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 int xenvif_rx_ring_slots_needed(struct xenvif *vif)
155 return DIV_ROUND_UP(vif->dev->gso_max_size, XEN_PAGE_SIZE) + 1;
157 return DIV_ROUND_UP(vif->dev->mtu, XEN_PAGE_SIZE);
160 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
165 needed = xenvif_rx_ring_slots_needed(queue->vif);
168 prod = queue->rx.sring->req_prod;
169 cons = queue->rx.req_cons;
171 if (prod - cons >= needed)
174 queue->rx.sring->req_event = prod + 1;
176 /* Make sure event is visible before we check prod
180 } while (queue->rx.sring->req_prod != prod);
185 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
189 spin_lock_irqsave(&queue->rx_queue.lock, flags);
191 __skb_queue_tail(&queue->rx_queue, skb);
193 queue->rx_queue_len += skb->len;
194 if (queue->rx_queue_len > queue->rx_queue_max)
195 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
197 spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
200 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
204 spin_lock_irq(&queue->rx_queue.lock);
206 skb = __skb_dequeue(&queue->rx_queue);
208 queue->rx_queue_len -= skb->len;
210 spin_unlock_irq(&queue->rx_queue.lock);
215 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
217 spin_lock_irq(&queue->rx_queue.lock);
219 if (queue->rx_queue_len < queue->rx_queue_max)
220 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
222 spin_unlock_irq(&queue->rx_queue.lock);
226 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
229 while ((skb = xenvif_rx_dequeue(queue)) != NULL)
233 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
238 skb = skb_peek(&queue->rx_queue);
241 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
243 xenvif_rx_dequeue(queue);
248 struct netrx_pending_operations {
249 unsigned copy_prod, copy_cons;
250 unsigned meta_prod, meta_cons;
251 struct gnttab_copy *copy;
252 struct xenvif_rx_meta *meta;
254 grant_ref_t copy_gref;
257 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
258 struct netrx_pending_operations *npo)
260 struct xenvif_rx_meta *meta;
261 struct xen_netif_rx_request *req;
263 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
265 meta = npo->meta + npo->meta_prod++;
266 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
272 npo->copy_gref = req->gref;
277 struct gop_frag_copy {
278 struct xenvif_queue *queue;
279 struct netrx_pending_operations *npo;
280 struct xenvif_rx_meta *meta;
287 static void xenvif_setup_copy_gop(unsigned long gfn,
290 struct gop_frag_copy *info)
292 struct gnttab_copy *copy_gop;
293 struct xen_page_foreign *foreign;
294 /* Convenient aliases */
295 struct xenvif_queue *queue = info->queue;
296 struct netrx_pending_operations *npo = info->npo;
297 struct page *page = info->page;
299 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
301 if (npo->copy_off == MAX_BUFFER_OFFSET)
302 info->meta = get_next_rx_buffer(queue, npo);
304 if (npo->copy_off + *len > MAX_BUFFER_OFFSET)
305 *len = MAX_BUFFER_OFFSET - npo->copy_off;
307 copy_gop = npo->copy + npo->copy_prod++;
308 copy_gop->flags = GNTCOPY_dest_gref;
309 copy_gop->len = *len;
311 foreign = xen_page_foreign(page);
313 copy_gop->source.domid = foreign->domid;
314 copy_gop->source.u.ref = foreign->gref;
315 copy_gop->flags |= GNTCOPY_source_gref;
317 copy_gop->source.domid = DOMID_SELF;
318 copy_gop->source.u.gmfn = gfn;
320 copy_gop->source.offset = offset;
322 copy_gop->dest.domid = queue->vif->domid;
323 copy_gop->dest.offset = npo->copy_off;
324 copy_gop->dest.u.ref = npo->copy_gref;
326 npo->copy_off += *len;
327 info->meta->size += *len;
329 /* Leave a gap for the GSO descriptor. */
330 if (info->head && ((1 << info->gso_type) & queue->vif->gso_mask))
331 queue->rx.req_cons++;
333 info->head = 0; /* There must be something in this buffer now */
336 static void xenvif_gop_frag_copy_grant(unsigned long gfn,
345 xenvif_setup_copy_gop(gfn, offset, &bytes, data);
352 * Set up the grant operations for this fragment. If it's a flipping
353 * interface, we also set up the unmap request from here.
355 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
356 struct netrx_pending_operations *npo,
357 struct page *page, unsigned long size,
358 unsigned long offset, int *head)
360 struct gop_frag_copy info = {
364 .gso_type = XEN_NETIF_GSO_TYPE_NONE,
368 if (skb_is_gso(skb)) {
369 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
370 info.gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
371 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
372 info.gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
375 /* Data must not cross a page boundary. */
376 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
378 info.meta = npo->meta + npo->meta_prod - 1;
380 /* Skip unused frames from start of page */
381 page += offset >> PAGE_SHIFT;
382 offset &= ~PAGE_MASK;
385 BUG_ON(offset >= PAGE_SIZE);
387 bytes = PAGE_SIZE - offset;
392 gnttab_foreach_grant_in_range(page, offset, bytes,
393 xenvif_gop_frag_copy_grant,
400 BUG_ON(!PageCompound(page));
409 * Prepare an SKB to be transmitted to the frontend.
411 * This function is responsible for allocating grant operations, meta
414 * It returns the number of meta structures consumed. The number of
415 * ring slots used is always equal to the number of meta slots used
416 * plus the number of GSO descriptors used. Currently, we use either
417 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
418 * frontend-side LRO).
