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"
49 struct vector_cmd_line_arg {
50 struct list_head list;
51 int unit;
52 char *arguments;
53 };
54
55 struct 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
63 static LIST_HEAD(vec_cmd_line);
64
65 static DEFINE_SPINLOCK(vector_devices_lock);
66 static LIST_HEAD(vector_devices);
67
68 static int driver_registered;
69
70 static void vector_eth_configure(int n, struct arglist *def);
71 static 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
83 static 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
vector_reset_stats(struct vector_private * vp)104 static 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
get_mtu(struct arglist * def)135 static 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
get_bpf_file(struct arglist * def)148 static char *get_bpf_file(struct arglist *def)
149 {
150 return uml_vector_fetch_arg(def, "bpffile");
151 }
152
get_bpf_flash(struct arglist * def)153 static 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
get_depth(struct arglist * def)165 static 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
get_headroom(struct arglist * def)177 static 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
get_req_size(struct arglist * def)189 static 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
get_transport_options(struct arglist * def)205 static 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
249 static 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
vector_advancehead(struct vector_queue * qi,int advance)258 static 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
vector_advancetail(struct vector_queue * qi,int advance)274 static 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
prep_msg(struct vector_private * vp,struct sk_buff * skb,struct iovec * iov)283 static 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;
315 drop:
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
vector_enqueue(struct vector_queue * qi,struct sk_buff * skb)324 static 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;
358 drop:
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
consume_vector_skbs(struct vector_queue * qi,int count)369 static 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
vector_send(struct vector_queue * qi)397 static 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
destroy_queue(struct vector_queue * qi)465 static 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 */
create_queue(struct vector_private * vp,int max_size,int header_size,int num_extra_frags)505 static 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;
581 out_skb_fail:
582 kfree(result->mmsg_vector);
583 out_mmsg_fail:
584 kfree(result);
585 return NULL;
586 out_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
prep_skb(struct vector_private * vp,struct user_msghdr * msg)600 static 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 }
650 done:
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
prep_queue_for_rx(struct vector_queue * qi)658 static 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
find_device(int n)689 static 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
vector_parse(char * str,int * index_out,char ** str_out,char ** error_out)706 static 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
vector_config(char * str,char ** error_out)737 static 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
vector_id(char ** str,int * start_out,int * end_out)770 static 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
vector_remove(int n,char ** error_out)785 static 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
809 static struct platform_driver uml_net_driver = {
810 .driver = {
811 .name = DRIVER_NAME,
812 },
813 };
814
815
vector_device_release(struct device * dev)816 static void vector_device_release(struct device *dev)
817 {
818 struct vector_device *device = dev_get_drvdata(dev);
819 struct net_device *netdev = device->dev;
820
821 list_del(&device->list);
822 kfree(device);
823 free_netdev(netdev);
824 }
825
826 /* Bog standard recv using recvmsg - not used normally unless the user
827 * explicitly specifies not to use recvmmsg vector RX.
828 */
829
vector_legacy_rx(struct vector_private * vp)830 static int vector_legacy_rx(struct vector_private *vp)
831 {
832 int pkt_len;
833 struct user_msghdr hdr;
834 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
835 int iovpos = 0;
836 struct sk_buff *skb;
837 int header_check;
838
839 hdr.msg_name = NULL;
840 hdr.msg_namelen = 0;
841 hdr.msg_iov = (struct iovec *) &iov;
842 hdr.msg_control = NULL;
843 hdr.msg_controllen = 0;
844 hdr.msg_flags = 0;
845
846 if (vp->header_size > 0) {
847 iov[0].iov_base = vp->header_rxbuffer;
848 iov[0].iov_len = vp->header_size;
849 }
850
851 skb = prep_skb(vp, &hdr);
852
853 if (skb == NULL) {
854 /* Read a packet into drop_buffer and don't do
855 * anything with it.
