Merge branch 'fix/hda' into for-linus
[pandora-kernel.git] / drivers / usb / host / ehci-q.c
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
45 static int
46 qtd_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
87 static inline void
88 qh_update (struct ehci_hcd *ehci, struct ehci_qh *qh, struct ehci_qtd *qtd)
89 {
90         /* writes to an active overlay are unsafe */
91         BUG_ON(qh->qh_state != QH_STATE_IDLE);
92
93         qh->hw_qtd_next = QTD_NEXT(ehci, qtd->qtd_dma);
94         qh->hw_alt_next = EHCI_LIST_END(ehci);
95
96         /* Except for control endpoints, we make hardware maintain data
97          * toggle (like OHCI) ... here (re)initialize the toggle in the QH,
98          * and set the pseudo-toggle in udev. Only usb_clear_halt() will
99          * ever clear it.
100          */
101         if (!(qh->hw_info1 & cpu_to_hc32(ehci, 1 << 14))) {
102                 unsigned        is_out, epnum;
103
104                 is_out = !(qtd->hw_token & cpu_to_hc32(ehci, 1 << 8));
105                 epnum = (hc32_to_cpup(ehci, &qh->hw_info1) >> 8) & 0x0f;
106                 if (unlikely (!usb_gettoggle (qh->dev, epnum, is_out))) {
107                         qh->hw_token &= ~cpu_to_hc32(ehci, QTD_TOGGLE);
108                         usb_settoggle (qh->dev, epnum, is_out, 1);
109                 }
110         }
111
112         /* HC must see latest qtd and qh data before we clear ACTIVE+HALT */
113         wmb ();
114         qh->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  */
121 static void
122 qh_refresh (struct ehci_hcd *ehci, struct ehci_qh *qh)
123 {
124         struct ehci_qtd *qtd;
125
126         if (list_empty (&qh->qtd_list))
127                 qtd = qh->dummy;
128         else {
129                 qtd = list_entry (qh->qtd_list.next,
130                                 struct ehci_qtd, qtd_list);
131                 /* first qtd may already be partially processed */
132                 if (cpu_to_hc32(ehci, qtd->qtd_dma) == qh->hw_current)
133                         qtd = NULL;
134         }
135
136         if (qtd)
137                 qh_update (ehci, qh, qtd);
138 }
139
140 /*-------------------------------------------------------------------------*/
141
142 static void qh_link_async(struct ehci_hcd *ehci, struct ehci_qh *qh);
143
144 static void ehci_clear_tt_buffer_complete(struct usb_hcd *hcd,
145                 struct usb_host_endpoint *ep)
146 {
147         struct ehci_hcd         *ehci = hcd_to_ehci(hcd);
148         struct ehci_qh          *qh = ep->hcpriv;
149         unsigned long           flags;
150
151         spin_lock_irqsave(&ehci->lock, flags);
152         qh->clearing_tt = 0;
153         if (qh->qh_state == QH_STATE_IDLE && !list_empty(&qh->qtd_list)
154                         && HC_IS_RUNNING(hcd->state))
155                 qh_link_async(ehci, qh);
156         spin_unlock_irqrestore(&ehci->lock, flags);
157 }
158
159 static void ehci_clear_tt_buffer(struct ehci_hcd *ehci, struct ehci_qh *qh,
160                 struct urb *urb, u32 token)
161 {
162
163         /* If an async split transaction gets an error or is unlinked,
164          * the TT buffer may be left in an indeterminate state.  We
165          * have to clear the TT buffer.
166          *
167          * Note: this routine is never called for Isochronous transfers.
168          */
169         if (urb->dev->tt && !usb_pipeint(urb->pipe) && !qh->clearing_tt) {
170 #ifdef DEBUG
171                 struct usb_device *tt = urb->dev->tt->hub;
172                 dev_dbg(&tt->dev,
173                         "clear tt buffer port %d, a%d ep%d t%08x\n",
174                         urb->dev->ttport, urb->dev->devnum,
175                         usb_pipeendpoint(urb->pipe), token);
176 #endif /* DEBUG */
177                 if (!ehci_is_TDI(ehci)
178                                 || urb->dev->tt->hub !=
179                                    ehci_to_hcd(ehci)->self.root_hub) {
180                         if (usb_hub_clear_tt_buffer(urb) == 0)
181                                 qh->clearing_tt = 1;
182                 } else {
183
184                         /* REVISIT ARC-derived cores don't clear the root
185                          * hub TT buffer in this way...
186                          */
187                 }
188         }
189 }
190
191 static int qtd_copy_status (
192         struct ehci_hcd *ehci,
193         struct urb *urb,
194         size_t length,
195         u32 token
196 )
197 {
198         int     status = -EINPROGRESS;
199
200         /* count IN/OUT bytes, not SETUP (even short packets) */
201         if (likely (QTD_PID (token) != 2))
202                 urb->actual_length += length - QTD_LENGTH (token);
203
204         /* don't modify error codes */
205         if (unlikely(urb->unlinked))
206                 return status;
207
208         /* force cleanup after short read; not always an error */
209         if (unlikely (IS_SHORT_READ (token)))
210                 status = -EREMOTEIO;
211
212         /* serious "can't proceed" faults reported by the hardware */
213         if (token & QTD_STS_HALT) {
214                 if (token & QTD_STS_BABBLE) {
215                         /* FIXME "must" disable babbling device's port too */
216                         status = -EOVERFLOW;
217                 /* CERR nonzero + halt --> stall */
218                 } else if (QTD_CERR(token)) {
219                         status = -EPIPE;
220
221                 /* In theory, more than one of the following bits can be set
222                  * since they are sticky and the transaction is retried.
223                  * Which to test first is rather arbitrary.
