2 * Back-end of the driver for virtual network devices. This portion of the
3 * driver exports a 'unified' network-device interface that can be accessed
4 * by any operating system that implements a compatible front end. A
5 * reference front-end implementation can be found in:
6 * drivers/net/xen-netfront.c
8 * Copyright (c) 2002-2005, K A Fraser
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public License version 2
12 * as published by the Free Software Foundation; or, when distributed
13 * separately from the Linux kernel or incorporated into other
14 * software packages, subject to the following license:
16 * Permission is hereby granted, free of charge, to any person obtaining a copy
17 * of this source file (the "Software"), to deal in the Software without
18 * restriction, including without limitation the rights to use, copy, modify,
19 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
20 * and to permit persons to whom the Software is furnished to do so, subject to
21 * the following conditions:
23 * The above copyright notice and this permission notice shall be included in
24 * all copies or substantial portions of the Software.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
27 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
28 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
29 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
30 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
31 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
37 #include <linux/kthread.h>
38 #include <linux/if_vlan.h>
39 #include <linux/udp.h>
40 #include <linux/highmem.h>
45 #include <xen/events.h>
46 #include <xen/interface/memory.h>
48 #include <asm/xen/hypercall.h>
49 #include <asm/xen/page.h>
51 /* Provide an option to disable split event channels at load time as
52 * event channels are limited resource. Split event channels are
55 bool separate_tx_rx_irq = 1;
56 module_param(separate_tx_rx_irq, bool, 0644);
58 /* The time that packets can stay on the guest Rx internal queue
59 * before they are dropped.
61 unsigned int rx_drain_timeout_msecs = 10000;
62 module_param(rx_drain_timeout_msecs, uint, 0444);
64 /* The length of time before the frontend is considered unresponsive
65 * because it isn't providing Rx slots.
67 unsigned int rx_stall_timeout_msecs = 60000;
68 module_param(rx_stall_timeout_msecs, uint, 0444);
70 unsigned int xenvif_max_queues;
71 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
72 MODULE_PARM_DESC(max_queues,
73 "Maximum number of queues per virtual interface");
76 * This is the maximum slots a skb can have. If a guest sends a skb
77 * which exceeds this limit it is considered malicious.
79 #define FATAL_SKB_SLOTS_DEFAULT 20
80 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
81 module_param(fatal_skb_slots, uint, 0444);
83 /* The amount to copy out of the first guest Tx slot into the skb's
84 * linear area. If the first slot has more data, it will be mapped
85 * and put into the first frag.
87 * This is sized to avoid pulling headers from the frags for most
90 #define XEN_NETBACK_TX_COPY_LEN 128
93 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
96 static void make_tx_response(struct xenvif_queue *queue,
97 struct xen_netif_tx_request *txp,
100 static inline int tx_work_todo(struct xenvif_queue *queue);
102 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
109 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
112 return page_to_pfn(queue->mmap_pages[idx]);
115 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
118 return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
121 #define callback_param(vif, pending_idx) \
122 (vif->pending_tx_info[pending_idx].callback_struct)
124 /* Find the containing VIF's structure from a pointer in pending_tx_info array
126 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
128 u16 pending_idx = ubuf->desc;
129 struct pending_tx_info *temp =
130 container_of(ubuf, struct pending_tx_info, callback_struct);
131 return container_of(temp - pending_idx,
136 static u16 frag_get_pending_idx(skb_frag_t *frag)
138 return (u16)frag->page_offset;
141 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
143 frag->page_offset = pending_idx;
146 static inline pending_ring_idx_t pending_index(unsigned i)
148 return i & (MAX_PENDING_REQS-1);
151 bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue, int needed)
156 prod = queue->rx.sring->req_prod;
157 cons = queue->rx.req_cons;
159 if (prod - cons >= needed)
162 queue->rx.sring->req_event = prod + 1;
164 /* Make sure event is visible before we check prod
168 } while (queue->rx.sring->req_prod != prod);
173 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
177 spin_lock_irqsave(&queue->rx_queue.lock, flags);
179 __skb_queue_tail(&queue->rx_queue, skb);
181 queue->rx_queue_len += skb->len;
182 if (queue->rx_queue_len > queue->rx_queue_max)
183 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
185 spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
188 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
192 spin_lock_irq(&queue->rx_queue.lock);
194 skb = __skb_dequeue(&queue->rx_queue);
196 queue->rx_queue_len -= skb->len;
198 spin_unlock_irq(&queue->rx_queue.lock);
203 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue)
205 spin_lock_irq(&queue->rx_queue.lock);
207 if (queue->rx_queue_len < queue->rx_queue_max)
208 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id));
210 spin_unlock_irq(&queue->rx_queue.lock);
214 static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
217 while ((skb = xenvif_rx_dequeue(queue)) != NULL)
221 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
226 skb = skb_peek(&queue->rx_queue);
229 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
231 xenvif_rx_dequeue(queue);
236 struct netrx_pending_operations {
237 unsigned copy_prod, copy_cons;
238 unsigned meta_prod, meta_cons;
239 struct gnttab_copy *copy;
240 struct xenvif_rx_meta *meta;
242 grant_ref_t copy_gref;
245 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
246 struct netrx_pending_operations *npo)
248 struct xenvif_rx_meta *meta;
249 struct xen_netif_rx_request *req;
251 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
253 meta = npo->meta + npo->meta_prod++;
254 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
260 npo->copy_gref = req->gref;
266 * Set up the grant operations for this fragment. If it's a flipping
267 * interface, we also set up the unmap request from here.
