1/*
2 * Copyright (C) 2001-2004 by David Brownell
3 *
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License as published by the
6 * Free Software Foundation; either version 2 of the License, or (at your
7 * option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
11 * or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
12 * for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software Foundation,
16 * Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
17 */
18
19/* this file is part of ehci-hcd.c */
20
21/*-------------------------------------------------------------------------*/
22
23/*
24 * EHCI hardware queue manipulation ... the core.  QH/QTD manipulation.
25 *
26 * Control, bulk, and interrupt traffic all use "qh" lists.  They list "qtd"
27 * entries describing USB transactions, max 16-20kB/entry (with 4kB-aligned
28 * buffers needed for the larger number).  We use one QH per endpoint, queue
29 * multiple urbs (all three types) per endpoint.  URBs may need several qtds.
30 *
31 * ISO traffic uses "ISO TD" (itd, and sitd) records, and (along with
32 * interrupts) needs careful scheduling.  Performance improvements can be
33 * an ongoing challenge.  That's in "ehci-sched.c".
34 *
35 * USB 1.1 devices are handled (a) by "companion" OHCI or UHCI root hubs,
36 * or otherwise through transaction translators (TTs) in USB 2.0 hubs using
37 * (b) special fields in qh entries or (c) split iso entries.  TTs will
38 * buffer low/full speed data so the host collects it at high speed.
39 */
40
41/*-------------------------------------------------------------------------*/
42
43/* fill a qtd, returning how much of the buffer we were able to queue up */
44
45static int
46qtd_fill(struct ehci_hcd *ehci, struct ehci_qtd *qtd, dma_addr_t buf,
47		  size_t len, int token, int maxpacket)
48{
49	int	i, count;
50	u64	addr = buf;
51
52	/* one buffer entry per 4K ... first might be short or unaligned */
53	qtd->hw_buf[0] = cpu_to_hc32(ehci, (u32)addr);
54	qtd->hw_buf_hi[0] = cpu_to_hc32(ehci, (u32)(addr >> 32));
55	count = 0x1000 - (buf & 0x0fff);	/* rest of that page */
56	if (likely (len < count))		/* ... iff needed */
57		count = len;
58	else {
59		buf +=  0x1000;
60		buf &= ~0x0fff;
61
62		/* per-qtd limit: from 16K to 20K (best alignment) */
63		for (i = 1; count < len && i < 5; i++) {
64			addr = buf;
65			qtd->hw_buf[i] = cpu_to_hc32(ehci, (u32)addr);
66			qtd->hw_buf_hi[i] = cpu_to_hc32(ehci,
67					(u32)(addr >> 32));
68			buf += 0x1000;
69			if ((count + 0x1000) < len)
70				count += 0x1000;
71			else
72				count = len;
73		}
74
75		/* short packets may only terminate transfers */
76		if (count != len)
77			count -= (count % maxpacket);
78	}
79	qtd->hw_token = cpu_to_hc32(ehci, (count << 16) | token);
80	qtd->length = count;
81
82	return count;
83}
84
85/*-------------------------------------------------------------------------*/
86
87static inline void
88qh_update (struct ehci_hcd *ehci, struct ehci_qh *qh, struct ehci_qtd *qtd)
89{
90	struct ehci_qh_hw *hw = qh->hw;
91
92	/* writes to an active overlay are unsafe */
93	WARN_ON(qh->qh_state != QH_STATE_IDLE);
94
95	hw->hw_qtd_next = QTD_NEXT(ehci, qtd->qtd_dma);
96	hw->hw_alt_next = EHCI_LIST_END(ehci);
97
98	/* Except for control endpoints, we make hardware maintain data
99	 * toggle (like OHCI) ... here (re)initialize the toggle in the QH,
100	 * and set the pseudo-toggle in udev. Only usb_clear_halt() will
101	 * ever clear it.
102	 */
103	if (!(hw->hw_info1 & cpu_to_hc32(ehci, QH_TOGGLE_CTL))) {
104		unsigned	is_out, epnum;
105
106		is_out = qh->is_out;
107		epnum = (hc32_to_cpup(ehci, &hw->hw_info1) >> 8) & 0x0f;
108		if (unlikely(!usb_gettoggle(qh->ps.udev, epnum, is_out))) {
109			hw->hw_token &= ~cpu_to_hc32(ehci, QTD_TOGGLE);
110			usb_settoggle(qh->ps.udev, epnum, is_out, 1);
111		}
112	}
113
114	hw->hw_token &= cpu_to_hc32(ehci, QTD_TOGGLE | QTD_STS_PING);
115}
116
117/* if it weren't for a common silicon quirk (writing the dummy into the qh
118 * overlay, so qh->hw_token wrongly becomes inactive/halted), only fault
119 * recovery (including urb dequeue) would need software changes to a QH...
120 */
121static void
122qh_refresh (struct ehci_hcd *ehci, struct ehci_qh *qh)
123{
124	struct ehci_qtd *qtd;
125
126	qtd = list_entry(qh->qtd_list.next, struct ehci_qtd, qtd_list);
127
128	/*
129	 * first qtd may already be partially processed.
130	 * If we come here during unlink, the QH overlay region
131	 * might have reference to the just unlinked qtd. The
132	 * qtd is updated in qh_completions(). Update the QH
133	 * overlay here.
134	 */
135	if (qh->hw->hw_token & ACTIVE_BIT(ehci))
136		qh->hw->hw_qtd_next = qtd->hw_next;
137	else
138		qh_update(ehci, qh, qtd);
139}
140
141/*-------------------------------------------------------------------------*/
142
143static void qh_link_async(struct ehci_hcd *ehci, struct ehci_qh *qh);
144
145static void ehci_clear_tt_buffer_complete(struct usb_hcd *hcd,
146		struct usb_host_endpoint *ep)
147{
148	struct ehci_hcd		*ehci = hcd_to_ehci(hcd);
149	struct ehci_qh		*qh = ep->hcpriv;
150	unsigned long		flags;
151
152	spin_lock_irqsave(&ehci->lock, flags);
153	qh->clearing_tt = 0;
154	if (qh->qh_state == QH_STATE_IDLE && !list_empty(&qh->qtd_list)
155			&& ehci->rh_state == EHCI_RH_RUNNING)
156		qh_link_async(ehci, qh);
157	spin_unlock_irqrestore(&ehci->lock, flags);
158}
159
160static void ehci_clear_tt_buffer(struct ehci_hcd *ehci, struct ehci_qh *qh,
161		struct urb *urb, u32 token)
162{
163
164	/* If an async split transaction gets an error or is unlinked,
165	 * the TT buffer may be left in an indeterminate state.  We
166	 * have to clear the TT buffer.
167	 *
168	 * Note: this routine is never called for Isochronous transfers.
169	 */
170	if (urb->dev->tt && !usb_pipeint(urb->pipe) && !qh->clearing_tt) {
171#ifdef CONFIG_DYNAMIC_DEBUG
172		struct usb_device *tt = urb->dev->tt->hub;
173		dev_dbg(&tt->dev,
174			"clear tt buffer port %d, a%d ep%d t%08x\n",
175			urb->dev->ttport, urb->dev->devnum,
176			usb_pipeendpoint(urb->pipe), token);
177#endif /* CONFIG_DYNAMIC_DEBUG */
178		if (!ehci_is_TDI(ehci)
179				|| urb->dev->tt->hub !=
180				   ehci_to_hcd(ehci)->self.root_hub) {
181			if (usb_hub_clear_tt_buffer(urb) == 0)
182				qh->clearing_tt = 1;
183		} else {
184
185			/* REVISIT ARC-derived cores don't clear the root
186			 * hub TT buffer in this way...