224                  */
225                 } else if (token & QTD_STS_MMF) {
226                         /* fs/ls interrupt xfer missed the complete-split */
227                         status = -EPROTO;
228                 } else if (token & QTD_STS_DBE) {
229                         status = (QTD_PID (token) == 1) /* IN ? */
230                                 ? -ENOSR  /* hc couldn't read data */
231                                 : -ECOMM; /* hc couldn't write data */
232                 } else if (token & QTD_STS_XACT) {
233                         /* timeout, bad CRC, wrong PID, etc */
234                         ehci_dbg(ehci, "devpath %s ep%d%s 3strikes\n",
235                                 urb->dev->devpath,
236                                 usb_pipeendpoint(urb->pipe),
237                                 usb_pipein(urb->pipe) ? "in" : "out");
238                         status = -EPROTO;
239                 } else {        /* unknown */
240                         status = -EPROTO;
241                 }
242
243                 ehci_vdbg (ehci,
244                         "dev%d ep%d%s qtd token %08x --> status %d\n",
245                         usb_pipedevice (urb->pipe),
246                         usb_pipeendpoint (urb->pipe),
247                         usb_pipein (urb->pipe) ? "in" : "out",
248                         token, status);
249         }
250
251         return status;
252 }
253
254 static void
255 ehci_urb_done(struct ehci_hcd *ehci, struct urb *urb, int status)
256 __releases(ehci->lock)
257 __acquires(ehci->lock)
258 {
259         if (likely (urb->hcpriv != NULL)) {
260                 struct ehci_qh  *qh = (struct ehci_qh *) urb->hcpriv;
261
262                 /* S-mask in a QH means it's an interrupt urb */
263                 if ((qh->hw_info2 & cpu_to_hc32(ehci, QH_SMASK)) != 0) {
264
265                         /* ... update hc-wide periodic stats (for usbfs) */
266                         ehci_to_hcd(ehci)->self.bandwidth_int_reqs--;
267                 }
268                 qh_put (qh);
269         }
270
271         if (unlikely(urb->unlinked)) {
272                 COUNT(ehci->stats.unlink);
273         } else {
274                 /* report non-error and short read status as zero */
275                 if (status == -EINPROGRESS || status == -EREMOTEIO)
276                         status = 0;
277                 COUNT(ehci->stats.complete);
278         }
279
280 #ifdef EHCI_URB_TRACE
281         ehci_dbg (ehci,
282                 "%s %s urb %p ep%d%s status %d len %d/%d\n",
283                 __func__, urb->dev->devpath, urb,
284                 usb_pipeendpoint (urb->pipe),
285                 usb_pipein (urb->pipe) ? "in" : "out",
286                 status,
287                 urb->actual_length, urb->transfer_buffer_length);
288 #endif
289
290         /* complete() can reenter this HCD */
291         usb_hcd_unlink_urb_from_ep(ehci_to_hcd(ehci), urb);
292         spin_unlock (&ehci->lock);
293         usb_hcd_giveback_urb(ehci_to_hcd(ehci), urb, status);
294         spin_lock (&ehci->lock);
295 }
296
297 static void start_unlink_async (struct ehci_hcd *ehci, struct ehci_qh *qh);
298 static void unlink_async (struct ehci_hcd *ehci, struct ehci_qh *qh);
299
300 static void intr_deschedule (struct ehci_hcd *ehci, struct ehci_qh *qh);
301 static int qh_schedule (struct ehci_hcd *ehci, struct ehci_qh *qh);
302
303 /*
304  * Process and free completed qtds for a qh, returning URBs to drivers.
305  * Chases up to qh->hw_current.  Returns number of completions called,
306  * indicating how much "real" work we did.
307  */
308 static unsigned
309 qh_completions (struct ehci_hcd *ehci, struct ehci_qh *qh)
310 {
311         struct ehci_qtd         *last = NULL, *end = qh->dummy;
312         struct list_head        *entry, *tmp;
313         int                     last_status = -EINPROGRESS;
314         int                     stopped;
315         unsigned                count = 0;
316         u8                      state;
317         __le32                  halt = HALT_BIT(ehci);
318
319         if (unlikely (list_empty (&qh->qtd_list)))
320                 return count;
321
322         /* completions (or tasks on other cpus) must never clobber HALT
323          * till we've gone through and cleaned everything up, even when
324          * they add urbs to this qh's queue or mark them for unlinking.
325          *
326          * NOTE:  unlinking expects to be done in queue order.
327          */
328         state = qh->qh_state;
329         qh->qh_state = QH_STATE_COMPLETING;
330         stopped = (state == QH_STATE_IDLE);
331
332         /* remove de-activated QTDs from front of queue.
333          * after faults (including short reads), cleanup this urb
334          * then let the queue advance.
335          * if queue is stopped, handles unlinks.
336          */
337         list_for_each_safe (entry, tmp, &qh->qtd_list) {
338                 struct ehci_qtd *qtd;
339                 struct urb      *urb;
340                 u32             token = 0;
341
342                 qtd = list_entry (entry, struct ehci_qtd, qtd_list);
343                 urb = qtd->urb;
344
345                 /* clean up any state from previous QTD ...*/
346                 if (last) {
347                         if (likely (last->urb != urb)) {
348                                 ehci_urb_done(ehci, last->urb, last_status);
349                                 count++;
350                                 last_status = -EINPROGRESS;
351                         }
352                         ehci_qtd_free (ehci, last);
353                         last = NULL;
354                 }
355
356                 /* ignore urbs submitted during completions we reported */
357                 if (qtd == end)
358                         break;
359
360                 /* hardware copies qtd out of qh overlay */
361                 rmb ();
362                 token = hc32_to_cpu(ehci, qtd->hw_token);
363
364                 /* always clean up qtds the hc de-activated */
365  retry_xacterr:
366                 if ((token & QTD_STS_ACTIVE) == 0) {
367
368                         /* on STALL, error, and short reads this urb must
369                          * complete and all its qtds must be recycled.
