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 /* This is a miniumum size for the linear area to avoid lots of
105 * calls to __pskb_pull_tail() as we set up checksum offsets. The
106 * value 128 was chosen as it covers all IPv4 and most likely
109 #define PKT_PROT_LEN 128
111 static u16 frag_get_pending_idx(skb_frag_t *frag)
113 return (u16)frag->page_offset;
116 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
118 frag->page_offset = pending_idx;
121 static inline pending_ring_idx_t pending_index(unsigned i)
123 return i & (MAX_PENDING_REQS-1);
126 bool xenvif_rx_ring_slots_available(struct xenvif *vif, int needed)
131 prod = vif->rx.sring->req_prod;
132 cons = vif->rx.req_cons;
134 if (prod - cons >= needed)
137 vif->rx.sring->req_event = prod + 1;
139 /* Make sure event is visible before we check prod
143 } while (vif->rx.sring->req_prod != prod);
149 * Returns true if we should start a new receive buffer instead of
150 * adding 'size' bytes to a buffer which currently contains 'offset'
153 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
155 /* simple case: we have completely filled the current buffer. */
156 if (offset == MAX_BUFFER_OFFSET)
160 * complex case: start a fresh buffer if the current frag
161 * would overflow the current buffer but only if:
162 * (i) this frag would fit completely in the next buffer
163 * and (ii) there is already some data in the current buffer
164 * and (iii) this is not the head buffer.
167 * - (i) stops us splitting a frag into two copies
168 * unless the frag is too large for a single buffer.
169 * - (ii) stops us from leaving a buffer pointlessly empty.
170 * - (iii) stops us leaving the first buffer
171 * empty. Strictly speaking this is already covered
172 * by (ii) but is explicitly checked because
173 * netfront relies on the first buffer being
174 * non-empty and can crash otherwise.
176 * This means we will effectively linearise small
177 * frags but do not needlessly split large buffers
178 * into multiple copies tend to give large frags their
179 * own buffers as before.
181 if ((offset + size > MAX_BUFFER_OFFSET) &&
182 (size <= MAX_BUFFER_OFFSET) && offset && !head)
188 struct netrx_pending_operations {
189 unsigned copy_prod, copy_cons;
190 unsigned meta_prod, meta_cons;
191 struct gnttab_copy *copy;
192 struct xenvif_rx_meta *meta;
194 grant_ref_t copy_gref;
197 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
198 struct netrx_pending_operations *npo)
200 struct xenvif_rx_meta *meta;
201 struct xen_netif_rx_request *req;
203 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
205 meta = npo->meta + npo->meta_prod++;
206 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
212 npo->copy_gref = req->gref;
218 * Set up the grant operations for this fragment. If it's a flipping
219 * interface, we also set up the unmap request from here.
221 static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
222 struct netrx_pending_operations *npo,
223 struct page *page, unsigned long size,
224 unsigned long offset, int *head)
226 struct gnttab_copy *copy_gop;
227 struct xenvif_rx_meta *meta;
231 /* Data must not cross a page boundary. */
232 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
234 meta = npo->meta + npo->meta_prod - 1;
236 /* Skip unused frames from start of page */
237 page += offset >> PAGE_SHIFT;
238 offset &= ~PAGE_MASK;
241 BUG_ON(offset >= PAGE_SIZE);
242 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
244 bytes = PAGE_SIZE - offset;
249 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
251 * Netfront requires there to be some data in the head
256 meta = get_next_rx_buffer(vif, npo);
259 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
260 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
262 copy_gop = npo->copy + npo->copy_prod++;
263 copy_gop->flags = GNTCOPY_dest_gref;
264 copy_gop->len = bytes;
266 copy_gop->source.domid = DOMID_SELF;
267 copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
268 copy_gop->source.offset = offset;
270 copy_gop->dest.domid = vif->domid;
271 copy_gop->dest.offset = npo->copy_off;
272 copy_gop->dest.u.ref = npo->copy_gref;
274 npo->copy_off += bytes;
281 if (offset == PAGE_SIZE && size) {
282 BUG_ON(!PageCompound(page));
287 /* Leave a gap for the GSO descriptor. */
288 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
289 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
290 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
291 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
293 gso_type = XEN_NETIF_GSO_TYPE_NONE;
295 if (*head && ((1 << gso_type) & vif->gso_mask))
298 *head = 0; /* There must be something in this buffer now. */
304 * Prepare an SKB to be transmitted to the frontend.
