Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
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.
9 */
10
11#include <linux/memblock.h>
12#include <linux/etherdevice.h>
13#include <linux/ethtool.h>
14#include <linux/inetdevice.h>
15#include <linux/init.h>
16#include <linux/list.h>
17#include <linux/netdevice.h>
18#include <linux/platform_device.h>
19#include <linux/rtnetlink.h>
20#include <linux/skbuff.h>
21#include <linux/slab.h>
22#include <linux/interrupt.h>
23#include <linux/firmware.h>
24#include <linux/fs.h>
25#include <asm/atomic.h>
26#include <uapi/linux/filter.h>
27#include <init.h>
28#include <irq_kern.h>
29#include <irq_user.h>
30#include <net_kern.h>
31#include <os.h>
32#include "mconsole_kern.h"
33#include "vector_user.h"
34#include "vector_kern.h"
35
36/*
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
45 */
46
47
48#define DRIVER_NAME "uml-vector"
49struct vector_cmd_line_arg {
50 struct list_head list;
51 int unit;
52 char *arguments;
53};
54
55struct vector_device {
56 struct list_head list;
57 struct net_device *dev;
58 struct platform_device pdev;
59 int unit;
60 int opened;
61};
62
63static LIST_HEAD(vec_cmd_line);
64
65static DEFINE_SPINLOCK(vector_devices_lock);
66static LIST_HEAD(vector_devices);
67
68static int driver_registered;
69
70static void vector_eth_configure(int n, struct arglist *def);
71static int vector_mmsg_rx(struct vector_private *vp, int budget);
72
73/* Argument accessors to set variables (and/or set default values)
74 * mtu, buffer sizing, default headroom, etc
75 */
76
77#define DEFAULT_HEADROOM 2
78#define SAFETY_MARGIN 32
79#define DEFAULT_VECTOR_SIZE 64
80#define TX_SMALL_PACKET 128
81#define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
82
83static const struct {
84 const char string[ETH_GSTRING_LEN];
85} ethtool_stats_keys[] = {
86 { "rx_queue_max" },
87 { "rx_queue_running_average" },
88 { "tx_queue_max" },
89 { "tx_queue_running_average" },
90 { "rx_encaps_errors" },
91 { "tx_timeout_count" },
92 { "tx_restart_queue" },
93 { "tx_kicks" },
94 { "tx_flow_control_xon" },
95 { "tx_flow_control_xoff" },
96 { "rx_csum_offload_good" },
97 { "rx_csum_offload_errors"},
98 { "sg_ok"},
99 { "sg_linearized"},
100};
101
102#define VECTOR_NUM_STATS ARRAY_SIZE(ethtool_stats_keys)
103
104static void vector_reset_stats(struct vector_private *vp)
105{
106 /* We reuse the existing queue locks for stats */
107
108 /* RX stats are modified with RX head_lock held
109 * in vector_poll.
110 */
111
112 spin_lock(&vp->rx_queue->head_lock);
113 vp->estats.rx_queue_max = 0;
114 vp->estats.rx_queue_running_average = 0;
115 vp->estats.rx_encaps_errors = 0;
116 vp->estats.sg_ok = 0;
117 vp->estats.sg_linearized = 0;
118 spin_unlock(&vp->rx_queue->head_lock);
119
120 /* TX stats are modified with TX head_lock held
121 * in vector_send.
122 */
123
124 spin_lock(&vp->tx_queue->head_lock);
125 vp->estats.tx_timeout_count = 0;
126 vp->estats.tx_restart_queue = 0;
127 vp->estats.tx_kicks = 0;
128 vp->estats.tx_flow_control_xon = 0;
129 vp->estats.tx_flow_control_xoff = 0;
130 vp->estats.tx_queue_max = 0;
131 vp->estats.tx_queue_running_average = 0;
132 spin_unlock(&vp->tx_queue->head_lock);
133}
134
135static int get_mtu(struct arglist *def)
136{
137 char *mtu = uml_vector_fetch_arg(def, "mtu");
138 long result;
139
140 if (mtu != NULL) {
141 if (kstrtoul(mtu, 10, &result) == 0)
142 if ((result < (1 << 16) - 1) && (result >= 576))
143 return result;
144 }
145 return ETH_MAX_PACKET;
146}
147
148static char *get_bpf_file(struct arglist *def)
149{
150 return uml_vector_fetch_arg(def, "bpffile");
151}
152
153static bool get_bpf_flash(struct arglist *def)
154{
155 char *allow = uml_vector_fetch_arg(def, "bpfflash");
156 long result;
157
158 if (allow != NULL) {
159 if (kstrtoul(allow, 10, &result) == 0)
160 return result > 0;
161 }
162 return false;
163}
164
165static int get_depth(struct arglist *def)
166{
167 char *mtu = uml_vector_fetch_arg(def, "depth");
168 long result;
169
170 if (mtu != NULL) {
171 if (kstrtoul(mtu, 10, &result) == 0)
172 return result;
173 }
174 return DEFAULT_VECTOR_SIZE;
175}
176
177static int get_headroom(struct arglist *def)
178{
179 char *mtu = uml_vector_fetch_arg(def, "headroom");
180 long result;
181
182 if (mtu != NULL) {
183 if (kstrtoul(mtu, 10, &result) == 0)
184 return result;
185 }
186 return DEFAULT_HEADROOM;
187}
188
189static int get_req_size(struct arglist *def)
190{
191 char *gro = uml_vector_fetch_arg(def, "gro");
192 long result;
193
194 if (gro != NULL) {
195 if (kstrtoul(gro, 10, &result) == 0) {
196 if (result > 0)
197 return 65536;
198 }
199 }
200 return get_mtu(def) + ETH_HEADER_OTHER +
201 get_headroom(def) + SAFETY_MARGIN;
202}
203
204
205static int get_transport_options(struct arglist *def)
206{
207 char *transport = uml_vector_fetch_arg(def, "transport");
208 char *vector = uml_vector_fetch_arg(def, "vec");
209
210 int vec_rx = VECTOR_RX;
211 int vec_tx = VECTOR_TX;
212 long parsed;
213 int result = 0;
214
215 if (transport == NULL)
216 return -EINVAL;
217
218 if (vector != NULL) {
219 if (kstrtoul(vector, 10, &parsed) == 0) {
220 if (parsed == 0) {
221 vec_rx = 0;
222 vec_tx = 0;
223 }
224 }
225 }
226
227 if (get_bpf_flash(def))
228 result = VECTOR_BPF_FLASH;
229
230 if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
231 return result;
232 if (strncmp(transport, TRANS_HYBRID, TRANS_HYBRID_LEN) == 0)
233 return (result | vec_rx | VECTOR_BPF);
234 if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
235 return (result | vec_rx | vec_tx | VECTOR_QDISC_BYPASS);
236 return (result | vec_rx | vec_tx);
237}
238
239
240/* A mini-buffer for packet drop read
241 * All of our supported transports are datagram oriented and we always
242 * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
243 * than the packet size it still counts as full packet read and will
244 * clean the incoming stream to keep sigio/epoll happy
245 */
246
247#define DROP_BUFFER_SIZE 32
248
249static char *drop_buffer;
250
251
252/*
253 * Advance the mmsg queue head by n = advance. Resets the queue to
254 * maximum enqueue/dequeue-at-once capacity if possible. Called by
255 * dequeuers. Caller must hold the head_lock!
256 */
257
258static int vector_advancehead(struct vector_queue *qi, int advance)
259{
260 qi->head =
261 (qi->head + advance)
262 % qi->max_depth;
263
264
265 atomic_sub(advance, &qi->queue_depth);
266 return atomic_read(&qi->queue_depth);
267}
268
269/* Advance the queue tail by n = advance.
270 * This is called by enqueuers which should hold the
271 * head lock already
272 */
273
274static int vector_advancetail(struct vector_queue *qi, int advance)
275{
276 qi->tail =
277 (qi->tail + advance)
278 % qi->max_depth;
279 atomic_add(advance, &qi->queue_depth);
280 return atomic_read(&qi->queue_depth);
281}
282
283static int prep_msg(struct vector_private *vp,
284 struct sk_buff *skb,
285 struct iovec *iov)
286{
287 int iov_index = 0;
288 int nr_frags, frag;
289 skb_frag_t *skb_frag;
290
291 nr_frags = skb_shinfo(skb)->nr_frags;
292 if (nr_frags > MAX_IOV_SIZE) {
293 if (skb_linearize(skb) != 0)
294 goto drop;
295 }
296 if (vp->header_size > 0) {
297 iov[iov_index].iov_len = vp->header_size;
298 vp->form_header(iov[iov_index].iov_base, skb, vp);
299 iov_index++;
300 }
301 iov[iov_index].iov_base = skb->data;
302 if (nr_frags > 0) {
303 iov[iov_index].iov_len = skb->len - skb->data_len;
304 vp->estats.sg_ok++;
305 } else
306 iov[iov_index].iov_len = skb->len;
307 iov_index++;
308 for (frag = 0; frag < nr_frags; frag++) {
309 skb_frag = &skb_shinfo(skb)->frags[frag];
310 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
311 iov[iov_index].iov_len = skb_frag_size(skb_frag);
312 iov_index++;
313 }
314 return iov_index;
315drop:
316 return -1;
317}
318/*
319 * Generic vector enqueue with support for forming headers using transport
320 * specific callback. Allows GRE, L2TPv3, RAW and other transports
321 * to use a common enqueue procedure in vector mode
322 */
323
324static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
325{
326 struct vector_private *vp = netdev_priv(qi->dev);
327 int queue_depth;
328 int packet_len;
329 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
330 int iov_count;
331
332 spin_lock(&qi->tail_lock);
333 queue_depth = atomic_read(&qi->queue_depth);
334
335 if (skb)
336 packet_len = skb->len;
337
338 if (queue_depth < qi->max_depth) {
339
340 *(qi->skbuff_vector + qi->tail) = skb;
341 mmsg_vector += qi->tail;
342 iov_count = prep_msg(
343 vp,
344 skb,
345 mmsg_vector->msg_hdr.msg_iov
346 );
347 if (iov_count < 1)
348 goto drop;
349 mmsg_vector->msg_hdr.msg_iovlen = iov_count;
350 mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
351 mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
352 wmb(); /* Make the packet visible to the NAPI poll thread */
353 queue_depth = vector_advancetail(qi, 1);
354 } else
355 goto drop;
356 spin_unlock(&qi->tail_lock);
357 return queue_depth;
358drop:
359 qi->dev->stats.tx_dropped++;
360 if (skb != NULL) {
361 packet_len = skb->len;
362 dev_consume_skb_any(skb);
363 netdev_completed_queue(qi->dev, 1, packet_len);
364 }
365 spin_unlock(&qi->tail_lock);
366 return queue_depth;
367}
368
369static int consume_vector_skbs(struct vector_queue *qi, int count)
370{
371 struct sk_buff *skb;
372 int skb_index;
373 int bytes_compl = 0;
374
375 for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
376 skb = *(qi->skbuff_vector + skb_index);
377 /* mark as empty to ensure correct destruction if
378 * needed
379 */
380 bytes_compl += skb->len;
381 *(qi->skbuff_vector + skb_index) = NULL;
382 dev_consume_skb_any(skb);
383 }
384 qi->dev->stats.tx_bytes += bytes_compl;
385 qi->dev->stats.tx_packets += count;
386 netdev_completed_queue(qi->dev, count, bytes_compl);
387 return vector_advancehead(qi, count);
388}
389
390/*
391 * Generic vector dequeue via sendmmsg with support for forming headers
392 * using transport specific callback. Allows GRE, L2TPv3, RAW and
393 * other transports to use a common dequeue procedure in vector mode
394 */
395
396
397static int vector_send(struct vector_queue *qi)
398{
399 struct vector_private *vp = netdev_priv(qi->dev);
400 struct mmsghdr *send_from;
401 int result = 0, send_len;
402
403 if (spin_trylock(&qi->head_lock)) {
404 /* update queue_depth to current value */
405 while (atomic_read(&qi->queue_depth) > 0) {
406 /* Calculate the start of the vector */
407 send_len = atomic_read(&qi->queue_depth);
408 send_from = qi->mmsg_vector;
409 send_from += qi->head;
410 /* Adjust vector size if wraparound */
411 if (send_len + qi->head > qi->max_depth)
412 send_len = qi->max_depth - qi->head;
413 /* Try to TX as many packets as possible */
414 if (send_len > 0) {
415 result = uml_vector_sendmmsg(
416 vp->fds->tx_fd,
417 send_from,
418 send_len,
419 0
420 );
421 vp->in_write_poll =
422 (result != send_len);
423 }
424 /* For some of the sendmmsg error scenarios
425 * we may end being unsure in the TX success
426 * for all packets. It is safer to declare
427 * them all TX-ed and blame the network.
428 */
429 if (result < 0) {
430 if (net_ratelimit())
431 netdev_err(vp->dev, "sendmmsg err=%i\n",
432 result);
433 vp->in_error = true;
434 result = send_len;
435 }
436 if (result > 0) {
437 consume_vector_skbs(qi, result);
438 /* This is equivalent to an TX IRQ.
