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>
43 #include <xen/events.h>
44 #include <xen/interface/memory.h>
46 #include <asm/xen/hypercall.h>
47 #include <asm/xen/page.h>
49 struct pending_tx_info {
50 struct xen_netif_tx_request req;
53 typedef unsigned int pending_ring_idx_t;
55 struct netbk_rx_meta {
61 #define MAX_PENDING_REQS 256
63 /* Discriminate from any valid pending_idx value. */
64 #define INVALID_PENDING_IDX 0xFFFF
66 #define MAX_BUFFER_OFFSET PAGE_SIZE
68 /* extra field used in struct page */
71 #if BITS_PER_LONG < 64
73 #define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
74 unsigned int group:GROUP_WIDTH;
75 unsigned int idx:IDX_WIDTH;
77 unsigned int group, idx;
85 struct task_struct *task;
87 struct sk_buff_head rx_queue;
88 struct sk_buff_head tx_queue;
90 struct timer_list net_timer;
92 struct page *mmap_pages[MAX_PENDING_REQS];
94 pending_ring_idx_t pending_prod;
95 pending_ring_idx_t pending_cons;
96 struct list_head net_schedule_list;
98 /* Protect the net_schedule_list in netif. */
99 spinlock_t net_schedule_list_lock;
101 atomic_t netfront_count;
103 struct pending_tx_info pending_tx_info[MAX_PENDING_REQS];
104 struct gnttab_copy tx_copy_ops[MAX_PENDING_REQS];
106 u16 pending_ring[MAX_PENDING_REQS];
109 * Given MAX_BUFFER_OFFSET of 4096 the worst case is that each
110 * head/fragment page uses 2 copy operations because it
111 * straddles two buffers in the frontend.
113 struct gnttab_copy grant_copy_op[2*XEN_NETIF_RX_RING_SIZE];
114 struct netbk_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];
117 static struct xen_netbk *xen_netbk;
118 static int xen_netbk_group_nr;
120 void xen_netbk_add_xenvif(struct xenvif *vif)
123 int min_netfront_count;
125 struct xen_netbk *netbk;
127 min_netfront_count = atomic_read(&xen_netbk[0].netfront_count);
128 for (i = 0; i < xen_netbk_group_nr; i++) {
129 int netfront_count = atomic_read(&xen_netbk[i].netfront_count);
130 if (netfront_count < min_netfront_count) {
132 min_netfront_count = netfront_count;
136 netbk = &xen_netbk[min_group];
139 atomic_inc(&netbk->netfront_count);
142 void xen_netbk_remove_xenvif(struct xenvif *vif)
144 struct xen_netbk *netbk = vif->netbk;
146 atomic_dec(&netbk->netfront_count);
149 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx);
150 static void make_tx_response(struct xenvif *vif,
151 struct xen_netif_tx_request *txp,
153 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
160 static inline unsigned long idx_to_pfn(struct xen_netbk *netbk,
163 return page_to_pfn(netbk->mmap_pages[idx]);
166 static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
169 return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
172 /* extra field used in struct page */
173 static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
176 unsigned int group = netbk - xen_netbk;
177 union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
179 BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
180 pg->mapping = ext.mapping;
183 static int get_page_ext(struct page *pg,
184 unsigned int *pgroup, unsigned int *pidx)
186 union page_ext ext = { .mapping = pg->mapping };
187 struct xen_netbk *netbk;
188 unsigned int group, idx;
190 group = ext.e.group - 1;
192 if (group < 0 || group >= xen_netbk_group_nr)
195 netbk = &xen_netbk[group];
199 if ((idx < 0) || (idx >= MAX_PENDING_REQS))
202 if (netbk->mmap_pages[idx] != pg)
212 * This is the amount of packet we copy rather than map, so that the
213 * guest can't fiddle with the contents of the headers while we do
214 * packet processing on them (netfilter, routing, etc).
216 #define PKT_PROT_LEN (ETH_HLEN + \
218 sizeof(struct iphdr) + MAX_IPOPTLEN + \
219 sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
221 static u16 frag_get_pending_idx(skb_frag_t *frag)
223 return (u16)frag->page_offset;
226 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
228 frag->page_offset = pending_idx;
231 static inline pending_ring_idx_t pending_index(unsigned i)
233 return i & (MAX_PENDING_REQS-1);
236 static inline pending_ring_idx_t nr_pending_reqs(struct xen_netbk *netbk)
238 return MAX_PENDING_REQS -
239 netbk->pending_prod + netbk->pending_cons;
242 static void xen_netbk_kick_thread(struct xen_netbk *netbk)
247 static int max_required_rx_slots(struct xenvif *vif)
249 int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
251 if (vif->can_sg || vif->gso || vif->gso_prefix)
252 max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
257 int xen_netbk_rx_ring_full(struct xenvif *vif)
259 RING_IDX peek = vif->rx_req_cons_peek;
260 RING_IDX needed = max_required_rx_slots(vif);
262 return ((vif->rx.sring->req_prod - peek) < needed) ||
263 ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
266 int xen_netbk_must_stop_queue(struct xenvif *vif)
268 if (!xen_netbk_rx_ring_full(vif))
271 vif->rx.sring->req_event = vif->rx_req_cons_peek +
272 max_required_rx_slots(vif);
273 mb(); /* request notification /then/ check the queue */
275 return xen_netbk_rx_ring_full(vif);
279 * Returns true if we should start a new receive buffer instead of
280 * adding 'size' bytes to a buffer which currently contains 'offset'
283 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
285 /* simple case: we have completely filled the current buffer. */
286 if (offset == MAX_BUFFER_OFFSET)
290 * complex case: start a fresh buffer if the current frag
291 * would overflow the current buffer but only if:
292 * (i) this frag would fit completely in the next buffer
293 * and (ii) there is already some data in the current buffer
294 * and (iii) this is not the head buffer.
