Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2017 - 2019 Cambridge Greys Limited
4 * Copyright (C) 2011 - 2014 Cisco Systems Inc
5 * Copyright (C) 2001 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
6 * Copyright (C) 2001 Lennert Buytenhek (buytenh@gnu.org) and
7 * James Leu (jleu@mindspring.net).
8 * Copyright (C) 2001 by various other people who didn't put their name here.
9 */
10
11#include <linux/memblock.h>
12#include <linux/etherdevice.h>
13#include <linux/ethtool.h>
14#include <linux/inetdevice.h>
15#include <linux/init.h>
16#include <linux/list.h>
17#include <linux/netdevice.h>
18#include <linux/platform_device.h>
19#include <linux/rtnetlink.h>
20#include <linux/skbuff.h>
21#include <linux/slab.h>
22#include <linux/interrupt.h>
23#include <linux/firmware.h>
24#include <linux/fs.h>
25#include <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"
48struct vector_cmd_line_arg {
49 struct list_head list;
50 int unit;
51 char *arguments;
52};
53
54struct 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
62static LIST_HEAD(vec_cmd_line);
63
64static DEFINE_SPINLOCK(vector_devices_lock);
65static LIST_HEAD(vector_devices);
66
67static int driver_registered;
68
69static void vector_eth_configure(int n, struct arglist *def);
70static 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
82static 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
103static 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
119static 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
132static char *get_bpf_file(struct arglist *def)
133{
134 return uml_vector_fetch_arg(def, "bpffile");
135}
136
137static 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 (allow > 0);
145 }
146 return false;
147}
148
149static 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
161static 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
173static 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
189static 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
233static 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
248static 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
278static 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
292static 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;
324drop:
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
333static 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;
368drop:
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
379static 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
407static 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
480static 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 */
520static 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;
596out_skb_fail:
597 kfree(result->mmsg_vector);
598out_mmsg_fail:
599 kfree(result);
600 return NULL;
601out_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
615static 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 }
665done:
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
673static 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
695static 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
712static 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
745static 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 return -EINVAL;
771 }
772
773 vector_eth_configure(n, parsed);
774 return 0;
775}
776
777static int vector_id(char **str, int *start_out, int *end_out)
778{
779 char *end;
780 int n;
781
782 n = simple_strtoul(*str, &end, 0);
783 if ((*end != '\0') || (end == *str))
784 return -1;
785
786 *start_out = n;
787 *end_out = n;
788 *str = end;
789 return n;
790}
791
792static int vector_remove(int n, char **error_out)
793{
794 struct vector_device *vec_d;
795 struct net_device *dev;
796 struct vector_private *vp;
797
798 vec_d = find_device(n);
799 if (vec_d == NULL)
800 return -ENODEV;
801 dev = vec_d->dev;
802 vp = netdev_priv(dev);
803 if (vp->fds != NULL)
804 return -EBUSY;
805 unregister_netdev(dev);
806 platform_device_unregister(&vec_d->pdev);
807 return 0;
808}
809
810/*
811 * There is no shared per-transport initialization code, so
812 * we will just initialize each interface one by one and
813 * add them to a list
814 */
815
816static struct platform_driver uml_net_driver = {
817 .driver = {
818 .name = DRIVER_NAME,
819 },
820};
821
822
823static void vector_device_release(struct device *dev)
824{
825 struct vector_device *device = dev_get_drvdata(dev);
826 struct net_device *netdev = device->dev;
827
828 list_del(&device->list);
829 kfree(device);
830 free_netdev(netdev);
831}
832
833/* Bog standard recv using recvmsg - not used normally unless the user
834 * explicitly specifies not to use recvmmsg vector RX.
835 */
836
837static int vector_legacy_rx(struct vector_private *vp)
838{
839 int pkt_len;
840 struct user_msghdr hdr;
841 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
842 int iovpos = 0;
843 struct sk_buff *skb;
844 int header_check;
845
846 hdr.msg_name = NULL;
847 hdr.msg_namelen = 0;
848 hdr.msg_iov = (struct iovec *) &iov;
849 hdr.msg_control = NULL;
850 hdr.msg_controllen = 0;
851 hdr.msg_flags = 0;
852
853 if (vp->header_size > 0) {
854 iov[0].iov_base = vp->header_rxbuffer;
855 iov[0].iov_len = vp->header_size;
856 }
857
858 skb = prep_skb(vp, &hdr);
859
860 if (skb == NULL) {
861 /* Read a packet into drop_buffer and don't do
862 * anything with it.
863 */
864 iov[iovpos].iov_base = drop_buffer;
865 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
866 hdr.msg_iovlen = 1;
867 vp->dev->stats.rx_dropped++;
868 }
869
870 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
871 if (pkt_len < 0) {
872 vp->in_error = true;
873 return pkt_len;
874 }
875
876 if (skb != NULL) {
877 if (pkt_len > vp->header_size) {
878 if (vp->header_size > 0) {
879 header_check = vp->verify_header(
880 vp->header_rxbuffer, skb, vp);
881 if (header_check < 0) {
882 dev_kfree_skb_irq(skb);
883 vp->dev->stats.rx_dropped++;
884 vp->estats.rx_encaps_errors++;
885 return 0;
886 }
887 if (header_check > 0) {
888 vp->estats.rx_csum_offload_good++;
889 skb->ip_summed = CHECKSUM_UNNECESSARY;
890 }
891 }
892 pskb_trim(skb, pkt_len - vp->rx_header_size);
893 skb->protocol = eth_type_trans(skb, skb->dev);
894 vp->dev->stats.rx_bytes += skb->len;
895 vp->dev->stats.rx_packets++;
896 napi_gro_receive(&vp->napi, skb);
897 } else {
898 dev_kfree_skb_irq(skb);
899 }
900 }
901 return pkt_len;
902}
903
904/*
905 * Packet at a time TX which falls back to vector TX if the
906 * underlying transport is busy.
907 */
908
909
910
911static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
912{
913 struct iovec iov[3 + MAX_IOV_SIZE];
914 int iov_count, pkt_len = 0;
915
916 iov[0].iov_base = vp->header_txbuffer;
917 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
918
919 if (iov_count < 1)
920 goto drop;
921
922 pkt_len = uml_vector_writev(
923 vp->fds->tx_fd,
924 (struct iovec *) &iov,
925 iov_count
926 );
927
928 if (pkt_len < 0)
929 goto drop;
930
931 netif_trans_update(vp->dev);
932 netif_wake_queue(vp->dev);
933
934 if (pkt_len > 0) {
935 vp->dev->stats.tx_bytes += skb->len;
936 vp->dev->stats.tx_packets++;
937 } else {
938 vp->dev->stats.tx_dropped++;
939 }
940 consume_skb(skb);
941 return pkt_len;
942drop:
943 vp->dev->stats.tx_dropped++;
944 consume_skb(skb);
945 if (pkt_len < 0)
946 vp->in_error = true;
947 return pkt_len;
948}
949
950/*
951 * Receive as many messages as we can in one call using the special
952 * mmsg vector matched to an skb vector which we prepared earlier.
953 */
954
955static int vector_mmsg_rx(struct vector_private *vp, int budget)
956{
957 int packet_count, i;
958 struct vector_queue *qi = vp->rx_queue;
959 struct sk_buff *skb;
960 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
961 void **skbuff_vector = qi->skbuff_vector;
962 int header_check;
963
964 /* Refresh the vector and make sure it is with new skbs and the
965 * iovs are updated to point to them.
966 */
967
968 prep_queue_for_rx(qi);
969
970 /* Fire the Lazy Gun - get as many packets as we can in one go. */
971
972 if (budget > qi->max_depth)
973 budget = qi->max_depth;
974
975 packet_count = uml_vector_recvmmsg(
976 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
977
978 if (packet_count < 0)
979 vp->in_error = true;
980
981 if (packet_count <= 0)
982 return packet_count;
983
984 /* We treat packet processing as enqueue, buffer refresh as dequeue
985 * The queue_depth tells us how many buffers have been used and how
986 * many do we need to prep the next time prep_queue_for_rx() is called.
987 */
988
989 qi->queue_depth = packet_count;
990
991 for (i = 0; i < packet_count; i++) {
992 skb = (*skbuff_vector);
993 if (mmsg_vector->msg_len > vp->header_size) {
994 if (vp->header_size > 0) {
995 header_check = vp->verify_header(
996 mmsg_vector->msg_hdr.msg_iov->iov_base,
997 skb,
998 vp
999 );
1000 if (header_check < 0) {
1001 /* Overlay header failed to verify - discard.
1002 * We can actually keep this skb and reuse it,
1003 * but that will make the prep logic too
1004 * complex.
1005 */
1006 dev_kfree_skb_irq(skb);
1007 vp->estats.rx_encaps_errors++;
1008 continue;
1009 }
1010 if (header_check > 0) {
1011 vp->estats.rx_csum_offload_good++;
1012 skb->ip_summed = CHECKSUM_UNNECESSARY;
1013 }
1014 }
1015 pskb_trim(skb,
1016 mmsg_vector->msg_len - vp->rx_header_size);
1017 skb->protocol = eth_type_trans(skb, skb->dev);
1018 /*
1019 * We do not need to lock on updating stats here
1020 * The interrupt loop is non-reentrant.
