xref: /openbmc/linux/arch/um/drivers/vector_kern.c (revision 92ce7359)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2017 - 2019 Cambridge Greys Limited
4  * Copyright (C) 2011 - 2014 Cisco Systems Inc
5  * Copyright (C) 2001 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
6  * Copyright (C) 2001 Lennert Buytenhek (buytenh@gnu.org) and
7  * James Leu (jleu@mindspring.net).
8  * Copyright (C) 2001 by various other people who didn't put their name here.
9  */
10 
11 #include <linux/memblock.h>
12 #include <linux/etherdevice.h>
13 #include <linux/ethtool.h>
14 #include <linux/inetdevice.h>
15 #include <linux/init.h>
16 #include <linux/list.h>
17 #include <linux/netdevice.h>
18 #include <linux/platform_device.h>
19 #include <linux/rtnetlink.h>
20 #include <linux/skbuff.h>
21 #include <linux/slab.h>
22 #include <linux/interrupt.h>
23 #include <linux/firmware.h>
24 #include <linux/fs.h>
25 #include <uapi/linux/filter.h>
26 #include <init.h>
27 #include <irq_kern.h>
28 #include <irq_user.h>
29 #include <net_kern.h>
30 #include <os.h>
31 #include "mconsole_kern.h"
32 #include "vector_user.h"
33 #include "vector_kern.h"
34 
35 /*
36  * Adapted from network devices with the following major changes:
37  * All transports are static - simplifies the code significantly
38  * Multiple FDs/IRQs per device
39  * Vector IO optionally used for read/write, falling back to legacy
40  * based on configuration and/or availability
41  * Configuration is no longer positional - L2TPv3 and GRE require up to
42  * 10 parameters, passing this as positional is not fit for purpose.
43  * Only socket transports are supported
44  */
45 
46 
47 #define DRIVER_NAME "uml-vector"
48 struct vector_cmd_line_arg {
49 	struct list_head list;
50 	int unit;
51 	char *arguments;
52 };
53 
54 struct vector_device {
55 	struct list_head list;
56 	struct net_device *dev;
57 	struct platform_device pdev;
58 	int unit;
59 	int opened;
60 };
61 
62 static LIST_HEAD(vec_cmd_line);
63 
64 static DEFINE_SPINLOCK(vector_devices_lock);
65 static LIST_HEAD(vector_devices);
66 
67 static int driver_registered;
68 
69 static void vector_eth_configure(int n, struct arglist *def);
70 static int vector_mmsg_rx(struct vector_private *vp, int budget);
71 
72 /* Argument accessors to set variables (and/or set default values)
73  * mtu, buffer sizing, default headroom, etc
74  */
75 
76 #define DEFAULT_HEADROOM 2
77 #define SAFETY_MARGIN 32
78 #define DEFAULT_VECTOR_SIZE 64
79 #define TX_SMALL_PACKET 128
80 #define MAX_IOV_SIZE (MAX_SKB_FRAGS + 1)
81 
82 static const struct {
83 	const char string[ETH_GSTRING_LEN];
84 } ethtool_stats_keys[] = {
85 	{ "rx_queue_max" },
86 	{ "rx_queue_running_average" },
87 	{ "tx_queue_max" },
88 	{ "tx_queue_running_average" },
89 	{ "rx_encaps_errors" },
90 	{ "tx_timeout_count" },
91 	{ "tx_restart_queue" },
92 	{ "tx_kicks" },
93 	{ "tx_flow_control_xon" },
94 	{ "tx_flow_control_xoff" },
95 	{ "rx_csum_offload_good" },
96 	{ "rx_csum_offload_errors"},
97 	{ "sg_ok"},
98 	{ "sg_linearized"},
99 };
100 
101 #define VECTOR_NUM_STATS	ARRAY_SIZE(ethtool_stats_keys)
102 
vector_reset_stats(struct vector_private * vp)103 static void vector_reset_stats(struct vector_private *vp)
104 {
105 	vp->estats.rx_queue_max = 0;
106 	vp->estats.rx_queue_running_average = 0;
107 	vp->estats.tx_queue_max = 0;
108 	vp->estats.tx_queue_running_average = 0;
109 	vp->estats.rx_encaps_errors = 0;
110 	vp->estats.tx_timeout_count = 0;
111 	vp->estats.tx_restart_queue = 0;
112 	vp->estats.tx_kicks = 0;
113 	vp->estats.tx_flow_control_xon = 0;
114 	vp->estats.tx_flow_control_xoff = 0;
115 	vp->estats.sg_ok = 0;
116 	vp->estats.sg_linearized = 0;
117 }
118 
get_mtu(struct arglist * def)119 static int get_mtu(struct arglist *def)
120 {
121 	char *mtu = uml_vector_fetch_arg(def, "mtu");
122 	long result;
123 
124 	if (mtu != NULL) {
125 		if (kstrtoul(mtu, 10, &result) == 0)
126 			if ((result < (1 << 16) - 1) && (result >= 576))
127 				return result;
128 	}
129 	return ETH_MAX_PACKET;
130 }
131 
get_bpf_file(struct arglist * def)132 static char *get_bpf_file(struct arglist *def)
133 {
134 	return uml_vector_fetch_arg(def, "bpffile");
135 }
136 
get_bpf_flash(struct arglist * def)137 static bool get_bpf_flash(struct arglist *def)
138 {
139 	char *allow = uml_vector_fetch_arg(def, "bpfflash");
140 	long result;
141 
142 	if (allow != NULL) {
143 		if (kstrtoul(allow, 10, &result) == 0)
144 			return result > 0;
145 	}
146 	return false;
147 }
148 
get_depth(struct arglist * def)149 static int get_depth(struct arglist *def)
150 {
151 	char *mtu = uml_vector_fetch_arg(def, "depth");
152 	long result;
153 
154 	if (mtu != NULL) {
155 		if (kstrtoul(mtu, 10, &result) == 0)
156 			return result;
157 	}
158 	return DEFAULT_VECTOR_SIZE;
159 }
160 
get_headroom(struct arglist * def)161 static int get_headroom(struct arglist *def)
162 {
163 	char *mtu = uml_vector_fetch_arg(def, "headroom");
164 	long result;
165 
166 	if (mtu != NULL) {
167 		if (kstrtoul(mtu, 10, &result) == 0)
168 			return result;
169 	}
170 	return DEFAULT_HEADROOM;
171 }
172 
get_req_size(struct arglist * def)173 static int get_req_size(struct arglist *def)
174 {
175 	char *gro = uml_vector_fetch_arg(def, "gro");
176 	long result;
177 
178 	if (gro != NULL) {
179 		if (kstrtoul(gro, 10, &result) == 0) {
180 			if (result > 0)
181 				return 65536;
182 		}
183 	}
184 	return get_mtu(def) + ETH_HEADER_OTHER +
185 		get_headroom(def) + SAFETY_MARGIN;
186 }
187 
188 
get_transport_options(struct arglist * def)189 static int get_transport_options(struct arglist *def)
190 {
191 	char *transport = uml_vector_fetch_arg(def, "transport");
192 	char *vector = uml_vector_fetch_arg(def, "vec");
193 
194 	int vec_rx = VECTOR_RX;
195 	int vec_tx = VECTOR_TX;
196 	long parsed;
197 	int result = 0;
198 
199 	if (transport == NULL)
200 		return -EINVAL;
201 
202 	if (vector != NULL) {
203 		if (kstrtoul(vector, 10, &parsed) == 0) {
204 			if (parsed == 0) {
205 				vec_rx = 0;
206 				vec_tx = 0;
207 			}
208 		}
209 	}
210 
211 	if (get_bpf_flash(def))
212 		result = VECTOR_BPF_FLASH;
213 
214 	if (strncmp(transport, TRANS_TAP, TRANS_TAP_LEN) == 0)
215 		return result;
216 	if (strncmp(transport, TRANS_HYBRID, TRANS_HYBRID_LEN) == 0)
217 		return (result | vec_rx | VECTOR_BPF);
218 	if (strncmp(transport, TRANS_RAW, TRANS_RAW_LEN) == 0)
219 		return (result | vec_rx | vec_tx | VECTOR_QDISC_BYPASS);
220 	return (result | vec_rx | vec_tx);
221 }
222 
223 
224 /* A mini-buffer for packet drop read
225  * All of our supported transports are datagram oriented and we always
226  * read using recvmsg or recvmmsg. If we pass a buffer which is smaller
227  * than the packet size it still counts as full packet read and will
228  * clean the incoming stream to keep sigio/epoll happy
229  */
230 
231 #define DROP_BUFFER_SIZE 32
232 
233 static char *drop_buffer;
234 
235 /* Array backed queues optimized for bulk enqueue/dequeue and
236  * 1:N (small values of N) or 1:1 enqueuer/dequeuer ratios.