370                          */
371                         if ((token & QTD_STS_HALT) != 0) {
372
373                                 /* retry transaction errors until we
374                                  * reach the software xacterr limit
375                                  */
376                                 if ((token & QTD_STS_XACT) &&
377                                                 QTD_CERR(token) == 0 &&
378                                                 --qh->xacterrs > 0 &&
379                                                 !urb->unlinked) {
380                                         ehci_dbg(ehci,
381         "detected XactErr len %zu/%zu retry %d\n",
382         qtd->length - QTD_LENGTH(token), qtd->length,
383         QH_XACTERR_MAX - qh->xacterrs);
384
385                                         /* reset the token in the qtd and the
386                                          * qh overlay (which still contains
387                                          * the qtd) so that we pick up from
388                                          * where we left off
389                                          */
390                                         token &= ~QTD_STS_HALT;
391                                         token |= QTD_STS_ACTIVE |
392                                                         (EHCI_TUNE_CERR << 10);
393                                         qtd->hw_token = cpu_to_hc32(ehci,
394                                                         token);
395                                         wmb();
396                                         qh->hw_token = cpu_to_hc32(ehci, token);
397                                         goto retry_xacterr;
398                                 }
399                                 stopped = 1;
400
401                         /* magic dummy for some short reads; qh won't advance.
402                          * that silicon quirk can kick in with this dummy too.
403                          *
404                          * other short reads won't stop the queue, including
405                          * control transfers (status stage handles that) or
406                          * most other single-qtd reads ... the queue stops if
407                          * URB_SHORT_NOT_OK was set so the driver submitting
408                          * the urbs could clean it up.
409                          */
410                         } else if (IS_SHORT_READ (token)
411                                         && !(qtd->hw_alt_next
412                                                 & EHCI_LIST_END(ehci))) {
413                                 stopped = 1;
414                                 goto halt;
415                         }
416
417                 /* stop scanning when we reach qtds the hc is using */
418                 } else if (likely (!stopped
419                                 && HC_IS_RUNNING (ehci_to_hcd(ehci)->state))) {
420                         break;
421
422                 /* scan the whole queue for unlinks whenever it stops */
423                 } else {
424                         stopped = 1;
425
426                         /* cancel everything if we halt, suspend, etc */
427                         if (!HC_IS_RUNNING(ehci_to_hcd(ehci)->state))
428                                 last_status = -ESHUTDOWN;
429
430                         /* this qtd is active; skip it unless a previous qtd
431                          * for its urb faulted, or its urb was canceled.
432                          */
433                         else if (last_status == -EINPROGRESS && !urb->unlinked)
434                                 continue;
435
436                         /* qh unlinked; token in overlay may be most current */
437                         if (state == QH_STATE_IDLE
438                                         && cpu_to_hc32(ehci, qtd->qtd_dma)
439                                                 == qh->hw_current) {
440                                 token = hc32_to_cpu(ehci, qh->hw_token);
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                         /* force halt for unlinked or blocked qh, so we'll
450                          * patch the qh later and so that completions can't
451                          * activate it while we "know" it's stopped.
452                          */
453                         if ((halt & qh->hw_token) == 0) {
454 halt:
455                                 qh->hw_token |= halt;
456                                 wmb ();
457                         }
458                 }
459
460                 /* unless we already know the urb's status, collect qtd status
461                  * and update count of bytes transferred.  in common short read
462                  * cases with only one data qtd (including control transfers),
463                  * queue processing won't halt.  but with two or more qtds (for
464                  * example, with a 32 KB transfer), when the first qtd gets a
465                  * short read the second must be removed by hand.
466                  */
467                 if (last_status == -EINPROGRESS) {
468                         last_status = qtd_copy_status(ehci, urb,
469                                         qtd->length, token);
470                         if (last_status == -EREMOTEIO
471                                         && (qtd->hw_alt_next
472                                                 & EHCI_LIST_END(ehci)))
473                                 last_status = -EINPROGRESS;
474
475                         /* As part of low/full-speed endpoint-halt processing
476                          * we must clear the TT buffer (11.17.5).
477                          */
478                         if (unlikely(last_status != -EINPROGRESS &&
479                                         last_status != -EREMOTEIO))
480                                 ehci_clear_tt_buffer(ehci, qh, urb, token);
481                 }
482
483                 /* if we're removing something not at the queue head,
484                  * patch the hardware queue pointer.
485                  */
486                 if (stopped && qtd->qtd_list.prev != &qh->qtd_list) {
487                         last = list_entry (qtd->qtd_list.prev,
488                                         struct ehci_qtd, qtd_list);
489                         last->hw_next = qtd->hw_next;
490                 }
491
492                 /* remove qtd; it's recycled after possible urb completion */
493                 list_del (&qtd->qtd_list);
494                 last = qtd;
495
496                 /* reinit the xacterr counter for the next qtd */
497                 qh->xacterrs = QH_XACTERR_MAX;
498         }
499
500         /* last urb's completion might still need calling */
501         if (likely (last != NULL)) {
502                 ehci_urb_done(ehci, last->urb, last_status);
503                 count++;
504                 ehci_qtd_free (ehci, last);
505         }
506
507         /* restore original state; caller must unlink or relink */
508         qh->qh_state = state;
509
510         /* be sure the hardware's done with the qh before refreshing
511          * it after fault cleanup, or recovering from silicon wrongly
512          * overlaying the dummy qtd (which reduces DMA chatter).
513          */
514         if (stopped != 0 || qh->hw_qtd_next == EHCI_LIST_END(ehci)) {
515                 switch (state) {
516                 case QH_STATE_IDLE:
517                         qh_refresh(ehci, qh);
518                         break;
519                 case QH_STATE_LINKED:
520                         /* We won't refresh a QH that's linked (after the HC
521                          * stopped the queue).  That avoids a race:
522                          *  - HC reads first part of QH;
523                          *  - CPU updates that first part and the token;
524                          *  - HC reads rest of that QH, including token
525                          * Result:  HC gets an inconsistent image, and then
526                          * DMAs to/from the wrong memory (corrupting it).
