wa-xfer.c

/*
 * WUSB Wire Adapter
 * Data transfer and URB enqueing
 *
 * Copyright (C) 2005-2006 Intel Corporation
 * Inaky Perez-Gonzalez <inaky.perez-gonzalez@intel.com>
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License version
 * 2 as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
 * 02110-1301, USA.
 *
 *
 * How transfers work: get a buffer, break it up in segments (segment
 * size is a multiple of the maxpacket size). For each segment issue a
 * segment request (struct wa_xfer_*), then send the data buffer if
 * out or nothing if in (all over the DTO endpoint).
 *
 * For each submitted segment request, a notification will come over
 * the NEP endpoint and a transfer result (struct xfer_result) will
 * arrive in the DTI URB. Read it, get the xfer ID, see if there is
 * data coming (inbound transfer), schedule a read and handle it.
 *
 * Sounds simple, it is a pain to implement.
 *
 *
 * ENTRY POINTS
 *
 *   FIXME
 *
 * LIFE CYCLE / STATE DIAGRAM
 *
 *   FIXME
 *
 * THIS CODE IS DISGUSTING
 *
 *   Warned you are; it's my second try and still not happy with it.
 *
 * NOTES:
 *
 *   - No iso
 *
 *   - Supports DMA xfers, control, bulk and maybe interrupt
 *
 *   - Does not recycle unused rpipes
 *
 *     An rpipe is assigned to an endpoint the first time it is used,
 *     and then it's there, assigned, until the endpoint is disabled
 *     (destroyed [{h,d}wahc_op_ep_disable()]. The assignment of the
 *     rpipe to the endpoint is done under the wa->rpipe_sem semaphore
 *     (should be a mutex).
 *
 *     Two methods it could be done:
 *
 *     (a) set up a timer every time an rpipe's use count drops to 1
 *         (which means unused) or when a transfer ends. Reset the
 *         timer when a xfer is queued. If the timer expires, release
 *         the rpipe [see rpipe_ep_disable()].
 *
 *     (b) when looking for free rpipes to attach [rpipe_get_by_ep()],
 *         when none are found go over the list, check their endpoint
 *         and their activity record (if no last-xfer-done-ts in the
 *         last x seconds) take it
 *
 *     However, due to the fact that we have a set of limited
 *     resources (max-segments-at-the-same-time per xfer,
 *     xfers-per-ripe, blocks-per-rpipe, rpipes-per-host), at the end
 *     we are going to have to rebuild all this based on an scheduler,
 *     to where we have a list of transactions to do and based on the
 *     availability of the different required components (blocks,
 *     rpipes, segment slots, etc), we go scheduling them. Painful.
 */
#include <linux/init.h>
#include <linux/spinlock.h>
#include <linux/slab.h>
#include <linux/hash.h>
#include <linux/ratelimit.h>
#include <linux/export.h>

#include "wa-hc.h"
#include "wusbhc.h"

enum {
	WA_SEGS_MAX = 255,
};

enum wa_seg_status {
	WA_SEG_NOTREADY,
	WA_SEG_READY,
	WA_SEG_DELAYED,
	WA_SEG_SUBMITTED,
	WA_SEG_PENDING,
	WA_SEG_DTI_PENDING,
	WA_SEG_DONE,
	WA_SEG_ERROR,
	WA_SEG_ABORTED,
};

static void wa_xfer_delayed_run(struct wa_rpipe *);

/*
 * Life cycle governed by 'struct urb' (the refcount of the struct is
 * that of the 'struct urb' and usb_free_urb() would free the whole
 * struct).
 */
struct wa_seg {
	struct urb urb;
	struct urb *dto_urb;		/* for data output? */
	struct list_head list_node;	/* for rpipe->req_list */
	struct wa_xfer *xfer;		/* out xfer */
	u8 index;			/* which segment we are */
	enum wa_seg_status status;
	ssize_t result;			/* bytes xfered or error */
	struct wa_xfer_hdr xfer_hdr;
	u8 xfer_extra[];		/* xtra space for xfer_hdr_ctl */
};

static void wa_seg_init(struct wa_seg *seg)
{
	/* usb_init_urb() repeats a lot of work, so we do it here */
	kref_init(&seg->urb.kref);
}

/*
 * Protected by xfer->lock
 *
 */
struct wa_xfer {
	struct kref refcnt;
	struct list_head list_node;
	spinlock_t lock;
	u32 id;

	struct wahc *wa;		/* Wire adapter we are plugged to */
	struct usb_host_endpoint *ep;
	struct urb *urb;		/* URB we are transferring for */
	struct wa_seg **seg;		/* transfer segments */
	u8 segs, segs_submitted, segs_done;
	unsigned is_inbound:1;
	unsigned is_dma:1;
	size_t seg_size;
	int result;

	gfp_t gfp;			/* allocation mask */

	struct wusb_dev *wusb_dev;	/* for activity timestamps */
};

static inline void wa_xfer_init(struct wa_xfer *xfer)
{
	kref_init(&xfer->refcnt);
	INIT_LIST_HEAD(&xfer->list_node);
	spin_lock_init(&xfer->lock);
}

/*
 * Destroy a transfer structure
 *
 * Note that the xfer->seg[index] thingies follow the URB life cycle,
 * so we need to put them, not free them.
 */
static void wa_xfer_destroy(struct kref *_xfer)
{
	struct wa_xfer *xfer = container_of(_xfer, struct wa_xfer, refcnt);
	if (xfer->seg) {
		unsigned cnt;
		for (cnt = 0; cnt < xfer->segs; cnt++) {
			if (xfer->is_inbound)
				usb_put_urb(xfer->seg[cnt]->dto_urb);
			usb_put_urb(&xfer->seg[cnt]->urb);
		}
	}
	kfree(xfer);
}

static void wa_xfer_get(struct wa_xfer *xfer)
{
	kref_get(&xfer->refcnt);
}

static void wa_xfer_put(struct wa_xfer *xfer)
{
	kref_put(&xfer->refcnt, wa_xfer_destroy);
}

/*
 * xfer is referenced
 *
 * xfer->lock has to be unlocked
 *
 * We take xfer->lock for setting the result; this is a barrier
 * against drivers/usb/core/hcd.c:unlink1() being called after we call
 * usb_hcd_giveback_urb() and wa_urb_dequeue() trying to get a
 * reference to the transfer.
 */
static void wa_xfer_giveback(struct wa_xfer *xfer)
{
	unsigned long flags;

	spin_lock_irqsave(&xfer->wa->xfer_list_lock, flags);
	list_del_init(&xfer->list_node);
	spin_unlock_irqrestore(&xfer->wa->xfer_list_lock, flags);
	/* FIXME: segmentation broken -- kills DWA */
	wusbhc_giveback_urb(xfer->wa->wusb, xfer->urb, xfer->result);
	wa_put(xfer->wa);
	wa_xfer_put(xfer);
}

/*
 * xfer is referenced
 *
 * xfer->lock has to be unlocked
 */
static void wa_xfer_completion(struct wa_xfer *xfer)
{
	if (xfer->wusb_dev)
		wusb_dev_put(xfer->wusb_dev);
	rpipe_put(xfer->ep->hcpriv);
	wa_xfer_giveback(xfer);
}