306 * This function is responsible for allocating grant operations, meta
309 * It returns the number of meta structures consumed. The number of
310 * ring slots used is always equal to the number of meta slots used
311 * plus the number of GSO descriptors used. Currently, we use either
312 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
313 * frontend-side LRO).
315 static int xenvif_gop_skb(struct sk_buff *skb,
316 struct netrx_pending_operations *npo)
318 struct xenvif *vif = netdev_priv(skb->dev);
319 int nr_frags = skb_shinfo(skb)->nr_frags;
321 struct xen_netif_rx_request *req;
322 struct xenvif_rx_meta *meta;
329 old_meta_prod = npo->meta_prod;
331 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
332 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
333 gso_size = skb_shinfo(skb)->gso_size;
334 } else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
335 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
336 gso_size = skb_shinfo(skb)->gso_size;
338 gso_type = XEN_NETIF_GSO_TYPE_NONE;
342 /* Set up a GSO prefix descriptor, if necessary */
343 if ((1 << gso_type) & vif->gso_prefix_mask) {
344 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
345 meta = npo->meta + npo->meta_prod++;
346 meta->gso_type = gso_type;
347 meta->gso_size = gso_size;
352 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
353 meta = npo->meta + npo->meta_prod++;
355 if ((1 << gso_type) & vif->gso_mask) {
356 meta->gso_type = gso_type;
357 meta->gso_size = gso_size;
359 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
366 npo->copy_gref = req->gref;
369 while (data < skb_tail_pointer(skb)) {
370 unsigned int offset = offset_in_page(data);
371 unsigned int len = PAGE_SIZE - offset;
373 if (data + len > skb_tail_pointer(skb))
374 len = skb_tail_pointer(skb) - data;
376 xenvif_gop_frag_copy(vif, skb, npo,
377 virt_to_page(data), len, offset, &head);
381 for (i = 0; i < nr_frags; i++) {
382 xenvif_gop_frag_copy(vif, skb, npo,
383 skb_frag_page(&skb_shinfo(skb)->frags[i]),
384 skb_frag_size(&skb_shinfo(skb)->frags[i]),
385 skb_shinfo(skb)->frags[i].page_offset,
389 return npo->meta_prod - old_meta_prod;
393 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
394 * used to set up the operations on the top of
395 * netrx_pending_operations, which have since been done. Check that
396 * they didn't give any errors and advance over them.
398 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
399 struct netrx_pending_operations *npo)
401 struct gnttab_copy *copy_op;
402 int status = XEN_NETIF_RSP_OKAY;
405 for (i = 0; i < nr_meta_slots; i++) {
406 copy_op = npo->copy + npo->copy_cons++;
407 if (copy_op->status != GNTST_okay) {
409 "Bad status %d from copy to DOM%d.\n",
410 copy_op->status, vif->domid);
411 status = XEN_NETIF_RSP_ERROR;
418 static void xenvif_add_frag_responses(struct xenvif *vif, int status,
419 struct xenvif_rx_meta *meta,
423 unsigned long offset;
425 /* No fragments used */
426 if (nr_meta_slots <= 1)
431 for (i = 0; i < nr_meta_slots; i++) {
433 if (i == nr_meta_slots - 1)
436 flags = XEN_NETRXF_more_data;
439 make_rx_response(vif, meta[i].id, status, offset,
440 meta[i].size, flags);
444 struct xenvif_rx_cb {
448 #define XENVIF_RX_CB(skb) ((struct xenvif_rx_cb *)(skb)->cb)
450 void xenvif_kick_thread(struct xenvif *vif)
455 static void xenvif_rx_action(struct xenvif *vif)
459 struct xen_netif_rx_response *resp;
460 struct sk_buff_head rxq;
464 unsigned long offset;
465 bool need_to_notify = false;
467 struct netrx_pending_operations npo = {
468 .copy = vif->grant_copy_op,
472 skb_queue_head_init(&rxq);
474 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
475 RING_IDX max_slots_needed;
478 /* We need a cheap worse case estimate for the number of
482 max_slots_needed = DIV_ROUND_UP(offset_in_page(skb->data) +
485 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
487 size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
488 max_slots_needed += DIV_ROUND_UP(size, PAGE_SIZE);
490 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4 ||
491 skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
494 /* If the skb may not fit then bail out now */
