2 * Copyright (C) 2017 - Cambridge Greys Limited
3 * Copyright (C) 2011 - 2014 Cisco Systems Inc
4 * Copyright (C) 2001 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
5 * Copyright (C) 2001 Lennert Buytenhek (buytenh@gnu.org) and
6 * James Leu (jleu@mindspring.net).
7 * Copyright (C) 2001 by various other people who didn't put their name here.
8 * Licensed under the GPL.
11 #include <linux/version.h>
12 #include <linux/memblock.h>
13 #include <linux/etherdevice.h>
14 #include <linux/ethtool.h>
15 #include <linux/inetdevice.h>
16 #include <linux/init.h>
17 #include <linux/list.h>
18 #include <linux/netdevice.h>
19 #include <linux/platform_device.h>
20 #include <linux/rtnetlink.h>
21 #include <linux/skbuff.h>
22 #include <linux/slab.h>
23 #include <linux/interrupt.h>
29 #include "mconsole_kern.h"
30 #include "vector_user.h"
31 #include "vector_kern.h"
34 * Adapted from network devices with the following major changes:
35 * All transports are static - simplifies the code significantly
36 * Multiple FDs/IRQs per device
37 * Vector IO optionally used for read/write, falling back to legacy
38 * based on configuration and/or availability
39 * Configuration is no longer positional - L2TPv3 and GRE require up to
40 * 10 parameters, passing this as positional is not fit for purpose.
41 * Only socket transports are supported
45 #define DRIVER_NAME "uml-vector"
46 #define DRIVER_VERSION "01"
47 struct vector_cmd_line_arg {
48 struct list_head list;
53 struct vector_device {
54 struct list_head list;
55 struct net_device *dev;
56 struct platform_device pdev;
61 static LIST_HEAD(vec_cmd_line);
63 static DEFINE_SPINLOCK(vector_devices_lock);
64 static LIST_HEAD(vector_devices);
66 static int driver_registered;
68 static void vector_eth_configure(int n, struct arglist *def);
70 /* Argument accessors to set variables (and/or set default values)
71 * mtu, buffer sizing, default headroom, etc
74 #define DEFAULT_HEADROOM 2
75 #define SAFETY_MARGIN 32
76 #define DEFAULT_VECTOR_SIZE 64
77 #define TX_SMALL_PACKET 128
78 #define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
81 const char string[ETH_GSTRING_LEN];
82 } ethtool_stats_keys[] = {
84 { "rx_queue_running_average" },
86 { "tx_queue_running_average" },
87 { "rx_encaps_errors" },
88 { "tx_timeout_count" },
89 { "tx_restart_queue" },
91 { "tx_flow_control_xon" },
92 { "tx_flow_control_xoff" },
93 { "rx_csum_offload_good" },
94 { "rx_csum_offload_errors"},
99 #define VECTOR_NUM_STATS ARRAY_SIZE(ethtool_stats_keys)
101 static void vector_reset_stats(struct vector_private *vp)
103 vp->estats.rx_queue_max = 0;
104 vp->estats.rx_queue_running_average = 0;
105 vp->estats.tx_queue_max = 0;
106 vp->estats.tx_queue_running_average = 0;
107 vp->estats.rx_encaps_errors = 0;
108 vp->estats.tx_timeout_count = 0;
109 vp->estats.tx_restart_queue = 0;
110 vp->estats.tx_kicks = 0;
111 vp->estats.tx_flow_control_xon = 0;
112 vp->estats.tx_flow_control_xoff = 0;
113 vp->estats.sg_ok = 0;
114 vp->estats.sg_linearized = 0;
117 static int get_mtu(struct arglist *def)
119 char *mtu = uml_vector_fetch_arg(def, "mtu");
123 if (kstrtoul(mtu, 10, &result) == 0)
126 return ETH_MAX_PACKET;
129 static int get_depth(struct arglist *def)
131 char *mtu = uml_vector_fetch_arg(def, "depth");
135 if (kstrtoul(mtu, 10, &result) == 0)
138 return DEFAULT_VECTOR_SIZE;
141 static int get_headroom(struct arglist *def)
143 char *mtu = uml_vector_fetch_arg(def, "headroom");
147 if (kstrtoul(mtu, 10, &result) == 0)
150 return DEFAULT_HEADROOM;
153 static int get_req_size(struct arglist *def)
155 char *gro = uml_vector_fetch_arg(def, "gro");
159 if (kstrtoul(gro, 10, &result) == 0) {
164 return get_mtu(def) + ETH_HEADER_OTHER +
165 get_headroom(def) + SAFETY_MARGIN;
169 static int get_transport_options(struct arglist *def)
171 char *transport = uml_vector_fetch_arg(def, "transport");
172 char *vector = uml_vector_fetch_arg(def, "vec");
174 int vec_rx = VECTOR_RX;
175 int vec_tx = VECTOR_TX;
178 if (vector != NULL) {
179 if (kstrtoul(vector, 10, &parsed) == 0) {
188 if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
189 return (vec_rx | VECTOR_BPF);
190 if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
191 return (vec_rx | vec_tx | VECTOR_QDISC_BYPASS);
192 return (vec_rx | vec_tx);
196 /* A mini-buffer for packet drop read
197 * All of our supported transports are datagram oriented and we always
198 * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
199 * than the packet size it still counts as full packet read and will
200 * clean the incoming stream to keep sigio/epoll happy
203 #define DROP_BUFFER_SIZE 32
205 static char *drop_buffer;
207 /* Array backed queues optimized for bulk enqueue/dequeue and
208 * 1:N (small values of N) or 1:1 enqueuer/dequeuer ratios.
