *
187 * ring->lock is held.
188 */
189static void ring_write_descriptors(struct tb_ring *ring)
190{
191 struct ring_frame *frame, *n;
192 struct ring_desc *descriptor;
193 list_for_each_entry_safe(frame, n, &ring->queue, list) {
194 if (ring_full(ring))
195 break;
196 list_move_tail(&frame->list, &ring->in_flight);
197 descriptor = &ring->descriptors[ring->head];
198 descriptor->phys = frame->buffer_phy;
199 descriptor->time = 0;
200 descriptor->flags = RING_DESC_POSTED | RING_DESC_INTERRUPT;
201 if (ring->is_tx) {
202 descriptor->length = frame->size;
203 descriptor->eof = frame->eof;
204 descriptor->sof = frame->sof;
205 }
206 ring->head = (ring->head + 1) % ring->size;
207 if (ring->is_tx)
208 ring_iowrite_prod(ring, ring->head);
209 else
210 ring_iowrite_cons(ring, ring->head);
211 }
212}
213
214/**
215 * ring_work() - progress completed frames
216 *
217 * If the ring is shutting down then all frames are marked as canceled and
218 * their callbacks are invoked.
219 *
220 * Otherwise we collect all completed frame from the ring buffer, write new
221 * frame to the ring buffer and invoke the callbacks for the completed frames.
222 */
223static void ring_work(struct work_struct *work)
224{
225 struct tb_ring *ring = container_of(work, typeof(*ring), work);
226 struct ring_frame *frame;
227 bool canceled = false;
228 unsigned long flags;
229 LIST_HEAD(done);
230
231 spin_lock_irqsave(&ring->lock, flags);
232
233 if (!ring->running) {
234 /* Move all frames to done and mark them as canceled. */
235 list_splice_tail_init(&ring->in_flight, &done);
236 list_splice_tail_init(&ring->queue, &done);
237 canceled = true;
238 goto invoke_callback;
239 }
240
241 while (!ring_empty(ring)) {
242 if (!(ring->descriptors[ring->tail].flags
243 & RING_DESC_COMPLETED))
244 break;
245 frame = list_first_entry(&ring->in_flight, typeof(*frame),
246 list);
247 list_move_tail(&frame->list, &done);
248 if (!ring->is_tx) {
249 frame->size = ring->descriptors[ring->tail].length;
250 frame->eof = ring->descriptors[ring->tail].eof;
251 frame->sof = ring->descriptors[ring->tail].sof;
252 frame->flags = ring->descriptors[ring->tail].flags;
253 }
254 ring->tail = (ring->tail + 1) % ring->size;
255 }
256 ring_write_descriptors(ring);
257
258invoke_callback:
259 /* allow callbacks to schedule new work */
260 spin_unlock_irqrestore(&ring->lock, flags);
261 while (!list_empty(&done)) {
262 frame = list_first_entry(&done, typeof(*frame), list);
263 /*
264 * The callback may reenqueue or delete frame.
265 * Do not hold on to it.
266 */
267 list_del_init(&frame->list);
268 if (frame->callback)
269 frame->callback(ring, frame, canceled);
270 }
271}
272
273int __tb_ring_enqueue(struct tb_ring *ring, struct ring_frame *frame)
274{
275 unsigned long flags;
276 int ret = 0;
277
278 spin_lock_irqsave(&ring->lock, flags);
279 if (ring->running) {
280 list_add_tail(&frame->list, &ring->queue);
281 ring_write_descriptors(ring);
282 } else {
283 ret = -ESHUTDOWN;
284 }
285 spin_unlock_irqrestore(&ring->lock, flags);
286 return ret;
287}
288EXPORT_SYMBOL_GPL(__tb_ring_enqueue);
289
290/**
291 * tb_ring_poll() - Poll one completed frame from the ring
292 * @ring: Ring to poll
293 *
294 * This function can be called when @start_poll callback of the @ring
295 * has been called. It will read one completed frame from the ring and
296 * return it to the caller. Returns %NULL if there is no more completed
297 * frames.
298 */
299struct ring_frame *tb_ring_poll(struct tb_ring *ring)
300{
301 struct ring_frame *frame = NULL;
302 unsigned long flags;
303
304 spin_lock_irqsave(&ring->lock, flags);
305 if (!ring->running)
306 goto unlock;
307 if (ring_empty(ring))
308 goto unlock;
309
310 if (ring->descriptors[ring->tail].flags & RING_DESC_COMPLETED) {
311 frame = list_first_entry(&ring->in_flight, typeof(*frame),
312 list);
313 list_del_init(&frame->list);
314
315 if (!ring->is_tx) {
316 frame->size = ring->descriptors[ring->tail].length;
317 frame->eof = ring->descriptors[ring->tail].eof;
318 frame->sof = ring->descriptors[ring->tail].sof;
319 frame->flags = ring->descriptors[ring->tail].flags;
320 }
321
322 ring->tail = (ring->tail + 1) % ring->size;
323 }
324
325unlock:
326 spin_unlock_irqrestore(&ring->lock, flags);
327 return frame;
328}
329EXPORT_SYMBOL_GPL(tb_ring_poll);
330
331static void __ring_interrupt_mask(struct tb_ring *ring, bool mask)
332{
333 int idx = ring_interrupt_index(ring);
334 int reg = REG_RING_INTERRUPT_BASE + idx / 32 * 4;
335 int bit = idx % 32;
336 u32 val;
337
338 val = ioread32(ring->nhi->iobase + reg);
339 if (mask)
340 val &= ~BIT(bit);
341 else
342 val |= BIT(bit);
343 iowrite32(val, ring->nhi->iobase + reg);
344}
345
346/* Both @nhi->lock and @ring->lock should be held */
347static void __ring_interrupt(struct tb_ring *ring)
348{
349 if (!ring->running)
350 return;
351
352 if (ring->start_poll) {
353 __ring_interrupt_mask(ring, true);
354 ring->start_poll(ring->poll_data);
355 } else {
356 schedule_work(&ring->work);
357 }
358}
359
360/**
361 * tb_ring_poll_complete() - Re-start interrupt for the ring
362 * @ring: Ring to re-start the interrupt
363 *
364 * This will re-start (unmask) the ring interrupt once the user is done
365 * with polling.
366 */
367void tb_ring_poll_complete(struct tb_ring *ring)
368{
369 unsigned long flags;
370
371 spin_lock_irqsave(&ring->nhi->lock, flags);
372 spin_lock(&ring->lock);
373 if (ring->start_poll)
374 __ring_interrupt_mask(ring, false);
375 spin_unlock(&ring->lock);
376 spin_unlock_irqrestore(&ring->nhi->lock, flags);
377}
378EXPORT_SYMBOL_GPL(tb_ring_poll_complete);
379
380static irqreturn_t ring_msix(int irq, void *data)
381{
382 struct tb_ring *ring = data;
383
384 spin_lock(&ring->nhi->lock);
385 spin_lock(&ring->lock);
386 __ring_interrupt(ring);
387 spin_unlock(&ring->lock);
388 spin_unlock(&ring->nhi->lock);
389
390 return IRQ_HANDLED;
391}
392
393static int ring_request_msix(struct tb_ring *ring, bool no_suspend)
394{
395 struct tb_nhi *nhi = ring->nhi;
396 unsigned long irqflags;
397 int ret;
398
399 if (!nhi->pdev->msix_enabled)
400 return 0;
401
402 ret = ida_simple_get(&nhi->msix_ida, 0, MSIX_MAX_VECS, GFP_KERNEL);
403 if (ret < 0)
404 return ret;
405
406 ring->vector = ret;
407
408 ring->irq = pci_irq_vector(ring->nhi->pdev, ring->vector);
409 if (ring->irq < 0)
410 return ring->irq;
411
412 irqflags = no_suspend ? IRQF_NO_SUSPEND : 0;
413 return request_irq(ring->irq, ring_msix, irqflags, "thunderbolt", ring);
414}
415
416static void ring_release_msix(struct tb_ring *ring)
417{
418 if (ring->irq <= 0)
419 return;
420
421 free_irq(ring->irq, ring);
422 ida_simple_remove(&ring->nhi->msix_ida, ring->vector);
423 ring->vector = 0;
424 ring->irq = 0;
425}
426
427static int nhi_alloc_hop(struct tb_nhi *nhi, struct tb_ring *ring)
428{
429 int ret = 0;
430
431 spin_lock_irq(&nhi->lock);
432
433 if (ring->hop < 0) {
434 unsigned int i;
435
436 /*
437 * Automatically allocate HopID from the non-reserved
438 * range 1 .. hop_count - 1.
439 */
440 for (i = RING_FIRST_USABLE_HOPID; i < nhi->hop_count; i++) {
441 if (ring->is_tx) {
442 if (!nhi->tx_rings[i]) {
443 ring->hop = i;
444 break;
445 }
446 } else {
447 if (!nhi->rx_rings[i]) {
448 ring->hop = i;
449 break;
450 }
451 }
452 }
453 }
454
455 if (ring->hop < 0 || ring->hop >= nhi->hop_count) {
456 dev_warn(&nhi->pdev->dev, "invalid hop: %d\n", ring->hop);
457 ret = -EINVAL;
458 goto err_unlock;
459 }
460 if (ring->is_tx && nhi->tx_rings[ring->hop]) {
461 dev_warn(&nhi->pdev->dev, "TX hop %d already allocated\n",
462 ring->hop);
463 ret = -EBUSY;
464 goto err_unlock;
465 } else if (!ring->is_tx && nhi->rx_rings[ring->hop]) {
466 dev_warn(&nhi->pdev->dev, "RX hop %d already allocated\n",
467 ring->hop);
468 ret = -EBUSY;
469 goto err_unlock;
470 }
471
472 if (ring->is_tx)
473 nhi->tx_rings[ring->hop] = ring;
474 else
475 nhi->rx_rings[ring->hop] = ring;
476
477err_unlock:
478 spin_unlock_irq(&nhi->lock);
479
480 return ret;
481}
482
483static struct tb_ring *tb_ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
484 bool transmit, unsigned int flags,
485 u16 sof_mask, u16 eof_mask,
486 void (*start_poll)(void *),
487 void *poll_data)
488{
489 struct tb_ring *ring = NULL;
490
491 dev_dbg(&nhi->pdev->dev, "allocating %s ring %d of size %d\n",
492 transmit ? "TX" : "RX", hop, size);
493
494 ring = kzalloc(sizeof(*ring), GFP_KERNEL);
495 if (!ring)
496 return NULL;
497
498 spin_lock_init(&ring->lock);
499 INIT_LIST_HEAD(&ring->queue);
500 INIT_LIST_HEAD(&ring->in_flight);
501 INIT_WORK(&ring->work, ring_work);
502
503 ring->nhi = nhi;
504 ring->hop = hop;
505 ring->is_tx = transmit;
506 ring->size = size;
507 ring->flags = flags;
508 ring->sof_mask = sof_mask;
509 ring->eof_mask = eof_mask;
510 ring->head = 0;
511 ring->tail = 0;
512 ring->running = false;
513 ring->start_poll = start_poll;
514 ring->poll_data = poll_data;
515
516 ring->descriptors = dma_alloc_coherent(&ring->nhi->pdev->dev,
517 size * sizeof(*ring->descriptors),
518 &ring->descriptors_dma, GFP_KERNEL | __GFP_ZERO);
519 if (!ring->descriptors)
520 goto err_free_ring;
521
522 if (ring_request_msix(ring, flags & RING_FLAG_NO_SUSPEND))
523 goto err_free_descs;
524
525 if (nhi_alloc_hop(nhi, ring))
526 goto err_release_msix;
527
528 return ring;
529
530err_release_msix:
531 ring_release_msix(ring);
532err_free_descs:
533 dma_free_coherent(&ring->nhi->pdev->dev,
534 ring->size * sizeof(*ring->descriptors),
535 ring->descriptors, ring->descriptors_dma);
536err_free_ring:
537 kfree(ring);
538
539 return NULL;
540}
541
542/**
543 * tb_ring_alloc_tx() - Allocate DMA ring for transmit
544 * @nhi: Pointer to the NHI the ring is to be allocated
545 * @hop: HopID (ring) to allocate
546 * @size: Number of entries in the ring
547 * @flags: Flags for the ring
548 */
549struct tb_ring *tb_ring_alloc_tx(struct tb_nhi *nhi, int hop, int size,
550 unsigned int flags)
551{
552 return tb_ring_alloc(nhi, hop, size, true, flags, 0, 0, NULL, NULL);
553}
554EXPORT_SYMBOL_GPL(tb_ring_alloc_tx);
555
556/**
557 * tb_ring_alloc_rx() - Allocate DMA ring for receive
558 * @nhi: Pointer to the NHI the ring is to be allocated
559 * @hop: HopID (ring) to allocate. Pass %-1 for automatic allocation.
