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>
44 #include <xen/events.h>
45 #include <xen/interface/memory.h>
47 #include <asm/xen/hypercall.h>
48 #include <asm/xen/page.h>
50 /* Provide an option to disable split event channels at load time as
51 * event channels are limited resource. Split event channels are
54 bool separate_tx_rx_irq = 1;
55 module_param(separate_tx_rx_irq, bool, 0644);
58 * This is the maximum slots a skb can have. If a guest sends a skb
59 * which exceeds this limit it is considered malicious.
61 #define FATAL_SKB_SLOTS_DEFAULT 20
62 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
63 module_param(fatal_skb_slots, uint, 0444);
66 * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
67 * one or more merged tx requests, otherwise it is the continuation of
68 * previous tx request.
70 static inline int pending_tx_is_head(struct xenvif *vif, RING_IDX idx)
72 return vif->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
75 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
78 static void make_tx_response(struct xenvif *vif,
79 struct xen_netif_tx_request *txp,
82 static inline int tx_work_todo(struct xenvif *vif);
83 static inline int rx_work_todo(struct xenvif *vif);
85 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
92 static inline unsigned long idx_to_pfn(struct xenvif *vif,
95 return page_to_pfn(vif->mmap_pages[idx]);
98 static inline unsigned long idx_to_kaddr(struct xenvif *vif,
101 return (unsigned long)pfn_to_kaddr(idx_to_pfn(vif, idx));
104 static inline struct xenvif* ubuf_to_vif(struct ubuf_info *ubuf)
108 /* This is a miniumum size for the linear area to avoid lots of
109 * calls to __pskb_pull_tail() as we set up checksum offsets. The
110 * value 128 was chosen as it covers all IPv4 and most likely
113 #define PKT_PROT_LEN 128
115 static u16 frag_get_pending_idx(skb_frag_t *frag)
117 return (u16)frag->page_offset;
120 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
122 frag->page_offset = pending_idx;
125 static inline pending_ring_idx_t pending_index(unsigned i)
127 return i & (MAX_PENDING_REQS-1);
130 bool xenvif_rx_ring_slots_available(struct xenvif *vif, int needed)
135 prod = vif->rx.sring->req_prod;
136 cons = vif->rx.req_cons;
138 if (prod - cons >= needed)
141 vif->rx.sring->req_event = prod + 1;
143 /* Make sure event is visible before we check prod
147 } while (vif->rx.sring->req_prod != prod);
153 * Returns true if we should start a new receive buffer instead of
154 * adding 'size' bytes to a buffer which currently contains 'offset'
157 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
159 /* simple case: we have completely filled the current buffer. */
160 if (offset == MAX_BUFFER_OFFSET)
164 * complex case: start a fresh buffer if the current frag
165 * would overflow the current buffer but only if:
166 * (i) this frag would fit completely in the next buffer
167 * and (ii) there is already some data in the current buffer
168 * and (iii) this is not the head buffer.
171 * - (i) stops us splitting a frag into two copies
172 * unless the frag is too large for a single buffer.
173 * - (ii) stops us from leaving a buffer pointlessly empty.
174 * - (iii) stops us leaving the first buffer
175 * empty. Strictly speaking this is already covered
176 * by (ii) but is explicitly checked because
177 * netfront relies on the first buffer being
178 * non-empty and can crash otherwise.
180 * This means we will effectively linearise small
181 * frags but do not needlessly split large buffers
182 * into multiple copies tend to give large frags their
183 * own buffers as before.
185 if ((offset + size > MAX_BUFFER_OFFSET) &&
186 (size <= MAX_BUFFER_OFFSET) && offset && !head)
192 struct netrx_pending_operations {
193 unsigned copy_prod, copy_cons;
194 unsigned meta_prod, meta_cons;
195 struct gnttab_copy *copy;
196 struct xenvif_rx_meta *meta;
198 grant_ref_t copy_gref;
201 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
202 struct netrx_pending_operations *npo)
204 struct xenvif_rx_meta *meta;
205 struct xen_netif_rx_request *req;
207 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
209 meta = npo->meta + npo->meta_prod++;
210 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
216 npo->copy_gref = req->gref;
222 * Set up the grant operations for this fragment. If it's a flipping
223 * interface, we also set up the unmap request from here.