420 static int xenvif_gop_skb(struct sk_buff *skb,
421 struct netrx_pending_operations *npo,
422 struct xenvif_queue *queue)
424 struct xenvif *vif = netdev_priv(skb->dev);
425 int nr_frags = skb_shinfo(skb)->nr_frags;
427 struct xen_netif_rx_request *req;
428 struct xenvif_rx_meta *meta;
434 old_meta_prod = npo->meta_prod;
436 gso_type = XEN_NETIF_GSO_TYPE_NONE;
437 if (skb_is_gso(skb)) {
438 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
439 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
440 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
441 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
444 /* Set up a GSO prefix descriptor, if necessary */
445 if ((1 << gso_type) & vif->gso_prefix_mask) {
446 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
447 meta = npo->meta + npo->meta_prod++;
448 meta->gso_type = gso_type;
449 meta->gso_size = skb_shinfo(skb)->gso_size;
454 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
455 meta = npo->meta + npo->meta_prod++;
457 if ((1 << gso_type) & vif->gso_mask) {
458 meta->gso_type = gso_type;
459 meta->gso_size = skb_shinfo(skb)->gso_size;
461 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
468 npo->copy_gref = req->gref;
471 while (data < skb_tail_pointer(skb)) {
472 unsigned int offset = offset_in_page(data);
473 unsigned int len = PAGE_SIZE - offset;
475 if (data + len > skb_tail_pointer(skb))
476 len = skb_tail_pointer(skb) - data;
478 xenvif_gop_frag_copy(queue, skb, npo,
479 virt_to_page(data), len, offset, &head);
483 for (i = 0; i < nr_frags; i++) {
484 xenvif_gop_frag_copy(queue, skb, npo,
485 skb_frag_page(&skb_shinfo(skb)->frags[i]),
486 skb_frag_size(&skb_shinfo(skb)->frags[i]),
487 skb_shinfo(skb)->frags[i].page_offset,
491 return npo->meta_prod - old_meta_prod;
495 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
496 * used to set up the operations on the top of
497 * netrx_pending_operations, which have since been done. Check that
498 * they didn't give any errors and advance over them.
500 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
501 struct netrx_pending_operations *npo)
503 struct gnttab_copy *copy_op;
504 int status = XEN_NETIF_RSP_OKAY;
507 for (i = 0; i < nr_meta_slots; i++) {
508 copy_op = npo->copy + npo->copy_cons++;
509 if (copy_op->status != GNTST_okay) {
511 "Bad status %d from copy to DOM%d.\n",
512 copy_op->status, vif->domid);
513 status = XEN_NETIF_RSP_ERROR;
520 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
521 struct xenvif_rx_meta *meta,
525 unsigned long offset;
527 /* No fragments used */
528 if (nr_meta_slots <= 1)
533 for (i = 0; i < nr_meta_slots; i++) {
535 if (i == nr_meta_slots - 1)
538 flags = XEN_NETRXF_more_data;
541 make_rx_response(queue, meta[i].id, status, offset,
542 meta[i].size, flags);
546 void xenvif_kick_thread(struct xenvif_queue *queue)
551 static void xenvif_rx_action(struct xenvif_queue *queue)
555 struct xen_netif_rx_response *resp;
556 struct sk_buff_head rxq;
560 unsigned long offset;
561 bool need_to_notify = false;
563 struct netrx_pending_operations npo = {
564 .copy = queue->grant_copy_op,
568 skb_queue_head_init(&rxq);
570 while (xenvif_rx_ring_slots_available(queue)
571 && (skb = xenvif_rx_dequeue(queue)) != NULL) {
572 queue->last_rx_time = jiffies;
574 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
576 __skb_queue_tail(&rxq, skb);
579 BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
584 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
585 gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
587 while ((skb = __skb_dequeue(&rxq)) != NULL) {
589 if ((1 << queue->meta[npo.meta_cons].gso_type) &
590 queue->vif->gso_prefix_mask) {
591 resp = RING_GET_RESPONSE(&queue->rx,
592 queue->rx.rsp_prod_pvt++);
594 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
596 resp->offset = queue->meta[npo.meta_cons].gso_size;
597 resp->id = queue->meta[npo.meta_cons].id;
598 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
601 XENVIF_RX_CB(skb)->meta_slots_used--;
605 queue->stats.tx_bytes += skb->len;
606 queue->stats.tx_packets++;
608 status = xenvif_check_gop(queue->vif,
609 XENVIF_RX_CB(skb)->meta_slots_used,
612 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
615 flags = XEN_NETRXF_more_data;
617 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
618 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
619 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
620 /* remote but checksummed. */
621 flags |= XEN_NETRXF_data_validated;
624 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
626 queue->meta[npo.meta_cons].size,
629 if ((1 << queue->meta[npo.meta_cons].gso_type) &
630 queue->vif->gso_mask) {
631 struct xen_netif_extra_info *gso =
632 (struct xen_netif_extra_info *)
633 RING_GET_RESPONSE(&queue->rx,
634 queue->rx.rsp_prod_pvt++);
636 resp->flags |= XEN_NETRXF_extra_info;
638 gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
639 gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
641 gso->u.gso.features = 0;
643 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
647 xenvif_add_frag_responses(queue, status,
648 queue->meta + npo.meta_cons + 1,
649 XENVIF_RX_CB(skb)->meta_slots_used);
651 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
653 need_to_notify |= !!ret;
655 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
661 notify_remote_via_irq(queue->rx_irq);
664 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
668 RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
671 napi_schedule(&queue->napi);
674 static void tx_add_credit(struct xenvif_queue *queue)
676 unsigned long max_burst, max_credit;
679 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
680 * Otherwise the interface can seize up due to insufficient credit.