439 * Restart the upper layers to feed us
440 * more packets.
441 */
442 if (result > vp->estats.tx_queue_max)
443 vp->estats.tx_queue_max = result;
444 vp->estats.tx_queue_running_average =
445 (vp->estats.tx_queue_running_average + result) >> 1;
446 }
447 netif_wake_queue(qi->dev);
448 /* if TX is busy, break out of the send loop,
449 * poll write IRQ will reschedule xmit for us.
450 */
451 if (result != send_len) {
452 vp->estats.tx_restart_queue++;
453 break;
454 }
455 }
456 spin_unlock(&qi->head_lock);
457 }
458 return atomic_read(&qi->queue_depth);
459}
460
461/* Queue destructor. Deliberately stateless so we can use
462 * it in queue cleanup if initialization fails.
463 */
464
465static void destroy_queue(struct vector_queue *qi)
466{
467 int i;
468 struct iovec *iov;
469 struct vector_private *vp = netdev_priv(qi->dev);
470 struct mmsghdr *mmsg_vector;
471
472 if (qi == NULL)
473 return;
474 /* deallocate any skbuffs - we rely on any unused to be
475 * set to NULL.
476 */
477 if (qi->skbuff_vector != NULL) {
478 for (i = 0; i < qi->max_depth; i++) {
479 if (*(qi->skbuff_vector + i) != NULL)
480 dev_kfree_skb_any(*(qi->skbuff_vector + i));
481 }
482 kfree(qi->skbuff_vector);
483 }
484 /* deallocate matching IOV structures including header buffs */
485 if (qi->mmsg_vector != NULL) {
486 mmsg_vector = qi->mmsg_vector;
487 for (i = 0; i < qi->max_depth; i++) {
488 iov = mmsg_vector->msg_hdr.msg_iov;
489 if (iov != NULL) {
490 if ((vp->header_size > 0) &&
491 (iov->iov_base != NULL))
492 kfree(iov->iov_base);
493 kfree(iov);
494 }
495 mmsg_vector++;
496 }
497 kfree(qi->mmsg_vector);
498 }
499 kfree(qi);
500}
501
502/*
503 * Queue constructor. Create a queue with a given side.
504 */
505static struct vector_queue *create_queue(
506 struct vector_private *vp,
507 int max_size,
508 int header_size,
509 int num_extra_frags)
510{
511 struct vector_queue *result;
512 int i;
513 struct iovec *iov;
514 struct mmsghdr *mmsg_vector;
515
516 result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
517 if (result == NULL)
518 return NULL;
519 result->max_depth = max_size;
520 result->dev = vp->dev;
521 result->mmsg_vector = kmalloc(
522 (sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
523 if (result->mmsg_vector == NULL)
524 goto out_mmsg_fail;
525 result->skbuff_vector = kmalloc(
526 (sizeof(void *) * max_size), GFP_KERNEL);
527 if (result->skbuff_vector == NULL)
528 goto out_skb_fail;
529
530 /* further failures can be handled safely by destroy_queue*/
531
532 mmsg_vector = result->mmsg_vector;
533 for (i = 0; i < max_size; i++) {
534 /* Clear all pointers - we use non-NULL as marking on
535 * what to free on destruction
536 */
537 *(result->skbuff_vector + i) = NULL;
538 mmsg_vector->msg_hdr.msg_iov = NULL;
539 mmsg_vector++;
540 }
541 mmsg_vector = result->mmsg_vector;
542 result->max_iov_frags = num_extra_frags;
543 for (i = 0; i < max_size; i++) {
544 if (vp->header_size > 0)
545 iov = kmalloc_array(3 + num_extra_frags,
546 sizeof(struct iovec),
547 GFP_KERNEL
548 );
549 else
550 iov = kmalloc_array(2 + num_extra_frags,
551 sizeof(struct iovec),
552 GFP_KERNEL
553 );
554 if (iov == NULL)
555 goto out_fail;
556 mmsg_vector->msg_hdr.msg_iov = iov;
557 mmsg_vector->msg_hdr.msg_iovlen = 1;
558 mmsg_vector->msg_hdr.msg_control = NULL;
559 mmsg_vector->msg_hdr.msg_controllen = 0;
560 mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
561 mmsg_vector->msg_hdr.msg_name = NULL;
562 mmsg_vector->msg_hdr.msg_namelen = 0;
563 if (vp->header_size > 0) {
564 iov->iov_base = kmalloc(header_size, GFP_KERNEL);
565 if (iov->iov_base == NULL)
566 goto out_fail;
567 iov->iov_len = header_size;
568 mmsg_vector->msg_hdr.msg_iovlen = 2;
569 iov++;
570 }
571 iov->iov_base = NULL;
572 iov->iov_len = 0;
573 mmsg_vector++;
574 }
575 spin_lock_init(&result->head_lock);
576 spin_lock_init(&result->tail_lock);
577 atomic_set(&result->queue_depth, 0);
578 result->head = 0;
579 result->tail = 0;
580 return result;
581out_skb_fail:
582 kfree(result->mmsg_vector);
583out_mmsg_fail:
584 kfree(result);
585 return NULL;
586out_fail:
587 destroy_queue(result);
588 return NULL;
589}
590
591/*
592 * We do not use the RX queue as a proper wraparound queue for now
593 * This is not necessary because the consumption via napi_gro_receive()
594 * happens in-line. While we can try using the return code of
595 * netif_rx() for flow control there are no drivers doing this today.
596 * For this RX specific use we ignore the tail/head locks and
597 * just read into a prepared queue filled with skbuffs.
598 */
599
600static struct sk_buff *prep_skb(
601 struct vector_private *vp,
602 struct user_msghdr *msg)
603{
604 int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
605 struct sk_buff *result;
606 int iov_index = 0, len;
607 struct iovec *iov = msg->msg_iov;
608 int err, nr_frags, frag;
609 skb_frag_t *skb_frag;
610
611 if (vp->req_size <= linear)
612 len = linear;
613 else
614 len = vp->req_size;
615 result = alloc_skb_with_frags(
616 linear,
617 len - vp->max_packet,
618 3,
619 &err,
620 GFP_ATOMIC
621 );
622 if (vp->header_size > 0)
623 iov_index++;
624 if (result == NULL) {
625 iov[iov_index].iov_base = NULL;
626 iov[iov_index].iov_len = 0;
627 goto done;
628 }
629 skb_reserve(result, vp->headroom);
630 result->dev = vp->dev;
631 skb_put(result, vp->max_packet);
632 result->data_len = len - vp->max_packet;
633 result->len += len - vp->max_packet;
634 skb_reset_mac_header(result);
635 result->ip_summed = CHECKSUM_NONE;
636 iov[iov_index].iov_base = result->data;
637 iov[iov_index].iov_len = vp->max_packet;
638 iov_index++;
639
640 nr_frags = skb_shinfo(result)->nr_frags;
641 for (frag = 0; frag < nr_frags; frag++) {
642 skb_frag = &skb_shinfo(result)->frags[frag];
643 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
644 if (iov[iov_index].iov_base != NULL)
645 iov[iov_index].iov_len = skb_frag_size(skb_frag);
646 else
647 iov[iov_index].iov_len = 0;
648 iov_index++;
649 }
650done:
651 msg->msg_iovlen = iov_index;
652 return result;
653}
654
655
656/* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs */
657
658static void prep_queue_for_rx(struct vector_queue *qi)
659{
660 struct vector_private *vp = netdev_priv(qi->dev);
661 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
662 void **skbuff_vector = qi->skbuff_vector;
663 int i, queue_depth;
664
665 queue_depth = atomic_read(&qi->queue_depth);
666
667 if (queue_depth == 0)
668 return;
669
670 /* RX is always emptied 100% during each cycle, so we do not
671 * have to do the tail wraparound math for it.
672 */
673
674 qi->head = qi->tail = 0;
675
676 for (i = 0; i < queue_depth; i++) {
677 /* it is OK if allocation fails - recvmmsg with NULL data in
678 * iov argument still performs an RX, just drops the packet
679 * This allows us stop faffing around with a "drop buffer"
680 */
681
682 *skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
683 skbuff_vector++;
684 mmsg_vector++;
685 }
686 atomic_set(&qi->queue_depth, 0);
687}
688
689static struct vector_device *find_device(int n)
690{
691 struct vector_device *device;
692 struct list_head *ele;
693
694 spin_lock(&vector_devices_lock);
695 list_for_each(ele, &vector_devices) {
696 device = list_entry(ele, struct vector_device, list);
697 if (device->unit == n)
698 goto out;
699 }
700 device = NULL;
701 out:
702 spin_unlock(&vector_devices_lock);
703 return device;
704}
705
706static int vector_parse(char *str, int *index_out, char **str_out,
707 char **error_out)
708{
709 int n, err;
710 char *start = str;
711
712 while ((*str != ':') && (strlen(str) > 1))
713 str++;
714 if (*str != ':') {
715 *error_out = "Expected ':' after device number";
716 return -EINVAL;
717 }
718 *str = '\0';
719
720 err = kstrtouint(start, 0, &n);
721 if (err < 0) {
722 *error_out = "Bad device number";
723 return err;
724 }
725
726 str++;
727 if (find_device(n)) {
728 *error_out = "Device already configured";
729 return -EINVAL;
730 }
731
732 *index_out = n;
733 *str_out = str;
734 return 0;
735}
736
737static int vector_config(char *str, char **error_out)
738{
739 int err, n;
740 char *params;
741 struct arglist *parsed;
742
743 err = vector_parse(str, &n, ¶ms, error_out);
744 if (err != 0)
745 return err;
746
747 /* This string is broken up and the pieces used by the underlying
748 * driver. We should copy it to make sure things do not go wrong
749 * later.
750 */
751
752 params = kstrdup(params, GFP_KERNEL);
753 if (params == NULL) {
754 *error_out = "vector_config failed to strdup string";
755 return -ENOMEM;
756 }
757
758 parsed = uml_parse_vector_ifspec(params);
759
760 if (parsed == NULL) {
761 *error_out = "vector_config failed to parse parameters";
762 kfree(params);
763 return -EINVAL;
764 }
765
766 vector_eth_configure(n, parsed);
767 return 0;
768}
769
770static int vector_id(char **str, int *start_out, int *end_out)
771{
772 char *end;
773 int n;
774
775 n = simple_strtoul(*str, &end, 0);
776 if ((*end != '\0') || (end == *str))
777 return -1;
778
779 *start_out = n;
780 *end_out = n;
781 *str = end;
782 return n;
783}
784
785static int vector_remove(int n, char **error_out)
786{
787 struct vector_device *vec_d;
788 struct net_device *dev;
789 struct vector_private *vp;
790
791 vec_d = find_device(n);
792 if (vec_d == NULL)
793 return -ENODEV;
794 dev = vec_d->dev;
795 vp = netdev_priv(dev);
796 if (vp->fds != NULL)
797 return -EBUSY;
798 unregister_netdev(dev);
799 platform_device_unregister(&vec_d->pdev);
800 return 0;
801}
802
803/*
804 * There is no shared per-transport initialization code, so
805 * we will just initialize each interface one by one and
806 * add them to a list
807 */
808
809static struct platform_driver uml_net_driver = {
810 .driver = {
811 .name = DRIVER_NAME,
812 },
813};
814
815
816static void vector_device_release(struct device *dev)
817{
818 struct vector_device *device =
819 container_of(dev, struct vector_device, pdev.dev);
820 struct net_device *netdev = device->dev;
821
822 list_del(&device->list);
823 kfree(device);
824 free_netdev(netdev);
825}
826
827/* Bog standard recv using recvmsg - not used normally unless the user
828 * explicitly specifies not to use recvmmsg vector RX.
829 */
830
831static int vector_legacy_rx(struct vector_private *vp)
832{
833 int pkt_len;
834 struct user_msghdr hdr;
835 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
836 int iovpos = 0;
837 struct sk_buff *skb;
838 int header_check;
839
840 hdr.msg_name = NULL;
841 hdr.msg_namelen = 0;
842 hdr.msg_iov = (struct iovec *) &iov;
843 hdr.msg_control = NULL;
844 hdr.msg_controllen = 0;
845 hdr.msg_flags = 0;
846
847 if (vp->header_size > 0) {
848 iov[0].iov_base = vp->header_rxbuffer;
849 iov[0].iov_len = vp->header_size;
850 }
851
852 skb = prep_skb(vp, &hdr);
853
854 if (skb == NULL) {
855 /* Read a packet into drop_buffer and don't do
856 * anything with it.