225 static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
226 struct netrx_pending_operations *npo,
227 struct page *page, unsigned long size,
228 unsigned long offset, int *head,
229 struct xenvif *foreign_vif,
230 grant_ref_t foreign_gref)
232 struct gnttab_copy *copy_gop;
233 struct xenvif_rx_meta *meta;
237 /* Data must not cross a page boundary. */
238 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
240 meta = npo->meta + npo->meta_prod - 1;
242 /* Skip unused frames from start of page */
243 page += offset >> PAGE_SHIFT;
244 offset &= ~PAGE_MASK;
247 BUG_ON(offset >= PAGE_SIZE);
248 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
250 bytes = PAGE_SIZE - offset;
255 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
257 * Netfront requires there to be some data in the head
262 meta = get_next_rx_buffer(vif, npo);
265 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
266 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
268 copy_gop = npo->copy + npo->copy_prod++;
269 copy_gop->flags = GNTCOPY_dest_gref;
270 copy_gop->len = bytes;
273 copy_gop->source.domid = foreign_vif->domid;
274 copy_gop->source.u.ref = foreign_gref;
275 copy_gop->flags |= GNTCOPY_source_gref;
277 copy_gop->source.domid = DOMID_SELF;
278 copy_gop->source.u.gmfn =
279 virt_to_mfn(page_address(page));
281 copy_gop->source.offset = offset;
283 copy_gop->dest.domid = vif->domid;
284 copy_gop->dest.offset = npo->copy_off;
285 copy_gop->dest.u.ref = npo->copy_gref;
287 npo->copy_off += bytes;
294 if (offset == PAGE_SIZE && size) {
295 BUG_ON(!PageCompound(page));
300 /* Leave a gap for the GSO descriptor. */
301 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
302 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
303 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
304 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
306 gso_type = XEN_NETIF_GSO_TYPE_NONE;
308 if (*head && ((1 << gso_type) & vif->gso_mask))
311 *head = 0; /* There must be something in this buffer now. */
317 * Prepare an SKB to be transmitted to the frontend.
319 * This function is responsible for allocating grant operations, meta
322 * It returns the number of meta structures consumed. The number of
323 * ring slots used is always equal to the number of meta slots used
324 * plus the number of GSO descriptors used. Currently, we use either
325 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
326 * frontend-side LRO).
328 static int xenvif_gop_skb(struct sk_buff *skb,
329 struct netrx_pending_operations *npo)
331 struct xenvif *vif = netdev_priv(skb->dev);
332 int nr_frags = skb_shinfo(skb)->nr_frags;
334 struct xen_netif_rx_request *req;
335 struct xenvif_rx_meta *meta;
341 struct ubuf_info *ubuf = skb_shinfo(skb)->destructor_arg;
342 grant_ref_t foreign_grefs[MAX_SKB_FRAGS];
343 struct xenvif *foreign_vif = NULL;
345 old_meta_prod = npo->meta_prod;
347 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
348 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
349 gso_size = skb_shinfo(skb)->gso_size;
350 } else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
351 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
352 gso_size = skb_shinfo(skb)->gso_size;
354 gso_type = XEN_NETIF_GSO_TYPE_NONE;
358 /* Set up a GSO prefix descriptor, if necessary */
359 if ((1 << gso_type) & vif->gso_prefix_mask) {
360 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
361 meta = npo->meta + npo->meta_prod++;
362 meta->gso_type = gso_type;
363 meta->gso_size = gso_size;
368 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
369 meta = npo->meta + npo->meta_prod++;
371 if ((1 << gso_type) & vif->gso_mask) {
372 meta->gso_type = gso_type;
373 meta->gso_size = gso_size;
375 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
382 npo->copy_gref = req->gref;
384 if ((skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) &&
385 (ubuf->callback == &xenvif_zerocopy_callback)) {
387 foreign_vif = ubuf_to_vif(ubuf);
390 u16 pending_idx = ubuf->desc;
392 foreign_vif->pending_tx_info[pending_idx].req.gref;
393 ubuf = (struct ubuf_info *) ubuf->ctx;
398 while (data < skb_tail_pointer(skb)) {
399 unsigned int offset = offset_in_page(data);
400 unsigned int len = PAGE_SIZE - offset;
402 if (data + len > skb_tail_pointer(skb))
403 len = skb_tail_pointer(skb) - data;
405 xenvif_gop_frag_copy(vif, skb, npo,
406 virt_to_page(data), len, offset, &head,
412 for (i = 0; i < nr_frags; i++) {
413 xenvif_gop_frag_copy(vif, skb, npo,
414 skb_frag_page(&skb_shinfo(skb)->frags[i]),
415 skb_frag_size(&skb_shinfo(skb)->frags[i]),
416 skb_shinfo(skb)->frags[i].page_offset,
422 return npo->meta_prod - old_meta_prod;
426 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
427 * used to set up the operations on the top of
428 * netrx_pending_operations, which have since been done. Check that
429 * they didn't give any errors and advance over them.