237  * For more details and full design rationale see
238  * http://foswiki.cambridgegreys.com/Main/EatYourTailAndEnjoyIt
239  */
240 
241 
242 /*
243  * Advance the mmsg queue head by n = advance. Resets the queue to
244  * maximum enqueue/dequeue-at-once capacity if possible. Called by
245  * dequeuers. Caller must hold the head_lock!
246  */
247 
vector_advancehead(struct vector_queue * qi,int advance)248 static int vector_advancehead(struct vector_queue *qi, int advance)
249 {
250 	int queue_depth;
251 
252 	qi->head =
253 		(qi->head + advance)
254 			% qi->max_depth;
255 
256 
257 	spin_lock(&qi->tail_lock);
258 	qi->queue_depth -= advance;
259 
260 	/* we are at 0, use this to
261 	 * reset head and tail so we can use max size vectors
262 	 */
263 
264 	if (qi->queue_depth == 0) {
265 		qi->head = 0;
266 		qi->tail = 0;
267 	}
268 	queue_depth = qi->queue_depth;
269 	spin_unlock(&qi->tail_lock);
270 	return queue_depth;
271 }
272 
273 /*	Advance the queue tail by n = advance.
274  *	This is called by enqueuers which should hold the
275  *	head lock already
276  */
277 
vector_advancetail(struct vector_queue * qi,int advance)278 static int vector_advancetail(struct vector_queue *qi, int advance)
279 {
280 	int queue_depth;
281 
282 	qi->tail =
283 		(qi->tail + advance)
284 			% qi->max_depth;
285 	spin_lock(&qi->head_lock);
286 	qi->queue_depth += advance;
287 	queue_depth = qi->queue_depth;
288 	spin_unlock(&qi->head_lock);
289 	return queue_depth;
290 }
291 
prep_msg(struct vector_private * vp,struct sk_buff * skb,struct iovec * iov)292 static int prep_msg(struct vector_private *vp,
293 	struct sk_buff *skb,
294 	struct iovec *iov)
295 {
296 	int iov_index = 0;
297 	int nr_frags, frag;
298 	skb_frag_t *skb_frag;
299 
300 	nr_frags = skb_shinfo(skb)->nr_frags;
301 	if (nr_frags > MAX_IOV_SIZE) {
302 		if (skb_linearize(skb) != 0)
303 			goto drop;
304 	}
305 	if (vp->header_size > 0) {
306 		iov[iov_index].iov_len = vp->header_size;
307 		vp->form_header(iov[iov_index].iov_base, skb, vp);
308 		iov_index++;
309 	}
310 	iov[iov_index].iov_base = skb->data;
311 	if (nr_frags > 0) {
312 		iov[iov_index].iov_len = skb->len - skb->data_len;
313 		vp->estats.sg_ok++;
314 	} else
315 		iov[iov_index].iov_len = skb->len;
316 	iov_index++;
317 	for (frag = 0; frag < nr_frags; frag++) {
318 		skb_frag = &skb_shinfo(skb)->frags[frag];
319 		iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
320 		iov[iov_index].iov_len = skb_frag_size(skb_frag);
321 		iov_index++;
322 	}
323 	return iov_index;
324 drop:
325 	return -1;
326 }
327 /*
328  * Generic vector enqueue with support for forming headers using transport
329  * specific callback. Allows GRE, L2TPv3, RAW and other transports
330  * to use a common enqueue procedure in vector mode
331  */
332 
vector_enqueue(struct vector_queue * qi,struct sk_buff * skb)333 static int vector_enqueue(struct vector_queue *qi, struct sk_buff *skb)
334 {
335 	struct vector_private *vp = netdev_priv(qi->dev);
336 	int queue_depth;
337 	int packet_len;
338 	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
339 	int iov_count;
340 
341 	spin_lock(&qi->tail_lock);
342 	spin_lock(&qi->head_lock);
343 	queue_depth = qi->queue_depth;
344 	spin_unlock(&qi->head_lock);
345 
346 	if (skb)
347 		packet_len = skb->len;
348 
349 	if (queue_depth < qi->max_depth) {
350 
351 		*(qi->skbuff_vector + qi->tail) = skb;
352 		mmsg_vector += qi->tail;
353 		iov_count = prep_msg(
354 			vp,
355 			skb,
356 			mmsg_vector->msg_hdr.msg_iov
357 		);
358 		if (iov_count < 1)
359 			goto drop;
360 		mmsg_vector->msg_hdr.msg_iovlen = iov_count;
361 		mmsg_vector->msg_hdr.msg_name = vp->fds->remote_addr;
362 		mmsg_vector->msg_hdr.msg_namelen = vp->fds->remote_addr_size;
363 		queue_depth = vector_advancetail(qi, 1);
364 	} else
365 		goto drop;
366 	spin_unlock(&qi->tail_lock);
367 	return queue_depth;
368 drop:
369 	qi->dev->stats.tx_dropped++;
370 	if (skb != NULL) {
371 		packet_len = skb->len;
372 		dev_consume_skb_any(skb);
373 		netdev_completed_queue(qi->dev, 1, packet_len);
374 	}
375 	spin_unlock(&qi->tail_lock);
376 	return queue_depth;
377 }
378 
consume_vector_skbs(struct vector_queue * qi,int count)379 static int consume_vector_skbs(struct vector_queue *qi, int count)
380 {
381 	struct sk_buff *skb;
382 	int skb_index;
383 	int bytes_compl = 0;
384 
385 	for (skb_index = qi->head; skb_index < qi->head + count; skb_index++) {
386 		skb = *(qi->skbuff_vector + skb_index);
387 		/* mark as empty to ensure correct destruction if
388 		 * needed
389 		 */
390 		bytes_compl += skb->len;
391 		*(qi->skbuff_vector + skb_index) = NULL;
392 		dev_consume_skb_any(skb);
393 	}
394 	qi->dev->stats.tx_bytes += bytes_compl;
395 	qi->dev->stats.tx_packets += count;
396 	netdev_completed_queue(qi->dev, count, bytes_compl);
397 	return vector_advancehead(qi, count);
398 }
399 
400 /*
401  * Generic vector deque via sendmmsg with support for forming headers
402  * using transport specific callback. Allows GRE, L2TPv3, RAW and
403  * other transports to use a common dequeue procedure in vector mode
404  */
405 
406 
vector_send(struct vector_queue * qi)407 static int vector_send(struct vector_queue *qi)
408 {
409 	struct vector_private *vp = netdev_priv(qi->dev);
410 	struct mmsghdr *send_from;
411 	int result = 0, send_len, queue_depth = qi->max_depth;
412 
413 	if (spin_trylock(&qi->head_lock)) {
414 		if (spin_trylock(&qi->tail_lock)) {
415 			/* update queue_depth to current value */
416 			queue_depth = qi->queue_depth;
417 			spin_unlock(&qi->tail_lock);
418 			while (queue_depth > 0) {
419 				/* Calculate the start of the vector */
420 				send_len = queue_depth;
421 				send_from = qi->mmsg_vector;
422 				send_from += qi->head;
423 				/* Adjust vector size if wraparound */
424 				if (send_len + qi->head > qi->max_depth)
425 					send_len = qi->max_depth - qi->head;
426 				/* Try to TX as many packets as possible */
427 				if (send_len > 0) {
428 					result = uml_vector_sendmmsg(
429 						 vp->fds->tx_fd,
430 						 send_from,
431 						 send_len,
432 						 0
433 					);
434 					vp->in_write_poll =
435 						(result != send_len);
436 				}
437 				/* For some of the sendmmsg error scenarios
438 				 * we may end being unsure in the TX success
439 				 * for all packets. It is safer to declare
440 				 * them all TX-ed and blame the network.
441 				 */
442 				if (result < 0) {
443 					if (net_ratelimit())
444 						netdev_err(vp->dev, "sendmmsg err=%i\n",
445 							result);
446 					vp->in_error = true;
447 					result = send_len;
448 				}
449 				if (result > 0) {
450 					queue_depth =
451 						consume_vector_skbs(qi, result);
452 					/* This is equivalent to an TX IRQ.
