Patchwork [16/29] i2400m: RX and TX data/control paths

login
register
mail settings
Submitter Inaky Perez-Gonzalez
Date Dec. 8, 2008, 7:09 p.m.
Message ID <0c36ac28174f7594738955696f091c3645f672ed.1228759689.git.inaky@linux.intel.com>
Download mbox | patch
Permalink /patch/12834/
State Not Applicable
Delegated to: David Miller
Headers show

Comments

Inaky Perez-Gonzalez - Dec. 8, 2008, 7:09 p.m.
Handling of TX/RX data to/from the i2400m device (IP packets, control
and diagnostics). On RX, this parses the received read transaction
from the device, breaks it in chunks and passes it to the
corresponding subsystems (network and control).

Transmission to the device is done through a software FIFO, as
data/control frames can be coalesced (while the device is reading the
previous tx transaction, others accumulate). A FIFO is used because at
the end it is resource-cheaper that scatter/gather over USB. As well,
most traffic is going to be download (vs upload).

Signed-off-by: Inaky Perez-Gonzalez <inaky@linux.intel.com>
---
 drivers/net/wimax/i2400m/rx.c |  536 +++++++++++++++++++++++++++
 drivers/net/wimax/i2400m/tx.c |  817 +++++++++++++++++++++++++++++++++++++++++
 2 files changed, 1353 insertions(+), 0 deletions(-)
 create mode 100644 drivers/net/wimax/i2400m/rx.c
 create mode 100644 drivers/net/wimax/i2400m/tx.c

Patch

diff --git a/drivers/net/wimax/i2400m/rx.c b/drivers/net/wimax/i2400m/rx.c
new file mode 100644
index 0000000..dbe9e5e
--- /dev/null
+++ b/drivers/net/wimax/i2400m/rx.c
@@ -0,0 +1,536 @@ 
+/*
+ * Intel Wireless WiMAX Connection 2400m
+ * Handle incoming traffic and deliver it to the control or data planes
+ *
+ *
+ * Copyright (C) 2007-2008 Intel Corporation. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ *   * Redistributions of source code must retain the above copyright
+ *     notice, this list of conditions and the following disclaimer.
+ *   * Redistributions in binary form must reproduce the above copyright
+ *     notice, this list of conditions and the following disclaimer in
+ *     the documentation and/or other materials provided with the
+ *     distribution.
+ *   * Neither the name of Intel Corporation nor the names of its
+ *     contributors may be used to endorse or promote products derived
+ *     from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ *
+ * Intel Corporation <linux-wimax@intel.com>
+ * Yanir Lubetkin <yanirx.lubetkin@intel.com>
+ *  - Initial implementation
+ * Inaky Perez-Gonzalez <inaky.perez-gonzalez@intel.com>
+ *  - Use skb_clone(), break up processing in chunks
+ *  - Split transport/device specific
+ *  - Make buffer size dynamic to exert less memory pressure
+ *
+ *
+ * This handles the RX path.
+ *
+ * We receive an RX message from the bus-specific driver, which
+ * contains one or more payloads that have potentially different
+ * destinataries (data or control paths).
+ *
+ * So we just take that payload from the transport specific code in
+ * the form of an skb, break it up in chunks (a cloned skb each in the
+ * case of network packets) and pass it to netdev or to the
+ * command/ack handler (and from there to the WiMAX stack).
+ *
+ * PROTOCOL FORMAT
+ *
+ * The format of the buffer is:
+ *
+ * HEADER                      (struct i2400m_msg_hdr)
+ * PAYLOAD DESCRIPTOR 0        (struct i2400m_pld)
+ * PAYLOAD DESCRIPTOR 1
+ * ...
+ * PAYLOAD DESCRIPTOR N
+ * PAYLOAD 0                   (raw bytes)
+ * PAYLOAD 1
+ * ...
+ * PAYLOAD N
+ *
+ * See tx.c for a deeper description on alignment requirements and
+ * other fun facts of it.
+ *
+ * ROADMAP
+ *
+ * i2400m_rx
+ *   i2400m_rx_msg_hdr_check
+ *   i2400m_rx_pl_descr_check
+ *   i2400m_rx_payload
+ *     i2400m_net_rx
+ *     i2400m_rx_ctl
+ *       i2400m_msg_size_check
+ *       i2400m_report_hook_work    [in a workqueue]
+ *         i2400m_report_hook
+ *       wimax_msg_to_user
+ *       i2400m_rx_ctl_ack
+ *         wimax_msg_to_user_alloc
+ *     i2400m_rx_trace
+ *       i2400m_msg_size_check
+ *       wimax_pipe_msg
+ */
+#include <linux/kernel.h>
+#include <linux/if_arp.h>
+#include <linux/netdevice.h>
+#include <linux/workqueue.h>
+#include "i2400m.h"
+
+
+#define D_SUBMODULE rx
+#include "debug-levels.h"
+
+struct i2400m_report_hook_args {
+	struct sk_buff *skb_rx;
+	const struct i2400m_l3l4_hdr *l3l4_hdr;
+	size_t size;
+};
+
+
+/*
+ * Execute i2400m_report_hook in a workqueue
+ *
+ * Unpacks arguments from the deferred call, executes it and then
+ * drops the references.
+ *
+ * Obvious NOTE: References are needed because we are a separate
+ *     thread; otherwise the buffer changes under us because it is
+ *     released by the original caller.
+ */
+static
+void i2400m_report_hook_work(struct work_struct *ws)
+{
+	struct i2400m_work *iw =
+		container_of(ws, struct i2400m_work, ws);
+	struct i2400m_report_hook_args *args = (void *) iw->pl;
+	i2400m_report_hook(iw->i2400m, args->l3l4_hdr, args->size);
+	kfree_skb(args->skb_rx);
+	i2400m_put(iw->i2400m);
+	kfree(iw);
+}
+
+
+/*
+ * Process an ack to a command
+ *
+ * @i2400m: device descriptor
+ * @payload: pointer to message
+ * @size: size of the message
+ *
+ * Pass the acknodledgment (in an skb) to the thread that is waiting
+ * for it in i2400m->msg_completion.
+ *
+ * We need to coordinate properly with the thread waiting for the
+ * ack. Check if it is waiting or if it is gone. We loose the spinlock
+ * to avoid allocating on atomic contexts (yeah, could use GFP_ATOMIC,
+ * but this is not so speed critical).
