2 * Back-end of the driver for virtual network devices. This portion of the
3 * driver exports a 'unified' network-device interface that can be accessed
4 * by any operating system that implements a compatible front end. A
5 * reference front-end implementation can be found in:
6 * drivers/net/xen-netfront.c
8 * Copyright (c) 2002-2005, K A Fraser
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public License version 2
12 * as published by the Free Software Foundation; or, when distributed
13 * separately from the Linux kernel or incorporated into other
14 * software packages, subject to the following license:
16 * Permission is hereby granted, free of charge, to any person obtaining a copy
17 * of this source file (the "Software"), to deal in the Software without
18 * restriction, including without limitation the rights to use, copy, modify,
19 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
20 * and to permit persons to whom the Software is furnished to do so, subject to
21 * the following conditions:
23 * The above copyright notice and this permission notice shall be included in
24 * all copies or substantial portions of the Software.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
27 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
28 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
29 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
30 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
31 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
37 #include <linux/kthread.h>
38 #include <linux/if_vlan.h>
39 #include <linux/udp.h>
40 #include <linux/highmem.h>
45 #include <xen/events.h>
46 #include <xen/interface/memory.h>
48 #include <asm/xen/hypercall.h>
49 #include <asm/xen/page.h>
51 /* Provide an option to disable split event channels at load time as
52 * event channels are limited resource. Split event channels are
55 bool separate_tx_rx_irq = 1;
56 module_param(separate_tx_rx_irq, bool, 0644);
58 /* The time that packets can stay on the guest Rx internal queue
59 * before they are dropped.
61 unsigned int rx_drain_timeout_msecs = 10000;
62 module_param(rx_drain_timeout_msecs, uint, 0444);
64 /* The length of time before the frontend is considered unresponsive
65 * because it isn't providing Rx slots.
67 unsigned int rx_stall_timeout_msecs = 60000;
68 module_param(rx_stall_timeout_msecs, uint, 0444);
70 unsigned int xenvif_max_queues;
71 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
72 MODULE_PARM_DESC(max_queues,
73 "Maximum number of queues per virtual interface");
76 * This is the maximum slots a skb can have. If a guest sends a skb
77 * which exceeds this limit it is considered malicious.
79 #define FATAL_SKB_SLOTS_DEFAULT 20
80 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
81 module_param(fatal_skb_slots, uint, 0444);
83 /* The amount to copy out of the first guest Tx slot into the skb's
84 * linear area. If the first slot has more data, it will be mapped
85 * and put into the first frag.
87 * This is sized to avoid pulling headers from the frags for most
90 #define XEN_NETBACK_TX_COPY_LEN 128
93 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
96 static void make_tx_response(struct xenvif_queue *queue,
97 struct xen_netif_tx_request *txp,
100 static inline int tx_work_todo(struct xenvif_queue *queue);
102 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
109 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
112 return page_to_pfn(queue->mmap_pages[idx]);
115 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
118 return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
121 #define callback_param(vif, pending_idx) \
122 (vif->pending_tx_info[pending_idx].callback_struct)
124 /* Find the containing VIF's structure from a pointer in pending_tx_info array
126 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
128 u16 pending_idx = ubuf->desc;
129 struct pending_tx_info *temp =
130 container_of(ubuf, struct pending_tx_info, callback_struct);
131 return container_of(temp - pending_idx,
136 static u16 frag_get_pending_idx(skb_frag_t *frag)
138 return (u16)frag->page_offset;
141 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
143 frag->page_offset = pending_idx;
146 static inline pending_ring_idx_t pending_index(unsigned i)
148 return i & (MAX_PENDING_REQS-1);
151 bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue, int needed)
156 prod = queue->rx.sring->req_prod;
157 cons = queue->rx.req_cons;
159 if (prod - cons >= needed)
162 queue->rx.sring->req_event = prod + 1;
164 /* Make sure event is visible before we check prod
168 } while (queue->rx.sring->req_prod != prod);
173 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
177 spin_lock_irqsave(&queue->rx_queue.lock, flags);
179 __skb_queue_tail(&queue->rx_queue, skb);
181 queue->rx_queue_len += skb->len;
182 if (queue->rx_queue_len > queue->rx_queue_max)
183 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
185 spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
188 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
192 spin_lock_irq(&queue->rx_queue.lock);
194 skb = __skb_dequeue(&queue->rx_queue);
196 queue->rx_queue_len -= skb->len;
198 spin_unlock_irq(&queue->rx_queue.lock);
203 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
205 spin_lock_irq(&queue->rx_queue.lock);
207 if (queue->rx_queue_len < queue->rx_queue_max)
208 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
210 spin_unlock_irq(&queue->rx_queue.lock);
214 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
217 while ((skb = xenvif_rx_dequeue(queue)) != NULL)
221 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
226 skb = skb_peek(&queue->rx_queue);
229 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
231 xenvif_rx_dequeue(queue);
237 * Returns true if we should start a new receive buffer instead of
238 * adding 'size' bytes to a buffer which currently contains 'offset'
241 static bool start_new_rx_buffer(int offset, unsigned long size, int head,
244 /* simple case: we have completely filled the current buffer. */
245 if (offset == MAX_BUFFER_OFFSET)
249 * complex case: start a fresh buffer if the current frag
250 * would overflow the current buffer but only if:
251 * (i) this frag would fit completely in the next buffer
252 * and (ii) there is already some data in the current buffer
253 * and (iii) this is not the head buffer.
254 * and (iv) there is no need to fully utilize the buffers
257 * - (i) stops us splitting a frag into two copies
258 * unless the frag is too large for a single buffer.
259 * - (ii) stops us from leaving a buffer pointlessly empty.
260 * - (iii) stops us leaving the first buffer
261 * empty. Strictly speaking this is already covered
262 * by (ii) but is explicitly checked because
263 * netfront relies on the first buffer being
264 * non-empty and can crash otherwise.
265 * - (iv) is needed for skbs which can use up more than MAX_SKB_FRAGS
268 * This means we will effectively linearise small
269 * frags but do not needlessly split large buffers
270 * into multiple copies tend to give large frags their
271 * own buffers as before.
