2 * Intel Wireless WiMAX Connection 2400m
3 * Glue with the networking stack
6 * Copyright (C) 2007 Intel Corporation <linux-wimax@intel.com>
7 * Yanir Lubetkin <yanirx.lubetkin@intel.com>
8 * Inaky Perez-Gonzalez <inaky.perez-gonzalez@intel.com>
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public License version
12 * 2 as published by the Free Software Foundation.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
25 * This implements an ethernet device for the i2400m.
27 * We fake being an ethernet device to simplify the support from user
28 * space and from the other side. The world is (sadly) configured to
29 * take in only Ethernet devices...
31 * Because of this, when using firmwares <= v1.3, there is an
32 * copy-each-rxed-packet overhead on the RX path. Each IP packet has
33 * to be reallocated to add an ethernet header (as there is no space
34 * in what we get from the device). This is a known drawback and
35 * firmwares >= 1.4 add header space that can be used to insert the
36 * ethernet header without having to reallocate and copy.
38 * TX error handling is tricky; because we have to FIFO/queue the
39 * buffers for transmission (as the hardware likes it aggregated), we
40 * just give the skb to the TX subsystem and by the time it is
41 * transmitted, we have long forgotten about it. So we just don't care
44 * Note that when the device is in idle mode with the basestation, we
45 * need to negotiate coming back up online. That involves negotiation
46 * and possible user space interaction. Thus, we defer to a workqueue
47 * to do all that. By default, we only queue a single packet and drop
48 * the rest, as potentially the time to go back from idle to normal is
53 * i2400m_open Called on ifconfig up
54 * i2400m_stop Called on ifconfig down
56 * i2400m_hard_start_xmit Called by the network stack to send a packet
57 * i2400m_net_wake_tx Wake up device from basestation-IDLE & TX
59 * i2400m_cmd_exit_idle
61 * i2400m_net_tx TX a data frame
64 * i2400m_change_mtu Called on ifconfig mtu XXX
66 * i2400m_tx_timeout Called when the device times out
68 * i2400m_net_rx Called by the RX code when a data frame is
69 * available (firmware <= 1.3)
70 * i2400m_net_erx Called by the RX code when a data frame is
71 * available (firmware >= 1.4).
72 * i2400m_netdev_setup Called to setup all the netdev stuff from
75 #include <linux/if_arp.h>
76 #include <linux/slab.h>
77 #include <linux/netdevice.h>
78 #include <linux/ethtool.h>
82 #define D_SUBMODULE netdev
83 #include "debug-levels.h"
86 /* netdev interface */
87 /* 20 secs? yep, this is the maximum timeout that the device
88 * might take to get out of IDLE / negotiate it with the base
89 * station. We add 1sec for good measure. */
90 I2400M_TX_TIMEOUT = 21 * HZ,
96 int i2400m_open(struct net_device *net_dev)
99 struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
100 struct device *dev = i2400m_dev(i2400m);
102 d_fnstart(3, dev, "(net_dev %p [i2400m %p])\n", net_dev, i2400m);
103 /* Make sure we wait until init is complete... */
104 mutex_lock(&i2400m->init_mutex);
109 mutex_unlock(&i2400m->init_mutex);
110 d_fnend(3, dev, "(net_dev %p [i2400m %p]) = %d\n",
111 net_dev, i2400m, result);
117 int i2400m_stop(struct net_device *net_dev)
119 struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
120 struct device *dev = i2400m_dev(i2400m);
122 d_fnstart(3, dev, "(net_dev %p [i2400m %p])\n", net_dev, i2400m);
123 i2400m_net_wake_stop(i2400m);
124 d_fnend(3, dev, "(net_dev %p [i2400m %p]) = 0\n", net_dev, i2400m);
130 * Wake up the device and transmit a held SKB, then restart the net queue
132 * When the device goes into basestation-idle mode, we need to tell it
133 * to exit that mode; it will negotiate with the base station, user
134 * space may have to intervene to rehandshake crypto and then tell us
135 * when it is ready to transmit the packet we have "queued". Still we
136 * need to give it sometime after it reports being ok.
138 * On error, there is not much we can do. If the error was on TX, we
139 * still wake the queue up to see if the next packet will be luckier.
