1 /* 2 * Copyright(c) 2015-2018 Intel Corporation. 3 * 4 * This file is provided under a dual BSD/GPLv2 license. When using or 5 * redistributing this file, you may do so under either license. 6 * 7 * GPL LICENSE SUMMARY 8 * 9 * This program is free software; you can redistribute it and/or modify 10 * it under the terms of version 2 of the GNU General Public License as 11 * published by the Free Software Foundation. 12 * 13 * This program is distributed in the hope that it will be useful, but 14 * WITHOUT ANY WARRANTY; without even the implied warranty of 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 16 * General Public License for more details. 17 * 18 * BSD LICENSE 19 * 20 * Redistribution and use in source and binary forms, with or without 21 * modification, are permitted provided that the following conditions 22 * are met: 23 * 24 * - Redistributions of source code must retain the above copyright 25 * notice, this list of conditions and the following disclaimer. 26 * - Redistributions in binary form must reproduce the above copyright 27 * notice, this list of conditions and the following disclaimer in 28 * the documentation and/or other materials provided with the 29 * distribution. 30 * - Neither the name of Intel Corporation nor the names of its 31 * contributors may be used to endorse or promote products derived 32 * from this software without specific prior written permission. 33 * 34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 45 * 46 */ 47 48 #include <linux/spinlock.h> 49 #include <linux/pci.h> 50 #include <linux/io.h> 51 #include <linux/delay.h> 52 #include <linux/netdevice.h> 53 #include <linux/vmalloc.h> 54 #include <linux/module.h> 55 #include <linux/prefetch.h> 56 #include <rdma/ib_verbs.h> 57 58 #include "hfi.h" 59 #include "trace.h" 60 #include "qp.h" 61 #include "sdma.h" 62 #include "debugfs.h" 63 #include "vnic.h" 64 #include "fault.h" 65 66 #undef pr_fmt 67 #define pr_fmt(fmt) DRIVER_NAME ": " fmt 68 69 /* 70 * The size has to be longer than this string, so we can append 71 * board/chip information to it in the initialization code. 72 */ 73 const char ib_hfi1_version[] = HFI1_DRIVER_VERSION "\n"; 74 75 DEFINE_MUTEX(hfi1_mutex); /* general driver use */ 76 77 unsigned int hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU; 78 module_param_named(max_mtu, hfi1_max_mtu, uint, S_IRUGO); 79 MODULE_PARM_DESC(max_mtu, "Set max MTU bytes, default is " __stringify( 80 HFI1_DEFAULT_MAX_MTU)); 81 82 unsigned int hfi1_cu = 1; 83 module_param_named(cu, hfi1_cu, uint, S_IRUGO); 84 MODULE_PARM_DESC(cu, "Credit return units"); 85 86 unsigned long hfi1_cap_mask = HFI1_CAP_MASK_DEFAULT; 87 static int hfi1_caps_set(const char *val, const struct kernel_param *kp); 88 static int hfi1_caps_get(char *buffer, const struct kernel_param *kp); 89 static const struct kernel_param_ops cap_ops = { 90 .set = hfi1_caps_set, 91 .get = hfi1_caps_get 92 }; 93 module_param_cb(cap_mask, &cap_ops, &hfi1_cap_mask, S_IWUSR | S_IRUGO); 94 MODULE_PARM_DESC(cap_mask, "Bit mask of enabled/disabled HW features"); 95 96 MODULE_LICENSE("Dual BSD/GPL"); 97 MODULE_DESCRIPTION("Intel Omni-Path Architecture driver"); 98 99 /* 100 * MAX_PKT_RCV is the max # if packets processed per receive interrupt. 101 */ 102 #define MAX_PKT_RECV 64 103 /* 104 * MAX_PKT_THREAD_RCV is the max # of packets processed before 105 * the qp_wait_list queue is flushed. 106 */ 107 #define MAX_PKT_RECV_THREAD (MAX_PKT_RECV * 4) 108 #define EGR_HEAD_UPDATE_THRESHOLD 16 109 110 struct hfi1_ib_stats hfi1_stats; 111 112 static int hfi1_caps_set(const char *val, const struct kernel_param *kp) 113 { 114 int ret = 0; 115 unsigned long *cap_mask_ptr = (unsigned long *)kp->arg, 116 cap_mask = *cap_mask_ptr, value, diff, 117 write_mask = ((HFI1_CAP_WRITABLE_MASK << HFI1_CAP_USER_SHIFT) | 118 HFI1_CAP_WRITABLE_MASK); 119 120 ret = kstrtoul(val, 0, &value); 121 if (ret) { 122 pr_warn("Invalid module parameter value for 'cap_mask'\n"); 123 goto done; 124 } 125 /* Get the changed bits (except the locked bit) */ 126 diff = value ^ (cap_mask & ~HFI1_CAP_LOCKED_SMASK); 127 128 /* Remove any bits that are not allowed to change after driver load */ 129 if (HFI1_CAP_LOCKED() && (diff & ~write_mask)) { 130 pr_warn("Ignoring non-writable capability bits %#lx\n", 131 diff & ~write_mask); 132 diff &= write_mask; 133 } 134 135 /* Mask off any reserved bits */ 136 diff &= ~HFI1_CAP_RESERVED_MASK; 137 /* Clear any previously set and changing bits */ 138 cap_mask &= ~diff; 139 /* Update the bits with the new capability */ 140 cap_mask |= (value & diff); 141 /* Check