netdev.c

Go to the documentation of this file.

/*
 * Intel Wireless WiMAX Connection 2400m
 * Glue with the networking stack
 *
 *
 * Copyright (C) 2007 Intel Corporation <[email protected]>
 * Yanir Lubetkin <[email protected]>
 * Inaky Perez-Gonzalez <[email protected]>
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License version
 * 2 as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
 * 02110-1301, USA.
 *
 *
 * This implements an ethernet device for the i2400m.
 *
 * We fake being an ethernet device to simplify the support from user
 * space and from the other side. The world is (sadly) configured to
 * take in only Ethernet devices...
 *
 * Because of this, when using firmwares <= v1.3, there is an
 * copy-each-rxed-packet overhead on the RX path. Each IP packet has
 * to be reallocated to add an ethernet header (as there is no space
 * in what we get from the device). This is a known drawback and
 * firmwares >= 1.4 add header space that can be used to insert the
 * ethernet header without having to reallocate and copy.
 *
 * TX error handling is tricky; because we have to FIFO/queue the
 * buffers for transmission (as the hardware likes it aggregated), we
 * just give the skb to the TX subsystem and by the time it is
 * transmitted, we have long forgotten about it. So we just don't care
 * too much about it.
 *
 * Note that when the device is in idle mode with the basestation, we
 * need to negotiate coming back up online. That involves negotiation
 * and possible user space interaction. Thus, we defer to a workqueue
 * to do all that. By default, we only queue a single packet and drop
 * the rest, as potentially the time to go back from idle to normal is
 * long.
 *
 * ROADMAP
 *
 * i2400m_open         Called on ifconfig up
 * i2400m_stop         Called on ifconfig down
 *
 * i2400m_hard_start_xmit Called by the network stack to send a packet
 *   i2400m_net_wake_tx   Wake up device from basestation-IDLE & TX
 *     i2400m_wake_tx_work
 *       i2400m_cmd_exit_idle
 *       i2400m_tx
 *   i2400m_net_tx        TX a data frame
 *     i2400m_tx
 *
 * i2400m_change_mtu      Called on ifconfig mtu XXX
 *
 * i2400m_tx_timeout      Called when the device times out
 *
 * i2400m_net_rx          Called by the RX code when a data frame is
 *                        available (firmware <= 1.3)
 * i2400m_net_erx         Called by the RX code when a data frame is
 *                        available (firmware >= 1.4).
 * i2400m_netdev_setup    Called to setup all the netdev stuff from
 *                        alloc_netdev.
 */
#include <linux/if_arp.h>
#include <linux/slab.h>
#include <linux/netdevice.h>
#include <linux/ethtool.h>
#include <linux/export.h>
#include "i2400m.h"


#define D_SUBMODULE netdev
#include "debug-levels.h"

enum {
/* netdev interface */
    /* 20 secs? yep, this is the maximum timeout that the device
     * might take to get out of IDLE / negotiate it with the base
     * station. We add 1sec for good measure. */
    I2400M_TX_TIMEOUT = 21 * HZ,
    /*
     * Experimentation has determined that, 20 to be a good value
     * for minimizing the jitter in the throughput.
     */
    I2400M_TX_QLEN = 20,
};


static
int i2400m_open(struct net_device *net_dev)
{
    int result;
    struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
    struct device *dev = i2400m_dev(i2400m);

    d_fnstart(3, dev, "(net_dev %p [i2400m %p])\n", net_dev, i2400m);
    /* Make sure we wait until init is complete... */
    mutex_lock(&i2400m->init_mutex);
    if (i2400m->updown)
        result = 0;
    else
        result = -EBUSY;
    mutex_unlock(&i2400m->init_mutex);
    d_fnend(3, dev, "(net_dev %p [i2400m %p]) = %d\n",
        net_dev, i2400m, result);
    return result;
}


static
int i2400m_stop(struct net_device *net_dev)
{
    struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
    struct device *dev = i2400m_dev(i2400m);

    d_fnstart(3, dev, "(net_dev %p [i2400m %p])\n", net_dev, i2400m);
    i2400m_net_wake_stop(i2400m);
    d_fnend(3, dev, "(net_dev %p [i2400m %p]) = 0\n", net_dev, i2400m);
    return 0;
}