856 */
857 iov[iovpos].iov_base = drop_buffer;
858 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
859 hdr.msg_iovlen = 1;
860 vp->dev->stats.rx_dropped++;
861 }
862
863 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
864 if (pkt_len < 0) {
865 vp->in_error = true;
866 return pkt_len;
867 }
868
869 if (skb != NULL) {
870 if (pkt_len > vp->header_size) {
871 if (vp->header_size > 0) {
872 header_check = vp->verify_header(
873 vp->header_rxbuffer, skb, vp);
874 if (header_check < 0) {
875 dev_kfree_skb_irq(skb);
876 vp->dev->stats.rx_dropped++;
877 vp->estats.rx_encaps_errors++;
878 return 0;
879 }
880 if (header_check > 0) {
881 vp->estats.rx_csum_offload_good++;
882 skb->ip_summed = CHECKSUM_UNNECESSARY;
883 }
884 }
885 pskb_trim(skb, pkt_len - vp->rx_header_size);
886 skb->protocol = eth_type_trans(skb, skb->dev);
887 vp->dev->stats.rx_bytes += skb->len;
888 vp->dev->stats.rx_packets++;
889 napi_gro_receive(&vp->napi, skb);
890 } else {
891 dev_kfree_skb_irq(skb);
892 }
893 }
894 return pkt_len;
895 }
896
897 /*
898 * Packet at a time TX which falls back to vector TX if the
899 * underlying transport is busy.
900 */
901
902
903
writev_tx(struct vector_private * vp,struct sk_buff * skb)904 static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
905 {
906 struct iovec iov[3 + MAX_IOV_SIZE];
907 int iov_count, pkt_len = 0;
908
909 iov[0].iov_base = vp->header_txbuffer;
910 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
911
912 if (iov_count < 1)
913 goto drop;
914
915 pkt_len = uml_vector_writev(
916 vp->fds->tx_fd,
917 (struct iovec *) &iov,
918 iov_count
919 );
920
921 if (pkt_len < 0)
922 goto drop;
923
924 netif_trans_update(vp->dev);
925 netif_wake_queue(vp->dev);
926
927 if (pkt_len > 0) {
928 vp->dev->stats.tx_bytes += skb->len;
929 vp->dev->stats.tx_packets++;
930 } else {
931 vp->dev->stats.tx_dropped++;
932 }
933 consume_skb(skb);
934 return pkt_len;
935 drop:
936 vp->dev->stats.tx_dropped++;
937 consume_skb(skb);
938 if (pkt_len < 0)
939 vp->in_error = true;
940 return pkt_len;
941 }
942
943 /*
944 * Receive as many messages as we can in one call using the special
945 * mmsg vector matched to an skb vector which we prepared earlier.
946 */
947
vector_mmsg_rx(struct vector_private * vp,int budget)948 static int vector_mmsg_rx(struct vector_private *vp, int budget)
949 {
950 int packet_count, i;
951 struct vector_queue *qi = vp->rx_queue;
952 struct sk_buff *skb;
953 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
954 void **skbuff_vector = qi->skbuff_vector;
955 int header_check;
956
957 /* Refresh the vector and make sure it is with new skbs and the
958 * iovs are updated to point to them.
959 */
960
961 prep_queue_for_rx(qi);
962
963 /* Fire the Lazy Gun - get as many packets as we can in one go. */
964
965 if (budget > qi->max_depth)
966 budget = qi->max_depth;
967
968 packet_count = uml_vector_recvmmsg(
969 vp->fds->rx_fd, qi->mmsg_vector, budget, 0);
970
971 if (packet_count < 0)
972 vp->in_error = true;
973
974 if (packet_count <= 0)
975 return packet_count;
976
977 /* We treat packet processing as enqueue, buffer refresh as dequeue
978 * The queue_depth tells us how many buffers have been used and how
979 * many do we need to prep the next time prep_queue_for_rx() is called.
980 */
981
982 atomic_add(packet_count, &qi->queue_depth);
983
984 for (i = 0; i < packet_count; i++) {
985 skb = (*skbuff_vector);
986 if (mmsg_vector->msg_len > vp->header_size) {
987 if (vp->header_size > 0) {
988 header_check = vp->verify_header(
989 mmsg_vector->msg_hdr.msg_iov->iov_base,
990 skb,
991 vp
992 );
993 if (header_check < 0) {
994 /* Overlay header failed to verify - discard.
995 * We can actually keep this skb and reuse it,
996 * but that will make the prep logic too
997 * complex.
