netback.c

   1/*
   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
   7 *
   8 * Copyright (c) 2002-2005, K A Fraser
   9 *
  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:
  15 *
  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:
  22 *
  23 * The above copyright notice and this permission notice shall be included in
  24 * all copies or substantial portions of the Software.
  25 *
  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
  32 * IN THE SOFTWARE.
  33 */
  34
  35#include "common.h"
  36
  37#include <linux/kthread.h>
  38#include <linux/if_vlan.h>
  39#include <linux/udp.h>
  40#include <linux/highmem.h>
  41#include <linux/skbuff_ref.h>
  42
  43#include <net/tcp.h>
  44
  45#include <xen/xen.h>
  46#include <xen/events.h>
  47#include <xen/interface/memory.h>
  48#include <xen/page.h>
  49
  50#include <asm/xen/hypercall.h>
  51
  52/* Provide an option to disable split event channels at load time as
  53 * event channels are limited resource. Split event channels are
  54 * enabled by default.
  55 */
  56bool separate_tx_rx_irq = true;
  57module_param(separate_tx_rx_irq, bool, 0644);
  58
  59/* The time that packets can stay on the guest Rx internal queue
  60 * before they are dropped.
  61 */
  62unsigned int rx_drain_timeout_msecs = 10000;
  63module_param(rx_drain_timeout_msecs, uint, 0444);
  64
  65/* The length of time before the frontend is considered unresponsive
  66 * because it isn't providing Rx slots.
  67 */
  68unsigned int rx_stall_timeout_msecs = 60000;
  69module_param(rx_stall_timeout_msecs, uint, 0444);
  70
  71#define MAX_QUEUES_DEFAULT 8
  72unsigned int xenvif_max_queues;
  73module_param_named(max_queues, xenvif_max_queues, uint, 0644);
  74MODULE_PARM_DESC(max_queues,
  75		 "Maximum number of queues per virtual interface");
  76
  77/*
  78 * This is the maximum slots a skb can have. If a guest sends a skb
  79 * which exceeds this limit it is considered malicious.
  80 */
  81#define FATAL_SKB_SLOTS_DEFAULT 20
  82static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
  83module_param(fatal_skb_slots, uint, 0444);
  84
  85/* The amount to copy out of the first guest Tx slot into the skb's
  86 * linear area.  If the first slot has more data, it will be mapped
  87 * and put into the first frag.
  88 *
  89 * This is sized to avoid pulling headers from the frags for most
  90 * TCP/IP packets.
  91 */
  92#define XEN_NETBACK_TX_COPY_LEN 128
  93
  94/* This is the maximum number of flows in the hash cache. */
  95#define XENVIF_HASH_CACHE_SIZE_DEFAULT 64
  96unsigned int xenvif_hash_cache_size = XENVIF_HASH_CACHE_SIZE_DEFAULT;
  97module_param_named(hash_cache_size, xenvif_hash_cache_size, uint, 0644);
  98MODULE_PARM_DESC(hash_cache_size, "Number of flows in the hash cache");
  99
 100/* The module parameter tells that we have to put data
 101 * for xen-netfront with the XDP_PACKET_HEADROOM offset
 102 * needed for XDP processing
 103 */
 104bool provides_xdp_headroom = true;
 105module_param(provides_xdp_headroom, bool, 0644);
 106
 107static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
 108			       s8 status);
 109
 110static void make_tx_response(struct xenvif_queue *queue,
 111			     const struct xen_netif_tx_request *txp,
 112			     unsigned int extra_count,
 113			     s8 status);
 114
 115static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx);
 116
 117static inline int tx_work_todo(struct xenvif_queue *queue);
 118
 119static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
 120				       u16 idx)
 121{
 122	return page_to_pfn(queue->mmap_pages[idx]);
 123}
 124
 125static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
 126					 u16 idx)
 127{
 128	return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
 129}
 130
 131#define callback_param(vif, pending_idx) \
 132	(vif->pending_tx_info[pending_idx].callback_struct)
 133
 134/* Find the containing VIF's structure from a pointer in pending_tx_info array
 135 */
 136static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info_msgzc *ubuf)
 137{
 138	u16 pending_idx = ubuf->desc;
 139	struct pending_tx_info *temp =
 140		container_of(ubuf, struct pending_tx_info, callback_struct);
 141	return container_of(temp - pending_idx,
 142			    struct xenvif_queue,
 143			    pending_tx_info[0]);
 144}
 145
 146static u16 frag_get_pending_idx(skb_frag_t *frag)
 147{
 148	return (u16)skb_frag_off(frag);
 149}
 150
 151static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
 152{
 153	skb_frag_off_set(frag, pending_idx);
 154}
 155
 156static inline pending_ring_idx_t pending_index(unsigned i)
 157{
 158	return i & (MAX_PENDING_REQS-1);
 159}
 160
 161void xenvif_kick_thread(struct xenvif_queue *queue)
 162{
 163	wake_up(&queue->wq);
 164}
 165
 166void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
 167{
 168	int more_to_do;
 169
 170	RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
 171
 172	if (more_to_do)
 173		napi_schedule(&queue->napi);
 174	else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
 175				     &queue->eoi_pending) &
 176		 (NETBK_TX_EOI | NETBK_COMMON_EOI))
 177		xen_irq_lateeoi(queue->tx_irq, 0);
 178}
 179
 180static void tx_add_credit(struct xenvif_queue *queue)
 181{
 182	unsigned long max_burst, max_credit;
 183
 184	/*
 185	 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
 186	 * Otherwise the interface can seize up due to insufficient credit.
 187	 */
 188	max_burst = max(131072UL, queue->credit_bytes);
 189
 190	/* Take care that adding a new chunk of credit doesn't wrap to zero. */
 191	max_credit = queue->remaining_credit + queue->credit_bytes;
 192	if (max_credit < queue->remaining_credit)
 193		max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
 194
 195	queue->remaining_credit = min(max_credit, max_burst);
 196	queue->rate_limited = false;
 197}
 198
 199void xenvif_tx_credit_callback(struct timer_list *t)
 200{
 201	struct xenvif_queue *queue = timer_container_of(queue, t,
 202							credit_timeout);
 203	tx_add_credit(queue);
 204	xenvif_napi_schedule_or_enable_events(queue);
 205}
 206
 207static void xenvif_tx_err(struct xenvif_queue *queue,
 208			  struct xen_netif_tx_request *txp,
 209			  unsigned int extra_count, RING_IDX end)
 210{
 211	RING_IDX cons = queue->tx.req_cons;
 212
 213	do {
 214		make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
 215		if (cons == end)
 216			break;
 217		RING_COPY_REQUEST(&queue->tx, cons++, txp);
 218		extra_count = 0; /* only the first frag can have extras */
 219	} while (1);
 220	queue->tx.req_cons = cons;
 221}
 222
 223static void xenvif_fatal_tx_err(struct xenvif *vif)
 224{
 225	netdev_err(vif->dev, "fatal error; disabling device\n");
 226	vif->disabled = true;
 227	/* Disable the vif from queue 0's kthread */
 228	if (vif->num_queues)
 229		xenvif_kick_thread(&vif->queues[0]);
 230}
 231
 232static int xenvif_count_requests(struct xenvif_queue *queue,
 233				 struct xen_netif_tx_request *first,
 234				 unsigned int extra_count,
 235				 struct xen_netif_tx_request *txp,
 236				 int work_to_do)
 237{
 238	RING_IDX cons = queue->tx.req_cons;
 239	int slots = 0;
 240	int drop_err = 0;
 241	int more_data;
 242
 243	if (!(first->flags & XEN_NETTXF_more_data))
 244		return 0;
 245
 246	do {
 247		struct xen_netif_tx_request dropped_tx = { 0 };
 248
 249		if (slots >= work_to_do) {
 250			netdev_err(queue->vif->dev,
 251				   "Asked for %d slots but exceeds this limit\n",
 252				   work_to_do);
 253			xenvif_fatal_tx_err(queue->vif);
 254			return -ENODATA;
 255		}
 256
 257		/* This guest is really using too many slots and
 258		 * considered malicious.
