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, len, err;
716 char *start = str;
717
718 len = strlen(str);
719
720 while ((*str != ':') && (strlen(str) > 1))
721 str++;
722 if (*str != ':') {
723 *error_out = "Expected ':' after device number";
724 return -EINVAL;
725 }
726 *str = '\0';
727
728 err = kstrtouint(start, 0, &n);
729 if (err < 0) {
730 *error_out = "Bad device number";
731 return err;
732 }
733
734 str++;
735 if (find_device(n)) {
736 *error_out = "Device already configured";
737 return -EINVAL;
738 }
739
740 *index_out = n;
741 *str_out = str;
742 return 0;
743 }
744
vector_config(char * str,char ** error_out)745 static int vector_config(char *str, char **error_out)
746 {
747 int err, n;
748 char *params;
749 struct arglist *parsed;
750
751 err = vector_parse(str, &n, ¶ms, error_out);
752 if (err != 0)
753 return err;
754
755 /* This string is broken up and the pieces used by the underlying
756 * driver. We should copy it to make sure things do not go wrong
757 * later.
758 */
759
760 params = kstrdup(params, GFP_KERNEL);
761 if (params == NULL) {
762 *error_out = "vector_config failed to strdup string";
763 return -ENOMEM;
764 }
765
766 parsed = uml_parse_vector_ifspec(params);
767
768 if (parsed == NULL) {
769 *error_out = "vector_config failed to parse parameters";
770 kfree(params);
771 return -EINVAL;
772 }
773
774 vector_eth_configure(n, parsed);
775 return 0;
776 }
777
vector_id(char ** str,int * start_out,int * end_out)778 static int vector_id(char **str, int *start_out, int *end_out)
779 {
780 char *end;
781 int n;
782
783 n = simple_strtoul(*str, &end, 0);
784 if ((*end != '\0') || (end == *str))
785 return -1;
786
787 *start_out = n;
788 *end_out = n;
789 *str = end;
790 return n;
791 }
792
vector_remove(int n,char ** error_out)793 static int vector_remove(int n, char **error_out)
794 {
795 struct vector_device *vec_d;
796 struct net_device *dev;
797 struct vector_private *vp;
798
799 vec_d = find_device(n);
800 if (vec_d == NULL)
801 return -ENODEV;
802 dev = vec_d->dev;
803 vp = netdev_priv(dev);
804 if (vp->fds != NULL)
805 return -EBUSY;
806 unregister_netdev(dev);
807 platform_device_unregister(&vec_d->pdev);
808 return 0;
809 }
810
811 /*
812 * There is no shared per-transport initialization code, so
813 * we will just initialize each interface one by one and
814 * add them to a list
815 */
816
817 static struct platform_driver uml_net_driver = {
818 .driver = {
819 .name = DRIVER_NAME,
820 },
821 };
822
823
vector_device_release(struct device * dev)824 static void vector_device_release(struct device *dev)
825 {
826 struct vector_device *device =
827 container_of(dev, struct vector_device, pdev.dev);
828 struct net_device *netdev = device->dev;
829
830 list_del(&device->list);
831 kfree(device);
832 free_netdev(netdev);
833 }
834
835 /* Bog standard recv using recvmsg - not used normally unless the user
836 * explicitly specifies not to use recvmmsg vector RX.
837 */
838
vector_legacy_rx(struct vector_private * vp)839 static int vector_legacy_rx(struct vector_private *vp)
840 {
841 int pkt_len;
842 struct user_msghdr hdr;
843 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
844 int iovpos = 0;
845 struct sk_buff *skb;
846 int header_check;
847
848 hdr.msg_name = NULL;
849 hdr.msg_namelen = 0;
850 hdr.msg_iov = (struct iovec *) &iov;
851 hdr.msg_control = NULL;
852 hdr.msg_controllen = 0;
853 hdr.msg_flags = 0;
854
855 if (vp->header_size > 0) {
856 iov[0].iov_base = vp->header_rxbuffer;
857 iov[0].iov_len = vp->header_size;
858 }
859
860 skb = prep_skb(vp, &hdr);
861
862 if (skb == NULL) {
863 /* Read a packet into drop_buffer and don't do
864 * anything with it.
