1 // SPDX-License-Identifier: GPL-2.0-only
3 * Thunderbolt driver - NHI driver
5 * The NHI (native host interface) is the pci device that allows us to send and
6 * receive frames from the thunderbolt bus.
8 * Copyright (c) 2014 Andreas Noever <andreas.noever@gmail.com>
9 * Copyright (C) 2018, Intel Corporation
12 #include <linux/pm_runtime.h>
13 #include <linux/slab.h>
14 #include <linux/errno.h>
15 #include <linux/pci.h>
16 #include <linux/dma-mapping.h>
17 #include <linux/interrupt.h>
18 #include <linux/iommu.h>
19 #include <linux/module.h>
20 #include <linux/delay.h>
21 #include <linux/property.h>
22 #include <linux/string_helpers.h>
28 #define RING_TYPE(ring) ((ring)->is_tx ? "TX ring" : "RX ring")
30 #define RING_FIRST_USABLE_HOPID 1
32 * Used with QUIRK_E2E to specify an unused HopID the Rx credits are
35 #define RING_E2E_RESERVED_HOPID RING_FIRST_USABLE_HOPID
37 * Minimal number of vectors when we use MSI-X. Two for control channel
38 * Rx/Tx and the rest four are for cross domain DMA paths.
40 #define MSIX_MIN_VECS 6
41 #define MSIX_MAX_VECS 16
43 #define NHI_MAILBOX_TIMEOUT 500 /* ms */
45 /* Host interface quirks */
46 #define QUIRK_AUTO_CLEAR_INT BIT(0)
47 #define QUIRK_E2E BIT(1)
49 static int ring_interrupt_index(struct tb_ring *ring)
53 bit += ring->nhi->hop_count;
58 * ring_interrupt_active() - activate/deactivate interrupts for a single ring
60 * ring->nhi->lock must be held.
62 static void ring_interrupt_active(struct tb_ring *ring, bool active)
64 int reg = REG_RING_INTERRUPT_BASE +
65 ring_interrupt_index(ring) / 32 * 4;
66 int bit = ring_interrupt_index(ring) & 31;
71 u32 step, shift, ivr, misc;
72 void __iomem *ivr_base;
78 index = ring->hop + ring->nhi->hop_count;
80 if (ring->nhi->quirks & QUIRK_AUTO_CLEAR_INT) {
82 * Ask the hardware to clear interrupt status
83 * bits automatically since we already know
84 * which interrupt was triggered.
86 misc = ioread32(ring->nhi->iobase + REG_DMA_MISC);
87 if (!(misc & REG_DMA_MISC_INT_AUTO_CLEAR)) {
88 misc |= REG_DMA_MISC_INT_AUTO_CLEAR;
89 iowrite32(misc, ring->nhi->iobase + REG_DMA_MISC);
93 ivr_base = ring->nhi->iobase + REG_INT_VEC_ALLOC_BASE;
94 step = index / REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
95 shift = index % REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
96 ivr = ioread32(ivr_base + step);
97 ivr &= ~(REG_INT_VEC_ALLOC_MASK << shift);
99 ivr |= ring->vector << shift;
100 iowrite32(ivr, ivr_base + step);
103 old = ioread32(ring->nhi->iobase + reg);
109 dev_dbg(&ring->nhi->pdev->dev,
110 "%s interrupt at register %#x bit %d (%#x -> %#x)\n",
111 active ? "enabling" : "disabling", reg, bit, old, new);
114 dev_WARN(&ring->nhi->pdev->dev,
115 "interrupt for %s %d is already %s\n",
116 RING_TYPE(ring), ring->hop,
117 active ? "enabled" : "disabled");
118 iowrite32(new, ring->nhi->iobase + reg);
122 * nhi_disable_interrupts() - disable interrupts for all rings
124 * Use only during init and shutdown.
126 static void nhi_disable_interrupts(struct tb_nhi *nhi)
129 /* disable interrupts */
130 for (i = 0; i < RING_INTERRUPT_REG_COUNT(nhi); i++)
131 iowrite32(0, nhi->iobase + REG_RING_INTERRUPT_BASE + 4 * i);
133 /* clear interrupt status bits */
134 for (i = 0; i < RING_NOTIFY_REG_COUNT(nhi); i++)
135 ioread32(nhi->iobase + REG_RING_NOTIFY_BASE + 4 * i);
138 /* ring helper methods */
140 static void __iomem *ring_desc_base(struct tb_ring *ring)
142 void __iomem *io = ring->nhi->iobase;
143 io += ring->is_tx ? REG_TX_RING_BASE : REG_RX_RING_BASE;
144 io += ring->hop * 16;
148 static void __iomem *ring_options_base(struct tb_ring *ring)
150 void __iomem *io = ring->nhi->iobase;
151 io += ring->is_tx ? REG_TX_OPTIONS_BASE : REG_RX_OPTIONS_BASE;
152 io += ring->hop * 32;
156 static void ring_iowrite_cons(struct tb_ring *ring, u16 cons)
159 * The other 16-bits in the register is read-only and writes to it
160 * are ignored by the hardware so we can save one ioread32() by
161 * filling the read-only bits with zeroes.
