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 * To avoid confusion, we define XEN_NETBK_LEGACY_SLOTS_MAX indicating
67 * the maximum slots a valid packet can use. Now this value is defined
68 * to be XEN_NETIF_NR_SLOTS_MIN, which is supposed to be supported by
71 #define XEN_NETBK_LEGACY_SLOTS_MAX XEN_NETIF_NR_SLOTS_MIN
74 * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
75 * one or more merged tx requests, otherwise it is the continuation of
76 * previous tx request.
78 static inline int pending_tx_is_head(struct xenvif *vif, RING_IDX idx)
80 return vif->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
83 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
86 static void make_tx_response(struct xenvif *vif,
87 struct xen_netif_tx_request *txp,
90 static inline int tx_work_todo(struct xenvif *vif);
91 static inline int rx_work_todo(struct xenvif *vif);
93 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
100 static inline unsigned long idx_to_pfn(struct xenvif *vif,
103 return page_to_pfn(vif->mmap_pages[idx]);
106 static inline unsigned long idx_to_kaddr(struct xenvif *vif,
109 return (unsigned long)pfn_to_kaddr(idx_to_pfn(vif, idx));
112 /* This is a miniumum size for the linear area to avoid lots of
113 * calls to __pskb_pull_tail() as we set up checksum offsets. The
114 * value 128 was chosen as it covers all IPv4 and most likely
117 #define PKT_PROT_LEN 128
119 static u16 frag_get_pending_idx(skb_frag_t *frag)
121 return (u16)frag->page_offset;
124 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
126 frag->page_offset = pending_idx;
129 static inline pending_ring_idx_t pending_index(unsigned i)
131 return i & (MAX_PENDING_REQS-1);
134 static inline pending_ring_idx_t nr_pending_reqs(struct xenvif *vif)
136 return MAX_PENDING_REQS -
137 vif->pending_prod + vif->pending_cons;
140 bool xenvif_rx_ring_slots_available(struct xenvif *vif, int needed)
145 prod = vif->rx.sring->req_prod;
146 cons = vif->rx.req_cons;
148 if (prod - cons >= needed)
151 vif->rx.sring->req_event = prod + 1;
153 /* Make sure event is visible before we check prod
157 } while (vif->rx.sring->req_prod != prod);
163 * Returns true if we should start a new receive buffer instead of
164 * adding 'size' bytes to a buffer which currently contains 'offset'
167 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
169 /* simple case: we have completely filled the current buffer. */
170 if (offset == MAX_BUFFER_OFFSET)
174 * complex case: start a fresh buffer if the current frag
175 * would overflow the current buffer but only if:
176 * (i) this frag would fit completely in the next buffer
177 * and (ii) there is already some data in the current buffer
178 * and (iii) this is not the head buffer.
181 * - (i) stops us splitting a frag into two copies
182 * unless the frag is too large for a single buffer.
183 * - (ii) stops us from leaving a buffer pointlessly empty.
184 * - (iii) stops us leaving the first buffer
185 * empty. Strictly speaking this is already covered
186 * by (ii) but is explicitly checked because
187 * netfront relies on the first buffer being
188 * non-empty and can crash otherwise.
190 * This means we will effectively linearise small
191 * frags but do not needlessly split large buffers
192 * into multiple copies tend to give large frags their
193 * own buffers as before.
195 if ((offset + size > MAX_BUFFER_OFFSET) &&
196 (size <= MAX_BUFFER_OFFSET) && offset && !head)
202 struct netrx_pending_operations {
203 unsigned copy_prod, copy_cons;
204 unsigned meta_prod, meta_cons;
205 struct gnttab_copy *copy;
206 struct xenvif_rx_meta *meta;
208 grant_ref_t copy_gref;
211 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
212 struct netrx_pending_operations *npo)
214 struct xenvif_rx_meta *meta;
215 struct xen_netif_rx_request *req;
217 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
219 meta = npo->meta + npo->meta_prod++;
220 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
226 npo->copy_gref = req->gref;
232 * Set up the grant operations for this fragment. If it's a flipping
233 * interface, we also set up the unmap request from here.