187			 */
188		}
189	}
190}
191
192static int qtd_copy_status (
193	struct ehci_hcd *ehci,
194	struct urb *urb,
195	size_t length,
196	u32 token
197)
198{
199	int	status = -EINPROGRESS;
200
201	/* count IN/OUT bytes, not SETUP (even short packets) */
202	if (likely (QTD_PID (token) != 2))
203		urb->actual_length += length - QTD_LENGTH (token);
204
205	/* don't modify error codes */
206	if (unlikely(urb->unlinked))
207		return status;
208
209	/* force cleanup after short read; not always an error */
210	if (unlikely (IS_SHORT_READ (token)))
211		status = -EREMOTEIO;
212
213	/* serious "can't proceed" faults reported by the hardware */
214	if (token & QTD_STS_HALT) {
215		if (token & QTD_STS_BABBLE) {
216			/* FIXME "must" disable babbling device's port too */
217			status = -EOVERFLOW;
218		/* CERR nonzero + halt --> stall */
219		} else if (QTD_CERR(token)) {
220			status = -EPIPE;
221
222		/* In theory, more than one of the following bits can be set
223		 * since they are sticky and the transaction is retried.
224		 * Which to test first is rather arbitrary.
225		 */
226		} else if (token & QTD_STS_MMF) {
227			/* fs/ls interrupt xfer missed the complete-split */
228			status = -EPROTO;
229		} else if (token & QTD_STS_DBE) {
230			status = (QTD_PID (token) == 1) /* IN ? */
231				? -ENOSR  /* hc couldn't read data */
232				: -ECOMM; /* hc couldn't write data */
233		} else if (token & QTD_STS_XACT) {
234			/* timeout, bad CRC, wrong PID, etc */
235			ehci_dbg(ehci, "devpath %s ep%d%s 3strikes\n",
236				urb->dev->devpath,
237				usb_pipeendpoint(urb->pipe),
238				usb_pipein(urb->pipe) ? "in" : "out");
239			status = -EPROTO;
240		} else {	/* unknown */
241			status = -EPROTO;
242		}
243	}
244
245	return status;
246}
247
248static void
249ehci_urb_done(struct ehci_hcd *ehci, struct urb *urb, int status)
250{
251	if (usb_pipetype(urb->pipe) == PIPE_INTERRUPT) {
252		/* ... update hc-wide periodic stats */
253		ehci_to_hcd(ehci)->self.bandwidth_int_reqs--;
254	}
255
256	if (unlikely(urb->unlinked)) {
257		COUNT(ehci->stats.unlink);
258	} else {
259		/* report non-error and short read status as zero */
260		if (status == -EINPROGRESS || status == -EREMOTEIO)
261			status = 0;
262		COUNT(ehci->stats.complete);
263	}
264
265#ifdef EHCI_URB_TRACE
266	ehci_dbg (ehci,
267		"%s %s urb %p ep%d%s status %d len %d/%d\n",
268		__func__, urb->dev->devpath, urb,
269		usb_pipeendpoint (urb->pipe),
270		usb_pipein (urb->pipe) ? "in" : "out",
271		status,
272		urb->actual_length, urb->transfer_buffer_length);
273#endif
274
275	usb_hcd_unlink_urb_from_ep(ehci_to_hcd(ehci), urb);
276	usb_hcd_giveback_urb(ehci_to_hcd(ehci), urb, status);
277}
278
279static int qh_schedule (struct ehci_hcd *ehci, struct ehci_qh *qh);
280
281/*
282 * Process and free completed qtds for a qh, returning URBs to drivers.
283 * Chases up to qh->hw_current.  Returns nonzero if the caller should
284 * unlink qh.
285 */
286static unsigned
287qh_completions (struct ehci_hcd *ehci, struct ehci_qh *qh)
288{
289	struct ehci_qtd		*last, *end = qh->dummy;
290	struct list_head	*entry, *tmp;
291	int			last_status;
292	int			stopped;
293	u8			state;
294	struct ehci_qh_hw	*hw = qh->hw;
295
296	/* completions (or tasks on other cpus) must never clobber HALT
297	 * till we've gone through and cleaned everything up, even when
298	 * they add urbs to this qh's queue or mark them for unlinking.
299	 *
300	 * NOTE:  unlinking expects to be done in queue order.
301	 *
302	 * It's a bug for qh->qh_state to be anything other than
303	 * QH_STATE_IDLE, unless our caller is scan_async() or
304	 * scan_intr().
305	 */
306	state = qh->qh_state;
307	qh->qh_state = QH_STATE_COMPLETING;
308	stopped = (state == QH_STATE_IDLE);
309
310 rescan:
311	last = NULL;
312	last_status = -EINPROGRESS;
313	qh->dequeue_during_giveback = 0;
314
315	/* remove de-activated QTDs from front of queue.
316	 * after faults (including short reads), cleanup this urb
317	 * then let the queue advance.
318	 * if queue is stopped, handles unlinks.
319	 */
320	list_for_each_safe (entry, tmp, &qh->qtd_list) {
321		struct ehci_qtd	*qtd;
322		struct urb	*urb;
323		u32		token = 0;
324
325		qtd = list_entry (entry, struct ehci_qtd, qtd_list);
326		urb = qtd->urb;
327
328		/* clean up any state from previous QTD ...*/
329		if (last) {
330			if (likely (last->urb != urb)) {
331				ehci_urb_done(ehci, last->urb, last_status);
332				last_status = -EINPROGRESS;
333			}
334			ehci_qtd_free (ehci, last);
335			last = NULL;
336		}
337
338		/* ignore urbs submitted during completions we reported */
339		if (qtd == end)
340			break;
341
342		/* hardware copies qtd out of qh overlay */
343		rmb ();
344		token = hc32_to_cpu(ehci, qtd->hw_token);
345
346		/* always clean up qtds the hc de-activated */
347 retry_xacterr:
348		if ((token & QTD_STS_ACTIVE) == 0) {
349
350			/* Report Data Buffer Error: non-fatal but useful */
351			if (token & QTD_STS_DBE)
352				ehci_dbg(ehci,
353					"detected DataBufferErr for urb %p ep%d%s len %d, qtd %p [qh %p]\n",
354					urb,
355					usb_endpoint_num(&urb->ep->desc),
356					usb_endpoint_dir_in(&urb->ep->desc) ? "in" : "out",
357					urb->transfer_buffer_length,
358					qtd,
359					qh);
360
361			/* on STALL, error, and short reads this urb must
362			 * complete and all its qtds must be recycled.