 259		 */
 260		if (unlikely(slots >= fatal_skb_slots)) {
 261			netdev_err(queue->vif->dev,
 262				   "Malicious frontend using %d slots, threshold %u\n",
 263				   slots, fatal_skb_slots);
 264			xenvif_fatal_tx_err(queue->vif);
 265			return -E2BIG;
 266		}
 267
 268		/* Xen network protocol had implicit dependency on
 269		 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
 270		 * the historical MAX_SKB_FRAGS value 18 to honor the
 271		 * same behavior as before. Any packet using more than
 272		 * 18 slots but less than fatal_skb_slots slots is
 273		 * dropped
 274		 */
 275		if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
 276			if (net_ratelimit())
 277				netdev_dbg(queue->vif->dev,
 278					   "Too many slots (%d) exceeding limit (%d), dropping packet\n",
 279					   slots, XEN_NETBK_LEGACY_SLOTS_MAX);
 280			drop_err = -E2BIG;
 281		}
 282
 283		if (drop_err)
 284			txp = &dropped_tx;
 285
 286		RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
 287
 288		/* If the guest submitted a frame >= 64 KiB then
 289		 * first->size overflowed and following slots will
 290		 * appear to be larger than the frame.
 291		 *
 292		 * This cannot be fatal error as there are buggy
 293		 * frontends that do this.
 294		 *
 295		 * Consume all slots and drop the packet.
 296		 */
 297		if (!drop_err && txp->size > first->size) {
 298			if (net_ratelimit())
 299				netdev_dbg(queue->vif->dev,
 300					   "Invalid tx request, slot size %u > remaining size %u\n",
 301					   txp->size, first->size);
 302			drop_err = -EIO;
 303		}
 304
 305		first->size -= txp->size;
 306		slots++;
 307
 308		if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
 309			netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
 310				 txp->offset, txp->size);
 311			xenvif_fatal_tx_err(queue->vif);
 312			return -EINVAL;
 313		}
 314
 315		more_data = txp->flags & XEN_NETTXF_more_data;
 316
 317		if (!drop_err)
 318			txp++;
 319
 320	} while (more_data);
 321
 322	if (drop_err) {
 323		xenvif_tx_err(queue, first, extra_count, cons + slots);
 324		return drop_err;
 325	}
 326
 327	return slots;
 328}
 329
 330
 331struct xenvif_tx_cb {
 332	u16 copy_pending_idx[XEN_NETBK_LEGACY_SLOTS_MAX + 1];
 333	u8 copy_count;
 334	u32 split_mask;
 335};
 336
 337#define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
 338#define copy_pending_idx(skb, i) (XENVIF_TX_CB(skb)->copy_pending_idx[i])
 339#define copy_count(skb) (XENVIF_TX_CB(skb)->copy_count)
 340
 341static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
 342					   u16 pending_idx,
 343					   struct xen_netif_tx_request *txp,
 344					   unsigned int extra_count,
 345					   struct gnttab_map_grant_ref *mop)
 346{
 347	queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
 348	gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
 349			  GNTMAP_host_map | GNTMAP_readonly,
 350			  txp->gref, queue->vif->domid);
 351
 352	memcpy(&queue->pending_tx_info[pending_idx].req, txp,
 353	       sizeof(*txp));
 354	queue->pending_tx_info[pending_idx].extra_count = extra_count;
 355}
 356
 357static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
 358{
 359	struct sk_buff *skb =
 360		alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
 361			  GFP_ATOMIC | __GFP_NOWARN);
 362
 363	BUILD_BUG_ON(sizeof(*XENVIF_TX_CB(skb)) > sizeof(skb->cb));
 364	if (unlikely(skb == NULL))
 365		return NULL;
 366
 367	/* Packets passed to netif_rx() must have some headroom. */
 368	skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
 369
 370	/* Initialize it here to avoid later surprises */
 371	skb_shinfo(skb)->destructor_arg = NULL;
 372
 373	return skb;
 374}
 375
 376static void xenvif_get_requests(struct xenvif_queue *queue,
 377				struct sk_buff *skb,
 378				struct xen_netif_tx_request *first,
 379				struct xen_netif_tx_request *txfrags,
 380			        unsigned *copy_ops,
 381			        unsigned *map_ops,
 382				unsigned int frag_overflow,
 383				struct sk_buff *nskb,
 384				unsigned int extra_count,
 385				unsigned int data_len)
 386{
 387	struct skb_shared_info *shinfo = skb_shinfo(skb);
 388	skb_frag_t *frags = shinfo->frags;
 389	u16 pending_idx;
 390	pending_ring_idx_t index;
 391	unsigned int nr_slots;
 392	struct gnttab_copy *cop = queue->tx_copy_ops + *copy_ops;
 393	struct gnttab_map_grant_ref *gop = queue->tx_map_ops + *map_ops;
 394	struct xen_netif_tx_request *txp = first;
 395
 396	nr_slots = shinfo->nr_frags + frag_overflow + 1;
 397
 398	copy_count(skb) = 0;
 399	XENVIF_TX_CB(skb)->split_mask = 0;
 400
 401	/* Create copy ops for exactly data_len bytes into the skb head. */
 402	__skb_put(skb, data_len);
 403	while (data_len > 0) {
 404		int amount = data_len > txp->size ? txp->size : data_len;
 405		bool split = false;
 406
 407		cop->source.u.ref = txp->gref;
 408		cop->source.domid = queue->vif->domid;
 409		cop->source.offset = txp->offset;
 410
 411		cop->dest.domid = DOMID_SELF;
 412		cop->dest.offset = (offset_in_page(skb->data +
 413						   skb_headlen(skb) -
 414						   data_len)) & ~XEN_PAGE_MASK;
 415		cop->dest.u.gmfn = virt_to_gfn(skb->data + skb_headlen(skb)
 416				               - data_len);
 417
 418		/* Don't cross local page boundary! */
 419		if (cop->dest.offset + amount > XEN_PAGE_SIZE) {
 420			amount = XEN_PAGE_SIZE - cop->dest.offset;
 421			XENVIF_TX_CB(skb)->split_mask |= 1U << copy_count(skb);
 422			split = true;
 423		}
 424
 425		cop->len = amount;
 426		cop->flags = GNTCOPY_source_gref;
 427
 428		index = pending_index(queue->pending_cons);
 429		pending_idx = queue->pending_ring[index];
 430		callback_param(queue, pending_idx).ctx = NULL;
 431		copy_pending_idx(skb, copy_count(skb)) = pending_idx;
 432		if (!split)
 433			copy_count(skb)++;
 434
 435		cop++;
 436		data_len -= amount;
 437
 438		if (amount == txp->size) {
 439			/* The copy op covered the full tx_request */
 440
 441			memcpy(&queue->pending_tx_info[pending_idx].req,
 442			       txp, sizeof(*txp));
 443			queue->pending_tx_info[pending_idx].extra_count =
 444				(txp == first) ? extra_count : 0;
 445
 446			if (txp == first)
 447				txp = txfrags;
 448			else
 449				txp++;
 450			queue->pending_cons++;
 451			nr_slots--;
 452		} else {
 453			/* The copy op partially covered the tx_request.