865 */
866 iov[iovpos].iov_base = drop_buffer;
867 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
868 hdr.msg_iovlen = 1;
869 vp->dev->stats.rx_dropped++;
870 }
871
872 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
873 if (pkt_len < 0) {
874 vp->in_error = true;
875 return pkt_len;
876 }
877
878 if (skb != NULL) {
879 if (pkt_len > vp->header_size) {
880 if (vp->header_size > 0) {
881 header_check = vp->verify_header(
882 vp->header_rxbuffer, skb, vp);
883 if (header_check < 0) {
884 dev_kfree_skb_irq(skb);
885 vp->dev->stats.rx_dropped++;
886 vp->estats.rx_encaps_errors++;
887 return 0;
888 }
889 if (header_check > 0) {
890 vp->estats.rx_csum_offload_good++;
891 skb->ip_summed = CHECKSUM_UNNECESSARY;
892 }
893 }
894 pskb_trim(skb, pkt_len - vp->rx_header_size);
895 skb->protocol = eth_type_trans(skb, skb->dev);
896 vp->dev->stats.rx_bytes += skb->len;
897 vp->dev->stats.rx_packets++;
898 napi_gro_receive(&vp->napi, skb);
899 } else {
900 dev_kfree_skb_irq(skb);
901 }
902 }
903 return pkt_len;
904 }
905
906 /*
907 * Packet at a time TX which falls back to vector TX if the
908 * underlying transport is busy.
909 */
910
911
912
writev_tx(struct vector_private * vp,struct sk_buff * skb)913 static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
914 {
915 struct iovec iov[3 + MAX_IOV_SIZE];
916 int iov_count, pkt_len = 0;
917
918 iov[0].iov_base = vp->header_txbuffer;
919 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
920
921 if (iov_count < 1)
922 goto drop;
923
924 pkt_len = uml_vector_writev(
925 vp->fds->tx_fd,
926 (struct iovec *) &iov,
927 iov_count
928 );
929
930 if (pkt_len < 0)
931 goto drop;
932
933 netif_trans_update(vp->dev);
934 netif_wake_queue(vp->dev);
935
936 if (pkt_len > 0) {
937 vp->dev->stats.tx_bytes += skb->len;
938 vp->dev->stats.tx_packets++;
939 } else {
940 vp->dev->stats.tx_dropped++;
941 }
942 consume_skb(skb);
943 return pkt_len;
944 drop:
945 vp->dev->stats.tx_dropped++;
946 consume_skb(skb);
947 if (pkt_len < 0)
948 vp->in_error = true;
949 return pkt_len;
950 }
951
952 /*
953 * Receive as many messages as we can in one call using the special
954 * mmsg vector matched to an skb vector which we prepared earlier.
955 */
956
vector_mmsg_rx(struct vector_private * vp,int budget)957 static int vector_mmsg_rx(struct vector_private *vp, int budget)
958 {
959 int packet_count, i;
960 struct vector_queue *qi = vp->rx_queue;
961 struct sk_buff *skb;
962 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
963 void **skbuff_vector = qi->skbuff_vector;
964 int header_check;
965
966 /* Refresh the vector and make sure it is with new skbs and the
967 * iovs are updated to point to them.
968 */
969
970 prep_queue_for_rx(qi);
971
972 /* Fire the Lazy Gun - get as many packets as we can in one go. */
973
974 if (budget > qi->max_depth)
975 budget = qi->max_depth;
976
977 packet_count = uml_vector_recvmmsg(
978 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
979
980 if (packet_count < 0)
981 vp->in_error = true;
982
983 if (packet_count <= 0)
984 return packet_count;
985
986 /* We treat packet processing as enqueue, buffer refresh as dequeue
987 * The queue_depth tells us how many buffers have been used and how
988 * many do we need to prep the next time prep_queue_for_rx() is called.
989 */
990
991 qi->queue_depth = packet_count;
992
993 for (i = 0; i < packet_count; i++) {
994 skb = (*skbuff_vector);
995 if (mmsg_vector->msg_len > vp->header_size) {
996 if (vp->header_size > 0) {
997 header_check = vp->verify_header(
998 mmsg_vector->msg_hdr.msg_iov->iov_base,
999 skb,
1000 vp
1001 );
1002 if (header_check < 0) {
1003 /* Overlay header failed to verify - discard.
1004 * We can actually keep this skb and reuse it,
1005 * but that will make the prep logic too
1006 * complex.
