ftrace: Synchronize the initialization of ftrace_opsgrafted

There's some internal state that ftrace_ops needs to have set, but since
it can be declared outside of the ftrace.c code, it calls
ftrace_ops_init() on the ops in every global function. The issue is that
if two tasks call it on the same ops at the same time it is possible to
have the initialization of one corrupt the initialization of the other
call.

Create a ops_mutex to use to synchronize every initialization of the
ftrace_ops. The mutex is taken within checking the ftrace_ops flag that
states it was initializied but the flag is checked again after the mutex
has been taken. Checking first outside the mutex allows it to shortcut
having to take the mutex. But then the check needs to be done again after
the mute is taken in case of races.

Cc: stable@vger.kernel.org
Link: https://patch.msgid.link/20260902095501.6b59af20@gandalf.local.home
Fixes: f04f24fb7e48d ("ftrace, kprobes: Fix a deadlock on ftrace_regex_lock")
Reported-by: sashiko-bot@kernel.org
Close: https://lore.kernel.org/all/20260829025528.49A831F000E9@smtp.kernel.org/
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

9 files changed, 4972 insertions, 0 deletions

diff --git a/drivers/char/xillybus/Kconfig b/drivers/char/xillybus/Kconfig
new file mode 100644
index 000000000..f51d53339
--- /dev/null
+++ b/drivers/char/xillybus/Kconfig
@@ -0,0 +1,52 @@
+# SPDX-License-Identifier: GPL-2.0-only
+#
+# Xillybus devices
+#
+
+config XILLYBUS_CLASS
+	tristate
+
+config XILLYBUS
+	tristate "Xillybus generic FPGA interface"
+	depends on PCI || OF
+	select CRC32
+	select XILLYBUS_CLASS
+	help
+	  Xillybus is a generic interface for peripherals designed on
+	  programmable logic (FPGA). The driver probes the hardware for
+	  its capabilities, and creates device files accordingly.
+
+	  If unsure, say N.
+
+if XILLYBUS
+
+config XILLYBUS_PCIE
+	tristate "Xillybus over PCIe"
+	depends on PCI_MSI
+	help
+	  Set to M if you want Xillybus to use PCI Express for communicating
+	  with the FPGA. The module will be called xillybus_pcie.
+
+config XILLYBUS_OF
+	tristate "Xillybus over Device Tree"
+	depends on OF && HAS_DMA && HAS_IOMEM
+	help
+	  Set to M if you want Xillybus to find its resources from the
+	  Open Firmware Flattened Device Tree. If the target is an embedded
+	  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.
diff --git a/drivers/char/xillybus/Makefile b/drivers/char/xillybus/Makefile
new file mode 100644
index 000000000..16f31d032
--- /dev/null
+++ b/drivers/char/xillybus/Makefile
@@ -0,0 +1,10 @@
+# SPDX-License-Identifier: GPL-2.0-only
+#
+# Makefile for Xillybus driver
+#
+
+obj-$(CONFIG_XILLYBUS_CLASS)	+= xillybus_class.o
+obj-$(CONFIG_XILLYBUS)		+= xillybus_core.o
+obj-$(CONFIG_XILLYBUS_PCIE)	+= xillybus_pcie.o
+obj-$(CONFIG_XILLYBUS_OF)	+= xillybus_of.o
+obj-$(CONFIG_XILLYUSB)		+= xillyusb.o
diff --git a/drivers/char/xillybus/xillybus.h b/drivers/char/xillybus/xillybus.h
new file mode 100644
index 000000000..51de7cbc5
--- /dev/null
+++ b/drivers/char/xillybus/xillybus.h
@@ -0,0 +1,126 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * linux/drivers/misc/xillybus.h
+ *
+ * Copyright 2011 Xillybus Ltd, http://xillybus.com
+ *
+ * Header file for the Xillybus FPGA/host framework.
+ */
+
+#ifndef __XILLYBUS_H
+#define __XILLYBUS_H
+
+#include <linux/list.h>
+#include <linux/device.h>
+#include <linux/dma-mapping.h>
+#include <linux/interrupt.h>
+#include <linux/sched.h>
+#include <linux/cdev.h>
+#include <linux/spinlock.h>
+#include <linux/mutex.h>
+#include <linux/workqueue.h>
+
+struct xilly_endpoint_hardware;
+
+struct xilly_buffer {
+	void *addr;
+	dma_addr_t dma_addr;
+	int end_offset; /* Counting elements, not bytes */
+};
+
+struct xilly_idt_handle {
+	unsigned char *chandesc;
+	unsigned char *names;
+	int names_len;
+	int entries;
+};
+
+/*
+ * Read-write confusion: wr_* and rd_* notation sticks to FPGA view, so
+ * wr_* buffers are those consumed by read(), since the FPGA writes to them
+ * and vice versa.
+ */
+
+struct xilly_channel {
+	struct xilly_endpoint *endpoint;
+	int chan_num;
+	int log2_element_size;
+	int seekable;
+
+	struct xilly_buffer **wr_buffers; /* FPGA writes, driver reads! */
+	int num_wr_buffers;
+	unsigned int wr_buf_size; /* In bytes */
+	int wr_fpga_buf_idx;
+	int wr_host_buf_idx;
+	int wr_host_buf_pos;
+	int wr_empty;
+	int wr_ready; /* Significant only when wr_empty == 1 */
+	int wr_sleepy;
+	int wr_eof;
+	int wr_hangup;
+	spinlock_t wr_spinlock;
+	struct mutex wr_mutex;
+	wait_queue_head_t wr_wait;
+	wait_queue_head_t wr_ready_wait;
+	int wr_ref_count;
+	int wr_synchronous;
+	int wr_allow_partial;
+	int wr_exclusive_open;
+	int wr_supports_nonempty;
+
+	struct xilly_buffer **rd_buffers; /* FPGA reads, driver writes! */
+	int num_rd_buffers;
+	unsigned int rd_buf_size; /* In bytes */
+	int rd_fpga_buf_idx;
+	int rd_host_buf_pos;
+	int rd_host_buf_idx;
+	int rd_full;
+	spinlock_t rd_spinlock;
+	struct mutex rd_mutex;
+	wait_queue_head_t rd_wait;
+	int rd_ref_count;
+	int rd_allow_partial;
+	int rd_synchronous;
+	int rd_exclusive_open;
+	struct delayed_work rd_workitem;
+	unsigned char rd_leftovers[4];
+};
+
+struct xilly_endpoint {
+	struct device *dev;
+	struct module *owner;
+
+	int dma_using_dac; /* =1 if 64-bit DMA is used, =0 otherwise. */
+	__iomem void *registers;
+	int fatal_error;
+
+	struct mutex register_mutex;
+	wait_queue_head_t ep_wait;
+
+	int num_channels; /* EXCLUDING message buffer */
+	struct xilly_channel **channels;
+	int msg_counter;
+	int failed_messages;
+	int idtlen;
+
+	u32 *msgbuf_addr;
+	dma_addr_t msgbuf_dma_addr;
+	unsigned int msg_buf_size;
+};
+
+struct xilly_mapping {
+	struct device *device;
+	dma_addr_t dma_addr;
+	size_t size;
+	int direction;
+};
+
+irqreturn_t xillybus_isr(int irq, void *data);
+
+struct xilly_endpoint *xillybus_init_endpoint(struct device *dev);
+
+int xillybus_endpoint_discovery(struct xilly_endpoint *endpoint);
+
+void xillybus_endpoint_remove(struct xilly_endpoint *endpoint);
+
+#endif /* __XILLYBUS_H */
diff --git a/drivers/char/xillybus/xillybus_class.c b/drivers/char/xillybus/xillybus_class.c
new file mode 100644
index 000000000..5e8f03b77
--- /dev/null
+++ b/drivers/char/xillybus/xillybus_class.c
@@ -0,0 +1,256 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright 2021 Xillybus Ltd, http://xillybus.com
+ *
+ * Driver for the Xillybus class
+ */
+
+#include <linux/types.h>
+#include <linux/module.h>
+#include <linux/device.h>
+#include <linux/fs.h>
+#include <linux/cdev.h>
+#include <linux/slab.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
+
+#include "xillybus_class.h"
+
+MODULE_DESCRIPTION("Driver for Xillybus class");
+MODULE_AUTHOR("Eli Billauer, Xillybus Ltd.");
+MODULE_ALIAS("xillybus_class");
+MODULE_LICENSE("GPL v2");
+
+static DEFINE_MUTEX(unit_mutex);
+static LIST_HEAD(unit_list);
+static const struct class xillybus_class = {
+	.name = "xillybus",
+};
+
+#define UNITNAMELEN 16
+
+struct xilly_unit {
+	struct list_head list_entry;
+	void *private_data;
+
+	struct cdev *cdev;
+	char name[UNITNAMELEN];
+	int major;
+	int lowest_minor;
+	int num_nodes;
+};
+
+int xillybus_init_chrdev(struct device *dev,
+			 const struct file_operations *fops,
+			 struct module *owner,
+			 void *private_data,
+			 unsigned char *idt, unsigned int len,
+			 int num_nodes,
+			 const char *prefix, bool enumerate)
+{
+	int rc;
+	dev_t mdev;
+	int i;
+	char devname[48];
+
+	struct device *device;
+	size_t namelen;
+	struct xilly_unit *unit, *u;
+
+	unit = kzalloc_obj(*unit);
+
+	if (!unit)
+		return -ENOMEM;
+
+	mutex_lock(&unit_mutex);
+
+	if (!enumerate)
+		snprintf(unit->name, UNITNAMELEN, "%s", prefix);
+
+	for (i = 0; enumerate; i++) {
+		snprintf(unit->name, UNITNAMELEN, "%s_%02d",
+			 prefix, i);
+
+		enumerate = false;
+		list_for_each_entry(u, &unit_list, list_entry)
+			if (!strcmp(unit->name, u->name)) {
+				enumerate = true;
+				break;
+			}
+	}
+
+	rc = alloc_chrdev_region(&mdev, 0, num_nodes, unit->name);
+
+	if (rc) {
+		dev_warn(dev, "Failed to obtain major/minors");
+		goto fail_obtain;
+	}
+
+	unit->major = MAJOR(mdev);
+	unit->lowest_minor = MINOR(mdev);
+	unit->num_nodes = num_nodes;
+	unit->private_data = private_data;
+
+	unit->cdev = cdev_alloc();
+	if (!unit->cdev) {
+		rc = -ENOMEM;
+		goto unregister_chrdev;
+	}
