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 #define MAX_QUEUES_DEFAULT 8
71 unsigned int xenvif_max_queues;
72 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
73 MODULE_PARM_DESC(max_queues,
74 "Maximum number of queues per virtual interface");
77 * This is the maximum slots a skb can have. If a guest sends a skb
78 * which exceeds this limit it is considered malicious.
80 #define FATAL_SKB_SLOTS_DEFAULT 20
81 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
82 module_param(fatal_skb_slots, uint, 0444);
84 /* The amount to copy out of the first guest Tx slot into the skb's
85 * linear area. If the first slot has more data, it will be mapped
86 * and put into the first frag.
88 * This is sized to avoid pulling headers from the frags for most
91 #define XEN_NETBACK_TX_COPY_LEN 128
94 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
97 static void make_tx_response(struct xenvif_queue *queue,
98 struct xen_netif_tx_request *txp,
100 static void push_tx_responses(struct xenvif_queue *queue);
102 static inline int tx_work_todo(struct xenvif_queue *queue);
104 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
111 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
114 return page_to_pfn(queue->mmap_pages[idx]);
117 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
120 return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
123 #define callback_param(vif, pending_idx) \
124 (vif->pending_tx_info[pending_idx].callback_struct)
126 /* Find the containing VIF's structure from a pointer in pending_tx_info array
128 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
130 u16 pending_idx = ubuf->desc;
131 struct pending_tx_info *temp =
132 container_of(ubuf, struct pending_tx_info, callback_struct);
133 return container_of(temp - pending_idx,
138 static u16 frag_get_pending_idx(skb_frag_t *frag)
140 return (u16)frag->page_offset;
143 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
145 frag->page_offset = pending_idx;
148 static inline pending_ring_idx_t pending_index(unsigned i)
150 return i & (MAX_PENDING_REQS-1);
153 static int xenvif_rx_ring_slots_needed(struct xenvif *vif)
156 return DIV_ROUND_UP(vif->dev->gso_max_size, XEN_PAGE_SIZE) + 1;
158 return DIV_ROUND_UP(vif->dev->mtu, XEN_PAGE_SIZE);
161 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
166 needed = xenvif_rx_ring_slots_needed(queue->vif);
169 prod = queue->rx.sring->req_prod;
170 cons = queue->rx.req_cons;
172 if (prod - cons >= needed)
175 queue->rx.sring->req_event = prod + 1;
177 /* Make sure event is visible before we check prod
181 } while (queue->rx.sring->req_prod != prod);
186 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
190 spin_lock_irqsave(&queue->rx_queue.lock, flags);
192 __skb_queue_tail(&queue->rx_queue, skb);
194 queue->rx_queue_len += skb->len;
195 if (queue->rx_queue_len > queue->rx_queue_max)
196 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
198 spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
201 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
205 spin_lock_irq(&queue->rx_queue.lock);
207 skb = __skb_dequeue(&queue->rx_queue);
209 queue->rx_queue_len -= skb->len;
211 spin_unlock_irq(&queue->rx_queue.lock);
216 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
218 spin_lock_irq(&queue->rx_queue.lock);
220 if (queue->rx_queue_len < queue->rx_queue_max)
221 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
223 spin_unlock_irq(&queue->rx_queue.lock);
227 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
230 while ((skb = xenvif_rx_dequeue(queue)) != NULL)
234 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
239 skb = skb_peek(&queue->rx_queue);
242 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
244 xenvif_rx_dequeue(queue);
249 struct netrx_pending_operations {
250 unsigned copy_prod, copy_cons;
251 unsigned meta_prod, meta_cons;
252 struct gnttab_copy *copy;
253 struct xenvif_rx_meta *meta;
255 grant_ref_t copy_gref;
258 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
259 struct netrx_pending_operations *npo)
261 struct xenvif_rx_meta *meta;
262 struct xen_netif_rx_request req;
264 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
266 meta = npo->meta + npo->meta_prod++;
267 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
273 npo->copy_gref = req.gref;
278 struct gop_frag_copy {
279 struct xenvif_queue *queue;
280 struct netrx_pending_operations *npo;
281 struct xenvif_rx_meta *meta;
288 static void xenvif_setup_copy_gop(unsigned long gfn,
291 struct gop_frag_copy *info)
293 struct gnttab_copy *copy_gop;
294 struct xen_page_foreign *foreign;
295 /* Convenient aliases */
296 struct xenvif_queue *queue = info->queue;
297 struct netrx_pending_operations *npo = info->npo;
298 struct page *page = info->page;
300 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
302 if (npo->copy_off == MAX_BUFFER_OFFSET)
303 info->meta = get_next_rx_buffer(queue, npo);
305 if (npo->copy_off + *len > MAX_BUFFER_OFFSET)
306 *len = MAX_BUFFER_OFFSET - npo->copy_off;
308 copy_gop = npo->copy + npo->copy_prod++;
309 copy_gop->flags = GNTCOPY_dest_gref;
310 copy_gop->len = *len;
312 foreign = xen_page_foreign(page);
314 copy_gop->source.domid = foreign->domid;
315 copy_gop->source.u.ref = foreign->gref;
316 copy_gop->flags |= GNTCOPY_source_gref;
318 copy_gop->source.domid = DOMID_SELF;
319 copy_gop->source.u.gmfn = gfn;
321 copy_gop->source.offset = offset;
323 copy_gop->dest.domid = queue->vif->domid;
324 copy_gop->dest.offset = npo->copy_off;
325 copy_gop->dest.u.ref = npo->copy_gref;
327 npo->copy_off += *len;
328 info->meta->size += *len;
330 /* Leave a gap for the GSO descriptor. */
331 if (info->head && ((1 << info->gso_type) & queue->vif->gso_mask))
332 queue->rx.req_cons++;
334 info->head = 0; /* There must be something in this buffer now */
337 static void xenvif_gop_frag_copy_grant(unsigned long gfn,
346 xenvif_setup_copy_gop(gfn, offset, &bytes, data);
353 * Set up the grant operations for this fragment. If it's a flipping
354 * interface, we also set up the unmap request from here.