163 iowrite32(cons, ring_desc_base(ring) + 8);
166 static void ring_iowrite_prod(struct tb_ring *ring, u16 prod)
168 /* See ring_iowrite_cons() above for explanation */
169 iowrite32(prod << 16, ring_desc_base(ring) + 8);
172 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)
174 iowrite32(value, ring_desc_base(ring) + offset);
177 static void ring_iowrite64desc(struct tb_ring *ring, u64 value, u32 offset)
179 iowrite32(value, ring_desc_base(ring) + offset);
180 iowrite32(value >> 32, ring_desc_base(ring) + offset + 4);
183 static void ring_iowrite32options(struct tb_ring *ring, u32 value, u32 offset)
185 iowrite32(value, ring_options_base(ring) + offset);
188 static bool ring_full(struct tb_ring *ring)
190 return ((ring->head + 1) % ring->size) == ring->tail;
193 static bool ring_empty(struct tb_ring *ring)
195 return ring->head == ring->tail;
199 * ring_write_descriptors() - post frames from ring->queue to the controller
201 * ring->lock is held.
203 static void ring_write_descriptors(struct tb_ring *ring)
205 struct ring_frame *frame, *n;
206 struct ring_desc *descriptor;
207 list_for_each_entry_safe(frame, n, &ring->queue, list) {
210 list_move_tail(&frame->list, &ring->in_flight);
211 descriptor = &ring->descriptors[ring->head];
212 descriptor->phys = frame->buffer_phy;
213 descriptor->time = 0;
214 descriptor->flags = RING_DESC_POSTED | RING_DESC_INTERRUPT;
216 descriptor->length = frame->size;
217 descriptor->eof = frame->eof;
218 descriptor->sof = frame->sof;
220 ring->head = (ring->head + 1) % ring->size;
222 ring_iowrite_prod(ring, ring->head);
224 ring_iowrite_cons(ring, ring->head);
229 * ring_work() - progress completed frames
231 * If the ring is shutting down then all frames are marked as canceled and
232 * their callbacks are invoked.
234 * Otherwise we collect all completed frame from the ring buffer, write new
235 * frame to the ring buffer and invoke the callbacks for the completed frames.
237 static void ring_work(struct work_struct *work)
239 struct tb_ring *ring = container_of(work, typeof(*ring), work);
240 struct ring_frame *frame;
241 bool canceled = false;
245 spin_lock_irqsave(&ring->lock, flags);
247 if (!ring->running) {
248 /* Move all frames to done and mark them as canceled. */
249 list_splice_tail_init(&ring->in_flight, &done);
250 list_splice_tail_init(&ring->queue, &done);
252 goto invoke_callback;
255 while (!ring_empty(ring)) {
256 if (!(ring->descriptors[ring->tail].flags
257 & RING_DESC_COMPLETED))
259 frame = list_first_entry(&ring->in_flight, typeof(*frame),
261 list_move_tail(&frame->list, &done);
263 frame->size = ring->descriptors[ring->tail].length;
264 frame->eof = ring->descriptors[ring->tail].eof;
265 frame->sof = ring->descriptors[ring->tail].sof;
266 frame->flags = ring->descriptors[ring->tail].flags;
268 ring->tail = (ring->tail + 1) % ring->size;
270 ring_write_descriptors(ring);
273 /* allow callbacks to schedule new work */
274 spin_unlock_irqrestore(&ring->lock, flags);
275 while (!list_empty(&done)) {
276 frame = list_first_entry(&done, typeof(*frame), list);
278 * The callback may reenqueue or delete frame.
279 * Do not hold on to it.
281 list_del_init(&frame->list);
283 frame->callback(ring, frame, canceled);
287 int __tb_ring_enqueue(struct tb_ring *ring, struct ring_frame *frame)
292 spin_lock_irqsave(&ring->lock, flags);
294 list_add_tail(&frame->list, &ring->queue);
295 ring_write_descriptors(ring);
299 spin_unlock_irqrestore(&ring->lock, flags);
302 EXPORT_SYMBOL_GPL(__tb_ring_enqueue);
305 * tb_ring_poll() - Poll one completed frame from the ring
306 * @ring: Ring to poll
308 * This function can be called when @start_poll callback of the @ring
309 * has been called. It will read one completed frame from the ring and
310 * return it to the caller. Returns %NULL if there is no more completed
313 struct ring_frame *tb_ring_poll(struct tb_ring *ring)
315 struct ring_frame *frame = NULL;
318 spin_lock_irqsave(&ring->lock, flags);
321 if (ring_empty(ring))
324 if (ring->descriptors[ring->tail].flags & RING_DESC_COMPLETED) {
325 frame = list_first_entry(&ring->in_flight, typeof(*frame),
327 list_del_init(&frame->list);
330 frame->size = ring->descriptors[ring->tail].length;
331 frame->eof = ring->descriptors[ring->tail].eof;
332 frame->sof = ring->descriptors[ring->tail].sof;
333 frame->flags = ring->descriptors[ring->tail].flags;
336 ring->tail = (ring->tail + 1) % ring->size;
340 spin_unlock_irqrestore(&ring->lock, flags);
343 EXPORT_SYMBOL_GPL(tb_ring_poll);
345 static void __ring_interrupt_mask(struct tb_ring *ring, bool mask)
347 int idx = ring_interrupt_index(ring);
348 int reg = REG_RING_INTERRUPT_BASE + idx / 32 * 4;
352 val = ioread32(ring->nhi->iobase + reg);
357 iowrite32(val, ring->nhi->iobase + reg);
360 /* Both @nhi->lock and @ring->lock should be held */
361 static void __ring_interrupt(struct tb_ring *ring)