235 static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
236 struct netrx_pending_operations *npo,
237 struct page *page, unsigned long size,
238 unsigned long offset, int *head)
240 struct gnttab_copy *copy_gop;
241 struct xenvif_rx_meta *meta;
245 /* Data must not cross a page boundary. */
246 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
248 meta = npo->meta + npo->meta_prod - 1;
250 /* Skip unused frames from start of page */
251 page += offset >> PAGE_SHIFT;
252 offset &= ~PAGE_MASK;
255 BUG_ON(offset >= PAGE_SIZE);
256 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
258 bytes = PAGE_SIZE - offset;
263 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
265 * Netfront requires there to be some data in the head
270 meta = get_next_rx_buffer(vif, npo);
273 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
274 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
276 copy_gop = npo->copy + npo->copy_prod++;
277 copy_gop->flags = GNTCOPY_dest_gref;
278 copy_gop->len = bytes;
280 copy_gop->source.domid = DOMID_SELF;
281 copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
282 copy_gop->source.offset = offset;
284 copy_gop->dest.domid = vif->domid;
285 copy_gop->dest.offset = npo->copy_off;
286 copy_gop->dest.u.ref = npo->copy_gref;
288 npo->copy_off += bytes;
295 if (offset == PAGE_SIZE && size) {
296 BUG_ON(!PageCompound(page));
301 /* Leave a gap for the GSO descriptor. */
302 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
303 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
304 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
305 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
307 gso_type = XEN_NETIF_GSO_TYPE_NONE;
309 if (*head && ((1 << gso_type) & vif->gso_mask))
312 *head = 0; /* There must be something in this buffer now. */
318 * Prepare an SKB to be transmitted to the frontend.
320 * This function is responsible for allocating grant operations, meta
323 * It returns the number of meta structures consumed. The number of
324 * ring slots used is always equal to the number of meta slots used
325 * plus the number of GSO descriptors used. Currently, we use either
326 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
327 * frontend-side LRO).
329 static int xenvif_gop_skb(struct sk_buff *skb,
330 struct netrx_pending_operations *npo)
332 struct xenvif *vif = netdev_priv(skb->dev);
333 int nr_frags = skb_shinfo(skb)->nr_frags;
335 struct xen_netif_rx_request *req;
336 struct xenvif_rx_meta *meta;
343 old_meta_prod = npo->meta_prod;
345 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
346 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
347 gso_size = skb_shinfo(skb)->gso_size;
348 } else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
349 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
350 gso_size = skb_shinfo(skb)->gso_size;
352 gso_type = XEN_NETIF_GSO_TYPE_NONE;
356 /* Set up a GSO prefix descriptor, if necessary */
357 if ((1 << gso_type) & vif->gso_prefix_mask) {
358 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
359 meta = npo->meta + npo->meta_prod++;
360 meta->gso_type = gso_type;
361 meta->gso_size = gso_size;
366 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
367 meta = npo->meta + npo->meta_prod++;
369 if ((1 << gso_type) & vif->gso_mask) {
370 meta->gso_type = gso_type;
371 meta->gso_size = gso_size;
373 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
380 npo->copy_gref = req->gref;
383 while (data < skb_tail_pointer(skb)) {
384 unsigned int offset = offset_in_page(data);
385 unsigned int len = PAGE_SIZE - offset;
387 if (data + len > skb_tail_pointer(skb))
388 len = skb_tail_pointer(skb) - data;
390 xenvif_gop_frag_copy(vif, skb, npo,
391 virt_to_page(data), len, offset, &head);
395 for (i = 0; i < nr_frags; i++) {
396 xenvif_gop_frag_copy(vif, skb, npo,
397 skb_frag_page(&skb_shinfo(skb)->frags[i]),
398 skb_frag_size(&skb_shinfo(skb)->frags[i]),
399 skb_shinfo(skb)->frags[i].page_offset,
403 return npo->meta_prod - old_meta_prod;
407 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
408 * used to set up the operations on the top of
409 * netrx_pending_operations, which have since been done. Check that
410 * they didn't give any errors and advance over them.
