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 <asm/atomic.h>
26#include <uapi/linux/filter.h>
27#include <init.h>
28#include <irq_kern.h>
29#include <irq_user.h>
30#include <net_kern.h>
31#include <os.h>
32#include "mconsole_kern.h"
33#include "vector_user.h"
34#include "vector_kern.h"
35
36/*
37 * Adapted from network devices with the following major changes:
38 * All transports are static - simplifies the code significantly
39 * Multiple FDs/IRQs per device
40 * Vector IO optionally used for read/write, falling back to legacy
41 * based on configuration and/or availability
42 * Configuration is no longer positional - L2TPv3 and GRE require up to
43 * 10 parameters, passing this as positional is not fit for purpose.
44 * Only socket transports are supported
45 */
46
47
48#define DRIVER_NAME "uml-vector"
49struct vector_cmd_line_arg {
50 struct list_head list;
51 int unit;
52 char *arguments;
53};
54
55struct vector_device {
56 struct list_head list;
57 struct net_device *dev;
58 struct platform_device pdev;
59 int unit;
60 int opened;
61};
62
63static LIST_HEAD(vec_cmd_line);
64
65static DEFINE_SPINLOCK(vector_devices_lock);
66static LIST_HEAD(vector_devices);
67
68static int driver_registered;
69
70static void vector_eth_configure(int n, struct arglist *def);
71static int vector_mmsg_rx(struct vector_private *vp, int budget);
72
73/* Argument accessors to set variables (and/or set default values)
74 * mtu, buffer sizing, default headroom, etc
75 */
76
77#define DEFAULT_HEADROOM 2
78#define SAFETY_MARGIN 32
79#define DEFAULT_VECTOR_SIZE 64
80#define TX_SMALL_PACKET 128
81#define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
82
83static const struct {
84 const char string[ETH_GSTRING_LEN];
85} ethtool_stats_keys[] = {
86 { "rx_queue_max" },
87 { "rx_queue_running_average" },
88 { "tx_queue_max" },
89 { "tx_queue_running_average" },
90 { "rx_encaps_errors" },
91 { "tx_timeout_count" },
92 { "tx_restart_queue" },
93 { "tx_kicks" },
94 { "tx_flow_control_xon" },
95 { "tx_flow_control_xoff" },
96 { "rx_csum_offload_good" },
97 { "rx_csum_offload_errors"},
98 { "sg_ok"},
99 { "sg_linearized"},
100};
101
102#define VECTOR_NUM_STATS ARRAY_SIZE(ethtool_stats_keys)
103
104static void vector_reset_stats(struct vector_private *vp)
105{
106 /* We reuse the existing queue locks for stats */
107
108 /* RX stats are modified with RX head_lock held
109 * in vector_poll.
110 */
111
112 spin_lock(&vp->rx_queue->head_lock);
113 vp->estats.rx_queue_max = 0;
114 vp->estats.rx_queue_running_average = 0;
115 vp->estats.rx_encaps_errors = 0;
116 vp->estats.sg_ok = 0;
117 vp->estats.sg_linearized = 0;
118 spin_unlock(&vp->rx_queue->head_lock);
119
120 /* TX stats are modified with TX head_lock held
121 * in vector_send.
122 */
123
124 spin_lock(&vp->tx_queue->head_lock);
125 vp->estats.tx_timeout_count = 0;
126 vp->estats.tx_restart_queue = 0;
127 vp->estats.tx_kicks = 0;
128 vp->estats.tx_flow_control_xon = 0;
129 vp->estats.tx_flow_control_xoff = 0;
130 vp->estats.tx_queue_max = 0;
131 vp->estats.tx_queue_running_average = 0;
132 spin_unlock(&vp->tx_queue->head_lock);
133}
134
135static int get_mtu(struct arglist *def)
136{
137 char *mtu = uml_vector_fetch_arg(def, "mtu");
138 long result;
139
140 if (mtu != NULL) {
141 if (kstrtoul(mtu, 10, &result) == 0)
142 if ((result < (1 << 16) - 1) && (result >= 576))
143 return result;
144 }
145 return ETH_MAX_PACKET;
146}
147
148static char *get_bpf_file(struct arglist *def)
149{
150 return uml_vector_fetch_arg(def, "bpffile");
151}
152
153static bool get_bpf_flash(struct arglist *def)
154{
155 char *allow = uml_vector_fetch_arg(def, "bpfflash");
156 long result;
157
158 if (allow != NULL) {
159 if (kstrtoul(allow, 10, &result) == 0)
160 return result > 0;
161 }
162 return false;
163}
164
165static int get_depth(struct arglist *def)
166{
167 char *mtu = uml_vector_fetch_arg(def, "depth");
168 long result;
169
170 if (mtu != NULL) {
171 if (kstrtoul(mtu, 10, &result) == 0)
172 return result;
173 }
174 return DEFAULT_VECTOR_SIZE;
175}
176
177static int get_headroom(struct arglist *def)
178{
179 char *mtu = uml_vector_fetch_arg(def, "headroom");
180 long result;
181
182 if (mtu != NULL) {
183 if (kstrtoul(mtu, 10, &result) == 0)
184 return result;
185 }
186 return DEFAULT_HEADROOM;
187}
188
189static int get_req_size(struct arglist *def)
190{
191 char *gro = uml_vector_fetch_arg(def, "gro");
192 long result;
193
194 if (gro != NULL) {
195 if (kstrtoul(gro, 10, &result) == 0) {
196 if (result > 0)
197 return 65536;
198 }
199 }
200 return get_mtu(def) + ETH_HEADER_OTHER +
201 get_headroom(def) + SAFETY_MARGIN;
202}
203
204
205static int get_transport_options(struct arglist *def)
206{
207 char *transport = uml_vector_fetch_arg(def, "transport");
208 char *vector = uml_vector_fetch_arg(def, "vec");
209
210 int vec_rx = VECTOR_RX;
211 int vec_tx = VECTOR_TX;
212 long parsed;
213 int result = 0;
214
215 if (transport == NULL)
216 return -EINVAL;
217
218 if (vector != NULL) {
219 if (kstrtoul(vector, 10, &parsed) == 0) {
220 if (parsed == 0) {
221 vec_rx = 0;
222 vec_tx = 0;
223 }
224 }
225 }
226
227 if (get_bpf_flash(def))
228 result = VECTOR_BPF_FLASH;
229
230 if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
231 return result;
232 if (strncmp(transport, TRANS_HYBRID, TRANS_HYBRID_LEN) == 0)
233 return (result | vec_rx | VECTOR_BPF);
234 if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
235 return (result | vec_rx | vec_tx | VECTOR_QDISC_BYPASS);
236 return (result | vec_rx | vec_tx);
237}
238
239
240/* A mini-buffer for packet drop read
241 * All of our supported transports are datagram oriented and we always
242 * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
243 * than the packet size it still counts as full packet read and will
244 * clean the incoming stream to keep sigio/epoll happy
245 */
246
247#define DROP_BUFFER_SIZE 32
248
249static char *drop_buffer;
250
251
252/*
253 * Advance the mmsg queue head by n = advance. Resets the queue to
254 * maximum enqueue/dequeue-at-once capacity if possible. Called by
255 * dequeuers. Caller must hold the head_lock!
