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 - Cambridge Greys Limited
4 * Copyright (C) 2011 - 2014 Cisco Systems Inc
5 * Copyright (C) 2001 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
6 * Copyright (C) 2001 Lennert Buytenhek (buytenh@gnu.org) and
7 * James Leu (jleu@mindspring.net).
8 * Copyright (C) 2001 by various other people who didn't put their name here.
9 */
10
11#include <linux/version.h>
12#include <linux/memblock.h>
13#include <linux/etherdevice.h>
14#include <linux/ethtool.h>
15#include <linux/inetdevice.h>
16#include <linux/init.h>
17#include <linux/list.h>
18#include <linux/netdevice.h>
19#include <linux/platform_device.h>
20#include <linux/rtnetlink.h>
21#include <linux/skbuff.h>
22#include <linux/slab.h>
23#include <linux/interrupt.h>
24#include <init.h>
25#include <irq_kern.h>
26#include <irq_user.h>
27#include <net_kern.h>
28#include <os.h>
29#include "mconsole_kern.h"
30#include "vector_user.h"
31#include "vector_kern.h"
32
33/*
34 * Adapted from network devices with the following major changes:
35 * All transports are static - simplifies the code significantly
36 * Multiple FDs/IRQs per device
37 * Vector IO optionally used for read/write, falling back to legacy
38 * based on configuration and/or availability
39 * Configuration is no longer positional - L2TPv3 and GRE require up to
40 * 10 parameters, passing this as positional is not fit for purpose.
41 * Only socket transports are supported
42 */
43
44
45#define DRIVER_NAME "uml-vector"
46#define DRIVER_VERSION "01"
47struct vector_cmd_line_arg {
48 struct list_head list;
49 int unit;
50 char *arguments;
51};
52
53struct vector_device {
54 struct list_head list;
55 struct net_device *dev;
56 struct platform_device pdev;
57 int unit;
58 int opened;
59};
60
61static LIST_HEAD(vec_cmd_line);
62
63static DEFINE_SPINLOCK(vector_devices_lock);
64static LIST_HEAD(vector_devices);
65
66static int driver_registered;
67
68static void vector_eth_configure(int n, struct arglist *def);
69
70/* Argument accessors to set variables (and/or set default values)
71 * mtu, buffer sizing, default headroom, etc
72 */
73
74#define DEFAULT_HEADROOM 2
75#define SAFETY_MARGIN 32
76#define DEFAULT_VECTOR_SIZE 64
77#define TX_SMALL_PACKET 128
78#define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
79#define MAX_ITERATIONS 64
80
81static const struct {
82 const char string[ETH_GSTRING_LEN];
83} ethtool_stats_keys[] = {
84 { "rx_queue_max" },
85 { "rx_queue_running_average" },
86 { "tx_queue_max" },
87 { "tx_queue_running_average" },
88 { "rx_encaps_errors" },
89 { "tx_timeout_count" },
90 { "tx_restart_queue" },
91 { "tx_kicks" },
92 { "tx_flow_control_xon" },
93 { "tx_flow_control_xoff" },
94 { "rx_csum_offload_good" },
95 { "rx_csum_offload_errors"},
96 { "sg_ok"},
97 { "sg_linearized"},
98};
99
100#define VECTOR_NUM_STATS ARRAY_SIZE(ethtool_stats_keys)
101
102static void vector_reset_stats(struct vector_private *vp)
103{
104 vp->estats.rx_queue_max = 0;
105 vp->estats.rx_queue_running_average = 0;
106 vp->estats.tx_queue_max = 0;
107 vp->estats.tx_queue_running_average = 0;
108 vp->estats.rx_encaps_errors = 0;
109 vp->estats.tx_timeout_count = 0;
110 vp->estats.tx_restart_queue = 0;
111 vp->estats.tx_kicks = 0;
112 vp->estats.tx_flow_control_xon = 0;
113 vp->estats.tx_flow_control_xoff = 0;
114 vp->estats.sg_ok = 0;
115 vp->estats.sg_linearized = 0;
116}
117
118static int get_mtu(struct arglist *def)
119{
120 char *mtu = uml_vector_fetch_arg(def, "mtu");
121 long result;
122
123 if (mtu != NULL) {
124 if (kstrtoul(mtu, 10, &result) == 0)
125 if ((result < (1 << 16) - 1) && (result >= 576))
126 return result;
127 }
128 return ETH_MAX_PACKET;
129}
130
131static int get_depth(struct arglist *def)
132{
133 char *mtu = uml_vector_fetch_arg(def, "depth");
134 long result;
135
136 if (mtu != NULL) {
137 if (kstrtoul(mtu, 10, &result) == 0)
138 return result;
139 }
140 return DEFAULT_VECTOR_SIZE;
141}
142
143static int get_headroom(struct arglist *def)
144{
145 char *mtu = uml_vector_fetch_arg(def, "headroom");
146 long result;
147
148 if (mtu != NULL) {
149 if (kstrtoul(mtu, 10, &result) == 0)
150 return result;
151 }
152 return DEFAULT_HEADROOM;
153}
154
155static int get_req_size(struct arglist *def)
156{
157 char *gro = uml_vector_fetch_arg(def, "gro");
158 long result;
159
160 if (gro != NULL) {
161 if (kstrtoul(gro, 10, &result) == 0) {
162 if (result > 0)
163 return 65536;
164 }
165 }
166 return get_mtu(def) + ETH_HEADER_OTHER +
167 get_headroom(def) + SAFETY_MARGIN;
168}
169
170
171static int get_transport_options(struct arglist *def)
172{
173 char *transport = uml_vector_fetch_arg(def, "transport");
174 char *vector = uml_vector_fetch_arg(def, "vec");
175
176 int vec_rx = VECTOR_RX;
177 int vec_tx = VECTOR_TX;
178 long parsed;
179
180 if (vector != NULL) {
181 if (kstrtoul(vector, 10, &parsed) == 0) {
182 if (parsed == 0) {
183 vec_rx = 0;
184 vec_tx = 0;
185 }
186 }
187 }
188
189
190 if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
191 return 0;
192 if (strncmp(transport, TRANS_HYBRID, TRANS_HYBRID_LEN) == 0)
193 return (vec_rx | VECTOR_BPF);
194 if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
195 return (vec_rx | vec_tx | VECTOR_QDISC_BYPASS);
196 return (vec_rx | vec_tx);
197}
198
199
200/* A mini-buffer for packet drop read
201 * All of our supported transports are datagram oriented and we always
202 * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
203 * than the packet size it still counts as full packet read and will
204 * clean the incoming stream to keep sigio/epoll happy
205 */
206
207#define DROP_BUFFER_SIZE 32
208
209static char *drop_buffer;
210
211/* Array backed queues optimized for bulk enqueue/dequeue and
212 * 1:N (small values of N) or 1:1 enqueuer/dequeuer ratios.