273 BUG_ON(size > MAX_BUFFER_OFFSET);
274 if ((offset + size > MAX_BUFFER_OFFSET) && offset && !head &&
281 struct netrx_pending_operations {
282 unsigned copy_prod, copy_cons;
283 unsigned meta_prod, meta_cons;
284 struct gnttab_copy *copy;
285 struct xenvif_rx_meta *meta;
287 grant_ref_t copy_gref;
290 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
291 struct netrx_pending_operations *npo)
293 struct xenvif_rx_meta *meta;
294 struct xen_netif_rx_request *req;
296 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
298 meta = npo->meta + npo->meta_prod++;
299 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
305 npo->copy_gref = req->gref;
311 * Set up the grant operations for this fragment. If it's a flipping
312 * interface, we also set up the unmap request from here.
314 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
315 struct netrx_pending_operations *npo,
316 struct page *page, unsigned long size,
317 unsigned long offset, int *head)
319 struct gnttab_copy *copy_gop;
320 struct xenvif_rx_meta *meta;
322 int gso_type = XEN_NETIF_GSO_TYPE_NONE;
324 /* Data must not cross a page boundary. */
325 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
327 meta = npo->meta + npo->meta_prod - 1;
329 /* Skip unused frames from start of page */
330 page += offset >> PAGE_SHIFT;
331 offset &= ~PAGE_MASK;
334 struct xen_page_foreign *foreign;
336 BUG_ON(offset >= PAGE_SIZE);
337 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
339 bytes = PAGE_SIZE - offset;
344 if (start_new_rx_buffer(npo->copy_off,
347 XENVIF_RX_CB(skb)->full_coalesce)) {
349 * Netfront requires there to be some data in the head
354 meta = get_next_rx_buffer(queue, npo);
357 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
358 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
360 copy_gop = npo->copy + npo->copy_prod++;
361 copy_gop->flags = GNTCOPY_dest_gref;
362 copy_gop->len = bytes;
364 foreign = xen_page_foreign(page);
366 copy_gop->source.domid = foreign->domid;
367 copy_gop->source.u.ref = foreign->gref;
368 copy_gop->flags |= GNTCOPY_source_gref;
370 copy_gop->source.domid = DOMID_SELF;
371 copy_gop->source.u.gmfn =
372 virt_to_mfn(page_address(page));
374 copy_gop->source.offset = offset;
376 copy_gop->dest.domid = queue->vif->domid;
377 copy_gop->dest.offset = npo->copy_off;
378 copy_gop->dest.u.ref = npo->copy_gref;
380 npo->copy_off += bytes;
387 if (offset == PAGE_SIZE && size) {
388 BUG_ON(!PageCompound(page));
393 /* Leave a gap for the GSO descriptor. */
394 if (skb_is_gso(skb)) {
395 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
396 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
397 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
398 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
401 if (*head && ((1 << gso_type) & queue->vif->gso_mask))
402 queue->rx.req_cons++;
404 *head = 0; /* There must be something in this buffer now. */
410 * Prepare an SKB to be transmitted to the frontend.
412 * This function is responsible for allocating grant operations, meta
415 * It returns the number of meta structures consumed. The number of
416 * ring slots used is always equal to the number of meta slots used
417 * plus the number of GSO descriptors used. Currently, we use either
418 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
419 * frontend-side LRO).
421 static int xenvif_gop_skb(struct sk_buff *skb,
422 struct netrx_pending_operations *npo,
423 struct xenvif_queue *queue)
425 struct xenvif *vif = netdev_priv(skb->dev);
426 int nr_frags = skb_shinfo(skb)->nr_frags;
428 struct xen_netif_rx_request *req;
429 struct xenvif_rx_meta *meta;
435 old_meta_prod = npo->meta_prod;
437 gso_type = XEN_NETIF_GSO_TYPE_NONE;
438 if (skb_is_gso(skb)) {
439 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
440 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
441 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
442 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
445 /* Set up a GSO prefix descriptor, if necessary */
446 if ((1 << gso_type) & vif->gso_prefix_mask) {
447 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
448 meta = npo->meta + npo->meta_prod++;
449 meta->gso_type = gso_type;
450 meta->gso_size = skb_shinfo(skb)->gso_size;
455 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
456 meta = npo->meta + npo->meta_prod++;
458 if ((1 << gso_type) & vif->gso_mask) {
459 meta->gso_type = gso_type;
460 meta->gso_size = skb_shinfo(skb)->gso_size;
462 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
469 npo->copy_gref = req->gref;
472 while (data < skb_tail_pointer(skb)) {
473 unsigned int offset = offset_in_page(data);
474 unsigned int len = PAGE_SIZE - offset;
476 if (data + len > skb_tail_pointer(skb))
477 len = skb_tail_pointer(skb) - data;
479 xenvif_gop_frag_copy(queue, skb, npo,
480 virt_to_page(data), len, offset, &head);
484 for (i = 0; i < nr_frags; i++) {
485 xenvif_gop_frag_copy(queue, skb, npo,
486 skb_frag_page(&skb_shinfo(skb)->frags[i]),
487 skb_frag_size(&skb_shinfo(skb)->frags[i]),
488 skb_shinfo(skb)->frags[i].page_offset,
492 return npo->meta_prod - old_meta_prod;
496 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
497 * used to set up the operations on the top of
498 * netrx_pending_operations, which have since been done. Check that
499 * they didn't give any errors and advance over them.
501 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
502 struct netrx_pending_operations *npo)
504 struct gnttab_copy *copy_op;
505 int status = XEN_NETIF_RSP_OKAY;
508 for (i = 0; i < nr_meta_slots; i++) {
509 copy_op = npo->copy + npo->copy_cons++;
510 if (copy_op->status != GNTST_okay) {
512 "Bad status %d from copy to DOM%d.\n",
513 copy_op->status, vif->domid);
514 status = XEN_NETIF_RSP_ERROR;
521 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
522 struct xenvif_rx_meta *meta,
526 unsigned long offset;
528 /* No fragments used */
529 if (nr_meta_slots <= 1)
534 for (i = 0; i < nr_meta_slots; i++) {
536 if (i == nr_meta_slots - 1)
539 flags = XEN_NETRXF_more_data;
542 make_rx_response(queue, meta[i].id, status, offset,
543 meta[i].size, flags);
547 void xenvif_kick_thread(struct xenvif_queue *queue)
552 static void xenvif_rx_action(struct xenvif_queue *queue)
556 struct xen_netif_rx_response *resp;
557 struct sk_buff_head rxq;
561 unsigned long offset;
562 bool need_to_notify = false;
564 struct netrx_pending_operations npo = {
565 .copy = queue->grant_copy_op,
569 skb_queue_head_init(&rxq);
571 while (xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX)
572 && (skb = xenvif_rx_dequeue(queue)) != NULL) {
573 RING_IDX max_slots_needed;
574 RING_IDX old_req_cons;
575 RING_IDX ring_slots_used;
578 queue->last_rx_time = jiffies;
580 /* We need a cheap worse case estimate for the number of
584 max_slots_needed = DIV_ROUND_UP(offset_in_page(skb->data) +
587 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
591 size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
592 offset = skb_shinfo(skb)->frags[i].page_offset;
594 /* For a worse-case estimate we need to factor in
595 * the fragment page offset as this will affect the
596 * number of times xenvif_gop_frag_copy() will
597 * call start_new_rx_buffer().