256 */
257
258static int vector_advancehead(struct vector_queue *qi, int advance)
259{
260 qi->head =
261 (qi->head + advance)
262 % qi->max_depth;
263
264
265 atomic_sub(advance, &qi->queue_depth);
266 return atomic_read(&qi->queue_depth);
267}
268
269/* Advance the queue tail by n = advance.
270 * This is called by enqueuers which should hold the
271 * head lock already
272 */
273
274static int vector_advancetail(struct vector_queue *qi, int advance)
275{
276 qi->tail =
277 (qi->tail + advance)
278 % qi->max_depth;
279 atomic_add(advance, &qi->queue_depth);
280 return atomic_read(&qi->queue_depth);
281}
282
283static int prep_msg(struct vector_private *vp,
284 struct sk_buff *skb,
285 struct iovec *iov)
286{
287 int iov_index = 0;
288 int nr_frags, frag;
289 skb_frag_t *skb_frag;
290
291 nr_frags = skb_shinfo(skb)->nr_frags;
292 if (nr_frags > MAX_IOV_SIZE) {
293 if (skb_linearize(skb) != 0)
294 goto drop;
295 }
296 if (vp->header_size > 0) {
297 iov[iov_index].iov_len = vp->header_size;
298 vp->form_header(iov[iov_index].iov_base, skb, vp);
299 iov_index++;
300 }
301 iov[iov_index].iov_base = skb->data;
302 if (nr_frags > 0) {
303 iov[iov_index].iov_len = skb->len - skb->data_len;
304 vp->estats.sg_ok++;
305 } else
306 iov[iov_index].iov_len = skb->len;
307 iov_index++;
308 for (frag = 0; frag < nr_frags; frag++) {
309 skb_frag = &skb_shinfo(skb)->frags[frag];
310 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
311 iov[iov_index].iov_len = skb_frag_size(skb_frag);
312 iov_index++;
313 }
314 return iov_index;
315drop:
316 return -1;
317}
318/*
319 * Generic vector enqueue with support for forming headers using transport
320 * specific callback. Allows GRE, L2TPv3, RAW and other transports
321 * to use a common enqueue procedure in vector mode
322 */
323
324static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
325{
326 struct vector_private *vp = netdev_priv(qi->dev);
327 int queue_depth;
328 int packet_len;
329 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
330 int iov_count;
331
332 spin_lock(&qi->tail_lock);
333 queue_depth = atomic_read(&qi->queue_depth);
334
335 if (skb)
336 packet_len = skb->len;
337
338 if (queue_depth < qi->max_depth) {
339
340 *(qi->skbuff_vector + qi->tail) = skb;
341 mmsg_vector += qi->tail;
342 iov_count = prep_msg(
343 vp,
344 skb,
345 mmsg_vector->msg_hdr.msg_iov
346 );
347 if (iov_count < 1)
348 goto drop;
349 mmsg_vector->msg_hdr.msg_iovlen = iov_count;
350 mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
351 mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
352 wmb(); /* Make the packet visible to the NAPI poll thread */
353 queue_depth = vector_advancetail(qi, 1);
354 } else
355 goto drop;
356 spin_unlock(&qi->tail_lock);
357 return queue_depth;
358drop:
359 qi->dev->stats.tx_dropped++;
360 if (skb != NULL) {
361 packet_len = skb->len;
362 dev_consume_skb_any(skb);
363 netdev_completed_queue(qi->dev, 1, packet_len);
364 }
365 spin_unlock(&qi->tail_lock);
366 return queue_depth;
367}
368
369static int consume_vector_skbs(struct vector_queue *qi, int count)
370{
371 struct sk_buff *skb;
372 int skb_index;
373 int bytes_compl = 0;
374
375 for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
376 skb = *(qi->skbuff_vector + skb_index);
377 /* mark as empty to ensure correct destruction if
378 * needed
379 */
380 bytes_compl += skb->len;
381 *(qi->skbuff_vector + skb_index) = NULL;
382 dev_consume_skb_any(skb);
383 }
384 qi->dev->stats.tx_bytes += bytes_compl;
385 qi->dev->stats.tx_packets += count;
386 netdev_completed_queue(qi->dev, count, bytes_compl);
387 return vector_advancehead(qi, count);
388}
389
390/*
391 * Generic vector dequeue via sendmmsg with support for forming headers
392 * using transport specific callback. Allows GRE, L2TPv3, RAW and
393 * other transports to use a common dequeue procedure in vector mode
394 */
395
396
397static int vector_send(struct vector_queue *qi)
398{
399 struct vector_private *vp = netdev_priv(qi->dev);
400 struct mmsghdr *send_from;
401 int result = 0, send_len;
402
403 if (spin_trylock(&qi->head_lock)) {
404 /* update queue_depth to current value */
405 while (atomic_read(&qi->queue_depth) > 0) {
406 /* Calculate the start of the vector */
407 send_len = atomic_read(&qi->queue_depth);
408 send_from = qi->mmsg_vector;
409 send_from += qi->head;
410 /* Adjust vector size if wraparound */
411 if (send_len + qi->head > qi->max_depth)
412 send_len = qi->max_depth - qi->head;
413 /* Try to TX as many packets as possible */
414 if (send_len > 0) {
415 result = uml_vector_sendmmsg(
416 vp->fds->tx_fd,
417 send_from,
418 send_len,
419 0
420 );
421 vp->in_write_poll =
422 (result != send_len);
423 }
424 /* For some of the sendmmsg error scenarios
425 * we may end being unsure in the TX success
426 * for all packets. It is safer to declare
427 * them all TX-ed and blame the network.
428 */
429 if (result < 0) {
430 if (net_ratelimit())
431 netdev_err(vp->dev, "sendmmsg err=%i\n",
432 result);
433 vp->in_error = true;
434 result = send_len;
435 }
436 if (result > 0) {
437 consume_vector_skbs(qi, result);
438 /* This is equivalent to an TX IRQ.
439 * Restart the upper layers to feed us
440 * more packets.
441 */
442 if (result > vp->estats.tx_queue_max)
443 vp->estats.tx_queue_max = result;
444 vp->estats.tx_queue_running_average =
445 (vp->estats.tx_queue_running_average + result) >> 1;
446 }
447 netif_wake_queue(qi->dev);
448 /* if TX is busy, break out of the send loop,
449 * poll write IRQ will reschedule xmit for us.