+	unit->cdev->ops = fops;
+	unit->cdev->owner = owner;
+
+	rc = cdev_add(unit->cdev, MKDEV(unit->major, unit->lowest_minor),
+		      unit->num_nodes);
+	if (rc) {
+		dev_err(dev, "Failed to add cdev.\n");
+		/* kobject_put() is normally done by cdev_del() */
+		kobject_put(&unit->cdev->kobj);
+		goto unregister_chrdev;
+	}
+
+	for (i = 0; i < num_nodes; i++) {
+		namelen = strnlen(idt, len);
+
+		if (namelen == len) {
+			dev_err(dev, "IDT's list of names is too short. This is exceptionally weird, because its CRC is OK\n");
+			rc = -ENODEV;
+			goto unroll_device_create;
+		}
+
+		snprintf(devname, sizeof(devname), "%s_%s",
+			 unit->name, idt);
+
+		len -= namelen + 1;
+		idt += namelen + 1;
+
+		device = device_create(&xillybus_class,
+				       NULL,
+				       MKDEV(unit->major,
+					     i + unit->lowest_minor),
+				       NULL,
+				       "%s", devname);
+
+		if (IS_ERR(device)) {
+			dev_err(dev, "Failed to create %s device. Aborting.\n",
+				devname);
+			rc = -ENODEV;
+			goto unroll_device_create;
+		}
+	}
+
+	if (len) {
+		dev_err(dev, "IDT's list of names is too long. This is exceptionally weird, because its CRC is OK\n");
+		rc = -ENODEV;
+		goto unroll_device_create;
+	}
+
+	list_add_tail(&unit->list_entry, &unit_list);
+
+	dev_info(dev, "Created %d device files.\n", num_nodes);
+
+	mutex_unlock(&unit_mutex);
+
+	return 0;
+
+unroll_device_create:
+	for (i--; i >= 0; i--)
+		device_destroy(&xillybus_class, MKDEV(unit->major,
+						     i + unit->lowest_minor));
+
+	cdev_del(unit->cdev);
+
+unregister_chrdev:
+	unregister_chrdev_region(MKDEV(unit->major, unit->lowest_minor),
+				 unit->num_nodes);
+
+fail_obtain:
+	mutex_unlock(&unit_mutex);
+
+	kfree(unit);
+
+	return rc;
+}
+EXPORT_SYMBOL(xillybus_init_chrdev);
+
+void xillybus_cleanup_chrdev(void *private_data,
+			     struct device *dev)
+{
+	int minor;
+	struct xilly_unit *unit = NULL, *iter;
+
+	mutex_lock(&unit_mutex);
+
+	list_for_each_entry(iter, &unit_list, list_entry)
+		if (iter->private_data == private_data) {
+			unit = iter;
+			break;
+		}
+
+	if (!unit) {
+		dev_err(dev, "Weird bug: Failed to find unit\n");
+		mutex_unlock(&unit_mutex);
+		return;
+	}
+
+	for (minor = unit->lowest_minor;
+	     minor < (unit->lowest_minor + unit->num_nodes);
+	     minor++)
+		device_destroy(&xillybus_class, MKDEV(unit->major, minor));
+
+	cdev_del(unit->cdev);
+
+	unregister_chrdev_region(MKDEV(unit->major, unit->lowest_minor),
+				 unit->num_nodes);
+
+	dev_info(dev, "Removed %d device files.\n",
+		 unit->num_nodes);
+
+	list_del(&unit->list_entry);
+	kfree(unit);
+
+	mutex_unlock(&unit_mutex);
+}
+EXPORT_SYMBOL(xillybus_cleanup_chrdev);
+
+int xillybus_find_inode(struct inode *inode,
+			void **private_data, int *index)
+{
+	int minor = iminor(inode);
+	int major = imajor(inode);
+	struct xilly_unit *unit = NULL, *iter;
+
+	mutex_lock(&unit_mutex);
+
+	list_for_each_entry(iter, &unit_list, list_entry)
+		if (iter->major == major &&
+		    minor >= iter->lowest_minor &&
+		    minor < (iter->lowest_minor + iter->num_nodes)) {
+			unit = iter;
+			break;
+		}
+
+	if (!unit) {
+		mutex_unlock(&unit_mutex);
+		return -ENODEV;
+	}
+
+	*private_data = unit->private_data;
+	*index = minor - unit->lowest_minor;
+
+	mutex_unlock(&unit_mutex);
+	return 0;
+}
+EXPORT_SYMBOL(xillybus_find_inode);
+
+static int __init xillybus_class_init(void)
+{
+	return class_register(&xillybus_class);
+}
+
+static void __exit xillybus_class_exit(void)
+{
+	class_unregister(&xillybus_class);
+}
+
+module_init(xillybus_class_init);
+module_exit(xillybus_class_exit);
diff --git a/drivers/char/xillybus/xillybus_class.h b/drivers/char/xillybus/xillybus_class.h
new file mode 100644
index 000000000..5dbfdfc95
--- /dev/null
+++ b/drivers/char/xillybus/xillybus_class.h
@@ -0,0 +1,30 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright 2021 Xillybus Ltd, http://www.xillybus.com
+ *
+ * Header file for the Xillybus class
+ */
+
+#ifndef __XILLYBUS_CLASS_H
+#define __XILLYBUS_CLASS_H
+
+#include <linux/types.h>
+#include <linux/device.h>
+#include <linux/fs.h>
+#include <linux/module.h>
+
+int xillybus_init_chrdev(struct device *dev,
+			 const struct file_operations *fops,
+			 struct module *owner,
+			 void *private_data,
+			 unsigned char *idt, unsigned int len,
+			 int num_nodes,
+			 const char *prefix, bool enumerate);
+
+void xillybus_cleanup_chrdev(void *private_data,
+			     struct device *dev);
+
+int xillybus_find_inode(struct inode *inode,
+			void **private_data, int *index);
+
+#endif /* __XILLYBUS_CLASS_H */
diff --git a/drivers/char/xillybus/xillybus_core.c b/drivers/char/xillybus/xillybus_core.c
new file mode 100644
index 000000000..952ef149a
--- /dev/null
+++ b/drivers/char/xillybus/xillybus_core.c
@@ -0,0 +1,1990 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * linux/drivers/misc/xillybus_core.c
+ *
+ * Copyright 2011 Xillybus Ltd, http://xillybus.com
+ *
+ * Driver for the Xillybus 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/list.h>
+#include <linux/device.h>
+#include <linux/module.h>
+#include <linux/io.h>
+#include <linux/dma-mapping.h>
+#include <linux/interrupt.h>
+#include <linux/sched.h>
+#include <linux/fs.h>
+#include <linux/spinlock.h>
+#include <linux/mutex.h>
+#include <linux/crc32.h>
+#include <linux/poll.h>
+#include <linux/delay.h>
+#include <linux/slab.h>
+#include <linux/workqueue.h>
+#include "xillybus.h"
+#include "xillybus_class.h"
+
+MODULE_DESCRIPTION("Xillybus core functions");
+MODULE_AUTHOR("Eli Billauer, Xillybus Ltd.");
+MODULE_ALIAS("xillybus_core");
+MODULE_LICENSE("GPL v2");
+
+/* General timeout is 100 ms, rx timeout is 10 ms */
+#define XILLY_RX_TIMEOUT (10*HZ/1000)
+#define XILLY_TIMEOUT (100*HZ/1000)
+
+#define fpga_msg_ctrl_reg              0x0008
+#define fpga_dma_control_reg           0x0020
+#define fpga_dma_bufno_reg             0x0024
+#define fpga_dma_bufaddr_lowaddr_reg   0x0028
+#define fpga_dma_bufaddr_highaddr_reg  0x002c
+#define fpga_buf_ctrl_reg              0x0030
+#define fpga_buf_offset_reg            0x0034
+#define fpga_endian_reg                0x0040
+
+#define XILLYMSG_OPCODE_RELEASEBUF 1
+#define XILLYMSG_OPCODE_QUIESCEACK 2
+#define XILLYMSG_OPCODE_FIFOEOF 3
+#define XILLYMSG_OPCODE_FATAL_ERROR 4
+#define XILLYMSG_OPCODE_NONEMPTY 5
+
+static const char xillyname[] = "xillybus";
+
+static struct workqueue_struct *xillybus_wq;
+
+/*
+ * Locking scheme: Mutexes protect invocations of character device methods.
+ * If both locks are taken, wr_mutex is taken first, rd_mutex second.
+ *
+ * wr_spinlock protects wr_*_buf_idx, wr_empty, wr_sleepy, wr_ready and the
+ * buffers' end_offset fields against changes made by IRQ handler (and in
+ * theory, other file request handlers, but the mutex handles that). Nothing
+ * else.
+ * They are held for short direct memory manipulations. Needless to say,
+ * no mutex locking is allowed when a spinlock is held.
+ *
+ * rd_spinlock does the same with rd_*_buf_idx, rd_empty and end_offset.
+ *
+ * register_mutex is endpoint-specific, and is held when non-atomic
+ * register operations are performed. wr_mutex and rd_mutex may be
+ * held when register_mutex is taken, but none of the spinlocks. Note that
+ * register_mutex doesn't protect against sporadic buf_ctrl_reg writes
+ * which are unrelated to buf_offset_reg, since they are harmless.
+ *
+ * Blocking on the wait queues is allowed with mutexes held, but not with
+ * spinlocks.
+ *
+ * Only interruptible blocking is allowed on mutexes and wait queues.
+ *
+ * All in all, the locking order goes (with skips allowed, of course):
+ * wr_mutex -> rd_mutex -> register_mutex -> wr_spinlock -> rd_spinlock
+ */
+
+static void malformed_message(struct xilly_endpoint *endpoint, u32 *buf)
+{
+	int opcode;
+	int msg_channel, msg_bufno, msg_data, msg_dir;
+
+	opcode = (buf[0] >> 24) & 0xff;
+	msg_dir = buf[0] & 1;
+	msg_channel = (buf[0] >> 1) & 0x7ff;
+	msg_bufno = (buf[0] >> 12) & 0x3ff;
+	msg_data = buf[1] & 0xfffffff;
+
+	dev_warn(endpoint->dev,
+		 "Malformed message (skipping): opcode=%d, channel=%03x, dir=%d, bufno=%03x, data=%07x\n",
+		 opcode, msg_channel, msg_dir, msg_bufno, msg_data);
+}
+
+/*
+ * xillybus_isr assumes the interrupt is allocated exclusively to it,
+ * which is the natural case MSI and several other hardware-oriented
+ * interrupts. Sharing is not allowed.