453 					 * Restart the upper layers to feed us
454 					 * more packets.
455 					 */
456 					if (result > vp->estats.tx_queue_max)
457 						vp->estats.tx_queue_max = result;
458 					vp->estats.tx_queue_running_average =
459 						(vp->estats.tx_queue_running_average + result) >> 1;
460 				}
461 				netif_wake_queue(qi->dev);
462 				/* if TX is busy, break out of the send loop,
463 				 *  poll write IRQ will reschedule xmit for us
464 				 */
465 				if (result != send_len) {
466 					vp->estats.tx_restart_queue++;
467 					break;
468 				}
469 			}
470 		}
471 		spin_unlock(&qi->head_lock);
472 	}
473 	return queue_depth;
474 }
475 
476 /* Queue destructor. Deliberately stateless so we can use
477  * it in queue cleanup if initialization fails.
478  */
479 
destroy_queue(struct vector_queue * qi)480 static void destroy_queue(struct vector_queue *qi)
481 {
482 	int i;
483 	struct iovec *iov;
484 	struct vector_private *vp = netdev_priv(qi->dev);
485 	struct mmsghdr *mmsg_vector;
486 
487 	if (qi == NULL)
488 		return;
489 	/* deallocate any skbuffs - we rely on any unused to be
490 	 * set to NULL.
491 	 */
492 	if (qi->skbuff_vector != NULL) {
493 		for (i = 0; i < qi->max_depth; i++) {
494 			if (*(qi->skbuff_vector + i) != NULL)
495 				dev_kfree_skb_any(*(qi->skbuff_vector + i));
496 		}
497 		kfree(qi->skbuff_vector);
498 	}
499 	/* deallocate matching IOV structures including header buffs */
500 	if (qi->mmsg_vector != NULL) {
501 		mmsg_vector = qi->mmsg_vector;
502 		for (i = 0; i < qi->max_depth; i++) {
503 			iov = mmsg_vector->msg_hdr.msg_iov;
504 			if (iov != NULL) {
505 				if ((vp->header_size > 0) &&
506 					(iov->iov_base != NULL))
507 					kfree(iov->iov_base);
508 				kfree(iov);
509 			}
510 			mmsg_vector++;
511 		}
512 		kfree(qi->mmsg_vector);
513 	}
514 	kfree(qi);
515 }
516 
517 /*
518  * Queue constructor. Create a queue with a given side.
519  */
create_queue(struct vector_private * vp,int max_size,int header_size,int num_extra_frags)520 static struct vector_queue *create_queue(
521 	struct vector_private *vp,
522 	int max_size,
523 	int header_size,
524 	int num_extra_frags)
525 {
526 	struct vector_queue *result;
527 	int i;
528 	struct iovec *iov;
529 	struct mmsghdr *mmsg_vector;
530 
531 	result = kmalloc(sizeof(struct vector_queue), GFP_KERNEL);
532 	if (result == NULL)
533 		return NULL;
534 	result->max_depth = max_size;
535 	result->dev = vp->dev;
536 	result->mmsg_vector = kmalloc(
537 		(sizeof(struct mmsghdr) * max_size), GFP_KERNEL);
538 	if (result->mmsg_vector == NULL)
539 		goto out_mmsg_fail;
540 	result->skbuff_vector = kmalloc(
541 		(sizeof(void *) * max_size), GFP_KERNEL);
542 	if (result->skbuff_vector == NULL)
543 		goto out_skb_fail;
544 
545 	/* further failures can be handled safely by destroy_queue*/
546 
547 	mmsg_vector = result->mmsg_vector;
548 	for (i = 0; i < max_size; i++) {
549 		/* Clear all pointers - we use non-NULL as marking on
550 		 * what to free on destruction
551 		 */
552 		*(result->skbuff_vector + i) = NULL;
553 		mmsg_vector->msg_hdr.msg_iov = NULL;
554 		mmsg_vector++;
555 	}
556 	mmsg_vector = result->mmsg_vector;
557 	result->max_iov_frags = num_extra_frags;
558 	for (i = 0; i < max_size; i++) {
559 		if (vp->header_size > 0)
560 			iov = kmalloc_array(3 + num_extra_frags,
561 					    sizeof(struct iovec),
562 					    GFP_KERNEL
563 			);
564 		else
565 			iov = kmalloc_array(2 + num_extra_frags,
566 					    sizeof(struct iovec),
567 					    GFP_KERNEL
568 			);
569 		if (iov == NULL)
570 			goto out_fail;
571 		mmsg_vector->msg_hdr.msg_iov = iov;
572 		mmsg_vector->msg_hdr.msg_iovlen = 1;
573 		mmsg_vector->msg_hdr.msg_control = NULL;
574 		mmsg_vector->msg_hdr.msg_controllen = 0;
575 		mmsg_vector->msg_hdr.msg_flags = MSG_DONTWAIT;
576 		mmsg_vector->msg_hdr.msg_name = NULL;
577 		mmsg_vector->msg_hdr.msg_namelen = 0;
578 		if (vp->header_size > 0) {
579 			iov->iov_base = kmalloc(header_size, GFP_KERNEL);
580 			if (iov->iov_base == NULL)
581 				goto out_fail;
582 			iov->iov_len = header_size;
583 			mmsg_vector->msg_hdr.msg_iovlen = 2;
584 			iov++;
585 		}
586 		iov->iov_base = NULL;
587 		iov->iov_len = 0;
588 		mmsg_vector++;
589 	}
590 	spin_lock_init(&result->head_lock);
591 	spin_lock_init(&result->tail_lock);
592 	result->queue_depth = 0;
593 	result->head = 0;
594 	result->tail = 0;
595 	return result;
596 out_skb_fail:
597 	kfree(result->mmsg_vector);
598 out_mmsg_fail:
599 	kfree(result);
600 	return NULL;
601 out_fail:
602 	destroy_queue(result);
603 	return NULL;
604 }
605 
606 /*
607  * We do not use the RX queue as a proper wraparound queue for now
608  * This is not necessary because the consumption via napi_gro_receive()
609  * happens in-line. While we can try using the return code of
610  * netif_rx() for flow control there are no drivers doing this today.
611  * For this RX specific use we ignore the tail/head locks and
612  * just read into a prepared queue filled with skbuffs.