+ */
+static
+void i2400m_rx_ctl_ack(struct i2400m *i2400m,
+		       const void *payload, size_t size)
+{
+	struct device *dev = i2400m_dev(i2400m);
+	struct wimax_dev *wimax_dev = &i2400m->wimax_dev;
+	unsigned long flags;
+	struct sk_buff *ack_skb;
+
+	/* Anyone waiting for an answer? */
+	spin_lock_irqsave(&i2400m->rx_lock, flags);
+	if (i2400m->ack_skb != ERR_PTR(-EINPROGRESS)) {
+		dev_err(dev, "Huh? reply to command with no waiters\n");
+		goto error_no_waiter;
+	}
+	spin_unlock_irqrestore(&i2400m->rx_lock, flags);
+
+	ack_skb = wimax_msg_to_user_alloc(
+		wimax_dev, payload, size, GFP_KERNEL);
+
+	/* Check waiter didn't time out waiting for the answer... */
+	spin_lock_irqsave(&i2400m->rx_lock, flags);
+	if (i2400m->ack_skb != ERR_PTR(-EINPROGRESS)) {
+		d_printf(1, dev, "Huh? waiter for command reply cancelled\n");
+		goto error_waiter_cancelled;
+	}
+	if (ack_skb == NULL) {
+		dev_err(dev, "CMD/GET/SET ack: cannot allocate SKB\n");
+		i2400m->ack_skb = ERR_PTR(-ENOMEM);
+	} else
+		i2400m->ack_skb = ack_skb;
+	spin_unlock_irqrestore(&i2400m->rx_lock, flags);
+	complete(&i2400m->msg_completion);
+	return;
+
+error_waiter_cancelled:
+	if (ack_skb)
+		kfree_skb(ack_skb);
+error_no_waiter:
+	spin_unlock_irqrestore(&i2400m->rx_lock, flags);
+	return;
+}
+
+
+/*
+ * Receive and process a control payload
+ *
+ * @i2400m: device descriptor
+ * @skb_rx: skb that contains the payload (for reference counting)
+ * @payload: pointer to message
+ * @size: size of the message
+ *
+ * There are two types of control RX messages: reports (asynchronous,
+ * like your every day interrupts) and 'acks' (reponses to a command,
+ * get or set request).
+ *
+ * If it is a report, we run hooks on it (to extract information for
+ * things we need to do in the driver) and then pass it over to the
+ * WiMAX stack to send it to user space.
+ *
+ * NOTE: report processing is done in a workqueue specific to the
+ *     generic driver, to avoid deadlocks in the system.
+ *
+ * If it is not a report, it is an ack to a previously executed
+ * command, set or get, so wake up whoever is waiting for it from
+ * i2400m_msg_to_dev(). i2400m_rx_ctl_ack() takes care of that.
+ *
+ * Note that the sizes we pass to other functions from here are the
+ * sizes of the _l3l4_hdr + payload, not full buffer sizes, as we have
+ * verified in _msg_size_check() that they are congruent.
+ *
+ * For reports: We can't clone the original skb where the data is
+ * because we need to send this up via netlink; netlink has to add
+ * headers and we can't overwrite what's preceeding the payload...as
+ * it is another message. So we just dup them.
+ */
+static
+void i2400m_rx_ctl(struct i2400m *i2400m, struct sk_buff *skb_rx,
+		   const void *payload, size_t size)
+{
+	int result;
+	struct device *dev = i2400m_dev(i2400m);
+	const struct i2400m_l3l4_hdr *l3l4_hdr = payload;
+	unsigned msg_type;
+
+	result = i2400m_msg_size_check(i2400m, l3l4_hdr, size);
+	if (result < 0) {
+		dev_err(dev, "HW BUG? device sent a bad message: %d\n",
+			result);
+		goto error_check;
+	}
+	msg_type = le16_to_cpu(l3l4_hdr->type);
+	d_printf(1, dev, "%s 0x%04x: %zu bytes\n",
+		 msg_type & I2400M_MT_REPORT_MASK ? "REPORT" : "CMD/SET/GET",
+		 msg_type, size);
+	d_dump(2, dev, l3l4_hdr, size);
+	if (msg_type & I2400M_MT_REPORT_MASK) {
+		/* These hooks have to be ran serialized; as well, the
+		 * handling might force the execution of commands, and
+		 * that might cause reentrancy issues with
+		 * bus-specific subdrivers and workqueues. So we run
+		 * it in a separate workqueue. */
+		struct i2400m_report_hook_args args = {
+			.skb_rx = skb_rx,
+			.l3l4_hdr = l3l4_hdr,
+			.size = size
+		};
+		if (unlikely(i2400m->ready == 0))	/* only send if up */
+			return;
+		skb_get(skb_rx);
+		i2400m_queue_work(i2400m, i2400m_report_hook_work,
+				  GFP_KERNEL, &args, sizeof(args));
+		result = wimax_msg_to_user(&i2400m->wimax_dev, l3l4_hdr, size,
+					   GFP_KERNEL);
+		if (result < 0)
+			dev_err(dev, "error sending report to userspace: %d\n",
+				result);
+	} else		/* an ack to a CMD, GET or SET */
+		i2400m_rx_ctl_ack(i2400m, payload, size);
+error_check:
+	return;
+}
+
+
+
+
+/*
+ * Receive and send up a trace
+ *
+ * @i2400m: device descriptor
+ * @skb_rx: skb that contains the trace (for reference counting)
+ * @payload: pointer to trace message inside the skb
+ * @size: size of the message
+ *
+ * THe i2400m might produce trace information (diagnostics) and we
+ * send them through a different kernel-to-user pipe (to avoid
+ * clogging it).
+ *
+ * As in i2400m_rx_ctl(), we can't clone the original skb where the
+ * data is because we need to send this up via netlink; netlink has to
+ * add headers and we can't overwrite what's preceeding the
+ * payload...as it is another message. So we just dup them.
+ */
+static
+void i2400m_rx_trace(struct i2400m *i2400m,
+		     const void *payload, size_t size)
+{
+	int result;
+	struct device *dev = i2400m_dev(i2400m);
+	struct wimax_dev *wimax_dev = &i2400m->wimax_dev;
+	const struct i2400m_l3l4_hdr *l3l4_hdr = payload;
+	unsigned msg_type;
+
+	result = i2400m_msg_size_check(i2400m, l3l4_hdr, size);
+	if (result < 0) {
+		dev_err(dev, "HW BUG? device sent a bad trace message: %d\n",
+			result);
+		goto error_check;
+	}
+	msg_type = le16_to_cpu(l3l4_hdr->type);
+	d_printf(1, dev, "Trace %s 0x%04x: %zu bytes\n",
+		 msg_type & I2400M_MT_REPORT_MASK ? "REPORT" : "CMD/SET/GET",
+		 msg_type, size);
+	d_dump(2, dev, l3l4_hdr, size);
+	if (unlikely(i2400m->ready == 0))	/* only send if up */
+		return;
+	result = wimax_pipe_msg(wimax_dev, i2400m->pipe_trace,
+				l3l4_hdr, size, GFP_KERNEL);
+	if (result < 0)
+		dev_err(dev, "error sending trace to userspace: %d\n",
+			result);
+error_check:
+	return;
+}
+
+
+/*
+ * Act on a received payload
+ *
+ * @i2400m: device instance
+ * @skb_rx: skb where the transaction was received
+ * @single: 1 if there is only one payload, 0 otherwise
+ * @pld: payload descriptor
+ * @payload: payload data
+ *
+ * Upon reception of a payload, look at its guts in the payload
+ * descriptor and decide what to do with it.