209 * For more details and full design rationale see
210 * http://foswiki.cambridgegreys.com/Main/EatYourTailAndEnjoyIt
215 * Advance the mmsg queue head by n = advance. Resets the queue to
216 * maximum enqueue/dequeue-at-once capacity if possible. Called by
217 * dequeuers. Caller must hold the head_lock!
220 static int vector_advancehead(struct vector_queue *qi, int advance)
229 spin_lock(&qi->tail_lock);
230 qi->queue_depth -= advance;
232 /* we are at 0, use this to
233 * reset head and tail so we can use max size vectors
236 if (qi->queue_depth == 0) {
240 queue_depth = qi->queue_depth;
241 spin_unlock(&qi->tail_lock);
245 /* Advance the queue tail by n = advance.
246 * This is called by enqueuers which should hold the
250 static int vector_advancetail(struct vector_queue *qi, int advance)
257 spin_lock(&qi->head_lock);
258 qi->queue_depth += advance;
259 queue_depth = qi->queue_depth;
260 spin_unlock(&qi->head_lock);
264 static int prep_msg(struct vector_private *vp,
270 skb_frag_t *skb_frag;
272 nr_frags = skb_shinfo(skb)->nr_frags;
273 if (nr_frags > MAX_IOV_SIZE) {
274 if (skb_linearize(skb) != 0)
277 if (vp->header_size > 0) {
278 iov[iov_index].iov_len = vp->header_size;
279 vp->form_header(iov[iov_index].iov_base, skb, vp);
282 iov[iov_index].iov_base = skb->data;
284 iov[iov_index].iov_len = skb->len - skb->data_len;
287 iov[iov_index].iov_len = skb->len;
289 for (frag = 0; frag < nr_frags; frag++) {
290 skb_frag = &skb_shinfo(skb)->frags[frag];
291 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
292 iov[iov_index].iov_len = skb_frag_size(skb_frag);
300 * Generic vector enqueue with support for forming headers using transport
301 * specific callback. Allows GRE, L2TPv3, RAW and other transports
302 * to use a common enqueue procedure in vector mode
305 static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
307 struct vector_private *vp = netdev_priv(qi->dev);
310 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
313 spin_lock(&qi->tail_lock);
314 spin_lock(&qi->head_lock);
315 queue_depth = qi->queue_depth;
316 spin_unlock(&qi->head_lock);
319 packet_len = skb->len;
321 if (queue_depth < qi->max_depth) {
323 *(qi->skbuff_vector + qi->tail) = skb;
324 mmsg_vector += qi->tail;
325 iov_count = prep_msg(
328 mmsg_vector->msg_hdr.msg_iov
332 mmsg_vector->msg_hdr.msg_iovlen = iov_count;
333 mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
334 mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
335 queue_depth = vector_advancetail(qi, 1);
338 spin_unlock(&qi->tail_lock);
341 qi->dev->stats.tx_dropped++;
343 packet_len = skb->len;
344 dev_consume_skb_any(skb);
345 netdev_completed_queue(qi->dev, 1, packet_len);
347 spin_unlock(&qi->tail_lock);
351 static int consume_vector_skbs(struct vector_queue *qi, int count)
357 for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
358 skb = *(qi->skbuff_vector + skb_index);
359 /* mark as empty to ensure correct destruction if
362 bytes_compl += skb->len;
363 *(qi->skbuff_vector + skb_index) = NULL;
364 dev_consume_skb_any(skb);
366 qi->dev->stats.tx_bytes += bytes_compl;
367 qi->dev->stats.tx_packets += count;
368 netdev_completed_queue(qi->dev, count, bytes_compl);
369 return vector_advancehead(qi, count);
373 * Generic vector deque via sendmmsg with support for forming headers
374 * using transport specific callback. Allows GRE, L2TPv3, RAW and
375 * other transports to use a common dequeue procedure in vector mode
379 static int vector_send(struct vector_queue *qi)
381 struct vector_private *vp = netdev_priv(qi->dev);
382 struct mmsghdr *send_from;
383 int result = 0, send_len, queue_depth = qi->max_depth;
385 if (spin_trylock(&qi->head_lock)) {
386 if (spin_trylock(&qi->tail_lock)) {
387 /* update queue_depth to current value */
388 queue_depth = qi->queue_depth;
389 spin_unlock(&qi->tail_lock);
390 while (queue_depth > 0) {
391 /* Calculate the start of the vector */
392 send_len = queue_depth;
393 send_from = qi->mmsg_vector;
394 send_from += qi->head;
395 /* Adjust vector size if wraparound */
396 if (send_len + qi->head > qi->max_depth)
397 send_len = qi->max_depth - qi->head;
398 /* Try to TX as many packets as possible */
400 result = uml_vector_sendmmsg(
407 (result != send_len);
409 /* For some of the sendmmsg error scenarios
410 * we may end being unsure in the TX success
411 * for all packets. It is safer to declare
412 * them all TX-ed and blame the network.