297 * - (i) stops us splitting a frag into two copies
298 * unless the frag is too large for a single buffer.
299 * - (ii) stops us from leaving a buffer pointlessly empty.
300 * - (iii) stops us leaving the first buffer
301 * empty. Strictly speaking this is already covered
302 * by (ii) but is explicitly checked because
303 * netfront relies on the first buffer being
304 * non-empty and can crash otherwise.
306 * This means we will effectively linearise small
307 * frags but do not needlessly split large buffers
308 * into multiple copies tend to give large frags their
309 * own buffers as before.
311 if ((offset + size > MAX_BUFFER_OFFSET) &&
312 (size <= MAX_BUFFER_OFFSET) && offset && !head)
319 * Figure out how many ring slots we're going to need to send @skb to
320 * the guest. This function is essentially a dry run of
321 * netbk_gop_frag_copy.
323 unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
328 count = DIV_ROUND_UP(skb_headlen(skb), PAGE_SIZE);
330 copy_off = skb_headlen(skb) % PAGE_SIZE;
332 if (skb_shinfo(skb)->gso_size)
335 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
336 unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
339 BUG_ON(copy_off > MAX_BUFFER_OFFSET);
341 if (start_new_rx_buffer(copy_off, size, 0)) {
347 if (copy_off + bytes > MAX_BUFFER_OFFSET)
348 bytes = MAX_BUFFER_OFFSET - copy_off;
357 struct netrx_pending_operations {
358 unsigned copy_prod, copy_cons;
359 unsigned meta_prod, meta_cons;
360 struct gnttab_copy *copy;
361 struct netbk_rx_meta *meta;
363 grant_ref_t copy_gref;
366 static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
367 struct netrx_pending_operations *npo)
369 struct netbk_rx_meta *meta;
370 struct xen_netif_rx_request *req;
372 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
374 meta = npo->meta + npo->meta_prod++;
380 npo->copy_gref = req->gref;
386 * Set up the grant operations for this fragment. If it's a flipping
387 * interface, we also set up the unmap request from here.
389 static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
390 struct netrx_pending_operations *npo,
391 struct page *page, unsigned long size,
392 unsigned long offset, int *head)
394 struct gnttab_copy *copy_gop;
395 struct netbk_rx_meta *meta;
397 * These variables are used iff get_page_ext returns true,
398 * in which case they are guaranteed to be initialized.
400 unsigned int uninitialized_var(group), uninitialized_var(idx);
401 int foreign = get_page_ext(page, &group, &idx);
404 /* Data must not cross a page boundary. */
405 BUG_ON(size + offset > PAGE_SIZE);
407 meta = npo->meta + npo->meta_prod - 1;
410 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
412 if (start_new_rx_buffer(npo->copy_off, size, *head)) {
414 * Netfront requires there to be some data in the head
419 meta = get_next_rx_buffer(vif, npo);
423 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
424 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
426 copy_gop = npo->copy + npo->copy_prod++;
427 copy_gop->flags = GNTCOPY_dest_gref;
429 struct xen_netbk *netbk = &xen_netbk[group];
430 struct pending_tx_info *src_pend;
432 src_pend = &netbk->pending_tx_info[idx];
434 copy_gop->source.domid = src_pend->vif->domid;
435 copy_gop->source.u.ref = src_pend->req.gref;
436 copy_gop->flags |= GNTCOPY_source_gref;
438 void *vaddr = page_address(page);
439 copy_gop->source.domid = DOMID_SELF;
440 copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
442 copy_gop->source.offset = offset;
443 copy_gop->dest.domid = vif->domid;
445 copy_gop->dest.offset = npo->copy_off;
446 copy_gop->dest.u.ref = npo->copy_gref;
447 copy_gop->len = bytes;
449 npo->copy_off += bytes;
455 /* Leave a gap for the GSO descriptor. */
456 if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
459 *head = 0; /* There must be something in this buffer now. */
465 * Prepare an SKB to be transmitted to the frontend.
467 * This function is responsible for allocating grant operations, meta
470 * It returns the number of meta structures consumed. The number of
471 * ring slots used is always equal to the number of meta slots used
472 * plus the number of GSO descriptors used. Currently, we use either
473 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
474 * frontend-side LRO).