363			 */
364			if ((token & QTD_STS_HALT) != 0) {
365
366				/* retry transaction errors until we
367				 * reach the software xacterr limit
368				 */
369				if ((token & QTD_STS_XACT) &&
370						QTD_CERR(token) == 0 &&
371						++qh->xacterrs < QH_XACTERR_MAX &&
372						!urb->unlinked) {
373					ehci_dbg(ehci,
374	"detected XactErr len %zu/%zu retry %d\n",
375	qtd->length - QTD_LENGTH(token), qtd->length, qh->xacterrs);
376
377					/* reset the token in the qtd and the
378					 * qh overlay (which still contains
379					 * the qtd) so that we pick up from
380					 * where we left off
381					 */
382					token &= ~QTD_STS_HALT;
383					token |= QTD_STS_ACTIVE |
384							(EHCI_TUNE_CERR << 10);
385					qtd->hw_token = cpu_to_hc32(ehci,
386							token);
387					wmb();
388					hw->hw_token = cpu_to_hc32(ehci,
389							token);
390					goto retry_xacterr;
391				}
392				stopped = 1;
393
394			/* magic dummy for some short reads; qh won't advance.
395			 * that silicon quirk can kick in with this dummy too.
396			 *
397			 * other short reads won't stop the queue, including
398			 * control transfers (status stage handles that) or
399			 * most other single-qtd reads ... the queue stops if
400			 * URB_SHORT_NOT_OK was set so the driver submitting
401			 * the urbs could clean it up.
402			 */
403			} else if (IS_SHORT_READ (token)
404					&& !(qtd->hw_alt_next
405						& EHCI_LIST_END(ehci))) {
406				stopped = 1;
407			}
408
409		/* stop scanning when we reach qtds the hc is using */
410		} else if (likely (!stopped
411				&& ehci->rh_state >= EHCI_RH_RUNNING)) {
412			break;
413
414		/* scan the whole queue for unlinks whenever it stops */
415		} else {
416			stopped = 1;
417
418			/* cancel everything if we halt, suspend, etc */
419			if (ehci->rh_state < EHCI_RH_RUNNING)
420				last_status = -ESHUTDOWN;
421
422			/* this qtd is active; skip it unless a previous qtd
423			 * for its urb faulted, or its urb was canceled.
424			 */
425			else if (last_status == -EINPROGRESS && !urb->unlinked)
426				continue;
427
428			/*
429			 * If this was the active qtd when the qh was unlinked
430			 * and the overlay's token is active, then the overlay
431			 * hasn't been written back to the qtd yet so use its
432			 * token instead of the qtd's.  After the qtd is
433			 * processed and removed, the overlay won't be valid
434			 * any more.
435			 */
436			if (state == QH_STATE_IDLE &&
437					qh->qtd_list.next == &qtd->qtd_list &&
438					(hw->hw_token & ACTIVE_BIT(ehci))) {
439				token = hc32_to_cpu(ehci, hw->hw_token);
440				hw->hw_token &= ~ACTIVE_BIT(ehci);
441
442				/* An unlink may leave an incomplete
443				 * async transaction in the TT buffer.
444				 * We have to clear it.
445				 */
446				ehci_clear_tt_buffer(ehci, qh, urb, token);
447			}
448		}
449
450		/* unless we already know the urb's status, collect qtd status
451		 * and update count of bytes transferred.  in common short read
452		 * cases with only one data qtd (including control transfers),
453		 * queue processing won't halt.  but with two or more qtds (for
454		 * example, with a 32 KB transfer), when the first qtd gets a
455		 * short read the second must be removed by hand.
456		 */
457		if (last_status == -EINPROGRESS) {
458			last_status = qtd_copy_status(ehci, urb,
459					qtd->length, token);
460			if (last_status == -EREMOTEIO
461					&& (qtd->hw_alt_next
462						& EHCI_LIST_END(ehci)))
463				last_status = -EINPROGRESS;
464
465			/* As part of low/full-speed endpoint-halt processing
466			 * we must clear the TT buffer (11.17.5).
467			 */
468			if (unlikely(last_status != -EINPROGRESS &&
469					last_status != -EREMOTEIO)) {
470				/* The TT's in some hubs malfunction when they
471				 * receive this request following a STALL (they
472				 * stop sending isochronous packets).  Since a
473				 * STALL can't leave the TT buffer in a busy
474				 * state (if you believe Figures 11-48 - 11-51
475				 * in the USB 2.0 spec), we won't clear the TT
476				 * buffer in this case.  Strictly speaking this
477				 * is a violation of the spec.
478				 */
479				if (last_status != -EPIPE)
480					ehci_clear_tt_buffer(ehci, qh, urb,
481							token);
482			}
483		}
484
485		/* if we're removing something not at the queue head,
486		 * patch the hardware queue pointer.
487		 */
488		if (stopped && qtd->qtd_list.prev != &qh->qtd_list) {
489			last = list_entry (qtd->qtd_list.prev,
490					struct ehci_qtd, qtd_list);
491			last->hw_next = qtd->hw_next;
492		}
493
494		/* remove qtd; it's recycled after possible urb completion */
495		list_del (&qtd->qtd_list);
496		last = qtd;
497
498		/* reinit the xacterr counter for the next qtd */
499		qh->xacterrs = 0;
500	}
501
502	/* last urb's completion might still need calling */
503	if (likely (last != NULL)) {
504		ehci_urb_done(ehci, last->urb, last_status);
505		ehci_qtd_free (ehci, last);
506	}
507
508	/* Do we need to rescan for URBs dequeued during a giveback? */
509	if (unlikely(qh->dequeue_during_giveback)) {
510		/* If the QH is already unlinked, do the rescan now. */
511		if (state == QH_STATE_IDLE)
512			goto rescan;
513
514		/* Otherwise the caller must unlink the QH. */
515	}
516
517	/* restore original state; caller must unlink or relink */
518	qh->qh_state = state;
519
520	/* be sure the hardware's done with the qh before refreshing
521	 * it after fault cleanup, or recovering from silicon wrongly
522	 * overlaying the dummy qtd (which reduces DMA chatter).
523	 *
524	 * We won't refresh a QH that's linked (after the HC
525	 * stopped the queue).  That avoids a race:
526	 *  - HC reads first part of QH;
527	 *  - CPU updates that first part and the token;
528	 *  - HC reads rest of that QH, including token
529	 * Result:  HC gets an inconsistent image, and then
530	 * DMAs to/from the wrong memory (corrupting it).
531	 *
532	 * That should be rare for interrupt transfers,
533	 * except maybe high bandwidth ...
534	 */
535	if (stopped != 0 || hw->hw_qtd_next == EHCI_LIST_END(ehci))
536		qh->exception = 1;
537
538	/* Let the caller know if the QH needs to be unlinked. */
539	return qh->exception;
540}
541
542/*-------------------------------------------------------------------------*/
543
544// high bandwidth multiplier, as encoded in highspeed endpoint descriptors
545#define hb_mult(wMaxPacketSize) (1 + (((wMaxPacketSize) >> 11) & 0x03))
546// ... and packet size, for any kind of endpoint descriptor
547#define max_packet(wMaxPacketSize) ((wMaxPacketSize) & 0x07ff)
548
549/*
550 * reverse of qh_urb_transaction:  free a list of TDs.
551 * used for cleanup after errors, before HC sees an URB's TDs.
552 */
553static void qtd_list_free (
554	struct ehci_hcd		*ehci,
555	struct urb		*urb,
556	struct list_head	*qtd_list
557) {
558	struct list_head	*entry, *temp;
559
560	list_for_each_safe (entry, temp, qtd_list) {
561		struct ehci_qtd	*qtd;
562
563		qtd = list_entry (entry, struct ehci_qtd, qtd_list);
564		list_del (&qtd->qtd_list);
565		ehci_qtd_free (ehci, qtd);
566	}
567}
568
569/*
570 * create a list of filled qtds for this URB; won't link into qh.