450 */
451 if (result != send_len) {
452 vp->estats.tx_restart_queue++;
453 break;
454 }
455 }
456 spin_unlock(&qi->head_lock);
457 }
458 return atomic_read(&qi->queue_depth);
459}
460
461/* Queue destructor. Deliberately stateless so we can use
462 * it in queue cleanup if initialization fails.
463 */
464
465static void destroy_queue(struct vector_queue *qi)
466{
467 int i;
468 struct iovec *iov;
469 struct vector_private *vp = netdev_priv(qi->dev);
470 struct mmsghdr *mmsg_vector;
471
472 if (qi == NULL)
473 return;
474 /* deallocate any skbuffs - we rely on any unused to be
475 * set to NULL.
476 */
477 if (qi->skbuff_vector != NULL) {
478 for (i = 0; i < qi->max_depth; i++) {
479 if (*(qi->skbuff_vector + i) != NULL)
480 dev_kfree_skb_any(*(qi->skbuff_vector + i));
481 }
482 kfree(qi->skbuff_vector);
483 }
484 /* deallocate matching IOV structures including header buffs */
485 if (qi->mmsg_vector != NULL) {
486 mmsg_vector = qi->mmsg_vector;
487 for (i = 0; i < qi->max_depth; i++) {
488 iov = mmsg_vector->msg_hdr.msg_iov;
489 if (iov != NULL) {
490 if ((vp->header_size > 0) &&
491 (iov->iov_base != NULL))
492 kfree(iov->iov_base);
493 kfree(iov);
494 }
495 mmsg_vector++;
496 }
497 kfree(qi->mmsg_vector);
498 }
499 kfree(qi);
500}
501
502/*
503 * Queue constructor. Create a queue with a given side.
504 */
505static struct vector_queue *create_queue(
506 struct vector_private *vp,
507 int max_size,
508 int header_size,
509 int num_extra_frags)
510{
511 struct vector_queue *result;
512 int i;
513 struct iovec *iov;
514 struct mmsghdr *mmsg_vector;
515
516 result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
517 if (result == NULL)
518 return NULL;
519 result->max_depth = max_size;
520 result->dev = vp->dev;
521 result->mmsg_vector = kmalloc(
522 (sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
523 if (result->mmsg_vector == NULL)
524 goto out_mmsg_fail;
525 result->skbuff_vector = kmalloc(
526 (sizeof(void *) * max_size), GFP_KERNEL);
527 if (result->skbuff_vector == NULL)
528 goto out_skb_fail;
529
530 /* further failures can be handled safely by destroy_queue*/
531
532 mmsg_vector = result->mmsg_vector;
533 for (i = 0; i < max_size; i++) {
534 /* Clear all pointers - we use non-NULL as marking on
535 * what to free on destruction
536 */
537 *(result->skbuff_vector + i) = NULL;
538 mmsg_vector->msg_hdr.msg_iov = NULL;
539 mmsg_vector++;
540 }
541 mmsg_vector = result->mmsg_vector;
542 result->max_iov_frags = num_extra_frags;
543 for (i = 0; i < max_size; i++) {
544 if (vp->header_size > 0)
545 iov = kmalloc_array(3 + num_extra_frags,
546 sizeof(struct iovec),
547 GFP_KERNEL
548 );
549 else
550 iov = kmalloc_array(2 + num_extra_frags,
551 sizeof(struct iovec),
552 GFP_KERNEL
553 );
554 if (iov == NULL)
555 goto out_fail;
556 mmsg_vector->msg_hdr.msg_iov = iov;
557 mmsg_vector->msg_hdr.msg_iovlen = 1;
558 mmsg_vector->msg_hdr.msg_control = NULL;
559 mmsg_vector->msg_hdr.msg_controllen = 0;
560 mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
561 mmsg_vector->msg_hdr.msg_name = NULL;
562 mmsg_vector->msg_hdr.msg_namelen = 0;
563 if (vp->header_size > 0) {
564 iov->iov_base = kmalloc(header_size, GFP_KERNEL);
565 if (iov->iov_base == NULL)
566 goto out_fail;
567 iov->iov_len = header_size;
568 mmsg_vector->msg_hdr.msg_iovlen = 2;
569 iov++;
570 }
571 iov->iov_base = NULL;
572 iov->iov_len = 0;
573 mmsg_vector++;
574 }
575 spin_lock_init(&result->head_lock);
576 spin_lock_init(&result->tail_lock);
577 atomic_set(&result->queue_depth, 0);
578 result->head = 0;
579 result->tail = 0;
580 return result;
581out_skb_fail:
582 kfree(result->mmsg_vector);
583out_mmsg_fail:
584 kfree(result);
585 return NULL;
586out_fail:
587 destroy_queue(result);
588 return NULL;
589}
590
591/*
592 * We do not use the RX queue as a proper wraparound queue for now
593 * This is not necessary because the consumption via napi_gro_receive()
594 * happens in-line. While we can try using the return code of
595 * netif_rx() for flow control there are no drivers doing this today.
596 * For this RX specific use we ignore the tail/head locks and
597 * just read into a prepared queue filled with skbuffs.
598 */
599
600static struct sk_buff *prep_skb(
601 struct vector_private *vp,
602 struct user_msghdr *msg)
603{
604 int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
605 struct sk_buff *result;
606 int iov_index = 0, len;
607 struct iovec *iov = msg->msg_iov;
608 int err, nr_frags, frag;
609 skb_frag_t *skb_frag;
610
611 if (vp->req_size <= linear)
612 len = linear;
613 else
614 len = vp->req_size;
615 result = alloc_skb_with_frags(
616 linear,
617 len - vp->max_packet,
618 3,
619 &err,
620 GFP_ATOMIC
621 );
622 if (vp->header_size > 0)
623 iov_index++;
624 if (result == NULL) {
625 iov[iov_index].iov_base = NULL;
626 iov[iov_index].iov_len = 0;
627 goto done;
628 }
629 skb_reserve(result, vp->headroom);
630 result->dev = vp->dev;
631 skb_put(result, vp->max_packet);
632 result->data_len = len - vp->max_packet;
633 result->len += len - vp->max_packet;
634 skb_reset_mac_header(result);
635 result->ip_summed = CHECKSUM_NONE;
636 iov[iov_index].iov_base = result->data;
637 iov[iov_index].iov_len = vp->max_packet;
638 iov_index++;
639
640 nr_frags = skb_shinfo(result)->nr_frags;
641 for (frag = 0; frag < nr_frags; frag++) {
642 skb_frag = &skb_shinfo(result)->frags[frag];
643 iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
644 if (iov[iov_index].iov_base != NULL)
645 iov[iov_index].iov_len = skb_frag_size(skb_frag);
646 else
647 iov[iov_index].iov_len = 0;
648 iov_index++;
649 }
650done:
651 msg->msg_iovlen = iov_index;
652 return result;
653}
654
655
656/* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs */
657
658static void prep_queue_for_rx(struct vector_queue *qi)
659{
660 struct vector_private *vp = netdev_priv(qi->dev);
661 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
662 void **skbuff_vector = qi->skbuff_vector;
663 int i, queue_depth;
664
665 queue_depth = atomic_read(&qi->queue_depth);
666
667 if (queue_depth == 0)
668 return;
669
670 /* RX is always emptied 100% during each cycle, so we do not
671 * have to do the tail wraparound math for it.