613  */
614 
prep_skb(struct vector_private * vp,struct user_msghdr * msg)615 static struct sk_buff *prep_skb(
616 	struct vector_private *vp,
617 	struct user_msghdr *msg)
618 {
619 	int linear = vp->max_packet + vp->headroom + SAFETY_MARGIN;
620 	struct sk_buff *result;
621 	int iov_index = 0, len;
622 	struct iovec *iov = msg->msg_iov;
623 	int err, nr_frags, frag;
624 	skb_frag_t *skb_frag;
625 
626 	if (vp->req_size <= linear)
627 		len = linear;
628 	else
629 		len = vp->req_size;
630 	result = alloc_skb_with_frags(
631 		linear,
632 		len - vp->max_packet,
633 		3,
634 		&err,
635 		GFP_ATOMIC
636 	);
637 	if (vp->header_size > 0)
638 		iov_index++;
639 	if (result == NULL) {
640 		iov[iov_index].iov_base = NULL;
641 		iov[iov_index].iov_len = 0;
642 		goto done;
643 	}
644 	skb_reserve(result, vp->headroom);
645 	result->dev = vp->dev;
646 	skb_put(result, vp->max_packet);
647 	result->data_len = len - vp->max_packet;
648 	result->len += len - vp->max_packet;
649 	skb_reset_mac_header(result);
650 	result->ip_summed = CHECKSUM_NONE;
651 	iov[iov_index].iov_base = result->data;
652 	iov[iov_index].iov_len = vp->max_packet;
653 	iov_index++;
654 
655 	nr_frags = skb_shinfo(result)->nr_frags;
656 	for (frag = 0; frag < nr_frags; frag++) {
657 		skb_frag = &skb_shinfo(result)->frags[frag];
658 		iov[iov_index].iov_base = skb_frag_address_safe(skb_frag);
659 		if (iov[iov_index].iov_base != NULL)
660 			iov[iov_index].iov_len = skb_frag_size(skb_frag);
661 		else
662 			iov[iov_index].iov_len = 0;
663 		iov_index++;
664 	}
665 done:
666 	msg->msg_iovlen = iov_index;
667 	return result;
668 }
669 
670 
671 /* Prepare queue for recvmmsg one-shot rx - fill with fresh sk_buffs*/
672 
prep_queue_for_rx(struct vector_queue * qi)673 static void prep_queue_for_rx(struct vector_queue *qi)
674 {
675 	struct vector_private *vp = netdev_priv(qi->dev);
676 	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
677 	void **skbuff_vector = qi->skbuff_vector;
678 	int i;
679 
680 	if (qi->queue_depth == 0)
681 		return;
682 	for (i = 0; i < qi->queue_depth; i++) {
683 		/* it is OK if allocation fails - recvmmsg with NULL data in
684 		 * iov argument still performs an RX, just drops the packet
685 		 * This allows us stop faffing around with a "drop buffer"
686 		 */
687 
688 		*skbuff_vector = prep_skb(vp, &mmsg_vector->msg_hdr);
689 		skbuff_vector++;
690 		mmsg_vector++;
691 	}
692 	qi->queue_depth = 0;
693 }
694 
find_device(int n)695 static struct vector_device *find_device(int n)
696 {
697 	struct vector_device *device;
698 	struct list_head *ele;
699 
700 	spin_lock(&vector_devices_lock);
701 	list_for_each(ele, &vector_devices) {
702 		device = list_entry(ele, struct vector_device, list);
703 		if (device->unit == n)
704 			goto out;
705 	}
706 	device = NULL;
707  out:
708 	spin_unlock(&vector_devices_lock);
709 	return device;
710 }
711 
vector_parse(char * str,int * index_out,char ** str_out,char ** error_out)712 static int vector_parse(char *str, int *index_out, char **str_out,
713 			char **error_out)
714 {
715 	int n, len, err;
716 	char *start = str;
717 
718 	len = strlen(str);
719 
720 	while ((*str != ':') && (strlen(str) > 1))
721 		str++;
722 	if (*str != ':') {
723 		*error_out = "Expected ':' after device number";
724 		return -EINVAL;
725 	}
726 	*str = '\0';
727 
728 	err = kstrtouint(start, 0, &n);
729 	if (err < 0) {
730 		*error_out = "Bad device number";
731 		return err;
732 	}
733 
734 	str++;
735 	if (find_device(n)) {
736 		*error_out = "Device already configured";
737 		return -EINVAL;
738 	}
739 
740 	*index_out = n;
741 	*str_out = str;
742 	return 0;
743 }
744 
vector_config(char * str,char ** error_out)745 static int vector_config(char *str, char **error_out)
746 {
747 	int err, n;
748 	char *params;
749 	struct arglist *parsed;
750 
751 	err = vector_parse(str, &n, &params, error_out);
752 	if (err != 0)
753 		return err;
754 
755 	/* This string is broken up and the pieces used by the underlying
756 	 * driver. We should copy it to make sure things do not go wrong
757 	 * later.
758 	 */
759 
760 	params = kstrdup(params, GFP_KERNEL);
761 	if (params == NULL) {
762 		*error_out = "vector_config failed to strdup string";
763 		return -ENOMEM;
764 	}
765 
766 	parsed = uml_parse_vector_ifspec(params);
767 
768 	if (parsed == NULL) {
769 		*error_out = "vector_config failed to parse parameters";
770 		kfree(params);
771 		return -EINVAL;
772 	}
773 
774 	vector_eth_configure(n, parsed);
775 	return 0;
776 }
777 
vector_id(char ** str,int * start_out,int * end_out)778 static int vector_id(char **str, int *start_out, int *end_out)
779 {
780 	char *end;
781 	int n;
782 
783 	n = simple_strtoul(*str, &end, 0);
784 	if ((*end != '\0') || (end == *str))
785 		return -1;
786 
787 	*start_out = n;
788 	*end_out = n;
789 	*str = end;
790 	return n;
791 }
792 
vector_remove(int n,char ** error_out)793 static int vector_remove(int n, char **error_out)
794 {
795 	struct vector_device *vec_d;
796 	struct net_device *dev;
797 	struct vector_private *vp;
798 
799 	vec_d = find_device(n);
800 	if (vec_d == NULL)
801 		return -ENODEV;
802 	dev = vec_d->dev;
803 	vp = netdev_priv(dev);
804 	if (vp->fds != NULL)
805 		return -EBUSY;
806 	unregister_netdev(dev);
807 	platform_device_unregister(&vec_d->pdev);
808 	return 0;
809 }
810 
811 /*
812  * There is no shared per-transport initialization code, so
813  * we will just initialize each interface one by one and
814  * add them to a list
815  */
816 
817 static struct platform_driver uml_net_driver = {
818 	.driver = {
819 		.name = DRIVER_NAME,
820 	},
821 };
822 
823 
vector_device_release(struct device * dev)824 static void vector_device_release(struct device *dev)
825 {
826 	struct vector_device *device = dev_get_drvdata(dev);
827 	struct net_device *netdev = device->dev;
828 
829 	list_del(&device->list);
830 	kfree(device);
831 	free_netdev(netdev);
832 }
833 
834 /* Bog standard recv using recvmsg - not used normally unless the user
835  * explicitly specifies not to use recvmmsg vector RX.
836  */
837 
vector_legacy_rx(struct vector_private * vp)838 static int vector_legacy_rx(struct vector_private *vp)
839 {
840 	int pkt_len;
841 	struct user_msghdr hdr;
842 	struct iovec iov[2 + MAX_IOV_SIZE]; /* header + data use case only */
843 	int iovpos = 0;
844 	struct sk_buff *skb;
845 	int header_check;
846 
847 	hdr.msg_name = NULL;
848 	hdr.msg_namelen = 0;
849 	hdr.msg_iov = (struct iovec *) &iov;
850 	hdr.msg_control = NULL;
851 	hdr.msg_controllen = 0;
852 	hdr.msg_flags = 0;
853 
854 	if (vp->header_size > 0) {
855 		iov[0].iov_base = vp->header_rxbuffer;
856 		iov[0].iov_len = vp->header_size;
857 	}
858 
859 	skb = prep_skb(vp, &hdr);
860 
861 	if (skb == NULL) {
862 		/* Read a packet into drop_buffer and don't do
863 		 * anything with it.
864 		 */
865 		iov[iovpos].iov_base = drop_buffer;
866 		iov[iovpos].iov_len = DROP_BUFFER_SIZE;
867 		hdr.msg_iovlen = 1;
868 		vp->dev->stats.rx_dropped++;
869 	}
870 
871 	pkt_len = uml_vector_recvmsg(vp->fds->rx_fd, &hdr, 0);
872 	if (pkt_len < 0) {
873 		vp->in_error = true;
874 		return pkt_len;
875 	}
876 
877 	if (skb != NULL) {
878 		if (pkt_len > vp->header_size) {
879 			if (vp->header_size > 0) {
880 				header_check = vp->verify_header(
881 					vp->header_rxbuffer, skb, vp);
882 				if (header_check < 0) {
883 					dev_kfree_skb_irq(skb);
884 					vp->dev->stats.rx_dropped++;
885 					vp->estats.rx_encaps_errors++;
886 					return 0;
887 				}
888 				if (header_check > 0) {
889 					vp->estats.rx_csum_offload_good++;
890 					skb->ip_summed = CHECKSUM_UNNECESSARY;
891 				}
892 			}
893 			pskb_trim(skb, pkt_len - vp->rx_header_size);
894 			skb->protocol = eth_type_trans(skb, skb->dev);
895 			vp->dev->stats.rx_bytes += skb->len;
896 			vp->dev->stats.rx_packets++;
897 			napi_gro_receive(&vp->napi, skb);
898 		} else {
899 			dev_kfree_skb_irq(skb);
900 		}
901 	}
902 	return pkt_len;
903 }
904 
905 /*
906  * Packet at a time TX which falls back to vector TX if the
907  * underlying transport is busy.
908  */
909 
910 
911 
writev_tx(struct vector_private * vp,struct sk_buff * skb)912 static int writev_tx(struct vector_private *vp, struct sk_buff *skb)
913 {
914 	struct iovec iov[3 + MAX_IOV_SIZE];
915 	int iov_count, pkt_len = 0;
916 
917 	iov[0].iov_base = vp->header_txbuffer;
918 	iov_count = prep_msg(vp, skb, (struct iovec *) &iov);
919 
920 	if (iov_count < 1)
921 		goto drop;
922 
923 	pkt_len = uml_vector_writev(
924 		vp->fds->tx_fd,
925 		(struct iovec *) &iov,
926 		iov_count
927 	);
928 
929 	if (pkt_len < 0)
930 		goto drop;
931 
932 	netif_trans_update(vp->dev);
933 	netif_wake_queue(vp->dev);
934 
935 	if (pkt_len > 0) {
936 		vp->dev->stats.tx_bytes += skb->len;
937 		vp->dev->stats.tx_packets++;
938 	} else {
939 		vp->dev->stats.tx_dropped++;
940 	}
941 	consume_skb(skb);
942 	return pkt_len;
943 drop:
944 	vp->dev->stats.tx_dropped++;
945 	consume_skb(skb);
946 	if (pkt_len < 0)
947 		vp->in_error = true;
948 	return pkt_len;
949 }
950 
951 /*
952  * Receive as many messages as we can in one call using the special
953  * mmsg vector matched to an skb vector which we prepared earlier.
