char: xillybus: Add driver for XillyUSB (Xillybus variant for USB)

The XillyUSB driver is the USB variant for the Xillybus FPGA IP core.
Even though it presents a nearly identical API on the FPGA and host,
it's almost a complete rewrite of the driver: The framework for exchanging
data on a USB bus is fundamentally different from doing the same with a
PCIe interface, which leaves very little in common between the existing
driver and the new one for XillyUSB.

Signed-off-by: Eli Billauer <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Greg Kroah-Hartman <[e-mail]>

drivers/char/xillybus/Kconfig

···
 	  system, say M. The module will be called xillybus_of.
 
 endif # if XILLYBUS
+
+# XILLYUSB doesn't depend on XILLYBUS
+
+config XILLYUSB
+	tristate "XillyUSB: Xillybus generic FPGA interface for USB"
+	depends on USB
+	select CRC32
+	select XILLYBUS_CLASS
+	help
+	  XillyUSB is the Xillybus variant which uses USB for communicating
+	  with the FPGA.
+
+	  Set to M if you want Xillybus to use USB for communicating with
+	  the FPGA. The module will be called xillyusb.

drivers/char/xillybus/Makefile

···
 obj-$(CONFIG_XILLYBUS)		+= xillybus_core.o
 obj-$(CONFIG_XILLYBUS_PCIE)	+= xillybus_pcie.o
 obj-$(CONFIG_XILLYBUS_OF)	+= xillybus_of.o
+obj-$(CONFIG_XILLYUSB)		+= xillyusb.o

drivers/char/xillybus/xillyusb.c

···
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright 2020 Xillybus Ltd, http://xillybus.com
+ *
+ * Driver for the XillyUSB FPGA/host framework.
+ *
+ * This driver interfaces with a special IP core in an FPGA, setting up
+ * a pipe between a hardware FIFO in the programmable logic and a device
+ * file in the host. The number of such pipes and their attributes are
+ * set up on the logic. This driver detects these automatically and
+ * creates the device files accordingly.
+ */
+
+#include <linux/types.h>
+#include <linux/slab.h>
+#include <linux/list.h>
+#include <linux/device.h>
+#include <linux/module.h>
+#include <asm/byteorder.h>
+#include <linux/io.h>
+#include <linux/interrupt.h>
+#include <linux/sched.h>
+#include <linux/fs.h>
+#include <linux/spinlock.h>
+#include <linux/mutex.h>
+#include <linux/workqueue.h>
+#include <linux/crc32.h>
+#include <linux/poll.h>
+#include <linux/delay.h>
+#include <linux/usb.h>
+
+#include "xillybus_class.h"
+
+MODULE_DESCRIPTION("Driver for XillyUSB FPGA IP Core");
+MODULE_AUTHOR("Eli Billauer, Xillybus Ltd.");
+MODULE_VERSION("1.1");
+MODULE_ALIAS("xillyusb");
+MODULE_LICENSE("GPL v2");
+
+#define XILLY_RX_TIMEOUT		(10 * HZ / 1000)
+#define XILLY_RESPONSE_TIMEOUT		(500 * HZ / 1000)
+
+#define BUF_SIZE_ORDER			4
+#define BUFNUM				8
+#define LOG2_IDT_FIFO_SIZE		16
+#define LOG2_INITIAL_FIFO_BUF_SIZE	16
+
+#define MSG_EP_NUM			1
+#define IN_EP_NUM			1
+
+static const char xillyname[] = "xillyusb";
+
+static unsigned int fifo_buf_order;
+
+#define USB_VENDOR_ID_XILINX		0x03fd
+#define USB_VENDOR_ID_ALTERA		0x09fb
+
+#define USB_PRODUCT_ID_XILLYUSB		0xebbe
+
+static const struct usb_device_id xillyusb_table[] = {
+	{ USB_DEVICE(USB_VENDOR_ID_XILINX, USB_PRODUCT_ID_XILLYUSB) },
+	{ USB_DEVICE(USB_VENDOR_ID_ALTERA, USB_PRODUCT_ID_XILLYUSB) },
+	{ }
+};
+
+MODULE_DEVICE_TABLE(usb, xillyusb_table);
+
+struct xillyusb_dev;
+
+struct xillyfifo {
+	unsigned int bufsize; /* In bytes, always a power of 2 */
+	unsigned int bufnum;
+	unsigned int size; /* Lazy: Equals bufsize * bufnum */
+	unsigned int buf_order;
+
+	int fill; /* Number of bytes in the FIFO */
+	spinlock_t lock;
+	wait_queue_head_t waitq;
+
+	unsigned int readpos;
+	unsigned int readbuf;
+	unsigned int writepos;
+	unsigned int writebuf;
+	char **mem;
+};
+
+struct xillyusb_channel;
+
+struct xillyusb_endpoint {
+	struct xillyusb_dev *xdev;
+
+	struct mutex ep_mutex; /* serialize operations on endpoint */
+
+	struct list_head buffers;
+	struct list_head filled_buffers;
+	spinlock_t buffers_lock; /* protect these two lists */
+
+	unsigned int order;
+	unsigned int buffer_size;
+
+	unsigned int fill_mask;
+
+	int outstanding_urbs;
+
+	struct usb_anchor anchor;
+
+	struct xillyfifo fifo;
+
+	struct work_struct workitem;
+
+	bool shutting_down;
+	bool drained;
+	bool wake_on_drain;
+
+	u8 ep_num;
+};
+
+struct xillyusb_channel {
+	struct xillyusb_dev *xdev;
+
+	struct xillyfifo *in_fifo;
+	struct xillyusb_endpoint *out_ep;
+	struct mutex lock; /* protect @out_ep, @in_fifo, bit fields below */
+
+	struct mutex in_mutex; /* serialize fops on FPGA to host stream */
+	struct mutex out_mutex; /* serialize fops on host to FPGA stream */
+	wait_queue_head_t flushq;
+
+	int chan_idx;
+
+	u32 in_consumed_bytes;
+	u32 in_current_checkpoint;
+	u32 out_bytes;
+
+	unsigned int in_log2_element_size;
+	unsigned int out_log2_element_size;
+	unsigned int in_log2_fifo_size;
+	unsigned int out_log2_fifo_size;
+
+	unsigned int read_data_ok; /* EOF not arrived (yet) */
+	unsigned int poll_used;
+	unsigned int flushing;
+	unsigned int flushed;
+	unsigned int canceled;
+
+	/* Bit fields protected by @lock except for initialization */
+	unsigned readable:1;
+	unsigned writable:1;
+	unsigned open_for_read:1;
+	unsigned open_for_write:1;
+	unsigned in_synchronous:1;
+	unsigned out_synchronous:1;
+	unsigned in_seekable:1;
+	unsigned out_seekable:1;
+};
+
+struct xillybuffer {
+	struct list_head entry;
+	struct xillyusb_endpoint *ep;
+	void *buf;
+	unsigned int len;
+};
+
+struct xillyusb_dev {
+	struct xillyusb_channel *channels;
+
+	struct usb_device	*udev;
+	struct device		*dev; /* For dev_err() and such */
+	struct kref		kref;
+	struct workqueue_struct	*workq;
+
+	int error;
+	spinlock_t error_lock; /* protect @error */
+	struct work_struct wakeup_workitem;
+
+	int num_channels;
+
+	struct xillyusb_endpoint *msg_ep;
+	struct xillyusb_endpoint *in_ep;
+
+	struct mutex msg_mutex; /* serialize opcode transmission */
+	int in_bytes_left;
+	int leftover_chan_num;
+	unsigned int in_counter;
+	struct mutex process_in_mutex; /* synchronize wakeup_all() */
+};
+
+/* FPGA to host opcodes */
+enum {
+	OPCODE_DATA = 0,
+	OPCODE_QUIESCE_ACK = 1,
+	OPCODE_EOF = 2,
+	OPCODE_REACHED_CHECKPOINT = 3,
+	OPCODE_CANCELED_CHECKPOINT = 4,
+};
+
+/* Host to FPGA opcodes */
+enum {
+	OPCODE_QUIESCE = 0,
+	OPCODE_REQ_IDT = 1,
+	OPCODE_SET_CHECKPOINT = 2,
+	OPCODE_CLOSE = 3,
+	OPCODE_SET_PUSH = 4,
+	OPCODE_UPDATE_PUSH = 5,
+	OPCODE_CANCEL_CHECKPOINT = 6,
+	OPCODE_SET_ADDR = 7,
+};
+
+/*
+ * fifo_write() and fifo_read() are NOT reentrant (i.e. concurrent multiple
+ * calls to each on the same FIFO is not allowed) however it's OK to have
+ * threads calling each of the two functions once on the same FIFO, and
+ * at the same time.