366 if (ring->start_poll) {
367 __ring_interrupt_mask(ring, true);
368 ring->start_poll(ring->poll_data);
370 schedule_work(&ring->work);
375 * tb_ring_poll_complete() - Re-start interrupt for the ring
376 * @ring: Ring to re-start the interrupt
378 * This will re-start (unmask) the ring interrupt once the user is done
381 void tb_ring_poll_complete(struct tb_ring *ring)
385 spin_lock_irqsave(&ring->nhi->lock, flags);
386 spin_lock(&ring->lock);
387 if (ring->start_poll)
388 __ring_interrupt_mask(ring, false);
389 spin_unlock(&ring->lock);
390 spin_unlock_irqrestore(&ring->nhi->lock, flags);
392 EXPORT_SYMBOL_GPL(tb_ring_poll_complete);
394 static void ring_clear_msix(const struct tb_ring *ring)
396 if (ring->nhi->quirks & QUIRK_AUTO_CLEAR_INT)
400 ioread32(ring->nhi->iobase + REG_RING_NOTIFY_BASE);
402 ioread32(ring->nhi->iobase + REG_RING_NOTIFY_BASE +
403 4 * (ring->nhi->hop_count / 32));
406 static irqreturn_t ring_msix(int irq, void *data)
408 struct tb_ring *ring = data;
410 spin_lock(&ring->nhi->lock);
411 ring_clear_msix(ring);
412 spin_lock(&ring->lock);
413 __ring_interrupt(ring);
414 spin_unlock(&ring->lock);
415 spin_unlock(&ring->nhi->lock);
420 static int ring_request_msix(struct tb_ring *ring, bool no_suspend)
422 struct tb_nhi *nhi = ring->nhi;
423 unsigned long irqflags;
426 if (!nhi->pdev->msix_enabled)
429 ret = ida_simple_get(&nhi->msix_ida, 0, MSIX_MAX_VECS, GFP_KERNEL);
435 ret = pci_irq_vector(ring->nhi->pdev, ring->vector);
441 irqflags = no_suspend ? IRQF_NO_SUSPEND : 0;
442 ret = request_irq(ring->irq, ring_msix, irqflags, "thunderbolt", ring);
449 ida_simple_remove(&nhi->msix_ida, ring->vector);
454 static void ring_release_msix(struct tb_ring *ring)
459 free_irq(ring->irq, ring);
460 ida_simple_remove(&ring->nhi->msix_ida, ring->vector);
465 static int nhi_alloc_hop(struct tb_nhi *nhi, struct tb_ring *ring)
467 unsigned int start_hop = RING_FIRST_USABLE_HOPID;
470 if (nhi->quirks & QUIRK_E2E) {
471 start_hop = RING_FIRST_USABLE_HOPID + 1;
472 if (ring->flags & RING_FLAG_E2E && !ring->is_tx) {
473 dev_dbg(&nhi->pdev->dev, "quirking E2E TX HopID %u -> %u\n",
474 ring->e2e_tx_hop, RING_E2E_RESERVED_HOPID);
475 ring->e2e_tx_hop = RING_E2E_RESERVED_HOPID;
479 spin_lock_irq(&nhi->lock);
485 * Automatically allocate HopID from the non-reserved
486 * range 1 .. hop_count - 1.
488 for (i = start_hop; i < nhi->hop_count; i++) {
490 if (!nhi->tx_rings[i]) {
495 if (!nhi->rx_rings[i]) {
503 if (ring->hop > 0 && ring->hop < start_hop) {
504 dev_warn(&nhi->pdev->dev, "invalid hop: %d\n", ring->hop);
508 if (ring->hop < 0 || ring->hop >= nhi->hop_count) {
509 dev_warn(&nhi->pdev->dev, "invalid hop: %d\n", ring->hop);
513 if (ring->is_tx && nhi->tx_rings[ring->hop]) {
514 dev_warn(&nhi->pdev->dev, "TX hop %d already allocated\n",
518 } else if (!ring->is_tx && nhi->rx_rings[ring->hop]) {
519 dev_warn(&nhi->pdev->dev, "RX hop %d already allocated\n",
526 nhi->tx_rings[ring->hop] = ring;
528 nhi->rx_rings[ring->hop] = ring;
531 spin_unlock_irq(&nhi->lock);
536 static struct tb_ring *tb_ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
537 bool transmit, unsigned int flags,
538 int e2e_tx_hop, u16 sof_mask, u16 eof_mask,
539 void (*start_poll)(void *),
542 struct tb_ring *ring = NULL;
544 dev_dbg(&nhi->pdev->dev, "allocating %s ring %d of size %d\n",
545 transmit ? "TX" : "RX", hop, size);
547 ring = kzalloc(sizeof(*ring), GFP_KERNEL);
551 spin_lock_init(&ring->lock);
552 INIT_LIST_HEAD(&ring->queue);
553 INIT_LIST_HEAD(&ring->in_flight);
554 INIT_WORK(&ring->work, ring_work);
558 ring->is_tx = transmit;
561 ring->e2e_tx_hop = e2e_tx_hop;
562 ring->sof_mask = sof_mask;
563 ring->eof_mask = eof_mask;
566 ring->running = false;
567 ring->start_poll = start_poll;
568 ring->poll_data = poll_data;
570 ring->descriptors = dma_alloc_coherent(&ring->nhi->pdev->dev,
571 size * sizeof(*ring->descriptors),
572 &ring->descriptors_dma, GFP_KERNEL | __GFP_ZERO);
573 if (!ring->descriptors)
576 if (ring_request_msix(ring, flags & RING_FLAG_NO_SUSPEND))
579 if (nhi_alloc_hop(nhi, ring))
580 goto err_release_msix;
585 ring_release_msix(ring);
587 dma_free_coherent(&ring->nhi->pdev->dev,
588 ring->size * sizeof(*ring->descriptors),
589 ring->descriptors, ring->descriptors_dma);
597 * tb_ring_alloc_tx() - Allocate DMA ring for transmit
598 * @nhi: Pointer to the NHI the ring is to be allocated
599 * @hop: HopID (ring) to allocate
600 * @size: Number of entries in the ring
601 * @flags: Flags for the ring
603 struct tb_ring *tb_ring_alloc_tx(struct tb_nhi *nhi, int hop, int size,
606 return tb_ring_alloc(nhi, hop, size, true, flags, 0, 0, 0, NULL, NULL);
608 EXPORT_SYMBOL_GPL(tb_ring_alloc_tx);
611 * tb_ring_alloc_rx() - Allocate DMA ring for receive
612 * @nhi: Pointer to the NHI the ring is to be allocated
613 * @hop: HopID (ring) to allocate. Pass %-1 for automatic allocation.