998 */
999 dev_kfree_skb_irq(skb);
1000 vp->estats.rx_encaps_errors++;
1001 continue;
1002 }
1003 if (header_check > 0) {
1004 vp->estats.rx_csum_offload_good++;
1005 skb->ip_summed = CHECKSUM_UNNECESSARY;
1006 }
1007 }
1008 pskb_trim(skb,
1009 mmsg_vector->msg_len - vp->rx_header_size);
1010 skb->protocol = eth_type_trans(skb, skb->dev);
1011 /*
1012 * We do not need to lock on updating stats here
1013 * The interrupt loop is non-reentrant.
1014 */
1015 vp->dev->stats.rx_bytes += skb->len;
1016 vp->dev->stats.rx_packets++;
1017 napi_gro_receive(&vp->napi, skb);
1018 } else {
1019 /* Overlay header too short to do anything - discard.
1020 * We can actually keep this skb and reuse it,
1021 * but that will make the prep logic too complex.
1022 */
1023 if (skb != NULL)
1024 dev_kfree_skb_irq(skb);
1025 }
1026 (*skbuff_vector) = NULL;
1027 /* Move to the next buffer element */
1028 mmsg_vector++;
1029 skbuff_vector++;
1030 }
1031 if (packet_count > 0) {
1032 if (vp->estats.rx_queue_max < packet_count)
1033 vp->estats.rx_queue_max = packet_count;
1034 vp->estats.rx_queue_running_average =
1035 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1036 }
1037 return packet_count;
1038 }
1039
vector_net_start_xmit(struct sk_buff * skb,struct net_device * dev)1040 static 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 (queue_depth < vp->tx_queue->max_depth && netdev_xmit_more()) {
1065 mod_timer(&vp->tl, vp->coalesce);
1066 return NETDEV_TX_OK;
1067 } else {
1068 queue_depth = vector_send(vp->tx_queue);
1069 if (queue_depth > 0)
1070 napi_schedule(&vp->napi);
1071 }
1072
1073 return NETDEV_TX_OK;
1074 }
1075
vector_rx_interrupt(int irq,void * dev_id)1076 static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1077 {
1078 struct net_device *dev = dev_id;
1079 struct vector_private *vp = netdev_priv(dev);
1080
1081 if (!netif_running(dev))
1082 return IRQ_NONE;
1083 napi_schedule(&vp->napi);
1084 return IRQ_HANDLED;
1085
1086 }
1087
vector_tx_interrupt(int irq,void * dev_id)1088 static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1089 {
1090 struct net_device *dev = dev_id;
1091 struct vector_private *vp = netdev_priv(dev);
1092
1093 if (!netif_running(dev))
1094 return IRQ_NONE;
1095 /* We need to pay attention to it only if we got
1096 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1097 * we ignore it. In the future, it may be worth
1098 * it to improve the IRQ controller a bit to make
1099 * tweaking the IRQ mask less costly
1100 */
1101
1102 napi_schedule(&vp->napi);
1103 return IRQ_HANDLED;
1104
1105 }
1106
1107 static int irq_rr;
1108
vector_net_close(struct net_device * dev)1109 static int vector_net_close(struct net_device *dev)
1110 {
1111 struct vector_private *vp = netdev_priv(dev);
1112
1113 netif_stop_queue(dev);
1114 del_timer(&vp->tl);
1115
1116 vp->opened = false;
1117
1118 if (vp->fds == NULL)
1119 return 0;
1120
1121 /* Disable and free all IRQS */
1122 if (vp->rx_irq > 0) {
1123 um_free_irq(vp->rx_irq, dev);
1124 vp->rx_irq = 0;
1125 }
1126 if (vp->tx_irq > 0) {
1127 um_free_irq(vp->tx_irq, dev);
1128 vp->tx_irq = 0;
1129 }
1130 napi_disable(&vp->napi);
1131 netif_napi_del(&vp->napi);
1132 if (vp->fds->rx_fd > 0) {
1133 if (vp->bpf)
1134 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1135 os_close_file(vp->fds->rx_fd);
1136 vp->fds->rx_fd = -1;
1137 }
1138 if (vp->fds->tx_fd > 0) {
1139 os_close_file(vp->fds->tx_fd);
1140 vp->fds->tx_fd = -1;
1141 }
1142 if (vp->bpf != NULL)
1143 kfree(vp->bpf->filter);
1144 kfree(vp->bpf);
1145 vp->bpf = NULL;
1146 kfree(vp->fds->remote_addr);
1147 kfree(vp->transport_data);
1148 kfree(vp->header_rxbuffer);
1149 kfree(vp->header_txbuffer);
1150 if (vp->rx_queue != NULL)
1151 destroy_queue(vp->rx_queue);
1152 if (vp->tx_queue != NULL)
1153 destroy_queue(vp->tx_queue);
1154 kfree(vp->fds);
1155 vp->fds = NULL;
1156 vp->in_error = false;
1157 return 0;
1158 }
1159
vector_poll(struct napi_struct * napi,int budget)1160 static int vector_poll(struct napi_struct *napi, int budget)
1161 {