1021 */
1022 vp->dev->stats.rx_bytes += skb->len;
1023 vp->dev->stats.rx_packets++;
1024 napi_gro_receive(&vp->napi, skb);
1025 } else {
1026 /* Overlay header too short to do anything - discard.
1027 * We can actually keep this skb and reuse it,
1028 * but that will make the prep logic too complex.
1029 */
1030 if (skb != NULL)
1031 dev_kfree_skb_irq(skb);
1032 }
1033 (*skbuff_vector) = NULL;
1034 /* Move to the next buffer element */
1035 mmsg_vector++;
1036 skbuff_vector++;
1037 }
1038 if (packet_count > 0) {
1039 if (vp->estats.rx_queue_max < packet_count)
1040 vp->estats.rx_queue_max = packet_count;
1041 vp->estats.rx_queue_running_average =
1042 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1043 }
1044 return packet_count;
1045}
1046
1047static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1048{
1049 struct vector_private *vp = netdev_priv(dev);
1050 int queue_depth = 0;
1051
1052 if (vp->in_error) {
1053 deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1054 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1055 deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1056 return NETDEV_TX_BUSY;
1057 }
1058
1059 if ((vp->options & VECTOR_TX) == 0) {
1060 writev_tx(vp, skb);
1061 return NETDEV_TX_OK;
1062 }
1063
1064 /* We do BQL only in the vector path, no point doing it in
1065 * packet at a time mode as there is no device queue
1066 */
1067
1068 netdev_sent_queue(vp->dev, skb->len);
1069 queue_depth = vector_enqueue(vp->tx_queue, skb);
1070
1071 if (queue_depth < vp->tx_queue->max_depth && netdev_xmit_more()) {
1072 mod_timer(&vp->tl, vp->coalesce);
1073 return NETDEV_TX_OK;
1074 } else {
1075 queue_depth = vector_send(vp->tx_queue);
1076 if (queue_depth > 0)
1077 napi_schedule(&vp->napi);
1078 }
1079
1080 return NETDEV_TX_OK;
1081}
1082
1083static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1084{
1085 struct net_device *dev = dev_id;
1086 struct vector_private *vp = netdev_priv(dev);
1087
1088 if (!netif_running(dev))
1089 return IRQ_NONE;
1090 napi_schedule(&vp->napi);
1091 return IRQ_HANDLED;
1092
1093}
1094
1095static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1096{
1097 struct net_device *dev = dev_id;
1098 struct vector_private *vp = netdev_priv(dev);
1099
1100 if (!netif_running(dev))
1101 return IRQ_NONE;
1102 /* We need to pay attention to it only if we got
1103 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1104 * we ignore it. In the future, it may be worth
1105 * it to improve the IRQ controller a bit to make
1106 * tweaking the IRQ mask less costly
1107 */
1108
1109 napi_schedule(&vp->napi);
1110 return IRQ_HANDLED;
1111
1112}
1113
1114static int irq_rr;
1115
1116static int vector_net_close(struct net_device *dev)
1117{
1118 struct vector_private *vp = netdev_priv(dev);
1119 unsigned long flags;
1120
1121 netif_stop_queue(dev);
1122 del_timer(&vp->tl);
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 spin_lock_irqsave(&vp->lock, flags);
1163 vp->opened = false;
1164 vp->in_error = false;
1165 spin_unlock_irqrestore(&vp->lock, flags);
1166 return 0;
1167}
1168
1169static int vector_poll(struct napi_struct *napi, int budget)
1170{
1171 struct vector_private *vp = container_of(napi, struct vector_private, napi);
1172 int work_done = 0;
1173 int err;
1174 bool tx_enqueued = false;
1175
1176 if ((vp->options & VECTOR_TX) != 0)
1177 tx_enqueued = (vector_send(vp->tx_queue) > 0);
1178 if ((vp->options & VECTOR_RX) > 0)
1179 err = vector_mmsg_rx(vp, budget);
1180 else {
1181 err = vector_legacy_rx(vp);
1182 if (err > 0)
1183 err = 1;
1184 }
1185 if (err > 0)
1186 work_done += err;
1187
1188 if (tx_enqueued || err > 0)
1189 napi_schedule(napi);
1190 if (work_done < budget)
1191 napi_complete_done(napi, work_done);
1192 return work_done;
1193}
1194
1195static void vector_reset_tx(struct work_struct *work)
1196{
1197 struct vector_private *vp =
1198 container_of(work, struct vector_private, reset_tx);
1199 netdev_reset_queue(vp->dev);
1200 netif_start_queue(vp->dev);
1201 netif_wake_queue(vp->dev);
1202}
1203
1204static int vector_net_open(struct net_device *dev)
1205{
1206 struct vector_private *vp = netdev_priv(dev);
1207 unsigned long flags;
1208 int err = -EINVAL;
1209 struct vector_device *vdevice;
1210
1211 spin_lock_irqsave(&vp->lock, flags);
1212 if (vp->opened) {
1213 spin_unlock_irqrestore(&vp->lock, flags);
1214 return -ENXIO;
1215 }
1216 vp->opened = true;
1217 spin_unlock_irqrestore(&vp->lock, flags);
1218
1219 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1220
1221 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1222
1223 if (vp->fds == NULL)
1224 goto out_close;
1225
1226 if (build_transport_data(vp) < 0)
1227 goto out_close;
1228
1229 if ((vp->options & VECTOR_RX) > 0) {
1230 vp->rx_queue = create_queue(
1231 vp,
1232 get_depth(vp->parsed),
1233 vp->rx_header_size,
1234 MAX_IOV_SIZE
1235 );
1236 vp->rx_queue->queue_depth = get_depth(vp->parsed);
1237 } else {
1238 vp->header_rxbuffer = kmalloc(
1239 vp->rx_header_size,
1240 GFP_KERNEL
1241 );
1242 if (vp->header_rxbuffer == NULL)
1243 goto out_close;
1244 }
1245 if ((vp->options & VECTOR_TX) > 0) {
1246 vp->tx_queue = create_queue(
1247 vp,
1248 get_depth(vp->parsed),
1249 vp->header_size,
1250 MAX_IOV_SIZE
1251 );
1252 } else {
1253 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1254 if (vp->header_txbuffer == NULL)
1255 goto out_close;
1256 }
1257
1258 netif_napi_add_weight(vp->dev, &vp->napi, vector_poll,
1259 get_depth(vp->parsed));
1260 napi_enable(&vp->napi);
1261
1262 /* READ IRQ */
1263 err = um_request_irq(
1264 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1265 IRQ_READ, vector_rx_interrupt,
1266 IRQF_SHARED, dev->name, dev);
1267 if (err < 0) {
1268 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1269 err = -ENETUNREACH;
1270 goto out_close;
1271 }
1272 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1273 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1274 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1275
1276 /* WRITE IRQ - we need it only if we have vector TX */
1277 if ((vp->options & VECTOR_TX) > 0) {
1278 err = um_request_irq(
1279 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1280 IRQ_WRITE, vector_tx_interrupt,
1281 IRQF_SHARED, dev->name, dev);
1282 if (err < 0) {
1283 netdev_err(dev,
1284 "vector_open: failed to get tx irq(%d)\n", err);
1285 err = -ENETUNREACH;
1286 goto out_close;
1287 }
1288 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1289 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1290 }
1291
1292 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1293 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1294 vp->options |= VECTOR_BPF;
1295 }
1296 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1297 vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1298
1299 if (vp->bpf != NULL)
1300 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1301
1302 netif_start_queue(dev);
1303 vector_reset_stats(vp);
1304
1305 /* clear buffer - it can happen that the host side of the interface
1306 * is full when we get here. In this case, new data is never queued,
1307 * SIGIOs never arrive, and the net never works.
1308 */
1309
1310 napi_schedule(&vp->napi);
1311
1312 vdevice = find_device(vp->unit);
1313 vdevice->opened = 1;
1314
1315 if ((vp->options & VECTOR_TX) != 0)
1316 add_timer(&vp->tl);
1317 return 0;
1318out_close:
1319 vector_net_close(dev);
1320 return err;
1321}
1322
1323
1324static void vector_net_set_multicast_list(struct net_device *dev)
1325{
1326 /* TODO: - we can do some BPF games here */
1327 return;
1328}
1329
1330static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1331{
1332 struct vector_private *vp = netdev_priv(dev);
1333
1334 vp->estats.tx_timeout_count++;
1335 netif_trans_update(dev);
1336 schedule_work(&vp->reset_tx);
1337}
1338
1339static netdev_features_t vector_fix_features(struct net_device *dev,
1340 netdev_features_t features)
1341{
1342 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1343 return features;
1344}
1345
1346static int vector_set_features(struct net_device *dev,
1347 netdev_features_t features)
1348{
1349 struct vector_private *vp = netdev_priv(dev);
1350 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1351 * no way to negotiate it on raw sockets, so we can change
1352 * only our side.