614 * @size: Number of entries in the ring
615 * @flags: Flags for the ring
616 * @e2e_tx_hop: Transmit HopID when E2E is enabled in @flags
617 * @sof_mask: Mask of PDF values that start a frame
618 * @eof_mask: Mask of PDF values that end a frame
619 * @start_poll: If not %NULL the ring will call this function when an
620 * interrupt is triggered and masked, instead of callback
622 * @poll_data: Optional data passed to @start_poll
624 struct tb_ring *tb_ring_alloc_rx(struct tb_nhi *nhi, int hop, int size,
625 unsigned int flags, int e2e_tx_hop,
626 u16 sof_mask, u16 eof_mask,
627 void (*start_poll)(void *), void *poll_data)
629 return tb_ring_alloc(nhi, hop, size, false, flags, e2e_tx_hop, sof_mask, eof_mask,
630 start_poll, poll_data);
632 EXPORT_SYMBOL_GPL(tb_ring_alloc_rx);
635 * tb_ring_start() - enable a ring
636 * @ring: Ring to start
638 * Must not be invoked in parallel with tb_ring_stop().
640 void tb_ring_start(struct tb_ring *ring)
645 spin_lock_irq(&ring->nhi->lock);
646 spin_lock(&ring->lock);
647 if (ring->nhi->going_away)
650 dev_WARN(&ring->nhi->pdev->dev, "ring already started\n");
653 dev_dbg(&ring->nhi->pdev->dev, "starting %s %d\n",
654 RING_TYPE(ring), ring->hop);
656 if (ring->flags & RING_FLAG_FRAME) {
659 flags = RING_FLAG_ENABLE;
661 frame_size = TB_FRAME_SIZE;
662 flags = RING_FLAG_ENABLE | RING_FLAG_RAW;
665 ring_iowrite64desc(ring, ring->descriptors_dma, 0);
667 ring_iowrite32desc(ring, ring->size, 12);
668 ring_iowrite32options(ring, 0, 4); /* time releated ? */
669 ring_iowrite32options(ring, flags, 0);
671 u32 sof_eof_mask = ring->sof_mask << 16 | ring->eof_mask;
673 ring_iowrite32desc(ring, (frame_size << 16) | ring->size, 12);
674 ring_iowrite32options(ring, sof_eof_mask, 4);
675 ring_iowrite32options(ring, flags, 0);
679 * Now that the ring valid bit is set we can configure E2E if
680 * enabled for the ring.
682 if (ring->flags & RING_FLAG_E2E) {
686 hop = ring->e2e_tx_hop << REG_RX_OPTIONS_E2E_HOP_SHIFT;
687 hop &= REG_RX_OPTIONS_E2E_HOP_MASK;
690 dev_dbg(&ring->nhi->pdev->dev,
691 "enabling E2E for %s %d with TX HopID %d\n",
692 RING_TYPE(ring), ring->hop, ring->e2e_tx_hop);
694 dev_dbg(&ring->nhi->pdev->dev, "enabling E2E for %s %d\n",
695 RING_TYPE(ring), ring->hop);
698 flags |= RING_FLAG_E2E_FLOW_CONTROL;
699 ring_iowrite32options(ring, flags, 0);
702 ring_interrupt_active(ring, true);
703 ring->running = true;
705 spin_unlock(&ring->lock);
706 spin_unlock_irq(&ring->nhi->lock);
708 EXPORT_SYMBOL_GPL(tb_ring_start);
711 * tb_ring_stop() - shutdown a ring
712 * @ring: Ring to stop
714 * Must not be invoked from a callback.
716 * This method will disable the ring. Further calls to
717 * tb_ring_tx/tb_ring_rx will return -ESHUTDOWN until ring_stop has been
720 * All enqueued frames will be canceled and their callbacks will be executed
721 * with frame->canceled set to true (on the callback thread). This method
722 * returns only after all callback invocations have finished.
724 void tb_ring_stop(struct tb_ring *ring)
726 spin_lock_irq(&ring->nhi->lock);
727 spin_lock(&ring->lock);
728 dev_dbg(&ring->nhi->pdev->dev, "stopping %s %d\n",
729 RING_TYPE(ring), ring->hop);
730 if (ring->nhi->going_away)
732 if (!ring->running) {
733 dev_WARN(&ring->nhi->pdev->dev, "%s %d already stopped\n",
734 RING_TYPE(ring), ring->hop);
737 ring_interrupt_active(ring, false);
739 ring_iowrite32options(ring, 0, 0);
740 ring_iowrite64desc(ring, 0, 0);
741 ring_iowrite32desc(ring, 0, 8);
742 ring_iowrite32desc(ring, 0, 12);
745 ring->running = false;
748 spin_unlock(&ring->lock);
749 spin_unlock_irq(&ring->nhi->lock);
752 * schedule ring->work to invoke callbacks on all remaining frames.