527                          *
528                          * That should be rare for interrupt transfers,
529                          * except maybe high bandwidth ...
530                          */
531                         if ((cpu_to_hc32(ehci, QH_SMASK)
532                                         & qh->hw_info2) != 0) {
533                                 intr_deschedule (ehci, qh);
534                                 (void) qh_schedule (ehci, qh);
535                         } else
536                                 unlink_async (ehci, qh);
537                         break;
538                 /* otherwise, unlink already started */
539                 }
540         }
541
542         return count;
543 }
544
545 /*-------------------------------------------------------------------------*/
546
547 // high bandwidth multiplier, as encoded in highspeed endpoint descriptors
548 #define hb_mult(wMaxPacketSize) (1 + (((wMaxPacketSize) >> 11) & 0x03))
549 // ... and packet size, for any kind of endpoint descriptor
550 #define max_packet(wMaxPacketSize) ((wMaxPacketSize) & 0x07ff)
551
552 /*
553  * reverse of qh_urb_transaction:  free a list of TDs.
554  * used for cleanup after errors, before HC sees an URB's TDs.
555  */
556 static void qtd_list_free (
557         struct ehci_hcd         *ehci,
558         struct urb              *urb,
559         struct list_head        *qtd_list
560 ) {
561         struct list_head        *entry, *temp;
562
563         list_for_each_safe (entry, temp, qtd_list) {
564                 struct ehci_qtd *qtd;
565
566                 qtd = list_entry (entry, struct ehci_qtd, qtd_list);
567                 list_del (&qtd->qtd_list);
568                 ehci_qtd_free (ehci, qtd);
569         }
570 }
571
572 /*
573  * create a list of filled qtds for this URB; won't link into qh.
574  */
575 static struct list_head *
576 qh_urb_transaction (
577         struct ehci_hcd         *ehci,
578         struct urb              *urb,
579         struct list_head        *head,
580         gfp_t                   flags
581 ) {
582         struct ehci_qtd         *qtd, *qtd_prev;
583         dma_addr_t              buf;
584         int                     len, maxpacket;
585         int                     is_input;
586         u32                     token;
587
588         /*
589          * URBs map to sequences of QTDs:  one logical transaction
590          */
591         qtd = ehci_qtd_alloc (ehci, flags);
592         if (unlikely (!qtd))
593                 return NULL;
594         list_add_tail (&qtd->qtd_list, head);
595         qtd->urb = urb;
596
597         token = QTD_STS_ACTIVE;
598         token |= (EHCI_TUNE_CERR << 10);
599         /* for split transactions, SplitXState initialized to zero */
600
601         len = urb->transfer_buffer_length;
602         is_input = usb_pipein (urb->pipe);
603         if (usb_pipecontrol (urb->pipe)) {
604                 /* SETUP pid */
605                 qtd_fill(ehci, qtd, urb->setup_dma,
606                                 sizeof (struct usb_ctrlrequest),
607                                 token | (2 /* "setup" */ << 8), 8);
608
609                 /* ... and always at least one more pid */
610                 token ^= QTD_TOGGLE;
611                 qtd_prev = qtd;
612                 qtd = ehci_qtd_alloc (ehci, flags);
613                 if (unlikely (!qtd))
614                         goto cleanup;
615                 qtd->urb = urb;
616                 qtd_prev->hw_next = QTD_NEXT(ehci, qtd->qtd_dma);
617                 list_add_tail (&qtd->qtd_list, head);
618
619                 /* for zero length DATA stages, STATUS is always IN */
620                 if (len == 0)
621                         token |= (1 /* "in" */ << 8);
622         }
623
624         /*
625          * data transfer stage:  buffer setup
626          */
627         buf = urb->transfer_dma;
628
629         if (is_input)
630                 token |= (1 /* "in" */ << 8);
631         /* else it's already initted to "out" pid (0 << 8) */
632
633         maxpacket = max_packet(usb_maxpacket(urb->dev, urb->pipe, !is_input));
634
635         /*
636          * buffer gets wrapped in one or more qtds;
637          * last one may be "short" (including zero len)
638          * and may serve as a control status ack
639          */
640         for (;;) {
641                 int this_qtd_len;
642
643                 this_qtd_len = qtd_fill(ehci, qtd, buf, len, token, maxpacket);
644                 len -= this_qtd_len;
645                 buf += this_qtd_len;
646
647                 /*
648                  * short reads advance to a "magic" dummy instead of the next
649                  * qtd ... that forces the queue to stop, for manual cleanup.
650                  * (this will usually be overridden later.)
651                  */
652                 if (is_input)
653                         qtd->hw_alt_next = ehci->async->hw_alt_next;
654
655                 /* qh makes control packets use qtd toggle; maybe switch it */
656                 if ((maxpacket & (this_qtd_len + (maxpacket - 1))) == 0)
657                         token ^= QTD_TOGGLE;
658
659                 if (likely (len <= 0))
660                         break;
661
662                 qtd_prev = qtd;
663                 qtd = ehci_qtd_alloc (ehci, flags);
664                 if (unlikely (!qtd))
665                         goto cleanup;
666                 qtd->urb = urb;
667                 qtd_prev->hw_next = QTD_NEXT(ehci, qtd->qtd_dma);
668                 list_add_tail (&qtd->qtd_list, head);
669         }
670
671         /*
672          * unless the caller requires manual cleanup after short reads,
673          * have the alt_next mechanism keep the queue running after the
674          * last data qtd (the only one, for control and most other cases).