/*
 * If transfer is done, wrap it up and return true
 *
 * xfer->lock has to be locked
 */
static unsigned __wa_xfer_is_done(struct wa_xfer *xfer)
{
	struct device *dev = &xfer->wa->usb_iface->dev;
	unsigned result, cnt;
	struct wa_seg *seg;
	struct urb *urb = xfer->urb;
	unsigned found_short = 0;

	result = xfer->segs_done == xfer->segs_submitted;
	if (result == 0)
		goto out;
	urb->actual_length = 0;
	for (cnt = 0; cnt < xfer->segs; cnt++) {
		seg = xfer->seg[cnt];
		switch (seg->status) {
		case WA_SEG_DONE:
			if (found_short && seg->result > 0) {
				dev_dbg(dev, "xfer %p#%u: bad short segments (%zu)\n",
					xfer, cnt, seg->result);
				urb->status = -EINVAL;
				goto out;
			}
			urb->actual_length += seg->result;
			if (seg->result < xfer->seg_size
			    && cnt != xfer->segs-1)
				found_short = 1;
			dev_dbg(dev, "xfer %p#%u: DONE short %d "
				"result %zu urb->actual_length %d\n",
				xfer, seg->index, found_short, seg->result,
				urb->actual_length);
			break;
		case WA_SEG_ERROR:
			xfer->result = seg->result;
			dev_dbg(dev, "xfer %p#%u: ERROR result %zu\n",
				xfer, seg->index, seg->result);
			goto out;
		case WA_SEG_ABORTED:
			dev_dbg(dev, "xfer %p#%u ABORTED: result %d\n",
				xfer, seg->index, urb->status);
			xfer->result = urb->status;
			goto out;
		default:
			dev_warn(dev, "xfer %p#%u: is_done bad state %d\n",
				 xfer, cnt, seg->status);
			xfer->result = -EINVAL;
			goto out;
		}
	}
	xfer->result = 0;
out:
	return result;
}

/*
 * Initialize a transfer's ID
 *
 * We need to use a sequential number; if we use the pointer or the
 * hash of the pointer, it can repeat over sequential transfers and
 * then it will confuse the HWA....wonder why in hell they put a 32
 * bit handle in there then.
 */
static void wa_xfer_id_init(struct wa_xfer *xfer)
{
	xfer->id = atomic_add_return(1, &xfer->wa->xfer_id_count);
}

/*
 * Return the xfer's ID associated with xfer
 *
 * Need to generate a
 */
static u32 wa_xfer_id(struct wa_xfer *xfer)
{
	return xfer->id;
}

/*
 * Search for a transfer list ID on the HCD's URB list
 *
 * For 32 bit architectures, we use the pointer itself; for 64 bits, a
 * 32-bit hash of the pointer.
 *
 * @returns NULL if not found.
 */
static struct wa_xfer *wa_xfer_get_by_id(struct wahc *wa, u32 id)
{
	unsigned long flags;
	struct wa_xfer *xfer_itr;
	spin_lock_irqsave(&wa->xfer_list_lock, flags);
	list_for_each_entry(xfer_itr, &wa->xfer_list, list_node) {
		if (id == xfer_itr->id) {
			wa_xfer_get(xfer_itr);
			goto out;
		}
	}
	xfer_itr = NULL;
out:
	spin_unlock_irqrestore(&wa->xfer_list_lock, flags);
	return xfer_itr;
}

struct wa_xfer_abort_buffer {
	struct urb urb;
	struct wa_xfer_abort cmd;
};

static void __wa_xfer_abort_cb(struct urb *urb)
{
	struct wa_xfer_abort_buffer *b = urb->context;
	usb_put_urb(&b->urb);
}

/*
 * Aborts an ongoing transaction
 *
 * Assumes the transfer is referenced and locked and in a submitted
 * state (mainly that there is an endpoint/rpipe assigned).
 *
 * The callback (see above) does nothing but freeing up the data by
 * putting the URB. Because the URB is allocated at the head of the
 * struct, the whole space we allocated is kfreed.
 *
 * We'll get an 'aborted transaction' xfer result on DTI, that'll
 * politely ignore because at this point the transaction has been
 * marked as aborted already.
 */
static void __wa_xfer_abort(struct wa_xfer *xfer)
{
	int result;
	struct device *dev = &xfer->wa->usb_iface->dev;
	struct wa_xfer_abort_buffer *b;
	struct wa_rpipe *rpipe = xfer->ep->hcpriv;

	b = kmalloc(sizeof(*b), GFP_ATOMIC);
	if (b == NULL)
		goto error_kmalloc;
	b->cmd.bLength =  sizeof(b->cmd);
	b->cmd.bRequestType = WA_XFER_ABORT;
	b->cmd.wRPipe = rpipe->descr.wRPipeIndex;
	b->cmd.dwTransferID = wa_xfer_id(xfer);

	usb_init_urb(&b->urb);
	usb_fill_bulk_urb(&b->urb, xfer->wa->usb_dev,
		usb_sndbulkpipe(xfer->wa->usb_dev,
				xfer->wa->dto_epd->bEndpointAddress),
		&b->cmd, sizeof(b->cmd), __wa_xfer_abort_cb, b);
	result = usb_submit_urb(&b->urb, GFP_ATOMIC);
	if (result < 0)
		goto error_submit;
	return;				/* callback frees! */


error_submit:
	if (printk_ratelimit())
		dev_err(dev, "xfer %p: Can't submit abort request: %d\n",
			xfer, result);
	kfree(b);
error_kmalloc:
	return;

}

/*
 *
 * @returns < 0 on error, transfer segment request size if ok
 */
static ssize_t __wa_xfer_setup_sizes(struct wa_xfer *xfer,
				     enum wa_xfer_type *pxfer_type)
{
	ssize_t result;
	struct device *dev = &xfer->wa->usb_iface->dev;
	size_t maxpktsize;
	struct urb *urb = xfer->urb;
	struct wa_rpipe *rpipe = xfer->ep->hcpriv;

	switch (rpipe->descr.bmAttribute & 0x3) {
	case USB_ENDPOINT_XFER_CONTROL:
		*pxfer_type = WA_XFER_TYPE_CTL;
		result = sizeof(struct wa_xfer_ctl);
		break;
	case USB_ENDPOINT_XFER_INT:
	case USB_ENDPOINT_XFER_BULK:
		*pxfer_type = WA_XFER_TYPE_BI;
		result = sizeof(struct wa_xfer_bi);
		break;
	case USB_ENDPOINT_XFER_ISOC:
		dev_err(dev, "FIXME: ISOC not implemented\n");
		result = -ENOSYS;
		goto error;
	default:
		/* never happens */
		BUG();
		result = -EINVAL;	/* shut gcc up */
	};
	xfer->is_inbound = urb->pipe & USB_DIR_IN ? 1 : 0;
	xfer->is_dma = urb->transfer_flags & URB_NO_TRANSFER_DMA_MAP ? 1 : 0;
	xfer->seg_size = le16_to_cpu(rpipe->descr.wBlocks)
		* 1 << (xfer->wa->wa_descr->bRPipeBlockSize - 1);
	/* Compute the segment size and make sure it is a multiple of
	 * the maxpktsize (WUSB1.0[8.3.3.1])...not really too much of
	 * a check (FIXME) */
	maxpktsize = le16_to_cpu(rpipe->descr.wMaxPacketSize);
	if (xfer->seg_size < maxpktsize) {
		dev_err(dev, "HW BUG? seg_size %zu smaller than maxpktsize "
			"%zu\n", xfer->seg_size, maxpktsize);
		result = -EINVAL;
		goto error;
	}
	xfer->seg_size = (xfer->seg_size / maxpktsize) * maxpktsize;
	xfer->segs = (urb->transfer_buffer_length + xfer->seg_size - 1)
		/ xfer->seg_size;
	if (xfer->segs >= WA_SEGS_MAX) {
		dev_err(dev, "BUG? ops, number of segments %d bigger than %d\n",
			(int)(urb->transfer_buffer_length / xfer->seg_size),
			WA_SEGS_MAX);
		result = -EINVAL;
		goto error;
	}
	if (xfer->segs == 0 && *pxfer_type == WA_XFER_TYPE_CTL)
		xfer->segs = 1;
error:
	return result;
}