356 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
357 struct netrx_pending_operations *npo,
358 struct page *page, unsigned long size,
359 unsigned long offset, int *head)
361 struct gop_frag_copy info = {
365 .gso_type = XEN_NETIF_GSO_TYPE_NONE,
369 if (skb_is_gso(skb)) {
370 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
371 info.gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
372 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
373 info.gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
376 /* Data must not cross a page boundary. */
377 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
379 info.meta = npo->meta + npo->meta_prod - 1;
381 /* Skip unused frames from start of page */
382 page += offset >> PAGE_SHIFT;
383 offset &= ~PAGE_MASK;
386 BUG_ON(offset >= PAGE_SIZE);
388 bytes = PAGE_SIZE - offset;
393 gnttab_foreach_grant_in_range(page, offset, bytes,
394 xenvif_gop_frag_copy_grant,
401 BUG_ON(!PageCompound(page));
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 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
448 meta = npo->meta + npo->meta_prod++;
449 meta->gso_type = gso_type;
450 meta->gso_size = skb_shinfo(skb)->gso_size;
455 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req);
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)
572 && (skb = xenvif_rx_dequeue(queue)) != NULL) {
573 queue->last_rx_time = jiffies;
575 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
577 __skb_queue_tail(&rxq, skb);
580 BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
585 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
586 gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
588 while ((skb = __skb_dequeue(&rxq)) != NULL) {
590 if ((1 << queue->meta[npo.meta_cons].gso_type) &
591 queue->vif->gso_prefix_mask) {
592 resp = RING_GET_RESPONSE(&queue->rx,
593 queue->rx.rsp_prod_pvt++);
595 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
597 resp->offset = queue->meta[npo.meta_cons].gso_size;
598 resp->id = queue->meta[npo.meta_cons].id;
599 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
602 XENVIF_RX_CB(skb)->meta_slots_used--;
606 queue->stats.tx_bytes += skb->len;
607 queue->stats.tx_packets++;
609 status = xenvif_check_gop(queue->vif,
610 XENVIF_RX_CB(skb)->meta_slots_used,
613 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
616 flags = XEN_NETRXF_more_data;
618 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
619 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
620 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
621 /* remote but checksummed. */
622 flags |= XEN_NETRXF_data_validated;
625 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
627 queue->meta[npo.meta_cons].size,
630 if ((1 << queue->meta[npo.meta_cons].gso_type) &
631 queue->vif->gso_mask) {
632 struct xen_netif_extra_info *gso =
633 (struct xen_netif_extra_info *)
634 RING_GET_RESPONSE(&queue->rx,
635 queue->rx.rsp_prod_pvt++);
637 resp->flags |= XEN_NETRXF_extra_info;
639 gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
640 gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
642 gso->u.gso.features = 0;
644 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
648 xenvif_add_frag_responses(queue, status,
649 queue->meta + npo.meta_cons + 1,
650 XENVIF_RX_CB(skb)->meta_slots_used);
652 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
654 need_to_notify |= !!ret;
656 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
662 notify_remote_via_irq(queue->rx_irq);
665 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
669 RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
672 napi_schedule(&queue->napi);
673 else if (xenvif_atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
674 &queue->eoi_pending) &
675 (NETBK_TX_EOI | NETBK_COMMON_EOI))
676 xen_irq_lateeoi(queue->tx_irq, 0);
679 static void tx_add_credit(struct xenvif_queue *queue)
681 unsigned long max_burst, max_credit;
684 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
685 * Otherwise the interface can seize up due to insufficient credit.
687 max_burst = max(131072UL, queue->credit_bytes);
689 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
690 max_credit = queue->remaining_credit + queue->credit_bytes;
691 if (max_credit < queue->remaining_credit)
692 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
694 queue->remaining_credit = min(max_credit, max_burst);
695 queue->rate_limited = false;
698 void xenvif_tx_credit_callback(unsigned long data)
700 struct xenvif_queue *queue = (struct xenvif_queue *)data;
701 tx_add_credit(queue);
702 xenvif_napi_schedule_or_enable_events(queue);
705 static void xenvif_tx_err(struct xenvif_queue *queue,
706 struct xen_netif_tx_request *txp, RING_IDX end)
708 RING_IDX cons = queue->tx.req_cons;
712 spin_lock_irqsave(&queue->response_lock, flags);
713 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
714 push_tx_responses(queue);
715 spin_unlock_irqrestore(&queue->response_lock, flags);
718 RING_COPY_REQUEST(&queue->tx, cons++, txp);
720 queue->tx.req_cons = cons;
723 static void xenvif_fatal_tx_err(struct xenvif *vif)
725 netdev_err(vif->dev, "fatal error; disabling device\n");
726 vif->disabled = true;
727 /* Disable the vif from queue 0's kthread */
729 xenvif_kick_thread(&vif->queues[0]);
732 static int xenvif_count_requests(struct xenvif_queue *queue,
733 struct xen_netif_tx_request *first,
734 struct xen_netif_tx_request *txp,
737 RING_IDX cons = queue->tx.req_cons;
742 if (!(first->flags & XEN_NETTXF_more_data))
746 struct xen_netif_tx_request dropped_tx = { 0 };
748 if (slots >= work_to_do) {
749 netdev_err(queue->vif->dev,
750 "Asked for %d slots but exceeds this limit\n",
752 xenvif_fatal_tx_err(queue->vif);
756 /* This guest is really using too many slots and
757 * considered malicious.