for any kernel/user restrictions */ 142 diff = (cap_mask & (HFI1_CAP_MUST_HAVE_KERN << HFI1_CAP_USER_SHIFT)) ^ 143 ((cap_mask & HFI1_CAP_MUST_HAVE_KERN) << HFI1_CAP_USER_SHIFT); 144 cap_mask &= ~diff; 145 /* Set the bitmask to the final set */ 146 *cap_mask_ptr = cap_mask; 147 done: 148 return ret; 149 } 150 151 static int hfi1_caps_get(char *buffer, const struct kernel_param *kp) 152 { 153 unsigned long cap_mask = *(unsigned long *)kp->arg; 154 155 cap_mask &= ~HFI1_CAP_LOCKED_SMASK; 156 cap_mask |= ((cap_mask & HFI1_CAP_K2U) << HFI1_CAP_USER_SHIFT); 157 158 return scnprintf(buffer, PAGE_SIZE, "0x%lx", cap_mask); 159 } 160 161 struct pci_dev *get_pci_dev(struct rvt_dev_info *rdi) 162 { 163 struct hfi1_ibdev *ibdev = container_of(rdi, struct hfi1_ibdev, rdi); 164 struct hfi1_devdata *dd = container_of(ibdev, 165 struct hfi1_devdata, verbs_dev); 166 return dd->pcidev; 167 } 168 169 /* 170 * Return count of units with at least one port ACTIVE. 171 */ 172 int hfi1_count_active_units(void) 173 { 174 struct hfi1_devdata *dd; 175 struct hfi1_pportdata *ppd; 176 unsigned long index, flags; 177 int pidx, nunits_active = 0; 178 179 xa_lock_irqsave(&hfi1_dev_table, flags); 180 xa_for_each(&hfi1_dev_table, index, dd) { 181 if (!(dd->flags & HFI1_PRESENT) || !dd->kregbase1) 182 continue; 183 for (pidx = 0; pidx < dd->num_pports; ++pidx) { 184 ppd = dd->pport + pidx; 185 if (ppd->lid && ppd->linkup) { 186 nunits_active++; 187 break; 188 } 189 } 190 } 191 xa_unlock_irqrestore(&hfi1_dev_table, flags); 192 return nunits_active; 193 } 194 195 /* 196 * Get address of eager buffer from it's index (allocated in chunks, not 197 * contiguous). 198 */ 199 static inline void *get_egrbuf(const struct hfi1_ctxtdata *rcd, u64 rhf, 200 u8 *update) 201 { 202 u32 idx = rhf_egr_index(rhf), offset = rhf_egr_buf_offset(rhf); 203 204 *update |= !(idx & (rcd->egrbufs.threshold - 1)) && !offset; 205 return (void *)(((u64)(rcd->egrbufs.rcvtids[idx].addr)) + 206 (offset * RCV_BUF_BLOCK_SIZE)); 207 } 208 209 static inline void *hfi1_get_header(struct hfi1_ctxtdata *rcd, 210 __le32 *rhf_addr) 211 { 212 u32 offset = rhf_hdrq_offset(rhf_to_cpu(rhf_addr)); 213 214 return (void *)(rhf_addr - rcd->rhf_offset + offset); 215 } 216 217 static inline struct ib_header *hfi1_get_msgheader(struct hfi1_ctxtdata *rcd, 218 __le32 *rhf_addr) 219 { 220 return (struct ib_header *)hfi1_get_header(rcd, rhf_addr); 221 } 222 223 static inline struct hfi1_16b_header 224 *hfi1_get_16B_header(struct hfi1_ctxtdata *rcd, 225 __le32 *rhf_addr) 226 { 227 return (struct hfi1_16b_header *)hfi1_get_header(rcd, rhf_addr); 228 } 229 230 /* 231 * Validate and encode the a given RcvArray Buffer size. 232 * The function will check whether the given size falls within 233 * allowed size ranges for the respective type and, optionally, 234 * return the proper encoding. 235 */ 236 int hfi1_rcvbuf_validate(u32 size, u8 type, u16 *encoded) 237 { 238 if (unlikely(!PAGE_ALIGNED(size))) 239 return 0; 240 if (unlikely(size < MIN_EAGER_BUFFER)) 241 return 0; 242 if (size > 243 (type == PT_EAGER ? MAX_EAGER_BUFFER : MAX_EXPECTED_BUFFER)) 244 return 0; 245 if (encoded) 246 *encoded = ilog2(size / PAGE_SIZE) + 1; 247 return 1; 248 } 249 250 static void rcv_hdrerr(struct hfi1_ctxtdata *rcd, struct hfi1_pportdata *ppd, 251 struct hfi1_packet *packet) 252 { 253 struct ib_header *rhdr = packet->hdr; 254 u32 rte = rhf_rcv_type_err(packet->rhf); 255 u32 mlid_base; 256 struct hfi1_ibport *ibp = rcd_to_iport(rcd); 257 struct hfi1_devdata *dd = ppd->dd; 258 struct hfi1_ibdev *verbs_dev = &dd->verbs_dev; 259 struct rvt_dev_info *rdi = &verbs_dev->rdi; 260 261 if ((packet->rhf & RHF_DC_ERR) && 262 hfi1_dbg_fault_suppress_err(verbs_dev)) 263 return; 264 265 if (packet->rhf & RHF_ICRC_ERR) 266 return; 267 268 if (packet->etype == RHF_RCV_TYPE_BYPASS) { 269 goto drop; 270 } else { 271 u8 lnh = ib_get_lnh(rhdr); 272 273 mlid_base = be16_to_cpu(IB_MULTICAST_LID_BASE); 274 if (lnh == HFI1_LRH_BTH) { 275 packet->ohdr = &rhdr->u.oth; 276 } else if (lnh == HFI1_LRH_GRH) { 277 packet->ohdr = &rhdr->u.l.oth; 278 packet->grh = &rhdr->u.l.grh; 279 } else { 280 goto drop; 281 } 282 } 283 284 if (packet->rhf & RHF_TID_ERR) { 285 /* For TIDERR and RC QPs preemptively schedule a NAK */ 286 u32 tlen = rhf_pkt_len(packet->rhf); /* in bytes */ 287 u32 dlid = ib_get_dlid(rhdr); 288 u32 qp_num; 289 290 /* Sanity check packet */ 291 if (tlen < 24) 292 goto drop; 293 294 /* Check for GRH */ 295 if (packet->grh) { 296 u32 vtf; 297 struct ib_grh *grh = packet->grh; 298 299 if (grh->next_hdr != IB_GRH_NEXT_HDR) 300 goto drop; 301 vtf = be32_to_cpu(grh->version_tclass_flow); 302 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION) 303 goto drop; 304 } 305 306 /* Get the destination QP number. */ 307 qp_num = ib_bth_get_qpn(packet->ohdr); 308 if (dlid < mlid_base) { 309 struct rvt_qp *qp; 310 unsigned long flags; 311 312 rcu_read_lock(); 313 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num); 314 if (!qp) { 315 rcu_read_unlock(); 316 goto drop; 317 } 318 319 /* 320 * Handle only RC QPs - for other QP types drop error 321 * packet. 