+ */
+
+irqreturn_t xillybus_isr(int irq, void *data)
+{
+	struct xilly_endpoint *ep = data;
+	u32 *buf;
+	unsigned int buf_size;
+	int i;
+	int opcode;
+	unsigned int msg_channel, msg_bufno, msg_data, msg_dir;
+	struct xilly_channel *channel;
+
+	buf = ep->msgbuf_addr;
+	buf_size = ep->msg_buf_size/sizeof(u32);
+
+	dma_sync_single_for_cpu(ep->dev, ep->msgbuf_dma_addr,
+				ep->msg_buf_size, DMA_FROM_DEVICE);
+
+	for (i = 0; i < buf_size; i += 2) {
+		if (((buf[i+1] >> 28) & 0xf) != ep->msg_counter) {
+			malformed_message(ep, &buf[i]);
+			dev_warn(ep->dev,
+				 "Sending a NACK on counter %x (instead of %x) on entry %d\n",
+				 ((buf[i+1] >> 28) & 0xf),
+				 ep->msg_counter,
+				 i/2);
+
+			if (++ep->failed_messages > 10) {
+				dev_err(ep->dev,
+					"Lost sync with interrupt messages. Stopping.\n");
+			} else {
+				dma_sync_single_for_device(ep->dev,
+							   ep->msgbuf_dma_addr,
+							   ep->msg_buf_size,
+							   DMA_FROM_DEVICE);
+
+				iowrite32(0x01,  /* Message NACK */
+					  ep->registers + fpga_msg_ctrl_reg);
+			}
+			return IRQ_HANDLED;
+		} else if (buf[i] & (1 << 22)) /* Last message */
+			break;
+	}
+
+	if (i >= buf_size) {
+		dev_err(ep->dev, "Bad interrupt message. Stopping.\n");
+		return IRQ_HANDLED;
+	}
+
+	buf_size = i + 2;
+
+	for (i = 0; i < buf_size; i += 2) { /* Scan through messages */
+		opcode = (buf[i] >> 24) & 0xff;
+
+		msg_dir = buf[i] & 1;
+		msg_channel = (buf[i] >> 1) & 0x7ff;
+		msg_bufno = (buf[i] >> 12) & 0x3ff;
+		msg_data = buf[i+1] & 0xfffffff;
+
+		switch (opcode) {
+		case XILLYMSG_OPCODE_RELEASEBUF:
+			if ((msg_channel > ep->num_channels) ||
+			    (msg_channel == 0)) {
+				malformed_message(ep, &buf[i]);
+				break;
+			}
+
+			channel = ep->channels[msg_channel];
+
+			if (msg_dir) { /* Write channel */
+				if (msg_bufno >= channel->num_wr_buffers) {
+					malformed_message(ep, &buf[i]);
+					break;
+				}
+				spin_lock(&channel->wr_spinlock);
+				channel->wr_buffers[msg_bufno]->end_offset =
+					msg_data;
+				channel->wr_fpga_buf_idx = msg_bufno;
+				channel->wr_empty = 0;
+				channel->wr_sleepy = 0;
+				spin_unlock(&channel->wr_spinlock);
+
+				wake_up_interruptible(&channel->wr_wait);
+
+			} else {
+				/* Read channel */
+
+				if (msg_bufno >= channel->num_rd_buffers) {
+					malformed_message(ep, &buf[i]);
+					break;
+				}
+
+				spin_lock(&channel->rd_spinlock);
+				channel->rd_fpga_buf_idx = msg_bufno;
+				channel->rd_full = 0;
+				spin_unlock(&channel->rd_spinlock);
+
+				wake_up_interruptible(&channel->rd_wait);
+				if (!channel->rd_synchronous)
+					queue_delayed_work(
+						xillybus_wq,
+						&channel->rd_workitem,
+						XILLY_RX_TIMEOUT);
+			}
+
+			break;
+		case XILLYMSG_OPCODE_NONEMPTY:
+			if ((msg_channel > ep->num_channels) ||
+			    (msg_channel == 0) || (!msg_dir) ||
+			    !ep->channels[msg_channel]->wr_supports_nonempty) {
+				malformed_message(ep, &buf[i]);
+				break;
+			}
+
+			channel = ep->channels[msg_channel];
+
+			if (msg_bufno >= channel->num_wr_buffers) {
+				malformed_message(ep, &buf[i]);
+				break;
+			}
+			spin_lock(&channel->wr_spinlock);
+			if (msg_bufno == channel->wr_host_buf_idx)
+				channel->wr_ready = 1;
+			spin_unlock(&channel->wr_spinlock);
+
+			wake_up_interruptible(&channel->wr_ready_wait);
+
+			break;
+		case XILLYMSG_OPCODE_QUIESCEACK:
+			ep->idtlen = msg_data;
+			wake_up_interruptible(&ep->ep_wait);
+
+			break;
+		case XILLYMSG_OPCODE_FIFOEOF:
+			if ((msg_channel > ep->num_channels) ||
+			    (msg_channel == 0) || (!msg_dir) ||
+			    !ep->channels[msg_channel]->num_wr_buffers) {
+				malformed_message(ep, &buf[i]);
+				break;
+			}
+			channel = ep->channels[msg_channel];
+			spin_lock(&channel->wr_spinlock);
+			channel->wr_eof = msg_bufno;
+			channel->wr_sleepy = 0;
+
+			channel->wr_hangup = channel->wr_empty &&
+				(channel->wr_host_buf_idx == msg_bufno);
+
+			spin_unlock(&channel->wr_spinlock);
+
+			wake_up_interruptible(&channel->wr_wait);
+
+			break;
+		case XILLYMSG_OPCODE_FATAL_ERROR:
+			ep->fatal_error = 1;
+			wake_up_interruptible(&ep->ep_wait); /* For select() */
+			dev_err(ep->dev,
+				"FPGA reported a fatal error. This means that the low-level communication with the device has failed. This hardware problem is most likely unrelated to Xillybus (neither kernel module nor FPGA core), but reports are still welcome. All I/O is aborted.\n");
+			break;
+		default:
+			malformed_message(ep, &buf[i]);
+			break;
+		}
+	}
+
+	dma_sync_single_for_device(ep->dev, ep->msgbuf_dma_addr,
+				   ep->msg_buf_size, DMA_FROM_DEVICE);
+
+	ep->msg_counter = (ep->msg_counter + 1) & 0xf;
+	ep->failed_messages = 0;
+	iowrite32(0x03, ep->registers + fpga_msg_ctrl_reg); /* Message ACK */
+
+	return IRQ_HANDLED;
+}
+EXPORT_SYMBOL(xillybus_isr);
+
+/*
+ * A few trivial memory management functions.
+ * NOTE: These functions are used only on probe and remove, and therefore
+ * no locks are applied!
+ */
+
+static void xillybus_autoflush(struct work_struct *work);
+
+struct xilly_alloc_state {
+	void *salami;
+	int left_of_salami;
+	int nbuffer;
+	enum dma_data_direction direction;
+	u32 regdirection;
+};
+
+static void xilly_unmap(void *ptr)
+{
+	struct xilly_mapping *data = ptr;
+
+	dma_unmap_single(data->device, data->dma_addr,
+			 data->size, data->direction);
+
+	kfree(ptr);
+}
+
+static int xilly_map_single(struct xilly_endpoint *ep,
+			    void *ptr,
+			    size_t size,
+			    int direction,
+			    dma_addr_t *ret_dma_handle
+	)
+{
+	dma_addr_t addr;
+	struct xilly_mapping *this;
+
+	this = kzalloc_obj(*this);
+	if (!this)
+		return -ENOMEM;
+
+	addr = dma_map_single(ep->dev, ptr, size, direction);
+
+	if (dma_mapping_error(ep->dev, addr)) {
+		kfree(this);
+		return -ENODEV;
+	}
+
+	this->device = ep->dev;
+	this->dma_addr = addr;
+	this->size = size;
+	this->direction = direction;
+
+	*ret_dma_handle = addr;
+
+	return devm_add_action_or_reset(ep->dev, xilly_unmap, this);
+}
+
+static int xilly_get_dma_buffers(struct xilly_endpoint *ep,
+				 struct xilly_alloc_state *s,
+				 struct xilly_buffer **buffers,
+				 int bufnum, int bytebufsize)
+{
+	int i, rc;
+	dma_addr_t dma_addr;
+	struct device *dev = ep->dev;
+	struct xilly_buffer *this_buffer = NULL; /* Init to silence warning */
+
+	if (buffers) { /* Not the message buffer */
+		this_buffer = devm_kcalloc(dev, bufnum,
+					   sizeof(struct xilly_buffer),
+					   GFP_KERNEL);
+		if (!this_buffer)
+			return -ENOMEM;
+	}
+
+	for (i = 0; i < bufnum; i++) {
+		/*
+		 * Buffers are expected in descending size order, so there
+		 * is either enough space for this buffer or none at all.
+		 */
+
+		if ((s->left_of_salami < bytebufsize) &&
+		    (s->left_of_salami > 0)) {
+			dev_err(ep->dev,
+				"Corrupt buffer allocation in IDT. Aborting.\n");
+			return -ENODEV;
+		}
+
+		if (s->left_of_salami == 0) {
+			int allocorder, allocsize;
+
+			allocsize = PAGE_SIZE;
+			allocorder = 0;
+			while (bytebufsize > allocsize) {
+				allocsize *= 2;
+				allocorder++;
+			}
+
+			s->salami = (void *) devm_get_free_pages(
+				dev,
+				GFP_KERNEL | __GFP_DMA32 | __GFP_ZERO,
+				allocorder);
+			if (!s->salami)
+				return -ENOMEM;
+
+			s->left_of_salami = allocsize;
+		}
+
+		rc = xilly_map_single(ep, s->salami,
+				      bytebufsize, s->direction,
+				      &dma_addr);
+		if (rc)
+			return rc;
+
+		iowrite32((u32) (dma_addr & 0xffffffff),
+			  ep->registers + fpga_dma_bufaddr_lowaddr_reg);
+		iowrite32(((u32) ((((u64) dma_addr) >> 32) & 0xffffffff)),
+			  ep->registers + fpga_dma_bufaddr_highaddr_reg);
+
+		if (buffers) { /* Not the message buffer */
+			this_buffer->addr = s->salami;
+			this_buffer->dma_addr = dma_addr;
+			buffers[i] = this_buffer++;
+
+			iowrite32(s->regdirection | s->nbuffer++,
+				  ep->registers + fpga_dma_bufno_reg);
+		} else {
+			ep->msgbuf_addr = s->salami;
+			ep->msgbuf_dma_addr = dma_addr;
+			ep->msg_buf_size = bytebufsize;
+
+			iowrite32(s->regdirection,
+				  ep->registers + fpga_dma_bufno_reg);
+		}
+
+		s->left_of_salami -= bytebufsize;
+		s->salami += bytebufsize;
+	}
+	return 0;
+}
+
+static int xilly_setupchannels(struct xilly_endpoint *ep,
+			       unsigned char *chandesc,
+			       int entries)
+{
+	struct device *dev = ep->dev;
+	int i, entry, rc;
+	struct xilly_channel *channel;
+	int channelnum, bufnum, bufsize, format, is_writebuf;
+	int bytebufsize;
+	int synchronous, allowpartial, exclusive_open, seekable;
+	int supports_nonempty;
+	int msg_buf_done = 0;
+
+	struct xilly_alloc_state rd_alloc = {
+		.salami = NULL,
+		.left_of_salami = 0,
+		.nbuffer = 1,
+		.direction = DMA_TO_DEVICE,
+		.regdirection = 0,
+	};
+
+	struct xilly_alloc_state wr_alloc = {
+		.salami = NULL,
+		.left_of_salami = 0,
+		.nbuffer = 1,
+		.direction = DMA_FROM_DEVICE,
+		.regdirection = 0x80000000,
+	};
+
+	channel = devm_kcalloc(dev, ep->num_channels,
+			       sizeof(struct xilly_channel), GFP_KERNEL);
+	if (!channel)
+		return -ENOMEM;
+
+	ep->channels = devm_kcalloc(dev, ep->num_channels + 1,
+				    sizeof(struct xilly_channel *),
+				    GFP_KERNEL);
+	if (!ep->channels)
+		return -ENOMEM;
+
+	ep->channels[0] = NULL; /* Channel 0 is message buf. */
+
+	/* Initialize all channels with defaults */
+
+	for (i = 1; i <= ep->num_channels; i++) {
+		channel->wr_buffers = NULL;
+		channel->rd_buffers = NULL;
+		channel->num_wr_buffers = 0;
+		channel->num_rd_buffers = 0;
+		channel->wr_fpga_buf_idx = -1;
+		channel->wr_host_buf_idx = 0;
+		channel->wr_host_buf_pos = 0;
+		channel->wr_empty = 1;
+		channel->wr_ready = 0;
+		channel->wr_sleepy = 1;
+		channel->rd_fpga_buf_idx = 0;
+		channel->rd_host_buf_idx = 0;
+		channel->rd_host_buf_pos = 0;
+		channel->rd_full = 0;
+		channel->wr_ref_count = 0;
+		channel->rd_ref_count = 0;
+
+		spin_lock_init(&channel->wr_spinlock);
+		spin_lock_init(&channel->rd_spinlock);
+		mutex_init(&channel->wr_mutex);
+		mutex_init(&channel->rd_mutex);
+		init_waitqueue_head(&channel->rd_wait);
+		init_waitqueue_head(&channel->wr_wait);
+		init_waitqueue_head(&channel->wr_ready_wait);
+
+		INIT_DELAYED_WORK(&channel->rd_workitem, xillybus_autoflush);
+
+		channel->endpoint = ep;
+		channel->chan_num = i;
+
+		channel->log2_element_size = 0;
+
+		ep->channels[i] = channel++;
+	}
+
+	for (entry = 0; entry < entries; entry++, chandesc += 4) {
+		struct xilly_buffer **buffers = NULL;
+
+		is_writebuf = chandesc[0] & 0x01;
+		channelnum = (chandesc[0] >> 1) | ((chandesc[1] & 0x0f) << 7);
+		format = (chandesc[1] >> 4) & 0x03;
+		allowpartial = (chandesc[1] >> 6) & 0x01;
+		synchronous = (chandesc[1] >> 7) & 0x01;
+		bufsize = 1 << (chandesc[2] & 0x1f);
+		bufnum = 1 << (chandesc[3] & 0x0f);
+		exclusive_open = (chandesc[2] >> 7) & 0x01;
+		seekable = (chandesc[2] >> 6) & 0x01;
+		supports_nonempty = (chandesc[2] >> 5) & 0x01;
+
+		if ((channelnum > ep->num_channels) ||
+		    ((channelnum == 0) && !is_writebuf)) {
+			dev_err(ep->dev,
+				"IDT requests channel out of range. Aborting.\n");
+			return -ENODEV;
+		}
+
+		channel = ep->channels[channelnum]; /* NULL for msg channel */
+
+		if (!is_writebuf || channelnum > 0) {
+			channel->log2_element_size = ((format > 2) ?