759 if (unlikely(slots >= fatal_skb_slots)) {
760 netdev_err(queue->vif->dev,
761 "Malicious frontend using %d slots, threshold %u\n",
762 slots, fatal_skb_slots);
763 xenvif_fatal_tx_err(queue->vif);
767 /* Xen network protocol had implicit dependency on
768 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
769 * the historical MAX_SKB_FRAGS value 18 to honor the
770 * same behavior as before. Any packet using more than
771 * 18 slots but less than fatal_skb_slots slots is
774 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
776 netdev_dbg(queue->vif->dev,
777 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
778 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
785 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
787 /* If the guest submitted a frame >= 64 KiB then
788 * first->size overflowed and following slots will
789 * appear to be larger than the frame.
791 * This cannot be fatal error as there are buggy
792 * frontends that do this.
794 * Consume all slots and drop the packet.
796 if (!drop_err && txp->size > first->size) {
798 netdev_dbg(queue->vif->dev,
799 "Invalid tx request, slot size %u > remaining size %u\n",
800 txp->size, first->size);
804 first->size -= txp->size;
807 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
808 netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
809 txp->offset, txp->size);
810 xenvif_fatal_tx_err(queue->vif);
814 more_data = txp->flags & XEN_NETTXF_more_data;
822 xenvif_tx_err(queue, first, cons + slots);
830 struct xenvif_tx_cb {
834 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
836 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
838 struct xen_netif_tx_request *txp,
839 struct gnttab_map_grant_ref *mop)
841 queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
842 gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
843 GNTMAP_host_map | GNTMAP_readonly,
844 txp->gref, queue->vif->domid);
846 memcpy(&queue->pending_tx_info[pending_idx].req, txp,
850 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
852 struct sk_buff *skb =
853 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
854 GFP_ATOMIC | __GFP_NOWARN);
855 if (unlikely(skb == NULL))
858 /* Packets passed to netif_rx() must have some headroom. */
859 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
861 /* Initialize it here to avoid later surprises */
862 skb_shinfo(skb)->destructor_arg = NULL;
867 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
869 struct xen_netif_tx_request *txp,
870 struct gnttab_map_grant_ref *gop,
871 unsigned int frag_overflow,
872 struct sk_buff *nskb)
874 struct skb_shared_info *shinfo = skb_shinfo(skb);
875 skb_frag_t *frags = shinfo->frags;
876 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
878 pending_ring_idx_t index;
879 unsigned int nr_slots;
881 nr_slots = shinfo->nr_frags;
883 /* Skip first skb fragment if it is on same page as header fragment. */
884 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
886 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
887 shinfo->nr_frags++, txp++, gop++) {
888 index = pending_index(queue->pending_cons++);
889 pending_idx = queue->pending_ring[index];
890 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
891 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
896 shinfo = skb_shinfo(nskb);
897 frags = shinfo->frags;
899 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
900 shinfo->nr_frags++, txp++, gop++) {
901 index = pending_index(queue->pending_cons++);
902 pending_idx = queue->pending_ring[index];
903 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
904 frag_set_pending_idx(&frags[shinfo->nr_frags],
908 skb_shinfo(skb)->frag_list = nskb;
914 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
916 grant_handle_t handle)
918 if (unlikely(queue->grant_tx_handle[pending_idx] !=
919 NETBACK_INVALID_HANDLE)) {
920 netdev_err(queue->vif->dev,
921 "Trying to overwrite active handle! pending_idx: 0x%x\n",
925 queue->grant_tx_handle[pending_idx] = handle;
928 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
931 if (unlikely(queue->grant_tx_handle[pending_idx] ==
932 NETBACK_INVALID_HANDLE)) {
933 netdev_err(queue->vif->dev,
934 "Trying to unmap invalid handle! pending_idx: 0x%x\n",
938 queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
941 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
943 struct gnttab_map_grant_ref **gopp_map,
944 struct gnttab_copy **gopp_copy)
946 struct gnttab_map_grant_ref *gop_map = *gopp_map;
947 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
948 /* This always points to the shinfo of the skb being checked, which
949 * could be either the first or the one on the frag_list
951 struct skb_shared_info *shinfo = skb_shinfo(skb);
952 /* If this is non-NULL, we are currently checking the frag_list skb, and
953 * this points to the shinfo of the first one
955 struct skb_shared_info *first_shinfo = NULL;
956 int nr_frags = shinfo->nr_frags;
957 const bool sharedslot = nr_frags &&
958 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
961 /* Check status of header. */
962 err = (*gopp_copy)->status;
965 netdev_dbg(queue->vif->dev,
966 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
967 (*gopp_copy)->status,
969 (*gopp_copy)->source.u.ref);
970 /* The first frag might still have this slot mapped */
972 xenvif_idx_release(queue, pending_idx,
973 XEN_NETIF_RSP_ERROR);
978 for (i = 0; i < nr_frags; i++, gop_map++) {
981 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
983 /* Check error status: if okay then remember grant handle. */
984 newerr = gop_map->status;
986 if (likely(!newerr)) {
987 xenvif_grant_handle_set(queue,
990 /* Had a previous error? Invalidate this fragment. */
992 xenvif_idx_unmap(queue, pending_idx);
993 /* If the mapping of the first frag was OK, but
994 * the header's copy failed, and they are
995 * sharing a slot, send an error
997 if (i == 0 && !first_shinfo && sharedslot)
998 xenvif_idx_release(queue, pending_idx,
999 XEN_NETIF_RSP_ERROR);
1001 xenvif_idx_release(queue, pending_idx,
1002 XEN_NETIF_RSP_OKAY);
1007 /* Error on this fragment: respond to client with an error. */
1008 if (net_ratelimit())
1009 netdev_dbg(queue->vif->dev,
1010 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1016 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
1018 /* Not the first error? Preceding frags already invalidated. */
1022 /* First error: if the header haven't shared a slot with the
1023 * first frag, release it as well.