269 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
270 struct netrx_pending_operations *npo,
271 struct page *page, unsigned long size,
272 unsigned long offset, int *head)
274 struct gnttab_copy *copy_gop;
275 struct xenvif_rx_meta *meta;
277 int gso_type = XEN_NETIF_GSO_TYPE_NONE;
279 /* Data must not cross a page boundary. */
280 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
282 meta = npo->meta + npo->meta_prod - 1;
284 /* Skip unused frames from start of page */
285 page += offset >> PAGE_SHIFT;
286 offset &= ~PAGE_MASK;
289 struct xen_page_foreign *foreign;
291 BUG_ON(offset >= PAGE_SIZE);
292 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
294 if (npo->copy_off == MAX_BUFFER_OFFSET)
295 meta = get_next_rx_buffer(queue, npo);
297 bytes = PAGE_SIZE - offset;
301 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
302 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
304 copy_gop = npo->copy + npo->copy_prod++;
305 copy_gop->flags = GNTCOPY_dest_gref;
306 copy_gop->len = bytes;
308 foreign = xen_page_foreign(page);
310 copy_gop->source.domid = foreign->domid;
311 copy_gop->source.u.ref = foreign->gref;
312 copy_gop->flags |= GNTCOPY_source_gref;
314 copy_gop->source.domid = DOMID_SELF;
315 copy_gop->source.u.gmfn =
316 virt_to_mfn(page_address(page));
318 copy_gop->source.offset = offset;
320 copy_gop->dest.domid = queue->vif->domid;
321 copy_gop->dest.offset = npo->copy_off;
322 copy_gop->dest.u.ref = npo->copy_gref;
324 npo->copy_off += bytes;
331 if (offset == PAGE_SIZE && size) {
332 BUG_ON(!PageCompound(page));
337 /* Leave a gap for the GSO descriptor. */
338 if (skb_is_gso(skb)) {
339 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
340 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
341 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
342 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
345 if (*head && ((1 << gso_type) & queue->vif->gso_mask))
346 queue->rx.req_cons++;
348 *head = 0; /* There must be something in this buffer now. */
354 * Prepare an SKB to be transmitted to the frontend.
356 * This function is responsible for allocating grant operations, meta
359 * It returns the number of meta structures consumed. The number of
360 * ring slots used is always equal to the number of meta slots used
361 * plus the number of GSO descriptors used. Currently, we use either
362 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
363 * frontend-side LRO).
365 static int xenvif_gop_skb(struct sk_buff *skb,
366 struct netrx_pending_operations *npo,
367 struct xenvif_queue *queue)
369 struct xenvif *vif = netdev_priv(skb->dev);
370 int nr_frags = skb_shinfo(skb)->nr_frags;
372 struct xen_netif_rx_request *req;
373 struct xenvif_rx_meta *meta;
379 old_meta_prod = npo->meta_prod;
381 gso_type = XEN_NETIF_GSO_TYPE_NONE;
382 if (skb_is_gso(skb)) {
383 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
384 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
385 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
386 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
389 /* Set up a GSO prefix descriptor, if necessary */
390 if ((1 << gso_type) & vif->gso_prefix_mask) {
391 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
392 meta = npo->meta + npo->meta_prod++;
393 meta->gso_type = gso_type;
394 meta->gso_size = skb_shinfo(skb)->gso_size;
399 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
400 meta = npo->meta + npo->meta_prod++;
402 if ((1 << gso_type) & vif->gso_mask) {
403 meta->gso_type = gso_type;
404 meta->gso_size = skb_shinfo(skb)->gso_size;
406 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
413 npo->copy_gref = req->gref;
416 while (data < skb_tail_pointer(skb)) {
417 unsigned int offset = offset_in_page(data);
418 unsigned int len = PAGE_SIZE - offset;
420 if (data + len > skb_tail_pointer(skb))
421 len = skb_tail_pointer(skb) - data;
423 xenvif_gop_frag_copy(queue, skb, npo,
424 virt_to_page(data), len, offset, &head);
428 for (i = 0; i < nr_frags; i++) {
429 xenvif_gop_frag_copy(queue, skb, npo,
430 skb_frag_page(&skb_shinfo(skb)->frags[i]),
431 skb_frag_size(&skb_shinfo(skb)->frags[i]),
432 skb_shinfo(skb)->frags[i].page_offset,
436 return npo->meta_prod - old_meta_prod;
440 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
441 * used to set up the operations on the top of
442 * netrx_pending_operations, which have since been done. Check that
443 * they didn't give any errors and advance over them.
445 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
446 struct netrx_pending_operations *npo)
448 struct gnttab_copy *copy_op;
449 int status = XEN_NETIF_RSP_OKAY;
452 for (i = 0; i < nr_meta_slots; i++) {
453 copy_op = npo->copy + npo->copy_cons++;
454 if (copy_op->status != GNTST_okay) {
456 "Bad status %d from copy to DOM%d.\n",
457 copy_op->status, vif->domid);
458 status = XEN_NETIF_RSP_ERROR;
465 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
466 struct xenvif_rx_meta *meta,
470 unsigned long offset;
472 /* No fragments used */
473 if (nr_meta_slots <= 1)
478 for (i = 0; i < nr_meta_slots; i++) {
480 if (i == nr_meta_slots - 1)
483 flags = XEN_NETRXF_more_data;
486 make_rx_response(queue, meta[i].id, status, offset,
487 meta[i].size, flags);
491 void xenvif_kick_thread(struct xenvif_queue *queue)
496 static void xenvif_rx_action(struct xenvif_queue *queue)
500 struct xen_netif_rx_response *resp;
501 struct sk_buff_head rxq;
505 unsigned long offset;
506 bool need_to_notify = false;
508 struct netrx_pending_operations npo = {
509 .copy = queue->grant_copy_op,
513 skb_queue_head_init(&rxq);
515 while (xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX)
516 && (skb = xenvif_rx_dequeue(queue)) != NULL) {
517 RING_IDX old_req_cons;
518 RING_IDX ring_slots_used;
520 queue->last_rx_time = jiffies;
522 old_req_cons = queue->rx.req_cons;
523 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
524 ring_slots_used = queue->rx.req_cons - old_req_cons;
526 __skb_queue_tail(&rxq, skb);
529 BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
534 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
535 gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
537 while ((skb = __skb_dequeue(&rxq)) != NULL) {
539 if ((1 << queue->meta[npo.meta_cons].gso_type) &
540 queue->vif->gso_prefix_mask) {
541 resp = RING_GET_RESPONSE(&queue->rx,
542 queue->rx.rsp_prod_pvt++);
544 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
546 resp->offset = queue->meta[npo.meta_cons].gso_size;
547 resp->id = queue->meta[npo.meta_cons].id;
548 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
551 XENVIF_RX_CB(skb)->meta_slots_used--;
555 queue->stats.tx_bytes += skb->len;
556 queue->stats.tx_packets++;
558 status = xenvif_check_gop(queue->vif,
559 XENVIF_RX_CB(skb)->meta_slots_used,
562 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
565 flags = XEN_NETRXF_more_data;
567 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
568 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
569 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
570 /* remote but checksummed. */
571 flags |= XEN_NETRXF_data_validated;
574 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
576 queue->meta[npo.meta_cons].size,
579 if ((1 << queue->meta[npo.meta_cons].gso_type) &
580 queue->vif->gso_mask) {
581 struct xen_netif_extra_info *gso =
582 (struct xen_netif_extra_info *)
583 RING_GET_RESPONSE(&queue->rx,
584 queue->rx.rsp_prod_pvt++);
586 resp->flags |= XEN_NETRXF_extra_info;
588 gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
589 gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
591 gso->u.gso.features = 0;
593 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
597 xenvif_add_frag_responses(queue, status,
598 queue->meta + npo.meta_cons + 1,
599 XENVIF_RX_CB(skb)->meta_slots_used);
601 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
603 need_to_notify |= !!ret;
605 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
611 notify_remote_via_irq(queue->rx_irq);
614 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
618 RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
621 napi_schedule(&queue->napi);
624 static void tx_add_credit(struct xenvif_queue *queue)
626 unsigned long max_burst, max_credit;
629 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
630 * Otherwise the interface can seize up due to insufficient credit.
