1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2017 - 2019 Cambridge Greys Limited
4 * Copyright (C) 2011 - 2014 Cisco Systems Inc
5 * Copyright (C) 2001 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
6 * Copyright (C) 2001 Lennert Buytenhek (buytenh@gnu.org) and
7 * James Leu (jleu@mindspring.net).
8 * Copyright (C) 2001 by various other people who didn't put their name here.
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>
24 #include <linux/firmware.h>
26 #include <uapi/linux/filter.h>
32 #include "mconsole_kern.h"
33 #include "vector_user.h"
34 #include "vector_kern.h"
37 * Adapted from network devices with the following major changes:
38 * All transports are static - simplifies the code significantly
39 * Multiple FDs/IRQs per device
40 * Vector IO optionally used for read/write, falling back to legacy
41 * based on configuration and/or availability
42 * Configuration is no longer positional - L2TPv3 and GRE require up to
43 * 10 parameters, passing this as positional is not fit for purpose.
44 * Only socket transports are supported
48 #define DRIVER_NAME "uml-vector"
49 struct vector_cmd_line_arg {
50 struct list_head list;
55 struct vector_device {
56 struct list_head list;
57 struct net_device *dev;
58 struct platform_device pdev;
63 static LIST_HEAD(vec_cmd_line);
65 static DEFINE_SPINLOCK(vector_devices_lock);
66 static LIST_HEAD(vector_devices);
68 static int driver_registered;
70 static void vector_eth_configure(int n, struct arglist *def);
72 /* Argument accessors to set variables (and/or set default values)
73 * mtu, buffer sizing, default headroom, etc
76 #define DEFAULT_HEADROOM 2
77 #define SAFETY_MARGIN 32
78 #define DEFAULT_VECTOR_SIZE 64
79 #define TX_SMALL_PACKET 128
80 #define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
81 #define MAX_ITERATIONS 64
84 const char string[ETH_GSTRING_LEN];
85 } ethtool_stats_keys[] = {
87 { "rx_queue_running_average" },
89 { "tx_queue_running_average" },
90 { "rx_encaps_errors" },
91 { "tx_timeout_count" },
92 { "tx_restart_queue" },
94 { "tx_flow_control_xon" },
95 { "tx_flow_control_xoff" },
96 { "rx_csum_offload_good" },
97 { "rx_csum_offload_errors"},
102 #define VECTOR_NUM_STATS ARRAY_SIZE(ethtool_stats_keys)
104 static void vector_reset_stats(struct vector_private *vp)
106 vp->estats.rx_queue_max = 0;
107 vp->estats.rx_queue_running_average = 0;
108 vp->estats.tx_queue_max = 0;
109 vp->estats.tx_queue_running_average = 0;
110 vp->estats.rx_encaps_errors = 0;
111 vp->estats.tx_timeout_count = 0;
112 vp->estats.tx_restart_queue = 0;
113 vp->estats.tx_kicks = 0;
114 vp->estats.tx_flow_control_xon = 0;
115 vp->estats.tx_flow_control_xoff = 0;
116 vp->estats.sg_ok = 0;
117 vp->estats.sg_linearized = 0;
120 static int get_mtu(struct arglist *def)
122 char *mtu = uml_vector_fetch_arg(def, "mtu");
126 if (kstrtoul(mtu, 10, &result) == 0)
127 if ((result < (1 << 16) - 1) && (result >= 576))
130 return ETH_MAX_PACKET;
133 static char *get_bpf_file(struct arglist *def)
135 return uml_vector_fetch_arg(def, "bpffile");
138 static bool get_bpf_flash(struct arglist *def)
140 char *allow = uml_vector_fetch_arg(def, "bpfflash");
144 if (kstrtoul(allow, 10, &result) == 0)
150 static int get_depth(struct arglist *def)
152 char *mtu = uml_vector_fetch_arg(def, "depth");
156 if (kstrtoul(mtu, 10, &result) == 0)
159 return DEFAULT_VECTOR_SIZE;
162 static int get_headroom(struct arglist *def)
164 char *mtu = uml_vector_fetch_arg(def, "headroom");
168 if (kstrtoul(mtu, 10, &result) == 0)
171 return DEFAULT_HEADROOM;
174 static int get_req_size(struct arglist *def)
176 char *gro = uml_vector_fetch_arg(def, "gro");
180 if (kstrtoul(gro, 10, &result) == 0) {
185 return get_mtu(def) + ETH_HEADER_OTHER +
186 get_headroom(def) + SAFETY_MARGIN;
190 static int get_transport_options(struct arglist *def)
192 char *transport = uml_vector_fetch_arg(def, "transport");
193 char *vector = uml_vector_fetch_arg(def, "vec");
195 int vec_rx = VECTOR_RX;
196 int vec_tx = VECTOR_TX;
200 if (transport == NULL)
203 if (vector != NULL) {
204 if (kstrtoul(vector, 10, &parsed) == 0) {
212 if (get_bpf_flash(def))
213 result = VECTOR_BPF_FLASH;
215 if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
217 if (strncmp(transport, TRANS_HYBRID, TRANS_HYBRID_LEN) == 0)
218 return (result | vec_rx | VECTOR_BPF);
219 if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
220 return (result | vec_rx | vec_tx | VECTOR_QDISC_BYPASS);
221 return (result | vec_rx | vec_tx);
225 /* A mini-buffer for packet drop read
226 * All of our supported transports are datagram oriented and we always
227 * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
228 * than the packet size it still counts as full packet read and will
229 * clean the incoming stream to keep sigio/epoll happy
232 #define DROP_BUFFER_SIZE 32
234 static char *drop_buffer;
236 /* Array backed queues optimized for bulk enqueue/dequeue and
237 * 1:N (small values of N) or 1:1 enqueuer/dequeuer ratios.