 454			 * The remainder will be mapped or copied in the next
 455			 * iteration.
 456			 */
 457			txp->offset += amount;
 458			txp->size -= amount;
 459		}
 460	}
 461
 462	for (shinfo->nr_frags = 0; nr_slots > 0 && shinfo->nr_frags < MAX_SKB_FRAGS;
 463	     nr_slots--) {
 464		if (unlikely(!txp->size)) {
 465			make_tx_response(queue, txp, 0, XEN_NETIF_RSP_OKAY);
 466			++txp;
 467			continue;
 468		}
 469
 470		index = pending_index(queue->pending_cons++);
 471		pending_idx = queue->pending_ring[index];
 472		xenvif_tx_create_map_op(queue, pending_idx, txp,
 473				        txp == first ? extra_count : 0, gop);
 474		frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
 475		++shinfo->nr_frags;
 476		++gop;
 477
 478		if (txp == first)
 479			txp = txfrags;
 480		else
 481			txp++;
 482	}
 483
 484	if (nr_slots > 0) {
 485
 486		shinfo = skb_shinfo(nskb);
 487		frags = shinfo->frags;
 488
 489		for (shinfo->nr_frags = 0; shinfo->nr_frags < nr_slots; ++txp) {
 490			if (unlikely(!txp->size)) {
 491				make_tx_response(queue, txp, 0,
 492						 XEN_NETIF_RSP_OKAY);
 493				continue;
 494			}
 495
 496			index = pending_index(queue->pending_cons++);
 497			pending_idx = queue->pending_ring[index];
 498			xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
 499						gop);
 500			frag_set_pending_idx(&frags[shinfo->nr_frags],
 501					     pending_idx);
 502			++shinfo->nr_frags;
 503			++gop;
 504		}
 505
 506		if (shinfo->nr_frags) {
 507			skb_shinfo(skb)->frag_list = nskb;
 508			nskb = NULL;
 509		}
 510	}
 511
 512	if (nskb) {
 513		/* A frag_list skb was allocated but it is no longer needed
 514		 * because enough slots were converted to copy ops above or some
 515		 * were empty.
 516		 */
 517		kfree_skb(nskb);
 518	}
 519
 520	(*copy_ops) = cop - queue->tx_copy_ops;
 521	(*map_ops) = gop - queue->tx_map_ops;
 522}
 523
 524static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
 525					   u16 pending_idx,
 526					   grant_handle_t handle)
 527{
 528	if (unlikely(queue->grant_tx_handle[pending_idx] !=
 529		     NETBACK_INVALID_HANDLE)) {
 530		netdev_err(queue->vif->dev,
 531			   "Trying to overwrite active handle! pending_idx: 0x%x\n",
 532			   pending_idx);
 533		BUG();
 534	}
 535	queue->grant_tx_handle[pending_idx] = handle;
 536}
 537
 538static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
 539					     u16 pending_idx)
 540{
 541	if (unlikely(queue->grant_tx_handle[pending_idx] ==
 542		     NETBACK_INVALID_HANDLE)) {
 543		netdev_err(queue->vif->dev,
 544			   "Trying to unmap invalid handle! pending_idx: 0x%x\n",
 545			   pending_idx);
 546		BUG();
 547	}
 548	queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
 549}
 550
 551static int xenvif_tx_check_gop(struct xenvif_queue *queue,
 552			       struct sk_buff *skb,
 553			       struct gnttab_map_grant_ref **gopp_map,
 554			       struct gnttab_copy **gopp_copy)
 555{
 556	struct gnttab_map_grant_ref *gop_map = *gopp_map;
 557	u16 pending_idx;
 558	/* This always points to the shinfo of the skb being checked, which
 559	 * could be either the first or the one on the frag_list
 560	 */
 561	struct skb_shared_info *shinfo = skb_shinfo(skb);
 562	/* If this is non-NULL, we are currently checking the frag_list skb, and
 563	 * this points to the shinfo of the first one
 564	 */
 565	struct skb_shared_info *first_shinfo = NULL;
 566	int nr_frags = shinfo->nr_frags;
 567	const bool sharedslot = nr_frags &&
 568				frag_get_pending_idx(&shinfo->frags[0]) ==
 569				    copy_pending_idx(skb, copy_count(skb) - 1);
 570	int i, err = 0;
 571
 572	for (i = 0; i < copy_count(skb); i++) {
 573		int newerr;
 574
 575		/* Check status of header. */
 576		pending_idx = copy_pending_idx(skb, i);
 577
 578		newerr = (*gopp_copy)->status;
 579
 580		/* Split copies need to be handled together. */
 581		if (XENVIF_TX_CB(skb)->split_mask & (1U << i)) {
 582			(*gopp_copy)++;
 583			if (!newerr)
 584				newerr = (*gopp_copy)->status;
 585		}
 586		if (likely(!newerr)) {
 587			/* The first frag might still have this slot mapped */
 588			if (i < copy_count(skb) - 1 || !sharedslot)
 589				xenvif_idx_release(queue, pending_idx,
 590						   XEN_NETIF_RSP_OKAY);
 591		} else {
 592			err = newerr;
 593			if (net_ratelimit())
 594				netdev_dbg(queue->vif->dev,
 595					   "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
 596					   (*gopp_copy)->status,
 597					   pending_idx,
 598					   (*gopp_copy)->source.u.ref);
 599			/* The first frag might still have this slot mapped */
 600			if (i < copy_count(skb) - 1 || !sharedslot)
 601				xenvif_idx_release(queue, pending_idx,
 602						   XEN_NETIF_RSP_ERROR);
 603		}
 604		(*gopp_copy)++;
 605	}
 606
 607check_frags:
 608	for (i = 0; i < nr_frags; i++, gop_map++) {
 609		int j, newerr;
 610
 611		pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
 612
 613		/* Check error status: if okay then remember grant handle. */
 614		newerr = gop_map->status;
 615
 616		if (likely(!newerr)) {
 617			xenvif_grant_handle_set(queue,
 618						pending_idx,
 619						gop_map->handle);
 620			/* Had a previous error? Invalidate this fragment. */
 621			if (unlikely(err)) {
 622				xenvif_idx_unmap(queue, pending_idx);
 623				/* If the mapping of the first frag was OK, but
 624				 * the header's copy failed, and they are
 625				 * sharing a slot, send an error
 626				 */
 627				if (i == 0 && !first_shinfo && sharedslot)
 628					xenvif_idx_release(queue, pending_idx,
 629							   XEN_NETIF_RSP_ERROR);
 630				else
 631					xenvif_idx_release(queue, pending_idx,
 632							   XEN_NETIF_RSP_OKAY);
 633			}
 634			continue;
 635		}
 636
 637		/* Error on this fragment: respond to client with an error. */
 638		if (net_ratelimit())
 639			netdev_dbg(queue->vif->dev,
 640				   "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
 641				   i,
 642				   gop_map->status,
 643				   pending_idx,
 644				   gop_map->ref);
 645
 646		xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
 647
 648		/* Not the first error? Preceding frags already invalidated. */