857 */
858 iov[iovpos].iov_base = drop_buffer;
859 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
860 hdr.msg_iovlen = 1;
861 vp->dev->stats.rx_dropped++;
862 }
863
864 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
865 if (pkt_len < 0) {
866 vp->in_error = true;
867 return pkt_len;
868 }
869
870 if (skb != NULL) {
871 if (pkt_len > vp->header_size) {
872 if (vp->header_size > 0) {
873 header_check = vp->verify_header(
874 vp->header_rxbuffer, skb, vp);
875 if (header_check < 0) {
876 dev_kfree_skb_irq(skb);
877 vp->dev->stats.rx_dropped++;
878 vp->estats.rx_encaps_errors++;
879 return 0;
880 }
881 if (header_check > 0) {
882 vp->estats.rx_csum_offload_good++;
883 skb->ip_summed = CHECKSUM_UNNECESSARY;
884 }
885 }
886 pskb_trim(skb, pkt_len - vp->rx_header_size);
887 skb->protocol = eth_type_trans(skb, skb->dev);
888 vp->dev->stats.rx_bytes += skb->len;
889 vp->dev->stats.rx_packets++;
890 napi_gro_receive(&vp->napi, skb);
891 } else {
892 dev_kfree_skb_irq(skb);
893 }
894 }
895 return pkt_len;
896}
897
898/*
899 * Packet at a time TX which falls back to vector TX if the
900 * underlying transport is busy.
901 */
902
903
904
905static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
906{
907 struct iovec iov[3 + MAX_IOV_SIZE];
908 int iov_count, pkt_len = 0;
909
910 iov[0].iov_base = vp->header_txbuffer;
911 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
912
913 if (iov_count < 1)
914 goto drop;
915
916 pkt_len = uml_vector_writev(
917 vp->fds->tx_fd,
918 (struct iovec *) &iov,
919 iov_count
920 );
921
922 if (pkt_len < 0)
923 goto drop;
924
925 netif_trans_update(vp->dev);
926 netif_wake_queue(vp->dev);
927
928 if (pkt_len > 0) {
929 vp->dev->stats.tx_bytes += skb->len;
930 vp->dev->stats.tx_packets++;
931 } else {
932 vp->dev->stats.tx_dropped++;
933 }
934 consume_skb(skb);
935 return pkt_len;
936drop:
937 vp->dev->stats.tx_dropped++;
938 consume_skb(skb);
939 if (pkt_len < 0)
940 vp->in_error = true;
941 return pkt_len;
942}
943
944/*
945 * Receive as many messages as we can in one call using the special
946 * mmsg vector matched to an skb vector which we prepared earlier.
947 */
948
949static int vector_mmsg_rx(struct vector_private *vp, int budget)
950{
951 int packet_count, i;
952 struct vector_queue *qi = vp->rx_queue;
953 struct sk_buff *skb;
954 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
955 void **skbuff_vector = qi->skbuff_vector;
956 int header_check;
957
958 /* Refresh the vector and make sure it is with new skbs and the
959 * iovs are updated to point to them.
960 */
961
962 prep_queue_for_rx(qi);
963
964 /* Fire the Lazy Gun - get as many packets as we can in one go. */
965
966 if (budget > qi->max_depth)
967 budget = qi->max_depth;
968
969 packet_count = uml_vector_recvmmsg(
970 vp->fds->rx_fd, qi->mmsg_vector, budget, 0);
971
972 if (packet_count < 0)
973 vp->in_error = true;
974
975 if (packet_count <= 0)
976 return packet_count;
977
978 /* We treat packet processing as enqueue, buffer refresh as dequeue
979 * The queue_depth tells us how many buffers have been used and how
980 * many do we need to prep the next time prep_queue_for_rx() is called.
981 */
982
983 atomic_add(packet_count, &qi->queue_depth);
984
985 for (i = 0; i < packet_count; i++) {
986 skb = (*skbuff_vector);
987 if (mmsg_vector->msg_len > vp->header_size) {
988 if (vp->header_size > 0) {
989 header_check = vp->verify_header(
990 mmsg_vector->msg_hdr.msg_iov->iov_base,
991 skb,
992 vp
993 );
994 if (header_check < 0) {
995 /* Overlay header failed to verify - discard.
996 * We can actually keep this skb and reuse it,
997 * but that will make the prep logic too
998 * complex.
999 */
1000 dev_kfree_skb_irq(skb);
1001 vp->estats.rx_encaps_errors++;
1002 continue;
1003 }
1004 if (header_check > 0) {
1005 vp->estats.rx_csum_offload_good++;
1006 skb->ip_summed = CHECKSUM_UNNECESSARY;
1007 }
1008 }
1009 pskb_trim(skb,
1010 mmsg_vector->msg_len - vp->rx_header_size);
1011 skb->protocol = eth_type_trans(skb, skb->dev);
1012 /*
1013 * We do not need to lock on updating stats here
1014 * The interrupt loop is non-reentrant.
1015 */
1016 vp->dev->stats.rx_bytes += skb->len;
1017 vp->dev->stats.rx_packets++;
1018 napi_gro_receive(&vp->napi, skb);
1019 } else {
1020 /* Overlay header too short to do anything - discard.
1021 * We can actually keep this skb and reuse it,
1022 * but that will make the prep logic too complex.
1023 */
1024 if (skb != NULL)
1025 dev_kfree_skb_irq(skb);
1026 }
1027 (*skbuff_vector) = NULL;
1028 /* Move to the next buffer element */
1029 mmsg_vector++;
1030 skbuff_vector++;
1031 }
1032 if (packet_count > 0) {
1033 if (vp->estats.rx_queue_max < packet_count)
1034 vp->estats.rx_queue_max = packet_count;
1035 vp->estats.rx_queue_running_average =
1036 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1037 }
1038 return packet_count;
1039}
1040
1041static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1042{
1043 struct vector_private *vp = netdev_priv(dev);
1044 int queue_depth = 0;
1045
1046 if (vp->in_error) {
1047 deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1048 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1049 deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1050 return NETDEV_TX_BUSY;
1051 }
1052
1053 if ((vp->options & VECTOR_TX) == 0) {
1054 writev_tx(vp, skb);
1055 return NETDEV_TX_OK;
1056 }
1057
1058 /* We do BQL only in the vector path, no point doing it in
1059 * packet at a time mode as there is no device queue
1060 */
1061
1062 netdev_sent_queue(vp->dev, skb->len);
1063 queue_depth = vector_enqueue(vp->tx_queue, skb);
1064
1065 if (queue_depth < vp->tx_queue->max_depth && netdev_xmit_more()) {
1066 mod_timer(&vp->tl, vp->coalesce);
1067 return NETDEV_TX_OK;
1068 } else {
1069 queue_depth = vector_send(vp->tx_queue);
1070 if (queue_depth > 0)
1071 napi_schedule(&vp->napi);
1072 }
1073
1074 return NETDEV_TX_OK;
1075}
1076
1077static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1078{
1079 struct net_device *dev = dev_id;
1080 struct vector_private *vp = netdev_priv(dev);
1081
1082 if (!netif_running(dev))
1083 return IRQ_NONE;
1084 napi_schedule(&vp->napi);
1085 return IRQ_HANDLED;
1086
1087}
1088
1089static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1090{
1091 struct net_device *dev = dev_id;
1092 struct vector_private *vp = netdev_priv(dev);
1093
1094 if (!netif_running(dev))
1095 return IRQ_NONE;
1096 /* We need to pay attention to it only if we got
1097 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1098 * we ignore it. In the future, it may be worth
1099 * it to improve the IRQ controller a bit to make
1100 * tweaking the IRQ mask less costly
1101 */
1102
1103 napi_schedule(&vp->napi);
1104 return IRQ_HANDLED;
1105
1106}
1107
1108static int irq_rr;
1109
1110static int vector_net_close(struct net_device *dev)
1111{
1112 struct vector_private *vp = netdev_priv(dev);
1113
1114 netif_stop_queue(dev);
1115 del_timer(&vp->tl);
1116
1117 vp->opened = false;
1118
1119 if (vp->fds == NULL)
1120 return 0;
1121
1122 /* Disable and free all IRQS */
1123 if (vp->rx_irq > 0) {
1124 um_free_irq(vp->rx_irq, dev);
1125 vp->rx_irq = 0;
1126 }
1127 if (vp->tx_irq > 0) {
1128 um_free_irq(vp->tx_irq, dev);
1129 vp->tx_irq = 0;
1130 }
1131 napi_disable(&vp->napi);
1132 netif_napi_del(&vp->napi);
1133 if (vp->fds->rx_fd > 0) {
1134 if (vp->bpf)
1135 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1136 os_close_file(vp->fds->rx_fd);
1137 vp->fds->rx_fd = -1;
1138 }
1139 if (vp->fds->tx_fd > 0) {
1140 os_close_file(vp->fds->tx_fd);
1141 vp->fds->tx_fd = -1;
1142 }
1143 if (vp->bpf != NULL)
1144 kfree(vp->bpf->filter);
1145 kfree(vp->bpf);
1146 vp->bpf = NULL;
1147 kfree(vp->fds->remote_addr);
1148 kfree(vp->transport_data);
1149 kfree(vp->header_rxbuffer);
1150 kfree(vp->header_txbuffer);
1151 if (vp->rx_queue != NULL)
1152 destroy_queue(vp->rx_queue);
1153 if (vp->tx_queue != NULL)
1154 destroy_queue(vp->tx_queue);
1155 kfree(vp->fds);
1156 vp->fds = NULL;
1157 vp->in_error = false;
1158 return 0;
1159}
1160
1161static int vector_poll(struct napi_struct *napi, int budget)
1162{
1163 struct vector_private *vp = container_of(napi, struct vector_private, napi);
1164 int work_done = 0;
1165 int err;
1166 bool tx_enqueued = false;
1167
1168 if ((vp->options & VECTOR_TX) != 0)
1169 tx_enqueued = (vector_send(vp->tx_queue) > 0);
1170 spin_lock(&vp->rx_queue->head_lock);
1171 if ((vp->options & VECTOR_RX) > 0)
1172 err = vector_mmsg_rx(vp, budget);
1173 else {
1174 err = vector_legacy_rx(vp);
1175 if (err > 0)
1176 err = 1;
1177 }
1178 spin_unlock(&vp->rx_queue->head_lock);
1179 if (err > 0)
1180 work_done += err;
1181
1182 if (tx_enqueued || err > 0)
1183 napi_schedule(napi);
1184 if (work_done <= budget)
1185 napi_complete_done(napi, work_done);
1186 return work_done;
1187}
1188
1189static void vector_reset_tx(struct work_struct *work)
1190{
1191 struct vector_private *vp =
1192 container_of(work, struct vector_private, reset_tx);
1193 netdev_reset_queue(vp->dev);
1194 netif_start_queue(vp->dev);
1195 netif_wake_queue(vp->dev);
1196}
1197
1198static int vector_net_open(struct net_device *dev)
1199{
1200 struct vector_private *vp = netdev_priv(dev);
1201 int err = -EINVAL;
1202 struct vector_device *vdevice;
1203
1204 if (vp->opened)
1205 return -ENXIO;
1206 vp->opened = true;
1207
1208 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1209
1210 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1211
1212 if (vp->fds == NULL)
1213 goto out_close;
1214
1215 if (build_transport_data(vp) < 0)
1216 goto out_close;
1217
1218 if ((vp->options & VECTOR_RX) > 0) {
1219 vp->rx_queue = create_queue(
1220 vp,
1221 get_depth(vp->parsed),
1222 vp->rx_header_size,
1223 MAX_IOV_SIZE
1224 );
1225 atomic_set(&vp->rx_queue->queue_depth, get_depth(vp->parsed));
1226 } else {
1227 vp->header_rxbuffer = kmalloc(
1228 vp->rx_header_size,
1229 GFP_KERNEL
1230 );
1231 if (vp->header_rxbuffer == NULL)
1232 goto out_close;
1233 }
1234 if ((vp->options & VECTOR_TX) > 0) {
1235 vp->tx_queue = create_queue(
1236 vp,
1237 get_depth(vp->parsed),
1238 vp->header_size,
1239 MAX_IOV_SIZE
1240 );
1241 } else {
1242 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1243 if (vp->header_txbuffer == NULL)
1244 goto out_close;
1245 }
1246
1247 netif_napi_add_weight(vp->dev, &vp->napi, vector_poll,
1248 get_depth(vp->parsed));
1249 napi_enable(&vp->napi);
1250
1251 /* READ IRQ */
1252 err = um_request_irq(
1253 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1254 IRQ_READ, vector_rx_interrupt,
1255 IRQF_SHARED, dev->name, dev);
1256 if (err < 0) {
1257 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1258 err = -ENETUNREACH;
1259 goto out_close;
1260 }
1261 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1262 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1263 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1264
1265 /* WRITE IRQ - we need it only if we have vector TX */
1266 if ((vp->options & VECTOR_TX) > 0) {
1267 err = um_request_irq(
1268 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1269 IRQ_WRITE, vector_tx_interrupt,
1270 IRQF_SHARED, dev->name, dev);
1271 if (err < 0) {
1272 netdev_err(dev,
1273 "vector_open: failed to get tx irq(%d)\n", err);
1274 err = -ENETUNREACH;
1275 goto out_close;
1276 }
1277 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1278 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1279 }
1280
1281 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1282 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1283 vp->options |= VECTOR_BPF;
1284 }
1285 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1286 vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1287
1288 if (vp->bpf != NULL)
1289 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1290
1291 netif_start_queue(dev);
1292 vector_reset_stats(vp);
1293
1294 /* clear buffer - it can happen that the host side of the interface
1295 * is full when we get here. In this case, new data is never queued,
1296 * SIGIOs never arrive, and the net never works.