431 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
432 struct netrx_pending_operations *npo)
434 struct gnttab_copy *copy_op;
435 int status = XEN_NETIF_RSP_OKAY;
438 for (i = 0; i < nr_meta_slots; i++) {
439 copy_op = npo->copy + npo->copy_cons++;
440 if (copy_op->status != GNTST_okay) {
442 "Bad status %d from copy to DOM%d.\n",
443 copy_op->status, vif->domid);
444 status = XEN_NETIF_RSP_ERROR;
451 static void xenvif_add_frag_responses(struct xenvif *vif, int status,
452 struct xenvif_rx_meta *meta,
456 unsigned long offset;
458 /* No fragments used */
459 if (nr_meta_slots <= 1)
464 for (i = 0; i < nr_meta_slots; i++) {
466 if (i == nr_meta_slots - 1)
469 flags = XEN_NETRXF_more_data;
472 make_rx_response(vif, meta[i].id, status, offset,
473 meta[i].size, flags);
477 struct xenvif_rx_cb {
481 #define XENVIF_RX_CB(skb) ((struct xenvif_rx_cb *)(skb)->cb)
483 void xenvif_kick_thread(struct xenvif *vif)
488 static void xenvif_rx_action(struct xenvif *vif)
492 struct xen_netif_rx_response *resp;
493 struct sk_buff_head rxq;
497 unsigned long offset;
498 bool need_to_notify = false;
500 struct netrx_pending_operations npo = {
501 .copy = vif->grant_copy_op,
505 skb_queue_head_init(&rxq);
507 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
508 RING_IDX max_slots_needed;
511 /* We need a cheap worse case estimate for the number of
515 max_slots_needed = DIV_ROUND_UP(offset_in_page(skb->data) +
518 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
520 size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
521 max_slots_needed += DIV_ROUND_UP(size, PAGE_SIZE);
523 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4 ||
524 skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
527 /* If the skb may not fit then bail out now */
528 if (!xenvif_rx_ring_slots_available(vif, max_slots_needed)) {
529 skb_queue_head(&vif->rx_queue, skb);
530 need_to_notify = true;
531 vif->rx_last_skb_slots = max_slots_needed;
534 vif->rx_last_skb_slots = 0;
536 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo);
537 BUG_ON(XENVIF_RX_CB(skb)->meta_slots_used > max_slots_needed);
539 __skb_queue_tail(&rxq, skb);
542 BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
547 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
548 gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
550 while ((skb = __skb_dequeue(&rxq)) != NULL) {
552 if ((1 << vif->meta[npo.meta_cons].gso_type) &
553 vif->gso_prefix_mask) {
554 resp = RING_GET_RESPONSE(&vif->rx,
555 vif->rx.rsp_prod_pvt++);
557 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
559 resp->offset = vif->meta[npo.meta_cons].gso_size;
560 resp->id = vif->meta[npo.meta_cons].id;
561 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
564 XENVIF_RX_CB(skb)->meta_slots_used--;
568 vif->dev->stats.tx_bytes += skb->len;
569 vif->dev->stats.tx_packets++;
571 status = xenvif_check_gop(vif,
572 XENVIF_RX_CB(skb)->meta_slots_used,
575 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
578 flags = XEN_NETRXF_more_data;
580 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
581 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
582 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
583 /* remote but checksummed. */
584 flags |= XEN_NETRXF_data_validated;
587 resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
589 vif->meta[npo.meta_cons].size,
592 if ((1 << vif->meta[npo.meta_cons].gso_type) &
594 struct xen_netif_extra_info *gso =
595 (struct xen_netif_extra_info *)
596 RING_GET_RESPONSE(&vif->rx,
597 vif->rx.rsp_prod_pvt++);
599 resp->flags |= XEN_NETRXF_extra_info;
601 gso->u.gso.type = vif->meta[npo.meta_cons].gso_type;
602 gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
604 gso->u.gso.features = 0;
606 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
610 xenvif_add_frag_responses(vif, status,
611 vif->meta + npo.meta_cons + 1,
612 XENVIF_RX_CB(skb)->meta_slots_used);
614 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
616 need_to_notify |= !!ret;
618 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
624 notify_remote_via_irq(vif->rx_irq);
627 void xenvif_check_rx_xenvif(struct xenvif *vif)
631 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
634 napi_schedule(&vif->napi);
637 static void tx_add_credit(struct xenvif *vif)
639 unsigned long max_burst, max_credit;
642 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
643 * Otherwise the interface can seize up due to insufficient credit.