416 netdev_err(vp->dev, "sendmmsg err=%i\n",
422 consume_vector_skbs(qi, result);
423 /* This is equivalent to an TX IRQ.
424 * Restart the upper layers to feed us
427 if (result > vp->estats.tx_queue_max)
428 vp->estats.tx_queue_max = result;
429 vp->estats.tx_queue_running_average =
430 (vp->estats.tx_queue_running_average + result) >> 1;
432 netif_trans_update(qi->dev);
433 netif_wake_queue(qi->dev);
434 /* if TX is busy, break out of the send loop,
435 * poll write IRQ will reschedule xmit for us
437 if (result != send_len) {
438 vp->estats.tx_restart_queue++;
443 spin_unlock(&qi->head_lock);
445 tasklet_schedule(&vp->tx_poll);
450 /* Queue destructor. Deliberately stateless so we can use
451 * it in queue cleanup if initialization fails.
454 static void destroy_queue(struct vector_queue *qi)
458 struct vector_private *vp = netdev_priv(qi->dev);
459 struct mmsghdr *mmsg_vector;
463 /* deallocate any skbuffs - we rely on any unused to be
466 if (qi->skbuff_vector != NULL) {
467 for (i = 0; i < qi->max_depth; i++) {
468 if (*(qi->skbuff_vector + i) != NULL)
469 dev_kfree_skb_any(*(qi->skbuff_vector + i));
471 kfree(qi->skbuff_vector);
473 /* deallocate matching IOV structures including header buffs */
474 if (qi->mmsg_vector != NULL) {
475 mmsg_vector = qi->mmsg_vector;
476 for (i = 0; i < qi->max_depth; i++) {
477 iov = mmsg_vector->msg_hdr.msg_iov;
479 if ((vp->header_size > 0) &&
480 (iov->iov_base != NULL))
481 kfree(iov->iov_base);
486 kfree(qi->mmsg_vector);
492 * Queue constructor. Create a queue with a given side.
494 static struct vector_queue *create_queue(
495 struct vector_private *vp,
500 struct vector_queue *result;
503 struct mmsghdr *mmsg_vector;
505 result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
508 result->max_depth = max_size;
509 result->dev = vp->dev;
510 result->mmsg_vector = kmalloc(
511 (sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
512 if (result->mmsg_vector == NULL)
514 result->skbuff_vector = kmalloc(
515 (sizeof(void *) * max_size), GFP_KERNEL);
516 if (result->skbuff_vector == NULL)
519 /* further failures can be handled safely by destroy_queue*/
521 mmsg_vector = result->mmsg_vector;
522 for (i = 0; i < max_size; i++) {
523 /* Clear all pointers - we use non-NULL as marking on
524 * what to free on destruction
526 *(result->skbuff_vector + i) = NULL;
527 mmsg_vector->msg_hdr.msg_iov = NULL;
530 mmsg_vector = result->mmsg_vector;
531 result->max_iov_frags = num_extra_frags;
532 for (i = 0; i < max_size; i++) {
533 if (vp->header_size > 0)
534 iov = kmalloc_array(3 + num_extra_frags,
535 sizeof(struct iovec),
539 iov = kmalloc_array(2 + num_extra_frags,
540 sizeof(struct iovec),
545 mmsg_vector->msg_hdr.msg_iov = iov;
546 mmsg_vector->msg_hdr.msg_iovlen = 1;
547 mmsg_vector->msg_hdr.msg_control = NULL;
548 mmsg_vector->msg_hdr.msg_controllen = 0;
549 mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
550 mmsg_vector->msg_hdr.msg_name = NULL;
551 mmsg_vector->msg_hdr.msg_namelen = 0;
552 if (vp->header_size > 0) {
553 iov->iov_base = kmalloc(header_size, GFP_KERNEL);
554 if (iov->iov_base == NULL)
556 iov->iov_len = header_size;
557 mmsg_vector->msg_hdr.msg_iovlen = 2;
560 iov->iov_base = NULL;
564 spin_lock_init(&result->head_lock);
565 spin_lock_init(&result->tail_lock);
566 result->queue_depth = 0;
571 kfree(result->mmsg_vector);
576 destroy_queue(result);
581 * We do not use the RX queue as a proper wraparound queue for now
582 * This is not necessary because the consumption via netif_rx()
583 * happens in-line. While we can try using the return code of
584 * netif_rx() for flow control there are no drivers doing this today.