632 max_burst = RING_GET_REQUEST(&queue->tx, queue->tx.req_cons)->size;
633 max_burst = min(max_burst, 131072UL);
634 max_burst = max(max_burst, queue->credit_bytes);
636 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
637 max_credit = queue->remaining_credit + queue->credit_bytes;
638 if (max_credit < queue->remaining_credit)
639 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
641 queue->remaining_credit = min(max_credit, max_burst);
644 static void tx_credit_callback(unsigned long data)
646 struct xenvif_queue *queue = (struct xenvif_queue *)data;
647 tx_add_credit(queue);
648 xenvif_napi_schedule_or_enable_events(queue);
651 static void xenvif_tx_err(struct xenvif_queue *queue,
652 struct xen_netif_tx_request *txp, RING_IDX end)
654 RING_IDX cons = queue->tx.req_cons;
658 spin_lock_irqsave(&queue->response_lock, flags);
659 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
660 spin_unlock_irqrestore(&queue->response_lock, flags);
663 txp = RING_GET_REQUEST(&queue->tx, cons++);
665 queue->tx.req_cons = cons;
668 static void xenvif_fatal_tx_err(struct xenvif *vif)
670 netdev_err(vif->dev, "fatal error; disabling device\n");
671 vif->disabled = true;
672 /* Disable the vif from queue 0's kthread */
674 xenvif_kick_thread(&vif->queues[0]);
677 static int xenvif_count_requests(struct xenvif_queue *queue,
678 struct xen_netif_tx_request *first,
679 struct xen_netif_tx_request *txp,
682 RING_IDX cons = queue->tx.req_cons;
687 if (!(first->flags & XEN_NETTXF_more_data))
691 struct xen_netif_tx_request dropped_tx = { 0 };
693 if (slots >= work_to_do) {
694 netdev_err(queue->vif->dev,
695 "Asked for %d slots but exceeds this limit\n",
697 xenvif_fatal_tx_err(queue->vif);
701 /* This guest is really using too many slots and
702 * considered malicious.
704 if (unlikely(slots >= fatal_skb_slots)) {
705 netdev_err(queue->vif->dev,
706 "Malicious frontend using %d slots, threshold %u\n",
707 slots, fatal_skb_slots);
708 xenvif_fatal_tx_err(queue->vif);
712 /* Xen network protocol had implicit dependency on
713 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
714 * the historical MAX_SKB_FRAGS value 18 to honor the
715 * same behavior as before. Any packet using more than
716 * 18 slots but less than fatal_skb_slots slots is
719 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
721 netdev_dbg(queue->vif->dev,
722 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
723 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
730 memcpy(txp, RING_GET_REQUEST(&queue->tx, cons + slots),
733 /* If the guest submitted a frame >= 64 KiB then
734 * first->size overflowed and following slots will
735 * appear to be larger than the frame.
737 * This cannot be fatal error as there are buggy
738 * frontends that do this.
740 * Consume all slots and drop the packet.
