1// SPDX-License-Identifier: GPL-2.0+
   2/*
   3 * Copyright (c) 2001-2004 by David Brownell
   4 * Copyright (c) 2003 Michal Sojka, for high-speed iso transfers
   5 */
   6
   7/* this file is part of ehci-hcd.c */
   8
   9/*-------------------------------------------------------------------------*/
  10
  11/*
  12 * EHCI scheduled transaction support:  interrupt, iso, split iso
  13 * These are called "periodic" transactions in the EHCI spec.
  14 *
  15 * Note that for interrupt transfers, the QH/QTD manipulation is shared
  16 * with the "asynchronous" transaction support (control/bulk transfers).
  17 * The only real difference is in how interrupt transfers are scheduled.
  18 *
  19 * For ISO, we make an "iso_stream" head to serve the same role as a QH.
  20 * It keeps track of every ITD (or SITD) that's linked, and holds enough
  21 * pre-calculated schedule data to make appending to the queue be quick.
  22 */
  23
  24static int ehci_get_frame(struct usb_hcd *hcd);
  25
  26/*
  27 * periodic_next_shadow - return "next" pointer on shadow list
  28 * @periodic: host pointer to qh/itd/sitd
  29 * @tag: hardware tag for type of this record
  30 */
  31static union ehci_shadow *
  32periodic_next_shadow(struct ehci_hcd *ehci, union ehci_shadow *periodic,
  33		__hc32 tag)
  34{
  35	switch (hc32_to_cpu(ehci, tag)) {
  36	case Q_TYPE_QH:
  37		return &periodic->qh->qh_next;
  38	case Q_TYPE_FSTN:
  39		return &periodic->fstn->fstn_next;
  40	case Q_TYPE_ITD:
  41		return &periodic->itd->itd_next;
  42	/* case Q_TYPE_SITD: */
  43	default:
  44		return &periodic->sitd->sitd_next;
  45	}
  46}
  47
  48static __hc32 *
  49shadow_next_periodic(struct ehci_hcd *ehci, union ehci_shadow *periodic,
  50		__hc32 tag)
  51{
  52	switch (hc32_to_cpu(ehci, tag)) {
  53	/* our ehci_shadow.qh is actually software part */
  54	case Q_TYPE_QH:
  55		return &periodic->qh->hw->hw_next;
  56	/* others are hw parts */
  57	default:
  58		return periodic->hw_next;
  59	}
  60}
  61
  62/* caller must hold ehci->lock */
  63static void periodic_unlink(struct ehci_hcd *ehci, unsigned frame, void *ptr)
  64{
  65	union ehci_shadow	*prev_p = &ehci->pshadow[frame];
  66	__hc32			*hw_p = &ehci->periodic[frame];
  67	union ehci_shadow	here = *prev_p;
  68
  69	/* find predecessor of "ptr"; hw and shadow lists are in sync */
  70	while (here.ptr && here.ptr != ptr) {
  71		prev_p = periodic_next_shadow(ehci, prev_p,
  72				Q_NEXT_TYPE(ehci, *hw_p));
  73		hw_p = shadow_next_periodic(ehci, &here,
  74				Q_NEXT_TYPE(ehci, *hw_p));
  75		here = *prev_p;
  76	}
  77	/* an interrupt entry (at list end) could have been shared */
  78	if (!here.ptr)
  79		return;
  80
  81	/* update shadow and hardware lists ... the old "next" pointers
  82	 * from ptr may still be in use, the caller updates them.
  83	 */
  84	*prev_p = *periodic_next_shadow(ehci, &here,
  85			Q_NEXT_TYPE(ehci, *hw_p));
  86
  87	if (!ehci->use_dummy_qh ||
  88	    *shadow_next_periodic(ehci, &here, Q_NEXT_TYPE(ehci, *hw_p))
  89			!= EHCI_LIST_END(ehci))
  90		*hw_p = *shadow_next_periodic(ehci, &here,
  91				Q_NEXT_TYPE(ehci, *hw_p));
  92	else
  93		*hw_p = cpu_to_hc32(ehci, ehci->dummy->qh_dma);
  94}
  95
  96/*-------------------------------------------------------------------------*/
  97
  98/* Bandwidth and TT management */
  99
 100/* Find the TT data structure for this device; create it if necessary */
 101static struct ehci_tt *find_tt(struct usb_device *udev)
 102{
 103	struct usb_tt		*utt = udev->tt;
 104	struct ehci_tt		*tt, **tt_index, **ptt;
 105	unsigned		port;
 106	bool			allocated_index = false;
 107
 108	if (!utt)
 109		return NULL;		/* Not below a TT */
 110
 111	/*
 112	 * Find/create our data structure.
 113	 * For hubs with a single TT, we get it directly.
 114	 * For hubs with multiple TTs, there's an extra level of pointers.
 115	 */
 116	tt_index = NULL;
 117	if (utt->multi) {
 118		tt_index = utt->hcpriv;
 119		if (!tt_index) {		/* Create the index array */
 120			tt_index = kcalloc(utt->hub->maxchild,
 121					   sizeof(*tt_index),
 122					   GFP_ATOMIC);
 123			if (!tt_index)
 124				return ERR_PTR(-ENOMEM);
 125			utt->hcpriv = tt_index;
 126			allocated_index = true;
 127		}
 128		port = udev->ttport - 1;
 129		ptt = &tt_index[port];
 130	} else {
 131		port = 0;
 132		ptt = (struct ehci_tt **) &utt->hcpriv;
 133	}
 134
 135	tt = *ptt;
 136	if (!tt) {				/* Create the ehci_tt */
 137		struct ehci_hcd		*ehci =
 138				hcd_to_ehci(bus_to_hcd(udev->bus));
 139
 140		tt = kzalloc(sizeof(*tt), GFP_ATOMIC);
 141		if (!tt) {
 142			if (allocated_index) {
 143				utt->hcpriv = NULL;
 144				kfree(tt_index);
 145			}
 146			return ERR_PTR(-ENOMEM);
 147		}
 148		list_add_tail(&tt->tt_list, &ehci->tt_list);
 149		INIT_LIST_HEAD(&tt->ps_list);
 150		tt->usb_tt = utt;
 151		tt->tt_port = port;
 152		*ptt = tt;
 153	}
 154
 155	return tt;
 156}
 157
 158/* Release the TT above udev, if it's not in use */
 159static void drop_tt(struct usb_device *udev)
 160{
 161	struct usb_tt		*utt = udev->tt;
 162	struct ehci_tt		*tt, **tt_index, **ptt;
 163	int			cnt, i;
 164
 165	if (!utt || !utt->hcpriv)
 166		return;		/* Not below a TT, or never allocated */
 167
 168	cnt = 0;
 169	if (utt->multi) {
 170		tt_index = utt->hcpriv;
 171		ptt = &tt_index[udev->ttport - 1];
 172
 173		/* How many entries are left in tt_index? */
 174		for (i = 0; i < utt->hub->maxchild; ++i)
 175			cnt += !!tt_index[i];
 176	} else {
 177		tt_index = NULL;
 178		ptt = (struct ehci_tt **) &utt->hcpriv;
 179	}
 180
 181	tt = *ptt;
 182	if (!tt || !list_empty(&tt->ps_list))
 183		return;		/* never allocated, or still in use */
 184
 185	list_del(&tt->tt_list);
 186	*ptt = NULL;
 187	kfree(tt);
 188	if (cnt == 1) {
 189		utt->hcpriv = NULL;
 190		kfree(tt_index);
 191	}
 192}
 193
 194static void bandwidth_dbg(struct ehci_hcd *ehci, int sign, char *type,
 195		struct ehci_per_sched *ps)
 196{
 197	dev_dbg(&ps->udev->dev,
 198			"ep %02x: %s %s @ %u+%u (%u.%u+%u) [%u/%u us] mask %04x\n",
 199			ps->ep->desc.bEndpointAddress,
 200			(sign >= 0 ? "reserve" : "release"), type,
 201			(ps->bw_phase << 3) + ps->phase_uf, ps->bw_uperiod,
 202			ps->phase, ps->phase_uf, ps->period,
 203			ps->usecs, ps->c_usecs, ps->cs_mask);
 204}
 205
 206static void reserve_release_intr_bandwidth(struct ehci_hcd *ehci,
 207		struct ehci_qh *qh, int sign)
 208{
 209	unsigned		start_uf;
 210	unsigned		i, j, m;
 211	int			usecs = qh->ps.usecs;
 212	int			c_usecs = qh->ps.c_usecs;
 213	int			tt_usecs = qh->ps.tt_usecs;
 214	struct ehci_tt		*tt;
 215
 216	if (qh->ps.phase == NO_FRAME)	/* Bandwidth wasn't reserved */
 217		return;
 218	start_uf = qh->ps.bw_phase << 3;
 219
 220	bandwidth_dbg(ehci, sign, "intr", &qh->ps);
 221
 222	if (sign < 0) {		/* Release bandwidth */
 223		usecs = -usecs;
 224		c_usecs = -c_usecs;
 225		tt_usecs = -tt_usecs;
 226	}
 227
 228	/* Entire transaction (high speed) or start-split (full/low speed) */
 229	for (i = start_uf + qh->ps.phase_uf; i < EHCI_BANDWIDTH_SIZE;
 230			i += qh->ps.bw_uperiod)
 231		ehci->bandwidth[i] += usecs;
 232
 233	/* Complete-split (full/low speed) */
 234	if (qh->ps.c_usecs) {
 235		/* NOTE: adjustments needed for FSTN */
 236		for (i = start_uf; i < EHCI_BANDWIDTH_SIZE;
 237				i += qh->ps.bw_uperiod) {
 238			for ((j = 2, m = 1 << (j+8)); j < 8; (++j, m <<= 1)) {
 239				if (qh->ps.cs_mask & m)
 240					ehci->bandwidth[i+j] += c_usecs;
 241			}
 242		}
 243	}
 244
 245	/* FS/LS bus bandwidth */
 246	if (tt_usecs) {
 247		/*
 248		 * find_tt() will not return any error here as we have
 249		 * already called find_tt() before calling this function
 250		 * and checked for any error return. The previous call
 251		 * would have created the data structure.
 252		 */
 253		tt = find_tt(qh->ps.udev);
 254		if (sign > 0)
 255			list_add_tail(&qh->ps.ps_list, &tt->ps_list);
 256		else
 257			list_del(&qh->ps.ps_list);
 258
 259		for (i = start_uf >> 3; i < EHCI_BANDWIDTH_FRAMES;
 260				i += qh->ps.bw_period)
 261			tt->bandwidth[i] += tt_usecs;
 262	}
 263}
 264
 265/*-------------------------------------------------------------------------*/
 266
 267static void compute_tt_budget(u8 budget_table[EHCI_BANDWIDTH_SIZE],
 268		struct ehci_tt *tt)
 269{
 270	struct ehci_per_sched	*ps;
 271	unsigned		uframe, uf, x;
 272	u8			*budget_line;
 273
 274	if (!tt)
 275		return;
 276	memset(budget_table, 0, EHCI_BANDWIDTH_SIZE);
 277
 278	/* Add up the contributions from all the endpoints using this TT */
 279	list_for_each_entry(ps, &tt->ps_list, ps_list) {
 280		for (uframe = ps->bw_phase << 3; uframe < EHCI_BANDWIDTH_SIZE;
 281				uframe += ps->bw_uperiod) {
 282			budget_line = &budget_table[uframe];
 283			x = ps->tt_usecs;
 284
 285			/* propagate the time forward */
 286			for (uf = ps->phase_uf; uf < 8; ++uf) {
 287				x += budget_line[uf];
 288
 289				/* Each microframe lasts 125 us */
 290				if (x <= 125) {
 291					budget_line[uf] = x;
 292					break;
 293				}
 294				budget_line[uf] = 125;
 295				x -= 125;
 296			}
 297		}
 298	}
 299}
 300
 301static int __maybe_unused same_tt(struct usb_device *dev1,
 302		struct usb_device *dev2)
 303{
 304	if (!dev1->tt || !dev2->tt)
 305		return 0;
 306	if (dev1->tt != dev2->tt)
 307		return 0;
 308	if (dev1->tt->multi)
 309		return dev1->ttport == dev2->ttport;
 310	else
 311		return 1;
 312}
 313
 314#ifdef CONFIG_USB_EHCI_TT_NEWSCHED
 315
 316static const unsigned char
 317max_tt_usecs[] = { 125, 125, 125, 125, 125, 125, 30, 0 };
 318
 319/* carryover low/fullspeed bandwidth that crosses uframe boundries */
 320static inline void carryover_tt_bandwidth(unsigned short tt_usecs[8])
 321{
 322	int i;
 323
 324	for (i = 0; i < 7; i++) {
 325		if (max_tt_usecs[i] < tt_usecs[i]) {
 326			tt_usecs[i+1] += tt_usecs[i] - max_tt_usecs[i];
 327			tt_usecs[i] = max_tt_usecs[i];
 328		}
 329	}
 330}
 331
 332/*
 333 * Return true if the device's tt's downstream bus is available for a
 334 * periodic transfer of the specified length (usecs), starting at the
 335 * specified frame/uframe.  Note that (as summarized in section 11.19
 336 * of the usb 2.0 spec) TTs can buffer multiple transactions for each
 337 * uframe.