495 if (!xenvif_rx_ring_slots_available(vif, max_slots_needed)) {
496 skb_queue_head(&vif->rx_queue, skb);
497 need_to_notify = true;
498 vif->rx_last_skb_slots = max_slots_needed;
501 vif->rx_last_skb_slots = 0;
503 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo);
504 BUG_ON(XENVIF_RX_CB(skb)->meta_slots_used > max_slots_needed);
506 __skb_queue_tail(&rxq, skb);
509 BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
514 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
515 gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
517 while ((skb = __skb_dequeue(&rxq)) != NULL) {
519 if ((1 << vif->meta[npo.meta_cons].gso_type) &
520 vif->gso_prefix_mask) {
521 resp = RING_GET_RESPONSE(&vif->rx,
522 vif->rx.rsp_prod_pvt++);
524 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
526 resp->offset = vif->meta[npo.meta_cons].gso_size;
527 resp->id = vif->meta[npo.meta_cons].id;
528 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
531 XENVIF_RX_CB(skb)->meta_slots_used--;
535 vif->dev->stats.tx_bytes += skb->len;
536 vif->dev->stats.tx_packets++;
538 status = xenvif_check_gop(vif,
539 XENVIF_RX_CB(skb)->meta_slots_used,
542 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
545 flags = XEN_NETRXF_more_data;
547 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
548 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
549 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
550 /* remote but checksummed. */
551 flags |= XEN_NETRXF_data_validated;
554 resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
556 vif->meta[npo.meta_cons].size,
559 if ((1 << vif->meta[npo.meta_cons].gso_type) &
561 struct xen_netif_extra_info *gso =
562 (struct xen_netif_extra_info *)
563 RING_GET_RESPONSE(&vif->rx,
564 vif->rx.rsp_prod_pvt++);
566 resp->flags |= XEN_NETRXF_extra_info;
568 gso->u.gso.type = vif->meta[npo.meta_cons].gso_type;
569 gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
571 gso->u.gso.features = 0;
573 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
577 xenvif_add_frag_responses(vif, status,
578 vif->meta + npo.meta_cons + 1,
579 XENVIF_RX_CB(skb)->meta_slots_used);
581 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
583 need_to_notify |= !!ret;
585 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
591 notify_remote_via_irq(vif->rx_irq);
594 void xenvif_check_rx_xenvif(struct xenvif *vif)
598 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
601 napi_schedule(&vif->napi);
604 static void tx_add_credit(struct xenvif *vif)
606 unsigned long max_burst, max_credit;
609 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
610 * Otherwise the interface can seize up due to insufficient credit.
612 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
613 max_burst = min(max_burst, 131072UL);
614 max_burst = max(max_burst, vif->credit_bytes);
616 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
617 max_credit = vif->remaining_credit + vif->credit_bytes;
618 if (max_credit < vif->remaining_credit)
619 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
621 vif->remaining_credit = min(max_credit, max_burst);
624 static void tx_credit_callback(unsigned long data)
626 struct xenvif *vif = (struct xenvif *)data;
628 xenvif_check_rx_xenvif(vif);
631 static void xenvif_tx_err(struct xenvif *vif,
632 struct xen_netif_tx_request *txp, RING_IDX end)
634 RING_IDX cons = vif->tx.req_cons;
637 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
640 txp = RING_GET_REQUEST(&vif->tx, cons++);
642 vif->tx.req_cons = cons;
645 static void xenvif_fatal_tx_err(struct xenvif *vif)
647 netdev_err(vif->dev, "fatal error; disabling device\n");
648 xenvif_carrier_off(vif);
651 static int xenvif_count_requests(struct xenvif *vif,
652 struct xen_netif_tx_request *first,
653 struct xen_netif_tx_request *txp,
656 RING_IDX cons = vif->tx.req_cons;
661 if (!(first->flags & XEN_NETTXF_more_data))
665 struct xen_netif_tx_request dropped_tx = { 0 };
667 if (slots >= work_to_do) {
669 "Asked for %d slots but exceeds this limit\n",
671 xenvif_fatal_tx_err(vif);
675 /* This guest is really using too many slots and
676 * considered malicious.
