2 * Copyright (c) 2016 Citrix Systems Inc.
3 * Copyright (c) 2002-2005, K A Fraser
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License version 2
7 * as published by the Free Software Foundation; or, when distributed
8 * separately from the Linux kernel or incorporated into other
9 * software packages, subject to the following license:
11 * Permission is hereby granted, free of charge, to any person obtaining a copy
12 * of this source file (the "Software"), to deal in the Software without
13 * restriction, including without limitation the rights to use, copy, modify,
14 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
15 * and to permit persons to whom the Software is furnished to do so, subject to
16 * the following conditions:
18 * The above copyright notice and this permission notice shall be included in
19 * all copies or substantial portions of the Software.
21 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
22 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
23 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
24 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
25 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
26 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
31 #include <linux/kthread.h>
34 #include <xen/events.h>
36 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
42 skb = skb_peek(&queue->rx_queue);
46 needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
53 prod = queue->rx.sring->req_prod;
54 cons = queue->rx.req_cons;
56 if (prod - cons >= needed)
59 queue->rx.sring->req_event = prod + 1;
61 /* Make sure event is visible before we check prod
65 } while (queue->rx.sring->req_prod != prod);
70 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
74 spin_lock_irqsave(&queue->rx_queue.lock, flags);
76 __skb_queue_tail(&queue->rx_queue, skb);
78 queue->rx_queue_len += skb->len;
79 if (queue->rx_queue_len > queue->rx_queue_max) {
80 struct net_device *dev = queue->vif->dev;
82 netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id));
85 spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
88 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
92 spin_lock_irq(&queue->rx_queue.lock);
94 skb = __skb_dequeue(&queue->rx_queue);
96 queue->rx_queue_len -= skb->len;
97 if (queue->rx_queue_len < queue->rx_queue_max) {
98 struct netdev_queue *txq;
100 txq = netdev_get_tx_queue(queue->vif->dev, queue->id);
101 netif_tx_wake_queue(txq);
105 spin_unlock_irq(&queue->rx_queue.lock);
110 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
114 while ((skb = xenvif_rx_dequeue(queue)) != NULL)
118 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
123 skb = skb_peek(&queue->rx_queue);
126 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
128 xenvif_rx_dequeue(queue);
133 static void xenvif_rx_copy_flush(struct xenvif_queue *queue)
138 gnttab_batch_copy(queue->rx_copy.op, queue->rx_copy.num);
140 for (i = 0; i < queue->rx_copy.num; i++) {
141 struct gnttab_copy *op;
143 op = &queue->rx_copy.op[i];
145 /* If the copy failed, overwrite the status field in
146 * the corresponding response.
148 if (unlikely(op->status != GNTST_okay)) {
149 struct xen_netif_rx_response *rsp;
151 rsp = RING_GET_RESPONSE(&queue->rx,
152 queue->rx_copy.idx[i]);
153 rsp->status = op->status;
157 queue->rx_copy.num = 0;
159 /* Push responses for all completed packets. */
160 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, notify);
162 notify_remote_via_irq(queue->rx_irq);
164 __skb_queue_purge(queue->rx_copy.completed);
167 static void xenvif_rx_copy_add(struct xenvif_queue *queue,
168 struct xen_netif_rx_request *req,
169 unsigned int offset, void *data, size_t len)
171 struct gnttab_copy *op;
173 struct xen_page_foreign *foreign;
175 if (queue->rx_copy.num == COPY_BATCH_SIZE)
176 xenvif_rx_copy_flush(queue);
178 op = &queue->rx_copy.op[queue->rx_copy.num];
180 page = virt_to_page(data);
182 op->flags = GNTCOPY_dest_gref;
184 foreign = xen_page_foreign(page);
186 op->source.domid = foreign->domid;
187 op->source.u.ref = foreign->gref;
188 op->flags |= GNTCOPY_source_gref;
190 op->source.u.gmfn = virt_to_gfn(data);
191 op->source.domid = DOMID_SELF;
194 op->source.offset = xen_offset_in_page(data);
195 op->dest.u.ref = req->gref;
196 op->dest.domid = queue->vif->domid;
197 op->dest.offset = offset;
200 queue->rx_copy.idx[queue->rx_copy.num] = queue->rx.req_cons;
201 queue->rx_copy.num++;
204 static unsigned int xenvif_gso_type(struct sk_buff *skb)
206 if (skb_is_gso(skb)) {
207 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
208 return XEN_NETIF_GSO_TYPE_TCPV4;
210 return XEN_NETIF_GSO_TYPE_TCPV6;
212 return XEN_NETIF_GSO_TYPE_NONE;
215 struct xenvif_pkt_state {
217 size_t remaining_len;
218 int frag; /* frag == -1 => skb->head */
219 unsigned int frag_offset;
220 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
221 unsigned int extra_count;
225 static void xenvif_rx_next_skb(struct xenvif_queue *queue,
226 struct xenvif_pkt_state *pkt)
229 unsigned int gso_type;
231 skb = xenvif_rx_dequeue(queue);
233 queue->stats.tx_bytes += skb->len;
234 queue->stats.tx_packets++;
236 /* Reset packet state. */
237 memset(pkt, 0, sizeof(struct xenvif_pkt_state));
240 pkt->remaining_len = skb->len;
243 gso_type = xenvif_gso_type(skb);
244 if ((1 << gso_type) & queue->vif->gso_mask) {
245 struct xen_netif_extra_info *extra;
247 extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
249 extra->u.gso.type = gso_type;
250 extra->u.gso.size = skb_shinfo(skb)->gso_size;
251 extra->u.gso.pad = 0;
252 extra->u.gso.features = 0;
253 extra->type = XEN_NETIF_EXTRA_TYPE_GSO;
260 struct xen_netif_extra_info *extra;
262 extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
264 extra->u.hash.algorithm =
265 XEN_NETIF_CTRL_HASH_ALGORITHM_TOEPLITZ;
269 skb->protocol == htons(ETH_P_IP) ?