+						      2 : format);
+
+			bytebufsize = bufsize *
+				(1 << channel->log2_element_size);
+
+			buffers = devm_kcalloc(dev, bufnum,
+					       sizeof(struct xilly_buffer *),
+					       GFP_KERNEL);
+			if (!buffers)
+				return -ENOMEM;
+		} else {
+			bytebufsize = bufsize << 2;
+		}
+
+		if (!is_writebuf) {
+			channel->num_rd_buffers = bufnum;
+			channel->rd_buf_size = bytebufsize;
+			channel->rd_allow_partial = allowpartial;
+			channel->rd_synchronous = synchronous;
+			channel->rd_exclusive_open = exclusive_open;
+			channel->seekable = seekable;
+
+			channel->rd_buffers = buffers;
+			rc = xilly_get_dma_buffers(ep, &rd_alloc, buffers,
+						   bufnum, bytebufsize);
+		} else if (channelnum > 0) {
+			channel->num_wr_buffers = bufnum;
+			channel->wr_buf_size = bytebufsize;
+
+			channel->seekable = seekable;
+			channel->wr_supports_nonempty = supports_nonempty;
+
+			channel->wr_allow_partial = allowpartial;
+			channel->wr_synchronous = synchronous;
+			channel->wr_exclusive_open = exclusive_open;
+
+			channel->wr_buffers = buffers;
+			rc = xilly_get_dma_buffers(ep, &wr_alloc, buffers,
+						   bufnum, bytebufsize);
+		} else {
+			rc = xilly_get_dma_buffers(ep, &wr_alloc, NULL,
+						   bufnum, bytebufsize);
+			msg_buf_done++;
+		}
+
+		if (rc)
+			return -ENOMEM;
+	}
+
+	if (!msg_buf_done) {
+		dev_err(ep->dev,
+			"Corrupt IDT: No message buffer. Aborting.\n");
+		return -ENODEV;
+	}
+	return 0;
+}
+
+static int xilly_scan_idt(struct xilly_endpoint *endpoint,
+			  struct xilly_idt_handle *idt_handle)
+{
+	int count = 0;
+	unsigned char *idt = endpoint->channels[1]->wr_buffers[0]->addr;
+	unsigned char *end_of_idt = idt + endpoint->idtlen - 4;
+	unsigned char *scan;
+	int len;
+
+	scan = idt + 1;
+	idt_handle->names = scan;
+
+	while ((scan <= end_of_idt) && *scan) {
+		while ((scan <= end_of_idt) && *scan++)
+			/* Do nothing, just scan thru string */;
+		count++;
+	}
+
+	idt_handle->names_len = scan - idt_handle->names;
+
+	scan++;
+
+	if (scan > end_of_idt) {
+		dev_err(endpoint->dev,
+			"IDT device name list overflow. Aborting.\n");
+		return -ENODEV;
+	}
+	idt_handle->chandesc = scan;
+
+	len = endpoint->idtlen - (3 + ((int) (scan - idt)));
+
+	if (len & 0x03) {
+		dev_err(endpoint->dev,
+			"Corrupt IDT device name list. Aborting.\n");
+		return -ENODEV;
+	}
+
+	idt_handle->entries = len >> 2;
+	endpoint->num_channels = count;
+
+	return 0;
+}
+
+static int xilly_obtain_idt(struct xilly_endpoint *endpoint)
+{
+	struct xilly_channel *channel;
+	unsigned char *version;
+	long t;
+
+	channel = endpoint->channels[1]; /* This should be generated ad-hoc */
+
+	channel->wr_sleepy = 1;
+
+	iowrite32(1 |
+		  (3 << 24), /* Opcode 3 for channel 0 = Send IDT */
+		  endpoint->registers + fpga_buf_ctrl_reg);
+
+	t = wait_event_interruptible_timeout(channel->wr_wait,
+					     (!channel->wr_sleepy),
+					     XILLY_TIMEOUT);
+
+	if (t <= 0) {
+		dev_err(endpoint->dev, "Failed to obtain IDT. Aborting.\n");
+
+		if (endpoint->fatal_error)
+			return -EIO;
+
+		return -ENODEV;
+	}
+
+	dma_sync_single_for_cpu(channel->endpoint->dev,
+				channel->wr_buffers[0]->dma_addr,
+				channel->wr_buf_size,
+				DMA_FROM_DEVICE);
+
+	if (channel->wr_buffers[0]->end_offset != endpoint->idtlen) {
+		dev_err(endpoint->dev,
+			"IDT length mismatch (%d != %d). Aborting.\n",
+			channel->wr_buffers[0]->end_offset, endpoint->idtlen);
+		return -ENODEV;
+	}
+
+	if (crc32_le(~0, channel->wr_buffers[0]->addr,
+		     endpoint->idtlen+1) != 0) {
+		dev_err(endpoint->dev, "IDT failed CRC check. Aborting.\n");
+		return -ENODEV;
+	}
+
+	version = channel->wr_buffers[0]->addr;
+
+	/* Check version number. Reject anything above 0x82. */
+	if (*version > 0x82) {
+		dev_err(endpoint->dev,
+			"No support for IDT version 0x%02x. Maybe the xillybus driver needs an upgrade. Aborting.\n",
+			*version);
+		return -ENODEV;
+	}
+
+	return 0;
+}
+
+static ssize_t xillybus_read(struct file *filp, char __user *userbuf,
+			     size_t count, loff_t *f_pos)
+{
+	ssize_t rc;
+	unsigned long flags;
+	int bytes_done = 0;
+	int no_time_left = 0;
+	long deadline, left_to_sleep;
+	struct xilly_channel *channel = filp->private_data;
+
+	int empty, reached_eof, exhausted, ready;
+	/* Initializations are there only to silence warnings */
+
+	int howmany = 0, bufpos = 0, bufidx = 0, bufferdone = 0;
+	int waiting_bufidx;
+
+	if (channel->endpoint->fatal_error)
+		return -EIO;
+
+	deadline = jiffies + 1 + XILLY_RX_TIMEOUT;
+
+	rc = mutex_lock_interruptible(&channel->wr_mutex);
+	if (rc)
+		return rc;
+
+	while (1) { /* Note that we may drop mutex within this loop */
+		int bytes_to_do = count - bytes_done;
+
+		spin_lock_irqsave(&channel->wr_spinlock, flags);
+
+		empty = channel->wr_empty;
+		ready = !empty || channel->wr_ready;
+
+		if (!empty) {
+			bufidx = channel->wr_host_buf_idx;
+			bufpos = channel->wr_host_buf_pos;
+			howmany = ((channel->wr_buffers[bufidx]->end_offset
+				    + 1) << channel->log2_element_size)
+				- bufpos;
+
+			/* Update wr_host_* to its post-operation state */
+			if (howmany > bytes_to_do) {
+				bufferdone = 0;
+
+				howmany = bytes_to_do;
+				channel->wr_host_buf_pos += howmany;
+			} else {
+				bufferdone = 1;
+
+				channel->wr_host_buf_pos = 0;
+
+				if (bufidx == channel->wr_fpga_buf_idx) {
+					channel->wr_empty = 1;
+					channel->wr_sleepy = 1;
+					channel->wr_ready = 0;
+				}
+
+				if (bufidx >= (channel->num_wr_buffers - 1))
+					channel->wr_host_buf_idx = 0;
+				else
+					channel->wr_host_buf_idx++;
+			}
+		}
+
+		/*
+		 * Marking our situation after the possible changes above,
+		 * for use after releasing the spinlock.
+		 *
+		 * empty = empty before change
+		 * exhasted = empty after possible change
+		 */
+
+		reached_eof = channel->wr_empty &&
+			(channel->wr_host_buf_idx == channel->wr_eof);
+		channel->wr_hangup = reached_eof;
+		exhausted = channel->wr_empty;
+		waiting_bufidx = channel->wr_host_buf_idx;
+
+		spin_unlock_irqrestore(&channel->wr_spinlock, flags);
+
+		if (!empty) { /* Go on, now without the spinlock */
+
+			if (bufpos == 0) /* Position zero means it's virgin */
+				dma_sync_single_for_cpu(channel->endpoint->dev,
+							channel->wr_buffers[bufidx]->dma_addr,
+							channel->wr_buf_size,
+							DMA_FROM_DEVICE);
+
+			if (copy_to_user(
+				    userbuf,
+				    channel->wr_buffers[bufidx]->addr
+				    + bufpos, howmany))
+				rc = -EFAULT;
+
+			userbuf += howmany;
+			bytes_done += howmany;
+
+			if (bufferdone) {
+				dma_sync_single_for_device(channel->endpoint->dev,
+							   channel->wr_buffers[bufidx]->dma_addr,
+							   channel->wr_buf_size,
+							   DMA_FROM_DEVICE);
+
+				/*
+				 * Tell FPGA the buffer is done with. It's an
+				 * atomic operation to the FPGA, so what
+				 * happens with other channels doesn't matter,
+				 * and the certain channel is protected with
+				 * the channel-specific mutex.
+				 */
+
+				iowrite32(1 | (channel->chan_num << 1) |
+					  (bufidx << 12),
+					  channel->endpoint->registers +
+					  fpga_buf_ctrl_reg);
+			}
+
+			if (rc) {
+				mutex_unlock(&channel->wr_mutex);
+				return rc;
+			}
+		}
+
+		/* This includes a zero-count return = EOF */
+		if ((bytes_done >= count) || reached_eof)
+			break;
+
+		if (!exhausted)
+			continue; /* More in RAM buffer(s)? Just go on. */
+
+		if ((bytes_done > 0) &&
+		    (no_time_left ||
+		     (channel->wr_synchronous && channel->wr_allow_partial)))
+			break;
+
+		/*
+		 * Nonblocking read: The "ready" flag tells us that the FPGA
+		 * has data to send. In non-blocking mode, if it isn't on,
+		 * just return. But if there is, we jump directly to the point
+		 * where we ask for the FPGA to send all it has, and wait
+		 * until that data arrives. So in a sense, we *do* block in
+		 * nonblocking mode, but only for a very short time.
+		 */
+
+		if (!no_time_left && (filp->f_flags & O_NONBLOCK)) {
+			if (bytes_done > 0)
+				break;
+
+			if (ready)
+				goto desperate;
+
+			rc = -EAGAIN;
+			break;
+		}
+
+		if (!no_time_left || (bytes_done > 0)) {
+			/*
+			 * Note that in case of an element-misaligned read
+			 * request, offsetlimit will include the last element,
+			 * which will be partially read from.
+			 */
+			int offsetlimit = ((count - bytes_done) - 1) >>
+				channel->log2_element_size;
+			int buf_elements = channel->wr_buf_size >>
+				channel->log2_element_size;
+
+			/*
+			 * In synchronous mode, always send an offset limit.
+			 * Just don't send a value too big.