141 * If _cmd_exit_idle() fails...well, it could be many things; most
142 * commonly it is that something else took the device out of IDLE mode
143 * (for example, the base station). In that case we get an -EILSEQ and
144 * we are just going to ignore that one. If the device is back to
145 * connected, then fine -- if it is someother state, the packet will
148 void i2400m_wake_tx_work(struct work_struct *ws)
151 struct i2400m *i2400m = container_of(ws, struct i2400m, wake_tx_ws);
152 struct net_device *net_dev = i2400m->wimax_dev.net_dev;
153 struct device *dev = i2400m_dev(i2400m);
154 struct sk_buff *skb = i2400m->wake_tx_skb;
157 spin_lock_irqsave(&i2400m->tx_lock, flags);
158 skb = i2400m->wake_tx_skb;
159 i2400m->wake_tx_skb = NULL;
160 spin_unlock_irqrestore(&i2400m->tx_lock, flags);
162 d_fnstart(3, dev, "(ws %p i2400m %p skb %p)\n", ws, i2400m, skb);
165 dev_err(dev, "WAKE&TX: skb dissapeared!\n");
168 /* If we have, somehow, lost the connection after this was
169 * queued, don't do anything; this might be the device got
170 * reset or just disconnected. */
171 if (unlikely(!netif_carrier_ok(net_dev)))
173 result = i2400m_cmd_exit_idle(i2400m);
174 if (result == -EILSEQ)
177 dev_err(dev, "WAKE&TX: device didn't get out of idle: "
178 "%d - resetting\n", result);
179 i2400m_reset(i2400m, I2400M_RT_BUS);
182 result = wait_event_timeout(i2400m->state_wq,
183 i2400m->state != I2400M_SS_IDLE,
184 net_dev->watchdog_timeo - HZ/2);
188 dev_err(dev, "WAKE&TX: error waiting for device to exit IDLE: "
189 "%d - resetting\n", result);
190 i2400m_reset(i2400m, I2400M_RT_BUS);
193 msleep(20); /* device still needs some time or it drops it */
194 result = i2400m_tx(i2400m, skb->data, skb->len, I2400M_PT_DATA);
196 netif_wake_queue(net_dev);
198 kfree_skb(skb); /* refcount transferred by _hard_start_xmit() */
201 d_fnend(3, dev, "(ws %p i2400m %p skb %p) = void [%d]\n",
202 ws, i2400m, skb, result);
207 * Prepare the data payload TX header
209 * The i2400m expects a 4 byte header in front of a data packet.
211 * Because we pretend to be an ethernet device, this packet comes with
212 * an ethernet header. Pull it and push our header.
215 void i2400m_tx_prep_header(struct sk_buff *skb)
217 struct i2400m_pl_data_hdr *pl_hdr;
218 skb_pull(skb, ETH_HLEN);
219 pl_hdr = (struct i2400m_pl_data_hdr *) skb_push(skb, sizeof(*pl_hdr));
220 pl_hdr->reserved = 0;
226 * Cleanup resources acquired during i2400m_net_wake_tx()
228 * This is called by __i2400m_dev_stop and means we have to make sure
229 * the workqueue is flushed from any pending work.
231 void i2400m_net_wake_stop(struct i2400m *i2400m)
233 struct device *dev = i2400m_dev(i2400m);
235 d_fnstart(3, dev, "(i2400m %p)\n", i2400m);
236 /* See i2400m_hard_start_xmit(), references are taken there
237 * and here we release them if the work was still
238 * pending. Note we can't differentiate work not pending vs
239 * never scheduled, so the NULL check does that. */
240 if (cancel_work_sync(&i2400m->wake_tx_ws) == 0
241 && i2400m->wake_tx_skb != NULL) {
243 struct sk_buff *wake_tx_skb;
244 spin_lock_irqsave(&i2400m->tx_lock, flags);
245 wake_tx_skb = i2400m->wake_tx_skb; /* compat help */
246 i2400m->wake_tx_skb = NULL; /* compat help */
247 spin_unlock_irqrestore(&i2400m->tx_lock, flags);
249 kfree_skb(wake_tx_skb);
251 d_fnend(3, dev, "(i2400m %p) = void\n", i2400m);
257 * TX an skb to an idle device
259 * When the device is in basestation-idle mode, we need to wake it up
260 * and then TX. So we queue a work_struct for doing so.
262 * We need to get an extra ref for the skb (so it is not dropped), as
263 * well as be careful not to queue more than one request (won't help
264 * at all). If more than one request comes or there are errors, we
265 * just drop the packets (see i2400m_hard_start_xmit()).
