1 /* 2 * This file is subject to the terms and conditions of the GNU General Public 3 * License. See the file "COPYING" in the main directory of this archive 4 * for more details. 5 * 6 * Copyright (C) 1999-2008 Silicon Graphics, Inc. All rights reserved. 7 */ 8 9 /* 10 * Cross Partition Network Interface (XPNET) support 11 * 12 * XPNET provides a virtual network layered on top of the Cross 13 * Partition communication layer. 14 * 15 * XPNET provides direct point-to-point and broadcast-like support 16 * for an ethernet-like device. The ethernet broadcast medium is 17 * replaced with a point-to-point message structure which passes 18 * pointers to a DMA-capable block that a remote partition should 19 * retrieve and pass to the upper level networking layer. 20 * 21 */ 22 23 #include <linux/module.h> 24 #include <linux/netdevice.h> 25 #include <linux/etherdevice.h> 26 #include "xp.h" 27 28 /* 29 * The message payload transferred by XPC. 30 * 31 * buf_pa is the physical address where the DMA should pull from. 32 * 33 * NOTE: for performance reasons, buf_pa should _ALWAYS_ begin on a 34 * cacheline boundary. To accomplish this, we record the number of 35 * bytes from the beginning of the first cacheline to the first useful 36 * byte of the skb (leadin_ignore) and the number of bytes from the 37 * last useful byte of the skb to the end of the last cacheline 38 * (tailout_ignore). 39 * 40 * size is the number of bytes to transfer which includes the skb->len 41 * (useful bytes of the senders skb) plus the leadin and tailout 42 */ 43 struct xpnet_message { 44 u16 version; /* Version for this message */ 45 u16 embedded_bytes; /* #of bytes embedded in XPC message */ 46 u32 magic; /* Special number indicating this is xpnet */ 47 unsigned long buf_pa; /* phys address of buffer to retrieve */ 48 u32 size; /* #of bytes in buffer */ 49 u8 leadin_ignore; /* #of bytes to ignore at the beginning */ 50 u8 tailout_ignore; /* #of bytes to ignore at the end */ 51 unsigned char data; /* body of small packets */ 52 }; 53 54 /* 55 * Determine the size of our message, the cacheline aligned size, 56 * and then the number of message will request from XPC. 57 * 58 * XPC expects each message to exist in an individual cacheline. 59 */ 60 #define XPNET_MSG_SIZE XPC_MSG_PAYLOAD_MAX_SIZE 61 #define XPNET_MSG_DATA_MAX \ 62 (XPNET_MSG_SIZE - offsetof(struct xpnet_message, data)) 63 #define XPNET_MSG_NENTRIES (PAGE_SIZE / XPC_MSG_MAX_SIZE) 64 65 #define XPNET_MAX_KTHREADS (XPNET_MSG_NENTRIES + 1) 66 #define XPNET_MAX_IDLE_KTHREADS (XPNET_MSG_NENTRIES + 1) 67 68 /* 69 * Version number of XPNET implementation. XPNET can always talk to versions 70 * with same major #, and never talk to versions with a different version. 71 */ 72 #define _XPNET_VERSION(_major, _minor) (((_major) << 4) | (_minor)) 73 #define XPNET_VERSION_MAJOR(_v) ((_v) >> 4) 74 #define XPNET_VERSION_MINOR(_v) ((_v) & 0xf) 75 76 #define XPNET_VERSION _XPNET_VERSION(1, 0) /* version 1.0 */ 77 #define XPNET_VERSION_EMBED _XPNET_VERSION(1, 1) /* version 1.1 */ 78 #define XPNET_MAGIC 0x88786984 /* "XNET" */ 79 80 #define XPNET_VALID_MSG(_m) \ 81 ((XPNET_VERSION_MAJOR(_m->version) == XPNET_VERSION_MAJOR(XPNET_VERSION)) \ 82 && (msg->magic == XPNET_MAGIC)) 83 84 #define XPNET_DEVICE_NAME "xp0" 85 86 /* 87 * When messages are queued with xpc_send_notify, a kmalloc'd buffer 88 * of the following type is passed as a notification cookie. When the 89 * notification function is called, we use the cookie to decide 90 * whether all outstanding message sends have completed. The skb can 91 * then be released. 