1026 xenvif_idx_release(queue,
1027 XENVIF_TX_CB(skb)->pending_idx,
1028 XEN_NETIF_RSP_OKAY);
1030 /* Invalidate preceding fragments of this skb. */
1031 for (j = 0; j < i; j++) {
1032 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1033 xenvif_idx_unmap(queue, pending_idx);
1034 xenvif_idx_release(queue, pending_idx,
1035 XEN_NETIF_RSP_OKAY);
1038 /* And if we found the error while checking the frag_list, unmap
1039 * the first skb's frags
1042 for (j = 0; j < first_shinfo->nr_frags; j++) {
1043 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1044 xenvif_idx_unmap(queue, pending_idx);
1045 xenvif_idx_release(queue, pending_idx,
1046 XEN_NETIF_RSP_OKAY);
1050 /* Remember the error: invalidate all subsequent fragments. */
1054 if (skb_has_frag_list(skb) && !first_shinfo) {
1055 first_shinfo = skb_shinfo(skb);
1056 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1057 nr_frags = shinfo->nr_frags;
1062 *gopp_map = gop_map;
1066 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1068 struct skb_shared_info *shinfo = skb_shinfo(skb);
1069 int nr_frags = shinfo->nr_frags;
1071 u16 prev_pending_idx = INVALID_PENDING_IDX;
1073 for (i = 0; i < nr_frags; i++) {
1074 skb_frag_t *frag = shinfo->frags + i;
1075 struct xen_netif_tx_request *txp;
1079 pending_idx = frag_get_pending_idx(frag);
1081 /* If this is not the first frag, chain it to the previous*/
1082 if (prev_pending_idx == INVALID_PENDING_IDX)
1083 skb_shinfo(skb)->destructor_arg =
1084 &callback_param(queue, pending_idx);
1086 callback_param(queue, prev_pending_idx).ctx =
1087 &callback_param(queue, pending_idx);
1089 callback_param(queue, pending_idx).ctx = NULL;
1090 prev_pending_idx = pending_idx;
1092 txp = &queue->pending_tx_info[pending_idx].req;
1093 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1094 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1095 skb->len += txp->size;
1096 skb->data_len += txp->size;
1097 skb->truesize += txp->size;
1099 /* Take an extra reference to offset network stack's put_page */
1100 get_page(queue->mmap_pages[pending_idx]);
1104 static int xenvif_get_extras(struct xenvif_queue *queue,
1105 struct xen_netif_extra_info *extras,
1108 struct xen_netif_extra_info extra;
1109 RING_IDX cons = queue->tx.req_cons;
1112 if (unlikely(work_to_do-- <= 0)) {
1113 netdev_err(queue->vif->dev, "Missing extra info\n");
1114 xenvif_fatal_tx_err(queue->vif);
1118 RING_COPY_REQUEST(&queue->tx, cons, &extra);
1119 if (unlikely(!extra.type ||
1120 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1121 queue->tx.req_cons = ++cons;
1122 netdev_err(queue->vif->dev,
1123 "Invalid extra type: %d\n", extra.type);
1124 xenvif_fatal_tx_err(queue->vif);
1128 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1129 queue->tx.req_cons = ++cons;
1130 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1135 static int xenvif_set_skb_gso(struct xenvif *vif,
1136 struct sk_buff *skb,
1137 struct xen_netif_extra_info *gso)
1139 if (!gso->u.gso.size) {
1140 netdev_err(vif->dev, "GSO size must not be zero.\n");
1141 xenvif_fatal_tx_err(vif);
1145 switch (gso->u.gso.type) {
1146 case XEN_NETIF_GSO_TYPE_TCPV4:
1147 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1149 case XEN_NETIF_GSO_TYPE_TCPV6:
1150 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1153 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1154 xenvif_fatal_tx_err(vif);
1158 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1159 /* gso_segs will be calculated later */
1164 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1166 bool recalculate_partial_csum = false;
1168 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1169 * peers can fail to set NETRXF_csum_blank when sending a GSO
1170 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1171 * recalculate the partial checksum.
1173 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1174 queue->stats.rx_gso_checksum_fixup++;
1175 skb->ip_summed = CHECKSUM_PARTIAL;
1176 recalculate_partial_csum = true;
1179 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1180 if (skb->ip_summed != CHECKSUM_PARTIAL)
1183 return skb_checksum_setup(skb, recalculate_partial_csum);
1186 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1188 u64 now = get_jiffies_64();
1189 u64 next_credit = queue->credit_window_start +
1190 msecs_to_jiffies(queue->credit_usec / 1000);
1192 /* Timer could already be pending in rare cases. */
1193 if (timer_pending(&queue->credit_timeout)) {
1194 queue->rate_limited = true;
1198 /* Passed the point where we can replenish credit? */
1199 if (time_after_eq64(now, next_credit)) {
1200 queue->credit_window_start = now;
1201 tx_add_credit(queue);
1204 /* Still too big to send right now? Set a callback. */
1205 if (size > queue->remaining_credit) {
1206 queue->credit_timeout.data =
1207 (unsigned long)queue;
1208 mod_timer(&queue->credit_timeout,
1210 queue->credit_window_start = next_credit;
1211 queue->rate_limited = true;
1219 /* No locking is required in xenvif_mcast_add/del() as they are
1220 * only ever invoked from NAPI poll. An RCU list is used because
1221 * xenvif_mcast_match() is called asynchronously, during start_xmit.