213 * For more details and full design rationale see
214 * http://foswiki.cambridgegreys.com/Main/EatYourTailAndEnjoyIt
215 */
216
217
218/*
219 * Advance the mmsg queue head by n = advance. Resets the queue to
220 * maximum enqueue/dequeue-at-once capacity if possible. Called by
221 * dequeuers. Caller must hold the head_lock!
222 */
223
224static int vector_advancehead(struct vector_queue *qi, int advance)
225{
226 int queue_depth;
227
228 qi->head =
229 (qi->head + advance)
230 % qi->max_depth;
231
232
233 spin_lock(&qi->tail_lock);
234 qi->queue_depth -= advance;
235
236 /* we are at 0, use this to
237 * reset head and tail so we can use max size vectors
238 */
239
240 if (qi->queue_depth == 0) {
241 qi->head = 0;
242 qi->tail = 0;
243 }
244 queue_depth = qi->queue_depth;
245 spin_unlock(&qi->tail_lock);
246 return queue_depth;
247}
248
249/* Advance the queue tail by n = advance.
250 * This is called by enqueuers which should hold the
251 * head lock already
252 */
253
254static int vector_advancetail(struct vector_queue *qi, int advance)
255{
256 int queue_depth;
257
258 qi->tail =
259 (qi->tail + advance)
260 % qi->max_depth;
261 spin_lock(&qi->head_lock);
262 qi->queue_depth += advance;
263 queue_depth = qi->queue_depth;
264 spin_unlock(&qi->head_lock);
265 return queue_depth;
266}
267
268static int prep_msg(struct vector_private *vp,
269 struct sk_buff *skb,
270 struct iovec *iov)
271{
272 int iov_index = 0;
273 int nr_frags, frag;
274 skb_frag_t *skb_frag;
275
276 nr_frags = skb_shinfo(skb)->nr_frags;
277 if (nr_frags > MAX_IOV_SIZE) {
278 if (skb_linearize(skb) != 0)
279 goto drop;
280 }
281 if (vp->header_size > 0) {
282 iov[iov_index].iov_len = vp->header_size;
283 vp->form_header(iov[iov_index].iov_base, skb, vp);
284 iov_index++;
285 }
286 iov[iov_index].iov_base = skb->data;
287 if (nr_frags > 0) {
288 iov[iov_index].iov_len = skb->len - skb->data_len;
289 vp->estats.sg_ok++;
290 } else
291 iov[iov_index].iov_len = skb->len;
292 iov_index++;
293 for (frag = 0; frag < nr_frags; frag++) {
294 skb_frag = &skb_shinfo(skb)->frags[frag];
295 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
296 iov[iov_index].iov_len = skb_frag_size(skb_frag);
297 iov_index++;
298 }
299 return iov_index;
300drop:
301 return -1;
302}
303/*
304 * Generic vector enqueue with support for forming headers using transport
305 * specific callback. Allows GRE, L2TPv3, RAW and other transports
306 * to use a common enqueue procedure in vector mode
307 */
308
309static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
310{
311 struct vector_private *vp = netdev_priv(qi->dev);
312 int queue_depth;
313 int packet_len;
314 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
315 int iov_count;
316
317 spin_lock(&qi->tail_lock);
318 spin_lock(&qi->head_lock);
319 queue_depth = qi->queue_depth;
320 spin_unlock(&qi->head_lock);
321
322 if (skb)
323 packet_len = skb->len;
324
325 if (queue_depth < qi->max_depth) {
326
327 *(qi->skbuff_vector + qi->tail) = skb;
328 mmsg_vector += qi->tail;
329 iov_count = prep_msg(
330 vp,
331 skb,
332 mmsg_vector->msg_hdr.msg_iov
333 );
334 if (iov_count < 1)
335 goto drop;
336 mmsg_vector->msg_hdr.msg_iovlen = iov_count;
337 mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
338 mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
339 queue_depth = vector_advancetail(qi, 1);
340 } else
341 goto drop;
342 spin_unlock(&qi->tail_lock);
343 return queue_depth;
344drop:
345 qi->dev->stats.tx_dropped++;
346 if (skb != NULL) {
347 packet_len = skb->len;
348 dev_consume_skb_any(skb);
349 netdev_completed_queue(qi->dev, 1, packet_len);
350 }
351 spin_unlock(&qi->tail_lock);
352 return queue_depth;
353}
354
355static int consume_vector_skbs(struct vector_queue *qi, int count)
356{
357 struct sk_buff *skb;
358 int skb_index;
359 int bytes_compl = 0;
360
361 for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
362 skb = *(qi->skbuff_vector + skb_index);
363 /* mark as empty to ensure correct destruction if
364 * needed
365 */
366 bytes_compl += skb->len;
367 *(qi->skbuff_vector + skb_index) = NULL;
368 dev_consume_skb_any(skb);
369 }
370 qi->dev->stats.tx_bytes += bytes_compl;
371 qi->dev->stats.tx_packets += count;
372 netdev_completed_queue(qi->dev, count, bytes_compl);
373 return vector_advancehead(qi, count);
374}
375
376/*
377 * Generic vector deque via sendmmsg with support for forming headers
378 * using transport specific callback. Allows GRE, L2TPv3, RAW and
379 * other transports to use a common dequeue procedure in vector mode
380 */
381
382
383static int vector_send(struct vector_queue *qi)
384{
385 struct vector_private *vp = netdev_priv(qi->dev);
386 struct mmsghdr *send_from;
387 int result = 0, send_len, queue_depth = qi->max_depth;
388
389 if (spin_trylock(&qi->head_lock)) {
390 if (spin_trylock(&qi->tail_lock)) {
391 /* update queue_depth to current value */
392 queue_depth = qi->queue_depth;
393 spin_unlock(&qi->tail_lock);
394 while (queue_depth > 0) {
395 /* Calculate the start of the vector */
396 send_len = queue_depth;
397 send_from = qi->mmsg_vector;
398 send_from += qi->head;
399 /* Adjust vector size if wraparound */
400 if (send_len + qi->head > qi->max_depth)
401 send_len = qi->max_depth - qi->head;
402 /* Try to TX as many packets as possible */
403 if (send_len > 0) {
404 result = uml_vector_sendmmsg(
405 vp->fds->tx_fd,
406 send_from,
407 send_len,
408 0
409 );
410 vp->in_write_poll =
411 (result != send_len);
412 }
413 /* For some of the sendmmsg error scenarios
414 * we may end being unsure in the TX success
415 * for all packets. It is safer to declare
416 * them all TX-ed and blame the network.
