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