 338 *
 339 * The uframe parameter is when the fullspeed/lowspeed transfer
 340 * should be executed in "B-frame" terms, which is the same as the
 341 * highspeed ssplit's uframe (which is in "H-frame" terms).  For example
 342 * a ssplit in "H-frame" 0 causes a transfer in "B-frame" 0.
 343 * See the EHCI spec sec 4.5 and fig 4.7.
 344 *
 345 * This checks if the full/lowspeed bus, at the specified starting uframe,
 346 * has the specified bandwidth available, according to rules listed
 347 * in USB 2.0 spec section 11.18.1 fig 11-60.
 348 *
 349 * This does not check if the transfer would exceed the max ssplit
 350 * limit of 16, specified in USB 2.0 spec section 11.18.4 requirement #4,
 351 * since proper scheduling limits ssplits to less than 16 per uframe.
 352 */
 353static int tt_available(
 354	struct ehci_hcd		*ehci,
 355	struct ehci_per_sched	*ps,
 356	struct ehci_tt		*tt,
 357	unsigned		frame,
 358	unsigned		uframe
 359)
 360{
 361	unsigned		period = ps->bw_period;
 362	unsigned		usecs = ps->tt_usecs;
 363
 364	if ((period == 0) || (uframe >= 7))	/* error */
 365		return 0;
 366
 367	for (frame &= period - 1; frame < EHCI_BANDWIDTH_FRAMES;
 368			frame += period) {
 369		unsigned	i, uf;
 370		unsigned short	tt_usecs[8];
 371
 372		if (tt->bandwidth[frame] + usecs > 900)
 373			return 0;
 374
 375		uf = frame << 3;
 376		for (i = 0; i < 8; (++i, ++uf))
 377			tt_usecs[i] = ehci->tt_budget[uf];
 378
 379		if (max_tt_usecs[uframe] <= tt_usecs[uframe])
 380			return 0;
 381
 382		/* special case for isoc transfers larger than 125us:
 383		 * the first and each subsequent fully used uframe
 384		 * must be empty, so as to not illegally delay
 385		 * already scheduled transactions
 386		 */
 387		if (usecs > 125) {
 388			int ufs = (usecs / 125);
 389
 390			for (i = uframe; i < (uframe + ufs) && i < 8; i++)
 391				if (tt_usecs[i] > 0)
 392					return 0;
 393		}
 394
 395		tt_usecs[uframe] += usecs;
 396
 397		carryover_tt_bandwidth(tt_usecs);
 398
 399		/* fail if the carryover pushed bw past the last uframe's limit */
 400		if (max_tt_usecs[7] < tt_usecs[7])
 401			return 0;
 402	}
 403
 404	return 1;
 405}
 406
 407#else
 408
 409/* return true iff the device's transaction translator is available
 410 * for a periodic transfer starting at the specified frame, using
 411 * all the uframes in the mask.
 412 */
 413static int tt_no_collision(
 414	struct ehci_hcd		*ehci,
 415	unsigned		period,
 416	struct usb_device	*dev,
 417	unsigned		frame,
 418	u32			uf_mask
 419)
 420{
 421	if (period == 0)	/* error */
 422		return 0;
 423
 424	/* note bandwidth wastage:  split never follows csplit
 425	 * (different dev or endpoint) until the next uframe.
 426	 * calling convention doesn't make that distinction.
 427	 */
 428	for (; frame < ehci->periodic_size; frame += period) {
 429		union ehci_shadow	here;
 430		__hc32			type;
 431		struct ehci_qh_hw	*hw;
 432
 433		here = ehci->pshadow[frame];
 434		type = Q_NEXT_TYPE(ehci, ehci->periodic[frame]);
 435		while (here.ptr) {
 436			switch (hc32_to_cpu(ehci, type)) {
 437			case Q_TYPE_ITD:
 438				type = Q_NEXT_TYPE(ehci, here.itd->hw_next);
 439				here = here.itd->itd_next;
 440				continue;
 441			case Q_TYPE_QH:
 442				hw = here.qh->hw;
 443				if (same_tt(dev, here.qh->ps.udev)) {
 444					u32		mask;
 445
 446					mask = hc32_to_cpu(ehci,
 447							hw->hw_info2);
 448					/* "knows" no gap is needed */
 449					mask |= mask >> 8;
 450					if (mask & uf_mask)
 451						break;
 452				}
 453				type = Q_NEXT_TYPE(ehci, hw->hw_next);
 454				here = here.qh->qh_next;
 455				continue;
 456			case Q_TYPE_SITD:
 457				if (same_tt(dev, here.sitd->urb->dev)) {
 458					u16		mask;
 459
 460					mask = hc32_to_cpu(ehci, here.sitd
 461								->hw_uframe);
 462					/* FIXME assumes no gap for IN! */
 463					mask |= mask >> 8;
 464					if (mask & uf_mask)
 465						break;
 466				}
 467				type = Q_NEXT_TYPE(ehci, here.sitd->hw_next);
 468				here = here.sitd->sitd_next;
 469				continue;
 470			/* case Q_TYPE_FSTN: */
 471			default:
 472				ehci_dbg(ehci,
 473					"periodic frame %d bogus type %d\n",
 474					frame, type);
 475			}
 476
 477			/* collision or error */
 478			return 0;
 479		}
 480	}
 481
 482	/* no collision */
 483	return 1;
 484}
 485
 486#endif /* CONFIG_USB_EHCI_TT_NEWSCHED */
 487
 488/*-------------------------------------------------------------------------*/
 489
 490static void enable_periodic(struct ehci_hcd *ehci)
 491{
 492	if (ehci->periodic_count++)
 493		return;
 494
 495	/* Stop waiting to turn off the periodic schedule */
 496	ehci->enabled_hrtimer_events &= ~BIT(EHCI_HRTIMER_DISABLE_PERIODIC);
 497
 498	/* Don't start the schedule until PSS is 0 */
 499	ehci_poll_PSS(ehci);
 500	turn_on_io_watchdog(ehci);
 501}
 502
 503static void disable_periodic(struct ehci_hcd *ehci)
 504{
 505	if (--ehci->periodic_count)
 506		return;
 507
 508	/* Don't turn off the schedule until PSS is 1 */
 509	ehci_poll_PSS(ehci);
 510}
 511
 512/*-------------------------------------------------------------------------*/
 513
 514/* periodic schedule slots have iso tds (normal or split) first, then a
 515 * sparse tree for active interrupt transfers.
 516 *
 517 * this just links in a qh; caller guarantees uframe masks are set right.
 518 * no FSTN support (yet; ehci 0.96+)
 519 */
 520static void qh_link_periodic(struct ehci_hcd *ehci, struct ehci_qh *qh)
 521{
 522	unsigned	i;
 523	unsigned	period = qh->ps.period;
 524
 525	dev_dbg(&qh->ps.udev->dev,
 526		"link qh%d-%04x/%p start %d [%d/%d us]\n",
 527		period, hc32_to_cpup(ehci, &qh->hw->hw_info2)
 528			& (QH_CMASK | QH_SMASK),
 529		qh, qh->ps.phase, qh->ps.usecs, qh->ps.c_usecs);
 530
 531	/* high bandwidth, or otherwise every microframe */
 532	if (period == 0)
 533		period = 1;
 534
 535	for (i = qh->ps.phase; i < ehci->periodic_size; i += period) {
 536		union ehci_shadow	*prev = &ehci->pshadow[i];
 537		__hc32			*hw_p = &ehci->periodic[i];
 538		union ehci_shadow	here = *prev;
 539		__hc32			type = 0;
 540
 541		/* skip the iso nodes at list head */
 542		while (here.ptr) {
 543			type = Q_NEXT_TYPE(ehci, *hw_p);
 544			if (type == cpu_to_hc32(ehci, Q_TYPE_QH))
 545				break;
 546			prev = periodic_next_shadow(ehci, prev, type);
 547			hw_p = shadow_next_periodic(ehci, &here, type);
 548			here = *prev;
 549		}
 550
 551		/* sorting each branch by period (slow-->fast)
 552		 * enables sharing interior tree nodes
 553		 */
 554		while (here.ptr && qh != here.qh) {
 555			if (qh->ps.period > here.qh->ps.period)
 556				break;
 557			prev = &here.qh->qh_next;
 558			hw_p = &here.qh->hw->hw_next;
 559			here = *prev;
 560		}
 561		/* link in this qh, unless some earlier pass did that */
 562		if (qh != here.qh) {
 563			qh->qh_next = here;
 564			if (here.qh)
 565				qh->hw->hw_next = *hw_p;
 566			wmb();
 567			prev->qh = qh;
 568			*hw_p = QH_NEXT(ehci, qh->qh_dma);
 569		}
 570	}
 571	qh->qh_state = QH_STATE_LINKED;
 572	qh->xacterrs = 0;
 573	qh->unlink_reason = 0;
 574
 575	/* update per-qh bandwidth for debugfs */
 576	ehci_to_hcd(ehci)->self.bandwidth_allocated += qh->ps.bw_period
 577		? ((qh->ps.usecs + qh->ps.c_usecs) / qh->ps.bw_period)
 578		: (qh->ps.usecs * 8);
 579
 580	list_add(&qh->intr_node, &ehci->intr_qh_list);
 581
 582	/* maybe enable periodic schedule processing */
 583	++ehci->intr_count;
 584	enable_periodic(ehci);
 585}
 586
 587static void qh_unlink_periodic(struct ehci_hcd *ehci, struct ehci_qh *qh)
 588{
 589	unsigned	i;
 590	unsigned	period;
 591
 592	/*
 593	 * If qh is for a low/full-speed device, simply unlinking it
 594	 * could interfere with an ongoing split transaction.  To unlink
 595	 * it safely would require setting the QH_INACTIVATE bit and
 596	 * waiting at least one frame, as described in EHCI 4.12.2.5.
 597	 *
 598	 * We won't bother with any of this.  Instead, we assume that the
 599	 * only reason for unlinking an interrupt QH while the current URB
 600	 * is still active is to dequeue all the URBs (flush the whole
 601	 * endpoint queue).
 602	 *
 603	 * If rebalancing the periodic schedule is ever implemented, this
 604	 * approach will no longer be valid.