/* Fill in the common request header and xfer-type specific data. */
static void __wa_xfer_setup_hdr0(struct wa_xfer *xfer,
				 struct wa_xfer_hdr *xfer_hdr0,
				 enum wa_xfer_type xfer_type,
				 size_t xfer_hdr_size)
{
	struct wa_rpipe *rpipe = xfer->ep->hcpriv;

	xfer_hdr0 = &xfer->seg[0]->xfer_hdr;
	xfer_hdr0->bLength = xfer_hdr_size;
	xfer_hdr0->bRequestType = xfer_type;
	xfer_hdr0->wRPipe = rpipe->descr.wRPipeIndex;
	xfer_hdr0->dwTransferID = wa_xfer_id(xfer);
	xfer_hdr0->bTransferSegment = 0;
	switch (xfer_type) {
	case WA_XFER_TYPE_CTL: {
		struct wa_xfer_ctl *xfer_ctl =
			container_of(xfer_hdr0, struct wa_xfer_ctl, hdr);
		xfer_ctl->bmAttribute = xfer->is_inbound ? 1 : 0;
		memcpy(&xfer_ctl->baSetupData, xfer->urb->setup_packet,
		       sizeof(xfer_ctl->baSetupData));
		break;
	}
	case WA_XFER_TYPE_BI:
		break;
	case WA_XFER_TYPE_ISO:
		printk(KERN_ERR "FIXME: ISOC not implemented\n");
	default:
		BUG();
	};
}

/*
 * Callback for the OUT data phase of the segment request
 *
 * Check wa_seg_cb(); most comments also apply here because this
 * function does almost the same thing and they work closely
 * together.
 *
 * If the seg request has failed but this DTO phase has succeeded,
 * wa_seg_cb() has already failed the segment and moved the
 * status to WA_SEG_ERROR, so this will go through 'case 0' and
 * effectively do nothing.
 */
static void wa_seg_dto_cb(struct urb *urb)
{
	struct wa_seg *seg = urb->context;
	struct wa_xfer *xfer = seg->xfer;
	struct wahc *wa;
	struct device *dev;
	struct wa_rpipe *rpipe;
	unsigned long flags;
	unsigned rpipe_ready = 0;
	u8 done = 0;

	switch (urb->status) {
	case 0:
		spin_lock_irqsave(&xfer->lock, flags);
		wa = xfer->wa;
		dev = &wa->usb_iface->dev;
		dev_dbg(dev, "xfer %p#%u: data out done (%d bytes)\n",
			xfer, seg->index, urb->actual_length);
		if (seg->status < WA_SEG_PENDING)
			seg->status = WA_SEG_PENDING;
		seg->result = urb->actual_length;
		spin_unlock_irqrestore(&xfer->lock, flags);
		break;
	case -ECONNRESET:	/* URB unlinked; no need to do anything */
	case -ENOENT:		/* as it was done by the who unlinked us */
		break;
	default:		/* Other errors ... */
		spin_lock_irqsave(&xfer->lock, flags);
		wa = xfer->wa;
		dev = &wa->usb_iface->dev;
		rpipe = xfer->ep->hcpriv;
		dev_dbg(dev, "xfer %p#%u: data out error %d\n",
			xfer, seg->index, urb->status);
		if (edc_inc(&wa->nep_edc, EDC_MAX_ERRORS,
			    EDC_ERROR_TIMEFRAME)){
			dev_err(dev, "DTO: URB max acceptable errors "
				"exceeded, resetting device\n");
			wa_reset_all(wa);
		}
		if (seg->status != WA_SEG_ERROR) {
			seg->status = WA_SEG_ERROR;
			seg->result = urb->status;
			xfer->segs_done++;
			__wa_xfer_abort(xfer);
			rpipe_ready = rpipe_avail_inc(rpipe);
			done = __wa_xfer_is_done(xfer);
		}
		spin_unlock_irqrestore(&xfer->lock, flags);
		if (done)
			wa_xfer_completion(xfer);
		if (rpipe_ready)
			wa_xfer_delayed_run(rpipe);
	}
}

/*
 * Callback for the segment request
 *
 * If successful transition state (unless already transitioned or
 * outbound transfer); otherwise, take a note of the error, mark this
 * segment done and try completion.
 *
 * Note we don't access until we are sure that the transfer hasn't
 * been cancelled (ECONNRESET, ENOENT), which could mean that
 * seg->xfer could be already gone.
 *
 * We have to check before setting the status to WA_SEG_PENDING
 * because sometimes the xfer result callback arrives before this
 * callback (geeeeeeze), so it might happen that we are already in
 * another state. As well, we don't set it if the transfer is inbound,
 * as in that case, wa_seg_dto_cb will do it when the OUT data phase
 * finishes.
 */
static void wa_seg_cb(struct urb *urb)
{
	struct wa_seg *seg = urb->context;
	struct wa_xfer *xfer = seg->xfer;
	struct wahc *wa;
	struct device *dev;
	struct wa_rpipe *rpipe;
	unsigned long flags;
	unsigned rpipe_ready;
	u8 done = 0;

	switch (urb->status) {
	case 0:
		spin_lock_irqsave(&xfer->lock, flags);
		wa = xfer->wa;
		dev = &wa->usb_iface->dev;
		dev_dbg(dev, "xfer %p#%u: request done\n", xfer, seg->index);
		if (xfer->is_inbound && seg->status < WA_SEG_PENDING)
			seg->status = WA_SEG_PENDING;
		spin_unlock_irqrestore(&xfer->lock, flags);
		break;
	case -ECONNRESET:	/* URB unlinked; no need to do anything */
	case -ENOENT:		/* as it was done by the who unlinked us */
		break;
	default:		/* Other errors ... */
		spin_lock_irqsave(&xfer->lock, flags);
		wa = xfer->wa;
		dev = &wa->usb_iface->dev;
		rpipe = xfer->ep->hcpriv;
		if (printk_ratelimit())
			dev_err(dev, "xfer %p#%u: request error %d\n",
				xfer, seg->index, urb->status);
		if (edc_inc(&wa->nep_edc, EDC_MAX_ERRORS,
			    EDC_ERROR_TIMEFRAME)){
			dev_err(dev, "DTO: URB max acceptable errors "
				"exceeded, resetting device\n");
			wa_reset_all(wa);
		}
		usb_unlink_urb(seg->dto_urb);
		seg->status = WA_SEG_ERROR;
		seg->result = urb->status;
		xfer->segs_done++;
		__wa_xfer_abort(xfer);
		rpipe_ready = rpipe_avail_inc(rpipe);
		done = __wa_xfer_is_done(xfer);
		spin_unlock_irqrestore(&xfer->lock, flags);
		if (done)
			wa_xfer_completion(xfer);
		if (rpipe_ready)
			wa_xfer_delayed_run(rpipe);
	}
}