412 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
413 struct netrx_pending_operations *npo)
415 struct gnttab_copy *copy_op;
416 int status = XEN_NETIF_RSP_OKAY;
419 for (i = 0; i < nr_meta_slots; i++) {
420 copy_op = npo->copy + npo->copy_cons++;
421 if (copy_op->status != GNTST_okay) {
423 "Bad status %d from copy to DOM%d.\n",
424 copy_op->status, vif->domid);
425 status = XEN_NETIF_RSP_ERROR;
432 static void xenvif_add_frag_responses(struct xenvif *vif, int status,
433 struct xenvif_rx_meta *meta,
437 unsigned long offset;
439 /* No fragments used */
440 if (nr_meta_slots <= 1)
445 for (i = 0; i < nr_meta_slots; i++) {
447 if (i == nr_meta_slots - 1)
450 flags = XEN_NETRXF_more_data;
453 make_rx_response(vif, meta[i].id, status, offset,
454 meta[i].size, flags);
458 struct skb_cb_overlay {
462 void xenvif_kick_thread(struct xenvif *vif)
467 static void xenvif_rx_action(struct xenvif *vif)
471 struct xen_netif_rx_response *resp;
472 struct sk_buff_head rxq;
476 unsigned long offset;
477 struct skb_cb_overlay *sco;
478 bool need_to_notify = false;
480 struct netrx_pending_operations npo = {
481 .copy = vif->grant_copy_op,
485 skb_queue_head_init(&rxq);
487 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
488 RING_IDX max_slots_needed;
491 /* We need a cheap worse case estimate for the number of
495 max_slots_needed = DIV_ROUND_UP(offset_in_page(skb->data) +
498 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
500 size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
501 max_slots_needed += DIV_ROUND_UP(size, PAGE_SIZE);
503 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4 ||
504 skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
507 /* If the skb may not fit then bail out now */
508 if (!xenvif_rx_ring_slots_available(vif, max_slots_needed)) {
509 skb_queue_head(&vif->rx_queue, skb);
510 need_to_notify = true;
511 vif->rx_last_skb_slots = max_slots_needed;
514 vif->rx_last_skb_slots = 0;
516 sco = (struct skb_cb_overlay *)skb->cb;
517 sco->meta_slots_used = xenvif_gop_skb(skb, &npo);
518 BUG_ON(sco->meta_slots_used > max_slots_needed);
520 __skb_queue_tail(&rxq, skb);
523 BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
528 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
529 gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
531 while ((skb = __skb_dequeue(&rxq)) != NULL) {
532 sco = (struct skb_cb_overlay *)skb->cb;
534 if ((1 << vif->meta[npo.meta_cons].gso_type) &
535 vif->gso_prefix_mask) {
536 resp = RING_GET_RESPONSE(&vif->rx,
537 vif->rx.rsp_prod_pvt++);
539 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
541 resp->offset = vif->meta[npo.meta_cons].gso_size;
542 resp->id = vif->meta[npo.meta_cons].id;
543 resp->status = sco->meta_slots_used;
546 sco->meta_slots_used--;
550 vif->dev->stats.tx_bytes += skb->len;
551 vif->dev->stats.tx_packets++;
553 status = xenvif_check_gop(vif, sco->meta_slots_used, &npo);
555 if (sco->meta_slots_used == 1)
558 flags = XEN_NETRXF_more_data;
560 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
561 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
562 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
563 /* remote but checksummed. */
564 flags |= XEN_NETRXF_data_validated;
567 resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
569 vif->meta[npo.meta_cons].size,
572 if ((1 << vif->meta[npo.meta_cons].gso_type) &
574 struct xen_netif_extra_info *gso =
575 (struct xen_netif_extra_info *)
576 RING_GET_RESPONSE(&vif->rx,
577 vif->rx.rsp_prod_pvt++);
579 resp->flags |= XEN_NETRXF_extra_info;
581 gso->u.gso.type = vif->meta[npo.meta_cons].gso_type;
582 gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
584 gso->u.gso.features = 0;
586 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
590 xenvif_add_frag_responses(vif, status,
591 vif->meta + npo.meta_cons + 1,
592 sco->meta_slots_used);
594 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
596 need_to_notify |= !!ret;
598 npo.meta_cons += sco->meta_slots_used;
604 notify_remote_via_irq(vif->rx_irq);
607 void xenvif_check_rx_xenvif(struct xenvif *vif)
611 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
614 napi_schedule(&vif->napi);
617 static void tx_add_credit(struct xenvif *vif)
619 unsigned long max_burst, max_credit;
622 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
623 * Otherwise the interface can seize up due to insufficient credit.