268 int i2400m_net_wake_tx(struct i2400m *i2400m, struct net_device *net_dev,
272 struct device *dev = i2400m_dev(i2400m);
275 d_fnstart(3, dev, "(skb %p net_dev %p)\n", skb, net_dev);
276 if (net_ratelimit()) {
277 d_printf(3, dev, "WAKE&NETTX: "
278 "skb %p sending %d bytes to radio\n",
280 d_dump(4, dev, skb->data, skb->len);
282 /* We hold a ref count for i2400m and skb, so when
283 * stopping() the device, we need to cancel that work
284 * and if pending, release those resources. */
286 spin_lock_irqsave(&i2400m->tx_lock, flags);
287 if (!work_pending(&i2400m->wake_tx_ws)) {
288 netif_stop_queue(net_dev);
290 i2400m->wake_tx_skb = skb_get(skb); /* transfer ref count */
291 i2400m_tx_prep_header(skb);
292 result = schedule_work(&i2400m->wake_tx_ws);
293 WARN_ON(result == 0);
295 spin_unlock_irqrestore(&i2400m->tx_lock, flags);
297 /* Yes, this happens even if we stopped the
298 * queue -- blame the queue disciplines that
299 * queue without looking -- I guess there is a reason
302 d_printf(1, dev, "NETTX: device exiting idle, "
303 "dropping skb %p, queue running %d\n",
304 skb, netif_queue_stopped(net_dev));
307 d_fnend(3, dev, "(skb %p net_dev %p) = %d\n", skb, net_dev, result);
313 * Transmit a packet to the base station on behalf of the network stack.
315 * Returns: 0 if ok, < 0 errno code on error.
317 * We need to pull the ethernet header and add the hardware header,
318 * which is currently set to all zeroes and reserved.
321 int i2400m_net_tx(struct i2400m *i2400m, struct net_device *net_dev,
325 struct device *dev = i2400m_dev(i2400m);
327 d_fnstart(3, dev, "(i2400m %p net_dev %p skb %p)\n",
328 i2400m, net_dev, skb);
329 /* FIXME: check eth hdr, only IPv4 is routed by the device as of now */
330 net_dev->trans_start = jiffies;
331 i2400m_tx_prep_header(skb);
332 d_printf(3, dev, "NETTX: skb %p sending %d bytes to radio\n",
334 d_dump(4, dev, skb->data, skb->len);
335 result = i2400m_tx(i2400m, skb->data, skb->len, I2400M_PT_DATA);
336 d_fnend(3, dev, "(i2400m %p net_dev %p skb %p) = %d\n",
337 i2400m, net_dev, skb, result);
343 * Transmit a packet to the base station on behalf of the network stack
346 * Returns: NETDEV_TX_OK (always, even in case of error)
348 * In case of error, we just drop it. Reasons:
350 * - we add a hw header to each skb, and if the network stack
351 * retries, we have no way to know if that skb has it or not.
353 * - network protocols have their own drop-recovery mechanisms
355 * - there is not much else we can do
357 * If the device is idle, we need to wake it up; that is an operation
358 * that will sleep. See i2400m_net_wake_tx() for details.
361 netdev_tx_t i2400m_hard_start_xmit(struct sk_buff *skb,
362 struct net_device *net_dev)
364 struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
365 struct device *dev = i2400m_dev(i2400m);
368 d_fnstart(3, dev, "(skb %p net_dev %p)\n", skb, net_dev);
369 if (skb_header_cloned(skb)) {
371 * Make tcpdump/wireshark happy -- if they are
372 * running, the skb is cloned and we will overwrite
373 * the mac fields in i2400m_tx_prep_header. Expand
374 * seems to fix this...
376 result = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
378 result = NETDEV_TX_BUSY;
383 if (i2400m->state == I2400M_SS_IDLE)
384 result = i2400m_net_wake_tx(i2400m, net_dev, skb);
386 result = i2400m_net_tx(i2400m, net_dev, skb);
388 net_dev->stats.tx_dropped++;
390 net_dev->stats.tx_packets++;
391 net_dev->stats.tx_bytes += skb->len;
393 result = NETDEV_TX_OK;
396 d_fnend(3, dev, "(skb %p net_dev %p) = %d\n", skb, net_dev, result);
402 int i2400m_change_mtu(struct net_device *net_dev, int new_mtu)
405 struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
406 struct device *dev = i2400m_dev(i2400m);
408 if (new_mtu >= I2400M_MAX_MTU) {
409 dev_err(dev, "Cannot change MTU to %d (max is %d)\n",
410 new_mtu, I2400M_MAX_MTU);
413 net_dev->mtu = new_mtu;
421 void i2400m_tx_timeout(struct net_device *net_dev)
424 * We might want to kick the device
426 * There is not much we can do though, as the device requires
427 * that we send the data aggregated. By the time we receive
428 * this, there might be data pending to be sent or not...