/*
 * Allocate the segs array and initialize each of them
 *
 * The segments are freed by wa_xfer_destroy() when the xfer use count
 * drops to zero; however, because each segment is given the same life
 * cycle as the USB URB it contains, it is actually freed by
 * usb_put_urb() on the contained USB URB (twisted, eh?).
 */
static int __wa_xfer_setup_segs(struct wa_xfer *xfer, size_t xfer_hdr_size)
{
	int result, cnt;
	size_t alloc_size = sizeof(*xfer->seg[0])
		- sizeof(xfer->seg[0]->xfer_hdr) + xfer_hdr_size;
	struct usb_device *usb_dev = xfer->wa->usb_dev;
	const struct usb_endpoint_descriptor *dto_epd = xfer->wa->dto_epd;
	struct wa_seg *seg;
	size_t buf_itr, buf_size, buf_itr_size;

	result = -ENOMEM;
	xfer->seg = kcalloc(xfer->segs, sizeof(xfer->seg[0]), GFP_ATOMIC);
	if (xfer->seg == NULL)
		goto error_segs_kzalloc;
	buf_itr = 0;
	buf_size = xfer->urb->transfer_buffer_length;
	for (cnt = 0; cnt < xfer->segs; cnt++) {
		seg = xfer->seg[cnt] = kzalloc(alloc_size, GFP_ATOMIC);
		if (seg == NULL)
			goto error_seg_kzalloc;
		wa_seg_init(seg);
		seg->xfer = xfer;
		seg->index = cnt;
		usb_fill_bulk_urb(&seg->urb, usb_dev,
				  usb_sndbulkpipe(usb_dev,
						  dto_epd->bEndpointAddress),
				  &seg->xfer_hdr, xfer_hdr_size,
				  wa_seg_cb, seg);
		buf_itr_size = buf_size > xfer->seg_size ?
			xfer->seg_size : buf_size;
		if (xfer->is_inbound == 0 && buf_size > 0) {
			seg->dto_urb = usb_alloc_urb(0, GFP_ATOMIC);
			if (seg->dto_urb == NULL)
				goto error_dto_alloc;
			usb_fill_bulk_urb(
				seg->dto_urb, usb_dev,
				usb_sndbulkpipe(usb_dev,
						dto_epd->bEndpointAddress),
				NULL, 0, wa_seg_dto_cb, seg);
			if (xfer->is_dma) {
				seg->dto_urb->transfer_dma =
					xfer->urb->transfer_dma + buf_itr;
				seg->dto_urb->transfer_flags |=
					URB_NO_TRANSFER_DMA_MAP;
			} else
				seg->dto_urb->transfer_buffer =
					xfer->urb->transfer_buffer + buf_itr;
			seg->dto_urb->transfer_buffer_length = buf_itr_size;
		}
		seg->status = WA_SEG_READY;
		buf_itr += buf_itr_size;
		buf_size -= buf_itr_size;
	}
	return 0;

error_dto_alloc:
	kfree(xfer->seg[cnt]);
	cnt--;
error_seg_kzalloc:
	/* use the fact that cnt is left at were it failed */
	for (; cnt >= 0; cnt--) {
		if (xfer->seg[cnt] && xfer->is_inbound == 0)
			usb_free_urb(xfer->seg[cnt]->dto_urb);
		kfree(xfer->seg[cnt]);
	}
error_segs_kzalloc:
	return result;
}

/*
 * Allocates all the stuff needed to submit a transfer
 *
 * Breaks the whole data buffer in a list of segments, each one has a
 * structure allocated to it and linked in xfer->seg[index]
 *
 * FIXME: merge setup_segs() and the last part of this function, no
 *        need to do two for loops when we could run everything in a
 *        single one
 */
static int __wa_xfer_setup(struct wa_xfer *xfer, struct urb *urb)
{
	int result;
	struct device *dev = &xfer->wa->usb_iface->dev;
	enum wa_xfer_type xfer_type = 0; /* shut up GCC */
	size_t xfer_hdr_size, cnt, transfer_size;
	struct wa_xfer_hdr *xfer_hdr0, *xfer_hdr;

	result = __wa_xfer_setup_sizes(xfer, &xfer_type);
	if (result < 0)
		goto error_setup_sizes;
	xfer_hdr_size = result;
	result = __wa_xfer_setup_segs(xfer, xfer_hdr_size);
	if (result < 0) {
		dev_err(dev, "xfer %p: Failed to allocate %d segments: %d\n",
			xfer, xfer->segs, result);
		goto error_setup_segs;
	}
	/* Fill the first header */
	xfer_hdr0 = &xfer->seg[0]->xfer_hdr;
	wa_xfer_id_init(xfer);
	__wa_xfer_setup_hdr0(xfer, xfer_hdr0, xfer_type, xfer_hdr_size);

	/* Fill remainig headers */
	xfer_hdr = xfer_hdr0;
	transfer_size = urb->transfer_buffer_length;
	xfer_hdr0->dwTransferLength = transfer_size > xfer->seg_size ?
		xfer->seg_size : transfer_size;
	transfer_size -=  xfer->seg_size;
	for (cnt = 1; cnt < xfer->segs; cnt++) {
		xfer_hdr = &xfer->seg[cnt]->xfer_hdr;
		memcpy(xfer_hdr, xfer_hdr0, xfer_hdr_size);
		xfer_hdr->bTransferSegment = cnt;
		xfer_hdr->dwTransferLength = transfer_size > xfer->seg_size ?
			cpu_to_le32(xfer->seg_size)
			: cpu_to_le32(transfer_size);
		xfer->seg[cnt]->status = WA_SEG_READY;
		transfer_size -=  xfer->seg_size;
	}
	xfer_hdr->bTransferSegment |= 0x80;	/* this is the last segment */
	result = 0;
error_setup_segs:
error_setup_sizes:
	return result;
}

/*
 *
 *
 * rpipe->seg_lock is held!
 */
static int __wa_seg_submit(struct wa_rpipe *rpipe, struct wa_xfer *xfer,
			   struct wa_seg *seg)
{
	int result;
	result = usb_submit_urb(&seg->urb, GFP_ATOMIC);
	if (result < 0) {
		printk(KERN_ERR "xfer %p#%u: REQ submit failed: %d\n",
		       xfer, seg->index, result);
		goto error_seg_submit;
	}
	if (seg->dto_urb) {
		result = usb_submit_urb(seg->dto_urb, GFP_ATOMIC);
		if (result < 0) {
			printk(KERN_ERR "xfer %p#%u: DTO submit failed: %d\n",
			       xfer, seg->index, result);
			goto error_dto_submit;
		}
	}
	seg->status = WA_SEG_SUBMITTED;
	rpipe_avail_dec(rpipe);
	return 0;

error_dto_submit:
	usb_unlink_urb(&seg->urb);
error_seg_submit:
	seg->status = WA_SEG_ERROR;
	seg->result = result;
	return result;
}

/*
 * Execute more queued request segments until the maximum concurrent allowed
 *
 * The ugly unlock/lock sequence on the error path is needed as the
 * xfer->lock normally nests the seg_lock and not viceversa.
 *
 */
static void wa_xfer_delayed_run(struct wa_rpipe *rpipe)
{
	int result;
	struct device *dev = &rpipe->wa->usb_iface->dev;
	struct wa_seg *seg;
	struct wa_xfer *xfer;
	unsigned long flags;

	spin_lock_irqsave(&rpipe->seg_lock, flags);
	while (atomic_read(&rpipe->segs_available) > 0
	      && !list_empty(&rpipe->seg_list)) {
		seg = list_entry(rpipe->seg_list.next, struct wa_seg,
				 list_node);
		list_del(&seg->list_node);
		xfer = seg->xfer;
		result = __wa_seg_submit(rpipe, xfer, seg);
		dev_dbg(dev, "xfer %p#%u submitted from delayed [%d segments available] %d\n",
			xfer, seg->index, atomic_read(&rpipe->segs_available), result);
		if (unlikely(result < 0)) {
			spin_unlock_irqrestore(&rpipe->seg_lock, flags);
			spin_lock_irqsave(&xfer->lock, flags);
			__wa_xfer_abort(xfer);
			xfer->segs_done++;
			spin_unlock_irqrestore(&xfer->lock, flags);
			spin_lock_irqsave(&rpipe->seg_lock, flags);
		}
	}
	spin_unlock_irqrestore(&rpipe->seg_lock, flags);
}