599 max_slots_needed += DIV_ROUND_UP(offset + size,
603 /* To avoid the estimate becoming too pessimal for some
604 * frontends that limit posted rx requests, cap the estimate
605 * at MAX_SKB_FRAGS. In this case netback will fully coalesce
606 * the skb into the provided slots.
608 if (max_slots_needed > MAX_SKB_FRAGS) {
609 max_slots_needed = MAX_SKB_FRAGS;
610 XENVIF_RX_CB(skb)->full_coalesce = true;
612 XENVIF_RX_CB(skb)->full_coalesce = false;
615 /* We may need one more slot for GSO metadata */
616 if (skb_is_gso(skb) &&
617 (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4 ||
618 skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6))
621 old_req_cons = queue->rx.req_cons;
622 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
623 ring_slots_used = queue->rx.req_cons - old_req_cons;
625 BUG_ON(ring_slots_used > max_slots_needed);
627 __skb_queue_tail(&rxq, skb);
630 BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
635 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
636 gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
638 while ((skb = __skb_dequeue(&rxq)) != NULL) {
640 if ((1 << queue->meta[npo.meta_cons].gso_type) &
641 queue->vif->gso_prefix_mask) {
642 resp = RING_GET_RESPONSE(&queue->rx,
643 queue->rx.rsp_prod_pvt++);
645 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
647 resp->offset = queue->meta[npo.meta_cons].gso_size;
648 resp->id = queue->meta[npo.meta_cons].id;
649 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
652 XENVIF_RX_CB(skb)->meta_slots_used--;
656 queue->stats.tx_bytes += skb->len;
657 queue->stats.tx_packets++;
659 status = xenvif_check_gop(queue->vif,
660 XENVIF_RX_CB(skb)->meta_slots_used,
663 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
666 flags = XEN_NETRXF_more_data;
668 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
669 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
670 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
671 /* remote but checksummed. */
672 flags |= XEN_NETRXF_data_validated;
675 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
677 queue->meta[npo.meta_cons].size,
680 if ((1 << queue->meta[npo.meta_cons].gso_type) &
681 queue->vif->gso_mask) {
682 struct xen_netif_extra_info *gso =
683 (struct xen_netif_extra_info *)
684 RING_GET_RESPONSE(&queue->rx,
685 queue->rx.rsp_prod_pvt++);
687 resp->flags |= XEN_NETRXF_extra_info;
689 gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
690 gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
692 gso->u.gso.features = 0;
694 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
698 xenvif_add_frag_responses(queue, status,
699 queue->meta + npo.meta_cons + 1,
700 XENVIF_RX_CB(skb)->meta_slots_used);
702 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
704 need_to_notify |= !!ret;
706 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
712 notify_remote_via_irq(queue->rx_irq);
715 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
719 RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
722 napi_schedule(&queue->napi);
725 static void tx_add_credit(struct xenvif_queue *queue)
727 unsigned long max_burst, max_credit;
730 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
731 * Otherwise the interface can seize up due to insufficient credit.
733 max_burst = RING_GET_REQUEST(&queue->tx, queue->tx.req_cons)->size;
734 max_burst = min(max_burst, 131072UL);
735 max_burst = max(max_burst, queue->credit_bytes);
737 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
738 max_credit = queue->remaining_credit + queue->credit_bytes;
739 if (max_credit < queue->remaining_credit)
740 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
742 queue->remaining_credit = min(max_credit, max_burst);
745 static void tx_credit_callback(unsigned long data)
747 struct xenvif_queue *queue = (struct xenvif_queue *)data;
748 tx_add_credit(queue);
749 xenvif_napi_schedule_or_enable_events(queue);
752 static void xenvif_tx_err(struct xenvif_queue *queue,
753 struct xen_netif_tx_request *txp, RING_IDX end)
755 RING_IDX cons = queue->tx.req_cons;
759 spin_lock_irqsave(&queue->response_lock, flags);
760 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
761 spin_unlock_irqrestore(&queue->response_lock, flags);
764 txp = RING_GET_REQUEST(&queue->tx, cons++);
766 queue->tx.req_cons = cons;
769 static void xenvif_fatal_tx_err(struct xenvif *vif)
771 netdev_err(vif->dev, "fatal error; disabling device\n");
772 vif->disabled = true;
773 /* Disable the vif from queue 0's kthread */
775 xenvif_kick_thread(&vif->queues[0]);
778 static int xenvif_count_requests(struct xenvif_queue *queue,
779 struct xen_netif_tx_request *first,
780 struct xen_netif_tx_request *txp,
783 RING_IDX cons = queue->tx.req_cons;
788 if (!(first->flags & XEN_NETTXF_more_data))
792 struct xen_netif_tx_request dropped_tx = { 0 };
794 if (slots >= work_to_do) {
795 netdev_err(queue->vif->dev,
796 "Asked for %d slots but exceeds this limit\n",
798 xenvif_fatal_tx_err(queue->vif);
802 /* This guest is really using too many slots and
803 * considered malicious.
805 if (unlikely(slots >= fatal_skb_slots)) {
806 netdev_err(queue->vif->dev,
807 "Malicious frontend using %d slots, threshold %u\n",
808 slots, fatal_skb_slots);
809 xenvif_fatal_tx_err(queue->vif);
813 /* Xen network protocol had implicit dependency on
814 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
815 * the historical MAX_SKB_FRAGS value 18 to honor the
816 * same behavior as before. Any packet using more than
817 * 18 slots but less than fatal_skb_slots slots is
820 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
822 netdev_dbg(queue->vif->dev,
823 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
824 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
831 memcpy(txp, RING_GET_REQUEST(&queue->tx, cons + slots),
834 /* If the guest submitted a frame >= 64 KiB then
835 * first->size overflowed and following slots will
836 * appear to be larger than the frame.
838 * This cannot be fatal error as there are buggy
839 * frontends that do this.