560 * @size: Number of entries in the ring
561 * @flags: Flags for the ring
562 * @sof_mask: Mask of PDF values that start a frame
563 * @eof_mask: Mask of PDF values that end a frame
564 * @start_poll: If not %NULL the ring will call this function when an
565 * interrupt is triggered and masked, instead of callback
566 * in each Rx frame.
567 * @poll_data: Optional data passed to @start_poll
568 */
569struct tb_ring *tb_ring_alloc_rx(struct tb_nhi *nhi, int hop, int size,
570 unsigned int flags, u16 sof_mask, u16 eof_mask,
571 void (*start_poll)(void *), void *poll_data)
572{
573 return tb_ring_alloc(nhi, hop, size, false, flags, sof_mask, eof_mask,
574 start_poll, poll_data);
575}
576EXPORT_SYMBOL_GPL(tb_ring_alloc_rx);
577
578/**
579 * tb_ring_start() - enable a ring
580 *
581 * Must not be invoked in parallel with tb_ring_stop().
582 */
583void tb_ring_start(struct tb_ring *ring)
584{
585 u16 frame_size;
586 u32 flags;
587
588 spin_lock_irq(&ring->nhi->lock);
589 spin_lock(&ring->lock);
590 if (ring->nhi->going_away)
591 goto err;
592 if (ring->running) {
593 dev_WARN(&ring->nhi->pdev->dev, "ring already started\n");
594 goto err;
595 }
596 dev_dbg(&ring->nhi->pdev->dev, "starting %s %d\n",
597 RING_TYPE(ring), ring->hop);
598
599 if (ring->flags & RING_FLAG_FRAME) {
600 /* Means 4096 */
601 frame_size = 0;
602 flags = RING_FLAG_ENABLE;
603 } else {
604 frame_size = TB_FRAME_SIZE;
605 flags = RING_FLAG_ENABLE | RING_FLAG_RAW;
606 }
607
608 ring_iowrite64desc(ring, ring->descriptors_dma, 0);
609 if (ring->is_tx) {
610 ring_iowrite32desc(ring, ring->size, 12);
611 ring_iowrite32options(ring, 0, 4); /* time releated ? */
612 ring_iowrite32options(ring, flags, 0);
613 } else {
614 u32 sof_eof_mask = ring->sof_mask << 16 | ring->eof_mask;
615
616 ring_iowrite32desc(ring, (frame_size << 16) | ring->size, 12);
617 ring_iowrite32options(ring, sof_eof_mask, 4);
618 ring_iowrite32options(ring, flags, 0);
619 }
620 ring_interrupt_active(ring, true);
621 ring->running = true;
622err:
623 spin_unlock(&ring->lock);
624 spin_unlock_irq(&ring->nhi->lock);
625}
626EXPORT_SYMBOL_GPL(tb_ring_start);
627
628/**
629 * tb_ring_stop() - shutdown a ring
630 *
631 * Must not be invoked from a callback.
632 *
633 * This method will disable the ring. Further calls to
634 * tb_ring_tx/tb_ring_rx will return -ESHUTDOWN until ring_stop has been
635 * called.
636 *
637 * All enqueued frames will be canceled and their callbacks will be executed
638 * with frame->canceled set to true (on the callback thread). This method
639 * returns only after all callback invocations have finished.
640 */
641void tb_ring_stop(struct tb_ring *ring)
642{
643 spin_lock_irq(&ring->nhi->lock);
644 spin_lock(&ring->lock);
645 dev_dbg(&ring->nhi->pdev->dev, "stopping %s %d\n",
646 RING_TYPE(ring), ring->hop);
647 if (ring->nhi->going_away)
648 goto err;
649 if (!ring->running) {
650 dev_WARN(&ring->nhi->pdev->dev, "%s %d already stopped\n",
651 RING_TYPE(ring), ring->hop);
652 goto err;
653 }
654 ring_interrupt_active(ring, false);
655
656 ring_iowrite32options(ring, 0, 0);
657 ring_iowrite64desc(ring, 0, 0);
658 ring_iowrite32desc(ring, 0, 8);
659 ring_iowrite32desc(ring, 0, 12);
660 ring->head = 0;
661 ring->tail = 0;
662 ring->running = false;
663
664err:
665 spin_unlock(&ring->lock);
666 spin_unlock_irq(&ring->nhi->lock);
667
668 /*
669 * schedule ring->work to invoke callbacks on all remaining frames.
670 */
671 schedule_work(&ring->work);
672 flush_work(&ring->work);
673}
674EXPORT_SYMBOL_GPL(tb_ring_stop);
675
676/*
677 * tb_ring_free() - free ring
678 *
679 * When this method returns all invocations of ring->callback will have
680 * finished.
681 *
682 * Ring must be stopped.
683 *
684 * Must NOT be called from ring_frame->callback!
685 */
686void tb_ring_free(struct tb_ring *ring)
687{
688 spin_lock_irq(&ring->nhi->lock);
689 /*
690 * Dissociate the ring from the NHI. This also ensures that
691 * nhi_interrupt_work cannot reschedule ring->work.
692 */
693 if (ring->is_tx)
694 ring->nhi->tx_rings[ring->hop] = NULL;
695 else
696 ring->nhi->rx_rings[ring->hop] = NULL;
697
698 if (ring->running) {
699 dev_WARN(&ring->nhi->pdev->dev, "%s %d still running\n",
700 RING_TYPE(ring), ring->hop);
701 }
702 spin_unlock_irq(&ring->nhi->lock);
703
704 ring_release_msix(ring);
705
706 dma_free_coherent(&ring->nhi->pdev->dev,
707 ring->size * sizeof(*ring->descriptors),
708 ring->descriptors, ring->descriptors_dma);
709
710 ring->descriptors = NULL;
711 ring->descriptors_dma = 0;
712
713
714 dev_dbg(&ring->nhi->pdev->dev, "freeing %s %d\n", RING_TYPE(ring),
715 ring->hop);
716
717 /**
718 * ring->work can no longer be scheduled (it is scheduled only
719 * by nhi_interrupt_work, ring_stop and ring_msix). Wait for it
720 * to finish before freeing the ring.
721 */
722 flush_work(&ring->work);
723 kfree(ring);
724}
725EXPORT_SYMBOL_GPL(tb_ring_free);
726
727/**
728 * nhi_mailbox_cmd() - Send a command through NHI mailbox
729 * @nhi: Pointer to the NHI structure
730 * @cmd: Command to send
731 * @data: Data to be send with the command
732 *
733 * Sends mailbox command to the firmware running on NHI. Returns %0 in
734 * case of success and negative errno in case of failure.
735 */
736int nhi_mailbox_cmd(struct tb_nhi *nhi, enum nhi_mailbox_cmd cmd, u32 data)
737{
738 ktime_t timeout;
739 u32 val;
740
741 iowrite32(data, nhi->iobase + REG_INMAIL_DATA);
742
743 val = ioread32(nhi->iobase + REG_INMAIL_CMD);
744 val &= ~(REG_INMAIL_CMD_MASK | REG_INMAIL_ERROR);
745 val |= REG_INMAIL_OP_REQUEST | cmd;
746 iowrite32(val, nhi->iobase + REG_INMAIL_CMD);
747
748 timeout = ktime_add_ms(ktime_get(), NHI_MAILBOX_TIMEOUT);
749 do {
750 val = ioread32(nhi->iobase + REG_INMAIL_CMD);
751 if (!(val & REG_INMAIL_OP_REQUEST))
752 break;
753 usleep_range(10, 20);
754 } while (ktime_before(ktime_get(), timeout));
755
756 if (val & REG_INMAIL_OP_REQUEST)
757 return -ETIMEDOUT;
758 if (val & REG_INMAIL_ERROR)
759 return -EIO;
760
761 return 0;
762}
763
764/**
765 * nhi_mailbox_mode() - Return current firmware operation mode
766 * @nhi: Pointer to the NHI structure
767 *
768 * The function reads current firmware operation mode using NHI mailbox
769 * registers and returns it to the caller.
770 */
771enum nhi_fw_mode nhi_mailbox_mode(struct tb_nhi *nhi)
772{
773 u32 val;
774
775 val = ioread32(nhi->iobase + REG_OUTMAIL_CMD);
776 val &= REG_OUTMAIL_CMD_OPMODE_MASK;
777 val >>= REG_OUTMAIL_CMD_OPMODE_SHIFT;
778
779 return (enum nhi_fw_mode)val;
780}
781
782static void nhi_interrupt_work(struct work_struct *work)
783{
784 struct tb_nhi *nhi = container_of(work, typeof(*nhi), interrupt_work);
785 int value = 0; /* Suppress uninitialized usage warning. */
786 int bit;
787 int hop = -1;
788 int type = 0; /* current interrupt type 0: TX, 1: RX, 2: RX overflow */
789 struct tb_ring *ring;
790
791 spin_lock_irq(&nhi->lock);
792
793 /*
794 * Starting at REG_RING_NOTIFY_BASE there are three status bitfields
795 * (TX, RX, RX overflow). We iterate over the bits and read a new
796 * dwords as required. The registers are cleared on read.