/*
 * Wake up the device and transmit a held SKB, then restart the net queue
 *
 * When the device goes into basestation-idle mode, we need to tell it
 * to exit that mode; it will negotiate with the base station, user
 * space may have to intervene to rehandshake crypto and then tell us
 * when it is ready to transmit the packet we have "queued". Still we
 * need to give it sometime after it reports being ok.
 *
 * On error, there is not much we can do. If the error was on TX, we
 * still wake the queue up to see if the next packet will be luckier.
 *
 * If _cmd_exit_idle() fails...well, it could be many things; most
 * commonly it is that something else took the device out of IDLE mode
 * (for example, the base station). In that case we get an -EILSEQ and
 * we are just going to ignore that one. If the device is back to
 * connected, then fine -- if it is someother state, the packet will
 * be dropped anyway.
 */
void i2400m_wake_tx_work(struct work_struct *ws)
{
    int result;
    struct i2400m *i2400m = container_of(ws, struct i2400m, wake_tx_ws);
    struct net_device *net_dev = i2400m->wimax_dev.net_dev;
    struct device *dev = i2400m_dev(i2400m);
    struct sk_buff *skb = i2400m->wake_tx_skb;
    unsigned long flags;

    spin_lock_irqsave(&i2400m->tx_lock, flags);
    skb = i2400m->wake_tx_skb;
    i2400m->wake_tx_skb = NULL;
    spin_unlock_irqrestore(&i2400m->tx_lock, flags);

    d_fnstart(3, dev, "(ws %p i2400m %p skb %p)\n", ws, i2400m, skb);
    result = -EINVAL;
    if (skb == NULL) {
        dev_err(dev, "WAKE&TX: skb disappeared!\n");
        goto out_put;
    }
    /* If we have, somehow, lost the connection after this was
     * queued, don't do anything; this might be the device got
     * reset or just disconnected. */
    if (unlikely(!netif_carrier_ok(net_dev)))
        goto out_kfree;
    result = i2400m_cmd_exit_idle(i2400m);
    if (result == -EILSEQ)
        result = 0;
    if (result < 0) {
        dev_err(dev, "WAKE&TX: device didn't get out of idle: "
            "%d - resetting\n", result);
        i2400m_reset(i2400m, I2400M_RT_BUS);
        goto error;
    }
    result = wait_event_timeout(i2400m->state_wq,
                    i2400m->state != I2400M_SS_IDLE,
                    net_dev->watchdog_timeo - HZ/2);
    if (result == 0)
        result = -ETIMEDOUT;
    if (result < 0) {
        dev_err(dev, "WAKE&TX: error waiting for device to exit IDLE: "
            "%d - resetting\n", result);
        i2400m_reset(i2400m, I2400M_RT_BUS);
        goto error;
    }
    msleep(20); /* device still needs some time or it drops it */
    result = i2400m_tx(i2400m, skb->data, skb->len, I2400M_PT_DATA);
error:
    netif_wake_queue(net_dev);
out_kfree:
    kfree_skb(skb); /* refcount transferred by _hard_start_xmit() */
out_put:
    i2400m_put(i2400m);
    d_fnend(3, dev, "(ws %p i2400m %p skb %p) = void [%d]\n",
        ws, i2400m, skb, result);
}


/*
 * Prepare the data payload TX header
 *
 * The i2400m expects a 4 byte header in front of a data packet.
 *
 * Because we pretend to be an ethernet device, this packet comes with
 * an ethernet header. Pull it and push our header.
 */
static
void i2400m_tx_prep_header(struct sk_buff *skb)
{
    struct i2400m_pl_data_hdr *pl_hdr;
    skb_pull(skb, ETH_HLEN);
    pl_hdr = (struct i2400m_pl_data_hdr *) skb_push(skb, sizeof(*pl_hdr));
    pl_hdr->reserved = 0;
}