417 */
418 if (result < 0) {
419 if (net_ratelimit())
420 netdev_err(vp->dev, "sendmmsg err=%i\n",
421 result);
422 vp->in_error = true;
423 result = send_len;
424 }
425 if (result > 0) {
426 queue_depth =
427 consume_vector_skbs(qi, result);
428 /* This is equivalent to an TX IRQ.
429 * Restart the upper layers to feed us
430 * more packets.
431 */
432 if (result > vp->estats.tx_queue_max)
433 vp->estats.tx_queue_max = result;
434 vp->estats.tx_queue_running_average =
435 (vp->estats.tx_queue_running_average + result) >> 1;
436 }
437 netif_trans_update(qi->dev);
438 netif_wake_queue(qi->dev);
439 /* if TX is busy, break out of the send loop,
440 * poll write IRQ will reschedule xmit for us
441 */
442 if (result != send_len) {
443 vp->estats.tx_restart_queue++;
444 break;
445 }
446 }
447 }
448 spin_unlock(&qi->head_lock);
449 } else {
450 tasklet_schedule(&vp->tx_poll);
451 }
452 return queue_depth;
453}
454
455/* Queue destructor. Deliberately stateless so we can use
456 * it in queue cleanup if initialization fails.
457 */
458
459static void destroy_queue(struct vector_queue *qi)
460{
461 int i;
462 struct iovec *iov;
463 struct vector_private *vp = netdev_priv(qi->dev);
464 struct mmsghdr *mmsg_vector;
465
466 if (qi == NULL)
467 return;
468 /* deallocate any skbuffs - we rely on any unused to be
469 * set to NULL.
470 */
471 if (qi->skbuff_vector != NULL) {
472 for (i = 0; i < qi->max_depth; i++) {
473 if (*(qi->skbuff_vector + i) != NULL)
474 dev_kfree_skb_any(*(qi->skbuff_vector + i));
475 }
476 kfree(qi->skbuff_vector);
477 }
478 /* deallocate matching IOV structures including header buffs */
479 if (qi->mmsg_vector != NULL) {
480 mmsg_vector = qi->mmsg_vector;
481 for (i = 0; i < qi->max_depth; i++) {
482 iov = mmsg_vector->msg_hdr.msg_iov;
483 if (iov != NULL) {
484 if ((vp->header_size > 0) &&
485 (iov->iov_base != NULL))
486 kfree(iov->iov_base);
487 kfree(iov);
488 }
489 mmsg_vector++;
490 }
491 kfree(qi->mmsg_vector);
492 }
493 kfree(qi);
494}
495
496/*
497 * Queue constructor. Create a queue with a given side.
498 */
499static struct vector_queue *create_queue(
500 struct vector_private *vp,
501 int max_size,
502 int header_size,
503 int num_extra_frags)
504{
505 struct vector_queue *result;
506 int i;
507 struct iovec *iov;
508 struct mmsghdr *mmsg_vector;
509
510 result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
511 if (result == NULL)
512 return NULL;
513 result->max_depth = max_size;
514 result->dev = vp->dev;
515 result->mmsg_vector = kmalloc(
516 (sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
517 if (result->mmsg_vector == NULL)
518 goto out_mmsg_fail;
519 result->skbuff_vector = kmalloc(
520 (sizeof(void *) * max_size), GFP_KERNEL);
521 if (result->skbuff_vector == NULL)
522 goto out_skb_fail;
523
524 /* further failures can be handled safely by destroy_queue*/
525
526 mmsg_vector = result->mmsg_vector;
527 for (i = 0; i < max_size; i++) {
528 /* Clear all pointers - we use non-NULL as marking on
529 * what to free on destruction
530 */
531 *(result->skbuff_vector + i) = NULL;
532 mmsg_vector->msg_hdr.msg_iov = NULL;
533 mmsg_vector++;
534 }
535 mmsg_vector = result->mmsg_vector;
536 result->max_iov_frags = num_extra_frags;
537 for (i = 0; i < max_size; i++) {
538 if (vp->header_size > 0)
539 iov = kmalloc_array(3 + num_extra_frags,
540 sizeof(struct iovec),
541 GFP_KERNEL
542 );
543 else
544 iov = kmalloc_array(2 + num_extra_frags,
545 sizeof(struct iovec),
546 GFP_KERNEL
547 );
548 if (iov == NULL)
549 goto out_fail;
550 mmsg_vector->msg_hdr.msg_iov = iov;
551 mmsg_vector->msg_hdr.msg_iovlen = 1;
552 mmsg_vector->msg_hdr.msg_control = NULL;
553 mmsg_vector->msg_hdr.msg_controllen = 0;
554 mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
555 mmsg_vector->msg_hdr.msg_name = NULL;
556 mmsg_vector->msg_hdr.msg_namelen = 0;
557 if (vp->header_size > 0) {
558 iov->iov_base = kmalloc(header_size, GFP_KERNEL);
559 if (iov->iov_base == NULL)
560 goto out_fail;
561 iov->iov_len = header_size;
562 mmsg_vector->msg_hdr.msg_iovlen = 2;
563 iov++;
564 }
565 iov->iov_base = NULL;
566 iov->iov_len = 0;
567 mmsg_vector++;
568 }
569 spin_lock_init(&result->head_lock);
570 spin_lock_init(&result->tail_lock);
571 result->queue_depth = 0;
572 result->head = 0;
573 result->tail = 0;
574 return result;
575out_skb_fail:
576 kfree(result->mmsg_vector);
577out_mmsg_fail:
578 kfree(result);
579 return NULL;
580out_fail:
581 destroy_queue(result);
582 return NULL;
583}
584
585/*
586 * We do not use the RX queue as a proper wraparound queue for now
587 * This is not necessary because the consumption via netif_rx()
588 * happens in-line. While we can try using the return code of
589 * netif_rx() for flow control there are no drivers doing this today.
590 * For this RX specific use we ignore the tail/head locks and
591 * just read into a prepared queue filled with skbuffs.