645 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
646 max_burst = min(max_burst, 131072UL);
647 max_burst = max(max_burst, vif->credit_bytes);
649 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
650 max_credit = vif->remaining_credit + vif->credit_bytes;
651 if (max_credit < vif->remaining_credit)
652 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
654 vif->remaining_credit = min(max_credit, max_burst);
657 static void tx_credit_callback(unsigned long data)
659 struct xenvif *vif = (struct xenvif *)data;
661 xenvif_check_rx_xenvif(vif);
664 static void xenvif_tx_err(struct xenvif *vif,
665 struct xen_netif_tx_request *txp, RING_IDX end)
667 RING_IDX cons = vif->tx.req_cons;
670 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
673 txp = RING_GET_REQUEST(&vif->tx, cons++);
675 vif->tx.req_cons = cons;
678 static void xenvif_fatal_tx_err(struct xenvif *vif)
680 netdev_err(vif->dev, "fatal error; disabling device\n");
681 xenvif_carrier_off(vif);
684 static int xenvif_count_requests(struct xenvif *vif,
685 struct xen_netif_tx_request *first,
686 struct xen_netif_tx_request *txp,
689 RING_IDX cons = vif->tx.req_cons;
694 if (!(first->flags & XEN_NETTXF_more_data))
698 struct xen_netif_tx_request dropped_tx = { 0 };
700 if (slots >= work_to_do) {
702 "Asked for %d slots but exceeds this limit\n",
704 xenvif_fatal_tx_err(vif);
708 /* This guest is really using too many slots and
709 * considered malicious.
711 if (unlikely(slots >= fatal_skb_slots)) {
713 "Malicious frontend using %d slots, threshold %u\n",
714 slots, fatal_skb_slots);
715 xenvif_fatal_tx_err(vif);
719 /* Xen network protocol had implicit dependency on
720 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
721 * the historical MAX_SKB_FRAGS value 18 to honor the
722 * same behavior as before. Any packet using more than
723 * 18 slots but less than fatal_skb_slots slots is
726 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
729 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
730 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
737 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
740 /* If the guest submitted a frame >= 64 KiB then
741 * first->size overflowed and following slots will
742 * appear to be larger than the frame.
744 * This cannot be fatal error as there are buggy
745 * frontends that do this.
747 * Consume all slots and drop the packet.
749 if (!drop_err && txp->size > first->size) {
752 "Invalid tx request, slot size %u > remaining size %u\n",
753 txp->size, first->size);
757 first->size -= txp->size;
760 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
761 netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
762 txp->offset, txp->size);
763 xenvif_fatal_tx_err(vif);
767 more_data = txp->flags & XEN_NETTXF_more_data;
775 xenvif_tx_err(vif, first, cons + slots);
782 static struct page *xenvif_alloc_page(struct xenvif *vif,
787 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
790 vif->mmap_pages[pending_idx] = page;
796 struct xenvif_tx_cb {
800 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
802 static struct gnttab_copy *xenvif_get_requests(struct xenvif *vif,
804 struct xen_netif_tx_request *txp,
805 struct gnttab_copy *gop)
807 struct skb_shared_info *shinfo = skb_shinfo(skb);
808 skb_frag_t *frags = shinfo->frags;
809 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
813 pending_ring_idx_t index, start_idx = 0;
815 unsigned int nr_slots;
816 struct pending_tx_info *first = NULL;
818 /* At this point shinfo->nr_frags is in fact the number of
819 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
821 nr_slots = shinfo->nr_frags;
823 /* Skip first skb fragment if it is on same page as header fragment. */
824 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
826 /* Coalesce tx requests, at this point the packet passed in
827 * should be <= 64K. Any packets larger than 64K have been
828 * handled in xenvif_count_requests().
830 for (shinfo->nr_frags = slot = start; slot < nr_slots;
831 shinfo->nr_frags++) {
832 struct pending_tx_info *pending_tx_info =
833 vif->pending_tx_info;
835 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
841 while (dst_offset < PAGE_SIZE && slot < nr_slots) {
842 gop->flags = GNTCOPY_source_gref;
844 gop->source.u.ref = txp->gref;
845 gop->source.domid = vif->domid;
846 gop->source.offset = txp->offset;
848 gop->dest.domid = DOMID_SELF;
850 gop->dest.offset = dst_offset;
851 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
853 if (dst_offset + txp->size > PAGE_SIZE) {
854 /* This page can only merge a portion
855 * of tx request. Do not increment any
856 * pointer / counter here. The txp
857 * will be dealt with in future
858 * rounds, eventually hitting the
861 gop->len = PAGE_SIZE - dst_offset;
862 txp->offset += gop->len;
863 txp->size -= gop->len;
864 dst_offset += gop->len; /* quit loop */
866 /* This tx request can be merged in the page */
867 gop->len = txp->size;
868 dst_offset += gop->len;
870 index = pending_index(vif->pending_cons++);
872 pending_idx = vif->pending_ring[index];
874 memcpy(&pending_tx_info[pending_idx].req, txp,
877 /* Poison these fields, corresponding
878 * fields for head tx req will be set
879 * to correct values after the loop.