238 * For more details and full design rationale see
239 * http://foswiki.cambridgegreys.com/Main/EatYourTailAndEnjoyIt
244 * Advance the mmsg queue head by n = advance. Resets the queue to
245 * maximum enqueue/dequeue-at-once capacity if possible. Called by
246 * dequeuers. Caller must hold the head_lock!
249 static int vector_advancehead(struct vector_queue *qi, int advance)
258 spin_lock(&qi->tail_lock);
259 qi->queue_depth -= advance;
261 /* we are at 0, use this to
262 * reset head and tail so we can use max size vectors
265 if (qi->queue_depth == 0) {
269 queue_depth = qi->queue_depth;
270 spin_unlock(&qi->tail_lock);
274 /* Advance the queue tail by n = advance.
275 * This is called by enqueuers which should hold the
279 static int vector_advancetail(struct vector_queue *qi, int advance)
286 spin_lock(&qi->head_lock);
287 qi->queue_depth += advance;
288 queue_depth = qi->queue_depth;
289 spin_unlock(&qi->head_lock);
293 static int prep_msg(struct vector_private *vp,
299 skb_frag_t *skb_frag;
301 nr_frags = skb_shinfo(skb)->nr_frags;
302 if (nr_frags > MAX_IOV_SIZE) {
303 if (skb_linearize(skb) != 0)
306 if (vp->header_size > 0) {
307 iov[iov_index].iov_len = vp->header_size;
308 vp->form_header(iov[iov_index].iov_base, skb, vp);
311 iov[iov_index].iov_base = skb->data;
313 iov[iov_index].iov_len = skb->len - skb->data_len;
316 iov[iov_index].iov_len = skb->len;
318 for (frag = 0; frag < nr_frags; frag++) {
319 skb_frag = &skb_shinfo(skb)->frags[frag];
320 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
321 iov[iov_index].iov_len = skb_frag_size(skb_frag);
329 * Generic vector enqueue with support for forming headers using transport
330 * specific callback. Allows GRE, L2TPv3, RAW and other transports
331 * to use a common enqueue procedure in vector mode
334 static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
336 struct vector_private *vp = netdev_priv(qi->dev);
339 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
342 spin_lock(&qi->tail_lock);
343 spin_lock(&qi->head_lock);
344 queue_depth = qi->queue_depth;
345 spin_unlock(&qi->head_lock);
348 packet_len = skb->len;
350 if (queue_depth < qi->max_depth) {
352 *(qi->skbuff_vector + qi->tail) = skb;
353 mmsg_vector += qi->tail;
354 iov_count = prep_msg(
357 mmsg_vector->msg_hdr.msg_iov
361 mmsg_vector->msg_hdr.msg_iovlen = iov_count;
362 mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
363 mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
364 queue_depth = vector_advancetail(qi, 1);
367 spin_unlock(&qi->tail_lock);
370 qi->dev->stats.tx_dropped++;
372 packet_len = skb->len;
373 dev_consume_skb_any(skb);
374 netdev_completed_queue(qi->dev, 1, packet_len);
376 spin_unlock(&qi->tail_lock);
380 static int consume_vector_skbs(struct vector_queue *qi, int count)
386 for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
387 skb = *(qi->skbuff_vector + skb_index);
388 /* mark as empty to ensure correct destruction if
391 bytes_compl += skb->len;
392 *(qi->skbuff_vector + skb_index) = NULL;
393 dev_consume_skb_any(skb);
395 qi->dev->stats.tx_bytes += bytes_compl;
396 qi->dev->stats.tx_packets += count;
397 netdev_completed_queue(qi->dev, count, bytes_compl);
398 return vector_advancehead(qi, count);
402 * Generic vector deque via sendmmsg with support for forming headers
403 * using transport specific callback. Allows GRE, L2TPv3, RAW and
404 * other transports to use a common dequeue procedure in vector mode
408 static int vector_send(struct vector_queue *qi)
410 struct vector_private *vp = netdev_priv(qi->dev);
411 struct mmsghdr *send_from;
412 int result = 0, send_len, queue_depth = qi->max_depth;
414 if (spin_trylock(&qi->head_lock)) {
415 if (spin_trylock(&qi->tail_lock)) {
416 /* update queue_depth to current value */
417 queue_depth = qi->queue_depth;
418 spin_unlock(&qi->tail_lock);
419 while (queue_depth > 0) {
420 /* Calculate the start of the vector */
421 send_len = queue_depth;
422 send_from = qi->mmsg_vector;
423 send_from += qi->head;
424 /* Adjust vector size if wraparound */
425 if (send_len + qi->head > qi->max_depth)
426 send_len = qi->max_depth - qi->head;
427 /* Try to TX as many packets as possible */
429 result = uml_vector_sendmmsg(
436 (result != send_len);
438 /* For some of the sendmmsg error scenarios
439 * we may end being unsure in the TX success
440 * for all packets. It is safer to declare
441 * them all TX-ed and blame the network.