797 */
798 for (bit = 0; bit < 3 * nhi->hop_count; bit++) {
799 if (bit % 32 == 0)
800 value = ioread32(nhi->iobase
801 + REG_RING_NOTIFY_BASE
802 + 4 * (bit / 32));
803 if (++hop == nhi->hop_count) {
804 hop = 0;
805 type++;
806 }
807 if ((value & (1 << (bit % 32))) == 0)
808 continue;
809 if (type == 2) {
810 dev_warn(&nhi->pdev->dev,
811 "RX overflow for ring %d\n",
812 hop);
813 continue;
814 }
815 if (type == 0)
816 ring = nhi->tx_rings[hop];
817 else
818 ring = nhi->rx_rings[hop];
819 if (ring == NULL) {
820 dev_warn(&nhi->pdev->dev,
821 "got interrupt for inactive %s ring %d\n",
822 type ? "RX" : "TX",
823 hop);
824 continue;
825 }
826
827 spin_lock(&ring->lock);
828 __ring_interrupt(ring);
829 spin_unlock(&ring->lock);
830 }
831 spin_unlock_irq(&nhi->lock);
832}
833
834static irqreturn_t nhi_msi(int irq, void *data)
835{
836 struct tb_nhi *nhi = data;
837 schedule_work(&nhi->interrupt_work);
838 return IRQ_HANDLED;
839}
840
841static int __nhi_suspend_noirq(struct device *dev, bool wakeup)
842{
843 struct pci_dev *pdev = to_pci_dev(dev);
844 struct tb *tb = pci_get_drvdata(pdev);
845 struct tb_nhi *nhi = tb->nhi;
846 int ret;
847
848 ret = tb_domain_suspend_noirq(tb);
849 if (ret)
850 return ret;
851
852 if (nhi->ops && nhi->ops->suspend_noirq) {
853 ret = nhi->ops->suspend_noirq(tb->nhi, wakeup);
854 if (ret)
855 return ret;
856 }
857
858 return 0;
859}
860
861static int nhi_suspend_noirq(struct device *dev)
862{
863 return __nhi_suspend_noirq(dev, device_may_wakeup(dev));
864}
865
866static bool nhi_wake_supported(struct pci_dev *pdev)
867{
868 u8 val;
869
870 /*
871 * If power rails are sustainable for wakeup from S4 this
872 * property is set by the BIOS.
873 */
874 if (device_property_read_u8(&pdev->dev, "WAKE_SUPPORTED", &val))
875 return !!val;
876
877 return true;
878}
879
880static int nhi_poweroff_noirq(struct device *dev)
881{
882 struct pci_dev *pdev = to_pci_dev(dev);
883 bool wakeup;
884
885 wakeup = device_may_wakeup(dev) && nhi_wake_supported(pdev);
886 return __nhi_suspend_noirq(dev, wakeup);
887}
888
889static void nhi_enable_int_throttling(struct tb_nhi *nhi)
890{
891 /* Throttling is specified in 256ns increments */
892 u32 throttle = DIV_ROUND_UP(128 * NSEC_PER_USEC, 256);
893 unsigned int i;
894
895 /*
896 * Configure interrupt throttling for all vectors even if we
897 * only use few.
898 */
899 for (i = 0; i < MSIX_MAX_VECS; i++) {
900 u32 reg = REG_INT_THROTTLING_RATE + i * 4;
901 iowrite32(throttle, nhi->iobase + reg);
902 }
903}
904
905static int nhi_resume_noirq(struct device *dev)
906{
907 struct pci_dev *pdev = to_pci_dev(dev);
908 struct tb *tb = pci_get_drvdata(pdev);
909 struct tb_nhi *nhi = tb->nhi;
910 int ret;
911
912 /*
913 * Check that the device is still there. It may be that the user
914 * unplugged last device which causes the host controller to go
915 * away on PCs.
916 */
917 if (!pci_device_is_present(pdev)) {
918 nhi->going_away = true;
919 } else {
920 if (nhi->ops && nhi->ops->resume_noirq) {
921 ret = nhi->ops->resume_noirq(nhi);
922 if (ret)
923 return ret;
924 }
925 nhi_enable_int_throttling(tb->nhi);
926 }
927
928 return tb_domain_resume_noirq(tb);
929}
930
931static int nhi_suspend(struct device *dev)
932{
933 struct pci_dev *pdev = to_pci_dev(dev);
934 struct tb *tb = pci_get_drvdata(pdev);
935
936 return tb_domain_suspend(tb);
937}
938
939static void nhi_complete(struct device *dev)
940{
941 struct pci_dev *pdev = to_pci_dev(dev);
942 struct tb *tb = pci_get_drvdata(pdev);
943
944 /*
945 * If we were runtime suspended when system suspend started,
946 * schedule runtime resume now. It should bring the domain back
947 * to functional state.
948 */
949 if (pm_runtime_suspended(&pdev->dev))
950 pm_runtime_resume(&pdev->dev);
951 else
952 tb_domain_complete(tb);
953}
954
955static int nhi_runtime_suspend(struct device *dev)
956{
957 struct pci_dev *pdev = to_pci_dev(dev);
958 struct tb *tb = pci_get_drvdata(pdev);
959 struct tb_nhi *nhi = tb->nhi;
960 int ret;
961
962 ret = tb_domain_runtime_suspend(tb);
963 if (ret)
964 return ret;
965
966 if (nhi->ops && nhi->ops->runtime_suspend) {
967 ret = nhi->ops->runtime_suspend(tb->nhi);
968 if (ret)
969 return ret;
970 }
971 return 0;
972}
973
974static int nhi_runtime_resume(struct device *dev)
975{
976 struct pci_dev *pdev = to_pci_dev(dev);
977 struct tb *tb = pci_get_drvdata(pdev);
978 struct tb_nhi *nhi = tb->nhi;
979 int ret;
980
981 if (nhi->ops && nhi->ops->runtime_resume) {
982 ret = nhi->ops->runtime_resume(nhi);
983 if (ret)
984 return ret;
985 }
986
987 nhi_enable_int_throttling(nhi);
988 return tb_domain_runtime_resume(tb);
989}
990
991static void nhi_shutdown(struct tb_nhi *nhi)
992{
993 int i;
994
995 dev_dbg(&nhi->pdev->dev, "shutdown\n");
996
997 for (i = 0; i < nhi->hop_count; i++) {
998 if (nhi->tx_rings[i])
999 dev_WARN(&nhi->pdev->dev,
1000 "TX ring %d is still active\n", i);
1001 if (nhi->rx_rings[i])
1002 dev_WARN(&nhi->pdev->dev,
1003 "RX ring %d is still active\n", i);
1004 }
1005 nhi_disable_interrupts(nhi);
1006 /*
1007 * We have to release the irq before calling flush_work. Otherwise an
1008 * already executing IRQ handler could call schedule_work again.
1009 */
1010 if (!nhi->pdev->msix_enabled) {
1011 devm_free_irq(&nhi->pdev->dev, nhi->pdev->irq, nhi);
1012 flush_work(&nhi->interrupt_work);
1013 }
1014 ida_destroy(&nhi->msix_ida);
1015
1016 if (nhi->ops && nhi->ops->shutdown)
1017 nhi->ops->shutdown(nhi);
1018}
1019
1020static int nhi_init_msi(struct tb_nhi *nhi)
1021{
1022 struct pci_dev *pdev = nhi->pdev;
1023 int res, irq, nvec;
1024
1025 /* In case someone left them on. */
1026 nhi_disable_interrupts(nhi);
1027
1028 nhi_enable_int_throttling(nhi);
1029
1030 ida_init(&nhi->msix_ida);
1031
1032 /*
1033 * The NHI has 16 MSI-X vectors or a single MSI. We first try to
1034 * get all MSI-X vectors and if we succeed, each ring will have
1035 * one MSI-X. If for some reason that does not work out, we
1036 * fallback to a single MSI.
1037 */
1038 nvec = pci_alloc_irq_vectors(pdev, MSIX_MIN_VECS, MSIX_MAX_VECS,
1039 PCI_IRQ_MSIX);
1040 if (nvec < 0) {
1041 nvec = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI);
1042 if (nvec < 0)
1043 return nvec;
1044
1045 INIT_WORK(&nhi->interrupt_work, nhi_interrupt_work);
1046
1047 irq = pci_irq_vector(nhi->pdev, 0);
1048 if (irq < 0)
1049 return irq;
1050
1051 res = devm_request_irq(&pdev->dev, irq, nhi_msi,
1052 IRQF_NO_SUSPEND, "thunderbolt", nhi);
1053 if (res) {
1054 dev_err(&pdev->dev, "request_irq failed, aborting\n");
1055 return res;
1056 }
1057 }
1058
1059 return 0;
1060}
1061
1062static bool nhi_imr_valid(struct pci_dev *pdev)
1063{
1064 u8 val;
1065
1066 if (!device_property_read_u8(&pdev->dev, "IMR_VALID", &val))
1067 return !!val;
1068
1069 return true;
1070}
1071
1072static int nhi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1073{
1074 struct tb_nhi *nhi;
1075 struct tb *tb;
1076 int res;
1077
1078 if (!nhi_imr_valid(pdev)) {
1079 dev_warn(&pdev->dev, "firmware image not valid, aborting\n");
1080 return -ENODEV;
1081 }
1082
1083 res = pcim_enable_device(pdev);
1084 if (res) {
1085 dev_err(&pdev->dev, "cannot enable PCI device, aborting\n");
1086 return res;
1087 }
1088
1089 res = pcim_iomap_regions(pdev, 1 << 0, "thunderbolt");
1090 if (res) {
1091 dev_err(&pdev->dev, "cannot obtain PCI resources, aborting\n");
1092 return res;
1093 }
1094
1095 nhi = devm_kzalloc(&pdev->dev, sizeof(*nhi), GFP_KERNEL);
1096 if (!nhi)
1097 return -ENOMEM;
1098
1099 nhi->pdev = pdev;
1100 nhi->ops = (const struct tb_nhi_ops *)id->driver_data;
1101 /* cannot fail - table is allocated bin pcim_iomap_regions */
1102 nhi->iobase = pcim_iomap_table(pdev)[0];
1103 nhi->hop_count = ioread32(nhi->iobase + REG_HOP_COUNT) & 0x3ff;
1104 dev_dbg(&pdev->dev, "total paths: %d\n", nhi->hop_count);
1105
1106 nhi->tx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1107 sizeof(*nhi->tx_rings), GFP_KERNEL);
1108 nhi->rx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1109 sizeof(*nhi->rx_rings), GFP_KERNEL);
1110 if (!nhi->tx_rings || !nhi->rx_rings)
1111 return -ENOMEM;
1112
1113 res = nhi_init_msi(nhi);
1114 if (res) {
1115 dev_err(&pdev->dev, "cannot enable MSI, aborting\n");
1116 return res;
1117 }
1118
1119 spin_lock_init(&nhi->lock);
1120
1121 res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1122 if (res)
1123 res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
1124 if (res) {
1125 dev_err(&pdev->dev, "failed to set DMA mask\n");
1126 return res;
1127 }
1128
1129 pci_set_master(pdev);
1130
1131 if (nhi->ops && nhi->ops->init) {
1132 res = nhi->ops->init(nhi);
1133 if (res)
1134 return res;
1135 }
1136
1137 tb = icm_probe(nhi);
1138 if (!tb)
1139 tb = tb_probe(nhi);
1140 if (!tb) {
1141 dev_err(&nhi->pdev->dev,
1142 "failed to determine connection manager, aborting\n");
1143 return -ENODEV;
1144 }
1145
1146 dev_dbg(&nhi->pdev->dev, "NHI initialized, starting thunderbolt\n");
1147
1148 res = tb_domain_add(tb);
1149 if (res) {
1150 /*
1151 * At this point the RX/TX rings might already have been
1152 * activated. Do a proper shutdown.