1224 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
1226 struct xenvif_mcast_addr *mcast;
1228 if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
1229 if (net_ratelimit())
1230 netdev_err(vif->dev,
1231 "Too many multicast addresses\n");
1235 mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
1239 ether_addr_copy(mcast->addr, addr);
1240 list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
1241 vif->fe_mcast_count++;
1246 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
1248 struct xenvif_mcast_addr *mcast;
1250 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1251 if (ether_addr_equal(addr, mcast->addr)) {
1252 --vif->fe_mcast_count;
1253 list_del_rcu(&mcast->entry);
1254 kfree_rcu(mcast, rcu);
1260 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
1262 struct xenvif_mcast_addr *mcast;
1265 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1266 if (ether_addr_equal(addr, mcast->addr)) {
1276 void xenvif_mcast_addr_list_free(struct xenvif *vif)
1278 /* No need for locking or RCU here. NAPI poll and TX queue
1281 while (!list_empty(&vif->fe_mcast_addr)) {
1282 struct xenvif_mcast_addr *mcast;
1284 mcast = list_first_entry(&vif->fe_mcast_addr,
1285 struct xenvif_mcast_addr,
1287 --vif->fe_mcast_count;
1288 list_del(&mcast->entry);
1293 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1298 struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
1299 struct sk_buff *skb, *nskb;
1301 unsigned int frag_overflow;
1303 while (skb_queue_len(&queue->tx_queue) < budget) {
1304 struct xen_netif_tx_request txreq;
1305 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1306 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1310 unsigned int data_len;
1311 pending_ring_idx_t index;
1313 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1314 XEN_NETIF_TX_RING_SIZE) {
1315 netdev_err(queue->vif->dev,
1316 "Impossible number of requests. "
1317 "req_prod %d, req_cons %d, size %ld\n",
1318 queue->tx.sring->req_prod, queue->tx.req_cons,
1319 XEN_NETIF_TX_RING_SIZE);
1320 xenvif_fatal_tx_err(queue->vif);
1324 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1328 idx = queue->tx.req_cons;
1329 rmb(); /* Ensure that we see the request before we copy it. */
1330 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
1332 /* Credit-based scheduling. */
1333 if (txreq.size > queue->remaining_credit &&
1334 tx_credit_exceeded(queue, txreq.size))
1337 queue->remaining_credit -= txreq.size;
1340 queue->tx.req_cons = ++idx;
1342 memset(extras, 0, sizeof(extras));
1343 if (txreq.flags & XEN_NETTXF_extra_info) {
1344 work_to_do = xenvif_get_extras(queue, extras,
1346 idx = queue->tx.req_cons;
1347 if (unlikely(work_to_do < 0))
1351 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
1352 struct xen_netif_extra_info *extra;
1354 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
1355 ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
1357 make_tx_response(queue, &txreq,
1359 XEN_NETIF_RSP_OKAY :
1360 XEN_NETIF_RSP_ERROR);
1361 push_tx_responses(queue);
1365 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
1366 struct xen_netif_extra_info *extra;
1368 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
1369 xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
1371 make_tx_response(queue, &txreq, XEN_NETIF_RSP_OKAY);
1372 push_tx_responses(queue);
1376 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1377 if (unlikely(ret < 0))
1382 if (unlikely(txreq.size < ETH_HLEN)) {
1383 netdev_dbg(queue->vif->dev,
1384 "Bad packet size: %d\n", txreq.size);
1385 xenvif_tx_err(queue, &txreq, idx);
1389 /* No crossing a page as the payload mustn't fragment. */
1390 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1391 netdev_err(queue->vif->dev,
1392 "txreq.offset: %u, size: %u, end: %lu\n",
1393 txreq.offset, txreq.size,
1394 (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
1395 xenvif_fatal_tx_err(queue->vif);
1399 index = pending_index(queue->pending_cons);
1400 pending_idx = queue->pending_ring[index];
1402 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1403 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1404 XEN_NETBACK_TX_COPY_LEN : txreq.size;
1406 skb = xenvif_alloc_skb(data_len);
1407 if (unlikely(skb == NULL)) {
1408 netdev_dbg(queue->vif->dev,
1409 "Can't allocate a skb in start_xmit.\n");
1410 xenvif_tx_err(queue, &txreq, idx);
1414 skb_shinfo(skb)->nr_frags = ret;
1415 if (data_len < txreq.size)
1416 skb_shinfo(skb)->nr_frags++;
1417 /* At this point shinfo->nr_frags is in fact the number of
1418 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1422 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1423 frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1424 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1425 skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1426 nskb = xenvif_alloc_skb(0);
1427 if (unlikely(nskb == NULL)) {
1428 skb_shinfo(skb)->nr_frags = 0;
1430 xenvif_tx_err(queue, &txreq, idx);
1431 if (net_ratelimit())
1432 netdev_err(queue->vif->dev,
1433 "Can't allocate the frag_list skb.\n");
1438 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1439 struct xen_netif_extra_info *gso;
1440 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1442 if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1443 /* Failure in xenvif_set_skb_gso is fatal. */
1444 skb_shinfo(skb)->nr_frags = 0;
1451 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1453 __skb_put(skb, data_len);
1454 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1455 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1456 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1458 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1459 virt_to_gfn(skb->data);
1460 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1461 queue->tx_copy_ops[*copy_ops].dest.offset =
1462 offset_in_page(skb->data) & ~XEN_PAGE_MASK;
1464 queue->tx_copy_ops[*copy_ops].len = data_len;
1465 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1469 if (data_len < txreq.size) {
