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>
49 #include <asm/xen/hypercall.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 = true;
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,
99 static void push_tx_responses(struct xenvif_queue *queue);
101 static inline int tx_work_todo(struct xenvif_queue *queue);
103 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
110 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
113 return page_to_pfn(queue->mmap_pages[idx]);
116 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
119 return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
122 #define callback_param(vif, pending_idx) \
123 (vif->pending_tx_info[pending_idx].callback_struct)
125 /* Find the containing VIF's structure from a pointer in pending_tx_info array
127 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
129 u16 pending_idx = ubuf->desc;
130 struct pending_tx_info *temp =
131 container_of(ubuf, struct pending_tx_info, callback_struct);
132 return container_of(temp - pending_idx,
137 static u16 frag_get_pending_idx(skb_frag_t *frag)
139 return (u16)frag->page_offset;
142 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
144 frag->page_offset = pending_idx;
147 static inline pending_ring_idx_t pending_index(unsigned i)
149 return i & (MAX_PENDING_REQS-1);
152 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
158 skb = skb_peek(&queue->rx_queue);
162 needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
167 prod = queue->rx.sring->req_prod;
168 cons = queue->rx.req_cons;
170 if (prod - cons >= needed)
173 queue->rx.sring->req_event = prod + 1;
175 /* Make sure event is visible before we check prod
179 } while (queue->rx.sring->req_prod != prod);
184 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
188 spin_lock_irqsave(&queue->rx_queue.lock, flags);
190 __skb_queue_tail(&queue->rx_queue, skb);
192 queue->rx_queue_len += skb->len;
193 if (queue->rx_queue_len > queue->rx_queue_max)
194 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
196 spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
199 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
203 spin_lock_irq(&queue->rx_queue.lock);
205 skb = __skb_dequeue(&queue->rx_queue);
207 queue->rx_queue_len -= skb->len;
209 spin_unlock_irq(&queue->rx_queue.lock);
214 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
216 spin_lock_irq(&queue->rx_queue.lock);
218 if (queue->rx_queue_len < queue->rx_queue_max)
219 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
221 spin_unlock_irq(&queue->rx_queue.lock);
225 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
228 while ((skb = xenvif_rx_dequeue(queue)) != NULL)
232 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
237 skb = skb_peek(&queue->rx_queue);
240 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
242 xenvif_rx_dequeue(queue);
247 struct netrx_pending_operations {
248 unsigned copy_prod, copy_cons;
249 unsigned meta_prod, meta_cons;
250 struct gnttab_copy *copy;
251 struct xenvif_rx_meta *meta;
253 grant_ref_t copy_gref;
256 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
257 struct netrx_pending_operations *npo)
259 struct xenvif_rx_meta *meta;
260 struct xen_netif_rx_request req;
262 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
264 meta = npo->meta + npo->meta_prod++;
265 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
271 npo->copy_gref = req.gref;
276 struct gop_frag_copy {
277 struct xenvif_queue *queue;
278 struct netrx_pending_operations *npo;
279 struct xenvif_rx_meta *meta;
286 static void xenvif_setup_copy_gop(unsigned long gfn,
289 struct gop_frag_copy *info)
291 struct gnttab_copy *copy_gop;
292 struct xen_page_foreign *foreign;
293 /* Convenient aliases */
294 struct xenvif_queue *queue = info->queue;
295 struct netrx_pending_operations *npo = info->npo;
296 struct page *page = info->page;
298 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
300 if (npo->copy_off == MAX_BUFFER_OFFSET)
301 info->meta = get_next_rx_buffer(queue, npo);
303 if (npo->copy_off + *len > MAX_BUFFER_OFFSET)
304 *len = MAX_BUFFER_OFFSET - npo->copy_off;
306 copy_gop = npo->copy + npo->copy_prod++;
307 copy_gop->flags = GNTCOPY_dest_gref;
308 copy_gop->len = *len;
310 foreign = xen_page_foreign(page);
312 copy_gop->source.domid = foreign->domid;
313 copy_gop->source.u.ref = foreign->gref;
314 copy_gop->flags |= GNTCOPY_source_gref;
316 copy_gop->source.domid = DOMID_SELF;
317 copy_gop->source.u.gmfn = gfn;
319 copy_gop->source.offset = offset;
321 copy_gop->dest.domid = queue->vif->domid;
322 copy_gop->dest.offset = npo->copy_off;
323 copy_gop->dest.u.ref = npo->copy_gref;
325 npo->copy_off += *len;
326 info->meta->size += *len;
328 /* Leave a gap for the GSO descriptor. */
329 if (info->head && ((1 << info->gso_type) & queue->vif->gso_mask))
330 queue->rx.req_cons++;
332 info->head = 0; /* There must be something in this buffer now */
335 static void xenvif_gop_frag_copy_grant(unsigned long gfn,
344 xenvif_setup_copy_gop(gfn, offset, &bytes, data);
351 * Set up the grant operations for this fragment. If it's a flipping
352 * interface, we also set up the unmap request from here.
354 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
355 struct netrx_pending_operations *npo,
356 struct page *page, unsigned long size,
357 unsigned long offset, int *head)
359 struct gop_frag_copy info = {
363 .gso_type = XEN_NETIF_GSO_TYPE_NONE,
367 if (skb_is_gso(skb)) {
368 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
369 info.gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
370 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
371 info.gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
374 /* Data must not cross a page boundary. */
375 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
377 info.meta = npo->meta + npo->meta_prod - 1;
379 /* Skip unused frames from start of page */
380 page += offset >> PAGE_SHIFT;
381 offset &= ~PAGE_MASK;
384 BUG_ON(offset >= PAGE_SIZE);
386 bytes = PAGE_SIZE - offset;
391 gnttab_foreach_grant_in_range(page, offset, bytes,
392 xenvif_gop_frag_copy_grant,
399 BUG_ON(!PageCompound(page));
408 * Prepare an SKB to be transmitted to the frontend.