 649		if (err)
 650			continue;
 651
 652		/* Invalidate preceding fragments of this skb. */
 653		for (j = 0; j < i; j++) {
 654			pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
 655			xenvif_idx_unmap(queue, pending_idx);
 656			xenvif_idx_release(queue, pending_idx,
 657					   XEN_NETIF_RSP_OKAY);
 658		}
 659
 660		/* And if we found the error while checking the frag_list, unmap
 661		 * the first skb's frags
 662		 */
 663		if (first_shinfo) {
 664			for (j = 0; j < first_shinfo->nr_frags; j++) {
 665				pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
 666				xenvif_idx_unmap(queue, pending_idx);
 667				xenvif_idx_release(queue, pending_idx,
 668						   XEN_NETIF_RSP_OKAY);
 669			}
 670		}
 671
 672		/* Remember the error: invalidate all subsequent fragments. */
 673		err = newerr;
 674	}
 675
 676	if (skb_has_frag_list(skb) && !first_shinfo) {
 677		first_shinfo = shinfo;
 678		shinfo = skb_shinfo(shinfo->frag_list);
 679		nr_frags = shinfo->nr_frags;
 680
 681		goto check_frags;
 682	}
 683
 684	*gopp_map = gop_map;
 685	return err;
 686}
 687
 688static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
 689{
 690	struct skb_shared_info *shinfo = skb_shinfo(skb);
 691	int nr_frags = shinfo->nr_frags;
 692	int i;
 693	u16 prev_pending_idx = INVALID_PENDING_IDX;
 694
 695	for (i = 0; i < nr_frags; i++) {
 696		skb_frag_t *frag = shinfo->frags + i;
 697		struct xen_netif_tx_request *txp;
 698		struct page *page;
 699		u16 pending_idx;
 700
 701		pending_idx = frag_get_pending_idx(frag);
 702
 703		/* If this is not the first frag, chain it to the previous*/
 704		if (prev_pending_idx == INVALID_PENDING_IDX)
 705			skb_shinfo(skb)->destructor_arg =
 706				&callback_param(queue, pending_idx);
 707		else
 708			callback_param(queue, prev_pending_idx).ctx =
 709				&callback_param(queue, pending_idx);
 710
 711		callback_param(queue, pending_idx).ctx = NULL;
 712		prev_pending_idx = pending_idx;
 713
 714		txp = &queue->pending_tx_info[pending_idx].req;
 715		page = virt_to_page((void *)idx_to_kaddr(queue, pending_idx));
 716		__skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
 717		skb->len += txp->size;
 718		skb->data_len += txp->size;
 719		skb->truesize += txp->size;
 720
 721		/* Take an extra reference to offset network stack's put_page */
 722		get_page(queue->mmap_pages[pending_idx]);
 723	}
 724}
 725
 726static int xenvif_get_extras(struct xenvif_queue *queue,
 727			     struct xen_netif_extra_info *extras,
 728			     unsigned int *extra_count,
 729			     int work_to_do)
 730{
 731	struct xen_netif_extra_info extra;
 732	RING_IDX cons = queue->tx.req_cons;
 733
 734	do {
 735		if (unlikely(work_to_do-- <= 0)) {
 736			netdev_err(queue->vif->dev, "Missing extra info\n");
 737			xenvif_fatal_tx_err(queue->vif);
 738			return -EBADR;
 739		}
 740
 741		RING_COPY_REQUEST(&queue->tx, cons, &extra);
 742
 743		queue->tx.req_cons = ++cons;
 744		(*extra_count)++;
 745
 746		if (unlikely(!extra.type ||
 747			     extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
 748			netdev_err(queue->vif->dev,
 749				   "Invalid extra type: %d\n", extra.type);
 750			xenvif_fatal_tx_err(queue->vif);
 751			return -EINVAL;
 752		}
 753
 754		memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
 755	} while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
 756
 757	return work_to_do;
 758}
 759
 760static int xenvif_set_skb_gso(struct xenvif *vif,
 761			      struct sk_buff *skb,
 762			      struct xen_netif_extra_info *gso)
 763{
 764	if (!gso->u.gso.size) {
 765		netdev_err(vif->dev, "GSO size must not be zero.\n");
 766		xenvif_fatal_tx_err(vif);
 767		return -EINVAL;
 768	}
 769
 770	switch (gso->u.gso.type) {
 771	case XEN_NETIF_GSO_TYPE_TCPV4:
 772		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
 773		break;
 774	case XEN_NETIF_GSO_TYPE_TCPV6:
 775		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
 776		break;
 777	default:
 778		netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
 779		xenvif_fatal_tx_err(vif);
 780		return -EINVAL;
 781	}
 782
 783	skb_shinfo(skb)->gso_size = gso->u.gso.size;
 784	/* gso_segs will be calculated later */
 785
 786	return 0;
 787}
 788
 789static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
 790{
 791	bool recalculate_partial_csum = false;
 792
 793	/* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
 794	 * peers can fail to set NETRXF_csum_blank when sending a GSO
 795	 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
 796	 * recalculate the partial checksum.
 797	 */
 798	if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
 799		queue->stats.rx_gso_checksum_fixup++;
 800		skb->ip_summed = CHECKSUM_PARTIAL;
 801		recalculate_partial_csum = true;
 802	}
 803
 804	/* A non-CHECKSUM_PARTIAL SKB does not require setup. */
 805	if (skb->ip_summed != CHECKSUM_PARTIAL)
 806		return 0;
 807
 808	return skb_checksum_setup(skb, recalculate_partial_csum);
 809}
 810
 811static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
 812{
 813	u64 now = get_jiffies_64();
 814	u64 next_credit = queue->credit_window_start +
 815		msecs_to_jiffies(queue->credit_usec / 1000);
 816
 817	/* Timer could already be pending in rare cases. */
 818	if (timer_pending(&queue->credit_timeout)) {
 819		queue->rate_limited = true;
 820		return true;
 821	}
 822
 823	/* Passed the point where we can replenish credit? */
 824	if (time_after_eq64(now, next_credit)) {
 825		queue->credit_window_start = now;
 826		tx_add_credit(queue);
 827	}
 828
 829	/* Still too big to send right now? Set a callback. */
 830	if (size > queue->remaining_credit) {
 831		mod_timer(&queue->credit_timeout,
 832			  next_credit);
 833		queue->credit_window_start = next_credit;
 834		queue->rate_limited = true;
 835
 836		return true;
 837	}
 838
 839	return false;
 840}
 841
 842/* No locking is required in xenvif_mcast_add/del() as they are
 843 * only ever invoked from NAPI poll. An RCU list is used because
 844 * xenvif_mcast_match() is called asynchronously, during start_xmit.