841 * Consume all slots and drop the packet.
843 if (!drop_err && txp->size > first->size) {
845 netdev_dbg(queue->vif->dev,
846 "Invalid tx request, slot size %u > remaining size %u\n",
847 txp->size, first->size);
851 first->size -= txp->size;
854 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
855 netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
856 txp->offset, txp->size);
857 xenvif_fatal_tx_err(queue->vif);
861 more_data = txp->flags & XEN_NETTXF_more_data;
869 xenvif_tx_err(queue, first, cons + slots);
877 struct xenvif_tx_cb {
881 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
883 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
885 struct xen_netif_tx_request *txp,
886 struct gnttab_map_grant_ref *mop)
888 queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
889 gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
890 GNTMAP_host_map | GNTMAP_readonly,
891 txp->gref, queue->vif->domid);
893 memcpy(&queue->pending_tx_info[pending_idx].req, txp,
897 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
899 struct sk_buff *skb =
900 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
901 GFP_ATOMIC | __GFP_NOWARN);
902 if (unlikely(skb == NULL))
905 /* Packets passed to netif_rx() must have some headroom. */
906 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
908 /* Initialize it here to avoid later surprises */
909 skb_shinfo(skb)->destructor_arg = NULL;
914 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
916 struct xen_netif_tx_request *txp,
917 struct gnttab_map_grant_ref *gop)
919 struct skb_shared_info *shinfo = skb_shinfo(skb);
920 skb_frag_t *frags = shinfo->frags;
921 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
923 pending_ring_idx_t index;
924 unsigned int nr_slots, frag_overflow = 0;
926 /* At this point shinfo->nr_frags is in fact the number of
927 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
929 if (shinfo->nr_frags > MAX_SKB_FRAGS) {
930 frag_overflow = shinfo->nr_frags - MAX_SKB_FRAGS;
931 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
932 shinfo->nr_frags = MAX_SKB_FRAGS;
934 nr_slots = shinfo->nr_frags;
936 /* Skip first skb fragment if it is on same page as header fragment. */
937 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
939 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
940 shinfo->nr_frags++, txp++, gop++) {
941 index = pending_index(queue->pending_cons++);
942 pending_idx = queue->pending_ring[index];
943 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
944 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
948 struct sk_buff *nskb = xenvif_alloc_skb(0);
949 if (unlikely(nskb == NULL)) {
951 netdev_err(queue->vif->dev,
952 "Can't allocate the frag_list skb.\n");
956 shinfo = skb_shinfo(nskb);
957 frags = shinfo->frags;
959 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
960 shinfo->nr_frags++, txp++, gop++) {
961 index = pending_index(queue->pending_cons++);
962 pending_idx = queue->pending_ring[index];
963 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
964 frag_set_pending_idx(&frags[shinfo->nr_frags],
968 skb_shinfo(skb)->frag_list = nskb;
974 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
976 grant_handle_t handle)
978 if (unlikely(queue->grant_tx_handle[pending_idx] !=
979 NETBACK_INVALID_HANDLE)) {
980 netdev_err(queue->vif->dev,
981 "Trying to overwrite active handle! pending_idx: %x\n",
985 queue->grant_tx_handle[pending_idx] = handle;
988 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
991 if (unlikely(queue->grant_tx_handle[pending_idx] ==
992 NETBACK_INVALID_HANDLE)) {
993 netdev_err(queue->vif->dev,
994 "Trying to unmap invalid handle! pending_idx: %x\n",
998 queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
1001 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
1002 struct sk_buff *skb,
1003 struct gnttab_map_grant_ref **gopp_map,
1004 struct gnttab_copy **gopp_copy)
1006 struct gnttab_map_grant_ref *gop_map = *gopp_map;
1007 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1008 /* This always points to the shinfo of the skb being checked, which
1009 * could be either the first or the one on the frag_list
1011 struct skb_shared_info *shinfo = skb_shinfo(skb);
1012 /* If this is non-NULL, we are currently checking the frag_list skb, and
1013 * this points to the shinfo of the first one
1015 struct skb_shared_info *first_shinfo = NULL;
1016 int nr_frags = shinfo->nr_frags;
1017 const bool sharedslot = nr_frags &&
1018 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
1021 /* Check status of header. */
1022 err = (*gopp_copy)->status;
1023 if (unlikely(err)) {
1024 if (net_ratelimit())
1025 netdev_dbg(queue->vif->dev,
1026 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
1027 (*gopp_copy)->status,
1029 (*gopp_copy)->source.u.ref);
1030 /* The first frag might still have this slot mapped */
1032 xenvif_idx_release(queue, pending_idx,
1033 XEN_NETIF_RSP_ERROR);
1038 for (i = 0; i < nr_frags; i++, gop_map++) {
1041 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
1043 /* Check error status: if okay then remember grant handle. */
1044 newerr = gop_map->status;
1046 if (likely(!newerr)) {
1047 xenvif_grant_handle_set(queue,
1050 /* Had a previous error? Invalidate this fragment. */
1051 if (unlikely(err)) {
1052 xenvif_idx_unmap(queue, pending_idx);
1053 /* If the mapping of the first frag was OK, but
1054 * the header's copy failed, and they are
1055 * sharing a slot, send an error
1057 if (i == 0 && sharedslot)
1058 xenvif_idx_release(queue, pending_idx,
1059 XEN_NETIF_RSP_ERROR);
1061 xenvif_idx_release(queue, pending_idx,
1062 XEN_NETIF_RSP_OKAY);
1067 /* Error on this fragment: respond to client with an error. */
1068 if (net_ratelimit())
1069 netdev_dbg(queue->vif->dev,
1070 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1076 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
1078 /* Not the first error? Preceding frags already invalidated. */
1082 /* First error: if the header haven't shared a slot with the
1083 * first frag, release it as well.