445 netdev_err(vp->dev, "sendmmsg err=%i\n",
452 consume_vector_skbs(qi, result);
453 /* This is equivalent to an TX IRQ.
454 * Restart the upper layers to feed us
457 if (result > vp->estats.tx_queue_max)
458 vp->estats.tx_queue_max = result;
459 vp->estats.tx_queue_running_average =
460 (vp->estats.tx_queue_running_average + result) >> 1;
462 netif_trans_update(qi->dev);
463 netif_wake_queue(qi->dev);
464 /* if TX is busy, break out of the send loop,
465 * poll write IRQ will reschedule xmit for us
467 if (result != send_len) {
468 vp->estats.tx_restart_queue++;
473 spin_unlock(&qi->head_lock);
475 tasklet_schedule(&vp->tx_poll);
480 /* Queue destructor. Deliberately stateless so we can use
481 * it in queue cleanup if initialization fails.
484 static void destroy_queue(struct vector_queue *qi)
488 struct vector_private *vp = netdev_priv(qi->dev);
489 struct mmsghdr *mmsg_vector;
493 /* deallocate any skbuffs - we rely on any unused to be
496 if (qi->skbuff_vector != NULL) {
497 for (i = 0; i < qi->max_depth; i++) {
498 if (*(qi->skbuff_vector + i) != NULL)
499 dev_kfree_skb_any(*(qi->skbuff_vector + i));
501 kfree(qi->skbuff_vector);
503 /* deallocate matching IOV structures including header buffs */
504 if (qi->mmsg_vector != NULL) {
505 mmsg_vector = qi->mmsg_vector;
506 for (i = 0; i < qi->max_depth; i++) {
507 iov = mmsg_vector->msg_hdr.msg_iov;
509 if ((vp->header_size > 0) &&
510 (iov->iov_base != NULL))
511 kfree(iov->iov_base);
516 kfree(qi->mmsg_vector);
522 * Queue constructor. Create a queue with a given side.
524 static struct vector_queue *create_queue(
525 struct vector_private *vp,
530 struct vector_queue *result;
533 struct mmsghdr *mmsg_vector;
535 result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
538 result->max_depth = max_size;
539 result->dev = vp->dev;
540 result->mmsg_vector = kmalloc(
541 (sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
542 if (result->mmsg_vector == NULL)
544 result->skbuff_vector = kmalloc(
545 (sizeof(void *) * max_size), GFP_KERNEL);
546 if (result->skbuff_vector == NULL)
549 /* further failures can be handled safely by destroy_queue*/
551 mmsg_vector = result->mmsg_vector;
552 for (i = 0; i < max_size; i++) {
553 /* Clear all pointers - we use non-NULL as marking on
554 * what to free on destruction
556 *(result->skbuff_vector + i) = NULL;
557 mmsg_vector->msg_hdr.msg_iov = NULL;
560 mmsg_vector = result->mmsg_vector;
561 result->max_iov_frags = num_extra_frags;
562 for (i = 0; i < max_size; i++) {
563 if (vp->header_size > 0)
564 iov = kmalloc_array(3 + num_extra_frags,
565 sizeof(struct iovec),
569 iov = kmalloc_array(2 + num_extra_frags,
570 sizeof(struct iovec),
575 mmsg_vector->msg_hdr.msg_iov = iov;
576 mmsg_vector->msg_hdr.msg_iovlen = 1;
577 mmsg_vector->msg_hdr.msg_control = NULL;
578 mmsg_vector->msg_hdr.msg_controllen = 0;
579 mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
580 mmsg_vector->msg_hdr.msg_name = NULL;
581 mmsg_vector->msg_hdr.msg_namelen = 0;
582 if (vp->header_size > 0) {
583 iov->iov_base = kmalloc(header_size, GFP_KERNEL);
584 if (iov->iov_base == NULL)
586 iov->iov_len = header_size;
587 mmsg_vector->msg_hdr.msg_iovlen = 2;
590 iov->iov_base = NULL;
594 spin_lock_init(&result->head_lock);
595 spin_lock_init(&result->tail_lock);
596 result->queue_depth = 0;
601 kfree(result->mmsg_vector);
606 destroy_queue(result);
611 * We do not use the RX queue as a proper wraparound queue for now
612 * This is not necessary because the consumption via netif_rx()
613 * happens in-line. While we can try using the return code of
614 * netif_rx() for flow control there are no drivers doing this today.
615 * For this RX specific use we ignore the tail/head locks and
616 * just read into a prepared queue filled with skbuffs.