1470 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1472 xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1475 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1476 INVALID_PENDING_IDX);
1477 memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1481 queue->pending_cons++;
1483 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1484 frag_overflow, nskb);
1486 __skb_queue_tail(&queue->tx_queue, skb);
1488 queue->tx.req_cons = idx;
1490 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1491 (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1495 (*map_ops) = gop - queue->tx_map_ops;
1499 /* Consolidate skb with a frag_list into a brand new one with local pages on
1500 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1502 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1504 unsigned int offset = skb_headlen(skb);
1505 skb_frag_t frags[MAX_SKB_FRAGS];
1507 struct ubuf_info *uarg;
1508 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1510 queue->stats.tx_zerocopy_sent += 2;
1511 queue->stats.tx_frag_overflow++;
1513 xenvif_fill_frags(queue, nskb);
1514 /* Subtract frags size, we will correct it later */
1515 skb->truesize -= skb->data_len;
1516 skb->len += nskb->len;
1517 skb->data_len += nskb->len;
1519 /* create a brand new frags array and coalesce there */
1520 for (i = 0; offset < skb->len; i++) {
1524 BUG_ON(i >= MAX_SKB_FRAGS);
1525 page = alloc_page(GFP_ATOMIC);
1528 skb->truesize += skb->data_len;
1529 for (j = 0; j < i; j++)
1530 put_page(frags[j].page.p);
1534 if (offset + PAGE_SIZE < skb->len)
1537 len = skb->len - offset;
1538 if (skb_copy_bits(skb, offset, page_address(page), len))
1542 frags[i].page.p = page;
1543 frags[i].page_offset = 0;
1544 skb_frag_size_set(&frags[i], len);
1547 /* Release all the original (foreign) frags. */
1548 for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1549 skb_frag_unref(skb, f);
1550 uarg = skb_shinfo(skb)->destructor_arg;
1551 /* increase inflight counter to offset decrement in callback */
1552 atomic_inc(&queue->inflight_packets);
1553 uarg->callback(uarg, true);
1554 skb_shinfo(skb)->destructor_arg = NULL;
1556 /* Fill the skb with the new (local) frags. */
1557 memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1558 skb_shinfo(skb)->nr_frags = i;
1559 skb->truesize += i * PAGE_SIZE;
1564 static int xenvif_tx_submit(struct xenvif_queue *queue)
1566 struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1567 struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1568 struct sk_buff *skb;
1571 while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1572 struct xen_netif_tx_request *txp;
1576 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1577 txp = &queue->pending_tx_info[pending_idx].req;
1579 /* Check the remap error code. */
1580 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1581 /* If there was an error, xenvif_tx_check_gop is
1582 * expected to release all the frags which were mapped,
1583 * so kfree_skb shouldn't do it again
1585 skb_shinfo(skb)->nr_frags = 0;
1586 if (skb_has_frag_list(skb)) {
1587 struct sk_buff *nskb =
1588 skb_shinfo(skb)->frag_list;
1589 skb_shinfo(nskb)->nr_frags = 0;
1595 data_len = skb->len;
1596 callback_param(queue, pending_idx).ctx = NULL;
1597 if (data_len < txp->size) {
1598 /* Append the packet payload as a fragment. */
1599 txp->offset += data_len;
1600 txp->size -= data_len;
1602 /* Schedule a response immediately. */
1603 xenvif_idx_release(queue, pending_idx,
1604 XEN_NETIF_RSP_OKAY);
1607 if (txp->flags & XEN_NETTXF_csum_blank)
1608 skb->ip_summed = CHECKSUM_PARTIAL;
1609 else if (txp->flags & XEN_NETTXF_data_validated)
1610 skb->ip_summed = CHECKSUM_UNNECESSARY;
1612 xenvif_fill_frags(queue, skb);
1614 if (unlikely(skb_has_frag_list(skb))) {
1615 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1616 xenvif_skb_zerocopy_prepare(queue, nskb);
1617 if (xenvif_handle_frag_list(queue, skb)) {
1618 if (net_ratelimit())
1619 netdev_err(queue->vif->dev,
1620 "Not enough memory to consolidate frag_list!\n");
1621 xenvif_skb_zerocopy_prepare(queue, skb);
1625 /* Copied all the bits from the frag list -- free it. */
1626 skb_frag_list_init(skb);
1630 skb->dev = queue->vif->dev;
1631 skb->protocol = eth_type_trans(skb, skb->dev);
1632 skb_reset_network_header(skb);
1634 if (checksum_setup(queue, skb)) {
1635 netdev_dbg(queue->vif->dev,
1636 "Can't setup checksum in net_tx_action\n");
1637 /* We have to set this flag to trigger the callback */
1638 if (skb_shinfo(skb)->destructor_arg)
1639 xenvif_skb_zerocopy_prepare(queue, skb);
1644 skb_probe_transport_header(skb, 0);
1646 /* If the packet is GSO then we will have just set up the
1647 * transport header offset in checksum_setup so it's now
1648 * straightforward to calculate gso_segs.
1650 if (skb_is_gso(skb)) {
1651 int mss = skb_shinfo(skb)->gso_size;
1652 int hdrlen = skb_transport_header(skb) -
1653 skb_mac_header(skb) +
1656 skb_shinfo(skb)->gso_segs =
1657 DIV_ROUND_UP(skb->len - hdrlen, mss);
1660 queue->stats.rx_bytes += skb->len;
1661 queue->stats.rx_packets++;
1665 /* Set this flag right before netif_receive_skb, otherwise
1666 * someone might think this packet already left netback, and
1667 * do a skb_copy_ubufs while we are still in control of the
1668 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1670 if (skb_shinfo(skb)->destructor_arg) {
1671 xenvif_skb_zerocopy_prepare(queue, skb);
1672 queue->stats.tx_zerocopy_sent++;
1675 netif_receive_skb(skb);
1681 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1683 unsigned long flags;
1684 pending_ring_idx_t index;
1685 struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1687 /* This is the only place where we grab this lock, to protect callbacks
1690 spin_lock_irqsave(&queue->callback_lock, flags);
1692 u16 pending_idx = ubuf->desc;
1693 ubuf = (struct ubuf_info *) ubuf->ctx;
1694 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1696 index = pending_index(queue->dealloc_prod);
1697 queue->dealloc_ring[index] = pending_idx;
1698 /* Sync with xenvif_tx_dealloc_action:
1699 * insert idx then incr producer.