592 */
593
594static struct sk_buff *prep_skb(
595 struct vector_private *vp,
596 struct user_msghdr *msg)
597{
598 int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
599 struct sk_buff *result;
600 int iov_index = 0, len;
601 struct iovec *iov = msg->msg_iov;
602 int err, nr_frags, frag;
603 skb_frag_t *skb_frag;
604
605 if (vp->req_size <= linear)
606 len = linear;
607 else
608 len = vp->req_size;
609 result = alloc_skb_with_frags(
610 linear,
611 len - vp->max_packet,
612 3,
613 &err,
614 GFP_ATOMIC
615 );
616 if (vp->header_size > 0)
617 iov_index++;
618 if (result == NULL) {
619 iov[iov_index].iov_base = NULL;
620 iov[iov_index].iov_len = 0;
621 goto done;
622 }
623 skb_reserve(result, vp->headroom);
624 result->dev = vp->dev;
625 skb_put(result, vp->max_packet);
626 result->data_len = len - vp->max_packet;
627 result->len += len - vp->max_packet;
628 skb_reset_mac_header(result);
629 result->ip_summed = CHECKSUM_NONE;
630 iov[iov_index].iov_base = result->data;
631 iov[iov_index].iov_len = vp->max_packet;
632 iov_index++;
633
634 nr_frags = skb_shinfo(result)->nr_frags;
635 for (frag = 0; frag < nr_frags; frag++) {
636 skb_frag = &skb_shinfo(result)->frags[frag];
637 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
638 if (iov[iov_index].iov_base != NULL)
639 iov[iov_index].iov_len = skb_frag_size(skb_frag);
640 else
641 iov[iov_index].iov_len = 0;
642 iov_index++;
643 }
644done:
645 msg->msg_iovlen = iov_index;
646 return result;
647}
648
649
650/* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs*/
651
652static void prep_queue_for_rx(struct vector_queue *qi)
653{
654 struct vector_private *vp = netdev_priv(qi->dev);
655 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
656 void **skbuff_vector = qi->skbuff_vector;
657 int i;
658
659 if (qi->queue_depth == 0)
660 return;
661 for (i = 0; i < qi->queue_depth; i++) {
662 /* it is OK if allocation fails - recvmmsg with NULL data in
663 * iov argument still performs an RX, just drops the packet
664 * This allows us stop faffing around with a "drop buffer"
665 */
666
667 *skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
668 skbuff_vector++;
669 mmsg_vector++;
670 }
671 qi->queue_depth = 0;
672}
673
674static struct vector_device *find_device(int n)
675{
676 struct vector_device *device;
677 struct list_head *ele;
678
679 spin_lock(&vector_devices_lock);
680 list_for_each(ele, &vector_devices) {
681 device = list_entry(ele, struct vector_device, list);
682 if (device->unit == n)
683 goto out;
684 }
685 device = NULL;
686 out:
687 spin_unlock(&vector_devices_lock);
688 return device;
689}
690
691static int vector_parse(char *str, int *index_out, char **str_out,
692 char **error_out)
693{
694 int n, len, err;
695 char *start = str;
696
697 len = strlen(str);
698
699 while ((*str != ':') && (strlen(str) > 1))
700 str++;
701 if (*str != ':') {
702 *error_out = "Expected ':' after device number";
703 return -EINVAL;
704 }
705 *str = '\0';
706
707 err = kstrtouint(start, 0, &n);
708 if (err < 0) {
709 *error_out = "Bad device number";
710 return err;
711 }
712
713 str++;
714 if (find_device(n)) {
715 *error_out = "Device already configured";
716 return -EINVAL;
717 }
718
719 *index_out = n;
720 *str_out = str;
721 return 0;
722}
723
724static int vector_config(char *str, char **error_out)
725{
726 int err, n;
727 char *params;
728 struct arglist *parsed;
729
730 err = vector_parse(str, &n, ¶ms, error_out);
731 if (err != 0)
732 return err;
733
734 /* This string is broken up and the pieces used by the underlying
735 * driver. We should copy it to make sure things do not go wrong
736 * later.
737 */
738
739 params = kstrdup(params, GFP_KERNEL);
740 if (params == NULL) {
741 *error_out = "vector_config failed to strdup string";
742 return -ENOMEM;
743 }
744
745 parsed = uml_parse_vector_ifspec(params);
746
747 if (parsed == NULL) {
748 *error_out = "vector_config failed to parse parameters";
749 return -EINVAL;
750 }
751
752 vector_eth_configure(n, parsed);
753 return 0;
754}
755
756static int vector_id(char **str, int *start_out, int *end_out)
757{
758 char *end;
759 int n;
760
761 n = simple_strtoul(*str, &end, 0);
762 if ((*end != '\0') || (end == *str))
763 return -1;
764
765 *start_out = n;
766 *end_out = n;
767 *str = end;
768 return n;
769}
770
771static int vector_remove(int n, char **error_out)
772{
773 struct vector_device *vec_d;
774 struct net_device *dev;
775 struct vector_private *vp;
776
777 vec_d = find_device(n);
778 if (vec_d == NULL)
779 return -ENODEV;
780 dev = vec_d->dev;
781 vp = netdev_priv(dev);
782 if (vp->fds != NULL)
783 return -EBUSY;
784 unregister_netdev(dev);
785 platform_device_unregister(&vec_d->pdev);
786 return 0;
787}
788
789/*
790 * There is no shared per-transport initialization code, so
791 * we will just initialize each interface one by one and
792 * add them to a list
793 */
794
795static struct platform_driver uml_net_driver = {
796 .driver = {
797 .name = DRIVER_NAME,
798 },
799};
800
801
802static void vector_device_release(struct device *dev)
803{
804 struct vector_device *device = dev_get_drvdata(dev);
805 struct net_device *netdev = device->dev;
806
807 list_del(&device->list);
808 kfree(device);
809 free_netdev(netdev);
810}
811
812/* Bog standard recv using recvmsg - not used normally unless the user
813 * explicitly specifies not to use recvmmsg vector RX.
814 */
815
816static int vector_legacy_rx(struct vector_private *vp)
817{
818 int pkt_len;
819 struct user_msghdr hdr;
820 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
821 int iovpos = 0;
822 struct sk_buff *skb;
823 int header_check;
824
825 hdr.msg_name = NULL;
826 hdr.msg_namelen = 0;
827 hdr.msg_iov = (struct iovec *) &iov;
828 hdr.msg_control = NULL;
829 hdr.msg_controllen = 0;
830 hdr.msg_flags = 0;
831
832 if (vp->header_size > 0) {
833 iov[0].iov_base = vp->header_rxbuffer;
834 iov[0].iov_len = vp->header_size;
835 }
836
837 skb = prep_skb(vp, &hdr);
838
839 if (skb == NULL) {
840 /* Read a packet into drop_buffer and don't do
841 * anything with it.