476 static int netbk_gop_skb(struct sk_buff *skb,
477 struct netrx_pending_operations *npo)
479 struct xenvif *vif = netdev_priv(skb->dev);
480 int nr_frags = skb_shinfo(skb)->nr_frags;
482 struct xen_netif_rx_request *req;
483 struct netbk_rx_meta *meta;
488 old_meta_prod = npo->meta_prod;
490 /* Set up a GSO prefix descriptor, if necessary */
491 if (skb_shinfo(skb)->gso_size && vif->gso_prefix) {
492 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
493 meta = npo->meta + npo->meta_prod++;
494 meta->gso_size = skb_shinfo(skb)->gso_size;
499 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
500 meta = npo->meta + npo->meta_prod++;
502 if (!vif->gso_prefix)
503 meta->gso_size = skb_shinfo(skb)->gso_size;
510 npo->copy_gref = req->gref;
513 while (data < skb_tail_pointer(skb)) {
514 unsigned int offset = offset_in_page(data);
515 unsigned int len = PAGE_SIZE - offset;
517 if (data + len > skb_tail_pointer(skb))
518 len = skb_tail_pointer(skb) - data;
520 netbk_gop_frag_copy(vif, skb, npo,
521 virt_to_page(data), len, offset, &head);
525 for (i = 0; i < nr_frags; i++) {
526 netbk_gop_frag_copy(vif, skb, npo,
527 skb_frag_page(&skb_shinfo(skb)->frags[i]),
528 skb_frag_size(&skb_shinfo(skb)->frags[i]),
529 skb_shinfo(skb)->frags[i].page_offset,
533 return npo->meta_prod - old_meta_prod;
537 * This is a twin to netbk_gop_skb. Assume that netbk_gop_skb was
538 * used to set up the operations on the top of
539 * netrx_pending_operations, which have since been done. Check that
540 * they didn't give any errors and advance over them.
542 static int netbk_check_gop(struct xenvif *vif, int nr_meta_slots,
543 struct netrx_pending_operations *npo)
545 struct gnttab_copy *copy_op;
546 int status = XEN_NETIF_RSP_OKAY;
549 for (i = 0; i < nr_meta_slots; i++) {
550 copy_op = npo->copy + npo->copy_cons++;
551 if (copy_op->status != GNTST_okay) {
553 "Bad status %d from copy to DOM%d.\n",
554 copy_op->status, vif->domid);
555 status = XEN_NETIF_RSP_ERROR;
562 static void netbk_add_frag_responses(struct xenvif *vif, int status,
563 struct netbk_rx_meta *meta,
567 unsigned long offset;
569 /* No fragments used */
570 if (nr_meta_slots <= 1)
575 for (i = 0; i < nr_meta_slots; i++) {
577 if (i == nr_meta_slots - 1)
580 flags = XEN_NETRXF_more_data;
583 make_rx_response(vif, meta[i].id, status, offset,
584 meta[i].size, flags);
588 struct skb_cb_overlay {
592 static void xen_netbk_rx_action(struct xen_netbk *netbk)
594 struct xenvif *vif = NULL, *tmp;
597 struct xen_netif_rx_response *resp;
598 struct sk_buff_head rxq;
604 unsigned long offset;
605 struct skb_cb_overlay *sco;
607 struct netrx_pending_operations npo = {
608 .copy = netbk->grant_copy_op,
612 skb_queue_head_init(&rxq);
616 while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
617 vif = netdev_priv(skb->dev);
618 nr_frags = skb_shinfo(skb)->nr_frags;
620 sco = (struct skb_cb_overlay *)skb->cb;
621 sco->meta_slots_used = netbk_gop_skb(skb, &npo);
623 count += nr_frags + 1;
625 __skb_queue_tail(&rxq, skb);
627 /* Filled the batch queue? */
628 if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
632 BUG_ON(npo.meta_prod > ARRAY_SIZE(netbk->meta));
637 BUG_ON(npo.copy_prod > ARRAY_SIZE(netbk->grant_copy_op));
638 gnttab_batch_copy(netbk->grant_copy_op, npo.copy_prod);
640 while ((skb = __skb_dequeue(&rxq)) != NULL) {
641 sco = (struct skb_cb_overlay *)skb->cb;
643 vif = netdev_priv(skb->dev);
645 if (netbk->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
646 resp = RING_GET_RESPONSE(&vif->rx,
647 vif->rx.rsp_prod_pvt++);
649 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
651 resp->offset = netbk->meta[npo.meta_cons].gso_size;
652 resp->id = netbk->meta[npo.meta_cons].id;
653 resp->status = sco->meta_slots_used;
656 sco->meta_slots_used--;
660 vif->dev->stats.tx_bytes += skb->len;
661 vif->dev->stats.tx_packets++;
663 status = netbk_check_gop(vif, sco->meta_slots_used, &npo);
665 if (sco->meta_slots_used == 1)
668 flags = XEN_NETRXF_more_data;
670 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
671 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
672 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
673 /* remote but checksummed. */
674 flags |= XEN_NETRXF_data_validated;
677 resp = make_rx_response(vif, netbk->meta[npo.meta_cons].id,
679 netbk->meta[npo.meta_cons].size,
682 if (netbk->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
683 struct xen_netif_extra_info *gso =
684 (struct xen_netif_extra_info *)
685 RING_GET_RESPONSE(&vif->rx,
686 vif->rx.rsp_prod_pvt++);
688 resp->flags |= XEN_NETRXF_extra_info;
690 gso->u.gso.size = netbk->meta[npo.meta_cons].gso_size;
691 gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
693 gso->u.gso.features = 0;
695 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
699 netbk_add_frag_responses(vif, status,
700 netbk->meta + npo.meta_cons + 1,
701 sco->meta_slots_used);
703 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
705 if (ret && list_empty(&vif->notify_list))