1353 */
1354 if (features & NETIF_F_GRO)
1355 /* All new frame buffers will be GRO-sized */
1356 vp->req_size = 65536;
1357 else
1358 /* All new frame buffers will be normal sized */
1359 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1360 return 0;
1361}
1362
1363#ifdef CONFIG_NET_POLL_CONTROLLER
1364static void vector_net_poll_controller(struct net_device *dev)
1365{
1366 disable_irq(dev->irq);
1367 vector_rx_interrupt(dev->irq, dev);
1368 enable_irq(dev->irq);
1369}
1370#endif
1371
1372static void vector_net_get_drvinfo(struct net_device *dev,
1373 struct ethtool_drvinfo *info)
1374{
1375 strscpy(info->driver, DRIVER_NAME, sizeof(info->driver));
1376}
1377
1378static int vector_net_load_bpf_flash(struct net_device *dev,
1379 struct ethtool_flash *efl)
1380{
1381 struct vector_private *vp = netdev_priv(dev);
1382 struct vector_device *vdevice;
1383 const struct firmware *fw;
1384 int result = 0;
1385
1386 if (!(vp->options & VECTOR_BPF_FLASH)) {
1387 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1388 return -1;
1389 }
1390
1391 spin_lock(&vp->lock);
1392
1393 if (vp->bpf != NULL) {
1394 if (vp->opened)
1395 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1396 kfree(vp->bpf->filter);
1397 vp->bpf->filter = NULL;
1398 } else {
1399 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC);
1400 if (vp->bpf == NULL) {
1401 netdev_err(dev, "failed to allocate memory for firmware\n");
1402 goto flash_fail;
1403 }
1404 }
1405
1406 vdevice = find_device(vp->unit);
1407
1408 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1409 goto flash_fail;
1410
1411 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC);
1412 if (!vp->bpf->filter)
1413 goto free_buffer;
1414
1415 vp->bpf->len = fw->size / sizeof(struct sock_filter);
1416 release_firmware(fw);
1417
1418 if (vp->opened)
1419 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1420
1421 spin_unlock(&vp->lock);
1422
1423 return result;
1424
1425free_buffer:
1426 release_firmware(fw);
1427
1428flash_fail:
1429 spin_unlock(&vp->lock);
1430 if (vp->bpf != NULL)
1431 kfree(vp->bpf->filter);
1432 kfree(vp->bpf);
1433 vp->bpf = NULL;
1434 return -1;
1435}
1436
1437static void vector_get_ringparam(struct net_device *netdev,
1438 struct ethtool_ringparam *ring,
1439 struct kernel_ethtool_ringparam *kernel_ring,
1440 struct netlink_ext_ack *extack)
1441{
1442 struct vector_private *vp = netdev_priv(netdev);
1443
1444 ring->rx_max_pending = vp->rx_queue->max_depth;
1445 ring->tx_max_pending = vp->tx_queue->max_depth;
1446 ring->rx_pending = vp->rx_queue->max_depth;
1447 ring->tx_pending = vp->tx_queue->max_depth;
1448}
1449
1450static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1451{
1452 switch (stringset) {
1453 case ETH_SS_TEST:
1454 *buf = '\0';
1455 break;
1456 case ETH_SS_STATS:
1457 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1458 break;
1459 default:
1460 WARN_ON(1);
1461 break;
1462 }
1463}
1464
1465static int vector_get_sset_count(struct net_device *dev, int sset)
1466{
1467 switch (sset) {
1468 case ETH_SS_TEST:
1469 return 0;
1470 case ETH_SS_STATS:
1471 return VECTOR_NUM_STATS;
1472 default:
1473 return -EOPNOTSUPP;
1474 }
1475}
1476
1477static void vector_get_ethtool_stats(struct net_device *dev,
1478 struct ethtool_stats *estats,
1479 u64 *tmp_stats)
1480{
1481 struct vector_private *vp = netdev_priv(dev);
1482
1483 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1484}
1485
1486static int vector_get_coalesce(struct net_device *netdev,
1487 struct ethtool_coalesce *ec,
1488 struct kernel_ethtool_coalesce *kernel_coal,
1489 struct netlink_ext_ack *extack)
1490{
1491 struct vector_private *vp = netdev_priv(netdev);
1492
1493 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1494 return 0;
1495}
1496
1497static int vector_set_coalesce(struct net_device *netdev,
1498 struct ethtool_coalesce *ec,
1499 struct kernel_ethtool_coalesce *kernel_coal,
1500 struct netlink_ext_ack *extack)
1501{
1502 struct vector_private *vp = netdev_priv(netdev);
1503
1504 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1505 if (vp->coalesce == 0)
1506 vp->coalesce = 1;
1507 return 0;
1508}
1509
1510static const struct ethtool_ops vector_net_ethtool_ops = {
1511 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1512 .get_drvinfo = vector_net_get_drvinfo,
1513 .get_link = ethtool_op_get_link,
1514 .get_ts_info = ethtool_op_get_ts_info,
1515 .get_ringparam = vector_get_ringparam,
1516 .get_strings = vector_get_strings,
1517 .get_sset_count = vector_get_sset_count,
1518 .get_ethtool_stats = vector_get_ethtool_stats,
1519 .get_coalesce = vector_get_coalesce,
1520 .set_coalesce = vector_set_coalesce,
1521 .flash_device = vector_net_load_bpf_flash,
1522};
1523
1524
1525static const struct net_device_ops vector_netdev_ops = {
1526 .ndo_open = vector_net_open,
1527 .ndo_stop = vector_net_close,
1528 .ndo_start_xmit = vector_net_start_xmit,
1529 .ndo_set_rx_mode = vector_net_set_multicast_list,
1530 .ndo_tx_timeout = vector_net_tx_timeout,
1531 .ndo_set_mac_address = eth_mac_addr,
1532 .ndo_validate_addr = eth_validate_addr,
1533 .ndo_fix_features = vector_fix_features,
1534 .ndo_set_features = vector_set_features,
1535#ifdef CONFIG_NET_POLL_CONTROLLER
1536 .ndo_poll_controller = vector_net_poll_controller,
1537#endif
1538};
1539
1540static void vector_timer_expire(struct timer_list *t)
1541{
1542 struct vector_private *vp = from_timer(vp, t, tl);
1543
1544 vp->estats.tx_kicks++;
1545 napi_schedule(&vp->napi);
1546}
1547
1548
1549
1550static void vector_eth_configure(
1551 int n,
1552 struct arglist *def
1553 )
1554{
1555 struct vector_device *device;
1556 struct net_device *dev;
1557 struct vector_private *vp;
1558 int err;
1559
1560 device = kzalloc(sizeof(*device), GFP_KERNEL);
1561 if (device == NULL) {
1562 printk(KERN_ERR "eth_configure failed to allocate struct "
1563 "vector_device\n");
1564 return;
1565 }
1566 dev = alloc_etherdev(sizeof(struct vector_private));
1567 if (dev == NULL) {
1568 printk(KERN_ERR "eth_configure: failed to allocate struct "
1569 "net_device for vec%d\n", n);
1570 goto out_free_device;
1571 }
1572
1573 dev->mtu = get_mtu(def);
1574
1575 INIT_LIST_HEAD(&device->list);
1576 device->unit = n;
1577
1578 /* If this name ends up conflicting with an existing registered
1579 * netdevice, that is OK, register_netdev{,ice}() will notice this
1580 * and fail.