625 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
626 max_burst = min(max_burst, 131072UL);
627 max_burst = max(max_burst, vif->credit_bytes);
629 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
630 max_credit = vif->remaining_credit + vif->credit_bytes;
631 if (max_credit < vif->remaining_credit)
632 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
634 vif->remaining_credit = min(max_credit, max_burst);
637 static void tx_credit_callback(unsigned long data)
639 struct xenvif *vif = (struct xenvif *)data;
641 xenvif_check_rx_xenvif(vif);
644 static void xenvif_tx_err(struct xenvif *vif,
645 struct xen_netif_tx_request *txp, RING_IDX end)
647 RING_IDX cons = vif->tx.req_cons;
650 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
653 txp = RING_GET_REQUEST(&vif->tx, cons++);
655 vif->tx.req_cons = cons;
658 static void xenvif_fatal_tx_err(struct xenvif *vif)
660 netdev_err(vif->dev, "fatal error; disabling device\n");
661 xenvif_carrier_off(vif);
664 static int xenvif_count_requests(struct xenvif *vif,
665 struct xen_netif_tx_request *first,
666 struct xen_netif_tx_request *txp,
669 RING_IDX cons = vif->tx.req_cons;
674 if (!(first->flags & XEN_NETTXF_more_data))
678 struct xen_netif_tx_request dropped_tx = { 0 };
680 if (slots >= work_to_do) {
682 "Asked for %d slots but exceeds this limit\n",
684 xenvif_fatal_tx_err(vif);
688 /* This guest is really using too many slots and
689 * considered malicious.
691 if (unlikely(slots >= fatal_skb_slots)) {
693 "Malicious frontend using %d slots, threshold %u\n",
694 slots, fatal_skb_slots);
695 xenvif_fatal_tx_err(vif);
699 /* Xen network protocol had implicit dependency on
700 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
701 * the historical MAX_SKB_FRAGS value 18 to honor the
702 * same behavior as before. Any packet using more than
703 * 18 slots but less than fatal_skb_slots slots is
706 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
709 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
710 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
717 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
720 /* If the guest submitted a frame >= 64 KiB then
721 * first->size overflowed and following slots will
722 * appear to be larger than the frame.
724 * This cannot be fatal error as there are buggy
725 * frontends that do this.
727 * Consume all slots and drop the packet.
729 if (!drop_err && txp->size > first->size) {
732 "Invalid tx request, slot size %u > remaining size %u\n",
733 txp->size, first->size);
737 first->size -= txp->size;
740 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
741 netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
742 txp->offset, txp->size);
743 xenvif_fatal_tx_err(vif);
747 more_data = txp->flags & XEN_NETTXF_more_data;
755 xenvif_tx_err(vif, first, cons + slots);
762 static struct page *xenvif_alloc_page(struct xenvif *vif,
767 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
770 vif->mmap_pages[pending_idx] = page;
775 static struct gnttab_copy *xenvif_get_requests(struct xenvif *vif,
777 struct xen_netif_tx_request *txp,
778 struct gnttab_copy *gop)
780 struct skb_shared_info *shinfo = skb_shinfo(skb);
781 skb_frag_t *frags = shinfo->frags;
782 u16 pending_idx = *((u16 *)skb->data);
786 pending_ring_idx_t index, start_idx = 0;
788 unsigned int nr_slots;
789 struct pending_tx_info *first = NULL;
791 /* At this point shinfo->nr_frags is in fact the number of
792 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
794 nr_slots = shinfo->nr_frags;
796 /* Skip first skb fragment if it is on same page as header fragment. */
797 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
799 /* Coalesce tx requests, at this point the packet passed in
800 * should be <= 64K. Any packets larger than 64K have been
801 * handled in xenvif_count_requests().
803 for (shinfo->nr_frags = slot = start; slot < nr_slots;
804 shinfo->nr_frags++) {
805 struct pending_tx_info *pending_tx_info =
806 vif->pending_tx_info;
808 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
814 while (dst_offset < PAGE_SIZE && slot < nr_slots) {
815 gop->flags = GNTCOPY_source_gref;
817 gop->source.u.ref = txp->gref;
818 gop->source.domid = vif->domid;
819 gop->source.offset = txp->offset;
821 gop->dest.domid = DOMID_SELF;
823 gop->dest.offset = dst_offset;
824 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
826 if (dst_offset + txp->size > PAGE_SIZE) {
827 /* This page can only merge a portion
828 * of tx request. Do not increment any
829 * pointer / counter here. The txp
830 * will be dealt with in future
831 * rounds, eventually hitting the
834 gop->len = PAGE_SIZE - dst_offset;
835 txp->offset += gop->len;
836 txp->size -= gop->len;
837 dst_offset += gop->len; /* quit loop */
839 /* This tx request can be merged in the page */
840 gop->len = txp->size;
841 dst_offset += gop->len;
843 index = pending_index(vif->pending_cons++);
845 pending_idx = vif->pending_ring[index];
847 memcpy(&pending_tx_info[pending_idx].req, txp,
850 /* Poison these fields, corresponding
851 * fields for head tx req will be set
852 * to correct values after the loop.