/*
 *
 * xfer->lock is taken
 *
 * On failure submitting we just stop submitting and return error;
 * wa_urb_enqueue_b() will execute the completion path
 */
static int __wa_xfer_submit(struct wa_xfer *xfer)
{
	int result;
	struct wahc *wa = xfer->wa;
	struct device *dev = &wa->usb_iface->dev;
	unsigned cnt;
	struct wa_seg *seg;
	unsigned long flags;
	struct wa_rpipe *rpipe = xfer->ep->hcpriv;
	size_t maxrequests = le16_to_cpu(rpipe->descr.wRequests);
	u8 available;
	u8 empty;

	spin_lock_irqsave(&wa->xfer_list_lock, flags);
	list_add_tail(&xfer->list_node, &wa->xfer_list);
	spin_unlock_irqrestore(&wa->xfer_list_lock, flags);

	BUG_ON(atomic_read(&rpipe->segs_available) > maxrequests);
	result = 0;
	spin_lock_irqsave(&rpipe->seg_lock, flags);
	for (cnt = 0; cnt < xfer->segs; cnt++) {
		available = atomic_read(&rpipe->segs_available);
		empty = list_empty(&rpipe->seg_list);
		seg = xfer->seg[cnt];
		dev_dbg(dev, "xfer %p#%u: available %u empty %u (%s)\n",
			xfer, cnt, available, empty,
			available == 0 || !empty ? "delayed" : "submitted");
		if (available == 0 || !empty) {
			dev_dbg(dev, "xfer %p#%u: delayed\n", xfer, cnt);
			seg->status = WA_SEG_DELAYED;
			list_add_tail(&seg->list_node, &rpipe->seg_list);
		} else {
			result = __wa_seg_submit(rpipe, xfer, seg);
			if (result < 0) {
				__wa_xfer_abort(xfer);
				goto error_seg_submit;
			}
		}
		xfer->segs_submitted++;
	}
error_seg_submit:
	spin_unlock_irqrestore(&rpipe->seg_lock, flags);
	return result;
}

/*
 * Second part of a URB/transfer enqueuement
 *
 * Assumes this comes from wa_urb_enqueue() [maybe through
 * wa_urb_enqueue_run()]. At this point:
 *
 * xfer->wa	filled and refcounted
 * xfer->ep	filled with rpipe refcounted if
 *              delayed == 0
 * xfer->urb 	filled and refcounted (this is the case when called
 *              from wa_urb_enqueue() as we come from usb_submit_urb()
 *              and when called by wa_urb_enqueue_run(), as we took an
 *              extra ref dropped by _run() after we return).
 * xfer->gfp	filled
 *
 * If we fail at __wa_xfer_submit(), then we just check if we are done
 * and if so, we run the completion procedure. However, if we are not
 * yet done, we do nothing and wait for the completion handlers from
 * the submitted URBs or from the xfer-result path to kick in. If xfer
 * result never kicks in, the xfer will timeout from the USB code and
 * dequeue() will be called.
 */
static void wa_urb_enqueue_b(struct wa_xfer *xfer)
{
	int result;
	unsigned long flags;
	struct urb *urb = xfer->urb;
	struct wahc *wa = xfer->wa;
	struct wusbhc *wusbhc = wa->wusb;
	struct wusb_dev *wusb_dev;
	unsigned done;

	result = rpipe_get_by_ep(wa, xfer->ep, urb, xfer->gfp);
	if (result < 0)
		goto error_rpipe_get;
	result = -ENODEV;
	/* FIXME: segmentation broken -- kills DWA */
	mutex_lock(&wusbhc->mutex);		/* get a WUSB dev */
	if (urb->dev == NULL) {
		mutex_unlock(&wusbhc->mutex);
		goto error_dev_gone;
	}
	wusb_dev = __wusb_dev_get_by_usb_dev(wusbhc, urb->dev);
	if (wusb_dev == NULL) {
		mutex_unlock(&wusbhc->mutex);
		goto error_dev_gone;
	}
	mutex_unlock(&wusbhc->mutex);

	spin_lock_irqsave(&xfer->lock, flags);
	xfer->wusb_dev = wusb_dev;
	result = urb->status;
	if (urb->status != -EINPROGRESS)
		goto error_dequeued;

	result = __wa_xfer_setup(xfer, urb);
	if (result < 0)
		goto error_xfer_setup;
	result = __wa_xfer_submit(xfer);
	if (result < 0)
		goto error_xfer_submit;
	spin_unlock_irqrestore(&xfer->lock, flags);
	return;

	/* this is basically wa_xfer_completion() broken up wa_xfer_giveback()
	 * does a wa_xfer_put() that will call wa_xfer_destroy() and clean
	 * upundo setup().
	 */
error_xfer_setup:
error_dequeued:
	spin_unlock_irqrestore(&xfer->lock, flags);
	/* FIXME: segmentation broken, kills DWA */
	if (wusb_dev)
		wusb_dev_put(wusb_dev);
error_dev_gone:
	rpipe_put(xfer->ep->hcpriv);
error_rpipe_get:
	xfer->result = result;
	wa_xfer_giveback(xfer);
	return;

error_xfer_submit:
	done = __wa_xfer_is_done(xfer);
	xfer->result = result;
	spin_unlock_irqrestore(&xfer->lock, flags);
	if (done)
		wa_xfer_completion(xfer);
}

/*
 * Execute the delayed transfers in the Wire Adapter @wa
 *
 * We need to be careful here, as dequeue() could be called in the
 * middle.  That's why we do the whole thing under the
 * wa->xfer_list_lock. If dequeue() jumps in, it first locks urb->lock
 * and then checks the list -- so as we would be acquiring in inverse
 * order, we just drop the lock once we have the xfer and reacquire it
 * later.
 */
void wa_urb_enqueue_run(struct work_struct *ws)
{
	struct wahc *wa = container_of(ws, struct wahc, xfer_work);
	struct wa_xfer *xfer, *next;
	struct urb *urb;

	spin_lock_irq(&wa->xfer_list_lock);
	list_for_each_entry_safe(xfer, next, &wa->xfer_delayed_list,
				 list_node) {
		list_del_init(&xfer->list_node);
		spin_unlock_irq(&wa->xfer_list_lock);

		urb = xfer->urb;
		wa_urb_enqueue_b(xfer);
		usb_put_urb(urb);	/* taken when queuing */

		spin_lock_irq(&wa->xfer_list_lock);
	}
	spin_unlock_irq(&wa->xfer_list_lock);
}
EXPORT_SYMBOL_GPL(wa_urb_enqueue_run);

/*
 * Submit a transfer to the Wire Adapter in a delayed way
 *
 * The process of enqueuing involves possible sleeps() [see
 * enqueue_b(), for the rpipe_get() and the mutex_lock()]. If we are
 * in an atomic section, we defer the enqueue_b() call--else we call direct.
 *
 * @urb: We own a reference to it done by the HCI Linux USB stack that
 *       will be given up by calling usb_hcd_giveback_urb() or by
 *       returning error from this function -> ergo we don't have to
 *       refcount it.
 */
int wa_urb_enqueue(struct wahc *wa, struct usb_host_endpoint *ep,
		   struct urb *urb, gfp_t gfp)
{
	int result;
	struct device *dev = &wa->usb_iface->dev;
	struct wa_xfer *xfer;
	unsigned long my_flags;
	unsigned cant_sleep = irqs_disabled() | in_atomic();

	if (urb->transfer_buffer == NULL
	    && !(urb->transfer_flags & URB_NO_TRANSFER_DMA_MAP)
	    && urb->transfer_buffer_length != 0) {
		dev_err(dev, "BUG? urb %p: NULL xfer buffer & NODMA\n", urb);
		dump_stack();
	}

	result = -ENOMEM;
	xfer = kzalloc(sizeof(*xfer), gfp);
	if (xfer == NULL)
		goto error_kmalloc;

	result = -ENOENT;
	if (urb->status != -EINPROGRESS)	/* cancelled */
		goto error_dequeued;		/* before starting? */
	wa_xfer_init(xfer);
	xfer->wa = wa_get(wa);
	xfer->urb = urb;
	xfer->gfp = gfp;
	xfer->ep = ep;
	urb->hcpriv = xfer;

	dev_dbg(dev, "xfer %p urb %p pipe 0x%02x [%d bytes] %s %s %s\n",
		xfer, urb, urb->pipe, urb->transfer_buffer_length,
		urb->transfer_flags & URB_NO_TRANSFER_DMA_MAP ? "dma" : "nodma",
		urb->pipe & USB_DIR_IN ? "inbound" : "outbound",
		cant_sleep ? "deferred" : "inline");

	if (cant_sleep) {
		usb_get_urb(urb);
		spin_lock_irqsave(&wa->xfer_list_lock, my_flags);
		list_add_tail(&xfer->list_node, &wa->xfer_delayed_list);
		spin_unlock_irqrestore(&wa->xfer_list_lock, my_flags);
		queue_work(wusbd, &wa->xfer_work);
	} else {
		wa_urb_enqueue_b(xfer);
	}
	return 0;

error_dequeued:
	kfree(xfer);
error_kmalloc:
	return result;
}
EXPORT_SYMBOL_GPL(wa_urb_enqueue);