754 schedule_work(&ring->work);
755 flush_work(&ring->work);
757 EXPORT_SYMBOL_GPL(tb_ring_stop);
760 * tb_ring_free() - free ring
762 * When this method returns all invocations of ring->callback will have
765 * Ring must be stopped.
767 * Must NOT be called from ring_frame->callback!
769 void tb_ring_free(struct tb_ring *ring)
771 spin_lock_irq(&ring->nhi->lock);
773 * Dissociate the ring from the NHI. This also ensures that
774 * nhi_interrupt_work cannot reschedule ring->work.
777 ring->nhi->tx_rings[ring->hop] = NULL;
779 ring->nhi->rx_rings[ring->hop] = NULL;
782 dev_WARN(&ring->nhi->pdev->dev, "%s %d still running\n",
783 RING_TYPE(ring), ring->hop);
785 spin_unlock_irq(&ring->nhi->lock);
787 ring_release_msix(ring);
789 dma_free_coherent(&ring->nhi->pdev->dev,
790 ring->size * sizeof(*ring->descriptors),
791 ring->descriptors, ring->descriptors_dma);
793 ring->descriptors = NULL;
794 ring->descriptors_dma = 0;
797 dev_dbg(&ring->nhi->pdev->dev, "freeing %s %d\n", RING_TYPE(ring),
801 * ring->work can no longer be scheduled (it is scheduled only
802 * by nhi_interrupt_work, ring_stop and ring_msix). Wait for it
803 * to finish before freeing the ring.
805 flush_work(&ring->work);
808 EXPORT_SYMBOL_GPL(tb_ring_free);
811 * nhi_mailbox_cmd() - Send a command through NHI mailbox
812 * @nhi: Pointer to the NHI structure
813 * @cmd: Command to send
814 * @data: Data to be send with the command
816 * Sends mailbox command to the firmware running on NHI. Returns %0 in
817 * case of success and negative errno in case of failure.
819 int nhi_mailbox_cmd(struct tb_nhi *nhi, enum nhi_mailbox_cmd cmd, u32 data)
824 iowrite32(data, nhi->iobase + REG_INMAIL_DATA);
826 val = ioread32(nhi->iobase + REG_INMAIL_CMD);
827 val &= ~(REG_INMAIL_CMD_MASK | REG_INMAIL_ERROR);
828 val |= REG_INMAIL_OP_REQUEST | cmd;
829 iowrite32(val, nhi->iobase + REG_INMAIL_CMD);
831 timeout = ktime_add_ms(ktime_get(), NHI_MAILBOX_TIMEOUT);
833 val = ioread32(nhi->iobase + REG_INMAIL_CMD);
834 if (!(val & REG_INMAIL_OP_REQUEST))
836 usleep_range(10, 20);
837 } while (ktime_before(ktime_get(), timeout));
839 if (val & REG_INMAIL_OP_REQUEST)
841 if (val & REG_INMAIL_ERROR)
848 * nhi_mailbox_mode() - Return current firmware operation mode
849 * @nhi: Pointer to the NHI structure
851 * The function reads current firmware operation mode using NHI mailbox
852 * registers and returns it to the caller.
854 enum nhi_fw_mode nhi_mailbox_mode(struct tb_nhi *nhi)
858 val = ioread32(nhi->iobase + REG_OUTMAIL_CMD);
859 val &= REG_OUTMAIL_CMD_OPMODE_MASK;
860 val >>= REG_OUTMAIL_CMD_OPMODE_SHIFT;
862 return (enum nhi_fw_mode)val;
865 static void nhi_interrupt_work(struct work_struct *work)
867 struct tb_nhi *nhi = container_of(work, typeof(*nhi), interrupt_work);
868 int value = 0; /* Suppress uninitialized usage warning. */
871 int type = 0; /* current interrupt type 0: TX, 1: RX, 2: RX overflow */
872 struct tb_ring *ring;
874 spin_lock_irq(&nhi->lock);
877 * Starting at REG_RING_NOTIFY_BASE there are three status bitfields
878 * (TX, RX, RX overflow). We iterate over the bits and read a new
879 * dwords as required. The registers are cleared on read.