881 vif->mmap_pages[pending_idx] = (void *)(~0UL);
882 pending_tx_info[pending_idx].head =
883 INVALID_PENDING_RING_IDX;
886 first = &pending_tx_info[pending_idx];
888 head_idx = pending_idx;
898 first->req.offset = 0;
899 first->req.size = dst_offset;
900 first->head = start_idx;
901 vif->mmap_pages[head_idx] = page;
902 frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
905 BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
909 /* Unwind, freeing all pages and sending error responses. */
910 while (shinfo->nr_frags-- > start) {
911 xenvif_idx_release(vif,
912 frag_get_pending_idx(&frags[shinfo->nr_frags]),
913 XEN_NETIF_RSP_ERROR);
915 /* The head too, if necessary. */
917 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
922 static int xenvif_tx_check_gop(struct xenvif *vif,
924 struct gnttab_copy **gopp)
926 struct gnttab_copy *gop = *gopp;
927 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
928 struct skb_shared_info *shinfo = skb_shinfo(skb);
929 struct pending_tx_info *tx_info;
930 int nr_frags = shinfo->nr_frags;
932 u16 peek; /* peek into next tx request */
934 /* Check status of header. */
937 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
939 /* Skip first skb fragment if it is on same page as header fragment. */
940 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
942 for (i = start; i < nr_frags; i++) {
944 pending_ring_idx_t head;
946 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
947 tx_info = &vif->pending_tx_info[pending_idx];
948 head = tx_info->head;
950 /* Check error status: if okay then remember grant handle. */
952 newerr = (++gop)->status;
955 peek = vif->pending_ring[pending_index(++head)];
956 } while (!pending_tx_is_head(vif, peek));
958 if (likely(!newerr)) {
959 /* Had a previous error? Invalidate this fragment. */
961 xenvif_idx_release(vif, pending_idx,
966 /* Error on this fragment: respond to client with an error. */
967 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
969 /* Not the first error? Preceding frags already invalidated. */
973 /* First error: invalidate header and preceding fragments. */
974 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
975 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
976 for (j = start; j < i; j++) {
977 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
978 xenvif_idx_release(vif, pending_idx,
982 /* Remember the error: invalidate all subsequent fragments. */
990 static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
992 struct skb_shared_info *shinfo = skb_shinfo(skb);
993 int nr_frags = shinfo->nr_frags;
996 for (i = 0; i < nr_frags; i++) {
997 skb_frag_t *frag = shinfo->frags + i;
998 struct xen_netif_tx_request *txp;
1002 pending_idx = frag_get_pending_idx(frag);
1004 txp = &vif->pending_tx_info[pending_idx].req;
1005 page = virt_to_page(idx_to_kaddr(vif, pending_idx));
1006 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1007 skb->len += txp->size;
1008 skb->data_len += txp->size;
1009 skb->truesize += txp->size;
1011 /* Take an extra reference to offset xenvif_idx_release */
1012 get_page(vif->mmap_pages[pending_idx]);
1013 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
1017 static int xenvif_get_extras(struct xenvif *vif,
1018 struct xen_netif_extra_info *extras,
1021 struct xen_netif_extra_info extra;
1022 RING_IDX cons = vif->tx.req_cons;
1025 if (unlikely(work_to_do-- <= 0)) {
1026 netdev_err(vif->dev, "Missing extra info\n");
1027 xenvif_fatal_tx_err(vif);
1031 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1033 if (unlikely(!extra.type ||
1034 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1035 vif->tx.req_cons = ++cons;
1036 netdev_err(vif->dev,
1037 "Invalid extra type: %d\n", extra.type);
1038 xenvif_fatal_tx_err(vif);
1042 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1043 vif->tx.req_cons = ++cons;
1044 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1049 static int xenvif_set_skb_gso(struct xenvif *vif,
1050 struct sk_buff *skb,
1051 struct xen_netif_extra_info *gso)
1053 if (!gso->u.gso.size) {
1054 netdev_err(vif->dev, "GSO size must not be zero.\n");
1055 xenvif_fatal_tx_err(vif);
1059 switch (gso->u.gso.type) {
1060 case XEN_NETIF_GSO_TYPE_TCPV4:
1061 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1063 case XEN_NETIF_GSO_TYPE_TCPV6:
1064 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1067 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1068 xenvif_fatal_tx_err(vif);
1072 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1073 /* gso_segs will be calculated later */
1078 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1080 bool recalculate_partial_csum = false;
1082 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1083 * peers can fail to set NETRXF_csum_blank when sending a GSO
1084 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1085 * recalculate the partial checksum.
