linux/drivers/net/xen-netback/rx.c
<<
>>
Prefs
   1/*
   2 * Copyright (c) 2016 Citrix Systems Inc.
   3 * Copyright (c) 2002-2005, K A Fraser
   4 *
   5 * This program is free software; you can redistribute it and/or
   6 * modify it under the terms of the GNU General Public License version 2
   7 * as published by the Free Software Foundation; or, when distributed
   8 * separately from the Linux kernel or incorporated into other
   9 * software packages, subject to the following license:
  10 *
  11 * Permission is hereby granted, free of charge, to any person obtaining a copy
  12 * of this source file (the "Software"), to deal in the Software without
  13 * restriction, including without limitation the rights to use, copy, modify,
  14 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  15 * and to permit persons to whom the Software is furnished to do so, subject to
  16 * the following conditions:
  17 *
  18 * The above copyright notice and this permission notice shall be included in
  19 * all copies or substantial portions of the Software.
  20 *
  21 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  22 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  23 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  24 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  25 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  26 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  27 * IN THE SOFTWARE.
  28 */
  29#include "common.h"
  30
  31#include <linux/kthread.h>
  32
  33#include <xen/xen.h>
  34#include <xen/events.h>
  35
  36static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue)
  37{
  38        RING_IDX prod, cons;
  39        struct sk_buff *skb;
  40        int needed;
  41
  42        skb = skb_peek(&queue->rx_queue);
  43        if (!skb)
  44                return false;
  45
  46        needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE);
  47        if (skb_is_gso(skb))
  48                needed++;
  49        if (skb->sw_hash)
  50                needed++;
  51
  52        do {
  53                prod = queue->rx.sring->req_prod;
  54                cons = queue->rx.req_cons;
  55
  56                if (prod - cons >= needed)
  57                        return true;
  58
  59                queue->rx.sring->req_event = prod + 1;
  60
  61                /* Make sure event is visible before we check prod
  62                 * again.
  63                 */
  64                mb();
  65        } while (queue->rx.sring->req_prod != prod);
  66
  67        return false;
  68}
  69
  70void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb)
  71{
  72        unsigned long flags;
  73
  74        spin_lock_irqsave(&queue->rx_queue.lock, flags);
  75
  76        __skb_queue_tail(&queue->rx_queue, skb);
  77
  78        queue->rx_queue_len += skb->len;
  79        if (queue->rx_queue_len > queue->rx_queue_max) {
  80                struct net_device *dev = queue->vif->dev;
  81
  82                netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id));
  83        }
  84
  85        spin_unlock_irqrestore(&queue->rx_queue.lock, flags);
  86}
  87
  88static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue)
  89{
  90        struct sk_buff *skb;
  91
  92        spin_lock_irq(&queue->rx_queue.lock);
  93
  94        skb = __skb_dequeue(&queue->rx_queue);
  95        if (skb) {
  96                queue->rx_queue_len -= skb->len;
  97                if (queue->rx_queue_len < queue->rx_queue_max) {
  98                        struct netdev_queue *txq;
  99
 100                        txq = netdev_get_tx_queue(queue->vif->dev, queue->id);
 101                        netif_tx_wake_queue(txq);
 102                }
 103        }
 104
 105        spin_unlock_irq(&queue->rx_queue.lock);
 106
 107        return skb;
 108}
 109
 110static void xenvif_rx_queue_purge(struct xenvif_queue *queue)
 111{
 112        struct sk_buff *skb;
 113
 114        while ((skb = xenvif_rx_dequeue(queue)) != NULL)
 115                kfree_skb(skb);
 116}
 117
 118static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
 119{
 120        struct sk_buff *skb;
 121
 122        for (;;) {
 123                skb = skb_peek(&queue->rx_queue);
 124                if (!skb)
 125                        break;
 126                if (time_before(jiffies, XENVIF_RX_CB(skb)->expires))
 127                        break;
 128                xenvif_rx_dequeue(queue);
 129                kfree_skb(skb);
 130        }
 131}
 132
 133static void xenvif_rx_copy_flush(struct xenvif_queue *queue)
 134{
 135        unsigned int i;
 136        int notify;
 137
 138        gnttab_batch_copy(queue->rx_copy.op, queue->rx_copy.num);
 139
 140        for (i = 0; i < queue->rx_copy.num; i++) {
 141                struct gnttab_copy *op;
 142
 143                op = &queue->rx_copy.op[i];
 144
 145                /* If the copy failed, overwrite the status field in
 146                 * the corresponding response.