706 list_add_tail(&vif->notify_list, ¬ify);
708 xenvif_notify_tx_completion(vif);
711 npo.meta_cons += sco->meta_slots_used;
715 list_for_each_entry_safe(vif, tmp, ¬ify, notify_list) {
716 notify_remote_via_irq(vif->irq);
717 list_del_init(&vif->notify_list);
720 /* More work to do? */
721 if (!skb_queue_empty(&netbk->rx_queue) &&
722 !timer_pending(&netbk->net_timer))
723 xen_netbk_kick_thread(netbk);
726 void xen_netbk_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
728 struct xen_netbk *netbk = vif->netbk;
730 skb_queue_tail(&netbk->rx_queue, skb);
732 xen_netbk_kick_thread(netbk);
735 static void xen_netbk_alarm(unsigned long data)
737 struct xen_netbk *netbk = (struct xen_netbk *)data;
738 xen_netbk_kick_thread(netbk);
741 static int __on_net_schedule_list(struct xenvif *vif)
743 return !list_empty(&vif->schedule_list);
746 /* Must be called with net_schedule_list_lock held */
747 static void remove_from_net_schedule_list(struct xenvif *vif)
749 if (likely(__on_net_schedule_list(vif))) {
750 list_del_init(&vif->schedule_list);
755 static struct xenvif *poll_net_schedule_list(struct xen_netbk *netbk)
757 struct xenvif *vif = NULL;
759 spin_lock_irq(&netbk->net_schedule_list_lock);
760 if (list_empty(&netbk->net_schedule_list))
763 vif = list_first_entry(&netbk->net_schedule_list,
764 struct xenvif, schedule_list);
770 remove_from_net_schedule_list(vif);
772 spin_unlock_irq(&netbk->net_schedule_list_lock);
776 void xen_netbk_schedule_xenvif(struct xenvif *vif)
779 struct xen_netbk *netbk = vif->netbk;
781 if (__on_net_schedule_list(vif))
784 spin_lock_irqsave(&netbk->net_schedule_list_lock, flags);
785 if (!__on_net_schedule_list(vif) &&
786 likely(xenvif_schedulable(vif))) {
787 list_add_tail(&vif->schedule_list, &netbk->net_schedule_list);
790 spin_unlock_irqrestore(&netbk->net_schedule_list_lock, flags);
794 if ((nr_pending_reqs(netbk) < (MAX_PENDING_REQS/2)) &&
795 !list_empty(&netbk->net_schedule_list))
796 xen_netbk_kick_thread(netbk);
799 void xen_netbk_deschedule_xenvif(struct xenvif *vif)
801 struct xen_netbk *netbk = vif->netbk;
802 spin_lock_irq(&netbk->net_schedule_list_lock);
803 remove_from_net_schedule_list(vif);
804 spin_unlock_irq(&netbk->net_schedule_list_lock);
807 void xen_netbk_check_rx_xenvif(struct xenvif *vif)
811 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
814 xen_netbk_schedule_xenvif(vif);
817 static void tx_add_credit(struct xenvif *vif)
819 unsigned long max_burst, max_credit;
822 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
823 * Otherwise the interface can seize up due to insufficient credit.
825 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
826 max_burst = min(max_burst, 131072UL);
827 max_burst = max(max_burst, vif->credit_bytes);
829 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
830 max_credit = vif->remaining_credit + vif->credit_bytes;
831 if (max_credit < vif->remaining_credit)
832 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
834 vif->remaining_credit = min(max_credit, max_burst);
837 static void tx_credit_callback(unsigned long data)
839 struct xenvif *vif = (struct xenvif *)data;
841 xen_netbk_check_rx_xenvif(vif);
844 static void netbk_tx_err(struct xenvif *vif,
845 struct xen_netif_tx_request *txp, RING_IDX end)
847 RING_IDX cons = vif->tx.req_cons;
850 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
853 txp = RING_GET_REQUEST(&vif->tx, cons++);
855 vif->tx.req_cons = cons;
856 xen_netbk_check_rx_xenvif(vif);
860 static int netbk_count_requests(struct xenvif *vif,
861 struct xen_netif_tx_request *first,
862 struct xen_netif_tx_request *txp,
865 RING_IDX cons = vif->tx.req_cons;
868 if (!(first->flags & XEN_NETTXF_more_data))
872 if (frags >= work_to_do) {
873 netdev_dbg(vif->dev, "Need more frags\n");
877 if (unlikely(frags >= MAX_SKB_FRAGS)) {
878 netdev_dbg(vif->dev, "Too many frags\n");
882 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + frags),
884 if (txp->size > first->size) {
885 netdev_dbg(vif->dev, "Frags galore\n");
889 first->size -= txp->size;
892 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
893 netdev_dbg(vif->dev, "txp->offset: %x, size: %u\n",
894 txp->offset, txp->size);
897 } while ((txp++)->flags & XEN_NETTXF_more_data);
901 static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
906 page = alloc_page(GFP_KERNEL|__GFP_COLD);
909 set_page_ext(page, netbk, pending_idx);
910 netbk->mmap_pages[pending_idx] = page;
914 static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
917 struct xen_netif_tx_request *txp,
918 struct gnttab_copy *gop)
920 struct skb_shared_info *shinfo = skb_shinfo(skb);
921 skb_frag_t *frags = shinfo->frags;
922 u16 pending_idx = *((u16 *)skb->data);
925 /* Skip first skb fragment if it is on same page as header fragment. */
926 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
928 for (i = start; i < shinfo->nr_frags; i++, txp++) {
930 pending_ring_idx_t index;
931 struct pending_tx_info *pending_tx_info =