672 */
673
674 qi->head = qi->tail = 0;
675
676 for (i = 0; i < queue_depth; i++) {
677 /* it is OK if allocation fails - recvmmsg with NULL data in
678 * iov argument still performs an RX, just drops the packet
679 * This allows us stop faffing around with a "drop buffer"
680 */
681
682 *skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
683 skbuff_vector++;
684 mmsg_vector++;
685 }
686 atomic_set(&qi->queue_depth, 0);
687}
688
689static struct vector_device *find_device(int n)
690{
691 struct vector_device *device;
692 struct list_head *ele;
693
694 spin_lock(&vector_devices_lock);
695 list_for_each(ele, &vector_devices) {
696 device = list_entry(ele, struct vector_device, list);
697 if (device->unit == n)
698 goto out;
699 }
700 device = NULL;
701 out:
702 spin_unlock(&vector_devices_lock);
703 return device;
704}
705
706static int vector_parse(char *str, int *index_out, char **str_out,
707 char **error_out)
708{
709 int n, err;
710 char *start = str;
711
712 while ((*str != ':') && (strlen(str) > 1))
713 str++;
714 if (*str != ':') {
715 *error_out = "Expected ':' after device number";
716 return -EINVAL;
717 }
718 *str = '\0';
719
720 err = kstrtouint(start, 0, &n);
721 if (err < 0) {
722 *error_out = "Bad device number";
723 return err;
724 }
725
726 str++;
727 if (find_device(n)) {
728 *error_out = "Device already configured";
729 return -EINVAL;
730 }
731
732 *index_out = n;
733 *str_out = str;
734 return 0;
735}
736
737static int vector_config(char *str, char **error_out)
738{
739 int err, n;
740 char *params;
741 struct arglist *parsed;
742
743 err = vector_parse(str, &n, ¶ms, error_out);
744 if (err != 0)
745 return err;
746
747 /* This string is broken up and the pieces used by the underlying
748 * driver. We should copy it to make sure things do not go wrong
749 * later.
750 */
751
752 params = kstrdup(params, GFP_KERNEL);
753 if (params == NULL) {
754 *error_out = "vector_config failed to strdup string";
755 return -ENOMEM;
756 }
757
758 parsed = uml_parse_vector_ifspec(params);
759
760 if (parsed == NULL) {
761 *error_out = "vector_config failed to parse parameters";
762 kfree(params);
763 return -EINVAL;
764 }
765
766 vector_eth_configure(n, parsed);
767 return 0;
768}
769
770static int vector_id(char **str, int *start_out, int *end_out)
771{
772 char *end;
773 int n;
774
775 n = simple_strtoul(*str, &end, 0);
776 if ((*end != '\0') || (end == *str))
777 return -1;
778
779 *start_out = n;
780 *end_out = n;
781 *str = end;
782 return n;
783}
784
785static int vector_remove(int n, char **error_out)
786{
787 struct vector_device *vec_d;
788 struct net_device *dev;
789 struct vector_private *vp;
790
791 vec_d = find_device(n);
792 if (vec_d == NULL)
793 return -ENODEV;
794 dev = vec_d->dev;
795 vp = netdev_priv(dev);
796 if (vp->fds != NULL)
797 return -EBUSY;
798 unregister_netdev(dev);
799 platform_device_unregister(&vec_d->pdev);
800 return 0;
801}
802
803/*
804 * There is no shared per-transport initialization code, so
805 * we will just initialize each interface one by one and
806 * add them to a list
807 */
808
809static struct platform_driver uml_net_driver = {
810 .driver = {
811 .name = DRIVER_NAME,
812 },
813};
814
815
816static void vector_device_release(struct device *dev)
817{
818 struct vector_device *device =
819 container_of(dev, struct vector_device, pdev.dev);
820 struct net_device *netdev = device->dev;
821
822 list_del(&device->list);
823 kfree(device);
824 free_netdev(netdev);
825}
826
827/* Bog standard recv using recvmsg - not used normally unless the user
828 * explicitly specifies not to use recvmmsg vector RX.
829 */
830
831static int vector_legacy_rx(struct vector_private *vp)
832{
833 int pkt_len;
834 struct user_msghdr hdr;
835 struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
836 int iovpos = 0;
837 struct sk_buff *skb;
838 int header_check;
839
840 hdr.msg_name = NULL;
841 hdr.msg_namelen = 0;
842 hdr.msg_iov = (struct iovec *) &iov;
843 hdr.msg_control = NULL;
844 hdr.msg_controllen = 0;
845 hdr.msg_flags = 0;
846
847 if (vp->header_size > 0) {
848 iov[0].iov_base = vp->header_rxbuffer;
849 iov[0].iov_len = vp->header_size;
850 }
851
852 skb = prep_skb(vp, &hdr);
853
854 if (skb == NULL) {
855 /* Read a packet into drop_buffer and don't do
856 * anything with it.
857 */
858 iov[iovpos].iov_base = drop_buffer;
859 iov[iovpos].iov_len = DROP_BUFFER_SIZE;
860 hdr.msg_iovlen = 1;
861 vp->dev->stats.rx_dropped++;
862 }
863
864 pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
865 if (pkt_len < 0) {
866 vp->in_error = true;
867 return pkt_len;
868 }
869
870 if (skb != NULL) {
871 if (pkt_len > vp->header_size) {
872 if (vp->header_size > 0) {
873 header_check = vp->verify_header(
874 vp->header_rxbuffer, skb, vp);
875 if (header_check < 0) {
876 dev_kfree_skb_irq(skb);
877 vp->dev->stats.rx_dropped++;
878 vp->estats.rx_encaps_errors++;
879 return 0;
880 }
881 if (header_check > 0) {
882 vp->estats.rx_csum_offload_good++;
883 skb->ip_summed = CHECKSUM_UNNECESSARY;
884 }
885 }
886 pskb_trim(skb, pkt_len - vp->rx_header_size);
887 skb->protocol = eth_type_trans(skb, skb->dev);
888 vp->dev->stats.rx_bytes += skb->len;
889 vp->dev->stats.rx_packets++;
890 napi_gro_receive(&vp->napi, skb);
891 } else {
892 dev_kfree_skb_irq(skb);
893 }
894 }
895 return pkt_len;
896}
897
898/*
899 * Packet at a time TX which falls back to vector TX if the
900 * underlying transport is busy.