 605	 */
 606
 607	/* high bandwidth, or otherwise part of every microframe */
 608	period = qh->ps.period ? : 1;
 609
 610	for (i = qh->ps.phase; i < ehci->periodic_size; i += period)
 611		periodic_unlink(ehci, i, qh);
 612
 613	/* update per-qh bandwidth for debugfs */
 614	ehci_to_hcd(ehci)->self.bandwidth_allocated -= qh->ps.bw_period
 615		? ((qh->ps.usecs + qh->ps.c_usecs) / qh->ps.bw_period)
 616		: (qh->ps.usecs * 8);
 617
 618	dev_dbg(&qh->ps.udev->dev,
 619		"unlink qh%d-%04x/%p start %d [%d/%d us]\n",
 620		qh->ps.period,
 621		hc32_to_cpup(ehci, &qh->hw->hw_info2) & (QH_CMASK | QH_SMASK),
 622		qh, qh->ps.phase, qh->ps.usecs, qh->ps.c_usecs);
 623
 624	/* qh->qh_next still "live" to HC */
 625	qh->qh_state = QH_STATE_UNLINK;
 626	qh->qh_next.ptr = NULL;
 627
 628	if (ehci->qh_scan_next == qh)
 629		ehci->qh_scan_next = list_entry(qh->intr_node.next,
 630				struct ehci_qh, intr_node);
 631	list_del(&qh->intr_node);
 632}
 633
 634static void cancel_unlink_wait_intr(struct ehci_hcd *ehci, struct ehci_qh *qh)
 635{
 636	if (qh->qh_state != QH_STATE_LINKED ||
 637			list_empty(&qh->unlink_node))
 638		return;
 639
 640	list_del_init(&qh->unlink_node);
 641
 642	/*
 643	 * TODO: disable the event of EHCI_HRTIMER_START_UNLINK_INTR for
 644	 * avoiding unnecessary CPU wakeup
 645	 */
 646}
 647
 648static void start_unlink_intr(struct ehci_hcd *ehci, struct ehci_qh *qh)
 649{
 650	/* If the QH isn't linked then there's nothing we can do. */
 651	if (qh->qh_state != QH_STATE_LINKED)
 652		return;
 653
 654	/* if the qh is waiting for unlink, cancel it now */
 655	cancel_unlink_wait_intr(ehci, qh);
 656
 657	qh_unlink_periodic(ehci, qh);
 658
 659	/* Make sure the unlinks are visible before starting the timer */
 660	wmb();
 661
 662	/*
 663	 * The EHCI spec doesn't say how long it takes the controller to
 664	 * stop accessing an unlinked interrupt QH.  The timer delay is
 665	 * 9 uframes; presumably that will be long enough.
 666	 */
 667	qh->unlink_cycle = ehci->intr_unlink_cycle;
 668
 669	/* New entries go at the end of the intr_unlink list */
 670	list_add_tail(&qh->unlink_node, &ehci->intr_unlink);
 671
 672	if (ehci->intr_unlinking)
 673		;	/* Avoid recursive calls */
 674	else if (ehci->rh_state < EHCI_RH_RUNNING)
 675		ehci_handle_intr_unlinks(ehci);
 676	else if (ehci->intr_unlink.next == &qh->unlink_node) {
 677		ehci_enable_event(ehci, EHCI_HRTIMER_UNLINK_INTR, true);
 678		++ehci->intr_unlink_cycle;
 679	}
 680}
 681
 682/*
 683 * It is common only one intr URB is scheduled on one qh, and
 684 * given complete() is run in tasklet context, introduce a bit
 685 * delay to avoid unlink qh too early.
 686 */
 687static void start_unlink_intr_wait(struct ehci_hcd *ehci,
 688				   struct ehci_qh *qh)
 689{
 690	qh->unlink_cycle = ehci->intr_unlink_wait_cycle;
 691
 692	/* New entries go at the end of the intr_unlink_wait list */
 693	list_add_tail(&qh->unlink_node, &ehci->intr_unlink_wait);
 694
 695	if (ehci->rh_state < EHCI_RH_RUNNING)
 696		ehci_handle_start_intr_unlinks(ehci);
 697	else if (ehci->intr_unlink_wait.next == &qh->unlink_node) {
 698		ehci_enable_event(ehci, EHCI_HRTIMER_START_UNLINK_INTR, true);
 699		++ehci->intr_unlink_wait_cycle;
 700	}
 701}
 702
 703static void end_unlink_intr(struct ehci_hcd *ehci, struct ehci_qh *qh)
 704{
 705	struct ehci_qh_hw	*hw = qh->hw;
 706	int			rc;
 707
 708	qh->qh_state = QH_STATE_IDLE;
 709	hw->hw_next = EHCI_LIST_END(ehci);
 710
 711	if (!list_empty(&qh->qtd_list))
 712		qh_completions(ehci, qh);
 713
 714	/* reschedule QH iff another request is queued */
 715	if (!list_empty(&qh->qtd_list) && ehci->rh_state == EHCI_RH_RUNNING) {
 716		rc = qh_schedule(ehci, qh);
 717		if (rc == 0) {
 718			qh_refresh(ehci, qh);
 719			qh_link_periodic(ehci, qh);
 720		}
 721
 722		/* An error here likely indicates handshake failure
 723		 * or no space left in the schedule.  Neither fault
 724		 * should happen often ...
 725		 *
 726		 * FIXME kill the now-dysfunctional queued urbs
 727		 */
 728		else {
 729			ehci_err(ehci, "can't reschedule qh %p, err %d\n",
 730					qh, rc);
 731		}
 732	}
 733
 734	/* maybe turn off periodic schedule */
 735	--ehci->intr_count;
 736	disable_periodic(ehci);
 737}
 738
 739/*-------------------------------------------------------------------------*/
 740
 741static int check_period(
 742	struct ehci_hcd *ehci,
 743	unsigned	frame,
 744	unsigned	uframe,
 745	unsigned	uperiod,
 746	unsigned	usecs
 747) {
 748	/* complete split running into next frame?
 749	 * given FSTN support, we could sometimes check...
 750	 */
 751	if (uframe >= 8)
 752		return 0;
 753
 754	/* convert "usecs we need" to "max already claimed" */
 755	usecs = ehci->uframe_periodic_max - usecs;
 756
 757	for (uframe += frame << 3; uframe < EHCI_BANDWIDTH_SIZE;
 758			uframe += uperiod) {
 759		if (ehci->bandwidth[uframe] > usecs)
 760			return 0;
 761	}
 762
 763	/* success! */
 764	return 1;
 765}
 766
 767static int check_intr_schedule(
 768	struct ehci_hcd		*ehci,
 769	unsigned		frame,
 770	unsigned		uframe,
 771	struct ehci_qh		*qh,
 772	unsigned		*c_maskp,
 773	struct ehci_tt		*tt
 774)
 775{
 776	int		retval = -ENOSPC;
 777	u8		mask = 0;
 778
 779	if (qh->ps.c_usecs && uframe >= 6)	/* FSTN territory? */
 780		goto done;
 781
 782	if (!check_period(ehci, frame, uframe, qh->ps.bw_uperiod, qh->ps.usecs))
 783		goto done;
 784	if (!qh->ps.c_usecs) {
 785		retval = 0;
 786		*c_maskp = 0;
 787		goto done;
 788	}
 789
 790#ifdef CONFIG_USB_EHCI_TT_NEWSCHED
 791	if (tt_available(ehci, &qh->ps, tt, frame, uframe)) {
 792		unsigned i;
 793
 794		/* TODO : this may need FSTN for SSPLIT in uframe 5. */
 795		for (i = uframe+2; i < 8 && i <= uframe+4; i++)
 796			if (!check_period(ehci, frame, i,
 797					qh->ps.bw_uperiod, qh->ps.c_usecs))
 798				goto done;
 799			else
 800				mask |= 1 << i;
 801
 802		retval = 0;
 803
 804		*c_maskp = mask;
 805	}
 806#else
 807	/* Make sure this tt's buffer is also available for CSPLITs.
 808	 * We pessimize a bit; probably the typical full speed case
 809	 * doesn't need the second CSPLIT.
 810	 *
 811	 * NOTE:  both SPLIT and CSPLIT could be checked in just
 812	 * one smart pass...
 813	 */
 814	mask = 0x03 << (uframe + qh->gap_uf);
 815	*c_maskp = mask;
 816
 817	mask |= 1 << uframe;
 818	if (tt_no_collision(ehci, qh->ps.bw_period, qh->ps.udev, frame, mask)) {
 819		if (!check_period(ehci, frame, uframe + qh->gap_uf + 1,
 820				qh->ps.bw_uperiod, qh->ps.c_usecs))
 821			goto done;
 822		if (!check_period(ehci, frame, uframe + qh->gap_uf,
 823				qh->ps.bw_uperiod, qh->ps.c_usecs))
 824			goto done;
 825		retval = 0;
 826	}
 827#endif
 828done:
 829	return retval;
 830}
 831
 832/* "first fit" scheduling policy used the first time through,
 833 * or when the previous schedule slot can't be re-used.
 834 */
 835static int qh_schedule(struct ehci_hcd *ehci, struct ehci_qh *qh)
 836{
 837	int		status = 0;
 838	unsigned	uframe;
 839	unsigned	c_mask;
 840	struct ehci_qh_hw	*hw = qh->hw;
 841	struct ehci_tt		*tt;
 842
 843	hw->hw_next = EHCI_LIST_END(ehci);
 844
 845	/* reuse the previous schedule slots, if we can */
 846	if (qh->ps.phase != NO_FRAME) {
 847		ehci_dbg(ehci, "reused qh %p schedule\n", qh);
 848		return 0;
 849	}
 850
 851	uframe = 0;
 852	c_mask = 0;
 853	tt = find_tt(qh->ps.udev);
 854	if (IS_ERR(tt)) {
 855		status = PTR_ERR(tt);
 856		goto done;
 857	}
 858	compute_tt_budget(ehci->tt_budget, tt);
 859
 860	/* else scan the schedule to find a group of slots such that all
 861	 * uframes have enough periodic bandwidth available.
 862	 */
 863	/* "normal" case, uframing flexible except with splits */
 864	if (qh->ps.bw_period) {
 865		int		i;
 866		unsigned	frame;
 867
 868		for (i = qh->ps.bw_period; i > 0; --i) {
 869			frame = ++ehci->random_frame & (qh->ps.bw_period - 1);
 870			for (uframe = 0; uframe < 8; uframe++) {
 871				status = check_intr_schedule(ehci,
 872						frame, uframe, qh, &c_mask, tt);
 873				if (status == 0)
 874					goto got_it;
 875			}
 876		}
 877
 878	/* qh->ps.bw_period == 0 means every uframe */
 879	} else {
 880		status = check_intr_schedule(ehci, 0, 0, qh, &c_mask, tt);
 881	}
 882	if (status)
 883		goto done;
 884
 885 got_it:
 886	qh->ps.phase = (qh->ps.period ? ehci->random_frame &
 887			(qh->ps.period - 1) : 0);
 888	qh->ps.bw_phase = qh->ps.phase & (qh->ps.bw_period - 1);
 889	qh->ps.phase_uf = uframe;
 890	qh->ps.cs_mask = qh->ps.period ?
