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