1162 struct vector_private *vp = container_of(napi, struct vector_private, napi);
1163 int work_done = 0;
1164 int err;
1165 bool tx_enqueued = false;
1166
1167 if ((vp->options & VECTOR_TX) != 0)
1168 tx_enqueued = (vector_send(vp->tx_queue) > 0);
1169 spin_lock(&vp->rx_queue->head_lock);
1170 if ((vp->options & VECTOR_RX) > 0)
1171 err = vector_mmsg_rx(vp, budget);
1172 else {
1173 err = vector_legacy_rx(vp);
1174 if (err > 0)
1175 err = 1;
1176 }
1177 spin_unlock(&vp->rx_queue->head_lock);
1178 if (err > 0)
1179 work_done += err;
1180
1181 if (tx_enqueued || err > 0)
1182 napi_schedule(napi);
1183 if (work_done <= budget)
1184 napi_complete_done(napi, work_done);
1185 return work_done;
1186 }
1187
vector_reset_tx(struct work_struct * work)1188 static void vector_reset_tx(struct work_struct *work)
1189 {
1190 struct vector_private *vp =
1191 container_of(work, struct vector_private, reset_tx);
1192 netdev_reset_queue(vp->dev);
1193 netif_start_queue(vp->dev);
1194 netif_wake_queue(vp->dev);
1195 }
1196
vector_net_open(struct net_device * dev)1197 static int vector_net_open(struct net_device *dev)
1198 {
1199 struct vector_private *vp = netdev_priv(dev);
1200 int err = -EINVAL;
1201 struct vector_device *vdevice;
1202
1203 if (vp->opened)
1204 return -ENXIO;
1205 vp->opened = true;
1206
1207 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1208
1209 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1210
1211 if (vp->fds == NULL)
1212 goto out_close;
1213
1214 if (build_transport_data(vp) < 0)
1215 goto out_close;
1216
1217 if ((vp->options & VECTOR_RX) > 0) {
1218 vp->rx_queue = create_queue(
1219 vp,
1220 get_depth(vp->parsed),
1221 vp->rx_header_size,
1222 MAX_IOV_SIZE
1223 );
1224 atomic_set(&vp->rx_queue->queue_depth, get_depth(vp->parsed));
1225 } else {
1226 vp->header_rxbuffer = kmalloc(
1227 vp->rx_header_size,
1228 GFP_KERNEL
1229 );
1230 if (vp->header_rxbuffer == NULL)
1231 goto out_close;
1232 }
1233 if ((vp->options & VECTOR_TX) > 0) {
1234 vp->tx_queue = create_queue(
1235 vp,
1236 get_depth(vp->parsed),
1237 vp->header_size,
1238 MAX_IOV_SIZE
1239 );
1240 } else {
1241 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1242 if (vp->header_txbuffer == NULL)
1243 goto out_close;
1244 }
1245
1246 netif_napi_add_weight(vp->dev, &vp->napi, vector_poll,
1247 get_depth(vp->parsed));
1248 napi_enable(&vp->napi);
1249
1250 /* READ IRQ */
1251 err = um_request_irq(
1252 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1253 IRQ_READ, vector_rx_interrupt,
1254 IRQF_SHARED, dev->name, dev);
1255 if (err < 0) {
1256 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1257 err = -ENETUNREACH;
1258 goto out_close;
1259 }
1260 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1261 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1262 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1263
1264 /* WRITE IRQ - we need it only if we have vector TX */
1265 if ((vp->options & VECTOR_TX) > 0) {
1266 err = um_request_irq(
1267 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1268 IRQ_WRITE, vector_tx_interrupt,
1269 IRQF_SHARED, dev->name, dev);
1270 if (err < 0) {
1271 netdev_err(dev,
1272 "vector_open: failed to get tx irq(%d)\n", err);
1273 err = -ENETUNREACH;
1274 goto out_close;
1275 }
1276 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1277 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1278 }
1279
1280 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1281 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1282 vp->options |= VECTOR_BPF;
1283 }
1284 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1285 vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1286
1287 if (vp->bpf != NULL)
1288 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1289
1290 netif_start_queue(dev);
1291 vector_reset_stats(vp);
1292
1293 /* clear buffer - it can happen that the host side of the interface
1294 * is full when we get here. In this case, new data is never queued,
1295 * SIGIOs never arrive, and the net never works.
