Merge branch 'work.quota-compat' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux-2.6-microblaze.git] / drivers / thunderbolt / nhi.c
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
38 static 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  */
51 static 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  */
112 static 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
126 static 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
134 static 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
142 static 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
152 static 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
158 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)
159 {
160         iowrite32(value, ring_desc_base(ring) + offset);
161 }
162
163 static 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
169 static void ring_iowrite32options(struct tb_ring *ring, u32 value, u32 offset)
170 {
171         iowrite32(value, ring_options_base(ring) + offset);
172 }
173
174 static bool ring_full(struct tb_ring *ring)
175 {
176         return ((ring->head + 1) % ring->size) == ring->tail;
177 }
178
179 static 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  */
189 static 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  */
223 static 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
258 invoke_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
273 int __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 }
288 EXPORT_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  */
299 struct 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
325 unlock:
326         spin_unlock_irqrestore(&ring->lock, flags);
327         return frame;
328 }
329 EXPORT_SYMBOL_GPL(tb_ring_poll);
330
331 static 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 */
347 static 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  */
367 void 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 }
378 EXPORT_SYMBOL_GPL(tb_ring_poll_complete);
379
380 static 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
393 static 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
416 static 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
427 static 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
477 err_unlock:
478         spin_unlock_irq(&nhi->lock);
479
480         return ret;
481 }
482
483 static 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
530 err_release_msix:
531         ring_release_msix(ring);
532 err_free_descs:
533         dma_free_coherent(&ring->nhi->pdev->dev,
534                           ring->size * sizeof(*ring->descriptors),
535                           ring->descriptors, ring->descriptors_dma);
536 err_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  */
549 struct 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 }
554 EXPORT_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  */
569 struct 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 }
576 EXPORT_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  */
583 void 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;
622 err:
623         spin_unlock(&ring->lock);
624         spin_unlock_irq(&ring->nhi->lock);
625 }
626 EXPORT_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  */
641 void 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
664 err:
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 }
674 EXPORT_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  */
686 void 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 }
725 EXPORT_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  */
736 int 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  */
771 enum 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
782 static 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
834 static 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
841 static 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
861 static int nhi_suspend_noirq(struct device *dev)
862 {
863         return __nhi_suspend_noirq(dev, device_may_wakeup(dev));
864 }
865
866 static 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
880 static 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
889 static 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
905 static 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
931 static 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
939 static 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
955 static 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
974 static 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
991 static 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
1020 static 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
1062 static 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
1072 static 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
1168 static 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  */
1186 static 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
1204 static 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
1260 MODULE_DEVICE_TABLE(pci, nhi_ids);
1261 MODULE_LICENSE("GPL");
1262
1263 static 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
1272 static 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
1285 static void __exit nhi_unload(void)
1286 {
1287         pci_unregister_driver(&nhi_driver);
1288         tb_domain_exit();
1289 }
1290
1291 rootfs_initcall(nhi_init);
1292 module_exit(nhi_unload);