932 netbk->pending_tx_info;
934 index = pending_index(netbk->pending_cons++);
935 pending_idx = netbk->pending_ring[index];
936 page = xen_netbk_alloc_page(netbk, skb, pending_idx);
940 gop->source.u.ref = txp->gref;
941 gop->source.domid = vif->domid;
942 gop->source.offset = txp->offset;
944 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
945 gop->dest.domid = DOMID_SELF;
946 gop->dest.offset = txp->offset;
948 gop->len = txp->size;
949 gop->flags = GNTCOPY_source_gref;
953 memcpy(&pending_tx_info[pending_idx].req, txp, sizeof(*txp));
955 pending_tx_info[pending_idx].vif = vif;
956 frag_set_pending_idx(&frags[i], pending_idx);
962 static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
964 struct gnttab_copy **gopp)
966 struct gnttab_copy *gop = *gopp;
967 u16 pending_idx = *((u16 *)skb->data);
968 struct pending_tx_info *pending_tx_info = netbk->pending_tx_info;
969 struct xenvif *vif = pending_tx_info[pending_idx].vif;
970 struct xen_netif_tx_request *txp;
971 struct skb_shared_info *shinfo = skb_shinfo(skb);
972 int nr_frags = shinfo->nr_frags;
975 /* Check status of header. */
978 pending_ring_idx_t index;
979 index = pending_index(netbk->pending_prod++);
980 txp = &pending_tx_info[pending_idx].req;
981 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
982 netbk->pending_ring[index] = pending_idx;
986 /* Skip first skb fragment if it is on same page as header fragment. */
987 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
989 for (i = start; i < nr_frags; i++) {
991 pending_ring_idx_t index;
993 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
995 /* Check error status: if okay then remember grant handle. */
996 newerr = (++gop)->status;
997 if (likely(!newerr)) {
998 /* Had a previous error? Invalidate this fragment. */
1000 xen_netbk_idx_release(netbk, pending_idx);
1004 /* Error on this fragment: respond to client with an error. */
1005 txp = &netbk->pending_tx_info[pending_idx].req;
1006 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
1007 index = pending_index(netbk->pending_prod++);
1008 netbk->pending_ring[index] = pending_idx;
1011 /* Not the first error? Preceding frags already invalidated. */
1015 /* First error: invalidate header and preceding fragments. */
1016 pending_idx = *((u16 *)skb->data);
1017 xen_netbk_idx_release(netbk, pending_idx);
1018 for (j = start; j < i; j++) {
1019 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1020 xen_netbk_idx_release(netbk, pending_idx);
1023 /* Remember the error: invalidate all subsequent fragments. */
1031 static void xen_netbk_fill_frags(struct xen_netbk *netbk, struct sk_buff *skb)
1033 struct skb_shared_info *shinfo = skb_shinfo(skb);
1034 int nr_frags = shinfo->nr_frags;
1037 for (i = 0; i < nr_frags; i++) {
1038 skb_frag_t *frag = shinfo->frags + i;
1039 struct xen_netif_tx_request *txp;
1043 pending_idx = frag_get_pending_idx(frag);
1045 txp = &netbk->pending_tx_info[pending_idx].req;
1046 page = virt_to_page(idx_to_kaddr(netbk, pending_idx));
1047 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1048 skb->len += txp->size;
1049 skb->data_len += txp->size;
1050 skb->truesize += txp->size;
1052 /* Take an extra reference to offset xen_netbk_idx_release */
1053 get_page(netbk->mmap_pages[pending_idx]);
1054 xen_netbk_idx_release(netbk, pending_idx);
1058 static int xen_netbk_get_extras(struct xenvif *vif,
1059 struct xen_netif_extra_info *extras,
1062 struct xen_netif_extra_info extra;
1063 RING_IDX cons = vif->tx.req_cons;
1066 if (unlikely(work_to_do-- <= 0)) {
1067 netdev_dbg(vif->dev, "Missing extra info\n");
1071 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1073 if (unlikely(!extra.type ||
1074 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1075 vif->tx.req_cons = ++cons;
1076 netdev_dbg(vif->dev,
1077 "Invalid extra type: %d\n", extra.type);
1081 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1082 vif->tx.req_cons = ++cons;
1083 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1088 static int netbk_set_skb_gso(struct xenvif *vif,
1089 struct sk_buff *skb,
1090 struct xen_netif_extra_info *gso)
1092 if (!gso->u.gso.size) {
1093 netdev_dbg(vif->dev, "GSO size must not be zero.\n");
1097 /* Currently only TCPv4 S.O. is supported. */
1098 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
1099 netdev_dbg(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1103 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1104 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1106 /* Header must be checked, and gso_segs computed. */
1107 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1108 skb_shinfo(skb)->gso_segs = 0;
1113 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1118 int recalculate_partial_csum = 0;
1121 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1122 * peers can fail to set NETRXF_csum_blank when sending a GSO
1123 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1124 * recalculate the partial checksum.