1296 */
1297
1298 napi_schedule(&vp->napi);
1299
1300 vdevice = find_device(vp->unit);
1301 vdevice->opened = 1;
1302
1303 if ((vp->options & VECTOR_TX) != 0)
1304 add_timer(&vp->tl);
1305 return 0;
1306 out_close:
1307 vector_net_close(dev);
1308 return err;
1309 }
1310
1311
vector_net_set_multicast_list(struct net_device * dev)1312 static void vector_net_set_multicast_list(struct net_device *dev)
1313 {
1314 /* TODO: - we can do some BPF games here */
1315 return;
1316 }
1317
vector_net_tx_timeout(struct net_device * dev,unsigned int txqueue)1318 static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1319 {
1320 struct vector_private *vp = netdev_priv(dev);
1321
1322 vp->estats.tx_timeout_count++;
1323 netif_trans_update(dev);
1324 schedule_work(&vp->reset_tx);
1325 }
1326
vector_fix_features(struct net_device * dev,netdev_features_t features)1327 static netdev_features_t vector_fix_features(struct net_device *dev,
1328 netdev_features_t features)
1329 {
1330 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1331 return features;
1332 }
1333
vector_set_features(struct net_device * dev,netdev_features_t features)1334 static int vector_set_features(struct net_device *dev,
1335 netdev_features_t features)
1336 {
1337 struct vector_private *vp = netdev_priv(dev);
1338 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1339 * no way to negotiate it on raw sockets, so we can change
1340 * only our side.
1341 */
1342 if (features & NETIF_F_GRO)
1343 /* All new frame buffers will be GRO-sized */
1344 vp->req_size = 65536;
1345 else
1346 /* All new frame buffers will be normal sized */
1347 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1348 return 0;
1349 }
1350
1351 #ifdef CONFIG_NET_POLL_CONTROLLER
vector_net_poll_controller(struct net_device * dev)1352 static void vector_net_poll_controller(struct net_device *dev)
1353 {
1354 disable_irq(dev->irq);
1355 vector_rx_interrupt(dev->irq, dev);
1356 enable_irq(dev->irq);
1357 }
1358 #endif
1359
vector_net_get_drvinfo(struct net_device * dev,struct ethtool_drvinfo * info)1360 static void vector_net_get_drvinfo(struct net_device *dev,
1361 struct ethtool_drvinfo *info)
1362 {
1363 strscpy(info->driver, DRIVER_NAME);
1364 }
1365
vector_net_load_bpf_flash(struct net_device * dev,struct ethtool_flash * efl)1366 static int vector_net_load_bpf_flash(struct net_device *dev,
1367 struct ethtool_flash *efl)
1368 {
1369 struct vector_private *vp = netdev_priv(dev);
1370 struct vector_device *vdevice;
1371 const struct firmware *fw;
1372 int result = 0;
1373
1374 if (!(vp->options & VECTOR_BPF_FLASH)) {
1375 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1376 return -1;
1377 }
1378
1379 if (vp->bpf != NULL) {
1380 if (vp->opened)
1381 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1382 kfree(vp->bpf->filter);
1383 vp->bpf->filter = NULL;
1384 } else {
1385 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC);
1386 if (vp->bpf == NULL) {
1387 netdev_err(dev, "failed to allocate memory for firmware\n");
1388 goto flash_fail;
1389 }
1390 }
1391
1392 vdevice = find_device(vp->unit);
1393
1394 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1395 goto flash_fail;
1396
1397 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC);
1398 if (!vp->bpf->filter)
1399 goto free_buffer;
1400
1401 vp->bpf->len = fw->size / sizeof(struct sock_filter);
1402 release_firmware(fw);
1403
1404 if (vp->opened)
1405 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1406