678 if (unlikely(slots >= fatal_skb_slots)) {
680 "Malicious frontend using %d slots, threshold %u\n",
681 slots, fatal_skb_slots);
682 xenvif_fatal_tx_err(vif);
686 /* Xen network protocol had implicit dependency on
687 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
688 * the historical MAX_SKB_FRAGS value 18 to honor the
689 * same behavior as before. Any packet using more than
690 * 18 slots but less than fatal_skb_slots slots is
693 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
696 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
697 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
704 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
707 /* If the guest submitted a frame >= 64 KiB then
708 * first->size overflowed and following slots will
709 * appear to be larger than the frame.
711 * This cannot be fatal error as there are buggy
712 * frontends that do this.
714 * Consume all slots and drop the packet.
716 if (!drop_err && txp->size > first->size) {
719 "Invalid tx request, slot size %u > remaining size %u\n",
720 txp->size, first->size);
724 first->size -= txp->size;
727 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
728 netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
729 txp->offset, txp->size);
730 xenvif_fatal_tx_err(vif);
734 more_data = txp->flags & XEN_NETTXF_more_data;
742 xenvif_tx_err(vif, first, cons + slots);
749 static struct page *xenvif_alloc_page(struct xenvif *vif,
754 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
757 vif->mmap_pages[pending_idx] = page;
763 struct xenvif_tx_cb {
767 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
769 static struct gnttab_copy *xenvif_get_requests(struct xenvif *vif,
771 struct xen_netif_tx_request *txp,
772 struct gnttab_copy *gop)
774 struct skb_shared_info *shinfo = skb_shinfo(skb);
775 skb_frag_t *frags = shinfo->frags;
776 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
780 pending_ring_idx_t index, start_idx = 0;
782 unsigned int nr_slots;
783 struct pending_tx_info *first = NULL;
785 /* At this point shinfo->nr_frags is in fact the number of
786 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
788 nr_slots = shinfo->nr_frags;
790 /* Skip first skb fragment if it is on same page as header fragment. */
791 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
793 /* Coalesce tx requests, at this point the packet passed in
794 * should be <= 64K. Any packets larger than 64K have been
795 * handled in xenvif_count_requests().
797 for (shinfo->nr_frags = slot = start; slot < nr_slots;
798 shinfo->nr_frags++) {
799 struct pending_tx_info *pending_tx_info =
800 vif->pending_tx_info;
802 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
808 while (dst_offset < PAGE_SIZE && slot < nr_slots) {
809 gop->flags = GNTCOPY_source_gref;
811 gop->source.u.ref = txp->gref;
812 gop->source.domid = vif->domid;
813 gop->source.offset = txp->offset;
815 gop->dest.domid = DOMID_SELF;
817 gop->dest.offset = dst_offset;
818 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
820 if (dst_offset + txp->size > PAGE_SIZE) {
821 /* This page can only merge a portion
822 * of tx request. Do not increment any
823 * pointer / counter here. The txp
824 * will be dealt with in future
825 * rounds, eventually hitting the
828 gop->len = PAGE_SIZE - dst_offset;
829 txp->offset += gop->len;
830 txp->size -= gop->len;
831 dst_offset += gop->len; /* quit loop */
833 /* This tx request can be merged in the page */
834 gop->len = txp->size;
835 dst_offset += gop->len;
837 index = pending_index(vif->pending_cons++);
839 pending_idx = vif->pending_ring[index];
841 memcpy(&pending_tx_info[pending_idx].req, txp,
844 /* Poison these fields, corresponding
845 * fields for head tx req will be set
846 * to correct values after the loop.