430 net_dev->stats.tx_errors++;
436 * Create a fake ethernet header
438 * For emulating an ethernet device, every received IP header has to
439 * be prefixed with an ethernet header. Fake it with the given
443 void i2400m_rx_fake_eth_header(struct net_device *net_dev,
444 void *_eth_hdr, __be16 protocol)
446 struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
447 struct ethhdr *eth_hdr = _eth_hdr;
449 memcpy(eth_hdr->h_dest, net_dev->dev_addr, sizeof(eth_hdr->h_dest));
450 memcpy(eth_hdr->h_source, i2400m->src_mac_addr,
451 sizeof(eth_hdr->h_source));
452 eth_hdr->h_proto = protocol;
457 * i2400m_net_rx - pass a network packet to the stack
459 * @i2400m: device instance
460 * @skb_rx: the skb where the buffer pointed to by @buf is
461 * @i: 1 if payload is the only one
462 * @buf: pointer to the buffer containing the data
463 * @len: buffer's length
465 * This is only used now for the v1.3 firmware. It will be deprecated
468 * Note that due to firmware limitations, we don't have space to add
469 * an ethernet header, so we need to copy each packet. Firmware
470 * versions >= v1.4 fix this [see i2400m_net_erx()].
472 * We just clone the skb and set it up so that it's skb->data pointer
473 * points to "buf" and it's length.
475 * Note that if the payload is the last (or the only one) in a
476 * multi-payload message, we don't clone the SKB but just reuse it.
478 * This function is normally run from a thread context. However, we
479 * still use netif_rx() instead of netif_receive_skb() as was
480 * recommended in the mailing list. Reason is in some stress tests
481 * when sending/receiving a lot of data we seem to hit a softlock in
482 * the kernel's TCP implementation [aroudn tcp_delay_timer()]. Using
483 * netif_rx() took care of the issue.
485 * This is, of course, still open to do more research on why running
486 * with netif_receive_skb() hits this softlock. FIXME.
488 * FIXME: currently we don't do any efforts at distinguishing if what
489 * we got was an IPv4 or IPv6 header, to setup the protocol field
492 void i2400m_net_rx(struct i2400m *i2400m, struct sk_buff *skb_rx,
493 unsigned i, const void *buf, int buf_len)
495 struct net_device *net_dev = i2400m->wimax_dev.net_dev;
496 struct device *dev = i2400m_dev(i2400m);
499 d_fnstart(2, dev, "(i2400m %p buf %p buf_len %d)\n",
500 i2400m, buf, buf_len);
502 skb = skb_get(skb_rx);
503 d_printf(2, dev, "RX: reusing first payload skb %p\n", skb);
504 skb_pull(skb, buf - (void *) skb->data);
505 skb_trim(skb, (void *) skb_end_pointer(skb) - buf);
507 /* Yes, this is bad -- a lot of overhead -- see
508 * comments at the top of the file */
509 skb = __netdev_alloc_skb(net_dev, buf_len, GFP_KERNEL);
511 dev_err(dev, "NETRX: no memory to realloc skb\n");
512 net_dev->stats.rx_dropped++;
513 goto error_skb_realloc;
515 memcpy(skb_put(skb, buf_len), buf, buf_len);
517 i2400m_rx_fake_eth_header(i2400m->wimax_dev.net_dev,
518 skb->data - ETH_HLEN,
519 cpu_to_be16(ETH_P_IP));
520 skb_set_mac_header(skb, -ETH_HLEN);
521 skb->dev = i2400m->wimax_dev.net_dev;
522 skb->protocol = htons(ETH_P_IP);
523 net_dev->stats.rx_packets++;
524 net_dev->stats.rx_bytes += buf_len;
525 d_printf(3, dev, "NETRX: receiving %d bytes to network stack\n",
527 d_dump(4, dev, buf, buf_len);
528 netif_rx_ni(skb); /* see notes in function header */
530 d_fnend(2, dev, "(i2400m %p buf %p buf_len %d) = void\n",
531 i2400m, buf, buf_len);
536 * i2400m_net_erx - pass a network packet to the stack (extended version)
538 * @i2400m: device descriptor
539 * @skb: the skb where the packet is - the skb should be set to point
540 * at the IP packet; this function will add ethernet headers if
544 * This is only used now for firmware >= v1.4. Note it is quite
545 * similar to i2400m_net_rx() (used only for v1.3 firmware).