 845 */
 846
 847static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
 848{
 849	struct xenvif_mcast_addr *mcast;
 850
 851	if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
 852		if (net_ratelimit())
 853			netdev_err(vif->dev,
 854				   "Too many multicast addresses\n");
 855		return -ENOSPC;
 856	}
 857
 858	mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
 859	if (!mcast)
 860		return -ENOMEM;
 861
 862	ether_addr_copy(mcast->addr, addr);
 863	list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
 864	vif->fe_mcast_count++;
 865
 866	return 0;
 867}
 868
 869static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
 870{
 871	struct xenvif_mcast_addr *mcast;
 872
 873	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 874		if (ether_addr_equal(addr, mcast->addr)) {
 875			--vif->fe_mcast_count;
 876			list_del_rcu(&mcast->entry);
 877			kfree_rcu(mcast, rcu);
 878			break;
 879		}
 880	}
 881}
 882
 883bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
 884{
 885	struct xenvif_mcast_addr *mcast;
 886
 887	rcu_read_lock();
 888	list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
 889		if (ether_addr_equal(addr, mcast->addr)) {
 890			rcu_read_unlock();
 891			return true;
 892		}
 893	}
 894	rcu_read_unlock();
 895
 896	return false;
 897}
 898
 899void xenvif_mcast_addr_list_free(struct xenvif *vif)
 900{
 901	/* No need for locking or RCU here. NAPI poll and TX queue
 902	 * are stopped.
 903	 */
 904	while (!list_empty(&vif->fe_mcast_addr)) {
 905		struct xenvif_mcast_addr *mcast;
 906
 907		mcast = list_first_entry(&vif->fe_mcast_addr,
 908					 struct xenvif_mcast_addr,
 909					 entry);
 910		--vif->fe_mcast_count;
 911		list_del(&mcast->entry);
 912		kfree(mcast);
 913	}
 914}
 915
 916static void xenvif_tx_build_gops(struct xenvif_queue *queue,
 917				     int budget,
 918				     unsigned *copy_ops,
 919				     unsigned *map_ops)
 920{
 921	struct sk_buff *skb, *nskb;
 922	int ret;
 923	unsigned int frag_overflow;
 924
 925	while (skb_queue_len(&queue->tx_queue) < budget) {
 926		struct xen_netif_tx_request txreq;
 927		struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
 928		struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
 929		unsigned int extra_count;
 930		RING_IDX idx;
 931		int work_to_do;
 932		unsigned int data_len;
 933
 934		if (queue->tx.sring->req_prod - queue->tx.req_cons >
 935		    XEN_NETIF_TX_RING_SIZE) {
 936			netdev_err(queue->vif->dev,
 937				   "Impossible number of requests. "
 938				   "req_prod %d, req_cons %d, size %ld\n",
 939				   queue->tx.sring->req_prod, queue->tx.req_cons,
 940				   XEN_NETIF_TX_RING_SIZE);
 941			xenvif_fatal_tx_err(queue->vif);
 942			break;
 943		}
 944
 945		work_to_do = XEN_RING_NR_UNCONSUMED_REQUESTS(&queue->tx);
 946		if (!work_to_do)
 947			break;
 948
 949		idx = queue->tx.req_cons;
 950		rmb(); /* Ensure that we see the request before we copy it. */
 951		RING_COPY_REQUEST(&queue->tx, idx, &txreq);
 952
 953		/* Credit-based scheduling. */
 954		if (txreq.size > queue->remaining_credit &&
 955		    tx_credit_exceeded(queue, txreq.size))
 956			break;
 957
 958		queue->remaining_credit -= txreq.size;
 959
 960		work_to_do--;
 961		queue->tx.req_cons = ++idx;
 962
 963		memset(extras, 0, sizeof(extras));
 964		extra_count = 0;
 965		if (txreq.flags & XEN_NETTXF_extra_info) {
 966			work_to_do = xenvif_get_extras(queue, extras,
 967						       &extra_count,
 968						       work_to_do);
 969			idx = queue->tx.req_cons;
 970			if (unlikely(work_to_do < 0))
 971				break;
 972		}
 973
 974		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
 975			struct xen_netif_extra_info *extra;
 976
 977			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
 978			ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
 979
 980			make_tx_response(queue, &txreq, extra_count,
 981					 (ret == 0) ?
 982					 XEN_NETIF_RSP_OKAY :
 983					 XEN_NETIF_RSP_ERROR);
 984			continue;
 985		}
 986
 987		if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
 988			struct xen_netif_extra_info *extra;
 989
 990			extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
 991			xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
 992
 993			make_tx_response(queue, &txreq, extra_count,
 994					 XEN_NETIF_RSP_OKAY);
 995			continue;
 996		}
 997
 998		data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN) ?
 999			XEN_NETBACK_TX_COPY_LEN : txreq.size;
1000
1001		ret = xenvif_count_requests(queue, &txreq, extra_count,
1002					    txfrags, work_to_do);
1003
1004		if (unlikely(ret < 0))
1005			break;
1006
1007		idx += ret;
1008
1009		if (unlikely(txreq.size < ETH_HLEN)) {
1010			netdev_dbg(queue->vif->dev,
1011				   "Bad packet size: %d\n", txreq.size);
1012			xenvif_tx_err(queue, &txreq, extra_count, idx);
1013			break;
1014		}
1015
1016		/* No crossing a page as the payload mustn't fragment. */
1017		if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1018			netdev_err(queue->vif->dev, "Cross page boundary, txreq.offset: %u, size: %u\n",
1019				   txreq.offset, txreq.size);
1020			xenvif_fatal_tx_err(queue->vif);
1021			break;
1022		}
1023
1024		if (ret >= XEN_NETBK_LEGACY_SLOTS_MAX - 1 && data_len < txreq.size)
1025			data_len = txreq.size;
1026
1027		skb = xenvif_alloc_skb(data_len);
1028		if (unlikely(skb == NULL)) {
1029			netdev_dbg(queue->vif->dev,
1030				   "Can't allocate a skb in start_xmit.\n");
1031			xenvif_tx_err(queue, &txreq, extra_count, idx);
1032			break;
1033		}
1034
1035		skb_shinfo(skb)->nr_frags = ret;
1036		/* At this point shinfo->nr_frags is in fact the number of
1037		 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1038		 */
1039		frag_overflow = 0;
1040		nskb = NULL;
1041		if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1042			frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1043			BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1044			skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1045			nskb = xenvif_alloc_skb(0);
1046			if (unlikely(nskb == NULL)) {
1047				skb_shinfo(skb)->nr_frags = 0;
1048				kfree_skb(skb);
1049				xenvif_tx_err(queue, &txreq, extra_count, idx);
1050				if (net_ratelimit())
1051					netdev_err(queue->vif->dev,
1052						   "Can't allocate the frag_list skb.\n");
1053				break;
1054			}
1055		}
1056
1057		if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1058			struct xen_netif_extra_info *gso;
1059			gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1060
1061			if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1062				/* Failure in xenvif_set_skb_gso is fatal. */
1063				skb_shinfo(skb)->nr_frags = 0;
1064				kfree_skb(skb);
1065				kfree_skb(nskb);
1066				break;
1067			}
1068		}
1069
1070		if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1071			struct xen_netif_extra_info *extra;
1072			enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1073
