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 * (C) Copyright 2020 Hewlett Packard Enterprise Development LP 7 * Copyright (C) 1999-2009 Silicon Graphics, Inc. All rights reserved. 8 */ 9 10 /* 11 * Cross Partition Network Interface (XPNET) support 12 * 13 * XPNET provides a virtual network layered on top of the Cross 14 * Partition communication layer. 15 * 16 * XPNET provides direct point-to-point and broadcast-like support 17 * for an ethernet-like device. The ethernet broadcast medium is 18 * replaced with a point-to-point message structure which passes 19 * pointers to a DMA-capable block that a remote partition should 20 * retrieve and pass to the upper level networking layer. 21 * 22 */ 23 24 #include <linux/slab.h> 25 #include <linux/module.h> 26 #include <linux/netdevice.h> 27 #include <linux/etherdevice.h> 28 #include "xp.h" 29 30 /* 31 * The message payload transferred by XPC. 32 * 33 * buf_pa is the physical address where the DMA should pull from. 34 * 35 * NOTE: for performance reasons, buf_pa should _ALWAYS_ begin on a 36 * cacheline boundary. To accomplish this, we record the number of 37 * bytes from the beginning of the first cacheline to the first useful 38 * byte of the skb (leadin_ignore) and the number of bytes from the 39 * last useful byte of the skb to the end of the last cacheline 40 * (tailout_ignore). 41 * 42 * size is the number of bytes to transfer which includes the skb->len 43 * (useful bytes of the senders skb) plus the leadin and tailout 44 */ 45 struct xpnet_message { 46 u16 version; /* Version for this message */ 47 u16 embedded_bytes; /* #of bytes embedded in XPC message */ 48 u32 magic; /* Special number indicating this is xpnet */ 49 unsigned long buf_pa; /* phys address of buffer to retrieve */ 50 u32 size; /* #of bytes in buffer */ 51 u8 leadin_ignore; /* #of bytes to ignore at the beginning */ 52 u8 tailout_ignore; /* #of bytes to ignore at the end */ 53 unsigned char data; /* body of small packets */ 54 }; 55 56 /* 57 * Determine the size of our message, the cacheline aligned size, 58 * and then the number of message will request from XPC. 59 * 60 * XPC expects each message to exist in an individual cacheline. 61 */ 62 #define XPNET_MSG_SIZE XPC_MSG_PAYLOAD_MAX_SIZE 63 #define XPNET_MSG_DATA_MAX \ 64 (XPNET_MSG_SIZE - offsetof(struct xpnet_message, data)) 65 #define XPNET_MSG_NENTRIES (PAGE_SIZE / XPC_MSG_MAX_SIZE) 66 67 #define XPNET_MAX_KTHREADS (XPNET_MSG_NENTRIES + 1) 68 #define XPNET_MAX_IDLE_KTHREADS (XPNET_MSG_NENTRIES + 1) 69 70 /* 71 * Version number of XPNET implementation. XPNET can always talk to versions 72 * with same major #, and never talk to versions with a different version. 73 */ 74 #define _XPNET_VERSION(_major, _minor) (((_major) << 4) | (_minor)) 75 #define XPNET_VERSION_MAJOR(_v) ((_v) >> 4) 76 #define XPNET_VERSION_MINOR(_v) ((_v) & 0xf) 77 78 #define XPNET_VERSION _XPNET_VERSION(1, 0) /* version 1.0 */ 79 #define XPNET_VERSION_EMBED _XPNET_VERSION(1, 1) /* version 1.1 */ 80 #define XPNET_MAGIC 0x88786984 /* "XNET" */ 81 82 #define XPNET_VALID_MSG(_m) \ 83 ((XPNET_VERSION_MAJOR(_m->version) == XPNET_VERSION_MAJOR(XPNET_VERSION)) \ 84 && (msg->magic == XPNET_MAGIC)) 85 86 #define XPNET_DEVICE_NAME "xp0" 87 88 /* 89 * When messages are queued with xpc_send_notify, a kmalloc'd buffer 90 * of the following type is passed as a notification cookie. When the 91 * notification function is called, we use the cookie to decide 92 * whether all outstanding message sends have completed. The skb can 93 * then be released. 