 891			(c_mask << 8) | (1 << uframe) :
 892			QH_SMASK;
 893
 894	/* reset S-frame and (maybe) C-frame masks */
 895	hw->hw_info2 &= cpu_to_hc32(ehci, ~(QH_CMASK | QH_SMASK));
 896	hw->hw_info2 |= cpu_to_hc32(ehci, qh->ps.cs_mask);
 897	reserve_release_intr_bandwidth(ehci, qh, 1);
 898
 899done:
 900	return status;
 901}
 902
 903static int intr_submit(
 904	struct ehci_hcd		*ehci,
 905	struct urb		*urb,
 906	struct list_head	*qtd_list,
 907	gfp_t			mem_flags
 908) {
 909	unsigned		epnum;
 910	unsigned long		flags;
 911	struct ehci_qh		*qh;
 912	int			status;
 913	struct list_head	empty;
 914
 915	/* get endpoint and transfer/schedule data */
 916	epnum = urb->ep->desc.bEndpointAddress;
 917
 918	spin_lock_irqsave(&ehci->lock, flags);
 919
 920	if (unlikely(!HCD_HW_ACCESSIBLE(ehci_to_hcd(ehci)))) {
 921		status = -ESHUTDOWN;
 922		goto done_not_linked;
 923	}
 924	status = usb_hcd_link_urb_to_ep(ehci_to_hcd(ehci), urb);
 925	if (unlikely(status))
 926		goto done_not_linked;
 927
 928	/* get qh and force any scheduling errors */
 929	INIT_LIST_HEAD(&empty);
 930	qh = qh_append_tds(ehci, urb, &empty, epnum, &urb->ep->hcpriv);
 931	if (qh == NULL) {
 932		status = -ENOMEM;
 933		goto done;
 934	}
 935	if (qh->qh_state == QH_STATE_IDLE) {
 936		status = qh_schedule(ehci, qh);
 937		if (status)
 938			goto done;
 939	}
 940
 941	/* then queue the urb's tds to the qh */
 942	qh = qh_append_tds(ehci, urb, qtd_list, epnum, &urb->ep->hcpriv);
 943	BUG_ON(qh == NULL);
 944
 945	/* stuff into the periodic schedule */
 946	if (qh->qh_state == QH_STATE_IDLE) {
 947		qh_refresh(ehci, qh);
 948		qh_link_periodic(ehci, qh);
 949	} else {
 950		/* cancel unlink wait for the qh */
 951		cancel_unlink_wait_intr(ehci, qh);
 952	}
 953
 954	/* ... update usbfs periodic stats */
 955	ehci_to_hcd(ehci)->self.bandwidth_int_reqs++;
 956
 957done:
 958	if (unlikely(status))
 959		usb_hcd_unlink_urb_from_ep(ehci_to_hcd(ehci), urb);
 960done_not_linked:
 961	spin_unlock_irqrestore(&ehci->lock, flags);
 962	if (status)
 963		qtd_list_free(ehci, urb, qtd_list);
 964
 965	return status;
 966}
 967
 968static void scan_intr(struct ehci_hcd *ehci)
 969{
 970	struct ehci_qh		*qh;
 971
 972	list_for_each_entry_safe(qh, ehci->qh_scan_next, &ehci->intr_qh_list,
 973			intr_node) {
 974
 975		/* clean any finished work for this qh */
 976		if (!list_empty(&qh->qtd_list)) {
 977			int temp;
 978
 979			/*
 980			 * Unlinks could happen here; completion reporting
 981			 * drops the lock.  That's why ehci->qh_scan_next
 982			 * always holds the next qh to scan; if the next qh
 983			 * gets unlinked then ehci->qh_scan_next is adjusted
 984			 * in qh_unlink_periodic().
 985			 */
 986			temp = qh_completions(ehci, qh);
 987			if (unlikely(temp))
 988				start_unlink_intr(ehci, qh);
 989			else if (unlikely(list_empty(&qh->qtd_list) &&
 990					qh->qh_state == QH_STATE_LINKED))
 991				start_unlink_intr_wait(ehci, qh);
 992		}
 993	}
 994}
 995
 996/*-------------------------------------------------------------------------*/
 997
 998/* ehci_iso_stream ops work with both ITD and SITD */
 999
1000static struct ehci_iso_stream *
1001iso_stream_alloc(gfp_t mem_flags)
1002{
1003	struct ehci_iso_stream *stream;
1004
1005	stream = kzalloc(sizeof(*stream), mem_flags);
1006	if (likely(stream != NULL)) {
1007		INIT_LIST_HEAD(&stream->td_list);
1008		INIT_LIST_HEAD(&stream->free_list);
1009		stream->next_uframe = NO_FRAME;
1010		stream->ps.phase = NO_FRAME;
1011	}
1012	return stream;
1013}
1014
1015static void
1016iso_stream_init(
1017	struct ehci_hcd		*ehci,
1018	struct ehci_iso_stream	*stream,
1019	struct urb		*urb
1020)
1021{
1022	static const u8 smask_out[] = { 0x01, 0x03, 0x07, 0x0f, 0x1f, 0x3f };
1023
1024	struct usb_device	*dev = urb->dev;
1025	u32			buf1;
1026	unsigned		epnum, maxp;
1027	int			is_input;
1028	unsigned		tmp;
1029
1030	/*
1031	 * this might be a "high bandwidth" highspeed endpoint,
1032	 * as encoded in the ep descriptor's wMaxPacket field
1033	 */
1034	epnum = usb_pipeendpoint(urb->pipe);
1035	is_input = usb_pipein(urb->pipe) ? USB_DIR_IN : 0;
1036	maxp = usb_endpoint_maxp(&urb->ep->desc);
1037	buf1 = is_input ? 1 << 11 : 0;
1038
1039	/* knows about ITD vs SITD */
1040	if (dev->speed == USB_SPEED_HIGH) {
1041		unsigned multi = usb_endpoint_maxp_mult(&urb->ep->desc);
1042
1043		stream->highspeed = 1;
1044
1045		buf1 |= maxp;
1046		maxp *= multi;
1047
1048		stream->buf0 = cpu_to_hc32(ehci, (epnum << 8) | dev->devnum);
1049		stream->buf1 = cpu_to_hc32(ehci, buf1);
1050		stream->buf2 = cpu_to_hc32(ehci, multi);
1051
1052		/* usbfs wants to report the average usecs per frame tied up
1053		 * when transfers on this endpoint are scheduled ...
1054		 */
1055		stream->ps.usecs = HS_USECS_ISO(maxp);
1056
1057		/* period for bandwidth allocation */
1058		tmp = min_t(unsigned, EHCI_BANDWIDTH_SIZE,
1059				1 << (urb->ep->desc.bInterval - 1));
1060
1061		/* Allow urb->interval to override */
1062		stream->ps.bw_uperiod = min_t(unsigned, tmp, urb->interval);
1063
1064		stream->uperiod = urb->interval;
1065		stream->ps.period = urb->interval >> 3;
1066		stream->bandwidth = stream->ps.usecs * 8 /
1067				stream->ps.bw_uperiod;
1068
1069	} else {
1070		u32		addr;
1071		int		think_time;
1072		int		hs_transfers;
1073
1074		addr = dev->ttport << 24;
1075		if (!ehci_is_TDI(ehci)
1076				|| (dev->tt->hub !=
1077					ehci_to_hcd(ehci)->self.root_hub))
1078			addr |= dev->tt->hub->devnum << 16;
1079		addr |= epnum << 8;
1080		addr |= dev->devnum;
1081		stream->ps.usecs = HS_USECS_ISO(maxp);
1082		think_time = dev->tt->think_time;
1083		stream->ps.tt_usecs = NS_TO_US(think_time + usb_calc_bus_time(
1084				dev->speed, is_input, 1, maxp));
1085		hs_transfers = max(1u, (maxp + 187) / 188);
1086		if (is_input) {
1087			u32	tmp;
1088
1089			addr |= 1 << 31;
1090			stream->ps.c_usecs = stream->ps.usecs;
1091			stream->ps.usecs = HS_USECS_ISO(1);
1092			stream->ps.cs_mask = 1;
1093
1094			/* c-mask as specified in USB 2.0 11.18.4 3.c */
1095			tmp = (1 << (hs_transfers + 2)) - 1;
1096			stream->ps.cs_mask |= tmp << (8 + 2);
1097		} else
1098			stream->ps.cs_mask = smask_out[hs_transfers - 1];
1099
1100		/* period for bandwidth allocation */
1101		tmp = min_t(unsigned, EHCI_BANDWIDTH_FRAMES,
1102				1 << (urb->ep->desc.bInterval - 1));
1103
1104		/* Allow urb->interval to override */
1105		stream->ps.bw_period = min_t(unsigned, tmp, urb->interval);
1106		stream->ps.bw_uperiod = stream->ps.bw_period << 3;
1107
1108		stream->ps.period = urb->interval;
1109		stream->uperiod = urb->interval << 3;
1110		stream->bandwidth = (stream->ps.usecs + stream->ps.c_usecs) /
1111				stream->ps.bw_period;
1112
1113		/* stream->splits gets created from cs_mask later */
1114		stream->address = cpu_to_hc32(ehci, addr);
1115	}
1116
1117	stream->ps.udev = dev;
1118	stream->ps.ep = urb->ep;
1119
1120	stream->bEndpointAddress = is_input | epnum;
1121	stream->maxp = maxp;
1122}
1123
1124static struct ehci_iso_stream *
1125iso_stream_find(struct ehci_hcd *ehci, struct urb *urb)
1126{
1127	unsigned		epnum;
1128	struct ehci_iso_stream	*stream;
1129	struct usb_host_endpoint *ep;
1130	unsigned long		flags;
1131
1132	epnum = usb_pipeendpoint (urb->pipe);
1133	if (usb_pipein(urb->pipe))
1134		ep = urb->dev->ep_in[epnum];
1135	else
1136		ep = urb->dev->ep_out[epnum];
1137
1138	spin_lock_irqsave(&ehci->lock, flags);
1139	stream = ep->hcpriv;
1140
1141	if (unlikely(stream == NULL)) {
1142		stream = iso_stream_alloc(GFP_ATOMIC);
1143		if (likely(stream != NULL)) {
1144			ep->hcpriv = stream;
1145			iso_stream_init(ehci, stream, urb);
1146		}
1147
1148	/* if dev->ep [epnum] is a QH, hw is set */
1149	} else if (unlikely(stream->hw != NULL)) {
1150		ehci_dbg(ehci, "dev %s ep%d%s, not iso??\n",
1151			urb->dev->devpath, epnum,
1152			usb_pipein(urb->pipe) ? "in" : "out");
1153		stream = NULL;
1154	}
1155
1156	spin_unlock_irqrestore(&ehci->lock, flags);
1157	return stream;
1158}
1159
1160/*-------------------------------------------------------------------------*/
1161
1162/* ehci_iso_sched ops can be ITD-only or SITD-only */
1163
1164static struct ehci_iso_sched *
1165iso_sched_alloc(unsigned packets, gfp_t mem_flags)
1166{
1167	struct ehci_iso_sched	*iso_sched;
 
1168
1169	iso_sched = kzalloc(struct_size(iso_sched, packet, packets), mem_flags);
 
1170	if (likely(iso_sched != NULL))
1171		INIT_LIST_HEAD(&iso_sched->td_list);
1172
1173	return iso_sched;
1174}
1175
1176static inline void
1177itd_sched_init(
1178	struct ehci_hcd		*ehci,
1179	struct ehci_iso_sched	*iso_sched,
1180	struct ehci_iso_stream	*stream,
1181	struct urb		*urb
1182)
1183{
1184	unsigned	i;
1185	dma_addr_t	dma = urb->transfer_dma;
1186
1187	/* how many uframes are needed for these transfers */
1188	iso_sched->span = urb->number_of_packets * stream->uperiod;
1189
1190	/* figure out per-uframe itd fields that we'll need later
1191	 * when we fit new itds into the schedule.
1192	 */
1193	for (i = 0; i < urb->number_of_packets; i++) {
1194		struct ehci_iso_packet	*uframe = &iso_sched->packet[i];
1195		unsigned		length;
1196		dma_addr_t		buf;
1197		u32			trans;
1198
1199		length = urb->iso_frame_desc[i].length;
1200		buf = dma + urb->iso_frame_desc[i].offset;
1201
1202		trans = EHCI_ISOC_ACTIVE;
1203		trans |= buf & 0x0fff;
1204		if (unlikely(((i + 1) == urb->number_of_packets))
1205				&& !(urb->transfer_flags & URB_NO_INTERRUPT))
1206			trans |= EHCI_ITD_IOC;
1207		trans |= length << 16;
1208		uframe->transaction = cpu_to_hc32(ehci, trans);
1209
1210		/* might need to cross a buffer page within a uframe */
1211		uframe->bufp = (buf & ~(u64)0x0fff);
1212		buf += length;
1213		if (unlikely((uframe->bufp != (buf & ~(u64)0x0fff))))
1214			uframe->cross = 1;
1215	}
1216}
1217
1218static void
1219iso_sched_free(
1220	struct ehci_iso_stream	*stream,
1221	struct ehci_iso_sched	*iso_sched
1222)
1223{
1224	if (!iso_sched)
1225		return;
1226	/* caller must hold ehci->lock! */
1227	list_splice(&iso_sched->td_list, &stream->free_list);
1228	kfree(iso_sched);
1229}
1230
1231static int
1232itd_urb_transaction(
1233	struct ehci_iso_stream	*stream,
1234	struct ehci_hcd		*ehci,
1235	struct urb		*urb,
1236	gfp_t			mem_flags
1237)
1238{
1239	struct ehci_itd		*itd;
1240	dma_addr_t		itd_dma;
1241	int			i;
1242	unsigned		num_itds;
1243	struct ehci_iso_sched	*sched;
1244	unsigned long		flags;
1245
1246	sched = iso_sched_alloc(urb->number_of_packets, mem_flags);
1247	if (unlikely(sched == NULL))
1248		return -ENOMEM;
1249
1250	itd_sched_init(ehci, sched, stream, urb);
1251
1252	if (urb->interval < 8)
1253		num_itds = 1 + (sched->span + 7) / 8;
1254	else
1255		num_itds = urb->number_of_packets;
1256
1257	/* allocate/init ITDs */
1258	spin_lock_irqsave(&ehci->lock, flags);
1259	for (i = 0; i < num_itds; i++) {
1260
1261		/*
1262		 * Use iTDs from the free list, but not iTDs that may
1263		 * still be in use by the hardware.