+ */
+static
+void i2400m_rx_payload(struct i2400m *i2400m, struct sk_buff *skb_rx,
+		       unsigned single, const struct i2400m_pld *pld,
+		       const void *payload)
+{
+	struct device *dev = i2400m_dev(i2400m);
+	size_t pl_size = i2400m_pld_size(pld);
+	enum i2400m_pt pl_type = i2400m_pld_type(pld);
+
+	switch (pl_type) {
+	case I2400M_PT_DATA:
+		d_printf(3, dev, "RX: data payload %zu bytes\n", pl_size);
+		i2400m_net_rx(i2400m, skb_rx, single, payload, pl_size);
+		break;
+	case I2400M_PT_CTRL:
+		i2400m_rx_ctl(i2400m, skb_rx, payload, pl_size);
+		break;
+	case I2400M_PT_TRACE:
+		i2400m_rx_trace(i2400m, payload, pl_size);
+		break;
+	default:	/* Anything else shouldn't come to the host */
+		if (printk_ratelimit())
+			dev_err(dev, "RX: HW BUG? unexpected payload type %u\n",
+				pl_type);
+	}
+}
+
+
+/*
+ * Check a received transaction's message header
+ *
+ * @i2400m: device descriptor
+ * @msg_hdr: message header
+ * @buf_size: size of the received buffer
+ *
+ * Check that the declarations done by a RX buffer message header are
+ * sane and consistent with the amount of data that was received.
+ */
+static
+int i2400m_rx_msg_hdr_check(struct i2400m *i2400m,
+			    const struct i2400m_msg_hdr *msg_hdr,
+			    size_t buf_size)
+{
+	int result = -EIO;
+	struct device *dev = i2400m_dev(i2400m);
+	if (buf_size < sizeof(*msg_hdr)) {
+		dev_err(dev, "RX: HW BUG? message with short header (%zu "
+			"vs %zu bytes expected)\n", buf_size, sizeof(*msg_hdr));
+		goto error;
+	}
+	if (msg_hdr->barker != cpu_to_le32(I2400M_D2H_MSG_BARKER)) {
+		dev_err(dev, "RX: HW BUG? message received with unknown "
+			"barker 0x%08x (buf_size %zu bytes)\n",
+			le32_to_cpu(msg_hdr->barker), buf_size);
+		goto error;
+	}
+	if (msg_hdr->num_pls == 0) {
+		dev_err(dev, "RX: HW BUG? zero payload packets in message\n");
+		goto error;
+	}
+	if (le16_to_cpu(msg_hdr->num_pls) > I2400M_MAX_PLS_IN_MSG) {
+		dev_err(dev, "RX: HW BUG? message contains more payload "
+			"than maximum; ignoring.\n");
+		goto error;
+	}
+	result = 0;
+error:
+	return result;
+}
+
+
+/*
+ * Check a payload descriptor against the received data
+ *
+ * @i2400m: device descriptor
+ * @pld: payload descriptor
+ * @pl_itr: offset (in bytes) in the received buffer the payload is
+ *          located
+ * @buf_size: size of the received buffer
+ *
+ * Given a payload descriptor (part of a RX buffer), check it is sane
+ * and that the data it declares fits in the buffer.
+ */
+static
+int i2400m_rx_pl_descr_check(struct i2400m *i2400m,
+			      const struct i2400m_pld *pld,
+			      size_t pl_itr, size_t buf_size)
+{
+	int result = -EIO;
+	struct device *dev = i2400m_dev(i2400m);
+	size_t pl_size = i2400m_pld_size(pld);
+	enum i2400m_pt pl_type = i2400m_pld_type(pld);
+
+	if (pl_size > i2400m->bus_pl_size_max) {
+		dev_err(dev, "RX: HW BUG? payload @%zu: size %zu is "
+			"bigger than maximum %zu; ignoring message\n",
+			pl_itr, pl_size, i2400m->bus_pl_size_max);
+		goto error;
+	}
+	if (pl_itr + pl_size > buf_size) {	/* enough? */
+		dev_err(dev, "RX: HW BUG? payload @%zu: size %zu "
+			"goes beyond the received buffer "
+			"size (%zu bytes); ignoring message\n",
+			pl_itr, pl_size, buf_size);
+		goto error;
+	}
+	if (pl_type >= I2400M_PT_ILLEGAL) {
+		dev_err(dev, "RX: HW BUG? illegal payload type %u; "
+			"ignoring message\n", pl_type);
+		goto error;
+	}
+	result = 0;
+error:
+	return result;
+}
+
+
+/**
+ * i2400m_rx - Receive a buffer of data from the device
+ *
+ * @i2400m: device descriptor
+ * @skb: skbuff where the data has been received
+ *
+ * Parse in a buffer of data that contains an RX message sent from the
+ * device. See the file header for the format. Run all checks on the
+ * buffer header, then run over each payload's descriptors, verify
+ * their consistency and act on each payload's contents.  If
+ * everything is succesful, update the device's statistics.
+ *
+ * Note: You need to set the skb to contain only the length of the
+ * received buffer; for that, use skb_trim(skb, RECEIVED_SIZE).
+ *
+ * Returns:
+ *
+ * 0 if ok, < 0 errno on error
+ *
+ * If ok, this function owns now the skb and the caller DOESN'T have
+ * to run kfree_skb() on it. However, on error, the caller still owns
+ * the skb and it is responsible for releasing it.
+ */
+int i2400m_rx(struct i2400m *i2400m, struct sk_buff *skb)
+{
+	int i, result;
+	struct device *dev = i2400m_dev(i2400m);
+	const struct i2400m_msg_hdr *msg_hdr;
+	size_t pl_itr, pl_size, skb_len;
+	unsigned long flags;
+	unsigned num_pls;
+
+	skb_len = skb->len;
+	d_fnstart(4, dev, "(i2400m %p skb %p [size %zu])\n",
+		  i2400m, skb, skb_len);
+	result = -EIO;
+	msg_hdr = (void *) skb->data;
+	result = i2400m_rx_msg_hdr_check(i2400m, msg_hdr, skb->len);
+	if (result < 0)
+		goto error_msg_hdr_check;
+	result = -EIO;
+	num_pls = le16_to_cpu(msg_hdr->num_pls);
+	pl_itr = sizeof(*msg_hdr) +	/* Check payload descriptor(s) */
+		num_pls * sizeof(msg_hdr->pld[0]);
+	pl_itr = ALIGN(pl_itr, I2400M_PL_PAD);
+	if (pl_itr > skb->len) {	/* got all the payload descriptors? */
+		dev_err(dev, "RX: HW BUG? message too short (%u bytes) for "
+			"%u payload descriptors (%zu each, total %zu)\n",
+			skb->len, num_pls, sizeof(msg_hdr->pld[0]), pl_itr);
+		goto error_pl_descr_short;
+	}
+	/* Walk each payload payload--check we really got it */
+	for (i = 0; i < num_pls; i++) {
+		/* work around old gcc warnings */
+		pl_size = i2400m_pld_size(&msg_hdr->pld[i]);
+		result = i2400m_rx_pl_descr_check(i2400m, &msg_hdr->pld[i],
+						  pl_itr, skb->len);
+		if (result < 0)
+			goto error_pl_descr_check;
+		i2400m_rx_payload(i2400m, skb, num_pls == 1, &msg_hdr->pld[i],
+				  skb->data + pl_itr);
+		pl_itr += ALIGN(pl_size, I2400M_PL_PAD);
+		cond_resched();		/* Don't monopolize */
+	}
+	kfree_skb(skb);
+	/* Update device statistics */
+	spin_lock_irqsave(&i2400m->rx_lock, flags);
+	i2400m->rx_pl_num += i;
+	if (i > i2400m->rx_pl_max)
+		i2400m->rx_pl_max = i;
+	if (i < i2400m->rx_pl_min)
+		i2400m->rx_pl_min = i;
+	i2400m->rx_num++;
+	i2400m->rx_size_acc += skb->len;
+	if (skb->len < i2400m->rx_size_min)
+		i2400m->rx_size_min = skb->len;
+	if (skb->len > i2400m->rx_size_max)
+		i2400m->rx_size_max = skb->len;
+	spin_unlock_irqrestore(&i2400m->rx_lock, flags);
+error_pl_descr_check:
+error_pl_descr_short:
+error_msg_hdr_check:
+	d_fnend(4, dev, "(i2400m %p skb %p [size %zu]) = %d\n",
+		i2400m, skb, skb_len, result);
+	return result;
+}
+EXPORT_SYMBOL_GPL(i2400m_rx);
diff --git a/drivers/net/wimax/i2400m/tx.c b/drivers/net/wimax/i2400m/tx.c
new file mode 100644
index 0000000..613a88f
--- /dev/null
+++ b/drivers/net/wimax/i2400m/tx.c
@@ -0,0 +1,817 @@ 
+/*
+ * Intel Wireless WiMAX Connection 2400m
+ * Generic (non-bus specific) TX handling
+ *
+ *
+ * Copyright (C) 2007-2008 Intel Corporation. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ *   * Redistributions of source code must retain the above copyright
+ *     notice, this list of conditions and the following disclaimer.