848 vif->mmap_pages[pending_idx] = (void *)(~0UL);
849 pending_tx_info[pending_idx].head =
850 INVALID_PENDING_RING_IDX;
853 first = &pending_tx_info[pending_idx];
855 head_idx = pending_idx;
865 first->req.offset = 0;
866 first->req.size = dst_offset;
867 first->head = start_idx;
868 vif->mmap_pages[head_idx] = page;
869 frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
872 BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
876 /* Unwind, freeing all pages and sending error responses. */
877 while (shinfo->nr_frags-- > start) {
878 xenvif_idx_release(vif,
879 frag_get_pending_idx(&frags[shinfo->nr_frags]),
880 XEN_NETIF_RSP_ERROR);
882 /* The head too, if necessary. */
884 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
889 static int xenvif_tx_check_gop(struct xenvif *vif,
891 struct gnttab_copy **gopp)
893 struct gnttab_copy *gop = *gopp;
894 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
895 struct skb_shared_info *shinfo = skb_shinfo(skb);
896 struct pending_tx_info *tx_info;
897 int nr_frags = shinfo->nr_frags;
899 u16 peek; /* peek into next tx request */
901 /* Check status of header. */
904 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
906 /* Skip first skb fragment if it is on same page as header fragment. */
907 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
909 for (i = start; i < nr_frags; i++) {
911 pending_ring_idx_t head;
913 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
914 tx_info = &vif->pending_tx_info[pending_idx];
915 head = tx_info->head;
917 /* Check error status: if okay then remember grant handle. */
919 newerr = (++gop)->status;
922 peek = vif->pending_ring[pending_index(++head)];
923 } while (!pending_tx_is_head(vif, peek));
925 if (likely(!newerr)) {
926 /* Had a previous error? Invalidate this fragment. */
928 xenvif_idx_release(vif, pending_idx,
933 /* Error on this fragment: respond to client with an error. */
934 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
936 /* Not the first error? Preceding frags already invalidated. */
940 /* First error: invalidate header and preceding fragments. */
941 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
942 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
943 for (j = start; j < i; j++) {
944 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
945 xenvif_idx_release(vif, pending_idx,
949 /* Remember the error: invalidate all subsequent fragments. */
957 static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
959 struct skb_shared_info *shinfo = skb_shinfo(skb);
960 int nr_frags = shinfo->nr_frags;
963 for (i = 0; i < nr_frags; i++) {
964 skb_frag_t *frag = shinfo->frags + i;
965 struct xen_netif_tx_request *txp;
969 pending_idx = frag_get_pending_idx(frag);
971 txp = &vif->pending_tx_info[pending_idx].req;
972 page = virt_to_page(idx_to_kaddr(vif, pending_idx));
973 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
974 skb->len += txp->size;
975 skb->data_len += txp->size;
976 skb->truesize += txp->size;
978 /* Take an extra reference to offset xenvif_idx_release */
979 get_page(vif->mmap_pages[pending_idx]);
980 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
984 static int xenvif_get_extras(struct xenvif *vif,
985 struct xen_netif_extra_info *extras,
988 struct xen_netif_extra_info extra;
989 RING_IDX cons = vif->tx.req_cons;
992 if (unlikely(work_to_do-- <= 0)) {
993 netdev_err(vif->dev, "Missing extra info\n");
994 xenvif_fatal_tx_err(vif);
998 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1000 if (unlikely(!extra.type ||
1001 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1002 vif->tx.req_cons = ++cons;
1003 netdev_err(vif->dev,
1004 "Invalid extra type: %d\n", extra.type);
1005 xenvif_fatal_tx_err(vif);
1009 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1010 vif->tx.req_cons = ++cons;
1011 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1016 static int xenvif_set_skb_gso(struct xenvif *vif,
1017 struct sk_buff *skb,
1018 struct xen_netif_extra_info *gso)
1020 if (!gso->u.gso.size) {
1021 netdev_err(vif->dev, "GSO size must not be zero.\n");
1022 xenvif_fatal_tx_err(vif);
1026 switch (gso->u.gso.type) {
1027 case XEN_NETIF_GSO_TYPE_TCPV4:
1028 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1030 case XEN_NETIF_GSO_TYPE_TCPV6:
1031 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1034 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1035 xenvif_fatal_tx_err(vif);
1039 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1040 /* gso_segs will be calculated later */
1045 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1047 bool recalculate_partial_csum = false;
1049 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1050 * peers can fail to set NETRXF_csum_blank when sending a GSO
1051 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1052 * recalculate the partial checksum.