675          */
676         if (likely ((urb->transfer_flags & URB_SHORT_NOT_OK) == 0
677                                 || usb_pipecontrol (urb->pipe)))
678                 qtd->hw_alt_next = EHCI_LIST_END(ehci);
679
680         /*
681          * control requests may need a terminating data "status" ack;
682          * bulk ones may need a terminating short packet (zero length).
683          */
684         if (likely (urb->transfer_buffer_length != 0)) {
685                 int     one_more = 0;
686
687                 if (usb_pipecontrol (urb->pipe)) {
688                         one_more = 1;
689                         token ^= 0x0100;        /* "in" <--> "out"  */
690                         token |= QTD_TOGGLE;    /* force DATA1 */
691                 } else if (usb_pipebulk (urb->pipe)
692                                 && (urb->transfer_flags & URB_ZERO_PACKET)
693                                 && !(urb->transfer_buffer_length % maxpacket)) {
694                         one_more = 1;
695                 }
696                 if (one_more) {
697                         qtd_prev = qtd;
698                         qtd = ehci_qtd_alloc (ehci, flags);
699                         if (unlikely (!qtd))
700                                 goto cleanup;
701                         qtd->urb = urb;
702                         qtd_prev->hw_next = QTD_NEXT(ehci, qtd->qtd_dma);
703                         list_add_tail (&qtd->qtd_list, head);
704
705                         /* never any data in such packets */
706                         qtd_fill(ehci, qtd, 0, 0, token, 0);
707                 }
708         }
709
710         /* by default, enable interrupt on urb completion */
711         if (likely (!(urb->transfer_flags & URB_NO_INTERRUPT)))
712                 qtd->hw_token |= cpu_to_hc32(ehci, QTD_IOC);
713         return head;
714
715 cleanup:
716         qtd_list_free (ehci, urb, head);
717         return NULL;
718 }
719
720 /*-------------------------------------------------------------------------*/
721
722 // Would be best to create all qh's from config descriptors,
723 // when each interface/altsetting is established.  Unlink
724 // any previous qh and cancel its urbs first; endpoints are
725 // implicitly reset then (data toggle too).
726 // That'd mean updating how usbcore talks to HCDs. (2.7?)
727
728
729 /*
730  * Each QH holds a qtd list; a QH is used for everything except iso.
731  *
732  * For interrupt urbs, the scheduler must set the microframe scheduling
733  * mask(s) each time the QH gets scheduled.  For highspeed, that's
734  * just one microframe in the s-mask.  For split interrupt transactions
735  * there are additional complications: c-mask, maybe FSTNs.
736  */
737 static struct ehci_qh *
738 qh_make (
739         struct ehci_hcd         *ehci,
740         struct urb              *urb,
741         gfp_t                   flags
742 ) {
743         struct ehci_qh          *qh = ehci_qh_alloc (ehci, flags);
744         u32                     info1 = 0, info2 = 0;
745         int                     is_input, type;
746         int                     maxp = 0;
747         struct usb_tt           *tt = urb->dev->tt;
748
749         if (!qh)
750                 return qh;
751
752         /*
753          * init endpoint/device data for this QH
754          */
755         info1 |= usb_pipeendpoint (urb->pipe) << 8;
756         info1 |= usb_pipedevice (urb->pipe) << 0;
757
758         is_input = usb_pipein (urb->pipe);
759         type = usb_pipetype (urb->pipe);
760         maxp = usb_maxpacket (urb->dev, urb->pipe, !is_input);
761
762         /* 1024 byte maxpacket is a hardware ceiling.  High bandwidth
763          * acts like up to 3KB, but is built from smaller packets.
764          */
765         if (max_packet(maxp) > 1024) {
766                 ehci_dbg(ehci, "bogus qh maxpacket %d\n", max_packet(maxp));
767                 goto done;
768         }
769
770         /* Compute interrupt scheduling parameters just once, and save.
771          * - allowing for high bandwidth, how many nsec/uframe are used?
772          * - split transactions need a second CSPLIT uframe; same question
773          * - splits also need a schedule gap (for full/low speed I/O)
774          * - qh has a polling interval
775          *
776          * For control/bulk requests, the HC or TT handles these.
777          */
778         if (type == PIPE_INTERRUPT) {
779                 qh->usecs = NS_TO_US(usb_calc_bus_time(USB_SPEED_HIGH,
780                                 is_input, 0,
781                                 hb_mult(maxp) * max_packet(maxp)));
782                 qh->start = NO_FRAME;
783
784                 if (urb->dev->speed == USB_SPEED_HIGH) {
785                         qh->c_usecs = 0;
786                         qh->gap_uf = 0;
787
788                         qh->period = urb->interval >> 3;
789                         if (qh->period == 0 && urb->interval != 1) {
790                                 /* NOTE interval 2 or 4 uframes could work.
791                                  * But interval 1 scheduling is simpler, and
792                                  * includes high bandwidth.
793                                  */
794                                 dbg ("intr period %d uframes, NYET!",
795                                                 urb->interval);
796                                 goto done;
797                         }
798                 } else {
799                         int             think_time;
800
801                         /* gap is f(FS/LS transfer times) */
802                         qh->gap_uf = 1 + usb_calc_bus_time (urb->dev->speed,
803                                         is_input, 0, maxp) / (125 * 1000);
804
805                         /* FIXME this just approximates SPLIT/CSPLIT times */
806                         if (is_input) {         // SPLIT, gap, CSPLIT+DATA
807                                 qh->c_usecs = qh->usecs + HS_USECS (0);
808                                 qh->usecs = HS_USECS (1);
809                         } else {                // SPLIT+DATA, gap, CSPLIT
810                                 qh->usecs += HS_USECS (1);
811                                 qh->c_usecs = HS_USECS (0);
812                         }
813
814                         think_time = tt ? tt->think_time : 0;
815                         qh->tt_usecs = NS_TO_US (think_time +
816                                         usb_calc_bus_time (urb->dev->speed,
817                                         is_input, 0, max_packet (maxp)));
818                         qh->period = urb->interval;
819                 }
820         }
821
822         /* support for tt scheduling, and access to toggles */
823         qh->dev = urb->dev;
824
825         /* using TT? */
826         switch (urb->dev->speed) {
827         case USB_SPEED_LOW:
828                 info1 |= (1 << 12);     /* EPS "low" */
829                 /* FALL THROUGH */
830
831         case USB_SPEED_FULL:
832                 /* EPS 0 means "full" */
833                 if (type != PIPE_INTERRUPT)
834                         info1 |= (EHCI_TUNE_RL_TT << 28);
835                 if (type == PIPE_CONTROL) {
836                         info1 |= (1 << 27);     /* for TT */
837                         info1 |= 1 << 14;       /* toggle from qtd */
838                 }
839                 info1 |= maxp << 16;
840
841                 info2 |= (EHCI_TUNE_MULT_TT << 30);
842
843                 /* Some Freescale processors have an erratum in which the
844                  * port number in the queue head was 0..N-1 instead of 1..N.