901 */
902
903
904
905static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
906{
907 struct iovec iov[3 + MAX_IOV_SIZE];
908 int iov_count, pkt_len = 0;
909
910 iov[0].iov_base = vp->header_txbuffer;
911 iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
912
913 if (iov_count < 1)
914 goto drop;
915
916 pkt_len = uml_vector_writev(
917 vp->fds->tx_fd,
918 (struct iovec *) &iov,
919 iov_count
920 );
921
922 if (pkt_len < 0)
923 goto drop;
924
925 netif_trans_update(vp->dev);
926 netif_wake_queue(vp->dev);
927
928 if (pkt_len > 0) {
929 vp->dev->stats.tx_bytes += skb->len;
930 vp->dev->stats.tx_packets++;
931 } else {
932 vp->dev->stats.tx_dropped++;
933 }
934 consume_skb(skb);
935 return pkt_len;
936drop:
937 vp->dev->stats.tx_dropped++;
938 consume_skb(skb);
939 if (pkt_len < 0)
940 vp->in_error = true;
941 return pkt_len;
942}
943
944/*
945 * Receive as many messages as we can in one call using the special
946 * mmsg vector matched to an skb vector which we prepared earlier.
947 */
948
949static int vector_mmsg_rx(struct vector_private *vp, int budget)
950{
951 int packet_count, i;
952 struct vector_queue *qi = vp->rx_queue;
953 struct sk_buff *skb;
954 struct mmsghdr *mmsg_vector = qi->mmsg_vector;
955 void **skbuff_vector = qi->skbuff_vector;
956 int header_check;
957
958 /* Refresh the vector and make sure it is with new skbs and the
959 * iovs are updated to point to them.
960 */
961
962 prep_queue_for_rx(qi);
963
964 /* Fire the Lazy Gun - get as many packets as we can in one go. */
965
966 if (budget > qi->max_depth)
967 budget = qi->max_depth;
968
969 packet_count = uml_vector_recvmmsg(
970 vp->fds->rx_fd, qi->mmsg_vector, budget, 0);
971
972 if (packet_count < 0)
973 vp->in_error = true;
974
975 if (packet_count <= 0)
976 return packet_count;
977
978 /* We treat packet processing as enqueue, buffer refresh as dequeue
979 * The queue_depth tells us how many buffers have been used and how
980 * many do we need to prep the next time prep_queue_for_rx() is called.
981 */
982
983 atomic_add(packet_count, &qi->queue_depth);
984
985 for (i = 0; i < packet_count; i++) {
986 skb = (*skbuff_vector);
987 if (mmsg_vector->msg_len > vp->header_size) {
988 if (vp->header_size > 0) {
989 header_check = vp->verify_header(
990 mmsg_vector->msg_hdr.msg_iov->iov_base,
991 skb,
992 vp
993 );
994 if (header_check < 0) {
995 /* Overlay header failed to verify - discard.
996 * We can actually keep this skb and reuse it,
997 * but that will make the prep logic too
998 * complex.
999 */
1000 dev_kfree_skb_irq(skb);
1001 vp->estats.rx_encaps_errors++;
1002 continue;
1003 }
1004 if (header_check > 0) {
1005 vp->estats.rx_csum_offload_good++;
1006 skb->ip_summed = CHECKSUM_UNNECESSARY;
1007 }
1008 }
1009 pskb_trim(skb,
1010 mmsg_vector->msg_len - vp->rx_header_size);
1011 skb->protocol = eth_type_trans(skb, skb->dev);
1012 /*
1013 * We do not need to lock on updating stats here
1014 * The interrupt loop is non-reentrant.
1015 */
1016 vp->dev->stats.rx_bytes += skb->len;
1017 vp->dev->stats.rx_packets++;
1018 napi_gro_receive(&vp->napi, skb);
1019 } else {
1020 /* Overlay header too short to do anything - discard.
1021 * We can actually keep this skb and reuse it,
1022 * but that will make the prep logic too complex.
1023 */
1024 if (skb != NULL)
1025 dev_kfree_skb_irq(skb);
1026 }
1027 (*skbuff_vector) = NULL;
1028 /* Move to the next buffer element */
1029 mmsg_vector++;
1030 skbuff_vector++;
1031 }
1032 if (packet_count > 0) {
1033 if (vp->estats.rx_queue_max < packet_count)
1034 vp->estats.rx_queue_max = packet_count;
1035 vp->estats.rx_queue_running_average =
1036 (vp->estats.rx_queue_running_average + packet_count) >> 1;
1037 }
1038 return packet_count;
1039}
1040
1041static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1042{
1043 struct vector_private *vp = netdev_priv(dev);
1044 int queue_depth = 0;
1045
1046 if (vp->in_error) {
1047 deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1048 if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1049 deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1050 return NETDEV_TX_BUSY;
1051 }
1052
1053 if ((vp->options & VECTOR_TX) == 0) {
1054 writev_tx(vp, skb);
1055 return NETDEV_TX_OK;
1056 }
1057
1058 /* We do BQL only in the vector path, no point doing it in
1059 * packet at a time mode as there is no device queue
1060 */
1061
1062 netdev_sent_queue(vp->dev, skb->len);
1063 queue_depth = vector_enqueue(vp->tx_queue, skb);
1064
1065 if (queue_depth < vp->tx_queue->max_depth && netdev_xmit_more()) {
1066 mod_timer(&vp->tl, vp->coalesce);
1067 return NETDEV_TX_OK;
1068 } else {
1069 queue_depth = vector_send(vp->tx_queue);
1070 if (queue_depth > 0)
1071 napi_schedule(&vp->napi);
1072 }
1073
1074 return NETDEV_TX_OK;
1075}
1076
1077static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1078{
1079 struct net_device *dev = dev_id;
1080 struct vector_private *vp = netdev_priv(dev);
1081
1082 if (!netif_running(dev))
1083 return IRQ_NONE;
1084 napi_schedule(&vp->napi);
1085 return IRQ_HANDLED;
1086
1087}
1088
1089static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1090{
1091 struct net_device *dev = dev_id;
1092 struct vector_private *vp = netdev_priv(dev);
1093
1094 if (!netif_running(dev))
1095 return IRQ_NONE;
1096 /* We need to pay attention to it only if we got
1097 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1098 * we ignore it. In the future, it may be worth
1099 * it to improve the IRQ controller a bit to make
1100 * tweaking the IRQ mask less costly
1101 */
1102
1103 napi_schedule(&vp->napi);
1104 return IRQ_HANDLED;
1105
1106}
1107
1108static int irq_rr;
1109
1110static int vector_net_close(struct net_device *dev)
1111{
1112 struct vector_private *vp = netdev_priv(dev);
1113
1114 netif_stop_queue(dev);
1115 del_timer(&vp->tl);
1116
1117 vp->opened = false;
1118
1119 if (vp->fds == NULL)
1120 return 0;
1121
1122 /* Disable and free all IRQS */
1123 if (vp->rx_irq > 0) {
1124 um_free_irq(vp->rx_irq, dev);
1125 vp->rx_irq = 0;
1126 }
1127 if (vp->tx_irq > 0) {