682 max_burst = max(131072UL, queue->credit_bytes);
684 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
685 max_credit = queue->remaining_credit + queue->credit_bytes;
686 if (max_credit < queue->remaining_credit)
687 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
689 queue->remaining_credit = min(max_credit, max_burst);
692 void xenvif_tx_credit_callback(unsigned long data)
694 struct xenvif_queue *queue = (struct xenvif_queue *)data;
695 tx_add_credit(queue);
696 xenvif_napi_schedule_or_enable_events(queue);
699 static void xenvif_tx_err(struct xenvif_queue *queue,
700 struct xen_netif_tx_request *txp, RING_IDX end)
702 RING_IDX cons = queue->tx.req_cons;
706 spin_lock_irqsave(&queue->response_lock, flags);
707 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
708 push_tx_responses(queue);
709 spin_unlock_irqrestore(&queue->response_lock, flags);
712 txp = RING_GET_REQUEST(&queue->tx, cons++);
714 queue->tx.req_cons = cons;
717 static void xenvif_fatal_tx_err(struct xenvif *vif)
719 netdev_err(vif->dev, "fatal error; disabling device\n");
720 vif->disabled = true;
721 /* Disable the vif from queue 0's kthread */
723 xenvif_kick_thread(&vif->queues[0]);
726 static int xenvif_count_requests(struct xenvif_queue *queue,
727 struct xen_netif_tx_request *first,
728 struct xen_netif_tx_request *txp,
731 RING_IDX cons = queue->tx.req_cons;
736 if (!(first->flags & XEN_NETTXF_more_data))
740 struct xen_netif_tx_request dropped_tx = { 0 };
742 if (slots >= work_to_do) {
743 netdev_err(queue->vif->dev,
744 "Asked for %d slots but exceeds this limit\n",
746 xenvif_fatal_tx_err(queue->vif);
750 /* This guest is really using too many slots and
751 * considered malicious.
753 if (unlikely(slots >= fatal_skb_slots)) {
754 netdev_err(queue->vif->dev,
755 "Malicious frontend using %d slots, threshold %u\n",
756 slots, fatal_skb_slots);
757 xenvif_fatal_tx_err(queue->vif);
761 /* Xen network protocol had implicit dependency on
762 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
763 * the historical MAX_SKB_FRAGS value 18 to honor the
764 * same behavior as before. Any packet using more than
765 * 18 slots but less than fatal_skb_slots slots is
768 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
770 netdev_dbg(queue->vif->dev,
771 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
772 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
779 memcpy(txp, RING_GET_REQUEST(&queue->tx, cons + slots),
782 /* If the guest submitted a frame >= 64 KiB then
783 * first->size overflowed and following slots will
784 * appear to be larger than the frame.
786 * This cannot be fatal error as there are buggy
787 * frontends that do this.
789 * Consume all slots and drop the packet.
791 if (!drop_err && txp->size > first->size) {
793 netdev_dbg(queue->vif->dev,
794 "Invalid tx request, slot size %u > remaining size %u\n",
795 txp->size, first->size);
799 first->size -= txp->size;
802 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
803 netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
804 txp->offset, txp->size);
805 xenvif_fatal_tx_err(queue->vif);
809 more_data = txp->flags & XEN_NETTXF_more_data;
817 xenvif_tx_err(queue, first, cons + slots);
825 struct xenvif_tx_cb {
829 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
831 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
833 struct xen_netif_tx_request *txp,
834 struct gnttab_map_grant_ref *mop)
836 queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
837 gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
838 GNTMAP_host_map | GNTMAP_readonly,
839 txp->gref, queue->vif->domid);
841 memcpy(&queue->pending_tx_info[pending_idx].req, txp,
845 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
847 struct sk_buff *skb =
848 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
849 GFP_ATOMIC | __GFP_NOWARN);
850 if (unlikely(skb == NULL))
853 /* Packets passed to netif_rx() must have some headroom. */
854 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
856 /* Initialize it here to avoid later surprises */
857 skb_shinfo(skb)->destructor_arg = NULL;
862 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
864 struct xen_netif_tx_request *txp,
865 struct gnttab_map_grant_ref *gop,
866 unsigned int frag_overflow,
867 struct sk_buff *nskb)
869 struct skb_shared_info *shinfo = skb_shinfo(skb);
870 skb_frag_t *frags = shinfo->frags;
871 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
873 pending_ring_idx_t index;
874 unsigned int nr_slots;
876 nr_slots = shinfo->nr_frags;
878 /* Skip first skb fragment if it is on same page as header fragment. */
879 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
881 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
882 shinfo->nr_frags++, txp++, gop++) {
883 index = pending_index(queue->pending_cons++);
884 pending_idx = queue->pending_ring[index];
885 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
886 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
891 shinfo = skb_shinfo(nskb);
892 frags = shinfo->frags;
894 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
895 shinfo->nr_frags++, txp++, gop++) {
896 index = pending_index(queue->pending_cons++);
897 pending_idx = queue->pending_ring[index];
898 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
899 frag_set_pending_idx(&frags[shinfo->nr_frags],
903 skb_shinfo(skb)->frag_list = nskb;
909 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
911 grant_handle_t handle)
913 if (unlikely(queue->grant_tx_handle[pending_idx] !=
914 NETBACK_INVALID_HANDLE)) {
915 netdev_err(queue->vif->dev,
916 "Trying to overwrite active handle! pending_idx: 0x%x\n",
920 queue->grant_tx_handle[pending_idx] = handle;
923 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
926 if (unlikely(queue->grant_tx_handle[pending_idx] ==
927 NETBACK_INVALID_HANDLE)) {
928 netdev_err(queue->vif->dev,
929 "Trying to unmap invalid handle! pending_idx: 0x%x\n",
933 queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
936 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
938 struct gnttab_map_grant_ref **gopp_map,
939 struct gnttab_copy **gopp_copy)
941 struct gnttab_map_grant_ref *gop_map = *gopp_map;
942 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
943 /* This always points to the shinfo of the skb being checked, which
944 * could be either the first or the one on the frag_list
946 struct skb_shared_info *shinfo = skb_shinfo(skb);
947 /* If this is non-NULL, we are currently checking the frag_list skb, and
948 * this points to the shinfo of the first one
950 struct skb_shared_info *first_shinfo = NULL;
951 int nr_frags = shinfo->nr_frags;
952 const bool sharedslot = nr_frags &&
953 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
956 /* Check status of header. */
957 err = (*gopp_copy)->status;
960 netdev_dbg(queue->vif->dev,
961 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
962 (*gopp_copy)->status,
964 (*gopp_copy)->source.u.ref);
965 /* The first frag might still have this slot mapped */
967 xenvif_idx_release(queue, pending_idx,
968 XEN_NETIF_RSP_ERROR);
973 for (i = 0; i < nr_frags; i++, gop_map++) {
976 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
978 /* Check error status: if okay then remember grant handle. */
979 newerr = gop_map->status;
981 if (likely(!newerr)) {
982 xenvif_grant_handle_set(queue,
985 /* Had a previous error? Invalidate this fragment. */
987 xenvif_idx_unmap(queue, pending_idx);
988 /* If the mapping of the first frag was OK, but
989 * the header's copy failed, and they are
990 * sharing a slot, send an error
992 if (i == 0 && sharedslot)
993 xenvif_idx_release(queue, pending_idx,
994 XEN_NETIF_RSP_ERROR);
996 xenvif_idx_release(queue, pending_idx,
1002 /* Error on this fragment: respond to client with an error. */
1003 if (net_ratelimit())
1004 netdev_dbg(queue->vif->dev,
1005 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1011 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
1013 /* Not the first error? Preceding frags already invalidated. */
1017 /* First error: if the header haven't shared a slot with the
1018 * first frag, release it as well.