845                  */
846                 if (ehci_has_fsl_portno_bug(ehci))
847                         info2 |= (urb->dev->ttport-1) << 23;
848                 else
849                         info2 |= urb->dev->ttport << 23;
850
851                 /* set the address of the TT; for TDI's integrated
852                  * root hub tt, leave it zeroed.
853                  */
854                 if (tt && tt->hub != ehci_to_hcd(ehci)->self.root_hub)
855                         info2 |= tt->hub->devnum << 16;
856
857                 /* NOTE:  if (PIPE_INTERRUPT) { scheduler sets c-mask } */
858
859                 break;
860
861         case USB_SPEED_HIGH:            /* no TT involved */
862                 info1 |= (2 << 12);     /* EPS "high" */
863                 if (type == PIPE_CONTROL) {
864                         info1 |= (EHCI_TUNE_RL_HS << 28);
865                         info1 |= 64 << 16;      /* usb2 fixed maxpacket */
866                         info1 |= 1 << 14;       /* toggle from qtd */
867                         info2 |= (EHCI_TUNE_MULT_HS << 30);
868                 } else if (type == PIPE_BULK) {
869                         info1 |= (EHCI_TUNE_RL_HS << 28);
870                         /* The USB spec says that high speed bulk endpoints
871                          * always use 512 byte maxpacket.  But some device
872                          * vendors decided to ignore that, and MSFT is happy
873                          * to help them do so.  So now people expect to use
874                          * such nonconformant devices with Linux too; sigh.
875                          */
876                         info1 |= max_packet(maxp) << 16;
877                         info2 |= (EHCI_TUNE_MULT_HS << 30);
878                 } else {                /* PIPE_INTERRUPT */
879                         info1 |= max_packet (maxp) << 16;
880                         info2 |= hb_mult (maxp) << 30;
881                 }
882                 break;
883         default:
884                 dbg ("bogus dev %p speed %d", urb->dev, urb->dev->speed);
885 done:
886                 qh_put (qh);
887                 return NULL;
888         }
889
890         /* NOTE:  if (PIPE_INTERRUPT) { scheduler sets s-mask } */
891
892         /* init as live, toggle clear, advance to dummy */
893         qh->qh_state = QH_STATE_IDLE;
894         qh->hw_info1 = cpu_to_hc32(ehci, info1);
895         qh->hw_info2 = cpu_to_hc32(ehci, info2);
896         usb_settoggle (urb->dev, usb_pipeendpoint (urb->pipe), !is_input, 1);
897         qh_refresh (ehci, qh);
898         return qh;
899 }
900
901 /*-------------------------------------------------------------------------*/
902
903 /* move qh (and its qtds) onto async queue; maybe enable queue.  */
904
905 static void qh_link_async (struct ehci_hcd *ehci, struct ehci_qh *qh)
906 {
907         __hc32          dma = QH_NEXT(ehci, qh->qh_dma);
908         struct ehci_qh  *head;
909
910         /* Don't link a QH if there's a Clear-TT-Buffer pending */
911         if (unlikely(qh->clearing_tt))
912                 return;
913
914         /* (re)start the async schedule? */
915         head = ehci->async;
916         timer_action_done (ehci, TIMER_ASYNC_OFF);
917         if (!head->qh_next.qh) {
918                 u32     cmd = ehci_readl(ehci, &ehci->regs->command);
919
920                 if (!(cmd & CMD_ASE)) {
921                         /* in case a clear of CMD_ASE didn't take yet */
922                         (void)handshake(ehci, &ehci->regs->status,
923                                         STS_ASS, 0, 150);
924                         cmd |= CMD_ASE | CMD_RUN;
925                         ehci_writel(ehci, cmd, &ehci->regs->command);
926                         ehci_to_hcd(ehci)->state = HC_STATE_RUNNING;
927                         /* posted write need not be known to HC yet ... */
928                 }
929         }
930
931         /* clear halt and/or toggle; and maybe recover from silicon quirk */
932         if (qh->qh_state == QH_STATE_IDLE)
933                 qh_refresh (ehci, qh);
934
935         /* splice right after start */
936         qh->qh_next = head->qh_next;
937         qh->hw_next = head->hw_next;
938         wmb ();
939
940         head->qh_next.qh = qh;
941         head->hw_next = dma;
942
943         qh->xacterrs = QH_XACTERR_MAX;
944         qh->qh_state = QH_STATE_LINKED;
945         /* qtd completions reported later by interrupt */
946 }
947
948 /*-------------------------------------------------------------------------*/
949
950 /*
951  * For control/bulk/interrupt, return QH with these TDs appended.
952  * Allocates and initializes the QH if necessary.
953  * Returns null if it can't allocate a QH it needs to.
954  * If the QH has TDs (urbs) already, that's great.