842 */
843 iov[iovpos].iov_base = drop_buffer;
844 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
845 hdr.msg_iovlen = 1;
846 vp->dev->stats.rx_dropped++;
847 }
848
849 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
850 if (pkt_len < 0) {
851 vp->in_error = true;
852 return pkt_len;
853 }
854
855 if (skb != NULL) {
856 if (pkt_len > vp->header_size) {
857 if (vp->header_size > 0) {
858 header_check = vp->verify_header(
859 vp->header_rxbuffer, skb, vp);
860 if (header_check < 0) {
861 dev_kfree_skb_irq(skb);
862 vp->dev->stats.rx_dropped++;
863 vp->estats.rx_encaps_errors++;
864 return 0;
865 }
866 if (header_check > 0) {
867 vp->estats.rx_csum_offload_good++;
868 skb->ip_summed = CHECKSUM_UNNECESSARY;
869 }
870 }
871 pskb_trim(skb, pkt_len - vp->rx_header_size);
872 skb->protocol = eth_type_trans(skb, skb->dev);
873 vp->dev->stats.rx_bytes += skb->len;
874 vp->dev->stats.rx_packets++;
875 netif_rx(skb);
876 } else {
877 dev_kfree_skb_irq(skb);
878 }
879 }
880 return pkt_len;
881}
882
883/*
884 * Packet at a time TX which falls back to vector TX if the
885 * underlying transport is busy.
886 */
887
888
889
890static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
891{
892 struct iovec iov[3 + MAX_IOV_SIZE];
893 int iov_count, pkt_len = 0;
894
895 iov[0].iov_base = vp->header_txbuffer;
896 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
897
898 if (iov_count < 1)
899 goto drop;
900
901 pkt_len = uml_vector_writev(
902 vp->fds->tx_fd,
903 (struct iovec *) &iov,
904 iov_count
905 );
906
907 if (pkt_len < 0)
908 goto drop;
909
910 netif_trans_update(vp->dev);
911 netif_wake_queue(vp->dev);
912
913 if (pkt_len > 0) {
914 vp->dev->stats.tx_bytes += skb->len;
915 vp->dev->stats.tx_packets++;
916 } else {
917 vp->dev->stats.tx_dropped++;
918 }
919 consume_skb(skb);
920 return pkt_len;
921drop:
922 vp->dev->stats.tx_dropped++;
923 consume_skb(skb);
924 if (pkt_len < 0)
925 vp->in_error = true;
926 return pkt_len;
927}
928
929/*
930 * Receive as many messages as we can in one call using the special
931 * mmsg vector matched to an skb vector which we prepared earlier.
932 */
933
934static int vector_mmsg_rx(struct vector_private *vp)
935{
936 int packet_count, i;
937 struct vector_queue *qi = vp->rx_queue;
938 struct sk_buff *skb;
939 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
940 void **skbuff_vector = qi->skbuff_vector;
941 int header_check;
942
943 /* Refresh the vector and make sure it is with new skbs and the
944 * iovs are updated to point to them.
945 */
946
947 prep_queue_for_rx(qi);
948
949 /* Fire the Lazy Gun - get as many packets as we can in one go. */
950
951 packet_count = uml_vector_recvmmsg(
952 vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
953
954 if (packet_count < 0)
955 vp->in_error = true;
956
957 if (packet_count <= 0)
958 return packet_count;
959
960 /* We treat packet processing as enqueue, buffer refresh as dequeue
961 * The queue_depth tells us how many buffers have been used and how
962 * many do we need to prep the next time prep_queue_for_rx() is called.
963 */
964
965 qi->queue_depth = packet_count;
966
967 for (i = 0; i < packet_count; i++) {
968 skb = (*skbuff_vector);
969 if (mmsg_vector->msg_len > vp->header_size) {
970 if (vp->header_size > 0) {
971 header_check = vp->verify_header(
972 mmsg_vector->msg_hdr.msg_iov->iov_base,
973 skb,
974 vp
975 );
976 if (header_check < 0) {
977 /* Overlay header failed to verify - discard.
978 * We can actually keep this skb and reuse it,
979 * but that will make the prep logic too
980 * complex.
981 */
982 dev_kfree_skb_irq(skb);
983 vp->estats.rx_encaps_errors++;
984 continue;
985 }
986 if (header_check > 0) {
987 vp->estats.rx_csum_offload_good++;
988 skb->ip_summed = CHECKSUM_UNNECESSARY;
989 }
990 }
991 pskb_trim(skb,
992 mmsg_vector->msg_len - vp->rx_header_size);
993 skb->protocol = eth_type_trans(skb, skb->dev);
994 /*
995 * We do not need to lock on updating stats here
996 * The interrupt loop is non-reentrant.
997 */
998 vp->dev->stats.rx_bytes += skb->len;
999 vp->dev->stats.rx_packets++;
1000 netif_rx(skb);
1001 } else {
1002 /* Overlay header too short to do anything - discard.
1003 * We can actually keep this skb and reuse it,
1004 * but that will make the prep logic too complex.
1005 */
1006 if (skb != NULL)
1007 dev_kfree_skb_irq(skb);
1008 }
1009 (*skbuff_vector) = NULL;
1010 /* Move to the next buffer element */
1011 mmsg_vector++;
1012 skbuff_vector++;
1013 }
1014 if (packet_count > 0) {
1015 if (vp->estats.rx_queue_max < packet_count)
1016 vp->estats.rx_queue_max = packet_count;
1017 vp->estats.rx_queue_running_average =
1018 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1019 }
1020 return packet_count;
1021}
1022
1023static void vector_rx(struct vector_private *vp)
1024{
1025 int err;
1026 int iter = 0;
1027
1028 if ((vp->options & VECTOR_RX) > 0)
1029 while (((err = vector_mmsg_rx(vp)) > 0) && (iter < MAX_ITERATIONS))
1030 iter++;
1031 else
1032 while (((err = vector_legacy_rx(vp)) > 0) && (iter < MAX_ITERATIONS))
1033 iter++;
1034 if ((err != 0) && net_ratelimit())
1035 netdev_err(vp->dev, "vector_rx: error(%d)\n", err);
1036 if (iter == MAX_ITERATIONS)
1037 netdev_err(vp->dev, "vector_rx: device stuck, remote end may have closed the connection\n");
1038}
1039
1040static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1041{
1042 struct vector_private *vp = netdev_priv(dev);
1043 int queue_depth = 0;
1044
1045 if (vp->in_error) {
1046 deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1047 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1048 deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1049 return NETDEV_TX_BUSY;
1050 }
1051
1052 if ((vp->options & VECTOR_TX) == 0) {
1053 writev_tx(vp, skb);
1054 return NETDEV_TX_OK;
1055 }
1056
1057 /* We do BQL only in the vector path, no point doing it in
1058 * packet at a time mode as there is no device queue
1059 */
1060
1061 netdev_sent_queue(vp->dev, skb->len);
1062 queue_depth = vector_enqueue(vp->tx_queue, skb);
1063
1064 /* if the device queue is full, stop the upper layers and
1065 * flush it.