571 */
572static struct list_head *
573qh_urb_transaction (
574	struct ehci_hcd		*ehci,
575	struct urb		*urb,
576	struct list_head	*head,
577	gfp_t			flags
578) {
579	struct ehci_qtd		*qtd, *qtd_prev;
580	dma_addr_t		buf;
581	int			len, this_sg_len, maxpacket;
582	int			is_input;
583	u32			token;
584	int			i;
585	struct scatterlist	*sg;
586
587	/*
588	 * URBs map to sequences of QTDs:  one logical transaction
589	 */
590	qtd = ehci_qtd_alloc (ehci, flags);
591	if (unlikely (!qtd))
592		return NULL;
593	list_add_tail (&qtd->qtd_list, head);
594	qtd->urb = urb;
595
596	token = QTD_STS_ACTIVE;
597	token |= (EHCI_TUNE_CERR << 10);
598	/* for split transactions, SplitXState initialized to zero */
599
600	len = urb->transfer_buffer_length;
601	is_input = usb_pipein (urb->pipe);
602	if (usb_pipecontrol (urb->pipe)) {
603		/* SETUP pid */
604		qtd_fill(ehci, qtd, urb->setup_dma,
605				sizeof (struct usb_ctrlrequest),
606				token | (2 /* "setup" */ << 8), 8);
607
608		/* ... and always at least one more pid */
609		token ^= QTD_TOGGLE;
610		qtd_prev = qtd;
611		qtd = ehci_qtd_alloc (ehci, flags);
612		if (unlikely (!qtd))
613			goto cleanup;
614		qtd->urb = urb;
615		qtd_prev->hw_next = QTD_NEXT(ehci, qtd->qtd_dma);
616		list_add_tail (&qtd->qtd_list, head);
617
618		/* for zero length DATA stages, STATUS is always IN */
619		if (len == 0)
620			token |= (1 /* "in" */ << 8);
621	}
622
623	/*
624	 * data transfer stage:  buffer setup
625	 */
626	i = urb->num_mapped_sgs;
627	if (len > 0 && i > 0) {
628		sg = urb->sg;
629		buf = sg_dma_address(sg);
630
631		/* urb->transfer_buffer_length may be smaller than the
632		 * size of the scatterlist (or vice versa)
633		 */
634		this_sg_len = min_t(int, sg_dma_len(sg), len);
635	} else {
636		sg = NULL;
637		buf = urb->transfer_dma;
638		this_sg_len = len;
639	}
640
641	if (is_input)
642		token |= (1 /* "in" */ << 8);
643	/* else it's already initted to "out" pid (0 << 8) */
644
645	maxpacket = max_packet(usb_maxpacket(urb->dev, urb->pipe, !is_input));
646
647	/*
648	 * buffer gets wrapped in one or more qtds;
649	 * last one may be "short" (including zero len)
650	 * and may serve as a control status ack
651	 */
652	for (;;) {
653		int this_qtd_len;
654
655		this_qtd_len = qtd_fill(ehci, qtd, buf, this_sg_len, token,
656				maxpacket);
657		this_sg_len -= this_qtd_len;
658		len -= this_qtd_len;
659		buf += this_qtd_len;
660
661		/*
662		 * short reads advance to a "magic" dummy instead of the next
663		 * qtd ... that forces the queue to stop, for manual cleanup.
664		 * (this will usually be overridden later.)
665		 */
666		if (is_input)
667			qtd->hw_alt_next = ehci->async->hw->hw_alt_next;
668
669		/* qh makes control packets use qtd toggle; maybe switch it */
670		if ((maxpacket & (this_qtd_len + (maxpacket - 1))) == 0)
671			token ^= QTD_TOGGLE;
672
673		if (likely(this_sg_len <= 0)) {
674			if (--i <= 0 || len <= 0)
675				break;
676			sg = sg_next(sg);
677			buf = sg_dma_address(sg);
678			this_sg_len = min_t(int, sg_dma_len(sg), len);
679		}
680
681		qtd_prev = qtd;
682		qtd = ehci_qtd_alloc (ehci, flags);
683		if (unlikely (!qtd))
684			goto cleanup;
685		qtd->urb = urb;
686		qtd_prev->hw_next = QTD_NEXT(ehci, qtd->qtd_dma);
687		list_add_tail (&qtd->qtd_list, head);
688	}
689
690	/*
691	 * unless the caller requires manual cleanup after short reads,
692	 * have the alt_next mechanism keep the queue running after the
693	 * last data qtd (the only one, for control and most other cases).
694	 */
695	if (likely ((urb->transfer_flags & URB_SHORT_NOT_OK) == 0
696				|| usb_pipecontrol (urb->pipe)))
697		qtd->hw_alt_next = EHCI_LIST_END(ehci);
698
699	/*
700	 * control requests may need a terminating data "status" ack;
701	 * other OUT ones may need a terminating short packet
702	 * (zero length).
703	 */
704	if (likely (urb->transfer_buffer_length != 0)) {
705		int	one_more = 0;
706
707		if (usb_pipecontrol (urb->pipe)) {
708			one_more = 1;
709			token ^= 0x0100;	/* "in" <--> "out"  */
710			token |= QTD_TOGGLE;	/* force DATA1 */
711		} else if (usb_pipeout(urb->pipe)
712				&& (urb->transfer_flags & URB_ZERO_PACKET)
713				&& !(urb->transfer_buffer_length % maxpacket)) {
714			one_more = 1;
715		}
716		if (one_more) {
717			qtd_prev = qtd;
718			qtd = ehci_qtd_alloc (ehci, flags);
719			if (unlikely (!qtd))
720				goto cleanup;
721			qtd->urb = urb;
722			qtd_prev->hw_next = QTD_NEXT(ehci, qtd->qtd_dma);
723			list_add_tail (&qtd->qtd_list, head);
724
725			/* never any data in such packets */
726			qtd_fill(ehci, qtd, 0, 0, token, 0);
727		}
728	}
729
730	/* by default, enable interrupt on urb completion */
731	if (likely (!(urb->transfer_flags & URB_NO_INTERRUPT)))
732		qtd->hw_token |= cpu_to_hc32(ehci, QTD_IOC);
733	return head;
734
735cleanup:
736	qtd_list_free (ehci, urb, head);
737	return NULL;
738}
739
740/*-------------------------------------------------------------------------*/
741
742// Would be best to create all qh's from config descriptors,
743// when each interface/altsetting is established.  Unlink
744// any previous qh and cancel its urbs first; endpoints are
745// implicitly reset then (data toggle too).
746// That'd mean updating how usbcore talks to HCDs. (2.7?)
747
748
749/*
750 * Each QH holds a qtd list; a QH is used for everything except iso.
751 *
752 * For interrupt urbs, the scheduler must set the microframe scheduling
753 * mask(s) each time the QH gets scheduled.  For highspeed, that's
754 * just one microframe in the s-mask.  For split interrupt transactions
755 * there are additional complications: c-mask, maybe FSTNs.