1087 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1088 vif->rx_gso_checksum_fixup++;
1089 skb->ip_summed = CHECKSUM_PARTIAL;
1090 recalculate_partial_csum = true;
1093 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1094 if (skb->ip_summed != CHECKSUM_PARTIAL)
1097 return skb_checksum_setup(skb, recalculate_partial_csum);
1100 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1102 u64 now = get_jiffies_64();
1103 u64 next_credit = vif->credit_window_start +
1104 msecs_to_jiffies(vif->credit_usec / 1000);
1106 /* Timer could already be pending in rare cases. */
1107 if (timer_pending(&vif->credit_timeout))
1110 /* Passed the point where we can replenish credit? */
1111 if (time_after_eq64(now, next_credit)) {
1112 vif->credit_window_start = now;
1116 /* Still too big to send right now? Set a callback. */
1117 if (size > vif->remaining_credit) {
1118 vif->credit_timeout.data =
1120 vif->credit_timeout.function =
1122 mod_timer(&vif->credit_timeout,
1124 vif->credit_window_start = next_credit;
1132 static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget)
1134 struct gnttab_copy *gop = vif->tx_copy_ops, *request_gop;
1135 struct sk_buff *skb;
1138 while (xenvif_tx_pending_slots_available(vif) &&
1139 (skb_queue_len(&vif->tx_queue) < budget)) {
1140 struct xen_netif_tx_request txreq;
1141 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1143 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1147 unsigned int data_len;
1148 pending_ring_idx_t index;
1150 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1151 XEN_NETIF_TX_RING_SIZE) {
1152 netdev_err(vif->dev,
1153 "Impossible number of requests. "
1154 "req_prod %d, req_cons %d, size %ld\n",
1155 vif->tx.sring->req_prod, vif->tx.req_cons,
1156 XEN_NETIF_TX_RING_SIZE);
1157 xenvif_fatal_tx_err(vif);
1161 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);
1165 idx = vif->tx.req_cons;
1166 rmb(); /* Ensure that we see the request before we copy it. */
1167 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1169 /* Credit-based scheduling. */
1170 if (txreq.size > vif->remaining_credit &&
1171 tx_credit_exceeded(vif, txreq.size))
1174 vif->remaining_credit -= txreq.size;
1177 vif->tx.req_cons = ++idx;
1179 memset(extras, 0, sizeof(extras));
1180 if (txreq.flags & XEN_NETTXF_extra_info) {
1181 work_to_do = xenvif_get_extras(vif, extras,
1183 idx = vif->tx.req_cons;
1184 if (unlikely(work_to_do < 0))
1188 ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
1189 if (unlikely(ret < 0))
1194 if (unlikely(txreq.size < ETH_HLEN)) {
1195 netdev_dbg(vif->dev,
1196 "Bad packet size: %d\n", txreq.size);
1197 xenvif_tx_err(vif, &txreq, idx);
1201 /* No crossing a page as the payload mustn't fragment. */
1202 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1203 netdev_err(vif->dev,
1204 "txreq.offset: %x, size: %u, end: %lu\n",
1205 txreq.offset, txreq.size,
1206 (txreq.offset&~PAGE_MASK) + txreq.size);
1207 xenvif_fatal_tx_err(vif);
1211 index = pending_index(vif->pending_cons);
1212 pending_idx = vif->pending_ring[index];
1214 data_len = (txreq.size > PKT_PROT_LEN &&
1215 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1216 PKT_PROT_LEN : txreq.size;
1218 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1219 GFP_ATOMIC | __GFP_NOWARN);
1220 if (unlikely(skb == NULL)) {
1221 netdev_dbg(vif->dev,
1222 "Can't allocate a skb in start_xmit.\n");
1223 xenvif_tx_err(vif, &txreq, idx);
1227 /* Packets passed to netif_rx() must have some headroom. */
1228 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1230 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1231 struct xen_netif_extra_info *gso;
1232 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1234 if (xenvif_set_skb_gso(vif, skb, gso)) {
1235 /* Failure in xenvif_set_skb_gso is fatal. */
1241 /* XXX could copy straight to head */
1242 page = xenvif_alloc_page(vif, pending_idx);
1245 xenvif_tx_err(vif, &txreq, idx);
1249 gop->source.u.ref = txreq.gref;
1250 gop->source.domid = vif->domid;
1251 gop->source.offset = txreq.offset;
1253 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1254 gop->dest.domid = DOMID_SELF;
1255 gop->dest.offset = txreq.offset;