1007 */
1008 dev_kfree_skb_irq(skb);
1009 vp->estats.rx_encaps_errors++;
1010 continue;
1011 }
1012 if (header_check > 0) {
1013 vp->estats.rx_csum_offload_good++;
1014 skb->ip_summed = CHECKSUM_UNNECESSARY;
1015 }
1016 }
1017 pskb_trim(skb,
1018 mmsg_vector->msg_len - vp->rx_header_size);
1019 skb->protocol = eth_type_trans(skb, skb->dev);
1020 /*
1021 * We do not need to lock on updating stats here
1022 * The interrupt loop is non-reentrant.
1023 */
1024 vp->dev->stats.rx_bytes += skb->len;
1025 vp->dev->stats.rx_packets++;
1026 napi_gro_receive(&vp->napi, skb);
1027 } else {
1028 /* Overlay header too short to do anything - discard.
1029 * We can actually keep this skb and reuse it,
1030 * but that will make the prep logic too complex.
1031 */
1032 if (skb != NULL)
1033 dev_kfree_skb_irq(skb);
1034 }
1035 (*skbuff_vector) = NULL;
1036 /* Move to the next buffer element */
1037 mmsg_vector++;
1038 skbuff_vector++;
1039 }
1040 if (packet_count > 0) {
1041 if (vp->estats.rx_queue_max < packet_count)
1042 vp->estats.rx_queue_max = packet_count;
1043 vp->estats.rx_queue_running_average =
1044 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1045 }
1046 return packet_count;
1047 }
1048
vector_net_start_xmit(struct sk_buff * skb,struct net_device * dev)1049 static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1050 {
1051 struct vector_private *vp = netdev_priv(dev);
1052 int queue_depth = 0;
1053
1054 if (vp->in_error) {
1055 deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1056 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1057 deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1058 return NETDEV_TX_BUSY;
1059 }
1060
1061 if ((vp->options & VECTOR_TX) == 0) {
1062 writev_tx(vp, skb);
1063 return NETDEV_TX_OK;
1064 }
1065
1066 /* We do BQL only in the vector path, no point doing it in
1067 * packet at a time mode as there is no device queue
1068 */
1069
1070 netdev_sent_queue(vp->dev, skb->len);
1071 queue_depth = vector_enqueue(vp->tx_queue, skb);
1072
1073 if (queue_depth < vp->tx_queue->max_depth && netdev_xmit_more()) {
1074 mod_timer(&vp->tl, vp->coalesce);
1075 return NETDEV_TX_OK;
1076 } else {
1077 queue_depth = vector_send(vp->tx_queue);
1078 if (queue_depth > 0)
1079 napi_schedule(&vp->napi);
1080 }
1081
1082 return NETDEV_TX_OK;
1083 }
1084
vector_rx_interrupt(int irq,void * dev_id)1085 static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1086 {
1087 struct net_device *dev = dev_id;
1088 struct vector_private *vp = netdev_priv(dev);
1089
1090 if (!netif_running(dev))
1091 return IRQ_NONE;
1092 napi_schedule(&vp->napi);
1093 return IRQ_HANDLED;
1094
1095 }
1096
vector_tx_interrupt(int irq,void * dev_id)1097 static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1098 {
1099 struct net_device *dev = dev_id;
1100 struct vector_private *vp = netdev_priv(dev);
1101
1102 if (!netif_running(dev))
1103 return IRQ_NONE;
1104 /* We need to pay attention to it only if we got
1105 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1106 * we ignore it. In the future, it may be worth
1107 * it to improve the IRQ controller a bit to make
1108 * tweaking the IRQ mask less costly
1109 */
1110
1111 napi_schedule(&vp->napi);
1112 return IRQ_HANDLED;
1113
1114 }
1115
1116 static int irq_rr;
1117
vector_net_close(struct net_device * dev)1118 static int vector_net_close(struct net_device *dev)
1119 {
1120 struct vector_private *vp = netdev_priv(dev);
1121 unsigned long flags;
1122
1123 netif_stop_queue(dev);
1124 del_timer(&vp->tl);
1125
1126 if (vp->fds == NULL)
1127 return 0;
1128
1129 /* Disable and free all IRQS */
1130 if (vp->rx_irq > 0) {
1131 um_free_irq(vp->rx_irq, dev);
1132 vp->rx_irq = 0;
1133 }
1134 if (vp->tx_irq > 0) {
1135 um_free_irq(vp->tx_irq, dev);