1153 */
1154 tb_domain_put(tb);
1155 nhi_shutdown(nhi);
1156 return res;
1157 }
1158 pci_set_drvdata(pdev, tb);
1159
1160 pm_runtime_allow(&pdev->dev);
1161 pm_runtime_set_autosuspend_delay(&pdev->dev, TB_AUTOSUSPEND_DELAY);
1162 pm_runtime_use_autosuspend(&pdev->dev);
1163 pm_runtime_put_autosuspend(&pdev->dev);
1164
1165 return 0;
1166}
1167
1168static void nhi_remove(struct pci_dev *pdev)
1169{
1170 struct tb *tb = pci_get_drvdata(pdev);
1171 struct tb_nhi *nhi = tb->nhi;
1172
1173 pm_runtime_get_sync(&pdev->dev);
1174 pm_runtime_dont_use_autosuspend(&pdev->dev);
1175 pm_runtime_forbid(&pdev->dev);
1176
1177 tb_domain_remove(tb);
1178 nhi_shutdown(nhi);
1179}
1180
1181/*
1182 * The tunneled pci bridges are siblings of us. Use resume_noirq to reenable
1183 * the tunnels asap. A corresponding pci quirk blocks the downstream bridges
1184 * resume_noirq until we are done.
1185 */
1186static const struct dev_pm_ops nhi_pm_ops = {
1187 .suspend_noirq = nhi_suspend_noirq,
1188 .resume_noirq = nhi_resume_noirq,
1189 .freeze_noirq = nhi_suspend_noirq, /*
1190 * we just disable hotplug, the
1191 * pci-tunnels stay alive.
1192 */
1193 .thaw_noirq = nhi_resume_noirq,
1194 .restore_noirq = nhi_resume_noirq,
1195 .suspend = nhi_suspend,
1196 .freeze = nhi_suspend,
1197 .poweroff_noirq = nhi_poweroff_noirq,
1198 .poweroff = nhi_suspend,
1199 .complete = nhi_complete,
1200 .runtime_suspend = nhi_runtime_suspend,
1201 .runtime_resume = nhi_runtime_resume,
1202};
1203
1204static struct pci_device_id nhi_ids[] = {
1205 /*
1206 * We have to specify class, the TB bridges use the same device and
1207 * vendor (sub)id on gen 1 and gen 2 controllers.
1208 */
1209 {
1210 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1211 .vendor = PCI_VENDOR_ID_INTEL,
1212 .device = PCI_DEVICE_ID_INTEL_LIGHT_RIDGE,
1213 .subvendor = 0x2222, .subdevice = 0x1111,
1214 },
1215 {
1216 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1217 .vendor = PCI_VENDOR_ID_INTEL,
1218 .device = PCI_DEVICE_ID_INTEL_CACTUS_RIDGE_4C,
1219 .subvendor = 0x2222, .subdevice = 0x1111,
1220 },
1221 {
1222 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1223 .vendor = PCI_VENDOR_ID_INTEL,
1224 .device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI,
1225 .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1226 },
1227 {
1228 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1229 .vendor = PCI_VENDOR_ID_INTEL,
1230 .device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI,
1231 .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1232 },
1233
1234 /* Thunderbolt 3 */
1235 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_2C_NHI) },
1236 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_4C_NHI) },
1237 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_USBONLY_NHI) },
1238 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_NHI) },
1239 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_USBONLY_NHI) },
1240 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_2C_NHI) },
1241 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_4C_NHI) },
1242 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_USBONLY_NHI) },
1243 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI) },
1244 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI) },
1245 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ICL_NHI0),
1246 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1247 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ICL_NHI1),
1248 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1249 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_NHI0),
1250 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1251 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_NHI1),
1252 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1253
1254 /* Any USB4 compliant host */
1255 { PCI_DEVICE_CLASS(PCI_CLASS_SERIAL_USB_USB4, ~0) },
1256
1257 { 0,}
1258};
1259
1260MODULE_DEVICE_TABLE(pci, nhi_ids);
1261MODULE_LICENSE("GPL");
1262
1263static struct pci_driver nhi_driver = {
1264 .name = "thunderbolt",
1265 .id_table = nhi_ids,
1266 .probe = nhi_probe,
1267 .remove = nhi_remove,
1268 .shutdown = nhi_remove,
1269 .driver.pm = &nhi_pm_ops,
1270};
1271
1272static int __init nhi_init(void)
1273{
1274 int ret;
1275
1276 ret = tb_domain_init();
1277 if (ret)
1278 return ret;
1279 ret = pci_register_driver(&nhi_driver);
1280 if (ret)
1281 tb_domain_exit();
1282 return ret;
1283}
1284
1285static void __exit nhi_unload(void)
1286{
1287 pci_unregister_driver(&nhi_driver);
1288 tb_domain_exit();
1289}
1290
1291rootfs_initcall(nhi_init);
1292module_exit(nhi_unload);
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
21#include "nhi.h"
22#include "nhi_regs.h"
23#include "tb.h"
24
25#define RING_TYPE(ring) ((ring)->is_tx ? "TX ring" : "RX ring")
26
27#define RING_FIRST_USABLE_HOPID 1
28
29/*
30 * Minimal number of vectors when we use MSI-X. Two for control channel
31 * Rx/Tx and the rest four are for cross domain DMA paths.
32 */
33#define MSIX_MIN_VECS 6
34#define MSIX_MAX_VECS 16
35
36#define NHI_MAILBOX_TIMEOUT 500 /* ms */
37
38static int ring_interrupt_index(struct tb_ring *ring)
39{
40 int bit = ring->hop;
41 if (!ring->is_tx)
42 bit += ring->nhi->hop_count;
43 return bit;
44}
45
46/*
47 * ring_interrupt_active() - activate/deactivate interrupts for a single ring
48 *
49 * ring->nhi->lock must be held.
50 */
51static void ring_interrupt_active(struct tb_ring *ring, bool active)
52{
53 int reg = REG_RING_INTERRUPT_BASE +
54 ring_interrupt_index(ring) / 32 * 4;
55 int bit = ring_interrupt_index(ring) & 31;
56 int mask = 1 << bit;
57 u32 old, new;
58
59 if (ring->irq > 0) {
60 u32 step, shift, ivr, misc;
61 void __iomem *ivr_base;
62 int index;
63
64 if (ring->is_tx)
65 index = ring->hop;
66 else
67 index = ring->hop + ring->nhi->hop_count;
68
69 /*
70 * Ask the hardware to clear interrupt status bits automatically
71 * since we already know which interrupt was triggered.
72 */
73 misc = ioread32(ring->nhi->iobase + REG_DMA_MISC);
74 if (!(misc & REG_DMA_MISC_INT_AUTO_CLEAR)) {
75 misc |= REG_DMA_MISC_INT_AUTO_CLEAR;
76 iowrite32(misc, ring->nhi->iobase + REG_DMA_MISC);
77 }
78
79 ivr_base = ring->nhi->iobase + REG_INT_VEC_ALLOC_BASE;
80 step = index / REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
81 shift = index % REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
82 ivr = ioread32(ivr_base + step);
83 ivr &= ~(REG_INT_VEC_ALLOC_MASK << shift);
84 if (active)
85 ivr |= ring->vector << shift;
86 iowrite32(ivr, ivr_base + step);
87 }
88
89 old = ioread32(ring->nhi->iobase + reg);
90 if (active)
91 new = old | mask;
92 else
93 new = old & ~mask;
94
95 dev_dbg(&ring->nhi->pdev->dev,
96 "%s interrupt at register %#x bit %d (%#x -> %#x)\n",
97 active ? "enabling" : "disabling", reg, bit, old, new);
98
99 if (new == old)
100 dev_WARN(&ring->nhi->pdev->dev,
101 "interrupt for %s %d is already %s\n",
102 RING_TYPE(ring), ring->hop,
103 active ? "enabled" : "disabled");
104 iowrite32(new, ring->nhi->iobase + reg);
105}
106
107/*
108 * nhi_disable_interrupts() - disable interrupts for all rings
109 *
110 * Use only during init and shutdown.
111 */
112static void nhi_disable_interrupts(struct tb_nhi *nhi)
113{
114 int i = 0;
115 /* disable interrupts */
116 for (i = 0; i < RING_INTERRUPT_REG_COUNT(nhi); i++)
117 iowrite32(0, nhi->iobase + REG_RING_INTERRUPT_BASE + 4 * i);
118
119 /* clear interrupt status bits */
120 for (i = 0; i < RING_NOTIFY_REG_COUNT(nhi); i++)
121 ioread32(nhi->iobase + REG_RING_NOTIFY_BASE + 4 * i);
122}
123
124/* ring helper methods */
125
126static void __iomem *ring_desc_base(struct tb_ring *ring)
127{
128 void __iomem *io = ring->nhi->iobase;
129 io += ring->is_tx ? REG_TX_RING_BASE : REG_RX_RING_BASE;
130 io += ring->hop * 16;
131 return io;
132}
133
134static void __iomem *ring_options_base(struct tb_ring *ring)
135{
136 void __iomem *io = ring->nhi->iobase;
137 io += ring->is_tx ? REG_TX_OPTIONS_BASE : REG_RX_OPTIONS_BASE;
138 io += ring->hop * 32;
139 return io;
140}
141
142static void ring_iowrite_cons(struct tb_ring *ring, u16 cons)
143{
144 /*
145 * The other 16-bits in the register is read-only and writes to it
146 * are ignored by the hardware so we can save one ioread32() by
147 * filling the read-only bits with zeroes.