1264		 */
1265		if (likely(!list_empty(&stream->free_list))) {
1266			itd = list_first_entry(&stream->free_list,
1267					struct ehci_itd, itd_list);
1268			if (itd->frame == ehci->now_frame)
1269				goto alloc_itd;
1270			list_del(&itd->itd_list);
1271			itd_dma = itd->itd_dma;
1272		} else {
1273 alloc_itd:
1274			spin_unlock_irqrestore(&ehci->lock, flags);
1275			itd = dma_pool_alloc(ehci->itd_pool, mem_flags,
1276					&itd_dma);
1277			spin_lock_irqsave(&ehci->lock, flags);
1278			if (!itd) {
1279				iso_sched_free(stream, sched);
1280				spin_unlock_irqrestore(&ehci->lock, flags);
1281				return -ENOMEM;
1282			}
1283		}
1284
1285		memset(itd, 0, sizeof(*itd));
1286		itd->itd_dma = itd_dma;
1287		itd->frame = NO_FRAME;
1288		list_add(&itd->itd_list, &sched->td_list);
1289	}
1290	spin_unlock_irqrestore(&ehci->lock, flags);
1291
1292	/* temporarily store schedule info in hcpriv */
1293	urb->hcpriv = sched;
1294	urb->error_count = 0;
1295	return 0;
1296}
1297
1298/*-------------------------------------------------------------------------*/
1299
1300static void reserve_release_iso_bandwidth(struct ehci_hcd *ehci,
1301		struct ehci_iso_stream *stream, int sign)
1302{
1303	unsigned		uframe;
1304	unsigned		i, j;
1305	unsigned		s_mask, c_mask, m;
1306	int			usecs = stream->ps.usecs;
1307	int			c_usecs = stream->ps.c_usecs;
1308	int			tt_usecs = stream->ps.tt_usecs;
1309	struct ehci_tt		*tt;
1310
1311	if (stream->ps.phase == NO_FRAME)	/* Bandwidth wasn't reserved */
1312		return;
1313	uframe = stream->ps.bw_phase << 3;
1314
1315	bandwidth_dbg(ehci, sign, "iso", &stream->ps);
1316
1317	if (sign < 0) {		/* Release bandwidth */
1318		usecs = -usecs;
1319		c_usecs = -c_usecs;
1320		tt_usecs = -tt_usecs;
1321	}
1322
1323	if (!stream->splits) {		/* High speed */
1324		for (i = uframe + stream->ps.phase_uf; i < EHCI_BANDWIDTH_SIZE;
1325				i += stream->ps.bw_uperiod)
1326			ehci->bandwidth[i] += usecs;
1327
1328	} else {			/* Full speed */
1329		s_mask = stream->ps.cs_mask;
1330		c_mask = s_mask >> 8;
1331
1332		/* NOTE: adjustment needed for frame overflow */
1333		for (i = uframe; i < EHCI_BANDWIDTH_SIZE;
1334				i += stream->ps.bw_uperiod) {
1335			for ((j = stream->ps.phase_uf, m = 1 << j); j < 8;
1336					(++j, m <<= 1)) {
1337				if (s_mask & m)
1338					ehci->bandwidth[i+j] += usecs;
1339				else if (c_mask & m)
1340					ehci->bandwidth[i+j] += c_usecs;
1341			}
1342		}
1343
1344		/*
1345		 * find_tt() will not return any error here as we have
1346		 * already called find_tt() before calling this function
1347		 * and checked for any error return. The previous call
1348		 * would have created the data structure.
1349		 */
1350		tt = find_tt(stream->ps.udev);
1351		if (sign > 0)
1352			list_add_tail(&stream->ps.ps_list, &tt->ps_list);
1353		else
1354			list_del(&stream->ps.ps_list);
1355
1356		for (i = uframe >> 3; i < EHCI_BANDWIDTH_FRAMES;
1357				i += stream->ps.bw_period)
1358			tt->bandwidth[i] += tt_usecs;
1359	}
1360}
1361
1362static inline int
1363itd_slot_ok(
1364	struct ehci_hcd		*ehci,
1365	struct ehci_iso_stream	*stream,
1366	unsigned		uframe
1367)
1368{
1369	unsigned		usecs;
1370
1371	/* convert "usecs we need" to "max already claimed" */
1372	usecs = ehci->uframe_periodic_max - stream->ps.usecs;
1373
1374	for (uframe &= stream->ps.bw_uperiod - 1; uframe < EHCI_BANDWIDTH_SIZE;
1375			uframe += stream->ps.bw_uperiod) {
1376		if (ehci->bandwidth[uframe] > usecs)
1377			return 0;
1378	}
1379	return 1;
1380}
1381
1382static inline int
1383sitd_slot_ok(
1384	struct ehci_hcd		*ehci,
1385	struct ehci_iso_stream	*stream,
1386	unsigned		uframe,
1387	struct ehci_iso_sched	*sched,
1388	struct ehci_tt		*tt
1389)
1390{
1391	unsigned		mask, tmp;
1392	unsigned		frame, uf;
1393
1394	mask = stream->ps.cs_mask << (uframe & 7);
1395
1396	/* for OUT, don't wrap SSPLIT into H-microframe 7 */
1397	if (((stream->ps.cs_mask & 0xff) << (uframe & 7)) >= (1 << 7))
1398		return 0;
1399
1400	/* for IN, don't wrap CSPLIT into the next frame */
1401	if (mask & ~0xffff)
1402		return 0;
1403
1404	/* check bandwidth */
1405	uframe &= stream->ps.bw_uperiod - 1;
1406	frame = uframe >> 3;
1407
1408#ifdef CONFIG_USB_EHCI_TT_NEWSCHED
1409	/* The tt's fullspeed bus bandwidth must be available.
1410	 * tt_available scheduling guarantees 10+% for control/bulk.
1411	 */
1412	uf = uframe & 7;
1413	if (!tt_available(ehci, &stream->ps, tt, frame, uf))
1414		return 0;
1415#else
1416	/* tt must be idle for start(s), any gap, and csplit.
1417	 * assume scheduling slop leaves 10+% for control/bulk.
1418	 */
1419	if (!tt_no_collision(ehci, stream->ps.bw_period,
1420			stream->ps.udev, frame, mask))
1421		return 0;
1422#endif
1423
1424	do {
1425		unsigned	max_used;
1426		unsigned	i;
1427
1428		/* check starts (OUT uses more than one) */
1429		uf = uframe;
1430		max_used = ehci->uframe_periodic_max - stream->ps.usecs;
1431		for (tmp = stream->ps.cs_mask & 0xff; tmp; tmp >>= 1, uf++) {
1432			if (ehci->bandwidth[uf] > max_used)
1433				return 0;
1434		}
1435
1436		/* for IN, check CSPLIT */
1437		if (stream->ps.c_usecs) {
1438			max_used = ehci->uframe_periodic_max -
1439					stream->ps.c_usecs;
1440			uf = uframe & ~7;
1441			tmp = 1 << (2+8);
1442			for (i = (uframe & 7) + 2; i < 8; (++i, tmp <<= 1)) {
1443				if ((stream->ps.cs_mask & tmp) == 0)
1444					continue;
1445				if (ehci->bandwidth[uf+i] > max_used)
1446					return 0;
1447			}
1448		}
1449
1450		uframe += stream->ps.bw_uperiod;
1451	} while (uframe < EHCI_BANDWIDTH_SIZE);
1452
1453	stream->ps.cs_mask <<= uframe & 7;
1454	stream->splits = cpu_to_hc32(ehci, stream->ps.cs_mask);
1455	return 1;
1456}
1457
1458/*
1459 * This scheduler plans almost as far into the future as it has actual
1460 * periodic schedule slots.  (Affected by TUNE_FLS, which defaults to
1461 * "as small as possible" to be cache-friendlier.)  That limits the size
1462 * transfers you can stream reliably; avoid more than 64 msec per urb.
1463 * Also avoid queue depths of less than ehci's worst irq latency (affected
1464 * by the per-urb URB_NO_INTERRUPT hint, the log2_irq_thresh module parameter,
1465 * and other factors); or more than about 230 msec total (for portability,
1466 * given EHCI_TUNE_FLS and the slop).  Or, write a smarter scheduler!
1467 */
1468
1469static int
1470iso_stream_schedule(
1471	struct ehci_hcd		*ehci,
1472	struct urb		*urb,
1473	struct ehci_iso_stream	*stream
1474)
1475{
1476	u32			now, base, next, start, period, span, now2;
1477	u32			wrap = 0, skip = 0;
1478	int			status = 0;
1479	unsigned		mod = ehci->periodic_size << 3;
1480	struct ehci_iso_sched	*sched = urb->hcpriv;
1481	bool			empty = list_empty(&stream->td_list);
1482	bool			new_stream = false;
1483
1484	period = stream->uperiod;
1485	span = sched->span;
1486	if (!stream->highspeed)
1487		span <<= 3;
1488
1489	/* Start a new isochronous stream? */
1490	if (unlikely(empty && !hcd_periodic_completion_in_progress(
1491			ehci_to_hcd(ehci), urb->ep))) {
1492
1493		/* Schedule the endpoint */
1494		if (stream->ps.phase == NO_FRAME) {
1495			int		done = 0;
1496			struct ehci_tt	*tt = find_tt(stream->ps.udev);
1497
1498			if (IS_ERR(tt)) {
1499				status = PTR_ERR(tt);
1500				goto fail;
1501			}
1502			compute_tt_budget(ehci->tt_budget, tt);
1503
1504			start = ((-(++ehci->random_frame)) << 3) & (period - 1);
1505
1506			/* find a uframe slot with enough bandwidth.
1507			 * Early uframes are more precious because full-speed
1508			 * iso IN transfers can't use late uframes,
1509			 * and therefore they should be allocated last.
1510			 */
1511			next = start;
1512			start += period;
1513			do {
1514				start--;
1515				/* check schedule: enough space? */
1516				if (stream->highspeed) {
1517					if (itd_slot_ok(ehci, stream, start))
1518						done = 1;
1519				} else {
1520					if ((start % 8) >= 6)
1521						continue;
1522					if (sitd_slot_ok(ehci, stream, start,
1523							sched, tt))
1524						done = 1;
1525				}
1526			} while (start > next && !done);
1527
1528			/* no room in the schedule */
1529			if (!done) {
1530				ehci_dbg(ehci, "iso sched full %p", urb);
1531				status = -ENOSPC;
1532				goto fail;
1533			}
1534			stream->ps.phase = (start >> 3) &
1535					(stream->ps.period - 1);
1536			stream->ps.bw_phase = stream->ps.phase &
1537					(stream->ps.bw_period - 1);
1538			stream->ps.phase_uf = start & 7;
1539			reserve_release_iso_bandwidth(ehci, stream, 1);
1540		}
1541
1542		/* New stream is already scheduled; use the upcoming slot */
1543		else {
1544			start = (stream->ps.phase << 3) + stream->ps.phase_uf;
1545		}
1546
1547		stream->next_uframe = start;
1548		new_stream = true;
1549	}
1550
1551	now = ehci_read_frame_index(ehci) & (mod - 1);
1552
1553	/* Take the isochronous scheduling threshold into account */
1554	if (ehci->i_thresh)
1555		next = now + ehci->i_thresh;	/* uframe cache */
1556	else
1557		next = (now + 2 + 7) & ~0x07;	/* full frame cache */
1558
1559	/* If needed, initialize last_iso_frame so that this URB will be seen */
1560	if (ehci->isoc_count == 0)
1561		ehci->last_iso_frame = now >> 3;
1562
1563	/*
1564	 * Use ehci->last_iso_frame as the base.  There can't be any
1565	 * TDs scheduled for earlier than that.