585 * For this RX specific use we ignore the tail/head locks and
586 * just read into a prepared queue filled with skbuffs.
589 static struct sk_buff *prep_skb(
590 struct vector_private *vp,
591 struct user_msghdr *msg)
593 int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
594 struct sk_buff *result;
595 int iov_index = 0, len;
596 struct iovec *iov = msg->msg_iov;
597 int err, nr_frags, frag;
598 skb_frag_t *skb_frag;
600 if (vp->req_size <= linear)
604 result = alloc_skb_with_frags(
606 len - vp->max_packet,
611 if (vp->header_size > 0)
613 if (result == NULL) {
614 iov[iov_index].iov_base = NULL;
615 iov[iov_index].iov_len = 0;
618 skb_reserve(result, vp->headroom);
619 result->dev = vp->dev;
620 skb_put(result, vp->max_packet);
621 result->data_len = len - vp->max_packet;
622 result->len += len - vp->max_packet;
623 skb_reset_mac_header(result);
624 result->ip_summed = CHECKSUM_NONE;
625 iov[iov_index].iov_base = result->data;
626 iov[iov_index].iov_len = vp->max_packet;
629 nr_frags = skb_shinfo(result)->nr_frags;
630 for (frag = 0; frag < nr_frags; frag++) {
631 skb_frag = &skb_shinfo(result)->frags[frag];
632 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
633 if (iov[iov_index].iov_base != NULL)
634 iov[iov_index].iov_len = skb_frag_size(skb_frag);
636 iov[iov_index].iov_len = 0;
640 msg->msg_iovlen = iov_index;
645 /* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs*/
647 static void prep_queue_for_rx(struct vector_queue *qi)
649 struct vector_private *vp = netdev_priv(qi->dev);
650 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
651 void **skbuff_vector = qi->skbuff_vector;
654 if (qi->queue_depth == 0)
656 for (i = 0; i < qi->queue_depth; i++) {
657 /* it is OK if allocation fails - recvmmsg with NULL data in
658 * iov argument still performs an RX, just drops the packet
659 * This allows us stop faffing around with a "drop buffer"
662 *skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
669 static struct vector_device *find_device(int n)
671 struct vector_device *device;
672 struct list_head *ele;
674 spin_lock(&vector_devices_lock);
675 list_for_each(ele, &vector_devices) {
676 device = list_entry(ele, struct vector_device, list);
677 if (device->unit == n)
682 spin_unlock(&vector_devices_lock);
686 static int vector_parse(char *str, int *index_out, char **str_out,
694 while ((*str != ':') && (strlen(str) > 1))
697 *error_out = "Expected ':' after device number";
702 err = kstrtouint(start, 0, &n);
704 *error_out = "Bad device number";
709 if (find_device(n)) {
710 *error_out = "Device already configured";
719 static int vector_config(char *str, char **error_out)
723 struct arglist *parsed;
725 err = vector_parse(str, &n, ¶ms, error_out);
729 /* This string is broken up and the pieces used by the underlying
730 * driver. We should copy it to make sure things do not go wrong
734 params = kstrdup(params, GFP_KERNEL);
735 if (params == NULL) {
736 *error_out = "vector_config failed to strdup string";
740 parsed = uml_parse_vector_ifspec(params);
742 if (parsed == NULL) {
743 *error_out = "vector_config failed to parse parameters";
747 vector_eth_configure(n, parsed);
751 static int vector_id(char **str, int *start_out, int *end_out)
756 n = simple_strtoul(*str, &end, 0);
757 if ((*end != '\0') || (end == *str))
766 static int vector_remove(int n, char **error_out)
768 struct vector_device *vec_d;
769 struct net_device *dev;
770 struct vector_private *vp;
772 vec_d = find_device(n);
776 vp = netdev_priv(dev);
779 unregister_netdev(dev);
780 platform_device_unregister(&vec_d->pdev);
785 * There is no shared per-transport initialization code, so
786 * we will just initialize each interface one by one and
790 static struct platform_driver uml_net_driver = {
797 static void vector_device_release(struct device *dev)
799 struct vector_device *device = dev_get_drvdata(dev);
800 struct net_device *netdev = device->dev;
802 list_del(&device->list);
807 /* Bog standard recv using recvmsg - not used normally unless the user
808 * explicitly specifies not to use recvmmsg vector RX.