148 */
149 iowrite32(cons, ring_desc_base(ring) + 8);
150}
151
152static void ring_iowrite_prod(struct tb_ring *ring, u16 prod)
153{
154 /* See ring_iowrite_cons() above for explanation */
155 iowrite32(prod << 16, ring_desc_base(ring) + 8);
156}
157
158static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)
159{
160 iowrite32(value, ring_desc_base(ring) + offset);
161}
162
163static void ring_iowrite64desc(struct tb_ring *ring, u64 value, u32 offset)
164{
165 iowrite32(value, ring_desc_base(ring) + offset);
166 iowrite32(value >> 32, ring_desc_base(ring) + offset + 4);
167}
168
169static void ring_iowrite32options(struct tb_ring *ring, u32 value, u32 offset)
170{
171 iowrite32(value, ring_options_base(ring) + offset);
172}
173
174static bool ring_full(struct tb_ring *ring)
175{
176 return ((ring->head + 1) % ring->size) == ring->tail;
177}
178
179static bool ring_empty(struct tb_ring *ring)
180{
181 return ring->head == ring->tail;
182}
183
184/*
185 * ring_write_descriptors() - post frames from ring->queue to the controller
186 *
187 * ring->lock is held.
188 */
189static void ring_write_descriptors(struct tb_ring *ring)
190{
191 struct ring_frame *frame, *n;
192 struct ring_desc *descriptor;
193 list_for_each_entry_safe(frame, n, &ring->queue, list) {
194 if (ring_full(ring))
195 break;
196 list_move_tail(&frame->list, &ring->in_flight);
197 descriptor = &ring->descriptors[ring->head];
198 descriptor->phys = frame->buffer_phy;
199 descriptor->time = 0;
200 descriptor->flags = RING_DESC_POSTED | RING_DESC_INTERRUPT;
201 if (ring->is_tx) {
202 descriptor->length = frame->size;
203 descriptor->eof = frame->eof;
204 descriptor->sof = frame->sof;
205 }
206 ring->head = (ring->head + 1) % ring->size;
207 if (ring->is_tx)
208 ring_iowrite_prod(ring, ring->head);
209 else
210 ring_iowrite_cons(ring, ring->head);
211 }
212}
213
214/*
215 * ring_work() - progress completed frames
216 *
217 * If the ring is shutting down then all frames are marked as canceled and
218 * their callbacks are invoked.
219 *
220 * Otherwise we collect all completed frame from the ring buffer, write new
221 * frame to the ring buffer and invoke the callbacks for the completed frames.
222 */
223static void ring_work(struct work_struct *work)
224{
225 struct tb_ring *ring = container_of(work, typeof(*ring), work);
226 struct ring_frame *frame;
227 bool canceled = false;
228 unsigned long flags;
229 LIST_HEAD(done);
230
231 spin_lock_irqsave(&ring->lock, flags);
232
233 if (!ring->running) {
234 /* Move all frames to done and mark them as canceled. */
235 list_splice_tail_init(&ring->in_flight, &done);
236 list_splice_tail_init(&ring->queue, &done);
237 canceled = true;
238 goto invoke_callback;
239 }
240
241 while (!ring_empty(ring)) {
242 if (!(ring->descriptors[ring->tail].flags
243 & RING_DESC_COMPLETED))
244 break;
245 frame = list_first_entry(&ring->in_flight, typeof(*frame),
246 list);
247 list_move_tail(&frame->list, &done);
248 if (!ring->is_tx) {
249 frame->size = ring->descriptors[ring->tail].length;
250 frame->eof = ring->descriptors[ring->tail].eof;
251 frame->sof = ring->descriptors[ring->tail].sof;
252 frame->flags = ring->descriptors[ring->tail].flags;
253 }
254 ring->tail = (ring->tail + 1) % ring->size;
255 }
256 ring_write_descriptors(ring);
257
258invoke_callback:
259 /* allow callbacks to schedule new work */
260 spin_unlock_irqrestore(&ring->lock, flags);
261 while (!list_empty(&done)) {
262 frame = list_first_entry(&done, typeof(*frame), list);
263 /*
264 * The callback may reenqueue or delete frame.
265 * Do not hold on to it.
266 */
267 list_del_init(&frame->list);
268 if (frame->callback)
269 frame->callback(ring, frame, canceled);
270 }
271}
272
273int __tb_ring_enqueue(struct tb_ring *ring, struct ring_frame *frame)
274{
275 unsigned long flags;
276 int ret = 0;
277
278 spin_lock_irqsave(&ring->lock, flags);
279 if (ring->running) {
280 list_add_tail(&frame->list, &ring->queue);
281 ring_write_descriptors(ring);
282 } else {
283 ret = -ESHUTDOWN;
284 }
285 spin_unlock_irqrestore(&ring->lock, flags);
286 return ret;
287}
288EXPORT_SYMBOL_GPL(__tb_ring_enqueue);
289
290/**
291 * tb_ring_poll() - Poll one completed frame from the ring
292 * @ring: Ring to poll
293 *
294 * This function can be called when @start_poll callback of the @ring
295 * has been called. It will read one completed frame from the ring and
296 * return it to the caller. Returns %NULL if there is no more completed
297 * frames.
298 */
299struct ring_frame *tb_ring_poll(struct tb_ring *ring)
300{
301 struct ring_frame *frame = NULL;
302 unsigned long flags;
303
304 spin_lock_irqsave(&ring->lock, flags);
305 if (!ring->running)
306 goto unlock;
307 if (ring_empty(ring))
308 goto unlock;
309
310 if (ring->descriptors[ring->tail].flags & RING_DESC_COMPLETED) {
311 frame = list_first_entry(&ring->in_flight, typeof(*frame),
312 list);
313 list_del_init(&frame->list);
314
315 if (!ring->is_tx) {
316 frame->size = ring->descriptors[ring->tail].length;
317 frame->eof = ring->descriptors[ring->tail].eof;
318 frame->sof = ring->descriptors[ring->tail].sof;
319 frame->flags = ring->descriptors[ring->tail].flags;
320 }
321
322 ring->tail = (ring->tail + 1) % ring->size;
323 }
324
325unlock:
326 spin_unlock_irqrestore(&ring->lock, flags);
327 return frame;
328}
329EXPORT_SYMBOL_GPL(tb_ring_poll);
330
331static void __ring_interrupt_mask(struct tb_ring *ring, bool mask)
332{
333 int idx = ring_interrupt_index(ring);
334 int reg = REG_RING_INTERRUPT_BASE + idx / 32 * 4;
335 int bit = idx % 32;
336 u32 val;
337
338 val = ioread32(ring->nhi->iobase + reg);
339 if (mask)
340 val &= ~BIT(bit);
341 else
342 val |= BIT(bit);
343 iowrite32(val, ring->nhi->iobase + reg);
344}
345
346/* Both @nhi->lock and @ring->lock should be held */
347static void __ring_interrupt(struct tb_ring *ring)
348{
349 if (!ring->running)
350 return;
351
352 if (ring->start_poll) {
353 __ring_interrupt_mask(ring, true);
354 ring->start_poll(ring->poll_data);
355 } else {
356 schedule_work(&ring->work);
357 }
358}
359
360/**
361 * tb_ring_poll_complete() - Re-start interrupt for the ring
362 * @ring: Ring to re-start the interrupt
363 *
364 * This will re-start (unmask) the ring interrupt once the user is done
365 * with polling.
366 */
367void tb_ring_poll_complete(struct tb_ring *ring)
368{
369 unsigned long flags;
370
371 spin_lock_irqsave(&ring->nhi->lock, flags);
372 spin_lock(&ring->lock);
373 if (ring->start_poll)
374 __ring_interrupt_mask(ring, false);
375 spin_unlock(&ring->lock);
376 spin_unlock_irqrestore(&ring->nhi->lock, flags);
377}
378EXPORT_SYMBOL_GPL(tb_ring_poll_complete);
379
380static irqreturn_t ring_msix(int irq, void *data)
381{
382 struct tb_ring *ring = data;
383
384 spin_lock(&ring->nhi->lock);
385 spin_lock(&ring->lock);
386 __ring_interrupt(ring);
387 spin_unlock(&ring->lock);
388 spin_unlock(&ring->nhi->lock);
389
390 return IRQ_HANDLED;
391}
392
393static int ring_request_msix(struct tb_ring *ring, bool no_suspend)
394{
395 struct tb_nhi *nhi = ring->nhi;
396 unsigned long irqflags;
397 int ret;
398
399 if (!nhi->pdev->msix_enabled)
400 return 0;
401
402 ret = ida_simple_get(&nhi->msix_ida, 0, MSIX_MAX_VECS, GFP_KERNEL);
403 if (ret < 0)
404 return ret;
405
406 ring->vector = ret;
407
408 ret = pci_irq_vector(ring->nhi->pdev, ring->vector);
409 if (ret < 0)
410 goto err_ida_remove;
411
412 ring->irq = ret;
413
414 irqflags = no_suspend ? IRQF_NO_SUSPEND : 0;
415 ret = request_irq(ring->irq, ring_msix, irqflags, "thunderbolt", ring);
416 if (ret)
417 goto err_ida_remove;
418
419 return 0;
420
421err_ida_remove:
422 ida_simple_remove(&nhi->msix_ida, ring->vector);
423
424 return ret;
425}
426
427static void ring_release_msix(struct tb_ring *ring)
428{
429 if (ring->irq <= 0)
430 return;
431
432 free_irq(ring->irq, ring);
433 ida_simple_remove(&ring->nhi->msix_ida, ring->vector);
434 ring->vector = 0;
435 ring->irq = 0;
436}
437
438static int nhi_alloc_hop(struct tb_nhi *nhi, struct tb_ring *ring)
439{
440 int ret = 0;
441
442 spin_lock_irq(&nhi->lock);
443
444 if (ring->hop < 0) {
445 unsigned int i;
446
447 /*
448 * Automatically allocate HopID from the non-reserved
449 * range 1 .. hop_count - 1.