547 * This function is normally run from a thread context. However, we
548 * still use netif_rx() instead of netif_receive_skb() as was
549 * recommended in the mailing list. Reason is in some stress tests
550 * when sending/receiving a lot of data we seem to hit a softlock in
551 * the kernel's TCP implementation [aroudn tcp_delay_timer()]. Using
552 * netif_rx() took care of the issue.
554 * This is, of course, still open to do more research on why running
555 * with netif_receive_skb() hits this softlock. FIXME.
557 void i2400m_net_erx(struct i2400m *i2400m, struct sk_buff *skb,
560 struct net_device *net_dev = i2400m->wimax_dev.net_dev;
561 struct device *dev = i2400m_dev(i2400m);
564 d_fnstart(2, dev, "(i2400m %p skb %p [%u] cs %d)\n",
565 i2400m, skb, skb->len, cs);
567 case I2400M_CS_IPV4_0:
570 i2400m_rx_fake_eth_header(i2400m->wimax_dev.net_dev,
571 skb->data - ETH_HLEN,
572 cpu_to_be16(ETH_P_IP));
573 skb_set_mac_header(skb, -ETH_HLEN);
574 skb->dev = i2400m->wimax_dev.net_dev;
575 skb->protocol = htons(ETH_P_IP);
576 net_dev->stats.rx_packets++;
577 net_dev->stats.rx_bytes += skb->len;
580 dev_err(dev, "ERX: BUG? CS type %u unsupported\n", cs);
584 d_printf(3, dev, "ERX: receiving %d bytes to the network stack\n",
586 d_dump(4, dev, skb->data, skb->len);
587 netif_rx_ni(skb); /* see notes in function header */
589 d_fnend(2, dev, "(i2400m %p skb %p [%u] cs %d) = void\n",
590 i2400m, skb, skb->len, cs);
593 static const struct net_device_ops i2400m_netdev_ops = {
594 .ndo_open = i2400m_open,
595 .ndo_stop = i2400m_stop,
596 .ndo_start_xmit = i2400m_hard_start_xmit,
597 .ndo_tx_timeout = i2400m_tx_timeout,
598 .ndo_change_mtu = i2400m_change_mtu,
601 static void i2400m_get_drvinfo(struct net_device *net_dev,
602 struct ethtool_drvinfo *info)
604 struct i2400m *i2400m = net_dev_to_i2400m(net_dev);
606 strncpy(info->driver, KBUILD_MODNAME, sizeof(info->driver) - 1);
607 strncpy(info->fw_version, i2400m->fw_name, sizeof(info->fw_version) - 1);
608 if (net_dev->dev.parent)
609 strncpy(info->bus_info, dev_name(net_dev->dev.parent),
610 sizeof(info->bus_info) - 1);
613 static const struct ethtool_ops i2400m_ethtool_ops = {
614 .get_drvinfo = i2400m_get_drvinfo,
615 .get_link = ethtool_op_get_link,
619 * i2400m_netdev_setup - Setup setup @net_dev's i2400m private data
621 * Called by alloc_netdev()
623 void i2400m_netdev_setup(struct net_device *net_dev)
625 d_fnstart(3, NULL, "(net_dev %p)\n", net_dev);
626 ether_setup(net_dev);
627 net_dev->mtu = I2400M_MAX_MTU;
628 net_dev->tx_queue_len = I2400M_TX_QLEN;
630 NETIF_F_VLAN_CHALLENGED
633 IFF_NOARP /* i2400m is apure IP device */
634 & (~IFF_BROADCAST /* i2400m is P2P */
636 net_dev->watchdog_timeo = I2400M_TX_TIMEOUT;
637 net_dev->netdev_ops = &i2400m_netdev_ops;
638 net_dev->ethtool_ops = &i2400m_ethtool_ops;
639 d_fnend(3, NULL, "(net_dev %p) = void\n", net_dev);
641 EXPORT_SYMBOL_GPL(i2400m_netdev_setup);