756 */
757static struct ehci_qh *
758qh_make (
759	struct ehci_hcd		*ehci,
760	struct urb		*urb,
761	gfp_t			flags
762) {
763	struct ehci_qh		*qh = ehci_qh_alloc (ehci, flags);
764	u32			info1 = 0, info2 = 0;
765	int			is_input, type;
766	int			maxp = 0;
767	struct usb_tt		*tt = urb->dev->tt;
768	struct ehci_qh_hw	*hw;
769
770	if (!qh)
771		return qh;
772
773	/*
774	 * init endpoint/device data for this QH
775	 */
776	info1 |= usb_pipeendpoint (urb->pipe) << 8;
777	info1 |= usb_pipedevice (urb->pipe) << 0;
778
779	is_input = usb_pipein (urb->pipe);
780	type = usb_pipetype (urb->pipe);
781	maxp = usb_maxpacket (urb->dev, urb->pipe, !is_input);
782
783	/* 1024 byte maxpacket is a hardware ceiling.  High bandwidth
784	 * acts like up to 3KB, but is built from smaller packets.
785	 */
786	if (max_packet(maxp) > 1024) {
787		ehci_dbg(ehci, "bogus qh maxpacket %d\n", max_packet(maxp));
788		goto done;
789	}
790
791	/* Compute interrupt scheduling parameters just once, and save.
792	 * - allowing for high bandwidth, how many nsec/uframe are used?
793	 * - split transactions need a second CSPLIT uframe; same question
794	 * - splits also need a schedule gap (for full/low speed I/O)
795	 * - qh has a polling interval
796	 *
797	 * For control/bulk requests, the HC or TT handles these.
798	 */
799	if (type == PIPE_INTERRUPT) {
800		unsigned	tmp;
801
802		qh->ps.usecs = NS_TO_US(usb_calc_bus_time(USB_SPEED_HIGH,
803				is_input, 0,
804				hb_mult(maxp) * max_packet(maxp)));
805		qh->ps.phase = NO_FRAME;
806
807		if (urb->dev->speed == USB_SPEED_HIGH) {
808			qh->ps.c_usecs = 0;
809			qh->gap_uf = 0;
810
811			if (urb->interval > 1 && urb->interval < 8) {
812				/* NOTE interval 2 or 4 uframes could work.
813				 * But interval 1 scheduling is simpler, and
814				 * includes high bandwidth.
815				 */
816				urb->interval = 1;
817			} else if (urb->interval > ehci->periodic_size << 3) {
818				urb->interval = ehci->periodic_size << 3;
819			}
820			qh->ps.period = urb->interval >> 3;
821
822			/* period for bandwidth allocation */
823			tmp = min_t(unsigned, EHCI_BANDWIDTH_SIZE,
824					1 << (urb->ep->desc.bInterval - 1));
825
826			/* Allow urb->interval to override */
827			qh->ps.bw_uperiod = min_t(unsigned, tmp, urb->interval);
828			qh->ps.bw_period = qh->ps.bw_uperiod >> 3;
829		} else {
830			int		think_time;
831
832			/* gap is f(FS/LS transfer times) */
833			qh->gap_uf = 1 + usb_calc_bus_time (urb->dev->speed,
834					is_input, 0, maxp) / (125 * 1000);
835
836			/* FIXME this just approximates SPLIT/CSPLIT times */
837			if (is_input) {		// SPLIT, gap, CSPLIT+DATA
838				qh->ps.c_usecs = qh->ps.usecs + HS_USECS(0);
839				qh->ps.usecs = HS_USECS(1);
840			} else {		// SPLIT+DATA, gap, CSPLIT
841				qh->ps.usecs += HS_USECS(1);
842				qh->ps.c_usecs = HS_USECS(0);
843			}
844
845			think_time = tt ? tt->think_time : 0;
846			qh->ps.tt_usecs = NS_TO_US(think_time +
847					usb_calc_bus_time (urb->dev->speed,
848					is_input, 0, max_packet (maxp)));
849			if (urb->interval > ehci->periodic_size)
850				urb->interval = ehci->periodic_size;
851			qh->ps.period = urb->interval;
852
853			/* period for bandwidth allocation */
854			tmp = min_t(unsigned, EHCI_BANDWIDTH_FRAMES,
855					urb->ep->desc.bInterval);
856			tmp = rounddown_pow_of_two(tmp);
857
858			/* Allow urb->interval to override */
859			qh->ps.bw_period = min_t(unsigned, tmp, urb->interval);
860			qh->ps.bw_uperiod = qh->ps.bw_period << 3;
861		}
862	}
863
864	/* support for tt scheduling, and access to toggles */
865	qh->ps.udev = urb->dev;
866	qh->ps.ep = urb->ep;
867
868	/* using TT? */
869	switch (urb->dev->speed) {
870	case USB_SPEED_LOW:
871		info1 |= QH_LOW_SPEED;
872		/* FALL THROUGH */
873
874	case USB_SPEED_FULL:
875		/* EPS 0 means "full" */
876		if (type != PIPE_INTERRUPT)
877			info1 |= (EHCI_TUNE_RL_TT << 28);
878		if (type == PIPE_CONTROL) {
879			info1 |= QH_CONTROL_EP;		/* for TT */
880			info1 |= QH_TOGGLE_CTL;		/* toggle from qtd */
881		}
882		info1 |= maxp << 16;
883
884		info2 |= (EHCI_TUNE_MULT_TT << 30);
885
886		/* Some Freescale processors have an erratum in which the
887		 * port number in the queue head was 0..N-1 instead of 1..N.
888		 */
889		if (ehci_has_fsl_portno_bug(ehci))
890			info2 |= (urb->dev->ttport-1) << 23;
891		else
892			info2 |= urb->dev->ttport << 23;
893
894		/* set the address of the TT; for TDI's integrated
895		 * root hub tt, leave it zeroed.
896		 */
897		if (tt && tt->hub != ehci_to_hcd(ehci)->self.root_hub)
898			info2 |= tt->hub->devnum << 16;
899
900		/* NOTE:  if (PIPE_INTERRUPT) { scheduler sets c-mask } */
901
902		break;
903
904	case USB_SPEED_HIGH:		/* no TT involved */
905		info1 |= QH_HIGH_SPEED;
906		if (type == PIPE_CONTROL) {
907			info1 |= (EHCI_TUNE_RL_HS << 28);
908			info1 |= 64 << 16;	/* usb2 fixed maxpacket */
909			info1 |= QH_TOGGLE_CTL;	/* toggle from qtd */
910			info2 |= (EHCI_TUNE_MULT_HS << 30);
911		} else if (type == PIPE_BULK) {
912			info1 |= (EHCI_TUNE_RL_HS << 28);
913			/* The USB spec says that high speed bulk endpoints
914			 * always use 512 byte maxpacket.  But some device
915			 * vendors decided to ignore that, and MSFT is happy
916			 * to help them do so.  So now people expect to use
917			 * such nonconformant devices with Linux too; sigh.