1297 */
1298
1299 napi_schedule(&vp->napi);
1300
1301 vdevice = find_device(vp->unit);
1302 vdevice->opened = 1;
1303
1304 if ((vp->options & VECTOR_TX) != 0)
1305 add_timer(&vp->tl);
1306 return 0;
1307out_close:
1308 vector_net_close(dev);
1309 return err;
1310}
1311
1312
1313static void vector_net_set_multicast_list(struct net_device *dev)
1314{
1315 /* TODO: - we can do some BPF games here */
1316 return;
1317}
1318
1319static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1320{
1321 struct vector_private *vp = netdev_priv(dev);
1322
1323 vp->estats.tx_timeout_count++;
1324 netif_trans_update(dev);
1325 schedule_work(&vp->reset_tx);
1326}
1327
1328static netdev_features_t vector_fix_features(struct net_device *dev,
1329 netdev_features_t features)
1330{
1331 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1332 return features;
1333}
1334
1335static int vector_set_features(struct net_device *dev,
1336 netdev_features_t features)
1337{
1338 struct vector_private *vp = netdev_priv(dev);
1339 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1340 * no way to negotiate it on raw sockets, so we can change
1341 * only our side.
1342 */
1343 if (features & NETIF_F_GRO)
1344 /* All new frame buffers will be GRO-sized */
1345 vp->req_size = 65536;
1346 else
1347 /* All new frame buffers will be normal sized */
1348 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1349 return 0;
1350}
1351
1352#ifdef CONFIG_NET_POLL_CONTROLLER
1353static void vector_net_poll_controller(struct net_device *dev)
1354{
1355 disable_irq(dev->irq);
1356 vector_rx_interrupt(dev->irq, dev);
1357 enable_irq(dev->irq);
1358}
1359#endif
1360
1361static void vector_net_get_drvinfo(struct net_device *dev,
1362 struct ethtool_drvinfo *info)
1363{
1364 strscpy(info->driver, DRIVER_NAME);
1365}
1366
1367static int vector_net_load_bpf_flash(struct net_device *dev,
1368 struct ethtool_flash *efl)
1369{
1370 struct vector_private *vp = netdev_priv(dev);
1371 struct vector_device *vdevice;
1372 const struct firmware *fw;
1373 int result = 0;
1374
1375 if (!(vp->options & VECTOR_BPF_FLASH)) {
1376 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1377 return -1;
1378 }
1379
1380 if (vp->bpf != NULL) {
1381 if (vp->opened)
1382 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1383 kfree(vp->bpf->filter);
1384 vp->bpf->filter = NULL;
1385 } else {
1386 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC);
1387 if (vp->bpf == NULL) {
1388 netdev_err(dev, "failed to allocate memory for firmware\n");
1389 goto flash_fail;
1390 }
1391 }
1392
1393 vdevice = find_device(vp->unit);
1394
1395 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1396 goto flash_fail;
1397
1398 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC);
1399 if (!vp->bpf->filter)
1400 goto free_buffer;
1401
1402 vp->bpf->len = fw->size / sizeof(struct sock_filter);
1403 release_firmware(fw);
1404
1405 if (vp->opened)
1406 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1407
1408 return result;
1409
1410free_buffer:
1411 release_firmware(fw);
1412
1413flash_fail:
1414 if (vp->bpf != NULL)
1415 kfree(vp->bpf->filter);
1416 kfree(vp->bpf);
1417 vp->bpf = NULL;
1418 return -1;
1419}
1420
1421static void vector_get_ringparam(struct net_device *netdev,
1422 struct ethtool_ringparam *ring,
1423 struct kernel_ethtool_ringparam *kernel_ring,
1424 struct netlink_ext_ack *extack)
1425{
1426 struct vector_private *vp = netdev_priv(netdev);
1427
1428 ring->rx_max_pending = vp->rx_queue->max_depth;
1429 ring->tx_max_pending = vp->tx_queue->max_depth;
1430 ring->rx_pending = vp->rx_queue->max_depth;
1431 ring->tx_pending = vp->tx_queue->max_depth;
1432}
1433
1434static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1435{
1436 switch (stringset) {
1437 case ETH_SS_TEST:
1438 *buf = '\0';
1439 break;
1440 case ETH_SS_STATS:
1441 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1442 break;
1443 default:
1444 WARN_ON(1);
1445 break;
1446 }
1447}
1448
1449static int vector_get_sset_count(struct net_device *dev, int sset)
1450{
1451 switch (sset) {
1452 case ETH_SS_TEST:
1453 return 0;
1454 case ETH_SS_STATS:
1455 return VECTOR_NUM_STATS;
1456 default:
1457 return -EOPNOTSUPP;
1458 }
1459}
1460
1461static void vector_get_ethtool_stats(struct net_device *dev,
1462 struct ethtool_stats *estats,
1463 u64 *tmp_stats)
1464{
1465 struct vector_private *vp = netdev_priv(dev);
1466
1467 /* Stats are modified in the dequeue portions of
1468 * rx/tx which are protected by the head locks
1469 * grabbing these locks here ensures they are up
1470 * to date.
1471 */
1472
1473 spin_lock(&vp->tx_queue->head_lock);
1474 spin_lock(&vp->rx_queue->head_lock);
1475 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1476 spin_unlock(&vp->rx_queue->head_lock);
1477 spin_unlock(&vp->tx_queue->head_lock);
1478}
1479
1480static int vector_get_coalesce(struct net_device *netdev,
1481 struct ethtool_coalesce *ec,
1482 struct kernel_ethtool_coalesce *kernel_coal,
1483 struct netlink_ext_ack *extack)
1484{
1485 struct vector_private *vp = netdev_priv(netdev);
1486
1487 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1488 return 0;
1489}
1490
1491static int vector_set_coalesce(struct net_device *netdev,
1492 struct ethtool_coalesce *ec,
1493 struct kernel_ethtool_coalesce *kernel_coal,
1494 struct netlink_ext_ack *extack)
1495{
1496 struct vector_private *vp = netdev_priv(netdev);
1497
1498 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1499 if (vp->coalesce == 0)
1500 vp->coalesce = 1;
1501 return 0;
1502}
1503
1504static const struct ethtool_ops vector_net_ethtool_ops = {
1505 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1506 .get_drvinfo = vector_net_get_drvinfo,
1507 .get_link = ethtool_op_get_link,
1508 .get_ts_info = ethtool_op_get_ts_info,
1509 .get_ringparam = vector_get_ringparam,
1510 .get_strings = vector_get_strings,
1511 .get_sset_count = vector_get_sset_count,
1512 .get_ethtool_stats = vector_get_ethtool_stats,
1513 .get_coalesce = vector_get_coalesce,
1514 .set_coalesce = vector_set_coalesce,
1515 .flash_device = vector_net_load_bpf_flash,
1516};
1517
1518
1519static const struct net_device_ops vector_netdev_ops = {
1520 .ndo_open = vector_net_open,
1521 .ndo_stop = vector_net_close,
1522 .ndo_start_xmit = vector_net_start_xmit,
1523 .ndo_set_rx_mode = vector_net_set_multicast_list,
1524 .ndo_tx_timeout = vector_net_tx_timeout,
1525 .ndo_set_mac_address = eth_mac_addr,
1526 .ndo_validate_addr = eth_validate_addr,
1527 .ndo_fix_features = vector_fix_features,
1528 .ndo_set_features = vector_set_features,
1529#ifdef CONFIG_NET_POLL_CONTROLLER
1530 .ndo_poll_controller = vector_net_poll_controller,
1531#endif
1532};
1533
1534static void vector_timer_expire(struct timer_list *t)
1535{
1536 struct vector_private *vp = from_timer(vp, t, tl);
1537
1538 vp->estats.tx_kicks++;
1539 napi_schedule(&vp->napi);
1540}
1541
1542
1543
1544static void vector_eth_configure(
1545 int n,
1546 struct arglist *def
1547 )
1548{
1549 struct vector_device *device;
1550 struct net_device *dev;
1551 struct vector_private *vp;
1552 int err;
1553
1554 device = kzalloc(sizeof(*device), GFP_KERNEL);
1555 if (device == NULL) {
1556 printk(KERN_ERR "eth_configure failed to allocate struct "
1557 "vector_device\n");
1558 return;
1559 }
1560 dev = alloc_etherdev(sizeof(struct vector_private));
1561 if (dev == NULL) {
1562 printk(KERN_ERR "eth_configure: failed to allocate struct "
1563 "net_device for vec%d\n", n);
1564 goto out_free_device;
1565 }
1566
1567 dev->mtu = get_mtu(def);
1568
1569 INIT_LIST_HEAD(&device->list);
1570 device->unit = n;
1571
1572 /* If this name ends up conflicting with an existing registered
1573 * netdevice, that is OK, register_netdev{,ice}() will notice this
1574 * and fail.
1575 */
1576 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1577 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1578 vp = netdev_priv(dev);
1579
1580 /* sysfs register */
1581 if (!driver_registered) {
1582 platform_driver_register(¨_net_driver);
1583 driver_registered = 1;
1584 }
1585 device->pdev.id = n;
1586 device->pdev.name = DRIVER_NAME;
1587 device->pdev.dev.release = vector_device_release;
1588 dev_set_drvdata(&device->pdev.dev, device);
1589 if (platform_device_register(&device->pdev))
1590 goto out_free_netdev;
1591 SET_NETDEV_DEV(dev, &device->pdev.dev);
1592
1593 device->dev = dev;
1594
1595 *vp = ((struct vector_private)
1596 {
1597 .list = LIST_HEAD_INIT(vp->list),
1598 .dev = dev,
1599 .unit = n,
1600 .options = get_transport_options(def),
1601 .rx_irq = 0,
1602 .tx_irq = 0,
1603 .parsed = def,
1604 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1605 /* TODO - we need to calculate headroom so that ip header
1606 * is 16 byte aligned all the time
1607 */
1608 .headroom = get_headroom(def),
1609 .form_header = NULL,
1610 .verify_header = NULL,
1611 .header_rxbuffer = NULL,
1612 .header_txbuffer = NULL,
1613 .header_size = 0,
1614 .rx_header_size = 0,
1615 .rexmit_scheduled = false,
1616 .opened = false,
1617 .transport_data = NULL,
1618 .in_write_poll = false,
1619 .coalesce = 2,
1620 .req_size = get_req_size(def),
1621 .in_error = false,
1622 .bpf = NULL
1623 });
1624
1625 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1626 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1627
1628 timer_setup(&vp->tl, vector_timer_expire, 0);
1629
1630 /* FIXME */
1631 dev->netdev_ops = &vector_netdev_ops;
1632 dev->ethtool_ops = &vector_net_ethtool_ops;
1633 dev->watchdog_timeo = (HZ >> 1);
1634 /* primary IRQ - fixme */
1635 dev->irq = 0; /* we will adjust this once opened */
1636
1637 rtnl_lock();
1638 err = register_netdevice(dev);
1639 rtnl_unlock();
1640 if (err)
1641 goto out_undo_user_init;
1642
1643 spin_lock(&vector_devices_lock);
1644 list_add(&device->list, &vector_devices);
1645 spin_unlock(&vector_devices_lock);
1646
1647 return;
1648
1649out_undo_user_init:
1650 return;
1651out_free_netdev:
1652 free_netdev(dev);
1653out_free_device:
1654 kfree(device);
1655}
1656
1657
1658
1659
1660/*
1661 * Invoked late in the init
1662 */
1663
1664static int __init vector_init(void)
1665{
1666 struct list_head *ele;
1667 struct vector_cmd_line_arg *def;
1668 struct arglist *parsed;
1669
1670 list_for_each(ele, &vec_cmd_line) {
1671 def = list_entry(ele, struct vector_cmd_line_arg, list);
1672 parsed = uml_parse_vector_ifspec(def->arguments);
1673 if (parsed != NULL)
1674 vector_eth_configure(def->unit, parsed);
1675 }
1676 return 0;
1677}
1678
1679
1680/* Invoked at initial argument parsing, only stores
1681 * arguments until a proper vector_init is called
1682 * later
1683 */
1684
1685static int __init vector_setup(char *str)
1686{
1687 char *error;
1688 int n, err;
1689 struct vector_cmd_line_arg *new;
1690
1691 err = vector_parse(str, &n, &str, &error);
1692 if (err) {
1693 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1694 str, error);
1695 return 1;
1696 }
1697 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1698 if (!new)
1699 panic("%s: Failed to allocate %zu bytes\n", __func__,
1700 sizeof(*new));
1701 INIT_LIST_HEAD(&new->list);
1702 new->unit = n;
1703 new->arguments = str;
1704 list_add_tail(&new->list, &vec_cmd_line);
1705 return 1;
1706}
1707
1708__setup("vec", vector_setup);
1709__uml_help(vector_setup,
1710"vec[0-9]+:<option>=<value>,<option>=<value>\n"
1711" Configure a vector io network device.\n\n"
1712);
1713
1714late_initcall(vector_init);
1715
1716static struct mc_device vector_mc = {
1717 .list = LIST_HEAD_INIT(vector_mc.list),
1718 .name = "vec",
1719 .config = vector_config,
1720 .get_config = NULL,
1721 .id = vector_id,
1722 .remove = vector_remove,
1723};
1724
1725#ifdef CONFIG_INET
1726static int vector_inetaddr_event(
1727 struct notifier_block *this,
1728 unsigned long event,
1729 void *ptr)
1730{
1731 return NOTIFY_DONE;
1732}
1733
1734static struct notifier_block vector_inetaddr_notifier = {
1735 .notifier_call = vector_inetaddr_event,
1736};
1737
1738static void inet_register(void)
1739{
1740 register_inetaddr_notifier(&vector_inetaddr_notifier);
1741}
1742#else
1743static inline void inet_register(void)
1744{
1745}
1746#endif
1747
1748static int vector_net_init(void)
1749{
1750 mconsole_register_dev(&vector_mc);
1751 inet_register();
1752 return 0;
1753}
1754
1755__initcall(vector_net_init);
1756
1757
1758
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2017 - 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.