94 */ 95 struct xpnet_pending_msg { 96 struct sk_buff *skb; 97 atomic_t use_count; 98 }; 99 100 static struct net_device *xpnet_device; 101 102 /* 103 * When we are notified of other partitions activating, we add them to 104 * our bitmask of partitions to which we broadcast. 105 */ 106 static unsigned long *xpnet_broadcast_partitions; 107 /* protect above */ 108 static DEFINE_SPINLOCK(xpnet_broadcast_lock); 109 110 /* 111 * Since the Block Transfer Engine (BTE) is being used for the transfer 112 * and it relies upon cache-line size transfers, we need to reserve at 113 * least one cache-line for head and tail alignment. The BTE is 114 * limited to 8MB transfers. 115 * 116 * Testing has shown that changing MTU to greater than 64KB has no effect 117 * on TCP as the two sides negotiate a Max Segment Size that is limited 118 * to 64K. Other protocols May use packets greater than this, but for 119 * now, the default is 64KB. 120 */ 121 #define XPNET_MAX_MTU (0x800000UL - L1_CACHE_BYTES) 122 /* 68 comes from min TCP+IP+MAC header */ 123 #define XPNET_MIN_MTU 68 124 /* 32KB has been determined to be the ideal */ 125 #define XPNET_DEF_MTU (0x8000UL) 126 127 /* 128 * The partid is encapsulated in the MAC address beginning in the following 129 * octet and it consists of two octets. 130 */ 131 #define XPNET_PARTID_OCTET 2 132 133 /* Define the XPNET debug device structures to be used with dev_dbg() et al */ 134 135 static struct device_driver xpnet_dbg_name = { 136 .name = "xpnet" 137 }; 138 139 static struct device xpnet_dbg_subname = { 140 .init_name = "", /* set to "" */ 141 .driver = &xpnet_dbg_name 142 }; 143 144 static struct device *xpnet = &xpnet_dbg_subname; 145 146 /* 147 * Packet was recevied by XPC and forwarded to us. 148 */ 149 static void 150 xpnet_receive(short partid, int channel, struct xpnet_message *msg) 151 { 152 struct sk_buff *skb; 153 void *dst; 154 enum xp_retval ret; 155 156 if (!XPNET_VALID_MSG(msg)) { 157 /* 158 * Packet with a different XPC version. Ignore. 159 */ 160 xpc_received(partid, channel, (void *)msg); 161 162 xpnet_device->stats.rx_errors++; 163 164 return; 165 } 166 dev_dbg(xpnet, "received 0x%lx, %d, %d, %d\n", msg->buf_pa, msg->size, 167 msg->leadin_ignore, msg->tailout_ignore); 168 169 /* reserve an extra cache line */ 170 skb = dev_alloc_skb(msg->size + L1_CACHE_BYTES); 171 if (!skb) { 172 dev_err(xpnet, "failed on dev_alloc_skb(%d)\n", 173 msg->size + L1_CACHE_BYTES); 174 175 xpc_received(partid, channel, (void *)msg); 176 177 xpnet_device->stats.rx_errors++; 178 179 return; 180 } 181 182 /* 183 * The allocated skb has some reserved space. 184 * In order to use xp_remote_memcpy(), we need to get the 185 * skb->data pointer moved forward. 