854 vif->mmap_pages[pending_idx] = (void *)(~0UL);
855 pending_tx_info[pending_idx].head =
856 INVALID_PENDING_RING_IDX;
859 first = &pending_tx_info[pending_idx];
861 head_idx = pending_idx;
871 first->req.offset = 0;
872 first->req.size = dst_offset;
873 first->head = start_idx;
874 vif->mmap_pages[head_idx] = page;
875 frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
878 BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
882 /* Unwind, freeing all pages and sending error responses. */
883 while (shinfo->nr_frags-- > start) {
884 xenvif_idx_release(vif,
885 frag_get_pending_idx(&frags[shinfo->nr_frags]),
886 XEN_NETIF_RSP_ERROR);
888 /* The head too, if necessary. */
890 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
895 static int xenvif_tx_check_gop(struct xenvif *vif,
897 struct gnttab_copy **gopp)
899 struct gnttab_copy *gop = *gopp;
900 u16 pending_idx = *((u16 *)skb->data);
901 struct skb_shared_info *shinfo = skb_shinfo(skb);
902 struct pending_tx_info *tx_info;
903 int nr_frags = shinfo->nr_frags;
905 u16 peek; /* peek into next tx request */
907 /* Check status of header. */
910 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
912 /* Skip first skb fragment if it is on same page as header fragment. */
913 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
915 for (i = start; i < nr_frags; i++) {
917 pending_ring_idx_t head;
919 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
920 tx_info = &vif->pending_tx_info[pending_idx];
921 head = tx_info->head;
923 /* Check error status: if okay then remember grant handle. */
925 newerr = (++gop)->status;
928 peek = vif->pending_ring[pending_index(++head)];
929 } while (!pending_tx_is_head(vif, peek));
931 if (likely(!newerr)) {
932 /* Had a previous error? Invalidate this fragment. */
934 xenvif_idx_release(vif, pending_idx,
939 /* Error on this fragment: respond to client with an error. */
940 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
942 /* Not the first error? Preceding frags already invalidated. */
946 /* First error: invalidate header and preceding fragments. */
947 pending_idx = *((u16 *)skb->data);
948 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
949 for (j = start; j < i; j++) {
950 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
951 xenvif_idx_release(vif, pending_idx,
955 /* Remember the error: invalidate all subsequent fragments. */
963 static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
965 struct skb_shared_info *shinfo = skb_shinfo(skb);
966 int nr_frags = shinfo->nr_frags;
969 for (i = 0; i < nr_frags; i++) {
970 skb_frag_t *frag = shinfo->frags + i;
971 struct xen_netif_tx_request *txp;
975 pending_idx = frag_get_pending_idx(frag);
977 txp = &vif->pending_tx_info[pending_idx].req;
978 page = virt_to_page(idx_to_kaddr(vif, pending_idx));
979 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
980 skb->len += txp->size;
981 skb->data_len += txp->size;
982 skb->truesize += txp->size;
984 /* Take an extra reference to offset xenvif_idx_release */
985 get_page(vif->mmap_pages[pending_idx]);
986 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
990 static int xenvif_get_extras(struct xenvif *vif,
991 struct xen_netif_extra_info *extras,
994 struct xen_netif_extra_info extra;
995 RING_IDX cons = vif->tx.req_cons;
998 if (unlikely(work_to_do-- <= 0)) {
999 netdev_err(vif->dev, "Missing extra info\n");
1000 xenvif_fatal_tx_err(vif);
1004 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1006 if (unlikely(!extra.type ||
1007 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1008 vif->tx.req_cons = ++cons;
1009 netdev_err(vif->dev,
1010 "Invalid extra type: %d\n", extra.type);
1011 xenvif_fatal_tx_err(vif);
1015 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1016 vif->tx.req_cons = ++cons;
1017 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1022 static int xenvif_set_skb_gso(struct xenvif *vif,
1023 struct sk_buff *skb,
1024 struct xen_netif_extra_info *gso)
1026 if (!gso->u.gso.size) {
1027 netdev_err(vif->dev, "GSO size must not be zero.\n");
1028 xenvif_fatal_tx_err(vif);
1032 switch (gso->u.gso.type) {
1033 case XEN_NETIF_GSO_TYPE_TCPV4:
1034 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1036 case XEN_NETIF_GSO_TYPE_TCPV6:
1037 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1040 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1041 xenvif_fatal_tx_err(vif);
1045 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1046 /* gso_segs will be calculated later */
1051 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1053 bool recalculate_partial_csum = false;
1055 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1056 * peers can fail to set NETRXF_csum_blank when sending a GSO
1057 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1058 * recalculate the partial checksum.