1257 gop->len = txreq.size;
1258 gop->flags = GNTCOPY_source_gref;
1262 memcpy(&vif->pending_tx_info[pending_idx].req,
1263 &txreq, sizeof(txreq));
1264 vif->pending_tx_info[pending_idx].head = index;
1265 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1267 __skb_put(skb, data_len);
1269 skb_shinfo(skb)->nr_frags = ret;
1270 if (data_len < txreq.size) {
1271 skb_shinfo(skb)->nr_frags++;
1272 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1275 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1276 INVALID_PENDING_IDX);
1279 vif->pending_cons++;
1281 request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
1282 if (request_gop == NULL) {
1284 xenvif_tx_err(vif, &txreq, idx);
1289 __skb_queue_tail(&vif->tx_queue, skb);
1291 vif->tx.req_cons = idx;
1293 if ((gop-vif->tx_copy_ops) >= ARRAY_SIZE(vif->tx_copy_ops))
1297 return gop - vif->tx_copy_ops;
1301 static int xenvif_tx_submit(struct xenvif *vif)
1303 struct gnttab_copy *gop = vif->tx_copy_ops;
1304 struct sk_buff *skb;
1307 while ((skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
1308 struct xen_netif_tx_request *txp;
1312 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1313 txp = &vif->pending_tx_info[pending_idx].req;
1315 /* Check the remap error code. */
1316 if (unlikely(xenvif_tx_check_gop(vif, skb, &gop))) {
1317 netdev_dbg(vif->dev, "netback grant failed.\n");
1318 skb_shinfo(skb)->nr_frags = 0;
1323 data_len = skb->len;
1325 (void *)(idx_to_kaddr(vif, pending_idx)|txp->offset),
1327 if (data_len < txp->size) {
1328 /* Append the packet payload as a fragment. */
1329 txp->offset += data_len;
1330 txp->size -= data_len;
1332 /* Schedule a response immediately. */
1333 xenvif_idx_release(vif, pending_idx,
1334 XEN_NETIF_RSP_OKAY);
1337 if (txp->flags & XEN_NETTXF_csum_blank)
1338 skb->ip_summed = CHECKSUM_PARTIAL;
1339 else if (txp->flags & XEN_NETTXF_data_validated)
1340 skb->ip_summed = CHECKSUM_UNNECESSARY;
1342 xenvif_fill_frags(vif, skb);
1344 if (skb_is_nonlinear(skb) && skb_headlen(skb) < PKT_PROT_LEN) {
1345 int target = min_t(int, skb->len, PKT_PROT_LEN);
1346 __pskb_pull_tail(skb, target - skb_headlen(skb));
1349 skb->dev = vif->dev;
1350 skb->protocol = eth_type_trans(skb, skb->dev);
1351 skb_reset_network_header(skb);
1353 if (checksum_setup(vif, skb)) {
1354 netdev_dbg(vif->dev,
1355 "Can't setup checksum in net_tx_action\n");
1360 skb_probe_transport_header(skb, 0);
1362 /* If the packet is GSO then we will have just set up the
1363 * transport header offset in checksum_setup so it's now
1364 * straightforward to calculate gso_segs.
1366 if (skb_is_gso(skb)) {
1367 int mss = skb_shinfo(skb)->gso_size;
1368 int hdrlen = skb_transport_header(skb) -
1369 skb_mac_header(skb) +
1372 skb_shinfo(skb)->gso_segs =
1373 DIV_ROUND_UP(skb->len - hdrlen, mss);
1376 vif->dev->stats.rx_bytes += skb->len;
1377 vif->dev->stats.rx_packets++;
1381 netif_receive_skb(skb);
1387 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1392 /* Called after netfront has transmitted */
1393 int xenvif_tx_action(struct xenvif *vif, int budget)
1398 if (unlikely(!tx_work_todo(vif)))
1401 nr_gops = xenvif_tx_build_gops(vif, budget);
1406 gnttab_batch_copy(vif->tx_copy_ops, nr_gops);
1408 work_done = xenvif_tx_submit(vif);
1413 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
1416 struct pending_tx_info *pending_tx_info;
1417 pending_ring_idx_t head;
1418 u16 peek; /* peek into next tx request */
1420 BUG_ON(vif->mmap_pages[pending_idx] == (void *)(~0UL));
1422 /* Already complete? */
1423 if (vif->mmap_pages[pending_idx] == NULL)
1426 pending_tx_info = &vif->pending_tx_info[pending_idx];
1428 head = pending_tx_info->head;
1430 BUG_ON(!pending_tx_is_head(vif, head));
1431 BUG_ON(vif->pending_ring[pending_index(head)] != pending_idx);
1434 pending_ring_idx_t index;
1435 pending_ring_idx_t idx = pending_index(head);
1436 u16 info_idx = vif->pending_ring[idx];
1438 pending_tx_info = &vif->pending_tx_info[info_idx];
1439 make_tx_response(vif, &pending_tx_info->req, status);
1441 /* Setting any number other than
1442 * INVALID_PENDING_RING_IDX indicates this slot is
1443 * starting a new packet / ending a previous packet.
