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 */ 55 bool separate_tx_rx_irq = true; 56 module_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 */ 61 unsigned int rx_drain_timeout_msecs = 10000; 62 module_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 */ 67 unsigned int rx_stall_timeout_msecs = 60000; 68 module_param(rx_stall_timeout_msecs, uint, 0444); 69 70 unsigned int xenvif_max_queues; 71 module_param_named(max_queues, xenvif_max_queues, uint, 0644); 72 MODULE_PARM_DESC(max_queues, 73 "Maximum number of queues per virtual interface"); 74 75 /* 76 * This is the maximum slots a skb can have. If a guest sends a skb 77 * which exceeds this limit it is considered malicious. 78 */ 79 #define FATAL_SKB_SLOTS_DEFAULT 20 80 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT; 81 module_param(fatal_skb_slots, uint, 0444); 82 83 /* The amount to copy out of the first guest Tx slot into the skb's 84 * linear area. If the first slot has more data, it will be mapped 85 * and put into the first frag. 86 * 87 * This is sized to avoid pulling headers from the frags for most 88 * TCP/IP packets. 89 */ 90 #define XEN_NETBACK_TX_COPY_LEN 128 91 92 93 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx, 94 u8 status); 95 96 static void make_tx_response(struct xenvif_queue *queue, 97 struct xen_netif_tx_request *txp, 98 unsigned int extra_count, 99 s8 st); 100 static void push_tx_responses(struct xenvif_queue *queue); 101 102 static inline int tx_work_todo(struct xenvif_queue *queue); 103 104 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue, 105 u16 id, 106 s8 st, 107 u16 offset, 108 u16 size, 109 u16 flags); 110 111 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue, 112 u16 idx) 113 { 114 return page_to_pfn(queue->mmap_pages[idx]); 115 } 116 117 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue, 118 u16 idx) 119 { 120 return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx)); 121 } 122 123 #define callback_param(vif, pending_idx) \ 124 (vif->pending_tx_info[pending_idx].callback_struct) 125 126 /* Find the containing VIF's structure from a pointer in pending_tx_info array 127 */ 128 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf) 129 { 130 u16 pending_idx = ubuf->desc; 131 struct pending_tx_info *temp = 132 container_of(ubuf, struct pending_tx_info, callback_struct); 133 return container_of(temp - pending_idx, 134 struct xenvif_queue, 135 pending_tx_info[0]); 136 } 137 138 static u16 frag_get_pending_idx(skb_frag_t *frag) 139 { 140 return (u16)frag->page_offset; 141 } 142 143 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx) 144 { 145 frag->page_offset = pending_idx; 146 } 147 148 static inline pending_ring_idx_t pending_index(unsigned i) 149 { 150 return i & (MAX_PENDING_REQS-1); 151 } 152 153 static bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue) 154 { 155 RING_IDX prod, cons; 156 struct sk_buff *skb; 157 int needed; 158 159 skb = skb_peek(&queue->rx_queue); 160 if (!skb) 161 return false; 162 163 needed = DIV_ROUND_UP(skb->len, XEN_PAGE_SIZE); 164 if (skb_is_gso(skb)) 165 needed++; 166 167 do { 168 prod = queue->rx.sring->req_prod; 169 cons = queue->rx.req_cons; 170 171 if (prod - cons >= needed) 172 return true; 173 174 queue->rx.sring->req_event = prod + 1; 175 176 /* Make sure event is visible before we check prod 177 * again. 178 */ 179 mb(); 180 } while (queue->rx.sring->req_prod != prod); 181 182 return false; 183 } 184 185 void xenvif_rx_queue_tail(struct xenvif_queue *queue, struct sk_buff *skb) 186 { 187 unsigned long flags; 188 189 spin_lock_irqsave(&queue->rx_queue.lock, flags); 190 191 __skb_queue_tail(&queue->rx_queue, skb); 192 193 queue->rx_queue_len += skb->len; 194 if (queue->rx_queue_len > queue->rx_queue_max) 195 netif_tx_stop_queue(netdev_get_tx_queue(queue->vif->dev, queue->id)); 196 197 spin_unlock_irqrestore(&queue->rx_queue.lock, flags); 198 } 199 200 static struct sk_buff *xenvif_rx_dequeue(struct xenvif_queue *queue) 201 { 202 struct sk_buff *skb; 203 204 spin_lock_irq(&queue->rx_queue.lock); 205 206 skb = __skb_dequeue(&queue->rx_queue); 207 if (skb) 208 queue->rx_queue_len -= skb->len; 209 210 spin_unlock_irq(&queue->rx_queue.lock); 211 212 return skb; 213 } 214 215 static void xenvif_rx_queue_maybe_wake(struct xenvif_queue *queue) 216 { 217 spin_lock_irq(&queue->rx_queue.lock); 218 219 if (queue->rx_queue_len < queue->rx_queue_max) 220 netif_tx_wake_queue(netdev_get_tx_queue(queue->vif->dev, queue->id)); 221 222 spin_unlock_irq(&queue->rx_queue.lock); 223 } 224 225 226 static void xenvif_rx_queue_purge(struct xenvif_queue *queue) 227 { 228 struct sk_buff *skb; 229 while ((skb = xenvif_rx_dequeue(queue)) != NULL) 230 kfree_skb(skb); 231 } 232 233 static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue) 234 { 235 struct sk_buff *skb; 236 237 for(;;) { 238 skb = skb_peek(&queue->rx_queue); 239 if (!skb) 240 break; 241 if (time_before(jiffies, XENVIF_RX_CB(skb)->expires)) 242 break; 243 xenvif_rx_dequeue(queue); 244 kfree_skb(skb); 245 } 246 } 247 248 struct netrx_pending_operations { 249 unsigned copy_prod, copy_cons; 250 unsigned meta_prod, meta_cons; 251 struct gnttab_copy *copy; 252 struct xenvif_rx_meta *meta; 253 int copy_off; 254 grant_ref_t copy_gref; 255 }; 256 257 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue, 258 struct netrx_pending_operations *npo) 259 { 260 struct xenvif_rx_meta *meta; 261 struct xen_netif_rx_request req; 262 263 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req); 264 265 meta = npo->meta + npo->meta_prod++; 266 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE; 267 meta->gso_size = 0; 268 meta->size = 0; 269 meta->id = req.id; 270 271 npo->copy_off = 0; 272 npo->copy_gref = req.gref; 273 274 return meta; 275 } 276 277 struct gop_frag_copy { 278 struct xenvif_queue *queue; 279 struct netrx_pending_operations *npo; 280 struct xenvif_rx_meta *meta; 281 int head; 282 int gso_type; 283 284 struct page *page; 285 }; 286 287 static void xenvif_setup_copy_gop(unsigned long gfn, 288 unsigned int offset, 289 unsigned int *len, 290 struct gop_frag_copy *info) 291 { 292 struct gnttab_copy *copy_gop; 293 struct xen_page_foreign *foreign; 294 /* Convenient aliases */ 295 struct xenvif_queue *queue = info->queue; 296 struct netrx_pending_operations *npo = info->npo; 297 struct page *page = info->page; 298 299 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET); 300 301 if (npo->copy_off == MAX_BUFFER_OFFSET) 302 info->meta = get_next_rx_buffer(queue, npo); 303 304 if (npo->copy_off + *len > MAX_BUFFER_OFFSET) 305 *len = MAX_BUFFER_OFFSET - npo->copy_off; 306 307 copy_gop = npo->copy + npo->copy_prod++; 308 copy_gop->flags = GNTCOPY_dest_gref; 309 copy_gop->len = *len; 310 311 foreign = xen_page_foreign(page); 312 if (foreign) { 313 copy_gop->source.domid = foreign->domid; 314 copy_gop->source.u.ref = foreign->gref; 315 copy_gop->flags |= GNTCOPY_source_gref; 316 } else { 317 copy_gop->source.domid = DOMID_SELF; 318 copy_gop->source.u.gmfn = gfn; 319 } 320 copy_gop->source.offset = offset; 321 322 copy_gop->dest.domid = queue->vif->domid; 323 copy_gop->dest.offset = npo->copy_off; 324 copy_gop->dest.u.ref = npo->copy_gref; 325 326 npo->copy_off += *len; 327 info->meta->size += *len; 328 329 /* Leave a gap for the GSO descriptor. */ 330 if (info->head && ((1 << info->gso_type) & queue->vif->gso_mask)) 331 queue->rx.req_cons++; 332 333 info->head = 0; /* There must be something in this buffer now */ 334 } 335 336 static void xenvif_gop_frag_copy_grant(unsigned long gfn, 337 unsigned offset, 338 unsigned int len, 339 void *data) 340 { 341 unsigned int bytes; 342 343 while (len) { 344 bytes = len; 345 xenvif_setup_copy_gop(gfn, offset, &bytes, data); 346 offset += bytes; 347 len -= bytes; 348 } 349 } 350 351 /* 352 * Set up the grant operations for this fragment. If it's a flipping 353 * interface, we also set up the unmap request from here. 354 */ 355 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb, 356 struct netrx_pending_operations *npo, 357 struct page *page, unsigned long size, 358 unsigned long offset, int *head) 359 { 360 struct gop_frag_copy info = { 361 .queue = queue, 362 .npo = npo, 363 .head = *head, 364 .gso_type = XEN_NETIF_GSO_TYPE_NONE, 365 }; 366 unsigned long bytes; 367 368 if (skb_is_gso(skb)) { 369 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) 370 info.gso_type = XEN_NETIF_GSO_TYPE_TCPV4; 371 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) 372 info.gso_type = XEN_NETIF_GSO_TYPE_TCPV6; 373 } 374 375 /* Data must not cross a page boundary. */ 376 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page)); 377 378 info.meta = npo->meta + npo->meta_prod - 1; 379 380 /* Skip unused frames from start of page */ 381 page += offset >> PAGE_SHIFT; 382 offset &= ~PAGE_MASK; 383 384 while (size > 0) { 385 BUG_ON(offset >= PAGE_SIZE); 386 387 bytes = PAGE_SIZE - offset; 388 if (bytes > size) 389 bytes = size; 390 391 info.page = page; 392 gnttab_foreach_grant_in_range(page, offset, bytes, 393 xenvif_gop_frag_copy_grant, 394 &info); 395 size -= bytes; 396 offset = 0; 397 398 /* Next page */ 399 if (size) { 400 BUG_ON(!PageCompound(page)); 401 page++; 402 } 403 } 404 405 *head = info.head; 406 } 407 408 /* 409 * Prepare an SKB to be transmitted to the frontend. 410 * 411 * This function is responsible for allocating grant operations, meta 412 * structures, etc. 413 * 414 * It returns the number of meta structures consumed. The number of 415 * ring slots used is always equal to the number of meta slots used 416 * plus the number of GSO descriptors used. Currently, we use either 417 * zero GSO descriptors (for non-GSO packets) or one descriptor (for 418 * frontend-side LRO). 419 */ 420 static int xenvif_gop_skb(struct sk_buff *skb, 421 struct netrx_pending_operations *npo, 422 struct xenvif_queue *queue) 423 { 424 struct xenvif *vif = netdev_priv(skb->dev); 425 int nr_frags = skb_shinfo(skb)->nr_frags; 426 int i; 427 struct xen_netif_rx_request req; 428 struct xenvif_rx_meta *meta; 429 unsigned char *data; 430 int head = 1; 431 int old_meta_prod; 432 int gso_type; 433 434 old_meta_prod = npo->meta_prod; 435 436 gso_type = XEN_NETIF_GSO_TYPE_NONE; 437 if (skb_is_gso(skb)) { 438 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) 439 gso_type = XEN_NETIF_GSO_TYPE_TCPV4; 440 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) 441 gso_type = XEN_NETIF_GSO_TYPE_TCPV6; 442 } 443 444 /* Set up a GSO prefix descriptor, if necessary */ 445 if ((1 << gso_type) & vif->gso_prefix_mask) { 446 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req); 447 meta = npo->meta + npo->meta_prod++; 448 meta->gso_type = gso_type; 449 meta->gso_size = skb_shinfo(skb)->gso_size; 450 meta->size = 0; 451 meta->id = req.id; 452 } 453 454 RING_COPY_REQUEST(&queue->rx, queue->rx.req_cons++, &req); 455 meta = npo->meta + npo->meta_prod++; 456 457 if ((1 << gso_type) & vif->gso_mask) { 458 meta->gso_type = gso_type; 459 meta->gso_size = skb_shinfo(skb)->gso_size; 460 } else { 461 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE; 462 meta->gso_size = 0; 463 } 464 465 meta->size = 0; 466 meta->id = req.id; 467 npo->copy_off = 0; 468 npo->copy_gref = req.gref; 469 470 data = skb->data; 471 while (data < skb_tail_pointer(skb)) { 472 unsigned int offset = offset_in_page(data); 473 unsigned int len = PAGE_SIZE - offset; 474 475 if (data + len > skb_tail_pointer(skb)) 476 len = skb_tail_pointer(skb) - data; 477 478 xenvif_gop_frag_copy(queue, skb, npo, 479 virt_to_page(data), len, offset, &head); 480 data += len; 481 } 482 483 for (i = 0; i < nr_frags; i++) { 484 xenvif_gop_frag_copy(queue, skb, npo, 485 skb_frag_page(&skb_shinfo(skb)->frags[i]), 486 skb_frag_size(&skb_shinfo(skb)->frags[i]), 487 skb_shinfo(skb)->frags[i].page_offset, 488 &head); 489 } 490 491 return npo->meta_prod - old_meta_prod; 492 } 493 494 /* 495 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was 496 * used to set up the operations on the top of 497 * netrx_pending_operations, which have since been done. Check that 498 * they didn't give any errors and advance over them. 499 */ 500 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots, 501 struct netrx_pending_operations *npo) 502 { 503 struct gnttab_copy *copy_op; 504 int status = XEN_NETIF_RSP_OKAY; 505 int i; 506 507 for (i = 0; i < nr_meta_slots; i++) { 508 copy_op = npo->copy + npo->copy_cons++; 509 if (copy_op->status != GNTST_okay) { 510 netdev_dbg(vif->dev, 511 "Bad status %d from copy to DOM%d.\n", 512 copy_op->status, vif->domid); 513 status = XEN_NETIF_RSP_ERROR; 514 } 515 } 516 517 return status; 518 } 519 520 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status, 521 struct xenvif_rx_meta *meta, 522 int nr_meta_slots) 523 { 524 int i; 525 unsigned long offset; 526 527 /* No fragments used */ 528 if (nr_meta_slots <= 1) 529 return; 530 531 nr_meta_slots--; 532 533 for (i = 0; i < nr_meta_slots; i++) { 534 int flags; 535 if (i == nr_meta_slots - 1) 536 flags = 0; 537 else 538 flags = XEN_NETRXF_more_data; 539 540 offset = 0; 541 make_rx_response(queue, meta[i].id, status, offset, 542 meta[i].size, flags); 543 } 544 } 545 546 void xenvif_kick_thread(struct xenvif_queue *queue) 547 { 548 wake_up(&queue->wq); 549 } 550 551 static void xenvif_rx_action(struct xenvif_queue *queue) 552 { 553 s8 status; 554 u16 flags; 555 struct xen_netif_rx_response *resp; 556 struct sk_buff_head rxq; 557 struct sk_buff *skb; 558 LIST_HEAD(notify); 559 int ret; 560 unsigned long offset; 561 bool need_to_notify = false; 562 563 struct netrx_pending_operations npo = { 564 .copy = queue->grant_copy_op, 565 .meta = queue->meta, 566 }; 567 568 skb_queue_head_init(&rxq); 569 570 while (xenvif_rx_ring_slots_available(queue) 571 && (skb = xenvif_rx_dequeue(queue)) != NULL) { 572 queue->last_rx_time = jiffies; 573 574 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue); 575 576 __skb_queue_tail(&rxq, skb); 577 } 578 579 BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta)); 580 581 if (!npo.copy_prod) 582 goto done; 583 584 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS); 585 gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod); 586 587 while ((skb = __skb_dequeue(&rxq)) != NULL) { 588 589 if ((1 << queue->meta[npo.meta_cons].gso_type) & 590 queue->vif->gso_prefix_mask) { 591 resp = RING_GET_RESPONSE(&queue->rx, 592 queue->rx.rsp_prod_pvt++); 593 594 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data; 595 596 resp->offset = queue->meta[npo.meta_cons].gso_size; 597 resp->id = queue->meta[npo.meta_cons].id; 598 resp->status = XENVIF_RX_CB(skb)->meta_slots_used; 599 600 npo.meta_cons++; 601 XENVIF_RX_CB(skb)->meta_slots_used--; 602 } 603 604 605 queue->stats.tx_bytes += skb->len; 606 queue->stats.tx_packets++; 607 608 status = xenvif_check_gop(queue->vif, 609 XENVIF_RX_CB(skb)->meta_slots_used, 610 &npo); 611 612 if (XENVIF_RX_CB(skb)->meta_slots_used == 1) 613 flags = 0; 614 else 615 flags = XEN_NETRXF_more_data; 616 617 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */ 618 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated; 619 else if (skb->ip_summed == CHECKSUM_UNNECESSARY) 620 /* remote but checksummed. */ 621 flags |= XEN_NETRXF_data_validated; 622 623 offset = 0; 624 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id, 