/*
 * Dequeue a URB and make sure uwb_hcd_giveback_urb() [completion
 * handler] is called.
 *
 * Until a transfer goes successfully through wa_urb_enqueue() it
 * needs to be dequeued with completion calling; when stuck in delayed
 * or before wa_xfer_setup() is called, we need to do completion.
 *
 *  not setup  If there is no hcpriv yet, that means that that enqueue
 *             still had no time to set the xfer up. Because
 *             urb->status should be other than -EINPROGRESS,
 *             enqueue() will catch that and bail out.
 *
 * If the transfer has gone through setup, we just need to clean it
 * up. If it has gone through submit(), we have to abort it [with an
 * asynch request] and then make sure we cancel each segment.
 *
 */
int wa_urb_dequeue(struct wahc *wa, struct urb *urb)
{
	unsigned long flags, flags2;
	struct wa_xfer *xfer;
	struct wa_seg *seg;
	struct wa_rpipe *rpipe;
	unsigned cnt;
	unsigned rpipe_ready = 0;

	xfer = urb->hcpriv;
	if (xfer == NULL) {
		/* NOthing setup yet enqueue will see urb->status !=
		 * -EINPROGRESS (by hcd layer) and bail out with
		 * error, no need to do completion
		 */
		BUG_ON(urb->status == -EINPROGRESS);
		goto out;
	}
	spin_lock_irqsave(&xfer->lock, flags);
	rpipe = xfer->ep->hcpriv;
	/* Check the delayed list -> if there, release and complete */
	spin_lock_irqsave(&wa->xfer_list_lock, flags2);
	if (!list_empty(&xfer->list_node) && xfer->seg == NULL)
		goto dequeue_delayed;
	spin_unlock_irqrestore(&wa->xfer_list_lock, flags2);
	if (xfer->seg == NULL)  	/* still hasn't reached */
		goto out_unlock;	/* setup(), enqueue_b() completes */
	/* Ok, the xfer is in flight already, it's been setup and submitted.*/
	__wa_xfer_abort(xfer);
	for (cnt = 0; cnt < xfer->segs; cnt++) {
		seg = xfer->seg[cnt];
		switch (seg->status) {
		case WA_SEG_NOTREADY:
		case WA_SEG_READY:
			printk(KERN_ERR "xfer %p#%u: dequeue bad state %u\n",
			       xfer, cnt, seg->status);
			WARN_ON(1);
			break;
		case WA_SEG_DELAYED:
			seg->status = WA_SEG_ABORTED;
			spin_lock_irqsave(&rpipe->seg_lock, flags2);
			list_del(&seg->list_node);
			xfer->segs_done++;
			rpipe_ready = rpipe_avail_inc(rpipe);
			spin_unlock_irqrestore(&rpipe->seg_lock, flags2);
			break;
		case WA_SEG_SUBMITTED:
			seg->status = WA_SEG_ABORTED;
			usb_unlink_urb(&seg->urb);
			if (xfer->is_inbound == 0)
				usb_unlink_urb(seg->dto_urb);
			xfer->segs_done++;
			rpipe_ready = rpipe_avail_inc(rpipe);
			break;
		case WA_SEG_PENDING:
			seg->status = WA_SEG_ABORTED;
			xfer->segs_done++;
			rpipe_ready = rpipe_avail_inc(rpipe);
			break;
		case WA_SEG_DTI_PENDING:
			usb_unlink_urb(wa->dti_urb);
			seg->status = WA_SEG_ABORTED;
			xfer->segs_done++;
			rpipe_ready = rpipe_avail_inc(rpipe);
			break;
		case WA_SEG_DONE:
		case WA_SEG_ERROR:
		case WA_SEG_ABORTED:
			break;
		}
	}
	xfer->result = urb->status;	/* -ENOENT or -ECONNRESET */
	__wa_xfer_is_done(xfer);
	spin_unlock_irqrestore(&xfer->lock, flags);
	wa_xfer_completion(xfer);
	if (rpipe_ready)
		wa_xfer_delayed_run(rpipe);
	return 0;

out_unlock:
	spin_unlock_irqrestore(&xfer->lock, flags);
out:
	return 0;

dequeue_delayed:
	list_del_init(&xfer->list_node);
	spin_unlock_irqrestore(&wa->xfer_list_lock, flags2);
	xfer->result = urb->status;
	spin_unlock_irqrestore(&xfer->lock, flags);
	wa_xfer_giveback(xfer);
	usb_put_urb(urb);		/* we got a ref in enqueue() */
	return 0;
}
EXPORT_SYMBOL_GPL(wa_urb_dequeue);

/*
 * Translation from WA status codes (WUSB1.0 Table 8.15) to errno
 * codes
 *
 * Positive errno values are internal inconsistencies and should be
 * flagged louder. Negative are to be passed up to the user in the
 * normal way.
 *
 * @status: USB WA status code -- high two bits are stripped.
 */
static int wa_xfer_status_to_errno(u8 status)
{
	int errno;
	u8 real_status = status;
	static int xlat[] = {
		[WA_XFER_STATUS_SUCCESS] = 		0,
		[WA_XFER_STATUS_HALTED] = 		-EPIPE,
		[WA_XFER_STATUS_DATA_BUFFER_ERROR] = 	-ENOBUFS,
		[WA_XFER_STATUS_BABBLE] = 		-EOVERFLOW,
		[WA_XFER_RESERVED] = 			EINVAL,
		[WA_XFER_STATUS_NOT_FOUND] =		0,
		[WA_XFER_STATUS_INSUFFICIENT_RESOURCE] = -ENOMEM,
		[WA_XFER_STATUS_TRANSACTION_ERROR] = 	-EILSEQ,
		[WA_XFER_STATUS_ABORTED] = 		-EINTR,
		[WA_XFER_STATUS_RPIPE_NOT_READY] = 	EINVAL,
		[WA_XFER_INVALID_FORMAT] = 		EINVAL,
		[WA_XFER_UNEXPECTED_SEGMENT_NUMBER] = 	EINVAL,
		[WA_XFER_STATUS_RPIPE_TYPE_MISMATCH] = 	EINVAL,
	};
	status &= 0x3f;

	if (status == 0)
		return 0;
	if (status >= ARRAY_SIZE(xlat)) {
		printk_ratelimited(KERN_ERR "%s(): BUG? "
			       "Unknown WA transfer status 0x%02x\n",
			       __func__, real_status);
		return -EINVAL;
	}
	errno = xlat[status];
	if (unlikely(errno > 0)) {
		printk_ratelimited(KERN_ERR "%s(): BUG? "
			       "Inconsistent WA status: 0x%02x\n",
			       __func__, real_status);
		errno = -errno;
	}
	return errno;
}