619 static struct sk_buff *prep_skb(
620 struct vector_private *vp,
621 struct user_msghdr *msg)
623 int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
624 struct sk_buff *result;
625 int iov_index = 0, len;
626 struct iovec *iov = msg->msg_iov;
627 int err, nr_frags, frag;
628 skb_frag_t *skb_frag;
630 if (vp->req_size <= linear)
634 result = alloc_skb_with_frags(
636 len - vp->max_packet,
641 if (vp->header_size > 0)
643 if (result == NULL) {
644 iov[iov_index].iov_base = NULL;
645 iov[iov_index].iov_len = 0;
648 skb_reserve(result, vp->headroom);
649 result->dev = vp->dev;
650 skb_put(result, vp->max_packet);
651 result->data_len = len - vp->max_packet;
652 result->len += len - vp->max_packet;
653 skb_reset_mac_header(result);
654 result->ip_summed = CHECKSUM_NONE;
655 iov[iov_index].iov_base = result->data;
656 iov[iov_index].iov_len = vp->max_packet;
659 nr_frags = skb_shinfo(result)->nr_frags;
660 for (frag = 0; frag < nr_frags; frag++) {
661 skb_frag = &skb_shinfo(result)->frags[frag];
662 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
663 if (iov[iov_index].iov_base != NULL)
664 iov[iov_index].iov_len = skb_frag_size(skb_frag);
666 iov[iov_index].iov_len = 0;
670 msg->msg_iovlen = iov_index;
675 /* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs*/
677 static void prep_queue_for_rx(struct vector_queue *qi)
679 struct vector_private *vp = netdev_priv(qi->dev);
680 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
681 void **skbuff_vector = qi->skbuff_vector;
684 if (qi->queue_depth == 0)
686 for (i = 0; i < qi->queue_depth; i++) {
687 /* it is OK if allocation fails - recvmmsg with NULL data in
688 * iov argument still performs an RX, just drops the packet
689 * This allows us stop faffing around with a "drop buffer"
692 *skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
699 static struct vector_device *find_device(int n)
701 struct vector_device *device;
702 struct list_head *ele;
704 spin_lock(&vector_devices_lock);
705 list_for_each(ele, &vector_devices) {
706 device = list_entry(ele, struct vector_device, list);
707 if (device->unit == n)
712 spin_unlock(&vector_devices_lock);
716 static int vector_parse(char *str, int *index_out, char **str_out,
724 while ((*str != ':') && (strlen(str) > 1))
727 *error_out = "Expected ':' after device number";
732 err = kstrtouint(start, 0, &n);
734 *error_out = "Bad device number";
739 if (find_device(n)) {
740 *error_out = "Device already configured";
749 static int vector_config(char *str, char **error_out)
753 struct arglist *parsed;
755 err = vector_parse(str, &n, ¶ms, error_out);
759 /* This string is broken up and the pieces used by the underlying
760 * driver. We should copy it to make sure things do not go wrong
764 params = kstrdup(params, GFP_KERNEL);
765 if (params == NULL) {
766 *error_out = "vector_config failed to strdup string";
770 parsed = uml_parse_vector_ifspec(params);
772 if (parsed == NULL) {
773 *error_out = "vector_config failed to parse parameters";
777 vector_eth_configure(n, parsed);
781 static int vector_id(char **str, int *start_out, int *end_out)
786 n = simple_strtoul(*str, &end, 0);
787 if ((*end != '\0') || (end == *str))
796 static int vector_remove(int n, char **error_out)
798 struct vector_device *vec_d;
799 struct net_device *dev;
800 struct vector_private *vp;
802 vec_d = find_device(n);
806 vp = netdev_priv(dev);
809 unregister_netdev(dev);
810 platform_device_unregister(&vec_d->pdev);
815 * There is no shared per-transport initialization code, so
816 * we will just initialize each interface one by one and
820 static struct platform_driver uml_net_driver = {
827 static void vector_device_release(struct device *dev)
829 struct vector_device *device = dev_get_drvdata(dev);
830 struct net_device *netdev = device->dev;
832 list_del(&device->list);
837 /* Bog standard recv using recvmsg - not used normally unless the user
838 * explicitly specifies not to use recvmmsg vector RX.
841 static int vector_legacy_rx(struct vector_private *vp)
844 struct user_msghdr hdr;
845 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
852 hdr.msg_iov = (struct iovec *) &iov;
853 hdr.msg_control = NULL;
854 hdr.msg_controllen = 0;
857 if (vp->header_size > 0) {
858 iov[0].iov_base = vp->header_rxbuffer;
859 iov[0].iov_len = vp->header_size;
862 skb = prep_skb(vp, &hdr);
865 /* Read a packet into drop_buffer and don't do
868 iov[iovpos].iov_base = drop_buffer;
869 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
871 vp->dev->stats.rx_dropped++;
874 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
881 if (pkt_len > vp->header_size) {
882 if (vp->header_size > 0) {
883 header_check = vp->verify_header(
884 vp->header_rxbuffer, skb, vp);
885 if (header_check < 0) {
886 dev_kfree_skb_irq(skb);
887 vp->dev->stats.rx_dropped++;
888 vp->estats.rx_encaps_errors++;
891 if (header_check > 0) {
892 vp->estats.rx_csum_offload_good++;
893 skb->ip_summed = CHECKSUM_UNNECESSARY;
896 pskb_trim(skb, pkt_len - vp->rx_header_size);
897 skb->protocol = eth_type_trans(skb, skb->dev);
898 vp->dev->stats.rx_bytes += skb->len;
899 vp->dev->stats.rx_packets++;
902 dev_kfree_skb_irq(skb);
909 * Packet at a time TX which falls back to vector TX if the
910 * underlying transport is busy.