410 * This function is responsible for allocating grant operations, meta
413 * It returns the number of meta structures consumed. The number of
414 * ring slots used is always equal to the number of meta slots used
415 * plus the number of GSO descriptors used. Currently, we use either
416 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
417 * frontend-side LRO).
419 static int xenvif_gop_skb(struct sk_buff *skb,
420 struct netrx_pending_operations *npo,
421 struct xenvif_queue *queue)
423 struct xenvif *vif = netdev_priv(skb->dev);
424 int nr_frags = skb_shinfo(skb)->nr_frags;
426 struct xen_netif_rx_request req;
427 struct xenvif_rx_meta *meta;
433 old_meta_prod = npo->meta_prod;
435 gso_type = XEN_NETIF_GSO_TYPE_NONE;
436 if (skb_is_gso(skb)) {
437 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
438 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
439 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
440 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
443 /* Set up a GSO prefix descriptor, if necessary */
444 if ((1 << gso_type) & vif->gso_prefix_mask) {
445 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
446 meta = npo->meta + npo->meta_prod++;
447 meta->gso_type = gso_type;
448 meta->gso_size = skb_shinfo(skb)->gso_size;
453 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
454 meta = npo->meta + npo->meta_prod++;
456 if ((1 << gso_type) & vif->gso_mask) {
457 meta->gso_type = gso_type;
458 meta->gso_size = skb_shinfo(skb)->gso_size;
460 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
467 npo->copy_gref = req.gref;
470 while (data < skb_tail_pointer(skb)) {
471 unsigned int offset = offset_in_page(data);
472 unsigned int len = PAGE_SIZE - offset;
474 if (data + len > skb_tail_pointer(skb))
475 len = skb_tail_pointer(skb) - data;
477 xenvif_gop_frag_copy(queue, skb, npo,
478 virt_to_page(data), len, offset, &head);
482 for (i = 0; i < nr_frags; i++) {
483 xenvif_gop_frag_copy(queue, skb, npo,
484 skb_frag_page(&skb_shinfo(skb)->frags[i]),
485 skb_frag_size(&skb_shinfo(skb)->frags[i]),
486 skb_shinfo(skb)->frags[i].page_offset,
490 return npo->meta_prod - old_meta_prod;
494 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
495 * used to set up the operations on the top of
496 * netrx_pending_operations, which have since been done. Check that
497 * they didn't give any errors and advance over them.
499 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
500 struct netrx_pending_operations *npo)
502 struct gnttab_copy *copy_op;
503 int status = XEN_NETIF_RSP_OKAY;
506 for (i = 0; i < nr_meta_slots; i++) {
507 copy_op = npo->copy + npo->copy_cons++;
508 if (copy_op->status != GNTST_okay) {
510 "Bad status %d from copy to DOM%d.\n",
511 copy_op->status, vif->domid);
512 status = XEN_NETIF_RSP_ERROR;
519 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
520 struct xenvif_rx_meta *meta,
524 unsigned long offset;
526 /* No fragments used */
527 if (nr_meta_slots <= 1)
532 for (i = 0; i < nr_meta_slots; i++) {
534 if (i == nr_meta_slots - 1)
537 flags = XEN_NETRXF_more_data;
540 make_rx_response(queue, meta[i].id, status, offset,
541 meta[i].size, flags);
545 void xenvif_kick_thread(struct xenvif_queue *queue)
550 static void xenvif_rx_action(struct xenvif_queue *queue)
554 struct xen_netif_rx_response *resp;
555 struct sk_buff_head rxq;
559 unsigned long offset;
560 bool need_to_notify = false;
562 struct netrx_pending_operations npo = {
563 .copy = queue->grant_copy_op,
567 skb_queue_head_init(&rxq);
569 while (xenvif_rx_ring_slots_available(queue)
570 && (skb = xenvif_rx_dequeue(queue)) != NULL) {
571 queue->last_rx_time = jiffies;
573 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
575 __skb_queue_tail(&rxq, skb);
578 BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
583 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
584 gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
586 while ((skb = __skb_dequeue(&rxq)) != NULL) {
588 if ((1 << queue->meta[npo.meta_cons].gso_type) &
589 queue->vif->gso_prefix_mask) {
590 resp = RING_GET_RESPONSE(&queue->rx,
591 queue->rx.rsp_prod_pvt++);
593 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
595 resp->offset = queue->meta[npo.meta_cons].gso_size;
596 resp->id = queue->meta[npo.meta_cons].id;
597 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
600 XENVIF_RX_CB(skb)->meta_slots_used--;
604 queue->stats.tx_bytes += skb->len;
605 queue->stats.tx_packets++;
607 status = xenvif_check_gop(queue->vif,
608 XENVIF_RX_CB(skb)->meta_slots_used,
611 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
614 flags = XEN_NETRXF_more_data;
616 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
617 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
618 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
619 /* remote but checksummed. */
620 flags |= XEN_NETRXF_data_validated;
623 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
625 queue->meta[npo.meta_cons].size,
628 if ((1 << queue->meta[npo.meta_cons].gso_type) &
629 queue->vif->gso_mask) {
630 struct xen_netif_extra_info *gso =
631 (struct xen_netif_extra_info *)
632 RING_GET_RESPONSE(&queue->rx,
633 queue->rx.rsp_prod_pvt++);
635 resp->flags |= XEN_NETRXF_extra_info;
637 gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
638 gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
640 gso->u.gso.features = 0;
642 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
646 xenvif_add_frag_responses(queue, status,
647 queue->meta + npo.meta_cons + 1,
648 XENVIF_RX_CB(skb)->meta_slots_used);
650 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
652 need_to_notify |= !!ret;
654 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
660 notify_remote_via_irq(queue->rx_irq);
663 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
667 RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
670 napi_schedule(&queue->napi);
673 static void tx_add_credit(struct xenvif_queue *queue)
675 unsigned long max_burst, max_credit;
678 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
679 * Otherwise the interface can seize up due to insufficient credit.