270 _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP :
271 _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP;
274 skb->protocol == htons(ETH_P_IP) ?
275 _XEN_NETIF_CTRL_HASH_TYPE_IPV4 :
276 _XEN_NETIF_CTRL_HASH_TYPE_IPV6;
278 *(uint32_t *)extra->u.hash.value = skb_get_hash_raw(skb);
280 extra->type = XEN_NETIF_EXTRA_TYPE_HASH;
287 static void xenvif_rx_complete(struct xenvif_queue *queue,
288 struct xenvif_pkt_state *pkt)
290 /* All responses are ready to be pushed. */
291 queue->rx.rsp_prod_pvt = queue->rx.req_cons;
293 __skb_queue_tail(queue->rx_copy.completed, pkt->skb);
296 static void xenvif_rx_next_chunk(struct xenvif_queue *queue,
297 struct xenvif_pkt_state *pkt,
298 unsigned int offset, void **data,
301 struct sk_buff *skb = pkt->skb;
303 size_t frag_len, chunk_len;
305 if (pkt->frag == -1) {
306 frag_data = skb->data;
307 frag_len = skb_headlen(skb);
309 skb_frag_t *frag = &skb_shinfo(skb)->frags[pkt->frag];
311 frag_data = skb_frag_address(frag);
312 frag_len = skb_frag_size(frag);
315 frag_data += pkt->frag_offset;
316 frag_len -= pkt->frag_offset;
318 chunk_len = min(frag_len, XEN_PAGE_SIZE - offset);
319 chunk_len = min(chunk_len,
320 XEN_PAGE_SIZE - xen_offset_in_page(frag_data));
322 pkt->frag_offset += chunk_len;
324 /* Advance to next frag? */
325 if (frag_len == chunk_len) {
327 pkt->frag_offset = 0;
334 static void xenvif_rx_data_slot(struct xenvif_queue *queue,
335 struct xenvif_pkt_state *pkt,
336 struct xen_netif_rx_request *req,
337 struct xen_netif_rx_response *rsp)
339 unsigned int offset = 0;
346 xenvif_rx_next_chunk(queue, pkt, offset, &data, &len);
347 xenvif_rx_copy_add(queue, req, offset, data, len);
350 pkt->remaining_len -= len;
352 } while (offset < XEN_PAGE_SIZE && pkt->remaining_len > 0);
354 if (pkt->remaining_len > 0)
355 flags = XEN_NETRXF_more_data;
359 if (pkt->slot == 0) {
360 struct sk_buff *skb = pkt->skb;
362 if (skb->ip_summed == CHECKSUM_PARTIAL)
363 flags |= XEN_NETRXF_csum_blank |
364 XEN_NETRXF_data_validated;
365 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
366 flags |= XEN_NETRXF_data_validated;
368 if (pkt->extra_count != 0)
369 flags |= XEN_NETRXF_extra_info;
375 rsp->status = (s16)offset;
378 static void xenvif_rx_extra_slot(struct xenvif_queue *queue,
379 struct xenvif_pkt_state *pkt,
380 struct xen_netif_rx_request *req,
381 struct xen_netif_rx_response *rsp)
383 struct xen_netif_extra_info *extra = (void *)rsp;
388 for (i = 0; i < ARRAY_SIZE(pkt->extras); i++) {
389 if (pkt->extras[i].type) {
390 *extra = pkt->extras[i];
392 if (pkt->extra_count != 0)
393 extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE;
395 pkt->extras[i].type = 0;
402 void xenvif_rx_skb(struct xenvif_queue *queue)
404 struct xenvif_pkt_state pkt;
406 xenvif_rx_next_skb(queue, &pkt);
409 struct xen_netif_rx_request *req;
410 struct xen_netif_rx_response *rsp;
412 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons);
413 rsp = RING_GET_RESPONSE(&queue->rx, queue->rx.req_cons);
415 /* Extras must go after the first data slot */
416 if (pkt.slot != 0 && pkt.extra_count != 0)
417 xenvif_rx_extra_slot(queue, &pkt, req, rsp);
419 xenvif_rx_data_slot(queue, &pkt, req, rsp);
421 queue->rx.req_cons++;
423 } while (pkt.remaining_len > 0 || pkt.extra_count != 0);
425 xenvif_rx_complete(queue, &pkt);
428 #define RX_BATCH_SIZE 64
430 void xenvif_rx_action(struct xenvif_queue *queue)
432 struct sk_buff_head completed_skbs;
433 unsigned int work_done = 0;
435 __skb_queue_head_init(&completed_skbs);
436 queue->rx_copy.completed = &completed_skbs;
438 while (xenvif_rx_ring_slots_available(queue) &&
439 work_done < RX_BATCH_SIZE) {
440 xenvif_rx_skb(queue);
444 /* Flush any pending copies and complete all skbs. */