1054 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1055 vif->rx_gso_checksum_fixup++;
1056 skb->ip_summed = CHECKSUM_PARTIAL;
1057 recalculate_partial_csum = true;
1060 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1061 if (skb->ip_summed != CHECKSUM_PARTIAL)
1064 return skb_checksum_setup(skb, recalculate_partial_csum);
1067 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1069 u64 now = get_jiffies_64();
1070 u64 next_credit = vif->credit_window_start +
1071 msecs_to_jiffies(vif->credit_usec / 1000);
1073 /* Timer could already be pending in rare cases. */
1074 if (timer_pending(&vif->credit_timeout))
1077 /* Passed the point where we can replenish credit? */
1078 if (time_after_eq64(now, next_credit)) {
1079 vif->credit_window_start = now;
1083 /* Still too big to send right now? Set a callback. */
1084 if (size > vif->remaining_credit) {
1085 vif->credit_timeout.data =
1087 vif->credit_timeout.function =
1089 mod_timer(&vif->credit_timeout,
1091 vif->credit_window_start = next_credit;
1099 static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget)
1101 struct gnttab_copy *gop = vif->tx_copy_ops, *request_gop;
1102 struct sk_buff *skb;
1105 while (xenvif_tx_pending_slots_available(vif) &&
1106 (skb_queue_len(&vif->tx_queue) < budget)) {
1107 struct xen_netif_tx_request txreq;
1108 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1110 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1114 unsigned int data_len;
1115 pending_ring_idx_t index;
1117 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1118 XEN_NETIF_TX_RING_SIZE) {
1119 netdev_err(vif->dev,
1120 "Impossible number of requests. "
1121 "req_prod %d, req_cons %d, size %ld\n",
1122 vif->tx.sring->req_prod, vif->tx.req_cons,
1123 XEN_NETIF_TX_RING_SIZE);
1124 xenvif_fatal_tx_err(vif);
1128 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);
1132 idx = vif->tx.req_cons;
1133 rmb(); /* Ensure that we see the request before we copy it. */
1134 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1136 /* Credit-based scheduling. */
1137 if (txreq.size > vif->remaining_credit &&
1138 tx_credit_exceeded(vif, txreq.size))
1141 vif->remaining_credit -= txreq.size;
1144 vif->tx.req_cons = ++idx;
1146 memset(extras, 0, sizeof(extras));
1147 if (txreq.flags & XEN_NETTXF_extra_info) {
1148 work_to_do = xenvif_get_extras(vif, extras,
1150 idx = vif->tx.req_cons;
1151 if (unlikely(work_to_do < 0))
1155 ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
1156 if (unlikely(ret < 0))
1161 if (unlikely(txreq.size < ETH_HLEN)) {
1162 netdev_dbg(vif->dev,
1163 "Bad packet size: %d\n", txreq.size);
1164 xenvif_tx_err(vif, &txreq, idx);
1168 /* No crossing a page as the payload mustn't fragment. */
1169 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1170 netdev_err(vif->dev,
1171 "txreq.offset: %x, size: %u, end: %lu\n",
1172 txreq.offset, txreq.size,
1173 (txreq.offset&~PAGE_MASK) + txreq.size);
1174 xenvif_fatal_tx_err(vif);
1178 index = pending_index(vif->pending_cons);
1179 pending_idx = vif->pending_ring[index];
1181 data_len = (txreq.size > PKT_PROT_LEN &&
1182 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1183 PKT_PROT_LEN : txreq.size;
1185 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1186 GFP_ATOMIC | __GFP_NOWARN);
1187 if (unlikely(skb == NULL)) {
1188 netdev_dbg(vif->dev,
1189 "Can't allocate a skb in start_xmit.\n");
1190 xenvif_tx_err(vif, &txreq, idx);
1194 /* Packets passed to netif_rx() must have some headroom. */
1195 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1197 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1198 struct xen_netif_extra_info *gso;
1199 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1201 if (xenvif_set_skb_gso(vif, skb, gso)) {
1202 /* Failure in xenvif_set_skb_gso is fatal. */
1208 /* XXX could copy straight to head */
1209 page = xenvif_alloc_page(vif, pending_idx);
1212 xenvif_tx_err(vif, &txreq, idx);
1216 gop->source.u.ref = txreq.gref;
1217 gop->source.domid = vif->domid;
1218 gop->source.offset = txreq.offset;
1220 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1221 gop->dest.domid = DOMID_SELF;