92 */ 93 struct xpnet_pending_msg { 94 struct sk_buff *skb; 95 atomic_t use_count; 96 }; 97 98 /* driver specific structure pointed to by the device structure */ 99 struct xpnet_dev_private { 100 struct net_device_stats stats; 101 }; 102 103 struct net_device *xpnet_device; 104 105 /* 106 * When we are notified of other partitions activating, we add them to 107 * our bitmask of partitions to which we broadcast. 108 */ 109 static unsigned long *xpnet_broadcast_partitions; 110 /* protect above */ 111 static DEFINE_SPINLOCK(xpnet_broadcast_lock); 112 113 /* 114 * Since the Block Transfer Engine (BTE) is being used for the transfer 115 * and it relies upon cache-line size transfers, we need to reserve at 116 * least one cache-line for head and tail alignment. The BTE is 117 * limited to 8MB transfers. 118 * 119 * Testing has shown that changing MTU to greater than 64KB has no effect 120 * on TCP as the two sides negotiate a Max Segment Size that is limited 121 * to 64K. Other protocols May use packets greater than this, but for 122 * now, the default is 64KB. 123 */ 124 #define XPNET_MAX_MTU (0x800000UL - L1_CACHE_BYTES) 125 /* 32KB has been determined to be the ideal */ 126 #define XPNET_DEF_MTU (0x8000UL) 127 128 /* 129 * The partid is encapsulated in the MAC address beginning in the following 130 * octet and it consists of two octets. 131 */ 132 #define XPNET_PARTID_OCTET 2 133 134 /* Define the XPNET debug device structures to be used with dev_dbg() et al */ 135 136 struct device_driver xpnet_dbg_name = { 137 .name = "xpnet" 138 }; 139 140 struct device xpnet_dbg_subname = { 141 .bus_id = {0}, /* set to "" */ 142 .driver = &xpnet_dbg_name 143 }; 144 145 struct device *xpnet = &xpnet_dbg_subname; 146 147 /* 148 * Packet was recevied by XPC and forwarded to us. 149 */ 150 static void 151 xpnet_receive(short partid, int channel, struct xpnet_message *msg) 152 { 153 struct sk_buff *skb; 154 void *dst; 155 enum xp_retval ret; 156 struct xpnet_dev_private *priv = 157 (struct xpnet_dev_private *)xpnet_device->priv; 158 159 if (!XPNET_VALID_MSG(msg)) { 160 /* 161 * Packet with a different XPC version. Ignore. 162 */ 163 xpc_received(partid, channel, (void *)msg); 164 165 priv->stats.rx_errors++; 166 167 return; 168 } 169 dev_dbg(xpnet, "received 0x%lx, %d, %d, %d\n", msg->buf_pa, msg->size, 170 msg->leadin_ignore, msg->tailout_ignore); 171 172 /* reserve an extra cache line */ 173 skb = dev_alloc_skb(msg->size + L1_CACHE_BYTES); 174 if (!skb) { 175 dev_err(xpnet, "failed on dev_alloc_skb(%d)\n", 176 msg->size + L1_CACHE_BYTES); 177 178 xpc_received(partid, channel, (void *)msg); 179 180 priv->stats.rx_errors++; 181 182 return; 183 } 184 185 /* 186 * The allocated skb has some reserved space. 187 * In order to use xp_remote_memcpy(), we need to get the 188 * skb->data pointer moved forward. 189 */ 190 skb_reserve(skb, (L1_CACHE_BYTES - ((u64)skb->data & 191 (L1_CACHE_BYTES - 1)) + 192 msg->leadin_ignore)); 193 194 /* 195 * Update the tail pointer to indicate data actually 196 * transferred. 