+			 */
+
+			if (channel->wr_synchronous) {
+				/* Don't request more than one buffer */
+				if (channel->wr_allow_partial &&
+				    (offsetlimit >= buf_elements))
+					offsetlimit = buf_elements - 1;
+
+				/* Don't request more than all buffers */
+				if (!channel->wr_allow_partial &&
+				    (offsetlimit >=
+				     (buf_elements * channel->num_wr_buffers)))
+					offsetlimit = buf_elements *
+						channel->num_wr_buffers - 1;
+			}
+
+			/*
+			 * In asynchronous mode, force early flush of a buffer
+			 * only if that will allow returning a full count. The
+			 * "offsetlimit < ( ... )" rather than "<=" excludes
+			 * requesting a full buffer, which would obviously
+			 * cause a buffer transmission anyhow
+			 */
+
+			if (channel->wr_synchronous ||
+			    (offsetlimit < (buf_elements - 1))) {
+				mutex_lock(&channel->endpoint->register_mutex);
+
+				iowrite32(offsetlimit,
+					  channel->endpoint->registers +
+					  fpga_buf_offset_reg);
+
+				iowrite32(1 | (channel->chan_num << 1) |
+					  (2 << 24) |  /* 2 = offset limit */
+					  (waiting_bufidx << 12),
+					  channel->endpoint->registers +
+					  fpga_buf_ctrl_reg);
+
+				mutex_unlock(&channel->endpoint->
+					     register_mutex);
+			}
+		}
+
+		/*
+		 * If partial completion is disallowed, there is no point in
+		 * timeout sleeping. Neither if no_time_left is set and
+		 * there's no data.
+		 */
+
+		if (!channel->wr_allow_partial ||
+		    (no_time_left && (bytes_done == 0))) {
+			/*
+			 * This do-loop will run more than once if another
+			 * thread reasserted wr_sleepy before we got the mutex
+			 * back, so we try again.
+			 */
+
+			do {
+				mutex_unlock(&channel->wr_mutex);
+
+				if (wait_event_interruptible(
+					    channel->wr_wait,
+					    (!channel->wr_sleepy)))
+					goto interrupted;
+
+				if (mutex_lock_interruptible(
+					    &channel->wr_mutex))
+					goto interrupted;
+			} while (channel->wr_sleepy);
+
+			continue;
+
+interrupted: /* Mutex is not held if got here */
+			if (channel->endpoint->fatal_error)
+				return -EIO;
+			if (bytes_done)
+				return bytes_done;
+			if (filp->f_flags & O_NONBLOCK)
+				return -EAGAIN; /* Don't admit snoozing */
+			return -EINTR;
+		}
+
+		left_to_sleep = deadline - ((long) jiffies);
+
+		/*
+		 * If our time is out, skip the waiting. We may miss wr_sleepy
+		 * being deasserted but hey, almost missing the train is like
+		 * missing it.
+		 */
+
+		if (left_to_sleep > 0) {
+			left_to_sleep =
+				wait_event_interruptible_timeout(
+					channel->wr_wait,
+					(!channel->wr_sleepy),
+					left_to_sleep);
+
+			if (left_to_sleep > 0) /* wr_sleepy deasserted */
+				continue;
+
+			if (left_to_sleep < 0) { /* Interrupt */
+				mutex_unlock(&channel->wr_mutex);
+				if (channel->endpoint->fatal_error)
+					return -EIO;
+				if (bytes_done)
+					return bytes_done;
+				return -EINTR;
+			}
+		}
+
+desperate:
+		no_time_left = 1; /* We're out of sleeping time. Desperate! */
+
+		if (bytes_done == 0) {
+			/*
+			 * Reaching here means that we allow partial return,
+			 * that we've run out of time, and that we have
+			 * nothing to return.
+			 * So tell the FPGA to send anything it has or gets.
+			 */
+
+			iowrite32(1 | (channel->chan_num << 1) |
+				  (3 << 24) |  /* Opcode 3, flush it all! */
+				  (waiting_bufidx << 12),
+				  channel->endpoint->registers +
+				  fpga_buf_ctrl_reg);
+		}
+
+		/*
+		 * Reaching here means that we *do* have data in the buffer,
+		 * but the "partial" flag disallows returning less than
+		 * required. And we don't have as much. So loop again,
+		 * which is likely to end up blocking indefinitely until
+		 * enough data has arrived.
+		 */
+	}
+
+	mutex_unlock(&channel->wr_mutex);
+
+	if (channel->endpoint->fatal_error)
+		return -EIO;
+
+	if (rc)
+		return rc;
+
+	return bytes_done;
+}
+
+/*
+ * The timeout argument takes values as follows:
+ *  >0 : Flush with timeout
+ * ==0 : Flush, and wait idefinitely for the flush to complete
+ *  <0 : Autoflush: Flush only if there's a single buffer occupied
+ */
+
+static int xillybus_myflush(struct xilly_channel *channel, long timeout)
+{
+	int rc;
+	unsigned long flags;
+
+	int end_offset_plus1;
+	int bufidx, bufidx_minus1;
+	int i;
+	int empty;
+	int new_rd_host_buf_pos;
+
+	if (channel->endpoint->fatal_error)
+		return -EIO;
+	rc = mutex_lock_interruptible(&channel->rd_mutex);
+	if (rc)
+		return rc;
+
+	/*
+	 * Don't flush a closed channel. This can happen when the work queued
+	 * autoflush thread fires off after the file has closed. This is not
+	 * an error, just something to dismiss.
+	 */
+
+	if (!channel->rd_ref_count)
+		goto done;
+
+	bufidx = channel->rd_host_buf_idx;
+
+	bufidx_minus1 = (bufidx == 0) ?
+		channel->num_rd_buffers - 1 :
+		bufidx - 1;
+
+	end_offset_plus1 = channel->rd_host_buf_pos >>
+		channel->log2_element_size;
+
+	new_rd_host_buf_pos = channel->rd_host_buf_pos -
+		(end_offset_plus1 << channel->log2_element_size);
+
+	/* Submit the current buffer if it's nonempty */
+	if (end_offset_plus1) {
+		unsigned char *tail = channel->rd_buffers[bufidx]->addr +
+			(end_offset_plus1 << channel->log2_element_size);
+
+		/* Copy  unflushed data, so we can put it in next buffer */
+		for (i = 0; i < new_rd_host_buf_pos; i++)
+			channel->rd_leftovers[i] = *tail++;
+
+		spin_lock_irqsave(&channel->rd_spinlock, flags);
+
+		/* Autoflush only if a single buffer is occupied */
+
+		if ((timeout < 0) &&
+		    (channel->rd_full ||
+		     (bufidx_minus1 != channel->rd_fpga_buf_idx))) {
+			spin_unlock_irqrestore(&channel->rd_spinlock, flags);
+			/*
+			 * A new work item may be queued by the ISR exactly
+			 * now, since the execution of a work item allows the
+			 * queuing of a new one while it's running.
+			 */
+			goto done;
+		}
+
+		/* The 4th element is never needed for data, so it's a flag */
+		channel->rd_leftovers[3] = (new_rd_host_buf_pos != 0);
+
+		/* Set up rd_full to reflect a certain moment's state */
+
+		if (bufidx == channel->rd_fpga_buf_idx)
+			channel->rd_full = 1;
+		spin_unlock_irqrestore(&channel->rd_spinlock, flags);
+
+		if (bufidx >= (channel->num_rd_buffers - 1))
+			channel->rd_host_buf_idx = 0;
+		else
+			channel->rd_host_buf_idx++;
+
+		dma_sync_single_for_device(channel->endpoint->dev,
+					   channel->rd_buffers[bufidx]->dma_addr,
+					   channel->rd_buf_size,
+					   DMA_TO_DEVICE);
+
+		mutex_lock(&channel->endpoint->register_mutex);
+
+		iowrite32(end_offset_plus1 - 1,
+			  channel->endpoint->registers + fpga_buf_offset_reg);
+
+		iowrite32((channel->chan_num << 1) | /* Channel ID */
+			  (2 << 24) |  /* Opcode 2, submit buffer */
+			  (bufidx << 12),
+			  channel->endpoint->registers + fpga_buf_ctrl_reg);
+
+		mutex_unlock(&channel->endpoint->register_mutex);
+	} else if (bufidx == 0) {
+		bufidx = channel->num_rd_buffers - 1;
+	} else {
+		bufidx--;
+	}
+
+	channel->rd_host_buf_pos = new_rd_host_buf_pos;
+
+	if (timeout < 0)
+		goto done; /* Autoflush */
+
+	/*
+	 * bufidx is now the last buffer written to (or equal to
+	 * rd_fpga_buf_idx if buffer was never written to), and
+	 * channel->rd_host_buf_idx the one after it.
+	 *
+	 * If bufidx == channel->rd_fpga_buf_idx we're either empty or full.
+	 */
+
+	while (1) { /* Loop waiting for draining of buffers */
+		spin_lock_irqsave(&channel->rd_spinlock, flags);
+
+		if (bufidx != channel->rd_fpga_buf_idx)
+			channel->rd_full = 1; /*
+					       * Not really full,
+					       * but needs waiting.
+					       */
+
+		empty = !channel->rd_full;
+
+		spin_unlock_irqrestore(&channel->rd_spinlock, flags);
+
+		if (empty)
+			break;
+
+		/*
+		 * Indefinite sleep with mutex taken. With data waiting for
+		 * flushing user should not be surprised if open() for write
+		 * sleeps.
+		 */
+		if (timeout == 0)
+			wait_event_interruptible(channel->rd_wait,
+						 (!channel->rd_full));
+
+		else if (wait_event_interruptible_timeout(
+				 channel->rd_wait,
+				 (!channel->rd_full),
+				 timeout) == 0) {
+			dev_warn(channel->endpoint->dev,
+				 "Timed out while flushing. Output data may be lost.\n");
+
+			rc = -ETIMEDOUT;
+			break;
+		}
+
+		if (channel->rd_full) {
+			rc = -EINTR;
+			break;
+		}
+	}
+
+done:
+	mutex_unlock(&channel->rd_mutex);
+
+	if (channel->endpoint->fatal_error)
+		return -EIO;
+
+	return rc;
+}
+
+static int xillybus_flush(struct file *filp, fl_owner_t id)
+{
+	if (!(filp->f_mode & FMODE_WRITE))
+		return 0;
+
+	return xillybus_myflush(filp->private_data, HZ); /* 1 second timeout */
+}
+
+static void xillybus_autoflush(struct work_struct *work)
+{
+	struct delayed_work *workitem = to_delayed_work(work);
+	struct xilly_channel *channel = container_of(
+		workitem, struct xilly_channel, rd_workitem);
+	int rc;
+
+	rc = xillybus_myflush(channel, -1);
+	if (rc == -EINTR)
+		dev_warn(channel->endpoint->dev,
+			 "Autoflush failed because work queue thread got a signal.\n");
+	else if (rc)
+		dev_err(channel->endpoint->dev,
+			"Autoflush failed under weird circumstances.\n");
+}
+
+static ssize_t xillybus_write(struct file *filp, const char __user *userbuf,
+			      size_t count, loff_t *f_pos)
+{
+	ssize_t rc;
+	unsigned long flags;
+	int bytes_done = 0;
+	struct xilly_channel *channel = filp->private_data;
+
+	int full, exhausted;
+	/* Initializations are there only to silence warnings */
+
+	int howmany = 0, bufpos = 0, bufidx = 0, bufferdone = 0;
+	int end_offset_plus1 = 0;
+
+	if (channel->endpoint->fatal_error)
+		return -EIO;
+
+	rc = mutex_lock_interruptible(&channel->rd_mutex);
+	if (rc)
+		return rc;
+
+	while (1) {
+		int bytes_to_do = count - bytes_done;
+
+		spin_lock_irqsave(&channel->rd_spinlock, flags);
+
+		full = channel->rd_full;
+
+		if (!full) {
+			bufidx = channel->rd_host_buf_idx;
+			bufpos = channel->rd_host_buf_pos;
+			howmany = channel->rd_buf_size - bufpos;
+
+			/*
+			 * Update rd_host_* to its state after this operation.