186 */ 187 skb_reserve(skb, (L1_CACHE_BYTES - ((u64)skb->data & 188 (L1_CACHE_BYTES - 1)) + 189 msg->leadin_ignore)); 190 191 /* 192 * Update the tail pointer to indicate data actually 193 * transferred. 194 */ 195 skb_put(skb, (msg->size - msg->leadin_ignore - msg->tailout_ignore)); 196 197 /* 198 * Move the data over from the other side. 199 */ 200 if ((XPNET_VERSION_MINOR(msg->version) == 1) && 201 (msg->embedded_bytes != 0)) { 202 dev_dbg(xpnet, "copying embedded message. memcpy(0x%p, 0x%p, " 203 "%lu)\n", skb->data, &msg->data, 204 (size_t)msg->embedded_bytes); 205 206 skb_copy_to_linear_data(skb, &msg->data, 207 (size_t)msg->embedded_bytes); 208 } else { 209 dst = (void *)((u64)skb->data & ~(L1_CACHE_BYTES - 1)); 210 dev_dbg(xpnet, "transferring buffer to the skb->data area;\n\t" 211 "xp_remote_memcpy(0x%p, 0x%p, %u)\n", dst, 212 (void *)msg->buf_pa, msg->size); 213 214 ret = xp_remote_memcpy(xp_pa(dst), msg->buf_pa, msg->size); 215 if (ret != xpSuccess) { 216 /* 217 * !!! Need better way of cleaning skb. Currently skb 218 * !!! appears in_use and we can't just call 219 * !!! dev_kfree_skb. 220 */ 221 dev_err(xpnet, "xp_remote_memcpy(0x%p, 0x%p, 0x%x) " 222 "returned error=0x%x\n", dst, 223 (void *)msg->buf_pa, msg->size, ret); 224 225 xpc_received(partid, channel, (void *)msg); 226 227 xpnet_device->stats.rx_errors++; 228 229 return; 230 } 231 } 232 233 dev_dbg(xpnet, "<skb->head=0x%p skb->data=0x%p skb->tail=0x%p " 234 "skb->end=0x%p skb->len=%d\n", (void *)skb->head, 235 (void *)skb->data, skb_tail_pointer(skb), skb_end_pointer(skb), 236 skb->len); 237 238 skb->protocol = eth_type_trans(skb, xpnet_device); 239 skb->ip_summed = CHECKSUM_UNNECESSARY; 240 241 dev_dbg(xpnet, "passing skb to network layer\n" 242 "\tskb->head=0x%p skb->data=0x%p skb->tail=0x%p " 243 "skb->end=0x%p skb->len=%d\n", 244 (void *)skb->head, (void *)skb->data, skb_tail_pointer(skb), 245 skb_end_pointer(skb), skb->len); 246 247 xpnet_device->stats.rx_packets++; 248 xpnet_device->stats.rx_bytes += skb->len + ETH_HLEN; 249 250 netif_rx(skb); 251 xpc_received(partid, channel, (void *)msg); 252 } 253 254 /* 255 * This is the handler which XPC calls during any sort of change in 256 * state or message reception on a connection. 257 */ 258 static void 259 xpnet_connection_activity(enum xp_retval reason, short partid, int channel, 260 void *data, void *key) 261 { 262 DBUG_ON(partid < 0 || partid >= xp_max_npartitions); 263 DBUG_ON(channel != XPC_NET_CHANNEL); 264 265 switch (reason) { 266 case xpMsgReceived: /* message received */ 267 DBUG_ON(data == NULL); 268 269 xpnet_receive(partid, channel, (struct xpnet_message *)data); 270 break; 271 272 case xpConnected: /* connection completed to a partition */ 273 spin_lock_bh(&xpnet_broadcast_lock); 274 __set_bit(partid, xpnet_broadcast_partitions); 275 spin_unlock_bh(&xpnet_broadcast_lock); 276 277 netif_carrier_on(xpnet_device); 278 279 dev_dbg(xpnet, "%s connected to partition %d\n", 280 xpnet_device->name, partid); 281 break; 282 283 default: 284 spin_lock_bh(&xpnet_broadcast_lock); 285 __clear_bit(partid, xpnet_broadcast_partitions); 286 spin_unlock_bh(&xpnet_broadcast_lock); 287 288 if (bitmap_empty((unsigned long *)xpnet_broadcast_partitions, 289 xp_max_npartitions)) { 290 netif_carrier_off(xpnet_device); 