/*
 * Cleanup resources acquired during i2400m_net_wake_tx()
 *
 * This is called by __i2400m_dev_stop and means we have to make sure
 * the workqueue is flushed from any pending work.
 */
void i2400m_net_wake_stop(struct i2400m *i2400m)
{
    struct device *dev = i2400m_dev(i2400m);

    d_fnstart(3, dev, "(i2400m %p)\n", i2400m);
    /* See i2400m_hard_start_xmit(), references are taken there
     * and here we release them if the work was still
     * pending. Note we can't differentiate work not pending vs
     * never scheduled, so the NULL check does that. */
    if (cancel_work_sync(&i2400m->wake_tx_ws) == 0
        && i2400m->wake_tx_skb != NULL) {
        unsigned long flags;
        struct sk_buff *wake_tx_skb;
        spin_lock_irqsave(&i2400m->tx_lock, flags);
        wake_tx_skb = i2400m->wake_tx_skb;  /* compat help */
        i2400m->wake_tx_skb = NULL; /* compat help */
        spin_unlock_irqrestore(&i2400m->tx_lock, flags);
        i2400m_put(i2400m);
        kfree_skb(wake_tx_skb);
    }
    d_fnend(3, dev, "(i2400m %p) = void\n", i2400m);
}


/*
 * TX an skb to an idle device
 *
 * When the device is in basestation-idle mode, we need to wake it up
 * and then TX. So we queue a work_struct for doing so.
 *
 * We need to get an extra ref for the skb (so it is not dropped), as
 * well as be careful not to queue more than one request (won't help
 * at all). If more than one request comes or there are errors, we
 * just drop the packets (see i2400m_hard_start_xmit()).
 */
static
int i2400m_net_wake_tx(struct i2400m *i2400m, struct net_device *net_dev,
               struct sk_buff *skb)
{
    int result;
    struct device *dev = i2400m_dev(i2400m);
    unsigned long flags;

    d_fnstart(3, dev, "(skb %p net_dev %p)\n", skb, net_dev);
    if (net_ratelimit()) {
        d_printf(3, dev, "WAKE&NETTX: "
             "skb %p sending %d bytes to radio\n",
             skb, skb->len);
        d_dump(4, dev, skb->data, skb->len);
    }
    /* We hold a ref count for i2400m and skb, so when
     * stopping() the device, we need to cancel that work
     * and if pending, release those resources. */
    result = 0;
    spin_lock_irqsave(&i2400m->tx_lock, flags);
    if (!work_pending(&i2400m->wake_tx_ws)) {
        netif_stop_queue(net_dev);
        i2400m_get(i2400m);
        i2400m->wake_tx_skb = skb_get(skb); /* transfer ref count */
        i2400m_tx_prep_header(skb);
        result = schedule_work(&i2400m->wake_tx_ws);
        WARN_ON(result == 0);
    }
    spin_unlock_irqrestore(&i2400m->tx_lock, flags);
    if (result == 0) {
        /* Yes, this happens even if we stopped the
         * queue -- blame the queue disciplines that
         * queue without looking -- I guess there is a reason
         * for that. */
        if (net_ratelimit())
            d_printf(1, dev, "NETTX: device exiting idle, "
                 "dropping skb %p, queue running %d\n",
                 skb, netif_queue_stopped(net_dev));
        result = -EBUSY;
    }
    d_fnend(3, dev, "(skb %p net_dev %p) = %d\n", skb, net_dev, result);
    return result;
}