450 */
451 for (i = RING_FIRST_USABLE_HOPID; i < nhi->hop_count; i++) {
452 if (ring->is_tx) {
453 if (!nhi->tx_rings[i]) {
454 ring->hop = i;
455 break;
456 }
457 } else {
458 if (!nhi->rx_rings[i]) {
459 ring->hop = i;
460 break;
461 }
462 }
463 }
464 }
465
466 if (ring->hop < 0 || ring->hop >= nhi->hop_count) {
467 dev_warn(&nhi->pdev->dev, "invalid hop: %d\n", ring->hop);
468 ret = -EINVAL;
469 goto err_unlock;
470 }
471 if (ring->is_tx && nhi->tx_rings[ring->hop]) {
472 dev_warn(&nhi->pdev->dev, "TX hop %d already allocated\n",
473 ring->hop);
474 ret = -EBUSY;
475 goto err_unlock;
476 } else if (!ring->is_tx && nhi->rx_rings[ring->hop]) {
477 dev_warn(&nhi->pdev->dev, "RX hop %d already allocated\n",
478 ring->hop);
479 ret = -EBUSY;
480 goto err_unlock;
481 }
482
483 if (ring->is_tx)
484 nhi->tx_rings[ring->hop] = ring;
485 else
486 nhi->rx_rings[ring->hop] = ring;
487
488err_unlock:
489 spin_unlock_irq(&nhi->lock);
490
491 return ret;
492}
493
494static struct tb_ring *tb_ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
495 bool transmit, unsigned int flags,
496 int e2e_tx_hop, u16 sof_mask, u16 eof_mask,
497 void (*start_poll)(void *),
498 void *poll_data)
499{
500 struct tb_ring *ring = NULL;
501
502 dev_dbg(&nhi->pdev->dev, "allocating %s ring %d of size %d\n",
503 transmit ? "TX" : "RX", hop, size);
504
505 ring = kzalloc(sizeof(*ring), GFP_KERNEL);
506 if (!ring)
507 return NULL;
508
509 spin_lock_init(&ring->lock);
510 INIT_LIST_HEAD(&ring->queue);
511 INIT_LIST_HEAD(&ring->in_flight);
512 INIT_WORK(&ring->work, ring_work);
513
514 ring->nhi = nhi;
515 ring->hop = hop;
516 ring->is_tx = transmit;
517 ring->size = size;
518 ring->flags = flags;
519 ring->e2e_tx_hop = e2e_tx_hop;
520 ring->sof_mask = sof_mask;
521 ring->eof_mask = eof_mask;
522 ring->head = 0;
523 ring->tail = 0;
524 ring->running = false;
525 ring->start_poll = start_poll;
526 ring->poll_data = poll_data;
527
528 ring->descriptors = dma_alloc_coherent(&ring->nhi->pdev->dev,
529 size * sizeof(*ring->descriptors),
530 &ring->descriptors_dma, GFP_KERNEL | __GFP_ZERO);
531 if (!ring->descriptors)
532 goto err_free_ring;
533
534 if (ring_request_msix(ring, flags & RING_FLAG_NO_SUSPEND))
535 goto err_free_descs;
536
537 if (nhi_alloc_hop(nhi, ring))
538 goto err_release_msix;
539
540 return ring;
541
542err_release_msix:
543 ring_release_msix(ring);
544err_free_descs:
545 dma_free_coherent(&ring->nhi->pdev->dev,
546 ring->size * sizeof(*ring->descriptors),
547 ring->descriptors, ring->descriptors_dma);
548err_free_ring:
549 kfree(ring);
550
551 return NULL;
552}
553
554/**
555 * tb_ring_alloc_tx() - Allocate DMA ring for transmit
556 * @nhi: Pointer to the NHI the ring is to be allocated
557 * @hop: HopID (ring) to allocate
558 * @size: Number of entries in the ring
559 * @flags: Flags for the ring
560 */
561struct tb_ring *tb_ring_alloc_tx(struct tb_nhi *nhi, int hop, int size,
562 unsigned int flags)
563{
564 return tb_ring_alloc(nhi, hop, size, true, flags, 0, 0, 0, NULL, NULL);
565}
566EXPORT_SYMBOL_GPL(tb_ring_alloc_tx);
567
568/**
569 * tb_ring_alloc_rx() - Allocate DMA ring for receive
570 * @nhi: Pointer to the NHI the ring is to be allocated
571 * @hop: HopID (ring) to allocate. Pass %-1 for automatic allocation.
572 * @size: Number of entries in the ring
573 * @flags: Flags for the ring
574 * @e2e_tx_hop: Transmit HopID when E2E is enabled in @flags
575 * @sof_mask: Mask of PDF values that start a frame
576 * @eof_mask: Mask of PDF values that end a frame
577 * @start_poll: If not %NULL the ring will call this function when an
578 * interrupt is triggered and masked, instead of callback
579 * in each Rx frame.
580 * @poll_data: Optional data passed to @start_poll
581 */
582struct tb_ring *tb_ring_alloc_rx(struct tb_nhi *nhi, int hop, int size,
583 unsigned int flags, int e2e_tx_hop,
584 u16 sof_mask, u16 eof_mask,
585 void (*start_poll)(void *), void *poll_data)
586{
587 return tb_ring_alloc(nhi, hop, size, false, flags, e2e_tx_hop, sof_mask, eof_mask,
588 start_poll, poll_data);
589}
590EXPORT_SYMBOL_GPL(tb_ring_alloc_rx);
591
592/**
593 * tb_ring_start() - enable a ring
594 * @ring: Ring to start
595 *
596 * Must not be invoked in parallel with tb_ring_stop().
597 */
598void tb_ring_start(struct tb_ring *ring)
599{
600 u16 frame_size;
601 u32 flags;
602
603 spin_lock_irq(&ring->nhi->lock);
604 spin_lock(&ring->lock);
605 if (ring->nhi->going_away)
606 goto err;
607 if (ring->running) {
608 dev_WARN(&ring->nhi->pdev->dev, "ring already started\n");
609 goto err;
610 }
611 dev_dbg(&ring->nhi->pdev->dev, "starting %s %d\n",
612 RING_TYPE(ring), ring->hop);
613
614 if (ring->flags & RING_FLAG_FRAME) {
615 /* Means 4096 */
616 frame_size = 0;
617 flags = RING_FLAG_ENABLE;
618 } else {
619 frame_size = TB_FRAME_SIZE;
620 flags = RING_FLAG_ENABLE | RING_FLAG_RAW;
621 }
622
623 ring_iowrite64desc(ring, ring->descriptors_dma, 0);
624 if (ring->is_tx) {
625 ring_iowrite32desc(ring, ring->size, 12);
626 ring_iowrite32options(ring, 0, 4); /* time releated ? */
627 ring_iowrite32options(ring, flags, 0);
628 } else {
629 u32 sof_eof_mask = ring->sof_mask << 16 | ring->eof_mask;
630
631 ring_iowrite32desc(ring, (frame_size << 16) | ring->size, 12);
632 ring_iowrite32options(ring, sof_eof_mask, 4);
633 ring_iowrite32options(ring, flags, 0);
634 }
635
636 /*
637 * Now that the ring valid bit is set we can configure E2E if
638 * enabled for the ring.
639 */
640 if (ring->flags & RING_FLAG_E2E) {
641 if (!ring->is_tx) {
642 u32 hop;
643
644 hop = ring->e2e_tx_hop << REG_RX_OPTIONS_E2E_HOP_SHIFT;
645 hop &= REG_RX_OPTIONS_E2E_HOP_MASK;
646 flags |= hop;
647
648 dev_dbg(&ring->nhi->pdev->dev,
649 "enabling E2E for %s %d with TX HopID %d\n",
650 RING_TYPE(ring), ring->hop, ring->e2e_tx_hop);
651 } else {
652 dev_dbg(&ring->nhi->pdev->dev, "enabling E2E for %s %d\n",
653 RING_TYPE(ring), ring->hop);
654 }
655
656 flags |= RING_FLAG_E2E_FLOW_CONTROL;
657 ring_iowrite32options(ring, flags, 0);
658 }
659
660 ring_interrupt_active(ring, true);
661 ring->running = true;
662err:
663 spin_unlock(&ring->lock);
664 spin_unlock_irq(&ring->nhi->lock);
665}
666EXPORT_SYMBOL_GPL(tb_ring_start);
667
668/**
669 * tb_ring_stop() - shutdown a ring
670 * @ring: Ring to stop
671 *
672 * Must not be invoked from a callback.
673 *
674 * This method will disable the ring. Further calls to
675 * tb_ring_tx/tb_ring_rx will return -ESHUTDOWN until ring_stop has been
676 * called.
677 *
678 * All enqueued frames will be canceled and their callbacks will be executed
679 * with frame->canceled set to true (on the callback thread). This method
680 * returns only after all callback invocations have finished.
681 */
682void tb_ring_stop(struct tb_ring *ring)
683{
684 spin_lock_irq(&ring->nhi->lock);
685 spin_lock(&ring->lock);
686 dev_dbg(&ring->nhi->pdev->dev, "stopping %s %d\n",
687 RING_TYPE(ring), ring->hop);
688 if (ring->nhi->going_away)
689 goto err;
690 if (!ring->running) {
691 dev_WARN(&ring->nhi->pdev->dev, "%s %d already stopped\n",
692 RING_TYPE(ring), ring->hop);
693 goto err;
694 }
695 ring_interrupt_active(ring, false);
696
697 ring_iowrite32options(ring, 0, 0);
698 ring_iowrite64desc(ring, 0, 0);
699 ring_iowrite32desc(ring, 0, 8);
700 ring_iowrite32desc(ring, 0, 12);
701 ring->head = 0;
702 ring->tail = 0;
703 ring->running = false;
704
705err:
706 spin_unlock(&ring->lock);
707 spin_unlock_irq(&ring->nhi->lock);
708
709 /*
710 * schedule ring->work to invoke callbacks on all remaining frames.
711 */
712 schedule_work(&ring->work);
713 flush_work(&ring->work);
714}
715EXPORT_SYMBOL_GPL(tb_ring_stop);
716
717/*
718 * tb_ring_free() - free ring
719 *
720 * When this method returns all invocations of ring->callback will have
721 * finished.
722 *
723 * Ring must be stopped.
724 *
725 * Must NOT be called from ring_frame->callback!
726 */
727void tb_ring_free(struct tb_ring *ring)
728{
729 spin_lock_irq(&ring->nhi->lock);
730 /*
731 * Dissociate the ring from the NHI. This also ensures that
732 * nhi_interrupt_work cannot reschedule ring->work.
733 */
734 if (ring->is_tx)
735 ring->nhi->tx_rings[ring->hop] = NULL;
736 else
737 ring->nhi->rx_rings[ring->hop] = NULL;
738
739 if (ring->running) {
740 dev_WARN(&ring->nhi->pdev->dev, "%s %d still running\n",
741 RING_TYPE(ring), ring->hop);
742 }
743 spin_unlock_irq(&ring->nhi->lock);
744
745 ring_release_msix(ring);
746
747 dma_free_coherent(&ring->nhi->pdev->dev,
748 ring->size * sizeof(*ring->descriptors),
749 ring->descriptors, ring->descriptors_dma);
750
751 ring->descriptors = NULL;
752 ring->descriptors_dma = 0;
753
754
755 dev_dbg(&ring->nhi->pdev->dev, "freeing %s %d\n", RING_TYPE(ring),
756 ring->hop);
757
758 /*
759 * ring->work can no longer be scheduled (it is scheduled only
760 * by nhi_interrupt_work, ring_stop and ring_msix). Wait for it
761 * to finish before freeing the ring.