+ */
+
+static int fifo_write(struct xillyfifo *fifo,
+		      const void *data, unsigned int len,
+		      int (*copier)(void *, const void *, int))
+{
+	unsigned int done = 0;
+	unsigned int todo = len;
+	unsigned int nmax;
+	unsigned int writepos = fifo->writepos;
+	unsigned int writebuf = fifo->writebuf;
+	unsigned long flags;
+	int rc;
+
+	nmax = fifo->size - READ_ONCE(fifo->fill);
+
+	while (1) {
+		unsigned int nrail = fifo->bufsize - writepos;
+		unsigned int n = min(todo, nmax);
+
+		if (n == 0) {
+			spin_lock_irqsave(&fifo->lock, flags);
+			fifo->fill += done;
+			spin_unlock_irqrestore(&fifo->lock, flags);
+
+			fifo->writepos = writepos;
+			fifo->writebuf = writebuf;
+
+			return done;
+		}
+
+		if (n > nrail)
+			n = nrail;
+
+		rc = (*copier)(fifo->mem[writebuf] + writepos, data + done, n);
+
+		if (rc)
+			return rc;
+
+		done += n;
+		todo -= n;
+
+		writepos += n;
+		nmax -= n;
+
+		if (writepos == fifo->bufsize) {
+			writepos = 0;
+			writebuf++;
+
+			if (writebuf == fifo->bufnum)
+				writebuf = 0;
+		}
+	}
+}
+
+static int fifo_read(struct xillyfifo *fifo,
+		     void *data, unsigned int len,
+		     int (*copier)(void *, const void *, int))
+{
+	unsigned int done = 0;
+	unsigned int todo = len;
+	unsigned int fill;
+	unsigned int readpos = fifo->readpos;
+	unsigned int readbuf = fifo->readbuf;
+	unsigned long flags;
+	int rc;
+
+	/*
+	 * The spinlock here is necessary, because otherwise fifo->fill
+	 * could have been increased by fifo_write() after writing data
+	 * to the buffer, but this data would potentially not have been
+	 * visible on this thread at the time the updated fifo->fill was.
+	 * That could lead to reading invalid data.
+	 */
+
+	spin_lock_irqsave(&fifo->lock, flags);
+	fill = fifo->fill;
+	spin_unlock_irqrestore(&fifo->lock, flags);
+
+	while (1) {
+		unsigned int nrail = fifo->bufsize - readpos;
+		unsigned int n = min(todo, fill);
+
+		if (n == 0) {
+			spin_lock_irqsave(&fifo->lock, flags);
+			fifo->fill -= done;
+			spin_unlock_irqrestore(&fifo->lock, flags);
+
+			fifo->readpos = readpos;
+			fifo->readbuf = readbuf;
+
+			return done;
+		}
+
+		if (n > nrail)
+			n = nrail;
+
+		rc = (*copier)(data + done, fifo->mem[readbuf] + readpos, n);
+
+		if (rc)
+			return rc;
+
+		done += n;
+		todo -= n;
+
+		readpos += n;
+		fill -= n;
+
+		if (readpos == fifo->bufsize) {
+			readpos = 0;
+			readbuf++;
+
+			if (readbuf == fifo->bufnum)
+				readbuf = 0;
+		}
+	}
+}
+
+/*
+ * These three wrapper functions are used as the @copier argument to
+ * fifo_write() and fifo_read(), so that they can work directly with
+ * user memory as well.
+ */
+
+static int xilly_copy_from_user(void *dst, const void *src, int n)
+{
+	if (copy_from_user(dst, (const void __user *)src, n))
+		return -EFAULT;
+
+	return 0;
+}
+
+static int xilly_copy_to_user(void *dst, const void *src, int n)
+{
+	if (copy_to_user((void __user *)dst, src, n))
+		return -EFAULT;
+
+	return 0;
+}
+
+static int xilly_memcpy(void *dst, const void *src, int n)
+{
+	memcpy(dst, src, n);
+
+	return 0;
+}
+
+static int fifo_init(struct xillyfifo *fifo,
+		     unsigned int log2_size)
+{
+	unsigned int log2_bufnum;
+	unsigned int buf_order;
+	int i;
+
+	unsigned int log2_fifo_buf_size;
+
+retry:
+	log2_fifo_buf_size = fifo_buf_order + PAGE_SHIFT;
+
+	if (log2_size > log2_fifo_buf_size) {
+		log2_bufnum = log2_size - log2_fifo_buf_size;
+		buf_order = fifo_buf_order;
+		fifo->bufsize = 1 << log2_fifo_buf_size;
+	} else {
+		log2_bufnum = 0;
+		buf_order = (log2_size > PAGE_SHIFT) ?
+			log2_size - PAGE_SHIFT : 0;
+		fifo->bufsize = 1 << log2_size;
+	}
+
+	fifo->bufnum = 1 << log2_bufnum;
+	fifo->size = fifo->bufnum * fifo->bufsize;
+	fifo->buf_order = buf_order;
+
+	fifo->mem = kmalloc_array(fifo->bufnum, sizeof(void *), GFP_KERNEL);
+
+	if (!fifo->mem)
+		return -ENOMEM;
+
+	for (i = 0; i < fifo->bufnum; i++) {
+		fifo->mem[i] = (void *)
+			__get_free_pages(GFP_KERNEL, buf_order);
+
+		if (!fifo->mem[i])
+			goto memfail;
+	}
+
+	fifo->fill = 0;
+	fifo->readpos = 0;
+	fifo->readbuf = 0;
+	fifo->writepos = 0;
+	fifo->writebuf = 0;
+	spin_lock_init(&fifo->lock);
+	init_waitqueue_head(&fifo->waitq);
+	return 0;
+
+memfail:
+	for (i--; i >= 0; i--)
+		free_pages((unsigned long)fifo->mem[i], buf_order);
+
+	kfree(fifo->mem);
+	fifo->mem = NULL;
+
+	if (fifo_buf_order) {
+		fifo_buf_order--;
+		goto retry;
+	} else {
+		return -ENOMEM;
+	}
+}
+
+static void fifo_mem_release(struct xillyfifo *fifo)
+{
+	int i;
+
+	if (!fifo->mem)
+		return;
+
+	for (i = 0; i < fifo->bufnum; i++)
+		free_pages((unsigned long)fifo->mem[i], fifo->buf_order);
+
+	kfree(fifo->mem);
+}
+
+/*
+ * When endpoint_quiesce() returns, the endpoint has no URBs submitted,
+ * won't accept any new URB submissions, and its related work item doesn't
+ * and won't run anymore.
+ */
+
+static void endpoint_quiesce(struct xillyusb_endpoint *ep)
+{
+	mutex_lock(&ep->ep_mutex);
+	ep->shutting_down = true;
+	mutex_unlock(&ep->ep_mutex);
+
+	usb_kill_anchored_urbs(&ep->anchor);
+	cancel_work_sync(&ep->workitem);
+}
+
+/*
+ * Note that endpoint_dealloc() also frees fifo memory (if allocated), even
+ * though endpoint_alloc doesn't allocate that memory.
+ */
+
+static void endpoint_dealloc(struct xillyusb_endpoint *ep)
+{
+	struct list_head *this, *next;
+
+	fifo_mem_release(&ep->fifo);
+
+	/* Join @filled_buffers with @buffers to free these entries too */
+	list_splice(&ep->filled_buffers, &ep->buffers);
+
+	list_for_each_safe(this, next, &ep->buffers) {
+		struct xillybuffer *xb =
+			list_entry(this, struct xillybuffer, entry);
+
+		free_pages((unsigned long)xb->buf, ep->order);
+		kfree(xb);
+	}
+
+	kfree(ep);
+}
+
+static struct xillyusb_endpoint
+*endpoint_alloc(struct xillyusb_dev *xdev,
+		u8 ep_num,
+		void (*work)(struct work_struct *),
+		unsigned int order,
+		int bufnum)
+{
+	int i;
+
+	struct xillyusb_endpoint *ep;
+
+	ep = kzalloc(sizeof(*ep), GFP_KERNEL);
+
+	if (!ep)
+		return NULL;
+
+	INIT_LIST_HEAD(&ep->buffers);
+	INIT_LIST_HEAD(&ep->filled_buffers);
+
+	spin_lock_init(&ep->buffers_lock);
+	mutex_init(&ep->ep_mutex);
+
+	init_usb_anchor(&ep->anchor);
+	INIT_WORK(&ep->workitem, work);
+
+	ep->order = order;
+	ep->buffer_size =  1 << (PAGE_SHIFT + order);
+	ep->outstanding_urbs = 0;
+	ep->drained = true;
+	ep->wake_on_drain = false;
+	ep->xdev = xdev;
+	ep->ep_num = ep_num;
+	ep->shutting_down = false;
+
+	for (i = 0; i < bufnum; i++) {
+		struct xillybuffer *xb;
+		unsigned long addr;
+
+		xb = kzalloc(sizeof(*xb), GFP_KERNEL);
+
+		if (!xb) {
+			endpoint_dealloc(ep);
+			return NULL;
+		}
+
+		addr = __get_free_pages(GFP_KERNEL, order);
+
+		if (!addr) {
+			kfree(xb);
+			endpoint_dealloc(ep);
+			return NULL;
+		}
+
+		xb->buf = (void *)addr;
+		xb->ep = ep;
+		list_add_tail(&xb->entry, &ep->buffers);
+	}
+	return ep;
+}
+
+static void cleanup_dev(struct kref *kref)
+{
+	struct xillyusb_dev *xdev =
+		container_of(kref, struct xillyusb_dev, kref);
+
+	if (xdev->in_ep)
+		endpoint_dealloc(xdev->in_ep);
+
+	if (xdev->msg_ep)
+		endpoint_dealloc(xdev->msg_ep);
+
+	if (xdev->workq)
+		destroy_workqueue(xdev->workq);
+
+	kfree(xdev->channels); /* Argument may be NULL, and that's fine */
+	kfree(xdev);
+}
+
+/*
+ * @process_in_mutex is taken to ensure that bulk_in_work() won't call
+ * process_bulk_in() after wakeup_all()'s execution: The latter zeroes all
+ * @read_data_ok entries, which will make process_bulk_in() report false
+ * errors if executed. The mechanism relies on that xdev->error is assigned
+ * a non-zero value by report_io_error() prior to queueing wakeup_all(),
+ * which prevents bulk_in_work() from calling process_bulk_in().