1581 */
1582 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1583 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1584 vp = netdev_priv(dev);
1585
1586 /* sysfs register */
1587 if (!driver_registered) {
1588 platform_driver_register(¨_net_driver);
1589 driver_registered = 1;
1590 }
1591 device->pdev.id = n;
1592 device->pdev.name = DRIVER_NAME;
1593 device->pdev.dev.release = vector_device_release;
1594 dev_set_drvdata(&device->pdev.dev, device);
1595 if (platform_device_register(&device->pdev))
1596 goto out_free_netdev;
1597 SET_NETDEV_DEV(dev, &device->pdev.dev);
1598
1599 device->dev = dev;
1600
1601 *vp = ((struct vector_private)
1602 {
1603 .list = LIST_HEAD_INIT(vp->list),
1604 .dev = dev,
1605 .unit = n,
1606 .options = get_transport_options(def),
1607 .rx_irq = 0,
1608 .tx_irq = 0,
1609 .parsed = def,
1610 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1611 /* TODO - we need to calculate headroom so that ip header
1612 * is 16 byte aligned all the time
1613 */
1614 .headroom = get_headroom(def),
1615 .form_header = NULL,
1616 .verify_header = NULL,
1617 .header_rxbuffer = NULL,
1618 .header_txbuffer = NULL,
1619 .header_size = 0,
1620 .rx_header_size = 0,
1621 .rexmit_scheduled = false,
1622 .opened = false,
1623 .transport_data = NULL,
1624 .in_write_poll = false,
1625 .coalesce = 2,
1626 .req_size = get_req_size(def),
1627 .in_error = false,
1628 .bpf = NULL
1629 });
1630
1631 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1632 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1633
1634 timer_setup(&vp->tl, vector_timer_expire, 0);
1635 spin_lock_init(&vp->lock);
1636
1637 /* FIXME */
1638 dev->netdev_ops = &vector_netdev_ops;
1639 dev->ethtool_ops = &vector_net_ethtool_ops;
1640 dev->watchdog_timeo = (HZ >> 1);
1641 /* primary IRQ - fixme */
1642 dev->irq = 0; /* we will adjust this once opened */
1643
1644 rtnl_lock();
1645 err = register_netdevice(dev);
1646 rtnl_unlock();
1647 if (err)
1648 goto out_undo_user_init;
1649
1650 spin_lock(&vector_devices_lock);
1651 list_add(&device->list, &vector_devices);
1652 spin_unlock(&vector_devices_lock);
1653
1654 return;
1655
1656out_undo_user_init:
1657 return;
1658out_free_netdev:
1659 free_netdev(dev);
1660out_free_device:
1661 kfree(device);
1662}
1663
1664
1665
1666
1667/*
1668 * Invoked late in the init
1669 */
1670
1671static int __init vector_init(void)
1672{
1673 struct list_head *ele;
1674 struct vector_cmd_line_arg *def;
1675 struct arglist *parsed;
1676
1677 list_for_each(ele, &vec_cmd_line) {
1678 def = list_entry(ele, struct vector_cmd_line_arg, list);
1679 parsed = uml_parse_vector_ifspec(def->arguments);
1680 if (parsed != NULL)
1681 vector_eth_configure(def->unit, parsed);
1682 }
1683 return 0;
1684}
1685
1686
1687/* Invoked at initial argument parsing, only stores
1688 * arguments until a proper vector_init is called
1689 * later
1690 */
1691
1692static int __init vector_setup(char *str)
1693{
1694 char *error;
1695 int n, err;
1696 struct vector_cmd_line_arg *new;
1697
1698 err = vector_parse(str, &n, &str, &error);
1699 if (err) {
1700 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1701 str, error);
1702 return 1;
1703 }
1704 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1705 if (!new)
1706 panic("%s: Failed to allocate %zu bytes\n", __func__,
1707 sizeof(*new));
1708 INIT_LIST_HEAD(&new->list);
1709 new->unit = n;
1710 new->arguments = str;
1711 list_add_tail(&new->list, &vec_cmd_line);
1712 return 1;
1713}
1714
1715__setup("vec", vector_setup);
1716__uml_help(vector_setup,
1717"vec[0-9]+:<option>=<value>,<option>=<value>\n"
1718" Configure a vector io network device.\n\n"
1719);
1720
1721late_initcall(vector_init);
1722
1723static struct mc_device vector_mc = {
1724 .list = LIST_HEAD_INIT(vector_mc.list),
1725 .name = "vec",
1726 .config = vector_config,
1727 .get_config = NULL,
1728 .id = vector_id,
1729 .remove = vector_remove,
1730};
1731
1732#ifdef CONFIG_INET
1733static int vector_inetaddr_event(
1734 struct notifier_block *this,
1735 unsigned long event,
1736 void *ptr)
1737{
1738 return NOTIFY_DONE;
1739}
1740
1741static struct notifier_block vector_inetaddr_notifier = {
1742 .notifier_call = vector_inetaddr_event,
1743};
1744
1745static void inet_register(void)
1746{
1747 register_inetaddr_notifier(&vector_inetaddr_notifier);
1748}
1749#else
1750static inline void inet_register(void)
1751{
1752}
1753#endif
1754
1755static int vector_net_init(void)
1756{
1757 mconsole_register_dev(&vector_mc);
1758 inet_register();
1759 return 0;
1760}
1761
1762__initcall(vector_net_init);
1763
1764
1765
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"
48struct vector_cmd_line_arg {
49 struct list_head list;
50 int unit;
51 char *arguments;
52};
53
54struct 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
62static LIST_HEAD(vec_cmd_line);
63
64static DEFINE_SPINLOCK(vector_devices_lock);
65static LIST_HEAD(vector_devices);
66
67static int driver_registered;
68
69static void vector_eth_configure(int n, struct arglist *def);
70
71/* Argument accessors to set variables (and/or set default values)
72 * mtu, buffer sizing, default headroom, etc
73 */
74
75#define DEFAULT_HEADROOM 2
76#define SAFETY_MARGIN 32
77#define DEFAULT_VECTOR_SIZE 64
78#define TX_SMALL_PACKET 128
79#define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
80#define MAX_ITERATIONS 64
81
82static 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
103static 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
119static 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
132static char *get_bpf_file(struct arglist *def)
133{
134 return uml_vector_fetch_arg(def, "bpffile");
135}
136
137static 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 (allow > 0);
145 }
146 return false;
147}
148
149static 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
161static 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
173static 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
189static 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
233static 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
248static 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
278static 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
292static 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;
324drop:
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
333static 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;
368drop:
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
379static 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
407static 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_trans_update(qi->dev);
462 netif_wake_queue(qi->dev);
463 /* if TX is busy, break out of the send loop,
464 * poll write IRQ will reschedule xmit for us
465 */
466 if (result != send_len) {
467 vp->estats.tx_restart_queue++;
468 break;
469 }
470 }
471 }
472 spin_unlock(&qi->head_lock);
473 } else {
474 tasklet_schedule(&vp->tx_poll);
475 }
476 return queue_depth;
477}
478
479/* Queue destructor. Deliberately stateless so we can use
480 * it in queue cleanup if initialization fails.
481 */
482
483static void destroy_queue(struct vector_queue *qi)
484{
485 int i;
486 struct iovec *iov;
487 struct vector_private *vp = netdev_priv(qi->dev);
488 struct mmsghdr *mmsg_vector;
489
490 if (qi == NULL)
491 return;
492 /* deallocate any skbuffs - we rely on any unused to be
493 * set to NULL.
494 */
495 if (qi->skbuff_vector != NULL) {
496 for (i = 0; i < qi->max_depth; i++) {
497 if (*(qi->skbuff_vector + i) != NULL)
498 dev_kfree_skb_any(*(qi->skbuff_vector + i));
499 }
500 kfree(qi->skbuff_vector);
501 }
502 /* deallocate matching IOV structures including header buffs */
503 if (qi->mmsg_vector != NULL) {
504 mmsg_vector = qi->mmsg_vector;
505 for (i = 0; i < qi->max_depth; i++) {
506 iov = mmsg_vector->msg_hdr.msg_iov;
507 if (iov != NULL) {
508 if ((vp->header_size > 0) &&
509 (iov->iov_base != NULL))
510 kfree(iov->iov_base);
511 kfree(iov);
512 }
513 mmsg_vector++;
514 }
515 kfree(qi->mmsg_vector);
516 }
517 kfree(qi);
518}
519
520/*
521 * Queue constructor. Create a queue with a given side.
522 */
523static struct vector_queue *create_queue(
524 struct vector_private *vp,
525 int max_size,
526 int header_size,
527 int num_extra_frags)
528{
529 struct vector_queue *result;
530 int i;
531 struct iovec *iov;
532 struct mmsghdr *mmsg_vector;
533
534 result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
535 if (result == NULL)
536 return NULL;
537 result->max_depth = max_size;
538 result->dev = vp->dev;
539 result->mmsg_vector = kmalloc(
540 (sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
541 if (result->mmsg_vector == NULL)
542 goto out_mmsg_fail;
543 result->skbuff_vector = kmalloc(
544 (sizeof(void *) * max_size), GFP_KERNEL);
545 if (result->skbuff_vector == NULL)
546 goto out_skb_fail;
547
548 /* further failures can be handled safely by destroy_queue*/
549
550 mmsg_vector = result->mmsg_vector;
551 for (i = 0; i < max_size; i++) {
552 /* Clear all pointers - we use non-NULL as marking on
553 * what to free on destruction
554 */
555 *(result->skbuff_vector + i) = NULL;
556 mmsg_vector->msg_hdr.msg_iov = NULL;
557 mmsg_vector++;
558 }
559 mmsg_vector = result->mmsg_vector;
560 result->max_iov_frags = num_extra_frags;
561 for (i = 0; i < max_size; i++) {
562 if (vp->header_size > 0)
563 iov = kmalloc_array(3 + num_extra_frags,
564 sizeof(struct iovec),
565 GFP_KERNEL
566 );
567 else
568 iov = kmalloc_array(2 + num_extra_frags,
569 sizeof(struct iovec),
570 GFP_KERNEL
571 );
572 if (iov == NULL)
573 goto out_fail;
574 mmsg_vector->msg_hdr.msg_iov = iov;
575 mmsg_vector->msg_hdr.msg_iovlen = 1;
576 mmsg_vector->msg_hdr.msg_control = NULL;
577 mmsg_vector->msg_hdr.msg_controllen = 0;
578 mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
579 mmsg_vector->msg_hdr.msg_name = NULL;
580 mmsg_vector->msg_hdr.msg_namelen = 0;
581 if (vp->header_size > 0) {
582 iov->iov_base = kmalloc(header_size, GFP_KERNEL);
583 if (iov->iov_base == NULL)
584 goto out_fail;
585 iov->iov_len = header_size;
586 mmsg_vector->msg_hdr.msg_iovlen = 2;
587 iov++;
588 }
589 iov->iov_base = NULL;
590 iov->iov_len = 0;
591 mmsg_vector++;
592 }
593 spin_lock_init(&result->head_lock);
594 spin_lock_init(&result->tail_lock);
595 result->queue_depth = 0;
596 result->head = 0;
597 result->tail = 0;
598 return result;
599out_skb_fail:
600 kfree(result->mmsg_vector);
601out_mmsg_fail:
602 kfree(result);
603 return NULL;
604out_fail:
605 destroy_queue(result);
606 return NULL;
607}
608
609/*
610 * We do not use the RX queue as a proper wraparound queue for now
611 * This is not necessary because the consumption via netif_rx()
612 * happens in-line. While we can try using the return code of
613 * netif_rx() for flow control there are no drivers doing this today.