954  */
955 
vector_mmsg_rx(struct vector_private * vp,int budget)956 static int vector_mmsg_rx(struct vector_private *vp, int budget)
957 {
958 	int packet_count, i;
959 	struct vector_queue *qi = vp->rx_queue;
960 	struct sk_buff *skb;
961 	struct mmsghdr *mmsg_vector = qi->mmsg_vector;
962 	void **skbuff_vector = qi->skbuff_vector;
963 	int header_check;
964 
965 	/* Refresh the vector and make sure it is with new skbs and the
966 	 * iovs are updated to point to them.
967 	 */
968 
969 	prep_queue_for_rx(qi);
970 
971 	/* Fire the Lazy Gun - get as many packets as we can in one go. */
972 
973 	if (budget > qi->max_depth)
974 		budget = qi->max_depth;
975 
976 	packet_count = uml_vector_recvmmsg(
977 		vp->fds->rx_fd, qi->mmsg_vector, qi->max_depth, 0);
978 
979 	if (packet_count < 0)
980 		vp->in_error = true;
981 
982 	if (packet_count <= 0)
983 		return packet_count;
984 
985 	/* We treat packet processing as enqueue, buffer refresh as dequeue
986 	 * The queue_depth tells us how many buffers have been used and how
987 	 * many do we need to prep the next time prep_queue_for_rx() is called.
988 	 */
989 
990 	qi->queue_depth = packet_count;
991 
992 	for (i = 0; i < packet_count; i++) {
993 		skb = (*skbuff_vector);
994 		if (mmsg_vector->msg_len > vp->header_size) {
995 			if (vp->header_size > 0) {
996 				header_check = vp->verify_header(
997 					mmsg_vector->msg_hdr.msg_iov->iov_base,
998 					skb,
999 					vp
1000 				);
1001 				if (header_check < 0) {
1002 				/* Overlay header failed to verify - discard.
1003 				 * We can actually keep this skb and reuse it,
1004 				 * but that will make the prep logic too
1005 				 * complex.
1006 				 */
1007 					dev_kfree_skb_irq(skb);
1008 					vp->estats.rx_encaps_errors++;
1009 					continue;
1010 				}
1011 				if (header_check > 0) {
1012 					vp->estats.rx_csum_offload_good++;
1013 					skb->ip_summed = CHECKSUM_UNNECESSARY;
1014 				}
1015 			}
1016 			pskb_trim(skb,
1017 				mmsg_vector->msg_len - vp->rx_header_size);
1018 			skb->protocol = eth_type_trans(skb, skb->dev);
1019 			/*
1020 			 * We do not need to lock on updating stats here
1021 			 * The interrupt loop is non-reentrant.
1022 			 */
1023 			vp->dev->stats.rx_bytes += skb->len;
1024 			vp->dev->stats.rx_packets++;
1025 			napi_gro_receive(&vp->napi, skb);
1026 		} else {
1027 			/* Overlay header too short to do anything - discard.
1028 			 * We can actually keep this skb and reuse it,
1029 			 * but that will make the prep logic too complex.
1030 			 */
1031 			if (skb != NULL)
1032 				dev_kfree_skb_irq(skb);
1033 		}
1034 		(*skbuff_vector) = NULL;
1035 		/* Move to the next buffer element */
1036 		mmsg_vector++;
1037 		skbuff_vector++;
1038 	}
1039 	if (packet_count > 0) {
1040 		if (vp->estats.rx_queue_max < packet_count)
1041 			vp->estats.rx_queue_max = packet_count;
1042 		vp->estats.rx_queue_running_average =
1043 			(vp->estats.rx_queue_running_average + packet_count) >> 1;
1044 	}
1045 	return packet_count;
1046 }
1047 
vector_net_start_xmit(struct sk_buff * skb,struct net_device * dev)1048 static int vector_net_start_xmit(struct sk_buff *skb, struct net_device *dev)
1049 {
1050 	struct vector_private *vp = netdev_priv(dev);
1051 	int queue_depth = 0;
1052 
1053 	if (vp->in_error) {
1054 		deactivate_fd(vp->fds->rx_fd, vp->rx_irq);
1055 		if ((vp->fds->rx_fd != vp->fds->tx_fd) && (vp->tx_irq != 0))
1056 			deactivate_fd(vp->fds->tx_fd, vp->tx_irq);
1057 		return NETDEV_TX_BUSY;
1058 	}
1059 
1060 	if ((vp->options & VECTOR_TX) == 0) {
1061 		writev_tx(vp, skb);
1062 		return NETDEV_TX_OK;
1063 	}
1064 
1065 	/* We do BQL only in the vector path, no point doing it in
1066 	 * packet at a time mode as there is no device queue
1067 	 */
1068 
1069 	netdev_sent_queue(vp->dev, skb->len);
1070 	queue_depth = vector_enqueue(vp->tx_queue, skb);
1071 
1072 	if (queue_depth < vp->tx_queue->max_depth && netdev_xmit_more()) {
1073 		mod_timer(&vp->tl, vp->coalesce);
1074 		return NETDEV_TX_OK;
1075 	} else {
1076 		queue_depth = vector_send(vp->tx_queue);
1077 		if (queue_depth > 0)
1078 			napi_schedule(&vp->napi);
1079 	}
1080 
1081 	return NETDEV_TX_OK;
1082 }
1083 
vector_rx_interrupt(int irq,void * dev_id)1084 static irqreturn_t vector_rx_interrupt(int irq, void *dev_id)
1085 {
1086 	struct net_device *dev = dev_id;
1087 	struct vector_private *vp = netdev_priv(dev);
1088 
1089 	if (!netif_running(dev))
1090 		return IRQ_NONE;
1091 	napi_schedule(&vp->napi);
1092 	return IRQ_HANDLED;
1093 
1094 }
1095 
vector_tx_interrupt(int irq,void * dev_id)1096 static irqreturn_t vector_tx_interrupt(int irq, void *dev_id)
1097 {
1098 	struct net_device *dev = dev_id;
1099 	struct vector_private *vp = netdev_priv(dev);
1100 
1101 	if (!netif_running(dev))
1102 		return IRQ_NONE;
1103 	/* We need to pay attention to it only if we got
1104 	 * -EAGAIN or -ENOBUFFS from sendmmsg. Otherwise
1105 	 * we ignore it. In the future, it may be worth
1106 	 * it to improve the IRQ controller a bit to make
1107 	 * tweaking the IRQ mask less costly
1108 	 */
1109 
1110 	napi_schedule(&vp->napi);
1111 	return IRQ_HANDLED;
1112 
1113 }
1114 
1115 static int irq_rr;
1116 
vector_net_close(struct net_device * dev)1117 static int vector_net_close(struct net_device *dev)
1118 {
1119 	struct vector_private *vp = netdev_priv(dev);
1120 	unsigned long flags;
1121 
1122 	netif_stop_queue(dev);
1123 	del_timer(&vp->tl);
1124 
1125 	if (vp->fds == NULL)
1126 		return 0;
1127 
1128 	/* Disable and free all IRQS */
1129 	if (vp->rx_irq > 0) {
1130 		um_free_irq(vp->rx_irq, dev);
1131 		vp->rx_irq = 0;
1132 	}
1133 	if (vp->tx_irq > 0) {
1134 		um_free_irq(vp->tx_irq, dev);
1135 		vp->tx_irq = 0;
1136 	}
1137 	napi_disable(&vp->napi);
1138 	netif_napi_del(&vp->napi);
1139 	if (vp->fds->rx_fd > 0) {
1140 		if (vp->bpf)
1141 			uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1142 		os_close_file(vp->fds->rx_fd);
1143 		vp->fds->rx_fd = -1;
1144 	}
1145 	if (vp->fds->tx_fd > 0) {
1146 		os_close_file(vp->fds->tx_fd);
1147 		vp->fds->tx_fd = -1;
1148 	}
1149 	if (vp->bpf != NULL)
1150 		kfree(vp->bpf->filter);
1151 	kfree(vp->bpf);
1152 	vp->bpf = NULL;
1153 	kfree(vp->fds->remote_addr);
1154 	kfree(vp->transport_data);
1155 	kfree(vp->header_rxbuffer);