9 */
10
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 <init.h>
25#include <irq_kern.h>
26#include <irq_user.h>
27#include <net_kern.h>
28#include <os.h>
29#include "mconsole_kern.h"
30#include "vector_user.h"
31#include "vector_kern.h"
32
33/*
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
42 */
43
44
45#define DRIVER_NAME "uml-vector"
46#define DRIVER_VERSION "01"
47struct vector_cmd_line_arg {
48 struct list_head list;
49 int unit;
50 char *arguments;
51};
52
53struct vector_device {
54 struct list_head list;
55 struct net_device *dev;
56 struct platform_device pdev;
57 int unit;
58 int opened;
59};
60
61static LIST_HEAD(vec_cmd_line);
62
63static DEFINE_SPINLOCK(vector_devices_lock);
64static LIST_HEAD(vector_devices);
65
66static int driver_registered;
67
68static void vector_eth_configure(int n, struct arglist *def);
69
70/* Argument accessors to set variables (and/or set default values)
71 * mtu, buffer sizing, default headroom, etc
72 */
73
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)
79#define MAX_ITERATIONS 64
80
81static const struct {
82 const char string[ETH_GSTRING_LEN];
83} ethtool_stats_keys[] = {
84 { "rx_queue_max" },
85 { "rx_queue_running_average" },
86 { "tx_queue_max" },
87 { "tx_queue_running_average" },
88 { "rx_encaps_errors" },
89 { "tx_timeout_count" },
90 { "tx_restart_queue" },
91 { "tx_kicks" },
92 { "tx_flow_control_xon" },
93 { "tx_flow_control_xoff" },
94 { "rx_csum_offload_good" },
95 { "rx_csum_offload_errors"},
96 { "sg_ok"},
97 { "sg_linearized"},
98};
99
100#define VECTOR_NUM_STATS ARRAY_SIZE(ethtool_stats_keys)
101
102static void vector_reset_stats(struct vector_private *vp)
103{
104 vp->estats.rx_queue_max = 0;
105 vp->estats.rx_queue_running_average = 0;
106 vp->estats.tx_queue_max = 0;
107 vp->estats.tx_queue_running_average = 0;
108 vp->estats.rx_encaps_errors = 0;
109 vp->estats.tx_timeout_count = 0;
110 vp->estats.tx_restart_queue = 0;
111 vp->estats.tx_kicks = 0;
112 vp->estats.tx_flow_control_xon = 0;
113 vp->estats.tx_flow_control_xoff = 0;
114 vp->estats.sg_ok = 0;
115 vp->estats.sg_linearized = 0;
116}
117
118static int get_mtu(struct arglist *def)
119{
120 char *mtu = uml_vector_fetch_arg(def, "mtu");
121 long result;
122
123 if (mtu != NULL) {
124 if (kstrtoul(mtu, 10, &result) == 0)
125 if ((result < (1 << 16) - 1) && (result >= 576))
126 return result;
127 }
128 return ETH_MAX_PACKET;
129}
130
131static int get_depth(struct arglist *def)
132{
133 char *mtu = uml_vector_fetch_arg(def, "depth");
134 long result;
135
136 if (mtu != NULL) {
137 if (kstrtoul(mtu, 10, &result) == 0)
138 return result;
139 }
140 return DEFAULT_VECTOR_SIZE;
141}
142
143static int get_headroom(struct arglist *def)
144{
145 char *mtu = uml_vector_fetch_arg(def, "headroom");
146 long result;
147
148 if (mtu != NULL) {
149 if (kstrtoul(mtu, 10, &result) == 0)
150 return result;
151 }
152 return DEFAULT_HEADROOM;
153}
154
155static int get_req_size(struct arglist *def)
156{
157 char *gro = uml_vector_fetch_arg(def, "gro");
158 long result;
159
160 if (gro != NULL) {
161 if (kstrtoul(gro, 10, &result) == 0) {
162 if (result > 0)
163 return 65536;
164 }
165 }
166 return get_mtu(def) + ETH_HEADER_OTHER +
167 get_headroom(def) + SAFETY_MARGIN;
168}
169
170
171static int get_transport_options(struct arglist *def)
172{
173 char *transport = uml_vector_fetch_arg(def, "transport");
174 char *vector = uml_vector_fetch_arg(def, "vec");
175
176 int vec_rx = VECTOR_RX;
177 int vec_tx = VECTOR_TX;
178 long parsed;
179
180 if (vector != NULL) {
181 if (kstrtoul(vector, 10, &parsed) == 0) {
182 if (parsed == 0) {
183 vec_rx = 0;
184 vec_tx = 0;
185 }
186 }
187 }
188
189
190 if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
191 return 0;
192 if (strncmp(transport, TRANS_HYBRID, TRANS_HYBRID_LEN) == 0)
193 return (vec_rx | VECTOR_BPF);
194 if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
195 return (vec_rx | vec_tx | VECTOR_QDISC_BYPASS);
196 return (vec_rx | vec_tx);
197}
198
199
200/* A mini-buffer for packet drop read
201 * All of our supported transports are datagram oriented and we always
202 * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
203 * than the packet size it still counts as full packet read and will
204 * clean the incoming stream to keep sigio/epoll happy
205 */
206
207#define DROP_BUFFER_SIZE 32
208
209static char *drop_buffer;
210
211/* Array backed queues optimized for bulk enqueue/dequeue and
212 * 1:N (small values of N) or 1:1 enqueuer/dequeuer ratios.
213 * For more details and full design rationale see
214 * http://foswiki.cambridgegreys.com/Main/EatYourTailAndEnjoyIt
215 */
216
217
218/*
219 * Advance the mmsg queue head by n = advance. Resets the queue to
220 * maximum enqueue/dequeue-at-once capacity if possible. Called by
221 * dequeuers. Caller must hold the head_lock!
222 */
223
224static int vector_advancehead(struct vector_queue *qi, int advance)
225{
226 int queue_depth;
227
228 qi->head =
229 (qi->head + advance)
230 % qi->max_depth;
231
232
233 spin_lock(&qi->tail_lock);
234 qi->queue_depth -= advance;
235
236 /* we are at 0, use this to
237 * reset head and tail so we can use max size vectors
238 */
239
240 if (qi->queue_depth == 0) {
241 qi->head = 0;
242 qi->tail = 0;
243 }
244 queue_depth = qi->queue_depth;
245 spin_unlock(&qi->tail_lock);
246 return queue_depth;
247}
248
249/* Advance the queue tail by n = advance.
250 * This is called by enqueuers which should hold the
251 * head lock already
252 */
253
254static int vector_advancetail(struct vector_queue *qi, int advance)
255{
256 int queue_depth;
257
258 qi->tail =
259 (qi->tail + advance)
260 % qi->max_depth;
261 spin_lock(&qi->head_lock);
262 qi->queue_depth += advance;
263 queue_depth = qi->queue_depth;
264 spin_unlock(&qi->head_lock);
265 return queue_depth;
266}
267
268static int prep_msg(struct vector_private *vp,
269 struct sk_buff *skb,
270 struct iovec *iov)
271{
272 int iov_index = 0;
273 int nr_frags, frag;
274 skb_frag_t *skb_frag;
275
276 nr_frags = skb_shinfo(skb)->nr_frags;
277 if (nr_frags > MAX_IOV_SIZE) {
278 if (skb_linearize(skb) != 0)
279 goto drop;
280 }
281 if (vp->header_size > 0) {
282 iov[iov_index].iov_len = vp->header_size;
283 vp->form_header(iov[iov_index].iov_base, skb, vp);
284 iov_index++;
285 }
286 iov[iov_index].iov_base = skb->data;
287 if (nr_frags > 0) {
288 iov[iov_index].iov_len = skb->len - skb->data_len;
289 vp->estats.sg_ok++;
290 } else
291 iov[iov_index].iov_len = skb->len;
292 iov_index++;
293 for (frag = 0; frag < nr_frags; frag++) {
294 skb_frag = &skb_shinfo(skb)->frags[frag];
295 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
296 iov[iov_index].iov_len = skb_frag_size(skb_frag);
297 iov_index++;
298 }
299 return iov_index;
300drop:
301 return -1;
302}
303/*
304 * Generic vector enqueue with support for forming headers using transport
305 * specific callback. Allows GRE, L2TPv3, RAW and other transports
306 * to use a common enqueue procedure in vector mode
307 */
308
309static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
310{
311 struct vector_private *vp = netdev_priv(qi->dev);
312 int queue_depth;
313 int packet_len;
314 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
315 int iov_count;
316
317 spin_lock(&qi->tail_lock);
318 spin_lock(&qi->head_lock);
319 queue_depth = qi->queue_depth;
320 spin_unlock(&qi->head_lock);
321
322 if (skb)
323 packet_len = skb->len;
324
325 if (queue_depth < qi->max_depth) {
326
327 *(qi->skbuff_vector + qi->tail) = skb;
328 mmsg_vector += qi->tail;
329 iov_count = prep_msg(
330 vp,
331 skb,
332 mmsg_vector->msg_hdr.msg_iov
333 );
334 if (iov_count < 1)
335 goto drop;
336 mmsg_vector->msg_hdr.msg_iovlen = iov_count;
337 mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
338 mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
339 queue_depth = vector_advancetail(qi, 1);
340 } else
341 goto drop;
342 spin_unlock(&qi->tail_lock);
343 return queue_depth;
344drop:
345 qi->dev->stats.tx_dropped++;
346 if (skb != NULL) {
347 packet_len = skb->len;
348 dev_consume_skb_any(skb);
349 netdev_completed_queue(qi->dev, 1, packet_len);
350 }
351 spin_unlock(&qi->tail_lock);
352 return queue_depth;
353}
354
355static int consume_vector_skbs(struct vector_queue *qi, int count)
356{
357 struct sk_buff *skb;
358 int skb_index;
359 int bytes_compl = 0;
360
361 for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
362 skb = *(qi->skbuff_vector + skb_index);
363 /* mark as empty to ensure correct destruction if
364 * needed
365 */
366 bytes_compl += skb->len;
367 *(qi->skbuff_vector + skb_index) = NULL;
368 dev_consume_skb_any(skb);
369 }
370 qi->dev->stats.tx_bytes += bytes_compl;
371 qi->dev->stats.tx_packets += count;
372 netdev_completed_queue(qi->dev, count, bytes_compl);
373 return vector_advancehead(qi, count);
374}
375
376/*
377 * Generic vector deque via sendmmsg with support for forming headers
378 * using transport specific callback. Allows GRE, L2TPv3, RAW and
379 * other transports to use a common dequeue procedure in vector mode
380 */
381
382
383static int vector_send(struct vector_queue *qi)
384{
385 struct vector_private *vp = netdev_priv(qi->dev);
386 struct mmsghdr *send_from;
387 int result = 0, send_len, queue_depth = qi->max_depth;
388
389 if (spin_trylock(&qi->head_lock)) {
390 if (spin_trylock(&qi->tail_lock)) {
391 /* update queue_depth to current value */
392 queue_depth = qi->queue_depth;
393 spin_unlock(&qi->tail_lock);
394 while (queue_depth > 0) {
395 /* Calculate the start of the vector */
396 send_len = queue_depth;
397 send_from = qi->mmsg_vector;
398 send_from += qi->head;
399 /* Adjust vector size if wraparound */
400 if (send_len + qi->head > qi->max_depth)
401 send_len = qi->max_depth - qi->head;
402 /* Try to TX as many packets as possible */
403 if (send_len > 0) {
404 result = uml_vector_sendmmsg(
405 vp->fds->tx_fd,
406 send_from,
407 send_len,
408 0
409 );
410 vp->in_write_poll =
411 (result != send_len);
412 }
413 /* For some of the sendmmsg error scenarios
414 * we may end being unsure in the TX success
415 * for all packets. It is safer to declare
416 * them all TX-ed and blame the network.