291 } 292 293 dev_dbg(xpnet, "%s disconnected from partition %d\n", 294 xpnet_device->name, partid); 295 break; 296 } 297 } 298 299 static int 300 xpnet_dev_open(struct net_device *dev) 301 { 302 enum xp_retval ret; 303 304 dev_dbg(xpnet, "calling xpc_connect(%d, 0x%p, NULL, %ld, %ld, %ld, " 305 "%ld)\n", XPC_NET_CHANNEL, xpnet_connection_activity, 306 (unsigned long)XPNET_MSG_SIZE, 307 (unsigned long)XPNET_MSG_NENTRIES, 308 (unsigned long)XPNET_MAX_KTHREADS, 309 (unsigned long)XPNET_MAX_IDLE_KTHREADS); 310 311 ret = xpc_connect(XPC_NET_CHANNEL, xpnet_connection_activity, NULL, 312 XPNET_MSG_SIZE, XPNET_MSG_NENTRIES, 313 XPNET_MAX_KTHREADS, XPNET_MAX_IDLE_KTHREADS); 314 if (ret != xpSuccess) { 315 dev_err(xpnet, "ifconfig up of %s failed on XPC connect, " 316 "ret=%d\n", dev->name, ret); 317 318 return -ENOMEM; 319 } 320 321 dev_dbg(xpnet, "ifconfig up of %s; XPC connected\n", dev->name); 322 323 return 0; 324 } 325 326 static int 327 xpnet_dev_stop(struct net_device *dev) 328 { 329 xpc_disconnect(XPC_NET_CHANNEL); 330 331 dev_dbg(xpnet, "ifconfig down of %s; XPC disconnected\n", dev->name); 332 333 return 0; 334 } 335 336 /* 337 * Notification that the other end has received the message and 338 * DMA'd the skb information. At this point, they are done with 339 * our side. When all recipients are done processing, we 340 * release the skb and then release our pending message structure. 341 */ 342 static void 343 xpnet_send_completed(enum xp_retval reason, short partid, int channel, 344 void *__qm) 345 { 346 struct xpnet_pending_msg *queued_msg = (struct xpnet_pending_msg *)__qm; 347 348 DBUG_ON(queued_msg == NULL); 349 350 dev_dbg(xpnet, "message to %d notified with reason %d\n", 351 partid, reason); 352 353 if (atomic_dec_return(&queued_msg->use_count) == 0) { 354 dev_dbg(xpnet, "all acks for skb->head=-x%p\n", 355 (void *)queued_msg->skb->head); 356 357 dev_kfree_skb_any(queued_msg->skb); 358 kfree(queued_msg); 359 } 360 } 361 362 static void 363 xpnet_send(struct sk_buff *skb, struct xpnet_pending_msg *queued_msg, 364 u64 start_addr, u64 end_addr, u16 embedded_bytes, int dest_partid) 365 { 366 u8 msg_buffer[XPNET_MSG_SIZE]; 367 struct xpnet_message *msg = (struct xpnet_message *)&msg_buffer; 368 u16 msg_size = sizeof(struct xpnet_message); 369 enum xp_retval ret; 370 371 msg->embedded_bytes = embedded_bytes; 372 if (unlikely(embedded_bytes != 0)) { 373 msg->version = XPNET_VERSION_EMBED; 374 dev_dbg(xpnet, "calling memcpy(0x%p, 0x%p, 0x%lx)\n", 375 &msg->data, skb->data, (size_t)embedded_bytes); 376 skb_copy_from_linear_data(skb, &msg->data, 377 (size_t)embedded_bytes); 378 msg_size += embedded_bytes - 1; 379 } else { 380 msg->version = XPNET_VERSION; 381 } 382 msg->magic = XPNET_MAGIC; 383 msg->size = end_addr - start_addr; 384 msg->leadin_ignore = (u64)skb->data - start_addr; 385 msg->tailout_ignore = end_addr - (u64)skb_tail_pointer(skb); 386 msg->buf_pa = xp_pa((void *)start_addr); 387 388 dev_dbg(xpnet, "sending XPC message to %d:%d\n" 389 "msg->buf_pa=0x%lx, msg->size=%u, " 390 "msg->leadin_ignore=%u, msg->tailout_ignore=%u\n", 391 dest_partid, XPC_NET_CHANNEL, msg->buf_pa, msg->size, 392 msg->leadin_ignore, msg->tailout_ignore); 