 147                 */
 148                if (unlikely(op->status != GNTST_okay)) {
 149                        struct xen_netif_rx_response *rsp;
 150
 151                        rsp = RING_GET_RESPONSE(&queue->rx,
 152                                                queue->rx_copy.idx[i]);
 153                        rsp->status = op->status;
 154                }
 155        }
 156
 157        queue->rx_copy.num = 0;
 158
 159        /* Push responses for all completed packets. */
 160        RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, notify);
 161        if (notify)
 162                notify_remote_via_irq(queue->rx_irq);
 163
 164        __skb_queue_purge(queue->rx_copy.completed);
 165}
 166
 167static void xenvif_rx_copy_add(struct xenvif_queue *queue,
 168                               struct xen_netif_rx_request *req,
 169                               unsigned int offset, void *data, size_t len)
 170{
 171        struct gnttab_copy *op;
 172        struct page *page;
 173        struct xen_page_foreign *foreign;
 174
 175        if (queue->rx_copy.num == COPY_BATCH_SIZE)
 176                xenvif_rx_copy_flush(queue);
 177
 178        op = &queue->rx_copy.op[queue->rx_copy.num];
 179
 180        page = virt_to_page(data);
 181
 182        op->flags = GNTCOPY_dest_gref;
 183
 184        foreign = xen_page_foreign(page);
 185        if (foreign) {
 186                op->source.domid = foreign->domid;
 187                op->source.u.ref = foreign->gref;
 188                op->flags |= GNTCOPY_source_gref;
 189        } else {
 190                op->source.u.gmfn = virt_to_gfn(data);
 191                op->source.domid  = DOMID_SELF;
 192        }
 193
 194        op->source.offset = xen_offset_in_page(data);
 195        op->dest.u.ref    = req->gref;
 196        op->dest.domid    = queue->vif->domid;
 197        op->dest.offset   = offset;
 198        op->len           = len;
 199
 200        queue->rx_copy.idx[queue->rx_copy.num] = queue->rx.req_cons;
 201        queue->rx_copy.num++;
 202}
 203
 204static unsigned int xenvif_gso_type(struct sk_buff *skb)
 205{
 206        if (skb_is_gso(skb)) {
 207                if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
 208                        return XEN_NETIF_GSO_TYPE_TCPV4;
 209                else
 210                        return XEN_NETIF_GSO_TYPE_TCPV6;
 211        }
 212        return XEN_NETIF_GSO_TYPE_NONE;
 213}
 214
 215struct xenvif_pkt_state {
 216        struct sk_buff *skb;
 217        size_t remaining_len;
 218        struct sk_buff *frag_iter;
 219        int frag; /* frag == -1 => frag_iter->head */
 220        unsigned int frag_offset;
 221        struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
 222        unsigned int extra_count;
 223        unsigned int slot;
 224};
 225
 226static void xenvif_rx_next_skb(struct xenvif_queue *queue,
 227                               struct xenvif_pkt_state *pkt)
 228{
 229        struct sk_buff *skb;
 230        unsigned int gso_type;
 231
 232        skb = xenvif_rx_dequeue(queue);
 233
 234        queue->stats.tx_bytes += skb->len;
 235        queue->stats.tx_packets++;
 236
 237        /* Reset packet state. */
 238        memset(pkt, 0, sizeof(struct xenvif_pkt_state));
 239
 240        pkt->skb = skb;
 241        pkt->frag_iter = skb;
 242        pkt->remaining_len = skb->len;
 243        pkt->frag = -1;
 244
 245        gso_type = xenvif_gso_type(skb);
 246        if ((1 << gso_type) & queue->vif->gso_mask) {
 247                struct xen_netif_extra_info *extra;
 248
 249                extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
 250
 251                extra->u.gso.type = gso_type;
 252                extra->u.gso.size = skb_shinfo(skb)->gso_size;
 253                extra->u.gso.pad = 0;
 254                extra->u.gso.features = 0;
 255                extra->type = XEN_NETIF_EXTRA_TYPE_GSO;
 256                extra->flags = 0;
 257
 258                pkt->extra_count++;
 259        }
 260
 261        if (skb->sw_hash) {
 262                struct xen_netif_extra_info *extra;
 263
 264                extra = &pkt->extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
 265
 266                extra->u.hash.algorithm =
 267                        XEN_NETIF_CTRL_HASH_ALGORITHM_TOEPLITZ;
 268
 269                if (skb->l4_hash)
 270                        extra->u.hash.type =
 271                                skb->protocol == htons(ETH_P_IP) ?