1126 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1127 vif->rx_gso_checksum_fixup++;
1128 skb->ip_summed = CHECKSUM_PARTIAL;
1129 recalculate_partial_csum = 1;
1132 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1133 if (skb->ip_summed != CHECKSUM_PARTIAL)
1136 if (skb->protocol != htons(ETH_P_IP))
1139 iph = (void *)skb->data;
1140 th = skb->data + 4 * iph->ihl;
1141 if (th >= skb_tail_pointer(skb))
1144 skb->csum_start = th - skb->head;
1145 switch (iph->protocol) {
1147 skb->csum_offset = offsetof(struct tcphdr, check);
1149 if (recalculate_partial_csum) {
1150 struct tcphdr *tcph = (struct tcphdr *)th;
1151 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1152 skb->len - iph->ihl*4,
1157 skb->csum_offset = offsetof(struct udphdr, check);
1159 if (recalculate_partial_csum) {
1160 struct udphdr *udph = (struct udphdr *)th;
1161 udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1162 skb->len - iph->ihl*4,
1167 if (net_ratelimit())
1168 netdev_err(vif->dev,
1169 "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
1174 if ((th + skb->csum_offset + 2) > skb_tail_pointer(skb))
1183 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1185 unsigned long now = jiffies;
1186 unsigned long next_credit =
1187 vif->credit_timeout.expires +
1188 msecs_to_jiffies(vif->credit_usec / 1000);
1190 /* Timer could already be pending in rare cases. */
1191 if (timer_pending(&vif->credit_timeout))
1194 /* Passed the point where we can replenish credit? */
1195 if (time_after_eq(now, next_credit)) {
1196 vif->credit_timeout.expires = now;
1200 /* Still too big to send right now? Set a callback. */
1201 if (size > vif->remaining_credit) {
1202 vif->credit_timeout.data =
1204 vif->credit_timeout.function =
1206 mod_timer(&vif->credit_timeout,
1215 static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
1217 struct gnttab_copy *gop = netbk->tx_copy_ops, *request_gop;
1218 struct sk_buff *skb;
1221 while (((nr_pending_reqs(netbk) + MAX_SKB_FRAGS) < MAX_PENDING_REQS) &&
1222 !list_empty(&netbk->net_schedule_list)) {
1224 struct xen_netif_tx_request txreq;
1225 struct xen_netif_tx_request txfrags[MAX_SKB_FRAGS];
1227 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1231 unsigned int data_len;
1232 pending_ring_idx_t index;
1234 /* Get a netif from the list with work to do. */
1235 vif = poll_net_schedule_list(netbk);
1239 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
1245 idx = vif->tx.req_cons;
1246 rmb(); /* Ensure that we see the request before we copy it. */
1247 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1249 /* Credit-based scheduling. */
1250 if (txreq.size > vif->remaining_credit &&
1251 tx_credit_exceeded(vif, txreq.size)) {
1256 vif->remaining_credit -= txreq.size;
1259 vif->tx.req_cons = ++idx;
1261 memset(extras, 0, sizeof(extras));
1262 if (txreq.flags & XEN_NETTXF_extra_info) {
1263 work_to_do = xen_netbk_get_extras(vif, extras,
1265 idx = vif->tx.req_cons;
1266 if (unlikely(work_to_do < 0)) {
1267 netbk_tx_err(vif, &txreq, idx);
1272 ret = netbk_count_requests(vif, &txreq, txfrags, work_to_do);
1273 if (unlikely(ret < 0)) {
1274 netbk_tx_err(vif, &txreq, idx - ret);
1279 if (unlikely(txreq.size < ETH_HLEN)) {
1280 netdev_dbg(vif->dev,
1281 "Bad packet size: %d\n", txreq.size);
1282 netbk_tx_err(vif, &txreq, idx);
1286 /* No crossing a page as the payload mustn't fragment. */
1287 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1288 netdev_dbg(vif->dev,
1289 "txreq.offset: %x, size: %u, end: %lu\n",
1290 txreq.offset, txreq.size,
1291 (txreq.offset&~PAGE_MASK) + txreq.size);
1292 netbk_tx_err(vif, &txreq, idx);
1296 index = pending_index(netbk->pending_cons);
1297 pending_idx = netbk->pending_ring[index];
1299 data_len = (txreq.size > PKT_PROT_LEN &&
1300 ret < MAX_SKB_FRAGS) ?