762 */
763 flush_work(&ring->work);
764 kfree(ring);
765}
766EXPORT_SYMBOL_GPL(tb_ring_free);
767
768/**
769 * nhi_mailbox_cmd() - Send a command through NHI mailbox
770 * @nhi: Pointer to the NHI structure
771 * @cmd: Command to send
772 * @data: Data to be send with the command
773 *
774 * Sends mailbox command to the firmware running on NHI. Returns %0 in
775 * case of success and negative errno in case of failure.
776 */
777int nhi_mailbox_cmd(struct tb_nhi *nhi, enum nhi_mailbox_cmd cmd, u32 data)
778{
779 ktime_t timeout;
780 u32 val;
781
782 iowrite32(data, nhi->iobase + REG_INMAIL_DATA);
783
784 val = ioread32(nhi->iobase + REG_INMAIL_CMD);
785 val &= ~(REG_INMAIL_CMD_MASK | REG_INMAIL_ERROR);
786 val |= REG_INMAIL_OP_REQUEST | cmd;
787 iowrite32(val, nhi->iobase + REG_INMAIL_CMD);
788
789 timeout = ktime_add_ms(ktime_get(), NHI_MAILBOX_TIMEOUT);
790 do {
791 val = ioread32(nhi->iobase + REG_INMAIL_CMD);
792 if (!(val & REG_INMAIL_OP_REQUEST))
793 break;
794 usleep_range(10, 20);
795 } while (ktime_before(ktime_get(), timeout));
796
797 if (val & REG_INMAIL_OP_REQUEST)
798 return -ETIMEDOUT;
799 if (val & REG_INMAIL_ERROR)
800 return -EIO;
801
802 return 0;
803}
804
805/**
806 * nhi_mailbox_mode() - Return current firmware operation mode
807 * @nhi: Pointer to the NHI structure
808 *
809 * The function reads current firmware operation mode using NHI mailbox
810 * registers and returns it to the caller.
811 */
812enum nhi_fw_mode nhi_mailbox_mode(struct tb_nhi *nhi)
813{
814 u32 val;
815
816 val = ioread32(nhi->iobase + REG_OUTMAIL_CMD);
817 val &= REG_OUTMAIL_CMD_OPMODE_MASK;
818 val >>= REG_OUTMAIL_CMD_OPMODE_SHIFT;
819
820 return (enum nhi_fw_mode)val;
821}
822
823static void nhi_interrupt_work(struct work_struct *work)
824{
825 struct tb_nhi *nhi = container_of(work, typeof(*nhi), interrupt_work);
826 int value = 0; /* Suppress uninitialized usage warning. */
827 int bit;
828 int hop = -1;
829 int type = 0; /* current interrupt type 0: TX, 1: RX, 2: RX overflow */
830 struct tb_ring *ring;
831
832 spin_lock_irq(&nhi->lock);
833
834 /*
835 * Starting at REG_RING_NOTIFY_BASE there are three status bitfields
836 * (TX, RX, RX overflow). We iterate over the bits and read a new
837 * dwords as required. The registers are cleared on read.
838 */
839 for (bit = 0; bit < 3 * nhi->hop_count; bit++) {
840 if (bit % 32 == 0)
841 value = ioread32(nhi->iobase
842 + REG_RING_NOTIFY_BASE
843 + 4 * (bit / 32));
844 if (++hop == nhi->hop_count) {
845 hop = 0;
846 type++;
847 }
848 if ((value & (1 << (bit % 32))) == 0)
849 continue;
850 if (type == 2) {
851 dev_warn(&nhi->pdev->dev,
852 "RX overflow for ring %d\n",
853 hop);
854 continue;
855 }
856 if (type == 0)
857 ring = nhi->tx_rings[hop];
858 else
859 ring = nhi->rx_rings[hop];
860 if (ring == NULL) {
861 dev_warn(&nhi->pdev->dev,
862 "got interrupt for inactive %s ring %d\n",
863 type ? "RX" : "TX",
864 hop);
865 continue;
866 }
867
868 spin_lock(&ring->lock);
869 __ring_interrupt(ring);
870 spin_unlock(&ring->lock);
871 }
872 spin_unlock_irq(&nhi->lock);
873}
874
875static irqreturn_t nhi_msi(int irq, void *data)
876{
877 struct tb_nhi *nhi = data;
878 schedule_work(&nhi->interrupt_work);
879 return IRQ_HANDLED;
880}
881
882static int __nhi_suspend_noirq(struct device *dev, bool wakeup)
883{
884 struct pci_dev *pdev = to_pci_dev(dev);
885 struct tb *tb = pci_get_drvdata(pdev);
886 struct tb_nhi *nhi = tb->nhi;
887 int ret;
888
889 ret = tb_domain_suspend_noirq(tb);
890 if (ret)
891 return ret;
892
893 if (nhi->ops && nhi->ops->suspend_noirq) {
894 ret = nhi->ops->suspend_noirq(tb->nhi, wakeup);
895 if (ret)
896 return ret;
897 }
898
899 return 0;
900}
901
902static int nhi_suspend_noirq(struct device *dev)
903{
904 return __nhi_suspend_noirq(dev, device_may_wakeup(dev));
905}
906
907static int nhi_freeze_noirq(struct device *dev)
908{
909 struct pci_dev *pdev = to_pci_dev(dev);
910 struct tb *tb = pci_get_drvdata(pdev);
911
912 return tb_domain_freeze_noirq(tb);
913}
914
915static int nhi_thaw_noirq(struct device *dev)
916{
917 struct pci_dev *pdev = to_pci_dev(dev);
918 struct tb *tb = pci_get_drvdata(pdev);
919
920 return tb_domain_thaw_noirq(tb);
921}
922
923static bool nhi_wake_supported(struct pci_dev *pdev)
924{
925 u8 val;
926
927 /*
928 * If power rails are sustainable for wakeup from S4 this
929 * property is set by the BIOS.
930 */
931 if (device_property_read_u8(&pdev->dev, "WAKE_SUPPORTED", &val))
932 return !!val;
933
934 return true;
935}
936
937static int nhi_poweroff_noirq(struct device *dev)
938{
939 struct pci_dev *pdev = to_pci_dev(dev);
940 bool wakeup;
941
942 wakeup = device_may_wakeup(dev) && nhi_wake_supported(pdev);
943 return __nhi_suspend_noirq(dev, wakeup);
944}
945
946static void nhi_enable_int_throttling(struct tb_nhi *nhi)
947{
948 /* Throttling is specified in 256ns increments */
949 u32 throttle = DIV_ROUND_UP(128 * NSEC_PER_USEC, 256);
950 unsigned int i;
951
952 /*
953 * Configure interrupt throttling for all vectors even if we
954 * only use few.
955 */
956 for (i = 0; i < MSIX_MAX_VECS; i++) {
957 u32 reg = REG_INT_THROTTLING_RATE + i * 4;
958 iowrite32(throttle, nhi->iobase + reg);
959 }
960}
961
962static int nhi_resume_noirq(struct device *dev)
963{
964 struct pci_dev *pdev = to_pci_dev(dev);
965 struct tb *tb = pci_get_drvdata(pdev);
966 struct tb_nhi *nhi = tb->nhi;
967 int ret;
968
969 /*
970 * Check that the device is still there. It may be that the user
971 * unplugged last device which causes the host controller to go
972 * away on PCs.
973 */
974 if (!pci_device_is_present(pdev)) {
975 nhi->going_away = true;
976 } else {
977 if (nhi->ops && nhi->ops->resume_noirq) {
978 ret = nhi->ops->resume_noirq(nhi);
979 if (ret)
980 return ret;
981 }
982 nhi_enable_int_throttling(tb->nhi);
983 }
984
985 return tb_domain_resume_noirq(tb);
986}
987
988static int nhi_suspend(struct device *dev)
989{
990 struct pci_dev *pdev = to_pci_dev(dev);
991 struct tb *tb = pci_get_drvdata(pdev);
992
993 return tb_domain_suspend(tb);
994}
995
996static void nhi_complete(struct device *dev)
997{
998 struct pci_dev *pdev = to_pci_dev(dev);
999 struct tb *tb = pci_get_drvdata(pdev);
1000
1001 /*
1002 * If we were runtime suspended when system suspend started,
1003 * schedule runtime resume now. It should bring the domain back
1004 * to functional state.
1005 */
1006 if (pm_runtime_suspended(&pdev->dev))
1007 pm_runtime_resume(&pdev->dev);
1008 else
1009 tb_domain_complete(tb);
1010}
1011
1012static int nhi_runtime_suspend(struct device *dev)
1013{
1014 struct pci_dev *pdev = to_pci_dev(dev);
1015 struct tb *tb = pci_get_drvdata(pdev);
1016 struct tb_nhi *nhi = tb->nhi;
1017 int ret;
1018
1019 ret = tb_domain_runtime_suspend(tb);
1020 if (ret)
1021 return ret;
1022
1023 if (nhi->ops && nhi->ops->runtime_suspend) {
1024 ret = nhi->ops->runtime_suspend(tb->nhi);
1025 if (ret)
1026 return ret;
1027 }
1028 return 0;
1029}
1030
1031static int nhi_runtime_resume(struct device *dev)
1032{
1033 struct pci_dev *pdev = to_pci_dev(dev);
1034 struct tb *tb = pci_get_drvdata(pdev);
1035 struct tb_nhi *nhi = tb->nhi;
1036 int ret;
1037
1038 if (nhi->ops && nhi->ops->runtime_resume) {
1039 ret = nhi->ops->runtime_resume(nhi);
1040 if (ret)
1041 return ret;
1042 }
1043
1044 nhi_enable_int_throttling(nhi);
1045 return tb_domain_runtime_resume(tb);
1046}
1047
1048static void nhi_shutdown(struct tb_nhi *nhi)
1049{
1050 int i;
1051
1052 dev_dbg(&nhi->pdev->dev, "shutdown\n");
1053
1054 for (i = 0; i < nhi->hop_count; i++) {
1055 if (nhi->tx_rings[i])
1056 dev_WARN(&nhi->pdev->dev,
1057 "TX ring %d is still active\n", i);
1058 if (nhi->rx_rings[i])
1059 dev_WARN(&nhi->pdev->dev,
1060 "RX ring %d is still active\n", i);
1061 }
1062 nhi_disable_interrupts(nhi);
1063 /*
1064 * We have to release the irq before calling flush_work. Otherwise an
1065 * already executing IRQ handler could call schedule_work again.