1074			extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1075
1076			switch (extra->u.hash.type) {
1077			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1078			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1079				type = PKT_HASH_TYPE_L3;
1080				break;
1081
1082			case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1083			case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1084				type = PKT_HASH_TYPE_L4;
1085				break;
1086
1087			default:
1088				break;
1089			}
1090
1091			if (type != PKT_HASH_TYPE_NONE)
1092				skb_set_hash(skb,
1093					     *(u32 *)extra->u.hash.value,
1094					     type);
1095		}
1096
1097		xenvif_get_requests(queue, skb, &txreq, txfrags, copy_ops,
1098				    map_ops, frag_overflow, nskb, extra_count,
1099				    data_len);
1100
1101		__skb_queue_tail(&queue->tx_queue, skb);
1102
1103		queue->tx.req_cons = idx;
1104	}
1105
1106	return;
1107}
1108
1109/* Consolidate skb with a frag_list into a brand new one with local pages on
1110 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1111 */
1112static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1113{
1114	unsigned int offset = skb_headlen(skb);
1115	skb_frag_t frags[MAX_SKB_FRAGS];
1116	int i, f;
1117	struct ubuf_info *uarg;
1118	struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1119
1120	queue->stats.tx_zerocopy_sent += 2;
1121	queue->stats.tx_frag_overflow++;
1122
1123	xenvif_fill_frags(queue, nskb);
1124	/* Subtract frags size, we will correct it later */
1125	skb->truesize -= skb->data_len;
1126	skb->len += nskb->len;
1127	skb->data_len += nskb->len;
1128
1129	/* create a brand new frags array and coalesce there */
1130	for (i = 0; offset < skb->len; i++) {
1131		struct page *page;
1132		unsigned int len;
1133
1134		BUG_ON(i >= MAX_SKB_FRAGS);
1135		page = alloc_page(GFP_ATOMIC);
1136		if (!page) {
1137			int j;
1138			skb->truesize += skb->data_len;
1139			for (j = 0; j < i; j++)
1140				put_page(skb_frag_page(&frags[j]));
1141			return -ENOMEM;
1142		}
1143
1144		if (offset + PAGE_SIZE < skb->len)
1145			len = PAGE_SIZE;
1146		else
1147			len = skb->len - offset;
1148		if (skb_copy_bits(skb, offset, page_address(page), len))
1149			BUG();
1150
1151		offset += len;
1152		skb_frag_fill_page_desc(&frags[i], page, 0, len);
1153	}
1154
1155	/* Release all the original (foreign) frags. */
1156	for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1157		skb_frag_unref(skb, f);
1158	uarg = skb_shinfo(skb)->destructor_arg;
1159	/* increase inflight counter to offset decrement in callback */
1160	atomic_inc(&queue->inflight_packets);
1161	uarg->ops->complete(NULL, uarg, true);
1162	skb_shinfo(skb)->destructor_arg = NULL;
1163
1164	/* Fill the skb with the new (local) frags. */
1165	memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1166	skb_shinfo(skb)->nr_frags = i;
1167	skb->truesize += i * PAGE_SIZE;
1168
1169	return 0;
1170}
1171
1172static int xenvif_tx_submit(struct xenvif_queue *queue)
1173{
1174	struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1175	struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1176	struct sk_buff *skb;
1177	int work_done = 0;
1178
1179	while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1180		struct xen_netif_tx_request *txp;
1181		u16 pending_idx;
1182
1183		pending_idx = copy_pending_idx(skb, 0);
1184		txp = &queue->pending_tx_info[pending_idx].req;
1185
1186		/* Check the remap error code. */
1187		if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1188			/* If there was an error, xenvif_tx_check_gop is
1189			 * expected to release all the frags which were mapped,
1190			 * so kfree_skb shouldn't do it again
1191			 */
1192			skb_shinfo(skb)->nr_frags = 0;
1193			if (skb_has_frag_list(skb)) {
1194				struct sk_buff *nskb =
1195						skb_shinfo(skb)->frag_list;
1196				skb_shinfo(nskb)->nr_frags = 0;
1197			}
1198			kfree_skb(skb);
1199			continue;
1200		}
1201
1202		if (txp->flags & XEN_NETTXF_csum_blank)
1203			skb->ip_summed = CHECKSUM_PARTIAL;
1204		else if (txp->flags & XEN_NETTXF_data_validated)
1205			skb->ip_summed = CHECKSUM_UNNECESSARY;
1206
1207		xenvif_fill_frags(queue, skb);
1208
1209		if (unlikely(skb_has_frag_list(skb))) {
1210			struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1211			xenvif_skb_zerocopy_prepare(queue, nskb);
1212			if (xenvif_handle_frag_list(queue, skb)) {
1213				if (net_ratelimit())
1214					netdev_err(queue->vif->dev,
1215						   "Not enough memory to consolidate frag_list!\n");
1216				xenvif_skb_zerocopy_prepare(queue, skb);
1217				kfree_skb(skb);
1218				continue;
1219			}
1220			/* Copied all the bits from the frag list -- free it. */
1221			skb_frag_list_init(skb);
1222			kfree_skb(nskb);
1223		}
1224
1225		skb->dev      = queue->vif->dev;
1226		skb->protocol = eth_type_trans(skb, skb->dev);
1227		skb_reset_network_header(skb);
1228
1229		if (checksum_setup(queue, skb)) {
1230			netdev_dbg(queue->vif->dev,
1231				   "Can't setup checksum in net_tx_action\n");
1232			/* We have to set this flag to trigger the callback */
1233			if (skb_shinfo(skb)->destructor_arg)
1234				xenvif_skb_zerocopy_prepare(queue, skb);
1235			kfree_skb(skb);
1236			continue;
1237		}
1238
1239		skb_probe_transport_header(skb);
1240
1241		/* If the packet is GSO then we will have just set up the
1242		 * transport header offset in checksum_setup so it's now
1243		 * straightforward to calculate gso_segs.
1244		 */
1245		if (skb_is_gso(skb)) {
1246			int mss, hdrlen;
1247
1248			/* GSO implies having the L4 header. */
1249			WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1250			if (unlikely(!skb_transport_header_was_set(skb))) {
1251				kfree_skb(skb);
1252				continue;
1253			}
1254
1255			mss = skb_shinfo(skb)->gso_size;
1256			hdrlen = skb_tcp_all_headers(skb);
1257
1258			skb_shinfo(skb)->gso_segs =
1259				DIV_ROUND_UP(skb->len - hdrlen, mss);
1260		}
1261
1262		queue->stats.rx_bytes += skb->len;
1263		queue->stats.rx_packets++;
1264
1265		work_done++;
1266
1267		/* Set this flag right before netif_receive_skb, otherwise
1268		 * someone might think this packet already left netback, and
1269		 * do a skb_copy_ubufs while we are still in control of the
1270		 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1271		 */
1272		if (skb_shinfo(skb)->destructor_arg) {
1273			xenvif_skb_zerocopy_prepare(queue, skb);
1274			queue->stats.tx_zerocopy_sent++;
1275		}
1276
1277		netif_receive_skb(skb);
1278	}
1279
1280	return work_done;
1281}
1282
1283static void xenvif_zerocopy_callback(struct sk_buff *skb,
1284				     struct ubuf_info *ubuf_base,
1285				     bool zerocopy_success)
1286{
1287	unsigned long flags;
1288	pending_ring_idx_t index;
1289	struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
1290	struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1291
1292	/* This is the only place where we grab this lock, to protect callbacks
1293	 * from each other.