1136 vp->tx_irq = 0;
1137 }
1138 napi_disable(&vp->napi);
1139 netif_napi_del(&vp->napi);
1140 if (vp->fds->rx_fd > 0) {
1141 if (vp->bpf)
1142 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1143 os_close_file(vp->fds->rx_fd);
1144 vp->fds->rx_fd = -1;
1145 }
1146 if (vp->fds->tx_fd > 0) {
1147 os_close_file(vp->fds->tx_fd);
1148 vp->fds->tx_fd = -1;
1149 }
1150 if (vp->bpf != NULL)
1151 kfree(vp->bpf->filter);
1152 kfree(vp->bpf);
1153 vp->bpf = NULL;
1154 kfree(vp->fds->remote_addr);
1155 kfree(vp->transport_data);
1156 kfree(vp->header_rxbuffer);
1157 kfree(vp->header_txbuffer);
1158 if (vp->rx_queue != NULL)
1159 destroy_queue(vp->rx_queue);
1160 if (vp->tx_queue != NULL)
1161 destroy_queue(vp->tx_queue);
1162 kfree(vp->fds);
1163 vp->fds = NULL;
1164 spin_lock_irqsave(&vp->lock, flags);
1165 vp->opened = false;
1166 vp->in_error = false;
1167 spin_unlock_irqrestore(&vp->lock, flags);
1168 return 0;
1169 }
1170
vector_poll(struct napi_struct * napi,int budget)1171 static int vector_poll(struct napi_struct *napi, int budget)
1172 {
1173 struct vector_private *vp = container_of(napi, struct vector_private, napi);
1174 int work_done = 0;
1175 int err;
1176 bool tx_enqueued = false;
1177
1178 if ((vp->options & VECTOR_TX) != 0)
1179 tx_enqueued = (vector_send(vp->tx_queue) > 0);
1180 if ((vp->options & VECTOR_RX) > 0)
1181 err = vector_mmsg_rx(vp, budget);
1182 else {
1183 err = vector_legacy_rx(vp);
1184 if (err > 0)
1185 err = 1;
1186 }
1187 if (err > 0)
1188 work_done += err;
1189
1190 if (tx_enqueued || err > 0)
1191 napi_schedule(napi);
1192 if (work_done < budget)
1193 napi_complete_done(napi, work_done);
1194 return work_done;
1195 }
1196
vector_reset_tx(struct work_struct * work)1197 static void vector_reset_tx(struct work_struct *work)
1198 {
1199 struct vector_private *vp =
1200 container_of(work, struct vector_private, reset_tx);
1201 netdev_reset_queue(vp->dev);
1202 netif_start_queue(vp->dev);
1203 netif_wake_queue(vp->dev);
1204 }
1205
vector_net_open(struct net_device * dev)1206 static int vector_net_open(struct net_device *dev)
1207 {
1208 struct vector_private *vp = netdev_priv(dev);
1209 unsigned long flags;
1210 int err = -EINVAL;
1211 struct vector_device *vdevice;
1212
1213 spin_lock_irqsave(&vp->lock, flags);
1214 if (vp->opened) {
1215 spin_unlock_irqrestore(&vp->lock, flags);
1216 return -ENXIO;
1217 }
1218 vp->opened = true;
1219 spin_unlock_irqrestore(&vp->lock, flags);
1220
1221 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1222
1223 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1224
1225 if (vp->fds == NULL)
1226 goto out_close;
1227
1228 if (build_transport_data(vp) < 0)
1229 goto out_close;
1230
1231 if ((vp->options & VECTOR_RX) > 0) {
1232 vp->rx_queue = create_queue(
1233 vp,
1234 get_depth(vp->parsed),
1235 vp->rx_header_size,
1236 MAX_IOV_SIZE
1237 );
1238 vp->rx_queue->queue_depth = get_depth(vp->parsed);
1239 } else {
1240 vp->header_rxbuffer = kmalloc(
1241 vp->rx_header_size,
1242 GFP_KERNEL
1243 );
1244 if (vp->header_rxbuffer == NULL)
1245 goto out_close;
1246 }
1247 if ((vp->options & VECTOR_TX) > 0) {
1248 vp->tx_queue = create_queue(
1249 vp,
1250 get_depth(vp->parsed),
1251 vp->header_size,
1252 MAX_IOV_SIZE
1253 );
1254 } else {
1255 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1256 if (vp->header_txbuffer == NULL)
1257 goto out_close;
1258 }
1259
1260 netif_napi_add_weight(vp->dev, &vp->napi, vector_poll,
1261 get_depth(vp->parsed));
1262 napi_enable(&vp->napi);
1263
1264 /* READ IRQ */
1265 err = um_request_irq(
1266 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1267 IRQ_READ, vector_rx_interrupt,
1268 IRQF_SHARED, dev->name, dev);
1269 if (err < 0) {
1270 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1271 err = -ENETUNREACH;