1445 pending_tx_info->head = 0;
1447 index = pending_index(vif->pending_prod++);
1448 vif->pending_ring[index] = vif->pending_ring[info_idx];
1450 peek = vif->pending_ring[pending_index(++head)];
1452 } while (!pending_tx_is_head(vif, peek));
1454 put_page(vif->mmap_pages[pending_idx]);
1455 vif->mmap_pages[pending_idx] = NULL;
1459 static void make_tx_response(struct xenvif *vif,
1460 struct xen_netif_tx_request *txp,
1463 RING_IDX i = vif->tx.rsp_prod_pvt;
1464 struct xen_netif_tx_response *resp;
1467 resp = RING_GET_RESPONSE(&vif->tx, i);
1471 if (txp->flags & XEN_NETTXF_extra_info)
1472 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1474 vif->tx.rsp_prod_pvt = ++i;
1475 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1477 notify_remote_via_irq(vif->tx_irq);
1480 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1487 RING_IDX i = vif->rx.rsp_prod_pvt;
1488 struct xen_netif_rx_response *resp;
1490 resp = RING_GET_RESPONSE(&vif->rx, i);
1491 resp->offset = offset;
1492 resp->flags = flags;
1494 resp->status = (s16)size;
1496 resp->status = (s16)st;
1498 vif->rx.rsp_prod_pvt = ++i;
1503 static inline int rx_work_todo(struct xenvif *vif)
1505 return !skb_queue_empty(&vif->rx_queue) &&
1506 xenvif_rx_ring_slots_available(vif, vif->rx_last_skb_slots);
1509 static inline int tx_work_todo(struct xenvif *vif)
1512 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)) &&
1513 xenvif_tx_pending_slots_available(vif))
1519 void xenvif_unmap_frontend_rings(struct xenvif *vif)
1522 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1525 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1529 int xenvif_map_frontend_rings(struct xenvif *vif,
1530 grant_ref_t tx_ring_ref,
1531 grant_ref_t rx_ring_ref)
1534 struct xen_netif_tx_sring *txs;
1535 struct xen_netif_rx_sring *rxs;
1539 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1540 tx_ring_ref, &addr);
1544 txs = (struct xen_netif_tx_sring *)addr;
1545 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1547 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1548 rx_ring_ref, &addr);
1552 rxs = (struct xen_netif_rx_sring *)addr;
1553 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1558 xenvif_unmap_frontend_rings(vif);
1562 void xenvif_stop_queue(struct xenvif *vif)
1564 if (!vif->can_queue)
1567 netif_stop_queue(vif->dev);
1570 static void xenvif_start_queue(struct xenvif *vif)
1572 if (xenvif_schedulable(vif))
1573 netif_wake_queue(vif->dev);
1576 int xenvif_kthread_guest_rx(void *data)
1578 struct xenvif *vif = data;
1579 struct sk_buff *skb;
1581 while (!kthread_should_stop()) {
1582 wait_event_interruptible(vif->wq,
1583 rx_work_todo(vif) ||
1584 kthread_should_stop());
1585 if (kthread_should_stop())
1588 if (!skb_queue_empty(&vif->rx_queue))
1589 xenvif_rx_action(vif);
1591 if (skb_queue_empty(&vif->rx_queue) &&
1592 netif_queue_stopped(vif->dev))
1593 xenvif_start_queue(vif);
1598 /* Bin any remaining skbs */
1599 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL)
1605 static int __init netback_init(void)
1612 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1613 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1614 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1615 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1618 rc = xenvif_xenbus_init();
1628 module_init(netback_init);
1630 static void __exit netback_fini(void)
1632 xenvif_xenbus_fini();
1634 module_exit(netback_fini);
1636 MODULE_LICENSE("Dual BSD/GPL");
1637 MODULE_ALIAS("xen-backend:vif");