+ *
+ * The fact that wakeup_all() and bulk_in_work() are queued on the same
+ * workqueue makes their concurrent execution very unlikely, however the
+ * kernel's API doesn't seem to ensure this strictly.
+ */
+
+static void wakeup_all(struct work_struct *work)
+{
+	int i;
+	struct xillyusb_dev *xdev = container_of(work, struct xillyusb_dev,
+						 wakeup_workitem);
+
+	mutex_lock(&xdev->process_in_mutex);
+
+	for (i = 0; i < xdev->num_channels; i++) {
+		struct xillyusb_channel *chan = &xdev->channels[i];
+
+		mutex_lock(&chan->lock);
+
+		if (chan->in_fifo) {
+			/*
+			 * Fake an EOF: Even if such arrives, it won't be
+			 * processed.
+			 */
+			chan->read_data_ok = 0;
+			wake_up_interruptible(&chan->in_fifo->waitq);
+		}
+
+		if (chan->out_ep)
+			wake_up_interruptible(&chan->out_ep->fifo.waitq);
+
+		mutex_unlock(&chan->lock);
+
+		wake_up_interruptible(&chan->flushq);
+	}
+
+	mutex_unlock(&xdev->process_in_mutex);
+
+	wake_up_interruptible(&xdev->msg_ep->fifo.waitq);
+
+	kref_put(&xdev->kref, cleanup_dev);
+}
+
+static void report_io_error(struct xillyusb_dev *xdev,
+			    int errcode)
+{
+	unsigned long flags;
+	bool do_once = false;
+
+	spin_lock_irqsave(&xdev->error_lock, flags);
+	if (!xdev->error) {
+		xdev->error = errcode;
+		do_once = true;
+	}
+	spin_unlock_irqrestore(&xdev->error_lock, flags);
+
+	if (do_once) {
+		kref_get(&xdev->kref); /* xdev is used by work item */
+		queue_work(xdev->workq, &xdev->wakeup_workitem);
+	}
+}
+
+/*
+ * safely_assign_in_fifo() changes the value of chan->in_fifo and ensures
+ * the previous pointer is never used after its return.
+ */
+
+static void safely_assign_in_fifo(struct xillyusb_channel *chan,
+				  struct xillyfifo *fifo)
+{
+	mutex_lock(&chan->lock);
+	chan->in_fifo = fifo;
+	mutex_unlock(&chan->lock);
+
+	flush_work(&chan->xdev->in_ep->workitem);
+}
+
+static void bulk_in_completer(struct urb *urb)
+{
+	struct xillybuffer *xb = urb->context;
+	struct xillyusb_endpoint *ep = xb->ep;
+	unsigned long flags;
+
+	if (urb->status) {
+		if (!(urb->status == -ENOENT ||
+		      urb->status == -ECONNRESET ||
+		      urb->status == -ESHUTDOWN))
+			report_io_error(ep->xdev, -EIO);
+
+		spin_lock_irqsave(&ep->buffers_lock, flags);
+		list_add_tail(&xb->entry, &ep->buffers);
+		ep->outstanding_urbs--;
+		spin_unlock_irqrestore(&ep->buffers_lock, flags);
+
+		return;
+	}
+
+	xb->len = urb->actual_length;
+
+	spin_lock_irqsave(&ep->buffers_lock, flags);
+	list_add_tail(&xb->entry, &ep->filled_buffers);
+	spin_unlock_irqrestore(&ep->buffers_lock, flags);
+
+	if (!ep->shutting_down)
+		queue_work(ep->xdev->workq, &ep->workitem);
+}
+
+static void bulk_out_completer(struct urb *urb)
+{
+	struct xillybuffer *xb = urb->context;
+	struct xillyusb_endpoint *ep = xb->ep;
+	unsigned long flags;
+
+	if (urb->status &&
+	    (!(urb->status == -ENOENT ||
+	       urb->status == -ECONNRESET ||
+	       urb->status == -ESHUTDOWN)))
+		report_io_error(ep->xdev, -EIO);
+
+	spin_lock_irqsave(&ep->buffers_lock, flags);
+	list_add_tail(&xb->entry, &ep->buffers);
+	ep->outstanding_urbs--;
+	spin_unlock_irqrestore(&ep->buffers_lock, flags);
+
+	if (!ep->shutting_down)
+		queue_work(ep->xdev->workq, &ep->workitem);
+}
+
+static void try_queue_bulk_in(struct xillyusb_endpoint *ep)
+{
+	struct xillyusb_dev *xdev = ep->xdev;
+	struct xillybuffer *xb;
+	struct urb *urb;
+
+	int rc;
+	unsigned long flags;
+	unsigned int bufsize = ep->buffer_size;
+
+	mutex_lock(&ep->ep_mutex);
+
+	if (ep->shutting_down || xdev->error)
+		goto done;
+
+	while (1) {
+		spin_lock_irqsave(&ep->buffers_lock, flags);
+
+		if (list_empty(&ep->buffers)) {
+			spin_unlock_irqrestore(&ep->buffers_lock, flags);
+			goto done;
+		}
+
+		xb = list_first_entry(&ep->buffers, struct xillybuffer, entry);
+		list_del(&xb->entry);
+		ep->outstanding_urbs++;
+
+		spin_unlock_irqrestore(&ep->buffers_lock, flags);
+
+		urb = usb_alloc_urb(0, GFP_KERNEL);
+		if (!urb) {
+			report_io_error(xdev, -ENOMEM);
+			goto relist;
+		}
+
+		usb_fill_bulk_urb(urb, xdev->udev,
+				  usb_rcvbulkpipe(xdev->udev, ep->ep_num),
+				  xb->buf, bufsize, bulk_in_completer, xb);
+
+		usb_anchor_urb(urb, &ep->anchor);
+
+		rc = usb_submit_urb(urb, GFP_KERNEL);
+
+		if (rc) {
+			report_io_error(xdev, (rc == -ENOMEM) ? -ENOMEM :
+					-EIO);
+			goto unanchor;
+		}
+
+		usb_free_urb(urb); /* This just decrements reference count */
+	}
+
+unanchor:
+	usb_unanchor_urb(urb);
+	usb_free_urb(urb);
+
+relist:
+	spin_lock_irqsave(&ep->buffers_lock, flags);
+	list_add_tail(&xb->entry, &ep->buffers);
+	ep->outstanding_urbs--;
+	spin_unlock_irqrestore(&ep->buffers_lock, flags);
+
+done:
+	mutex_unlock(&ep->ep_mutex);
+}
+
+static void try_queue_bulk_out(struct xillyusb_endpoint *ep)
+{
+	struct xillyfifo *fifo = &ep->fifo;
+	struct xillyusb_dev *xdev = ep->xdev;
+	struct xillybuffer *xb;
+	struct urb *urb;
+
+	int rc;
+	unsigned int fill;
+	unsigned long flags;
+	bool do_wake = false;
+
+	mutex_lock(&ep->ep_mutex);
+
+	if (ep->shutting_down || xdev->error)
+		goto done;
+
+	fill = READ_ONCE(fifo->fill) & ep->fill_mask;
+
+	while (1) {
+		int count;
+		unsigned int max_read;
+
+		spin_lock_irqsave(&ep->buffers_lock, flags);
+
+		/*
+		 * Race conditions might have the FIFO filled while the
+		 * endpoint is marked as drained here. That doesn't matter,
+		 * because the sole purpose of @drained is to ensure that
+		 * certain data has been sent on the USB channel before
+		 * shutting it down. Hence knowing that the FIFO appears
+		 * to be empty with no outstanding URBs at some moment
+		 * is good enough.