955  */
956 static struct ehci_qh *qh_append_tds (
957         struct ehci_hcd         *ehci,
958         struct urb              *urb,
959         struct list_head        *qtd_list,
960         int                     epnum,
961         void                    **ptr
962 )
963 {
964         struct ehci_qh          *qh = NULL;
965         __hc32                  qh_addr_mask = cpu_to_hc32(ehci, 0x7f);
966
967         qh = (struct ehci_qh *) *ptr;
968         if (unlikely (qh == NULL)) {
969                 /* can't sleep here, we have ehci->lock... */
970                 qh = qh_make (ehci, urb, GFP_ATOMIC);
971                 *ptr = qh;
972         }
973         if (likely (qh != NULL)) {
974                 struct ehci_qtd *qtd;
975
976                 if (unlikely (list_empty (qtd_list)))
977                         qtd = NULL;
978                 else
979                         qtd = list_entry (qtd_list->next, struct ehci_qtd,
980                                         qtd_list);
981
982                 /* control qh may need patching ... */
983                 if (unlikely (epnum == 0)) {
984
985                         /* usb_reset_device() briefly reverts to address 0 */
986                         if (usb_pipedevice (urb->pipe) == 0)
987                                 qh->hw_info1 &= ~qh_addr_mask;
988                 }
989
990                 /* just one way to queue requests: swap with the dummy qtd.
991                  * only hc or qh_refresh() ever modify the overlay.
992                  */
993                 if (likely (qtd != NULL)) {
994                         struct ehci_qtd         *dummy;
995                         dma_addr_t              dma;
996                         __hc32                  token;
997
998                         /* to avoid racing the HC, use the dummy td instead of
999                          * the first td of our list (becomes new dummy).  both
1000                          * tds stay deactivated until we're done, when the
1001                          * HC is allowed to fetch the old dummy (4.10.2).
1002                          */
1003                         token = qtd->hw_token;
1004                         qtd->hw_token = HALT_BIT(ehci);
1005                         wmb ();
1006                         dummy = qh->dummy;
1007
1008                         dma = dummy->qtd_dma;
1009                         *dummy = *qtd;
1010                         dummy->qtd_dma = dma;
1011
1012                         list_del (&qtd->qtd_list);
1013                         list_add (&dummy->qtd_list, qtd_list);
1014                         list_splice_tail(qtd_list, &qh->qtd_list);
1015
1016                         ehci_qtd_init(ehci, qtd, qtd->qtd_dma);
1017                         qh->dummy = qtd;
1018
1019                         /* hc must see the new dummy at list end */
1020                         dma = qtd->qtd_dma;
1021                         qtd = list_entry (qh->qtd_list.prev,
1022                                         struct ehci_qtd, qtd_list);
1023                         qtd->hw_next = QTD_NEXT(ehci, dma);
1024
1025                         /* let the hc process these next qtds */
1026                         wmb ();
1027                         dummy->hw_token = token;
1028
1029                         urb->hcpriv = qh_get (qh);
1030                 }
1031         }
1032         return qh;
1033 }
1034
1035 /*-------------------------------------------------------------------------*/
1036
1037 static int
1038 submit_async (
1039         struct ehci_hcd         *ehci,
1040         struct urb              *urb,
1041         struct list_head        *qtd_list,
1042         gfp_t                   mem_flags
1043 ) {
1044         struct ehci_qtd         *qtd;
1045         int                     epnum;
1046         unsigned long           flags;
1047         struct ehci_qh          *qh = NULL;
1048         int                     rc;
1049
1050         qtd = list_entry (qtd_list->next, struct ehci_qtd, qtd_list);
1051         epnum = urb->ep->desc.bEndpointAddress;
1052
1053 #ifdef EHCI_URB_TRACE
1054         ehci_dbg (ehci,
1055                 "%s %s urb %p ep%d%s len %d, qtd %p [qh %p]\n",
1056                 __func__, urb->dev->devpath, urb,
1057                 epnum & 0x0f, (epnum & USB_DIR_IN) ? "in" : "out",
1058                 urb->transfer_buffer_length,
1059                 qtd, urb->ep->hcpriv);
1060 #endif
1061
1062         spin_lock_irqsave (&ehci->lock, flags);
1063         if (unlikely(!test_bit(HCD_FLAG_HW_ACCESSIBLE,
1064                                &ehci_to_hcd(ehci)->flags))) {
1065                 rc = -ESHUTDOWN;
1066                 goto done;
1067         }
1068         rc = usb_hcd_link_urb_to_ep(ehci_to_hcd(ehci), urb);
1069         if (unlikely(rc))
1070                 goto done;
1071
1072         qh = qh_append_tds(ehci, urb, qtd_list, epnum, &urb->ep->hcpriv);
1073         if (unlikely(qh == NULL)) {
1074                 usb_hcd_unlink_urb_from_ep(ehci_to_hcd(ehci), urb);
1075                 rc = -ENOMEM;
1076                 goto done;
1077         }
1078
1079         /* Control/bulk operations through TTs don't need scheduling,
1080          * the HC and TT handle it when the TT has a buffer ready.
1081          */
1082         if (likely (qh->qh_state == QH_STATE_IDLE))
1083                 qh_link_async (ehci, qh_get (qh));
1084  done:
1085         spin_unlock_irqrestore (&ehci->lock, flags);
1086         if (unlikely (qh == NULL))
1087                 qtd_list_free (ehci, urb, qtd_list);
1088         return rc;
1089 }
1090
1091 /*-------------------------------------------------------------------------*/
1092
1093 /* the async qh for the qtds being reclaimed are now unlinked from the HC */
1094
1095 static void end_unlink_async (struct ehci_hcd *ehci)
1096 {
1097         struct ehci_qh          *qh = ehci->reclaim;
1098         struct ehci_qh          *next;
1099
1100         iaa_watchdog_done(ehci);
1101
1102         // qh->hw_next = cpu_to_hc32(qh->qh_dma);
1103         qh->qh_state = QH_STATE_IDLE;
1104         qh->qh_next.qh = NULL;
1105         qh_put (qh);                    // refcount from reclaim
1106
1107         /* other unlink(s) may be pending (in QH_STATE_UNLINK_WAIT) */
1108         next = qh->reclaim;
1109         ehci->reclaim = next;
1110         qh->reclaim = NULL;
1111
1112         qh_completions (ehci, qh);
1113
1114         if (!list_empty (&qh->qtd_list)
1115                         && HC_IS_RUNNING (ehci_to_hcd(ehci)->state))
1116                 qh_link_async (ehci, qh);
1117         else {
1118                 qh_put (qh);            // refcount from async list
1119
1120                 /* it's not free to turn the async schedule on/off; leave it
1121                  * active but idle for a while once it empties.