1156 	kfree(vp->header_txbuffer);
1157 	if (vp->rx_queue != NULL)
1158 		destroy_queue(vp->rx_queue);
1159 	if (vp->tx_queue != NULL)
1160 		destroy_queue(vp->tx_queue);
1161 	kfree(vp->fds);
1162 	vp->fds = NULL;
1163 	spin_lock_irqsave(&vp->lock, flags);
1164 	vp->opened = false;
1165 	vp->in_error = false;
1166 	spin_unlock_irqrestore(&vp->lock, flags);
1167 	return 0;
1168 }
1169 
vector_poll(struct napi_struct * napi,int budget)1170 static int vector_poll(struct napi_struct *napi, int budget)
1171 {
1172 	struct vector_private *vp = container_of(napi, struct vector_private, napi);
1173 	int work_done = 0;
1174 	int err;
1175 	bool tx_enqueued = false;
1176 
1177 	if ((vp->options & VECTOR_TX) != 0)
1178 		tx_enqueued = (vector_send(vp->tx_queue) > 0);
1179 	if ((vp->options & VECTOR_RX) > 0)
1180 		err = vector_mmsg_rx(vp, budget);
1181 	else {
1182 		err = vector_legacy_rx(vp);
1183 		if (err > 0)
1184 			err = 1;
1185 	}
1186 	if (err > 0)
1187 		work_done += err;
1188 
1189 	if (tx_enqueued || err > 0)
1190 		napi_schedule(napi);
1191 	if (work_done < budget)
1192 		napi_complete_done(napi, work_done);
1193 	return work_done;
1194 }
1195 
vector_reset_tx(struct work_struct * work)1196 static void vector_reset_tx(struct work_struct *work)
1197 {
1198 	struct vector_private *vp =
1199 		container_of(work, struct vector_private, reset_tx);
1200 	netdev_reset_queue(vp->dev);
1201 	netif_start_queue(vp->dev);
1202 	netif_wake_queue(vp->dev);
1203 }
1204 
vector_net_open(struct net_device * dev)1205 static int vector_net_open(struct net_device *dev)
1206 {
1207 	struct vector_private *vp = netdev_priv(dev);
1208 	unsigned long flags;
1209 	int err = -EINVAL;
1210 	struct vector_device *vdevice;
1211 
1212 	spin_lock_irqsave(&vp->lock, flags);
1213 	if (vp->opened) {
1214 		spin_unlock_irqrestore(&vp->lock, flags);
1215 		return -ENXIO;
1216 	}
1217 	vp->opened = true;
1218 	spin_unlock_irqrestore(&vp->lock, flags);
1219 
1220 	vp->bpf = uml_vector_user_bpf(get_bpf_file(vp->parsed));
1221 
1222 	vp->fds = uml_vector_user_open(vp->unit, vp->parsed);
1223 
1224 	if (vp->fds == NULL)
1225 		goto out_close;
1226 
1227 	if (build_transport_data(vp) < 0)
1228 		goto out_close;
1229 
1230 	if ((vp->options & VECTOR_RX) > 0) {
1231 		vp->rx_queue = create_queue(
1232 			vp,
1233 			get_depth(vp->parsed),
1234 			vp->rx_header_size,
1235 			MAX_IOV_SIZE
1236 		);
1237 		vp->rx_queue->queue_depth = get_depth(vp->parsed);
1238 	} else {
1239 		vp->header_rxbuffer = kmalloc(
1240 			vp->rx_header_size,
1241 			GFP_KERNEL
1242 		);
1243 		if (vp->header_rxbuffer == NULL)
1244 			goto out_close;
1245 	}
1246 	if ((vp->options & VECTOR_TX) > 0) {
1247 		vp->tx_queue = create_queue(
1248 			vp,
1249 			get_depth(vp->parsed),
1250 			vp->header_size,
1251 			MAX_IOV_SIZE
1252 		);
1253 	} else {
1254 		vp->header_txbuffer = kmalloc(vp->header_size, GFP_KERNEL);
1255 		if (vp->header_txbuffer == NULL)
1256 			goto out_close;
1257 	}
1258 
1259 	netif_napi_add_weight(vp->dev, &vp->napi, vector_poll,
1260 			      get_depth(vp->parsed));
1261 	napi_enable(&vp->napi);
1262 
1263 	/* READ IRQ */
1264 	err = um_request_irq(
1265 		irq_rr + VECTOR_BASE_IRQ, vp->fds->rx_fd,
1266 			IRQ_READ, vector_rx_interrupt,
1267 			IRQF_SHARED, dev->name, dev);
1268 	if (err < 0) {
1269 		netdev_err(dev, "vector_open: failed to get rx irq(%d)\n", err);
1270 		err = -ENETUNREACH;
1271 		goto out_close;
1272 	}
1273 	vp->rx_irq = irq_rr + VECTOR_BASE_IRQ;
1274 	dev->irq = irq_rr + VECTOR_BASE_IRQ;
1275 	irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1276 
1277 	/* WRITE IRQ - we need it only if we have vector TX */
1278 	if ((vp->options & VECTOR_TX) > 0) {
1279 		err = um_request_irq(
1280 			irq_rr + VECTOR_BASE_IRQ, vp->fds->tx_fd,
1281 				IRQ_WRITE, vector_tx_interrupt,
1282 				IRQF_SHARED, dev->name, dev);
1283 		if (err < 0) {
1284 			netdev_err(dev,
1285 				"vector_open: failed to get tx irq(%d)\n", err);
1286 			err = -ENETUNREACH;
1287 			goto out_close;
1288 		}
1289 		vp->tx_irq = irq_rr + VECTOR_BASE_IRQ;
1290 		irq_rr = (irq_rr + 1) % VECTOR_IRQ_SPACE;
1291 	}
1292 
1293 	if ((vp->options & VECTOR_QDISC_BYPASS) != 0) {
1294 		if (!uml_raw_enable_qdisc_bypass(vp->fds->rx_fd))
1295 			vp->options |= VECTOR_BPF;
1296 	}
1297 	if (((vp->options & VECTOR_BPF) != 0) && (vp->bpf == NULL))
1298 		vp->bpf = uml_vector_default_bpf(dev->dev_addr);
1299 
1300 	if (vp->bpf != NULL)
1301 		uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1302 
1303 	netif_start_queue(dev);
1304 	vector_reset_stats(vp);
1305 
1306 	/* clear buffer - it can happen that the host side of the interface
1307 	 * is full when we get here. In this case, new data is never queued,
1308 	 * SIGIOs never arrive, and the net never works.
1309 	 */
1310 
1311 	napi_schedule(&vp->napi);
1312 
1313 	vdevice = find_device(vp->unit);
1314 	vdevice->opened = 1;
1315 
1316 	if ((vp->options & VECTOR_TX) != 0)
1317 		add_timer(&vp->tl);
1318 	return 0;
1319 out_close:
1320 	vector_net_close(dev);
1321 	return err;
1322 }
1323 
1324 
vector_net_set_multicast_list(struct net_device * dev)1325 static void vector_net_set_multicast_list(struct net_device *dev)
1326 {
1327 	/* TODO: - we can do some BPF games here */
1328 	return;
1329 }
1330 
vector_net_tx_timeout(struct net_device * dev,unsigned int txqueue)1331 static void vector_net_tx_timeout(struct net_device *dev, unsigned int txqueue)
1332 {
1333 	struct vector_private *vp = netdev_priv(dev);
1334 
1335 	vp->estats.tx_timeout_count++;
1336 	netif_trans_update(dev);
1337 	schedule_work(&vp->reset_tx);
1338 }
1339 
vector_fix_features(struct net_device * dev,netdev_features_t features)1340 static netdev_features_t vector_fix_features(struct net_device *dev,
1341 	netdev_features_t features)
1342 {
1343 	features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
1344 	return features;
1345 }
1346 
vector_set_features(struct net_device * dev,netdev_features_t features)1347 static int vector_set_features(struct net_device *dev,
1348 	netdev_features_t features)
1349 {
1350 	struct vector_private *vp = netdev_priv(dev);
1351 	/* Adjust buffer sizes for GSO/GRO. Unfortunately, there is
1352 	 * no way to negotiate it on raw sockets, so we can change
1353 	 * only our side.