625 status, offset, 626 queue->meta[npo.meta_cons].size, 627 flags); 628 629 if ((1 << queue->meta[npo.meta_cons].gso_type) & 630 queue->vif->gso_mask) { 631 struct xen_netif_extra_info *gso = 632 (struct xen_netif_extra_info *) 633 RING_GET_RESPONSE(&queue->rx, 634 queue->rx.rsp_prod_pvt++); 635 636 resp->flags |= XEN_NETRXF_extra_info; 637 638 gso->u.gso.type = queue->meta[npo.meta_cons].gso_type; 639 gso->u.gso.size = queue->meta[npo.meta_cons].gso_size; 640 gso->u.gso.pad = 0; 641 gso->u.gso.features = 0; 642 643 gso->type = XEN_NETIF_EXTRA_TYPE_GSO; 644 gso->flags = 0; 645 } 646 647 xenvif_add_frag_responses(queue, status, 648 queue->meta + npo.meta_cons + 1, 649 XENVIF_RX_CB(skb)->meta_slots_used); 650 651 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret); 652 653 need_to_notify |= !!ret; 654 655 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used; 656 dev_kfree_skb(skb); 657 } 658 659 done: 660 if (need_to_notify) 661 notify_remote_via_irq(queue->rx_irq); 662 } 663 664 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue) 665 { 666 int more_to_do; 667 668 RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do); 669 670 if (more_to_do) 671 napi_schedule(&queue->napi); 672 } 673 674 static void tx_add_credit(struct xenvif_queue *queue) 675 { 676 unsigned long max_burst, max_credit; 677 678 /* 679 * Allow a burst big enough to transmit a jumbo packet of up to 128kB. 680 * Otherwise the interface can seize up due to insufficient credit. 681 */ 682 max_burst = max(131072UL, queue->credit_bytes); 683 684 /* Take care that adding a new chunk of credit doesn't wrap to zero. */ 685 max_credit = queue->remaining_credit + queue->credit_bytes; 686 if (max_credit < queue->remaining_credit) 687 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */ 688 689 queue->remaining_credit = min(max_credit, max_burst); 690 } 691 692 void xenvif_tx_credit_callback(unsigned long data) 693 { 694 struct xenvif_queue *queue = (struct xenvif_queue *)data; 695 tx_add_credit(queue); 696 xenvif_napi_schedule_or_enable_events(queue); 697 } 698 699 static void xenvif_tx_err(struct xenvif_queue *queue, 700 struct xen_netif_tx_request *txp, 701 unsigned int extra_count, RING_IDX end) 702 { 703 RING_IDX cons = queue->tx.req_cons; 704 unsigned long flags; 705 706 do { 707 spin_lock_irqsave(&queue->response_lock, flags); 708 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR); 709 push_tx_responses(queue); 710 spin_unlock_irqrestore(&queue->response_lock, flags); 711 if (cons == end) 712 break; 713 RING_COPY_REQUEST(&queue->tx, cons++, txp); 714 } while (1); 715 queue->tx.req_cons = cons; 716 } 717 718 static void xenvif_fatal_tx_err(struct xenvif *vif) 719 { 720 netdev_err(vif->dev, "fatal error; disabling device\n"); 721 vif->disabled = true; 722 /* Disable the vif from queue 0's kthread */ 723 if (vif->queues) 724 xenvif_kick_thread(&vif->queues[0]); 725 } 726 727 static int xenvif_count_requests(struct xenvif_queue *queue, 728 struct xen_netif_tx_request *first, 729 unsigned int extra_count, 730 struct xen_netif_tx_request *txp, 731 int work_to_do) 732 { 733 RING_IDX cons = queue->tx.req_cons; 734 int slots = 0; 735 int drop_err = 0; 736 int more_data; 737 738 if (!(first->flags & XEN_NETTXF_more_data)) 739 return 0; 740 741 do { 742 struct xen_netif_tx_request dropped_tx = { 0 }; 743 744 if (slots >= work_to_do) { 745 netdev_err(queue->vif->dev, 746 "Asked for %d slots but exceeds this limit\n", 747 work_to_do); 748 xenvif_fatal_tx_err(queue->vif); 749 return -ENODATA; 750 } 751 752 /* This guest is really using too many slots and 753 * considered malicious. 754 */ 755 if (unlikely(slots >= fatal_skb_slots)) { 756 netdev_err(queue->vif->dev, 757 "Malicious frontend using %d slots, threshold %u\n", 758 slots, fatal_skb_slots); 759 xenvif_fatal_tx_err(queue->vif); 760 return -E2BIG; 761 } 762 763 /* Xen network protocol had implicit dependency on 764 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to 765 * the historical MAX_SKB_FRAGS value 18 to honor the 766 * same behavior as before. Any packet using more than 767 * 18 slots but less than fatal_skb_slots slots is 768 * dropped 769 */ 770 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) { 771 if (net_ratelimit()) 772 netdev_dbg(queue->vif->dev, 773 "Too many slots (%d) exceeding limit (%d), dropping packet\n", 774 slots, XEN_NETBK_LEGACY_SLOTS_MAX); 775 drop_err = -E2BIG; 776 } 777 778 if (drop_err) 779 txp = &dropped_tx; 780 781 RING_COPY_REQUEST(&queue->tx, cons + slots, txp); 782 783 /* If the guest submitted a frame >= 64 KiB then 784 * first->size overflowed and following slots will 785 * appear to be larger than the frame. 786 * 787 * This cannot be fatal error as there are buggy 788 * frontends that do this. 789 * 790 * Consume all slots and drop the packet. 791 */ 792 if (!drop_err && txp->size > first->size) { 793 if (net_ratelimit()) 794 netdev_dbg(queue->vif->dev, 795 "Invalid tx request, slot size %u > remaining size %u\n", 796 txp->size, first->size); 797 drop_err = -EIO; 798 } 799 800 first->size -= txp->size; 801 slots++; 802 803 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) { 804 netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n", 805 txp->offset, txp->size); 806 xenvif_fatal_tx_err(queue->vif); 807 return -EINVAL; 808 } 809 810 more_data = txp->flags & XEN_NETTXF_more_data; 811 812 if (!drop_err) 813 txp++; 814 815 } while (more_data); 816 817 if (drop_err) { 818 xenvif_tx_err(queue, first, extra_count, cons + slots); 819 return drop_err; 820 } 821 822 return slots; 823 } 824 825 826 struct xenvif_tx_cb { 827 u16 pending_idx; 828 }; 829 830 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb) 831 832 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue, 833 u16 pending_idx, 834 struct xen_netif_tx_request *txp, 835 unsigned int extra_count, 836 struct gnttab_map_grant_ref *mop) 837 { 838 queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx]; 839 gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx), 840 GNTMAP_host_map | GNTMAP_readonly, 841 txp->gref, queue->vif->domid); 842 843 memcpy(&queue->pending_tx_info[pending_idx].req, txp, 844 sizeof(*txp)); 845 queue->pending_tx_info[pending_idx].extra_count = extra_count; 846 } 847 848 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size) 849 { 850 struct sk_buff *skb = 851 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN, 852 GFP_ATOMIC | __GFP_NOWARN); 853 if (unlikely(skb == NULL)) 854 return NULL; 855 856 /* Packets passed to netif_rx() must have some headroom. */ 857 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN); 858 859 /* Initialize it here to avoid later surprises */ 860 skb_shinfo(skb)->destructor_arg = NULL; 861 862 return skb; 863 } 864 865 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue, 866 struct sk_buff *skb, 867 struct xen_netif_tx_request *txp, 868 struct gnttab_map_grant_ref *gop, 869 unsigned int frag_overflow, 870 struct sk_buff *nskb) 871 { 872 struct skb_shared_info *shinfo = skb_shinfo(skb); 873 skb_frag_t *frags = shinfo->frags; 874 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx; 875 int start; 876 pending_ring_idx_t index; 877 unsigned int nr_slots; 878 879 nr_slots = shinfo->nr_frags; 880 881 /* Skip first skb fragment if it is on same page as header fragment. */ 882 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx); 883 884 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots; 885 shinfo->nr_frags++, txp++, gop++) { 886 index = pending_index(queue->pending_cons++); 887 pending_idx = queue->pending_ring[index]; 888 xenvif_tx_create_map_op(queue, pending_idx, txp, 