681 max_burst = max(131072UL, queue->credit_bytes);
683 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
684 max_credit = queue->remaining_credit + queue->credit_bytes;
685 if (max_credit < queue->remaining_credit)
686 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
688 queue->remaining_credit = min(max_credit, max_burst);
691 void xenvif_tx_credit_callback(unsigned long data)
693 struct xenvif_queue *queue = (struct xenvif_queue *)data;
694 tx_add_credit(queue);
695 xenvif_napi_schedule_or_enable_events(queue);
698 static void xenvif_tx_err(struct xenvif_queue *queue,
699 struct xen_netif_tx_request *txp, RING_IDX end)
701 RING_IDX cons = queue->tx.req_cons;
705 spin_lock_irqsave(&queue->response_lock, flags);
706 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
707 push_tx_responses(queue);
708 spin_unlock_irqrestore(&queue->response_lock, flags);
711 RING_COPY_REQUEST(&queue->tx, cons++, txp);
713 queue->tx.req_cons = cons;
716 static void xenvif_fatal_tx_err(struct xenvif *vif)
718 netdev_err(vif->dev, "fatal error; disabling device\n");
719 vif->disabled = true;
720 /* Disable the vif from queue 0's kthread */
722 xenvif_kick_thread(&vif->queues[0]);
725 static int xenvif_count_requests(struct xenvif_queue *queue,
726 struct xen_netif_tx_request *first,
727 struct xen_netif_tx_request *txp,
730 RING_IDX cons = queue->tx.req_cons;
735 if (!(first->flags & XEN_NETTXF_more_data))
739 struct xen_netif_tx_request dropped_tx = { 0 };
741 if (slots >= work_to_do) {
742 netdev_err(queue->vif->dev,
743 "Asked for %d slots but exceeds this limit\n",
745 xenvif_fatal_tx_err(queue->vif);
749 /* This guest is really using too many slots and
750 * considered malicious.
752 if (unlikely(slots >= fatal_skb_slots)) {
753 netdev_err(queue->vif->dev,
754 "Malicious frontend using %d slots, threshold %u\n",
755 slots, fatal_skb_slots);
756 xenvif_fatal_tx_err(queue->vif);
760 /* Xen network protocol had implicit dependency on
761 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
762 * the historical MAX_SKB_FRAGS value 18 to honor the
763 * same behavior as before. Any packet using more than
764 * 18 slots but less than fatal_skb_slots slots is
767 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
769 netdev_dbg(queue->vif->dev,
770 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
771 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
778 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
780 /* If the guest submitted a frame >= 64 KiB then
781 * first->size overflowed and following slots will
782 * appear to be larger than the frame.
784 * This cannot be fatal error as there are buggy
785 * frontends that do this.
787 * Consume all slots and drop the packet.
789 if (!drop_err && txp->size > first->size) {
791 netdev_dbg(queue->vif->dev,
792 "Invalid tx request, slot size %u > remaining size %u\n",
793 txp->size, first->size);
797 first->size -= txp->size;
800 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
801 netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
802 txp->offset, txp->size);
803 xenvif_fatal_tx_err(queue->vif);
807 more_data = txp->flags & XEN_NETTXF_more_data;
815 xenvif_tx_err(queue, first, cons + slots);
823 struct xenvif_tx_cb {
827 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
829 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
831 struct xen_netif_tx_request *txp,
832 struct gnttab_map_grant_ref *mop)
834 queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
835 gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
836 GNTMAP_host_map | GNTMAP_readonly,
837 txp->gref, queue->vif->domid);
839 memcpy(&queue->pending_tx_info[pending_idx].req, txp,
843 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
845 struct sk_buff *skb =
846 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
847 GFP_ATOMIC | __GFP_NOWARN);
848 if (unlikely(skb == NULL))
851 /* Packets passed to netif_rx() must have some headroom. */
852 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
854 /* Initialize it here to avoid later surprises */
855 skb_shinfo(skb)->destructor_arg = NULL;
860 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
862 struct xen_netif_tx_request *txp,
863 struct gnttab_map_grant_ref *gop,
864 unsigned int frag_overflow,
865 struct sk_buff *nskb)
867 struct skb_shared_info *shinfo = skb_shinfo(skb);
868 skb_frag_t *frags = shinfo->frags;
869 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
871 pending_ring_idx_t index;
872 unsigned int nr_slots;
874 nr_slots = shinfo->nr_frags;
876 /* Skip first skb fragment if it is on same page as header fragment. */
877 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
879 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
880 shinfo->nr_frags++, txp++, gop++) {
881 index = pending_index(queue->pending_cons++);
882 pending_idx = queue->pending_ring[index];
883 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
884 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
889 shinfo = skb_shinfo(nskb);
890 frags = shinfo->frags;
892 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
893 shinfo->nr_frags++, txp++, gop++) {
894 index = pending_index(queue->pending_cons++);
895 pending_idx = queue->pending_ring[index];
896 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
897 frag_set_pending_idx(&frags[shinfo->nr_frags],
901 skb_shinfo(skb)->frag_list = nskb;
907 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
909 grant_handle_t handle)
911 if (unlikely(queue->grant_tx_handle[pending_idx] !=
912 NETBACK_INVALID_HANDLE)) {
913 netdev_err(queue->vif->dev,
914 "Trying to overwrite active handle! pending_idx: 0x%x\n",
918 queue->grant_tx_handle[pending_idx] = handle;
921 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
924 if (unlikely(queue->grant_tx_handle[pending_idx] ==
925 NETBACK_INVALID_HANDLE)) {
926 netdev_err(queue->vif->dev,
927 "Trying to unmap invalid handle! pending_idx: 0x%x\n",
931 queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
934 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
936 struct gnttab_map_grant_ref **gopp_map,
937 struct gnttab_copy **gopp_copy)
939 struct gnttab_map_grant_ref *gop_map = *gopp_map;
940 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
941 /* This always points to the shinfo of the skb being checked, which
942 * could be either the first or the one on the frag_list
944 struct skb_shared_info *shinfo = skb_shinfo(skb);
945 /* If this is non-NULL, we are currently checking the frag_list skb, and
946 * this points to the shinfo of the first one
948 struct skb_shared_info *first_shinfo = NULL;
949 int nr_frags = shinfo->nr_frags;
950 const bool sharedslot = nr_frags &&
951 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
954 /* Check status of header. */
955 err = (*gopp_copy)->status;
958 netdev_dbg(queue->vif->dev,
959 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
960 (*gopp_copy)->status,
962 (*gopp_copy)->source.u.ref);
963 /* The first frag might still have this slot mapped */
965 xenvif_idx_release(queue, pending_idx,
966 XEN_NETIF_RSP_ERROR);
971 for (i = 0; i < nr_frags; i++, gop_map++) {
974 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
976 /* Check error status: if okay then remember grant handle. */
977 newerr = gop_map->status;
979 if (likely(!newerr)) {
980 xenvif_grant_handle_set(queue,
983 /* Had a previous error? Invalidate this fragment. */
985 xenvif_idx_unmap(queue, pending_idx);
986 /* If the mapping of the first frag was OK, but
987 * the header's copy failed, and they are
988 * sharing a slot, send an error
990 if (i == 0 && sharedslot)
991 xenvif_idx_release(queue, pending_idx,
992 XEN_NETIF_RSP_ERROR);
994 xenvif_idx_release(queue, pending_idx,
1000 /* Error on this fragment: respond to client with an error. */
1001 if (net_ratelimit())
1002 netdev_dbg(queue->vif->dev,
1003 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1009 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
1011 /* Not the first error? Preceding frags already invalidated. */
1015 /* First error: if the header haven't shared a slot with the
1016 * first frag, release it as well.