1272 goto out_close;
1273 }
1274 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1275 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1276 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1277
1278 /* WRITE IRQ - we need it only if we have vector TX */
1279 if ((vp->options & VECTOR_TX) > 0) {
1280 err = um_request_irq(
1281 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1282 IRQ_WRITE, vector_tx_interrupt,
1283 IRQF_SHARED, dev->name, dev);
1284 if (err < 0) {
1285 netdev_err(dev,
1286 "vector_open: failed to get tx irq(%d)\n", err);
1287 err = -ENETUNREACH;
1288 goto out_close;
1289 }
1290 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1291 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1292 }
1293
1294 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1295 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1296 vp->options |= VECTOR_BPF;
1297 }
1298 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1299 vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1300
1301 if (vp->bpf != NULL)
1302 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1303
1304 netif_start_queue(dev);
1305 vector_reset_stats(vp);
1306
1307 /* clear buffer - it can happen that the host side of the interface
1308 * is full when we get here. In this case, new data is never queued,
1309 * SIGIOs never arrive, and the net never works.
1310 */
1311
1312 napi_schedule(&vp->napi);
1313
1314 vdevice = find_device(vp->unit);
1315 vdevice->opened = 1;
1316
1317 if ((vp->options & VECTOR_TX) != 0)
1318 add_timer(&vp->tl);
1319 return 0;
1320 out_close:
1321 vector_net_close(dev);
1322 return err;
1323 }
1324
1325
vector_net_set_multicast_list(struct net_device * dev)1326 static void vector_net_set_multicast_list(struct net_device *dev)
1327 {
1328 /* TODO: - we can do some BPF games here */
1329 return;
1330 }
1331
vector_net_tx_timeout(struct net_device * dev,unsigned int txqueue)1332 static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1333 {
1334 struct vector_private *vp = netdev_priv(dev);
1335
1336 vp->estats.tx_timeout_count++;
1337 netif_trans_update(dev);
1338 schedule_work(&vp->reset_tx);
1339 }
1340
vector_fix_features(struct net_device * dev,netdev_features_t features)1341 static netdev_features_t vector_fix_features(struct net_device *dev,
1342 netdev_features_t features)
1343 {
1344 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1345 return features;
1346 }
1347
vector_set_features(struct net_device * dev,netdev_features_t features)1348 static int vector_set_features(struct net_device *dev,
1349 netdev_features_t features)
1350 {
1351 struct vector_private *vp = netdev_priv(dev);
1352 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1353 * no way to negotiate it on raw sockets, so we can change
1354 * only our side.
1355 */
1356 if (features & NETIF_F_GRO)
1357 /* All new frame buffers will be GRO-sized */
1358 vp->req_size = 65536;
1359 else
1360 /* All new frame buffers will be normal sized */
1361 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1362 return 0;
1363 }
1364
1365 #ifdef CONFIG_NET_POLL_CONTROLLER
vector_net_poll_controller(struct net_device * dev)1366 static void vector_net_poll_controller(struct net_device *dev)
1367 {
1368 disable_irq(dev->irq);
1369 vector_rx_interrupt(dev->irq, dev);
1370 enable_irq(dev->irq);
1371 }
1372 #endif
1373
vector_net_get_drvinfo(struct net_device * dev,struct ethtool_drvinfo * info)1374 static void vector_net_get_drvinfo(struct net_device *dev,
1375 struct ethtool_drvinfo *info)
1376 {
1377 strscpy(info->driver, DRIVER_NAME, sizeof(info->driver));
1378 }
1379
vector_net_load_bpf_flash(struct net_device * dev,struct ethtool_flash * efl)1380 static int vector_net_load_bpf_flash(struct net_device *dev,