 272                                _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP :
 273                                _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP;
 274                else
 275                        extra->u.hash.type =
 276                                skb->protocol == htons(ETH_P_IP) ?
 277                                _XEN_NETIF_CTRL_HASH_TYPE_IPV4 :
 278                                _XEN_NETIF_CTRL_HASH_TYPE_IPV6;
 279
 280                *(uint32_t *)extra->u.hash.value = skb_get_hash_raw(skb);
 281
 282                extra->type = XEN_NETIF_EXTRA_TYPE_HASH;
 283                extra->flags = 0;
 284
 285                pkt->extra_count++;
 286        }
 287}
 288
 289static void xenvif_rx_complete(struct xenvif_queue *queue,
 290                               struct xenvif_pkt_state *pkt)
 291{
 292        /* All responses are ready to be pushed. */
 293        queue->rx.rsp_prod_pvt = queue->rx.req_cons;
 294
 295        __skb_queue_tail(queue->rx_copy.completed, pkt->skb);
 296}
 297
 298static void xenvif_rx_next_frag(struct xenvif_pkt_state *pkt)
 299{
 300        struct sk_buff *frag_iter = pkt->frag_iter;
 301        unsigned int nr_frags = skb_shinfo(frag_iter)->nr_frags;
 302
 303        pkt->frag++;
 304        pkt->frag_offset = 0;
 305
 306        if (pkt->frag >= nr_frags) {
 307                if (frag_iter == pkt->skb)
 308                        pkt->frag_iter = skb_shinfo(frag_iter)->frag_list;
 309                else
 310                        pkt->frag_iter = frag_iter->next;
 311
 312                pkt->frag = -1;
 313        }
 314}
 315
 316static void xenvif_rx_next_chunk(struct xenvif_queue *queue,
 317                                 struct xenvif_pkt_state *pkt,
 318                                 unsigned int offset, void **data,
 319                                 size_t *len)
 320{
 321        struct sk_buff *frag_iter = pkt->frag_iter;
 322        void *frag_data;
 323        size_t frag_len, chunk_len;
 324
 325        BUG_ON(!frag_iter);
 326
 327        if (pkt->frag == -1) {
 328                frag_data = frag_iter->data;
 329                frag_len = skb_headlen(frag_iter);
 330        } else {
 331                skb_frag_t *frag = &skb_shinfo(frag_iter)->frags[pkt->frag];
 332
 333                frag_data = skb_frag_address(frag);
 334                frag_len = skb_frag_size(frag);
 335        }
 336
 337        frag_data += pkt->frag_offset;
 338        frag_len -= pkt->frag_offset;
 339
 340        chunk_len = min_t(size_t, frag_len, XEN_PAGE_SIZE - offset);
 341        chunk_len = min_t(size_t, chunk_len, XEN_PAGE_SIZE -
 342                                             xen_offset_in_page(frag_data));
 343
 344        pkt->frag_offset += chunk_len;
 345
 346        /* Advance to next frag? */
 347        if (frag_len == chunk_len)
 348                xenvif_rx_next_frag(pkt);
 349
 350        *data = frag_data;
 351        *len = chunk_len;
 352}
 353
 354static void xenvif_rx_data_slot(struct xenvif_queue *queue,
 355                                struct xenvif_pkt_state *pkt,
 356                                struct xen_netif_rx_request *req,
 357                                struct xen_netif_rx_response *rsp)
 358{
 359        unsigned int offset = 0;
 360        unsigned int flags;
 361
 362        do {
 363                size_t len;
 364                void *data;
 365
 366                xenvif_rx_next_chunk(queue, pkt, offset, &data, &len);
 367                xenvif_rx_copy_add(queue, req, offset, data, len);
 368
 369                offset += len;
 370                pkt->remaining_len -= len;
 371
 372        } while (offset < XEN_PAGE_SIZE && pkt->remaining_len > 0);
 373
 374        if (pkt->remaining_len > 0)
 375                flags = XEN_NETRXF_more_data;
 376        else
 377                flags = 0;
 378
 379        if (pkt->slot == 0) {
 380                struct sk_buff *skb = pkt->skb;
 381
 382                if (skb->ip_summed == CHECKSUM_PARTIAL)
 383                        flags |= XEN_NETRXF_csum_blank |
 384                                 XEN_NETRXF_data_validated;
 385                else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 386                        flags |= XEN_NETRXF_data_validated;