+			 * count=0 means committing the buffer immediately,
+			 * which is like flushing, but not necessarily block.
+			 */
+
+			if ((howmany > bytes_to_do) &&
+			    (count ||
+			     ((bufpos >> channel->log2_element_size) == 0))) {
+				bufferdone = 0;
+
+				howmany = bytes_to_do;
+				channel->rd_host_buf_pos += howmany;
+			} else {
+				bufferdone = 1;
+
+				if (count) {
+					end_offset_plus1 =
+						channel->rd_buf_size >>
+						channel->log2_element_size;
+					channel->rd_host_buf_pos = 0;
+				} else {
+					unsigned char *tail;
+					int i;
+
+					howmany = 0;
+
+					end_offset_plus1 = bufpos >>
+						channel->log2_element_size;
+
+					channel->rd_host_buf_pos -=
+						end_offset_plus1 <<
+						channel->log2_element_size;
+
+					tail = channel->
+						rd_buffers[bufidx]->addr +
+						(end_offset_plus1 <<
+						 channel->log2_element_size);
+
+					for (i = 0;
+					     i < channel->rd_host_buf_pos;
+					     i++)
+						channel->rd_leftovers[i] =
+							*tail++;
+				}
+
+				if (bufidx == channel->rd_fpga_buf_idx)
+					channel->rd_full = 1;
+
+				if (bufidx >= (channel->num_rd_buffers - 1))
+					channel->rd_host_buf_idx = 0;
+				else
+					channel->rd_host_buf_idx++;
+			}
+		}
+
+		/*
+		 * Marking our situation after the possible changes above,
+		 * for use  after releasing the spinlock.
+		 *
+		 * full = full before change
+		 * exhasted = full after possible change
+		 */
+
+		exhausted = channel->rd_full;
+
+		spin_unlock_irqrestore(&channel->rd_spinlock, flags);
+
+		if (!full) { /* Go on, now without the spinlock */
+			unsigned char *head =
+				channel->rd_buffers[bufidx]->addr;
+			int i;
+
+			if ((bufpos == 0) || /* Zero means it's virgin */
+			    (channel->rd_leftovers[3] != 0)) {
+				dma_sync_single_for_cpu(channel->endpoint->dev,
+							channel->rd_buffers[bufidx]->dma_addr,
+							channel->rd_buf_size,
+							DMA_TO_DEVICE);
+
+				/* Virgin, but leftovers are due */
+				for (i = 0; i < bufpos; i++)
+					*head++ = channel->rd_leftovers[i];
+
+				channel->rd_leftovers[3] = 0; /* Clear flag */
+			}
+
+			if (copy_from_user(
+				    channel->rd_buffers[bufidx]->addr + bufpos,
+				    userbuf, howmany))
+				rc = -EFAULT;
+
+			userbuf += howmany;
+			bytes_done += howmany;
+
+			if (bufferdone) {
+				dma_sync_single_for_device(channel->endpoint->dev,
+							   channel->rd_buffers[bufidx]->dma_addr,
+							   channel->rd_buf_size,
+							   DMA_TO_DEVICE);
+
+				mutex_lock(&channel->endpoint->register_mutex);
+
+				iowrite32(end_offset_plus1 - 1,
+					  channel->endpoint->registers +
+					  fpga_buf_offset_reg);
+
+				iowrite32((channel->chan_num << 1) |
+					  (2 << 24) |  /* 2 = submit buffer */
+					  (bufidx << 12),
+					  channel->endpoint->registers +
+					  fpga_buf_ctrl_reg);
+
+				mutex_unlock(&channel->endpoint->
+					     register_mutex);
+
+				channel->rd_leftovers[3] =
+					(channel->rd_host_buf_pos != 0);
+			}
+
+			if (rc) {
+				mutex_unlock(&channel->rd_mutex);
+
+				if (channel->endpoint->fatal_error)
+					return -EIO;
+
+				if (!channel->rd_synchronous)
+					queue_delayed_work(
+						xillybus_wq,
+						&channel->rd_workitem,
+						XILLY_RX_TIMEOUT);
+
+				return rc;
+			}
+		}
+
+		if (bytes_done >= count)
+			break;
+
+		if (!exhausted)
+			continue; /* If there's more space, just go on */
+
+		if ((bytes_done > 0) && channel->rd_allow_partial)
+			break;
+
+		/*
+		 * Indefinite sleep with mutex taken. With data waiting for
+		 * flushing, user should not be surprised if open() for write
+		 * sleeps.
+		 */
+
+		if (filp->f_flags & O_NONBLOCK) {
+			rc = -EAGAIN;
+			break;
+		}
+
+		if (wait_event_interruptible(channel->rd_wait,
+					     (!channel->rd_full))) {
+			mutex_unlock(&channel->rd_mutex);
+
+			if (channel->endpoint->fatal_error)
+				return -EIO;
+
+			if (bytes_done)
+				return bytes_done;
+			return -EINTR;
+		}
+	}
+
+	mutex_unlock(&channel->rd_mutex);
+
+	if (!channel->rd_synchronous)
+		queue_delayed_work(xillybus_wq,
+				   &channel->rd_workitem,
+				   XILLY_RX_TIMEOUT);
+
+	if (channel->endpoint->fatal_error)
+		return -EIO;
+
+	if (rc)
+		return rc;
+
+	if ((channel->rd_synchronous) && (bytes_done > 0)) {
+		rc = xillybus_myflush(filp->private_data, 0); /* No timeout */
+
+		if (rc && (rc != -EINTR))
+			return rc;
+	}
+
+	return bytes_done;
+}
+
+static int xillybus_open(struct inode *inode, struct file *filp)
+{
+	int rc;
+	unsigned long flags;
+	struct xilly_endpoint *endpoint;
+	struct xilly_channel *channel;
+	int index;
+
+	rc = xillybus_find_inode(inode, (void **)&endpoint, &index);
+	if (rc)
+		return rc;
+
+	if (endpoint->fatal_error)
+		return -EIO;
+
+	channel = endpoint->channels[1 + index];
+	filp->private_data = channel;
+
+	/*
+	 * It gets complicated because:
+	 * 1. We don't want to take a mutex we don't have to
+	 * 2. We don't want to open one direction if the other will fail.
+	 */
+
+	if ((filp->f_mode & FMODE_READ) && (!channel->num_wr_buffers))
+		return -ENODEV;
+
+	if ((filp->f_mode & FMODE_WRITE) && (!channel->num_rd_buffers))
+		return -ENODEV;
+
+	if ((filp->f_mode & FMODE_READ) && (filp->f_flags & O_NONBLOCK) &&
+	    (channel->wr_synchronous || !channel->wr_allow_partial ||
+	     !channel->wr_supports_nonempty)) {
+		dev_err(endpoint->dev,
+			"open() failed: O_NONBLOCK not allowed for read on this device\n");
+		return -ENODEV;
+	}
+
+	if ((filp->f_mode & FMODE_WRITE) && (filp->f_flags & O_NONBLOCK) &&
+	    (channel->rd_synchronous || !channel->rd_allow_partial)) {
+		dev_err(endpoint->dev,
+			"open() failed: O_NONBLOCK not allowed for write on this device\n");
+		return -ENODEV;
+	}
+
+	/*
+	 * Note: open() may block on getting mutexes despite O_NONBLOCK.
+	 * This shouldn't occur normally, since multiple open of the same
+	 * file descriptor is almost always prohibited anyhow
+	 * (*_exclusive_open is normally set in real-life systems).
+	 */
+
+	if (filp->f_mode & FMODE_READ) {
+		rc = mutex_lock_interruptible(&channel->wr_mutex);
+		if (rc)
+			return rc;
+	}
+
+	if (filp->f_mode & FMODE_WRITE) {
+		rc = mutex_lock_interruptible(&channel->rd_mutex);
+		if (rc)
+			goto unlock_wr;
+	}
+
+	if ((filp->f_mode & FMODE_READ) &&
+	    (channel->wr_ref_count != 0) &&
+	    (channel->wr_exclusive_open)) {
+		rc = -EBUSY;
+		goto unlock;
+	}
+
+	if ((filp->f_mode & FMODE_WRITE) &&
+	    (channel->rd_ref_count != 0) &&
+	    (channel->rd_exclusive_open)) {
+		rc = -EBUSY;
+		goto unlock;
+	}
+
+	if (filp->f_mode & FMODE_READ) {
+		if (channel->wr_ref_count == 0) { /* First open of file */
+			/* Move the host to first buffer */
+			spin_lock_irqsave(&channel->wr_spinlock, flags);
+			channel->wr_host_buf_idx = 0;
+			channel->wr_host_buf_pos = 0;
+			channel->wr_fpga_buf_idx = -1;
+			channel->wr_empty = 1;
+			channel->wr_ready = 0;
+			channel->wr_sleepy = 1;
+			channel->wr_eof = -1;
+			channel->wr_hangup = 0;
+
+			spin_unlock_irqrestore(&channel->wr_spinlock, flags);
+
+			iowrite32(1 | (channel->chan_num << 1) |
+				  (4 << 24) |  /* Opcode 4, open channel */
+				  ((channel->wr_synchronous & 1) << 23),
+				  channel->endpoint->registers +
+				  fpga_buf_ctrl_reg);
+		}
+
+		channel->wr_ref_count++;
+	}
+
+	if (filp->f_mode & FMODE_WRITE) {
+		if (channel->rd_ref_count == 0) { /* First open of file */
+			/* Move the host to first buffer */
+			spin_lock_irqsave(&channel->rd_spinlock, flags);
+			channel->rd_host_buf_idx = 0;
+			channel->rd_host_buf_pos = 0;
+			channel->rd_leftovers[3] = 0; /* No leftovers. */
+			channel->rd_fpga_buf_idx = channel->num_rd_buffers - 1;
+			channel->rd_full = 0;
+
+			spin_unlock_irqrestore(&channel->rd_spinlock, flags);
+
+			iowrite32((channel->chan_num << 1) |
+				  (4 << 24),   /* Opcode 4, open channel */
+				  channel->endpoint->registers +
+				  fpga_buf_ctrl_reg);
+		}
+
+		channel->rd_ref_count++;
+	}
+
+unlock:
+	if (filp->f_mode & FMODE_WRITE)
+		mutex_unlock(&channel->rd_mutex);
+unlock_wr:
+	if (filp->f_mode & FMODE_READ)
+		mutex_unlock(&channel->wr_mutex);
+
+	if (!rc && (!channel->seekable))
+		return nonseekable_open(inode, filp);
+
+	return rc;
+}
+
+static int xillybus_release(struct inode *inode, struct file *filp)
+{
+	unsigned long flags;
+	struct xilly_channel *channel = filp->private_data;
+
+	int buf_idx;
+	int eof;
+
+	if (channel->endpoint->fatal_error)
+		return -EIO;
+
+	if (filp->f_mode & FMODE_WRITE) {
+		mutex_lock(&channel->rd_mutex);
+
+		channel->rd_ref_count--;
+
+		if (channel->rd_ref_count == 0) {
+			/*
+			 * We rely on the kernel calling flush()
+			 * before we get here.