322 */ 323 spin_lock_irqsave(&qp->r_lock, flags); 324 325 /* Check for valid receive state. */ 326 if (!(ib_rvt_state_ops[qp->state] & 327 RVT_PROCESS_RECV_OK)) { 328 ibp->rvp.n_pkt_drops++; 329 } 330 331 switch (qp->ibqp.qp_type) { 332 case IB_QPT_RC: 333 hfi1_rc_hdrerr(rcd, packet, qp); 334 break; 335 default: 336 /* For now don't handle any other QP types */ 337 break; 338 } 339 340 spin_unlock_irqrestore(&qp->r_lock, flags); 341 rcu_read_unlock(); 342 } /* Unicast QP */ 343 } /* Valid packet with TIDErr */ 344 345 /* handle "RcvTypeErr" flags */ 346 switch (rte) { 347 case RHF_RTE_ERROR_OP_CODE_ERR: 348 { 349 void *ebuf = NULL; 350 u8 opcode; 351 352 if (rhf_use_egr_bfr(packet->rhf)) 353 ebuf = packet->ebuf; 354 355 if (!ebuf) 356 goto drop; /* this should never happen */ 357 358 opcode = ib_bth_get_opcode(packet->ohdr); 359 if (opcode == IB_OPCODE_CNP) { 360 /* 361 * Only in pre-B0 h/w is the CNP_OPCODE handled 362 * via this code path. 363 */ 364 struct rvt_qp *qp = NULL; 365 u32 lqpn, rqpn; 366 u16 rlid; 367 u8 svc_type, sl, sc5; 368 369 sc5 = hfi1_9B_get_sc5(rhdr, packet->rhf); 370 sl = ibp->sc_to_sl[sc5]; 371 372 lqpn = ib_bth_get_qpn(packet->ohdr); 373 rcu_read_lock(); 374 qp = rvt_lookup_qpn(rdi, &ibp->rvp, lqpn); 375 if (!qp) { 376 rcu_read_unlock(); 377 goto drop; 378 } 379 380 switch (qp->ibqp.qp_type) { 381 case IB_QPT_UD: 382 rlid = 0; 383 rqpn = 0; 384 svc_type = IB_CC_SVCTYPE_UD; 385 break; 386 case IB_QPT_UC: 387 rlid = ib_get_slid(rhdr); 388 rqpn = qp->remote_qpn; 389 svc_type = IB_CC_SVCTYPE_UC; 390 break; 391 default: 392 rcu_read_unlock(); 393 goto drop; 394 } 395 396 process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type); 397 rcu_read_unlock(); 398 } 399 400 packet->rhf &= ~RHF_RCV_TYPE_ERR_SMASK; 401 break; 402 } 403 default: 404 break; 405 } 406 407 drop: 408 return; 409 } 410 411 static inline void init_packet(struct hfi1_ctxtdata *rcd, 412 struct hfi1_packet *packet) 413 { 414 packet->rsize = get_hdrqentsize(rcd); /* words */ 415 packet->maxcnt = get_hdrq_cnt(rcd) * packet->rsize; /* words */ 416 packet->rcd = rcd; 417 packet->updegr = 0; 418 packet->etail = -1; 419 packet->rhf_addr = get_rhf_addr(rcd); 420 packet->rhf = rhf_to_cpu(packet->rhf_addr); 421 packet->rhqoff = hfi1_rcd_head(rcd); 422 packet->numpkt = 0; 423 } 424 425 /* We support only two types - 9B and 16B for now */ 426 static const hfi1_handle_cnp hfi1_handle_cnp_tbl[2] = { 427 [HFI1_PKT_TYPE_9B] = &return_cnp, 428 [HFI1_PKT_TYPE_16B] = &return_cnp_16B 429 }; 430 431 /** 432 * hfi1_process_ecn_slowpath - Process FECN or BECN bits 433 * @qp: The packet's destination QP 434 * @pkt: The packet itself. 435 * @prescan: Is the caller the RXQ prescan 436 * 437 * Process the packet's FECN or BECN bits. By now, the packet 438 * has already been evaluated whether processing of those bit should 439 * be done. 440 * The significance of the @prescan argument is that if the caller 441 * is the RXQ prescan, a CNP will be send out instead of waiting for the 442 * normal packet processing to send an ACK with BECN set (or a CNP). 443 */ 444 bool hfi1_process_ecn_slowpath(struct rvt_qp *qp, struct hfi1_packet *pkt, 445 bool prescan) 446 { 447 struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 448 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); 449 struct ib_other_headers *ohdr = pkt->ohdr; 450 struct ib_grh *grh = pkt->grh; 451 u32 rqpn = 0; 452 u16 pkey; 453 u32 rlid, slid, dlid = 0; 454 u8 hdr_type, sc, svc_type, opcode; 455 bool is_mcast = false, ignore_fecn = false, do_cnp = false, 456 fecn, becn; 457 458 /* can be called from prescan */ 459 if (pkt->etype == RHF_RCV_TYPE_BYPASS) { 460 pkey = hfi1_16B_get_pkey(pkt->hdr); 461 sc = hfi1_16B_get_sc(pkt->hdr); 462 dlid = hfi1_16B_get_dlid(pkt->hdr); 463 slid = hfi1_16B_get_slid(pkt->hdr); 464 is_mcast = hfi1_is_16B_mcast(dlid); 465 opcode = ib_bth_get_opcode(ohdr); 466 hdr_type = HFI1_PKT_TYPE_16B; 467 fecn = hfi1_16B_get_fecn(pkt->hdr); 468 becn = hfi1_16B_get_becn(pkt->hdr); 469 } else { 470 pkey = ib_bth_get_pkey(ohdr); 471 sc = hfi1_9B_get_sc5(pkt->hdr, pkt->rhf); 472 dlid = qp->ibqp.qp_type != IB_QPT_UD ? ib_get_dlid(pkt->hdr) : 473 ppd->lid; 474 slid = ib_get_slid(pkt->hdr); 475 is_mcast = (dlid > be16_to_cpu(IB_MULTICAST_LID_BASE)) && 476 (dlid != be16_to_cpu(IB_LID_PERMISSIVE)); 477 opcode = ib_bth_get_opcode(ohdr); 478 hdr_type = HFI1_PKT_TYPE_9B; 479 fecn = ib_bth_get_fecn(ohdr); 480 becn = ib_bth_get_becn(ohdr); 481 } 482 483 switch (qp->ibqp.qp_type) { 484 case IB_QPT_UD: 485 rlid = slid; 486 rqpn = ib_get_sqpn(pkt->ohdr); 487 svc_type = IB_CC_SVCTYPE_UD; 488 break; 489 case IB_QPT_SMI: 490 case IB_QPT_GSI: 491 rlid = slid; 492 rqpn = ib_get_sqpn(pkt->ohdr); 493 svc_type = IB_CC_SVCTYPE_UD; 494 break; 495 case IB_QPT_UC: 496 rlid = rdma_ah_get_dlid(&qp->remote_ah_attr); 497 rqpn = qp->remote_qpn; 498 svc_type = IB_CC_SVCTYPE_UC; 499 break; 500 case IB_QPT_RC: 501 rlid = rdma_ah_get_dlid(&qp->remote_ah_attr); 502 rqpn = qp->remote_qpn; 503 svc_type = IB_CC_SVCTYPE_RC; 504 break; 505 default: 506 return false; 507 } 508 509 ignore_fecn = is_mcast || (opcode == IB_OPCODE_CNP) || 510 (opcode == IB_OPCODE_RC_ACKNOWLEDGE); 511 /* 512 * ACKNOWLEDGE packets do not get a CNP but this will be 513 * guarded by ignore_fecn above. 514 */ 515 do_cnp = prescan || 516 (opcode >= IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST && 517 opcode <= IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE) || 518 opcode == TID_OP(READ_RESP) || 519 opcode == TID_OP(ACK); 520 521 /* Call appropriate CNP handler */ 522 if (!ignore_fecn && do_cnp && fecn) 523 hfi1_handle_cnp_tbl[hdr_type](ibp, qp, rqpn, pkey, 524 dlid, rlid, sc, grh); 525 526 if (becn) { 527 u32 lqpn = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK; 528 u8 sl = ibp->sc_to_sl[sc]; 529 530 process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type); 531 } 532 return !ignore_fecn && fecn; 533 } 534 535 struct ps_mdata { 536 struct hfi1_ctxtdata *rcd; 537 u32 rsize; 538 u32 maxcnt; 539 u32 ps_head; 540 u32 ps_tail; 541 u32 ps_seq; 542 }; 543 544 static inline void init_ps_mdata(struct ps_mdata *mdata, 545 struct hfi1_packet *packet) 546 { 547 struct hfi1_ctxtdata *rcd = packet->rcd; 548 549 mdata->rcd = rcd; 550 mdata->rsize = packet->rsize; 551 mdata->maxcnt = packet->maxcnt; 552 mdata->ps_head = packet->rhqoff; 553 554 if (get_dma_rtail_setting(rcd)) { 555 mdata->ps_tail = get_rcvhdrtail(rcd); 556 if (rcd->ctxt == HFI1_CTRL_CTXT) 557 mdata->ps_seq = hfi1_seq_cnt(rcd); 558 else 559 mdata->ps_seq = 0; /* not used with DMA_RTAIL */ 560 } else { 561 mdata->ps_tail = 0; /* used only with DMA_RTAIL*/ 562 mdata->ps_seq = hfi1_seq_cnt(rcd); 563 } 564 } 565 566 static inline int ps_done(struct ps_mdata *mdata, u64 rhf, 567 struct hfi1_ctxtdata *rcd) 568 { 569 if (get_dma_rtail_setting(rcd)) 570 return mdata->ps_head == mdata->ps_tail; 571 return mdata->ps_seq != rhf_rcv_seq(rhf); 572 } 573 574 static inline int ps_skip(struct ps_mdata *mdata, u64 rhf, 575 struct hfi1_ctxtdata *rcd) 576 { 577 /* 578 * Control context can potentially receive an invalid rhf. 579 * Drop such packets. 580 */ 581 if ((rcd->ctxt == HFI1_CTRL_CTXT) && (mdata->ps_head != mdata->ps_tail)) 582 return mdata->ps_seq != rhf_rcv_seq(rhf); 583 584 return 0; 585 } 586 587 static inline void update_ps_mdata(struct ps_mdata *mdata, 588 struct hfi1_ctxtdata *rcd) 589 { 590 mdata->ps_head += mdata->rsize; 591 if (mdata->ps_head >= mdata->maxcnt) 592 mdata->ps_head = 0; 593 594 /* Control context must do seq counting */ 595 if (!get_dma_rtail_setting(rcd) || 596 rcd->ctxt == HFI1_CTRL_CTXT) 597 mdata->ps_seq = hfi1_seq_incr_wrap(mdata->ps_seq); 598 } 599 600 /* 601 * prescan_rxq - search through the receive queue looking for packets 602 * containing Excplicit Congestion Notifications (FECNs, or BECNs). 603 * When an ECN is found, process the Congestion Notification, and toggle 604 * it off. 605 * This is declared as a macro to allow quick checking of the port to avoid 606 * the overhead of a function call if not enabled. 