614 * For this RX specific use we ignore the tail/head locks and
615 * just read into a prepared queue filled with skbuffs.
616 */
617
618static struct sk_buff *prep_skb(
619 struct vector_private *vp,
620 struct user_msghdr *msg)
621{
622 int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
623 struct sk_buff *result;
624 int iov_index = 0, len;
625 struct iovec *iov = msg->msg_iov;
626 int err, nr_frags, frag;
627 skb_frag_t *skb_frag;
628
629 if (vp->req_size <= linear)
630 len = linear;
631 else
632 len = vp->req_size;
633 result = alloc_skb_with_frags(
634 linear,
635 len - vp->max_packet,
636 3,
637 &err,
638 GFP_ATOMIC
639 );
640 if (vp->header_size > 0)
641 iov_index++;
642 if (result == NULL) {
643 iov[iov_index].iov_base = NULL;
644 iov[iov_index].iov_len = 0;
645 goto done;
646 }
647 skb_reserve(result, vp->headroom);
648 result->dev = vp->dev;
649 skb_put(result, vp->max_packet);
650 result->data_len = len - vp->max_packet;
651 result->len += len - vp->max_packet;
652 skb_reset_mac_header(result);
653 result->ip_summed = CHECKSUM_NONE;
654 iov[iov_index].iov_base = result->data;
655 iov[iov_index].iov_len = vp->max_packet;
656 iov_index++;
657
658 nr_frags = skb_shinfo(result)->nr_frags;
659 for (frag = 0; frag < nr_frags; frag++) {
660 skb_frag = &skb_shinfo(result)->frags[frag];
661 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
662 if (iov[iov_index].iov_base != NULL)
663 iov[iov_index].iov_len = skb_frag_size(skb_frag);
664 else
665 iov[iov_index].iov_len = 0;
666 iov_index++;
667 }
668done:
669 msg->msg_iovlen = iov_index;
670 return result;
671}
672
673
674/* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs*/
675
676static void prep_queue_for_rx(struct vector_queue *qi)
677{
678 struct vector_private *vp = netdev_priv(qi->dev);
679 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
680 void **skbuff_vector = qi->skbuff_vector;
681 int i;
682
683 if (qi->queue_depth == 0)
684 return;
685 for (i = 0; i < qi->queue_depth; i++) {
686 /* it is OK if allocation fails - recvmmsg with NULL data in
687 * iov argument still performs an RX, just drops the packet
688 * This allows us stop faffing around with a "drop buffer"
689 */
690
691 *skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
692 skbuff_vector++;
693 mmsg_vector++;
694 }
695 qi->queue_depth = 0;
696}
697
698static struct vector_device *find_device(int n)
699{
700 struct vector_device *device;
701 struct list_head *ele;
702
703 spin_lock(&vector_devices_lock);
704 list_for_each(ele, &vector_devices) {
705 device = list_entry(ele, struct vector_device, list);
706 if (device->unit == n)
707 goto out;
708 }
709 device = NULL;
710 out:
711 spin_unlock(&vector_devices_lock);
712 return device;
713}
714
715static int vector_parse(char *str, int *index_out, char **str_out,
716 char **error_out)
717{
718 int n, len, err;
719 char *start = str;
720
721 len = strlen(str);
722
723 while ((*str != ':') && (strlen(str) > 1))
724 str++;
725 if (*str != ':') {
726 *error_out = "Expected ':' after device number";
727 return -EINVAL;
728 }
729 *str = '\0';
730
731 err = kstrtouint(start, 0, &n);
732 if (err < 0) {
733 *error_out = "Bad device number";
734 return err;
735 }
736
737 str++;
738 if (find_device(n)) {
739 *error_out = "Device already configured";
740 return -EINVAL;
741 }
742
743 *index_out = n;
744 *str_out = str;
745 return 0;
746}
747
748static int vector_config(char *str, char **error_out)
749{
750 int err, n;
751 char *params;
752 struct arglist *parsed;
753
754 err = vector_parse(str, &n, ¶ms, error_out);
755 if (err != 0)
756 return err;
757
758 /* This string is broken up and the pieces used by the underlying
759 * driver. We should copy it to make sure things do not go wrong
760 * later.
761 */
762
763 params = kstrdup(params, GFP_KERNEL);
764 if (params == NULL) {
765 *error_out = "vector_config failed to strdup string";
766 return -ENOMEM;
767 }
768
769 parsed = uml_parse_vector_ifspec(params);
770
771 if (parsed == NULL) {
772 *error_out = "vector_config failed to parse parameters";
773 return -EINVAL;
774 }
775
776 vector_eth_configure(n, parsed);
777 return 0;
778}
779
780static int vector_id(char **str, int *start_out, int *end_out)
781{
782 char *end;
783 int n;
784
785 n = simple_strtoul(*str, &end, 0);
786 if ((*end != '\0') || (end == *str))
787 return -1;
788
789 *start_out = n;
790 *end_out = n;
791 *str = end;
792 return n;
793}
794
795static int vector_remove(int n, char **error_out)
796{
797 struct vector_device *vec_d;
798 struct net_device *dev;
799 struct vector_private *vp;
800
801 vec_d = find_device(n);
802 if (vec_d == NULL)
803 return -ENODEV;
804 dev = vec_d->dev;
805 vp = netdev_priv(dev);
806 if (vp->fds != NULL)
807 return -EBUSY;
808 unregister_netdev(dev);
809 platform_device_unregister(&vec_d->pdev);
810 return 0;
811}
812
813/*
814 * There is no shared per-transport initialization code, so
815 * we will just initialize each interface one by one and
816 * add them to a list
817 */
818
819static struct platform_driver uml_net_driver = {
820 .driver = {
821 .name = DRIVER_NAME,
822 },
823};
824
825
826static void vector_device_release(struct device *dev)
827{
828 struct vector_device *device = dev_get_drvdata(dev);
829 struct net_device *netdev = device->dev;
830
831 list_del(&device->list);
832 kfree(device);
833 free_netdev(netdev);
834}
835
836/* Bog standard recv using recvmsg - not used normally unless the user
837 * explicitly specifies not to use recvmmsg vector RX.
838 */
839
840static int vector_legacy_rx(struct vector_private *vp)
841{
842 int pkt_len;
843 struct user_msghdr hdr;
844 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
845 int iovpos = 0;
846 struct sk_buff *skb;
847 int header_check;
848
849 hdr.msg_name = NULL;
850 hdr.msg_namelen = 0;
851 hdr.msg_iov = (struct iovec *) &iov;
852 hdr.msg_control = NULL;
853 hdr.msg_controllen = 0;
854 hdr.msg_flags = 0;
855
856 if (vp->header_size > 0) {
857 iov[0].iov_base = vp->header_rxbuffer;
858 iov[0].iov_len = vp->header_size;
859 }
860
861 skb = prep_skb(vp, &hdr);
862
863 if (skb == NULL) {
864 /* Read a packet into drop_buffer and don't do
865 * anything with it.
866 */
867 iov[iovpos].iov_base = drop_buffer;
868 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
869 hdr.msg_iovlen = 1;
870 vp->dev->stats.rx_dropped++;
871 }
872
873 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
874 if (pkt_len < 0) {
875 vp->in_error = true;
876 return pkt_len;
877 }
878
879 if (skb != NULL) {
880 if (pkt_len > vp->header_size) {
881 if (vp->header_size > 0) {
882 header_check = vp->verify_header(
883 vp->header_rxbuffer, skb, vp);
884 if (header_check < 0) {
885 dev_kfree_skb_irq(skb);
886 vp->dev->stats.rx_dropped++;
887 vp->estats.rx_encaps_errors++;
888 return 0;
889 }
890 if (header_check > 0) {
891 vp->estats.rx_csum_offload_good++;
892 skb->ip_summed = CHECKSUM_UNNECESSARY;
893 }
894 }
895 pskb_trim(skb, pkt_len - vp->rx_header_size);
896 skb->protocol = eth_type_trans(skb, skb->dev);
897 vp->dev->stats.rx_bytes += skb->len;
898 vp->dev->stats.rx_packets++;
899 netif_rx(skb);
900 } else {
901 dev_kfree_skb_irq(skb);
902 }
903 }
904 return pkt_len;
905}
906
907/*
908 * Packet at a time TX which falls back to vector TX if the
909 * underlying transport is busy.