1019 xenvif_idx_release(queue,
1020 XENVIF_TX_CB(skb)->pending_idx,
1021 XEN_NETIF_RSP_OKAY);
1023 /* Invalidate preceding fragments of this skb. */
1024 for (j = 0; j < i; j++) {
1025 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1026 xenvif_idx_unmap(queue, pending_idx);
1027 xenvif_idx_release(queue, pending_idx,
1028 XEN_NETIF_RSP_OKAY);
1031 /* And if we found the error while checking the frag_list, unmap
1032 * the first skb's frags
1035 for (j = 0; j < first_shinfo->nr_frags; j++) {
1036 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1037 xenvif_idx_unmap(queue, pending_idx);
1038 xenvif_idx_release(queue, pending_idx,
1039 XEN_NETIF_RSP_OKAY);
1043 /* Remember the error: invalidate all subsequent fragments. */
1047 if (skb_has_frag_list(skb) && !first_shinfo) {
1048 first_shinfo = skb_shinfo(skb);
1049 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1050 nr_frags = shinfo->nr_frags;
1055 *gopp_map = gop_map;
1059 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1061 struct skb_shared_info *shinfo = skb_shinfo(skb);
1062 int nr_frags = shinfo->nr_frags;
1064 u16 prev_pending_idx = INVALID_PENDING_IDX;
1066 for (i = 0; i < nr_frags; i++) {
1067 skb_frag_t *frag = shinfo->frags + i;
1068 struct xen_netif_tx_request *txp;
1072 pending_idx = frag_get_pending_idx(frag);
1074 /* If this is not the first frag, chain it to the previous*/
1075 if (prev_pending_idx == INVALID_PENDING_IDX)
1076 skb_shinfo(skb)->destructor_arg =
1077 &callback_param(queue, pending_idx);
1079 callback_param(queue, prev_pending_idx).ctx =
1080 &callback_param(queue, pending_idx);
1082 callback_param(queue, pending_idx).ctx = NULL;
1083 prev_pending_idx = pending_idx;
1085 txp = &queue->pending_tx_info[pending_idx].req;
1086 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1087 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1088 skb->len += txp->size;
1089 skb->data_len += txp->size;
1090 skb->truesize += txp->size;
1092 /* Take an extra reference to offset network stack's put_page */
1093 get_page(queue->mmap_pages[pending_idx]);
1097 static int xenvif_get_extras(struct xenvif_queue *queue,
1098 struct xen_netif_extra_info *extras,
1101 struct xen_netif_extra_info extra;
1102 RING_IDX cons = queue->tx.req_cons;
1105 if (unlikely(work_to_do-- <= 0)) {
1106 netdev_err(queue->vif->dev, "Missing extra info\n");
1107 xenvif_fatal_tx_err(queue->vif);
1111 RING_COPY_REQUEST(&queue->tx, cons, &extra);
1112 if (unlikely(!extra.type ||
1113 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1114 queue->tx.req_cons = ++cons;
1115 netdev_err(queue->vif->dev,
1116 "Invalid extra type: %d\n", extra.type);
1117 xenvif_fatal_tx_err(queue->vif);
1121 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1122 queue->tx.req_cons = ++cons;
1123 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1128 static int xenvif_set_skb_gso(struct xenvif *vif,
1129 struct sk_buff *skb,
1130 struct xen_netif_extra_info *gso)
1132 if (!gso->u.gso.size) {
1133 netdev_err(vif->dev, "GSO size must not be zero.\n");
1134 xenvif_fatal_tx_err(vif);
1138 switch (gso->u.gso.type) {
1139 case XEN_NETIF_GSO_TYPE_TCPV4:
1140 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1142 case XEN_NETIF_GSO_TYPE_TCPV6:
1143 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1146 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1147 xenvif_fatal_tx_err(vif);
1151 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1152 /* gso_segs will be calculated later */
1157 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1159 bool recalculate_partial_csum = false;
1161 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1162 * peers can fail to set NETRXF_csum_blank when sending a GSO
1163 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1164 * recalculate the partial checksum.