417 */
418 if (result < 0) {
419 if (net_ratelimit())
420 netdev_err(vp->dev, "sendmmsg err=%i\n",
421 result);
422 vp->in_error = true;
423 result = send_len;
424 }
425 if (result > 0) {
426 queue_depth =
427 consume_vector_skbs(qi, result);
428 /* This is equivalent to an TX IRQ.
429 * Restart the upper layers to feed us
430 * more packets.
431 */
432 if (result > vp->estats.tx_queue_max)
433 vp->estats.tx_queue_max = result;
434 vp->estats.tx_queue_running_average =
435 (vp->estats.tx_queue_running_average + result) >> 1;
436 }
437 netif_trans_update(qi->dev);
438 netif_wake_queue(qi->dev);
439 /* if TX is busy, break out of the send loop,
440 * poll write IRQ will reschedule xmit for us
441 */
442 if (result != send_len) {
443 vp->estats.tx_restart_queue++;
444 break;
445 }
446 }
447 }
448 spin_unlock(&qi->head_lock);
449 } else {
450 tasklet_schedule(&vp->tx_poll);
451 }
452 return queue_depth;
453}
454
455/* Queue destructor. Deliberately stateless so we can use
456 * it in queue cleanup if initialization fails.
457 */
458
459static void destroy_queue(struct vector_queue *qi)
460{
461 int i;
462 struct iovec *iov;
463 struct vector_private *vp = netdev_priv(qi->dev);
464 struct mmsghdr *mmsg_vector;
465
466 if (qi == NULL)
467 return;
468 /* deallocate any skbuffs - we rely on any unused to be
469 * set to NULL.
470 */
471 if (qi->skbuff_vector != NULL) {
472 for (i = 0; i < qi->max_depth; i++) {
473 if (*(qi->skbuff_vector + i) != NULL)
474 dev_kfree_skb_any(*(qi->skbuff_vector + i));
475 }
476 kfree(qi->skbuff_vector);
477 }
478 /* deallocate matching IOV structures including header buffs */
479 if (qi->mmsg_vector != NULL) {
480 mmsg_vector = qi->mmsg_vector;
481 for (i = 0; i < qi->max_depth; i++) {
482 iov = mmsg_vector->msg_hdr.msg_iov;
483 if (iov != NULL) {
484 if ((vp->header_size > 0) &&
485 (iov->iov_base != NULL))
486 kfree(iov->iov_base);
487 kfree(iov);
488 }
489 mmsg_vector++;
490 }
491 kfree(qi->mmsg_vector);
492 }
493 kfree(qi);
494}
495
496/*
497 * Queue constructor. Create a queue with a given side.
498 */
499static struct vector_queue *create_queue(
500 struct vector_private *vp,
501 int max_size,
502 int header_size,
503 int num_extra_frags)
504{
505 struct vector_queue *result;
506 int i;
507 struct iovec *iov;
508 struct mmsghdr *mmsg_vector;
509
510 result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
511 if (result == NULL)
512 return NULL;
513 result->max_depth = max_size;
514 result->dev = vp->dev;
515 result->mmsg_vector = kmalloc(
516 (sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
517 if (result->mmsg_vector == NULL)
518 goto out_mmsg_fail;
519 result->skbuff_vector = kmalloc(
520 (sizeof(void *) * max_size), GFP_KERNEL);
521 if (result->skbuff_vector == NULL)
522 goto out_skb_fail;
523
524 /* further failures can be handled safely by destroy_queue*/
525
526 mmsg_vector = result->mmsg_vector;
527 for (i = 0; i < max_size; i++) {
528 /* Clear all pointers - we use non-NULL as marking on
529 * what to free on destruction
530 */
531 *(result->skbuff_vector + i) = NULL;
532 mmsg_vector->msg_hdr.msg_iov = NULL;
533 mmsg_vector++;
534 }
535 mmsg_vector = result->mmsg_vector;
536 result->max_iov_frags = num_extra_frags;
537 for (i = 0; i < max_size; i++) {
538 if (vp->header_size > 0)
539 iov = kmalloc_array(3 + num_extra_frags,
540 sizeof(struct iovec),
541 GFP_KERNEL
542 );
543 else
544 iov = kmalloc_array(2 + num_extra_frags,
545 sizeof(struct iovec),
546 GFP_KERNEL
547 );
548 if (iov == NULL)
549 goto out_fail;
550 mmsg_vector->msg_hdr.msg_iov = iov;
551 mmsg_vector->msg_hdr.msg_iovlen = 1;
552 mmsg_vector->msg_hdr.msg_control = NULL;
553 mmsg_vector->msg_hdr.msg_controllen = 0;
554 mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
555 mmsg_vector->msg_hdr.msg_name = NULL;
556 mmsg_vector->msg_hdr.msg_namelen = 0;
557 if (vp->header_size > 0) {
558 iov->iov_base = kmalloc(header_size, GFP_KERNEL);
559 if (iov->iov_base == NULL)
560 goto out_fail;
561 iov->iov_len = header_size;
562 mmsg_vector->msg_hdr.msg_iovlen = 2;
563 iov++;
564 }
565 iov->iov_base = NULL;
566 iov->iov_len = 0;
567 mmsg_vector++;
568 }
569 spin_lock_init(&result->head_lock);
570 spin_lock_init(&result->tail_lock);
571 result->queue_depth = 0;
572 result->head = 0;
573 result->tail = 0;
574 return result;
575out_skb_fail:
576 kfree(result->mmsg_vector);
577out_mmsg_fail:
578 kfree(result);
579 return NULL;
580out_fail:
581 destroy_queue(result);
582 return NULL;
583}
584
585/*
586 * We do not use the RX queue as a proper wraparound queue for now
587 * This is not necessary because the consumption via netif_rx()
588 * happens in-line. While we can try using the return code of
589 * netif_rx() for flow control there are no drivers doing this today.
590 * For this RX specific use we ignore the tail/head locks and
591 * just read into a prepared queue filled with skbuffs.
592 */
593
594static struct sk_buff *prep_skb(
595 struct vector_private *vp,
596 struct user_msghdr *msg)
597{
598 int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
599 struct sk_buff *result;
600 int iov_index = 0, len;
601 struct iovec *iov = msg->msg_iov;
602 int err, nr_frags, frag;
603 skb_frag_t *skb_frag;
604
605 if (vp->req_size <= linear)
606 len = linear;
607 else
608 len = vp->req_size;
609 result = alloc_skb_with_frags(
610 linear,
611 len - vp->max_packet,
612 3,
613 &err,
614 GFP_ATOMIC
615 );
616 if (vp->header_size > 0)
617 iov_index++;
618 if (result == NULL) {
619 iov[iov_index].iov_base = NULL;
620 iov[iov_index].iov_len = 0;
621 goto done;
622 }
623 skb_reserve(result, vp->headroom);
624 result->dev = vp->dev;
625 skb_put(result, vp->max_packet);
626 result->data_len = len - vp->max_packet;
627 result->len += len - vp->max_packet;
628 skb_reset_mac_header(result);
629 result->ip_summed = CHECKSUM_NONE;
630 iov[iov_index].iov_base = result->data;
631 iov[iov_index].iov_len = vp->max_packet;
632 iov_index++;
633
634 nr_frags = skb_shinfo(result)->nr_frags;
635 for (frag = 0; frag < nr_frags; frag++) {
636 skb_frag = &skb_shinfo(result)->frags[frag];
637 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
638 if (iov[iov_index].iov_base != NULL)
639 iov[iov_index].iov_len = skb_frag_size(skb_frag);
640 else
641 iov[iov_index].iov_len = 0;
642 iov_index++;
643 }
644done:
645 msg->msg_iovlen = iov_index;
646 return result;
647}
648
649
650/* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs*/
651
652static void prep_queue_for_rx(struct vector_queue *qi)
653{
654 struct vector_private *vp = netdev_priv(qi->dev);
655 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
656 void **skbuff_vector = qi->skbuff_vector;
657 int i;
658
659 if (qi->queue_depth == 0)
660 return;
661 for (i = 0; i < qi->queue_depth; i++) {
662 /* it is OK if allocation fails - recvmmsg with NULL data in
663 * iov argument still performs an RX, just drops the packet
664 * This allows us stop faffing around with a "drop buffer"
665 */
666
667 *skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
668 skbuff_vector++;
669 mmsg_vector++;
670 }
671 qi->queue_depth = 0;
672}
673
674static struct vector_device *find_device(int n)
675{
676 struct vector_device *device;
677 struct list_head *ele;
678
679 spin_lock(&vector_devices_lock);
680 list_for_each(ele, &vector_devices) {
681 device = list_entry(ele, struct vector_device, list);
682 if (device->unit == n)
683 goto out;
684 }
685 device = NULL;
686 out:
687 spin_unlock(&vector_devices_lock);
688 return device;
689}
690
691static int vector_parse(char *str, int *index_out, char **str_out,
692 char **error_out)
693{
694 int n, len, err;
695 char *start = str;
696
697 len = strlen(str);
698
699 while ((*str != ':') && (strlen(str) > 1))
700 str++;
701 if (*str != ':') {
702 *error_out = "Expected ':' after device number";
703 return -EINVAL;
704 }
705 *str = '\0';
706
707 err = kstrtouint(start, 0, &n);
708 if (err < 0) {
709 *error_out = "Bad device number";
710 return err;
711 }
712
713 str++;
714 if (find_device(n)) {
715 *error_out = "Device already configured";
716 return -EINVAL;
717 }
718
719 *index_out = n;
720 *str_out = str;
721 return 0;
722}
723
724static int vector_config(char *str, char **error_out)
725{
726 int err, n;
727 char *params;
728 struct arglist *parsed;
729
730 err = vector_parse(str, &n, ¶ms, error_out);
731 if (err != 0)
732 return err;
733
734 /* This string is broken up and the pieces used by the underlying
735 * driver. We should copy it to make sure things do not go wrong
736 * later.
737 */
738
739 params = kstrdup(params, GFP_KERNEL);
740 if (params == NULL) {
741 *error_out = "vector_config failed to strdup string";
742 return -ENOMEM;
743 }
744
745 parsed = uml_parse_vector_ifspec(params);
746
747 if (parsed == NULL) {
748 *error_out = "vector_config failed to parse parameters";
749 return -EINVAL;
750 }
751
752 vector_eth_configure(n, parsed);
753 return 0;
754}
755
756static int vector_id(char **str, int *start_out, int *end_out)
757{
758 char *end;
759 int n;
760
761 n = simple_strtoul(*str, &end, 0);
762 if ((*end != '\0') || (end == *str))
763 return -1;
764
765 *start_out = n;
766 *end_out = n;
767 *str = end;
768 return n;
769}
770
771static int vector_remove(int n, char **error_out)
772{
773 struct vector_device *vec_d;
774 struct net_device *dev;
775 struct vector_private *vp;
776
777 vec_d = find_device(n);
778 if (vec_d == NULL)
779 return -ENODEV;
780 dev = vec_d->dev;
781 vp = netdev_priv(dev);
782 if (vp->fds != NULL)
783 return -EBUSY;
784 unregister_netdev(dev);
785 platform_device_unregister(&vec_d->pdev);
786 return 0;
787}
788
789/*
790 * There is no shared per-transport initialization code, so
791 * we will just initialize each interface one by one and
792 * add them to a list
793 */
794
795static struct platform_driver uml_net_driver = {
796 .driver = {
797 .name = DRIVER_NAME,
798 },
799};
800
801
802static void vector_device_release(struct device *dev)
803{
804 struct vector_device *device = dev_get_drvdata(dev);
805 struct net_device *netdev = device->dev;
806
807 list_del(&device->list);
808 kfree(device);
809 free_netdev(netdev);
810}
811
812/* Bog standard recv using recvmsg - not used normally unless the user
813 * explicitly specifies not to use recvmmsg vector RX.
814 */
815
816static int vector_legacy_rx(struct vector_private *vp)
817{
818 int pkt_len;
819 struct user_msghdr hdr;
820 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
821 int iovpos = 0;
822 struct sk_buff *skb;
823 int header_check;
824
825 hdr.msg_name = NULL;
826 hdr.msg_namelen = 0;
827 hdr.msg_iov = (struct iovec *) &iov;
828 hdr.msg_control = NULL;
829 hdr.msg_controllen = 0;
830 hdr.msg_flags = 0;
831
832 if (vp->header_size > 0) {
833 iov[0].iov_base = vp->header_rxbuffer;
834 iov[0].iov_len = vp->header_size;
835 }
836
837 skb = prep_skb(vp, &hdr);
838
839 if (skb == NULL) {
840 /* Read a packet into drop_buffer and don't do
841 * anything with it.