607 */ 608 #define prescan_rxq(rcd, packet) \ 609 do { \ 610 if (rcd->ppd->cc_prescan) \ 611 __prescan_rxq(packet); \ 612 } while (0) 613 static void __prescan_rxq(struct hfi1_packet *packet) 614 { 615 struct hfi1_ctxtdata *rcd = packet->rcd; 616 struct ps_mdata mdata; 617 618 init_ps_mdata(&mdata, packet); 619 620 while (1) { 621 struct hfi1_ibport *ibp = rcd_to_iport(rcd); 622 __le32 *rhf_addr = (__le32 *)rcd->rcvhdrq + mdata.ps_head + 623 packet->rcd->rhf_offset; 624 struct rvt_qp *qp; 625 struct ib_header *hdr; 626 struct rvt_dev_info *rdi = &rcd->dd->verbs_dev.rdi; 627 u64 rhf = rhf_to_cpu(rhf_addr); 628 u32 etype = rhf_rcv_type(rhf), qpn, bth1; 629 u8 lnh; 630 631 if (ps_done(&mdata, rhf, rcd)) 632 break; 633 634 if (ps_skip(&mdata, rhf, rcd)) 635 goto next; 636 637 if (etype != RHF_RCV_TYPE_IB) 638 goto next; 639 640 packet->hdr = hfi1_get_msgheader(packet->rcd, rhf_addr); 641 hdr = packet->hdr; 642 lnh = ib_get_lnh(hdr); 643 644 if (lnh == HFI1_LRH_BTH) { 645 packet->ohdr = &hdr->u.oth; 646 packet->grh = NULL; 647 } else if (lnh == HFI1_LRH_GRH) { 648 packet->ohdr = &hdr->u.l.oth; 649 packet->grh = &hdr->u.l.grh; 650 } else { 651 goto next; /* just in case */ 652 } 653 654 if (!hfi1_may_ecn(packet)) 655 goto next; 656 657 bth1 = be32_to_cpu(packet->ohdr->bth[1]); 658 qpn = bth1 & RVT_QPN_MASK; 659 rcu_read_lock(); 660 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qpn); 661 662 if (!qp) { 663 rcu_read_unlock(); 664 goto next; 665 } 666 667 hfi1_process_ecn_slowpath(qp, packet, true); 668 rcu_read_unlock(); 669 670 /* turn off BECN, FECN */ 671 bth1 &= ~(IB_FECN_SMASK | IB_BECN_SMASK); 672 packet->ohdr->bth[1] = cpu_to_be32(bth1); 673 next: 674 update_ps_mdata(&mdata, rcd); 675 } 676 } 677 678 static void process_rcv_qp_work(struct hfi1_packet *packet) 679 { 680 struct rvt_qp *qp, *nqp; 681 struct hfi1_ctxtdata *rcd = packet->rcd; 682 683 /* 684 * Iterate over all QPs waiting to respond. 685 * The list won't change since the IRQ is only run on one CPU. 686 */ 687 list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) { 688 list_del_init(&qp->rspwait); 689 if (qp->r_flags & RVT_R_RSP_NAK) { 690 qp->r_flags &= ~RVT_R_RSP_NAK; 691 packet->qp = qp; 692 hfi1_send_rc_ack(packet, 0); 693 } 694 if (qp->r_flags & RVT_R_RSP_SEND) { 695 unsigned long flags; 696 697 qp->r_flags &= ~RVT_R_RSP_SEND; 698 spin_lock_irqsave(&qp->s_lock, flags); 699 if (ib_rvt_state_ops[qp->state] & 700 RVT_PROCESS_OR_FLUSH_SEND) 701 hfi1_schedule_send(qp); 702 spin_unlock_irqrestore(&qp->s_lock, flags); 703 } 704 rvt_put_qp(qp); 705 } 706 } 707 708 static noinline int max_packet_exceeded(struct hfi1_packet *packet, int thread) 709 { 710 if (thread) { 711 if ((packet->numpkt & (MAX_PKT_RECV_THREAD - 1)) == 0) 712 /* allow defered processing */ 713 process_rcv_qp_work(packet); 714 cond_resched(); 715 return RCV_PKT_OK; 716 } else { 717 this_cpu_inc(*packet->rcd->dd->rcv_limit); 718 return RCV_PKT_LIMIT; 719 } 720 } 721 722 static inline int check_max_packet(struct hfi1_packet *packet, int thread) 723 { 724 int ret = RCV_PKT_OK; 725 726 if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0)) 727 ret = max_packet_exceeded(packet, thread); 728 return ret; 729 } 730 731 static noinline int skip_rcv_packet(struct hfi1_packet *packet, int thread) 732 { 733 int ret; 734 735 packet->rcd->dd->ctx0_seq_drop++; 736 /* Set up for the next packet */ 737 packet->rhqoff += packet->rsize; 738 if (packet->rhqoff >= packet->maxcnt) 739 packet->rhqoff = 0; 740 741 packet->numpkt++; 742 ret = check_max_packet(packet, thread); 743 744 packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff + 745 packet->rcd->rhf_offset; 746 packet->rhf = rhf_to_cpu(packet->rhf_addr); 747 748 return ret; 749 } 750 751 static inline int process_rcv_packet(struct hfi1_packet *packet, int thread) 752 { 753 int ret; 754 755 packet->etype = rhf_rcv_type(packet->rhf); 756 757 /* total length */ 758 packet->tlen = rhf_pkt_len(packet->rhf); /* in bytes */ 759 /* retrieve eager buffer details */ 760 packet->ebuf = NULL; 761 if (rhf_use_egr_bfr(packet->rhf)) { 762 packet->etail = rhf_egr_index(packet->rhf); 763 packet->ebuf = get_egrbuf(packet->rcd, packet->rhf, 764 &packet->updegr); 765 /* 766 * Prefetch the contents of the eager buffer. It is 767 * OK to send a negative length to prefetch_range(). 768 * The +2 is the size of the RHF. 769 */ 770 prefetch_range(packet->ebuf, 771 packet->tlen - ((get_hdrqentsize(packet->rcd) - 772 (rhf_hdrq_offset(packet->rhf) 773 + 2)) * 4)); 774 } 775 776 /* 777 * Call a type specific handler for the packet. We 778 * should be able to trust that etype won't be beyond 779 * the range of valid indexes. If so something is really 780 * wrong and we can probably just let things come 781 * crashing down. There is no need to eat another 782 * comparison in this performance critical code. 783 */ 784 packet->rcd->rhf_rcv_function_map[packet->etype](packet); 785 packet->numpkt++; 786 787 /* Set up for the next packet */ 788 packet->rhqoff += packet->rsize; 789 if (packet->rhqoff >= packet->maxcnt) 790 packet->rhqoff = 0; 791 792 ret = check_max_packet(packet, thread); 793 794 packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff + 795 packet->rcd->rhf_offset; 796 packet->rhf = rhf_to_cpu(packet->rhf_addr); 797 798 return ret; 799 } 800 801 static inline void process_rcv_update(int last, struct hfi1_packet *packet) 802 { 803 /* 804 * Update head regs etc., every 16 packets, if not last pkt, 805 * to help prevent rcvhdrq overflows, when many packets 806 * are processed and queue is nearly full. 807 * Don't request an interrupt for intermediate updates. 