1381 struct ethtool_flash *efl)
1382 {
1383 struct vector_private *vp = netdev_priv(dev);
1384 struct vector_device *vdevice;
1385 const struct firmware *fw;
1386 int result = 0;
1387
1388 if (!(vp->options & VECTOR_BPF_FLASH)) {
1389 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1390 return -1;
1391 }
1392
1393 spin_lock(&vp->lock);
1394
1395 if (vp->bpf != NULL) {
1396 if (vp->opened)
1397 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1398 kfree(vp->bpf->filter);
1399 vp->bpf->filter = NULL;
1400 } else {
1401 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC);
1402 if (vp->bpf == NULL) {
1403 netdev_err(dev, "failed to allocate memory for firmware\n");
1404 goto flash_fail;
1405 }
1406 }
1407
1408 vdevice = find_device(vp->unit);
1409
1410 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1411 goto flash_fail;
1412
1413 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC);
1414 if (!vp->bpf->filter)
1415 goto free_buffer;
1416
1417 vp->bpf->len = fw->size / sizeof(struct sock_filter);
1418 release_firmware(fw);
1419
1420 if (vp->opened)
1421 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1422
1423 spin_unlock(&vp->lock);
1424
1425 return result;
1426
1427 free_buffer:
1428 release_firmware(fw);
1429
1430 flash_fail:
1431 spin_unlock(&vp->lock);
1432 if (vp->bpf != NULL)
1433 kfree(vp->bpf->filter);
1434 kfree(vp->bpf);
1435 vp->bpf = NULL;
1436 return -1;
1437 }
1438
vector_get_ringparam(struct net_device * netdev,struct ethtool_ringparam * ring,struct kernel_ethtool_ringparam * kernel_ring,struct netlink_ext_ack * extack)1439 static void vector_get_ringparam(struct net_device *netdev,
1440 struct ethtool_ringparam *ring,
1441 struct kernel_ethtool_ringparam *kernel_ring,
1442 struct netlink_ext_ack *extack)
1443 {
1444 struct vector_private *vp = netdev_priv(netdev);
1445
1446 ring->rx_max_pending = vp->rx_queue->max_depth;
1447 ring->tx_max_pending = vp->tx_queue->max_depth;
1448 ring->rx_pending = vp->rx_queue->max_depth;
1449 ring->tx_pending = vp->tx_queue->max_depth;
1450 }
1451
vector_get_strings(struct net_device * dev,u32 stringset,u8 * buf)1452 static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1453 {
1454 switch (stringset) {
1455 case ETH_SS_TEST:
1456 *buf = '\0';
1457 break;
1458 case ETH_SS_STATS:
1459 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1460 break;
1461 default:
1462 WARN_ON(1);
1463 break;
1464 }
1465 }
1466
vector_get_sset_count(struct net_device * dev,int sset)1467 static int vector_get_sset_count(struct net_device *dev, int sset)
1468 {
1469 switch (sset) {
1470 case ETH_SS_TEST:
1471 return 0;
1472 case ETH_SS_STATS:
1473 return VECTOR_NUM_STATS;
1474 default:
1475 return -EOPNOTSUPP;
1476 }
1477 }
1478
vector_get_ethtool_stats(struct net_device * dev,struct ethtool_stats * estats,u64 * tmp_stats)1479 static void vector_get_ethtool_stats(struct net_device *dev,
1480 struct ethtool_stats *estats,
1481 u64 *tmp_stats)
1482 {
1483 struct vector_private *vp = netdev_priv(dev);
1484
1485 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1486 }
1487
vector_get_coalesce(struct net_device * netdev,struct ethtool_coalesce * ec,struct kernel_ethtool_coalesce * kernel_coal,struct netlink_ext_ack * extack)1488 static int vector_get_coalesce(struct net_device *netdev,
1489 struct ethtool_coalesce *ec,
1490 struct kernel_ethtool_coalesce *kernel_coal,
1491 struct netlink_ext_ack *extack)
1492 {
1493 struct vector_private *vp = netdev_priv(netdev);
1494
1495 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1496 return 0;
1497 }
1498
vector_set_coalesce(struct net_device * netdev,struct ethtool_coalesce * ec,struct kernel_ethtool_coalesce * kernel_coal,struct netlink_ext_ack * extack)1499 static int vector_set_coalesce(struct net_device *netdev,