1060 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1061 vif->rx_gso_checksum_fixup++;
1062 skb->ip_summed = CHECKSUM_PARTIAL;
1063 recalculate_partial_csum = true;
1066 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1067 if (skb->ip_summed != CHECKSUM_PARTIAL)
1070 return skb_checksum_setup(skb, recalculate_partial_csum);
1073 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1075 u64 now = get_jiffies_64();
1076 u64 next_credit = vif->credit_window_start +
1077 msecs_to_jiffies(vif->credit_usec / 1000);
1079 /* Timer could already be pending in rare cases. */
1080 if (timer_pending(&vif->credit_timeout))
1083 /* Passed the point where we can replenish credit? */
1084 if (time_after_eq64(now, next_credit)) {
1085 vif->credit_window_start = now;
1089 /* Still too big to send right now? Set a callback. */
1090 if (size > vif->remaining_credit) {
1091 vif->credit_timeout.data =
1093 vif->credit_timeout.function =
1095 mod_timer(&vif->credit_timeout,
1097 vif->credit_window_start = next_credit;
1105 static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget)
1107 struct gnttab_copy *gop = vif->tx_copy_ops, *request_gop;
1108 struct sk_buff *skb;
1111 while ((nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
1112 < MAX_PENDING_REQS) &&
1113 (skb_queue_len(&vif->tx_queue) < budget)) {
1114 struct xen_netif_tx_request txreq;
1115 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1117 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1121 unsigned int data_len;
1122 pending_ring_idx_t index;
1124 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1125 XEN_NETIF_TX_RING_SIZE) {
1126 netdev_err(vif->dev,
1127 "Impossible number of requests. "
1128 "req_prod %d, req_cons %d, size %ld\n",
1129 vif->tx.sring->req_prod, vif->tx.req_cons,
1130 XEN_NETIF_TX_RING_SIZE);
1131 xenvif_fatal_tx_err(vif);
1135 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);
1139 idx = vif->tx.req_cons;
1140 rmb(); /* Ensure that we see the request before we copy it. */
1141 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1143 /* Credit-based scheduling. */
1144 if (txreq.size > vif->remaining_credit &&
1145 tx_credit_exceeded(vif, txreq.size))
1148 vif->remaining_credit -= txreq.size;
1151 vif->tx.req_cons = ++idx;
1153 memset(extras, 0, sizeof(extras));
1154 if (txreq.flags & XEN_NETTXF_extra_info) {
1155 work_to_do = xenvif_get_extras(vif, extras,
1157 idx = vif->tx.req_cons;
1158 if (unlikely(work_to_do < 0))
1162 ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
1163 if (unlikely(ret < 0))
1168 if (unlikely(txreq.size < ETH_HLEN)) {
1169 netdev_dbg(vif->dev,
1170 "Bad packet size: %d\n", txreq.size);
1171 xenvif_tx_err(vif, &txreq, idx);
1175 /* No crossing a page as the payload mustn't fragment. */
1176 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1177 netdev_err(vif->dev,
1178 "txreq.offset: %x, size: %u, end: %lu\n",
1179 txreq.offset, txreq.size,
1180 (txreq.offset&~PAGE_MASK) + txreq.size);
1181 xenvif_fatal_tx_err(vif);
1185 index = pending_index(vif->pending_cons);
1186 pending_idx = vif->pending_ring[index];
1188 data_len = (txreq.size > PKT_PROT_LEN &&
1189 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1190 PKT_PROT_LEN : txreq.size;
1192 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1193 GFP_ATOMIC | __GFP_NOWARN);
1194 if (unlikely(skb == NULL)) {
1195 netdev_dbg(vif->dev,
1196 "Can't allocate a skb in start_xmit.\n");
1197 xenvif_tx_err(vif, &txreq, idx);
1201 /* Packets passed to netif_rx() must have some headroom. */
1202 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1204 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1205 struct xen_netif_extra_info *gso;
1206 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1208 if (xenvif_set_skb_gso(vif, skb, gso)) {
1209 /* Failure in xenvif_set_skb_gso is fatal. */
1215 /* XXX could copy straight to head */
1216 page = xenvif_alloc_page(vif, pending_idx);
1219 xenvif_tx_err(vif, &txreq, idx);
1223 gop->source.u.ref = txreq.gref;
1224 gop->source.domid = vif->domid;
1225 gop->source.offset = txreq.offset;
1227 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1228 gop->dest.domid = DOMID_SELF;