1407 return result;
1408
1409 free_buffer:
1410 release_firmware(fw);
1411
1412 flash_fail:
1413 if (vp->bpf != NULL)
1414 kfree(vp->bpf->filter);
1415 kfree(vp->bpf);
1416 vp->bpf = NULL;
1417 return -1;
1418 }
1419
vector_get_ringparam(struct net_device * netdev,struct ethtool_ringparam * ring,struct kernel_ethtool_ringparam * kernel_ring,struct netlink_ext_ack * extack)1420 static void vector_get_ringparam(struct net_device *netdev,
1421 struct ethtool_ringparam *ring,
1422 struct kernel_ethtool_ringparam *kernel_ring,
1423 struct netlink_ext_ack *extack)
1424 {
1425 struct vector_private *vp = netdev_priv(netdev);
1426
1427 ring->rx_max_pending = vp->rx_queue->max_depth;
1428 ring->tx_max_pending = vp->tx_queue->max_depth;
1429 ring->rx_pending = vp->rx_queue->max_depth;
1430 ring->tx_pending = vp->tx_queue->max_depth;
1431 }
1432
vector_get_strings(struct net_device * dev,u32 stringset,u8 * buf)1433 static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1434 {
1435 switch (stringset) {
1436 case ETH_SS_TEST:
1437 *buf = '\0';
1438 break;
1439 case ETH_SS_STATS:
1440 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1441 break;
1442 default:
1443 WARN_ON(1);
1444 break;
1445 }
1446 }
1447
vector_get_sset_count(struct net_device * dev,int sset)1448 static int vector_get_sset_count(struct net_device *dev, int sset)
1449 {
1450 switch (sset) {
1451 case ETH_SS_TEST:
1452 return 0;
1453 case ETH_SS_STATS:
1454 return VECTOR_NUM_STATS;
1455 default:
1456 return -EOPNOTSUPP;
1457 }
1458 }
1459
vector_get_ethtool_stats(struct net_device * dev,struct ethtool_stats * estats,u64 * tmp_stats)1460 static void vector_get_ethtool_stats(struct net_device *dev,
1461 struct ethtool_stats *estats,
1462 u64 *tmp_stats)
1463 {
1464 struct vector_private *vp = netdev_priv(dev);
1465
1466 /* Stats are modified in the dequeue portions of
1467 * rx/tx which are protected by the head locks
1468 * grabbing these locks here ensures they are up
1469 * to date.
1470 */
1471
1472 spin_lock(&vp->tx_queue->head_lock);
1473 spin_lock(&vp->rx_queue->head_lock);
1474 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1475 spin_unlock(&vp->rx_queue->head_lock);
1476 spin_unlock(&vp->tx_queue->head_lock);
1477 }
1478
vector_get_coalesce(struct net_device * netdev,struct ethtool_coalesce * ec,struct kernel_ethtool_coalesce * kernel_coal,struct netlink_ext_ack * extack)1479 static int vector_get_coalesce(struct net_device *netdev,
1480 struct ethtool_coalesce *ec,
1481 struct kernel_ethtool_coalesce *kernel_coal,
1482 struct netlink_ext_ack *extack)
1483 {
1484 struct vector_private *vp = netdev_priv(netdev);
1485
1486 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1487 return 0;
1488 }
1489
vector_set_coalesce(struct net_device * netdev,struct ethtool_coalesce * ec,struct kernel_ethtool_coalesce * kernel_coal,struct netlink_ext_ack * extack)1490 static int vector_set_coalesce(struct net_device *netdev,
1491 struct ethtool_coalesce *ec,
1492 struct kernel_ethtool_coalesce *kernel_coal,
1493 struct netlink_ext_ack *extack)
1494 {
1495 struct vector_private *vp = netdev_priv(netdev);
1496
1497 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1498 if (vp->coalesce == 0)
1499 vp->coalesce = 1;
1500 return 0;
1501 }
1502
1503 static const struct ethtool_ops vector_net_ethtool_ops = {
1504 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1505 .get_drvinfo = vector_net_get_drvinfo,
1506 .get_link = ethtool_op_get_link,