1222 gop->dest.offset = txreq.offset;
1224 gop->len = txreq.size;
1225 gop->flags = GNTCOPY_source_gref;
1229 memcpy(&vif->pending_tx_info[pending_idx].req,
1230 &txreq, sizeof(txreq));
1231 vif->pending_tx_info[pending_idx].head = index;
1232 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1234 __skb_put(skb, data_len);
1236 skb_shinfo(skb)->nr_frags = ret;
1237 if (data_len < txreq.size) {
1238 skb_shinfo(skb)->nr_frags++;
1239 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1242 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1243 INVALID_PENDING_IDX);
1246 vif->pending_cons++;
1248 request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
1249 if (request_gop == NULL) {
1251 xenvif_tx_err(vif, &txreq, idx);
1256 __skb_queue_tail(&vif->tx_queue, skb);
1258 vif->tx.req_cons = idx;
1260 if ((gop-vif->tx_copy_ops) >= ARRAY_SIZE(vif->tx_copy_ops))
1264 return gop - vif->tx_copy_ops;
1268 static int xenvif_tx_submit(struct xenvif *vif)
1270 struct gnttab_copy *gop = vif->tx_copy_ops;
1271 struct sk_buff *skb;
1274 while ((skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
1275 struct xen_netif_tx_request *txp;
1279 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1280 txp = &vif->pending_tx_info[pending_idx].req;
1282 /* Check the remap error code. */
1283 if (unlikely(xenvif_tx_check_gop(vif, skb, &gop))) {
1284 netdev_dbg(vif->dev, "netback grant failed.\n");
1285 skb_shinfo(skb)->nr_frags = 0;
1290 data_len = skb->len;
1292 (void *)(idx_to_kaddr(vif, pending_idx)|txp->offset),
1294 if (data_len < txp->size) {
1295 /* Append the packet payload as a fragment. */
1296 txp->offset += data_len;
1297 txp->size -= data_len;
1299 /* Schedule a response immediately. */
1300 xenvif_idx_release(vif, pending_idx,
1301 XEN_NETIF_RSP_OKAY);
1304 if (txp->flags & XEN_NETTXF_csum_blank)
1305 skb->ip_summed = CHECKSUM_PARTIAL;
1306 else if (txp->flags & XEN_NETTXF_data_validated)
1307 skb->ip_summed = CHECKSUM_UNNECESSARY;
1309 xenvif_fill_frags(vif, skb);
1311 if (skb_is_nonlinear(skb) && skb_headlen(skb) < PKT_PROT_LEN) {
1312 int target = min_t(int, skb->len, PKT_PROT_LEN);
1313 __pskb_pull_tail(skb, target - skb_headlen(skb));
1316 skb->dev = vif->dev;
1317 skb->protocol = eth_type_trans(skb, skb->dev);
1318 skb_reset_network_header(skb);
1320 if (checksum_setup(vif, skb)) {
1321 netdev_dbg(vif->dev,
1322 "Can't setup checksum in net_tx_action\n");
1327 skb_probe_transport_header(skb, 0);
1329 /* If the packet is GSO then we will have just set up the
1330 * transport header offset in checksum_setup so it's now
1331 * straightforward to calculate gso_segs.
1333 if (skb_is_gso(skb)) {
1334 int mss = skb_shinfo(skb)->gso_size;
1335 int hdrlen = skb_transport_header(skb) -
1336 skb_mac_header(skb) +
1339 skb_shinfo(skb)->gso_segs =
1340 DIV_ROUND_UP(skb->len - hdrlen, mss);
1343 vif->dev->stats.rx_bytes += skb->len;
1344 vif->dev->stats.rx_packets++;
1348 netif_receive_skb(skb);
1354 /* Called after netfront has transmitted */
1355 int xenvif_tx_action(struct xenvif *vif, int budget)
1360 if (unlikely(!tx_work_todo(vif)))
1363 nr_gops = xenvif_tx_build_gops(vif, budget);
1368 gnttab_batch_copy(vif->tx_copy_ops, nr_gops);
1370 work_done = xenvif_tx_submit(vif);
1375 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
1378 struct pending_tx_info *pending_tx_info;
1379 pending_ring_idx_t head;
1380 u16 peek; /* peek into next tx request */
1382 BUG_ON(vif->mmap_pages[pending_idx] == (void *)(~0UL));
1384 /* Already complete? */
1385 if (vif->mmap_pages[pending_idx] == NULL)
1388 pending_tx_info = &vif->pending_tx_info[pending_idx];
1390 head = pending_tx_info->head;
1392 BUG_ON(!pending_tx_is_head(vif, head));
1393 BUG_ON(vif->pending_ring[pending_index(head)] != pending_idx);
1396 pending_ring_idx_t index;
1397 pending_ring_idx_t idx = pending_index(head);
1398 u16 info_idx = vif->pending_ring[idx];
1400 pending_tx_info = &vif->pending_tx_info[info_idx];
1401 make_tx_response(vif, &pending_tx_info->req, status);
1403 /* Setting any number other than
1404 * INVALID_PENDING_RING_IDX indicates this slot is
1405 * starting a new packet / ending a previous packet.