918			 */
919			info1 |= max_packet(maxp) << 16;
920			info2 |= (EHCI_TUNE_MULT_HS << 30);
921		} else {		/* PIPE_INTERRUPT */
922			info1 |= max_packet (maxp) << 16;
923			info2 |= hb_mult (maxp) << 30;
924		}
925		break;
926	default:
927		ehci_dbg(ehci, "bogus dev %p speed %d\n", urb->dev,
928			urb->dev->speed);
929done:
930		qh_destroy(ehci, qh);
931		return NULL;
932	}
933
934	/* NOTE:  if (PIPE_INTERRUPT) { scheduler sets s-mask } */
935
936	/* init as live, toggle clear */
937	qh->qh_state = QH_STATE_IDLE;
938	hw = qh->hw;
939	hw->hw_info1 = cpu_to_hc32(ehci, info1);
940	hw->hw_info2 = cpu_to_hc32(ehci, info2);
941	qh->is_out = !is_input;
942	usb_settoggle (urb->dev, usb_pipeendpoint (urb->pipe), !is_input, 1);
943	return qh;
944}
945
946/*-------------------------------------------------------------------------*/
947
948static void enable_async(struct ehci_hcd *ehci)
949{
950	if (ehci->async_count++)
951		return;
952
953	/* Stop waiting to turn off the async schedule */
954	ehci->enabled_hrtimer_events &= ~BIT(EHCI_HRTIMER_DISABLE_ASYNC);
955
956	/* Don't start the schedule until ASS is 0 */
957	ehci_poll_ASS(ehci);
958	turn_on_io_watchdog(ehci);
959}
960
961static void disable_async(struct ehci_hcd *ehci)
962{
963	if (--ehci->async_count)
964		return;
965
966	/* The async schedule and unlink lists are supposed to be empty */
967	WARN_ON(ehci->async->qh_next.qh || !list_empty(&ehci->async_unlink) ||
968			!list_empty(&ehci->async_idle));
969
970	/* Don't turn off the schedule until ASS is 1 */
971	ehci_poll_ASS(ehci);
972}
973
974/* move qh (and its qtds) onto async queue; maybe enable queue.  */
975
976static void qh_link_async (struct ehci_hcd *ehci, struct ehci_qh *qh)
977{
978	__hc32		dma = QH_NEXT(ehci, qh->qh_dma);
979	struct ehci_qh	*head;
980
981	/* Don't link a QH if there's a Clear-TT-Buffer pending */
982	if (unlikely(qh->clearing_tt))
983		return;
984
985	WARN_ON(qh->qh_state != QH_STATE_IDLE);
986
987	/* clear halt and/or toggle; and maybe recover from silicon quirk */
988	qh_refresh(ehci, qh);
989
990	/* splice right after start */
991	head = ehci->async;
992	qh->qh_next = head->qh_next;
993	qh->hw->hw_next = head->hw->hw_next;
994	wmb ();
995
996	head->qh_next.qh = qh;
997	head->hw->hw_next = dma;
998
999	qh->qh_state = QH_STATE_LINKED;
1000	qh->xacterrs = 0;
1001	qh->exception = 0;
1002	/* qtd completions reported later by interrupt */
1003
1004	enable_async(ehci);
1005}
1006
1007/*-------------------------------------------------------------------------*/
1008
1009/*
1010 * For control/bulk/interrupt, return QH with these TDs appended.
1011 * Allocates and initializes the QH if necessary.
1012 * Returns null if it can't allocate a QH it needs to.
1013 * If the QH has TDs (urbs) already, that's great.
1014 */
1015static struct ehci_qh *qh_append_tds (
1016	struct ehci_hcd		*ehci,
1017	struct urb		*urb,
1018	struct list_head	*qtd_list,
1019	int			epnum,
1020	void			**ptr
1021)
1022{
1023	struct ehci_qh		*qh = NULL;
1024	__hc32			qh_addr_mask = cpu_to_hc32(ehci, 0x7f);
1025
1026	qh = (struct ehci_qh *) *ptr;
1027	if (unlikely (qh == NULL)) {
1028		/* can't sleep here, we have ehci->lock... */
1029		qh = qh_make (ehci, urb, GFP_ATOMIC);
1030		*ptr = qh;
1031	}
1032	if (likely (qh != NULL)) {
1033		struct ehci_qtd	*qtd;
1034
1035		if (unlikely (list_empty (qtd_list)))
1036			qtd = NULL;
1037		else
1038			qtd = list_entry (qtd_list->next, struct ehci_qtd,
1039					qtd_list);
1040
1041		/* control qh may need patching ... */
1042		if (unlikely (epnum == 0)) {
1043
1044                        /* usb_reset_device() briefly reverts to address 0 */
1045                        if (usb_pipedevice (urb->pipe) == 0)
1046				qh->hw->hw_info1 &= ~qh_addr_mask;
1047		}
1048
1049		/* just one way to queue requests: swap with the dummy qtd.
1050		 * only hc or qh_refresh() ever modify the overlay.
1051		 */
1052		if (likely (qtd != NULL)) {
1053			struct ehci_qtd		*dummy;
1054			dma_addr_t		dma;
1055			__hc32			token;
1056
1057			/* to avoid racing the HC, use the dummy td instead of
1058			 * the first td of our list (becomes new dummy).  both
1059			 * tds stay deactivated until we're done, when the
1060			 * HC is allowed to fetch the old dummy (4.10.2).
1061			 */
1062			token = qtd->hw_token;
1063			qtd->hw_token = HALT_BIT(ehci);
1064
1065			dummy = qh->dummy;
1066
1067			dma = dummy->qtd_dma;
1068			*dummy = *qtd;
1069			dummy->qtd_dma = dma;
1070
1071			list_del (&qtd->qtd_list);
1072			list_add (&dummy->qtd_list, qtd_list);
1073			list_splice_tail(qtd_list, &qh->qtd_list);
1074
1075			ehci_qtd_init(ehci, qtd, qtd->qtd_dma);
1076			qh->dummy = qtd;
1077
1078			/* hc must see the new dummy at list end */
1079			dma = qtd->qtd_dma;
1080			qtd = list_entry (qh->qtd_list.prev,
1081					struct ehci_qtd, qtd_list);
1082			qtd->hw_next = QTD_NEXT(ehci, dma);
1083
1084			/* let the hc process these next qtds */
1085			wmb ();
1086			dummy->hw_token = token;
1087
1088			urb->hcpriv = qh;
1089		}
1090	}
1091	return qh;
1092}
1093
1094/*-------------------------------------------------------------------------*/
1095
1096static int
1097submit_async (
1098	struct ehci_hcd		*ehci,
1099	struct urb		*urb,
1100	struct list_head	*qtd_list,
1101	gfp_t			mem_flags
1102) {
1103	int			epnum;
1104	unsigned long		flags;
1105	struct ehci_qh		*qh = NULL;
1106	int			rc;
1107
1108	epnum = urb->ep->desc.bEndpointAddress;
1109
1110#ifdef EHCI_URB_TRACE
1111	{
1112		struct ehci_qtd *qtd;
1113		qtd = list_entry(qtd_list->next, struct ehci_qtd, qtd_list);
1114		ehci_dbg(ehci,
1115			 "%s %s urb %p ep%d%s len %d, qtd %p [qh %p]\n",
1116			 __func__, urb->dev->devpath, urb,
1117			 epnum & 0x0f, (epnum & USB_DIR_IN) ? "in" : "out",
1118			 urb->transfer_buffer_length,
1119			 qtd, urb->ep->hcpriv);
1120	}
1121#endif
1122
1123	spin_lock_irqsave (&ehci->lock, flags);
1124	if (unlikely(!HCD_HW_ACCESSIBLE(ehci_to_hcd(ehci)))) {
1125		rc = -ESHUTDOWN;
1126		goto done;
1127	}
1128	rc = usb_hcd_link_urb_to_ep(ehci_to_hcd(ehci), urb);
1129	if (unlikely(rc))
1130		goto done;
1131
1132	qh = qh_append_tds(ehci, urb, qtd_list, epnum, &urb->ep->hcpriv);
1133	if (unlikely(qh == NULL)) {
1134		usb_hcd_unlink_urb_from_ep(ehci_to_hcd(ehci), urb);
1135		rc = -ENOMEM;
1136		goto done;
1137	}
1138
1139	/* Control/bulk operations through TTs don't need scheduling,
1140	 * the HC and TT handle it when the TT has a buffer ready.