1021 xenvif_idx_release(queue,
1022 XENVIF_TX_CB(skb)->pending_idx,
1023 XEN_NETIF_RSP_OKAY);
1025 /* Invalidate preceding fragments of this skb. */
1026 for (j = 0; j < i; j++) {
1027 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1028 xenvif_idx_unmap(queue, pending_idx);
1029 xenvif_idx_release(queue, pending_idx,
1030 XEN_NETIF_RSP_OKAY);
1033 /* And if we found the error while checking the frag_list, unmap
1034 * the first skb's frags
1037 for (j = 0; j < first_shinfo->nr_frags; j++) {
1038 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1039 xenvif_idx_unmap(queue, pending_idx);
1040 xenvif_idx_release(queue, pending_idx,
1041 XEN_NETIF_RSP_OKAY);
1045 /* Remember the error: invalidate all subsequent fragments. */
1049 if (skb_has_frag_list(skb) && !first_shinfo) {
1050 first_shinfo = skb_shinfo(skb);
1051 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1052 nr_frags = shinfo->nr_frags;
1057 *gopp_map = gop_map;
1061 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1063 struct skb_shared_info *shinfo = skb_shinfo(skb);
1064 int nr_frags = shinfo->nr_frags;
1066 u16 prev_pending_idx = INVALID_PENDING_IDX;
1068 for (i = 0; i < nr_frags; i++) {
1069 skb_frag_t *frag = shinfo->frags + i;
1070 struct xen_netif_tx_request *txp;
1074 pending_idx = frag_get_pending_idx(frag);
1076 /* If this is not the first frag, chain it to the previous*/
1077 if (prev_pending_idx == INVALID_PENDING_IDX)
1078 skb_shinfo(skb)->destructor_arg =
1079 &callback_param(queue, pending_idx);
1081 callback_param(queue, prev_pending_idx).ctx =
1082 &callback_param(queue, pending_idx);
1084 callback_param(queue, pending_idx).ctx = NULL;
1085 prev_pending_idx = pending_idx;
1087 txp = &queue->pending_tx_info[pending_idx].req;
1088 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1089 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1090 skb->len += txp->size;
1091 skb->data_len += txp->size;
1092 skb->truesize += txp->size;
1094 /* Take an extra reference to offset network stack's put_page */
1095 get_page(queue->mmap_pages[pending_idx]);
1099 static int xenvif_get_extras(struct xenvif_queue *queue,
1100 struct xen_netif_extra_info *extras,
1103 struct xen_netif_extra_info extra;
1104 RING_IDX cons = queue->tx.req_cons;
1107 if (unlikely(work_to_do-- <= 0)) {
1108 netdev_err(queue->vif->dev, "Missing extra info\n");
1109 xenvif_fatal_tx_err(queue->vif);
1113 memcpy(&extra, RING_GET_REQUEST(&queue->tx, cons),
1115 if (unlikely(!extra.type ||
1116 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1117 queue->tx.req_cons = ++cons;
1118 netdev_err(queue->vif->dev,
1119 "Invalid extra type: %d\n", extra.type);
1120 xenvif_fatal_tx_err(queue->vif);
1124 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1125 queue->tx.req_cons = ++cons;
1126 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1131 static int xenvif_set_skb_gso(struct xenvif *vif,
1132 struct sk_buff *skb,
1133 struct xen_netif_extra_info *gso)
1135 if (!gso->u.gso.size) {
1136 netdev_err(vif->dev, "GSO size must not be zero.\n");
1137 xenvif_fatal_tx_err(vif);
1141 switch (gso->u.gso.type) {
1142 case XEN_NETIF_GSO_TYPE_TCPV4:
1143 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1145 case XEN_NETIF_GSO_TYPE_TCPV6:
1146 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1149 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1150 xenvif_fatal_tx_err(vif);
1154 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1155 /* gso_segs will be calculated later */
1160 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1162 bool recalculate_partial_csum = false;
1164 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1165 * peers can fail to set NETRXF_csum_blank when sending a GSO
1166 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1167 * recalculate the partial checksum.
1169 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1170 queue->stats.rx_gso_checksum_fixup++;
1171 skb->ip_summed = CHECKSUM_PARTIAL;
1172 recalculate_partial_csum = true;
1175 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1176 if (skb->ip_summed != CHECKSUM_PARTIAL)
1179 return skb_checksum_setup(skb, recalculate_partial_csum);
1182 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1184 u64 now = get_jiffies_64();
1185 u64 next_credit = queue->credit_window_start +
1186 msecs_to_jiffies(queue->credit_usec / 1000);
1188 /* Timer could already be pending in rare cases. */
1189 if (timer_pending(&queue->credit_timeout))
1192 /* Passed the point where we can replenish credit? */
1193 if (time_after_eq64(now, next_credit)) {
1194 queue->credit_window_start = now;
1195 tx_add_credit(queue);
1198 /* Still too big to send right now? Set a callback. */
1199 if (size > queue->remaining_credit) {
1200 queue->credit_timeout.data =
1201 (unsigned long)queue;
1202 mod_timer(&queue->credit_timeout,
1204 queue->credit_window_start = next_credit;
1212 /* No locking is required in xenvif_mcast_add/del() as they are
1213 * only ever invoked from NAPI poll. An RCU list is used because
1214 * xenvif_mcast_match() is called asynchronously, during start_xmit.