+		 */
+
+		if (!fill) {
+			ep->drained = !ep->outstanding_urbs;
+			if (ep->drained && ep->wake_on_drain)
+				do_wake = true;
+
+			spin_unlock_irqrestore(&ep->buffers_lock, flags);
+			goto done;
+		}
+
+		ep->drained = false;
+
+		if ((fill < ep->buffer_size && ep->outstanding_urbs) ||
+		    list_empty(&ep->buffers)) {
+			spin_unlock_irqrestore(&ep->buffers_lock, flags);
+			goto done;
+		}
+
+		xb = list_first_entry(&ep->buffers, struct xillybuffer, entry);
+		list_del(&xb->entry);
+		ep->outstanding_urbs++;
+
+		spin_unlock_irqrestore(&ep->buffers_lock, flags);
+
+		max_read = min(fill, ep->buffer_size);
+
+		count = fifo_read(&ep->fifo, xb->buf, max_read, xilly_memcpy);
+
+		/*
+		 * xilly_memcpy always returns 0 => fifo_read can't fail =>
+		 * count > 0
+		 */
+
+		urb = usb_alloc_urb(0, GFP_KERNEL);
+		if (!urb) {
+			report_io_error(xdev, -ENOMEM);
+			goto relist;
+		}
+
+		usb_fill_bulk_urb(urb, xdev->udev,
+				  usb_sndbulkpipe(xdev->udev, ep->ep_num),
+				  xb->buf, count, bulk_out_completer, xb);
+
+		usb_anchor_urb(urb, &ep->anchor);
+
+		rc = usb_submit_urb(urb, GFP_KERNEL);
+
+		if (rc) {
+			report_io_error(xdev, (rc == -ENOMEM) ? -ENOMEM :
+					-EIO);
+			goto unanchor;
+		}
+
+		usb_free_urb(urb); /* This just decrements reference count */
+
+		fill -= count;
+		do_wake = true;
+	}
+
+unanchor:
+	usb_unanchor_urb(urb);
+	usb_free_urb(urb);
+
+relist:
+	spin_lock_irqsave(&ep->buffers_lock, flags);
+	list_add_tail(&xb->entry, &ep->buffers);
+	ep->outstanding_urbs--;
+	spin_unlock_irqrestore(&ep->buffers_lock, flags);
+
+done:
+	mutex_unlock(&ep->ep_mutex);
+
+	if (do_wake)
+		wake_up_interruptible(&fifo->waitq);
+}
+
+static void bulk_out_work(struct work_struct *work)
+{
+	struct xillyusb_endpoint *ep = container_of(work,
+						    struct xillyusb_endpoint,
+						    workitem);
+	try_queue_bulk_out(ep);
+}
+
+static int process_in_opcode(struct xillyusb_dev *xdev,
+			     int opcode,
+			     int chan_num)
+{
+	struct xillyusb_channel *chan;
+	struct device *dev = xdev->dev;
+	int chan_idx = chan_num >> 1;
+
+	if (chan_idx >= xdev->num_channels) {
+		dev_err(dev, "Received illegal channel ID %d from FPGA\n",
+			chan_num);
+		return -EIO;
+	}
+
+	chan = &xdev->channels[chan_idx];
+
+	switch (opcode) {
+	case OPCODE_EOF:
+		if (!chan->read_data_ok) {
+			dev_err(dev, "Received unexpected EOF for channel %d\n",
+				chan_num);
+			return -EIO;
+		}
+
+		/*
+		 * A write memory barrier ensures that the FIFO's fill level
+		 * is visible before read_data_ok turns zero, so the data in
+		 * the FIFO isn't missed by the consumer.
+		 */
+		smp_wmb();
+		WRITE_ONCE(chan->read_data_ok, 0);
+		wake_up_interruptible(&chan->in_fifo->waitq);
+		break;
+
+	case OPCODE_REACHED_CHECKPOINT:
+		chan->flushing = 0;
+		wake_up_interruptible(&chan->flushq);
+		break;
+
+	case OPCODE_CANCELED_CHECKPOINT:
+		chan->canceled = 1;
+		wake_up_interruptible(&chan->flushq);
+		break;
+
+	default:
+		dev_err(dev, "Received illegal opcode %d from FPGA\n",
+			opcode);
+		return -EIO;
+	}
+
+	return 0;
+}
+
+static int process_bulk_in(struct xillybuffer *xb)
+{
+	struct xillyusb_endpoint *ep = xb->ep;
+	struct xillyusb_dev *xdev = ep->xdev;
+	struct device *dev = xdev->dev;
+	int dws = xb->len >> 2;
+	__le32 *p = xb->buf;
+	u32 ctrlword;
+	struct xillyusb_channel *chan;
+	struct xillyfifo *fifo;
+	int chan_num = 0, opcode;
+	int chan_idx;
+	int bytes, count, dwconsume;
+	int in_bytes_left = 0;
+	int rc;
+
+	if ((dws << 2) != xb->len) {
+		dev_err(dev, "Received BULK IN transfer with %d bytes, not a multiple of 4\n",
+			xb->len);
+		return -EIO;
+	}
+
+	if (xdev->in_bytes_left) {
+		bytes = min(xdev->in_bytes_left, dws << 2);
+		in_bytes_left = xdev->in_bytes_left - bytes;
+		chan_num = xdev->leftover_chan_num;
+		goto resume_leftovers;
+	}
+
+	while (dws) {
+		ctrlword = le32_to_cpu(*p++);
+		dws--;
+
+		chan_num = ctrlword & 0xfff;
+		count = (ctrlword >> 12) & 0x3ff;
+		opcode = (ctrlword >> 24) & 0xf;
+
+		if (opcode != OPCODE_DATA) {
+			unsigned int in_counter = xdev->in_counter++ & 0x3ff;
+
+			if (count != in_counter) {
+				dev_err(dev, "Expected opcode counter %d, got %d\n",
+					in_counter, count);
+				return -EIO;
+			}
+
+			rc = process_in_opcode(xdev, opcode, chan_num);
+
+			if (rc)
+				return rc;
+
+			continue;
+		}
+
+		bytes = min(count + 1, dws << 2);
+		in_bytes_left = count + 1 - bytes;
+
+resume_leftovers:
+		chan_idx = chan_num >> 1;
+
+		if (!(chan_num & 1) || chan_idx >= xdev->num_channels ||
+		    !xdev->channels[chan_idx].read_data_ok) {
+			dev_err(dev, "Received illegal channel ID %d from FPGA\n",
+				chan_num);
+			return -EIO;
+		}
+		chan = &xdev->channels[chan_idx];
+
+		fifo = chan->in_fifo;
+
+		if (unlikely(!fifo))
+			return -EIO; /* We got really unexpected data */
+
+		if (bytes != fifo_write(fifo, p, bytes, xilly_memcpy)) {
+			dev_err(dev, "Misbehaving FPGA overflew an upstream FIFO!\n");
+			return -EIO;
+		}
+
+		wake_up_interruptible(&fifo->waitq);
+
+		dwconsume = (bytes + 3) >> 2;
+		dws -= dwconsume;
+		p += dwconsume;
+	}
+
+	xdev->in_bytes_left = in_bytes_left;
+	xdev->leftover_chan_num = chan_num;
+	return 0;
+}
+
+static void bulk_in_work(struct work_struct *work)
+{
+	struct xillyusb_endpoint *ep =
+		container_of(work, struct xillyusb_endpoint, workitem);
+	struct xillyusb_dev *xdev = ep->xdev;
+	unsigned long flags;
+	struct xillybuffer *xb;
+	bool consumed = false;
+	int rc = 0;
+
+	mutex_lock(&xdev->process_in_mutex);
+
+	spin_lock_irqsave(&ep->buffers_lock, flags);
+
+	while (1) {
+		if (rc || list_empty(&ep->filled_buffers)) {
+			spin_unlock_irqrestore(&ep->buffers_lock, flags);
+			mutex_unlock(&xdev->process_in_mutex);
+
+			if (rc)
+				report_io_error(xdev, rc);
+			else if (consumed)
+				try_queue_bulk_in(ep);
+
+			return;
+		}
+
+		xb = list_first_entry(&ep->filled_buffers, struct xillybuffer,
+				      entry);
+		list_del(&xb->entry);
+
+		spin_unlock_irqrestore(&ep->buffers_lock, flags);
+
+		consumed = true;
+
+		if (!xdev->error)
+			rc = process_bulk_in(xb);
+
+		spin_lock_irqsave(&ep->buffers_lock, flags);
+		list_add_tail(&xb->entry, &ep->buffers);
+		ep->outstanding_urbs--;
+	}
+}
+
+static int xillyusb_send_opcode(struct xillyusb_dev *xdev,
+				int chan_num, char opcode, u32 data)
+{
+	struct xillyusb_endpoint *ep = xdev->msg_ep;
+	struct xillyfifo *fifo = &ep->fifo;
+	__le32 msg[2];
+
+	int rc = 0;
+
+	msg[0] = cpu_to_le32((chan_num & 0xfff) |
+			     ((opcode & 0xf) << 24));
+	msg[1] = cpu_to_le32(data);
+
+	mutex_lock(&xdev->msg_mutex);
+
+	/*
+	 * The wait queue is woken with the interruptible variant, so the
+	 * wait function matches, however returning because of an interrupt
+	 * will mess things up considerably, in particular when the caller is
+	 * the release method. And the xdev->error part prevents being stuck
+	 * forever in the event of a bizarre hardware bug: Pull the USB plug.
+	 */
+
+	while (wait_event_interruptible(fifo->waitq,
+					fifo->fill <= (fifo->size - 8) ||
+					xdev->error))
+		; /* Empty loop */
+
+	if (xdev->error) {
+		rc = xdev->error;
+		goto unlock_done;
+	}
+
+	fifo_write(fifo, (void *)msg, 8, xilly_memcpy);
+
+	try_queue_bulk_out(ep);
+
+unlock_done:
+	mutex_unlock(&xdev->msg_mutex);
+
+	return rc;
+}
+
+/*
+ * Note that flush_downstream() merely waits for the data to arrive to
+ * the application logic at the FPGA -- unlike PCIe Xillybus' counterpart,
+ * it does nothing to make it happen (and neither is it necessary).
+ *
+ * This function is not reentrant for the same @chan, but this is covered
+ * by the fact that for any given @chan, it's called either by the open,
+ * write, llseek and flush fops methods, which can't run in parallel (and the
+ * write + flush and llseek method handlers are protected with out_mutex).
+ *
+ * chan->flushed is there to avoid multiple flushes at the same position,
+ * in particular as a result of programs that close the file descriptor
+ * e.g. after a dup2() for redirection.