0, gop); 889 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx); 890 } 891 892 if (frag_overflow) { 893 894 shinfo = skb_shinfo(nskb); 895 frags = shinfo->frags; 896 897 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow; 898 shinfo->nr_frags++, txp++, gop++) { 899 index = pending_index(queue->pending_cons++); 900 pending_idx = queue->pending_ring[index]; 901 xenvif_tx_create_map_op(queue, pending_idx, txp, 0, 902 gop); 903 frag_set_pending_idx(&frags[shinfo->nr_frags], 904 pending_idx); 905 } 906 907 skb_shinfo(skb)->frag_list = nskb; 908 } 909 910 return gop; 911 } 912 913 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue, 914 u16 pending_idx, 915 grant_handle_t handle) 916 { 917 if (unlikely(queue->grant_tx_handle[pending_idx] != 918 NETBACK_INVALID_HANDLE)) { 919 netdev_err(queue->vif->dev, 920 "Trying to overwrite active handle! pending_idx: 0x%x\n", 921 pending_idx); 922 BUG(); 923 } 924 queue->grant_tx_handle[pending_idx] = handle; 925 } 926 927 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue, 928 u16 pending_idx) 929 { 930 if (unlikely(queue->grant_tx_handle[pending_idx] == 931 NETBACK_INVALID_HANDLE)) { 932 netdev_err(queue->vif->dev, 933 "Trying to unmap invalid handle! pending_idx: 0x%x\n", 934 pending_idx); 935 BUG(); 936 } 937 queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE; 938 } 939 940 static int xenvif_tx_check_gop(struct xenvif_queue *queue, 941 struct sk_buff *skb, 942 struct gnttab_map_grant_ref **gopp_map, 943 struct gnttab_copy **gopp_copy) 944 { 945 struct gnttab_map_grant_ref *gop_map = *gopp_map; 946 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx; 947 /* This always points to the shinfo of the skb being checked, which 948 * could be either the first or the one on the frag_list 949 */ 950 struct skb_shared_info *shinfo = skb_shinfo(skb); 951 /* If this is non-NULL, we are currently checking the frag_list skb, and 952 * this points to the shinfo of the first one 953 */ 954 struct skb_shared_info *first_shinfo = NULL; 955 int nr_frags = shinfo->nr_frags; 956 const bool sharedslot = nr_frags && 957 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx; 958 int i, err; 959 960 /* Check status of header. */ 961 err = (*gopp_copy)->status; 962 if (unlikely(err)) { 963 if (net_ratelimit()) 964 netdev_dbg(queue->vif->dev, 965 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n", 966 (*gopp_copy)->status, 967 pending_idx, 968 (*gopp_copy)->source.u.ref); 969 /* The first frag might still have this slot mapped */ 970 if (!sharedslot) 971 xenvif_idx_release(queue, pending_idx, 972 XEN_NETIF_RSP_ERROR); 973 } 974 (*gopp_copy)++; 975 976 check_frags: 977 for (i = 0; i < nr_frags; i++, gop_map++) { 978 int j, newerr; 979 980 pending_idx = frag_get_pending_idx(&shinfo->frags[i]); 981 982 /* Check error status: if okay then remember grant handle. */ 983 newerr = gop_map->status; 984 985 if (likely(!newerr)) { 986 xenvif_grant_handle_set(queue, 987 pending_idx, 988 gop_map->handle); 989 /* Had a previous error? Invalidate this fragment. */ 990 if (unlikely(err)) { 991 xenvif_idx_unmap(queue, pending_idx); 992 /* If the mapping of the first frag was OK, but 993 * the header's copy failed, and they are 994 * sharing a slot, send an error 995 */ 996 if (i == 0 && sharedslot) 997 xenvif_idx_release(queue, pending_idx, 998 XEN_NETIF_RSP_ERROR); 999 else 1000 xenvif_idx_release(queue, pending_idx, 1001 XEN_NETIF_RSP_OKAY); 1002 } 1003 continue; 1004 } 1005 1006 /* Error on this fragment: respond to client with an error. */ 1007 if (net_ratelimit()) 1008 netdev_dbg(queue->vif->dev, 1009 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n", 1010 i, 1011 gop_map->status, 1012 pending_idx, 1013 gop_map->ref); 1014 1015 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR); 1016 1017 /* Not the first error? Preceding frags already invalidated. */ 1018 if (err) 1019 continue; 1020 1021 /* First error: if the header haven't shared a slot with the 1022 * first frag, release it as well. 1023 */ 1024 if (!sharedslot) 1025 xenvif_idx_release(queue, 1026 XENVIF_TX_CB(skb)->pending_idx, 1027 XEN_NETIF_RSP_OKAY); 1028 1029 /* Invalidate preceding fragments of this skb. */ 1030 for (j = 0; j < i; j++) { 1031 pending_idx = frag_get_pending_idx(&shinfo->frags[j]); 1032 xenvif_idx_unmap(queue, pending_idx); 1033 xenvif_idx_release(queue, pending_idx, 1034 XEN_NETIF_RSP_OKAY); 1035 } 1036 1037 /* And if we found the error while checking the frag_list, unmap 1038 * the first skb's frags 1039 */ 1040 if (first_shinfo) { 1041 for (j = 0; j < first_shinfo->nr_frags; j++) { 1042 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]); 1043 xenvif_idx_unmap(queue, pending_idx); 1044 xenvif_idx_release(queue, pending_idx, 1045 XEN_NETIF_RSP_OKAY); 1046 } 1047 } 1048 1049 /* Remember the error: invalidate all subsequent fragments. */ 1050 err = newerr; 1051 } 1052 1053 if (skb_has_frag_list(skb) && !first_shinfo) { 1054 first_shinfo = skb_shinfo(skb); 1055 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list); 1056 nr_frags = shinfo->nr_frags; 1057 1058 goto check_frags; 1059 } 1060 1061 *gopp_map = gop_map; 1062 return err; 1063 } 1064 1065 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb) 1066 { 1067 struct skb_shared_info *shinfo = skb_shinfo(skb); 1068 int nr_frags = shinfo->nr_frags; 1069 int i; 1070 u16 prev_pending_idx = INVALID_PENDING_IDX; 1071 1072 for (i = 0; i < nr_frags; i++) { 1073 skb_frag_t *frag = shinfo->frags + i; 1074 struct xen_netif_tx_request *txp; 1075 struct page *page; 1076 u16 pending_idx; 1077 1078 pending_idx = frag_get_pending_idx(frag); 1079 1080 /* If this is not the first frag, chain it to the previous*/ 1081 if (prev_pending_idx == INVALID_PENDING_IDX) 1082 skb_shinfo(skb)->destructor_arg = 1083 &callback_param(queue, pending_idx); 1084 else 1085 callback_param(queue, prev_pending_idx).ctx = 1086 &callback_param(queue, pending_idx); 1087 1088 callback_param(queue, pending_idx).ctx = NULL; 1089 prev_pending_idx = pending_idx; 1090 1091 txp = &queue->pending_tx_info[pending_idx].req; 1092 page = virt_to_page(idx_to_kaddr(queue, pending_idx)); 1093 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size); 1094 skb->len += txp->size; 1095 skb->data_len += txp->size; 1096 skb->truesize += txp->size; 1097 1098 /* Take an extra reference to offset network stack's put_page */ 1099 get_page(queue->mmap_pages[pending_idx]); 1100 } 1101 } 1102 1103 static int xenvif_get_extras(struct xenvif_queue *queue, 1104 struct xen_netif_extra_info *extras, 1105 unsigned int *extra_count, 1106 int work_to_do) 1107 { 1108 struct xen_netif_extra_info extra; 1109 RING_IDX cons = queue->tx.req_cons; 1110 1111 do { 1112 if (unlikely(work_to_do-- <= 0)) { 1113 netdev_err(queue->vif->dev, "Missing extra info\n"); 1114 xenvif_fatal_tx_err(queue->vif); 1115 return -EBADR; 1116 } 1117 1118 RING_COPY_REQUEST(&queue->tx, cons, &extra); 1119 1120 queue->tx.req_cons = ++cons; 1121 (*extra_count)++; 1122 1123 if (unlikely(!extra.type || 1124 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) { 1125 netdev_err(queue->vif->dev, 1126 "Invalid extra type: %d\n", extra.type); 1127 xenvif_fatal_tx_err(queue->vif); 1128 return -EINVAL; 1129 } 1130 1131 memcpy(&extras[extra.type - 1], &extra, sizeof(extra)); 1132 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE); 1133 1134 return work_to_do; 1135 } 1136 1137 static int xenvif_set_skb_gso(struct xenvif *vif, 1138 struct sk_buff *skb, 1139 struct xen_netif_extra_info *gso) 1140 { 1141 if (!gso->u.gso.size) { 1142 netdev_err(vif->dev, "GSO size must not be zero.