+ *   * Redistributions in binary form must reproduce the above copyright
+ *     notice, this list of conditions and the following disclaimer in
+ *     the documentation and/or other materials provided with the
+ *     distribution.
+ *   * Neither the name of Intel Corporation nor the names of its
+ *     contributors may be used to endorse or promote products derived
+ *     from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ *
+ * Intel Corporation <linux-wimax@intel.com>
+ * Yanir Lubetkin <yanirx.lubetkin@intel.com>
+ *  - Initial implementation
+ *
+ * Intel Corporation <linux-wimax@intel.com>
+ * Inaky Perez-Gonzalez <inaky.perez-gonzalez@intel.com>
+ *  - Rewritten to use a single FIFO to lower the memory allocation
+ *    pressure and optimize cache hits when copying to the queue, as
+ *    well as splitting out bus-specific code.
+ *
+ *
+ * Implements data transmission to the device; this is done through a
+ * software FIFO, as data/control frames can be coalesced (while the
+ * device is reading the previous tx transaction, others accumulate).
+ *
+ * A FIFO is used because at the end it is resource-cheaper that trying
+ * to implement scatter/gather over USB. As well, most traffic is going
+ * to be download (vs upload).
+ *
+ * The format for sending/receiving data to/from the i2400m is
+ * described in detail in rx.c:PROTOCOL FORMAT. In here we implement
+ * the transmission of that. This is split between a bus-independent
+ * part that just prepares everything and a bus-specific part that
+ * does the actual transmission over the bus to the device (in the
+ * bus-specific driver).
+ *
+ *
+ * The general format of a device-host transaction is MSG-HDR, PLD1,
+ * PLD2...PLDN, PL1, PL2,...PLN, PADDING.
+ *
+ * Because we need the send payload descriptors and then payloads and
+ * because it is kind of expensive to do scatterlists in USB (one URB
+ * per node), it becomes cheaper to append all the data to a FIFO
+ * (copying to a FIFO potentially in cache is cheaper).
+ *
+ * Then the bus-specific code takes the parts of that FIFO that are
+ * written and passes them to the device.
+ *
+ * So the concepts to keep in mind there are:
+ *
+ * We use a FIFO to queue the data in a linear buffer. We first append
+ * a MSG-HDR, space for I2400M_TX_PLD_MAX payload descriptors and then
+ * go appending payloads until we run out of space or of payload
+ * descriptors. Then we append padding to make the whole transaction a
+ * multiple of i2400m->bus_tx_block_size (as defined by the bus layer).
+ *
+ * - A TX message: a combination of a message header, payload
+ *   descriptors and payloads.
+ *
+ *     Open: it is marked as active (i2400m->tx_msg is valid) and we
+ *       can keep adding payloads to it.
+ *
+ *     Closed: we are not appending more payloads to this TX message
+ *       (exahusted space in the queue, too many payloads or
+ *       whichever).  We have appended padding so the whole message
+ *       length is aligned to i2400m->bus_tx_block_size (as set by the
+ *       bus/transport layer).
+ *
+ * - Most of the time we keep a TX message open to which we append
+ *   payloads.
+ *
+ * - If we are going to append and there is no more space (we are at
+ *   the end of the FIFO), we close the message, mark the rest of the
+ *   FIFO space unusable (skip_tail), create a new message at the
+ *   beginning of the FIFO (if there is space) and append the message
+ *   there.
+ *
+ *   This is because we need to give linear TX messages to the bus
+ *   engine. So we don't write a message to the remaining FIFO space
+ *   until the tail and continue at the head of it.
+ *
+ * - We overload one of the fields in the message header to use it as
+ *   'size' of the TX message, so we can iterate over them. It also
+ *   contains a flag that indicates if we have to skip it or not.
+ *   When we send the buffer, we update that to its real on-the-wire
+ *   value.
+ *
+ * - The MSG-HDR PLD1...PLD2 stuff has to be a size multiple of 16.
+ *
+ *   It follows that if MSG-HDR says we have N messages, the whole
+ *   header + descriptors is 16 + 4*N; for those to be a multiple of
+ *   16, it follows that N can be 4, 8, 12, ... (32, 48, 64, 80...
+ *   bytes).
+ *
+ *   So if we have only 1 payload, we have to submit a header that in
+ *   all truth has space for 4.
+ *
+ *   The implication is that we reserve space for 12 (64 bytes); but
+ *   if we fill up only (eg) 2, our header becomes 32 bytes only. So
+ *   the TX engine has to shift those 32 bytes of msg header and 2
+ *   payloads and padding so that right after it the payloads start
+ *   and the TX engine has to know about that.
+ *
+ *   It is cheaper to move the header up than the whole payloads down.
+ *
+ *   We do this in i2400m_tx_close(). See 'i2400m_msg_hdr->offset'.
+ *
+ * - Each payload has to be size-padded to 16 bytes; before appending
+ *   it, we just do it.
+ *
+ * - The whole message has to be padded to i2400m->bus_tx_block_size;
+ *   we do this at close time. Thus, when reserving space for the
+ *   payload, we always make sure there is also free space for this
+ *   padding that sooner or later will happen.
+ *
+ * When we append a message, we tell the bus specific code to kick in
+ * TXs. It will TX (in parallel) until the buffer is exhausted--hence
+ * the lockin we do. The TX code will only send a TX message at the
+ * time (which remember, might contain more than one payload). Of
+ * course, when the bus-specific driver attempts to TX a message that
+ * is still open, it gets closed first.
+ *
+ * Gee, this is messy; well a picture. In the example below we have a
+ * partially full FIFO, with a closed message ready to be delivered
+ * (with a moved message header to make sure it is size-aligned to
+ * 16), TAIL room that was unusable (and thus is marked with a message
+ * header that says 'skip this') and at the head of the buffer, an
+ * imcomplete message with a couple of payloads.