742 if (!drop_err && txp->size > first->size) {
744 netdev_dbg(queue->vif->dev,
745 "Invalid tx request, slot size %u > remaining size %u\n",
746 txp->size, first->size);
750 first->size -= txp->size;
753 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
754 netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
755 txp->offset, txp->size);
756 xenvif_fatal_tx_err(queue->vif);
760 more_data = txp->flags & XEN_NETTXF_more_data;
768 xenvif_tx_err(queue, first, cons + slots);
776 struct xenvif_tx_cb {
780 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
782 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
784 struct xen_netif_tx_request *txp,
785 struct gnttab_map_grant_ref *mop)
787 queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
788 gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
789 GNTMAP_host_map | GNTMAP_readonly,
790 txp->gref, queue->vif->domid);
792 memcpy(&queue->pending_tx_info[pending_idx].req, txp,
796 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
798 struct sk_buff *skb =
799 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
800 GFP_ATOMIC | __GFP_NOWARN);
801 if (unlikely(skb == NULL))
804 /* Packets passed to netif_rx() must have some headroom. */
805 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
807 /* Initialize it here to avoid later surprises */
808 skb_shinfo(skb)->destructor_arg = NULL;
813 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
815 struct xen_netif_tx_request *txp,
816 struct gnttab_map_grant_ref *gop)
818 struct skb_shared_info *shinfo = skb_shinfo(skb);
819 skb_frag_t *frags = shinfo->frags;
820 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
822 pending_ring_idx_t index;
823 unsigned int nr_slots, frag_overflow = 0;
825 /* At this point shinfo->nr_frags is in fact the number of
826 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
828 if (shinfo->nr_frags > MAX_SKB_FRAGS) {
829 frag_overflow = shinfo->nr_frags - MAX_SKB_FRAGS;
830 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
831 shinfo->nr_frags = MAX_SKB_FRAGS;
833 nr_slots = shinfo->nr_frags;
835 /* Skip first skb fragment if it is on same page as header fragment. */
836 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
838 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
839 shinfo->nr_frags++, txp++, gop++) {
840 index = pending_index(queue->pending_cons++);
841 pending_idx = queue->pending_ring[index];
842 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
843 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
847 struct sk_buff *nskb = xenvif_alloc_skb(0);
848 if (unlikely(nskb == NULL)) {
850 netdev_err(queue->vif->dev,
851 "Can't allocate the frag_list skb.\n");
855 shinfo = skb_shinfo(nskb);
856 frags = shinfo->frags;
858 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
859 shinfo->nr_frags++, txp++, gop++) {
860 index = pending_index(queue->pending_cons++);
861 pending_idx = queue->pending_ring[index];
862 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
863 frag_set_pending_idx(&frags[shinfo->nr_frags],
867 skb_shinfo(skb)->frag_list = nskb;
873 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
875 grant_handle_t handle)
877 if (unlikely(queue->grant_tx_handle[pending_idx] !=
878 NETBACK_INVALID_HANDLE)) {
879 netdev_err(queue->vif->dev,
880 "Trying to overwrite active handle! pending_idx: %x\n",
884 queue->grant_tx_handle[pending_idx] = handle;
887 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
890 if (unlikely(queue->grant_tx_handle[pending_idx] ==
891 NETBACK_INVALID_HANDLE)) {
892 netdev_err(queue->vif->dev,
893 "Trying to unmap invalid handle! pending_idx: %x\n",
897 queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
900 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
902 struct gnttab_map_grant_ref **gopp_map,
903 struct gnttab_copy **gopp_copy)
905 struct gnttab_map_grant_ref *gop_map = *gopp_map;
906 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
907 /* This always points to the shinfo of the skb being checked, which
908 * could be either the first or the one on the frag_list
910 struct skb_shared_info *shinfo = skb_shinfo(skb);
911 /* If this is non-NULL, we are currently checking the frag_list skb, and
912 * this points to the shinfo of the first one
914 struct skb_shared_info *first_shinfo = NULL;
915 int nr_frags = shinfo->nr_frags;
916 const bool sharedslot = nr_frags &&
917 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
920 /* Check status of header. */
921 err = (*gopp_copy)->status;
924 netdev_dbg(queue->vif->dev,
925 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
926 (*gopp_copy)->status,
928 (*gopp_copy)->source.u.ref);
929 /* The first frag might still have this slot mapped */
931 xenvif_idx_release(queue, pending_idx,
932 XEN_NETIF_RSP_ERROR);
937 for (i = 0; i < nr_frags; i++, gop_map++) {
940 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
942 /* Check error status: if okay then remember grant handle. */
943 newerr = gop_map->status;
945 if (likely(!newerr)) {
946 xenvif_grant_handle_set(queue,
949 /* Had a previous error? Invalidate this fragment. */
951 xenvif_idx_unmap(queue, pending_idx);
952 /* If the mapping of the first frag was OK, but
953 * the header's copy failed, and they are
954 * sharing a slot, send an error
956 if (i == 0 && sharedslot)
957 xenvif_idx_release(queue, pending_idx,
958 XEN_NETIF_RSP_ERROR);
960 xenvif_idx_release(queue, pending_idx,
966 /* Error on this fragment: respond to client with an error. */
968 netdev_dbg(queue->vif->dev,
969 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
975 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
977 /* Not the first error? Preceding frags already invalidated. */
981 /* First error: if the header haven't shared a slot with the
982 * first frag, release it as well.
985 xenvif_idx_release(queue,
986 XENVIF_TX_CB(skb)->pending_idx,
989 /* Invalidate preceding fragments of this skb. */
990 for (j = 0; j < i; j++) {
991 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
992 xenvif_idx_unmap(queue, pending_idx);
993 xenvif_idx_release(queue, pending_idx,
997 /* And if we found the error while checking the frag_list, unmap
998 * the first skb's frags
1001 for (j = 0; j < first_shinfo->nr_frags; j++) {
1002 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1003 xenvif_idx_unmap(queue, pending_idx);
1004 xenvif_idx_release(queue, pending_idx,
1005 XEN_NETIF_RSP_OKAY);
1009 /* Remember the error: invalidate all subsequent fragments. */
1013 if (skb_has_frag_list(skb) && !first_shinfo) {
1014 first_shinfo = skb_shinfo(skb);
1015 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1016 nr_frags = shinfo->nr_frags;
1021 *gopp_map = gop_map;
1025 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1027 struct skb_shared_info *shinfo = skb_shinfo(skb);
1028 int nr_frags = shinfo->nr_frags;
1030 u16 prev_pending_idx = INVALID_PENDING_IDX;
1032 for (i = 0; i < nr_frags; i++) {
1033 skb_frag_t *frag = shinfo->frags + i;
1034 struct xen_netif_tx_request *txp;
1038 pending_idx = frag_get_pending_idx(frag);
1040 /* If this is not the first frag, chain it to the previous*/
1041 if (prev_pending_idx == INVALID_PENDING_IDX)
1042 skb_shinfo(skb)->destructor_arg =
1043 &callback_param(queue, pending_idx);
1045 callback_param(queue, prev_pending_idx).ctx =
1046 &callback_param(queue, pending_idx);
1048 callback_param(queue, pending_idx).ctx = NULL;
1049 prev_pending_idx = pending_idx;
1051 txp = &queue->pending_tx_info[pending_idx].req;
1052 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1053 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1054 skb->len += txp->size;
1055 skb->data_len += txp->size;
1056 skb->truesize += txp->size;
1058 /* Take an extra reference to offset network stack's put_page */
1059 get_page(queue->mmap_pages[pending_idx]);
1063 static int xenvif_get_extras(struct xenvif_queue *queue,
1064 struct xen_netif_extra_info *extras,
1067 struct xen_netif_extra_info extra;
1068 RING_IDX cons = queue->tx.req_cons;
1071 if (unlikely(work_to_do-- <= 0)) {
1072 netdev_err(queue->vif->dev, "Missing extra info\n");
1073 xenvif_fatal_tx_err(queue->vif);
1077 memcpy(&extra, RING_GET_REQUEST(&queue->tx, cons),
1079 if (unlikely(!extra.type ||
1080 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1081 queue->tx.req_cons = ++cons;
1082 netdev_err(queue->vif->dev,
1083 "Invalid extra type: %d\n", extra.type);
1084 xenvif_fatal_tx_err(queue->vif);
1088 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1089 queue->tx.req_cons = ++cons;
1090 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1095 static int xenvif_set_skb_gso(struct xenvif *vif,
1096 struct sk_buff *skb,
1097 struct xen_netif_extra_info *gso)
1099 if (!gso->u.gso.size) {
1100 netdev_err(vif->dev, "GSO size must not be zero.\n");
1101 xenvif_fatal_tx_err(vif);
1105 switch (gso->u.gso.type) {
1106 case XEN_NETIF_GSO_TYPE_TCPV4:
1107 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1109 case XEN_NETIF_GSO_TYPE_TCPV6:
1110 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1113 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1114 xenvif_fatal_tx_err(vif);
1118 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1119 /* gso_segs will be calculated later */
1124 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1126 bool recalculate_partial_csum = false;
1128 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1129 * peers can fail to set NETRXF_csum_blank when sending a GSO
1130 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1131 * recalculate the partial checksum.