393 394 atomic_inc(&queued_msg->use_count); 395 396 ret = xpc_send_notify(dest_partid, XPC_NET_CHANNEL, XPC_NOWAIT, msg, 397 msg_size, xpnet_send_completed, queued_msg); 398 if (unlikely(ret != xpSuccess)) 399 atomic_dec(&queued_msg->use_count); 400 } 401 402 /* 403 * Network layer has formatted a packet (skb) and is ready to place it 404 * "on the wire". Prepare and send an xpnet_message to all partitions 405 * which have connected with us and are targets of this packet. 406 * 407 * MAC-NOTE: For the XPNET driver, the MAC address contains the 408 * destination partid. If the destination partid octets are 0xffff, 409 * this packet is to be broadcast to all connected partitions. 410 */ 411 static netdev_tx_t 412 xpnet_dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev) 413 { 414 struct xpnet_pending_msg *queued_msg; 415 u64 start_addr, end_addr; 416 short dest_partid; 417 u16 embedded_bytes = 0; 418 419 dev_dbg(xpnet, ">skb->head=0x%p skb->data=0x%p skb->tail=0x%p " 420 "skb->end=0x%p skb->len=%d\n", (void *)skb->head, 421 (void *)skb->data, skb_tail_pointer(skb), skb_end_pointer(skb), 422 skb->len); 423 424 if (skb->data[0] == 0x33) { 425 dev_kfree_skb(skb); 426 return NETDEV_TX_OK; /* nothing needed to be done */ 427 } 428 429 /* 430 * The xpnet_pending_msg tracks how many outstanding 431 * xpc_send_notifies are relying on this skb. When none 432 * remain, release the skb. 433 */ 434 queued_msg = kmalloc(sizeof(struct xpnet_pending_msg), GFP_ATOMIC); 435 if (queued_msg == NULL) { 436 dev_warn(xpnet, "failed to kmalloc %ld bytes; dropping " 437 "packet\n", sizeof(struct xpnet_pending_msg)); 438 439 dev->stats.tx_errors++; 440 dev_kfree_skb(skb); 441 return NETDEV_TX_OK; 442 } 443 444 /* get the beginning of the first cacheline and end of last */ 445 start_addr = ((u64)skb->data & ~(L1_CACHE_BYTES - 1)); 446 end_addr = L1_CACHE_ALIGN((u64)skb_tail_pointer(skb)); 447 448 /* calculate how many bytes to embed in the XPC message */ 449 if (unlikely(skb->len <= XPNET_MSG_DATA_MAX)) { 450 /* skb->data does fit so embed */ 451 embedded_bytes = skb->len; 452 } 453 454 /* 455 * Since the send occurs asynchronously, we set the count to one 456 * and begin sending. Any sends that happen to complete before 457 * we are done sending will not free the skb. We will be left 458 * with that task during exit. This also handles the case of 459 * a packet destined for a partition which is no longer up. 460 */ 461 atomic_set(&queued_msg->use_count, 1); 462 queued_msg->skb = skb; 463 464 if (skb->data[0] == 0xff) { 465 /* we are being asked to broadcast to all partitions */ 466 for_each_set_bit(dest_partid, xpnet_broadcast_partitions, 467 xp_max_npartitions) { 468 469 xpnet_send(skb, queued_msg, start_addr, end_addr, 470 embedded_bytes, dest_partid); 471 } 472 } else { 473 dest_partid = (short)skb->data[XPNET_PARTID_OCTET + 1]; 474 dest_partid |= (short)skb->data[XPNET_PARTID_OCTET + 0] << 8; 475 476 if (dest_partid >= 0 && 477 dest_partid < xp_max_npartitions && 478 test_bit(dest_partid, xpnet_broadcast_partitions) != 0) { 479 480 xpnet_send(skb, queued_msg, start_addr, end_addr, 481 embedded_bytes, dest_partid); 482 } 483 } 484 485 dev->stats.tx_packets++; 486 dev->stats.tx_bytes += skb->len; 487 488 if (atomic_dec_return(&queued_msg->use_count) == 0) { 489 dev_kfree_skb(skb); 490 kfree(queued_msg); 491 } 492 493 return NETDEV_TX_OK; 494 } 495 496 /* 497 * Deal with transmit timeouts coming from the network layer. 