+			 */
+
+			iowrite32((channel->chan_num << 1) | /* Channel ID */
+				  (5 << 24),  /* Opcode 5, close channel */
+				  channel->endpoint->registers +
+				  fpga_buf_ctrl_reg);
+		}
+		mutex_unlock(&channel->rd_mutex);
+	}
+
+	if (filp->f_mode & FMODE_READ) {
+		mutex_lock(&channel->wr_mutex);
+
+		channel->wr_ref_count--;
+
+		if (channel->wr_ref_count == 0) {
+			iowrite32(1 | (channel->chan_num << 1) |
+				  (5 << 24),  /* Opcode 5, close channel */
+				  channel->endpoint->registers +
+				  fpga_buf_ctrl_reg);
+
+			/*
+			 * This is crazily cautious: We make sure that not
+			 * only that we got an EOF (be it because we closed
+			 * the channel or because of a user's EOF), but verify
+			 * that it's one beyond the last buffer arrived, so
+			 * we have no leftover buffers pending before wrapping
+			 * up (which can only happen in asynchronous channels,
+			 * BTW)
+			 */
+
+			while (1) {
+				spin_lock_irqsave(&channel->wr_spinlock,
+						  flags);
+				buf_idx = channel->wr_fpga_buf_idx;
+				eof = channel->wr_eof;
+				channel->wr_sleepy = 1;
+				spin_unlock_irqrestore(&channel->wr_spinlock,
+						       flags);
+
+				/*
+				 * Check if eof points at the buffer after
+				 * the last one the FPGA submitted. Note that
+				 * no EOF is marked by negative eof.
+				 */
+
+				buf_idx++;
+				if (buf_idx == channel->num_wr_buffers)
+					buf_idx = 0;
+
+				if (buf_idx == eof)
+					break;
+
+				/*
+				 * Steal extra 100 ms if awaken by interrupt.
+				 * This is a simple workaround for an
+				 * interrupt pending when entering, which would
+				 * otherwise result in declaring the hardware
+				 * non-responsive.
+				 */
+
+				if (wait_event_interruptible(
+					    channel->wr_wait,
+					    (!channel->wr_sleepy)))
+					msleep(100);
+
+				if (channel->wr_sleepy) {
+					mutex_unlock(&channel->wr_mutex);
+					dev_warn(channel->endpoint->dev,
+						 "Hardware failed to respond to close command, therefore left in messy state.\n");
+					return -EINTR;
+				}
+			}
+		}
+
+		mutex_unlock(&channel->wr_mutex);
+	}
+
+	return 0;
+}
+
+static loff_t xillybus_llseek(struct file *filp, loff_t offset, int whence)
+{
+	struct xilly_channel *channel = filp->private_data;
+	loff_t pos = filp->f_pos;
+	int rc = 0;
+
+	/*
+	 * 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.
+	 */
+
+	if (channel->endpoint->fatal_error)
+		return -EIO;
+
+	mutex_lock(&channel->wr_mutex);
+	mutex_lock(&channel->rd_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 << channel->log2_element_size) - 1)) {
+		rc = -EINVAL;
+		goto end;
+	}
+
+	mutex_lock(&channel->endpoint->register_mutex);
+
+	iowrite32(pos >> channel->log2_element_size,
+		  channel->endpoint->registers + fpga_buf_offset_reg);
+
+	iowrite32((channel->chan_num << 1) |
+		  (6 << 24),  /* Opcode 6, set address */
+		  channel->endpoint->registers + fpga_buf_ctrl_reg);
+
+	mutex_unlock(&channel->endpoint->register_mutex);
+
+end:
+	mutex_unlock(&channel->rd_mutex);
+	mutex_unlock(&channel->wr_mutex);
+
+	if (rc) /* Return error after releasing mutexes */
+		return rc;
+
+	filp->f_pos = pos;
+
+	/*
+	 * Since seekable devices are allowed only when the channel is
+	 * synchronous, we assume that there is no data pending in either
+	 * direction (which holds true as long as no concurrent access on the
+	 * file descriptor takes place).
+	 * The only thing we may need to throw away is leftovers from partial
+	 * write() flush.
+	 */
+
+	channel->rd_leftovers[3] = 0;
+
+	return pos;
+}
+
+static __poll_t xillybus_poll(struct file *filp, poll_table *wait)
+{
+	struct xilly_channel *channel = filp->private_data;
+	__poll_t mask = 0;
+	unsigned long flags;
+
+	poll_wait(filp, &channel->endpoint->ep_wait, wait);
+
+	/*
+	 * poll() won't play ball regarding read() channels which
+	 * aren't asynchronous and support the nonempty message. Allowing
+	 * that will create situations where data has been delivered at
+	 * the FPGA, and users expecting select() to wake up, which it may
+	 * not.
+	 */
+
+	if (!channel->wr_synchronous && channel->wr_supports_nonempty) {
+		poll_wait(filp, &channel->wr_wait, wait);
+		poll_wait(filp, &channel->wr_ready_wait, wait);
+
+		spin_lock_irqsave(&channel->wr_spinlock, flags);
+		if (!channel->wr_empty || channel->wr_ready)
+			mask |= EPOLLIN | EPOLLRDNORM;
+
+		if (channel->wr_hangup)
+			/*
+			 * Not EPOLLHUP, because its behavior is in the
+			 * mist, and EPOLLIN does what we want: Wake up
+			 * the read file descriptor so it sees EOF.
+			 */
+			mask |=  EPOLLIN | EPOLLRDNORM;
+		spin_unlock_irqrestore(&channel->wr_spinlock, flags);
+	}
+
+	/*
+	 * If partial data write is disallowed on a write() channel,
+	 * it's pointless to ever signal OK to write, because is could
+	 * block despite some space being available.
+	 */
+
+	if (channel->rd_allow_partial) {
+		poll_wait(filp, &channel->rd_wait, wait);
+
+		spin_lock_irqsave(&channel->rd_spinlock, flags);
+		if (!channel->rd_full)
+			mask |= EPOLLOUT | EPOLLWRNORM;
+		spin_unlock_irqrestore(&channel->rd_spinlock, flags);
+	}
+
+	if (channel->endpoint->fatal_error)
+		mask |= EPOLLERR;
+
+	return mask;
+}
+
+static const struct file_operations xillybus_fops = {
+	.owner      = THIS_MODULE,
+	.read       = xillybus_read,
+	.write      = xillybus_write,
+	.open       = xillybus_open,
+	.flush      = xillybus_flush,
+	.release    = xillybus_release,
+	.llseek     = xillybus_llseek,
+	.poll       = xillybus_poll,
+};
+
+struct xilly_endpoint *xillybus_init_endpoint(struct device *dev)
+{
+	struct xilly_endpoint *endpoint;
+
+	endpoint = devm_kzalloc(dev, sizeof(*endpoint), GFP_KERNEL);
+	if (!endpoint)
+		return NULL;
+
+	endpoint->dev = dev;
+	endpoint->msg_counter = 0x0b;
+	endpoint->failed_messages = 0;
+	endpoint->fatal_error = 0;
+
+	init_waitqueue_head(&endpoint->ep_wait);
+	mutex_init(&endpoint->register_mutex);
+
+	return endpoint;
+}
+EXPORT_SYMBOL(xillybus_init_endpoint);
+
+static int xilly_quiesce(struct xilly_endpoint *endpoint)
+{
+	long t;
+
+	endpoint->idtlen = -1;
+
+	iowrite32((u32) (endpoint->dma_using_dac & 0x0001),
+		  endpoint->registers + fpga_dma_control_reg);
+
+	t = wait_event_interruptible_timeout(endpoint->ep_wait,
+					     (endpoint->idtlen >= 0),
+					     XILLY_TIMEOUT);
+	if (t <= 0) {
+		dev_err(endpoint->dev,
+			"Failed to quiesce the device on exit.\n");
+		return -ENODEV;
+	}
+	return 0;
+}
+
+int xillybus_endpoint_discovery(struct xilly_endpoint *endpoint)
+{
+	int rc;
+	long t;
+
+	void *bootstrap_resources;
+	int idtbuffersize = (1 << PAGE_SHIFT);
+	struct device *dev = endpoint->dev;
+
+	/*
+	 * The bogus IDT is used during bootstrap for allocating the initial
+	 * message buffer, and then the message buffer and space for the IDT
+	 * itself. The initial message buffer is of a single page's size, but
+	 * it's soon replaced with a more modest one (and memory is freed).
+	 */
+
+	unsigned char bogus_idt[8] = { 1, 224, (PAGE_SHIFT)-2, 0,
+				       3, 192, PAGE_SHIFT, 0 };
+	struct xilly_idt_handle idt_handle;
+
+	/*
+	 * Writing the value 0x00000001 to Endianness register signals which
+	 * endianness this processor is using, so the FPGA can swap words as
+	 * necessary.
+	 */
+
+	iowrite32(1, endpoint->registers + fpga_endian_reg);
+
+	/* Bootstrap phase I: Allocate temporary message buffer */
+
+	bootstrap_resources = devres_open_group(dev, NULL, GFP_KERNEL);
+	if (!bootstrap_resources)
+		return -ENOMEM;
+
+	endpoint->num_channels = 0;
+
+	rc = xilly_setupchannels(endpoint, bogus_idt, 1);
+	if (rc)
+		return rc;
+
+	/* Clear the message subsystem (and counter in particular) */
+	iowrite32(0x04, endpoint->registers + fpga_msg_ctrl_reg);
+
+	endpoint->idtlen = -1;
+
+	/*
+	 * Set DMA 32/64 bit mode, quiesce the device (?!) and get IDT
+	 * buffer size.
+	 */
+	iowrite32((u32) (endpoint->dma_using_dac & 0x0001),
+		  endpoint->registers + fpga_dma_control_reg);
+
+	t = wait_event_interruptible_timeout(endpoint->ep_wait,
+					     (endpoint->idtlen >= 0),
+					     XILLY_TIMEOUT);
+	if (t <= 0) {
+		dev_err(endpoint->dev, "No response from FPGA. Aborting.\n");
+		return -ENODEV;
+	}
+
+	/* Enable DMA */
+	iowrite32((u32) (0x0002 | (endpoint->dma_using_dac & 0x0001)),
+		  endpoint->registers + fpga_dma_control_reg);
+
+	/* Bootstrap phase II: Allocate buffer for IDT and obtain it */
+	while (endpoint->idtlen >= idtbuffersize) {
+		idtbuffersize *= 2;
+		bogus_idt[6]++;
+	}
+
+	endpoint->num_channels = 1;
+
+	rc = xilly_setupchannels(endpoint, bogus_idt, 2);
+	if (rc)
+		goto failed_idt;
+
+	rc = xilly_obtain_idt(endpoint);
+	if (rc)
+		goto failed_idt;
+
+	rc = xilly_scan_idt(endpoint, &idt_handle);
+	if (rc)
+		goto failed_idt;
+
+	devres_close_group(dev, bootstrap_resources);
+
+	/* Bootstrap phase III: Allocate buffers according to IDT */
+
+	rc = xilly_setupchannels(endpoint,
+				 idt_handle.chandesc,
+				 idt_handle.entries);
+	if (rc)
+		goto failed_idt;
+
+	rc = xillybus_init_chrdev(dev, &xillybus_fops,
+				  endpoint->owner, endpoint,
+				  idt_handle.names,
+				  idt_handle.names_len,
+				  endpoint->num_channels,
+				  xillyname, false);
+
+	if (rc)
+		goto failed_idt;
+
+	devres_release_group(dev, bootstrap_resources);
+
+	return 0;
+
+failed_idt:
+	xilly_quiesce(endpoint);
+	flush_workqueue(xillybus_wq);
+
+	return rc;
+}
+EXPORT_SYMBOL(xillybus_endpoint_discovery);
+
+void xillybus_endpoint_remove(struct xilly_endpoint *endpoint)
+{
+	xillybus_cleanup_chrdev(endpoint, endpoint->dev);
+
+	xilly_quiesce(endpoint);
+
+	/*
+	 * Flushing is done upon endpoint release to prevent access to memory
+	 * just about to be released. This makes the quiesce complete.