915 static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
917 struct iovec iov[3 + MAX_IOV_SIZE];
918 int iov_count, pkt_len = 0;
920 iov[0].iov_base = vp->header_txbuffer;
921 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
926 pkt_len = uml_vector_writev(
928 (struct iovec *) &iov,
935 netif_trans_update(vp->dev);
936 netif_wake_queue(vp->dev);
939 vp->dev->stats.tx_bytes += skb->len;
940 vp->dev->stats.tx_packets++;
942 vp->dev->stats.tx_dropped++;
947 vp->dev->stats.tx_dropped++;
955 * Receive as many messages as we can in one call using the special
956 * mmsg vector matched to an skb vector which we prepared earlier.
959 static int vector_mmsg_rx(struct vector_private *vp)
962 struct vector_queue *qi = vp->rx_queue;
964 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
965 void **skbuff_vector = qi->skbuff_vector;
968 /* Refresh the vector and make sure it is with new skbs and the
969 * iovs are updated to point to them.
972 prep_queue_for_rx(qi);
974 /* Fire the Lazy Gun - get as many packets as we can in one go. */
976 packet_count = uml_vector_recvmmsg(
977 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
979 if (packet_count < 0)
982 if (packet_count <= 0)
985 /* We treat packet processing as enqueue, buffer refresh as dequeue
986 * The queue_depth tells us how many buffers have been used and how
987 * many do we need to prep the next time prep_queue_for_rx() is called.
990 qi->queue_depth = packet_count;
992 for (i = 0; i < packet_count; i++) {
993 skb = (*skbuff_vector);
994 if (mmsg_vector->msg_len > vp->header_size) {
995 if (vp->header_size > 0) {
996 header_check = vp->verify_header(
997 mmsg_vector->msg_hdr.msg_iov->iov_base,
1001 if (header_check < 0) {
1002 /* Overlay header failed to verify - discard.
1003 * We can actually keep this skb and reuse it,
1004 * but that will make the prep logic too
1007 dev_kfree_skb_irq(skb);
1008 vp->estats.rx_encaps_errors++;
1011 if (header_check > 0) {
1012 vp->estats.rx_csum_offload_good++;
1013 skb->ip_summed = CHECKSUM_UNNECESSARY;
1017 mmsg_vector->msg_len - vp->rx_header_size);
1018 skb->protocol = eth_type_trans(skb, skb->dev);
1020 * We do not need to lock on updating stats here
1021 * The interrupt loop is non-reentrant.
1023 vp->dev->stats.rx_bytes += skb->len;
1024 vp->dev->stats.rx_packets++;
1027 /* Overlay header too short to do anything - discard.
1028 * We can actually keep this skb and reuse it,
1029 * but that will make the prep logic too complex.
1032 dev_kfree_skb_irq(skb);
1034 (*skbuff_vector) = NULL;
1035 /* Move to the next buffer element */
1039 if (packet_count > 0) {
1040 if (vp->estats.rx_queue_max < packet_count)
1041 vp->estats.rx_queue_max = packet_count;
1042 vp->estats.rx_queue_running_average =
1043 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1045 return packet_count;
1048 static void vector_rx(struct vector_private *vp)
1053 if ((vp->options & VECTOR_RX) > 0)
1054 while (((err = vector_mmsg_rx(vp)) > 0) && (iter < MAX_ITERATIONS))
1057 while (((err = vector_legacy_rx(vp)) > 0) && (iter < MAX_ITERATIONS))
1059 if ((err != 0) && net_ratelimit())
1060 netdev_err(vp->dev, "vector_rx: error(%d)\n", err);
1061 if (iter == MAX_ITERATIONS)
1062 netdev_err(vp->dev, "vector_rx: device stuck, remote end may have closed the connection\n");
1065 static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1067 struct vector_private *vp = netdev_priv(dev);
1068 int queue_depth = 0;
1071 deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1072 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1073 deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1074 return NETDEV_TX_BUSY;
1077 if ((vp->options & VECTOR_TX) == 0) {
1079 return NETDEV_TX_OK;
1082 /* We do BQL only in the vector path, no point doing it in
1083 * packet at a time mode as there is no device queue
1086 netdev_sent_queue(vp->dev, skb->len);
1087 queue_depth = vector_enqueue(vp->tx_queue, skb);
1089 /* if the device queue is full, stop the upper layers and
1093 if (queue_depth >= vp->tx_queue->max_depth - 1) {
1094 vp->estats.tx_kicks++;
1095 netif_stop_queue(dev);
1096 vector_send(vp->tx_queue);
1097 return NETDEV_TX_OK;
1099 if (netdev_xmit_more()) {
1100 mod_timer(&vp->tl, vp->coalesce);
1101 return NETDEV_TX_OK;
1103 if (skb->len < TX_SMALL_PACKET) {
1104 vp->estats.tx_kicks++;
1105 vector_send(vp->tx_queue);
1107 tasklet_schedule(&vp->tx_poll);
1108 return NETDEV_TX_OK;