/*
 * Process a xfer result completion message
 *
 * inbound transfers: need to schedule a DTI read
 *
 * FIXME: this functio needs to be broken up in parts
 */
static void wa_xfer_result_chew(struct wahc *wa, struct wa_xfer *xfer)
{
	int result;
	struct device *dev = &wa->usb_iface->dev;
	unsigned long flags;
	u8 seg_idx;
	struct wa_seg *seg;
	struct wa_rpipe *rpipe;
	struct wa_xfer_result *xfer_result = wa->xfer_result;
	u8 done = 0;
	u8 usb_status;
	unsigned rpipe_ready = 0;

	spin_lock_irqsave(&xfer->lock, flags);
	seg_idx = xfer_result->bTransferSegment & 0x7f;
	if (unlikely(seg_idx >= xfer->segs))
		goto error_bad_seg;
	seg = xfer->seg[seg_idx];
	rpipe = xfer->ep->hcpriv;
	usb_status = xfer_result->bTransferStatus;
	dev_dbg(dev, "xfer %p#%u: bTransferStatus 0x%02x (seg %u)\n",
		xfer, seg_idx, usb_status, seg->status);
	if (seg->status == WA_SEG_ABORTED
	    || seg->status == WA_SEG_ERROR)	/* already handled */
		goto segment_aborted;
	if (seg->status == WA_SEG_SUBMITTED)	/* ops, got here */
		seg->status = WA_SEG_PENDING;	/* before wa_seg{_dto}_cb() */
	if (seg->status != WA_SEG_PENDING) {
		if (printk_ratelimit())
			dev_err(dev, "xfer %p#%u: Bad segment state %u\n",
				xfer, seg_idx, seg->status);
		seg->status = WA_SEG_PENDING;	/* workaround/"fix" it */
	}
	if (usb_status & 0x80) {
		seg->result = wa_xfer_status_to_errno(usb_status);
		dev_err(dev, "DTI: xfer %p#%u failed (0x%02x)\n",
			xfer, seg->index, usb_status);
		goto error_complete;
	}
	/* FIXME: we ignore warnings, tally them for stats */
	if (usb_status & 0x40) 		/* Warning?... */
		usb_status = 0;		/* ... pass */
	if (xfer->is_inbound) {	/* IN data phase: read to buffer */
		seg->status = WA_SEG_DTI_PENDING;
		BUG_ON(wa->buf_in_urb->status == -EINPROGRESS);
		if (xfer->is_dma) {
			wa->buf_in_urb->transfer_dma =
				xfer->urb->transfer_dma
				+ seg_idx * xfer->seg_size;
			wa->buf_in_urb->transfer_flags
				|= URB_NO_TRANSFER_DMA_MAP;
		} else {
			wa->buf_in_urb->transfer_buffer =
				xfer->urb->transfer_buffer
				+ seg_idx * xfer->seg_size;
			wa->buf_in_urb->transfer_flags
				&= ~URB_NO_TRANSFER_DMA_MAP;
		}
		wa->buf_in_urb->transfer_buffer_length =
			le32_to_cpu(xfer_result->dwTransferLength);
		wa->buf_in_urb->context = seg;
		result = usb_submit_urb(wa->buf_in_urb, GFP_ATOMIC);
		if (result < 0)
			goto error_submit_buf_in;
	} else {
		/* OUT data phase, complete it -- */
		seg->status = WA_SEG_DONE;
		seg->result = le32_to_cpu(xfer_result->dwTransferLength);
		xfer->segs_done++;
		rpipe_ready = rpipe_avail_inc(rpipe);
		done = __wa_xfer_is_done(xfer);
	}
	spin_unlock_irqrestore(&xfer->lock, flags);
	if (done)
		wa_xfer_completion(xfer);
	if (rpipe_ready)
		wa_xfer_delayed_run(rpipe);
	return;

error_submit_buf_in:
	if (edc_inc(&wa->dti_edc, EDC_MAX_ERRORS, EDC_ERROR_TIMEFRAME)) {
		dev_err(dev, "DTI: URB max acceptable errors "
			"exceeded, resetting device\n");
		wa_reset_all(wa);
	}
	if (printk_ratelimit())
		dev_err(dev, "xfer %p#%u: can't submit DTI data phase: %d\n",
			xfer, seg_idx, result);
	seg->result = result;
error_complete:
	seg->status = WA_SEG_ERROR;
	xfer->segs_done++;
	rpipe_ready = rpipe_avail_inc(rpipe);
	__wa_xfer_abort(xfer);
	done = __wa_xfer_is_done(xfer);
	spin_unlock_irqrestore(&xfer->lock, flags);
	if (done)
		wa_xfer_completion(xfer);
	if (rpipe_ready)
		wa_xfer_delayed_run(rpipe);
	return;

error_bad_seg:
	spin_unlock_irqrestore(&xfer->lock, flags);
	wa_urb_dequeue(wa, xfer->urb);
	if (printk_ratelimit())
		dev_err(dev, "xfer %p#%u: bad segment\n", xfer, seg_idx);
	if (edc_inc(&wa->dti_edc, EDC_MAX_ERRORS, EDC_ERROR_TIMEFRAME)) {
		dev_err(dev, "DTI: URB max acceptable errors "
			"exceeded, resetting device\n");
		wa_reset_all(wa);
	}
	return;

segment_aborted:
	/* nothing to do, as the aborter did the completion */
	spin_unlock_irqrestore(&xfer->lock, flags);
}

/*
 * Callback for the IN data phase
 *
 * If successful transition state; otherwise, take a note of the
 * error, mark this segment done and try completion.
 *
 * Note we don't access until we are sure that the transfer hasn't
 * been cancelled (ECONNRESET, ENOENT), which could mean that
 * seg->xfer could be already gone.
 */
static void wa_buf_in_cb(struct urb *urb)
{
	struct wa_seg *seg = urb->context;
	struct wa_xfer *xfer = seg->xfer;
	struct wahc *wa;
	struct device *dev;
	struct wa_rpipe *rpipe;
	unsigned rpipe_ready;
	unsigned long flags;
	u8 done = 0;

	switch (urb->status) {
	case 0:
		spin_lock_irqsave(&xfer->lock, flags);
		wa = xfer->wa;
		dev = &wa->usb_iface->dev;
		rpipe = xfer->ep->hcpriv;
		dev_dbg(dev, "xfer %p#%u: data in done (%zu bytes)\n",
			xfer, seg->index, (size_t)urb->actual_length);
		seg->status = WA_SEG_DONE;
		seg->result = urb->actual_length;
		xfer->segs_done++;
		rpipe_ready = rpipe_avail_inc(rpipe);
		done = __wa_xfer_is_done(xfer);
		spin_unlock_irqrestore(&xfer->lock, flags);
		if (done)
			wa_xfer_completion(xfer);
		if (rpipe_ready)
			wa_xfer_delayed_run(rpipe);
		break;
	case -ECONNRESET:	/* URB unlinked; no need to do anything */
	case -ENOENT:		/* as it was done by the who unlinked us */
		break;
	default:		/* Other errors ... */
		spin_lock_irqsave(&xfer->lock, flags);
		wa = xfer->wa;
		dev = &wa->usb_iface->dev;
		rpipe = xfer->ep->hcpriv;
		if (printk_ratelimit())
			dev_err(dev, "xfer %p#%u: data in error %d\n",
				xfer, seg->index, urb->status);
		if (edc_inc(&wa->nep_edc, EDC_MAX_ERRORS,
			    EDC_ERROR_TIMEFRAME)){
			dev_err(dev, "DTO: URB max acceptable errors "
				"exceeded, resetting device\n");
			wa_reset_all(wa);
		}
		seg->status = WA_SEG_ERROR;
		seg->result = urb->status;
		xfer->segs_done++;
		rpipe_ready = rpipe_avail_inc(rpipe);
		__wa_xfer_abort(xfer);
		done = __wa_xfer_is_done(xfer);
		spin_unlock_irqrestore(&xfer->lock, flags);
		if (done)
			wa_xfer_completion(xfer);
		if (rpipe_ready)
			wa_xfer_delayed_run(rpipe);
	}
}