1507 .get_ts_info = ethtool_op_get_ts_info,
1508 .get_ringparam = vector_get_ringparam,
1509 .get_strings = vector_get_strings,
1510 .get_sset_count = vector_get_sset_count,
1511 .get_ethtool_stats = vector_get_ethtool_stats,
1512 .get_coalesce = vector_get_coalesce,
1513 .set_coalesce = vector_set_coalesce,
1514 .flash_device = vector_net_load_bpf_flash,
1515 };
1516
1517
1518 static const struct net_device_ops vector_netdev_ops = {
1519 .ndo_open = vector_net_open,
1520 .ndo_stop = vector_net_close,
1521 .ndo_start_xmit = vector_net_start_xmit,
1522 .ndo_set_rx_mode = vector_net_set_multicast_list,
1523 .ndo_tx_timeout = vector_net_tx_timeout,
1524 .ndo_set_mac_address = eth_mac_addr,
1525 .ndo_validate_addr = eth_validate_addr,
1526 .ndo_fix_features = vector_fix_features,
1527 .ndo_set_features = vector_set_features,
1528 #ifdef CONFIG_NET_POLL_CONTROLLER
1529 .ndo_poll_controller = vector_net_poll_controller,
1530 #endif
1531 };
1532
vector_timer_expire(struct timer_list * t)1533 static void vector_timer_expire(struct timer_list *t)
1534 {
1535 struct vector_private *vp = from_timer(vp, t, tl);
1536
1537 vp->estats.tx_kicks++;
1538 napi_schedule(&vp->napi);
1539 }
1540
1541
1542
vector_eth_configure(int n,struct arglist * def)1543 static void vector_eth_configure(
1544 int n,
1545 struct arglist *def
1546 )
1547 {
1548 struct vector_device *device;
1549 struct net_device *dev;
1550 struct vector_private *vp;
1551 int err;
1552
1553 device = kzalloc(sizeof(*device), GFP_KERNEL);
1554 if (device == NULL) {
1555 printk(KERN_ERR "eth_configure failed to allocate struct "
1556 "vector_device\n");
1557 return;
1558 }
1559 dev = alloc_etherdev(sizeof(struct vector_private));
1560 if (dev == NULL) {
1561 printk(KERN_ERR "eth_configure: failed to allocate struct "
1562 "net_device for vec%d\n", n);
1563 goto out_free_device;
1564 }
1565
1566 dev->mtu = get_mtu(def);
1567
1568 INIT_LIST_HEAD(&device->list);
1569 device->unit = n;
1570
1571 /* If this name ends up conflicting with an existing registered
1572 * netdevice, that is OK, register_netdev{,ice}() will notice this
1573 * and fail.
1574 */
1575 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1576 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1577 vp = netdev_priv(dev);
1578
1579 /* sysfs register */
1580 if (!driver_registered) {
1581 platform_driver_register(¨_net_driver);
1582 driver_registered = 1;
1583 }
1584 device->pdev.id = n;
1585 device->pdev.name = DRIVER_NAME;
1586 device->pdev.dev.release = vector_device_release;
1587 dev_set_drvdata(&device->pdev.dev, device);
1588 if (platform_device_register(&device->pdev))
1589 goto out_free_netdev;
1590 SET_NETDEV_DEV(dev, &device->pdev.dev);
1591
1592 device->dev = dev;
1593
1594 *vp = ((struct vector_private)
1595 {
1596 .list = LIST_HEAD_INIT(vp->list),
1597 .dev = dev,
1598 .unit = n,
1599 .options = get_transport_options(def),
1600 .rx_irq = 0,
1601 .tx_irq = 0,
1602 .parsed = def,
1603 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1604 /* TODO - we need to calculate headroom so that ip header
1605 * is 16 byte aligned all the time
1606 */
1607 .headroom = get_headroom(def),
1608 .form_header = NULL,
1609 .verify_header = NULL,
1610 .header_rxbuffer = NULL,
1611 .header_txbuffer = NULL,
1612 .header_size = 0,
1613 .rx_header_size = 0,
1614 .rexmit_scheduled = false,
1615 .opened = false,
1616 .transport_data = NULL,
1617 .in_write_poll = false,
1618 .coalesce = 2,