/*
 * Transmit a packet to the base station on behalf of the network stack.
 *
 * Returns: 0 if ok, < 0 errno code on error.
 *
 * We need to pull the ethernet header and add the hardware header,
 * which is currently set to all zeroes and reserved.
 */
static
int i2400m_net_tx(struct i2400m *i2400m, struct net_device *net_dev,
          struct sk_buff *skb)
{
    int result;
    struct device *dev = i2400m_dev(i2400m);

    d_fnstart(3, dev, "(i2400m %p net_dev %p skb %p)\n",
          i2400m, net_dev, skb);
    /* FIXME: check eth hdr, only IPv4 is routed by the device as of now */
    net_dev->trans_start = jiffies;
    i2400m_tx_prep_header(skb);
    d_printf(3, dev, "NETTX: skb %p sending %d bytes to radio\n",
         skb, skb->len);
    d_dump(4, dev, skb->data, skb->len);
    result = i2400m_tx(i2400m, skb->data, skb->len, I2400M_PT_DATA);
    d_fnend(3, dev, "(i2400m %p net_dev %p skb %p) = %d\n",
        i2400m, net_dev, skb, result);
    return result;
}


/*
 * Transmit a packet to the base station on behalf of the network stack
 *
 *
 * Returns: NETDEV_TX_OK (always, even in case of error)
 *
 * In case of error, we just drop it. Reasons:
 *
 *  - we add a hw header to each skb, and if the network stack
 *    retries, we have no way to know if that skb has it or not.
 *
 *  - network protocols have their own drop-recovery mechanisms
 *
 *  - there is not much else we can do
 *
 * If the device is idle, we need to wake it up; that is an operation
 * that will sleep. See i2400m_net_wake_tx() for details.
 */
static
netdev_tx_t i2400m_hard_start_xmit(struct sk_buff *skb,
                     struct net_device *net_dev)
{
    struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
    struct device *dev = i2400m_dev(i2400m);
    int result = -1;

    d_fnstart(3, dev, "(skb %p net_dev %p)\n", skb, net_dev);

    if (skb_header_cloned(skb) && 
        pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
        goto drop;

    if (i2400m->state == I2400M_SS_IDLE)
        result = i2400m_net_wake_tx(i2400m, net_dev, skb);
    else
        result = i2400m_net_tx(i2400m, net_dev, skb);
    if (result <  0) {
drop:
        net_dev->stats.tx_dropped++;
    } else {
        net_dev->stats.tx_packets++;
        net_dev->stats.tx_bytes += skb->len;
    }
    dev_kfree_skb(skb);
    d_fnend(3, dev, "(skb %p net_dev %p) = %d\n", skb, net_dev, result);
    return NETDEV_TX_OK;
}


static
int i2400m_change_mtu(struct net_device *net_dev, int new_mtu)
{
    int result;
    struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
    struct device *dev = i2400m_dev(i2400m);

    if (new_mtu >= I2400M_MAX_MTU) {
        dev_err(dev, "Cannot change MTU to %d (max is %d)\n",
            new_mtu, I2400M_MAX_MTU);
        result = -EINVAL;
    } else {
        net_dev->mtu = new_mtu;
        result = 0;
    }
    return result;
}


static
void i2400m_tx_timeout(struct net_device *net_dev)
{
    /*
     * We might want to kick the device
     *
     * There is not much we can do though, as the device requires
     * that we send the data aggregated. By the time we receive
     * this, there might be data pending to be sent or not...
     */
    net_dev->stats.tx_errors++;
}


/*
 * Create a fake ethernet header
 *
 * For emulating an ethernet device, every received IP header has to
 * be prefixed with an ethernet header. Fake it with the given
 * protocol.
 */
static
void i2400m_rx_fake_eth_header(struct net_device *net_dev,
                   void *_eth_hdr, __be16 protocol)
{
    struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
    struct ethhdr *eth_hdr = _eth_hdr;

    memcpy(eth_hdr->h_dest, net_dev->dev_addr, sizeof(eth_hdr->h_dest));
    memcpy(eth_hdr->h_source, i2400m->src_mac_addr,
           sizeof(eth_hdr->h_source));
    eth_hdr->h_proto = protocol;
}