808 */ 809 if (!last && !(packet->numpkt & 0xf)) { 810 update_usrhead(packet->rcd, packet->rhqoff, packet->updegr, 811 packet->etail, 0, 0); 812 packet->updegr = 0; 813 } 814 packet->grh = NULL; 815 } 816 817 static inline void finish_packet(struct hfi1_packet *packet) 818 { 819 /* 820 * Nothing we need to free for the packet. 821 * 822 * The only thing we need to do is a final update and call for an 823 * interrupt 824 */ 825 update_usrhead(packet->rcd, hfi1_rcd_head(packet->rcd), packet->updegr, 826 packet->etail, rcv_intr_dynamic, packet->numpkt); 827 } 828 829 /* 830 * Handle receive interrupts when using the no dma rtail option. 831 */ 832 int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata *rcd, int thread) 833 { 834 int last = RCV_PKT_OK; 835 struct hfi1_packet packet; 836 837 init_packet(rcd, &packet); 838 if (last_rcv_seq(rcd, rhf_rcv_seq(packet.rhf))) { 839 last = RCV_PKT_DONE; 840 goto bail; 841 } 842 843 prescan_rxq(rcd, &packet); 844 845 while (last == RCV_PKT_OK) { 846 last = process_rcv_packet(&packet, thread); 847 if (hfi1_seq_incr(rcd, rhf_rcv_seq(packet.rhf))) 848 last = RCV_PKT_DONE; 849 process_rcv_update(last, &packet); 850 } 851 process_rcv_qp_work(&packet); 852 hfi1_set_rcd_head(rcd, packet.rhqoff); 853 bail: 854 finish_packet(&packet); 855 return last; 856 } 857 858 int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata *rcd, int thread) 859 { 860 u32 hdrqtail; 861 int last = RCV_PKT_OK; 862 struct hfi1_packet packet; 863 864 init_packet(rcd, &packet); 865 hdrqtail = get_rcvhdrtail(rcd); 866 if (packet.rhqoff == hdrqtail) { 867 last = RCV_PKT_DONE; 868 goto bail; 869 } 870 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */ 871 872 prescan_rxq(rcd, &packet); 873 874 while (last == RCV_PKT_OK) { 875 last = process_rcv_packet(&packet, thread); 876 if (packet.rhqoff == hdrqtail) 877 last = RCV_PKT_DONE; 878 process_rcv_update(last, &packet); 879 } 880 process_rcv_qp_work(&packet); 881 hfi1_set_rcd_head(rcd, packet.rhqoff); 882 bail: 883 finish_packet(&packet); 884 return last; 885 } 886 887 static void set_all_fastpath(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd) 888 { 889 u16 i; 890 891 /* 892 * For dynamically allocated kernel contexts (like vnic) switch 893 * interrupt handler only for that context. Otherwise, switch 894 * interrupt handler for all statically allocated kernel contexts. 895 */ 896 if (rcd->ctxt >= dd->first_dyn_alloc_ctxt && !rcd->is_vnic) { 897 hfi1_rcd_get(rcd); 898 hfi1_set_fast(rcd); 899 hfi1_rcd_put(rcd); 900 return; 901 } 902 903 for (i = HFI1_CTRL_CTXT + 1; i < dd->num_rcv_contexts; i++) { 904 rcd = hfi1_rcd_get_by_index(dd, i); 905 if (rcd && (i < dd->first_dyn_alloc_ctxt || rcd->is_vnic)) 906 hfi1_set_fast(rcd); 907 hfi1_rcd_put(rcd); 908 } 909 } 910 911 void set_all_slowpath(struct hfi1_devdata *dd) 912 { 913 struct hfi1_ctxtdata *rcd; 914 u16 i; 915 916 /* HFI1_CTRL_CTXT must always use the slow path interrupt handler */ 917 for (i = HFI1_CTRL_CTXT + 1; i < dd->num_rcv_contexts; i++) { 918 rcd = hfi1_rcd_get_by_index(dd, i); 919 if (!rcd) 920 continue; 921 if (i < dd->first_dyn_alloc_ctxt || rcd->is_vnic) 922 rcd->do_interrupt = rcd->slow_handler; 923 924 hfi1_rcd_put(rcd); 925 } 926 } 927 928 static bool __set_armed_to_active(struct hfi1_packet *packet) 929 { 930 u8 etype = rhf_rcv_type(packet->rhf); 931 u8 sc = SC15_PACKET; 932 933 if (etype == RHF_RCV_TYPE_IB) { 934 struct ib_header *hdr = hfi1_get_msgheader(packet->rcd, 935 packet->rhf_addr); 936 sc = hfi1_9B_get_sc5(hdr, packet->rhf); 937 } else if (etype == RHF_RCV_TYPE_BYPASS) { 938 struct hfi1_16b_header *hdr = hfi1_get_16B_header( 939 packet->rcd, 940 packet->rhf_addr); 941 sc = hfi1_16B_get_sc(hdr); 942 } 943 if (sc != SC15_PACKET) { 944 int hwstate = driver_lstate(packet->rcd->ppd); 945 struct work_struct *lsaw = 946 &packet->rcd->ppd->linkstate_active_work; 947 948 if (hwstate != IB_PORT_ACTIVE) { 949 dd_dev_info(packet->rcd->dd, 950 "Unexpected link state %s\n", 951 opa_lstate_name(hwstate)); 952 return false; 953 } 954 955 queue_work(packet->rcd->ppd->link_wq, lsaw); 956 return true; 957 } 958 return false; 959 } 960 961 /** 962 * armed to active - the fast path for armed to active 963 * @packet: the packet structure 964 * 965 * Return true if packet processing needs to bail. 966 */ 967 static bool set_armed_to_active(struct hfi1_packet *packet) 968 { 969 if (likely(packet->rcd->ppd->host_link_state != HLS_UP_ARMED)) 970 return false; 971 return __set_armed_to_active(packet); 972 } 973 974 /* 975 * handle_receive_interrupt - receive a packet 976 * @rcd: the context 977 * 978 * Called from interrupt handler for errors or receive interrupt. 