1111 static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1113 struct net_device *dev = dev_id;
1114 struct vector_private *vp = netdev_priv(dev);
1116 if (!netif_running(dev))
1123 static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1125 struct net_device *dev = dev_id;
1126 struct vector_private *vp = netdev_priv(dev);
1128 if (!netif_running(dev))
1130 /* We need to pay attention to it only if we got
1131 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1132 * we ignore it. In the future, it may be worth
1133 * it to improve the IRQ controller a bit to make
1134 * tweaking the IRQ mask less costly
1137 if (vp->in_write_poll)
1138 tasklet_schedule(&vp->tx_poll);
1145 static int vector_net_close(struct net_device *dev)
1147 struct vector_private *vp = netdev_priv(dev);
1148 unsigned long flags;
1150 netif_stop_queue(dev);
1153 if (vp->fds == NULL)
1156 /* Disable and free all IRQS */
1157 if (vp->rx_irq > 0) {
1158 um_free_irq(vp->rx_irq, dev);
1161 if (vp->tx_irq > 0) {
1162 um_free_irq(vp->tx_irq, dev);
1165 tasklet_kill(&vp->tx_poll);
1166 if (vp->fds->rx_fd > 0) {
1168 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1169 os_close_file(vp->fds->rx_fd);
1170 vp->fds->rx_fd = -1;
1172 if (vp->fds->tx_fd > 0) {
1173 os_close_file(vp->fds->tx_fd);
1174 vp->fds->tx_fd = -1;
1176 if (vp->bpf != NULL)
1177 kfree(vp->bpf->filter);
1180 kfree(vp->fds->remote_addr);
1181 kfree(vp->transport_data);
1182 kfree(vp->header_rxbuffer);
1183 kfree(vp->header_txbuffer);
1184 if (vp->rx_queue != NULL)
1185 destroy_queue(vp->rx_queue);
1186 if (vp->tx_queue != NULL)
1187 destroy_queue(vp->tx_queue);
1190 spin_lock_irqsave(&vp->lock, flags);
1192 vp->in_error = false;
1193 spin_unlock_irqrestore(&vp->lock, flags);
1199 static void vector_tx_poll(unsigned long data)
1201 struct vector_private *vp = (struct vector_private *)data;
1203 vp->estats.tx_kicks++;
1204 vector_send(vp->tx_queue);
1206 static void vector_reset_tx(struct work_struct *work)
1208 struct vector_private *vp =
1209 container_of(work, struct vector_private, reset_tx);
1210 netdev_reset_queue(vp->dev);
1211 netif_start_queue(vp->dev);
1212 netif_wake_queue(vp->dev);
1215 static int vector_net_open(struct net_device *dev)
1217 struct vector_private *vp = netdev_priv(dev);
1218 unsigned long flags;
1220 struct vector_device *vdevice;
1222 spin_lock_irqsave(&vp->lock, flags);
1224 spin_unlock_irqrestore(&vp->lock, flags);
1228 spin_unlock_irqrestore(&vp->lock, flags);
1230 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1232 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1234 if (vp->fds == NULL)
1237 if (build_transport_data(vp) < 0)
1240 if ((vp->options & VECTOR_RX) > 0) {
1241 vp->rx_queue = create_queue(
1243 get_depth(vp->parsed),
1247 vp->rx_queue->queue_depth = get_depth(vp->parsed);
1249 vp->header_rxbuffer = kmalloc(
1253 if (vp->header_rxbuffer == NULL)
1256 if ((vp->options & VECTOR_TX) > 0) {
1257 vp->tx_queue = create_queue(
1259 get_depth(vp->parsed),
1264 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1265 if (vp->header_txbuffer == NULL)
1270 err = um_request_irq(
1271 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1272 IRQ_READ, vector_rx_interrupt,
1273 IRQF_SHARED, dev->name, dev);
1275 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1279 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1280 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1281 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1283 /* WRITE IRQ - we need it only if we have vector TX */
1284 if ((vp->options & VECTOR_TX) > 0) {
1285 err = um_request_irq(
1286 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1287 IRQ_WRITE, vector_tx_interrupt,
1288 IRQF_SHARED, dev->name, dev);
1291 "vector_open: failed to get tx irq(%d)\n", err);
1295 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1296 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1299 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1300 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1301 vp->options |= VECTOR_BPF;
1303 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1304 vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1306 if (vp->bpf != NULL)
1307 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1309 netif_start_queue(dev);
1311 /* clear buffer - it can happen that the host side of the interface
1312 * is full when we get here. In this case, new data is never queued,
1313 * SIGIOs never arrive, and the net never works.