1229 gop->dest.offset = txreq.offset;
1231 gop->len = txreq.size;
1232 gop->flags = GNTCOPY_source_gref;
1236 memcpy(&vif->pending_tx_info[pending_idx].req,
1237 &txreq, sizeof(txreq));
1238 vif->pending_tx_info[pending_idx].head = index;
1239 *((u16 *)skb->data) = pending_idx;
1241 __skb_put(skb, data_len);
1243 skb_shinfo(skb)->nr_frags = ret;
1244 if (data_len < txreq.size) {
1245 skb_shinfo(skb)->nr_frags++;
1246 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1249 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1250 INVALID_PENDING_IDX);
1253 vif->pending_cons++;
1255 request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
1256 if (request_gop == NULL) {
1258 xenvif_tx_err(vif, &txreq, idx);
1263 __skb_queue_tail(&vif->tx_queue, skb);
1265 vif->tx.req_cons = idx;
1267 if ((gop-vif->tx_copy_ops) >= ARRAY_SIZE(vif->tx_copy_ops))
1271 return gop - vif->tx_copy_ops;
1275 static int xenvif_tx_submit(struct xenvif *vif)
1277 struct gnttab_copy *gop = vif->tx_copy_ops;
1278 struct sk_buff *skb;
1281 while ((skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
1282 struct xen_netif_tx_request *txp;
1286 pending_idx = *((u16 *)skb->data);
1287 txp = &vif->pending_tx_info[pending_idx].req;
1289 /* Check the remap error code. */
1290 if (unlikely(xenvif_tx_check_gop(vif, skb, &gop))) {
1291 netdev_dbg(vif->dev, "netback grant failed.\n");
1292 skb_shinfo(skb)->nr_frags = 0;
1297 data_len = skb->len;
1299 (void *)(idx_to_kaddr(vif, pending_idx)|txp->offset),
1301 if (data_len < txp->size) {
1302 /* Append the packet payload as a fragment. */
1303 txp->offset += data_len;
1304 txp->size -= data_len;
1306 /* Schedule a response immediately. */
1307 xenvif_idx_release(vif, pending_idx,
1308 XEN_NETIF_RSP_OKAY);
1311 if (txp->flags & XEN_NETTXF_csum_blank)
1312 skb->ip_summed = CHECKSUM_PARTIAL;
1313 else if (txp->flags & XEN_NETTXF_data_validated)
1314 skb->ip_summed = CHECKSUM_UNNECESSARY;
1316 xenvif_fill_frags(vif, skb);
1318 if (skb_is_nonlinear(skb) && skb_headlen(skb) < PKT_PROT_LEN) {
1319 int target = min_t(int, skb->len, PKT_PROT_LEN);
1320 __pskb_pull_tail(skb, target - skb_headlen(skb));
1323 skb->dev = vif->dev;
1324 skb->protocol = eth_type_trans(skb, skb->dev);
1325 skb_reset_network_header(skb);
1327 if (checksum_setup(vif, skb)) {
1328 netdev_dbg(vif->dev,
1329 "Can't setup checksum in net_tx_action\n");
1334 skb_probe_transport_header(skb, 0);
1336 /* If the packet is GSO then we will have just set up the
1337 * transport header offset in checksum_setup so it's now
1338 * straightforward to calculate gso_segs.
1340 if (skb_is_gso(skb)) {
1341 int mss = skb_shinfo(skb)->gso_size;
1342 int hdrlen = skb_transport_header(skb) -
1343 skb_mac_header(skb) +
1346 skb_shinfo(skb)->gso_segs =
1347 DIV_ROUND_UP(skb->len - hdrlen, mss);
1350 vif->dev->stats.rx_bytes += skb->len;
1351 vif->dev->stats.rx_packets++;
1355 netif_receive_skb(skb);
1361 /* Called after netfront has transmitted */
1362 int xenvif_tx_action(struct xenvif *vif, int budget)
1367 if (unlikely(!tx_work_todo(vif)))
1370 nr_gops = xenvif_tx_build_gops(vif, budget);
1375 gnttab_batch_copy(vif->tx_copy_ops, nr_gops);
1377 work_done = xenvif_tx_submit(vif);
1382 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
1385 struct pending_tx_info *pending_tx_info;
1386 pending_ring_idx_t head;
1387 u16 peek; /* peek into next tx request */
1389 BUG_ON(vif->mmap_pages[pending_idx] == (void *)(~0UL));
1391 /* Already complete? */
1392 if (vif->mmap_pages[pending_idx] == NULL)
1395 pending_tx_info = &vif->pending_tx_info[pending_idx];
1397 head = pending_tx_info->head;
1399 BUG_ON(!pending_tx_is_head(vif, head));
1400 BUG_ON(vif->pending_ring[pending_index(head)] != pending_idx);
1403 pending_ring_idx_t index;
1404 pending_ring_idx_t idx = pending_index(head);
1405 u16 info_idx = vif->pending_ring[idx];
1407 pending_tx_info = &vif->pending_tx_info[info_idx];
1408 make_tx_response(vif, &pending_tx_info->req, status);
1410 /* Setting any number other than
1411 * INVALID_PENDING_RING_IDX indicates this slot is
1412 * starting a new packet / ending a previous packet.