1066 */
1067 if (!nhi->pdev->msix_enabled) {
1068 devm_free_irq(&nhi->pdev->dev, nhi->pdev->irq, nhi);
1069 flush_work(&nhi->interrupt_work);
1070 }
1071 ida_destroy(&nhi->msix_ida);
1072
1073 if (nhi->ops && nhi->ops->shutdown)
1074 nhi->ops->shutdown(nhi);
1075}
1076
1077static int nhi_init_msi(struct tb_nhi *nhi)
1078{
1079 struct pci_dev *pdev = nhi->pdev;
1080 int res, irq, nvec;
1081
1082 /* In case someone left them on. */
1083 nhi_disable_interrupts(nhi);
1084
1085 nhi_enable_int_throttling(nhi);
1086
1087 ida_init(&nhi->msix_ida);
1088
1089 /*
1090 * The NHI has 16 MSI-X vectors or a single MSI. We first try to
1091 * get all MSI-X vectors and if we succeed, each ring will have
1092 * one MSI-X. If for some reason that does not work out, we
1093 * fallback to a single MSI.
1094 */
1095 nvec = pci_alloc_irq_vectors(pdev, MSIX_MIN_VECS, MSIX_MAX_VECS,
1096 PCI_IRQ_MSIX);
1097 if (nvec < 0) {
1098 nvec = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI);
1099 if (nvec < 0)
1100 return nvec;
1101
1102 INIT_WORK(&nhi->interrupt_work, nhi_interrupt_work);
1103
1104 irq = pci_irq_vector(nhi->pdev, 0);
1105 if (irq < 0)
1106 return irq;
1107
1108 res = devm_request_irq(&pdev->dev, irq, nhi_msi,
1109 IRQF_NO_SUSPEND, "thunderbolt", nhi);
1110 if (res) {
1111 dev_err(&pdev->dev, "request_irq failed, aborting\n");
1112 return res;
1113 }
1114 }
1115
1116 return 0;
1117}
1118
1119static bool nhi_imr_valid(struct pci_dev *pdev)
1120{
1121 u8 val;
1122
1123 if (!device_property_read_u8(&pdev->dev, "IMR_VALID", &val))
1124 return !!val;
1125
1126 return true;
1127}
1128
1129static struct tb *nhi_select_cm(struct tb_nhi *nhi)
1130{
1131 struct tb *tb;
1132
1133 /*
1134 * USB4 case is simple. If we got control of any of the
1135 * capabilities, we use software CM.
1136 */
1137 if (tb_acpi_is_native())
1138 return tb_probe(nhi);
1139
1140 /*
1141 * Either firmware based CM is running (we did not get control
1142 * from the firmware) or this is pre-USB4 PC so try first
1143 * firmware CM and then fallback to software CM.
1144 */
1145 tb = icm_probe(nhi);
1146 if (!tb)
1147 tb = tb_probe(nhi);
1148
1149 return tb;
1150}
1151
1152static int nhi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1153{
1154 struct tb_nhi *nhi;
1155 struct tb *tb;
1156 int res;
1157
1158 if (!nhi_imr_valid(pdev)) {
1159 dev_warn(&pdev->dev, "firmware image not valid, aborting\n");
1160 return -ENODEV;
1161 }
1162
1163 res = pcim_enable_device(pdev);
1164 if (res) {
1165 dev_err(&pdev->dev, "cannot enable PCI device, aborting\n");
1166 return res;
1167 }
1168
1169 res = pcim_iomap_regions(pdev, 1 << 0, "thunderbolt");
1170 if (res) {
1171 dev_err(&pdev->dev, "cannot obtain PCI resources, aborting\n");
1172 return res;
1173 }
1174
1175 nhi = devm_kzalloc(&pdev->dev, sizeof(*nhi), GFP_KERNEL);
1176 if (!nhi)
1177 return -ENOMEM;
1178
1179 nhi->pdev = pdev;
1180 nhi->ops = (const struct tb_nhi_ops *)id->driver_data;
1181 /* cannot fail - table is allocated bin pcim_iomap_regions */
1182 nhi->iobase = pcim_iomap_table(pdev)[0];
1183 nhi->hop_count = ioread32(nhi->iobase + REG_HOP_COUNT) & 0x3ff;
1184 dev_dbg(&pdev->dev, "total paths: %d\n", nhi->hop_count);
1185
1186 nhi->tx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1187 sizeof(*nhi->tx_rings), GFP_KERNEL);
1188 nhi->rx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1189 sizeof(*nhi->rx_rings), GFP_KERNEL);
1190 if (!nhi->tx_rings || !nhi->rx_rings)
1191 return -ENOMEM;
1192
1193 res = nhi_init_msi(nhi);
1194 if (res) {
1195 dev_err(&pdev->dev, "cannot enable MSI, aborting\n");
1196 return res;
1197 }
1198
1199 spin_lock_init(&nhi->lock);
1200
1201 res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1202 if (res)
1203 res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
1204 if (res) {
1205 dev_err(&pdev->dev, "failed to set DMA mask\n");
1206 return res;
1207 }
1208
1209 pci_set_master(pdev);
1210
1211 if (nhi->ops && nhi->ops->init) {
1212 res = nhi->ops->init(nhi);
1213 if (res)
1214 return res;
1215 }
1216
1217 tb = nhi_select_cm(nhi);
1218 if (!tb) {
1219 dev_err(&nhi->pdev->dev,
1220 "failed to determine connection manager, aborting\n");
1221 return -ENODEV;
1222 }
1223
1224 dev_dbg(&nhi->pdev->dev, "NHI initialized, starting thunderbolt\n");
1225
1226 res = tb_domain_add(tb);
1227 if (res) {
1228 /*
1229 * At this point the RX/TX rings might already have been
1230 * activated. Do a proper shutdown.
1231 */
1232 tb_domain_put(tb);
1233 nhi_shutdown(nhi);
1234 return res;
1235 }
1236 pci_set_drvdata(pdev, tb);
1237
1238 device_wakeup_enable(&pdev->dev);
1239
1240 pm_runtime_allow(&pdev->dev);
1241 pm_runtime_set_autosuspend_delay(&pdev->dev, TB_AUTOSUSPEND_DELAY);
1242 pm_runtime_use_autosuspend(&pdev->dev);
1243 pm_runtime_put_autosuspend(&pdev->dev);
1244
1245 return 0;
1246}
1247
1248static void nhi_remove(struct pci_dev *pdev)
1249{
1250 struct tb *tb = pci_get_drvdata(pdev);
1251 struct tb_nhi *nhi = tb->nhi;
1252
1253 pm_runtime_get_sync(&pdev->dev);
1254 pm_runtime_dont_use_autosuspend(&pdev->dev);
1255 pm_runtime_forbid(&pdev->dev);
1256
1257 tb_domain_remove(tb);
1258 nhi_shutdown(nhi);
1259}
1260
1261/*
1262 * The tunneled pci bridges are siblings of us. Use resume_noirq to reenable
1263 * the tunnels asap. A corresponding pci quirk blocks the downstream bridges
1264 * resume_noirq until we are done.
1265 */
1266static const struct dev_pm_ops nhi_pm_ops = {
1267 .suspend_noirq = nhi_suspend_noirq,
1268 .resume_noirq = nhi_resume_noirq,
1269 .freeze_noirq = nhi_freeze_noirq, /*
1270 * we just disable hotplug, the
1271 * pci-tunnels stay alive.
1272 */
1273 .thaw_noirq = nhi_thaw_noirq,
1274 .restore_noirq = nhi_resume_noirq,
1275 .suspend = nhi_suspend,
1276 .poweroff_noirq = nhi_poweroff_noirq,
1277 .poweroff = nhi_suspend,
1278 .complete = nhi_complete,
1279 .runtime_suspend = nhi_runtime_suspend,
1280 .runtime_resume = nhi_runtime_resume,
1281};
1282
1283static struct pci_device_id nhi_ids[] = {
1284 /*
1285 * We have to specify class, the TB bridges use the same device and
1286 * vendor (sub)id on gen 1 and gen 2 controllers.
1287 */
1288 {
1289 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1290 .vendor = PCI_VENDOR_ID_INTEL,
1291 .device = PCI_DEVICE_ID_INTEL_LIGHT_RIDGE,
1292 .subvendor = 0x2222, .subdevice = 0x1111,
1293 },
1294 {
1295 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1296 .vendor = PCI_VENDOR_ID_INTEL,
1297 .device = PCI_DEVICE_ID_INTEL_CACTUS_RIDGE_4C,
1298 .subvendor = 0x2222, .subdevice = 0x1111,
1299 },
1300 {
1301 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1302 .vendor = PCI_VENDOR_ID_INTEL,
1303 .device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI,
1304 .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1305 },
1306 {
1307 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1308 .vendor = PCI_VENDOR_ID_INTEL,
1309 .device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI,
1310 .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1311 },
1312
1313 /* Thunderbolt 3 */
1314 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_2C_NHI) },
1315 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_4C_NHI) },
1316 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_USBONLY_NHI) },
1317 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_NHI) },
1318 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_USBONLY_NHI) },
1319 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_2C_NHI) },
1320 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_4C_NHI) },
1321 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_USBONLY_NHI) },
1322 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI) },
1323 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI) },
1324 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ICL_NHI0),
1325 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1326 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ICL_NHI1),
1327 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1328 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_NHI0),
1329 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1330 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_NHI1),
1331 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1332 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_H_NHI0),
1333 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1334 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_H_NHI1),
1335 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1336 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ADL_NHI0),
1337 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1338 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ADL_NHI1),
1339 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1340
1341 /* Any USB4 compliant host */
1342 { PCI_DEVICE_CLASS(PCI_CLASS_SERIAL_USB_USB4, ~0) },
1343
1344 { 0,}
1345};
1346
1347MODULE_DEVICE_TABLE(pci, nhi_ids);
1348MODULE_LICENSE("GPL");
1349
1350static struct pci_driver nhi_driver = {
1351 .name = "thunderbolt",
1352 .id_table = nhi_ids,
1353 .probe = nhi_probe,
1354 .remove = nhi_remove,
1355 .shutdown = nhi_remove,
1356 .driver.pm = &nhi_pm_ops,
1357};
1358
1359static int __init nhi_init(void)
1360{
1361 int ret;
1362
1363 ret = tb_domain_init();
1364 if (ret)
1365 return ret;
1366 ret = pci_register_driver(&nhi_driver);
1367 if (ret)
1368 tb_domain_exit();
1369 return ret;
1370}
1371
1372static void __exit nhi_unload(void)
1373{
1374 pci_unregister_driver(&nhi_driver);
1375 tb_domain_exit();
1376}
1377
1378rootfs_initcall(nhi_init);
1379module_exit(nhi_unload);