842 */
843 iov[iovpos].iov_base = drop_buffer;
844 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
845 hdr.msg_iovlen = 1;
846 vp->dev->stats.rx_dropped++;
847 }
848
849 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
850 if (pkt_len < 0) {
851 vp->in_error = true;
852 return pkt_len;
853 }
854
855 if (skb != NULL) {
856 if (pkt_len > vp->header_size) {
857 if (vp->header_size > 0) {
858 header_check = vp->verify_header(
859 vp->header_rxbuffer, skb, vp);
860 if (header_check < 0) {
861 dev_kfree_skb_irq(skb);
862 vp->dev->stats.rx_dropped++;
863 vp->estats.rx_encaps_errors++;
864 return 0;
865 }
866 if (header_check > 0) {
867 vp->estats.rx_csum_offload_good++;
868 skb->ip_summed = CHECKSUM_UNNECESSARY;
869 }
870 }
871 pskb_trim(skb, pkt_len - vp->rx_header_size);
872 skb->protocol = eth_type_trans(skb, skb->dev);
873 vp->dev->stats.rx_bytes += skb->len;
874 vp->dev->stats.rx_packets++;
875 netif_rx(skb);
876 } else {
877 dev_kfree_skb_irq(skb);
878 }
879 }
880 return pkt_len;
881}
882
883/*
884 * Packet at a time TX which falls back to vector TX if the
885 * underlying transport is busy.
886 */
887
888
889
890static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
891{
892 struct iovec iov[3 + MAX_IOV_SIZE];
893 int iov_count, pkt_len = 0;
894
895 iov[0].iov_base = vp->header_txbuffer;
896 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
897
898 if (iov_count < 1)
899 goto drop;
900
901 pkt_len = uml_vector_writev(
902 vp->fds->tx_fd,
903 (struct iovec *) &iov,
904 iov_count
905 );
906
907 if (pkt_len < 0)
908 goto drop;
909
910 netif_trans_update(vp->dev);
911 netif_wake_queue(vp->dev);
912
913 if (pkt_len > 0) {
914 vp->dev->stats.tx_bytes += skb->len;
915 vp->dev->stats.tx_packets++;
916 } else {
917 vp->dev->stats.tx_dropped++;
918 }
919 consume_skb(skb);
920 return pkt_len;
921drop:
922 vp->dev->stats.tx_dropped++;
923 consume_skb(skb);
924 if (pkt_len < 0)
925 vp->in_error = true;
926 return pkt_len;
927}
928
929/*
930 * Receive as many messages as we can in one call using the special
931 * mmsg vector matched to an skb vector which we prepared earlier.
932 */
933
934static int vector_mmsg_rx(struct vector_private *vp)
935{
936 int packet_count, i;
937 struct vector_queue *qi = vp->rx_queue;
938 struct sk_buff *skb;
939 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
940 void **skbuff_vector = qi->skbuff_vector;
941 int header_check;
942
943 /* Refresh the vector and make sure it is with new skbs and the
944 * iovs are updated to point to them.
945 */
946
947 prep_queue_for_rx(qi);
948
949 /* Fire the Lazy Gun - get as many packets as we can in one go. */
950
951 packet_count = uml_vector_recvmmsg(
952 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
953
954 if (packet_count < 0)
955 vp->in_error = true;
956
957 if (packet_count <= 0)
958 return packet_count;
959
960 /* We treat packet processing as enqueue, buffer refresh as dequeue
961 * The queue_depth tells us how many buffers have been used and how
962 * many do we need to prep the next time prep_queue_for_rx() is called.
963 */
964
965 qi->queue_depth = packet_count;
966
967 for (i = 0; i < packet_count; i++) {
968 skb = (*skbuff_vector);
969 if (mmsg_vector->msg_len > vp->header_size) {
970 if (vp->header_size > 0) {
971 header_check = vp->verify_header(
972 mmsg_vector->msg_hdr.msg_iov->iov_base,
973 skb,
974 vp
975 );
976 if (header_check < 0) {
977 /* Overlay header failed to verify - discard.
978 * We can actually keep this skb and reuse it,
979 * but that will make the prep logic too
980 * complex.
981 */
982 dev_kfree_skb_irq(skb);
983 vp->estats.rx_encaps_errors++;
984 continue;
985 }
986 if (header_check > 0) {
987 vp->estats.rx_csum_offload_good++;
988 skb->ip_summed = CHECKSUM_UNNECESSARY;
989 }
990 }
991 pskb_trim(skb,
992 mmsg_vector->msg_len - vp->rx_header_size);
993 skb->protocol = eth_type_trans(skb, skb->dev);
994 /*
995 * We do not need to lock on updating stats here
996 * The interrupt loop is non-reentrant.
997 */
998 vp->dev->stats.rx_bytes += skb->len;
999 vp->dev->stats.rx_packets++;
1000 netif_rx(skb);
1001 } else {
1002 /* Overlay header too short to do anything - discard.
1003 * We can actually keep this skb and reuse it,
1004 * but that will make the prep logic too complex.
1005 */
1006 if (skb != NULL)
1007 dev_kfree_skb_irq(skb);
1008 }
1009 (*skbuff_vector) = NULL;
1010 /* Move to the next buffer element */
1011 mmsg_vector++;
1012 skbuff_vector++;
1013 }
1014 if (packet_count > 0) {
1015 if (vp->estats.rx_queue_max < packet_count)
1016 vp->estats.rx_queue_max = packet_count;
1017 vp->estats.rx_queue_running_average =
1018 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1019 }
1020 return packet_count;
1021}
1022
1023static void vector_rx(struct vector_private *vp)
1024{
1025 int err;
1026 int iter = 0;
1027
1028 if ((vp->options & VECTOR_RX) > 0)
1029 while (((err = vector_mmsg_rx(vp)) > 0) && (iter < MAX_ITERATIONS))
1030 iter++;
1031 else
1032 while (((err = vector_legacy_rx(vp)) > 0) && (iter < MAX_ITERATIONS))
1033 iter++;
1034 if ((err != 0) && net_ratelimit())
1035 netdev_err(vp->dev, "vector_rx: error(%d)\n", err);
1036 if (iter == MAX_ITERATIONS)
1037 netdev_err(vp->dev, "vector_rx: device stuck, remote end may have closed the connection\n");
1038}
1039
1040static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1041{
1042 struct vector_private *vp = netdev_priv(dev);
1043 int queue_depth = 0;
1044
1045 if (vp->in_error) {
1046 deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1047 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1048 deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1049 return NETDEV_TX_BUSY;
1050 }
1051
1052 if ((vp->options & VECTOR_TX) == 0) {
1053 writev_tx(vp, skb);
1054 return NETDEV_TX_OK;
1055 }
1056
1057 /* We do BQL only in the vector path, no point doing it in
1058 * packet at a time mode as there is no device queue
1059 */
1060
1061 netdev_sent_queue(vp->dev, skb->len);
1062 queue_depth = vector_enqueue(vp->tx_queue, skb);
1063
1064 /* if the device queue is full, stop the upper layers and
1065 * flush it.
1066 */
1067
1068 if (queue_depth >= vp->tx_queue->max_depth - 1) {
1069 vp->estats.tx_kicks++;
1070 netif_stop_queue(dev);
1071 vector_send(vp->tx_queue);
1072 return NETDEV_TX_OK;
1073 }
1074 if (netdev_xmit_more()) {
1075 mod_timer(&vp->tl, vp->coalesce);
1076 return NETDEV_TX_OK;
1077 }
1078 if (skb->len < TX_SMALL_PACKET) {
1079 vp->estats.tx_kicks++;
1080 vector_send(vp->tx_queue);
1081 } else
1082 tasklet_schedule(&vp->tx_poll);
1083 return NETDEV_TX_OK;
1084}
1085
1086static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1087{
1088 struct net_device *dev = dev_id;
1089 struct vector_private *vp = netdev_priv(dev);
1090
1091 if (!netif_running(dev))
1092 return IRQ_NONE;
1093 vector_rx(vp);
1094 return IRQ_HANDLED;
1095
1096}
1097
1098static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1099{
1100 struct net_device *dev = dev_id;
1101 struct vector_private *vp = netdev_priv(dev);
1102
1103 if (!netif_running(dev))
1104 return IRQ_NONE;
1105 /* We need to pay attention to it only if we got
1106 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1107 * we ignore it. In the future, it may be worth
1108 * it to improve the IRQ controller a bit to make
1109 * tweaking the IRQ mask less costly
1110 */
1111
1112 if (vp->in_write_poll)
1113 tasklet_schedule(&vp->tx_poll);
1114 return IRQ_HANDLED;
1115
1116}
1117
1118static int irq_rr;
1119
1120static int vector_net_close(struct net_device *dev)
1121{
1122 struct vector_private *vp = netdev_priv(dev);
1123 unsigned long flags;
1124
1125 netif_stop_queue(dev);
1126 del_timer(&vp->tl);
1127
1128 if (vp->fds == NULL)
1129 return 0;
1130
1131 /* Disable and free all IRQS */
1132 if (vp->rx_irq > 0) {
1133 um_free_irq(vp->rx_irq, dev);
1134 vp->rx_irq = 0;
1135 }
1136 if (vp->tx_irq > 0) {
1137 um_free_irq(vp->tx_irq, dev);
1138 vp->tx_irq = 0;
1139 }
1140 tasklet_kill(&vp->tx_poll);
1141 if (vp->fds->rx_fd > 0) {
1142 os_close_file(vp->fds->rx_fd);
1143 vp->fds->rx_fd = -1;
1144 }
1145 if (vp->fds->tx_fd > 0) {
1146 os_close_file(vp->fds->tx_fd);
1147 vp->fds->tx_fd = -1;
1148 }
1149 kfree(vp->bpf);
1150 kfree(vp->fds->remote_addr);
1151 kfree(vp->transport_data);
1152 kfree(vp->header_rxbuffer);
1153 kfree(vp->header_txbuffer);
1154 if (vp->rx_queue != NULL)
1155 destroy_queue(vp->rx_queue);
1156 if (vp->tx_queue != NULL)
1157 destroy_queue(vp->tx_queue);
1158 kfree(vp->fds);
1159 vp->fds = NULL;
1160 spin_lock_irqsave(&vp->lock, flags);
1161 vp->opened = false;
1162 vp->in_error = false;
1163 spin_unlock_irqrestore(&vp->lock, flags);
1164 return 0;
1165}
1166
1167/* TX tasklet */
1168
1169static void vector_tx_poll(unsigned long data)
1170{
1171 struct vector_private *vp = (struct vector_private *)data;
1172
1173 vp->estats.tx_kicks++;
1174 vector_send(vp->tx_queue);
1175}
1176static void vector_reset_tx(struct work_struct *work)
1177{
1178 struct vector_private *vp =
1179 container_of(work, struct vector_private, reset_tx);
1180 netdev_reset_queue(vp->dev);
1181 netif_start_queue(vp->dev);
1182 netif_wake_queue(vp->dev);
1183}
1184static int vector_net_open(struct net_device *dev)
1185{
1186 struct vector_private *vp = netdev_priv(dev);
1187 unsigned long flags;
1188 int err = -EINVAL;
1189 struct vector_device *vdevice;
1190
1191 spin_lock_irqsave(&vp->lock, flags);
1192 if (vp->opened) {
1193 spin_unlock_irqrestore(&vp->lock, flags);
1194 return -ENXIO;
1195 }
1196 vp->opened = true;
1197 spin_unlock_irqrestore(&vp->lock, flags);
1198
1199 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1200
1201 if (vp->fds == NULL)
1202 goto out_close;
1203
1204 if (build_transport_data(vp) < 0)
1205 goto out_close;
1206
1207 if ((vp->options & VECTOR_RX) > 0) {
1208 vp->rx_queue = create_queue(
1209 vp,
1210 get_depth(vp->parsed),
1211 vp->rx_header_size,
1212 MAX_IOV_SIZE
1213 );
1214 vp->rx_queue->queue_depth = get_depth(vp->parsed);
1215 } else {
1216 vp->header_rxbuffer = kmalloc(
1217 vp->rx_header_size,
1218 GFP_KERNEL
1219 );
1220 if (vp->header_rxbuffer == NULL)
1221 goto out_close;
1222 }
1223 if ((vp->options & VECTOR_TX) > 0) {
1224 vp->tx_queue = create_queue(
1225 vp,
1226 get_depth(vp->parsed),
1227 vp->header_size,
1228 MAX_IOV_SIZE
1229 );
1230 } else {
1231 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1232 if (vp->header_txbuffer == NULL)
1233 goto out_close;
1234 }
1235
1236 /* READ IRQ */
1237 err = um_request_irq(
1238 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1239 IRQ_READ, vector_rx_interrupt,
1240 IRQF_SHARED, dev->name, dev);
1241 if (err != 0) {
1242 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1243 err = -ENETUNREACH;
1244 goto out_close;
1245 }
1246 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1247 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1248 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1249
1250 /* WRITE IRQ - we need it only if we have vector TX */
1251 if ((vp->options & VECTOR_TX) > 0) {
1252 err = um_request_irq(
1253 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1254 IRQ_WRITE, vector_tx_interrupt,
1255 IRQF_SHARED, dev->name, dev);
1256 if (err != 0) {
1257 netdev_err(dev,
1258 "vector_open: failed to get tx irq(%d)\n", err);
1259 err = -ENETUNREACH;
1260 goto out_close;
1261 }
1262 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1263 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1264 }
1265
1266 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1267 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1268 vp->options |= VECTOR_BPF;
1269 }
1270 if ((vp->options & VECTOR_BPF) != 0)
1271 vp->bpf = uml_vector_default_bpf(vp->fds->rx_fd, dev->dev_addr);
1272
1273 netif_start_queue(dev);
1274
1275 /* clear buffer - it can happen that the host side of the interface
1276 * is full when we get here. In this case, new data is never queued,
1277 * SIGIOs never arrive, and the net never works.