1294	 */
1295	spin_lock_irqsave(&queue->callback_lock, flags);
1296	do {
1297		u16 pending_idx = ubuf->desc;
1298		ubuf = (struct ubuf_info_msgzc *) ubuf->ctx;
1299		BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1300			MAX_PENDING_REQS);
1301		index = pending_index(queue->dealloc_prod);
1302		queue->dealloc_ring[index] = pending_idx;
1303		/* Sync with xenvif_tx_dealloc_action:
1304		 * insert idx then incr producer.
1305		 */
1306		smp_wmb();
1307		queue->dealloc_prod++;
1308	} while (ubuf);
1309	spin_unlock_irqrestore(&queue->callback_lock, flags);
1310
1311	if (likely(zerocopy_success))
1312		queue->stats.tx_zerocopy_success++;
1313	else
1314		queue->stats.tx_zerocopy_fail++;
1315	xenvif_skb_zerocopy_complete(queue);
1316}
1317
1318const struct ubuf_info_ops xenvif_ubuf_ops = {
1319	.complete = xenvif_zerocopy_callback,
1320};
1321
1322static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1323{
1324	struct gnttab_unmap_grant_ref *gop;
1325	pending_ring_idx_t dc, dp;
1326	u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1327	unsigned int i = 0;
1328
1329	dc = queue->dealloc_cons;
1330	gop = queue->tx_unmap_ops;
1331
1332	/* Free up any grants we have finished using */
1333	do {
1334		dp = queue->dealloc_prod;
1335
1336		/* Ensure we see all indices enqueued by all
1337		 * xenvif_zerocopy_callback().
1338		 */
1339		smp_rmb();
1340
1341		while (dc != dp) {
1342			BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1343			pending_idx =
1344				queue->dealloc_ring[pending_index(dc++)];
1345
1346			pending_idx_release[gop - queue->tx_unmap_ops] =
1347				pending_idx;
1348			queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1349				queue->mmap_pages[pending_idx];
1350			gnttab_set_unmap_op(gop,
1351					    idx_to_kaddr(queue, pending_idx),
1352					    GNTMAP_host_map,
1353					    queue->grant_tx_handle[pending_idx]);
1354			xenvif_grant_handle_reset(queue, pending_idx);
1355			++gop;
1356		}
1357
1358	} while (dp != queue->dealloc_prod);
1359
1360	queue->dealloc_cons = dc;
1361
1362	if (gop - queue->tx_unmap_ops > 0) {
1363		int ret;
1364		ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1365					NULL,
1366					queue->pages_to_unmap,
1367					gop - queue->tx_unmap_ops);
1368		if (ret) {
1369			netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1370				   gop - queue->tx_unmap_ops, ret);
1371			for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1372				if (gop[i].status != GNTST_okay)
1373					netdev_err(queue->vif->dev,
1374						   " host_addr: 0x%llx handle: 0x%x status: %d\n",
1375						   gop[i].host_addr,
1376						   gop[i].handle,
1377						   gop[i].status);
1378			}
1379			BUG();
1380		}
1381	}
1382
1383	for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1384		xenvif_idx_release(queue, pending_idx_release[i],
1385				   XEN_NETIF_RSP_OKAY);
1386}
1387
1388
1389/* Called after netfront has transmitted */
1390int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1391{
1392	unsigned nr_mops = 0, nr_cops = 0;
1393	int work_done, ret;
1394
1395	if (unlikely(!tx_work_todo(queue)))
1396		return 0;
1397
1398	xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1399
1400	if (nr_cops == 0)
1401		return 0;
1402
1403	gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1404	if (nr_mops != 0) {
1405		ret = gnttab_map_refs(queue->tx_map_ops,
1406				      NULL,
1407				      queue->pages_to_map,
1408				      nr_mops);
1409		if (ret) {
1410			unsigned int i;
1411
1412			netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1413				   nr_mops, ret);
1414			for (i = 0; i < nr_mops; ++i)
1415				WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1416				             GNTST_okay);
1417		}
1418	}
1419
1420	work_done = xenvif_tx_submit(queue);
1421
1422	return work_done;
1423}
1424
1425static void _make_tx_response(struct xenvif_queue *queue,
1426			     const struct xen_netif_tx_request *txp,
1427			     unsigned int extra_count,
1428			     s8 status)
1429{
1430	RING_IDX i = queue->tx.rsp_prod_pvt;
1431	struct xen_netif_tx_response *resp;
1432
1433	resp = RING_GET_RESPONSE(&queue->tx, i);
1434	resp->id     = txp->id;
1435	resp->status = status;
1436
1437	while (extra_count-- != 0)
1438		RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1439
1440	queue->tx.rsp_prod_pvt = ++i;
1441}
1442
1443static void push_tx_responses(struct xenvif_queue *queue)
1444{
1445	int notify;
1446
1447	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1448	if (notify)
1449		notify_remote_via_irq(queue->tx_irq);
1450}
1451
1452static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1453			       s8 status)
1454{
1455	struct pending_tx_info *pending_tx_info;
1456	pending_ring_idx_t index;
1457	unsigned long flags;
1458
1459	pending_tx_info = &queue->pending_tx_info[pending_idx];
1460
1461	spin_lock_irqsave(&queue->response_lock, flags);
1462
1463	_make_tx_response(queue, &pending_tx_info->req,
1464			  pending_tx_info->extra_count, status);
1465
1466	/* Release the pending index before pusing the Tx response so
1467	 * its available before a new Tx request is pushed by the
1468	 * frontend.