1217 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
1219 struct xenvif_mcast_addr *mcast;
1221 if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
1222 if (net_ratelimit())
1223 netdev_err(vif->dev,
1224 "Too many multicast addresses\n");
1228 mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
1232 ether_addr_copy(mcast->addr, addr);
1233 list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
1234 vif->fe_mcast_count++;
1239 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
1241 struct xenvif_mcast_addr *mcast;
1243 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1244 if (ether_addr_equal(addr, mcast->addr)) {
1245 --vif->fe_mcast_count;
1246 list_del_rcu(&mcast->entry);
1247 kfree_rcu(mcast, rcu);
1253 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
1255 struct xenvif_mcast_addr *mcast;
1258 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
1259 if (ether_addr_equal(addr, mcast->addr)) {
1269 void xenvif_mcast_addr_list_free(struct xenvif *vif)
1271 /* No need for locking or RCU here. NAPI poll and TX queue
1274 while (!list_empty(&vif->fe_mcast_addr)) {
1275 struct xenvif_mcast_addr *mcast;
1277 mcast = list_first_entry(&vif->fe_mcast_addr,
1278 struct xenvif_mcast_addr,
1280 --vif->fe_mcast_count;
1281 list_del(&mcast->entry);
1286 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1291 struct gnttab_map_grant_ref *gop = queue->tx_map_ops;
1292 struct sk_buff *skb, *nskb;
1294 unsigned int frag_overflow;
1296 while (skb_queue_len(&queue->tx_queue) < budget) {
1297 struct xen_netif_tx_request txreq;
1298 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1299 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1303 unsigned int data_len;
1304 pending_ring_idx_t index;
1306 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1307 XEN_NETIF_TX_RING_SIZE) {
1308 netdev_err(queue->vif->dev,
1309 "Impossible number of requests. "
1310 "req_prod %d, req_cons %d, size %ld\n",
1311 queue->tx.sring->req_prod, queue->tx.req_cons,
1312 XEN_NETIF_TX_RING_SIZE);
1313 xenvif_fatal_tx_err(queue->vif);
1317 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1321 idx = queue->tx.req_cons;
1322 rmb(); /* Ensure that we see the request before we copy it. */
1323 memcpy(&txreq, RING_GET_REQUEST(&queue->tx, idx), sizeof(txreq));
1325 /* Credit-based scheduling. */
1326 if (txreq.size > queue->remaining_credit &&
1327 tx_credit_exceeded(queue, txreq.size))
1330 queue->remaining_credit -= txreq.size;
1333 queue->tx.req_cons = ++idx;
1335 memset(extras, 0, sizeof(extras));
1336 if (txreq.flags & XEN_NETTXF_extra_info) {
1337 work_to_do = xenvif_get_extras(queue, extras,
1339 idx = queue->tx.req_cons;
1340 if (unlikely(work_to_do < 0))
1344 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
1345 struct xen_netif_extra_info *extra;
1347 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
1348 ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
1350 make_tx_response(queue, &txreq,
1352 XEN_NETIF_RSP_OKAY :
1353 XEN_NETIF_RSP_ERROR);
1354 push_tx_responses(queue);
1358 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
1359 struct xen_netif_extra_info *extra;
1361 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
1362 xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
1364 make_tx_response(queue, &txreq, XEN_NETIF_RSP_OKAY);
1365 push_tx_responses(queue);
1369 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1370 if (unlikely(ret < 0))
1375 if (unlikely(txreq.size < ETH_HLEN)) {
1376 netdev_dbg(queue->vif->dev,
1377 "Bad packet size: %d\n", txreq.size);
1378 xenvif_tx_err(queue, &txreq, idx);
1382 /* No crossing a page as the payload mustn't fragment. */
1383 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1384 netdev_err(queue->vif->dev,
1385 "txreq.offset: %u, size: %u, end: %lu\n",
1386 txreq.offset, txreq.size,
1387 (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size);
1388 xenvif_fatal_tx_err(queue->vif);
1392 index = pending_index(queue->pending_cons);
1393 pending_idx = queue->pending_ring[index];
1395 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1396 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1397 XEN_NETBACK_TX_COPY_LEN : txreq.size;
1399 skb = xenvif_alloc_skb(data_len);
1400 if (unlikely(skb == NULL)) {
1401 netdev_dbg(queue->vif->dev,
1402 "Can't allocate a skb in start_xmit.\n");
1403 xenvif_tx_err(queue, &txreq, idx);
1407 skb_shinfo(skb)->nr_frags = ret;
1408 if (data_len < txreq.size)
1409 skb_shinfo(skb)->nr_frags++;
1410 /* At this point shinfo->nr_frags is in fact the number of
1411 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1415 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1416 frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1417 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1418 skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1419 nskb = xenvif_alloc_skb(0);
1420 if (unlikely(nskb == NULL)) {
1422 xenvif_tx_err(queue, &txreq, idx);
1423 if (net_ratelimit())
1424 netdev_err(queue->vif->dev,
1425 "Can't allocate the frag_list skb.\n");
1430 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1431 struct xen_netif_extra_info *gso;
1432 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1434 if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1435 /* Failure in xenvif_set_skb_gso is fatal. */
1442 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1444 __skb_put(skb, data_len);
1445 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1446 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1447 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1449 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1450 virt_to_gfn(skb->data);
1451 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1452 queue->tx_copy_ops[*copy_ops].dest.offset =
1453 offset_in_page(skb->data) & ~XEN_PAGE_MASK;
1455 queue->tx_copy_ops[*copy_ops].len = data_len;
1456 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1460 if (data_len < txreq.size) {