910 */
911
912
913
914static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
915{
916 struct iovec iov[3 + MAX_IOV_SIZE];
917 int iov_count, pkt_len = 0;
918
919 iov[0].iov_base = vp->header_txbuffer;
920 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
921
922 if (iov_count < 1)
923 goto drop;
924
925 pkt_len = uml_vector_writev(
926 vp->fds->tx_fd,
927 (struct iovec *) &iov,
928 iov_count
929 );
930
931 if (pkt_len < 0)
932 goto drop;
933
934 netif_trans_update(vp->dev);
935 netif_wake_queue(vp->dev);
936
937 if (pkt_len > 0) {
938 vp->dev->stats.tx_bytes += skb->len;
939 vp->dev->stats.tx_packets++;
940 } else {
941 vp->dev->stats.tx_dropped++;
942 }
943 consume_skb(skb);
944 return pkt_len;
945drop:
946 vp->dev->stats.tx_dropped++;
947 consume_skb(skb);
948 if (pkt_len < 0)
949 vp->in_error = true;
950 return pkt_len;
951}
952
953/*
954 * Receive as many messages as we can in one call using the special
955 * mmsg vector matched to an skb vector which we prepared earlier.
956 */
957
958static int vector_mmsg_rx(struct vector_private *vp)
959{
960 int packet_count, i;
961 struct vector_queue *qi = vp->rx_queue;
962 struct sk_buff *skb;
963 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
964 void **skbuff_vector = qi->skbuff_vector;
965 int header_check;
966
967 /* Refresh the vector and make sure it is with new skbs and the
968 * iovs are updated to point to them.
969 */
970
971 prep_queue_for_rx(qi);
972
973 /* Fire the Lazy Gun - get as many packets as we can in one go. */
974
975 packet_count = uml_vector_recvmmsg(
976 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
977
978 if (packet_count < 0)
979 vp->in_error = true;
980
981 if (packet_count <= 0)
982 return packet_count;
983
984 /* We treat packet processing as enqueue, buffer refresh as dequeue
985 * The queue_depth tells us how many buffers have been used and how
986 * many do we need to prep the next time prep_queue_for_rx() is called.
987 */
988
989 qi->queue_depth = packet_count;
990
991 for (i = 0; i < packet_count; i++) {
992 skb = (*skbuff_vector);
993 if (mmsg_vector->msg_len > vp->header_size) {
994 if (vp->header_size > 0) {
995 header_check = vp->verify_header(
996 mmsg_vector->msg_hdr.msg_iov->iov_base,
997 skb,
998 vp
999 );
1000 if (header_check < 0) {
1001 /* Overlay header failed to verify - discard.
1002 * We can actually keep this skb and reuse it,
1003 * but that will make the prep logic too
1004 * complex.
1005 */
1006 dev_kfree_skb_irq(skb);
1007 vp->estats.rx_encaps_errors++;
1008 continue;
1009 }
1010 if (header_check > 0) {
1011 vp->estats.rx_csum_offload_good++;
1012 skb->ip_summed = CHECKSUM_UNNECESSARY;
1013 }
1014 }
1015 pskb_trim(skb,
1016 mmsg_vector->msg_len - vp->rx_header_size);
1017 skb->protocol = eth_type_trans(skb, skb->dev);
1018 /*
1019 * We do not need to lock on updating stats here
1020 * The interrupt loop is non-reentrant.
1021 */
1022 vp->dev->stats.rx_bytes += skb->len;
1023 vp->dev->stats.rx_packets++;
1024 netif_rx(skb);
1025 } else {
1026 /* Overlay header too short to do anything - discard.
1027 * We can actually keep this skb and reuse it,
1028 * but that will make the prep logic too complex.
1029 */
1030 if (skb != NULL)
1031 dev_kfree_skb_irq(skb);
1032 }
1033 (*skbuff_vector) = NULL;
1034 /* Move to the next buffer element */
1035 mmsg_vector++;
1036 skbuff_vector++;
1037 }
1038 if (packet_count > 0) {
1039 if (vp->estats.rx_queue_max < packet_count)
1040 vp->estats.rx_queue_max = packet_count;
1041 vp->estats.rx_queue_running_average =
1042 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1043 }
1044 return packet_count;
1045}
1046
1047static void vector_rx(struct vector_private *vp)
1048{
1049 int err;
1050 int iter = 0;
1051
1052 if ((vp->options & VECTOR_RX) > 0)
1053 while (((err = vector_mmsg_rx(vp)) > 0) && (iter < MAX_ITERATIONS))
1054 iter++;
1055 else
1056 while (((err = vector_legacy_rx(vp)) > 0) && (iter < MAX_ITERATIONS))
1057 iter++;
1058 if ((err != 0) && net_ratelimit())
1059 netdev_err(vp->dev, "vector_rx: error(%d)\n", err);
1060 if (iter == MAX_ITERATIONS)
1061 netdev_err(vp->dev, "vector_rx: device stuck, remote end may have closed the connection\n");
1062}
1063
1064static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1065{
1066 struct vector_private *vp = netdev_priv(dev);
1067 int queue_depth = 0;
1068
1069 if (vp->in_error) {
1070 deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1071 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1072 deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1073 return NETDEV_TX_BUSY;
1074 }
1075
1076 if ((vp->options & VECTOR_TX) == 0) {
1077 writev_tx(vp, skb);
1078 return NETDEV_TX_OK;
1079 }
1080
1081 /* We do BQL only in the vector path, no point doing it in
1082 * packet at a time mode as there is no device queue
1083 */
1084
1085 netdev_sent_queue(vp->dev, skb->len);
1086 queue_depth = vector_enqueue(vp->tx_queue, skb);
1087
1088 /* if the device queue is full, stop the upper layers and
1089 * flush it.
1090 */
1091
1092 if (queue_depth >= vp->tx_queue->max_depth - 1) {
1093 vp->estats.tx_kicks++;
1094 netif_stop_queue(dev);
1095 vector_send(vp->tx_queue);
1096 return NETDEV_TX_OK;
1097 }
1098 if (netdev_xmit_more()) {
1099 mod_timer(&vp->tl, vp->coalesce);
1100 return NETDEV_TX_OK;
1101 }
1102 if (skb->len < TX_SMALL_PACKET) {
1103 vp->estats.tx_kicks++;
1104 vector_send(vp->tx_queue);
1105 } else
1106 tasklet_schedule(&vp->tx_poll);
1107 return NETDEV_TX_OK;
1108}
1109
1110static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1111{
1112 struct net_device *dev = dev_id;
1113 struct vector_private *vp = netdev_priv(dev);
1114
1115 if (!netif_running(dev))
1116 return IRQ_NONE;
1117 vector_rx(vp);
1118 return IRQ_HANDLED;
1119
1120}
1121
1122static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1123{
1124 struct net_device *dev = dev_id;
1125 struct vector_private *vp = netdev_priv(dev);
1126
1127 if (!netif_running(dev))
1128 return IRQ_NONE;
1129 /* We need to pay attention to it only if we got
1130 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1131 * we ignore it. In the future, it may be worth
1132 * it to improve the IRQ controller a bit to make
1133 * tweaking the IRQ mask less costly
1134 */
1135
1136 if (vp->in_write_poll)
1137 tasklet_schedule(&vp->tx_poll);
1138 return IRQ_HANDLED;
1139
1140}
1141
1142static int irq_rr;
1143
1144static int vector_net_close(struct net_device *dev)
1145{
1146 struct vector_private *vp = netdev_priv(dev);
1147 unsigned long flags;
1148
1149 netif_stop_queue(dev);
1150 del_timer(&vp->tl);
1151
1152 if (vp->fds == NULL)
1153 return 0;
1154
1155 /* Disable and free all IRQS */
1156 if (vp->rx_irq > 0) {
1157 um_free_irq(vp->rx_irq, dev);
1158 vp->rx_irq = 0;
1159 }
1160 if (vp->tx_irq > 0) {
1161 um_free_irq(vp->tx_irq, dev);
1162 vp->tx_irq = 0;
1163 }
1164 tasklet_kill(&vp->tx_poll);
1165 if (vp->fds->rx_fd > 0) {
1166 if (vp->bpf)
1167 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1168 os_close_file(vp->fds->rx_fd);
1169 vp->fds->rx_fd = -1;
1170 }
1171 if (vp->fds->tx_fd > 0) {
1172 os_close_file(vp->fds->tx_fd);
1173 vp->fds->tx_fd = -1;
1174 }
1175 if (vp->bpf != NULL)
1176 kfree(vp->bpf->filter);
1177 kfree(vp->bpf);
1178 vp->bpf = NULL;
1179 kfree(vp->fds->remote_addr);
1180 kfree(vp->transport_data);
1181 kfree(vp->header_rxbuffer);
1182 kfree(vp->header_txbuffer);
1183 if (vp->rx_queue != NULL)
1184 destroy_queue(vp->rx_queue);
1185 if (vp->tx_queue != NULL)
1186 destroy_queue(vp->tx_queue);
1187 kfree(vp->fds);
1188 vp->fds = NULL;
1189 spin_lock_irqsave(&vp->lock, flags);
1190 vp->opened = false;
1191 vp->in_error = false;
1192 spin_unlock_irqrestore(&vp->lock, flags);