1500 struct ethtool_coalesce *ec,
1501 struct kernel_ethtool_coalesce *kernel_coal,
1502 struct netlink_ext_ack *extack)
1503 {
1504 struct vector_private *vp = netdev_priv(netdev);
1505
1506 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1507 if (vp->coalesce == 0)
1508 vp->coalesce = 1;
1509 return 0;
1510 }
1511
1512 static const struct ethtool_ops vector_net_ethtool_ops = {
1513 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1514 .get_drvinfo = vector_net_get_drvinfo,
1515 .get_link = ethtool_op_get_link,
1516 .get_ts_info = ethtool_op_get_ts_info,
1517 .get_ringparam = vector_get_ringparam,
1518 .get_strings = vector_get_strings,
1519 .get_sset_count = vector_get_sset_count,
1520 .get_ethtool_stats = vector_get_ethtool_stats,
1521 .get_coalesce = vector_get_coalesce,
1522 .set_coalesce = vector_set_coalesce,
1523 .flash_device = vector_net_load_bpf_flash,
1524 };
1525
1526
1527 static const struct net_device_ops vector_netdev_ops = {
1528 .ndo_open = vector_net_open,
1529 .ndo_stop = vector_net_close,
1530 .ndo_start_xmit = vector_net_start_xmit,
1531 .ndo_set_rx_mode = vector_net_set_multicast_list,
1532 .ndo_tx_timeout = vector_net_tx_timeout,
1533 .ndo_set_mac_address = eth_mac_addr,
1534 .ndo_validate_addr = eth_validate_addr,
1535 .ndo_fix_features = vector_fix_features,
1536 .ndo_set_features = vector_set_features,
1537 #ifdef CONFIG_NET_POLL_CONTROLLER
1538 .ndo_poll_controller = vector_net_poll_controller,
1539 #endif
1540 };
1541
vector_timer_expire(struct timer_list * t)1542 static void vector_timer_expire(struct timer_list *t)
1543 {
1544 struct vector_private *vp = from_timer(vp, t, tl);
1545
1546 vp->estats.tx_kicks++;
1547 napi_schedule(&vp->napi);
1548 }
1549
1550
1551
vector_eth_configure(int n,struct arglist * def)1552 static void vector_eth_configure(
1553 int n,
1554 struct arglist *def
1555 )
1556 {
1557 struct vector_device *device;
1558 struct net_device *dev;
1559 struct vector_private *vp;
1560 int err;
1561
1562 device = kzalloc(sizeof(*device), GFP_KERNEL);
1563 if (device == NULL) {
1564 printk(KERN_ERR "eth_configure failed to allocate struct "
1565 "vector_device\n");
1566 return;
1567 }
1568 dev = alloc_etherdev(sizeof(struct vector_private));
1569 if (dev == NULL) {
1570 printk(KERN_ERR "eth_configure: failed to allocate struct "
1571 "net_device for vec%d\n", n);
1572 goto out_free_device;
1573 }
1574
1575 dev->mtu = get_mtu(def);
1576
1577 INIT_LIST_HEAD(&device->list);
1578 device->unit = n;
1579
1580 /* If this name ends up conflicting with an existing registered
1581 * netdevice, that is OK, register_netdev{,ice}() will notice this
1582 * and fail.
1583 */
1584 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1585 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1586 vp = netdev_priv(dev);
1587
1588 /* sysfs register */
1589 if (!driver_registered) {
1590 platform_driver_register(¨_net_driver);
1591 driver_registered = 1;
1592 }
1593 device->pdev.id = n;
1594 device->pdev.name = DRIVER_NAME;
1595 device->pdev.dev.release = vector_device_release;
1596 dev_set_drvdata(&device->pdev.dev, device);
1597 if (platform_device_register(&device->pdev))
1598 goto out_free_netdev;
1599 SET_NETDEV_DEV(dev, &device->pdev.dev);
1600
1601 device->dev = dev;
1602
1603 *vp = ((struct vector_private)
1604 {
1605 .list = LIST_HEAD_INIT(vp->list),
1606 .dev = dev,
1607 .unit = n,
1608 .options = get_transport_options(def),
1609 .rx_irq = 0,
1610 .tx_irq = 0,
1611 .parsed = def,
1612 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1613 /* TODO - we need to calculate headroom so that ip header
1614 * is 16 byte aligned all the time
1615 */
1616 .headroom = get_headroom(def),
1617 .form_header = NULL,