445 xenvif_rx_copy_flush(queue);
448 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
452 prod = queue->rx.sring->req_prod;
453 cons = queue->rx.req_cons;
455 return !queue->stalled &&
458 queue->last_rx_time + queue->vif->stall_timeout);
461 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
465 prod = queue->rx.sring->req_prod;
466 cons = queue->rx.req_cons;
468 return queue->stalled && prod - cons >= 1;
471 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
473 return xenvif_rx_ring_slots_available(queue) ||
474 (queue->vif->stall_timeout &&
475 (xenvif_rx_queue_stalled(queue) ||
476 xenvif_rx_queue_ready(queue))) ||
477 kthread_should_stop() ||
478 queue->vif->disabled;
481 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
486 skb = skb_peek(&queue->rx_queue);
488 return MAX_SCHEDULE_TIMEOUT;
490 timeout = XENVIF_RX_CB(skb)->expires - jiffies;
491 return timeout < 0 ? 0 : timeout;
494 /* Wait until the guest Rx thread has work.
496 * The timeout needs to be adjusted based on the current head of the
497 * queue (and not just the head at the beginning). In particular, if
498 * the queue is initially empty an infinite timeout is used and this
499 * needs to be reduced when a skb is queued.
501 * This cannot be done with wait_event_timeout() because it only
502 * calculates the timeout once.
504 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
508 if (xenvif_have_rx_work(queue))
514 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
515 if (xenvif_have_rx_work(queue))
517 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
521 finish_wait(&queue->wq, &wait);
524 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
526 struct xenvif *vif = queue->vif;
528 queue->stalled = true;
530 /* At least one queue has stalled? Disable the carrier. */
531 spin_lock(&vif->lock);
532 if (vif->stalled_queues++ == 0) {
533 netdev_info(vif->dev, "Guest Rx stalled");
534 netif_carrier_off(vif->dev);
536 spin_unlock(&vif->lock);
539 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
541 struct xenvif *vif = queue->vif;
543 queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
544 queue->stalled = false;
546 /* All queues are ready? Enable the carrier. */
547 spin_lock(&vif->lock);
548 if (--vif->stalled_queues == 0) {
549 netdev_info(vif->dev, "Guest Rx ready");
550 netif_carrier_on(vif->dev);
552 spin_unlock(&vif->lock);
555 int xenvif_kthread_guest_rx(void *data)
557 struct xenvif_queue *queue = data;
558 struct xenvif *vif = queue->vif;
560 if (!vif->stall_timeout)
561 xenvif_queue_carrier_on(queue);
564 xenvif_wait_for_rx_work(queue);
566 if (kthread_should_stop())
569 /* This frontend is found to be rogue, disable it in
570 * kthread context. Currently this is only set when
571 * netback finds out frontend sends malformed packet,
572 * but we cannot disable the interface in softirq
573 * context so we defer it here, if this thread is
574 * associated with queue 0.
576 if (unlikely(vif->disabled && queue->id == 0)) {
577 xenvif_carrier_off(vif);
581 if (!skb_queue_empty(&queue->rx_queue))
582 xenvif_rx_action(queue);
584 /* If the guest hasn't provided any Rx slots for a
585 * while it's probably not responsive, drop the
586 * carrier so packets are dropped earlier.
588 if (vif->stall_timeout) {
589 if (xenvif_rx_queue_stalled(queue))
590 xenvif_queue_carrier_off(queue);
591 else if (xenvif_rx_queue_ready(queue))
592 xenvif_queue_carrier_on(queue);
595 /* Queued packets may have foreign pages from other
596 * domains. These cannot be queued indefinitely as
597 * this would starve guests of grant refs and transmit
600 xenvif_rx_queue_drop_expired(queue);
605 /* Bin any remaining skbs */
606 xenvif_rx_queue_purge(queue);