1141	 */
1142	if (likely (qh->qh_state == QH_STATE_IDLE))
1143		qh_link_async(ehci, qh);
1144 done:
1145	spin_unlock_irqrestore (&ehci->lock, flags);
1146	if (unlikely (qh == NULL))
1147		qtd_list_free (ehci, urb, qtd_list);
1148	return rc;
1149}
1150
1151/*-------------------------------------------------------------------------*/
1152#ifdef CONFIG_USB_HCD_TEST_MODE
1153/*
1154 * This function creates the qtds and submits them for the
1155 * SINGLE_STEP_SET_FEATURE Test.
1156 * This is done in two parts: first SETUP req for GetDesc is sent then
1157 * 15 seconds later, the IN stage for GetDesc starts to req data from dev
1158 *
1159 * is_setup : i/p arguement decides which of the two stage needs to be
1160 * performed; TRUE - SETUP and FALSE - IN+STATUS
1161 * Returns 0 if success
1162 */
1163static int submit_single_step_set_feature(
1164	struct usb_hcd  *hcd,
1165	struct urb      *urb,
1166	int             is_setup
1167) {
1168	struct ehci_hcd		*ehci = hcd_to_ehci(hcd);
1169	struct list_head	qtd_list;
1170	struct list_head	*head;
1171
1172	struct ehci_qtd		*qtd, *qtd_prev;
1173	dma_addr_t		buf;
1174	int			len, maxpacket;
1175	u32			token;
1176
1177	INIT_LIST_HEAD(&qtd_list);
1178	head = &qtd_list;
1179
1180	/* URBs map to sequences of QTDs:  one logical transaction */
1181	qtd = ehci_qtd_alloc(ehci, GFP_KERNEL);
1182	if (unlikely(!qtd))
1183		return -1;
1184	list_add_tail(&qtd->qtd_list, head);
1185	qtd->urb = urb;
1186
1187	token = QTD_STS_ACTIVE;
1188	token |= (EHCI_TUNE_CERR << 10);
1189
1190	len = urb->transfer_buffer_length;
1191	/*
1192	 * Check if the request is to perform just the SETUP stage (getDesc)
1193	 * as in SINGLE_STEP_SET_FEATURE test, DATA stage (IN) happens
1194	 * 15 secs after the setup
1195	 */
1196	if (is_setup) {
1197		/* SETUP pid */
1198		qtd_fill(ehci, qtd, urb->setup_dma,
1199				sizeof(struct usb_ctrlrequest),
1200				token | (2 /* "setup" */ << 8), 8);
1201
1202		submit_async(ehci, urb, &qtd_list, GFP_ATOMIC);
1203		return 0; /*Return now; we shall come back after 15 seconds*/
1204	}
1205
1206	/*
1207	 * IN: data transfer stage:  buffer setup : start the IN txn phase for
1208	 * the get_Desc SETUP which was sent 15seconds back
1209	 */
1210	token ^= QTD_TOGGLE;   /*We need to start IN with DATA-1 Pid-sequence*/
1211	buf = urb->transfer_dma;
1212
1213	token |= (1 /* "in" */ << 8);  /*This is IN stage*/
1214
1215	maxpacket = max_packet(usb_maxpacket(urb->dev, urb->pipe, 0));
1216
1217	qtd_fill(ehci, qtd, buf, len, token, maxpacket);
1218
1219	/*
1220	 * Our IN phase shall always be a short read; so keep the queue running
1221	 * and let it advance to the next qtd which zero length OUT status
1222	 */
1223	qtd->hw_alt_next = EHCI_LIST_END(ehci);
1224
1225	/* STATUS stage for GetDesc control request */
1226	token ^= 0x0100;        /* "in" <--> "out"  */
1227	token |= QTD_TOGGLE;    /* force DATA1 */
1228
1229	qtd_prev = qtd;
1230	qtd = ehci_qtd_alloc(ehci, GFP_ATOMIC);
1231	if (unlikely(!qtd))
1232		goto cleanup;
1233	qtd->urb = urb;
1234	qtd_prev->hw_next = QTD_NEXT(ehci, qtd->qtd_dma);
1235	list_add_tail(&qtd->qtd_list, head);
1236
1237	/* dont fill any data in such packets */
1238	qtd_fill(ehci, qtd, 0, 0, token, 0);
1239
1240	/* by default, enable interrupt on urb completion */
1241	if (likely(!(urb->transfer_flags & URB_NO_INTERRUPT)))
1242		qtd->hw_token |= cpu_to_hc32(ehci, QTD_IOC);
1243
1244	submit_async(ehci, urb, &qtd_list, GFP_KERNEL);
1245
1246	return 0;
1247
1248cleanup:
1249	qtd_list_free(ehci, urb, head);
1250	return -1;
1251}
1252#endif /* CONFIG_USB_HCD_TEST_MODE */
1253
1254/*-------------------------------------------------------------------------*/
1255
1256static void single_unlink_async(struct ehci_hcd *ehci, struct ehci_qh *qh)
1257{
1258	struct ehci_qh		*prev;
1259
1260	/* Add to the end of the list of QHs waiting for the next IAAD */
1261	qh->qh_state = QH_STATE_UNLINK_WAIT;
1262	list_add_tail(&qh->unlink_node, &ehci->async_unlink);
1263
1264	/* Unlink it from the schedule */
1265	prev = ehci->async;
1266	while (prev->qh_next.qh != qh)
1267		prev = prev->qh_next.qh;
1268
1269	prev->hw->hw_next = qh->hw->hw_next;
1270	prev->qh_next = qh->qh_next;
1271	if (ehci->qh_scan_next == qh)
1272		ehci->qh_scan_next = qh->qh_next.qh;
1273}
1274
1275static void start_iaa_cycle(struct ehci_hcd *ehci)
1276{
1277	/* Do nothing if an IAA cycle is already running */
1278	if (ehci->iaa_in_progress)
1279		return;
1280	ehci->iaa_in_progress = true;
1281
1282	/* If the controller isn't running, we don't have to wait for it */
1283	if (unlikely(ehci->rh_state < EHCI_RH_RUNNING)) {
1284		end_unlink_async(ehci);
1285
1286	/* Otherwise start a new IAA cycle */
1287	} else if (likely(ehci->rh_state == EHCI_RH_RUNNING)) {
1288
1289		/* Make sure the unlinks are all visible to the hardware */
1290		wmb();
1291
1292		ehci_writel(ehci, ehci->command | CMD_IAAD,
1293				&ehci->regs->command);
1294		ehci_readl(ehci, &ehci->regs->command);
1295		ehci_enable_event(ehci, EHCI_HRTIMER_IAA_WATCHDOG, true);
1296	}
1297}
1298
1299/* the async qh for the qtds being unlinked are now gone from the HC */
1300
1301static void end_unlink_async(struct ehci_hcd *ehci)
1302{
1303	struct ehci_qh		*qh;
1304	bool			early_exit;
1305
1306	if (ehci->has_synopsys_hc_bug)
1307		ehci_writel(ehci, (u32) ehci->async->qh_dma,
1308			    &ehci->regs->async_next);
1309
1310	/* The current IAA cycle has ended */
1311	ehci->iaa_in_progress = false;
1312
1313	if (list_empty(&ehci->async_unlink))
1314		return;
1315	qh = list_first_entry(&ehci->async_unlink, struct ehci_qh,
1316			unlink_node);	/* QH whose IAA cycle just ended */
1317
1318	/*
1319	 * If async_unlinking is set then this routine is already running,
1320	 * either on the stack or on another CPU.