+ *
+ * N   ___________________________________________________
+ *    |                                                   |
+ *    |     TAIL room                                     |
+ *    |                                                   |
+ *    |  msg_hdr to skip (size |= 0x80000)                |
+ *    |---------------------------------------------------|-------
+ *    |                                                   |  /|\
+ *    |                                                   |   |
+ *    |  TX message padding                               |   |
+ *    |                                                   |   |
+ *    |                                                   |   |
+ *    |- - - - - - - - - - - - - - - - - - - - - - - - - -|   |
+ *    |                                                   |   |
+ *    |  payload 1                                        |   |
+ *    |                                                   | N * tx_block_size
+ *    |                                                   |   |
+ *    |- - - - - - - - - - - - - - - - - - - - - - - - - -|   |
+ *    |                                                   |   |
+ *    |  payload 1                                        |   |
+ *    |                                                   |   |
+ *    |                                                   |   |
+ *    |- - - - - - - - - - - - - - - - - - - - - - - - - -|- -|- - - -
+ *    |  padding 3                  /|\                   |   |   /|\
+ *    |  padding 2                   |                    |   |    |
+ *    |  pld 1                32 bytes (2 * 16)           |   |    |
+ *    |  pld 0                       |                    |   |    |
+ *    |  moved msg_hdr              \|/                   |  \|/   |
+ *    |- - - - - - - - - - - - - - - - - - - - - - - - - -|- - -   |
+ *    |                                                   |    _PLD_SIZE
+ *    |  unused                                           |        |
+ *    |                                                   |        |
+ *    |- - - - - - - - - - - - - - - - - - - - - - - - - -|        |
+ *    |  msg_hdr (size X)       [this message is closed]  |       \|/
+ *    |===================================================|========== <=== OUT
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |          Free rooom                               |
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |===================================================|========== <=== IN
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |                                                   |
+ *    |  payload 1                                        |
+ *    |                                                   |
+ *    |                                                   |
+ *    |- - - - - - - - - - - - - - - - - - - - - - - - - -|
+ *    |                                                   |
+ *    |  payload 0                                        |
+ *    |                                                   |
+ *    |                                                   |
+ *    |- - - - - - - - - - - - - - - - - - - - - - - - - -|
+ *    |  pld 11                     /|\                   |
+ *    |  ...                         |                    |
+ *    |  pld 1                64 bytes (2 * 16)           |
+ *    |  pld 0                       |                    |
+ *    |  msg_hdr (size X)           \|/ [message is open] |
+ * 0   ---------------------------------------------------
+ *
+ *
+ * ROADMAP
+ *
+ * i2400m_tx_setup()           Called by i2400m_setup
+ * i2400m_tx_release()         Called by i2400m_release()
+ *
+ *  i2400m_tx()                 Called to send data or control frames
+ *    i2400m_tx_fifo_push()     Allocates append-space in the FIFO
+ *    i2400m_tx_new()           Opens a new message in the FIFO
+ *    i2400m_tx_fits()          Checks if a new payload fits in the message
+ *    i2400m_tx_close()         Closes an open message in the FIFO
+ *    i2400m_tx_skip_tail()     Marks unusable FIFO tail space
+ *    i2400m->bus_tx_kick()
+ *
+ * Now i2400m->bus_tx_kick() is the the bus-specific driver backend
+ * implementation; that would do:
+ *
+ * i2400m->bus_tx_kick()
+ *   i2400m_tx_msg_get()	Gets first message ready to go
+ *   ...sends it...
+ *   i2400m_tx_msg_sent()       Ack the message is sent; repeat from
+ *                              _tx_msg_get() until it returns NULL
+ *                               (FIFO empty).
+ */
+#include <linux/netdevice.h>
+#include "i2400m.h"
+
+
+#define D_SUBMODULE tx
+#include "debug-levels.h"
+
+enum {
+	/**
+	 * TX Buffer size
+	 *
+	 * Doc says maximum transaction is 16KiB. If we had 16KiB en
+	 * route and 16KiB being queued, it boils down to needing
+	 * 32KiB.
+	 */
+	I2400M_TX_BUF_SIZE = 32768,
+	/**
+	 * Message header and payload descriptors have to be 16
+	 * aligned (16 + 4 * N = 16 * M). If we take that average sent
+	 * packets are MTU size (~1400-~1500) it follows that we could
+	 * fit at most 10-11 payloads in one transaction. To meet the
+	 * alignment requirement, that means we need to leave space
+	 * for 12 (64 bytes). To simplify, we leave space for that. If
+	 * at the end there are less, we pad up to the nearest
+	 * multiple of 16.
+	 */
+	I2400M_TX_PLD_MAX = 12,
+	I2400M_TX_PLD_SIZE = sizeof(struct i2400m_msg_hdr)
+	+ I2400M_TX_PLD_MAX * sizeof(struct i2400m_pld),
+	I2400M_TX_SKIP = 0x80000000,
+};
+
+#define TAIL_FULL ((void *)~(unsigned long)NULL)
+
+/*
+ * Allocate @size bytes in the TX fifo, return a pointer to it
+ *
+ * @i2400m: device descriptor
+ * @size: size of the buffer we need to allocate
+ * @padding: ensure that there is at least this many bytes of free
+ *     contiguous space in the fifo. This is needed because later on
+ *     we might need to add padding.
+ *
+ * Returns:
+ *
+ *     Pointer to the allocated space. NULL if there is no
+ *     space. TAIL_FULL if there is no space at the tail but there is at
+ *     the head (Case B below).
+ *
+ * These are the two basic cases we need to keep an eye for -- it is
+ * much better explained in linux/kernel/kfifo.c, but this code
+ * basically does the same. No rocket science here.
+ *
+ *       Case A               Case B
+ * N  ___________          ___________
+ *   | tail room |        |   data    |
+ *   |           |        |           |
+ *   |<-  IN   ->|        |<-  OUT  ->|
+ *   |           |        |           |
+ *   |   data    |        |   room    |
+ *   |           |        |           |
+ *   |<-  OUT  ->|        |<-  IN   ->|
+ *   |           |        |           |
+ *   | head room |        |   data    |
+ * 0  -----------          -----------
+ *
+ * We allocate only *contiguous* space.
+ *
+ * We can allocate only from 'room'. In Case B, it is simple; in case
+ * A, we only try from the tail room; if it is not enough, we just
+ * fail and return TAIL_FULL and let the caller figure out if we wants to
+ * skip the tail room and try to allocate from the head.