1354 	 */
1355 	if (features & NETIF_F_GRO)
1356 		/* All new frame buffers will be GRO-sized */
1357 		vp->req_size = 65536;
1358 	else
1359 		/* All new frame buffers will be normal sized */
1360 		vp->req_size = vp->max_packet + vp->headroom + SAFETY_MARGIN;
1361 	return 0;
1362 }
1363 
1364 #ifdef CONFIG_NET_POLL_CONTROLLER
vector_net_poll_controller(struct net_device * dev)1365 static void vector_net_poll_controller(struct net_device *dev)
1366 {
1367 	disable_irq(dev->irq);
1368 	vector_rx_interrupt(dev->irq, dev);
1369 	enable_irq(dev->irq);
1370 }
1371 #endif
1372 
vector_net_get_drvinfo(struct net_device * dev,struct ethtool_drvinfo * info)1373 static void vector_net_get_drvinfo(struct net_device *dev,
1374 				struct ethtool_drvinfo *info)
1375 {
1376 	strscpy(info->driver, DRIVER_NAME, sizeof(info->driver));
1377 }
1378 
vector_net_load_bpf_flash(struct net_device * dev,struct ethtool_flash * efl)1379 static int vector_net_load_bpf_flash(struct net_device *dev,
1380 				struct ethtool_flash *efl)
1381 {
1382 	struct vector_private *vp = netdev_priv(dev);
1383 	struct vector_device *vdevice;
1384 	const struct firmware *fw;
1385 	int result = 0;
1386 
1387 	if (!(vp->options & VECTOR_BPF_FLASH)) {
1388 		netdev_err(dev, "loading firmware not permitted: %s\n", efl->data);
1389 		return -1;
1390 	}
1391 
1392 	spin_lock(&vp->lock);
1393 
1394 	if (vp->bpf != NULL) {
1395 		if (vp->opened)
1396 			uml_vector_detach_bpf(vp->fds->rx_fd, vp->bpf);
1397 		kfree(vp->bpf->filter);
1398 		vp->bpf->filter = NULL;
1399 	} else {
1400 		vp->bpf = kmalloc(sizeof(struct sock_fprog), GFP_ATOMIC);
1401 		if (vp->bpf == NULL) {
1402 			netdev_err(dev, "failed to allocate memory for firmware\n");
1403 			goto flash_fail;
1404 		}
1405 	}
1406 
1407 	vdevice = find_device(vp->unit);
1408 
1409 	if (request_firmware(&fw, efl->data, &vdevice->pdev.dev))
1410 		goto flash_fail;
1411 
1412 	vp->bpf->filter = kmemdup(fw->data, fw->size, GFP_ATOMIC);
1413 	if (!vp->bpf->filter)
1414 		goto free_buffer;
1415 
1416 	vp->bpf->len = fw->size / sizeof(struct sock_filter);
1417 	release_firmware(fw);
1418 
1419 	if (vp->opened)
1420 		result = uml_vector_attach_bpf(vp->fds->rx_fd, vp->bpf);
1421 
1422 	spin_unlock(&vp->lock);
1423 
1424 	return result;
1425 
1426 free_buffer:
1427 	release_firmware(fw);
1428 
1429 flash_fail:
1430 	spin_unlock(&vp->lock);
1431 	if (vp->bpf != NULL)
1432 		kfree(vp->bpf->filter);
1433 	kfree(vp->bpf);
1434 	vp->bpf = NULL;
1435 	return -1;
1436 }
1437 
vector_get_ringparam(struct net_device * netdev,struct ethtool_ringparam * ring,struct kernel_ethtool_ringparam * kernel_ring,struct netlink_ext_ack * extack)1438 static void vector_get_ringparam(struct net_device *netdev,
1439 				 struct ethtool_ringparam *ring,
1440 				 struct kernel_ethtool_ringparam *kernel_ring,
1441 				 struct netlink_ext_ack *extack)
1442 {
1443 	struct vector_private *vp = netdev_priv(netdev);
1444 
1445 	ring->rx_max_pending = vp->rx_queue->max_depth;
1446 	ring->tx_max_pending = vp->tx_queue->max_depth;
1447 	ring->rx_pending = vp->rx_queue->max_depth;
1448 	ring->tx_pending = vp->tx_queue->max_depth;
1449 }
1450 
vector_get_strings(struct net_device * dev,u32 stringset,u8 * buf)1451 static void vector_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
1452 {
1453 	switch (stringset) {
1454 	case ETH_SS_TEST:
1455 		*buf = '\0';
1456 		break;
1457 	case ETH_SS_STATS:
1458 		memcpy(buf, &ethtool_stats_keys, sizeof(ethtool_stats_keys));
1459 		break;
1460 	default:
1461 		WARN_ON(1);
1462 		break;
1463 	}
1464 }
1465 
vector_get_sset_count(struct net_device * dev,int sset)1466 static int vector_get_sset_count(struct net_device *dev, int sset)
1467 {
1468 	switch (sset) {
1469 	case ETH_SS_TEST:
1470 		return 0;
1471 	case ETH_SS_STATS:
1472 		return VECTOR_NUM_STATS;
1473 	default:
1474 		return -EOPNOTSUPP;
1475 	}
1476 }
1477 
vector_get_ethtool_stats(struct net_device * dev,struct ethtool_stats * estats,u64 * tmp_stats)1478 static void vector_get_ethtool_stats(struct net_device *dev,
1479 	struct ethtool_stats *estats,
1480 	u64 *tmp_stats)
1481 {
1482 	struct vector_private *vp = netdev_priv(dev);
1483 
1484 	memcpy(tmp_stats, &vp->estats, sizeof(struct vector_estats));
1485 }
1486 
vector_get_coalesce(struct net_device * netdev,struct ethtool_coalesce * ec,struct kernel_ethtool_coalesce * kernel_coal,struct netlink_ext_ack * extack)1487 static int vector_get_coalesce(struct net_device *netdev,
1488 			       struct ethtool_coalesce *ec,
1489 			       struct kernel_ethtool_coalesce *kernel_coal,
1490 			       struct netlink_ext_ack *extack)
1491 {
1492 	struct vector_private *vp = netdev_priv(netdev);
1493 
1494 	ec->tx_coalesce_usecs = (vp->coalesce * 1000000) / HZ;
1495 	return 0;
1496 }
1497 
vector_set_coalesce(struct net_device * netdev,struct ethtool_coalesce * ec,struct kernel_ethtool_coalesce * kernel_coal,struct netlink_ext_ack * extack)1498 static int vector_set_coalesce(struct net_device *netdev,
1499 			       struct ethtool_coalesce *ec,
1500 			       struct kernel_ethtool_coalesce *kernel_coal,
1501 			       struct netlink_ext_ack *extack)
1502 {
1503 	struct vector_private *vp = netdev_priv(netdev);
1504 
1505 	vp->coalesce = (ec->tx_coalesce_usecs * HZ) / 1000000;
1506 	if (vp->coalesce == 0)
1507 		vp->coalesce = 1;
1508 	return 0;
1509 }
1510 
1511 static const struct ethtool_ops vector_net_ethtool_ops = {
1512 	.supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
1513 	.get_drvinfo	= vector_net_get_drvinfo,
1514 	.get_link	= ethtool_op_get_link,
1515 	.get_ts_info	= ethtool_op_get_ts_info,
1516 	.get_ringparam	= vector_get_ringparam,
1517 	.get_strings	= vector_get_strings,
1518 	.get_sset_count	= vector_get_sset_count,
1519 	.get_ethtool_stats = vector_get_ethtool_stats,
1520 	.get_coalesce	= vector_get_coalesce,
1521 	.set_coalesce	= vector_set_coalesce,
1522 	.flash_device	= vector_net_load_bpf_flash,
1523 };
1524 
1525 
1526 static const struct net_device_ops vector_netdev_ops = {
1527 	.ndo_open		= vector_net_open,
1528 	.ndo_stop		= vector_net_close,
1529 	.ndo_start_xmit		= vector_net_start_xmit,
1530 	.ndo_set_rx_mode	= vector_net_set_multicast_list,
1531 	.ndo_tx_timeout		= vector_net_tx_timeout,
1532 	.ndo_set_mac_address	= eth_mac_addr,
1533 	.ndo_validate_addr	= eth_validate_addr,
1534 	.ndo_fix_features	= vector_fix_features,
1535 	.ndo_set_features	= vector_set_features,
1536 #ifdef CONFIG_NET_POLL_CONTROLLER
1537 	.ndo_poll_controller = vector_net_poll_controller,
1538 #endif
1539 };
1540 
vector_timer_expire(struct timer_list * t)1541 static void vector_timer_expire(struct timer_list *t)
1542 {
1543 	struct vector_private *vp = from_timer(vp, t, tl);
1544 
1545 	vp->estats.tx_kicks++;
1546 	napi_schedule(&vp->napi);
1547 }
1548 
1549 
1550 
vector_eth_configure(int n,struct arglist * def)1551 static void vector_eth_configure(
1552 		int n,
1553 		struct arglist *def
1554 	)
1555 {
1556 	struct vector_device *device;
1557 	struct net_device *dev;
1558 	struct vector_private *vp;
1559 	int err;
1560 
1561 	device = kzalloc(sizeof(*device), GFP_KERNEL);
1562 	if (device == NULL) {
1563 		printk(KERN_ERR "eth_configure failed to allocate struct "
1564 				 "vector_device\n");
1565 		return;
1566 	}
1567 	dev = alloc_etherdev(sizeof(struct vector_private));
1568 	if (dev == NULL) {
1569 		printk(KERN_ERR "eth_configure: failed to allocate struct "
1570 				 "net_device for vec%d\n", n);
1571 		goto out_free_device;
1572 	}
1573 
1574 	dev->mtu = get_mtu(def);
1575 
1576 	INIT_LIST_HEAD(&device->list);
1577 	device->unit = n;
1578 
1579 	/* If this name ends up conflicting with an existing registered
1580 	 * netdevice, that is OK, register_netdev{,ice}() will notice this
1581 	 * and fail.