+ */
+
+static int flush_downstream(struct xillyusb_channel *chan,
+			    long timeout,
+			    bool interruptible)
+{
+	struct xillyusb_dev *xdev = chan->xdev;
+	int chan_num = chan->chan_idx << 1;
+	long deadline, left_to_sleep;
+	int rc;
+
+	if (chan->flushed)
+		return 0;
+
+	deadline = jiffies + 1 + timeout;
+
+	if (chan->flushing) {
+		long cancel_deadline = jiffies + 1 + XILLY_RESPONSE_TIMEOUT;
+
+		chan->canceled = 0;
+		rc = xillyusb_send_opcode(xdev, chan_num,
+					  OPCODE_CANCEL_CHECKPOINT, 0);
+
+		if (rc)
+			return rc; /* Only real error, never -EINTR */
+
+		/* Ignoring interrupts. Cancellation must be handled */
+		while (!chan->canceled) {
+			left_to_sleep = cancel_deadline - ((long)jiffies);
+
+			if (left_to_sleep <= 0) {
+				report_io_error(xdev, -EIO);
+				return -EIO;
+			}
+
+			rc = wait_event_interruptible_timeout(chan->flushq,
+							      chan->canceled ||
+							      xdev->error,
+							      left_to_sleep);
+
+			if (xdev->error)
+				return xdev->error;
+		}
+	}
+
+	chan->flushing = 1;
+
+	/*
+	 * The checkpoint is given in terms of data elements, not bytes. As
+	 * a result, if less than an element's worth of data is stored in the
+	 * FIFO, it's not flushed, including the flush before closing, which
+	 * means that such data is lost. This is consistent with PCIe Xillybus.
+	 */
+
+	rc = xillyusb_send_opcode(xdev, chan_num,
+				  OPCODE_SET_CHECKPOINT,
+				  chan->out_bytes >>
+				  chan->out_log2_element_size);
+
+	if (rc)
+		return rc; /* Only real error, never -EINTR */
+
+	if (!timeout) {
+		while (chan->flushing) {
+			rc = wait_event_interruptible(chan->flushq,
+						      !chan->flushing ||
+						      xdev->error);
+			if (xdev->error)
+				return xdev->error;
+
+			if (interruptible && rc)
+				return -EINTR;
+		}
+
+		goto done;
+	}
+
+	while (chan->flushing) {
+		left_to_sleep = deadline - ((long)jiffies);
+
+		if (left_to_sleep <= 0)
+			return -ETIMEDOUT;
+
+		rc = wait_event_interruptible_timeout(chan->flushq,
+						      !chan->flushing ||
+						      xdev->error,
+						      left_to_sleep);
+
+		if (xdev->error)
+			return xdev->error;
+
+		if (interruptible && rc < 0)
+			return -EINTR;
+	}
+
+done:
+	chan->flushed = 1;
+	return 0;
+}
+
+/* request_read_anything(): Ask the FPGA for any little amount of data */
+static int request_read_anything(struct xillyusb_channel *chan,
+				 char opcode)
+{
+	struct xillyusb_dev *xdev = chan->xdev;
+	unsigned int sh = chan->in_log2_element_size;
+	int chan_num = (chan->chan_idx << 1) | 1;
+	u32 mercy = chan->in_consumed_bytes + (2 << sh) - 1;
+
+	return xillyusb_send_opcode(xdev, chan_num, opcode, mercy >> sh);
+}
+
+static int xillyusb_open(struct inode *inode, struct file *filp)
+{
+	struct xillyusb_dev *xdev;
+	struct xillyusb_channel *chan;
+	struct xillyfifo *in_fifo = NULL;
+	struct xillyusb_endpoint *out_ep = NULL;
+	int rc;
+	int index;
+
+	rc = xillybus_find_inode(inode, (void **)&xdev, &index);
+	if (rc)
+		return rc;
+
+	chan = &xdev->channels[index];
+	filp->private_data = chan;
+
+	mutex_lock(&chan->lock);
+
+	rc = -ENODEV;
+
+	if (xdev->error)
+		goto unmutex_fail;
+
+	if (((filp->f_mode & FMODE_READ) && !chan->readable) ||
+	    ((filp->f_mode & FMODE_WRITE) && !chan->writable))
+		goto unmutex_fail;
+
+	if ((filp->f_flags & O_NONBLOCK) && (filp->f_mode & FMODE_READ) &&
+	    chan->in_synchronous) {
+		dev_err(xdev->dev,
+			"open() failed: O_NONBLOCK not allowed for read on this device\n");
+		goto unmutex_fail;
+	}
+
+	if ((filp->f_flags & O_NONBLOCK) && (filp->f_mode & FMODE_WRITE) &&
+	    chan->out_synchronous) {
+		dev_err(xdev->dev,
+			"open() failed: O_NONBLOCK not allowed for write on this device\n");
+		goto unmutex_fail;
+	}
+
+	rc = -EBUSY;
+
+	if (((filp->f_mode & FMODE_READ) && chan->open_for_read) ||
+	    ((filp->f_mode & FMODE_WRITE) && chan->open_for_write))
+		goto unmutex_fail;
+
+	kref_get(&xdev->kref);
+
+	if (filp->f_mode & FMODE_READ)
+		chan->open_for_read = 1;
+
+	if (filp->f_mode & FMODE_WRITE)
+		chan->open_for_write = 1;
+
+	mutex_unlock(&chan->lock);
+
+	if (filp->f_mode & FMODE_WRITE) {
+		out_ep = endpoint_alloc(xdev,
+					(chan->chan_idx + 2) | USB_DIR_OUT,
+					bulk_out_work, BUF_SIZE_ORDER, BUFNUM);
+
+		if (!out_ep) {
+			rc = -ENOMEM;
+			goto unopen;
+		}
+
+		rc = fifo_init(&out_ep->fifo, chan->out_log2_fifo_size);
+
+		if (rc)
+			goto late_unopen;
+
+		out_ep->fill_mask = -(1 << chan->out_log2_element_size);
+		chan->out_bytes = 0;
+		chan->flushed = 0;
+
+		/*
+		 * Sending a flush request to a previously closed stream
+		 * effectively opens it, and also waits until the command is
+		 * confirmed by the FPGA. The latter is necessary because the
+		 * data is sent through a separate BULK OUT endpoint, and the
+		 * xHCI controller is free to reorder transmissions.
+		 *
+		 * This can't go wrong unless there's a serious hardware error
+		 * (or the computer is stuck for 500 ms?)
+		 */
+		rc = flush_downstream(chan, XILLY_RESPONSE_TIMEOUT, false);
+
+		if (rc == -ETIMEDOUT) {
+			rc = -EIO;
+			report_io_error(xdev, rc);
+		}
+
+		if (rc)
+			goto late_unopen;
+	}
+
+	if (filp->f_mode & FMODE_READ) {
+		in_fifo = kzalloc(sizeof(*in_fifo), GFP_KERNEL);
+
+		if (!in_fifo) {
+			rc = -ENOMEM;
+			goto late_unopen;
+		}
+
+		rc = fifo_init(in_fifo, chan->in_log2_fifo_size);
+
+		if (rc) {
+			kfree(in_fifo);
+			goto late_unopen;
+		}
+	}
+
+	mutex_lock(&chan->lock);
+	if (in_fifo) {
+		chan->in_fifo = in_fifo;
+		chan->read_data_ok = 1;
+	}
+	if (out_ep)
+		chan->out_ep = out_ep;
+	mutex_unlock(&chan->lock);
+
+	if (in_fifo) {
+		u32 in_checkpoint = 0;
+
+		if (!chan->in_synchronous)
+			in_checkpoint = in_fifo->size >>
+				chan->in_log2_element_size;
+
+		chan->in_consumed_bytes = 0;
+		chan->poll_used = 0;
+		chan->in_current_checkpoint = in_checkpoint;
+		rc = xillyusb_send_opcode(xdev, (chan->chan_idx << 1) | 1,
+					  OPCODE_SET_CHECKPOINT,
+					  in_checkpoint);
+
+		if (rc) /* Failure guarantees that opcode wasn't sent */
+			goto unfifo;
+
+		/*
+		 * In non-blocking mode, request the FPGA to send any data it
+		 * has right away. Otherwise, the first read() will always
+		 * return -EAGAIN, which is OK strictly speaking, but ugly.
+		 * Checking and unrolling if this fails isn't worth the
+		 * effort -- the error is propagated to the first read()
+		 * anyhow.