1318 vector_reset_stats(vp);
1319 vdevice = find_device(vp->unit);
1320 vdevice->opened = 1;
1322 if ((vp->options & VECTOR_TX) != 0)
1326 vector_net_close(dev);
1331 static void vector_net_set_multicast_list(struct net_device *dev)
1333 /* TODO: - we can do some BPF games here */
1337 static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1339 struct vector_private *vp = netdev_priv(dev);
1341 vp->estats.tx_timeout_count++;
1342 netif_trans_update(dev);
1343 schedule_work(&vp->reset_tx);
1346 static netdev_features_t vector_fix_features(struct net_device *dev,
1347 netdev_features_t features)
1349 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1353 static int vector_set_features(struct net_device *dev,
1354 netdev_features_t features)
1356 struct vector_private *vp = netdev_priv(dev);
1357 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1358 * no way to negotiate it on raw sockets, so we can change
1361 if (features & NETIF_F_GRO)
1362 /* All new frame buffers will be GRO-sized */
1363 vp->req_size = 65536;
1365 /* All new frame buffers will be normal sized */
1366 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1370 #ifdef CONFIG_NET_POLL_CONTROLLER
1371 static void vector_net_poll_controller(struct net_device *dev)
1373 disable_irq(dev->irq);
1374 vector_rx_interrupt(dev->irq, dev);
1375 enable_irq(dev->irq);
1379 static void vector_net_get_drvinfo(struct net_device *dev,
1380 struct ethtool_drvinfo *info)
1382 strlcpy(info->driver, DRIVER_NAME, sizeof(info->driver));
1385 static int vector_net_load_bpf_flash(struct net_device *dev,
1386 struct ethtool_flash *efl)
1388 struct vector_private *vp = netdev_priv(dev);
1389 struct vector_device *vdevice;
1390 const struct firmware *fw;
1393 if (!(vp->options & VECTOR_BPF_FLASH)) {
1394 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1398 spin_lock(&vp->lock);
1400 if (vp->bpf != NULL) {
1402 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1403 kfree(vp->bpf->filter);
1404 vp->bpf->filter = NULL;
1406 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_KERNEL);
1407 if (vp->bpf == NULL) {
1408 netdev_err(dev, "failed to allocate memory for firmware\n");
1413 vdevice = find_device(vp->unit);
1415 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1418 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_KERNEL);
1419 if (!vp->bpf->filter)
1422 vp->bpf->len = fw->size / sizeof(struct sock_filter);
1423 release_firmware(fw);
1426 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1428 spin_unlock(&vp->lock);
1433 release_firmware(fw);
1436 spin_unlock(&vp->lock);
1437 if (vp->bpf != NULL)
1438 kfree(vp->bpf->filter);
1444 static void vector_get_ringparam(struct net_device *netdev,
1445 struct ethtool_ringparam *ring)
1447 struct vector_private *vp = netdev_priv(netdev);
1449 ring->rx_max_pending = vp->rx_queue->max_depth;
1450 ring->tx_max_pending = vp->tx_queue->max_depth;
1451 ring->rx_pending = vp->rx_queue->max_depth;
1452 ring->tx_pending = vp->tx_queue->max_depth;
1455 static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1457 switch (stringset) {
1462 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1470 static int vector_get_sset_count(struct net_device *dev, int sset)
1476 return VECTOR_NUM_STATS;
1482 static void vector_get_ethtool_stats(struct net_device *dev,
1483 struct ethtool_stats *estats,
1486 struct vector_private *vp = netdev_priv(dev);
1488 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1491 static int vector_get_coalesce(struct net_device *netdev,
1492 struct ethtool_coalesce *ec)
1494 struct vector_private *vp = netdev_priv(netdev);
1496 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1500 static int vector_set_coalesce(struct net_device *netdev,
1501 struct ethtool_coalesce *ec)
1503 struct vector_private *vp = netdev_priv(netdev);
1505 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1506 if (vp->coalesce == 0)
1511 static const struct ethtool_ops vector_net_ethtool_ops = {
1512 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1513 .get_drvinfo = vector_net_get_drvinfo,
1514 .get_link = ethtool_op_get_link,
1515 .get_ts_info = ethtool_op_get_ts_info,
1516 .get_ringparam = vector_get_ringparam,
1517 .get_strings = vector_get_strings,
1518 .get_sset_count = vector_get_sset_count,
1519 .get_ethtool_stats = vector_get_ethtool_stats,
1520 .get_coalesce = vector_get_coalesce,
1521 .set_coalesce = vector_set_coalesce,
1522 .flash_device = vector_net_load_bpf_flash,
1526 static const struct net_device_ops vector_netdev_ops = {
1527 .ndo_open = vector_net_open,
1528 .ndo_stop = vector_net_close,
1529 .ndo_start_xmit = vector_net_start_xmit,
1530 .ndo_set_rx_mode = vector_net_set_multicast_list,