1086 xenvif_idx_release(queue,
1087 XENVIF_TX_CB(skb)->pending_idx,
1088 XEN_NETIF_RSP_OKAY);
1090 /* Invalidate preceding fragments of this skb. */
1091 for (j = 0; j < i; j++) {
1092 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1093 xenvif_idx_unmap(queue, pending_idx);
1094 xenvif_idx_release(queue, pending_idx,
1095 XEN_NETIF_RSP_OKAY);
1098 /* And if we found the error while checking the frag_list, unmap
1099 * the first skb's frags
1102 for (j = 0; j < first_shinfo->nr_frags; j++) {
1103 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1104 xenvif_idx_unmap(queue, pending_idx);
1105 xenvif_idx_release(queue, pending_idx,
1106 XEN_NETIF_RSP_OKAY);
1110 /* Remember the error: invalidate all subsequent fragments. */
1114 if (skb_has_frag_list(skb) && !first_shinfo) {
1115 first_shinfo = skb_shinfo(skb);
1116 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1117 nr_frags = shinfo->nr_frags;
1122 *gopp_map = gop_map;
1126 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1128 struct skb_shared_info *shinfo = skb_shinfo(skb);
1129 int nr_frags = shinfo->nr_frags;
1131 u16 prev_pending_idx = INVALID_PENDING_IDX;
1133 for (i = 0; i < nr_frags; i++) {
1134 skb_frag_t *frag = shinfo->frags + i;
1135 struct xen_netif_tx_request *txp;
1139 pending_idx = frag_get_pending_idx(frag);
1141 /* If this is not the first frag, chain it to the previous*/
1142 if (prev_pending_idx == INVALID_PENDING_IDX)
1143 skb_shinfo(skb)->destructor_arg =
1144 &callback_param(queue, pending_idx);
1146 callback_param(queue, prev_pending_idx).ctx =
1147 &callback_param(queue, pending_idx);
1149 callback_param(queue, pending_idx).ctx = NULL;
1150 prev_pending_idx = pending_idx;
1152 txp = &queue->pending_tx_info[pending_idx].req;
1153 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1154 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1155 skb->len += txp->size;
1156 skb->data_len += txp->size;
1157 skb->truesize += txp->size;
1159 /* Take an extra reference to offset network stack's put_page */
1160 get_page(queue->mmap_pages[pending_idx]);
1164 static int xenvif_get_extras(struct xenvif_queue *queue,
1165 struct xen_netif_extra_info *extras,
1168 struct xen_netif_extra_info extra;
1169 RING_IDX cons = queue->tx.req_cons;
1172 if (unlikely(work_to_do-- <= 0)) {
1173 netdev_err(queue->vif->dev, "Missing extra info\n");
1174 xenvif_fatal_tx_err(queue->vif);
1178 memcpy(&extra, RING_GET_REQUEST(&queue->tx, cons),
1180 if (unlikely(!extra.type ||
1181 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1182 queue->tx.req_cons = ++cons;
1183 netdev_err(queue->vif->dev,
1184 "Invalid extra type: %d\n", extra.type);
1185 xenvif_fatal_tx_err(queue->vif);
1189 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1190 queue->tx.req_cons = ++cons;
1191 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1196 static int xenvif_set_skb_gso(struct xenvif *vif,
1197 struct sk_buff *skb,
1198 struct xen_netif_extra_info *gso)
1200 if (!gso->u.gso.size) {
1201 netdev_err(vif->dev, "GSO size must not be zero.\n");
1202 xenvif_fatal_tx_err(vif);
1206 switch (gso->u.gso.type) {
1207 case XEN_NETIF_GSO_TYPE_TCPV4:
1208 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1210 case XEN_NETIF_GSO_TYPE_TCPV6:
1211 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1214 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1215 xenvif_fatal_tx_err(vif);
1219 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1220 /* gso_segs will be calculated later */
1225 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1227 bool recalculate_partial_csum = false;
1229 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1230 * peers can fail to set NETRXF_csum_blank when sending a GSO
1231 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1232 * recalculate the partial checksum.
1234 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1235 queue->stats.rx_gso_checksum_fixup++;
1236 skb->ip_summed = CHECKSUM_PARTIAL;
1237 recalculate_partial_csum = true;
1240 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1241 if (skb->ip_summed != CHECKSUM_PARTIAL)
1244 return skb_checksum_setup(skb, recalculate_partial_csum);
1247 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1249 u64 now = get_jiffies_64();
1250 u64 next_credit = queue->credit_window_start +
1251 msecs_to_jiffies(queue->credit_usec / 1000);
1253 /* Timer could already be pending in rare cases. */
1254 if (timer_pending(&queue->credit_timeout))
1257 /* Passed the point where we can replenish credit? */
1258 if (time_after_eq64(now, next_credit)) {
1259 queue->credit_window_start = now;
1260 tx_add_credit(queue);
1263 /* Still too big to send right now? Set a callback. */
1264 if (size > queue->remaining_credit) {
1265 queue->credit_timeout.data =
1266 (unsigned long)queue;
1267 queue->credit_timeout.function =
1269 mod_timer(&queue->credit_timeout,
1271 queue->credit_window_start = next_credit;
1279 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1284 struct gnttab_map_grant_ref *gop = queue->tx_map_ops, *request_gop;
1285 struct sk_buff *skb;
1288 while (skb_queue_len(&queue->tx_queue) < budget) {
1289 struct xen_netif_tx_request txreq;
1290 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1291 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1295 unsigned int data_len;
1296 pending_ring_idx_t index;
1298 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1299 XEN_NETIF_TX_RING_SIZE) {
1300 netdev_err(queue->vif->dev,
1301 "Impossible number of requests. "
1302 "req_prod %d, req_cons %d, size %ld\n",
1303 queue->tx.sring->req_prod, queue->tx.req_cons,
1304 XEN_NETIF_TX_RING_SIZE);
1305 xenvif_fatal_tx_err(queue->vif);
1309 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1313 idx = queue->tx.req_cons;
1314 rmb(); /* Ensure that we see the request before we copy it. */
1315 memcpy(&txreq, RING_GET_REQUEST(&queue->tx, idx), sizeof(txreq));
1317 /* Credit-based scheduling. */
1318 if (txreq.size > queue->remaining_credit &&
1319 tx_credit_exceeded(queue, txreq.size))
1322 queue->remaining_credit -= txreq.size;
1325 queue->tx.req_cons = ++idx;
1327 memset(extras, 0, sizeof(extras));
1328 if (txreq.flags & XEN_NETTXF_extra_info) {
1329 work_to_do = xenvif_get_extras(queue, extras,
1331 idx = queue->tx.req_cons;
1332 if (unlikely(work_to_do < 0))
1336 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1337 if (unlikely(ret < 0))
1342 if (unlikely(txreq.size < ETH_HLEN)) {