1166 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1167 queue->stats.rx_gso_checksum_fixup++;
1168 skb->ip_summed = CHECKSUM_PARTIAL;
1169 recalculate_partial_csum = true;
1172 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1173 if (skb->ip_summed != CHECKSUM_PARTIAL)
1176 return skb_checksum_setup(skb, recalculate_partial_csum);
1179 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1181 u64 now = get_jiffies_64();
1182 u64 next_credit = queue->credit_window_start +
1183 msecs_to_jiffies(queue->credit_usec / 1000);
1185 /* Timer could already be pending in rare cases. */
1186 if (timer_pending(&queue->credit_timeout))
1189 /* Passed the point where we can replenish credit? */
1190 if (time_after_eq64(now, next_credit)) {
1191 queue->credit_window_start = now;
1192 tx_add_credit(queue);
1195 /* Still too big to send right now? Set a callback. */
1196 if (size > queue->remaining_credit) {
1197 queue->credit_timeout.data =
1198 (unsigned long)queue;
1199 mod_timer(&queue->credit_timeout,
1201 queue->credit_window_start = next_credit;
1209 /* No locking is required in xenvif_mcast_add/del() as they are
1210 * only ever invoked from NAPI poll. An RCU list is used because
1211 * xenvif_mcast_match() is called asynchronously, during start_xmit.
1214 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
1216 struct xenvif_mcast_addr *mcast;
1218 if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
1219 if (net_ratelimit())
1220 netdev_err(vif->dev,
1221 "Too many multicast addresses\n");
1225 mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
1229 ether_addr_copy(mcast->addr, addr);
1230 list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
1231 vif->fe_mcast_count++;
1236 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
1238 struct xenvif_mcast_addr *mcast;
1240 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1241 if (ether_addr_equal(addr, mcast->addr)) {
1242 --vif->fe_mcast_count;
1243 list_del_rcu(&mcast->entry);
1244 kfree_rcu(mcast, rcu);
1250 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
1252 struct xenvif_mcast_addr *mcast;
1255 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1256 if (ether_addr_equal(addr, mcast->addr)) {
1266 void xenvif_mcast_addr_list_free(struct xenvif *vif)
1268 /* No need for locking or RCU here. NAPI poll and TX queue
1271 while (!list_empty(&vif->fe_mcast_addr)) {
1272 struct xenvif_mcast_addr *mcast;
1274 mcast = list_first_entry(&vif->fe_mcast_addr,
1275 struct xenvif_mcast_addr,
1277 --vif->fe_mcast_count;
1278 list_del(&mcast->entry);
1283 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1288 struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
1289 struct sk_buff *skb, *nskb;
1291 unsigned int frag_overflow;
1293 while (skb_queue_len(&queue->tx_queue) < budget) {
1294 struct xen_netif_tx_request txreq;
1295 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1296 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1300 unsigned int data_len;
1301 pending_ring_idx_t index;
1303 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1304 XEN_NETIF_TX_RING_SIZE) {
1305 netdev_err(queue->vif->dev,
1306 "Impossible number of requests. "
1307 "req_prod %d, req_cons %d, size %ld\n",
1308 queue->tx.sring->req_prod, queue->tx.req_cons,
1309 XEN_NETIF_TX_RING_SIZE);
1310 xenvif_fatal_tx_err(queue->vif);
1314 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1318 idx = queue->tx.req_cons;
1319 rmb(); /* Ensure that we see the request before we copy it. */
1320 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
1322 /* Credit-based scheduling. */
1323 if (txreq.size > queue->remaining_credit &&
1324 tx_credit_exceeded(queue, txreq.size))
1327 queue->remaining_credit -= txreq.size;
1330 queue->tx.req_cons = ++idx;
1332 memset(extras, 0, sizeof(extras));
1333 if (txreq.flags & XEN_NETTXF_extra_info) {
1334 work_to_do = xenvif_get_extras(queue, extras,
1336 idx = queue->tx.req_cons;
1337 if (unlikely(work_to_do < 0))
1341 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
1342 struct xen_netif_extra_info *extra;
1344 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
1345 ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
1347 make_tx_response(queue, &txreq,
1349 XEN_NETIF_RSP_OKAY :
1350 XEN_NETIF_RSP_ERROR);
1351 push_tx_responses(queue);
1355 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
1356 struct xen_netif_extra_info *extra;
1358 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
1359 xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
1361 make_tx_response(queue, &txreq, XEN_NETIF_RSP_OKAY);
1362 push_tx_responses(queue);
1366 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1367 if (unlikely(ret < 0))
1372 if (unlikely(txreq.size < ETH_HLEN)) {
1373 netdev_dbg(queue->vif->dev,
1374 "Bad packet size: %d\n", txreq.size);
1375 xenvif_tx_err(queue, &txreq, idx);
1379 /* No crossing a page as the payload mustn't fragment. */
1380 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1381 netdev_err(queue->vif->dev,
1382 "txreq.offset: %u, size: %u, end: %lu\n",
1383 txreq.offset, txreq.size,
1384 (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
1385 xenvif_fatal_tx_err(queue->vif);
1389 index = pending_index(queue->pending_cons);
1390 pending_idx = queue->pending_ring[index];
1392 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1393 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1394 XEN_NETBACK_TX_COPY_LEN : txreq.size;
1396 skb = xenvif_alloc_skb(data_len);
1397 if (unlikely(skb == NULL)) {
1398 netdev_dbg(queue->vif->dev,
1399 "Can't allocate a skb in start_xmit.\n");
1400 xenvif_tx_err(queue, &txreq, idx);
1404 skb_shinfo(skb)->nr_frags = ret;
1405 if (data_len < txreq.size)
1406 skb_shinfo(skb)->nr_frags++;
1407 /* At this point shinfo->nr_frags is in fact the number of