1469	 */
1470	index = pending_index(queue->pending_prod++);
1471	queue->pending_ring[index] = pending_idx;
1472
1473	push_tx_responses(queue);
1474
1475	spin_unlock_irqrestore(&queue->response_lock, flags);
1476}
1477
1478static void make_tx_response(struct xenvif_queue *queue,
1479			     const struct xen_netif_tx_request *txp,
1480			     unsigned int extra_count,
1481			     s8 status)
1482{
1483	unsigned long flags;
1484
1485	spin_lock_irqsave(&queue->response_lock, flags);
1486
1487	_make_tx_response(queue, txp, extra_count, status);
1488	push_tx_responses(queue);
1489
1490	spin_unlock_irqrestore(&queue->response_lock, flags);
1491}
1492
1493static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1494{
1495	int ret;
1496	struct gnttab_unmap_grant_ref tx_unmap_op;
1497
1498	gnttab_set_unmap_op(&tx_unmap_op,
1499			    idx_to_kaddr(queue, pending_idx),
1500			    GNTMAP_host_map,
1501			    queue->grant_tx_handle[pending_idx]);
1502	xenvif_grant_handle_reset(queue, pending_idx);
1503
1504	ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1505				&queue->mmap_pages[pending_idx], 1);
1506	if (ret) {
1507		netdev_err(queue->vif->dev,
1508			   "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1509			   ret,
1510			   pending_idx,
1511			   tx_unmap_op.host_addr,
1512			   tx_unmap_op.handle,
1513			   tx_unmap_op.status);
1514		BUG();
1515	}
1516}
1517
1518static inline int tx_work_todo(struct xenvif_queue *queue)
1519{
1520	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1521		return 1;
1522
1523	return 0;
1524}
1525
1526static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1527{
1528	return queue->dealloc_cons != queue->dealloc_prod;
1529}
1530
1531void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1532{
1533	if (queue->tx.sring)
1534		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1535					queue->tx.sring);
1536	if (queue->rx.sring)
1537		xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1538					queue->rx.sring);
1539}
1540
1541int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1542				   grant_ref_t tx_ring_ref,
1543				   grant_ref_t rx_ring_ref)
1544{
1545	void *addr;
1546	struct xen_netif_tx_sring *txs;
1547	struct xen_netif_rx_sring *rxs;
1548	RING_IDX rsp_prod, req_prod;
1549	int err;
1550
1551	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1552				     &tx_ring_ref, 1, &addr);
1553	if (err)
1554		goto err;
1555
1556	txs = (struct xen_netif_tx_sring *)addr;
1557	rsp_prod = READ_ONCE(txs->rsp_prod);
1558	req_prod = READ_ONCE(txs->req_prod);
1559
1560	BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1561
1562	err = -EIO;
1563	if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1564		goto err;
1565
1566	err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1567				     &rx_ring_ref, 1, &addr);
1568	if (err)
1569		goto err;
1570
1571	rxs = (struct xen_netif_rx_sring *)addr;
1572	rsp_prod = READ_ONCE(rxs->rsp_prod);
1573	req_prod = READ_ONCE(rxs->req_prod);
1574
1575	BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1576
1577	err = -EIO;
1578	if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1579		goto err;
1580
1581	return 0;
1582
1583err:
1584	xenvif_unmap_frontend_data_rings(queue);
1585	return err;
1586}
1587
1588static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1589{
1590	/* Dealloc thread must remain running until all inflight
1591	 * packets complete.
1592	 */
1593	return kthread_should_stop() &&
1594		!atomic_read(&queue->inflight_packets);
1595}
1596
1597int xenvif_dealloc_kthread(void *data)
1598{
1599	struct xenvif_queue *queue = data;
1600
1601	for (;;) {
1602		wait_event_interruptible(queue->dealloc_wq,
1603					 tx_dealloc_work_todo(queue) ||
1604					 xenvif_dealloc_kthread_should_stop(queue));
1605		if (xenvif_dealloc_kthread_should_stop(queue))
1606			break;
1607
1608		xenvif_tx_dealloc_action(queue);
1609		cond_resched();
1610	}
1611
1612	/* Unmap anything remaining*/
1613	if (tx_dealloc_work_todo(queue))
1614		xenvif_tx_dealloc_action(queue);
1615
1616	return 0;
1617}
1618
1619static void make_ctrl_response(struct xenvif *vif,
1620			       const struct xen_netif_ctrl_request *req,
1621			       u32 status, u32 data)
1622{
1623	RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1624	struct xen_netif_ctrl_response rsp = {
1625		.id = req->id,
1626		.type = req->type,
1627		.status = status,
1628		.data = data,
1629	};
1630
1631	*RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1632	vif->ctrl.rsp_prod_pvt = ++idx;
1633}
1634
1635static void push_ctrl_response(struct xenvif *vif)
1636{
1637	int notify;
1638
1639	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1640	if (notify)
1641		notify_remote_via_irq(vif->ctrl_irq);
1642}
1643
1644static void process_ctrl_request(struct xenvif *vif,
1645				 const struct xen_netif_ctrl_request *req)
1646{
1647	u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1648	u32 data = 0;
1649
1650	switch (req->type) {
1651	case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1652		status = xenvif_set_hash_alg(vif, req->data[0]);
1653		break;
1654
1655	case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1656		status = xenvif_get_hash_flags(vif, &data);
1657		break;
1658
1659	case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1660		status = xenvif_set_hash_flags(vif, req->data[0]);
1661		break;
1662
1663	case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1664		status = xenvif_set_hash_key(vif, req->data[0],
1665					     req->data[1]);
1666		break;
1667
1668	case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1669		status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1670		data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1671		break;
1672
1673	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1674		status = xenvif_set_hash_mapping_size(vif,
1675						      req->data[0]);
1676		break;
1677
1678	case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1679		status = xenvif_set_hash_mapping(vif, req->data[0],
1680						 req->data[1],
1681						 req->data[2]);
1682		break;
1683
1684	default:
1685		break;
1686	}
1687
1688	make_ctrl_response(vif, req, status, data);
1689	push_ctrl_response(vif);
1690}
1691
1692static void xenvif_ctrl_action(struct xenvif *vif)
1693{
1694	for (;;) {
1695		RING_IDX req_prod, req_cons;
1696
1697		req_prod = vif->ctrl.sring->req_prod;
1698		req_cons = vif->ctrl.req_cons;
1699
1700		/* Make sure we can see requests before we process them. */
1701		rmb();
1702
1703		if (req_cons == req_prod)
1704			break;
1705
1706		while (req_cons != req_prod) {
1707			struct xen_netif_ctrl_request req;
1708
1709			RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1710			req_cons++;
1711
1712			process_ctrl_request(vif, &req);
1713		}
1714
1715		vif->ctrl.req_cons = req_cons;
1716		vif->ctrl.sring->req_event = req_cons + 1;
1717	}
1718}
1719
1720static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1721{
1722	if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1723		return true;
1724
1725	return false;
1726}
1727
1728irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1729{
1730	struct xenvif *vif = data;
1731	unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1732
1733	while (xenvif_ctrl_work_todo(vif)) {
1734		xenvif_ctrl_action(vif);
1735		eoi_flag = 0;
1736	}
1737
1738	xen_irq_lateeoi(irq, eoi_flag);
1739
1740	return IRQ_HANDLED;
1741}
1742
1743static int __init netback_init(void)
1744{
1745	int rc = 0;
1746
1747	if (!xen_domain())
1748		return -ENODEV;
1749
1750	/* Allow as many queues as there are CPUs but max. 8 if user has not
1751	 * specified a value.
1752	 */
1753	if (xenvif_max_queues == 0)
1754		xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1755					  num_online_cpus());
1756
1757	if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1758		pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1759			fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1760		fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1761	}
1762
1763	rc = xenvif_xenbus_init();
1764	if (rc)
1765		goto failed_init;
1766
1767#ifdef CONFIG_DEBUG_FS
1768	xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1769#endif /* CONFIG_DEBUG_FS */
1770
1771	return 0;
1772
1773failed_init:
1774	return rc;
1775}
1776
1777module_init(netback_init);
1778
1779static void __exit netback_fini(void)
1780{
1781#ifdef CONFIG_DEBUG_FS
1782	debugfs_remove_recursive(xen_netback_dbg_root);
1783#endif /* CONFIG_DEBUG_FS */
1784	xenvif_xenbus_fini();
1785}
1786module_exit(netback_fini);
1787
1788MODULE_DESCRIPTION("Xen backend network device module");
1789MODULE_LICENSE("Dual BSD/GPL");
1790MODULE_ALIAS("xen-backend:vif");