1133 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1134 queue->stats.rx_gso_checksum_fixup++;
1135 skb->ip_summed = CHECKSUM_PARTIAL;
1136 recalculate_partial_csum = true;
1139 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1140 if (skb->ip_summed != CHECKSUM_PARTIAL)
1143 return skb_checksum_setup(skb, recalculate_partial_csum);
1146 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1148 u64 now = get_jiffies_64();
1149 u64 next_credit = queue->credit_window_start +
1150 msecs_to_jiffies(queue->credit_usec / 1000);
1152 /* Timer could already be pending in rare cases. */
1153 if (timer_pending(&queue->credit_timeout))
1156 /* Passed the point where we can replenish credit? */
1157 if (time_after_eq64(now, next_credit)) {
1158 queue->credit_window_start = now;
1159 tx_add_credit(queue);
1162 /* Still too big to send right now? Set a callback. */
1163 if (size > queue->remaining_credit) {
1164 queue->credit_timeout.data =
1165 (unsigned long)queue;
1166 queue->credit_timeout.function =
1168 mod_timer(&queue->credit_timeout,
1170 queue->credit_window_start = next_credit;
1178 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1183 struct gnttab_map_grant_ref *gop = queue->tx_map_ops, *request_gop;
1184 struct sk_buff *skb;
1187 while (skb_queue_len(&queue->tx_queue) < budget) {
1188 struct xen_netif_tx_request txreq;
1189 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1190 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1194 unsigned int data_len;
1195 pending_ring_idx_t index;
1197 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1198 XEN_NETIF_TX_RING_SIZE) {
1199 netdev_err(queue->vif->dev,
1200 "Impossible number of requests. "
1201 "req_prod %d, req_cons %d, size %ld\n",
1202 queue->tx.sring->req_prod, queue->tx.req_cons,
1203 XEN_NETIF_TX_RING_SIZE);
1204 xenvif_fatal_tx_err(queue->vif);
1208 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1212 idx = queue->tx.req_cons;
1213 rmb(); /* Ensure that we see the request before we copy it. */
1214 memcpy(&txreq, RING_GET_REQUEST(&queue->tx, idx), sizeof(txreq));
1216 /* Credit-based scheduling. */
1217 if (txreq.size > queue->remaining_credit &&
1218 tx_credit_exceeded(queue, txreq.size))
1221 queue->remaining_credit -= txreq.size;
1224 queue->tx.req_cons = ++idx;
1226 memset(extras, 0, sizeof(extras));
1227 if (txreq.flags & XEN_NETTXF_extra_info) {
1228 work_to_do = xenvif_get_extras(queue, extras,
1230 idx = queue->tx.req_cons;
1231 if (unlikely(work_to_do < 0))
1235 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1236 if (unlikely(ret < 0))
1241 if (unlikely(txreq.size < ETH_HLEN)) {
1242 netdev_dbg(queue->vif->dev,
1243 "Bad packet size: %d\n", txreq.size);
1244 xenvif_tx_err(queue, &txreq, idx);
1248 /* No crossing a page as the payload mustn't fragment. */
1249 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1250 netdev_err(queue->vif->dev,
1251 "txreq.offset: %x, size: %u, end: %lu\n",
1252 txreq.offset, txreq.size,
1253 (txreq.offset&~PAGE_MASK) + txreq.size);
1254 xenvif_fatal_tx_err(queue->vif);
1258 index = pending_index(queue->pending_cons);
1259 pending_idx = queue->pending_ring[index];
1261 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN &&
1262 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1263 XEN_NETBACK_TX_COPY_LEN : txreq.size;
1265 skb = xenvif_alloc_skb(data_len);
1266 if (unlikely(skb == NULL)) {
1267 netdev_dbg(queue->vif->dev,
1268 "Can't allocate a skb in start_xmit.\n");
1269 xenvif_tx_err(queue, &txreq, idx);
1273 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1274 struct xen_netif_extra_info *gso;
1275 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1277 if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1278 /* Failure in xenvif_set_skb_gso is fatal. */
1284 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1286 __skb_put(skb, data_len);
1287 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1288 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1289 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1291 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1292 virt_to_mfn(skb->data);
1293 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1294 queue->tx_copy_ops[*copy_ops].dest.offset =
1295 offset_in_page(skb->data);
1297 queue->tx_copy_ops[*copy_ops].len = data_len;
1298 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1302 skb_shinfo(skb)->nr_frags = ret;
1303 if (data_len < txreq.size) {
1304 skb_shinfo(skb)->nr_frags++;
1305 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1307 xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1310 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1311 INVALID_PENDING_IDX);
1312 memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1316 queue->pending_cons++;
1318 request_gop = xenvif_get_requests(queue, skb, txfrags, gop);
1319 if (request_gop == NULL) {
1321 xenvif_tx_err(queue, &txreq, idx);
1326 __skb_queue_tail(&queue->tx_queue, skb);
1328 queue->tx.req_cons = idx;
1330 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1331 (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1335 (*map_ops) = gop - queue->tx_map_ops;
1339 /* Consolidate skb with a frag_list into a brand new one with local pages on