\n"); 1143 xenvif_fatal_tx_err(vif); 1144 return -EINVAL; 1145 } 1146 1147 switch (gso->u.gso.type) { 1148 case XEN_NETIF_GSO_TYPE_TCPV4: 1149 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4; 1150 break; 1151 case XEN_NETIF_GSO_TYPE_TCPV6: 1152 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6; 1153 break; 1154 default: 1155 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type); 1156 xenvif_fatal_tx_err(vif); 1157 return -EINVAL; 1158 } 1159 1160 skb_shinfo(skb)->gso_size = gso->u.gso.size; 1161 /* gso_segs will be calculated later */ 1162 1163 return 0; 1164 } 1165 1166 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb) 1167 { 1168 bool recalculate_partial_csum = false; 1169 1170 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy 1171 * peers can fail to set NETRXF_csum_blank when sending a GSO 1172 * frame. In this case force the SKB to CHECKSUM_PARTIAL and 1173 * recalculate the partial checksum. 1174 */ 1175 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) { 1176 queue->stats.rx_gso_checksum_fixup++; 1177 skb->ip_summed = CHECKSUM_PARTIAL; 1178 recalculate_partial_csum = true; 1179 } 1180 1181 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */ 1182 if (skb->ip_summed != CHECKSUM_PARTIAL) 1183 return 0; 1184 1185 return skb_checksum_setup(skb, recalculate_partial_csum); 1186 } 1187 1188 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size) 1189 { 1190 u64 now = get_jiffies_64(); 1191 u64 next_credit = queue->credit_window_start + 1192 msecs_to_jiffies(queue->credit_usec / 1000); 1193 1194 /* Timer could already be pending in rare cases. */ 1195 if (timer_pending(&queue->credit_timeout)) 1196 return true; 1197 1198 /* Passed the point where we can replenish credit? */ 1199 if (time_after_eq64(now, next_credit)) { 1200 queue->credit_window_start = now; 1201 tx_add_credit(queue); 1202 } 1203 1204 /* Still too big to send right now? Set a callback. */ 1205 if (size > queue->remaining_credit) { 1206 queue->credit_timeout.data = 1207 (unsigned long)queue; 1208 mod_timer(&queue->credit_timeout, 1209 next_credit); 1210 queue->credit_window_start = next_credit; 1211 1212 return true; 1213 } 1214 1215 return false; 1216 } 1217 1218 /* No locking is required in xenvif_mcast_add/del() as they are 1219 * only ever invoked from NAPI poll. An RCU list is used because 1220 * xenvif_mcast_match() is called asynchronously, during start_xmit. 1221 */ 1222 1223 static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr) 1224 { 1225 struct xenvif_mcast_addr *mcast; 1226 1227 if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) { 1228 if (net_ratelimit()) 1229 netdev_err(vif->dev, 1230 "Too many multicast addresses\n"); 1231 return -ENOSPC; 1232 } 1233 1234 mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC); 1235 if (!mcast) 1236 return -ENOMEM; 1237 1238 ether_addr_copy(mcast->addr, addr); 1239 list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr); 1240 vif->fe_mcast_count++; 1241 1242 return 0; 1243 } 1244 1245 static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr) 1246 { 1247 struct xenvif_mcast_addr *mcast; 1248 1249 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) { 1250 if (ether_addr_equal(addr, mcast->addr)) { 1251 --vif->fe_mcast_count; 1252 list_del_rcu(&mcast->entry); 1253 kfree_rcu(mcast, rcu); 1254 break; 1255 } 1256 } 1257 } 1258 1259 bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr) 1260 { 1261 struct xenvif_mcast_addr *mcast; 1262 1263 rcu_read_lock(); 1264 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) { 1265 if (ether_addr_equal(addr, mcast->addr)) { 1266 rcu_read_unlock(); 1267 return true; 1268 } 1269 } 1270 rcu_read_unlock(); 1271 1272 return false; 1273 } 1274 1275 void xenvif_mcast_addr_list_free(struct xenvif *vif) 1276 { 1277 /* No need for locking or RCU here. NAPI poll and TX queue 1278 * are stopped. 1279 */ 1280 while (!list_empty(&vif->fe_mcast_addr)) { 1281 struct xenvif_mcast_addr *mcast; 1282 1283 mcast = list_first_entry(&vif->fe_mcast_addr, 1284 struct xenvif_mcast_addr, 1285 entry); 1286 --vif->fe_mcast_count; 1287 list_del(&mcast->entry); 1288 kfree(mcast); 1289 } 1290 } 1291 1292 static void xenvif_tx_build_gops(struct xenvif_queue *queue, 1293 int budget, 1294 unsigned *copy_ops, 1295 unsigned *map_ops) 1296 { 1297 struct gnttab_map_grant_ref *gop = queue->tx_map_ops; 1298 struct sk_buff *skb, *nskb; 1299 int ret; 1300 unsigned int frag_overflow; 1301 1302 while (skb_queue_len(&queue->tx_queue) < budget) { 1303 struct xen_netif_tx_request txreq; 1304 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX]; 1305 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1]; 1306 unsigned int extra_count; 1307 u16 pending_idx; 1308 RING_IDX idx; 1309 int work_to_do; 1310 unsigned int data_len; 1311 pending_ring_idx_t index; 1312 1313 if (queue->tx.sring->req_prod - queue->tx.req_cons > 1314 XEN_NETIF_TX_RING_SIZE) { 1315 netdev_err(queue->vif->dev, 1316 "Impossible number of requests. " 1317 "req_prod %d, req_cons %d, size %ld\n", 1318 queue->tx.sring->req_prod, queue->tx.req_cons, 1319 XEN_NETIF_TX_RING_SIZE); 1320 xenvif_fatal_tx_err(queue->vif); 1321 break; 1322 } 1323 1324 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx); 1325 if (!work_to_do) 1326 break; 1327 1328 idx = queue->tx.req_cons; 1329 rmb(); /* Ensure that we see the request before we copy it. */ 1330 RING_COPY_REQUEST(&queue->tx, idx, &txreq); 1331 1332 /* Credit-based scheduling. */ 1333 if (txreq.size > queue->remaining_credit && 1334 tx_credit_exceeded(queue, txreq.size)) 1335 break; 1336 1337 queue->remaining_credit -= txreq.size; 1338 1339 work_to_do--; 1340 queue->tx.req_cons = ++idx; 1341 1342 memset(extras, 0, sizeof(extras)); 1343 extra_count = 0; 1344 if (txreq.flags & XEN_NETTXF_extra_info) { 1345 work_to_do = xenvif_get_extras(queue, extras, 1346 &extra_count, 1347 work_to_do); 1348 idx = queue->tx.req_cons; 1349 if (unlikely(work_to_do < 0)) 1350 break; 1351 } 1352 1353 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) { 1354 struct xen_netif_extra_info *extra; 1355 1356 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1]; 1357 ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr); 1358 1359 make_tx_response(queue, &txreq, extra_count, 1360 (ret == 0) ? 1361 XEN_NETIF_RSP_OKAY : 1362 XEN_NETIF_RSP_ERROR); 1363 push_tx_responses(queue); 1364 continue; 1365 } 1366 1367 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) { 1368 struct xen_netif_extra_info *extra; 1369 1370 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1]; 1371 xenvif_mcast_del(queue->vif, extra->u.mcast.addr); 1372 1373 make_tx_response(queue, &txreq, extra_count, 1374 XEN_NETIF_RSP_OKAY); 1375 push_tx_responses(queue); 1376 continue; 1377 } 1378 1379 ret = xenvif_count_requests(queue, &txreq, extra_count, 1380 txfrags, work_to_do); 1381 if (unlikely(ret < 0)) 1382 break; 1383 1384 idx += ret; 1385 1386 if (unlikely(txreq.size < ETH_HLEN)) { 1387 netdev_dbg(queue->vif->dev, 1388 "Bad packet size: %d\n", txreq.size); 1389 xenvif_tx_err(queue, &txreq, extra_count, idx); 1390 break; 1391 } 1392 1393 /* No crossing a page as the payload mustn't fragment. */ 1394 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) { 1395 netdev_err(queue->vif->dev, 1396 "txreq.offset: %u, size: %u, end: %lu\n", 1397 txreq.offset, txreq.size, 1398 (unsigned long)(txreq.offset&~XEN_PAGE_MASK) + txreq.size); 1399 xenvif_fatal_tx_err(queue->vif); 1400 break; 1401 } 1402 1403 index = pending_index(queue->pending_cons); 1404 pending_idx = queue->pending_ring[index]; 1405 1406 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN && 1407 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ? 