1066 */
1067
1068 if (queue_depth >= vp->tx_queue->max_depth - 1) {
1069 vp->estats.tx_kicks++;
1070 netif_stop_queue(dev);
1071 vector_send(vp->tx_queue);
1072 return NETDEV_TX_OK;
1073 }
1074 if (netdev_xmit_more()) {
1075 mod_timer(&vp->tl, vp->coalesce);
1076 return NETDEV_TX_OK;
1077 }
1078 if (skb->len < TX_SMALL_PACKET) {
1079 vp->estats.tx_kicks++;
1080 vector_send(vp->tx_queue);
1081 } else
1082 tasklet_schedule(&vp->tx_poll);
1083 return NETDEV_TX_OK;
1084}
1085
1086static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1087{
1088 struct net_device *dev = dev_id;
1089 struct vector_private *vp = netdev_priv(dev);
1090
1091 if (!netif_running(dev))
1092 return IRQ_NONE;
1093 vector_rx(vp);
1094 return IRQ_HANDLED;
1095
1096}
1097
1098static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1099{
1100 struct net_device *dev = dev_id;
1101 struct vector_private *vp = netdev_priv(dev);
1102
1103 if (!netif_running(dev))
1104 return IRQ_NONE;
1105 /* We need to pay attention to it only if we got
1106 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1107 * we ignore it. In the future, it may be worth
1108 * it to improve the IRQ controller a bit to make
1109 * tweaking the IRQ mask less costly
1110 */
1111
1112 if (vp->in_write_poll)
1113 tasklet_schedule(&vp->tx_poll);
1114 return IRQ_HANDLED;
1115
1116}
1117
1118static int irq_rr;
1119
1120static int vector_net_close(struct net_device *dev)
1121{
1122 struct vector_private *vp = netdev_priv(dev);
1123 unsigned long flags;
1124
1125 netif_stop_queue(dev);
1126 del_timer(&vp->tl);
1127
1128 if (vp->fds == NULL)
1129 return 0;
1130
1131 /* Disable and free all IRQS */
1132 if (vp->rx_irq > 0) {
1133 um_free_irq(vp->rx_irq, dev);
1134 vp->rx_irq = 0;
1135 }
1136 if (vp->tx_irq > 0) {
1137 um_free_irq(vp->tx_irq, dev);
1138 vp->tx_irq = 0;
1139 }
1140 tasklet_kill(&vp->tx_poll);
1141 if (vp->fds->rx_fd > 0) {
1142 os_close_file(vp->fds->rx_fd);
1143 vp->fds->rx_fd = -1;
1144 }
1145 if (vp->fds->tx_fd > 0) {
1146 os_close_file(vp->fds->tx_fd);
1147 vp->fds->tx_fd = -1;
1148 }
1149 kfree(vp->bpf);
1150 kfree(vp->fds->remote_addr);
1151 kfree(vp->transport_data);
1152 kfree(vp->header_rxbuffer);
1153 kfree(vp->header_txbuffer);
1154 if (vp->rx_queue != NULL)
1155 destroy_queue(vp->rx_queue);
1156 if (vp->tx_queue != NULL)
1157 destroy_queue(vp->tx_queue);
1158 kfree(vp->fds);
1159 vp->fds = NULL;
1160 spin_lock_irqsave(&vp->lock, flags);
1161 vp->opened = false;
1162 vp->in_error = false;
1163 spin_unlock_irqrestore(&vp->lock, flags);
1164 return 0;
1165}
1166
1167/* TX tasklet */
1168
1169static void vector_tx_poll(unsigned long data)
1170{
1171 struct vector_private *vp = (struct vector_private *)data;
1172
1173 vp->estats.tx_kicks++;
1174 vector_send(vp->tx_queue);
1175}
1176static void vector_reset_tx(struct work_struct *work)
1177{
1178 struct vector_private *vp =
1179 container_of(work, struct vector_private, reset_tx);
1180 netdev_reset_queue(vp->dev);
1181 netif_start_queue(vp->dev);
1182 netif_wake_queue(vp->dev);
1183}
1184static int vector_net_open(struct net_device *dev)
1185{
1186 struct vector_private *vp = netdev_priv(dev);
1187 unsigned long flags;
1188 int err = -EINVAL;
1189 struct vector_device *vdevice;
1190
1191 spin_lock_irqsave(&vp->lock, flags);
1192 if (vp->opened) {
1193 spin_unlock_irqrestore(&vp->lock, flags);
1194 return -ENXIO;
1195 }
1196 vp->opened = true;
1197 spin_unlock_irqrestore(&vp->lock, flags);
1198
1199 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1200
1201 if (vp->fds == NULL)
1202 goto out_close;
1203
1204 if (build_transport_data(vp) < 0)
1205 goto out_close;
1206
1207 if ((vp->options & VECTOR_RX) > 0) {
1208 vp->rx_queue = create_queue(
1209 vp,
1210 get_depth(vp->parsed),
1211 vp->rx_header_size,
1212 MAX_IOV_SIZE
1213 );
1214 vp->rx_queue->queue_depth = get_depth(vp->parsed);
1215 } else {
1216 vp->header_rxbuffer = kmalloc(
1217 vp->rx_header_size,
1218 GFP_KERNEL
1219 );
1220 if (vp->header_rxbuffer == NULL)
1221 goto out_close;
1222 }
1223 if ((vp->options & VECTOR_TX) > 0) {
1224 vp->tx_queue = create_queue(
1225 vp,
1226 get_depth(vp->parsed),
1227 vp->header_size,
1228 MAX_IOV_SIZE
1229 );
1230 } else {
1231 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1232 if (vp->header_txbuffer == NULL)
1233 goto out_close;
1234 }
1235
1236 /* READ IRQ */
1237 err = um_request_irq(
1238 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1239 IRQ_READ, vector_rx_interrupt,
1240 IRQF_SHARED, dev->name, dev);
1241 if (err != 0) {
1242 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1243 err = -ENETUNREACH;
1244 goto out_close;
1245 }
1246 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1247 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1248 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1249
1250 /* WRITE IRQ - we need it only if we have vector TX */
1251 if ((vp->options & VECTOR_TX) > 0) {
1252 err = um_request_irq(
1253 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1254 IRQ_WRITE, vector_tx_interrupt,
1255 IRQF_SHARED, dev->name, dev);
1256 if (err != 0) {
1257 netdev_err(dev,
1258 "vector_open: failed to get tx irq(%d)\n", err);
1259 err = -ENETUNREACH;
1260 goto out_close;
1261 }
1262 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1263 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1264 }
1265
1266 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1267 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1268 vp->options |= VECTOR_BPF;
1269 }
1270 if ((vp->options & VECTOR_BPF) != 0)
1271 vp->bpf = uml_vector_default_bpf(vp->fds->rx_fd, dev->dev_addr);
1272
1273 netif_start_queue(dev);
1274
1275 /* clear buffer - it can happen that the host side of the interface
1276 * is full when we get here. In this case, new data is never queued,
1277 * SIGIOs never arrive, and the net never works.