1414 pending_tx_info->head = 0;
1416 index = pending_index(vif->pending_prod++);
1417 vif->pending_ring[index] = vif->pending_ring[info_idx];
1419 peek = vif->pending_ring[pending_index(++head)];
1421 } while (!pending_tx_is_head(vif, peek));
1423 put_page(vif->mmap_pages[pending_idx]);
1424 vif->mmap_pages[pending_idx] = NULL;
1428 static void make_tx_response(struct xenvif *vif,
1429 struct xen_netif_tx_request *txp,
1432 RING_IDX i = vif->tx.rsp_prod_pvt;
1433 struct xen_netif_tx_response *resp;
1436 resp = RING_GET_RESPONSE(&vif->tx, i);
1440 if (txp->flags & XEN_NETTXF_extra_info)
1441 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1443 vif->tx.rsp_prod_pvt = ++i;
1444 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1446 notify_remote_via_irq(vif->tx_irq);
1449 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1456 RING_IDX i = vif->rx.rsp_prod_pvt;
1457 struct xen_netif_rx_response *resp;
1459 resp = RING_GET_RESPONSE(&vif->rx, i);
1460 resp->offset = offset;
1461 resp->flags = flags;
1463 resp->status = (s16)size;
1465 resp->status = (s16)st;
1467 vif->rx.rsp_prod_pvt = ++i;
1472 static inline int rx_work_todo(struct xenvif *vif)
1474 return !skb_queue_empty(&vif->rx_queue) &&
1475 xenvif_rx_ring_slots_available(vif, vif->rx_last_skb_slots);
1478 static inline int tx_work_todo(struct xenvif *vif)
1481 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)) &&
1482 (nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
1483 < MAX_PENDING_REQS))
1489 void xenvif_unmap_frontend_rings(struct xenvif *vif)
1492 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1495 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1499 int xenvif_map_frontend_rings(struct xenvif *vif,
1500 grant_ref_t tx_ring_ref,
1501 grant_ref_t rx_ring_ref)
1504 struct xen_netif_tx_sring *txs;
1505 struct xen_netif_rx_sring *rxs;
1509 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1510 tx_ring_ref, &addr);
1514 txs = (struct xen_netif_tx_sring *)addr;
1515 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1517 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1518 rx_ring_ref, &addr);
1522 rxs = (struct xen_netif_rx_sring *)addr;
1523 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1528 xenvif_unmap_frontend_rings(vif);
1532 void xenvif_stop_queue(struct xenvif *vif)
1534 if (!vif->can_queue)
1537 netif_stop_queue(vif->dev);
1540 static void xenvif_start_queue(struct xenvif *vif)
1542 if (xenvif_schedulable(vif))
1543 netif_wake_queue(vif->dev);
1546 int xenvif_kthread(void *data)
1548 struct xenvif *vif = data;
1549 struct sk_buff *skb;
1551 while (!kthread_should_stop()) {
1552 wait_event_interruptible(vif->wq,
1553 rx_work_todo(vif) ||
1554 kthread_should_stop());
1555 if (kthread_should_stop())
1558 if (!skb_queue_empty(&vif->rx_queue))
1559 xenvif_rx_action(vif);
1561 if (skb_queue_empty(&vif->rx_queue) &&
1562 netif_queue_stopped(vif->dev))
1563 xenvif_start_queue(vif);
1568 /* Bin any remaining skbs */
1569 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL)
1575 static int __init netback_init(void)
1582 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1583 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1584 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1585 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1588 rc = xenvif_xenbus_init();
1598 module_init(netback_init);
1600 static void __exit netback_fini(void)
1602 xenvif_xenbus_fini();
1604 module_exit(netback_fini);
1606 MODULE_LICENSE("Dual BSD/GPL");
1607 MODULE_ALIAS("xen-backend:vif");