[PATCH] USB: ehci update: 3/3, highspeed iso rewrite

This is an updated version of a patch submitted to me from
Michal Sojka <sojkam1@fel.cvut.cz>, basically providing a
much-needed rewrite of the highspeed ISO support.  I updated
the scheduling and made it a closer match to how OHCI works;
and also tested it a bunch.

So far it seems most of the requests for highspeed ISO support
have been for realtime data collection -- custom apps, nothing
a mainstream kernel would ship with.   The USB Video class is
now defined; highspeed video will also need these updates.

Key changes:

   - Define an "iso_stream" head for iso endpoints.  This acts
     enough like a QH that endpoint_disable() works.  It holds the
     queue of ITDs, and the endpoint's current schedule state.
     And it's easy to find (spinlocked array access), no search.

   - Uses a temporary "itd_sched" while submitting each URB, with
     not-yet-linked ITDs and request-specific metadata.  There's
     a per-stream cache of ITDs, so resubmitting ISO urbs (to
     achieve a "ring" of transfers) is typically cheap.

   - Scheduling for most URBs is almost a NOP:  just a sanity
     check to make sure there's no need to reschedule, and then
     just link into the schedule at the current schedule slot.
     (The previous code was a gross hack that didn't even work
     reasonably with more than two URBs queued.)

   - Is a reasonable model to use with full speed ISO transfers.
     (They need additional TT scheduling hooks, most of which
     are already written but not merged.)

   - Handles several cases the previous code didn't, including
     high bandwidth transfers (loads up to 24 MByte/sec)

   - Has had more testing than the old code, including 20+ hour
     successful IN+OUT runs, more varied transfer intervals and
     maxpacket sizes.  (Using net2280 and a gadgetfs driver.)

So it's worth replacing the existing code with this; there
aren't too many rough edges, and it's much more fixable than
the previous version.


p.s. Many thanks, Michal!

drivers/usb/host/ehci-dbg.c

@@ -140,6 +140,36 @@ dbg_qh (const char *label, struct ehci_hcd *ehci, struct ehci_qh *qh)
 	dbg_qtd ("overlay", ehci, (struct ehci_qtd *) &qh->hw_qtd_next);
 }
 
+static void __attribute__((__unused__))
+dbg_itd (const char *label, struct ehci_hcd *ehci, struct ehci_itd *itd) 
+{
+	ehci_dbg (ehci, "%s [%d] itd %p, next %08x, urb %p\n",
+		label, itd->frame, itd, le32_to_cpu(itd->hw_next), itd->urb);
+	ehci_dbg (ehci,
+		"  trans: %08x %08x %08x %08x %08x %08x %08x %08x\n", 
+		le32_to_cpu(itd->hw_transaction[0]),
+		le32_to_cpu(itd->hw_transaction[1]),
+		le32_to_cpu(itd->hw_transaction[2]),
+		le32_to_cpu(itd->hw_transaction[3]),
+		le32_to_cpu(itd->hw_transaction[4]),
+		le32_to_cpu(itd->hw_transaction[5]),
+		le32_to_cpu(itd->hw_transaction[6]),
+		le32_to_cpu(itd->hw_transaction[7]));
+	ehci_dbg (ehci,
+		"  buf:   %08x %08x %08x %08x %08x %08x %08x\n", 
+		le32_to_cpu(itd->hw_bufp[0]),
+		le32_to_cpu(itd->hw_bufp[1]),
+		le32_to_cpu(itd->hw_bufp[2]),
+		le32_to_cpu(itd->hw_bufp[3]),
+		le32_to_cpu(itd->hw_bufp[4]),
+		le32_to_cpu(itd->hw_bufp[5]),
+		le32_to_cpu(itd->hw_bufp[6]));
+	ehci_dbg (ehci, "  index: %d %d %d %d %d %d %d %d\n",
+		itd->index[0], itd->index[1], itd->index[2],
+		itd->index[3], itd->index[4], itd->index[5],
+		itd->index[6], itd->index[7]);
+}
+
 static int __attribute__((__unused__))
 dbg_status_buf (char *buf, unsigned len, char *label, u32 status)
 {

drivers/usb/host/ehci-hcd.c

@@ -67,6 +67,9 @@
  *
  * HISTORY:
  *
+ * 2003-12-29 Rewritten high speed iso transfer support (by Michal Sojka,
+ *	<sojkam@centrum.cz>, updates by DB).
+ *
  * 2002-11-29	Correct handling for hw async_next register.
  * 2002-08-06	Handling for bulk and interrupt transfers is mostly shared;
  *	only scheduling is different, no arbitrary limitations.
@@ -90,7 +93,7 @@
  * 2001-June	Works with usb-storage and NEC EHCI on 2.4
  */
 