1193 return 0;
1194}
1195
1196/* TX tasklet */
1197
1198static void vector_tx_poll(struct tasklet_struct *t)
1199{
1200 struct vector_private *vp = from_tasklet(vp, t, tx_poll);
1201
1202 vp->estats.tx_kicks++;
1203 vector_send(vp->tx_queue);
1204}
1205static void vector_reset_tx(struct work_struct *work)
1206{
1207 struct vector_private *vp =
1208 container_of(work, struct vector_private, reset_tx);
1209 netdev_reset_queue(vp->dev);
1210 netif_start_queue(vp->dev);
1211 netif_wake_queue(vp->dev);
1212}
1213
1214static int vector_net_open(struct net_device *dev)
1215{
1216 struct vector_private *vp = netdev_priv(dev);
1217 unsigned long flags;
1218 int err = -EINVAL;
1219 struct vector_device *vdevice;
1220
1221 spin_lock_irqsave(&vp->lock, flags);
1222 if (vp->opened) {
1223 spin_unlock_irqrestore(&vp->lock, flags);
1224 return -ENXIO;
1225 }
1226 vp->opened = true;
1227 spin_unlock_irqrestore(&vp->lock, flags);
1228
1229 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1230
1231 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1232
1233 if (vp->fds == NULL)
1234 goto out_close;
1235
1236 if (build_transport_data(vp) < 0)
1237 goto out_close;
1238
1239 if ((vp->options & VECTOR_RX) > 0) {
1240 vp->rx_queue = create_queue(
1241 vp,
1242 get_depth(vp->parsed),
1243 vp->rx_header_size,
1244 MAX_IOV_SIZE
1245 );
1246 vp->rx_queue->queue_depth = get_depth(vp->parsed);
1247 } else {
1248 vp->header_rxbuffer = kmalloc(
1249 vp->rx_header_size,
1250 GFP_KERNEL
1251 );
1252 if (vp->header_rxbuffer == NULL)
1253 goto out_close;
1254 }
1255 if ((vp->options & VECTOR_TX) > 0) {
1256 vp->tx_queue = create_queue(
1257 vp,
1258 get_depth(vp->parsed),
1259 vp->header_size,
1260 MAX_IOV_SIZE
1261 );
1262 } else {
1263 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1264 if (vp->header_txbuffer == NULL)
1265 goto out_close;
1266 }
1267
1268 /* READ IRQ */
1269 err = um_request_irq(
1270 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1271 IRQ_READ, vector_rx_interrupt,
1272 IRQF_SHARED, dev->name, dev);
1273 if (err < 0) {
1274 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1275 err = -ENETUNREACH;
1276 goto out_close;
1277 }
1278 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1279 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1280 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1281
1282 /* WRITE IRQ - we need it only if we have vector TX */
1283 if ((vp->options & VECTOR_TX) > 0) {
1284 err = um_request_irq(
1285 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1286 IRQ_WRITE, vector_tx_interrupt,
1287 IRQF_SHARED, dev->name, dev);
1288 if (err < 0) {
1289 netdev_err(dev,
1290 "vector_open: failed to get tx irq(%d)\n", err);
1291 err = -ENETUNREACH;
1292 goto out_close;
1293 }
1294 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1295 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1296 }
1297
1298 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1299 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1300 vp->options |= VECTOR_BPF;
1301 }
1302 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1303 vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1304
1305 if (vp->bpf != NULL)
1306 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1307
1308 netif_start_queue(dev);
1309
1310 /* clear buffer - it can happen that the host side of the interface
1311 * is full when we get here. In this case, new data is never queued,
1312 * SIGIOs never arrive, and the net never works.
1313 */
1314
1315 vector_rx(vp);
1316
1317 vector_reset_stats(vp);
1318 vdevice = find_device(vp->unit);
1319 vdevice->opened = 1;
1320
1321 if ((vp->options & VECTOR_TX) != 0)
1322 add_timer(&vp->tl);
1323 return 0;
1324out_close:
1325 vector_net_close(dev);
1326 return err;
1327}
1328
1329
1330static void vector_net_set_multicast_list(struct net_device *dev)
1331{
1332 /* TODO: - we can do some BPF games here */
1333 return;
1334}
1335
1336static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1337{
1338 struct vector_private *vp = netdev_priv(dev);
1339
1340 vp->estats.tx_timeout_count++;
1341 netif_trans_update(dev);
1342 schedule_work(&vp->reset_tx);
1343}
1344
1345static netdev_features_t vector_fix_features(struct net_device *dev,
1346 netdev_features_t features)
1347{
1348 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1349 return features;
1350}
1351
1352static int vector_set_features(struct net_device *dev,
1353 netdev_features_t features)
1354{
1355 struct vector_private *vp = netdev_priv(dev);
1356 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1357 * no way to negotiate it on raw sockets, so we can change
1358 * only our side.
1359 */
1360 if (features & NETIF_F_GRO)
1361 /* All new frame buffers will be GRO-sized */
1362 vp->req_size = 65536;
1363 else
1364 /* All new frame buffers will be normal sized */
1365 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1366 return 0;
1367}
1368
1369#ifdef CONFIG_NET_POLL_CONTROLLER
1370static void vector_net_poll_controller(struct net_device *dev)
1371{
1372 disable_irq(dev->irq);
1373 vector_rx_interrupt(dev->irq, dev);
1374 enable_irq(dev->irq);
1375}
1376#endif
1377
1378static void vector_net_get_drvinfo(struct net_device *dev,
1379 struct ethtool_drvinfo *info)
1380{
1381 strlcpy(info->driver, DRIVER_NAME, sizeof(info->driver));
1382}
1383
1384static int vector_net_load_bpf_flash(struct net_device *dev,
1385 struct ethtool_flash *efl)
1386{
1387 struct vector_private *vp = netdev_priv(dev);
1388 struct vector_device *vdevice;
1389 const struct firmware *fw;
1390 int result = 0;
1391
1392 if (!(vp->options & VECTOR_BPF_FLASH)) {
1393 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1394 return -1;
1395 }
1396
1397 spin_lock(&vp->lock);
1398
1399 if (vp->bpf != NULL) {
1400 if (vp->opened)
1401 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1402 kfree(vp->bpf->filter);
1403 vp->bpf->filter = NULL;
1404 } else {
1405 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC);
1406 if (vp->bpf == NULL) {
1407 netdev_err(dev, "failed to allocate memory for firmware\n");
1408 goto flash_fail;
1409 }
1410 }
1411
1412 vdevice = find_device(vp->unit);
1413
1414 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1415 goto flash_fail;
1416
1417 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC);
1418 if (!vp->bpf->filter)
1419 goto free_buffer;
1420
1421 vp->bpf->len = fw->size / sizeof(struct sock_filter);
1422 release_firmware(fw);
1423
1424 if (vp->opened)
1425 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1426
1427 spin_unlock(&vp->lock);
1428
1429 return result;
1430
1431free_buffer:
1432 release_firmware(fw);
1433
1434flash_fail:
1435 spin_unlock(&vp->lock);
1436 if (vp->bpf != NULL)
1437 kfree(vp->bpf->filter);
1438 kfree(vp->bpf);
1439 vp->bpf = NULL;
1440 return -1;
1441}
1442
1443static void vector_get_ringparam(struct net_device *netdev,
1444 struct ethtool_ringparam *ring)
1445{
1446 struct vector_private *vp = netdev_priv(netdev);
1447
1448 ring->rx_max_pending = vp->rx_queue->max_depth;
1449 ring->tx_max_pending = vp->tx_queue->max_depth;
1450 ring->rx_pending = vp->rx_queue->max_depth;
1451 ring->tx_pending = vp->tx_queue->max_depth;
1452}
1453
1454static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1455{
1456 switch (stringset) {
1457 case ETH_SS_TEST:
1458 *buf = '\0';
1459 break;
1460 case ETH_SS_STATS:
1461 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1462 break;
1463 default:
1464 WARN_ON(1);
1465 break;
1466 }
1467}
1468
1469static int vector_get_sset_count(struct net_device *dev, int sset)
1470{
1471 switch (sset) {
1472 case ETH_SS_TEST:
1473 return 0;
1474 case ETH_SS_STATS:
1475 return VECTOR_NUM_STATS;
1476 default:
1477 return -EOPNOTSUPP;
1478 }
1479}
1480
1481static void vector_get_ethtool_stats(struct net_device *dev,
1482 struct ethtool_stats *estats,
1483 u64 *tmp_stats)
1484{
1485 struct vector_private *vp = netdev_priv(dev);
1486
1487 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1488}
1489
1490static int vector_get_coalesce(struct net_device *netdev,
1491 struct ethtool_coalesce *ec)
1492{
1493 struct vector_private *vp = netdev_priv(netdev);
1494
1495 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1496 return 0;
1497}
1498
1499static int vector_set_coalesce(struct net_device *netdev,
1500 struct ethtool_coalesce *ec)
1501{
1502 struct vector_private *vp = netdev_priv(netdev);
1503
1504 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1505 if (vp->coalesce == 0)
1506 vp->coalesce = 1;
1507 return 0;
1508}
1509
1510static const struct ethtool_ops vector_net_ethtool_ops = {
1511 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1512 .get_drvinfo = vector_net_get_drvinfo,
1513 .get_link = ethtool_op_get_link,
1514 .get_ts_info = ethtool_op_get_ts_info,
1515 .get_ringparam = vector_get_ringparam,
1516 .get_strings = vector_get_strings,
1517 .get_sset_count = vector_get_sset_count,
1518 .get_ethtool_stats = vector_get_ethtool_stats,
1519 .get_coalesce = vector_get_coalesce,
1520 .set_coalesce = vector_set_coalesce,
1521 .flash_device = vector_net_load_bpf_flash,
1522};
1523
1524
1525static const struct net_device_ops vector_netdev_ops = {
1526 .ndo_open = vector_net_open,
1527 .ndo_stop = vector_net_close,
1528 .ndo_start_xmit = vector_net_start_xmit,
1529 .ndo_set_rx_mode = vector_net_set_multicast_list,
1530 .ndo_tx_timeout = vector_net_tx_timeout,
1531 .ndo_set_mac_address = eth_mac_addr,
1532 .ndo_validate_addr = eth_validate_addr,
1533 .ndo_fix_features = vector_fix_features,
1534 .ndo_set_features = vector_set_features,
1535#ifdef CONFIG_NET_POLL_CONTROLLER
1536 .ndo_poll_controller = vector_net_poll_controller,
1537#endif
1538};
1539
1540
1541static void vector_timer_expire(struct timer_list *t)
1542{
1543 struct vector_private *vp = from_timer(vp, t, tl);
1544
1545 vp->estats.tx_kicks++;
1546 vector_send(vp->tx_queue);
1547}
1548
1549static void vector_eth_configure(
1550 int n,
1551 struct arglist *def
1552 )
1553{
1554 struct vector_device *device;
1555 struct net_device *dev;
1556 struct vector_private *vp;
1557 int err;
1558
1559 device = kzalloc(sizeof(*device), GFP_KERNEL);
1560 if (device == NULL) {
1561 printk(KERN_ERR "eth_configure failed to allocate struct "
1562 "vector_device\n");
1563 return;
1564 }
1565 dev = alloc_etherdev(sizeof(struct vector_private));
1566 if (dev == NULL) {
1567 printk(KERN_ERR "eth_configure: failed to allocate struct "
1568 "net_device for vec%d\n", n);
1569 goto out_free_device;
1570 }
1571
1572 dev->mtu = get_mtu(def);
1573
1574 INIT_LIST_HEAD(&device->list);
1575 device->unit = n;
1576
1577 /* If this name ends up conflicting with an existing registered
1578 * netdevice, that is OK, register_netdev{,ice}() will notice this
1579 * and fail.
1580 */
1581 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1582 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1583 vp = netdev_priv(dev);
1584
1585 /* sysfs register */
1586 if (!driver_registered) {
1587 platform_driver_register(¨_net_driver);
1588 driver_registered = 1;
1589 }
1590 device->pdev.id = n;
1591 device->pdev.name = DRIVER_NAME;
1592 device->pdev.dev.release = vector_device_release;
1593 dev_set_drvdata(&device->pdev.dev, device);
1594 if (platform_device_register(&device->pdev))
1595 goto out_free_netdev;
1596 SET_NETDEV_DEV(dev, &device->pdev.dev);
1597
1598 device->dev = dev;
1599
1600 *vp = ((struct vector_private)
1601 {
1602 .list = LIST_HEAD_INIT(vp->list),
1603 .dev = dev,
1604 .unit = n,
1605 .options = get_transport_options(def),
1606 .rx_irq = 0,
1607 .tx_irq = 0,
1608 .parsed = def,
1609 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1610 /* TODO - we need to calculate headroom so that ip header
1611 * is 16 byte aligned all the time
1612 */
1613 .headroom = get_headroom(def),
1614 .form_header = NULL,
1615 .verify_header = NULL,
1616 .header_rxbuffer = NULL,
1617 .header_txbuffer = NULL,
1618 .header_size = 0,
1619 .rx_header_size = 0,
1620 .rexmit_scheduled = false,
1621 .opened = false,
1622 .transport_data = NULL,
1623 .in_write_poll = false,
1624 .coalesce = 2,
1625 .req_size = get_req_size(def),
1626 .in_error = false,
1627 .bpf = NULL
1628 });
1629
1630 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1631 tasklet_setup(&vp->tx_poll, vector_tx_poll);
1632 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1633
1634 timer_setup(&vp->tl, vector_timer_expire, 0);
1635 spin_lock_init(&vp->lock);
1636
1637 /* FIXME */
1638 dev->netdev_ops = &vector_netdev_ops;
1639 dev->ethtool_ops = &vector_net_ethtool_ops;
1640 dev->watchdog_timeo = (HZ >> 1);
1641 /* primary IRQ - fixme */
1642 dev->irq = 0; /* we will adjust this once opened */
1643
1644 rtnl_lock();
1645 err = register_netdevice(dev);
1646 rtnl_unlock();
1647 if (err)
1648 goto out_undo_user_init;
1649
1650 spin_lock(&vector_devices_lock);
1651 list_add(&device->list, &vector_devices);
1652 spin_unlock(&vector_devices_lock);
1653
1654 return;
1655
1656out_undo_user_init:
1657 return;
1658out_free_netdev:
1659 free_netdev(dev);
1660out_free_device:
1661 kfree(device);
1662}
1663
1664
1665
1666
1667/*
1668 * Invoked late in the init
1669 */
1670
1671static int __init vector_init(void)
1672{
1673 struct list_head *ele;
1674 struct vector_cmd_line_arg *def;
1675 struct arglist *parsed;
1676
1677 list_for_each(ele, &vec_cmd_line) {
1678 def = list_entry(ele, struct vector_cmd_line_arg, list);
1679 parsed = uml_parse_vector_ifspec(def->arguments);
1680 if (parsed != NULL)
1681 vector_eth_configure(def->unit, parsed);
1682 }
1683 return 0;
1684}
1685
1686
1687/* Invoked at initial argument parsing, only stores
1688 * arguments until a proper vector_init is called
1689 * later
1690 */
1691
1692static int __init vector_setup(char *str)
1693{
1694 char *error;
1695 int n, err;
1696 struct vector_cmd_line_arg *new;
1697
1698 err = vector_parse(str, &n, &str, &error);
1699 if (err) {
1700 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1701 str, error);
1702 return 1;
1703 }
1704 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1705 if (!new)
1706 panic("%s: Failed to allocate %zu bytes\n", __func__,
1707 sizeof(*new));
1708 INIT_LIST_HEAD(&new->list);
1709 new->unit = n;
1710 new->arguments = str;
1711 list_add_tail(&new->list, &vec_cmd_line);
1712 return 1;
1713}
1714
1715__setup("vec", vector_setup);
1716__uml_help(vector_setup,
1717"vec[0-9]+:<option>=<value>,<option>=<value>\n"
1718" Configure a vector io network device.\n\n"
1719);
1720
1721late_initcall(vector_init);
1722
1723static struct mc_device vector_mc = {
1724 .list = LIST_HEAD_INIT(vector_mc.list),
1725 .name = "vec",
1726 .config = vector_config,
1727 .get_config = NULL,
1728 .id = vector_id,
1729 .remove = vector_remove,
1730};
1731
1732#ifdef CONFIG_INET
1733static int vector_inetaddr_event(
1734 struct notifier_block *this,
1735 unsigned long event,
1736 void *ptr)
1737{
1738 return NOTIFY_DONE;
1739}
1740
1741static struct notifier_block vector_inetaddr_notifier = {
1742 .notifier_call = vector_inetaddr_event,
1743};
1744
1745static void inet_register(void)
1746{
1747 register_inetaddr_notifier(&vector_inetaddr_notifier);
1748}
1749#else
1750static inline void inet_register(void)
1751{
1752}
1753#endif
1754
1755static int vector_net_init(void)
1756{
1757 mconsole_register_dev(&vector_mc);
1758 inet_register();
1759 return 0;
1760}
1761
1762__initcall(vector_net_init);
1763
1764
1765