1278 */
1279
1280 vector_rx(vp);
1281
1282 vector_reset_stats(vp);
1283 vdevice = find_device(vp->unit);
1284 vdevice->opened = 1;
1285
1286 if ((vp->options & VECTOR_TX) != 0)
1287 add_timer(&vp->tl);
1288 return 0;
1289out_close:
1290 vector_net_close(dev);
1291 return err;
1292}
1293
1294
1295static void vector_net_set_multicast_list(struct net_device *dev)
1296{
1297 /* TODO: - we can do some BPF games here */
1298 return;
1299}
1300
1301static void vector_net_tx_timeout(struct net_device *dev)
1302{
1303 struct vector_private *vp = netdev_priv(dev);
1304
1305 vp->estats.tx_timeout_count++;
1306 netif_trans_update(dev);
1307 schedule_work(&vp->reset_tx);
1308}
1309
1310static netdev_features_t vector_fix_features(struct net_device *dev,
1311 netdev_features_t features)
1312{
1313 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1314 return features;
1315}
1316
1317static int vector_set_features(struct net_device *dev,
1318 netdev_features_t features)
1319{
1320 struct vector_private *vp = netdev_priv(dev);
1321 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1322 * no way to negotiate it on raw sockets, so we can change
1323 * only our side.
1324 */
1325 if (features & NETIF_F_GRO)
1326 /* All new frame buffers will be GRO-sized */
1327 vp->req_size = 65536;
1328 else
1329 /* All new frame buffers will be normal sized */
1330 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1331 return 0;
1332}
1333
1334#ifdef CONFIG_NET_POLL_CONTROLLER
1335static void vector_net_poll_controller(struct net_device *dev)
1336{
1337 disable_irq(dev->irq);
1338 vector_rx_interrupt(dev->irq, dev);
1339 enable_irq(dev->irq);
1340}
1341#endif
1342
1343static void vector_net_get_drvinfo(struct net_device *dev,
1344 struct ethtool_drvinfo *info)
1345{
1346 strlcpy(info->driver, DRIVER_NAME, sizeof(info->driver));
1347 strlcpy(info->version, DRIVER_VERSION, sizeof(info->version));
1348}
1349
1350static void vector_get_ringparam(struct net_device *netdev,
1351 struct ethtool_ringparam *ring)
1352{
1353 struct vector_private *vp = netdev_priv(netdev);
1354
1355 ring->rx_max_pending = vp->rx_queue->max_depth;
1356 ring->tx_max_pending = vp->tx_queue->max_depth;
1357 ring->rx_pending = vp->rx_queue->max_depth;
1358 ring->tx_pending = vp->tx_queue->max_depth;
1359}
1360
1361static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1362{
1363 switch (stringset) {
1364 case ETH_SS_TEST:
1365 *buf = '\0';
1366 break;
1367 case ETH_SS_STATS:
1368 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1369 break;
1370 default:
1371 WARN_ON(1);
1372 break;
1373 }
1374}
1375
1376static int vector_get_sset_count(struct net_device *dev, int sset)
1377{
1378 switch (sset) {
1379 case ETH_SS_TEST:
1380 return 0;
1381 case ETH_SS_STATS:
1382 return VECTOR_NUM_STATS;
1383 default:
1384 return -EOPNOTSUPP;
1385 }
1386}
1387
1388static void vector_get_ethtool_stats(struct net_device *dev,
1389 struct ethtool_stats *estats,
1390 u64 *tmp_stats)
1391{
1392 struct vector_private *vp = netdev_priv(dev);
1393
1394 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1395}
1396
1397static int vector_get_coalesce(struct net_device *netdev,
1398 struct ethtool_coalesce *ec)
1399{
1400 struct vector_private *vp = netdev_priv(netdev);
1401
1402 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1403 return 0;
1404}
1405
1406static int vector_set_coalesce(struct net_device *netdev,
1407 struct ethtool_coalesce *ec)
1408{
1409 struct vector_private *vp = netdev_priv(netdev);
1410
1411 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1412 if (vp->coalesce == 0)
1413 vp->coalesce = 1;
1414 return 0;
1415}
1416
1417static const struct ethtool_ops vector_net_ethtool_ops = {
1418 .get_drvinfo = vector_net_get_drvinfo,
1419 .get_link = ethtool_op_get_link,
1420 .get_ts_info = ethtool_op_get_ts_info,
1421 .get_ringparam = vector_get_ringparam,
1422 .get_strings = vector_get_strings,
1423 .get_sset_count = vector_get_sset_count,
1424 .get_ethtool_stats = vector_get_ethtool_stats,
1425 .get_coalesce = vector_get_coalesce,
1426 .set_coalesce = vector_set_coalesce,
1427};
1428
1429
1430static const struct net_device_ops vector_netdev_ops = {
1431 .ndo_open = vector_net_open,
1432 .ndo_stop = vector_net_close,
1433 .ndo_start_xmit = vector_net_start_xmit,
1434 .ndo_set_rx_mode = vector_net_set_multicast_list,
1435 .ndo_tx_timeout = vector_net_tx_timeout,
1436 .ndo_set_mac_address = eth_mac_addr,
1437 .ndo_validate_addr = eth_validate_addr,
1438 .ndo_fix_features = vector_fix_features,
1439 .ndo_set_features = vector_set_features,
1440#ifdef CONFIG_NET_POLL_CONTROLLER
1441 .ndo_poll_controller = vector_net_poll_controller,
1442#endif
1443};
1444
1445
1446static void vector_timer_expire(struct timer_list *t)
1447{
1448 struct vector_private *vp = from_timer(vp, t, tl);
1449
1450 vp->estats.tx_kicks++;
1451 vector_send(vp->tx_queue);
1452}
1453
1454static void vector_eth_configure(
1455 int n,
1456 struct arglist *def
1457 )
1458{
1459 struct vector_device *device;
1460 struct net_device *dev;
1461 struct vector_private *vp;
1462 int err;
1463
1464 device = kzalloc(sizeof(*device), GFP_KERNEL);
1465 if (device == NULL) {
1466 printk(KERN_ERR "eth_configure failed to allocate struct "
1467 "vector_device\n");
1468 return;
1469 }
1470 dev = alloc_etherdev(sizeof(struct vector_private));
1471 if (dev == NULL) {
1472 printk(KERN_ERR "eth_configure: failed to allocate struct "
1473 "net_device for vec%d\n", n);
1474 goto out_free_device;
1475 }
1476
1477 dev->mtu = get_mtu(def);
1478
1479 INIT_LIST_HEAD(&device->list);
1480 device->unit = n;
1481
1482 /* If this name ends up conflicting with an existing registered
1483 * netdevice, that is OK, register_netdev{,ice}() will notice this
1484 * and fail.