+	 */
+	flush_workqueue(xillybus_wq);
+}
+EXPORT_SYMBOL(xillybus_endpoint_remove);
+
+static int __init xillybus_init(void)
+{
+	xillybus_wq = alloc_workqueue(xillyname, WQ_UNBOUND, 0);
+	if (!xillybus_wq)
+		return -ENOMEM;
+
+	return 0;
+}
+
+static void __exit xillybus_exit(void)
+{
+	/* flush_workqueue() was called for each endpoint released */
+	destroy_workqueue(xillybus_wq);
+}
+
+module_init(xillybus_init);
+module_exit(xillybus_exit);
diff --git a/drivers/char/xillybus/xillybus_of.c b/drivers/char/xillybus/xillybus_of.c
new file mode 100644
index 000000000..46e1046ab
--- /dev/null
+++ b/drivers/char/xillybus/xillybus_of.c
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * linux/drivers/misc/xillybus_of.c
+ *
+ * Copyright 2011 Xillybus Ltd, http://xillybus.com
+ *
+ * Driver for the Xillybus FPGA/host framework using Open Firmware.
+ */
+
+#include <linux/module.h>
+#include <linux/device.h>
+#include <linux/slab.h>
+#include <linux/platform_device.h>
+#include <linux/of.h>
+#include <linux/err.h>
+#include "xillybus.h"
+
+MODULE_DESCRIPTION("Xillybus driver for Open Firmware");
+MODULE_AUTHOR("Eli Billauer, Xillybus Ltd.");
+MODULE_ALIAS("xillybus_of");
+MODULE_LICENSE("GPL v2");
+
+static const char xillyname[] = "xillybus_of";
+
+/* Match table for of_platform binding */
+static const struct of_device_id xillybus_of_match[] = {
+	{ .compatible = "xillybus,xillybus-1.00.a", },
+	{ .compatible = "xlnx,xillybus-1.00.a", }, /* Deprecated */
+	{}
+};
+
+MODULE_DEVICE_TABLE(of, xillybus_of_match);
+
+static int xilly_drv_probe(struct platform_device *op)
+{
+	struct device *dev = &op->dev;
+	struct xilly_endpoint *endpoint;
+	int rc;
+	int irq;
+
+	endpoint = xillybus_init_endpoint(dev);
+
+	if (!endpoint)
+		return -ENOMEM;
+
+	dev_set_drvdata(dev, endpoint);
+
+	endpoint->owner = THIS_MODULE;
+
+	endpoint->registers = devm_platform_ioremap_resource(op, 0);
+	if (IS_ERR(endpoint->registers))
+		return PTR_ERR(endpoint->registers);
+
+	irq = platform_get_irq(op, 0);
+
+	rc = devm_request_irq(dev, irq, xillybus_isr, 0, xillyname, endpoint);
+
+	if (rc)
+		return -ENODEV;
+
+	return xillybus_endpoint_discovery(endpoint);
+}
+
+static void xilly_drv_remove(struct platform_device *op)
+{
+	struct device *dev = &op->dev;
+	struct xilly_endpoint *endpoint = dev_get_drvdata(dev);
+
+	xillybus_endpoint_remove(endpoint);
+}
+
+static struct platform_driver xillybus_platform_driver = {
+	.probe = xilly_drv_probe,
+	.remove = xilly_drv_remove,
+	.driver = {
+		.name = xillyname,
+		.of_match_table = xillybus_of_match,
+	},
+};
+
+module_platform_driver(xillybus_platform_driver);
diff --git a/drivers/char/xillybus/xillybus_pcie.c b/drivers/char/xillybus/xillybus_pcie.c
new file mode 100644
index 000000000..32064b6c7
--- /dev/null
+++ b/drivers/char/xillybus/xillybus_pcie.c
@@ -0,0 +1,124 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * linux/drivers/misc/xillybus_pcie.c
+ *
+ * Copyright 2011 Xillybus Ltd, http://xillybus.com
+ *
+ * Driver for the Xillybus FPGA/host framework using PCI Express.
+ */
+
+#include <linux/module.h>
+#include <linux/pci.h>
+#include <linux/slab.h>
+#include "xillybus.h"
+
+MODULE_DESCRIPTION("Xillybus driver for PCIe");
+MODULE_AUTHOR("Eli Billauer, Xillybus Ltd.");
+MODULE_ALIAS("xillybus_pcie");
+MODULE_LICENSE("GPL v2");
+
+#define PCI_DEVICE_ID_XILLYBUS		0xebeb
+
+#define PCI_VENDOR_ID_ACTEL		0x11aa
+#define PCI_VENDOR_ID_LATTICE		0x1204
+
+static const char xillyname[] = "xillybus_pcie";
+
+static const struct pci_device_id xillyids[] = {
+	{PCI_DEVICE(PCI_VENDOR_ID_XILINX, PCI_DEVICE_ID_XILLYBUS)},
+	{PCI_DEVICE(PCI_VENDOR_ID_ALTERA, PCI_DEVICE_ID_XILLYBUS)},
+	{PCI_DEVICE(PCI_VENDOR_ID_ACTEL, PCI_DEVICE_ID_XILLYBUS)},
+	{PCI_DEVICE(PCI_VENDOR_ID_LATTICE, PCI_DEVICE_ID_XILLYBUS)},
+	{ /* End: all zeroes */ }
+};
+
+static int xilly_probe(struct pci_dev *pdev,
+		       const struct pci_device_id *ent)
+{
+	struct xilly_endpoint *endpoint;
+	int rc;
+
+	endpoint = xillybus_init_endpoint(&pdev->dev);
+
+	if (!endpoint)
+		return -ENOMEM;
+
+	pci_set_drvdata(pdev, endpoint);
+
+	endpoint->owner = THIS_MODULE;
+
+	rc = pcim_enable_device(pdev);
+	if (rc) {
+		dev_err(endpoint->dev,
+			"pcim_enable_device() failed. Aborting.\n");
+		return rc;
+	}
+
+	/* L0s has caused packet drops. No power saving, thank you. */
+
+	pci_disable_link_state(pdev, PCIE_LINK_STATE_L0S);
+
+	if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM)) {
+		dev_err(endpoint->dev,
+			"Incorrect BAR configuration. Aborting.\n");
+		return -ENODEV;
+	}
+
+	rc = pcim_iomap_regions(pdev, 0x01, xillyname);
+	if (rc) {
+		dev_err(endpoint->dev,
+			"pcim_iomap_regions() failed. Aborting.\n");
+		return rc;
+	}
+
+	endpoint->registers = pcim_iomap_table(pdev)[0];
+
+	pci_set_master(pdev);
+
+	/* Set up a single MSI interrupt */
+	if (pci_enable_msi(pdev)) {
+		dev_err(endpoint->dev,
+			"Failed to enable MSI interrupts. Aborting.\n");
+		return -ENODEV;
+	}
+	rc = devm_request_irq(&pdev->dev, pdev->irq, xillybus_isr, 0,
+			      xillyname, endpoint);
+	if (rc)
+		return -ENODEV;
+
+	/*
+	 * Some (old and buggy?) hardware drops 64-bit addressed PCIe packets,
+	 * even when the PCIe driver claims that a 64-bit mask is OK. On the
+	 * other hand, on some architectures, 64-bit addressing is mandatory.
+	 * So go for the 64-bit mask only when failing is the other option.
+	 */
+
+	if (!dma_set_mask(&pdev->dev, DMA_BIT_MASK(32))) {
+		endpoint->dma_using_dac = 0;
+	} else if (!dma_set_mask(&pdev->dev, DMA_BIT_MASK(64))) {
+		endpoint->dma_using_dac = 1;
+	} else {
+		dev_err(endpoint->dev, "Failed to set DMA mask. Aborting.\n");
+		return -ENODEV;
+	}
+
+	return xillybus_endpoint_discovery(endpoint);
+}
+
+static void xilly_remove(struct pci_dev *pdev)
+{
+	struct xilly_endpoint *endpoint = pci_get_drvdata(pdev);
+
+	xillybus_endpoint_remove(endpoint);
+}
+
+MODULE_DEVICE_TABLE(pci, xillyids);
+
+static struct pci_driver xillybus_driver = {
+	.name = xillyname,
+	.id_table = xillyids,
+	.probe = xilly_probe,
+	.remove = xilly_remove,
+};
+
+module_pci_driver(xillybus_driver);
diff --git a/drivers/char/xillybus/xillyusb.c b/drivers/char/xillybus/xillyusb.c
new file mode 100644
index 000000000..34e7ad3bc
--- /dev/null
+++ b/drivers/char/xillybus/xillyusb.c
@@ -0,0 +1,2303 @@
+// 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_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;
+static struct workqueue_struct *wakeup_wq;
+
+#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() */
+};
+
+/*
+ * kref_mutex is used in xillyusb_open() to prevent the xillyusb_dev
+ * struct from being freed during the gap between being found by
+ * xillybus_find_inode() and having its reference count incremented.
+ */
+
+static DEFINE_MUTEX(kref_mutex);
+
+/* 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_obj(*ep);
+
+	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_obj(*xb);
+
+		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);
+
+	usb_put_dev(xdev->udev);
+	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().
+ */
+
+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(wakeup_wq, &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 overflowed 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;
+
+	mutex_lock(&kref_mutex);
+
+	rc = xillybus_find_inode(inode, (void **)&xdev, &index);
+	if (rc) {
+		mutex_unlock(&kref_mutex);
+		return rc;
+	}
+
+	kref_get(&xdev->kref);
+	mutex_unlock(&kref_mutex);
+
+	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;
+
+	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_obj(*in_fifo);
+
+		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:
+	kref_put(&xdev->kref, cleanup_dev);
+	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)
+{
+	struct usb_device *udev = xdev->udev;
+
+	/* Verify that device has the two fundamental bulk in/out endpoints */
+	if (usb_pipe_type_check(udev, usb_sndbulkpipe(udev, MSG_EP_NUM)) ||
+	    usb_pipe_type_check(udev, usb_rcvbulkpipe(udev, IN_EP_NUM)))
+		return -ENODEV;
+
+	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 */
+	xdev->msg_ep = NULL;
+	return -ENOMEM;
+}
+
+static int setup_channels(struct xillyusb_dev *xdev,
+			  __le16 *chandesc,
+			  int num_channels)
+{
+	struct usb_device *udev = xdev->udev;
+	struct xillyusb_channel *chan, *new_channels;
+	int i;
+
+	chan = kzalloc_objs(*chan, num_channels);
+	if (!chan)
+		return -ENOMEM;
+
+	new_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 */
+			if (usb_pipe_type_check(udev,
+						usb_sndbulkpipe(udev, i + 2))) {
+				dev_err(xdev->dev,
+					"Missing BULK OUT endpoint %d\n",
+					i + 2);
+				kfree(new_channels);
+				return -ENODEV;
+			}
+
+			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;
+		}
+	}
+
+	xdev->channels = new_channels;
+	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
+	 * device's workqueue and the wakeup work item. This rarely
+	 * does anything.
+	 */
+	flush_workqueue(xdev->workq);
+	flush_work(&xdev->wakeup_workitem);
+
+	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_obj(*xdev);
+	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 | WQ_UNBOUND, 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;
+
+	mutex_lock(&kref_mutex);
+	kref_put(&xdev->kref, cleanup_dev);
+	mutex_unlock(&kref_mutex);
+}
+
+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;
+
+	wakeup_wq = alloc_workqueue(xillyname, WQ_UNBOUND, 0);
+	if (!wakeup_wq)
+		return -ENOMEM;
+
+	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);
+
+	if (rc)
+		destroy_workqueue(wakeup_wq);
+
+	return rc;
+}
+
+static void __exit xillyusb_exit(void)
+{
+	usb_deregister(&xillyusb_driver);
+
+	destroy_workqueue(wakeup_wq);
+}
+
+module_init(xillyusb_init);
+module_exit(xillyusb_exit);