1408 XEN_NETBACK_TX_COPY_LEN : txreq.size; 1409 1410 skb = xenvif_alloc_skb(data_len); 1411 if (unlikely(skb == NULL)) { 1412 netdev_dbg(queue->vif->dev, 1413 "Can't allocate a skb in start_xmit.\n"); 1414 xenvif_tx_err(queue, &txreq, extra_count, idx); 1415 break; 1416 } 1417 1418 skb_shinfo(skb)->nr_frags = ret; 1419 if (data_len < txreq.size) 1420 skb_shinfo(skb)->nr_frags++; 1421 /* At this point shinfo->nr_frags is in fact the number of 1422 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX. 1423 */ 1424 frag_overflow = 0; 1425 nskb = NULL; 1426 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) { 1427 frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS; 1428 BUG_ON(frag_overflow > MAX_SKB_FRAGS); 1429 skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS; 1430 nskb = xenvif_alloc_skb(0); 1431 if (unlikely(nskb == NULL)) { 1432 kfree_skb(skb); 1433 xenvif_tx_err(queue, &txreq, extra_count, idx); 1434 if (net_ratelimit()) 1435 netdev_err(queue->vif->dev, 1436 "Can't allocate the frag_list skb.\n"); 1437 break; 1438 } 1439 } 1440 1441 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) { 1442 struct xen_netif_extra_info *gso; 1443 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1]; 1444 1445 if (xenvif_set_skb_gso(queue->vif, skb, gso)) { 1446 /* Failure in xenvif_set_skb_gso is fatal. */ 1447 kfree_skb(skb); 1448 kfree_skb(nskb); 1449 break; 1450 } 1451 } 1452 1453 XENVIF_TX_CB(skb)->pending_idx = pending_idx; 1454 1455 __skb_put(skb, data_len); 1456 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref; 1457 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid; 1458 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset; 1459 1460 queue->tx_copy_ops[*copy_ops].dest.u.gmfn = 1461 virt_to_gfn(skb->data); 1462 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF; 1463 queue->tx_copy_ops[*copy_ops].dest.offset = 1464 offset_in_page(skb->data) & ~XEN_PAGE_MASK; 1465 1466 queue->tx_copy_ops[*copy_ops].len = data_len; 1467 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref; 1468 1469 (*copy_ops)++; 1470 1471 if (data_len < txreq.size) { 1472 frag_set_pending_idx(&skb_shinfo(skb)->frags[0], 1473 pending_idx); 1474 xenvif_tx_create_map_op(queue, pending_idx, &txreq, 1475 extra_count, gop); 1476 gop++; 1477 } else { 1478 frag_set_pending_idx(&skb_shinfo(skb)->frags[0], 1479 INVALID_PENDING_IDX); 1480 memcpy(&queue->pending_tx_info[pending_idx].req, 1481 &txreq, sizeof(txreq)); 1482 queue->pending_tx_info[pending_idx].extra_count = 1483 extra_count; 1484 } 1485 1486 queue->pending_cons++; 1487 1488 gop = xenvif_get_requests(queue, skb, txfrags, gop, 1489 frag_overflow, nskb); 1490 1491 __skb_queue_tail(&queue->tx_queue, skb); 1492 1493 queue->tx.req_cons = idx; 1494 1495 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) || 1496 (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops))) 1497 break; 1498 } 1499 1500 (*map_ops) = gop - queue->tx_map_ops; 1501 return; 1502 } 1503 1504 /* Consolidate skb with a frag_list into a brand new one with local pages on 1505 * frags. Returns 0 or -ENOMEM if can't allocate new pages. 1506 */ 1507 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb) 1508 { 1509 unsigned int offset = skb_headlen(skb); 1510 skb_frag_t frags[MAX_SKB_FRAGS]; 1511 int i, f; 1512 struct ubuf_info *uarg; 1513 struct sk_buff *nskb = skb_shinfo(skb)->frag_list; 1514 1515 queue->stats.tx_zerocopy_sent += 2; 1516 queue->stats.tx_frag_overflow++; 1517 1518 xenvif_fill_frags(queue, nskb); 1519 /* Subtract frags size, we will correct it later */ 1520 skb->truesize -= skb->data_len; 1521 skb->len += nskb->len; 1522 skb->data_len += nskb->len; 1523 1524 /* create a brand new frags array and coalesce there */ 1525 for (i = 0; offset < skb->len; i++) { 1526 struct page *page; 1527 unsigned int len; 1528 1529 BUG_ON(i >= MAX_SKB_FRAGS); 1530 page = alloc_page(GFP_ATOMIC); 1531 if (!page) { 1532 int j; 1533 skb->truesize += skb->data_len; 1534 for (j = 0; j < i; j++) 1535 put_page(frags[j].page.p); 1536 return -ENOMEM; 1537 } 1538 1539 if (offset + PAGE_SIZE < skb->len) 1540 len = PAGE_SIZE; 1541 else 1542 len = skb->len - offset; 1543 if (skb_copy_bits(skb, offset, page_address(page), len)) 1544 BUG(); 1545 1546 offset += len; 1547 frags[i].page.p = page; 1548 frags[i].page_offset = 0; 1549 skb_frag_size_set(&frags[i], len); 1550 } 1551 1552 /* Copied all the bits from the frag list -- free it. */ 1553 skb_frag_list_init(skb); 1554 xenvif_skb_zerocopy_prepare(queue, nskb); 1555 kfree_skb(nskb); 1556 1557 /* Release all the original (foreign) frags. */ 1558 for (f = 0; f < skb_shinfo(skb)->nr_frags; f++) 1559 skb_frag_unref(skb, f); 1560 uarg = skb_shinfo(skb)->destructor_arg; 1561 /* increase inflight counter to offset decrement in callback */ 1562 atomic_inc(&queue->inflight_packets); 1563 uarg->callback(uarg, true); 1564 skb_shinfo(skb)->destructor_arg = NULL; 1565 1566 /* Fill the skb with the new (local) frags. */ 1567 memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t)); 1568 skb_shinfo(skb)->nr_frags = i; 1569 skb->truesize += i * PAGE_SIZE; 1570 1571 return 0; 1572 } 1573 1574 static int xenvif_tx_submit(struct xenvif_queue *queue) 1575 { 1576 struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops; 1577 struct gnttab_copy *gop_copy = queue->tx_copy_ops; 1578 struct sk_buff *skb; 1579 int work_done = 0; 1580 1581 while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) { 1582 struct xen_netif_tx_request *txp; 1583 u16 pending_idx; 1584 unsigned data_len; 1585 1586 pending_idx = XENVIF_TX_CB(skb)->pending_idx; 1587 txp = &queue->pending_tx_info[pending_idx].req; 1588 1589 /* Check the remap error code. */ 1590 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) { 1591 /* If there was an error, xenvif_tx_check_gop is 1592 * expected to release all the frags which were mapped, 1593 * so kfree_skb shouldn't do it again 1594 */ 1595 skb_shinfo(skb)->nr_frags = 0; 1596 if (skb_has_frag_list(skb)) { 1597 struct sk_buff *nskb = 1598 skb_shinfo(skb)->frag_list; 1599 skb_shinfo(nskb)->nr_frags = 0; 1600 } 1601 kfree_skb(skb); 1602 continue; 1603 } 1604 1605 data_len = skb->len; 1606 callback_param(queue, pending_idx).ctx = NULL; 1607 if (data_len < txp->size) { 1608 /* Append the packet payload as a fragment. */ 1609 txp->offset += data_len; 1610 txp->size -= data_len; 1611 } else { 1612 /* Schedule a response immediately. */ 1613 xenvif_idx_release(queue, pending_idx, 1614 XEN_NETIF_RSP_OKAY); 1615 } 1616 1617 if (txp->flags & XEN_NETTXF_csum_blank) 1618 skb->ip_summed = CHECKSUM_PARTIAL; 1619 else if (txp->flags & XEN_NETTXF_data_validated) 1620 skb->ip_summed = CHECKSUM_UNNECESSARY; 1621 1622 xenvif_fill_frags(queue, skb); 1623 1624 if (unlikely(skb_has_frag_list(skb))) { 1625 if (xenvif_handle_frag_list(queue, skb)) { 1626 if (net_ratelimit()) 1627 netdev_err(queue->vif->dev, 1628 "Not enough memory to consolidate frag_list!\n"); 1629 xenvif_skb_zerocopy_prepare(queue, skb); 1630 kfree_skb(skb); 1631 continue; 1632 } 1633 } 1634 1635 skb->dev = queue->vif->dev; 1636 skb->protocol = eth_type_trans(skb, skb->dev); 1637 skb_reset_network_header(skb); 1638 1639 if (checksum_setup(queue, skb)) { 1640 netdev_dbg(queue->vif->dev, 1641 "Can't setup checksum in net_tx_action\n"); 1642 /* We have to set this flag to trigger the callback */ 1643 if (skb_shinfo(skb)->destructor_arg) 1644 xenvif_skb_zerocopy_prepare(queue, skb); 1645 kfree_skb(skb); 1646 continue; 1647 } 1648 1649 skb_probe_transport_header(skb, 0); 1650 1651 /* If the packet is GSO then we will have just set up the 1652 * transport header offset in checksum_setup so it's now 1653 * straightforward to calculate gso_segs. 1654 */ 1655 if (skb_is_gso(skb)) { 1656 int mss = skb_shinfo(skb)->gso_size; 1657 int hdrlen = skb_transport_header(skb) - 1658 skb_mac_header(skb) + 1659 tcp_hdrlen(skb); 1660 1661 skb_shinfo(skb)->gso_segs = 1662 DIV_ROUND_UP(skb->len - hdrlen, mss); 1663 } 1664 1665 queue->stats.rx_bytes += skb->len; 1666 queue->stats.rx_packets++; 1667 1668 work_done++; 1669 1670 /* Set this flag right before netif_receive_skb, otherwise 1671 * someone might think this packet already left netback, and 1672 * do a skb_copy_ubufs while we are still in control of the 1673 * skb. E.g. the __pskb_pull_tail earlier can do such thing. 