+		 */
+		if (filp->f_flags & O_NONBLOCK)
+			request_read_anything(chan, OPCODE_SET_PUSH);
+	}
+
+	return 0;
+
+unfifo:
+	chan->read_data_ok = 0;
+	safely_assign_in_fifo(chan, NULL);
+	fifo_mem_release(in_fifo);
+	kfree(in_fifo);
+
+	if (out_ep) {
+		mutex_lock(&chan->lock);
+		chan->out_ep = NULL;
+		mutex_unlock(&chan->lock);
+	}
+
+late_unopen:
+	if (out_ep)
+		endpoint_dealloc(out_ep);
+
+unopen:
+	mutex_lock(&chan->lock);
+
+	if (filp->f_mode & FMODE_READ)
+		chan->open_for_read = 0;
+
+	if (filp->f_mode & FMODE_WRITE)
+		chan->open_for_write = 0;
+
+	mutex_unlock(&chan->lock);
+
+	kref_put(&xdev->kref, cleanup_dev);
+
+	return rc;
+
+unmutex_fail:
+	mutex_unlock(&chan->lock);
+	return rc;
+}
+
+static ssize_t xillyusb_read(struct file *filp, char __user *userbuf,
+			     size_t count, loff_t *f_pos)
+{
+	struct xillyusb_channel *chan = filp->private_data;
+	struct xillyusb_dev *xdev = chan->xdev;
+	struct xillyfifo *fifo = chan->in_fifo;
+	int chan_num = (chan->chan_idx << 1) | 1;
+
+	long deadline, left_to_sleep;
+	int bytes_done = 0;
+	bool sent_set_push = false;
+	int rc;
+
+	deadline = jiffies + 1 + XILLY_RX_TIMEOUT;
+
+	rc = mutex_lock_interruptible(&chan->in_mutex);
+
+	if (rc)
+		return rc;
+
+	while (1) {
+		u32 fifo_checkpoint_bytes, complete_checkpoint_bytes;
+		u32 complete_checkpoint, fifo_checkpoint;
+		u32 checkpoint;
+		s32 diff, leap;
+		unsigned int sh = chan->in_log2_element_size;
+		bool checkpoint_for_complete;
+
+		rc = fifo_read(fifo, (__force void *)userbuf + bytes_done,
+			       count - bytes_done, xilly_copy_to_user);
+
+		if (rc < 0)
+			break;
+
+		bytes_done += rc;
+		chan->in_consumed_bytes += rc;
+
+		left_to_sleep = deadline - ((long)jiffies);
+
+		/*
+		 * Some 32-bit arithmetic that may wrap. Note that
+		 * complete_checkpoint is rounded up to the closest element
+		 * boundary, because the read() can't be completed otherwise.
+		 * fifo_checkpoint_bytes is rounded down, because it protects
+		 * in_fifo from overflowing.
+		 */
+
+		fifo_checkpoint_bytes = chan->in_consumed_bytes + fifo->size;
+		complete_checkpoint_bytes =
+			chan->in_consumed_bytes + count - bytes_done;
+
+		fifo_checkpoint = fifo_checkpoint_bytes >> sh;
+		complete_checkpoint =
+			(complete_checkpoint_bytes + (1 << sh) - 1) >> sh;
+
+		diff = (fifo_checkpoint - complete_checkpoint) << sh;
+
+		if (chan->in_synchronous && diff >= 0) {
+			checkpoint = complete_checkpoint;
+			checkpoint_for_complete = true;
+		} else {
+			checkpoint = fifo_checkpoint;
+			checkpoint_for_complete = false;
+		}
+
+		leap = (checkpoint - chan->in_current_checkpoint) << sh;
+
+		/*
+		 * To prevent flooding of OPCODE_SET_CHECKPOINT commands as
+		 * data is consumed, it's issued only if it moves the
+		 * checkpoint by at least an 8th of the FIFO's size, or if
+		 * it's necessary to complete the number of bytes requested by
+		 * the read() call.
+		 *
+		 * chan->read_data_ok is checked to spare an unnecessary
+		 * submission after receiving EOF, however it's harmless if
+		 * such slips away.
+		 */
+
+		if (chan->read_data_ok &&
+		    (leap > (fifo->size >> 3) ||
+		     (checkpoint_for_complete && leap > 0))) {
+			chan->in_current_checkpoint = checkpoint;
+			rc = xillyusb_send_opcode(xdev, chan_num,
+						  OPCODE_SET_CHECKPOINT,
+						  checkpoint);
+
+			if (rc)
+				break;
+		}
+
+		if (bytes_done == count ||
+		    (left_to_sleep <= 0 && bytes_done))
+			break;
+
+		/*
+		 * Reaching here means that the FIFO was empty when
+		 * fifo_read() returned, but not necessarily right now. Error
+		 * and EOF are checked and reported only now, so that no data
+		 * that managed its way to the FIFO is lost.
+		 */
+
+		if (!READ_ONCE(chan->read_data_ok)) { /* FPGA has sent EOF */
+			/* Has data slipped into the FIFO since fifo_read()? */
+			smp_rmb();
+			if (READ_ONCE(fifo->fill))
+				continue;
+
+			rc = 0;
+			break;
+		}
+
+		if (xdev->error) {
+			rc = xdev->error;
+			break;
+		}
+
+		if (filp->f_flags & O_NONBLOCK) {
+			rc = -EAGAIN;
+			break;
+		}
+
+		if (!sent_set_push) {
+			rc = xillyusb_send_opcode(xdev, chan_num,
+						  OPCODE_SET_PUSH,
+						  complete_checkpoint);
+
+			if (rc)
+				break;
+
+			sent_set_push = true;
+		}
+
+		if (left_to_sleep > 0) {
+			/*
+			 * Note that when xdev->error is set (e.g. when the
+			 * device is unplugged), read_data_ok turns zero and
+			 * fifo->waitq is awaken.
+			 * Therefore no special attention to xdev->error.
+			 */
+
+			rc = wait_event_interruptible_timeout
+				(fifo->waitq,
+				 fifo->fill || !chan->read_data_ok,
+				 left_to_sleep);
+		} else { /* bytes_done == 0 */
+			/* Tell FPGA to send anything it has */
+			rc = request_read_anything(chan, OPCODE_UPDATE_PUSH);
+
+			if (rc)
+				break;
+
+			rc = wait_event_interruptible
+				(fifo->waitq,
+				 fifo->fill || !chan->read_data_ok);
+		}
+
+		if (rc < 0) {
+			rc = -EINTR;
+			break;
+		}
+	}
+
+	if (((filp->f_flags & O_NONBLOCK) || chan->poll_used) &&
+	    !READ_ONCE(fifo->fill))
+		request_read_anything(chan, OPCODE_SET_PUSH);
+
+	mutex_unlock(&chan->in_mutex);
+
+	if (bytes_done)
+		return bytes_done;
+
+	return rc;
+}
+
+static int xillyusb_flush(struct file *filp, fl_owner_t id)
+{
+	struct xillyusb_channel *chan = filp->private_data;
+	int rc;
+
+	if (!(filp->f_mode & FMODE_WRITE))
+		return 0;
+
+	rc = mutex_lock_interruptible(&chan->out_mutex);
+
+	if (rc)
+		return rc;
+
+	/*
+	 * One second's timeout on flushing. Interrupts are ignored, because if
+	 * the user pressed CTRL-C, that interrupt will still be in flight by
+	 * the time we reach here, and the opportunity to flush is lost.
+	 */
+	rc = flush_downstream(chan, HZ, false);
+
+	mutex_unlock(&chan->out_mutex);
+
+	if (rc == -ETIMEDOUT) {
+		/* The things you do to use dev_warn() and not pr_warn() */
+		struct xillyusb_dev *xdev = chan->xdev;
+
+		mutex_lock(&chan->lock);
+		if (!xdev->error)
+			dev_warn(xdev->dev,
+				 "Timed out while flushing. Output data may be lost.\n");
+		mutex_unlock(&chan->lock);
+	}
+
+	return rc;
+}
+
+static ssize_t xillyusb_write(struct file *filp, const char __user *userbuf,
+			      size_t count, loff_t *f_pos)
+{
+	struct xillyusb_channel *chan = filp->private_data;
+	struct xillyusb_dev *xdev = chan->xdev;
+	struct xillyfifo *fifo = &chan->out_ep->fifo;
+	int rc;
+
+	rc = mutex_lock_interruptible(&chan->out_mutex);
+
+	if (rc)
+		return rc;
+
+	while (1) {
+		if (xdev->error) {
+			rc = xdev->error;
+			break;
+		}
+
+		if (count == 0)
+			break;
+
+		rc = fifo_write(fifo, (__force void *)userbuf, count,
+				xilly_copy_from_user);
+
+		if (rc != 0)
+			break;
+
+		if (filp->f_flags & O_NONBLOCK) {
+			rc = -EAGAIN;
+			break;
+		}
+
+		if (wait_event_interruptible
+		    (fifo->waitq,
+		     fifo->fill != fifo->size || xdev->error)) {
+			rc = -EINTR;
+			break;
+		}
+	}
+
+	if (rc < 0)
+		goto done;
+
+	chan->out_bytes += rc;
+
+	if (rc) {
+		try_queue_bulk_out(chan->out_ep);
+		chan->flushed = 0;
+	}
+
+	if (chan->out_synchronous) {
+		int flush_rc = flush_downstream(chan, 0, true);
+
+		if (flush_rc && !rc)
+			rc = flush_rc;
+	}
+
+done:
+	mutex_unlock(&chan->out_mutex);
+
+	return rc;
+}
+
+static int xillyusb_release(struct inode *inode, struct file *filp)
+{
+	struct xillyusb_channel *chan = filp->private_data;
+	struct xillyusb_dev *xdev = chan->xdev;
+	int rc_read = 0, rc_write = 0;
+
+	if (filp->f_mode & FMODE_READ) {
+		struct xillyfifo *in_fifo = chan->in_fifo;
+
+		rc_read = xillyusb_send_opcode(xdev, (chan->chan_idx << 1) | 1,
+					       OPCODE_CLOSE, 0);
+		/*
+		 * If rc_read is nonzero, xdev->error indicates a global
+		 * device error. The error is reported later, so that
+		 * resources are freed.