498 */ 499 static void 500 xpnet_dev_tx_timeout(struct net_device *dev, unsigned int txqueue) 501 { 502 dev->stats.tx_errors++; 503 } 504 505 static const struct net_device_ops xpnet_netdev_ops = { 506 .ndo_open = xpnet_dev_open, 507 .ndo_stop = xpnet_dev_stop, 508 .ndo_start_xmit = xpnet_dev_hard_start_xmit, 509 .ndo_tx_timeout = xpnet_dev_tx_timeout, 510 .ndo_set_mac_address = eth_mac_addr, 511 .ndo_validate_addr = eth_validate_addr, 512 }; 513 514 static int __init 515 xpnet_init(void) 516 { 517 u8 addr[ETH_ALEN]; 518 int result; 519 520 if (!is_uv_system()) 521 return -ENODEV; 522 523 dev_info(xpnet, "registering network device %s\n", XPNET_DEVICE_NAME); 524 525 xpnet_broadcast_partitions = kcalloc(BITS_TO_LONGS(xp_max_npartitions), 526 sizeof(long), 527 GFP_KERNEL); 528 if (xpnet_broadcast_partitions == NULL) 529 return -ENOMEM; 530 531 /* 532 * use ether_setup() to init the majority of our device 533 * structure and then override the necessary pieces. 534 */ 535 xpnet_device = alloc_netdev(0, XPNET_DEVICE_NAME, NET_NAME_UNKNOWN, 536 ether_setup); 537 if (xpnet_device == NULL) { 538 kfree(xpnet_broadcast_partitions); 539 return -ENOMEM; 540 } 541 542 netif_carrier_off(xpnet_device); 543 544 xpnet_device->netdev_ops = &xpnet_netdev_ops; 545 xpnet_device->mtu = XPNET_DEF_MTU; 546 xpnet_device->min_mtu = XPNET_MIN_MTU; 547 xpnet_device->max_mtu = XPNET_MAX_MTU; 548 549 memset(addr, 0, sizeof(addr)); 550 /* 551 * Multicast assumes the LSB of the first octet is set for multicast 552 * MAC addresses. We chose the first octet of the MAC to be unlikely 553 * to collide with any vendor's officially issued MAC. 554 */ 555 addr[0] = 0x02; /* locally administered, no OUI */ 556 557 addr[XPNET_PARTID_OCTET + 1] = xp_partition_id; 558 addr[XPNET_PARTID_OCTET + 0] = (xp_partition_id >> 8); 559 eth_hw_addr_set(xpnet_device, addr); 560 561 /* 562 * ether_setup() sets this to a multicast device. We are 563 * really not supporting multicast at this time. 564 */ 565 xpnet_device->flags &= ~IFF_MULTICAST; 566 567 /* 568 * No need to checksum as it is a DMA transfer. The BTE will 569 * report an error if the data is not retrievable and the 570 * packet will be dropped. 571 */ 572 xpnet_device->features = NETIF_F_HW_CSUM; 573 574 result = register_netdev(xpnet_device); 575 if (result != 0) { 576 free_netdev(xpnet_device); 577 kfree(xpnet_broadcast_partitions); 578 } 579 580 return result; 581 } 582 583 module_init(xpnet_init); 584 585 static void __exit 586 xpnet_exit(void) 587 { 588 dev_info(xpnet, "unregistering network device %s\n", 589 xpnet_device[0].name); 590 591 unregister_netdev(xpnet_device); 592 free_netdev(xpnet_device); 593 kfree(xpnet_broadcast_partitions); 594 } 595 596 module_exit(xpnet_exit); 597 598 MODULE_AUTHOR("Silicon Graphics, Inc."); 599 MODULE_DESCRIPTION("Cross Partition Network adapter (XPNET)"); 600 MODULE_LICENSE("GPL"); 601