1340 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1342 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1344 unsigned int offset = skb_headlen(skb);
1345 skb_frag_t frags[MAX_SKB_FRAGS];
1347 struct ubuf_info *uarg;
1348 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1350 queue->stats.tx_zerocopy_sent += 2;
1351 queue->stats.tx_frag_overflow++;
1353 xenvif_fill_frags(queue, nskb);
1354 /* Subtract frags size, we will correct it later */
1355 skb->truesize -= skb->data_len;
1356 skb->len += nskb->len;
1357 skb->data_len += nskb->len;
1359 /* create a brand new frags array and coalesce there */
1360 for (i = 0; offset < skb->len; i++) {
1364 BUG_ON(i >= MAX_SKB_FRAGS);
1365 page = alloc_page(GFP_ATOMIC);
1368 skb->truesize += skb->data_len;
1369 for (j = 0; j < i; j++)
1370 put_page(frags[j].page.p);
1374 if (offset + PAGE_SIZE < skb->len)
1377 len = skb->len - offset;
1378 if (skb_copy_bits(skb, offset, page_address(page), len))
1382 frags[i].page.p = page;
1383 frags[i].page_offset = 0;
1384 skb_frag_size_set(&frags[i], len);
1386 /* swap out with old one */
1387 memcpy(skb_shinfo(skb)->frags,
1389 i * sizeof(skb_frag_t));
1390 skb_shinfo(skb)->nr_frags = i;
1391 skb->truesize += i * PAGE_SIZE;
1393 /* remove traces of mapped pages and frag_list */
1394 skb_frag_list_init(skb);
1395 uarg = skb_shinfo(skb)->destructor_arg;
1396 /* increase inflight counter to offset decrement in callback */
1397 atomic_inc(&queue->inflight_packets);
1398 uarg->callback(uarg, true);
1399 skb_shinfo(skb)->destructor_arg = NULL;
1401 xenvif_skb_zerocopy_prepare(queue, nskb);
1407 static int xenvif_tx_submit(struct xenvif_queue *queue)
1409 struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1410 struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1411 struct sk_buff *skb;
1414 while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1415 struct xen_netif_tx_request *txp;
1419 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1420 txp = &queue->pending_tx_info[pending_idx].req;
1422 /* Check the remap error code. */
1423 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1424 /* If there was an error, xenvif_tx_check_gop is
1425 * expected to release all the frags which were mapped,
1426 * so kfree_skb shouldn't do it again
1428 skb_shinfo(skb)->nr_frags = 0;
1429 if (skb_has_frag_list(skb)) {
1430 struct sk_buff *nskb =
1431 skb_shinfo(skb)->frag_list;
1432 skb_shinfo(nskb)->nr_frags = 0;
1438 data_len = skb->len;
1439 callback_param(queue, pending_idx).ctx = NULL;
1440 if (data_len < txp->size) {
1441 /* Append the packet payload as a fragment. */
1442 txp->offset += data_len;
1443 txp->size -= data_len;
1445 /* Schedule a response immediately. */
1446 xenvif_idx_release(queue, pending_idx,
1447 XEN_NETIF_RSP_OKAY);
1450 if (txp->flags & XEN_NETTXF_csum_blank)
1451 skb->ip_summed = CHECKSUM_PARTIAL;
1452 else if (txp->flags & XEN_NETTXF_data_validated)
1453 skb->ip_summed = CHECKSUM_UNNECESSARY;
1455 xenvif_fill_frags(queue, skb);
1457 if (unlikely(skb_has_frag_list(skb))) {
1458 if (xenvif_handle_frag_list(queue, skb)) {
1459 if (net_ratelimit())
1460 netdev_err(queue->vif->dev,
1461 "Not enough memory to consolidate frag_list!\n");
1462 xenvif_skb_zerocopy_prepare(queue, skb);
1468 skb->dev = queue->vif->dev;
1469 skb->protocol = eth_type_trans(skb, skb->dev);
1470 skb_reset_network_header(skb);
1472 if (checksum_setup(queue, skb)) {
1473 netdev_dbg(queue->vif->dev,
1474 "Can't setup checksum in net_tx_action\n");
1475 /* We have to set this flag to trigger the callback */
1476 if (skb_shinfo(skb)->destructor_arg)
1477 xenvif_skb_zerocopy_prepare(queue, skb);
1482 skb_probe_transport_header(skb, 0);
1484 /* If the packet is GSO then we will have just set up the
1485 * transport header offset in checksum_setup so it's now
1486 * straightforward to calculate gso_segs.
1488 if (skb_is_gso(skb)) {
1489 int mss = skb_shinfo(skb)->gso_size;
1490 int hdrlen = skb_transport_header(skb) -
1491 skb_mac_header(skb) +
1494 skb_shinfo(skb)->gso_segs =
1495 DIV_ROUND_UP(skb->len - hdrlen, mss);
1498 queue->stats.rx_bytes += skb->len;
1499 queue->stats.rx_packets++;
1503 /* Set this flag right before netif_receive_skb, otherwise
1504 * someone might think this packet already left netback, and
1505 * do a skb_copy_ubufs while we are still in control of the
1506 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1508 if (skb_shinfo(skb)->destructor_arg) {
1509 xenvif_skb_zerocopy_prepare(queue, skb);
1510 queue->stats.tx_zerocopy_sent++;
1513 netif_receive_skb(skb);
1519 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1521 unsigned long flags;
1522 pending_ring_idx_t index;
1523 struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1525 /* This is the only place where we grab this lock, to protect callbacks
1528 spin_lock_irqsave(&queue->callback_lock, flags);
1530 u16 pending_idx = ubuf->desc;
1531 ubuf = (struct ubuf_info *) ubuf->ctx;
1532 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1534 index = pending_index(queue->dealloc_prod);
1535 queue->dealloc_ring[index] = pending_idx;
1536 /* Sync with xenvif_tx_dealloc_action:
1537 * insert idx then incr producer.