1566	 */
1567	base = ehci->last_iso_frame << 3;
1568	next = (next - base) & (mod - 1);
1569	start = (stream->next_uframe - base) & (mod - 1);
1570
1571	if (unlikely(new_stream))
1572		goto do_ASAP;
1573
1574	/*
1575	 * Typical case: reuse current schedule, stream may still be active.
1576	 * Hopefully there are no gaps from the host falling behind
1577	 * (irq delays etc).  If there are, the behavior depends on
1578	 * whether URB_ISO_ASAP is set.
1579	 */
1580	now2 = (now - base) & (mod - 1);
1581
1582	/* Is the schedule about to wrap around? */
1583	if (unlikely(!empty && start < period)) {
1584		ehci_dbg(ehci, "request %p would overflow (%u-%u < %u mod %u)\n",
1585				urb, stream->next_uframe, base, period, mod);
1586		status = -EFBIG;
1587		goto fail;
1588	}
1589
1590	/* Is the next packet scheduled after the base time? */
1591	if (likely(!empty || start <= now2 + period)) {
1592
1593		/* URB_ISO_ASAP: make sure that start >= next */
1594		if (unlikely(start < next &&
1595				(urb->transfer_flags & URB_ISO_ASAP)))
1596			goto do_ASAP;
1597
1598		/* Otherwise use start, if it's not in the past */
1599		if (likely(start >= now2))
1600			goto use_start;
1601
1602	/* Otherwise we got an underrun while the queue was empty */
1603	} else {
1604		if (urb->transfer_flags & URB_ISO_ASAP)
1605			goto do_ASAP;
1606		wrap = mod;
1607		now2 += mod;
1608	}
1609
1610	/* How many uframes and packets do we need to skip? */
1611	skip = (now2 - start + period - 1) & -period;
1612	if (skip >= span) {		/* Entirely in the past? */
1613		ehci_dbg(ehci, "iso underrun %p (%u+%u < %u) [%u]\n",
1614				urb, start + base, span - period, now2 + base,
1615				base);
1616
1617		/* Try to keep the last TD intact for scanning later */
1618		skip = span - period;
1619
1620		/* Will it come before the current scan position? */
1621		if (empty) {
1622			skip = span;	/* Skip the entire URB */
1623			status = 1;	/* and give it back immediately */
1624			iso_sched_free(stream, sched);
1625			sched = NULL;
1626		}
1627	}
1628	urb->error_count = skip / period;
1629	if (sched)
1630		sched->first_packet = urb->error_count;
1631	goto use_start;
1632
1633 do_ASAP:
1634	/* Use the first slot after "next" */
1635	start = next + ((start - next) & (period - 1));
1636
1637 use_start:
1638	/* Tried to schedule too far into the future? */
1639	if (unlikely(start + span - period >= mod + wrap)) {
1640		ehci_dbg(ehci, "request %p would overflow (%u+%u >= %u)\n",
1641				urb, start, span - period, mod + wrap);
1642		status = -EFBIG;
1643		goto fail;
1644	}
1645
1646	start += base;
1647	stream->next_uframe = (start + skip) & (mod - 1);
1648
1649	/* report high speed start in uframes; full speed, in frames */
1650	urb->start_frame = start & (mod - 1);
1651	if (!stream->highspeed)
1652		urb->start_frame >>= 3;
1653	return status;
1654
1655 fail:
1656	iso_sched_free(stream, sched);
1657	urb->hcpriv = NULL;
1658	return status;
1659}
1660
1661/*-------------------------------------------------------------------------*/
1662
1663static inline void
1664itd_init(struct ehci_hcd *ehci, struct ehci_iso_stream *stream,
1665		struct ehci_itd *itd)
1666{
1667	int i;
1668
1669	/* it's been recently zeroed */
1670	itd->hw_next = EHCI_LIST_END(ehci);
1671	itd->hw_bufp[0] = stream->buf0;
1672	itd->hw_bufp[1] = stream->buf1;
1673	itd->hw_bufp[2] = stream->buf2;
1674
1675	for (i = 0; i < 8; i++)
1676		itd->index[i] = -1;
1677
1678	/* All other fields are filled when scheduling */
1679}
1680
1681static inline void
1682itd_patch(
1683	struct ehci_hcd		*ehci,
1684	struct ehci_itd		*itd,
1685	struct ehci_iso_sched	*iso_sched,
1686	unsigned		index,
1687	u16			uframe
1688)
1689{
1690	struct ehci_iso_packet	*uf = &iso_sched->packet[index];
1691	unsigned		pg = itd->pg;
1692
1693	/* BUG_ON(pg == 6 && uf->cross); */
1694
1695	uframe &= 0x07;
1696	itd->index[uframe] = index;
1697
1698	itd->hw_transaction[uframe] = uf->transaction;
1699	itd->hw_transaction[uframe] |= cpu_to_hc32(ehci, pg << 12);
1700	itd->hw_bufp[pg] |= cpu_to_hc32(ehci, uf->bufp & ~(u32)0);
1701	itd->hw_bufp_hi[pg] |= cpu_to_hc32(ehci, (u32)(uf->bufp >> 32));
1702
1703	/* iso_frame_desc[].offset must be strictly increasing */
1704	if (unlikely(uf->cross)) {
1705		u64	bufp = uf->bufp + 4096;
1706
1707		itd->pg = ++pg;
1708		itd->hw_bufp[pg] |= cpu_to_hc32(ehci, bufp & ~(u32)0);
1709		itd->hw_bufp_hi[pg] |= cpu_to_hc32(ehci, (u32)(bufp >> 32));
1710	}
1711}
1712
1713static inline void
1714itd_link(struct ehci_hcd *ehci, unsigned frame, struct ehci_itd *itd)
1715{
1716	union ehci_shadow	*prev = &ehci->pshadow[frame];
1717	__hc32			*hw_p = &ehci->periodic[frame];
1718	union ehci_shadow	here = *prev;
1719	__hc32			type = 0;
1720
1721	/* skip any iso nodes which might belong to previous microframes */
1722	while (here.ptr) {
1723		type = Q_NEXT_TYPE(ehci, *hw_p);
1724		if (type == cpu_to_hc32(ehci, Q_TYPE_QH))
1725			break;
1726		prev = periodic_next_shadow(ehci, prev, type);
1727		hw_p = shadow_next_periodic(ehci, &here, type);
1728		here = *prev;
1729	}
1730
1731	itd->itd_next = here;
1732	itd->hw_next = *hw_p;
1733	prev->itd = itd;
1734	itd->frame = frame;
1735	wmb();
1736	*hw_p = cpu_to_hc32(ehci, itd->itd_dma | Q_TYPE_ITD);
1737}
1738
1739/* fit urb's itds into the selected schedule slot; activate as needed */
1740static void itd_link_urb(
1741	struct ehci_hcd		*ehci,
1742	struct urb		*urb,
1743	unsigned		mod,
1744	struct ehci_iso_stream	*stream
1745)
1746{
1747	int			packet;
1748	unsigned		next_uframe, uframe, frame;
1749	struct ehci_iso_sched	*iso_sched = urb->hcpriv;
1750	struct ehci_itd		*itd;
1751
1752	next_uframe = stream->next_uframe & (mod - 1);
1753
1754	if (unlikely(list_empty(&stream->td_list)))
1755		ehci_to_hcd(ehci)->self.bandwidth_allocated
1756				+= stream->bandwidth;
1757
1758	if (ehci_to_hcd(ehci)->self.bandwidth_isoc_reqs == 0) {
1759		if (ehci->amd_pll_fix == 1)
1760			usb_amd_quirk_pll_disable();
1761	}
1762
1763	ehci_to_hcd(ehci)->self.bandwidth_isoc_reqs++;
1764
1765	/* fill iTDs uframe by uframe */
1766	for (packet = iso_sched->first_packet, itd = NULL;
1767			packet < urb->number_of_packets;) {
1768		if (itd == NULL) {
1769			/* ASSERT:  we have all necessary itds */
1770			/* BUG_ON(list_empty(&iso_sched->td_list)); */
1771
1772			/* ASSERT:  no itds for this endpoint in this uframe */
1773
1774			itd = list_entry(iso_sched->td_list.next,
1775					struct ehci_itd, itd_list);
1776			list_move_tail(&itd->itd_list, &stream->td_list);
1777			itd->stream = stream;
1778			itd->urb = urb;
1779			itd_init(ehci, stream, itd);
1780		}
1781
1782		uframe = next_uframe & 0x07;
1783		frame = next_uframe >> 3;
1784
1785		itd_patch(ehci, itd, iso_sched, packet, uframe);
1786
1787		next_uframe += stream->uperiod;
1788		next_uframe &= mod - 1;
1789		packet++;
1790
1791		/* link completed itds into the schedule */
1792		if (((next_uframe >> 3) != frame)
1793				|| packet == urb->number_of_packets) {
1794			itd_link(ehci, frame & (ehci->periodic_size - 1), itd);
1795			itd = NULL;
1796		}
1797	}
1798	stream->next_uframe = next_uframe;
1799
1800	/* don't need that schedule data any more */
1801	iso_sched_free(stream, iso_sched);
1802	urb->hcpriv = stream;
1803
1804	++ehci->isoc_count;
1805	enable_periodic(ehci);
1806}
1807
1808#define	ISO_ERRS (EHCI_ISOC_BUF_ERR | EHCI_ISOC_BABBLE | EHCI_ISOC_XACTERR)
1809
1810/* Process and recycle a completed ITD.  Return true iff its urb completed,
1811 * and hence its completion callback probably added things to the hardware
1812 * schedule.
1813 *
1814 * Note that we carefully avoid recycling this descriptor until after any
1815 * completion callback runs, so that it won't be reused quickly.  That is,
1816 * assuming (a) no more than two urbs per frame on this endpoint, and also
1817 * (b) only this endpoint's completions submit URBs.  It seems some silicon
1818 * corrupts things if you reuse completed descriptors very quickly...