1674 */ 1675 if (skb_shinfo(skb)->destructor_arg) { 1676 xenvif_skb_zerocopy_prepare(queue, skb); 1677 queue->stats.tx_zerocopy_sent++; 1678 } 1679 1680 netif_receive_skb(skb); 1681 } 1682 1683 return work_done; 1684 } 1685 1686 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success) 1687 { 1688 unsigned long flags; 1689 pending_ring_idx_t index; 1690 struct xenvif_queue *queue = ubuf_to_queue(ubuf); 1691 1692 /* This is the only place where we grab this lock, to protect callbacks 1693 * from each other. 1694 */ 1695 spin_lock_irqsave(&queue->callback_lock, flags); 1696 do { 1697 u16 pending_idx = ubuf->desc; 1698 ubuf = (struct ubuf_info *) ubuf->ctx; 1699 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >= 1700 MAX_PENDING_REQS); 1701 index = pending_index(queue->dealloc_prod); 1702 queue->dealloc_ring[index] = pending_idx; 1703 /* Sync with xenvif_tx_dealloc_action: 1704 * insert idx then incr producer. 1705 */ 1706 smp_wmb(); 1707 queue->dealloc_prod++; 1708 } while (ubuf); 1709 spin_unlock_irqrestore(&queue->callback_lock, flags); 1710 1711 if (likely(zerocopy_success)) 1712 queue->stats.tx_zerocopy_success++; 1713 else 1714 queue->stats.tx_zerocopy_fail++; 1715 xenvif_skb_zerocopy_complete(queue); 1716 } 1717 1718 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue) 1719 { 1720 struct gnttab_unmap_grant_ref *gop; 1721 pending_ring_idx_t dc, dp; 1722 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS]; 1723 unsigned int i = 0; 1724 1725 dc = queue->dealloc_cons; 1726 gop = queue->tx_unmap_ops; 1727 1728 /* Free up any grants we have finished using */ 1729 do { 1730 dp = queue->dealloc_prod; 1731 1732 /* Ensure we see all indices enqueued by all 1733 * xenvif_zerocopy_callback(). 1734 */ 1735 smp_rmb(); 1736 1737 while (dc != dp) { 1738 BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS); 1739 pending_idx = 1740 queue->dealloc_ring[pending_index(dc++)]; 1741 1742 pending_idx_release[gop - queue->tx_unmap_ops] = 1743 pending_idx; 1744 queue->pages_to_unmap[gop - queue->tx_unmap_ops] = 1745 queue->mmap_pages[pending_idx]; 1746 gnttab_set_unmap_op(gop, 1747 idx_to_kaddr(queue, pending_idx), 1748 GNTMAP_host_map, 1749 queue->grant_tx_handle[pending_idx]); 1750 xenvif_grant_handle_reset(queue, pending_idx); 1751 ++gop; 1752 } 1753 1754 } while (dp != queue->dealloc_prod); 1755 1756 queue->dealloc_cons = dc; 1757 1758 if (gop - queue->tx_unmap_ops > 0) { 1759 int ret; 1760 ret = gnttab_unmap_refs(queue->tx_unmap_ops, 1761 NULL, 1762 queue->pages_to_unmap, 1763 gop - queue->tx_unmap_ops); 1764 if (ret) { 1765 netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n", 1766 gop - queue->tx_unmap_ops, ret); 1767 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) { 1768 if (gop[i].status != GNTST_okay) 1769 netdev_err(queue->vif->dev, 1770 " host_addr: 0x%llx handle: 0x%x status: %d\n", 1771 gop[i].host_addr, 1772 gop[i].handle, 1773 gop[i].status); 1774 } 1775 BUG(); 1776 } 1777 } 1778 1779 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) 1780 xenvif_idx_release(queue, pending_idx_release[i], 1781 XEN_NETIF_RSP_OKAY); 1782 } 1783 1784 1785 /* Called after netfront has transmitted */ 1786 int xenvif_tx_action(struct xenvif_queue *queue, int budget) 1787 { 1788 unsigned nr_mops, nr_cops = 0; 1789 int work_done, ret; 1790 1791 if (unlikely(!tx_work_todo(queue))) 1792 return 0; 1793 1794 xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops); 1795 1796 if (nr_cops == 0) 1797 return 0; 1798 1799 gnttab_batch_copy(queue->tx_copy_ops, nr_cops); 1800 if (nr_mops != 0) { 1801 ret = gnttab_map_refs(queue->tx_map_ops, 1802 NULL, 1803 queue->pages_to_map, 1804 nr_mops); 1805 BUG_ON(ret); 1806 } 1807 1808 work_done = xenvif_tx_submit(queue); 1809 1810 return work_done; 1811 } 1812 1813 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx, 1814 u8 status) 1815 { 1816 struct pending_tx_info *pending_tx_info; 1817 pending_ring_idx_t index; 1818 unsigned long flags; 1819 1820 pending_tx_info = &queue->pending_tx_info[pending_idx]; 1821 1822 spin_lock_irqsave(&queue->response_lock, flags); 1823 1824 make_tx_response(queue, &pending_tx_info->req, 1825 pending_tx_info->extra_count, status); 1826 1827 /* Release the pending index before pusing the Tx response so 1828 * its available before a new Tx request is pushed by the 1829 * frontend. 1830 */ 1831 index = pending_index(queue->pending_prod++); 1832 queue->pending_ring[index] = pending_idx; 1833 1834 push_tx_responses(queue); 1835 1836 spin_unlock_irqrestore(&queue->response_lock, flags); 1837 } 1838 1839 1840 static void make_tx_response(struct xenvif_queue *queue, 1841 struct xen_netif_tx_request *txp, 1842 unsigned int extra_count, 1843 s8 st) 1844 { 1845 RING_IDX i = queue->tx.rsp_prod_pvt; 1846 struct xen_netif_tx_response *resp; 1847 1848 resp = RING_GET_RESPONSE(&queue->tx, i); 1849 resp->id = txp->id; 1850 resp->status = st; 1851 1852 while (extra_count-- != 0) 1853 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL; 1854 1855 queue->tx.rsp_prod_pvt = ++i; 1856 } 1857 1858 static void push_tx_responses(struct xenvif_queue *queue) 1859 { 1860 int notify; 1861 1862 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify); 1863 if (notify) 1864 notify_remote_via_irq(queue->tx_irq); 1865 } 1866 1867 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue, 1868 u16 id, 1869 s8 st, 1870 u16 offset, 1871 u16 size, 1872 u16 flags) 1873 { 1874 RING_IDX i = queue->rx.rsp_prod_pvt; 1875 struct xen_netif_rx_response *resp; 1876 1877 resp = RING_GET_RESPONSE(&queue->rx, i); 1878 resp->offset = offset; 1879 resp->flags = flags; 1880 resp->id = id; 1881 resp->status = (s16)size; 1882 if (st < 0) 1883 resp->status = (s16)st; 1884 1885 queue->rx.rsp_prod_pvt = ++i; 1886 1887 return resp; 1888 } 1889 1890 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx) 1891 { 1892 int ret; 1893 struct gnttab_unmap_grant_ref tx_unmap_op; 1894 1895 gnttab_set_unmap_op(&tx_unmap_op, 1896 idx_to_kaddr(queue, pending_idx), 1897 GNTMAP_host_map, 1898 queue->grant_tx_handle[pending_idx]); 1899 xenvif_grant_handle_reset(queue, pending_idx); 1900 1901 ret = gnttab_unmap_refs(&tx_unmap_op, NULL, 1902 &queue->mmap_pages[pending_idx], 1); 1903 if (ret) { 1904 netdev_err(queue->vif->dev, 1905 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n", 1906 ret, 1907 pending_idx, 1908 tx_unmap_op.host_addr, 1909 tx_unmap_op.handle, 1910 tx_unmap_op.status); 1911 BUG(); 1912 } 1913 } 1914 1915 static inline int tx_work_todo(struct xenvif_queue *queue) 1916 { 1917 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx))) 1918 return 1; 1919 1920 return 0; 1921 } 1922 1923 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue) 1924 { 1925 return queue->dealloc_cons != queue->dealloc_prod; 1926 } 1927 1928 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue) 1929 { 1930 if (queue->tx.sring) 1931 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif), 1932 queue->tx.sring); 1933 if (queue->rx.sring) 1934 