1618 .verify_header = NULL,
1619 .header_rxbuffer = NULL,
1620 .header_txbuffer = NULL,
1621 .header_size = 0,
1622 .rx_header_size = 0,
1623 .rexmit_scheduled = false,
1624 .opened = false,
1625 .transport_data = NULL,
1626 .in_write_poll = false,
1627 .coalesce = 2,
1628 .req_size = get_req_size(def),
1629 .in_error = false,
1630 .bpf = NULL
1631 });
1632
1633 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1634 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1635
1636 timer_setup(&vp->tl, vector_timer_expire, 0);
1637 spin_lock_init(&vp->lock);
1638
1639 /* FIXME */
1640 dev->netdev_ops = &vector_netdev_ops;
1641 dev->ethtool_ops = &vector_net_ethtool_ops;
1642 dev->watchdog_timeo = (HZ >> 1);
1643 /* primary IRQ - fixme */
1644 dev->irq = 0; /* we will adjust this once opened */
1645
1646 rtnl_lock();
1647 err = register_netdevice(dev);
1648 rtnl_unlock();
1649 if (err)
1650 goto out_undo_user_init;
1651
1652 spin_lock(&vector_devices_lock);
1653 list_add(&device->list, &vector_devices);
1654 spin_unlock(&vector_devices_lock);
1655
1656 return;
1657
1658 out_undo_user_init:
1659 return;
1660 out_free_netdev:
1661 free_netdev(dev);
1662 out_free_device:
1663 kfree(device);
1664 }
1665
1666
1667
1668
1669 /*
1670 * Invoked late in the init
1671 */
1672
vector_init(void)1673 static int __init vector_init(void)
1674 {
1675 struct list_head *ele;
1676 struct vector_cmd_line_arg *def;
1677 struct arglist *parsed;
1678
1679 list_for_each(ele, &vec_cmd_line) {
1680 def = list_entry(ele, struct vector_cmd_line_arg, list);
1681 parsed = uml_parse_vector_ifspec(def->arguments);
1682 if (parsed != NULL)
1683 vector_eth_configure(def->unit, parsed);
1684 }
1685 return 0;
1686 }
1687
1688
1689 /* Invoked at initial argument parsing, only stores
1690 * arguments until a proper vector_init is called
1691 * later
1692 */
1693
vector_setup(char * str)1694 static int __init vector_setup(char *str)
1695 {
1696 char *error;
1697 int n, err;
1698 struct vector_cmd_line_arg *new;
1699
1700 err = vector_parse(str, &n, &str, &error);
1701 if (err) {
1702 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1703 str, error);
1704 return 1;
1705 }
1706 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1707 if (!new)
1708 panic("%s: Failed to allocate %zu bytes\n", __func__,
1709 sizeof(*new));
1710 INIT_LIST_HEAD(&new->list);
1711 new->unit = n;
1712 new->arguments = str;
1713 list_add_tail(&new->list, &vec_cmd_line);
1714 return 1;
1715 }
1716
1717 __setup("vec", vector_setup);
1718 __uml_help(vector_setup,
1719 "vec[0-9]+:<option>=<value>,<option>=<value>\n"
1720 " Configure a vector io network device.\n\n"
1721 );
1722
1723 late_initcall(vector_init);
1724
1725 static struct mc_device vector_mc = {
1726 .list = LIST_HEAD_INIT(vector_mc.list),
1727 .name = "vec",
1728 .config = vector_config,
1729 .get_config = NULL,
1730 .id = vector_id,
1731 .remove = vector_remove,
1732 };
1733
1734 #ifdef CONFIG_INET
vector_inetaddr_event(struct notifier_block * this,unsigned long event,void * ptr)1735 static int vector_inetaddr_event(
1736 struct notifier_block *this,
1737 unsigned long event,
1738 void *ptr)
1739 {
1740 return NOTIFY_DONE;
1741 }
1742
1743 static struct notifier_block vector_inetaddr_notifier = {
1744 .notifier_call = vector_inetaddr_event,
1745 };
1746
inet_register(void)1747 static void inet_register(void)
1748 {
1749 register_inetaddr_notifier(&vector_inetaddr_notifier);
1750 }
1751 #else
inet_register(void)1752 static inline void inet_register(void)
1753 {
1754 }
1755 #endif
1756
vector_net_init(void)1757 static int vector_net_init(void)
1758 {
1759 mconsole_register_dev(&vector_mc);
1760 inet_register();
1761 return 0;
1762 }
1763
1764 __initcall(vector_net_init);
1765
1766
1767
1768