1407 pending_tx_info->head = 0;
1409 index = pending_index(vif->pending_prod++);
1410 vif->pending_ring[index] = vif->pending_ring[info_idx];
1412 peek = vif->pending_ring[pending_index(++head)];
1414 } while (!pending_tx_is_head(vif, peek));
1416 put_page(vif->mmap_pages[pending_idx]);
1417 vif->mmap_pages[pending_idx] = NULL;
1421 static void make_tx_response(struct xenvif *vif,
1422 struct xen_netif_tx_request *txp,
1425 RING_IDX i = vif->tx.rsp_prod_pvt;
1426 struct xen_netif_tx_response *resp;
1429 resp = RING_GET_RESPONSE(&vif->tx, i);
1433 if (txp->flags & XEN_NETTXF_extra_info)
1434 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1436 vif->tx.rsp_prod_pvt = ++i;
1437 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1439 notify_remote_via_irq(vif->tx_irq);
1442 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1449 RING_IDX i = vif->rx.rsp_prod_pvt;
1450 struct xen_netif_rx_response *resp;
1452 resp = RING_GET_RESPONSE(&vif->rx, i);
1453 resp->offset = offset;
1454 resp->flags = flags;
1456 resp->status = (s16)size;
1458 resp->status = (s16)st;
1460 vif->rx.rsp_prod_pvt = ++i;
1465 static inline int rx_work_todo(struct xenvif *vif)
1467 return !skb_queue_empty(&vif->rx_queue) &&
1468 xenvif_rx_ring_slots_available(vif, vif->rx_last_skb_slots);
1471 static inline int tx_work_todo(struct xenvif *vif)
1474 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)) &&
1475 xenvif_tx_pending_slots_available(vif))
1481 void xenvif_unmap_frontend_rings(struct xenvif *vif)
1484 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1487 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1491 int xenvif_map_frontend_rings(struct xenvif *vif,
1492 grant_ref_t tx_ring_ref,
1493 grant_ref_t rx_ring_ref)
1496 struct xen_netif_tx_sring *txs;
1497 struct xen_netif_rx_sring *rxs;
1501 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1502 tx_ring_ref, &addr);
1506 txs = (struct xen_netif_tx_sring *)addr;
1507 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1509 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1510 rx_ring_ref, &addr);
1514 rxs = (struct xen_netif_rx_sring *)addr;
1515 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1520 xenvif_unmap_frontend_rings(vif);
1524 void xenvif_stop_queue(struct xenvif *vif)
1526 if (!vif->can_queue)
1529 netif_stop_queue(vif->dev);
1532 static void xenvif_start_queue(struct xenvif *vif)
1534 if (xenvif_schedulable(vif))
1535 netif_wake_queue(vif->dev);
1538 int xenvif_kthread_guest_rx(void *data)
1540 struct xenvif *vif = data;
1541 struct sk_buff *skb;
1543 while (!kthread_should_stop()) {
1544 wait_event_interruptible(vif->wq,
1545 rx_work_todo(vif) ||
1546 kthread_should_stop());
1547 if (kthread_should_stop())
1550 if (!skb_queue_empty(&vif->rx_queue))
1551 xenvif_rx_action(vif);
1553 if (skb_queue_empty(&vif->rx_queue) &&
1554 netif_queue_stopped(vif->dev))
1555 xenvif_start_queue(vif);
1560 /* Bin any remaining skbs */
1561 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL)
1567 static int __init netback_init(void)
1574 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1575 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1576 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1577 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1580 rc = xenvif_xenbus_init();
1590 module_init(netback_init);
1592 static void __exit netback_fini(void)
1594 xenvif_xenbus_fini();
1596 module_exit(netback_fini);
1598 MODULE_LICENSE("Dual BSD/GPL");
1599 MODULE_ALIAS("xen-backend:vif");