1343 netdev_dbg(queue->vif->dev,
1344 "Bad packet size: %d\n", txreq.size);
1345 xenvif_tx_err(queue, &txreq, idx);
1349 /* No crossing a page as the payload mustn't fragment. */
1350 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1351 netdev_err(queue->vif->dev,
1352 "txreq.offset: %x, size: %u, end: %lu\n",
1353 txreq.offset, txreq.size,
1354 (txreq.offset&~PAGE_MASK) + txreq.size);
1355 xenvif_fatal_tx_err(queue->vif);
1359 index = pending_index(queue->pending_cons);
1360 pending_idx = queue->pending_ring[index];
1362 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1363 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1364 XEN_NETBACK_TX_COPY_LEN : txreq.size;
1366 skb = xenvif_alloc_skb(data_len);
1367 if (unlikely(skb == NULL)) {
1368 netdev_dbg(queue->vif->dev,
1369 "Can't allocate a skb in start_xmit.\n");
1370 xenvif_tx_err(queue, &txreq, idx);
1374 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1375 struct xen_netif_extra_info *gso;
1376 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1378 if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1379 /* Failure in xenvif_set_skb_gso is fatal. */
1385 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1387 __skb_put(skb, data_len);
1388 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1389 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1390 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1392 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1393 virt_to_mfn(skb->data);
1394 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1395 queue->tx_copy_ops[*copy_ops].dest.offset =
1396 offset_in_page(skb->data);
1398 queue->tx_copy_ops[*copy_ops].len = data_len;
1399 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1403 skb_shinfo(skb)->nr_frags = ret;
1404 if (data_len < txreq.size) {
1405 skb_shinfo(skb)->nr_frags++;
1406 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1408 xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1411 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1412 INVALID_PENDING_IDX);
1413 memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1417 queue->pending_cons++;
1419 request_gop = xenvif_get_requests(queue, skb, txfrags, gop);
1420 if (request_gop == NULL) {
1422 xenvif_tx_err(queue, &txreq, idx);
1427 __skb_queue_tail(&queue->tx_queue, skb);
1429 queue->tx.req_cons = idx;
1431 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1432 (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1436 (*map_ops) = gop - queue->tx_map_ops;
1440 /* Consolidate skb with a frag_list into a brand new one with local pages on
1441 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1443 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1445 unsigned int offset = skb_headlen(skb);
1446 skb_frag_t frags[MAX_SKB_FRAGS];
1448 struct ubuf_info *uarg;
1449 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1451 queue->stats.tx_zerocopy_sent += 2;
1452 queue->stats.tx_frag_overflow++;
1454 xenvif_fill_frags(queue, nskb);
1455 /* Subtract frags size, we will correct it later */
1456 skb->truesize -= skb->data_len;
1457 skb->len += nskb->len;
1458 skb->data_len += nskb->len;
1460 /* create a brand new frags array and coalesce there */
1461 for (i = 0; offset < skb->len; i++) {
1465 BUG_ON(i >= MAX_SKB_FRAGS);
1466 page = alloc_page(GFP_ATOMIC);
1469 skb->truesize += skb->data_len;
1470 for (j = 0; j < i; j++)
1471 put_page(frags[j].page.p);
1475 if (offset + PAGE_SIZE < skb->len)
1478 len = skb->len - offset;
1479 if (skb_copy_bits(skb, offset, page_address(page), len))
1483 frags[i].page.p = page;
1484 frags[i].page_offset = 0;
1485 skb_frag_size_set(&frags[i], len);
1487 /* swap out with old one */
1488 memcpy(skb_shinfo(skb)->frags,
1490 i * sizeof(skb_frag_t));
1491 skb_shinfo(skb)->nr_frags = i;
1492 skb->truesize += i * PAGE_SIZE;
1494 /* remove traces of mapped pages and frag_list */
1495 skb_frag_list_init(skb);
1496 uarg = skb_shinfo(skb)->destructor_arg;
1497 /* increase inflight counter to offset decrement in callback */
1498 atomic_inc(&queue->inflight_packets);
1499 uarg->callback(uarg, true);
1500 skb_shinfo(skb)->destructor_arg = NULL;
1502 xenvif_skb_zerocopy_prepare(queue, nskb);
1508 static int xenvif_tx_submit(struct xenvif_queue *queue)
1510 struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1511 struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1512 struct sk_buff *skb;
1515 while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1516 struct xen_netif_tx_request *txp;
1520 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1521 txp = &queue->pending_tx_info[pending_idx].req;
1523 /* Check the remap error code. */
1524 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1525 /* If there was an error, xenvif_tx_check_gop is
1526 * expected to release all the frags which were mapped,
1527 * so kfree_skb shouldn't do it again
1529 skb_shinfo(skb)->nr_frags = 0;
1530 if (skb_has_frag_list(skb)) {
1531 struct sk_buff *nskb =
1532 skb_shinfo(skb)->frag_list;
1533 skb_shinfo(nskb)->nr_frags = 0;
1539 data_len = skb->len;
1540 callback_param(queue, pending_idx).ctx = NULL;
1541 if (data_len < txp->size) {
1542 /* Append the packet payload as a fragment. */
1543 txp->offset += data_len;
1544 txp->size -= data_len;
1546 /* Schedule a response immediately. */
1547 xenvif_idx_release(queue, pending_idx,
1548 XEN_NETIF_RSP_OKAY);
1551 if (txp->flags & XEN_NETTXF_csum_blank)
1552 skb->ip_summed = CHECKSUM_PARTIAL;
1553 else if (txp->flags & XEN_NETTXF_data_validated)
1554 skb->ip_summed = CHECKSUM_UNNECESSARY;
1556 xenvif_fill_frags(queue, skb);
1558 if (unlikely(skb_has_frag_list(skb))) {
1559 if (xenvif_handle_frag_list(queue, skb)) {
1560 if (net_ratelimit())
1561 netdev_err(queue->vif->dev,
1562 "Not enough memory to consolidate frag_list!\n");
1563 xenvif_skb_zerocopy_prepare(queue, skb);
1569 skb->dev = queue->vif->dev;
1570 skb->protocol = eth_type_trans(skb, skb->dev);
1571 skb_reset_network_header(skb);
1573 if (checksum_setup(queue, skb)) {
1574 netdev_dbg(queue->vif->dev,
1575 "Can't setup checksum in net_tx_action\n");
1576 /* We have to set this flag to trigger the callback */
1577 if (skb_shinfo(skb)->destructor_arg)
1578 xenvif_skb_zerocopy_prepare(queue, skb);
1583 skb_probe_transport_header(skb, 0);
1585 /* If the packet is GSO then we will have just set up the
1586 * transport header offset in checksum_setup so it's now
1587 * straightforward to calculate gso_segs.