+ *
+ * Note:
+ *
+ *     Assumes i2400m->tx_lock is taken, and we use that as a barrier
+ *
+ *     The indexes keep increasing and we reset them to zero when we
+ *     pop data off the queue
+ */
+static
+void *i2400m_tx_fifo_push(struct i2400m *i2400m, size_t size, size_t padding)
+{
+	struct device *dev = i2400m_dev(i2400m);
+	size_t room, tail_room, needed_size;
+	void *ptr;
+
+	needed_size = size + padding;
+	room = I2400M_TX_BUF_SIZE - (i2400m->tx_in - i2400m->tx_out);
+	if (room < needed_size)	{ /* this takes care of Case B */
+		d_printf(2, dev, "fifo push %zu/%zu: no space\n",
+			 size, padding);
+		return NULL;
+	}
+	/* Is there space at the tail? */
+	tail_room = I2400M_TX_BUF_SIZE - i2400m->tx_in % I2400M_TX_BUF_SIZE;
+	if (tail_room < needed_size) {
+		if (i2400m->tx_out % I2400M_TX_BUF_SIZE
+		    < i2400m->tx_in % I2400M_TX_BUF_SIZE) {
+			d_printf(2, dev, "fifo push %zu/%zu: tail full\n",
+				 size, padding);
+			return TAIL_FULL;	/* There might be head space */
+		} else {
+			d_printf(2, dev, "fifo push %zu/%zu: no head space\n",
+				 size, padding);
+			return NULL;	/* There is no space */
+		}
+	}
+	ptr = i2400m->tx_buf + i2400m->tx_in % I2400M_TX_BUF_SIZE;
+	d_printf(2, dev, "fifo push %zu/%zu: at @%zu\n", size, padding,
+		 i2400m->tx_in % I2400M_TX_BUF_SIZE);
+	i2400m->tx_in += size;
+	return ptr;
+}
+
+
+/*
+ * Mark the tail of the FIFO buffer as 'to-skip'
+ *
+ * We should never hit the BUG_ON() because all the sizes we push to
+ * the FIFO are padded to be a multiple of 16 -- the size of *msg
+ * (I2400M_PL_PAD for the payloads, I2400M_TX_PLD_SIZE for the
+ * header).
+ *
+ * Note:
+ *
+ *     Assumes i2400m->tx_lock is taken, and we use that as a barrier
+ */
+static
+void i2400m_tx_skip_tail(struct i2400m *i2400m)
+{
+	struct device *dev = i2400m_dev(i2400m);
+	size_t tx_in = i2400m->tx_in % I2400M_TX_BUF_SIZE;
+	size_t tail_room = I2400M_TX_BUF_SIZE - tx_in;
+	struct i2400m_msg_hdr *msg = i2400m->tx_buf + tx_in;
+	BUG_ON(tail_room < sizeof(*msg));
+	msg->size = tail_room | I2400M_TX_SKIP;
+	d_printf(2, dev, "skip tail: skipping %zu bytes @%zu\n",
+		 tail_room, tx_in);
+	i2400m->tx_in += tail_room;
+}
+
+
+/*
+ * Check if a skb will fit in the TX queue's current active TX
+ * message (if there are still descriptors left unused).
+ *
+ * Returns:
+ *     0 if the message won't fit, 1 if it will.
+ *
+ * Note:
+ *
+ *     Assumes a TX message is active (i2400m->tx_msg).
+ *
+ *     Assumes i2400m->tx_lock is taken, and we use that as a barrier
+ */
+static
+unsigned i2400m_tx_fits(struct i2400m *i2400m)
+{
+	struct i2400m_msg_hdr *msg_hdr = i2400m->tx_msg;
+	return le16_to_cpu(msg_hdr->num_pls) < I2400M_TX_PLD_MAX;
+
+}
+
+
+/*
+ * Start a new TX message header in the queue.
+ *
+ * Reserve memory from the base FIFO engine and then just initialize
+ * the message header.
+ *
+ * We allocate the biggest TX message header we might need (one that'd
+ * fit I2400M_TX_PLD_MAX payloads) -- when it is closed it will be
+ * 'ironed it out' and the unneeded parts removed.
+ *
+ * NOTE:
+ *
+ *     Assumes that the previous message is CLOSED (eg: either
+ *     there was none or 'i2400m_tx_close()' was called on it).
+ *
+ *     Assumes i2400m->tx_lock is taken, and we use that as a barrier
+ */
+static
+void i2400m_tx_new(struct i2400m *i2400m)
+{
+	struct device *dev = i2400m_dev(i2400m);
+	struct i2400m_msg_hdr *tx_msg;
+	BUG_ON(i2400m->tx_msg != NULL);
+try_head:
+	tx_msg = i2400m_tx_fifo_push(i2400m, I2400M_TX_PLD_SIZE, 0);
+	if (tx_msg == NULL)
+		goto out;
+	else if (tx_msg == TAIL_FULL) {
+		i2400m_tx_skip_tail(i2400m);
+		d_printf(2, dev, "new TX message: tail full, trying head\n");
+		goto try_head;
+	}
+	memset(tx_msg, 0, I2400M_TX_PLD_SIZE);
+	tx_msg->size = I2400M_TX_PLD_SIZE;
+out:
+	i2400m->tx_msg = tx_msg;
+	d_printf(2, dev, "new TX message: %p @%zu\n",
+		 tx_msg, (void *) tx_msg - i2400m->tx_buf);
+}
+
+
+/*
+ * Finalize the current TX message header
+ *
+ * Sets the message header to be at the proper location depending on
+ * how many descriptors we have (check documentation at the file's
+ * header for more info on that).
+ *
+ * Appends padding bytes to make sure the whole TX message (counting
+ * from the 'relocated' message header) is aligned to
+ * tx_block_size. We assume the _append() code has left enough space
+ * in the FIFO for that. If there are no payloads, just pass, as it
+ * won't be transferred.
+ *
+ * The amount of padding bytes depends on how many payloads are in the
+ * TX message, as the "msg header and payload descriptors" will be
+ * shifted up in the buffer.
+ */
+static
+void i2400m_tx_close(struct i2400m *i2400m)
+{
+	struct device *dev = i2400m_dev(i2400m);
+	struct i2400m_msg_hdr *tx_msg = i2400m->tx_msg;
+	struct i2400m_msg_hdr *tx_msg_moved;
+	size_t aligned_size, padding, hdr_size;
+	void *pad_buf;
+
+	if (tx_msg->size & I2400M_TX_SKIP)	/* a skipper? nothing to do */
+		goto out;
+
+	/* Relocate the message header
+	 *
+	 * Find the current header size, align it to 16 and if we need
+	 * to move it so the tail is next to the payloads, move it and
+	 * set the offset.
+	 *
+	 * If it moved, this header is good only for transmission; the
+	 * original one (it is kept if we moved) is still used to
+	 * figure out where the next TX message starts (and where the
+	 * offset to the moved header is).
+	 */
+	hdr_size = sizeof(*tx_msg)
+		+ le16_to_cpu(tx_msg->num_pls) * sizeof(tx_msg->pld[0]);
+	hdr_size = ALIGN(hdr_size, I2400M_PL_PAD);
+	tx_msg->offset = I2400M_TX_PLD_SIZE - hdr_size;
+	tx_msg_moved = (void *) tx_msg + tx_msg->offset;
+	memmove(tx_msg_moved, tx_msg, hdr_size);
+	tx_msg_moved->size -= tx_msg->offset;
+	/*
+	 * Now figure out how much we have to add to the (moved!)
+	 * message so the size is a multiple of i2400m->bus_tx_block_size.