+		 *
+		 * Looping on wait_event_interruptible() kinda breaks the idea
+		 * of being interruptible, and this should have been
+		 * wait_event(). Only it's being waken with
+		 * wake_up_interruptible() for the sake of other uses. If
+		 * there's a global device error, chan->read_data_ok is
+		 * deasserted and the wait queue is awaken, so this is covered.
+		 */
+
+		while (wait_event_interruptible(in_fifo->waitq,
+						!chan->read_data_ok))
+			; /* Empty loop */
+
+		safely_assign_in_fifo(chan, NULL);
+		fifo_mem_release(in_fifo);
+		kfree(in_fifo);
+
+		mutex_lock(&chan->lock);
+		chan->open_for_read = 0;
+		mutex_unlock(&chan->lock);
+	}
+
+	if (filp->f_mode & FMODE_WRITE) {
+		struct xillyusb_endpoint *ep = chan->out_ep;
+		/*
+		 * chan->flushing isn't zeroed. If the pre-release flush timed
+		 * out, a cancel request will be sent before the next
+		 * OPCODE_SET_CHECKPOINT (i.e. when the file is opened again).
+		 * This is despite that the FPGA forgets about the checkpoint
+		 * request as the file closes. Still, in an exceptional race
+		 * condition, the FPGA could send an OPCODE_REACHED_CHECKPOINT
+		 * just before closing that would reach the host after the
+		 * file has re-opened.
+		 */
+
+		mutex_lock(&chan->lock);
+		chan->out_ep = NULL;
+		mutex_unlock(&chan->lock);
+
+		endpoint_quiesce(ep);
+		endpoint_dealloc(ep);
+
+		/* See comments on rc_read above */
+		rc_write = xillyusb_send_opcode(xdev, chan->chan_idx << 1,
+						OPCODE_CLOSE, 0);
+
+		mutex_lock(&chan->lock);
+		chan->open_for_write = 0;
+		mutex_unlock(&chan->lock);
+	}
+
+	kref_put(&xdev->kref, cleanup_dev);
+
+	return rc_read ? rc_read : rc_write;
+}
+
+/*
+ * Xillybus' API allows device nodes to be seekable, giving the user
+ * application access to a RAM array on the FPGA (or logic emulating it).
+ */
+
+static loff_t xillyusb_llseek(struct file *filp, loff_t offset, int whence)
+{
+	struct xillyusb_channel *chan = filp->private_data;
+	struct xillyusb_dev *xdev = chan->xdev;
+	loff_t pos = filp->f_pos;
+	int rc = 0;
+	unsigned int log2_element_size = chan->readable ?
+		chan->in_log2_element_size : chan->out_log2_element_size;
+
+	/*
+	 * Take both mutexes not allowing interrupts, since it seems like
+	 * common applications don't expect an -EINTR here. Besides, multiple
+	 * access to a single file descriptor on seekable devices is a mess
+	 * anyhow.
+	 */
+
+	mutex_lock(&chan->out_mutex);
+	mutex_lock(&chan->in_mutex);
+
+	switch (whence) {
+	case SEEK_SET:
+		pos = offset;
+		break;
+	case SEEK_CUR:
+		pos += offset;
+		break;
+	case SEEK_END:
+		pos = offset; /* Going to the end => to the beginning */
+		break;
+	default:
+		rc = -EINVAL;
+		goto end;
+	}
+
+	/* In any case, we must finish on an element boundary */
+	if (pos & ((1 << log2_element_size) - 1)) {
+		rc = -EINVAL;
+		goto end;
+	}
+
+	rc = xillyusb_send_opcode(xdev, chan->chan_idx << 1,
+				  OPCODE_SET_ADDR,
+				  pos >> log2_element_size);
+
+	if (rc)
+		goto end;
+
+	if (chan->writable) {
+		chan->flushed = 0;
+		rc = flush_downstream(chan, HZ, false);
+	}
+
+end:
+	mutex_unlock(&chan->out_mutex);
+	mutex_unlock(&chan->in_mutex);
+
+	if (rc) /* Return error after releasing mutexes */
+		return rc;
+
+	filp->f_pos = pos;
+
+	return pos;
+}
+
+static __poll_t xillyusb_poll(struct file *filp, poll_table *wait)
+{
+	struct xillyusb_channel *chan = filp->private_data;
+	__poll_t mask = 0;
+
+	if (chan->in_fifo)
+		poll_wait(filp, &chan->in_fifo->waitq, wait);
+
+	if (chan->out_ep)
+		poll_wait(filp, &chan->out_ep->fifo.waitq, wait);
+
+	/*
+	 * If this is the first time poll() is called, and the file is
+	 * readable, set the relevant flag. Also tell the FPGA to send all it
+	 * has, to kickstart the mechanism that ensures there's always some
+	 * data in in_fifo unless the stream is dry end-to-end. Note that the
+	 * first poll() may not return a EPOLLIN, even if there's data on the
+	 * FPGA. Rather, the data will arrive soon, and trigger the relevant
+	 * wait queue.
+	 */
+
+	if (!chan->poll_used && chan->in_fifo) {
+		chan->poll_used = 1;
+		request_read_anything(chan, OPCODE_SET_PUSH);
+	}
+
+	/*
+	 * poll() won't play ball regarding read() channels which
+	 * are synchronous. Allowing that will create situations where data has
+	 * been delivered at the FPGA, and users expecting select() to wake up,
+	 * which it may not. So make it never work.
+	 */
+
+	if (chan->in_fifo && !chan->in_synchronous &&
+	    (READ_ONCE(chan->in_fifo->fill) || !chan->read_data_ok))
+		mask |= EPOLLIN | EPOLLRDNORM;
+
+	if (chan->out_ep &&
+	    (READ_ONCE(chan->out_ep->fifo.fill) != chan->out_ep->fifo.size))
+		mask |= EPOLLOUT | EPOLLWRNORM;
+
+	if (chan->xdev->error)
+		mask |= EPOLLERR;
+
+	return mask;
+}
+
+static const struct file_operations xillyusb_fops = {
+	.owner      = THIS_MODULE,
+	.read       = xillyusb_read,
+	.write      = xillyusb_write,
+	.open       = xillyusb_open,
+	.flush      = xillyusb_flush,
+	.release    = xillyusb_release,
+	.llseek     = xillyusb_llseek,
+	.poll       = xillyusb_poll,
+};
+
+static int xillyusb_setup_base_eps(struct xillyusb_dev *xdev)
+{
+	xdev->msg_ep = endpoint_alloc(xdev, MSG_EP_NUM | USB_DIR_OUT,
+				      bulk_out_work, 1, 2);
+	if (!xdev->msg_ep)
+		return -ENOMEM;
+
+	if (fifo_init(&xdev->msg_ep->fifo, 13)) /* 8 kiB */
+		goto dealloc;
+
+	xdev->msg_ep->fill_mask = -8; /* 8 bytes granularity */
+
+	xdev->in_ep = endpoint_alloc(xdev, IN_EP_NUM | USB_DIR_IN,
+				     bulk_in_work, BUF_SIZE_ORDER, BUFNUM);
+	if (!xdev->in_ep)
+		goto dealloc;
+
+	try_queue_bulk_in(xdev->in_ep);
+
+	return 0;
+
+dealloc:
+	endpoint_dealloc(xdev->msg_ep); /* Also frees FIFO mem if allocated */
+	return -ENOMEM;
+}
+
+static int setup_channels(struct xillyusb_dev *xdev,
+			  __le16 *chandesc,
+			  int num_channels)
+{
+	struct xillyusb_channel *chan;
+	int i;
+
+	chan = kcalloc(num_channels, sizeof(*chan), GFP_KERNEL);
+	if (!chan)
+		return -ENOMEM;
+
+	xdev->channels = chan;
+
+	for (i = 0; i < num_channels; i++, chan++) {
+		unsigned int in_desc = le16_to_cpu(*chandesc++);
+		unsigned int out_desc = le16_to_cpu(*chandesc++);
+
+		chan->xdev = xdev;
+		mutex_init(&chan->in_mutex);
+		mutex_init(&chan->out_mutex);
+		mutex_init(&chan->lock);
+		init_waitqueue_head(&chan->flushq);
+
+		chan->chan_idx = i;
+
+		if (in_desc & 0x80) { /* Entry is valid */
+			chan->readable = 1;
+			chan->in_synchronous = !!(in_desc & 0x40);
+			chan->in_seekable = !!(in_desc & 0x20);
+			chan->in_log2_element_size = in_desc & 0x0f;
+			chan->in_log2_fifo_size = ((in_desc >> 8) & 0x1f) + 16;
+		}
+
+		/*
+		 * A downstream channel should never exist above index 13,
+		 * as it would request a nonexistent BULK endpoint > 15.
+		 * In the peculiar case that it does, it's ignored silently.