1122                  */
1123                 if (HC_IS_RUNNING (ehci_to_hcd(ehci)->state)
1124                                 && ehci->async->qh_next.qh == NULL)
1125                         timer_action (ehci, TIMER_ASYNC_OFF);
1126         }
1127
1128         if (next) {
1129                 ehci->reclaim = NULL;
1130                 start_unlink_async (ehci, next);
1131         }
1132 }
1133
1134 /* makes sure the async qh will become idle */
1135 /* caller must own ehci->lock */
1136
1137 static void start_unlink_async (struct ehci_hcd *ehci, struct ehci_qh *qh)
1138 {
1139         int             cmd = ehci_readl(ehci, &ehci->regs->command);
1140         struct ehci_qh  *prev;
1141
1142 #ifdef DEBUG
1143         assert_spin_locked(&ehci->lock);
1144         if (ehci->reclaim
1145                         || (qh->qh_state != QH_STATE_LINKED
1146                                 && qh->qh_state != QH_STATE_UNLINK_WAIT)
1147                         )
1148                 BUG ();
1149 #endif
1150
1151         /* stop async schedule right now? */
1152         if (unlikely (qh == ehci->async)) {
1153                 /* can't get here without STS_ASS set */
1154                 if (ehci_to_hcd(ehci)->state != HC_STATE_HALT
1155                                 && !ehci->reclaim) {
1156                         /* ... and CMD_IAAD clear */
1157                         ehci_writel(ehci, cmd & ~CMD_ASE,
1158                                     &ehci->regs->command);
1159                         wmb ();
1160                         // handshake later, if we need to
1161                         timer_action_done (ehci, TIMER_ASYNC_OFF);
1162                 }
1163                 return;
1164         }
1165
1166         qh->qh_state = QH_STATE_UNLINK;
1167         ehci->reclaim = qh = qh_get (qh);
1168
1169         prev = ehci->async;
1170         while (prev->qh_next.qh != qh)
1171                 prev = prev->qh_next.qh;
1172
1173         prev->hw_next = qh->hw_next;
1174         prev->qh_next = qh->qh_next;
1175         wmb ();
1176
1177         /* If the controller isn't running, we don't have to wait for it */
1178         if (unlikely(!HC_IS_RUNNING(ehci_to_hcd(ehci)->state))) {
1179                 /* if (unlikely (qh->reclaim != 0))
1180                  *      this will recurse, probably not much
1181                  */
1182                 end_unlink_async (ehci);
1183                 return;
1184         }
1185
1186         cmd |= CMD_IAAD;
1187         ehci_writel(ehci, cmd, &ehci->regs->command);
1188         (void)ehci_readl(ehci, &ehci->regs->command);
1189         iaa_watchdog_start(ehci);
1190 }
1191
1192 /*-------------------------------------------------------------------------*/
1193
1194 static void scan_async (struct ehci_hcd *ehci)
1195 {
1196         struct ehci_qh          *qh;
1197         enum ehci_timer_action  action = TIMER_IO_WATCHDOG;
1198
1199         ehci->stamp = ehci_readl(ehci, &ehci->regs->frame_index);
1200         timer_action_done (ehci, TIMER_ASYNC_SHRINK);
1201 rescan:
1202         qh = ehci->async->qh_next.qh;
1203         if (likely (qh != NULL)) {
1204                 do {
1205                         /* clean any finished work for this qh */
1206                         if (!list_empty (&qh->qtd_list)
1207                                         && qh->stamp != ehci->stamp) {
1208                                 int temp;
1209
1210                                 /* unlinks could happen here; completion
1211                                  * reporting drops the lock.  rescan using
1212                                  * the latest schedule, but don't rescan
1213                                  * qhs we already finished (no looping).
1214                                  */
1215                                 qh = qh_get (qh);
1216                                 qh->stamp = ehci->stamp;
1217                                 temp = qh_completions (ehci, qh);
1218                                 qh_put (qh);
1219                                 if (temp != 0) {
1220                                         goto rescan;
1221                                 }
1222                         }
1223
1224                         /* unlink idle entries, reducing DMA usage as well
1225                          * as HCD schedule-scanning costs.  delay for any qh
1226                          * we just scanned, there's a not-unusual case that it
1227                          * doesn't stay idle for long.
1228                          * (plus, avoids some kind of re-activation race.)
1229                          */
1230                         if (list_empty(&qh->qtd_list)
1231                                         && qh->qh_state == QH_STATE_LINKED) {
1232                                 if (!ehci->reclaim
1233                                         && ((ehci->stamp - qh->stamp) & 0x1fff)
1234                                                 >= (EHCI_SHRINK_FRAMES * 8))
1235                                         start_unlink_async(ehci, qh);
1236                                 else
1237                                         action = TIMER_ASYNC_SHRINK;
1238                         }
1239
1240                         qh = qh->qh_next.qh;
1241                 } while (qh);
1242         }
1243         if (action == TIMER_ASYNC_SHRINK)
1244                 timer_action (ehci, TIMER_ASYNC_SHRINK);
1245 }