+	 */
+	aligned_size = ALIGN(tx_msg_moved->size, i2400m->bus_tx_block_size);
+	padding = aligned_size - tx_msg_moved->size;
+	if (padding > 0) {
+		pad_buf = i2400m_tx_fifo_push(i2400m, padding, 0);
+		if (unlikely(WARN_ON(pad_buf == NULL
+				     || pad_buf == TAIL_FULL))) {
+			/* This should not happen -- append should verify
+			 * there is always space left at least to append
+			 * tx_block_size */
+			dev_err(dev,
+				"SW BUG! Possible data leakage from memory the "
+				"device should not read for padding - "
+				"size %lu aligned_size %zu tx_buf %p in "
+				"%zu out %zu\n",
+				(unsigned long) tx_msg_moved->size,
+				aligned_size, i2400m->tx_buf, i2400m->tx_in,
+				i2400m->tx_out);
+		} else
+			memset(pad_buf, 0xad, padding);
+	}
+	tx_msg_moved->padding = cpu_to_le16(padding);
+	tx_msg_moved->size += padding;
+	if (tx_msg != tx_msg_moved)
+		tx_msg->size += padding;
+out:
+	i2400m->tx_msg = NULL;
+}
+
+
+/**
+ * i2400m_tx - send the data in a buffer to the device
+ *
+ * @buf: pointer to the buffer to transmit
+ *
+ * @buf_len: buffer size
+ *
+ * @pl_type: type of the payload we are sending.
+ *
+ * Returns:
+ *     0 if ok, < 0 errno code on error (-ENOSPC, if there is no more
+ *     room for the message in the queue).
+ *
+ * Appends the buffer to the TX FIFO and notifies the bus-specific
+ * part of the driver that there is new data ready to transmit.
+ * Once this function returns, the buffer has been copied, so it can
+ * be reused.
+ *
+ * The steps followed to append are explained in detail in the file
+ * header.
+ *
+ * Whenever we write to a message, we increase msg->size, so it
+ * reflects exactly how big the message is. This is needed so that if
+ * we concatenate two messages before they can be sent, the code that
+ * sends the messages can find the boundaries (and it will replace the
+ * size with the real barker before sending).
+ *
+ * Note:
+ *
+ *     Cold and warm reset payloads need to be sent as a single
+ *     payload, so we handle that.
+ */
+int i2400m_tx(struct i2400m *i2400m, const void *buf, size_t buf_len,
+	      enum i2400m_pt pl_type)
+{
+	int result = -ENOSPC;
+	struct device *dev = i2400m_dev(i2400m);
+	unsigned long flags;
+	size_t padded_len;
+	void *ptr;
+	unsigned is_singleton = pl_type == I2400M_PT_RESET_WARM
+		|| pl_type == I2400M_PT_RESET_COLD;
+
+	d_fnstart(3, dev, "(i2400m %p skb %p [%zu bytes] pt %u)\n",
+		  i2400m, buf, buf_len, pl_type);
+	padded_len = ALIGN(buf_len, I2400M_PL_PAD);
+	d_printf(5, dev, "padded_len %zd buf_len %zd\n", padded_len, buf_len);
+	/* If there is no current TX message, create one; if the
+	 * current one is out of payload slots or we have a singleton,
+	 * close it and start a new one */
+	spin_lock_irqsave(&i2400m->tx_lock, flags);
+try_new:
+	if (unlikely(i2400m->tx_msg == NULL))
+		i2400m_tx_new(i2400m);
+	else if (unlikely(!i2400m_tx_fits(i2400m)
+			  || (is_singleton && i2400m->tx_msg->num_pls != 0))) {
+		d_printf(2, dev, "closing TX message (fits %u singleton "
+			 "%u num_pls %u)\n", i2400m_tx_fits(i2400m),
+			 is_singleton, i2400m->tx_msg->num_pls);
+		i2400m_tx_close(i2400m);
+		i2400m_tx_new(i2400m);
+	}
+	if (i2400m->tx_msg->size + padded_len > I2400M_TX_BUF_SIZE / 2) {
+		d_printf(2, dev, "TX: message too big, going new\n");
+		i2400m_tx_close(i2400m);
+		i2400m_tx_new(i2400m);
+	}
+	if (i2400m->tx_msg == NULL)
+		goto error_tx_new;
+	/* So we have a current message header; now append space for
+	 * the message -- if there is not enough, try the head */
+	ptr = i2400m_tx_fifo_push(i2400m, padded_len,
+				  i2400m->bus_tx_block_size);
+	if (ptr == TAIL_FULL) {	/* Tail is full, try head */
+		d_printf(2, dev, "pl append: tail full\n");
+		i2400m_tx_close(i2400m);
+		i2400m_tx_skip_tail(i2400m);
+		goto try_new;
+	} else if (ptr == NULL) {	/* All full */
+		result = -ENOSPC;
+		d_printf(2, dev, "pl append: all full\n");
+	} else {			/* Got space, copy it, set padding */
+		struct i2400m_msg_hdr *tx_msg = i2400m->tx_msg;
+		unsigned num_pls = le16_to_cpu(tx_msg->num_pls);
+		memcpy(ptr, buf, buf_len);
+		memset(ptr + buf_len, 0xad, padded_len - buf_len);
+		i2400m_pld_set(&tx_msg->pld[num_pls], buf_len, pl_type);
+		d_printf(3, dev, "pld 0x%08x (type 0x%1x len 0x%04zx\n",
+			 le32_to_cpu(tx_msg->pld[num_pls].val),
+			 pl_type, buf_len);
+		tx_msg->num_pls = le16_to_cpu(num_pls+1);
+		tx_msg->size += padded_len;
+		d_printf(2, dev, "TX: appended %zu b (up to %u b) pl #%u \n",
+			padded_len, tx_msg->size, num_pls+1);
+		d_printf(2, dev,
+			 "TX: appended hdr @%zu %zu b pl #%u @%zu %zu/%zu b\n",
+			 (void *)tx_msg - i2400m->tx_buf, (size_t)tx_msg->size,
+			 num_pls+1, ptr - i2400m->tx_buf, buf_len, padded_len);
+		result = 0;
+		if (is_singleton)
+			i2400m_tx_close(i2400m);
+	}
+error_tx_new:
+	spin_unlock_irqrestore(&i2400m->tx_lock, flags);
+	i2400m->bus_tx_kick(i2400m);	/* always kick, might free up space */
+	d_fnend(3, dev, "(i2400m %p skb %p [%zu bytes] pt %u) = %d\n",
+		i2400m, buf, buf_len, pl_type, result);
+	return result;
+}
+EXPORT_SYMBOL_GPL(i2400m_tx);
+
+
+/**
+ * i2400m_tx_msg_get - Get the first TX message in the FIFO to start sending it
+ *
+ * @i2400m: device descriptors
+ * @bus_size: where to place the size of the TX message
+ *
+ * Called by the bus-specific driver to get the first TX message at
+ * the FIF that is ready for transmission.
+ *
+ * It sets the state in @i2400m to indicate the bus-specific driver is
+ * transfering that message (i2400m->tx_msg_size).
+ *
+ * Once the transfer is completed, call i2400m_tx_msg_sent().
+ *
+ * Notes:
+ *
+ *     The size of the TX message to be transmitted might be smaller than
+ *     that of the TX message in the FIFO (in case the header was
+ *     shorter). Hence, we copy it in @bus_size, for the bus layer to
+ *     use. We keep the message's size in i2400m->tx_msg_size so that
+ *     when the bus later is done transferring we know how much to
+ *     advance the fifo.