1128 um_free_irq(vp->tx_irq, dev);
1129 vp->tx_irq = 0;
1130 }
1131 napi_disable(&vp->napi);
1132 netif_napi_del(&vp->napi);
1133 if (vp->fds->rx_fd > 0) {
1134 if (vp->bpf)
1135 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1136 os_close_file(vp->fds->rx_fd);
1137 vp->fds->rx_fd = -1;
1138 }
1139 if (vp->fds->tx_fd > 0) {
1140 os_close_file(vp->fds->tx_fd);
1141 vp->fds->tx_fd = -1;
1142 }
1143 if (vp->bpf != NULL)
1144 kfree(vp->bpf->filter);
1145 kfree(vp->bpf);
1146 vp->bpf = NULL;
1147 kfree(vp->fds->remote_addr);
1148 kfree(vp->transport_data);
1149 kfree(vp->header_rxbuffer);
1150 kfree(vp->header_txbuffer);
1151 if (vp->rx_queue != NULL)
1152 destroy_queue(vp->rx_queue);
1153 if (vp->tx_queue != NULL)
1154 destroy_queue(vp->tx_queue);
1155 kfree(vp->fds);
1156 vp->fds = NULL;
1157 vp->in_error = false;
1158 return 0;
1159}
1160
1161static int vector_poll(struct napi_struct *napi, int budget)
1162{
1163 struct vector_private *vp = container_of(napi, struct vector_private, napi);
1164 int work_done = 0;
1165 int err;
1166 bool tx_enqueued = false;
1167
1168 if ((vp->options & VECTOR_TX) != 0)
1169 tx_enqueued = (vector_send(vp->tx_queue) > 0);
1170 spin_lock(&vp->rx_queue->head_lock);
1171 if ((vp->options & VECTOR_RX) > 0)
1172 err = vector_mmsg_rx(vp, budget);
1173 else {
1174 err = vector_legacy_rx(vp);
1175 if (err > 0)
1176 err = 1;
1177 }
1178 spin_unlock(&vp->rx_queue->head_lock);
1179 if (err > 0)
1180 work_done += err;
1181
1182 if (tx_enqueued || err > 0)
1183 napi_schedule(napi);
1184 if (work_done <= budget)
1185 napi_complete_done(napi, work_done);
1186 return work_done;
1187}
1188
1189static void vector_reset_tx(struct work_struct *work)
1190{
1191 struct vector_private *vp =
1192 container_of(work, struct vector_private, reset_tx);
1193 netdev_reset_queue(vp->dev);
1194 netif_start_queue(vp->dev);
1195 netif_wake_queue(vp->dev);
1196}
1197
1198static int vector_net_open(struct net_device *dev)
1199{
1200 struct vector_private *vp = netdev_priv(dev);
1201 int err = -EINVAL;
1202 struct vector_device *vdevice;
1203
1204 if (vp->opened)
1205 return -ENXIO;
1206 vp->opened = true;
1207
1208 vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1209
1210 vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1211
1212 if (vp->fds == NULL)
1213 goto out_close;
1214
1215 if (build_transport_data(vp) < 0)
1216 goto out_close;
1217
1218 if ((vp->options & VECTOR_RX) > 0) {
1219 vp->rx_queue = create_queue(
1220 vp,
1221 get_depth(vp->parsed),
1222 vp->rx_header_size,
1223 MAX_IOV_SIZE
1224 );
1225 atomic_set(&vp->rx_queue->queue_depth, get_depth(vp->parsed));
1226 } else {
1227 vp->header_rxbuffer = kmalloc(
1228 vp->rx_header_size,
1229 GFP_KERNEL
1230 );
1231 if (vp->header_rxbuffer == NULL)
1232 goto out_close;
1233 }
1234 if ((vp->options & VECTOR_TX) > 0) {
1235 vp->tx_queue = create_queue(
1236 vp,
1237 get_depth(vp->parsed),
1238 vp->header_size,
1239 MAX_IOV_SIZE
1240 );
1241 } else {
1242 vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1243 if (vp->header_txbuffer == NULL)
1244 goto out_close;
1245 }
1246
1247 netif_napi_add_weight(vp->dev, &vp->napi, vector_poll,
1248 get_depth(vp->parsed));
1249 napi_enable(&vp->napi);
1250
1251 /* READ IRQ */
1252 err = um_request_irq(
1253 irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1254 IRQ_READ, vector_rx_interrupt,
1255 IRQF_SHARED, dev->name, dev);
1256 if (err < 0) {
1257 netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1258 err = -ENETUNREACH;
1259 goto out_close;
1260 }
1261 vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1262 dev->irq = irq_rr + VECTOR_BASE_IRQ;
1263 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1264
1265 /* WRITE IRQ - we need it only if we have vector TX */
1266 if ((vp->options & VECTOR_TX) > 0) {
1267 err = um_request_irq(
1268 irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1269 IRQ_WRITE, vector_tx_interrupt,
1270 IRQF_SHARED, dev->name, dev);
1271 if (err < 0) {
1272 netdev_err(dev,
1273 "vector_open: failed to get tx irq(%d)\n", err);
1274 err = -ENETUNREACH;
1275 goto out_close;
1276 }
1277 vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1278 irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1279 }
1280
1281 if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1282 if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1283 vp->options |= VECTOR_BPF;
1284 }
1285 if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1286 vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1287
1288 if (vp->bpf != NULL)
1289 uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1290
1291 netif_start_queue(dev);
1292 vector_reset_stats(vp);
1293
1294 /* clear buffer - it can happen that the host side of the interface
1295 * is full when we get here. In this case, new data is never queued,
1296 * SIGIOs never arrive, and the net never works.
1297 */
1298
1299 napi_schedule(&vp->napi);
1300
1301 vdevice = find_device(vp->unit);
1302 vdevice->opened = 1;
1303
1304 if ((vp->options & VECTOR_TX) != 0)
1305 add_timer(&vp->tl);
1306 return 0;
1307out_close:
1308 vector_net_close(dev);
1309 return err;
1310}
1311
1312
1313static void vector_net_set_multicast_list(struct net_device *dev)
1314{
1315 /* TODO: - we can do some BPF games here */
1316 return;
1317}
1318
1319static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1320{
1321 struct vector_private *vp = netdev_priv(dev);
1322
1323 vp->estats.tx_timeout_count++;
1324 netif_trans_update(dev);
1325 schedule_work(&vp->reset_tx);
1326}
1327
1328static netdev_features_t vector_fix_features(struct net_device *dev,
1329 netdev_features_t features)
1330{
1331 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1332 return features;
1333}
1334
1335static int vector_set_features(struct net_device *dev,
1336 netdev_features_t features)
1337{
1338 struct vector_private *vp = netdev_priv(dev);
1339 /* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1340 * no way to negotiate it on raw sockets, so we can change
1341 * only our side.