1582 	 */
1583 	snprintf(dev->name, sizeof(dev->name), "vec%d", n);
1584 	uml_net_setup_etheraddr(dev, uml_vector_fetch_arg(def, "mac"));
1585 	vp = netdev_priv(dev);
1586 
1587 	/* sysfs register */
1588 	if (!driver_registered) {
1589 		platform_driver_register(&uml_net_driver);
1590 		driver_registered = 1;
1591 	}
1592 	device->pdev.id = n;
1593 	device->pdev.name = DRIVER_NAME;
1594 	device->pdev.dev.release = vector_device_release;
1595 	dev_set_drvdata(&device->pdev.dev, device);
1596 	if (platform_device_register(&device->pdev))
1597 		goto out_free_netdev;
1598 	SET_NETDEV_DEV(dev, &device->pdev.dev);
1599 
1600 	device->dev = dev;
1601 
1602 	*vp = ((struct vector_private)
1603 		{
1604 		.list			= LIST_HEAD_INIT(vp->list),
1605 		.dev			= dev,
1606 		.unit			= n,
1607 		.options		= get_transport_options(def),
1608 		.rx_irq			= 0,
1609 		.tx_irq			= 0,
1610 		.parsed			= def,
1611 		.max_packet		= get_mtu(def) + ETH_HEADER_OTHER,
1612 		/* TODO - we need to calculate headroom so that ip header
1613 		 * is 16 byte aligned all the time
1614 		 */
1615 		.headroom		= get_headroom(def),
1616 		.form_header		= NULL,
1617 		.verify_header		= NULL,
1618 		.header_rxbuffer	= NULL,
1619 		.header_txbuffer	= NULL,
1620 		.header_size		= 0,
1621 		.rx_header_size		= 0,
1622 		.rexmit_scheduled	= false,
1623 		.opened			= false,
1624 		.transport_data		= NULL,
1625 		.in_write_poll		= false,
1626 		.coalesce		= 2,
1627 		.req_size		= get_req_size(def),
1628 		.in_error		= false,
1629 		.bpf			= NULL
1630 	});
1631 
1632 	dev->features = dev->hw_features = (NETIF_F_SG | NETIF_F_FRAGLIST);
1633 	INIT_WORK(&vp->reset_tx, vector_reset_tx);
1634 
1635 	timer_setup(&vp->tl, vector_timer_expire, 0);
1636 	spin_lock_init(&vp->lock);
1637 
1638 	/* FIXME */
1639 	dev->netdev_ops = &vector_netdev_ops;
1640 	dev->ethtool_ops = &vector_net_ethtool_ops;
1641 	dev->watchdog_timeo = (HZ >> 1);
1642 	/* primary IRQ - fixme */
1643 	dev->irq = 0; /* we will adjust this once opened */
1644 
1645 	rtnl_lock();
1646 	err = register_netdevice(dev);
1647 	rtnl_unlock();
1648 	if (err)
1649 		goto out_undo_user_init;
1650 
1651 	spin_lock(&vector_devices_lock);
1652 	list_add(&device->list, &vector_devices);
1653 	spin_unlock(&vector_devices_lock);
1654 
1655 	return;
1656 
1657 out_undo_user_init:
1658 	return;
1659 out_free_netdev:
1660 	free_netdev(dev);
1661 out_free_device:
1662 	kfree(device);
1663 }
1664 
1665 
1666 
1667 
1668 /*
1669  * Invoked late in the init
1670  */
1671 
vector_init(void)1672 static int __init vector_init(void)
1673 {
1674 	struct list_head *ele;
1675 	struct vector_cmd_line_arg *def;
1676 	struct arglist *parsed;
1677 
1678 	list_for_each(ele, &vec_cmd_line) {
1679 		def = list_entry(ele, struct vector_cmd_line_arg, list);
1680 		parsed = uml_parse_vector_ifspec(def->arguments);
1681 		if (parsed != NULL)
1682 			vector_eth_configure(def->unit, parsed);
1683 	}
1684 	return 0;
1685 }
1686 
1687 
1688 /* Invoked at initial argument parsing, only stores
1689  * arguments until a proper vector_init is called
1690  * later
1691  */
1692 
vector_setup(char * str)1693 static int __init vector_setup(char *str)
1694 {
1695 	char *error;
1696 	int n, err;
1697 	struct vector_cmd_line_arg *new;
1698 
1699 	err = vector_parse(str, &n, &str, &error);
1700 	if (err) {
1701 		printk(KERN_ERR "vector_setup - Couldn't parse '%s' : %s\n",
1702 				 str, error);
1703 		return 1;
1704 	}
1705 	new = memblock_alloc(sizeof(*new), SMP_CACHE_BYTES);
1706 	if (!new)
1707 		panic("%s: Failed to allocate %zu bytes\n", __func__,
1708 		      sizeof(*new));
1709 	INIT_LIST_HEAD(&new->list);
1710 	new->unit = n;
1711 	new->arguments = str;
1712 	list_add_tail(&new->list, &vec_cmd_line);
1713 	return 1;
1714 }
1715 
1716 __setup("vec", vector_setup);
1717 __uml_help(vector_setup,
1718 "vec[0-9]+:<option>=<value>,<option>=<value>\n"
1719 "	 Configure a vector io network device.\n\n"
1720 );
1721 
1722 late_initcall(vector_init);
1723 
1724 static struct mc_device vector_mc = {
1725 	.list		= LIST_HEAD_INIT(vector_mc.list),
1726 	.name		= "vec",
1727 	.config		= vector_config,
1728 	.get_config	= NULL,
1729 	.id		= vector_id,
1730 	.remove		= vector_remove,
1731 };
1732 
1733 #ifdef CONFIG_INET
vector_inetaddr_event(struct notifier_block * this,unsigned long event,void * ptr)1734 static int vector_inetaddr_event(
1735 	struct notifier_block *this,
1736 	unsigned long event,
1737 	void *ptr)
1738 {
1739 	return NOTIFY_DONE;
1740 }
1741 
1742 static struct notifier_block vector_inetaddr_notifier = {
1743 	.notifier_call		= vector_inetaddr_event,
1744 };
1745 
inet_register(void)1746 static void inet_register(void)
1747 {
1748 	register_inetaddr_notifier(&vector_inetaddr_notifier);
1749 }
1750 #else
inet_register(void)1751 static inline void inet_register(void)
1752 {
1753 }
1754 #endif
1755 
vector_net_init(void)1756 static int vector_net_init(void)
1757 {
1758 	mconsole_register_dev(&vector_mc);
1759 	inet_register();
1760 	return 0;
1761 }
1762 
1763 __initcall(vector_net_init);
1764 
1765 
1766 
1767