-#define DRIVER_VERSION "2003-Jun-13"
+#define DRIVER_VERSION "2003-Dec-29"
 #define DRIVER_AUTHOR "David Brownell"
 #define DRIVER_DESC "USB 2.0 'Enhanced' Host Controller (EHCI) Driver"
 
@@ -901,10 +904,19 @@ ehci_endpoint_disable (struct usb_hcd *hcd, struct hcd_dev *dev, int ep)
 	if (!qh)
 		goto done;
 
+	/* endpoints can be iso streams.  for now, we don't
+	 * accelerate iso completions ... so spin a while.
+	 */
+	if (qh->hw_info1 == 0) {
+		ehci_vdbg (ehci, "iso delay\n");
+		goto idle_timeout;
+	}
+
 	if (!HCD_IS_RUNNING (ehci->hcd.state))
 		qh->qh_state = QH_STATE_IDLE;
 	switch (qh->qh_state) {
 	case QH_STATE_UNLINK:		/* wait for hw to finish? */
+idle_timeout:
 		spin_unlock_irqrestore (&ehci->lock, flags);
 		set_current_state (TASK_UNINTERRUPTIBLE);
 		schedule_timeout (1);

drivers/usb/host/ehci.h

@@ -386,6 +386,63 @@ struct ehci_qh {
 
 /*-------------------------------------------------------------------------*/
 
+/* description of one iso highspeed transaction (up to 3 KB data) */
+struct ehci_iso_uframe {
+	/* These will be copied to iTD when scheduling */
+	u64			bufp;		/* itd->hw_bufp{,_hi}[pg] |= */
+	u32			transaction;	/* itd->hw_transaction[i] |= */
+	u8			cross;		/* buf crosses pages */
+};
+
+/* temporary schedule data for highspeed packets from iso urbs
+ * each packet is one uframe's usb transactions, in some itd,
+ * beginning at stream->next_uframe
+ */
+struct ehci_itd_sched {
+	struct list_head	itd_list;
+	unsigned		span;
+	struct ehci_iso_uframe	packet [0];
+};
+
+/*
+ * ehci_iso_stream - groups all (s)itds for this endpoint.
+ * acts like a qh would, if EHCI had them for ISO.
+ */
+struct ehci_iso_stream {
+	/* first two fields match QH, but info1 == 0 */
+	u32			hw_next;
+	u32			hw_info1;
+
+	u32			refcount;
+	u8			bEndpointAddress;
+	struct list_head	itd_list;	/* queued itds */
+	struct list_head	free_itd_list;	/* list of unused itds */
+	struct hcd_dev		*dev;
+
+	/* output of (re)scheduling */
+	unsigned long		start;		/* jiffies */
+	unsigned long		rescheduled;
+	int			next_uframe;
+
+	/* the rest is derived from the endpoint descriptor,
+	 * trusting urb->interval == (1 << (epdesc->bInterval - 1)),
+	 * including the extra info for hw_bufp[0..2]
+	 */
+	u8			interval;
+	u8			usecs;		
+	u16			maxp;
+	unsigned		bandwidth;
+
+	/* This is used to initialize iTD's hw_bufp fields */
+	u32			buf0;		
+	u32			buf1;		
+	u32			buf2;
+
+	/* ... sITD won't use buf[012], and needs TT access ... */
+};
+
+/*-------------------------------------------------------------------------*/
+
 /*
  * EHCI Specification 0.95 Section 3.3
  * Fig 3-4 "Isochronous Transaction Descriptor (iTD)"
@@ -413,14 +470,14 @@ struct ehci_itd {
 	union ehci_shadow	itd_next;	/* ptr to periodic q entry */
 
 	struct urb		*urb;
-	struct list_head	itd_list;	/* list of urb frames' itds */
-	dma_addr_t		buf_dma;	/* frame's buffer address */
-
-	/* for now, only one hw_transaction per itd */
-	u32			transaction;
-	u16			index;		/* in urb->iso_frame_desc */
-	u16			uframe;		/* in periodic schedule */
-	u16			usecs;
+	struct ehci_iso_stream	*stream;	/* endpoint's queue */
+	struct list_head	itd_list;	/* list of stream's itds */
+
+	/* any/all hw_transactions here may be used by that urb */
+	unsigned		frame;		/* where scheduled */
+	unsigned		pg;
+	unsigned		index[8];	/* in urb->iso_frame_desc */
+	u8			usecs[8];
 } __attribute__ ((aligned (32)));
 
 /*-------------------------------------------------------------------------*/