1461 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1463 xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1466 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1467 INVALID_PENDING_IDX);
1468 memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1472 queue->pending_cons++;
1474 gop = xenvif_get_requests(queue, skb, txfrags, gop,
1475 frag_overflow, nskb);
1477 __skb_queue_tail(&queue->tx_queue, skb);
1479 queue->tx.req_cons = idx;
1481 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1482 (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1486 (*map_ops) = gop - queue->tx_map_ops;
1490 /* Consolidate skb with a frag_list into a brand new one with local pages on
1491 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1493 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1495 unsigned int offset = skb_headlen(skb);
1496 skb_frag_t frags[MAX_SKB_FRAGS];
1498 struct ubuf_info *uarg;
1499 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1501 queue->stats.tx_zerocopy_sent += 2;
1502 queue->stats.tx_frag_overflow++;
1504 xenvif_fill_frags(queue, nskb);
1505 /* Subtract frags size, we will correct it later */
1506 skb->truesize -= skb->data_len;
1507 skb->len += nskb->len;
1508 skb->data_len += nskb->len;
1510 /* create a brand new frags array and coalesce there */
1511 for (i = 0; offset < skb->len; i++) {
1515 BUG_ON(i >= MAX_SKB_FRAGS);
1516 page = alloc_page(GFP_ATOMIC);
1519 skb->truesize += skb->data_len;
1520 for (j = 0; j < i; j++)
1521 put_page(frags[j].page.p);
1525 if (offset + PAGE_SIZE < skb->len)
1528 len = skb->len - offset;
1529 if (skb_copy_bits(skb, offset, page_address(page), len))
1533 frags[i].page.p = page;
1534 frags[i].page_offset = 0;
1535 skb_frag_size_set(&frags[i], len);
1538 /* Copied all the bits from the frag list -- free it. */
1539 skb_frag_list_init(skb);
1540 xenvif_skb_zerocopy_prepare(queue, nskb);
1543 /* Release all the original (foreign) frags. */
1544 for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1545 skb_frag_unref(skb, f);
1546 uarg = skb_shinfo(skb)->destructor_arg;
1547 /* increase inflight counter to offset decrement in callback */
1548 atomic_inc(&queue->inflight_packets);
1549 uarg->callback(uarg, true);
1550 skb_shinfo(skb)->destructor_arg = NULL;
1552 /* Fill the skb with the new (local) frags. */
1553 memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1554 skb_shinfo(skb)->nr_frags = i;
1555 skb->truesize += i * PAGE_SIZE;
1560 static int xenvif_tx_submit(struct xenvif_queue *queue)
1562 struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1563 struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1564 struct sk_buff *skb;
1567 while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1568 struct xen_netif_tx_request *txp;
1572 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1573 txp = &queue->pending_tx_info[pending_idx].req;
1575 /* Check the remap error code. */
1576 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1577 /* If there was an error, xenvif_tx_check_gop is
1578 * expected to release all the frags which were mapped,
1579 * so kfree_skb shouldn't do it again
1581 skb_shinfo(skb)->nr_frags = 0;
1582 if (skb_has_frag_list(skb)) {
1583 struct sk_buff *nskb =
1584 skb_shinfo(skb)->frag_list;
1585 skb_shinfo(nskb)->nr_frags = 0;
1591 data_len = skb->len;
1592 callback_param(queue, pending_idx).ctx = NULL;
1593 if (data_len < txp->size) {
1594 /* Append the packet payload as a fragment. */
1595 txp->offset += data_len;
1596 txp->size -= data_len;
1598 /* Schedule a response immediately. */
1599 xenvif_idx_release(queue, pending_idx,
1600 XEN_NETIF_RSP_OKAY);
1603 if (txp->flags & XEN_NETTXF_csum_blank)
1604 skb->ip_summed = CHECKSUM_PARTIAL;
1605 else if (txp->flags & XEN_NETTXF_data_validated)
1606 skb->ip_summed = CHECKSUM_UNNECESSARY;
1608 xenvif_fill_frags(queue, skb);
1610 if (unlikely(skb_has_frag_list(skb))) {
1611 if (xenvif_handle_frag_list(queue, skb)) {
1612 if (net_ratelimit())
1613 netdev_err(queue->vif->dev,
1614 "Not enough memory to consolidate frag_list!\n");
1615 xenvif_skb_zerocopy_prepare(queue, skb);
1621 skb->dev = queue->vif->dev;
1622 skb->protocol = eth_type_trans(skb, skb->dev);
1623 skb_reset_network_header(skb);
1625 if (checksum_setup(queue, skb)) {
1626 netdev_dbg(queue->vif->dev,
1627 "Can't setup checksum in net_tx_action\n");
1628 /* We have to set this flag to trigger the callback */
1629 if (skb_shinfo(skb)->destructor_arg)
1630 xenvif_skb_zerocopy_prepare(queue, skb);
1635 skb_probe_transport_header(skb, 0);
1637 /* If the packet is GSO then we will have just set up the
1638 * transport header offset in checksum_setup so it's now
1639 * straightforward to calculate gso_segs.
1641 if (skb_is_gso(skb)) {
1642 int mss = skb_shinfo(skb)->gso_size;
1643 int hdrlen = skb_transport_header(skb) -
1644 skb_mac_header(skb) +
1647 skb_shinfo(skb)->gso_segs =
1648 DIV_ROUND_UP(skb->len - hdrlen, mss);
1651 queue->stats.rx_bytes += skb->len;
1652 queue->stats.rx_packets++;
1656 /* Set this flag right before netif_receive_skb, otherwise
1657 * someone might think this packet already left netback, and
1658 * do a skb_copy_ubufs while we are still in control of the
1659 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1661 if (skb_shinfo(skb)->destructor_arg) {
1662 xenvif_skb_zerocopy_prepare(queue, skb);
1663 queue->stats.tx_zerocopy_sent++;
1666 netif_receive_skb(skb);
1672 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1674 unsigned long flags;
1675 pending_ring_idx_t index;
1676 struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1678 /* This is the only place where we grab this lock, to protect callbacks
1681 spin_lock_irqsave(&queue->callback_lock, flags);
1683 u16 pending_idx = ubuf->desc;
1684 ubuf = (struct ubuf_info *) ubuf->ctx;
1685 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1687 index = pending_index(queue->dealloc_prod);
1688 queue->dealloc_ring[index] = pending_idx;
1689 /* Sync with xenvif_tx_dealloc_action:
1690 * insert idx then incr producer.