1408 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1412 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1413 frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1414 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1415 skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1416 nskb = xenvif_alloc_skb(0);
1417 if (unlikely(nskb == NULL)) {
1419 xenvif_tx_err(queue, &txreq, idx);
1420 if (net_ratelimit())
1421 netdev_err(queue->vif->dev,
1422 "Can't allocate the frag_list skb.\n");
1427 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1428 struct xen_netif_extra_info *gso;
1429 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1431 if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1432 /* Failure in xenvif_set_skb_gso is fatal. */
1439 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1441 __skb_put(skb, data_len);
1442 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1443 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1444 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1446 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1447 virt_to_gfn(skb->data);
1448 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1449 queue->tx_copy_ops[*copy_ops].dest.offset =
1450 offset_in_page(skb->data) & ~XEN_PAGE_MASK;
1452 queue->tx_copy_ops[*copy_ops].len = data_len;
1453 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1457 if (data_len < txreq.size) {
1458 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1460 xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1463 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1464 INVALID_PENDING_IDX);
1465 memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1469 queue->pending_cons++;
1471 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1472 frag_overflow, nskb);
1474 __skb_queue_tail(&queue->tx_queue, skb);
1476 queue->tx.req_cons = idx;
1478 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1479 (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1483 (*map_ops) = gop - queue->tx_map_ops;
1487 /* Consolidate skb with a frag_list into a brand new one with local pages on
1488 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1490 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1492 unsigned int offset = skb_headlen(skb);
1493 skb_frag_t frags[MAX_SKB_FRAGS];
1495 struct ubuf_info *uarg;
1496 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1498 queue->stats.tx_zerocopy_sent += 2;
1499 queue->stats.tx_frag_overflow++;
1501 xenvif_fill_frags(queue, nskb);
1502 /* Subtract frags size, we will correct it later */
1503 skb->truesize -= skb->data_len;
1504 skb->len += nskb->len;
1505 skb->data_len += nskb->len;
1507 /* create a brand new frags array and coalesce there */
1508 for (i = 0; offset < skb->len; i++) {
1512 BUG_ON(i >= MAX_SKB_FRAGS);
1513 page = alloc_page(GFP_ATOMIC);
1516 skb->truesize += skb->data_len;
1517 for (j = 0; j < i; j++)
1518 put_page(frags[j].page.p);
1522 if (offset + PAGE_SIZE < skb->len)
1525 len = skb->len - offset;
1526 if (skb_copy_bits(skb, offset, page_address(page), len))
1530 frags[i].page.p = page;
1531 frags[i].page_offset = 0;
1532 skb_frag_size_set(&frags[i], len);
1535 /* Copied all the bits from the frag list -- free it. */
1536 skb_frag_list_init(skb);
1537 xenvif_skb_zerocopy_prepare(queue, nskb);
1540 /* Release all the original (foreign) frags. */
1541 for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1542 skb_frag_unref(skb, f);
1543 uarg = skb_shinfo(skb)->destructor_arg;
1544 /* increase inflight counter to offset decrement in callback */
1545 atomic_inc(&queue->inflight_packets);
1546 uarg->callback(uarg, true);
1547 skb_shinfo(skb)->destructor_arg = NULL;
1549 /* Fill the skb with the new (local) frags. */
1550 memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1551 skb_shinfo(skb)->nr_frags = i;
1552 skb->truesize += i * PAGE_SIZE;
1557 static int xenvif_tx_submit(struct xenvif_queue *queue)
1559 struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1560 struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1561 struct sk_buff *skb;
1564 while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1565 struct xen_netif_tx_request *txp;
1569 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1570 txp = &queue->pending_tx_info[pending_idx].req;
1572 /* Check the remap error code. */
1573 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1574 /* If there was an error, xenvif_tx_check_gop is
1575 * expected to release all the frags which were mapped,
1576 * so kfree_skb shouldn't do it again
1578 skb_shinfo(skb)->nr_frags = 0;
1579 if (skb_has_frag_list(skb)) {
1580 struct sk_buff *nskb =
1581 skb_shinfo(skb)->frag_list;
1582 skb_shinfo(nskb)->nr_frags = 0;
1588 data_len = skb->len;
1589 callback_param(queue, pending_idx).ctx = NULL;
1590 if (data_len < txp->size) {
1591 /* Append the packet payload as a fragment. */
1592 txp->offset += data_len;
1593 txp->size -= data_len;
1595 /* Schedule a response immediately. */
1596 xenvif_idx_release(queue, pending_idx,
1597 XEN_NETIF_RSP_OKAY);
1600 if (txp->flags & XEN_NETTXF_csum_blank)
1601 skb->ip_summed = CHECKSUM_PARTIAL;
1602 else if (txp->flags & XEN_NETTXF_data_validated)
1603 skb->ip_summed = CHECKSUM_UNNECESSARY;
1605 xenvif_fill_frags(queue, skb);
1607 if (unlikely(skb_has_frag_list(skb))) {
1608 if (xenvif_handle_frag_list(queue, skb)) {
1609 if (net_ratelimit())
1610 netdev_err(queue->vif->dev,
1611 "Not enough memory to consolidate frag_list!\n");
1612 xenvif_skb_zerocopy_prepare(queue, skb);
1618 skb->dev = queue->vif->dev;
1619 skb->protocol = eth_type_trans(skb, skb->dev);
1620 skb_reset_network_header(skb);
1622 if (checksum_setup(queue, skb)) {
1623 netdev_dbg(queue->vif->dev,
1624 "Can't setup checksum in net_tx_action\n");
1625 /* We have to set this flag to trigger the callback */
1626 if (skb_shinfo(skb)->destructor_arg)
1627 xenvif_skb_zerocopy_prepare(queue, skb);
1632 skb_probe_transport_header(skb, 0);
1634 /* If the packet is GSO then we will have just set up the
1635 * transport header offset in checksum_setup so it's now
1636 * straightforward to calculate gso_segs.