1301 PKT_PROT_LEN : txreq.size;
1303 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1304 GFP_ATOMIC | __GFP_NOWARN);
1305 if (unlikely(skb == NULL)) {
1306 netdev_dbg(vif->dev,
1307 "Can't allocate a skb in start_xmit.\n");
1308 netbk_tx_err(vif, &txreq, idx);
1312 /* Packets passed to netif_rx() must have some headroom. */
1313 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1315 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1316 struct xen_netif_extra_info *gso;
1317 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1319 if (netbk_set_skb_gso(vif, skb, gso)) {
1321 netbk_tx_err(vif, &txreq, idx);
1326 /* XXX could copy straight to head */
1327 page = xen_netbk_alloc_page(netbk, skb, pending_idx);
1330 netbk_tx_err(vif, &txreq, idx);
1334 gop->source.u.ref = txreq.gref;
1335 gop->source.domid = vif->domid;
1336 gop->source.offset = txreq.offset;
1338 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1339 gop->dest.domid = DOMID_SELF;
1340 gop->dest.offset = txreq.offset;
1342 gop->len = txreq.size;
1343 gop->flags = GNTCOPY_source_gref;
1347 memcpy(&netbk->pending_tx_info[pending_idx].req,
1348 &txreq, sizeof(txreq));
1349 netbk->pending_tx_info[pending_idx].vif = vif;
1350 *((u16 *)skb->data) = pending_idx;
1352 __skb_put(skb, data_len);
1354 skb_shinfo(skb)->nr_frags = ret;
1355 if (data_len < txreq.size) {
1356 skb_shinfo(skb)->nr_frags++;
1357 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1360 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1361 INVALID_PENDING_IDX);
1364 netbk->pending_cons++;
1366 request_gop = xen_netbk_get_requests(netbk, vif,
1368 if (request_gop == NULL) {
1370 netbk_tx_err(vif, &txreq, idx);
1375 __skb_queue_tail(&netbk->tx_queue, skb);
1377 vif->tx.req_cons = idx;
1378 xen_netbk_check_rx_xenvif(vif);
1380 if ((gop-netbk->tx_copy_ops) >= ARRAY_SIZE(netbk->tx_copy_ops))
1384 return gop - netbk->tx_copy_ops;
1387 static void xen_netbk_tx_submit(struct xen_netbk *netbk)
1389 struct gnttab_copy *gop = netbk->tx_copy_ops;
1390 struct sk_buff *skb;
1392 while ((skb = __skb_dequeue(&netbk->tx_queue)) != NULL) {
1393 struct xen_netif_tx_request *txp;
1398 pending_idx = *((u16 *)skb->data);
1399 vif = netbk->pending_tx_info[pending_idx].vif;
1400 txp = &netbk->pending_tx_info[pending_idx].req;
1402 /* Check the remap error code. */
1403 if (unlikely(xen_netbk_tx_check_gop(netbk, skb, &gop))) {
1404 netdev_dbg(vif->dev, "netback grant failed.\n");
1405 skb_shinfo(skb)->nr_frags = 0;
1410 data_len = skb->len;
1412 (void *)(idx_to_kaddr(netbk, pending_idx)|txp->offset),
1414 if (data_len < txp->size) {
1415 /* Append the packet payload as a fragment. */
1416 txp->offset += data_len;
1417 txp->size -= data_len;
1419 /* Schedule a response immediately. */
1420 xen_netbk_idx_release(netbk, pending_idx);
1423 if (txp->flags & XEN_NETTXF_csum_blank)
1424 skb->ip_summed = CHECKSUM_PARTIAL;
1425 else if (txp->flags & XEN_NETTXF_data_validated)
1426 skb->ip_summed = CHECKSUM_UNNECESSARY;
1428 xen_netbk_fill_frags(netbk, skb);
1431 * If the initial fragment was < PKT_PROT_LEN then
1432 * pull through some bytes from the other fragments to
1433 * increase the linear region to PKT_PROT_LEN bytes.