/*
 * Handle an incoming transfer result buffer
 *
 * Given a transfer result buffer, it completes the transfer (possibly
 * scheduling and buffer in read) and then resubmits the DTI URB for a
 * new transfer result read.
 *
 *
 * The xfer_result DTI URB state machine
 *
 * States: OFF | RXR (Read-Xfer-Result) | RBI (Read-Buffer-In)
 *
 * We start in OFF mode, the first xfer_result notification [through
 * wa_handle_notif_xfer()] moves us to RXR by posting the DTI-URB to
 * read.
 *
 * We receive a buffer -- if it is not a xfer_result, we complain and
 * repost the DTI-URB. If it is a xfer_result then do the xfer seg
 * request accounting. If it is an IN segment, we move to RBI and post
 * a BUF-IN-URB to the right buffer. The BUF-IN-URB callback will
 * repost the DTI-URB and move to RXR state. if there was no IN
 * segment, it will repost the DTI-URB.
 *
 * We go back to OFF when we detect a ENOENT or ESHUTDOWN (or too many
 * errors) in the URBs.
 */
static void wa_xfer_result_cb(struct urb *urb)
{
	int result;
	struct wahc *wa = urb->context;
	struct device *dev = &wa->usb_iface->dev;
	struct wa_xfer_result *xfer_result;
	u32 xfer_id;
	struct wa_xfer *xfer;
	u8 usb_status;

	BUG_ON(wa->dti_urb != urb);
	switch (wa->dti_urb->status) {
	case 0:
		/* We have a xfer result buffer; check it */
		dev_dbg(dev, "DTI: xfer result %d bytes at %p\n",
			urb->actual_length, urb->transfer_buffer);
		if (wa->dti_urb->actual_length != sizeof(*xfer_result)) {
			dev_err(dev, "DTI Error: xfer result--bad size "
				"xfer result (%d bytes vs %zu needed)\n",
				urb->actual_length, sizeof(*xfer_result));
			break;
		}
		xfer_result = wa->xfer_result;
		if (xfer_result->hdr.bLength != sizeof(*xfer_result)) {
			dev_err(dev, "DTI Error: xfer result--"
				"bad header length %u\n",
				xfer_result->hdr.bLength);
			break;
		}
		if (xfer_result->hdr.bNotifyType != WA_XFER_RESULT) {
			dev_err(dev, "DTI Error: xfer result--"
				"bad header type 0x%02x\n",
				xfer_result->hdr.bNotifyType);
			break;
		}
		usb_status = xfer_result->bTransferStatus & 0x3f;
		if (usb_status == WA_XFER_STATUS_ABORTED
		    || usb_status == WA_XFER_STATUS_NOT_FOUND)
			/* taken care of already */
			break;
		xfer_id = xfer_result->dwTransferID;
		xfer = wa_xfer_get_by_id(wa, xfer_id);
		if (xfer == NULL) {
			/* FIXME: transaction might have been cancelled */
			dev_err(dev, "DTI Error: xfer result--"
				"unknown xfer 0x%08x (status 0x%02x)\n",
				xfer_id, usb_status);
			break;
		}
		wa_xfer_result_chew(wa, xfer);
		wa_xfer_put(xfer);
		break;
	case -ENOENT:		/* (we killed the URB)...so, no broadcast */
	case -ESHUTDOWN:	/* going away! */
		dev_dbg(dev, "DTI: going down! %d\n", urb->status);
		goto out;
	default:
		/* Unknown error */
		if (edc_inc(&wa->dti_edc, EDC_MAX_ERRORS,
			    EDC_ERROR_TIMEFRAME)) {
			dev_err(dev, "DTI: URB max acceptable errors "
				"exceeded, resetting device\n");
			wa_reset_all(wa);
			goto out;
		}
		if (printk_ratelimit())
			dev_err(dev, "DTI: URB error %d\n", urb->status);
		break;
	}
	/* Resubmit the DTI URB */
	result = usb_submit_urb(wa->dti_urb, GFP_ATOMIC);
	if (result < 0) {
		dev_err(dev, "DTI Error: Could not submit DTI URB (%d), "
			"resetting\n", result);
		wa_reset_all(wa);
	}
out:
	return;
}

/*
 * Transfer complete notification
 *
 * Called from the notif.c code. We get a notification on EP2 saying
 * that some endpoint has some transfer result data available. We are
 * about to read it.
 *
 * To speed up things, we always have a URB reading the DTI URB; we
 * don't really set it up and start it until the first xfer complete
 * notification arrives, which is what we do here.
 *
 * Follow up in wa_xfer_result_cb(), as that's where the whole state
 * machine starts.
 *
 * So here we just initialize the DTI URB for reading transfer result
 * notifications and also the buffer-in URB, for reading buffers. Then
 * we just submit the DTI URB.
 *
 * @wa shall be referenced
 */
void wa_handle_notif_xfer(struct wahc *wa, struct wa_notif_hdr *notif_hdr)
{
	int result;
	struct device *dev = &wa->usb_iface->dev;
	struct wa_notif_xfer *notif_xfer;
	const struct usb_endpoint_descriptor *dti_epd = wa->dti_epd;

	notif_xfer = container_of(notif_hdr, struct wa_notif_xfer, hdr);
	BUG_ON(notif_hdr->bNotifyType != WA_NOTIF_TRANSFER);

	if ((0x80 | notif_xfer->bEndpoint) != dti_epd->bEndpointAddress) {
		/* FIXME: hardcoded limitation, adapt */
		dev_err(dev, "BUG: DTI ep is %u, not %u (hack me)\n",
			notif_xfer->bEndpoint, dti_epd->bEndpointAddress);
		goto error;
	}
	if (wa->dti_urb != NULL)	/* DTI URB already started */
		goto out;

	wa->dti_urb = usb_alloc_urb(0, GFP_KERNEL);
	if (wa->dti_urb == NULL) {
		dev_err(dev, "Can't allocate DTI URB\n");
		goto error_dti_urb_alloc;
	}
	usb_fill_bulk_urb(
		wa->dti_urb, wa->usb_dev,
		usb_rcvbulkpipe(wa->usb_dev, 0x80 | notif_xfer->bEndpoint),
		wa->xfer_result, wa->xfer_result_size,
		wa_xfer_result_cb, wa);

	wa->buf_in_urb = usb_alloc_urb(0, GFP_KERNEL);
	if (wa->buf_in_urb == NULL) {
		dev_err(dev, "Can't allocate BUF-IN URB\n");
		goto error_buf_in_urb_alloc;
	}
	usb_fill_bulk_urb(
		wa->buf_in_urb, wa->usb_dev,
		usb_rcvbulkpipe(wa->usb_dev, 0x80 | notif_xfer->bEndpoint),
		NULL, 0, wa_buf_in_cb, wa);
	result = usb_submit_urb(wa->dti_urb, GFP_KERNEL);
	if (result < 0) {
		dev_err(dev, "DTI Error: Could not submit DTI URB (%d), "
			"resetting\n", result);
		goto error_dti_urb_submit;
	}
out:
	return;

error_dti_urb_submit:
	usb_put_urb(wa->buf_in_urb);
error_buf_in_urb_alloc:
	usb_put_urb(wa->dti_urb);
	wa->dti_urb = NULL;
error_dti_urb_alloc:
error:
	wa_reset_all(wa);
}