881 for (bit = 0; bit < 3 * nhi->hop_count; bit++) {
883 value = ioread32(nhi->iobase
884 + REG_RING_NOTIFY_BASE
886 if (++hop == nhi->hop_count) {
890 if ((value & (1 << (bit % 32))) == 0)
893 dev_warn(&nhi->pdev->dev,
894 "RX overflow for ring %d\n",
899 ring = nhi->tx_rings[hop];
901 ring = nhi->rx_rings[hop];
903 dev_warn(&nhi->pdev->dev,
904 "got interrupt for inactive %s ring %d\n",
910 spin_lock(&ring->lock);
911 __ring_interrupt(ring);
912 spin_unlock(&ring->lock);
914 spin_unlock_irq(&nhi->lock);
917 static irqreturn_t nhi_msi(int irq, void *data)
919 struct tb_nhi *nhi = data;
920 schedule_work(&nhi->interrupt_work);
924 static int __nhi_suspend_noirq(struct device *dev, bool wakeup)
926 struct pci_dev *pdev = to_pci_dev(dev);
927 struct tb *tb = pci_get_drvdata(pdev);
928 struct tb_nhi *nhi = tb->nhi;
931 ret = tb_domain_suspend_noirq(tb);
935 if (nhi->ops && nhi->ops->suspend_noirq) {
936 ret = nhi->ops->suspend_noirq(tb->nhi, wakeup);
944 static int nhi_suspend_noirq(struct device *dev)
946 return __nhi_suspend_noirq(dev, device_may_wakeup(dev));
949 static int nhi_freeze_noirq(struct device *dev)
951 struct pci_dev *pdev = to_pci_dev(dev);
952 struct tb *tb = pci_get_drvdata(pdev);
954 return tb_domain_freeze_noirq(tb);
957 static int nhi_thaw_noirq(struct device *dev)
959 struct pci_dev *pdev = to_pci_dev(dev);
960 struct tb *tb = pci_get_drvdata(pdev);
962 return tb_domain_thaw_noirq(tb);
965 static bool nhi_wake_supported(struct pci_dev *pdev)
970 * If power rails are sustainable for wakeup from S4 this
971 * property is set by the BIOS.
973 if (device_property_read_u8(&pdev->dev, "WAKE_SUPPORTED", &val))
979 static int nhi_poweroff_noirq(struct device *dev)
981 struct pci_dev *pdev = to_pci_dev(dev);
984 wakeup = device_may_wakeup(dev) && nhi_wake_supported(pdev);
985 return __nhi_suspend_noirq(dev, wakeup);
988 static void nhi_enable_int_throttling(struct tb_nhi *nhi)
990 /* Throttling is specified in 256ns increments */
991 u32 throttle = DIV_ROUND_UP(128 * NSEC_PER_USEC, 256);
995 * Configure interrupt throttling for all vectors even if we
998 for (i = 0; i < MSIX_MAX_VECS; i++) {
999 u32 reg = REG_INT_THROTTLING_RATE + i * 4;
1000 iowrite32(throttle, nhi->iobase + reg);
1004 static int nhi_resume_noirq(struct device *dev)
1006 struct pci_dev *pdev = to_pci_dev(dev);
1007 struct tb *tb = pci_get_drvdata(pdev);
1008 struct tb_nhi *nhi = tb->nhi;
1012 * Check that the device is still there. It may be that the user
1013 * unplugged last device which causes the host controller to go
1016 if (!pci_device_is_present(pdev)) {
1017 nhi->going_away = true;
1019 if (nhi->ops && nhi->ops->resume_noirq) {
1020 ret = nhi->ops->resume_noirq(nhi);
1024 nhi_enable_int_throttling(tb->nhi);
1027 return tb_domain_resume_noirq(tb);
1030 static int nhi_suspend(struct device *dev)
1032 struct pci_dev *pdev = to_pci_dev(dev);
1033 struct tb *tb = pci_get_drvdata(pdev);
1035 return tb_domain_suspend(tb);
1038 static void nhi_complete(struct device *dev)
1040 struct pci_dev *pdev = to_pci_dev(dev);
1041 struct tb *tb = pci_get_drvdata(pdev);
1044 * If we were runtime suspended when system suspend started,
1045 * schedule runtime resume now. It should bring the domain back
1046 * to functional state.
1048 if (pm_runtime_suspended(&pdev->dev))
1049 pm_runtime_resume(&pdev->dev);
1051 tb_domain_complete(tb);
1054 static int nhi_runtime_suspend(struct device *dev)
1056 struct pci_dev *pdev = to_pci_dev(dev);
1057 struct tb *tb = pci_get_drvdata(pdev);
1058 struct tb_nhi *nhi = tb->nhi;
1061 ret = tb_domain_runtime_suspend(tb);
1065 if (nhi->ops && nhi->ops->runtime_suspend) {
1066 ret = nhi->ops->runtime_suspend(tb->nhi);
1073 static int nhi_runtime_resume(struct device *dev)
1075 struct pci_dev *pdev = to_pci_dev(dev);
1076 struct tb *tb = pci_get_drvdata(pdev);
1077 struct tb_nhi *nhi = tb->nhi;
1080 if (nhi->ops && nhi->ops->runtime_resume) {
1081 ret = nhi->ops->runtime_resume(nhi);
1086 nhi_enable_int_throttling(nhi);
1087 return tb_domain_runtime_resume(tb);
1090 static void nhi_shutdown(struct tb_nhi *nhi)
1094 dev_dbg(&nhi->pdev->dev, "shutdown\n");
1096 for (i = 0; i < nhi->hop_count; i++) {
1097 if (nhi->tx_rings[i])
1098 dev_WARN(&nhi->pdev->dev,
1099 "TX ring %d is still active\n", i);
1100 if (nhi->rx_rings[i])
1101 dev_WARN(&nhi->pdev->dev,
1102 "RX ring %d is still active\n", i);
1104 nhi_disable_interrupts(nhi);
1106 * We have to release the irq before calling flush_work. Otherwise an
1107 * already executing IRQ handler could call schedule_work again.