+ *
+ *     We collect statistics here as all the data is available and we
+ *     assume it is going to work [see i2400m_tx_msg_sent()].
+ */
+struct i2400m_msg_hdr *i2400m_tx_msg_get(struct i2400m *i2400m,
+					 size_t *bus_size)
+{
+	struct device *dev = i2400m_dev(i2400m);
+	struct i2400m_msg_hdr *tx_msg, *tx_msg_moved;
+	unsigned long flags, pls;
+
+	d_fnstart(3, dev, "(i2400m %p bus_size %p)\n", i2400m, bus_size);
+	spin_lock_irqsave(&i2400m->tx_lock, flags);
+skip:
+	tx_msg_moved = NULL;
+	if (i2400m->tx_in == i2400m->tx_out) {	/* Empty FIFO? */
+		i2400m->tx_in = 0;
+		i2400m->tx_out = 0;
+		d_printf(2, dev, "TX: FIFO empty: resetting\n");
+		goto out_unlock;
+	}
+	tx_msg = i2400m->tx_buf + i2400m->tx_out % I2400M_TX_BUF_SIZE;
+	if (tx_msg->size & I2400M_TX_SKIP) {	/* skip? */
+		d_printf(2, dev, "TX: skip: msg @%zu (%zu b)\n",
+			 i2400m->tx_out % I2400M_TX_BUF_SIZE,
+			 (size_t) tx_msg->size & ~I2400M_TX_SKIP);
+		i2400m->tx_out += tx_msg->size & ~I2400M_TX_SKIP;
+		goto skip;
+	}
+
+	if (tx_msg->num_pls == 0) {		/* No payloads? */
+		if (tx_msg == i2400m->tx_msg) {	/* open, we are done */
+			d_printf(2, dev,
+				 "TX: FIFO empty: open msg w/o payloads @%zu\n",
+				 (void *) tx_msg - i2400m->tx_buf);
+			tx_msg = NULL;
+			goto out_unlock;
+		} else {			/* closed, skip it */
+			d_printf(2, dev,
+				 "TX: skip msg w/o payloads @%zu (%zu b)\n",
+				 (void *) tx_msg - i2400m->tx_buf,
+				 (size_t) tx_msg->size);
+			i2400m->tx_out += tx_msg->size & ~I2400M_TX_SKIP;
+			goto skip;
+		}
+	}
+	if (tx_msg == i2400m->tx_msg)		/* open msg? */
+		i2400m_tx_close(i2400m);
+
+	/* Now we have a valid TX message (with payloads) to TX */
+	tx_msg_moved = (void *) tx_msg + tx_msg->offset;
+	i2400m->tx_msg_size = tx_msg->size;
+	*bus_size = tx_msg_moved->size;
+	d_printf(2, dev, "TX: pid %d msg hdr at @%zu offset +@%zu "
+		 "size %zu bus_size %zu\n",
+		 current->pid, (void *) tx_msg - i2400m->tx_buf,
+		 (size_t) tx_msg->offset, (size_t) tx_msg->size,
+		 (size_t) tx_msg_moved->size);
+	tx_msg_moved->barker = le32_to_cpu(I2400M_H2D_PREVIEW_BARKER);
+	tx_msg_moved->sequence = le32_to_cpu(i2400m->tx_sequence++);
+
+	pls = le32_to_cpu(tx_msg_moved->num_pls);
+	i2400m->tx_pl_num += pls;		/* Update stats */
+	if (pls > i2400m->tx_pl_max)
+		i2400m->tx_pl_max = pls;
+	if (pls < i2400m->tx_pl_min)
+		i2400m->tx_pl_min = pls;
+	i2400m->tx_num++;
+	i2400m->tx_size_acc += *bus_size;
+	if (*bus_size < i2400m->tx_size_min)
+		i2400m->tx_size_min = *bus_size;
+	if (*bus_size > i2400m->tx_size_max)
+		i2400m->tx_size_max = *bus_size;
+out_unlock:
+	spin_unlock_irqrestore(&i2400m->tx_lock, flags);
+	d_fnstart(3, dev, "(i2400m %p bus_size %p [%zu]) = %p\n",
+		  i2400m, bus_size, *bus_size, tx_msg_moved);
+	return tx_msg_moved;
+}
+EXPORT_SYMBOL_GPL(i2400m_tx_msg_get);
+
+
+/**
+ * i2400m_tx_msg_sent - indicate the transmission of a TX message
+ *
+ * @i2400m: device descriptor
+ *
+ * Called by the bus-specific driver when a message has been sent;
+ * this pops it from the FIFO; and as there is space, start the queue
+ * in case it was stopped.
+ *
+ * Should be called even if the message send failed and we are
+ * dropping this TX message.
+ */
+void i2400m_tx_msg_sent(struct i2400m *i2400m)
+{
+	unsigned n;
+	unsigned long flags;
+	struct device *dev = i2400m_dev(i2400m);
+
+	d_fnstart(3, dev, "(i2400m %p)\n", i2400m);
+	spin_lock_irqsave(&i2400m->tx_lock, flags);
+	i2400m->tx_out += i2400m->tx_msg_size;
+	d_printf(2, dev, "TX: sent %zu b\n", (size_t) i2400m->tx_msg_size);
+	i2400m->tx_msg_size = 0;
+	BUG_ON(i2400m->tx_out > i2400m->tx_in);
+	/* level them FIFO markers off */
+	n = i2400m->tx_out / I2400M_TX_BUF_SIZE;
+	i2400m->tx_out %= I2400M_TX_BUF_SIZE;
+	i2400m->tx_in -= n * I2400M_TX_BUF_SIZE;
+	netif_start_queue(i2400m->wimax_dev.net_dev);
+	spin_unlock_irqrestore(&i2400m->tx_lock, flags);
+	d_fnend(3, dev, "(i2400m %p) = void\n", i2400m);
+}
+EXPORT_SYMBOL_GPL(i2400m_tx_msg_sent);
+
+
+/**
+ * i2400m_tx_setup - Initialize the TX queue and infrastructure
+ *
+ * Make sure we reset the TX sequence to zero, as when this function
+ * is called, the firmware has been just restarted.
+ */
+int i2400m_tx_setup(struct i2400m *i2400m)
+{
+	int result;
+
+	/* Do this here only once -- can't do on
+	 * i2400m_hard_start_xmit() as we'll cause race conditions if
+	 * the WS was scheduled on another CPU */
+	INIT_WORK(&i2400m->wake_tx_ws, i2400m_wake_tx_work);
+
+	i2400m->tx_sequence = 0;
+	i2400m->tx_buf = kmalloc(I2400M_TX_BUF_SIZE, GFP_KERNEL);
+	if (i2400m->tx_buf == NULL)
+		result = -ENOMEM;
+	else
+		result = 0;
+	/* Huh? the bus layer has to define this... */
+	BUG_ON(i2400m->bus_tx_block_size == 0);
+	return result;
+
+}
+
+
+/**
+ * i2400m_tx_release - Tear down the TX queue and infrastructure
+ */
+void i2400m_tx_release(struct i2400m *i2400m)
+{
+	kfree(i2400m->tx_buf);
+}