811 static int vector_legacy_rx(struct vector_private *vp)
814 struct user_msghdr hdr;
815 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
822 hdr.msg_iov = (struct iovec *) &iov;
823 hdr.msg_control = NULL;
824 hdr.msg_controllen = 0;
827 if (vp->header_size > 0) {
828 iov[0].iov_base = vp->header_rxbuffer;
829 iov[0].iov_len = vp->header_size;
832 skb = prep_skb(vp, &hdr);
835 /* Read a packet into drop_buffer and don't do
838 iov[iovpos].iov_base = drop_buffer;
839 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
841 vp->dev->stats.rx_dropped++;
844 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
847 if (pkt_len > vp->header_size) {
848 if (vp->header_size > 0) {
849 header_check = vp->verify_header(
850 vp->header_rxbuffer, skb, vp);
851 if (header_check < 0) {
852 dev_kfree_skb_irq(skb);
853 vp->dev->stats.rx_dropped++;
854 vp->estats.rx_encaps_errors++;
857 if (header_check > 0) {
858 vp->estats.rx_csum_offload_good++;
859 skb->ip_summed = CHECKSUM_UNNECESSARY;
862 pskb_trim(skb, pkt_len - vp->rx_header_size);
863 skb->protocol = eth_type_trans(skb, skb->dev);
864 vp->dev->stats.rx_bytes += skb->len;
865 vp->dev->stats.rx_packets++;
868 dev_kfree_skb_irq(skb);
875 * Packet at a time TX which falls back to vector TX if the
876 * underlying transport is busy.
881 static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
883 struct iovec iov[3 + MAX_IOV_SIZE];
884 int iov_count, pkt_len = 0;
886 iov[0].iov_base = vp->header_txbuffer;
887 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
891 pkt_len = uml_vector_writev(
893 (struct iovec *) &iov,
897 netif_trans_update(vp->dev);
898 netif_wake_queue(vp->dev);
901 vp->dev->stats.tx_bytes += skb->len;
902 vp->dev->stats.tx_packets++;
904 vp->dev->stats.tx_dropped++;
909 vp->dev->stats.tx_dropped++;
915 * Receive as many messages as we can in one call using the special
916 * mmsg vector matched to an skb vector which we prepared earlier.
919 static int vector_mmsg_rx(struct vector_private *vp)
922 struct vector_queue *qi = vp->rx_queue;
924 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
925 void **skbuff_vector = qi->skbuff_vector;
928 /* Refresh the vector and make sure it is with new skbs and the
929 * iovs are updated to point to them.
932 prep_queue_for_rx(qi);
934 /* Fire the Lazy Gun - get as many packets as we can in one go. */
936 packet_count = uml_vector_recvmmsg(
937 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
939 if (packet_count <= 0)
942 /* We treat packet processing as enqueue, buffer refresh as dequeue
943 * The queue_depth tells us how many buffers have been used and how
944 * many do we need to prep the next time prep_queue_for_rx() is called.
947 qi->queue_depth = packet_count;
949 for (i = 0; i < packet_count; i++) {
950 skb = (*skbuff_vector);
951 if (mmsg_vector->msg_len > vp->header_size) {
952 if (vp->header_size > 0) {
953 header_check = vp->verify_header(
954 mmsg_vector->msg_hdr.msg_iov->iov_base,
958 if (header_check < 0) {
959 /* Overlay header failed to verify - discard.
960 * We can actually keep this skb and reuse it,
961 * but that will make the prep logic too
964 dev_kfree_skb_irq(skb);
965 vp->estats.rx_encaps_errors++;
968 if (header_check > 0) {
969 vp->estats.rx_csum_offload_good++;
970 skb->ip_summed = CHECKSUM_UNNECESSARY;
974 mmsg_vector->msg_len - vp->rx_header_size);
975 skb->protocol = eth_type_trans(skb, skb->dev);
977 * We do not need to lock on updating stats here
978 * The interrupt loop is non-reentrant.
980 vp->dev->stats.rx_bytes += skb->len;
981 vp->dev->stats.rx_packets++;
984 /* Overlay header too short to do anything - discard.
985 * We can actually keep this skb and reuse it,
986 * but that will make the prep logic too complex.