1321	 */
1322	early_exit = ehci->async_unlinking;
1323
1324	/* If the controller isn't running, process all the waiting QHs */
1325	if (ehci->rh_state < EHCI_RH_RUNNING)
1326		list_splice_tail_init(&ehci->async_unlink, &ehci->async_idle);
1327
1328	/*
1329	 * Intel (?) bug: The HC can write back the overlay region even
1330	 * after the IAA interrupt occurs.  In self-defense, always go
1331	 * through two IAA cycles for each QH.
1332	 */
1333	else if (qh->qh_state == QH_STATE_UNLINK_WAIT) {
1334		qh->qh_state = QH_STATE_UNLINK;
1335		early_exit = true;
1336	}
1337
1338	/* Otherwise process only the first waiting QH (NVIDIA bug?) */
1339	else
1340		list_move_tail(&qh->unlink_node, &ehci->async_idle);
1341
1342	/* Start a new IAA cycle if any QHs are waiting for it */
1343	if (!list_empty(&ehci->async_unlink))
1344		start_iaa_cycle(ehci);
1345
1346	/*
1347	 * Don't allow nesting or concurrent calls,
1348	 * or wait for the second IAA cycle for the next QH.
1349	 */
1350	if (early_exit)
1351		return;
1352
1353	/* Process the idle QHs */
1354	ehci->async_unlinking = true;
1355	while (!list_empty(&ehci->async_idle)) {
1356		qh = list_first_entry(&ehci->async_idle, struct ehci_qh,
1357				unlink_node);
1358		list_del(&qh->unlink_node);
1359
1360		qh->qh_state = QH_STATE_IDLE;
1361		qh->qh_next.qh = NULL;
1362
1363		if (!list_empty(&qh->qtd_list))
1364			qh_completions(ehci, qh);
1365		if (!list_empty(&qh->qtd_list) &&
1366				ehci->rh_state == EHCI_RH_RUNNING)
1367			qh_link_async(ehci, qh);
1368		disable_async(ehci);
1369	}
1370	ehci->async_unlinking = false;
1371}
1372
1373static void start_unlink_async(struct ehci_hcd *ehci, struct ehci_qh *qh);
1374
1375static void unlink_empty_async(struct ehci_hcd *ehci)
1376{
1377	struct ehci_qh		*qh;
1378	struct ehci_qh		*qh_to_unlink = NULL;
1379	int			count = 0;
1380
1381	/* Find the last async QH which has been empty for a timer cycle */
1382	for (qh = ehci->async->qh_next.qh; qh; qh = qh->qh_next.qh) {
1383		if (list_empty(&qh->qtd_list) &&
1384				qh->qh_state == QH_STATE_LINKED) {
1385			++count;
1386			if (qh->unlink_cycle != ehci->async_unlink_cycle)
1387				qh_to_unlink = qh;
1388		}
1389	}
1390
1391	/* If nothing else is being unlinked, unlink the last empty QH */
1392	if (list_empty(&ehci->async_unlink) && qh_to_unlink) {
1393		start_unlink_async(ehci, qh_to_unlink);
1394		--count;
1395	}
1396
1397	/* Other QHs will be handled later */
1398	if (count > 0) {
1399		ehci_enable_event(ehci, EHCI_HRTIMER_ASYNC_UNLINKS, true);
1400		++ehci->async_unlink_cycle;
1401	}
1402}
1403
1404/* The root hub is suspended; unlink all the async QHs */
1405static void __maybe_unused unlink_empty_async_suspended(struct ehci_hcd *ehci)
1406{
1407	struct ehci_qh		*qh;
1408
1409	while (ehci->async->qh_next.qh) {
1410		qh = ehci->async->qh_next.qh;
1411		WARN_ON(!list_empty(&qh->qtd_list));
1412		single_unlink_async(ehci, qh);
1413	}
1414	start_iaa_cycle(ehci);
1415}
1416
1417/* makes sure the async qh will become idle */
1418/* caller must own ehci->lock */
1419
1420static void start_unlink_async(struct ehci_hcd *ehci, struct ehci_qh *qh)
1421{
1422	/* If the QH isn't linked then there's nothing we can do. */
1423	if (qh->qh_state != QH_STATE_LINKED)
1424		return;
1425
1426	single_unlink_async(ehci, qh);
1427	start_iaa_cycle(ehci);
1428}
1429
1430/*-------------------------------------------------------------------------*/
1431
1432static void scan_async (struct ehci_hcd *ehci)
1433{
1434	struct ehci_qh		*qh;
1435	bool			check_unlinks_later = false;
1436
1437	ehci->qh_scan_next = ehci->async->qh_next.qh;
1438	while (ehci->qh_scan_next) {
1439		qh = ehci->qh_scan_next;
1440		ehci->qh_scan_next = qh->qh_next.qh;
1441
1442		/* clean any finished work for this qh */
1443		if (!list_empty(&qh->qtd_list)) {
1444			int temp;
1445
1446			/*
1447			 * Unlinks could happen here; completion reporting
1448			 * drops the lock.  That's why ehci->qh_scan_next
1449			 * always holds the next qh to scan; if the next qh
1450			 * gets unlinked then ehci->qh_scan_next is adjusted
1451			 * in single_unlink_async().
1452			 */
1453			temp = qh_completions(ehci, qh);
1454			if (unlikely(temp)) {
1455				start_unlink_async(ehci, qh);
1456			} else if (list_empty(&qh->qtd_list)
1457					&& qh->qh_state == QH_STATE_LINKED) {
1458				qh->unlink_cycle = ehci->async_unlink_cycle;
1459				check_unlinks_later = true;
1460			}
1461		}
1462	}
1463
1464	/*
1465	 * Unlink empty entries, reducing DMA usage as well
1466	 * as HCD schedule-scanning costs.  Delay for any qh
1467	 * we just scanned, there's a not-unusual case that it
1468	 * doesn't stay idle for long.
1469	 */
1470	if (check_unlinks_later && ehci->rh_state == EHCI_RH_RUNNING &&
1471			!(ehci->enabled_hrtimer_events &
1472				BIT(EHCI_HRTIMER_ASYNC_UNLINKS))) {
1473		ehci_enable_event(ehci, EHCI_HRTIMER_ASYNC_UNLINKS, true);
1474		++ehci->async_unlink_cycle;
1475	}
1476}
1477