1638 if (skb_is_gso(skb)) {
1639 int mss = skb_shinfo(skb)->gso_size;
1640 int hdrlen = skb_transport_header(skb) -
1641 skb_mac_header(skb) +
1644 skb_shinfo(skb)->gso_segs =
1645 DIV_ROUND_UP(skb->len - hdrlen, mss);
1648 queue->stats.rx_bytes += skb->len;
1649 queue->stats.rx_packets++;
1653 /* Set this flag right before netif_receive_skb, otherwise
1654 * someone might think this packet already left netback, and
1655 * do a skb_copy_ubufs while we are still in control of the
1656 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1658 if (skb_shinfo(skb)->destructor_arg) {
1659 xenvif_skb_zerocopy_prepare(queue, skb);
1660 queue->stats.tx_zerocopy_sent++;
1663 netif_receive_skb(skb);
1669 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1671 unsigned long flags;
1672 pending_ring_idx_t index;
1673 struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1675 /* This is the only place where we grab this lock, to protect callbacks
1678 spin_lock_irqsave(&queue->callback_lock, flags);
1680 u16 pending_idx = ubuf->desc;
1681 ubuf = (struct ubuf_info *) ubuf->ctx;
1682 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1684 index = pending_index(queue->dealloc_prod);
1685 queue->dealloc_ring[index] = pending_idx;
1686 /* Sync with xenvif_tx_dealloc_action:
1687 * insert idx then incr producer.
1690 queue->dealloc_prod++;
1692 spin_unlock_irqrestore(&queue->callback_lock, flags);
1694 if (likely(zerocopy_success))
1695 queue->stats.tx_zerocopy_success++;
1697 queue->stats.tx_zerocopy_fail++;
1698 xenvif_skb_zerocopy_complete(queue);
1701 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1703 struct gnttab_unmap_grant_ref *gop;
1704 pending_ring_idx_t dc, dp;
1705 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1708 dc = queue->dealloc_cons;
1709 gop = queue->tx_unmap_ops;
1711 /* Free up any grants we have finished using */
1713 dp = queue->dealloc_prod;
1715 /* Ensure we see all indices enqueued by all
1716 * xenvif_zerocopy_callback().
1721 BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1723 queue->dealloc_ring[pending_index(dc++)];
1725 pending_idx_release[gop - queue->tx_unmap_ops] =
1727 queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1728 queue->mmap_pages[pending_idx];
1729 gnttab_set_unmap_op(gop,
1730 idx_to_kaddr(queue, pending_idx),
1732 queue->grant_tx_handle[pending_idx]);
1733 xenvif_grant_handle_reset(queue, pending_idx);
1737 } while (dp != queue->dealloc_prod);
1739 queue->dealloc_cons = dc;
1741 if (gop - queue->tx_unmap_ops > 0) {
1743 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1745 queue->pages_to_unmap,
1746 gop - queue->tx_unmap_ops);
1748 netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1749 gop - queue->tx_unmap_ops, ret);
1750 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1751 if (gop[i].status != GNTST_okay)
1752 netdev_err(queue->vif->dev,
1753 " host_addr: 0x%llx handle: 0x%x status: %d\n",
1762 for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1763 xenvif_idx_release(queue, pending_idx_release[i],
1764 XEN_NETIF_RSP_OKAY);
1768 /* Called after netfront has transmitted */
1769 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1771 unsigned nr_mops, nr_cops = 0;
1774 if (unlikely(!tx_work_todo(queue)))
1777 xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1782 gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1784 ret = gnttab_map_refs(queue->tx_map_ops,
1786 queue->pages_to_map,
1791 work_done = xenvif_tx_submit(queue);
1796 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1799 struct pending_tx_info *pending_tx_info;
1800 pending_ring_idx_t index;
1801 unsigned long flags;
1803 pending_tx_info = &queue->pending_tx_info[pending_idx];
1805 spin_lock_irqsave(&queue->response_lock, flags);
1807 make_tx_response(queue, &pending_tx_info->req, status);
1809 /* Release the pending index before pusing the Tx response so
1810 * its available before a new Tx request is pushed by the
1813 index = pending_index(queue->pending_prod++);
1814 queue->pending_ring[index] = pending_idx;
1816 push_tx_responses(queue);
1818 spin_unlock_irqrestore(&queue->response_lock, flags);
1822 static void make_tx_response(struct xenvif_queue *queue,
1823 struct xen_netif_tx_request *txp,
1826 RING_IDX i = queue->tx.rsp_prod_pvt;
1827 struct xen_netif_tx_response *resp;
1829 resp = RING_GET_RESPONSE(&queue->tx, i);
1833 if (txp->flags & XEN_NETTXF_extra_info)
1834 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1836 queue->tx.rsp_prod_pvt = ++i;
1839 static void push_tx_responses(struct xenvif_queue *queue)
1843 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1845 notify_remote_via_irq(queue->tx_irq);
1848 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1855 RING_IDX i = queue->rx.rsp_prod_pvt;
1856 struct xen_netif_rx_response *resp;
1858 resp = RING_GET_RESPONSE(&queue->rx, i);
1859 resp->offset = offset;
1860 resp->flags = flags;
1862 resp->status = (s16)size;
1864 resp->status = (s16)st;
1866 queue->rx.rsp_prod_pvt = ++i;
1871 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1874 struct gnttab_unmap_grant_ref tx_unmap_op;
1876 gnttab_set_unmap_op(&tx_unmap_op,
1877 idx_to_kaddr(queue, pending_idx),
1879 queue->grant_tx_handle[pending_idx]);
1880 xenvif_grant_handle_reset(queue, pending_idx);
1882 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1883 &queue->mmap_pages[pending_idx], 1);
1885 netdev_err(queue->vif->dev,
1886 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1889 tx_unmap_op.host_addr,
1891 tx_unmap_op.status);
1896 static inline int tx_work_todo(struct xenvif_queue *queue)
1898 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1904 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1906 return queue->dealloc_cons != queue->dealloc_prod;
1909 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1911 if (queue->tx.sring)
1912 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1914 if (queue->rx.sring)
1915 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1919 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1920 grant_ref_t tx_ring_ref,
1921 grant_ref_t rx_ring_ref)
1924 struct xen_netif_tx_sring *txs;