989 dev_kfree_skb_irq(skb);
991 (*skbuff_vector) = NULL;
992 /* Move to the next buffer element */
996 if (packet_count > 0) {
997 if (vp->estats.rx_queue_max < packet_count)
998 vp->estats.rx_queue_max = packet_count;
999 vp->estats.rx_queue_running_average =
1000 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1002 return packet_count;
1005 static void vector_rx(struct vector_private *vp)
1009 if ((vp->options & VECTOR_RX) > 0)
1010 while ((err = vector_mmsg_rx(vp)) > 0)
1013 while ((err = vector_legacy_rx(vp)) > 0)
1015 if ((err != 0) && net_ratelimit())
1016 netdev_err(vp->dev, "vector_rx: error(%d)\n", err);
1019 static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1021 struct vector_private *vp = netdev_priv(dev);
1022 int queue_depth = 0;
1024 if ((vp->options & VECTOR_TX) == 0) {
1026 return NETDEV_TX_OK;
1029 /* We do BQL only in the vector path, no point doing it in
1030 * packet at a time mode as there is no device queue
1033 netdev_sent_queue(vp->dev, skb->len);
1034 queue_depth = vector_enqueue(vp->tx_queue, skb);
1036 /* if the device queue is full, stop the upper layers and
1040 if (queue_depth >= vp->tx_queue->max_depth - 1) {
1041 vp->estats.tx_kicks++;
1042 netif_stop_queue(dev);
1043 vector_send(vp->tx_queue);
1044 return NETDEV_TX_OK;
1046 if (skb->xmit_more) {
1047 mod_timer(&vp->tl, vp->coalesce);
1048 return NETDEV_TX_OK;
1050 if (skb->len < TX_SMALL_PACKET) {
1051 vp->estats.tx_kicks++;
1052 vector_send(vp->tx_queue);
1054 tasklet_schedule(&vp->tx_poll);
1055 return NETDEV_TX_OK;
1058 static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1060 struct net_device *dev = dev_id;
1061 struct vector_private *vp = netdev_priv(dev);
1063 if (!netif_running(dev))
1070 static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1072 struct net_device *dev = dev_id;
1073 struct vector_private *vp = netdev_priv(dev);
1075 if (!netif_running(dev))
1077 /* We need to pay attention to it only if we got
1078 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1079 * we ignore it. In the future, it may be worth
1080 * it to improve the IRQ controller a bit to make
1081 * tweaking the IRQ mask less costly
1084 if (vp->in_write_poll)
1085 tasklet_schedule(&vp->tx_poll);
1092 static int vector_net_close(struct net_device *dev)
1094 struct vector_private *vp = netdev_priv(dev);
1095 unsigned long flags;
1097 netif_stop_queue(dev);
1100 if (vp->fds == NULL)
1103 /* Disable and free all IRQS */
1104 if (vp->rx_irq > 0) {
1105 um_free_irq(vp->rx_irq, dev);
1108 if (vp->tx_irq > 0) {
1109 um_free_irq(vp->tx_irq, dev);
1112 tasklet_kill(&vp->tx_poll);
1113 if (vp->fds->rx_fd > 0) {
1114 os_close_file(vp->fds->rx_fd);
1115 vp->fds->rx_fd = -1;
1117 if (vp->fds->tx_fd > 0) {
1118 os_close_file(vp->fds->tx_fd);
1119 vp->fds->tx_fd = -1;
1121 if (vp->bpf != NULL)
1123 if (vp->fds->remote_addr != NULL)
1124 kfree(vp->fds->remote_addr);
1125 if (vp->transport_data != NULL)
1126 kfree(vp->transport_data);
1127 if (vp->header_rxbuffer != NULL)
1128 kfree(vp->header_rxbuffer);
1129 if (vp->header_txbuffer != NULL)
1130 kfree(vp->header_txbuffer);
1131 if (vp->rx_queue != NULL)
1132 destroy_queue(vp->rx_queue);
1133 if (vp->tx_queue != NULL)
1134 destroy_queue(vp->tx_queue);
1137 spin_lock_irqsave(&vp->lock, flags);
1139 spin_unlock_irqrestore(&vp->lock, flags);
1145 static void vector_tx_poll(unsigned long data)
1147 struct vector_private *vp = (struct vector_private *)data;
1149 vp->estats.tx_kicks++;
1150 vector_send(vp->tx_queue);
1152 static void vector_reset_tx(struct work_struct *work)
1154 struct vector_private *vp =
1155 container_of(work, struct vector_private, reset_tx);
1156 netdev_reset_queue(vp->dev);
1157 netif_start_queue(vp->dev);
1158 netif_wake_queue(vp->dev);
1160 static int vector_net_open(struct net_device *dev)
1162 struct vector_private *vp = netdev_priv(dev);
1163 unsigned long flags;
1165 struct vector_device *vdevice;
1167 spin_lock_irqsave(&vp->lock, flags);
1169 spin_unlock_irqrestore(&vp->lock, flags);
1173 spin_unlock_irqrestore(&vp->lock, flags);
1175 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1177 if (vp->fds == NULL)
1180 if (build_transport_data(vp) < 0)
1183 if ((vp->options & VECTOR_RX) > 0) {
1184 vp->rx_queue = create_queue(
1186 get_depth(vp->parsed),
1190 vp->rx_queue->queue_depth = get_depth(vp->parsed);
1192 vp->header_rxbuffer = kmalloc(
1196 if (vp->header_rxbuffer == NULL)
1199 if ((vp->options & VECTOR_TX) > 0) {
1200 vp->tx_queue = create_queue(
1202 get_depth(vp->parsed),
1207 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1208 if (vp->header_txbuffer == NULL)
1213 err = um_request_irq(
1214 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1215 IRQ_READ, vector_rx_interrupt,
1216 IRQF_SHARED, dev->name, dev);
1218 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1222 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1223 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1224 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1226 /* WRITE IRQ - we need it only if we have vector TX */
1227 if ((vp->options & VECTOR_TX) > 0) {
1228 err = um_request_irq(
1229 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1230 IRQ_WRITE, vector_tx_interrupt,
1231 IRQF_SHARED, dev->name, dev);
1234 "vector_open: failed to get tx irq(%d)\n", err);
1238 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1239 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1242 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1243 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1244 vp->options |= VECTOR_BPF;
1246 if ((vp->options & VECTOR_BPF) != 0)
1247 vp->bpf = uml_vector_default_bpf(vp->fds->rx_fd, dev->dev_addr);
1249 netif_start_queue(dev);
1251 /* clear buffer - it can happen that the host side of the interface
1252 * is full when we get here. In this case, new data is never queued,
1253 * SIGIOs never arrive, and the net never works.