1619 .req_size = get_req_size(def),
1620 .in_error = false,
1621 .bpf = NULL
1622 });
1623
1624 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1625 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1626
1627 timer_setup(&vp->tl, vector_timer_expire, 0);
1628
1629 /* FIXME */
1630 dev->netdev_ops = &vector_netdev_ops;
1631 dev->ethtool_ops = &vector_net_ethtool_ops;
1632 dev->watchdog_timeo = (HZ >> 1);
1633 /* primary IRQ - fixme */
1634 dev->irq = 0; /* we will adjust this once opened */
1635
1636 rtnl_lock();
1637 err = register_netdevice(dev);
1638 rtnl_unlock();
1639 if (err)
1640 goto out_undo_user_init;
1641
1642 spin_lock(&vector_devices_lock);
1643 list_add(&device->list, &vector_devices);
1644 spin_unlock(&vector_devices_lock);
1645
1646 return;
1647
1648 out_undo_user_init:
1649 return;
1650 out_free_netdev:
1651 free_netdev(dev);
1652 out_free_device:
1653 kfree(device);
1654 }
1655
1656
1657
1658
1659 /*
1660 * Invoked late in the init
1661 */
1662
vector_init(void)1663 static int __init vector_init(void)
1664 {
1665 struct list_head *ele;
1666 struct vector_cmd_line_arg *def;
1667 struct arglist *parsed;
1668
1669 list_for_each(ele, &vec_cmd_line) {
1670 def = list_entry(ele, struct vector_cmd_line_arg, list);
1671 parsed = uml_parse_vector_ifspec(def->arguments);
1672 if (parsed != NULL)
1673 vector_eth_configure(def->unit, parsed);
1674 }
1675 return 0;
1676 }
1677
1678
1679 /* Invoked at initial argument parsing, only stores
1680 * arguments until a proper vector_init is called
1681 * later
1682 */
1683
vector_setup(char * str)1684 static int __init vector_setup(char *str)
1685 {
1686 char *error;
1687 int n, err;
1688 struct vector_cmd_line_arg *new;
1689
1690 err = vector_parse(str, &n, &str, &error);
1691 if (err) {
1692 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1693 str, error);
1694 return 1;
1695 }
1696 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1697 if (!new)
1698 panic("%s: Failed to allocate %zu bytes\n", __func__,
1699 sizeof(*new));
1700 INIT_LIST_HEAD(&new->list);
1701 new->unit = n;
1702 new->arguments = str;
1703 list_add_tail(&new->list, &vec_cmd_line);
1704 return 1;
1705 }
1706
1707 __setup("vec", vector_setup);
1708 __uml_help(vector_setup,
1709 "vec[0-9]+:<option>=<value>,<option>=<value>\n"
1710 " Configure a vector io network device.\n\n"
1711 );
1712
1713 late_initcall(vector_init);
1714
1715 static struct mc_device vector_mc = {
1716 .list = LIST_HEAD_INIT(vector_mc.list),
1717 .name = "vec",
1718 .config = vector_config,
1719 .get_config = NULL,
1720 .id = vector_id,
1721 .remove = vector_remove,
1722 };
1723
1724 #ifdef CONFIG_INET
vector_inetaddr_event(struct notifier_block * this,unsigned long event,void * ptr)1725 static int vector_inetaddr_event(
1726 struct notifier_block *this,
1727 unsigned long event,
1728 void *ptr)
1729 {
1730 return NOTIFY_DONE;
1731 }
1732
1733 static struct notifier_block vector_inetaddr_notifier = {
1734 .notifier_call = vector_inetaddr_event,
1735 };
1736
inet_register(void)1737 static void inet_register(void)
1738 {
1739 register_inetaddr_notifier(&vector_inetaddr_notifier);
1740 }
1741 #else
inet_register(void)1742 static inline void inet_register(void)
1743 {
1744 }
1745 #endif
1746
vector_net_init(void)1747 static int vector_net_init(void)
1748 {
1749 mconsole_register_dev(&vector_mc);
1750 inet_register();
1751 return 0;
1752 }
1753
1754 __initcall(vector_net_init);
1755
1756
1757
1758