+		 */
+
+		if ((out_desc & 0x80) && i < 14) { /* Entry is valid */
+			chan->writable = 1;
+			chan->out_synchronous = !!(out_desc & 0x40);
+			chan->out_seekable = !!(out_desc & 0x20);
+			chan->out_log2_element_size = out_desc & 0x0f;
+			chan->out_log2_fifo_size =
+				((out_desc >> 8) & 0x1f) + 16;
+		}
+	}
+
+	return 0;
+}
+
+static int xillyusb_discovery(struct usb_interface *interface)
+{
+	int rc;
+	struct xillyusb_dev *xdev = usb_get_intfdata(interface);
+	__le16 bogus_chandesc[2];
+	struct xillyfifo idt_fifo;
+	struct xillyusb_channel *chan;
+	unsigned int idt_len, names_offset;
+	unsigned char *idt;
+	int num_channels;
+
+	rc = xillyusb_send_opcode(xdev, ~0, OPCODE_QUIESCE, 0);
+
+	if (rc) {
+		dev_err(&interface->dev, "Failed to send quiesce request. Aborting.\n");
+		return rc;
+	}
+
+	/* Phase I: Set up one fake upstream channel and obtain IDT */
+
+	/* Set up a fake IDT with one async IN stream */
+	bogus_chandesc[0] = cpu_to_le16(0x80);
+	bogus_chandesc[1] = cpu_to_le16(0);
+
+	rc = setup_channels(xdev, bogus_chandesc, 1);
+
+	if (rc)
+		return rc;
+
+	rc = fifo_init(&idt_fifo, LOG2_IDT_FIFO_SIZE);
+
+	if (rc)
+		return rc;
+
+	chan = xdev->channels;
+
+	chan->in_fifo = &idt_fifo;
+	chan->read_data_ok = 1;
+
+	xdev->num_channels = 1;
+
+	rc = xillyusb_send_opcode(xdev, ~0, OPCODE_REQ_IDT, 0);
+
+	if (rc) {
+		dev_err(&interface->dev, "Failed to send IDT request. Aborting.\n");
+		goto unfifo;
+	}
+
+	rc = wait_event_interruptible_timeout(idt_fifo.waitq,
+					      !chan->read_data_ok,
+					      XILLY_RESPONSE_TIMEOUT);
+
+	if (xdev->error) {
+		rc = xdev->error;
+		goto unfifo;
+	}
+
+	if (rc < 0) {
+		rc = -EINTR; /* Interrupt on probe method? Interesting. */
+		goto unfifo;
+	}
+
+	if (chan->read_data_ok) {
+		rc = -ETIMEDOUT;
+		dev_err(&interface->dev, "No response from FPGA. Aborting.\n");
+		goto unfifo;
+	}
+
+	idt_len = READ_ONCE(idt_fifo.fill);
+	idt = kmalloc(idt_len, GFP_KERNEL);
+
+	if (!idt) {
+		rc = -ENOMEM;
+		goto unfifo;
+	}
+
+	fifo_read(&idt_fifo, idt, idt_len, xilly_memcpy);
+
+	if (crc32_le(~0, idt, idt_len) != 0) {
+		dev_err(&interface->dev, "IDT failed CRC check. Aborting.\n");
+		rc = -ENODEV;
+		goto unidt;
+	}
+
+	if (*idt > 0x90) {
+		dev_err(&interface->dev, "No support for IDT version 0x%02x. Maybe the xillyusb driver needs an upgrade. Aborting.\n",
+			(int)*idt);
+		rc = -ENODEV;
+		goto unidt;
+	}
+
+	/* Phase II: Set up the streams as defined in IDT */
+
+	num_channels = le16_to_cpu(*((__le16 *)(idt + 1)));
+	names_offset = 3 + num_channels * 4;
+	idt_len -= 4; /* Exclude CRC */
+
+	if (idt_len < names_offset) {
+		dev_err(&interface->dev, "IDT too short. This is exceptionally weird, because its CRC is OK\n");
+		rc = -ENODEV;
+		goto unidt;
+	}
+
+	rc = setup_channels(xdev, (void *)idt + 3, num_channels);
+
+	if (rc)
+		goto unidt;
+
+	/*
+	 * Except for wildly misbehaving hardware, or if it was disconnected
+	 * just after responding with the IDT, there is no reason for any
+	 * work item to be running now. To be sure that xdev->channels
+	 * is updated on anything that might run in parallel, flush the
+	 * workqueue, which rarely does anything.
+	 */
+	flush_workqueue(xdev->workq);
+
+	xdev->num_channels = num_channels;
+
+	fifo_mem_release(&idt_fifo);
+	kfree(chan);
+
+	rc = xillybus_init_chrdev(&interface->dev, &xillyusb_fops,
+				  THIS_MODULE, xdev,
+				  idt + names_offset,
+				  idt_len - names_offset,
+				  num_channels,
+				  xillyname, true);
+
+	kfree(idt);
+
+	return rc;
+
+unidt:
+	kfree(idt);
+
+unfifo:
+	safely_assign_in_fifo(chan, NULL);
+	fifo_mem_release(&idt_fifo);
+
+	return rc;
+}
+
+static int xillyusb_probe(struct usb_interface *interface,
+			  const struct usb_device_id *id)
+{
+	struct xillyusb_dev *xdev;
+	int rc;
+
+	xdev = kzalloc(sizeof(*xdev), GFP_KERNEL);
+	if (!xdev)
+		return -ENOMEM;
+
+	kref_init(&xdev->kref);
+	mutex_init(&xdev->process_in_mutex);
+	mutex_init(&xdev->msg_mutex);
+
+	xdev->udev = usb_get_dev(interface_to_usbdev(interface));
+	xdev->dev = &interface->dev;
+	xdev->error = 0;
+	spin_lock_init(&xdev->error_lock);
+	xdev->in_counter = 0;
+	xdev->in_bytes_left = 0;
+	xdev->workq = alloc_workqueue(xillyname, WQ_HIGHPRI, 0);
+
+	if (!xdev->workq) {
+		dev_err(&interface->dev, "Failed to allocate work queue\n");
+		rc = -ENOMEM;
+		goto fail;
+	}
+
+	INIT_WORK(&xdev->wakeup_workitem, wakeup_all);
+
+	usb_set_intfdata(interface, xdev);
+
+	rc = xillyusb_setup_base_eps(xdev);
+	if (rc)
+		goto fail;
+
+	rc = xillyusb_discovery(interface);
+	if (rc)
+		goto latefail;
+
+	return 0;
+
+latefail:
+	endpoint_quiesce(xdev->in_ep);
+	endpoint_quiesce(xdev->msg_ep);
+
+fail:
+	usb_set_intfdata(interface, NULL);
+	kref_put(&xdev->kref, cleanup_dev);
+	return rc;
+}
+
+static void xillyusb_disconnect(struct usb_interface *interface)
+{
+	struct xillyusb_dev *xdev = usb_get_intfdata(interface);
+	struct xillyusb_endpoint *msg_ep = xdev->msg_ep;
+	struct xillyfifo *fifo = &msg_ep->fifo;
+	int rc;
+	int i;
+
+	xillybus_cleanup_chrdev(xdev, &interface->dev);
+
+	/*
+	 * Try to send OPCODE_QUIESCE, which will fail silently if the device
+	 * was disconnected, but makes sense on module unload.
+	 */
+
+	msg_ep->wake_on_drain = true;
+	xillyusb_send_opcode(xdev, ~0, OPCODE_QUIESCE, 0);
+
+	/*
+	 * If the device has been disconnected, sending the opcode causes
+	 * a global device error with xdev->error, if such error didn't
+	 * occur earlier. Hence timing out means that the USB link is fine,
+	 * but somehow the message wasn't sent. Should never happen.
+	 */
+
+	rc = wait_event_interruptible_timeout(fifo->waitq,
+					      msg_ep->drained || xdev->error,
+					      XILLY_RESPONSE_TIMEOUT);
+
+	if (!rc)
+		dev_err(&interface->dev,
+			"Weird timeout condition on sending quiesce request.\n");
+
+	report_io_error(xdev, -ENODEV); /* Discourage further activity */
+
+	/*
+	 * This device driver is declared with soft_unbind set, or else
+	 * sending OPCODE_QUIESCE above would always fail. The price is
+	 * that the USB framework didn't kill outstanding URBs, so it has
+	 * to be done explicitly before returning from this call.
+	 */
+
+	for (i = 0; i < xdev->num_channels; i++) {
+		struct xillyusb_channel *chan = &xdev->channels[i];
+
+		/*
+		 * Lock taken to prevent chan->out_ep from changing. It also
+		 * ensures xillyusb_open() and xillyusb_flush() don't access
+		 * xdev->dev after being nullified below.
+		 */
+		mutex_lock(&chan->lock);
+		if (chan->out_ep)
+			endpoint_quiesce(chan->out_ep);
+		mutex_unlock(&chan->lock);
+	}
+
+	endpoint_quiesce(xdev->in_ep);
+	endpoint_quiesce(xdev->msg_ep);
+
+	usb_set_intfdata(interface, NULL);
+
+	xdev->dev = NULL;
+
+	kref_put(&xdev->kref, cleanup_dev);
+}
+
+static struct usb_driver xillyusb_driver = {
+	.name = xillyname,
+	.id_table = xillyusb_table,
+	.probe = xillyusb_probe,
+	.disconnect = xillyusb_disconnect,
+	.soft_unbind = 1,
+};
+
+static int __init xillyusb_init(void)
+{
+	int rc = 0;
+
+	if (LOG2_INITIAL_FIFO_BUF_SIZE > PAGE_SHIFT)
+		fifo_buf_order = LOG2_INITIAL_FIFO_BUF_SIZE - PAGE_SHIFT;
+	else
+		fifo_buf_order = 0;
+
+	rc = usb_register(&xillyusb_driver);
+
+	return rc;
+}
+
+static void __exit xillyusb_exit(void)
+{
+	usb_deregister(&xillyusb_driver);
+}
+
+module_init(xillyusb_init);
+module_exit(xillyusb_exit);