1485 */
1486 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1487 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1488 vp = netdev_priv(dev);
1489
1490 /* sysfs register */
1491 if (!driver_registered) {
1492 platform_driver_register(¨_net_driver);
1493 driver_registered = 1;
1494 }
1495 device->pdev.id = n;
1496 device->pdev.name = DRIVER_NAME;
1497 device->pdev.dev.release = vector_device_release;
1498 dev_set_drvdata(&device->pdev.dev, device);
1499 if (platform_device_register(&device->pdev))
1500 goto out_free_netdev;
1501 SET_NETDEV_DEV(dev, &device->pdev.dev);
1502
1503 device->dev = dev;
1504
1505 *vp = ((struct vector_private)
1506 {
1507 .list = LIST_HEAD_INIT(vp->list),
1508 .dev = dev,
1509 .unit = n,
1510 .options = get_transport_options(def),
1511 .rx_irq = 0,
1512 .tx_irq = 0,
1513 .parsed = def,
1514 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1515 /* TODO - we need to calculate headroom so that ip header
1516 * is 16 byte aligned all the time
1517 */
1518 .headroom = get_headroom(def),
1519 .form_header = NULL,
1520 .verify_header = NULL,
1521 .header_rxbuffer = NULL,
1522 .header_txbuffer = NULL,
1523 .header_size = 0,
1524 .rx_header_size = 0,
1525 .rexmit_scheduled = false,
1526 .opened = false,
1527 .transport_data = NULL,
1528 .in_write_poll = false,
1529 .coalesce = 2,
1530 .req_size = get_req_size(def),
1531 .in_error = false
1532 });
1533
1534 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1535 tasklet_init(&vp->tx_poll, vector_tx_poll, (unsigned long)vp);
1536 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1537
1538 timer_setup(&vp->tl, vector_timer_expire, 0);
1539 spin_lock_init(&vp->lock);
1540
1541 /* FIXME */
1542 dev->netdev_ops = &vector_netdev_ops;
1543 dev->ethtool_ops = &vector_net_ethtool_ops;
1544 dev->watchdog_timeo = (HZ >> 1);
1545 /* primary IRQ - fixme */
1546 dev->irq = 0; /* we will adjust this once opened */
1547
1548 rtnl_lock();
1549 err = register_netdevice(dev);
1550 rtnl_unlock();
1551 if (err)
1552 goto out_undo_user_init;
1553
1554 spin_lock(&vector_devices_lock);
1555 list_add(&device->list, &vector_devices);
1556 spin_unlock(&vector_devices_lock);
1557
1558 return;
1559
1560out_undo_user_init:
1561 return;
1562out_free_netdev:
1563 free_netdev(dev);
1564out_free_device:
1565 kfree(device);
1566}
1567
1568
1569
1570
1571/*
1572 * Invoked late in the init
1573 */
1574
1575static int __init vector_init(void)
1576{
1577 struct list_head *ele;
1578 struct vector_cmd_line_arg *def;
1579 struct arglist *parsed;
1580
1581 list_for_each(ele, &vec_cmd_line) {
1582 def = list_entry(ele, struct vector_cmd_line_arg, list);
1583 parsed = uml_parse_vector_ifspec(def->arguments);
1584 if (parsed != NULL)
1585 vector_eth_configure(def->unit, parsed);
1586 }
1587 return 0;
1588}
1589
1590
1591/* Invoked at initial argument parsing, only stores
1592 * arguments until a proper vector_init is called
1593 * later
1594 */
1595
1596static int __init vector_setup(char *str)
1597{
1598 char *error;
1599 int n, err;
1600 struct vector_cmd_line_arg *new;
1601
1602 err = vector_parse(str, &n, &str, &error);
1603 if (err) {
1604 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1605 str, error);
1606 return 1;
1607 }
1608 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1609 if (!new)
1610 panic("%s: Failed to allocate %zu bytes\n", __func__,
1611 sizeof(*new));
1612 INIT_LIST_HEAD(&new->list);
1613 new->unit = n;
1614 new->arguments = str;
1615 list_add_tail(&new->list, &vec_cmd_line);
1616 return 1;
1617}
1618
1619__setup("vec", vector_setup);
1620__uml_help(vector_setup,
1621"vec[0-9]+:<option>=<value>,<option>=<value>\n"
1622" Configure a vector io network device.\n\n"
1623);
1624
1625late_initcall(vector_init);
1626
1627static struct mc_device vector_mc = {
1628 .list = LIST_HEAD_INIT(vector_mc.list),
1629 .name = "vec",
1630 .config = vector_config,
1631 .get_config = NULL,
1632 .id = vector_id,
1633 .remove = vector_remove,
1634};
1635
1636#ifdef CONFIG_INET
1637static int vector_inetaddr_event(
1638 struct notifier_block *this,
1639 unsigned long event,
1640 void *ptr)
1641{
1642 return NOTIFY_DONE;
1643}
1644
1645static struct notifier_block vector_inetaddr_notifier = {
1646 .notifier_call = vector_inetaddr_event,
1647};
1648
1649static void inet_register(void)
1650{
1651 register_inetaddr_notifier(&vector_inetaddr_notifier);
1652}
1653#else
1654static inline void inet_register(void)
1655{
1656}
1657#endif
1658
1659static int vector_net_init(void)
1660{
1661 mconsole_register_dev(&vector_mc);
1662 inet_register();
1663 return 0;
1664}
1665
1666__initcall(vector_net_init);
1667
1668
1669