1540 queue->dealloc_prod++;
1542 wake_up(&queue->dealloc_wq);
1543 spin_unlock_irqrestore(&queue->callback_lock, flags);
1545 if (likely(zerocopy_success))
1546 queue->stats.tx_zerocopy_success++;
1548 queue->stats.tx_zerocopy_fail++;
1549 xenvif_skb_zerocopy_complete(queue);
1552 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1554 struct gnttab_unmap_grant_ref *gop;
1555 pending_ring_idx_t dc, dp;
1556 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1559 dc = queue->dealloc_cons;
1560 gop = queue->tx_unmap_ops;
1562 /* Free up any grants we have finished using */
1564 dp = queue->dealloc_prod;
1566 /* Ensure we see all indices enqueued by all
1567 * xenvif_zerocopy_callback().
1572 BUG_ON(gop - queue->tx_unmap_ops > MAX_PENDING_REQS);
1574 queue->dealloc_ring[pending_index(dc++)];
1576 pending_idx_release[gop-queue->tx_unmap_ops] =
1578 queue->pages_to_unmap[gop-queue->tx_unmap_ops] =
1579 queue->mmap_pages[pending_idx];
1580 gnttab_set_unmap_op(gop,
1581 idx_to_kaddr(queue, pending_idx),
1583 queue->grant_tx_handle[pending_idx]);
1584 xenvif_grant_handle_reset(queue, pending_idx);
1588 } while (dp != queue->dealloc_prod);
1590 queue->dealloc_cons = dc;
1592 if (gop - queue->tx_unmap_ops > 0) {
1594 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1596 queue->pages_to_unmap,
1597 gop - queue->tx_unmap_ops);
1599 netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tx ret %d\n",
1600 gop - queue->tx_unmap_ops, ret);
1601 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1602 if (gop[i].status != GNTST_okay)
1603 netdev_err(queue->vif->dev,
1604 " host_addr: %llx handle: %x status: %d\n",
1613 for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1614 xenvif_idx_release(queue, pending_idx_release[i],
1615 XEN_NETIF_RSP_OKAY);
1619 /* Called after netfront has transmitted */
1620 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1622 unsigned nr_mops, nr_cops = 0;
1625 if (unlikely(!tx_work_todo(queue)))
1628 xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1633 gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1635 ret = gnttab_map_refs(queue->tx_map_ops,
1637 queue->pages_to_map,
1642 work_done = xenvif_tx_submit(queue);
1647 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1650 struct pending_tx_info *pending_tx_info;
1651 pending_ring_idx_t index;
1652 unsigned long flags;
1654 pending_tx_info = &queue->pending_tx_info[pending_idx];
1655 spin_lock_irqsave(&queue->response_lock, flags);
1656 make_tx_response(queue, &pending_tx_info->req, status);
1657 index = pending_index(queue->pending_prod);
1658 queue->pending_ring[index] = pending_idx;
1659 /* TX shouldn't use the index before we give it back here */
1661 queue->pending_prod++;
1662 spin_unlock_irqrestore(&queue->response_lock, flags);
1666 static void make_tx_response(struct xenvif_queue *queue,
1667 struct xen_netif_tx_request *txp,
1670 RING_IDX i = queue->tx.rsp_prod_pvt;
1671 struct xen_netif_tx_response *resp;
1674 resp = RING_GET_RESPONSE(&queue->tx, i);
1678 if (txp->flags & XEN_NETTXF_extra_info)
1679 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1681 queue->tx.rsp_prod_pvt = ++i;
1682 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1684 notify_remote_via_irq(queue->tx_irq);
1687 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1694 RING_IDX i = queue->rx.rsp_prod_pvt;
1695 struct xen_netif_rx_response *resp;
1697 resp = RING_GET_RESPONSE(&queue->rx, i);
1698 resp->offset = offset;
1699 resp->flags = flags;
1701 resp->status = (s16)size;
1703 resp->status = (s16)st;
1705 queue->rx.rsp_prod_pvt = ++i;
1710 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1713 struct gnttab_unmap_grant_ref tx_unmap_op;
1715 gnttab_set_unmap_op(&tx_unmap_op,
1716 idx_to_kaddr(queue, pending_idx),
1718 queue->grant_tx_handle[pending_idx]);
1719 xenvif_grant_handle_reset(queue, pending_idx);
1721 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1722 &queue->mmap_pages[pending_idx], 1);
1724 netdev_err(queue->vif->dev,
1725 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: %x status: %d\n",
1728 tx_unmap_op.host_addr,
1730 tx_unmap_op.status);
1735 static inline int tx_work_todo(struct xenvif_queue *queue)
1737 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1743 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1745 return queue->dealloc_cons != queue->dealloc_prod;
1748 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1750 if (queue->tx.sring)
1751 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1753 if (queue->rx.sring)
1754 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1758 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1759 grant_ref_t tx_ring_ref,
1760 grant_ref_t rx_ring_ref)
1763 struct xen_netif_tx_sring *txs;
1764 struct xen_netif_rx_sring *rxs;
1768 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1769 tx_ring_ref, &addr);
1773 txs = (struct xen_netif_tx_sring *)addr;
1774 BACK_RING_INIT(&queue->tx, txs, PAGE_SIZE);
1776 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1777 rx_ring_ref, &addr);
1781 rxs = (struct xen_netif_rx_sring *)addr;
1782 BACK_RING_INIT(&queue->rx, rxs, PAGE_SIZE);
1787 xenvif_unmap_frontend_rings(queue);
1791 static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
1793 struct xenvif *vif = queue->vif;
1795 queue->stalled = true;
1797 /* At least one queue has stalled? Disable the carrier. */
1798 spin_lock(&vif->lock);
1799 if (vif->stalled_queues++ == 0) {