979 * This is the slow path interrupt handler. 980 */ 981 int handle_receive_interrupt(struct hfi1_ctxtdata *rcd, int thread) 982 { 983 struct hfi1_devdata *dd = rcd->dd; 984 u32 hdrqtail; 985 int needset, last = RCV_PKT_OK; 986 struct hfi1_packet packet; 987 int skip_pkt = 0; 988 989 /* Control context will always use the slow path interrupt handler */ 990 needset = (rcd->ctxt == HFI1_CTRL_CTXT) ? 0 : 1; 991 992 init_packet(rcd, &packet); 993 994 if (!get_dma_rtail_setting(rcd)) { 995 if (last_rcv_seq(rcd, rhf_rcv_seq(packet.rhf))) { 996 last = RCV_PKT_DONE; 997 goto bail; 998 } 999 hdrqtail = 0; 1000 } else { 1001 hdrqtail = get_rcvhdrtail(rcd); 1002 if (packet.rhqoff == hdrqtail) { 1003 last = RCV_PKT_DONE; 1004 goto bail; 1005 } 1006 smp_rmb(); /* prevent speculative reads of dma'ed hdrq */ 1007 1008 /* 1009 * Control context can potentially receive an invalid 1010 * rhf. Drop such packets. 1011 */ 1012 if (rcd->ctxt == HFI1_CTRL_CTXT) 1013 if (last_rcv_seq(rcd, rhf_rcv_seq(packet.rhf))) 1014 skip_pkt = 1; 1015 } 1016 1017 prescan_rxq(rcd, &packet); 1018 1019 while (last == RCV_PKT_OK) { 1020 if (hfi1_need_drop(dd)) { 1021 /* On to the next packet */ 1022 packet.rhqoff += packet.rsize; 1023 packet.rhf_addr = (__le32 *)rcd->rcvhdrq + 1024 packet.rhqoff + 1025 rcd->rhf_offset; 1026 packet.rhf = rhf_to_cpu(packet.rhf_addr); 1027 1028 } else if (skip_pkt) { 1029 last = skip_rcv_packet(&packet, thread); 1030 skip_pkt = 0; 1031 } else { 1032 if (set_armed_to_active(&packet)) 1033 goto bail; 1034 last = process_rcv_packet(&packet, thread); 1035 } 1036 1037 if (!get_dma_rtail_setting(rcd)) { 1038 if (hfi1_seq_incr(rcd, rhf_rcv_seq(packet.rhf))) 1039 last = RCV_PKT_DONE; 1040 } else { 1041 if (packet.rhqoff == hdrqtail) 1042 last = RCV_PKT_DONE; 1043 /* 1044 * Control context can potentially receive an invalid 1045 * rhf. Drop such packets. 1046 */ 1047 if (rcd->ctxt == HFI1_CTRL_CTXT) { 1048 bool lseq; 1049 1050 lseq = hfi1_seq_incr(rcd, 1051 rhf_rcv_seq(packet.rhf)); 1052 if (!last && lseq) 1053 skip_pkt = 1; 1054 } 1055 } 1056 1057 if (needset) { 1058 needset = false; 1059 set_all_fastpath(dd, rcd); 1060 } 1061 process_rcv_update(last, &packet); 1062 } 1063 1064 process_rcv_qp_work(&packet); 1065 hfi1_set_rcd_head(rcd, packet.rhqoff); 1066 1067 bail: 1068 /* 1069 * Always write head at end, and setup rcv interrupt, even 1070 * if no packets were processed. 1071 */ 1072 finish_packet(&packet); 1073 return last; 1074 } 1075 1076 /* 1077 * We may discover in the interrupt that the hardware link state has 1078 * changed from ARMED to ACTIVE (due to the arrival of a non-SC15 packet), 1079 * and we need to update the driver's notion of the link state. We cannot 1080 * run set_link_state from interrupt context, so we queue this function on 1081 * a workqueue. 1082 * 1083 * We delay the regular interrupt processing until after the state changes 1084 * so that the link will be in the correct state by the time any application 1085 * we wake up attempts to send a reply to any message it received. 1086 * (Subsequent receive interrupts may possibly force the wakeup before we 1087 * update the link state.) 1088 * 1089 * The rcd is freed in hfi1_free_ctxtdata after hfi1_postinit_cleanup invokes 1090 * dd->f_cleanup(dd) to disable the interrupt handler and flush workqueues, 1091 * so we're safe from use-after-free of the rcd. 1092 */ 1093 void receive_interrupt_work(struct work_struct *work) 1094 { 1095 struct hfi1_pportdata *ppd = container_of(work, struct hfi1_pportdata, 1096 linkstate_active_work); 1097 struct hfi1_devdata *dd = ppd->dd; 1098 struct hfi1_ctxtdata *rcd; 1099 u16 i; 1100 1101 /* Received non-SC15 packet implies neighbor_normal */ 1102 ppd->neighbor_normal = 1; 1103 set_link_state(ppd, HLS_UP_ACTIVE); 1104 1105 /* 1106 * Interrupt all statically allocated kernel contexts that could 1107 * have had an interrupt during auto activation. 1108 */ 1109 for (i = HFI1_CTRL_CTXT; i < dd->first_dyn_alloc_ctxt; i++) { 1110 rcd = hfi1_rcd_get_by_index(dd, i); 1111 if (rcd) 1112 force_recv_intr(rcd); 1113 hfi1_rcd_put(rcd); 1114 } 1115 } 1116 1117 /* 1118 * Convert a given MTU size to the on-wire MAD packet enumeration. 