1819 */
1820static bool itd_complete(struct ehci_hcd *ehci, struct ehci_itd *itd)
1821{
1822	struct urb				*urb = itd->urb;
1823	struct usb_iso_packet_descriptor	*desc;
1824	u32					t;
1825	unsigned				uframe;
1826	int					urb_index = -1;
1827	struct ehci_iso_stream			*stream = itd->stream;
1828	bool					retval = false;
1829
1830	/* for each uframe with a packet */
1831	for (uframe = 0; uframe < 8; uframe++) {
1832		if (likely(itd->index[uframe] == -1))
1833			continue;
1834		urb_index = itd->index[uframe];
1835		desc = &urb->iso_frame_desc[urb_index];
1836
1837		t = hc32_to_cpup(ehci, &itd->hw_transaction[uframe]);
1838		itd->hw_transaction[uframe] = 0;
1839
1840		/* report transfer status */
1841		if (unlikely(t & ISO_ERRS)) {
1842			urb->error_count++;
1843			if (t & EHCI_ISOC_BUF_ERR)
1844				desc->status = usb_pipein(urb->pipe)
1845					? -ENOSR  /* hc couldn't read */
1846					: -ECOMM; /* hc couldn't write */
1847			else if (t & EHCI_ISOC_BABBLE)
1848				desc->status = -EOVERFLOW;
1849			else /* (t & EHCI_ISOC_XACTERR) */
1850				desc->status = -EPROTO;
1851
1852			/* HC need not update length with this error */
1853			if (!(t & EHCI_ISOC_BABBLE)) {
1854				desc->actual_length = EHCI_ITD_LENGTH(t);
1855				urb->actual_length += desc->actual_length;
1856			}
1857		} else if (likely((t & EHCI_ISOC_ACTIVE) == 0)) {
1858			desc->status = 0;
1859			desc->actual_length = EHCI_ITD_LENGTH(t);
1860			urb->actual_length += desc->actual_length;
1861		} else {
1862			/* URB was too late */
1863			urb->error_count++;
1864		}
1865	}
1866
1867	/* handle completion now? */
1868	if (likely((urb_index + 1) != urb->number_of_packets))
1869		goto done;
1870
1871	/*
1872	 * ASSERT: it's really the last itd for this urb
1873	 * list_for_each_entry (itd, &stream->td_list, itd_list)
1874	 *	 BUG_ON(itd->urb == urb);
1875	 */
1876
1877	/* give urb back to the driver; completion often (re)submits */
1878	ehci_urb_done(ehci, urb, 0);
1879	retval = true;
1880	urb = NULL;
1881
1882	--ehci->isoc_count;
1883	disable_periodic(ehci);
1884
1885	ehci_to_hcd(ehci)->self.bandwidth_isoc_reqs--;
1886	if (ehci_to_hcd(ehci)->self.bandwidth_isoc_reqs == 0) {
1887		if (ehci->amd_pll_fix == 1)
1888			usb_amd_quirk_pll_enable();
1889	}
1890
1891	if (unlikely(list_is_singular(&stream->td_list)))
1892		ehci_to_hcd(ehci)->self.bandwidth_allocated
1893				-= stream->bandwidth;
1894
1895done:
1896	itd->urb = NULL;
1897
1898	/* Add to the end of the free list for later reuse */
1899	list_move_tail(&itd->itd_list, &stream->free_list);
1900
1901	/* Recycle the iTDs when the pipeline is empty (ep no longer in use) */
1902	if (list_empty(&stream->td_list)) {
1903		list_splice_tail_init(&stream->free_list,
1904				&ehci->cached_itd_list);
1905		start_free_itds(ehci);
1906	}
1907
1908	return retval;
1909}
1910
1911/*-------------------------------------------------------------------------*/
1912
1913static int itd_submit(struct ehci_hcd *ehci, struct urb *urb,
1914	gfp_t mem_flags)
1915{
1916	int			status = -EINVAL;
1917	unsigned long		flags;
1918	struct ehci_iso_stream	*stream;
1919
1920	/* Get iso_stream head */
1921	stream = iso_stream_find(ehci, urb);
1922	if (unlikely(stream == NULL)) {
1923		ehci_dbg(ehci, "can't get iso stream\n");
1924		return -ENOMEM;
1925	}
1926	if (unlikely(urb->interval != stream->uperiod)) {
1927		ehci_dbg(ehci, "can't change iso interval %d --> %d\n",
1928			stream->uperiod, urb->interval);
1929		goto done;
1930	}
1931
1932#ifdef EHCI_URB_TRACE
1933	ehci_dbg(ehci,
1934		"%s %s urb %p ep%d%s len %d, %d pkts %d uframes [%p]\n",
1935		__func__, urb->dev->devpath, urb,
1936		usb_pipeendpoint(urb->pipe),
1937		usb_pipein(urb->pipe) ? "in" : "out",
1938		urb->transfer_buffer_length,
1939		urb->number_of_packets, urb->interval,
1940		stream);
1941#endif
1942
1943	/* allocate ITDs w/o locking anything */
1944	status = itd_urb_transaction(stream, ehci, urb, mem_flags);
1945	if (unlikely(status < 0)) {
1946		ehci_dbg(ehci, "can't init itds\n");
1947		goto done;
1948	}
1949
1950	/* schedule ... need to lock */
1951	spin_lock_irqsave(&ehci->lock, flags);
1952	if (unlikely(!HCD_HW_ACCESSIBLE(ehci_to_hcd(ehci)))) {
1953		status = -ESHUTDOWN;
1954		goto done_not_linked;
1955	}
1956	status = usb_hcd_link_urb_to_ep(ehci_to_hcd(ehci), urb);
1957	if (unlikely(status))
1958		goto done_not_linked;
1959	status = iso_stream_schedule(ehci, urb, stream);
1960	if (likely(status == 0)) {
1961		itd_link_urb(ehci, urb, ehci->periodic_size << 3, stream);
1962	} else if (status > 0) {
1963		status = 0;
1964		ehci_urb_done(ehci, urb, 0);
1965	} else {
1966		usb_hcd_unlink_urb_from_ep(ehci_to_hcd(ehci), urb);
1967	}
1968 done_not_linked:
1969	spin_unlock_irqrestore(&ehci->lock, flags);
1970 done:
1971	return status;
1972}
1973
1974/*-------------------------------------------------------------------------*/
1975
1976/*
1977 * "Split ISO TDs" ... used for USB 1.1 devices going through the
1978 * TTs in USB 2.0 hubs.  These need microframe scheduling.
1979 */
1980
1981static inline void
1982sitd_sched_init(
1983	struct ehci_hcd		*ehci,
1984	struct ehci_iso_sched	*iso_sched,
1985	struct ehci_iso_stream	*stream,
1986	struct urb		*urb
1987)
1988{
1989	unsigned	i;
1990	dma_addr_t	dma = urb->transfer_dma;
1991
1992	/* how many frames are needed for these transfers */
1993	iso_sched->span = urb->number_of_packets * stream->ps.period;
1994
1995	/* figure out per-frame sitd fields that we'll need later
1996	 * when we fit new sitds into the schedule.
1997	 */
1998	for (i = 0; i < urb->number_of_packets; i++) {
1999		struct ehci_iso_packet	*packet = &iso_sched->packet[i];
2000		unsigned		length;
2001		dma_addr_t		buf;
2002		u32			trans;
2003
2004		length = urb->iso_frame_desc[i].length & 0x03ff;
2005		buf = dma + urb->iso_frame_desc[i].offset;
2006
2007		trans = SITD_STS_ACTIVE;
2008		if (((i + 1) == urb->number_of_packets)
2009				&& !(urb->transfer_flags & URB_NO_INTERRUPT))
2010			trans |= SITD_IOC;
2011		trans |= length << 16;
2012		packet->transaction = cpu_to_hc32(ehci, trans);
2013
2014		/* might need to cross a buffer page within a td */
2015		packet->bufp = buf;
2016		packet->buf1 = (buf + length) & ~0x0fff;
2017		if (packet->buf1 != (buf & ~(u64)0x0fff))
2018			packet->cross = 1;
2019
2020		/* OUT uses multiple start-splits */
2021		if (stream->bEndpointAddress & USB_DIR_IN)
2022			continue;
2023		length = (length + 187) / 188;
2024		if (length > 1) /* BEGIN vs ALL */
2025			length |= 1 << 3;
2026		packet->buf1 |= length;
2027	}
2028}
2029
2030static int
2031sitd_urb_transaction(
2032	struct ehci_iso_stream	*stream,
2033	struct ehci_hcd		*ehci,
2034	struct urb		*urb,
2035	gfp_t			mem_flags
2036)
2037{
2038	struct ehci_sitd	*sitd;
2039	dma_addr_t		sitd_dma;
2040	int			i;
2041	struct ehci_iso_sched	*iso_sched;
2042	unsigned long		flags;
2043
2044	iso_sched = iso_sched_alloc(urb->number_of_packets, mem_flags);
2045	if (iso_sched == NULL)
2046		return -ENOMEM;
2047
2048	sitd_sched_init(ehci, iso_sched, stream, urb);
2049
2050	/* allocate/init sITDs */
2051	spin_lock_irqsave(&ehci->lock, flags);
2052	for (i = 0; i < urb->number_of_packets; i++) {
2053
2054		/* NOTE:  for now, we don't try to handle wraparound cases
2055		 * for IN (using sitd->hw_backpointer, like a FSTN), which
2056		 * means we never need two sitds for full speed packets.
2057		 */
2058
2059		/*
2060		 * Use siTDs from the free list, but not siTDs that may
2061		 * still be in use by the hardware.
2062		 */
2063		if (likely(!list_empty(&stream->free_list))) {
2064			sitd = list_first_entry(&stream->free_list,
2065					 struct ehci_sitd, sitd_list);
2066			if (sitd->frame == ehci->now_frame)
2067				goto alloc_sitd;
2068			list_del(&sitd->sitd_list);
2069			sitd_dma = sitd->sitd_dma;
2070		} else {
2071 alloc_sitd:
2072			spin_unlock_irqrestore(&ehci->lock, flags);
2073			sitd = dma_pool_alloc(ehci->sitd_pool, mem_flags,
2074					&sitd_dma);
2075			spin_lock_irqsave(&ehci->lock, flags);
2076			if (!sitd) {
2077				iso_sched_free(stream, iso_sched);
2078				spin_unlock_irqrestore(&ehci->lock, flags);
2079				return -ENOMEM;
2080			}
2081		}
2082
2083		memset(sitd, 0, sizeof(*sitd));
2084		sitd->sitd_dma = sitd_dma;
2085		sitd->frame = NO_FRAME;
2086		list_add(&sitd->sitd_list, &iso_sched->td_list);
2087	}
2088
2089	/* temporarily store schedule info in hcpriv */
2090	urb->hcpriv = iso_sched;
2091	urb->error_count = 0;
2092
2093	spin_unlock_irqrestore(&ehci->lock, flags);
2094	return 0;
2095}
2096
2097/*-------------------------------------------------------------------------*/
2098
2099static inline void
2100sitd_patch(
2101	struct ehci_hcd		*ehci,
2102	struct ehci_iso_stream	*stream,
2103	struct ehci_sitd	*sitd,
2104	struct ehci_iso_sched	*iso_sched,
2105	unsigned		index
2106)
2107{
2108	struct ehci_iso_packet	*uf = &iso_sched->packet[index];
2109	u64			bufp;
2110
2111	sitd->hw_next = EHCI_LIST_END(ehci);
2112	sitd->hw_fullspeed_ep = stream->address;
2113	sitd->hw_uframe = stream->splits;
2114	sitd->hw_results = uf->transaction;
2115	sitd->hw_backpointer = EHCI_LIST_END(ehci);
2116
2117	bufp = uf->bufp;
2118	sitd->hw_buf[0] = cpu_to_hc32(ehci, bufp);
2119	sitd->hw_buf_hi[0] = cpu_to_hc32(ehci, bufp >> 32);
2120
2121	sitd->hw_buf[1] = cpu_to_hc32(ehci, uf->buf1);
2122	if (uf->cross)
2123		bufp += 4096;
2124	sitd->hw_buf_hi[1] = cpu_to_hc32(ehci, bufp >> 32);
2125	sitd->index = index;
2126}
2127
2128static inline void
2129sitd_link(struct ehci_hcd *ehci, unsigned frame, struct ehci_sitd *sitd)
2130{
2131	/* note: sitd ordering could matter (CSPLIT then SSPLIT) */
2132	sitd->sitd_next = ehci->pshadow[frame];
2133	sitd->hw_next = ehci->periodic[frame];
2134	ehci->pshadow[frame].sitd = sitd;
2135	sitd->frame = frame;
2136	wmb();
2137	ehci->periodic[frame] = cpu_to_hc32(ehci, sitd->sitd_dma | Q_TYPE_SITD);
2138}
2139
2140/* fit urb's sitds into the selected schedule slot; activate as needed */
2141static void sitd_link_urb(
2142	struct ehci_hcd		*ehci,
2143	struct urb		*urb,
2144	unsigned		mod,
2145	struct ehci_iso_stream	*stream
2146)
2147{
2148	int			packet;
2149	unsigned		next_uframe;
2150	struct ehci_iso_sched	*sched = urb->hcpriv;
2151	struct ehci_sitd	*sitd;
2152
2153	next_uframe = stream->next_uframe;
2154
2155	if (list_empty(&stream->td_list))
2156		/* usbfs ignores TT bandwidth */
2157		ehci_to_hcd(ehci)->self.bandwidth_allocated
2158				+= stream->bandwidth;
2159
2160	if (ehci_to_hcd(ehci)->self.bandwidth_isoc_reqs == 0) {
2161		if (ehci->amd_pll_fix == 1)
2162			usb_amd_quirk_pll_disable();
2163	}
2164
2165	ehci_to_hcd(ehci)->self.bandwidth_isoc_reqs++;
2166
2167	/* fill sITDs frame by frame */
2168	for (packet = sched->first_packet, sitd = NULL;
2169			packet < urb->number_of_packets;
2170			packet++) {
2171
2172		/* ASSERT:  we have all necessary sitds */
2173		BUG_ON(list_empty(&sched->td_list));
2174
2175		/* ASSERT:  no itds for this endpoint in this frame */
2176
2177		sitd = list_entry(sched->td_list.next,
2178				struct ehci_sitd, sitd_list);
2179		list_move_tail(&sitd->sitd_list, &stream->td_list);
2180		sitd->stream = stream;
2181		sitd->urb = urb;
2182
2183		sitd_patch(ehci, stream, sitd, sched, packet);
2184		sitd_link(ehci, (next_uframe >> 3) & (ehci->periodic_size - 1),
2185				sitd);
2186
2187		next_uframe += stream->uperiod;
2188	}
2189	stream->next_uframe = next_uframe & (mod - 1);
2190
2191	/* don't need that schedule data any more */
2192	iso_sched_free(stream, sched);
2193	urb->hcpriv = stream;
2194
2195	++ehci->isoc_count;
2196	enable_periodic(ehci);
2197}
2198
2199/*-------------------------------------------------------------------------*/
2200
2201#define	SITD_ERRS (SITD_STS_ERR | SITD_STS_DBE | SITD_STS_BABBLE \
2202				| SITD_STS_XACT | SITD_STS_MMF)
2203
2204/* Process and recycle a completed SITD.  Return true iff its urb completed,
2205 * and hence its completion callback probably added things to the hardware
2206 * schedule.