1109 if (!nhi->pdev->msix_enabled) {
1110 devm_free_irq(&nhi->pdev->dev, nhi->pdev->irq, nhi);
1111 flush_work(&nhi->interrupt_work);
1113 ida_destroy(&nhi->msix_ida);
1115 if (nhi->ops && nhi->ops->shutdown)
1116 nhi->ops->shutdown(nhi);
1119 static void nhi_check_quirks(struct tb_nhi *nhi)
1121 if (nhi->pdev->vendor == PCI_VENDOR_ID_INTEL) {
1123 * Intel hardware supports auto clear of the interrupt
1124 * status register right after interrupt is being
1127 nhi->quirks |= QUIRK_AUTO_CLEAR_INT;
1129 switch (nhi->pdev->device) {
1130 case PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI:
1131 case PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI:
1133 * Falcon Ridge controller needs the end-to-end
1134 * flow control workaround to avoid losing Rx
1135 * packets when RING_FLAG_E2E is set.
1137 nhi->quirks |= QUIRK_E2E;
1143 static int nhi_check_iommu_pdev(struct pci_dev *pdev, void *data)
1145 if (!pdev->external_facing ||
1146 !device_iommu_capable(&pdev->dev, IOMMU_CAP_PRE_BOOT_PROTECTION))
1148 *(bool *)data = true;
1149 return 1; /* Stop walking */
1152 static void nhi_check_iommu(struct tb_nhi *nhi)
1154 struct pci_bus *bus = nhi->pdev->bus;
1155 bool port_ok = false;
1158 * Ideally what we'd do here is grab every PCI device that
1159 * represents a tunnelling adapter for this NHI and check their
1160 * status directly, but unfortunately USB4 seems to make it
1161 * obnoxiously difficult to reliably make any correlation.
1163 * So for now we'll have to bodge it... Hoping that the system
1164 * is at least sane enough that an adapter is in the same PCI
1165 * segment as its NHI, if we can find *something* on that segment
1166 * which meets the requirements for Kernel DMA Protection, we'll
1167 * take that to imply that firmware is aware and has (hopefully)
1168 * done the right thing in general. We need to know that the PCI
1169 * layer has seen the ExternalFacingPort property which will then
1170 * inform the IOMMU layer to enforce the complete "untrusted DMA"
1171 * flow, but also that the IOMMU driver itself can be trusted not
1172 * to have been subverted by a pre-boot DMA attack.
1177 pci_walk_bus(bus, nhi_check_iommu_pdev, &port_ok);
1179 nhi->iommu_dma_protection = port_ok;
1180 dev_dbg(&nhi->pdev->dev, "IOMMU DMA protection is %s\n",
1181 str_enabled_disabled(port_ok));
1184 static int nhi_init_msi(struct tb_nhi *nhi)
1186 struct pci_dev *pdev = nhi->pdev;
1187 struct device *dev = &pdev->dev;
1190 /* In case someone left them on. */
1191 nhi_disable_interrupts(nhi);
1193 nhi_enable_int_throttling(nhi);
1195 ida_init(&nhi->msix_ida);
1198 * The NHI has 16 MSI-X vectors or a single MSI. We first try to
1199 * get all MSI-X vectors and if we succeed, each ring will have
1200 * one MSI-X. If for some reason that does not work out, we
1201 * fallback to a single MSI.
1203 nvec = pci_alloc_irq_vectors(pdev, MSIX_MIN_VECS, MSIX_MAX_VECS,
1206 nvec = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI);
1210 INIT_WORK(&nhi->interrupt_work, nhi_interrupt_work);
1212 irq = pci_irq_vector(nhi->pdev, 0);
1216 res = devm_request_irq(&pdev->dev, irq, nhi_msi,
1217 IRQF_NO_SUSPEND, "thunderbolt", nhi);
1219 return dev_err_probe(dev, res, "request_irq failed, aborting\n");
1225 static bool nhi_imr_valid(struct pci_dev *pdev)
1229 if (!device_property_read_u8(&pdev->dev, "IMR_VALID", &val))
1235 static struct tb *nhi_select_cm(struct tb_nhi *nhi)
1240 * USB4 case is simple. If we got control of any of the
1241 * capabilities, we use software CM.
1243 if (tb_acpi_is_native())
1244 return tb_probe(nhi);
1247 * Either firmware based CM is running (we did not get control
1248 * from the firmware) or this is pre-USB4 PC so try first
1249 * firmware CM and then fallback to software CM.
1251 tb = icm_probe(nhi);
1258 static int nhi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1260 struct device *dev = &pdev->dev;
1265 if (!nhi_imr_valid(pdev))
1266 return dev_err_probe(dev, -ENODEV, "firmware image not valid, aborting\n");
1268 res = pcim_enable_device(pdev);
1270 return dev_err_probe(dev, res, "cannot enable PCI device, aborting\n");
1272 res = pcim_iomap_regions(pdev, 1 << 0, "thunderbolt");
1274 return dev_err_probe(dev, res, "cannot obtain PCI resources, aborting\n");
1276 nhi = devm_kzalloc(&pdev->dev, sizeof(*nhi), GFP_KERNEL);
1281 nhi->ops = (const struct tb_nhi_ops *)id->driver_data;
1282 /* cannot fail - table is allocated in pcim_iomap_regions */
1283 nhi->iobase = pcim_iomap_table(pdev)[0];
1284 nhi->hop_count = ioread32(nhi->iobase + REG_HOP_COUNT) & 0x3ff;
1285 dev_dbg(dev, "total paths: %d\n", nhi->hop_count);
1287 nhi->tx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1288 sizeof(*nhi->tx_rings), GFP_KERNEL);
1289 nhi->rx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
1290 sizeof(*nhi->rx_rings), GFP_KERNEL);
1291 if (!nhi->tx_rings || !nhi->rx_rings)
1294 nhi_check_quirks(nhi);
1295 nhi_check_iommu(nhi);
1297 res = nhi_init_msi(nhi);
1299 return dev_err_probe(dev, res, "cannot enable MSI, aborting\n");
1301 spin_lock_init(&nhi->lock);
1303 res = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1305 return dev_err_probe(dev, res, "failed to set DMA mask\n");
1307 pci_set_master(pdev);
1309 if (nhi->ops && nhi->ops->init) {
1310 res = nhi->ops->init(nhi);
1315 tb = nhi_select_cm(nhi);
1317 return dev_err_probe(dev, -ENODEV,
1318 "failed to determine connection manager, aborting\n");
1320 dev_dbg(dev, "NHI initialized, starting thunderbolt\n");
1322 res = tb_domain_add(tb);
1325 * At this point the RX/TX rings might already have been
1326 * activated. Do a proper shutdown.