1435 if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
1436 int target = min_t(int, skb->len, PKT_PROT_LEN);
1437 __pskb_pull_tail(skb, target - skb_headlen(skb));
1440 skb->dev = vif->dev;
1441 skb->protocol = eth_type_trans(skb, skb->dev);
1443 if (checksum_setup(vif, skb)) {
1444 netdev_dbg(vif->dev,
1445 "Can't setup checksum in net_tx_action\n");
1450 vif->dev->stats.rx_bytes += skb->len;
1451 vif->dev->stats.rx_packets++;
1453 xenvif_receive_skb(vif, skb);
1457 /* Called after netfront has transmitted */
1458 static void xen_netbk_tx_action(struct xen_netbk *netbk)
1462 nr_gops = xen_netbk_tx_build_gops(netbk);
1467 gnttab_batch_copy(netbk->tx_copy_ops, nr_gops);
1469 xen_netbk_tx_submit(netbk);
1472 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx)
1475 struct pending_tx_info *pending_tx_info;
1476 pending_ring_idx_t index;
1478 /* Already complete? */
1479 if (netbk->mmap_pages[pending_idx] == NULL)
1482 pending_tx_info = &netbk->pending_tx_info[pending_idx];
1484 vif = pending_tx_info->vif;
1486 make_tx_response(vif, &pending_tx_info->req, XEN_NETIF_RSP_OKAY);
1488 index = pending_index(netbk->pending_prod++);
1489 netbk->pending_ring[index] = pending_idx;
1493 netbk->mmap_pages[pending_idx]->mapping = 0;
1494 put_page(netbk->mmap_pages[pending_idx]);
1495 netbk->mmap_pages[pending_idx] = NULL;
1498 static void make_tx_response(struct xenvif *vif,
1499 struct xen_netif_tx_request *txp,
1502 RING_IDX i = vif->tx.rsp_prod_pvt;
1503 struct xen_netif_tx_response *resp;
1506 resp = RING_GET_RESPONSE(&vif->tx, i);
1510 if (txp->flags & XEN_NETTXF_extra_info)
1511 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1513 vif->tx.rsp_prod_pvt = ++i;
1514 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1516 notify_remote_via_irq(vif->irq);
1519 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1526 RING_IDX i = vif->rx.rsp_prod_pvt;
1527 struct xen_netif_rx_response *resp;
1529 resp = RING_GET_RESPONSE(&vif->rx, i);
1530 resp->offset = offset;
1531 resp->flags = flags;
1533 resp->status = (s16)size;
1535 resp->status = (s16)st;
1537 vif->rx.rsp_prod_pvt = ++i;
1542 static inline int rx_work_todo(struct xen_netbk *netbk)
1544 return !skb_queue_empty(&netbk->rx_queue);
1547 static inline int tx_work_todo(struct xen_netbk *netbk)
1550 if (((nr_pending_reqs(netbk) + MAX_SKB_FRAGS) < MAX_PENDING_REQS) &&
1551 !list_empty(&netbk->net_schedule_list))
1557 static int xen_netbk_kthread(void *data)
1559 struct xen_netbk *netbk = data;
1560 while (!kthread_should_stop()) {
1561 wait_event_interruptible(netbk->wq,
1562 rx_work_todo(netbk) ||
1563 tx_work_todo(netbk) ||
1564 kthread_should_stop());
1567 if (kthread_should_stop())
1570 if (rx_work_todo(netbk))
1571 xen_netbk_rx_action(netbk);
1573 if (tx_work_todo(netbk))
1574 xen_netbk_tx_action(netbk);
1580 void xen_netbk_unmap_frontend_rings(struct xenvif *vif)
1583 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1586 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1590 int xen_netbk_map_frontend_rings(struct xenvif *vif,
1591 grant_ref_t tx_ring_ref,
1592 grant_ref_t rx_ring_ref)
1595 struct xen_netif_tx_sring *txs;
1596 struct xen_netif_rx_sring *rxs;
1600 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1601 tx_ring_ref, &addr);
1605 txs = (struct xen_netif_tx_sring *)addr;
1606 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1608 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1609 rx_ring_ref, &addr);
1613 rxs = (struct xen_netif_rx_sring *)addr;
1614 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1616 vif->rx_req_cons_peek = 0;
1621 xen_netbk_unmap_frontend_rings(vif);
1625 static int __init netback_init(void)
1634 xen_netbk_group_nr = num_online_cpus();
1635 xen_netbk = vzalloc(sizeof(struct xen_netbk) * xen_netbk_group_nr);
1639 for (group = 0; group < xen_netbk_group_nr; group++) {
1640 struct xen_netbk *netbk = &xen_netbk[group];
1641 skb_queue_head_init(&netbk->rx_queue);
1642 skb_queue_head_init(&netbk->tx_queue);
1644 init_timer(&netbk->net_timer);
1645 netbk->net_timer.data = (unsigned long)netbk;
1646 netbk->net_timer.function = xen_netbk_alarm;
1648 netbk->pending_cons = 0;
1649 netbk->pending_prod = MAX_PENDING_REQS;
1650 for (i = 0; i < MAX_PENDING_REQS; i++)
1651 netbk->pending_ring[i] = i;
1653 init_waitqueue_head(&netbk->wq);
1654 netbk->task = kthread_create(xen_netbk_kthread,
1656 "netback/%u", group);
1658 if (IS_ERR(netbk->task)) {
1659 printk(KERN_ALERT "kthread_create() fails at netback\n");
1660 del_timer(&netbk->net_timer);
1661 rc = PTR_ERR(netbk->task);
1665 kthread_bind(netbk->task, group);
1667 INIT_LIST_HEAD(&netbk->net_schedule_list);
1669 spin_lock_init(&netbk->net_schedule_list_lock);
1671 atomic_set(&netbk->netfront_count, 0);
1673 wake_up_process(netbk->task);
1676 rc = xenvif_xenbus_init();
1683 while (--group >= 0) {
1684 struct xen_netbk *netbk = &xen_netbk[group];
1685 for (i = 0; i < MAX_PENDING_REQS; i++) {
1686 if (netbk->mmap_pages[i])
1687 __free_page(netbk->mmap_pages[i]);
1689 del_timer(&netbk->net_timer);
1690 kthread_stop(netbk->task);
1697 module_init(netback_init);
1699 MODULE_LICENSE("Dual BSD/GPL");
1700 MODULE_ALIAS("xen-backend:vif");