1925 struct xen_netif_rx_sring *rxs;
1929 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1930 &tx_ring_ref, 1, &addr);
1934 txs = (struct xen_netif_tx_sring *)addr;
1935 BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1937 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1938 &rx_ring_ref, 1, &addr);
1942 rxs = (struct xen_netif_rx_sring *)addr;
1943 BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1948 xenvif_unmap_frontend_rings(queue);
1952 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1954 struct xenvif *vif = queue->vif;
1956 queue->stalled = true;
1958 /* At least one queue has stalled? Disable the carrier. */
1959 spin_lock(&vif->lock);
1960 if (vif->stalled_queues++ == 0) {
1961 netdev_info(vif->dev, "Guest Rx stalled");
1962 netif_carrier_off(vif->dev);
1964 spin_unlock(&vif->lock);
1967 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1969 struct xenvif *vif = queue->vif;
1971 queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1972 queue->stalled = false;
1974 /* All queues are ready? Enable the carrier. */
1975 spin_lock(&vif->lock);
1976 if (--vif->stalled_queues == 0) {
1977 netdev_info(vif->dev, "Guest Rx ready");
1978 netif_carrier_on(vif->dev);
1980 spin_unlock(&vif->lock);
1983 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
1985 RING_IDX prod, cons;
1987 prod = queue->rx.sring->req_prod;
1988 cons = queue->rx.req_cons;
1990 return !queue->stalled && prod - cons < 1
1991 && time_after(jiffies,
1992 queue->last_rx_time + queue->vif->stall_timeout);
1995 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
1997 RING_IDX prod, cons;
1999 prod = queue->rx.sring->req_prod;
2000 cons = queue->rx.req_cons;
2002 return queue->stalled && prod - cons >= 1;
2005 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
2007 return xenvif_rx_ring_slots_available(queue)
2008 || (queue->vif->stall_timeout &&
2009 (xenvif_rx_queue_stalled(queue)
2010 || xenvif_rx_queue_ready(queue)))
2011 || kthread_should_stop()
2012 || queue->vif->disabled;
2015 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
2017 struct sk_buff *skb;
2020 skb = skb_peek(&queue->rx_queue);
2022 return MAX_SCHEDULE_TIMEOUT;
2024 timeout = XENVIF_RX_CB(skb)->expires - jiffies;
2025 return timeout < 0 ? 0 : timeout;
2028 /* Wait until the guest Rx thread has work.
2030 * The timeout needs to be adjusted based on the current head of the
2031 * queue (and not just the head at the beginning). In particular, if
2032 * the queue is initially empty an infinite timeout is used and this
2033 * needs to be reduced when a skb is queued.
2035 * This cannot be done with wait_event_timeout() because it only
2036 * calculates the timeout once.
2038 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
2042 if (xenvif_have_rx_work(queue))
2048 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
2049 if (xenvif_have_rx_work(queue))
2051 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
2055 finish_wait(&queue->wq, &wait);
2058 int xenvif_kthread_guest_rx(void *data)
2060 struct xenvif_queue *queue = data;
2061 struct xenvif *vif = queue->vif;
2063 if (!vif->stall_timeout)
2064 xenvif_queue_carrier_on(queue);
2067 xenvif_wait_for_rx_work(queue);
2069 if (kthread_should_stop())
2072 /* This frontend is found to be rogue, disable it in
2073 * kthread context. Currently this is only set when
2074 * netback finds out frontend sends malformed packet,
2075 * but we cannot disable the interface in softirq
2076 * context so we defer it here, if this thread is
2077 * associated with queue 0.
2079 if (unlikely(vif->disabled && queue->id == 0)) {
2080 xenvif_carrier_off(vif);
2084 if (!skb_queue_empty(&queue->rx_queue))
2085 xenvif_rx_action(queue);
2087 /* If the guest hasn't provided any Rx slots for a
2088 * while it's probably not responsive, drop the
2089 * carrier so packets are dropped earlier.
2091 if (vif->stall_timeout) {
2092 if (xenvif_rx_queue_stalled(queue))
2093 xenvif_queue_carrier_off(queue);
2094 else if (xenvif_rx_queue_ready(queue))
2095 xenvif_queue_carrier_on(queue);
2098 /* Queued packets may have foreign pages from other
2099 * domains. These cannot be queued indefinitely as
2100 * this would starve guests of grant refs and transmit
2103 xenvif_rx_queue_drop_expired(queue);
2105 xenvif_rx_queue_maybe_wake(queue);
2110 /* Bin any remaining skbs */
2111 xenvif_rx_queue_purge(queue);
2116 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2118 /* Dealloc thread must remain running until all inflight
2121 return kthread_should_stop() &&
2122 !atomic_read(&queue->inflight_packets);
2125 int xenvif_dealloc_kthread(void *data)
2127 struct xenvif_queue *queue = data;
2130 wait_event_interruptible(queue->dealloc_wq,
2131 tx_dealloc_work_todo(queue) ||
2132 xenvif_dealloc_kthread_should_stop(queue));
2133 if (xenvif_dealloc_kthread_should_stop(queue))
2136 xenvif_tx_dealloc_action(queue);
2140 /* Unmap anything remaining*/
2141 if (tx_dealloc_work_todo(queue))
2142 xenvif_tx_dealloc_action(queue);
2147 static int __init netback_init(void)
2154 /* Allow as many queues as there are CPUs if user has not
2155 * specified a value.
2157 if (xenvif_max_queues == 0)
2158 xenvif_max_queues = num_online_cpus();
2160 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2161 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2162 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2163 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2166 rc = xenvif_xenbus_init();
2170 #ifdef CONFIG_DEBUG_FS
2171 xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2172 if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2173 pr_warn("Init of debugfs returned %ld!\n",
2174 PTR_ERR(xen_netback_dbg_root));
2175 #endif /* CONFIG_DEBUG_FS */
2183 module_init(netback_init);
2185 static void __exit netback_fini(void)
2187 #ifdef CONFIG_DEBUG_FS
2188 if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2189 debugfs_remove_recursive(xen_netback_dbg_root);
2190 #endif /* CONFIG_DEBUG_FS */
2191 xenvif_xenbus_fini();
2193 module_exit(netback_fini);
2195 MODULE_LICENSE("Dual BSD/GPL");
2196 MODULE_ALIAS("xen-backend:vif");