1589 if (skb_is_gso(skb)) {
1590 int mss = skb_shinfo(skb)->gso_size;
1591 int hdrlen = skb_transport_header(skb) -
1592 skb_mac_header(skb) +
1595 skb_shinfo(skb)->gso_segs =
1596 DIV_ROUND_UP(skb->len - hdrlen, mss);
1599 queue->stats.rx_bytes += skb->len;
1600 queue->stats.rx_packets++;
1604 /* Set this flag right before netif_receive_skb, otherwise
1605 * someone might think this packet already left netback, and
1606 * do a skb_copy_ubufs while we are still in control of the
1607 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1609 if (skb_shinfo(skb)->destructor_arg) {
1610 xenvif_skb_zerocopy_prepare(queue, skb);
1611 queue->stats.tx_zerocopy_sent++;
1614 netif_receive_skb(skb);
1620 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1622 unsigned long flags;
1623 pending_ring_idx_t index;
1624 struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1626 /* This is the only place where we grab this lock, to protect callbacks
1629 spin_lock_irqsave(&queue->callback_lock, flags);
1631 u16 pending_idx = ubuf->desc;
1632 ubuf = (struct ubuf_info *) ubuf->ctx;
1633 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1635 index = pending_index(queue->dealloc_prod);
1636 queue->dealloc_ring[index] = pending_idx;
1637 /* Sync with xenvif_tx_dealloc_action:
1638 * insert idx then incr producer.
1641 queue->dealloc_prod++;
1643 wake_up(&queue->dealloc_wq);
1644 spin_unlock_irqrestore(&queue->callback_lock, flags);
1646 if (likely(zerocopy_success))
1647 queue->stats.tx_zerocopy_success++;
1649 queue->stats.tx_zerocopy_fail++;
1650 xenvif_skb_zerocopy_complete(queue);
1653 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1655 struct gnttab_unmap_grant_ref *gop;
1656 pending_ring_idx_t dc, dp;
1657 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1660 dc = queue->dealloc_cons;
1661 gop = queue->tx_unmap_ops;
1663 /* Free up any grants we have finished using */
1665 dp = queue->dealloc_prod;
1667 /* Ensure we see all indices enqueued by all
1668 * xenvif_zerocopy_callback().
1673 BUG_ON(gop - queue->tx_unmap_ops > MAX_PENDING_REQS);
1675 queue->dealloc_ring[pending_index(dc++)];
1677 pending_idx_release[gop-queue->tx_unmap_ops] =
1679 queue->pages_to_unmap[gop-queue->tx_unmap_ops] =
1680 queue->mmap_pages[pending_idx];
1681 gnttab_set_unmap_op(gop,
1682 idx_to_kaddr(queue, pending_idx),
1684 queue->grant_tx_handle[pending_idx]);
1685 xenvif_grant_handle_reset(queue, pending_idx);
1689 } while (dp != queue->dealloc_prod);
1691 queue->dealloc_cons = dc;
1693 if (gop - queue->tx_unmap_ops > 0) {
1695 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1697 queue->pages_to_unmap,
1698 gop - queue->tx_unmap_ops);
1700 netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tx ret %d\n",
1701 gop - queue->tx_unmap_ops, ret);
1702 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1703 if (gop[i].status != GNTST_okay)
1704 netdev_err(queue->vif->dev,
1705 " host_addr: %llx handle: %x status: %d\n",
1714 for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1715 xenvif_idx_release(queue, pending_idx_release[i],
1716 XEN_NETIF_RSP_OKAY);
1720 /* Called after netfront has transmitted */
1721 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1723 unsigned nr_mops, nr_cops = 0;
1726 if (unlikely(!tx_work_todo(queue)))
1729 xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1734 gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1736 ret = gnttab_map_refs(queue->tx_map_ops,
1738 queue->pages_to_map,
1743 work_done = xenvif_tx_submit(queue);
1748 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1751 struct pending_tx_info *pending_tx_info;
1752 pending_ring_idx_t index;
1753 unsigned long flags;
1755 pending_tx_info = &queue->pending_tx_info[pending_idx];
1756 spin_lock_irqsave(&queue->response_lock, flags);
1757 make_tx_response(queue, &pending_tx_info->req, status);
1758 index = pending_index(queue->pending_prod);
1759 queue->pending_ring[index] = pending_idx;
1760 /* TX shouldn't use the index before we give it back here */
1762 queue->pending_prod++;
1763 spin_unlock_irqrestore(&queue->response_lock, flags);
1767 static void make_tx_response(struct xenvif_queue *queue,
1768 struct xen_netif_tx_request *txp,
1771 RING_IDX i = queue->tx.rsp_prod_pvt;
1772 struct xen_netif_tx_response *resp;
1775 resp = RING_GET_RESPONSE(&queue->tx, i);
1779 if (txp->flags & XEN_NETTXF_extra_info)
1780 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1782 queue->tx.rsp_prod_pvt = ++i;
1783 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1785 notify_remote_via_irq(queue->tx_irq);
1788 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1795 RING_IDX i = queue->rx.rsp_prod_pvt;
1796 struct xen_netif_rx_response *resp;
1798 resp = RING_GET_RESPONSE(&queue->rx, i);
1799 resp->offset = offset;
1800 resp->flags = flags;
1802 resp->status = (s16)size;
1804 resp->status = (s16)st;
1806 queue->rx.rsp_prod_pvt = ++i;
1811 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1814 struct gnttab_unmap_grant_ref tx_unmap_op;
1816 gnttab_set_unmap_op(&tx_unmap_op,
1817 idx_to_kaddr(queue, pending_idx),
1819 queue->grant_tx_handle[pending_idx]);
1820 xenvif_grant_handle_reset(queue, pending_idx);
1822 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1823 &queue->mmap_pages[pending_idx], 1);
1825 netdev_err(queue->vif->dev,
1826 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: %x status: %d\n",
1829 tx_unmap_op.host_addr,
1831 tx_unmap_op.status);
1836 static inline int tx_work_todo(struct xenvif_queue *queue)
1838 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1844 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1846 return queue->dealloc_cons != queue->dealloc_prod;
1849 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1851 if (queue->tx.sring)
1852 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1854 if (queue->rx.sring)
1855 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1859 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1860 grant_ref_t tx_ring_ref,
1861 grant_ref_t rx_ring_ref)
1864 struct xen_netif_tx_sring *txs;
1865 struct xen_netif_rx_sring *rxs;