1693 queue->dealloc_prod++;
1695 spin_unlock_irqrestore(&queue->callback_lock, flags);
1697 if (likely(zerocopy_success))
1698 queue->stats.tx_zerocopy_success++;
1700 queue->stats.tx_zerocopy_fail++;
1701 xenvif_skb_zerocopy_complete(queue);
1704 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1706 struct gnttab_unmap_grant_ref *gop;
1707 pending_ring_idx_t dc, dp;
1708 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1711 dc = queue->dealloc_cons;
1712 gop = queue->tx_unmap_ops;
1714 /* Free up any grants we have finished using */
1716 dp = queue->dealloc_prod;
1718 /* Ensure we see all indices enqueued by all
1719 * xenvif_zerocopy_callback().
1724 BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1726 queue->dealloc_ring[pending_index(dc++)];
1728 pending_idx_release[gop - queue->tx_unmap_ops] =
1730 queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1731 queue->mmap_pages[pending_idx];
1732 gnttab_set_unmap_op(gop,
1733 idx_to_kaddr(queue, pending_idx),
1735 queue->grant_tx_handle[pending_idx]);
1736 xenvif_grant_handle_reset(queue, pending_idx);
1740 } while (dp != queue->dealloc_prod);
1742 queue->dealloc_cons = dc;
1744 if (gop - queue->tx_unmap_ops > 0) {
1746 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1748 queue->pages_to_unmap,
1749 gop - queue->tx_unmap_ops);
1751 netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1752 gop - queue->tx_unmap_ops, ret);
1753 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1754 if (gop[i].status != GNTST_okay)
1755 netdev_err(queue->vif->dev,
1756 " host_addr: 0x%llx handle: 0x%x status: %d\n",
1765 for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1766 xenvif_idx_release(queue, pending_idx_release[i],
1767 XEN_NETIF_RSP_OKAY);
1771 /* Called after netfront has transmitted */
1772 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1774 unsigned nr_mops, nr_cops = 0;
1777 if (unlikely(!tx_work_todo(queue)))
1780 xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1785 gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1787 ret = gnttab_map_refs(queue->tx_map_ops,
1789 queue->pages_to_map,
1794 work_done = xenvif_tx_submit(queue);
1799 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1802 struct pending_tx_info *pending_tx_info;
1803 pending_ring_idx_t index;
1804 unsigned long flags;
1806 pending_tx_info = &queue->pending_tx_info[pending_idx];
1808 spin_lock_irqsave(&queue->response_lock, flags);
1810 make_tx_response(queue, &pending_tx_info->req, status);
1812 /* Release the pending index before pusing the Tx response so
1813 * its available before a new Tx request is pushed by the
1816 index = pending_index(queue->pending_prod++);
1817 queue->pending_ring[index] = pending_idx;
1819 push_tx_responses(queue);
1821 spin_unlock_irqrestore(&queue->response_lock, flags);
1825 static void make_tx_response(struct xenvif_queue *queue,
1826 struct xen_netif_tx_request *txp,
1829 RING_IDX i = queue->tx.rsp_prod_pvt;
1830 struct xen_netif_tx_response *resp;
1832 resp = RING_GET_RESPONSE(&queue->tx, i);
1836 if (txp->flags & XEN_NETTXF_extra_info)
1837 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1839 queue->tx.rsp_prod_pvt = ++i;
1842 static void push_tx_responses(struct xenvif_queue *queue)
1846 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1848 notify_remote_via_irq(queue->tx_irq);
1851 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1858 RING_IDX i = queue->rx.rsp_prod_pvt;
1859 struct xen_netif_rx_response *resp;
1861 resp = RING_GET_RESPONSE(&queue->rx, i);
1862 resp->offset = offset;
1863 resp->flags = flags;
1865 resp->status = (s16)size;
1867 resp->status = (s16)st;
1869 queue->rx.rsp_prod_pvt = ++i;
1874 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1877 struct gnttab_unmap_grant_ref tx_unmap_op;
1879 gnttab_set_unmap_op(&tx_unmap_op,
1880 idx_to_kaddr(queue, pending_idx),
1882 queue->grant_tx_handle[pending_idx]);
1883 xenvif_grant_handle_reset(queue, pending_idx);
1885 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1886 &queue->mmap_pages[pending_idx], 1);
1888 netdev_err(queue->vif->dev,
1889 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1892 tx_unmap_op.host_addr,
1894 tx_unmap_op.status);
1899 static inline int tx_work_todo(struct xenvif_queue *queue)
1901 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1907 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1909 return queue->dealloc_cons != queue->dealloc_prod;
1912 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1914 if (queue->tx.sring)
1915 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1917 if (queue->rx.sring)
1918 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1922 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1923 grant_ref_t tx_ring_ref,
1924 grant_ref_t rx_ring_ref)
1927 struct xen_netif_tx_sring *txs;
1928 struct xen_netif_rx_sring *rxs;
1932 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1933 &tx_ring_ref, 1, &addr);
1937 txs = (struct xen_netif_tx_sring *)addr;
1938 BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1940 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1941 &rx_ring_ref, 1, &addr);
1945 rxs = (struct xen_netif_rx_sring *)addr;
1946 BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1951 xenvif_unmap_frontend_rings(queue);
1955 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1957 struct xenvif *vif = queue->vif;
1959 queue->stalled = true;
1961 /* At least one queue has stalled? Disable the carrier. */
1962 spin_lock(&vif->lock);