1342 */
1343 if (features & NETIF_F_GRO)
1344 /* All new frame buffers will be GRO-sized */
1345 vp->req_size = 65536;
1346 else
1347 /* All new frame buffers will be normal sized */
1348 vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1349 return 0;
1350}
1351
1352#ifdef CONFIG_NET_POLL_CONTROLLER
1353static void vector_net_poll_controller(struct net_device *dev)
1354{
1355 disable_irq(dev->irq);
1356 vector_rx_interrupt(dev->irq, dev);
1357 enable_irq(dev->irq);
1358}
1359#endif
1360
1361static void vector_net_get_drvinfo(struct net_device *dev,
1362 struct ethtool_drvinfo *info)
1363{
1364 strscpy(info->driver, DRIVER_NAME);
1365}
1366
1367static int vector_net_load_bpf_flash(struct net_device *dev,
1368 struct ethtool_flash *efl)
1369{
1370 struct vector_private *vp = netdev_priv(dev);
1371 struct vector_device *vdevice;
1372 const struct firmware *fw;
1373 int result = 0;
1374
1375 if (!(vp->options & VECTOR_BPF_FLASH)) {
1376 netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1377 return -1;
1378 }
1379
1380 if (vp->bpf != NULL) {
1381 if (vp->opened)
1382 uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1383 kfree(vp->bpf->filter);
1384 vp->bpf->filter = NULL;
1385 } else {
1386 vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC);
1387 if (vp->bpf == NULL) {
1388 netdev_err(dev, "failed to allocate memory for firmware\n");
1389 goto flash_fail;
1390 }
1391 }
1392
1393 vdevice = find_device(vp->unit);
1394
1395 if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1396 goto flash_fail;
1397
1398 vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC);
1399 if (!vp->bpf->filter)
1400 goto free_buffer;
1401
1402 vp->bpf->len = fw->size / sizeof(struct sock_filter);
1403 release_firmware(fw);
1404
1405 if (vp->opened)
1406 result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1407
1408 return result;
1409
1410free_buffer:
1411 release_firmware(fw);
1412
1413flash_fail:
1414 if (vp->bpf != NULL)
1415 kfree(vp->bpf->filter);
1416 kfree(vp->bpf);
1417 vp->bpf = NULL;
1418 return -1;
1419}
1420
1421static void vector_get_ringparam(struct net_device *netdev,
1422 struct ethtool_ringparam *ring,
1423 struct kernel_ethtool_ringparam *kernel_ring,
1424 struct netlink_ext_ack *extack)
1425{
1426 struct vector_private *vp = netdev_priv(netdev);
1427
1428 ring->rx_max_pending = vp->rx_queue->max_depth;
1429 ring->tx_max_pending = vp->tx_queue->max_depth;
1430 ring->rx_pending = vp->rx_queue->max_depth;
1431 ring->tx_pending = vp->tx_queue->max_depth;
1432}
1433
1434static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1435{
1436 switch (stringset) {
1437 case ETH_SS_TEST:
1438 *buf = '\0';
1439 break;
1440 case ETH_SS_STATS:
1441 memcpy(buf, ðtool_stats_keys, sizeof(ethtool_stats_keys));
1442 break;
1443 default:
1444 WARN_ON(1);
1445 break;
1446 }
1447}
1448
1449static int vector_get_sset_count(struct net_device *dev, int sset)
1450{
1451 switch (sset) {
1452 case ETH_SS_TEST:
1453 return 0;
1454 case ETH_SS_STATS:
1455 return VECTOR_NUM_STATS;
1456 default:
1457 return -EOPNOTSUPP;
1458 }
1459}
1460
1461static void vector_get_ethtool_stats(struct net_device *dev,
1462 struct ethtool_stats *estats,
1463 u64 *tmp_stats)
1464{
1465 struct vector_private *vp = netdev_priv(dev);
1466
1467 /* Stats are modified in the dequeue portions of
1468 * rx/tx which are protected by the head locks
1469 * grabbing these locks here ensures they are up
1470 * to date.
1471 */
1472
1473 spin_lock(&vp->tx_queue->head_lock);
1474 spin_lock(&vp->rx_queue->head_lock);
1475 memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1476 spin_unlock(&vp->rx_queue->head_lock);
1477 spin_unlock(&vp->tx_queue->head_lock);
1478}
1479
1480static int vector_get_coalesce(struct net_device *netdev,
1481 struct ethtool_coalesce *ec,
1482 struct kernel_ethtool_coalesce *kernel_coal,
1483 struct netlink_ext_ack *extack)
1484{
1485 struct vector_private *vp = netdev_priv(netdev);
1486
1487 ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1488 return 0;
1489}
1490
1491static int vector_set_coalesce(struct net_device *netdev,
1492 struct ethtool_coalesce *ec,
1493 struct kernel_ethtool_coalesce *kernel_coal,
1494 struct netlink_ext_ack *extack)
1495{
1496 struct vector_private *vp = netdev_priv(netdev);
1497
1498 vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1499 if (vp->coalesce == 0)
1500 vp->coalesce = 1;
1501 return 0;
1502}
1503
1504static const struct ethtool_ops vector_net_ethtool_ops = {
1505 .supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1506 .get_drvinfo = vector_net_get_drvinfo,
1507 .get_link = ethtool_op_get_link,
1508 .get_ts_info = ethtool_op_get_ts_info,
1509 .get_ringparam = vector_get_ringparam,
1510 .get_strings = vector_get_strings,
1511 .get_sset_count = vector_get_sset_count,
1512 .get_ethtool_stats = vector_get_ethtool_stats,
1513 .get_coalesce = vector_get_coalesce,
1514 .set_coalesce = vector_set_coalesce,
1515 .flash_device = vector_net_load_bpf_flash,
1516};
1517
1518
1519static const struct net_device_ops vector_netdev_ops = {
1520 .ndo_open = vector_net_open,
1521 .ndo_stop = vector_net_close,
1522 .ndo_start_xmit = vector_net_start_xmit,
1523 .ndo_set_rx_mode = vector_net_set_multicast_list,
1524 .ndo_tx_timeout = vector_net_tx_timeout,
1525 .ndo_set_mac_address = eth_mac_addr,
1526 .ndo_validate_addr = eth_validate_addr,
1527 .ndo_fix_features = vector_fix_features,
1528 .ndo_set_features = vector_set_features,
1529#ifdef CONFIG_NET_POLL_CONTROLLER
1530 .ndo_poll_controller = vector_net_poll_controller,
1531#endif
1532};
1533
1534static void vector_timer_expire(struct timer_list *t)
1535{
1536 struct vector_private *vp = from_timer(vp, t, tl);
1537
1538 vp->estats.tx_kicks++;
1539 napi_schedule(&vp->napi);
1540}
1541
1542
1543
1544static void vector_eth_configure(
1545 int n,
1546 struct arglist *def
1547 )
1548{
1549 struct vector_device *device;
1550 struct net_device *dev;
1551 struct vector_private *vp;
1552 int err;
1553
1554 device = kzalloc(sizeof(*device), GFP_KERNEL);
1555 if (device == NULL) {
1556 printk(KERN_ERR "eth_configure failed to allocate struct "
1557 "vector_device\n");
1558 return;
1559 }
1560 dev = alloc_etherdev(sizeof(struct vector_private));
1561 if (dev == NULL) {
1562 printk(KERN_ERR "eth_configure: failed to allocate struct "
1563 "net_device for vec%d\n", n);
1564 goto out_free_device;
1565 }
1566
1567 dev->mtu = get_mtu(def);
1568
1569 INIT_LIST_HEAD(&device->list);
1570 device->unit = n;
1571
1572 /* If this name ends up conflicting with an existing registered
1573 * netdevice, that is OK, register_netdev{,ice}() will notice this
1574 * and fail.