1869 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1870 tx_ring_ref, &addr);
1874 txs = (struct xen_netif_tx_sring *)addr;
1875 BACK_RING_INIT(&queue->tx, txs, PAGE_SIZE);
1877 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1878 rx_ring_ref, &addr);
1882 rxs = (struct xen_netif_rx_sring *)addr;
1883 BACK_RING_INIT(&queue->rx, rxs, PAGE_SIZE);
1888 xenvif_unmap_frontend_rings(queue);
1892 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1894 struct xenvif *vif = queue->vif;
1896 queue->stalled = true;
1898 /* At least one queue has stalled? Disable the carrier. */
1899 spin_lock(&vif->lock);
1900 if (vif->stalled_queues++ == 0) {
1901 netdev_info(vif->dev, "Guest Rx stalled");
1902 netif_carrier_off(vif->dev);
1904 spin_unlock(&vif->lock);
1907 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1909 struct xenvif *vif = queue->vif;
1911 queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1912 queue->stalled = false;
1914 /* All queues are ready? Enable the carrier. */
1915 spin_lock(&vif->lock);
1916 if (--vif->stalled_queues == 0) {
1917 netdev_info(vif->dev, "Guest Rx ready");
1918 netif_carrier_on(vif->dev);
1920 spin_unlock(&vif->lock);
1923 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
1925 RING_IDX prod, cons;
1927 prod = queue->rx.sring->req_prod;
1928 cons = queue->rx.req_cons;
1930 return !queue->stalled
1931 && prod - cons < XEN_NETBK_RX_SLOTS_MAX
1932 && time_after(jiffies,
1933 queue->last_rx_time + queue->vif->stall_timeout);
1936 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
1938 RING_IDX prod, cons;
1940 prod = queue->rx.sring->req_prod;
1941 cons = queue->rx.req_cons;
1943 return queue->stalled
1944 && prod - cons >= XEN_NETBK_RX_SLOTS_MAX;
1947 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
1949 return (!skb_queue_empty(&queue->rx_queue)
1950 && xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX))
1951 || (queue->vif->stall_timeout &&
1952 (xenvif_rx_queue_stalled(queue)
1953 || xenvif_rx_queue_ready(queue)))
1954 || kthread_should_stop()
1955 || queue->vif->disabled;
1958 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
1960 struct sk_buff *skb;
1963 skb = skb_peek(&queue->rx_queue);
1965 return MAX_SCHEDULE_TIMEOUT;
1967 timeout = XENVIF_RX_CB(skb)->expires - jiffies;
1968 return timeout < 0 ? 0 : timeout;
1971 /* Wait until the guest Rx thread has work.
1973 * The timeout needs to be adjusted based on the current head of the
1974 * queue (and not just the head at the beginning). In particular, if
1975 * the queue is initially empty an infinite timeout is used and this
1976 * needs to be reduced when a skb is queued.
1978 * This cannot be done with wait_event_timeout() because it only
1979 * calculates the timeout once.
1981 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
1985 if (xenvif_have_rx_work(queue))
1991 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
1992 if (xenvif_have_rx_work(queue))
1994 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
1998 finish_wait(&queue->wq, &wait);
2001 int xenvif_kthread_guest_rx(void *data)
2003 struct xenvif_queue *queue = data;
2004 struct xenvif *vif = queue->vif;
2006 if (!vif->stall_timeout)
2007 xenvif_queue_carrier_on(queue);
2010 xenvif_wait_for_rx_work(queue);
2012 if (kthread_should_stop())
2015 /* This frontend is found to be rogue, disable it in
2016 * kthread context. Currently this is only set when
2017 * netback finds out frontend sends malformed packet,
2018 * but we cannot disable the interface in softirq
2019 * context so we defer it here, if this thread is
2020 * associated with queue 0.
2022 if (unlikely(vif->disabled && queue->id == 0)) {
2023 xenvif_carrier_off(vif);
2027 if (!skb_queue_empty(&queue->rx_queue))
2028 xenvif_rx_action(queue);
2030 /* If the guest hasn't provided any Rx slots for a
2031 * while it's probably not responsive, drop the
2032 * carrier so packets are dropped earlier.
2034 if (vif->stall_timeout) {
2035 if (xenvif_rx_queue_stalled(queue))
2036 xenvif_queue_carrier_off(queue);
2037 else if (xenvif_rx_queue_ready(queue))
2038 xenvif_queue_carrier_on(queue);
2041 /* Queued packets may have foreign pages from other
2042 * domains. These cannot be queued indefinitely as
2043 * this would starve guests of grant refs and transmit
2046 xenvif_rx_queue_drop_expired(queue);
2048 xenvif_rx_queue_maybe_wake(queue);
2053 /* Bin any remaining skbs */
2054 xenvif_rx_queue_purge(queue);
2059 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2061 /* Dealloc thread must remain running until all inflight
2064 return kthread_should_stop() &&
2065 !atomic_read(&queue->inflight_packets);
2068 int xenvif_dealloc_kthread(void *data)
2070 struct xenvif_queue *queue = data;
2073 wait_event_interruptible(queue->dealloc_wq,
2074 tx_dealloc_work_todo(queue) ||
2075 xenvif_dealloc_kthread_should_stop(queue));
2076 if (xenvif_dealloc_kthread_should_stop(queue))
2079 xenvif_tx_dealloc_action(queue);
2083 /* Unmap anything remaining*/
2084 if (tx_dealloc_work_todo(queue))
2085 xenvif_tx_dealloc_action(queue);
2090 static int __init netback_init(void)
2097 /* Allow as many queues as there are CPUs, by default */
2098 xenvif_max_queues = num_online_cpus();
2100 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2101 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2102 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2103 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2106 rc = xenvif_xenbus_init();
2110 #ifdef CONFIG_DEBUG_FS
2111 xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2112 if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2113 pr_warn("Init of debugfs returned %ld!\n",
2114 PTR_ERR(xen_netback_dbg_root));
2115 #endif /* CONFIG_DEBUG_FS */
2123 module_init(netback_init);
2125 static void __exit netback_fini(void)
2127 #ifdef CONFIG_DEBUG_FS
2128 if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2129 debugfs_remove_recursive(xen_netback_dbg_root);
2130 #endif /* CONFIG_DEBUG_FS */
2131 xenvif_xenbus_fini();
2133 module_exit(netback_fini);
2135 MODULE_LICENSE("Dual BSD/GPL");
2136 MODULE_ALIAS("xen-backend:vif");