1963 if (vif->stalled_queues++ == 0) {
1964 netdev_info(vif->dev, "Guest Rx stalled");
1965 netif_carrier_off(vif->dev);
1967 spin_unlock(&vif->lock);
1970 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1972 struct xenvif *vif = queue->vif;
1974 queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1975 queue->stalled = false;
1977 /* All queues are ready? Enable the carrier. */
1978 spin_lock(&vif->lock);
1979 if (--vif->stalled_queues == 0) {
1980 netdev_info(vif->dev, "Guest Rx ready");
1981 netif_carrier_on(vif->dev);
1983 spin_unlock(&vif->lock);
1986 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
1988 RING_IDX prod, cons;
1990 prod = queue->rx.sring->req_prod;
1991 cons = queue->rx.req_cons;
1993 return !queue->stalled && prod - cons < 1
1994 && time_after(jiffies,
1995 queue->last_rx_time + queue->vif->stall_timeout);
1998 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
2000 RING_IDX prod, cons;
2002 prod = queue->rx.sring->req_prod;
2003 cons = queue->rx.req_cons;
2005 return queue->stalled && prod - cons >= 1;
2008 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
2010 return (!skb_queue_empty(&queue->rx_queue)
2011 && xenvif_rx_ring_slots_available(queue))
2012 || (queue->vif->stall_timeout &&
2013 (xenvif_rx_queue_stalled(queue)
2014 || xenvif_rx_queue_ready(queue)))
2015 || kthread_should_stop()
2016 || queue->vif->disabled;
2019 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
2021 struct sk_buff *skb;
2024 skb = skb_peek(&queue->rx_queue);
2026 return MAX_SCHEDULE_TIMEOUT;
2028 timeout = XENVIF_RX_CB(skb)->expires - jiffies;
2029 return timeout < 0 ? 0 : timeout;
2032 /* Wait until the guest Rx thread has work.
2034 * The timeout needs to be adjusted based on the current head of the
2035 * queue (and not just the head at the beginning). In particular, if
2036 * the queue is initially empty an infinite timeout is used and this
2037 * needs to be reduced when a skb is queued.
2039 * This cannot be done with wait_event_timeout() because it only
2040 * calculates the timeout once.
2042 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
2046 if (xenvif_have_rx_work(queue))
2052 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
2053 if (xenvif_have_rx_work(queue))
2055 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
2059 finish_wait(&queue->wq, &wait);
2062 int xenvif_kthread_guest_rx(void *data)
2064 struct xenvif_queue *queue = data;
2065 struct xenvif *vif = queue->vif;
2067 if (!vif->stall_timeout)
2068 xenvif_queue_carrier_on(queue);
2071 xenvif_wait_for_rx_work(queue);
2073 if (kthread_should_stop())
2076 /* This frontend is found to be rogue, disable it in
2077 * kthread context. Currently this is only set when
2078 * netback finds out frontend sends malformed packet,
2079 * but we cannot disable the interface in softirq
2080 * context so we defer it here, if this thread is
2081 * associated with queue 0.
2083 if (unlikely(vif->disabled && queue->id == 0)) {
2084 xenvif_carrier_off(vif);
2088 if (!skb_queue_empty(&queue->rx_queue))
2089 xenvif_rx_action(queue);
2091 /* If the guest hasn't provided any Rx slots for a
2092 * while it's probably not responsive, drop the
2093 * carrier so packets are dropped earlier.
2095 if (vif->stall_timeout) {
2096 if (xenvif_rx_queue_stalled(queue))
2097 xenvif_queue_carrier_off(queue);
2098 else if (xenvif_rx_queue_ready(queue))
2099 xenvif_queue_carrier_on(queue);
2102 /* Queued packets may have foreign pages from other
2103 * domains. These cannot be queued indefinitely as
2104 * this would starve guests of grant refs and transmit
2107 xenvif_rx_queue_drop_expired(queue);
2109 xenvif_rx_queue_maybe_wake(queue);
2114 /* Bin any remaining skbs */
2115 xenvif_rx_queue_purge(queue);
2120 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2122 /* Dealloc thread must remain running until all inflight
2125 return kthread_should_stop() &&
2126 !atomic_read(&queue->inflight_packets);
2129 int xenvif_dealloc_kthread(void *data)
2131 struct xenvif_queue *queue = data;
2134 wait_event_interruptible(queue->dealloc_wq,
2135 tx_dealloc_work_todo(queue) ||
2136 xenvif_dealloc_kthread_should_stop(queue));
2137 if (xenvif_dealloc_kthread_should_stop(queue))
2140 xenvif_tx_dealloc_action(queue);
2144 /* Unmap anything remaining*/
2145 if (tx_dealloc_work_todo(queue))
2146 xenvif_tx_dealloc_action(queue);
2151 static int __init netback_init(void)
2158 /* Allow as many queues as there are CPUs if user has not
2159 * specified a value.
2161 if (xenvif_max_queues == 0)
2162 xenvif_max_queues = num_online_cpus();
2164 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2165 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2166 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2167 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2170 rc = xenvif_xenbus_init();
2174 #ifdef CONFIG_DEBUG_FS
2175 xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2176 if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2177 pr_warn("Init of debugfs returned %ld!\n",
2178 PTR_ERR(xen_netback_dbg_root));
2179 #endif /* CONFIG_DEBUG_FS */
2187 module_init(netback_init);
2189 static void __exit netback_fini(void)
2191 #ifdef CONFIG_DEBUG_FS
2192 if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2193 debugfs_remove_recursive(xen_netback_dbg_root);
2194 #endif /* CONFIG_DEBUG_FS */
2195 xenvif_xenbus_fini();
2197 module_exit(netback_fini);
2199 MODULE_LICENSE("Dual BSD/GPL");
2200 MODULE_ALIAS("xen-backend:vif");