2207 *
2208 * Note that we carefully avoid recycling this descriptor until after any
2209 * completion callback runs, so that it won't be reused quickly.  That is,
2210 * assuming (a) no more than two urbs per frame on this endpoint, and also
2211 * (b) only this endpoint's completions submit URBs.  It seems some silicon
2212 * corrupts things if you reuse completed descriptors very quickly...
2213 */
2214static bool sitd_complete(struct ehci_hcd *ehci, struct ehci_sitd *sitd)
2215{
2216	struct urb				*urb = sitd->urb;
2217	struct usb_iso_packet_descriptor	*desc;
2218	u32					t;
2219	int					urb_index;
2220	struct ehci_iso_stream			*stream = sitd->stream;
2221	bool					retval = false;
2222
2223	urb_index = sitd->index;
2224	desc = &urb->iso_frame_desc[urb_index];
2225	t = hc32_to_cpup(ehci, &sitd->hw_results);
2226
2227	/* report transfer status */
2228	if (unlikely(t & SITD_ERRS)) {
2229		urb->error_count++;
2230		if (t & SITD_STS_DBE)
2231			desc->status = usb_pipein(urb->pipe)
2232				? -ENOSR  /* hc couldn't read */
2233				: -ECOMM; /* hc couldn't write */
2234		else if (t & SITD_STS_BABBLE)
2235			desc->status = -EOVERFLOW;
2236		else /* XACT, MMF, etc */
2237			desc->status = -EPROTO;
2238	} else if (unlikely(t & SITD_STS_ACTIVE)) {
2239		/* URB was too late */
2240		urb->error_count++;
2241	} else {
2242		desc->status = 0;
2243		desc->actual_length = desc->length - SITD_LENGTH(t);
2244		urb->actual_length += desc->actual_length;
2245	}
2246
2247	/* handle completion now? */
2248	if ((urb_index + 1) != urb->number_of_packets)
2249		goto done;
2250
2251	/*
2252	 * ASSERT: it's really the last sitd for this urb
2253	 * list_for_each_entry (sitd, &stream->td_list, sitd_list)
2254	 *	 BUG_ON(sitd->urb == urb);
2255	 */
2256
2257	/* give urb back to the driver; completion often (re)submits */
2258	ehci_urb_done(ehci, urb, 0);
2259	retval = true;
2260	urb = NULL;
2261
2262	--ehci->isoc_count;
2263	disable_periodic(ehci);
2264
2265	ehci_to_hcd(ehci)->self.bandwidth_isoc_reqs--;
2266	if (ehci_to_hcd(ehci)->self.bandwidth_isoc_reqs == 0) {
2267		if (ehci->amd_pll_fix == 1)
2268			usb_amd_quirk_pll_enable();
2269	}
2270
2271	if (list_is_singular(&stream->td_list))
2272		ehci_to_hcd(ehci)->self.bandwidth_allocated
2273				-= stream->bandwidth;
2274
2275done:
2276	sitd->urb = NULL;
2277
2278	/* Add to the end of the free list for later reuse */
2279	list_move_tail(&sitd->sitd_list, &stream->free_list);
2280
2281	/* Recycle the siTDs when the pipeline is empty (ep no longer in use) */
2282	if (list_empty(&stream->td_list)) {
2283		list_splice_tail_init(&stream->free_list,
2284				&ehci->cached_sitd_list);
2285		start_free_itds(ehci);
2286	}
2287
2288	return retval;
2289}
2290
2291
2292static int sitd_submit(struct ehci_hcd *ehci, struct urb *urb,
2293	gfp_t mem_flags)
2294{
2295	int			status = -EINVAL;
2296	unsigned long		flags;
2297	struct ehci_iso_stream	*stream;
2298
2299	/* Get iso_stream head */
2300	stream = iso_stream_find(ehci, urb);
2301	if (stream == NULL) {
2302		ehci_dbg(ehci, "can't get iso stream\n");
2303		return -ENOMEM;
2304	}
2305	if (urb->interval != stream->ps.period) {
2306		ehci_dbg(ehci, "can't change iso interval %d --> %d\n",
2307			stream->ps.period, urb->interval);
2308		goto done;
2309	}
2310
2311#ifdef EHCI_URB_TRACE
2312	ehci_dbg(ehci,
2313		"submit %p dev%s ep%d%s-iso len %d\n",
2314		urb, urb->dev->devpath,
2315		usb_pipeendpoint(urb->pipe),
2316		usb_pipein(urb->pipe) ? "in" : "out",
2317		urb->transfer_buffer_length);
2318#endif
2319
2320	/* allocate SITDs */
2321	status = sitd_urb_transaction(stream, ehci, urb, mem_flags);
2322	if (status < 0) {
2323		ehci_dbg(ehci, "can't init sitds\n");
2324		goto done;
2325	}
2326
2327	/* schedule ... need to lock */
2328	spin_lock_irqsave(&ehci->lock, flags);
2329	if (unlikely(!HCD_HW_ACCESSIBLE(ehci_to_hcd(ehci)))) {
2330		status = -ESHUTDOWN;
2331		goto done_not_linked;
2332	}
2333	status = usb_hcd_link_urb_to_ep(ehci_to_hcd(ehci), urb);
2334	if (unlikely(status))
2335		goto done_not_linked;
2336	status = iso_stream_schedule(ehci, urb, stream);
2337	if (likely(status == 0)) {
2338		sitd_link_urb(ehci, urb, ehci->periodic_size << 3, stream);
2339	} else if (status > 0) {
2340		status = 0;
2341		ehci_urb_done(ehci, urb, 0);
2342	} else {
2343		usb_hcd_unlink_urb_from_ep(ehci_to_hcd(ehci), urb);
2344	}
2345 done_not_linked:
2346	spin_unlock_irqrestore(&ehci->lock, flags);
2347 done:
2348	return status;
2349}
2350
2351/*-------------------------------------------------------------------------*/
2352
2353static void scan_isoc(struct ehci_hcd *ehci)
2354{
2355	unsigned		uf, now_frame, frame;
2356	unsigned		fmask = ehci->periodic_size - 1;
2357	bool			modified, live;
2358	union ehci_shadow	q, *q_p;
2359	__hc32			type, *hw_p;
2360
2361	/*
2362	 * When running, scan from last scan point up to "now"
2363	 * else clean up by scanning everything that's left.
2364	 * Touches as few pages as possible:  cache-friendly.
2365	 */
2366	if (ehci->rh_state >= EHCI_RH_RUNNING) {
2367		uf = ehci_read_frame_index(ehci);
2368		now_frame = (uf >> 3) & fmask;
2369		live = true;
2370	} else  {
2371		now_frame = (ehci->last_iso_frame - 1) & fmask;
2372		live = false;
2373	}
2374	ehci->now_frame = now_frame;
2375
2376	frame = ehci->last_iso_frame;
2377
2378restart:
2379	/* Scan each element in frame's queue for completions */
2380	q_p = &ehci->pshadow[frame];
2381	hw_p = &ehci->periodic[frame];
2382	q.ptr = q_p->ptr;
2383	type = Q_NEXT_TYPE(ehci, *hw_p);
2384	modified = false;
2385
2386	while (q.ptr != NULL) {
2387		switch (hc32_to_cpu(ehci, type)) {
2388		case Q_TYPE_ITD:
2389			/*
2390			 * If this ITD is still active, leave it for
2391			 * later processing ... check the next entry.
2392			 * No need to check for activity unless the
2393			 * frame is current.
2394			 */
2395			if (frame == now_frame && live) {
2396				rmb();
2397				for (uf = 0; uf < 8; uf++) {
2398					if (q.itd->hw_transaction[uf] &
2399							ITD_ACTIVE(ehci))
2400						break;
2401				}
2402				if (uf < 8) {
2403					q_p = &q.itd->itd_next;
2404					hw_p = &q.itd->hw_next;
2405					type = Q_NEXT_TYPE(ehci,
2406							q.itd->hw_next);
2407					q = *q_p;
2408					break;
2409				}
2410			}
2411
2412			/*
2413			 * Take finished ITDs out of the schedule
2414			 * and process them:  recycle, maybe report
2415			 * URB completion.  HC won't cache the
2416			 * pointer for much longer, if at all.
2417			 */
2418			*q_p = q.itd->itd_next;
2419			if (!ehci->use_dummy_qh ||
2420					q.itd->hw_next != EHCI_LIST_END(ehci))
2421				*hw_p = q.itd->hw_next;
2422			else
2423				*hw_p = cpu_to_hc32(ehci, ehci->dummy->qh_dma);
2424			type = Q_NEXT_TYPE(ehci, q.itd->hw_next);
2425			wmb();
2426			modified = itd_complete(ehci, q.itd);
2427			q = *q_p;
2428			break;
2429		case Q_TYPE_SITD:
2430			/*
2431			 * If this SITD is still active, leave it for
2432			 * later processing ... check the next entry.
2433			 * No need to check for activity unless the
2434			 * frame is current.
2435			 */
2436			if (((frame == now_frame) ||
2437					(((frame + 1) & fmask) == now_frame))
2438				&& live
2439				&& (q.sitd->hw_results & SITD_ACTIVE(ehci))) {
2440
2441				q_p = &q.sitd->sitd_next;
2442				hw_p = &q.sitd->hw_next;
2443				type = Q_NEXT_TYPE(ehci, q.sitd->hw_next);
2444				q = *q_p;
2445				break;
2446			}
2447
2448			/*
2449			 * Take finished SITDs out of the schedule
2450			 * and process them:  recycle, maybe report
2451			 * URB completion.
2452			 */
2453			*q_p = q.sitd->sitd_next;
2454			if (!ehci->use_dummy_qh ||
2455					q.sitd->hw_next != EHCI_LIST_END(ehci))
2456				*hw_p = q.sitd->hw_next;
2457			else
2458				*hw_p = cpu_to_hc32(ehci, ehci->dummy->qh_dma);
2459			type = Q_NEXT_TYPE(ehci, q.sitd->hw_next);
2460			wmb();
2461			modified = sitd_complete(ehci, q.sitd);
2462			q = *q_p;
2463			break;
2464		default:
2465			ehci_dbg(ehci, "corrupt type %d frame %d shadow %p\n",
2466					type, frame, q.ptr);
2467			/* BUG(); */
2468			fallthrough;
2469		case Q_TYPE_QH:
2470		case Q_TYPE_FSTN:
2471			/* End of the iTDs and siTDs */
2472			q.ptr = NULL;
2473			break;
2474		}
2475
2476		/* Assume completion callbacks modify the queue */
2477		if (unlikely(modified && ehci->isoc_count > 0))
2478			goto restart;
2479	}
2480
2481	/* Stop when we have reached the current frame */
2482	if (frame == now_frame)
2483		return;
2484
2485	/* The last frame may still have active siTDs */
2486	ehci->last_iso_frame = frame;
2487	frame = (frame + 1) & fmask;
2488
2489	goto restart;
2490}