1258 vector_reset_stats(vp);
1259 vdevice = find_device(vp->unit);
1260 vdevice->opened = 1;
1262 if ((vp->options & VECTOR_TX) != 0)
1266 vector_net_close(dev);
1271 static void vector_net_set_multicast_list(struct net_device *dev)
1273 /* TODO: - we can do some BPF games here */
1277 static void vector_net_tx_timeout(struct net_device *dev)
1279 struct vector_private *vp = netdev_priv(dev);
1281 vp->estats.tx_timeout_count++;
1282 netif_trans_update(dev);
1283 schedule_work(&vp->reset_tx);
1286 static netdev_features_t vector_fix_features(struct net_device *dev,
1287 netdev_features_t features)
1289 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1293 static int vector_set_features(struct net_device *dev,
1294 netdev_features_t features)
1296 struct vector_private *vp = netdev_priv(dev);
1297 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1298 * no way to negotiate it on raw sockets, so we can change
1301 if (features & NETIF_F_GRO)
1302 /* All new frame buffers will be GRO-sized */
1303 vp->req_size = 65536;
1305 /* All new frame buffers will be normal sized */
1306 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1310 #ifdef CONFIG_NET_POLL_CONTROLLER
1311 static void vector_net_poll_controller(struct net_device *dev)
1313 disable_irq(dev->irq);
1314 vector_rx_interrupt(dev->irq, dev);
1315 enable_irq(dev->irq);
1319 static void vector_net_get_drvinfo(struct net_device *dev,
1320 struct ethtool_drvinfo *info)
1322 strlcpy(info->driver, DRIVER_NAME, sizeof(info->driver));
1323 strlcpy(info->version, DRIVER_VERSION, sizeof(info->version));
1326 static void vector_get_ringparam(struct net_device *netdev,
1327 struct ethtool_ringparam *ring)
1329 struct vector_private *vp = netdev_priv(netdev);
1331 ring->rx_max_pending = vp->rx_queue->max_depth;
1332 ring->tx_max_pending = vp->tx_queue->max_depth;
1333 ring->rx_pending = vp->rx_queue->max_depth;
1334 ring->tx_pending = vp->tx_queue->max_depth;
1337 static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1339 switch (stringset) {
1344 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1352 static int vector_get_sset_count(struct net_device *dev, int sset)
1358 return VECTOR_NUM_STATS;
1364 static void vector_get_ethtool_stats(struct net_device *dev,
1365 struct ethtool_stats *estats,
1368 struct vector_private *vp = netdev_priv(dev);
1370 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1373 static int vector_get_coalesce(struct net_device *netdev,
1374 struct ethtool_coalesce *ec)
1376 struct vector_private *vp = netdev_priv(netdev);
1378 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1382 static int vector_set_coalesce(struct net_device *netdev,
1383 struct ethtool_coalesce *ec)
1385 struct vector_private *vp = netdev_priv(netdev);
1387 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1388 if (vp->coalesce == 0)
1393 static const struct ethtool_ops vector_net_ethtool_ops = {
1394 .get_drvinfo = vector_net_get_drvinfo,
1395 .get_link = ethtool_op_get_link,
1396 .get_ts_info = ethtool_op_get_ts_info,
1397 .get_ringparam = vector_get_ringparam,
1398 .get_strings = vector_get_strings,
1399 .get_sset_count = vector_get_sset_count,
1400 .get_ethtool_stats = vector_get_ethtool_stats,
1401 .get_coalesce = vector_get_coalesce,
1402 .set_coalesce = vector_set_coalesce,
1406 static const struct net_device_ops vector_netdev_ops = {
1407 .ndo_open = vector_net_open,
1408 .ndo_stop = vector_net_close,
1409 .ndo_start_xmit = vector_net_start_xmit,
1410 .ndo_set_rx_mode = vector_net_set_multicast_list,
1411 .ndo_tx_timeout = vector_net_tx_timeout,
1412 .ndo_set_mac_address = eth_mac_addr,
1413 .ndo_validate_addr = eth_validate_addr,
1414 .ndo_fix_features = vector_fix_features,
1415 .ndo_set_features = vector_set_features,
1416 #ifdef CONFIG_NET_POLL_CONTROLLER
1417 .ndo_poll_controller = vector_net_poll_controller,
1422 static void vector_timer_expire(struct timer_list *t)
1424 struct vector_private *vp = from_timer(vp, t, tl);
1426 vp->estats.tx_kicks++;
1427 vector_send(vp->tx_queue);
1430 static void vector_eth_configure(