1702 queue->dealloc_prod++;
1704 spin_unlock_irqrestore(&queue->callback_lock, flags);
1706 if (likely(zerocopy_success))
1707 queue->stats.tx_zerocopy_success++;
1709 queue->stats.tx_zerocopy_fail++;
1710 xenvif_skb_zerocopy_complete(queue);
1713 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1715 struct gnttab_unmap_grant_ref *gop;
1716 pending_ring_idx_t dc, dp;
1717 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1720 dc = queue->dealloc_cons;
1721 gop = queue->tx_unmap_ops;
1723 /* Free up any grants we have finished using */
1725 dp = queue->dealloc_prod;
1727 /* Ensure we see all indices enqueued by all
1728 * xenvif_zerocopy_callback().
1733 BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1735 queue->dealloc_ring[pending_index(dc++)];
1737 pending_idx_release[gop - queue->tx_unmap_ops] =
1739 queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1740 queue->mmap_pages[pending_idx];
1741 gnttab_set_unmap_op(gop,
1742 idx_to_kaddr(queue, pending_idx),
1744 queue->grant_tx_handle[pending_idx]);
1745 xenvif_grant_handle_reset(queue, pending_idx);
1749 } while (dp != queue->dealloc_prod);
1751 queue->dealloc_cons = dc;
1753 if (gop - queue->tx_unmap_ops > 0) {
1755 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1757 queue->pages_to_unmap,
1758 gop - queue->tx_unmap_ops);
1760 netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1761 gop - queue->tx_unmap_ops, ret);
1762 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1763 if (gop[i].status != GNTST_okay)
1764 netdev_err(queue->vif->dev,
1765 " host_addr: 0x%llx handle: 0x%x status: %d\n",
1774 for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1775 xenvif_idx_release(queue, pending_idx_release[i],
1776 XEN_NETIF_RSP_OKAY);
1780 /* Called after netfront has transmitted */
1781 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1783 unsigned nr_mops, nr_cops = 0;
1786 if (unlikely(!tx_work_todo(queue)))
1789 xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1794 gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1796 ret = gnttab_map_refs(queue->tx_map_ops,
1798 queue->pages_to_map,
1803 netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1805 for (i = 0; i < nr_mops; ++i)
1806 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1811 work_done = xenvif_tx_submit(queue);
1816 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1819 struct pending_tx_info *pending_tx_info;
1820 pending_ring_idx_t index;
1821 unsigned long flags;
1823 pending_tx_info = &queue->pending_tx_info[pending_idx];
1825 spin_lock_irqsave(&queue->response_lock, flags);
1827 make_tx_response(queue, &pending_tx_info->req, status);
1829 /* Release the pending index before pusing the Tx response so
1830 * its available before a new Tx request is pushed by the
1833 index = pending_index(queue->pending_prod++);
1834 queue->pending_ring[index] = pending_idx;
1836 push_tx_responses(queue);
1838 spin_unlock_irqrestore(&queue->response_lock, flags);
1842 static void make_tx_response(struct xenvif_queue *queue,
1843 struct xen_netif_tx_request *txp,
1846 RING_IDX i = queue->tx.rsp_prod_pvt;
1847 struct xen_netif_tx_response *resp;
1849 resp = RING_GET_RESPONSE(&queue->tx, i);
1853 if (txp->flags & XEN_NETTXF_extra_info)
1854 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1856 queue->tx.rsp_prod_pvt = ++i;
1859 static void push_tx_responses(struct xenvif_queue *queue)
1863 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1865 notify_remote_via_irq(queue->tx_irq);
1868 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1875 RING_IDX i = queue->rx.rsp_prod_pvt;
1876 struct xen_netif_rx_response *resp;
1878 resp = RING_GET_RESPONSE(&queue->rx, i);
1879 resp->offset = offset;
1880 resp->flags = flags;
1882 resp->status = (s16)size;
1884 resp->status = (s16)st;
1886 queue->rx.rsp_prod_pvt = ++i;
1891 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1894 struct gnttab_unmap_grant_ref tx_unmap_op;
1896 gnttab_set_unmap_op(&tx_unmap_op,
1897 idx_to_kaddr(queue, pending_idx),
1899 queue->grant_tx_handle[pending_idx]);
1900 xenvif_grant_handle_reset(queue, pending_idx);
1902 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1903 &queue->mmap_pages[pending_idx], 1);
1905 netdev_err(queue->vif->dev,
1906 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1909 tx_unmap_op.host_addr,
1911 tx_unmap_op.status);
1916 static inline int tx_work_todo(struct xenvif_queue *queue)
1918 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1924 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1926 return queue->dealloc_cons != queue->dealloc_prod;
1929 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1931 if (queue->tx.sring)
1932 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1934 if (queue->rx.sring)
1935 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1939 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1940 grant_ref_t tx_ring_ref,
1941 grant_ref_t rx_ring_ref)
1944 struct xen_netif_tx_sring *txs;
1945 struct xen_netif_rx_sring *rxs;
1949 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1950 &tx_ring_ref, 1, &addr);
1954 txs = (struct xen_netif_tx_sring *)addr;
1955 BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1957 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1958 &rx_ring_ref, 1, &addr);
1962 rxs = (struct xen_netif_rx_sring *)addr;
1963 BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1968 xenvif_unmap_frontend_rings(queue);
1972 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1974 struct xenvif *vif = queue->vif;
1976 queue->stalled = true;
1978 /* At least one queue has stalled? Disable the carrier. */
1979 spin_lock(&vif->lock);
1980 if (vif->stalled_queues++ == 0) {
1981 netdev_info(vif->dev, "Guest Rx stalled");
1982 netif_carrier_off(vif->dev);
1984 spin_unlock(&vif->lock);
1987 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1989 struct xenvif *vif = queue->vif;
1991 queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1992 queue->stalled = false;
1994 /* All queues are ready? Enable the carrier. */
1995 spin_lock(&vif->lock);
1996 if (--vif->stalled_queues == 0) {
1997 netdev_info(vif->dev, "Guest Rx ready");
1998 netif_carrier_on(vif->dev);
2000 spin_unlock(&vif->lock);
2003 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
2005 RING_IDX prod, cons;
2007 prod = queue->rx.sring->req_prod;
2008 cons = queue->rx.req_cons;
2010 return !queue->stalled && prod - cons < 1
2011 && time_after(jiffies,
2012 queue->last_rx_time + queue->vif->stall_timeout);
2015 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
2017 RING_IDX prod, cons;
2019 prod = queue->rx.sring->req_prod;
2020 cons = queue->rx.req_cons;
2022 return queue->stalled && prod - cons >= 1;
2025 bool xenvif_have_rx_work(struct xenvif_queue *queue, bool test_kthread)
2027 return (!skb_queue_empty(&queue->rx_queue)
2028 && xenvif_rx_ring_slots_available(queue))
2029 || (queue->vif->stall_timeout &&
2030 (xenvif_rx_queue_stalled(queue)
2031 || xenvif_rx_queue_ready(queue)))
2032 || (test_kthread && kthread_should_stop())
2033 || queue->vif->disabled;
2036 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
2038 struct sk_buff *skb;
2041 skb = skb_peek(&queue->rx_queue);
2043 return MAX_SCHEDULE_TIMEOUT;
2045 timeout = XENVIF_RX_CB(skb)->expires - jiffies;
2046 return timeout < 0 ? 0 : timeout;
2049 /* Wait until the guest Rx thread has work.