1531 .ndo_tx_timeout = vector_net_tx_timeout,
1532 .ndo_set_mac_address = eth_mac_addr,
1533 .ndo_validate_addr = eth_validate_addr,
1534 .ndo_fix_features = vector_fix_features,
1535 .ndo_set_features = vector_set_features,
1536 #ifdef CONFIG_NET_POLL_CONTROLLER
1537 .ndo_poll_controller = vector_net_poll_controller,
1542 static void vector_timer_expire(struct timer_list *t)
1544 struct vector_private *vp = from_timer(vp, t, tl);
1546 vp->estats.tx_kicks++;
1547 vector_send(vp->tx_queue);
1550 static void vector_eth_configure(
1555 struct vector_device *device;
1556 struct net_device *dev;
1557 struct vector_private *vp;
1560 device = kzalloc(sizeof(*device), GFP_KERNEL);
1561 if (device == NULL) {
1562 printk(KERN_ERR "eth_configure failed to allocate struct "
1566 dev = alloc_etherdev(sizeof(struct vector_private));
1568 printk(KERN_ERR "eth_configure: failed to allocate struct "
1569 "net_device for vec%d\n", n);
1570 goto out_free_device;
1573 dev->mtu = get_mtu(def);
1575 INIT_LIST_HEAD(&device->list);
1578 /* If this name ends up conflicting with an existing registered
1579 * netdevice, that is OK, register_netdev{,ice}() will notice this
1582 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1583 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1584 vp = netdev_priv(dev);
1586 /* sysfs register */
1587 if (!driver_registered) {
1588 platform_driver_register(¨_net_driver);
1589 driver_registered = 1;
1591 device->pdev.id = n;
1592 device->pdev.name = DRIVER_NAME;
1593 device->pdev.dev.release = vector_device_release;
1594 dev_set_drvdata(&device->pdev.dev, device);
1595 if (platform_device_register(&device->pdev))
1596 goto out_free_netdev;
1597 SET_NETDEV_DEV(dev, &device->pdev.dev);
1601 *vp = ((struct vector_private)
1603 .list = LIST_HEAD_INIT(vp->list),
1606 .options = get_transport_options(def),
1610 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1611 /* TODO - we need to calculate headroom so that ip header
1612 * is 16 byte aligned all the time
1614 .headroom = get_headroom(def),
1615 .form_header = NULL,
1616 .verify_header = NULL,
1617 .header_rxbuffer = NULL,
1618 .header_txbuffer = NULL,
1620 .rx_header_size = 0,
1621 .rexmit_scheduled = false,
1623 .transport_data = NULL,
1624 .in_write_poll = false,
1626 .req_size = get_req_size(def),
1631 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1632 tasklet_init(&vp->tx_poll, vector_tx_poll, (unsigned long)vp);
1633 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1635 timer_setup(&vp->tl, vector_timer_expire, 0);
1636 spin_lock_init(&vp->lock);
1639 dev->netdev_ops = &vector_netdev_ops;
1640 dev->ethtool_ops = &vector_net_ethtool_ops;
1641 dev->watchdog_timeo = (HZ >> 1);
1642 /* primary IRQ - fixme */
1643 dev->irq = 0; /* we will adjust this once opened */
1646 err = register_netdevice(dev);
1649 goto out_undo_user_init;
1651 spin_lock(&vector_devices_lock);
1652 list_add(&device->list, &vector_devices);
1653 spin_unlock(&vector_devices_lock);
1669 * Invoked late in the init
1672 static int __init vector_init(void)
1674 struct list_head *ele;
1675 struct vector_cmd_line_arg *def;
1676 struct arglist *parsed;
1678 list_for_each(ele, &vec_cmd_line) {
1679 def = list_entry(ele, struct vector_cmd_line_arg, list);
1680 parsed = uml_parse_vector_ifspec(def->arguments);
1682 vector_eth_configure(def->unit, parsed);
1688 /* Invoked at initial argument parsing, only stores
1689 * arguments until a proper vector_init is called
1693 static int __init vector_setup(char *str)
1697 struct vector_cmd_line_arg *new;
1699 err = vector_parse(str, &n, &str, &error);
1701 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1705 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1707 panic("%s: Failed to allocate %zu bytes\n", __func__,
1709 INIT_LIST_HEAD(&new->list);
1711 new->arguments = str;
1712 list_add_tail(&new->list, &vec_cmd_line);
1716 __setup("vec", vector_setup);
1717 __uml_help(vector_setup,
1718 "vec[0-9]+:<option>=<value>,<option>=<value>\n"
1719 " Configure a vector io network device.\n\n"
1722 late_initcall(vector_init);
1724 static struct mc_device vector_mc = {
1725 .list = LIST_HEAD_INIT(vector_mc.list),
1727 .config = vector_config,
1730 .remove = vector_remove,
1734 static int vector_inetaddr_event(
1735 struct notifier_block *this,
1736 unsigned long event,
1742 static struct notifier_block vector_inetaddr_notifier = {
1743 .notifier_call = vector_inetaddr_event,
1746 static void inet_register(void)
1748 register_inetaddr_notifier(&vector_inetaddr_notifier);
1751 static inline void inet_register(void)
1756 static int vector_net_init(void)
1758 mconsole_register_dev(&vector_mc);
1763 __initcall(vector_net_init);