/*
 * i2400m_net_rx - pass a network packet to the stack
 *
 * @i2400m: device instance
 * @skb_rx: the skb where the buffer pointed to by @buf is
 * @i: 1 if payload is the only one
 * @buf: pointer to the buffer containing the data
 * @len: buffer's length
 *
 * This is only used now for the v1.3 firmware. It will be deprecated
 * in >= 2.6.31.
 *
 * Note that due to firmware limitations, we don't have space to add
 * an ethernet header, so we need to copy each packet. Firmware
 * versions >= v1.4 fix this [see i2400m_net_erx()].
 *
 * We just clone the skb and set it up so that it's skb->data pointer
 * points to "buf" and it's length.
 *
 * Note that if the payload is the last (or the only one) in a
 * multi-payload message, we don't clone the SKB but just reuse it.
 *
 * This function is normally run from a thread context. However, we
 * still use netif_rx() instead of netif_receive_skb() as was
 * recommended in the mailing list. Reason is in some stress tests
 * when sending/receiving a lot of data we seem to hit a softlock in
 * the kernel's TCP implementation [aroudn tcp_delay_timer()]. Using
 * netif_rx() took care of the issue.
 *
 * This is, of course, still open to do more research on why running
 * with netif_receive_skb() hits this softlock. FIXME.
 *
 * FIXME: currently we don't do any efforts at distinguishing if what
 * we got was an IPv4 or IPv6 header, to setup the protocol field
 * correctly.
 */
void i2400m_net_rx(struct i2400m *i2400m, struct sk_buff *skb_rx,
           unsigned i, const void *buf, int buf_len)
{
    struct net_device *net_dev = i2400m->wimax_dev.net_dev;
    struct device *dev = i2400m_dev(i2400m);
    struct sk_buff *skb;

    d_fnstart(2, dev, "(i2400m %p buf %p buf_len %d)\n",
          i2400m, buf, buf_len);
    if (i) {
        skb = skb_get(skb_rx);
        d_printf(2, dev, "RX: reusing first payload skb %p\n", skb);
        skb_pull(skb, buf - (void *) skb->data);
        skb_trim(skb, (void *) skb_end_pointer(skb) - buf);
    } else {
        /* Yes, this is bad -- a lot of overhead -- see
         * comments at the top of the file */
        skb = __netdev_alloc_skb(net_dev, buf_len, GFP_KERNEL);
        if (skb == NULL) {
            dev_err(dev, "NETRX: no memory to realloc skb\n");
            net_dev->stats.rx_dropped++;
            goto error_skb_realloc;
        }
        memcpy(skb_put(skb, buf_len), buf, buf_len);
    }
    i2400m_rx_fake_eth_header(i2400m->wimax_dev.net_dev,
                  skb->data - ETH_HLEN,
                  cpu_to_be16(ETH_P_IP));
    skb_set_mac_header(skb, -ETH_HLEN);
    skb->dev = i2400m->wimax_dev.net_dev;
    skb->protocol = htons(ETH_P_IP);
    net_dev->stats.rx_packets++;
    net_dev->stats.rx_bytes += buf_len;
    d_printf(3, dev, "NETRX: receiving %d bytes to network stack\n",
        buf_len);
    d_dump(4, dev, buf, buf_len);
    netif_rx_ni(skb);   /* see notes in function header */
error_skb_realloc:
    d_fnend(2, dev, "(i2400m %p buf %p buf_len %d) = void\n",
        i2400m, buf, buf_len);
}