1332 pci_set_drvdata(pdev, tb);
1334 device_wakeup_enable(&pdev->dev);
1336 pm_runtime_allow(&pdev->dev);
1337 pm_runtime_set_autosuspend_delay(&pdev->dev, TB_AUTOSUSPEND_DELAY);
1338 pm_runtime_use_autosuspend(&pdev->dev);
1339 pm_runtime_put_autosuspend(&pdev->dev);
1344 static void nhi_remove(struct pci_dev *pdev)
1346 struct tb *tb = pci_get_drvdata(pdev);
1347 struct tb_nhi *nhi = tb->nhi;
1349 pm_runtime_get_sync(&pdev->dev);
1350 pm_runtime_dont_use_autosuspend(&pdev->dev);
1351 pm_runtime_forbid(&pdev->dev);
1353 tb_domain_remove(tb);
1358 * The tunneled pci bridges are siblings of us. Use resume_noirq to reenable
1359 * the tunnels asap. A corresponding pci quirk blocks the downstream bridges
1360 * resume_noirq until we are done.
1362 static const struct dev_pm_ops nhi_pm_ops = {
1363 .suspend_noirq = nhi_suspend_noirq,
1364 .resume_noirq = nhi_resume_noirq,
1365 .freeze_noirq = nhi_freeze_noirq, /*
1366 * we just disable hotplug, the
1367 * pci-tunnels stay alive.
1369 .thaw_noirq = nhi_thaw_noirq,
1370 .restore_noirq = nhi_resume_noirq,
1371 .suspend = nhi_suspend,
1372 .poweroff_noirq = nhi_poweroff_noirq,
1373 .poweroff = nhi_suspend,
1374 .complete = nhi_complete,
1375 .runtime_suspend = nhi_runtime_suspend,
1376 .runtime_resume = nhi_runtime_resume,
1379 static struct pci_device_id nhi_ids[] = {
1381 * We have to specify class, the TB bridges use the same device and
1382 * vendor (sub)id on gen 1 and gen 2 controllers.
1385 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1386 .vendor = PCI_VENDOR_ID_INTEL,
1387 .device = PCI_DEVICE_ID_INTEL_LIGHT_RIDGE,
1388 .subvendor = 0x2222, .subdevice = 0x1111,
1391 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1392 .vendor = PCI_VENDOR_ID_INTEL,
1393 .device = PCI_DEVICE_ID_INTEL_CACTUS_RIDGE_4C,
1394 .subvendor = 0x2222, .subdevice = 0x1111,
1397 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1398 .vendor = PCI_VENDOR_ID_INTEL,
1399 .device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI,
1400 .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1403 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
1404 .vendor = PCI_VENDOR_ID_INTEL,
1405 .device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI,
1406 .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
1410 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_2C_NHI) },
1411 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_4C_NHI) },
1412 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_USBONLY_NHI) },
1413 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_NHI) },
1414 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_LP_USBONLY_NHI) },
1415 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_2C_NHI) },
1416 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_4C_NHI) },
1417 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ALPINE_RIDGE_C_USBONLY_NHI) },
1418 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_2C_NHI) },
1419 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TITAN_RIDGE_4C_NHI) },
1420 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ICL_NHI0),
1421 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1422 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ICL_NHI1),
1423 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1425 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_NHI0),
1426 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1427 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_NHI1),
1428 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1429 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_H_NHI0),
1430 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1431 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_TGL_H_NHI1),
1432 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1433 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ADL_NHI0),
1434 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1435 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_ADL_NHI1),
1436 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1437 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_RPL_NHI0),
1438 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1439 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_RPL_NHI1),
1440 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1441 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_MTL_M_NHI0),
1442 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1443 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_MTL_P_NHI0),
1444 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1445 { PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_MTL_P_NHI1),
1446 .driver_data = (kernel_ulong_t)&icl_nhi_ops },
1448 /* Any USB4 compliant host */
1449 { PCI_DEVICE_CLASS(PCI_CLASS_SERIAL_USB_USB4, ~0) },
1454 MODULE_DEVICE_TABLE(pci, nhi_ids);
1455 MODULE_LICENSE("GPL");
1457 static struct pci_driver nhi_driver = {
1458 .name = "thunderbolt",
1459 .id_table = nhi_ids,
1461 .remove = nhi_remove,
1462 .shutdown = nhi_remove,
1463 .driver.pm = &nhi_pm_ops,
1466 static int __init nhi_init(void)
1470 ret = tb_domain_init();
1473 ret = pci_register_driver(&nhi_driver);
1479 static void __exit nhi_unload(void)
1481 pci_unregister_driver(&nhi_driver);
1485 rootfs_initcall(nhi_init);
1486 module_exit(nhi_unload);