 387
 388                if (pkt->extra_count != 0)
 389                        flags |= XEN_NETRXF_extra_info;
 390        }
 391
 392        rsp->offset = 0;
 393        rsp->flags = flags;
 394        rsp->id = req->id;
 395        rsp->status = (s16)offset;
 396}
 397
 398static void xenvif_rx_extra_slot(struct xenvif_queue *queue,
 399                                 struct xenvif_pkt_state *pkt,
 400                                 struct xen_netif_rx_request *req,
 401                                 struct xen_netif_rx_response *rsp)
 402{
 403        struct xen_netif_extra_info *extra = (void *)rsp;
 404        unsigned int i;
 405
 406        pkt->extra_count--;
 407
 408        for (i = 0; i < ARRAY_SIZE(pkt->extras); i++) {
 409                if (pkt->extras[i].type) {
 410                        *extra = pkt->extras[i];
 411
 412                        if (pkt->extra_count != 0)
 413                                extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE;
 414
 415                        pkt->extras[i].type = 0;
 416                        return;
 417                }
 418        }
 419        BUG();
 420}
 421
 422void xenvif_rx_skb(struct xenvif_queue *queue)
 423{
 424        struct xenvif_pkt_state pkt;
 425
 426        xenvif_rx_next_skb(queue, &pkt);
 427
 428        queue->last_rx_time = jiffies;
 429
 430        do {
 431                struct xen_netif_rx_request *req;
 432                struct xen_netif_rx_response *rsp;
 433
 434                req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons);
 435                rsp = RING_GET_RESPONSE(&queue->rx, queue->rx.req_cons);
 436
 437                /* Extras must go after the first data slot */
 438                if (pkt.slot != 0 && pkt.extra_count != 0)
 439                        xenvif_rx_extra_slot(queue, &pkt, req, rsp);
 440                else
 441                        xenvif_rx_data_slot(queue, &pkt, req, rsp);
 442
 443                queue->rx.req_cons++;
 444                pkt.slot++;
 445        } while (pkt.remaining_len > 0 || pkt.extra_count != 0);
 446
 447        xenvif_rx_complete(queue, &pkt);
 448}
 449
 450#define RX_BATCH_SIZE 64
 451
 452void xenvif_rx_action(struct xenvif_queue *queue)
 453{
 454        struct sk_buff_head completed_skbs;
 455        unsigned int work_done = 0;
 456
 457        __skb_queue_head_init(&completed_skbs);
 458        queue->rx_copy.completed = &completed_skbs;
 459
 460        while (xenvif_rx_ring_slots_available(queue) &&
 461               work_done < RX_BATCH_SIZE) {
 462                xenvif_rx_skb(queue);
 463                work_done++;
 464        }
 465
 466        /* Flush any pending copies and complete all skbs. */
 467        xenvif_rx_copy_flush(queue);
 468}
 469
 470static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)
 471{
 472        RING_IDX prod, cons;
 473
 474        prod = queue->rx.sring->req_prod;
 475        cons = queue->rx.req_cons;
 476
 477        return !queue->stalled &&
 478                prod - cons < 1 &&
 479                time_after(jiffies,
 480                           queue->last_rx_time + queue->vif->stall_timeout);
 481}
 482
 483static bool xenvif_rx_queue_ready(struct xenvif_queue *queue)
 484{
 485        RING_IDX prod, cons;
 486
 487        prod = queue->rx.sring->req_prod;
 488        cons = queue->rx.req_cons;
 489
 490        return queue->stalled && prod - cons >= 1;
 491}
 492
 493static bool xenvif_have_rx_work(struct xenvif_queue *queue)
 494{
 495        return xenvif_rx_ring_slots_available(queue) ||
 496                (queue->vif->stall_timeout &&
 497                 (xenvif_rx_queue_stalled(queue) ||
 498                  xenvif_rx_queue_ready(queue))) ||
 499                kthread_should_stop() ||
 500                queue->vif->disabled;
 501}
 502
 503static long xenvif_rx_queue_timeout(struct xenvif_queue *queue)
 504{
 505        struct sk_buff *skb;
 506        long timeout;
 507
 508        skb = skb_peek(&queue->rx_queue);
 509        if (!skb)
 510                return MAX_SCHEDULE_TIMEOUT;
 511
 512        timeout = XENVIF_RX_CB(skb)->expires - jiffies;
 513        return timeout < 0 ? 0 : timeout;
 514}
 515
 516/* Wait until the guest Rx thread has work.