/*
 * i2400m_net_erx - pass a network packet to the stack (extended version)
 *
 * @i2400m: device descriptor
 * @skb: the skb where the packet is - the skb should be set to point
 *     at the IP packet; this function will add ethernet headers if
 *     needed.
 * @cs: packet type
 *
 * This is only used now for firmware >= v1.4. Note it is quite
 * similar to i2400m_net_rx() (used only for v1.3 firmware).
 *
 * This function is normally run from a thread context. However, we
 * still use netif_rx() instead of netif_receive_skb() as was
 * recommended in the mailing list. Reason is in some stress tests
 * when sending/receiving a lot of data we seem to hit a softlock in
 * the kernel's TCP implementation [aroudn tcp_delay_timer()]. Using
 * netif_rx() took care of the issue.
 *
 * This is, of course, still open to do more research on why running
 * with netif_receive_skb() hits this softlock. FIXME.
 */
void i2400m_net_erx(struct i2400m *i2400m, struct sk_buff *skb,
            enum i2400m_cs cs)
{
    struct net_device *net_dev = i2400m->wimax_dev.net_dev;
    struct device *dev = i2400m_dev(i2400m);
    int protocol;

    d_fnstart(2, dev, "(i2400m %p skb %p [%u] cs %d)\n",
          i2400m, skb, skb->len, cs);
    switch(cs) {
    case I2400M_CS_IPV4_0:
    case I2400M_CS_IPV4:
        protocol = ETH_P_IP;
        i2400m_rx_fake_eth_header(i2400m->wimax_dev.net_dev,
                      skb->data - ETH_HLEN,
                      cpu_to_be16(ETH_P_IP));
        skb_set_mac_header(skb, -ETH_HLEN);
        skb->dev = i2400m->wimax_dev.net_dev;
        skb->protocol = htons(ETH_P_IP);
        net_dev->stats.rx_packets++;
        net_dev->stats.rx_bytes += skb->len;
        break;
    default:
        dev_err(dev, "ERX: BUG? CS type %u unsupported\n", cs);
        goto error;

    }
    d_printf(3, dev, "ERX: receiving %d bytes to the network stack\n",
         skb->len);
    d_dump(4, dev, skb->data, skb->len);
    netif_rx_ni(skb);   /* see notes in function header */
error:
    d_fnend(2, dev, "(i2400m %p skb %p [%u] cs %d) = void\n",
        i2400m, skb, skb->len, cs);
}

static const struct net_device_ops i2400m_netdev_ops = {
    .ndo_open = i2400m_open,
    .ndo_stop = i2400m_stop,
    .ndo_start_xmit = i2400m_hard_start_xmit,
    .ndo_tx_timeout = i2400m_tx_timeout,
    .ndo_change_mtu = i2400m_change_mtu,
};

static void i2400m_get_drvinfo(struct net_device *net_dev,
                   struct ethtool_drvinfo *info)
{
    struct i2400m *i2400m = net_dev_to_i2400m(net_dev);

    strncpy(info->driver, KBUILD_MODNAME, sizeof(info->driver) - 1);
    strncpy(info->fw_version,
            i2400m->fw_name ? : "", sizeof(info->fw_version) - 1);
    if (net_dev->dev.parent)
        strncpy(info->bus_info, dev_name(net_dev->dev.parent),
            sizeof(info->bus_info) - 1);
}

static const struct ethtool_ops i2400m_ethtool_ops = {
    .get_drvinfo = i2400m_get_drvinfo,
    .get_link = ethtool_op_get_link,
};

void i2400m_netdev_setup(struct net_device *net_dev)
{
    d_fnstart(3, NULL, "(net_dev %p)\n", net_dev);
    ether_setup(net_dev);
    net_dev->mtu = I2400M_MAX_MTU;
    net_dev->tx_queue_len = I2400M_TX_QLEN;
    net_dev->features =
          NETIF_F_VLAN_CHALLENGED
        | NETIF_F_HIGHDMA;
    net_dev->flags =
        IFF_NOARP       /* i2400m is apure IP device */
        & (~IFF_BROADCAST   /* i2400m is P2P */
           & ~IFF_MULTICAST);
    net_dev->watchdog_timeo = I2400M_TX_TIMEOUT;
    net_dev->netdev_ops = &i2400m_netdev_ops;
    net_dev->ethtool_ops = &i2400m_ethtool_ops;
    d_fnend(3, NULL, "(net_dev %p) = void\n", net_dev);
}
EXPORT_SYMBOL_GPL(i2400m_netdev_setup);