1435 struct vector_device *device;
1436 struct net_device *dev;
1437 struct vector_private *vp;
1440 device = kzalloc(sizeof(*device), GFP_KERNEL);
1441 if (device == NULL) {
1442 printk(KERN_ERR "eth_configure failed to allocate struct "
1446 dev = alloc_etherdev(sizeof(struct vector_private));
1448 printk(KERN_ERR "eth_configure: failed to allocate struct "
1449 "net_device for vec%d\n", n);
1450 goto out_free_device;
1453 dev->mtu = get_mtu(def);
1455 INIT_LIST_HEAD(&device->list);
1458 /* If this name ends up conflicting with an existing registered
1459 * netdevice, that is OK, register_netdev{,ice}() will notice this
1462 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1463 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1464 vp = netdev_priv(dev);
1466 /* sysfs register */
1467 if (!driver_registered) {
1468 platform_driver_register(¨_net_driver);
1469 driver_registered = 1;
1471 device->pdev.id = n;
1472 device->pdev.name = DRIVER_NAME;
1473 device->pdev.dev.release = vector_device_release;
1474 dev_set_drvdata(&device->pdev.dev, device);
1475 if (platform_device_register(&device->pdev))
1476 goto out_free_netdev;
1477 SET_NETDEV_DEV(dev, &device->pdev.dev);
1481 *vp = ((struct vector_private)
1483 .list = LIST_HEAD_INIT(vp->list),
1486 .options = get_transport_options(def),
1490 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1491 /* TODO - we need to calculate headroom so that ip header
1492 * is 16 byte aligned all the time
1494 .headroom = get_headroom(def),
1495 .form_header = NULL,
1496 .verify_header = NULL,
1497 .header_rxbuffer = NULL,
1498 .header_txbuffer = NULL,
1500 .rx_header_size = 0,
1501 .rexmit_scheduled = false,
1503 .transport_data = NULL,
1504 .in_write_poll = false,
1506 .req_size = get_req_size(def)
1509 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1510 tasklet_init(&vp->tx_poll, vector_tx_poll, (unsigned long)vp);
1511 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1513 timer_setup(&vp->tl, vector_timer_expire, 0);
1514 spin_lock_init(&vp->lock);
1517 dev->netdev_ops = &vector_netdev_ops;
1518 dev->ethtool_ops = &vector_net_ethtool_ops;
1519 dev->watchdog_timeo = (HZ >> 1);
1520 /* primary IRQ - fixme */
1521 dev->irq = 0; /* we will adjust this once opened */
1524 err = register_netdevice(dev);
1527 goto out_undo_user_init;
1529 spin_lock(&vector_devices_lock);
1530 list_add(&device->list, &vector_devices);
1531 spin_unlock(&vector_devices_lock);
1547 * Invoked late in the init
1550 static int __init vector_init(void)
1552 struct list_head *ele;
1553 struct vector_cmd_line_arg *def;
1554 struct arglist *parsed;
1556 list_for_each(ele, &vec_cmd_line) {
1557 def = list_entry(ele, struct vector_cmd_line_arg, list);
1558 parsed = uml_parse_vector_ifspec(def->arguments);
1560 vector_eth_configure(def->unit, parsed);
1566 /* Invoked at initial argument parsing, only stores
1567 * arguments until a proper vector_init is called
1571 static int __init vector_setup(char *str)
1575 struct vector_cmd_line_arg *new;
1577 err = vector_parse(str, &n, &str, &error);
1579 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1583 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1584 INIT_LIST_HEAD(&new->list);
1586 new->arguments = str;
1587 list_add_tail(&new->list, &vec_cmd_line);
1591 __setup("vec", vector_setup);
1592 __uml_help(vector_setup,
1593 "vec[0-9]+:<option>=<value>,<option>=<value>\n"
1594 " Configure a vector io network device.\n\n"
1597 late_initcall(vector_init);
1599 static struct mc_device vector_mc = {
1600 .list = LIST_HEAD_INIT(vector_mc.list),
1602 .config = vector_config,
1605 .remove = vector_remove,
1609 static int vector_inetaddr_event(
1610 struct notifier_block *this,
1611 unsigned long event,
1617 static struct notifier_block vector_inetaddr_notifier = {
1618 .notifier_call = vector_inetaddr_event,
1621 static void inet_register(void)
1623 register_inetaddr_notifier(&vector_inetaddr_notifier);
1626 static inline void inet_register(void)
1631 static int vector_net_init(void)
1633 mconsole_register_dev(&vector_mc);
1638 __initcall(vector_net_init);