 517 *
 518 * The timeout needs to be adjusted based on the current head of the
 519 * queue (and not just the head at the beginning).  In particular, if
 520 * the queue is initially empty an infinite timeout is used and this
 521 * needs to be reduced when a skb is queued.
 522 *
 523 * This cannot be done with wait_event_timeout() because it only
 524 * calculates the timeout once.
 525 */
 526static void xenvif_wait_for_rx_work(struct xenvif_queue *queue)
 527{
 528        DEFINE_WAIT(wait);
 529
 530        if (xenvif_have_rx_work(queue))
 531                return;
 532
 533        for (;;) {
 534                long ret;
 535
 536                prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE);
 537                if (xenvif_have_rx_work(queue))
 538                        break;
 539                ret = schedule_timeout(xenvif_rx_queue_timeout(queue));
 540                if (!ret)
 541                        break;
 542        }
 543        finish_wait(&queue->wq, &wait);
 544}
 545
 546static void xenvif_queue_carrier_off(struct xenvif_queue *queue)
 547{
 548        struct xenvif *vif = queue->vif;
 549
 550        queue->stalled = true;
 551
 552        /* At least one queue has stalled? Disable the carrier. */
 553        spin_lock(&vif->lock);
 554        if (vif->stalled_queues++ == 0) {
 555                netdev_info(vif->dev, "Guest Rx stalled");
 556                netif_carrier_off(vif->dev);
 557        }
 558        spin_unlock(&vif->lock);
 559}
 560
 561static void xenvif_queue_carrier_on(struct xenvif_queue *queue)
 562{
 563        struct xenvif *vif = queue->vif;
 564
 565        queue->last_rx_time = jiffies; /* Reset Rx stall detection. */
 566        queue->stalled = false;
 567
 568        /* All queues are ready? Enable the carrier. */
 569        spin_lock(&vif->lock);
 570        if (--vif->stalled_queues == 0) {
 571                netdev_info(vif->dev, "Guest Rx ready");
 572                netif_carrier_on(vif->dev);
 573        }
 574        spin_unlock(&vif->lock);
 575}
 576
 577int xenvif_kthread_guest_rx(void *data)
 578{
 579        struct xenvif_queue *queue = data;
 580        struct xenvif *vif = queue->vif;
 581
 582        if (!vif->stall_timeout)
 583                xenvif_queue_carrier_on(queue);
 584
 585        for (;;) {
 586                xenvif_wait_for_rx_work(queue);
 587
 588                if (kthread_should_stop())
 589                        break;
 590
 591                /* This frontend is found to be rogue, disable it in
 592                 * kthread context. Currently this is only set when
 593                 * netback finds out frontend sends malformed packet,
 594                 * but we cannot disable the interface in softirq
 595                 * context so we defer it here, if this thread is
 596                 * associated with queue 0.
 597                 */
 598                if (unlikely(vif->disabled && queue->id == 0)) {
 599                        xenvif_carrier_off(vif);
 600                        break;
 601                }
 602
 603                if (!skb_queue_empty(&queue->rx_queue))
 604                        xenvif_rx_action(queue);
 605
 606                /* If the guest hasn't provided any Rx slots for a
 607                 * while it's probably not responsive, drop the
 608                 * carrier so packets are dropped earlier.
 609                 */
 610                if (vif->stall_timeout) {
 611                        if (xenvif_rx_queue_stalled(queue))
 612                                xenvif_queue_carrier_off(queue);
 613                        else if (xenvif_rx_queue_ready(queue))
 614                                xenvif_queue_carrier_on(queue);
 615                }
 616
 617                /* Queued packets may have foreign pages from other
 618                 * domains.  These cannot be queued indefinitely as
 619                 * this would starve guests of grant refs and transmit
 620                 * slots.
 621                 */
 622                xenvif_rx_queue_drop_expired(queue);
 623
 624                cond_resched();
 625        }
 626
 627        /* Bin any remaining skbs */
 628        xenvif_rx_queue_purge(queue);
 629
 630        return 0;
 631}
 632