197 */ 198 skb_put(skb, (msg->size - msg->leadin_ignore - msg->tailout_ignore)); 199 200 /* 201 * Move the data over from the other side. 202 */ 203 if ((XPNET_VERSION_MINOR(msg->version) == 1) && 204 (msg->embedded_bytes != 0)) { 205 dev_dbg(xpnet, "copying embedded message. memcpy(0x%p, 0x%p, " 206 "%lu)\n", skb->data, &msg->data, 207 (size_t)msg->embedded_bytes); 208 209 skb_copy_to_linear_data(skb, &msg->data, 210 (size_t)msg->embedded_bytes); 211 } else { 212 dst = (void *)((u64)skb->data & ~(L1_CACHE_BYTES - 1)); 213 dev_dbg(xpnet, "transferring buffer to the skb->data area;\n\t" 214 "xp_remote_memcpy(0x%p, 0x%p, %hu)\n", dst, 215 (void *)msg->buf_pa, msg->size); 216 217 ret = xp_remote_memcpy(xp_pa(dst), msg->buf_pa, msg->size); 218 if (ret != xpSuccess) { 219 /* 220 * !!! Need better way of cleaning skb. Currently skb 221 * !!! appears in_use and we can't just call 222 * !!! dev_kfree_skb. 223 */ 224 dev_err(xpnet, "xp_remote_memcpy(0x%p, 0x%p, 0x%hx) " 225 "returned error=0x%x\n", dst, 226 (void *)msg->buf_pa, msg->size, ret); 227 228 xpc_received(partid, channel, (void *)msg); 229 230 priv->stats.rx_errors++; 231 232 return; 233 } 234 } 235 236 dev_dbg(xpnet, "<skb->head=0x%p skb->data=0x%p skb->tail=0x%p " 237 "skb->end=0x%p skb->len=%d\n", (void *)skb->head, 238 (void *)skb->data, skb_tail_pointer(skb), skb_end_pointer(skb), 239 skb->len); 240 241 skb->protocol = eth_type_trans(skb, xpnet_device); 242 skb->ip_summed = CHECKSUM_UNNECESSARY; 243 244 dev_dbg(xpnet, "passing skb to network layer\n" 245 KERN_DEBUG "\tskb->head=0x%p skb->data=0x%p skb->tail=0x%p " 246 "skb->end=0x%p skb->len=%d\n", 247 (void *)skb->head, (void *)skb->data, skb_tail_pointer(skb), 248 skb_end_pointer(skb), skb->len); 249 250 xpnet_device->last_rx = jiffies; 251 priv->stats.rx_packets++; 252 priv->stats.rx_bytes += skb->len + ETH_HLEN; 253 254 netif_rx_ni(skb); 255 xpc_received(partid, channel, (void *)msg); 256 } 257 258 /* 259 * This is the handler which XPC calls during any sort of change in 260 * state or message reception on a connection. 261 */ 262 static void 263 xpnet_connection_activity(enum xp_retval reason, short partid, int channel, 264 void *data, void *key) 265 { 266 DBUG_ON(partid < 0 || partid >= xp_max_npartitions); 267 DBUG_ON(channel != XPC_NET_CHANNEL); 268 269 switch (reason) { 270 case xpMsgReceived: /* message received */ 271 DBUG_ON(data == NULL); 272 273 xpnet_receive(partid, channel, (struct xpnet_message *)data); 274 break; 275 276 case xpConnected: /* connection completed to a partition */ 277 spin_lock_bh(&xpnet_broadcast_lock); 278 __set_bit(partid, xpnet_broadcast_partitions); 279 spin_unlock_bh(&xpnet_broadcast_lock); 280 281 netif_carrier_on(xpnet_device); 282 283 dev_dbg(xpnet, "%s connected to partition %d\n", 284 xpnet_device->name, partid); 285 break; 286 287 default: 288 spin_lock_bh(&xpnet_broadcast_lock); 289 __clear_bit(partid, xpnet_broadcast_partitions); 290 spin_unlock_bh(&xpnet_broadcast_lock); 291 292 if (bitmap_empty((unsigned long *)xpnet_broadcast_partitions, 293 xp_max_npartitions)) { 294 netif_carrier_off(xpnet_device); 295 } 296 297 dev_dbg(xpnet, "%s disconnected from partition %d\n", 298 xpnet_device->name, partid); 299 break; 300 } 301 } 302 303 static int 304 xpnet_dev_open(struct net_device *dev) 305 { 306 enum xp_retval ret; 307 308 dev_dbg(xpnet, "calling xpc_connect(%d, 0x%p, NULL, %ld, %ld, %ld, " 309 "%ld)\n", XPC_NET_CHANNEL, xpnet_connection_activity, 310 (unsigned long)XPNET_MSG_SIZE, 311 (unsigned long)XPNET_MSG_NENTRIES, 312 (unsigned long)XPNET_MAX_KTHREADS, 313 (unsigned long)XPNET_MAX_IDLE_KTHREADS); 314 315 ret = xpc_connect(XPC_NET_CHANNEL, xpnet_connection_activity, NULL, 316 XPNET_MSG_SIZE, XPNET_MSG_NENTRIES, 317 XPNET_MAX_KTHREADS, XPNET_MAX_IDLE_KTHREADS); 318 if (ret != xpSuccess) { 319 dev_err(xpnet, "ifconfig up of %s failed on XPC connect, " 320 "ret=%d\n", dev->name, ret); 321 322 return -ENOMEM; 323 } 324 325 dev_dbg(xpnet, "ifconfig up of %s; XPC connected\n", dev->name); 326 327 return 0; 328 } 329 330 static int 331 xpnet_dev_stop(struct net_device *dev) 332 { 333 xpc_disconnect(XPC_NET_CHANNEL); 334 335 dev_dbg(xpnet, "ifconfig down of %s; XPC disconnected\n", dev->name); 336 337 return 0; 338 } 339 340 static int 341 xpnet_dev_change_mtu(struct net_device *dev, int new_mtu) 342 { 343 /* 68 comes from min TCP+IP+MAC header */ 344 if ((new_mtu < 68) || (new_mtu > XPNET_MAX_MTU)) { 345 dev_err(xpnet, "ifconfig %s mtu %d failed; value must be " 346 "between 68 and %ld\n", dev->name, new_mtu, 347 XPNET_MAX_MTU); 348 return -EINVAL; 349 } 350 351 dev->mtu = new_mtu; 352 dev_dbg(xpnet, "ifconfig %s mtu set to %d\n", dev->name, new_mtu); 353 return 0; 354 } 355 356 /* 357 * Required for the net_device structure. 358 */ 359 static int 360 xpnet_dev_set_config(struct net_device *dev, struct ifmap *new_map) 361 { 362 return 0; 363 } 364 365 /* 366 * Return statistics to the caller. 367 */ 368 static struct net_device_stats * 369 xpnet_dev_get_stats(struct net_device *dev) 370 { 371 struct xpnet_dev_private *priv; 372 373 priv = (struct xpnet_dev_private *)dev->priv; 374 375 return &priv->stats; 376 } 377 378 /* 379 * Notification that the other end has received the message and 380 * DMA'd the skb information. At this point, they are done with 381 * our side. When all recipients are done processing, we 382 * release the skb and then release our pending message structure. 383 */ 384 static void 385 xpnet_send_completed(enum xp_retval reason, short partid, int channel, 386 void *__qm) 387 { 388 struct xpnet_pending_msg *queued_msg = (struct xpnet_pending_msg *)__qm; 389 390 DBUG_ON(queued_msg == NULL); 391 392 dev_dbg(xpnet, "message to %d notified with reason %d\n", 393 partid, reason); 394 395 if (atomic_dec_return(&queued_msg->use_count) == 0) { 396 dev_dbg(xpnet, "all acks for skb->head=-x%p\n", 397 (void *)queued_msg->skb->head); 398 399 dev_kfree_skb_any(queued_msg->skb); 400 kfree(queued_msg); 401 } 402 } 403 404 static void 405 xpnet_send(struct sk_buff *skb, struct xpnet_pending_msg *queued_msg, 406 u64 start_addr, u64 end_addr, u16 embedded_bytes, int dest_partid) 407 { 408 u8 msg_buffer[XPNET_MSG_SIZE]; 409 struct xpnet_message *msg = (struct xpnet_message *)&msg_buffer; 410 u16 msg_size = sizeof(struct xpnet_message); 411 enum xp_retval ret; 412 413 msg->embedded_bytes = embedded_bytes; 414 if (unlikely(embedded_bytes != 0)) { 415 msg->version = XPNET_VERSION_EMBED; 416 dev_dbg(xpnet, "calling memcpy(0x%p, 0x%p, 0x%lx)\n", 417 &msg->data, skb->data, (size_t)embedded_bytes); 418 skb_copy_from_linear_data(skb, &msg->data, 419 (size_t)embedded_bytes); 420 msg_size += embedded_bytes - 1; 421 } else { 422 msg->version = XPNET_VERSION; 423 } 424 msg->magic = XPNET_MAGIC; 425 msg->size = end_addr - start_addr; 426 msg->leadin_ignore = (u64)skb->data - start_addr; 427 msg->tailout_ignore = end_addr - (u64)skb_tail_pointer(skb); 428 msg->buf_pa = xp_pa((void *)start_addr); 429 430 dev_dbg(xpnet, "sending XPC message to %d:%d\n" 431 KERN_DEBUG "msg->buf_pa=0x%lx, msg->size=%u, " 432 "msg->leadin_ignore=%u, msg->tailout_ignore=%u\n", 433 dest_partid, XPC_NET_CHANNEL, msg->buf_pa, msg->size, 434 msg->leadin_ignore, msg->tailout_ignore); 435 436 atomic_inc(&queued_msg->use_count); 437 438 ret = xpc_send_notify(dest_partid, XPC_NET_CHANNEL, XPC_NOWAIT, msg, 439 msg_size, xpnet_send_completed, queued_msg); 440 if (unlikely(ret != xpSuccess)) 441 atomic_dec(&queued_msg->use_count); 442 } 443 444 /* 445 * Network layer has formatted a packet (skb) and is ready to place it 446 * "on the wire". Prepare and send an xpnet_message to all partitions 447 * which have connected with us and are targets of this packet. 448 * 449 * MAC-NOTE: For the XPNET driver, the MAC address contains the 450 * destination partid. If the destination partid octets are 0xffff, 451 * this packet is to be broadcast to all connected partitions. 452 */ 453 static int 454 xpnet_dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev) 455 { 456 struct xpnet_pending_msg *queued_msg; 457 u64 start_addr, end_addr; 458 short dest_partid; 459 struct xpnet_dev_private *priv = (struct xpnet_dev_private *)dev->priv; 460 u16 embedded_bytes = 0; 461 462 dev_dbg(xpnet, ">skb->head=0x%p skb->data=0x%p skb->tail=0x%p " 463 "skb->end=0x%p skb->len=%d\n", (void *)skb->head, 464 (void *)skb->data, skb_tail_pointer(skb), skb_end_pointer(skb), 465 skb->len); 466 467 if (skb->data[0] == 0x33) { 468 dev_kfree_skb(skb); 469 return 0; /* nothing needed to be done */ 470 } 471 472 /* 473 * The xpnet_pending_msg tracks how many outstanding 474 * xpc_send_notifies are relying on this skb. When none 475 * remain, release the skb. 476 */ 477 queued_msg = kmalloc(sizeof(struct xpnet_pending_msg), GFP_ATOMIC); 478 if (queued_msg == NULL) { 479 dev_warn(xpnet, "failed to kmalloc %ld bytes; dropping " 480 "packet\n", sizeof(struct xpnet_pending_msg)); 481 482 priv->stats.tx_errors++; 483 return -ENOMEM; 484 } 485 486 /* get the beginning of the first cacheline and end of last */ 487 start_addr = ((u64)skb->data & ~(L1_CACHE_BYTES - 1)); 488 end_addr = L1_CACHE_ALIGN((u64)skb_tail_pointer(skb)); 489 490 /* calculate how many bytes to embed in the XPC message */ 491 if (unlikely(skb->len <= XPNET_MSG_DATA_MAX)) { 492 /* skb->data does fit so embed */ 493 embedded_bytes = skb->len; 494 } 495 496 /* 497 * Since the send occurs asynchronously, we set the count to one 498 * and begin sending. Any sends that happen to complete before 499 * we are done sending will not free the skb. We will be left 500 * with that task during exit. This also handles the case of 501 * a packet destined for a partition which is no longer up. 502 */ 503 atomic_set(&queued_msg->use_count, 1); 504 queued_msg->skb = skb; 505 506 if (skb->data[0] == 0xff) { 507 /* we are being asked to broadcast to all partitions */ 508 for_each_bit(dest_partid, xpnet_broadcast_partitions, 509 xp_max_npartitions) { 510 511 xpnet_send(skb, queued_msg, start_addr, end_addr, 512 embedded_bytes, dest_partid); 513 } 514 } else { 515 dest_partid = (short)skb->data[XPNET_PARTID_OCTET + 1]; 516 dest_partid |= (short)skb->data[XPNET_PARTID_OCTET + 0] << 8; 517 518 if (dest_partid >= 0 && 519 dest_partid < xp_max_npartitions && 520 test_bit(dest_partid, xpnet_broadcast_partitions) != 0) { 521 522 xpnet_send(skb, queued_msg, start_addr, end_addr, 523 embedded_bytes, dest_partid); 524 } 525 } 526 527 if (atomic_dec_return(&queued_msg->use_count) == 0) { 528 dev_kfree_skb(skb); 529 kfree(queued_msg); 530 } 531 532 priv->stats.tx_packets++; 533 priv->stats.tx_bytes += skb->len; 534 535 return 0; 536 } 537 538 /* 539 * Deal with transmit timeouts coming from the network layer. 540 */ 541 static void 542 xpnet_dev_tx_timeout(struct net_device *dev) 543 { 544 struct xpnet_dev_private *priv; 545 546 priv = (struct xpnet_dev_private *)dev->priv; 547 548 priv->stats.tx_errors++; 549 return; 550 } 551 552 static int __init 553 xpnet_init(void) 554 { 555 int result; 556 557 if (!is_shub() && !is_uv()) 558 return -ENODEV; 559 560 dev_info(xpnet, "registering network device %s\n", XPNET_DEVICE_NAME); 561 562 xpnet_broadcast_partitions = kzalloc(BITS_TO_LONGS(xp_max_npartitions) * 563 sizeof(long), GFP_KERNEL); 564 if (xpnet_broadcast_partitions == NULL) 565 return -ENOMEM; 566 567 /* 568 * use ether_setup() to init the majority of our device 569 * structure and then override the necessary pieces. 570 */ 571 xpnet_device = alloc_netdev(sizeof(struct xpnet_dev_private), 572 XPNET_DEVICE_NAME, ether_setup); 573 if (xpnet_device == NULL) { 574 kfree(xpnet_broadcast_partitions); 575 return -ENOMEM; 576 } 577 578 netif_carrier_off(xpnet_device); 579 580 xpnet_device->mtu = XPNET_DEF_MTU; 581 xpnet_device->change_mtu = xpnet_dev_change_mtu; 582 xpnet_device->open = xpnet_dev_open; 583 xpnet_device->get_stats = xpnet_dev_get_stats; 584 xpnet_device->stop = xpnet_dev_stop; 585 xpnet_device->hard_start_xmit = xpnet_dev_hard_start_xmit; 586 xpnet_device->tx_timeout = xpnet_dev_tx_timeout; 587 xpnet_device->set_config = xpnet_dev_set_config; 588 589 /* 590 * Multicast assumes the LSB of the first octet is set for multicast 591 * MAC addresses. We chose the first octet of the MAC to be unlikely 592 * to collide with any vendor's officially issued MAC. 593 */ 594 xpnet_device->dev_addr[0] = 0x02; /* locally administered, no OUI */ 595 596 xpnet_device->dev_addr[XPNET_PARTID_OCTET + 1] = xp_partition_id; 597 xpnet_device->dev_addr[XPNET_PARTID_OCTET + 0] = (xp_partition_id >> 8); 598 599 /* 600 * ether_setup() sets this to a multicast device. We are 601 * really not supporting multicast at this time. 602 */ 603 xpnet_device->flags &= ~IFF_MULTICAST; 604 605 /* 606 * No need to checksum as it is a DMA transfer. The BTE will 607 * report an error if the data is not retrievable and the 608 * packet will be dropped. 609 */ 610 xpnet_device->features = NETIF_F_NO_CSUM; 611 612 result = register_netdev(xpnet_device); 613 if (result != 0) { 614 free_netdev(xpnet_device); 615 kfree(xpnet_broadcast_partitions); 616 } 617 618 return result; 619 } 620 621 module_init(xpnet_init); 622 623 static void __exit 624 xpnet_exit(void) 625 { 626 dev_info(xpnet, "unregistering network device %s\n", 627 xpnet_device[0].name); 628 629 unregister_netdev(xpnet_device); 630 free_netdev(xpnet_device); 631 kfree(xpnet_broadcast_partitions); 632 } 633 634 module_exit(xpnet_exit); 635 636 MODULE_AUTHOR("Silicon Graphics, Inc."); 637 MODULE_DESCRIPTION("Cross Partition Network adapter (XPNET)"); 638 MODULE_LICENSE("GPL"); 639