1800 netdev_info(vif->dev, "Guest Rx stalled");
1801 netif_carrier_off(vif->dev);
1803 spin_unlock(&vif->lock);
1806 static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
1808 struct xenvif *vif = queue->vif;
1810 queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
1811 queue->stalled = false;
1813 /* All queues are ready? Enable the carrier. */
1814 spin_lock(&vif->lock);
1815 if (--vif->stalled_queues == 0) {
1816 netdev_info(vif->dev, "Guest Rx ready");
1817 netif_carrier_on(vif->dev);
1819 spin_unlock(&vif->lock);
1822 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
1824 RING_IDX prod, cons;
1826 prod = queue->rx.sring->req_prod;
1827 cons = queue->rx.req_cons;
1829 return !queue->stalled
1830 && prod - cons < XEN_NETBK_RX_SLOTS_MAX
1831 && time_after(jiffies,
1832 queue->last_rx_time + queue->vif->stall_timeout);
1835 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
1837 RING_IDX prod, cons;
1839 prod = queue->rx.sring->req_prod;
1840 cons = queue->rx.req_cons;
1842 return queue->stalled
1843 && prod - cons >= XEN_NETBK_RX_SLOTS_MAX;
1846 static bool xenvif_have_rx_work(struct xenvif_queue *queue)
1848 return (!skb_queue_empty(&queue->rx_queue)
1849 && xenvif_rx_ring_slots_available(queue, XEN_NETBK_RX_SLOTS_MAX))
1850 || (queue->vif->stall_timeout &&
1851 (xenvif_rx_queue_stalled(queue)
1852 || xenvif_rx_queue_ready(queue)))
1853 || kthread_should_stop()
1854 || queue->vif->disabled;
1857 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
1859 struct sk_buff *skb;
1862 skb = skb_peek(&queue->rx_queue);
1864 return MAX_SCHEDULE_TIMEOUT;
1866 timeout = XENVIF_RX_CB(skb)->expires - jiffies;
1867 return timeout < 0 ? 0 : timeout;
1870 /* Wait until the guest Rx thread has work.
1872 * The timeout needs to be adjusted based on the current head of the
1873 * queue (and not just the head at the beginning). In particular, if
1874 * the queue is initially empty an infinite timeout is used and this
1875 * needs to be reduced when a skb is queued.
1877 * This cannot be done with wait_event_timeout() because it only
1878 * calculates the timeout once.
1880 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
1884 if (xenvif_have_rx_work(queue))
1890 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
1891 if (xenvif_have_rx_work(queue))
1893 ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
1897 finish_wait(&queue->wq, &wait);
1900 int xenvif_kthread_guest_rx(void *data)
1902 struct xenvif_queue *queue = data;
1903 struct xenvif *vif = queue->vif;
1905 if (!vif->stall_timeout)
1906 xenvif_queue_carrier_on(queue);
1909 xenvif_wait_for_rx_work(queue);
1911 if (kthread_should_stop())
1914 /* This frontend is found to be rogue, disable it in
1915 * kthread context. Currently this is only set when
1916 * netback finds out frontend sends malformed packet,
1917 * but we cannot disable the interface in softirq
1918 * context so we defer it here, if this thread is
1919 * associated with queue 0.
1921 if (unlikely(vif->disabled && queue->id == 0)) {
1922 xenvif_carrier_off(vif);
1926 if (!skb_queue_empty(&queue->rx_queue))
1927 xenvif_rx_action(queue);
1929 /* If the guest hasn't provided any Rx slots for a
1930 * while it's probably not responsive, drop the
1931 * carrier so packets are dropped earlier.
1933 if (vif->stall_timeout) {
1934 if (xenvif_rx_queue_stalled(queue))
1935 xenvif_queue_carrier_off(queue);
1936 else if (xenvif_rx_queue_ready(queue))
1937 xenvif_queue_carrier_on(queue);
1940 /* Queued packets may have foreign pages from other
1941 * domains. These cannot be queued indefinitely as
1942 * this would starve guests of grant refs and transmit
1945 xenvif_rx_queue_drop_expired(queue);
1947 xenvif_rx_queue_maybe_wake(queue);
1952 /* Bin any remaining skbs */
1953 xenvif_rx_queue_purge(queue);
1958 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1960 /* Dealloc thread must remain running until all inflight
1963 return kthread_should_stop() &&
1964 !atomic_read(&queue->inflight_packets);
1967 int xenvif_dealloc_kthread(void *data)
1969 struct xenvif_queue *queue = data;
1972 wait_event_interruptible(queue->dealloc_wq,
1973 tx_dealloc_work_todo(queue) ||
1974 xenvif_dealloc_kthread_should_stop(queue));
1975 if (xenvif_dealloc_kthread_should_stop(queue))
1978 xenvif_tx_dealloc_action(queue);
1982 /* Unmap anything remaining*/
1983 if (tx_dealloc_work_todo(queue))
1984 xenvif_tx_dealloc_action(queue);
1989 static int __init netback_init(void)
1996 /* Allow as many queues as there are CPUs, by default */
1997 xenvif_max_queues = num_online_cpus();
1999 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2000 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2001 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2002 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2005 rc = xenvif_xenbus_init();
2009 #ifdef CONFIG_DEBUG_FS
2010 xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2011 if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2012 pr_warn("Init of debugfs returned %ld!\n",
2013 PTR_ERR(xen_netback_dbg_root));
2014 #endif /* CONFIG_DEBUG_FS */
2022 module_init(netback_init);
2024 static void __exit netback_fini(void)
2026 #ifdef CONFIG_DEBUG_FS
2027 if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2028 debugfs_remove_recursive(xen_netback_dbg_root);
2029 #endif /* CONFIG_DEBUG_FS */
2030 xenvif_xenbus_fini();
2032 module_exit(netback_fini);
2034 MODULE_LICENSE("Dual BSD/GPL");
2035 MODULE_ALIAS("xen-backend:vif");