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif), 1935 queue->rx.sring); 1936 } 1937 1938 int xenvif_map_frontend_rings(struct xenvif_queue *queue, 1939 grant_ref_t tx_ring_ref, 1940 grant_ref_t rx_ring_ref) 1941 { 1942 void *addr; 1943 struct xen_netif_tx_sring *txs; 1944 struct xen_netif_rx_sring *rxs; 1945 1946 int err = -ENOMEM; 1947 1948 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif), 1949 &tx_ring_ref, 1, &addr); 1950 if (err) 1951 goto err; 1952 1953 txs = (struct xen_netif_tx_sring *)addr; 1954 BACK_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE); 1955 1956 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif), 1957 &rx_ring_ref, 1, &addr); 1958 if (err) 1959 goto err; 1960 1961 rxs = (struct xen_netif_rx_sring *)addr; 1962 BACK_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE); 1963 1964 return 0; 1965 1966 err: 1967 xenvif_unmap_frontend_rings(queue); 1968 return err; 1969 } 1970 1971 static void xenvif_queue_carrier_off(struct xenvif_queue *queue) 1972 { 1973 struct xenvif *vif = queue->vif; 1974 1975 queue->stalled = true; 1976 1977 /* At least one queue has stalled? Disable the carrier. */ 1978 spin_lock(&vif->lock); 1979 if (vif->stalled_queues++ == 0) { 1980 netdev_info(vif->dev, "Guest Rx stalled"); 1981 netif_carrier_off(vif->dev); 1982 } 1983 spin_unlock(&vif->lock); 1984 } 1985 1986 static void xenvif_queue_carrier_on(struct xenvif_queue *queue) 1987 { 1988 struct xenvif *vif = queue->vif; 1989 1990 queue->last_rx_time = jiffies; /* Reset Rx stall detection. */ 1991 queue->stalled = false; 1992 1993 /* All queues are ready? Enable the carrier. */ 1994 spin_lock(&vif->lock); 1995 if (--vif->stalled_queues == 0) { 1996 netdev_info(vif->dev, "Guest Rx ready"); 1997 netif_carrier_on(vif->dev); 1998 } 1999 spin_unlock(&vif->lock); 2000 } 2001 2002 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue) 2003 { 2004 RING_IDX prod, cons; 2005 2006 prod = queue->rx.sring->req_prod; 2007 cons = queue->rx.req_cons; 2008 2009 return !queue->stalled && prod - cons < 1 2010 && time_after(jiffies, 2011 queue->last_rx_time + queue->vif->stall_timeout); 2012 } 2013 2014 static bool xenvif_rx_queue_ready(struct xenvif_queue *queue) 2015 { 2016 RING_IDX prod, cons; 2017 2018 prod = queue->rx.sring->req_prod; 2019 cons = queue->rx.req_cons; 2020 2021 return queue->stalled && prod - cons >= 1; 2022 } 2023 2024 static bool xenvif_have_rx_work(struct xenvif_queue *queue) 2025 { 2026 return xenvif_rx_ring_slots_available(queue) 2027 || (queue->vif->stall_timeout && 2028 (xenvif_rx_queue_stalled(queue) 2029 || xenvif_rx_queue_ready(queue))) 2030 || kthread_should_stop() 2031 || queue->vif->disabled; 2032 } 2033 2034 static long xenvif_rx_queue_timeout(struct xenvif_queue *queue) 2035 { 2036 struct sk_buff *skb; 2037 long timeout; 2038 2039 skb = skb_peek(&queue->rx_queue); 2040 if (!skb) 2041 return MAX_SCHEDULE_TIMEOUT; 2042 2043 timeout = XENVIF_RX_CB(skb)->expires - jiffies; 2044 return timeout < 0 ? 0 : timeout; 2045 } 2046 2047 /* Wait until the guest Rx thread has work. 2048 * 2049 * The timeout needs to be adjusted based on the current head of the 2050 * queue (and not just the head at the beginning). In particular, if 2051 * the queue is initially empty an infinite timeout is used and this 2052 * needs to be reduced when a skb is queued. 2053 * 2054 * This cannot be done with wait_event_timeout() because it only 2055 * calculates the timeout once. 2056 */ 2057 static void xenvif_wait_for_rx_work(struct xenvif_queue *queue) 2058 { 2059 DEFINE_WAIT(wait); 2060 2061 if (xenvif_have_rx_work(queue)) 2062 return; 2063 2064 for (;;) { 2065 long ret; 2066 2067 prepare_to_wait(&queue->wq, &wait, TASK_INTERRUPTIBLE); 2068 if (xenvif_have_rx_work(queue)) 2069 break; 2070 ret = schedule_timeout(xenvif_rx_queue_timeout(queue)); 2071 if (!ret) 2072 break; 2073 } 2074 finish_wait(&queue->wq, &wait); 2075 } 2076 2077 int xenvif_kthread_guest_rx(void *data) 2078 { 2079 struct xenvif_queue *queue = data; 2080 struct xenvif *vif = queue->vif; 2081 2082 if (!vif->stall_timeout) 2083 xenvif_queue_carrier_on(queue); 2084 2085 for (;;) { 2086 xenvif_wait_for_rx_work(queue); 2087 2088 if (kthread_should_stop()) 2089 break; 2090 2091 /* This frontend is found to be rogue, disable it in 2092 * kthread context. Currently this is only set when 2093 * netback finds out frontend sends malformed packet, 2094 * but we cannot disable the interface in softirq 2095 * context so we defer it here, if this thread is 2096 * associated with queue 0. 2097 */ 2098 if (unlikely(vif->disabled && queue->id == 0)) { 2099 xenvif_carrier_off(vif); 2100 break; 2101 } 2102 2103 if (!skb_queue_empty(&queue->rx_queue)) 2104 xenvif_rx_action(queue); 2105 2106 /* If the guest hasn't provided any Rx slots for a 2107 * while it's probably not responsive, drop the 2108 * carrier so packets are dropped earlier. 2109 */ 2110 if (vif->stall_timeout) { 2111 if (xenvif_rx_queue_stalled(queue)) 2112 xenvif_queue_carrier_off(queue); 2113 else if (xenvif_rx_queue_ready(queue)) 2114 xenvif_queue_carrier_on(queue); 2115 } 2116 2117 /* Queued packets may have foreign pages from other 2118 * domains. These cannot be queued indefinitely as 2119 * this would starve guests of grant refs and transmit 2120 * slots. 2121 */ 2122 xenvif_rx_queue_drop_expired(queue); 2123 2124 xenvif_rx_queue_maybe_wake(queue); 2125 2126 cond_resched(); 2127 } 2128 2129 /* Bin any remaining skbs */ 2130 xenvif_rx_queue_purge(queue); 2131 2132 return 0; 2133 } 2134 2135 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue) 2136 { 2137 /* Dealloc thread must remain running until all inflight 2138 * packets complete. 2139 */ 2140 return kthread_should_stop() && 2141 !atomic_read(&queue->inflight_packets); 2142 } 2143 2144 int xenvif_dealloc_kthread(void *data) 2145 { 2146 struct xenvif_queue *queue = data; 2147 2148 for (;;) { 2149 wait_event_interruptible(queue->dealloc_wq, 2150 tx_dealloc_work_todo(queue) || 2151 xenvif_dealloc_kthread_should_stop(queue)); 2152 if (xenvif_dealloc_kthread_should_stop(queue)) 2153 break; 2154 2155 xenvif_tx_dealloc_action(queue); 2156 cond_resched(); 2157 } 2158 2159 /* Unmap anything remaining*/ 2160 if (tx_dealloc_work_todo(queue)) 2161 xenvif_tx_dealloc_action(queue); 2162 2163 return 0; 2164 } 2165 2166 static int __init netback_init(void) 2167 { 2168 int rc = 0; 2169 2170 if (!xen_domain()) 2171 return -ENODEV; 2172 2173 /* Allow as many queues as there are CPUs if user has not 2174 * specified a value. 2175 */ 2176 if (xenvif_max_queues == 0) 2177 xenvif_max_queues = num_online_cpus(); 2178 2179 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) { 2180 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n", 2181 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX); 2182 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX; 2183 } 2184 2185 rc = xenvif_xenbus_init(); 2186 if (rc) 2187 goto failed_init; 2188 2189 #ifdef CONFIG_DEBUG_FS 2190 xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL); 2191 if (IS_ERR_OR_NULL(xen_netback_dbg_root)) 2192 pr_warn("Init of debugfs returned %ld!\n", 2193 PTR_ERR(xen_netback_dbg_root)); 2194 #endif /* CONFIG_DEBUG_FS */ 2195 2196 return 0; 2197 2198 failed_init: 2199 return rc; 2200 } 2201 2202 module_init(netback_init); 2203 2204 static void __exit netback_fini(void) 2205 { 2206 #ifdef CONFIG_DEBUG_FS 2207 if (!IS_ERR_OR_NULL(xen_netback_dbg_root)) 2208 debugfs_remove_recursive(xen_netback_dbg_root); 2209 #endif /* CONFIG_DEBUG_FS */ 2210 xenvif_xenbus_fini(); 2211 } 2212 module_exit(netback_fini); 2213 2214 MODULE_LICENSE("Dual BSD/GPL"); 2215 MODULE_ALIAS("xen-backend:vif"); 2216