1575 */
1576 snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1577 uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1578 vp = netdev_priv(dev);
1579
1580 /* sysfs register */
1581 if (!driver_registered) {
1582 platform_driver_register(¨_net_driver);
1583 driver_registered = 1;
1584 }
1585 device->pdev.id = n;
1586 device->pdev.name = DRIVER_NAME;
1587 device->pdev.dev.release = vector_device_release;
1588 dev_set_drvdata(&device->pdev.dev, device);
1589 if (platform_device_register(&device->pdev))
1590 goto out_free_netdev;
1591 SET_NETDEV_DEV(dev, &device->pdev.dev);
1592
1593 device->dev = dev;
1594
1595 *vp = ((struct vector_private)
1596 {
1597 .list = LIST_HEAD_INIT(vp->list),
1598 .dev = dev,
1599 .unit = n,
1600 .options = get_transport_options(def),
1601 .rx_irq = 0,
1602 .tx_irq = 0,
1603 .parsed = def,
1604 .max_packet = get_mtu(def) + ETH_HEADER_OTHER,
1605 /* TODO - we need to calculate headroom so that ip header
1606 * is 16 byte aligned all the time
1607 */
1608 .headroom = get_headroom(def),
1609 .form_header = NULL,
1610 .verify_header = NULL,
1611 .header_rxbuffer = NULL,
1612 .header_txbuffer = NULL,
1613 .header_size = 0,
1614 .rx_header_size = 0,
1615 .rexmit_scheduled = false,
1616 .opened = false,
1617 .transport_data = NULL,
1618 .in_write_poll = false,
1619 .coalesce = 2,
1620 .req_size = get_req_size(def),
1621 .in_error = false,
1622 .bpf = NULL
1623 });
1624
1625 dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1626 INIT_WORK(&vp->reset_tx, vector_reset_tx);
1627
1628 timer_setup(&vp->tl, vector_timer_expire, 0);
1629
1630 /* FIXME */
1631 dev->netdev_ops = &vector_netdev_ops;
1632 dev->ethtool_ops = &vector_net_ethtool_ops;
1633 dev->watchdog_timeo = (HZ >> 1);
1634 /* primary IRQ - fixme */
1635 dev->irq = 0; /* we will adjust this once opened */
1636
1637 rtnl_lock();
1638 err = register_netdevice(dev);
1639 rtnl_unlock();
1640 if (err)
1641 goto out_undo_user_init;
1642
1643 spin_lock(&vector_devices_lock);
1644 list_add(&device->list, &vector_devices);
1645 spin_unlock(&vector_devices_lock);
1646
1647 return;
1648
1649out_undo_user_init:
1650 return;
1651out_free_netdev:
1652 free_netdev(dev);
1653out_free_device:
1654 kfree(device);
1655}
1656
1657
1658
1659
1660/*
1661 * Invoked late in the init
1662 */
1663
1664static int __init vector_init(void)
1665{
1666 struct list_head *ele;
1667 struct vector_cmd_line_arg *def;
1668 struct arglist *parsed;
1669
1670 list_for_each(ele, &vec_cmd_line) {
1671 def = list_entry(ele, struct vector_cmd_line_arg, list);
1672 parsed = uml_parse_vector_ifspec(def->arguments);
1673 if (parsed != NULL)
1674 vector_eth_configure(def->unit, parsed);
1675 }
1676 return 0;
1677}
1678
1679
1680/* Invoked at initial argument parsing, only stores
1681 * arguments until a proper vector_init is called
1682 * later
1683 */
1684
1685static int __init vector_setup(char *str)
1686{
1687 char *error;
1688 int n, err;
1689 struct vector_cmd_line_arg *new;
1690
1691 err = vector_parse(str, &n, &str, &error);
1692 if (err) {
1693 printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1694 str, error);
1695 return 1;
1696 }
1697 new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1698 if (!new)
1699 panic("%s: Failed to allocate %zu bytes\n", __func__,
1700 sizeof(*new));
1701 INIT_LIST_HEAD(&new->list);
1702 new->unit = n;
1703 new->arguments = str;
1704 list_add_tail(&new->list, &vec_cmd_line);
1705 return 1;
1706}
1707
1708__setup("vec", vector_setup);
1709__uml_help(vector_setup,
1710"vec[0-9]+:<option>=<value>,<option>=<value>\n"
1711" Configure a vector io network device.\n\n"
1712);
1713
1714late_initcall(vector_init);
1715
1716static struct mc_device vector_mc = {
1717 .list = LIST_HEAD_INIT(vector_mc.list),
1718 .name = "vec",
1719 .config = vector_config,
1720 .get_config = NULL,
1721 .id = vector_id,
1722 .remove = vector_remove,
1723};
1724
1725#ifdef CONFIG_INET
1726static int vector_inetaddr_event(
1727 struct notifier_block *this,
1728 unsigned long event,
1729 void *ptr)
1730{
1731 return NOTIFY_DONE;
1732}
1733
1734static struct notifier_block vector_inetaddr_notifier = {
1735 .notifier_call = vector_inetaddr_event,
1736};
1737
1738static void inet_register(void)
1739{
1740 register_inetaddr_notifier(&vector_inetaddr_notifier);
1741}
1742#else
1743static inline void inet_register(void)
1744{
1745}
1746#endif
1747
1748static int vector_net_init(void)
1749{
1750 mconsole_register_dev(&vector_mc);
1751 inet_register();
1752 return 0;
1753}
1754
1755__initcall(vector_net_init);
1756
1757
1758