1278 */
1279
1280 vector_rx(vp);
1281
1282 vector_reset_stats(vp);
1283 vdevice = find_device(vp->unit);
1284 vdevice->opened = 1;
1285
1286 if ((vp->options & VECTOR_TX) != 0)
1287 add_timer(&vp->tl);
1288 return 0;
1289out_close:
1290 vector_net_close(dev);
1291 return err;
1292}
1293
1294
1295static void vector_net_set_multicast_list(struct net_device *dev)
1296{
1297 /* TODO: - we can do some BPF games here */
1298 return;
1299}
1300
1301static void vector_net_tx_timeout(struct net_device *dev)
1302{
1303 struct vector_private *vp = netdev_priv(dev);
1304
1305 vp->estats.tx_timeout_count++;
1306 netif_trans_update(dev);
1307 schedule_work(&vp->reset_tx);
1308}
1309
1310static netdev_features_t vector_fix_features(struct net_device *dev,
1311 netdev_features_t features)
1312{
1313 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1314 return features;
1315}
1316
1317static int vector_set_features(struct net_device *dev,
1318 netdev_features_t features)
1319{
1320 struct vector_private *vp = netdev_priv(dev);
1321 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1322 * no way to negotiate it on raw sockets, so we can change
1323 * only our side.
1324 */
1325 if (features & NETIF_F_GRO)
1326 /* All new frame buffers will be GRO-sized */
1327 vp->req_size = 65536;
1328 else
1329 /* All new frame buffers will be normal sized */
1330 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1331 return 0;
1332}
1333
1334#ifdef CONFIG_NET_POLL_CONTROLLER
1335static void vector_net_poll_controller(struct net_device *dev)
1336{
1337 disable_irq(dev->irq);
1338 vector_rx_interrupt(dev->irq, dev);
1339 enable_irq(dev->irq);
1340}
1341#endif
1342
1343static void vector_net_get_drvinfo(struct net_device *dev,
1344 struct ethtool_drvinfo *info)
1345{
1346 strlcpy(info->driver, DRIVER_NAME, sizeof(info->driver));
1347 strlcpy(info->version, DRIVER_VERSION, sizeof(info->version));
1348}
1349
1350static void vector_get_ringparam(struct net_device *netdev,
1351 struct ethtool_ringparam *ring)
1352{
1353 struct vector_private *vp = netdev_priv(netdev);
1354
1355 ring->rx_max_pending = vp->rx_queue->max_depth;
1356 ring->tx_max_pending = vp->tx_queue->max_depth;
1357 ring->rx_pending = vp->rx_queue->max_depth;
1358 ring->tx_pending = vp->tx_queue->max_depth;
1359}
1360
1361static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1362{
1363 switch (stringset) {
1364 case ETH_SS_TEST:
1365 *buf = '\0';
1366 break;
1367 case ETH_SS_STATS:
1368 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1369 break;
1370 default:
1371 WARN_ON(1);
1372 break;
1373 }
1374}
1375
1376static int vector_get_sset_count(struct net_device *dev, int sset)
1377{
1378 switch (sset) {
1379 case ETH_SS_TEST:
1380 return 0;
1381 case ETH_SS_STATS:
1382 return VECTOR_NUM_STATS;
1383 default:
1384 return -EOPNOTSUPP;
1385 }
1386}
1387
1388static void vector_get_ethtool_stats(struct net_device *dev,
1389 struct ethtool_stats *estats,
1390 u64 *tmp_stats)
1391{
1392 struct vector_private *vp = netdev_priv(dev);
1393
1394 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1395}
1396
1397static int vector_get_coalesce(struct net_device *netdev,
1398 struct ethtool_coalesce *ec)
1399{
1400 struct vector_private *vp = netdev_priv(netdev);
1401
1402 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1403 return 0;
1404}
1405
1406static int vector_set_coalesce(struct net_device *netdev,
1407 struct ethtool_coalesce *ec)
1408{
1409 struct vector_private *vp = netdev_priv(netdev);
1410
1411 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1412 if (vp->coalesce == 0)
1413 vp->coalesce = 1;
1414 return 0;
1415}
1416
1417static const struct ethtool_ops vector_net_ethtool_ops = {
1418 .get_drvinfo = vector_net_get_drvinfo,
1419 .get_link = ethtool_op_get_link,
1420 .get_ts_info = ethtool_op_get_ts_info,
1421 .get_ringparam = vector_get_ringparam,
1422 .get_strings = vector_get_strings,
1423 .get_sset_count = vector_get_sset_count,
1424 .get_ethtool_stats = vector_get_ethtool_stats,
1425 .get_coalesce = vector_get_coalesce,
1426 .set_coalesce = vector_set_coalesce,
1427};
1428
1429
1430static const struct net_device_ops vector_netdev_ops = {
1431 .ndo_open = vector_net_open,
1432 .ndo_stop = vector_net_close,
1433 .ndo_start_xmit = vector_net_start_xmit,
1434 .ndo_set_rx_mode = vector_net_set_multicast_list,
1435 .ndo_tx_timeout = vector_net_tx_timeout,
1436 .ndo_set_mac_address = eth_mac_addr,
1437 .ndo_validate_addr = eth_validate_addr,
1438 .ndo_fix_features = vector_fix_features,
1439 .ndo_set_features = vector_set_features,
1440#ifdef CONFIG_NET_POLL_CONTROLLER
1441 .ndo_poll_controller = vector_net_poll_controller,
1442#endif
1443};
1444
1445
1446static void vector_timer_expire(struct timer_list *t)
1447{
1448 struct vector_private *vp = from_timer(vp, t, tl);
1449
1450 vp->estats.tx_kicks++;
1451 vector_send(vp->tx_queue);
1452}
1453
1454static void vector_eth_configure(
1455 int n,
1456 struct arglist *def
1457 )
1458{
1459 struct vector_device *device;
1460 struct net_device *dev;
1461 struct vector_private *vp;
1462 int err;
1463
1464 device = kzalloc(sizeof(*device), GFP_KERNEL);
1465 if (device == NULL) {
1466 printk(KERN_ERR "eth_configure failed to allocate struct "
1467 "vector_device\n");
1468 return;
1469 }
1470 dev = alloc_etherdev(sizeof(struct vector_private));
1471 if (dev == NULL) {
1472 printk(KERN_ERR "eth_configure: failed to allocate struct "
1473 "net_device for vec%d\n", n);
1474 goto out_free_device;
1475 }
1476
1477 dev->mtu = get_mtu(def);
1478
1479 INIT_LIST_HEAD(&device->list);
1480 device->unit = n;
1481
1482 /* If this name ends up conflicting with an existing registered
1483 * netdevice, that is OK, register_netdev{,ice}() will notice this
1484 * and fail.
1485 */
1486 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1487 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1488 vp = netdev_priv(dev);
1489
1490 /* sysfs register */
1491 if (!driver_registered) {
1492 platform_driver_register(¨_net_driver);
1493 driver_registered = 1;
1494 }
1495 device->pdev.id = n;
1496 device->pdev.name = DRIVER_NAME;
1497 device->pdev.dev.release = vector_device_release;
1498 dev_set_drvdata(&device->pdev.dev, device);
1499 if (platform_device_register(&device->pdev))
1500 goto out_free_netdev;
1501 SET_NETDEV_DEV(dev, &device->pdev.dev);
1502
1503 device->dev = dev;
1504
1505 *vp = ((struct vector_private)
1506 {
1507 .list = LIST_HEAD_INIT(vp->list),
1508 .dev = dev,
1509 .unit = n,
1510 .options = get_transport_options(def),
1511 .rx_irq = 0,
1512 .tx_irq = 0,
1513 .parsed = def,
1514 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1515 /* TODO - we need to calculate headroom so that ip header
1516 * is 16 byte aligned all the time
1517 */
1518 .headroom = get_headroom(def),
1519 .form_header = NULL,
1520 .verify_header = NULL,
1521 .header_rxbuffer = NULL,
1522 .header_txbuffer = NULL,
1523 .header_size = 0,
1524 .rx_header_size = 0,
1525 .rexmit_scheduled = false,
1526 .opened = false,
1527 .transport_data = NULL,
1528 .in_write_poll = false,
1529 .coalesce = 2,
1530 .req_size = get_req_size(def),
1531 .in_error = false
1532 });
1533
1534 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1535 tasklet_init(&vp->tx_poll, vector_tx_poll, (unsigned long)vp);
1536 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1537
1538 timer_setup(&vp->tl, vector_timer_expire, 0);
1539 spin_lock_init(&vp->lock);
1540
1541 /* FIXME */
1542 dev->netdev_ops = &vector_netdev_ops;
1543 dev->ethtool_ops = &vector_net_ethtool_ops;
1544 dev->watchdog_timeo = (HZ >> 1);
1545 /* primary IRQ - fixme */
1546 dev->irq = 0; /* we will adjust this once opened */
1547
1548 rtnl_lock();
1549 err = register_netdevice(dev);
1550 rtnl_unlock();
1551 if (err)
1552 goto out_undo_user_init;
1553
1554 spin_lock(&vector_devices_lock);
1555 list_add(&device->list, &vector_devices);
1556 spin_unlock(&vector_devices_lock);
1557
1558 return;
1559
1560out_undo_user_init:
1561 return;
1562out_free_netdev:
1563 free_netdev(dev);
1564out_free_device:
1565 kfree(device);
1566}
1567
1568
1569
1570
1571/*
1572 * Invoked late in the init
1573 */
1574
1575static int __init vector_init(void)
1576{
1577 struct list_head *ele;
1578 struct vector_cmd_line_arg *def;
1579 struct arglist *parsed;
1580
1581 list_for_each(ele, &vec_cmd_line) {
1582 def = list_entry(ele, struct vector_cmd_line_arg, list);
1583 parsed = uml_parse_vector_ifspec(def->arguments);
1584 if (parsed != NULL)
1585 vector_eth_configure(def->unit, parsed);
1586 }
1587 return 0;
1588}
1589
1590
1591/* Invoked at initial argument parsing, only stores
1592 * arguments until a proper vector_init is called
1593 * later
1594 */
1595
1596static int __init vector_setup(char *str)
1597{
1598 char *error;
1599 int n, err;
1600 struct vector_cmd_line_arg *new;
1601
1602 err = vector_parse(str, &n, &str, &error);
1603 if (err) {
1604 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1605 str, error);
1606 return 1;
1607 }
1608 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1609 if (!new)
1610 panic("%s: Failed to allocate %zu bytes\n", __func__,
1611 sizeof(*new));
1612 INIT_LIST_HEAD(&new->list);
1613 new->unit = n;
1614 new->arguments = str;
1615 list_add_tail(&new->list, &vec_cmd_line);
1616 return 1;
1617}
1618
1619__setup("vec", vector_setup);
1620__uml_help(vector_setup,
1621"vec[0-9]+:<option>=<value>,<option>=<value>\n"
1622" Configure a vector io network device.\n\n"
1623);
1624
1625late_initcall(vector_init);
1626
1627static struct mc_device vector_mc = {
1628 .list = LIST_HEAD_INIT(vector_mc.list),
1629 .name = "vec",
1630 .config = vector_config,
1631 .get_config = NULL,
1632 .id = vector_id,
1633 .remove = vector_remove,
1634};
1635
1636#ifdef CONFIG_INET
1637static int vector_inetaddr_event(
1638 struct notifier_block *this,
1639 unsigned long event,
1640 void *ptr)
1641{
1642 return NOTIFY_DONE;
1643}
1644
1645static struct notifier_block vector_inetaddr_notifier = {
1646 .notifier_call = vector_inetaddr_event,
1647};
1648
1649static void inet_register(void)
1650{
1651 register_inetaddr_notifier(&vector_inetaddr_notifier);
1652}
1653#else
1654static inline void inet_register(void)
1655{
1656}
1657#endif
1658
1659static int vector_net_init(void)
1660{
1661 mconsole_register_dev(&vector_mc);
1662 inet_register();
1663 return 0;
1664}
1665
1666__initcall(vector_net_init);
1667
1668
1669