2051 * The timeout needs to be adjusted based on the current head of the
2052 * queue (and not just the head at the beginning). In particular, if
2053 * the queue is initially empty an infinite timeout is used and this
2054 * needs to be reduced when a skb is queued.
2056 * This cannot be done with wait_event_timeout() because it only
2057 * calculates the timeout once.
2059 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
2063 if (xenvif_have_rx_work(queue, true))
2069 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
2070 if (xenvif_have_rx_work(queue, true))
2072 if (xenvif_atomic_fetch_andnot(NETBK_RX_EOI | NETBK_COMMON_EOI,
2073 &queue->eoi_pending) &
2074 (NETBK_RX_EOI | NETBK_COMMON_EOI))
2075 xen_irq_lateeoi(queue->rx_irq, 0);
2077 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
2081 finish_wait(&queue->wq, &wait);
2084 int xenvif_kthread_guest_rx(void *data)
2086 struct xenvif_queue *queue = data;
2087 struct xenvif *vif = queue->vif;
2089 if (!vif->stall_timeout)
2090 xenvif_queue_carrier_on(queue);
2093 xenvif_wait_for_rx_work(queue);
2095 if (kthread_should_stop())
2098 /* This frontend is found to be rogue, disable it in
2099 * kthread context. Currently this is only set when
2100 * netback finds out frontend sends malformed packet,
2101 * but we cannot disable the interface in softirq
2102 * context so we defer it here, if this thread is
2103 * associated with queue 0.
2105 if (unlikely(vif->disabled && queue->id == 0)) {
2106 xenvif_carrier_off(vif);
2110 if (!skb_queue_empty(&queue->rx_queue))
2111 xenvif_rx_action(queue);
2113 /* If the guest hasn't provided any Rx slots for a
2114 * while it's probably not responsive, drop the
2115 * carrier so packets are dropped earlier.
2117 if (vif->stall_timeout) {
2118 if (xenvif_rx_queue_stalled(queue))
2119 xenvif_queue_carrier_off(queue);
2120 else if (xenvif_rx_queue_ready(queue))
2121 xenvif_queue_carrier_on(queue);
2124 /* Queued packets may have foreign pages from other
2125 * domains. These cannot be queued indefinitely as
2126 * this would starve guests of grant refs and transmit
2129 xenvif_rx_queue_drop_expired(queue);
2131 xenvif_rx_queue_maybe_wake(queue);
2136 /* Bin any remaining skbs */
2137 xenvif_rx_queue_purge(queue);
2142 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2144 /* Dealloc thread must remain running until all inflight
2147 return kthread_should_stop() &&
2148 !atomic_read(&queue->inflight_packets);
2151 int xenvif_dealloc_kthread(void *data)
2153 struct xenvif_queue *queue = data;
2156 wait_event_interruptible(queue->dealloc_wq,
2157 tx_dealloc_work_todo(queue) ||
2158 xenvif_dealloc_kthread_should_stop(queue));
2159 if (xenvif_dealloc_kthread_should_stop(queue))
2162 xenvif_tx_dealloc_action(queue);
2166 /* Unmap anything remaining*/
2167 if (tx_dealloc_work_todo(queue))
2168 xenvif_tx_dealloc_action(queue);
2173 static int __init netback_init(void)
2180 /* Allow as many queues as there are CPUs but max. 8 if user has not
2181 * specified a value.
2183 if (xenvif_max_queues == 0)
2184 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
2187 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2188 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2189 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2190 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2193 rc = xenvif_xenbus_init();
2197 #ifdef CONFIG_DEBUG_FS
2198 xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2199 if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2200 pr_warn("Init of debugfs returned %ld!\n",
2201 PTR_ERR(xen_netback_dbg_root));
2202 #endif /* CONFIG_DEBUG_FS */
2210 module_init(netback_init);
2212 static void __exit netback_fini(void)
2214 #ifdef CONFIG_DEBUG_FS
2215 if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2216 debugfs_remove_recursive(xen_netback_dbg_root);
2217 #endif /* CONFIG_DEBUG_FS */
2218 xenvif_xenbus_fini();
2220 module_exit(netback_fini);
2222 MODULE_LICENSE("Dual BSD/GPL");
2223 MODULE_ALIAS("xen-backend:vif");