1119 * Return -1 if the size is invalid. 1120 */ 1121 int mtu_to_enum(u32 mtu, int default_if_bad) 1122 { 1123 switch (mtu) { 1124 case 0: return OPA_MTU_0; 1125 case 256: return OPA_MTU_256; 1126 case 512: return OPA_MTU_512; 1127 case 1024: return OPA_MTU_1024; 1128 case 2048: return OPA_MTU_2048; 1129 case 4096: return OPA_MTU_4096; 1130 case 8192: return OPA_MTU_8192; 1131 case 10240: return OPA_MTU_10240; 1132 } 1133 return default_if_bad; 1134 } 1135 1136 u16 enum_to_mtu(int mtu) 1137 { 1138 switch (mtu) { 1139 case OPA_MTU_0: return 0; 1140 case OPA_MTU_256: return 256; 1141 case OPA_MTU_512: return 512; 1142 case OPA_MTU_1024: return 1024; 1143 case OPA_MTU_2048: return 2048; 1144 case OPA_MTU_4096: return 4096; 1145 case OPA_MTU_8192: return 8192; 1146 case OPA_MTU_10240: return 10240; 1147 default: return 0xffff; 1148 } 1149 } 1150 1151 /* 1152 * set_mtu - set the MTU 1153 * @ppd: the per port data 1154 * 1155 * We can handle "any" incoming size, the issue here is whether we 1156 * need to restrict our outgoing size. We do not deal with what happens 1157 * to programs that are already running when the size changes. 1158 */ 1159 int set_mtu(struct hfi1_pportdata *ppd) 1160 { 1161 struct hfi1_devdata *dd = ppd->dd; 1162 int i, drain, ret = 0, is_up = 0; 1163 1164 ppd->ibmtu = 0; 1165 for (i = 0; i < ppd->vls_supported; i++) 1166 if (ppd->ibmtu < dd->vld[i].mtu) 1167 ppd->ibmtu = dd->vld[i].mtu; 1168 ppd->ibmaxlen = ppd->ibmtu + lrh_max_header_bytes(ppd->dd); 1169 1170 mutex_lock(&ppd->hls_lock); 1171 if (ppd->host_link_state == HLS_UP_INIT || 1172 ppd->host_link_state == HLS_UP_ARMED || 1173 ppd->host_link_state == HLS_UP_ACTIVE) 1174 is_up = 1; 1175 1176 drain = !is_ax(dd) && is_up; 1177 1178 if (drain) 1179 /* 1180 * MTU is specified per-VL. To ensure that no packet gets 1181 * stuck (due, e.g., to the MTU for the packet's VL being 1182 * reduced), empty the per-VL FIFOs before adjusting MTU. 1183 */ 1184 ret = stop_drain_data_vls(dd); 1185 1186 if (ret) { 1187 dd_dev_err(dd, "%s: cannot stop/drain VLs - refusing to change per-VL MTUs\n", 1188 __func__); 1189 goto err; 1190 } 1191 1192 hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_MTU, 0); 1193 1194 if (drain) 1195 open_fill_data_vls(dd); /* reopen all VLs */ 1196 1197 err: 1198 mutex_unlock(&ppd->hls_lock); 1199 1200 return ret; 1201 } 1202 1203 int hfi1_set_lid(struct hfi1_pportdata *ppd, u32 lid, u8 lmc) 1204 { 1205 struct hfi1_devdata *dd = ppd->dd; 1206 1207 ppd->lid = lid; 1208 ppd->lmc = lmc; 1209 hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_LIDLMC, 0); 1210 1211 dd_dev_info(dd, "port %u: got a lid: 0x%x\n", ppd->port, lid); 1212 1213 return 0; 1214 } 1215 1216 void shutdown_led_override(struct hfi1_pportdata *ppd) 1217 { 1218 struct hfi1_devdata *dd = ppd->dd; 1219 1220 /* 1221 * This pairs with the memory barrier in hfi1_start_led_override to 1222 * ensure that we read the correct state of LED beaconing represented 1223 * by led_override_timer_active 1224 */ 1225 smp_rmb(); 1226 if (atomic_read(&ppd->led_override_timer_active)) { 1227 del_timer_sync(&ppd->led_override_timer); 1228 atomic_set(&ppd->led_override_timer_active, 0); 1229 /* Ensure the atomic_set is visible to all CPUs */ 1230 smp_wmb(); 1231 } 1232 1233 /* Hand control of the LED to the DC for normal operation */ 1234 write_csr(dd, DCC_CFG_LED_CNTRL, 0); 1235 } 1236 1237 static void run_led_override(struct timer_list *t) 1238 { 1239 struct hfi1_pportdata *ppd = from_timer(ppd, t, led_override_timer); 1240 struct hfi1_devdata *dd = ppd->dd; 1241 unsigned long timeout; 1242 int phase_idx; 1243 1244 if (!(dd->flags & HFI1_INITTED)) 1245 return; 1246 1247 phase_idx = ppd->led_override_phase & 1; 1248 1249 setextled(dd, phase_idx); 1250 1251 timeout = ppd->led_override_vals[phase_idx]; 1252 1253 /* Set up for next phase */ 1254 ppd->led_override_phase = !ppd->led_override_phase; 1255 1256 mod_timer(&ppd->led_override_timer, jiffies + timeout); 1257 } 1258 1259 /* 1260 * To have the LED blink in a particular pattern, provide timeon and timeoff 1261 * in milliseconds. 1262 * To turn off custom blinking and return to normal operation, use 1263 * shutdown_led_override() 1264 */ 1265 void hfi1_start_led_override(struct hfi1_pportdata *ppd, unsigned int timeon, 1266 unsigned int timeoff) 1267 { 1268 if (!(ppd->dd->flags & HFI1_INITTED)) 1269 return; 1270 1271 /* Convert to jiffies for direct use in timer */ 1272 ppd->led_override_vals[0] = msecs_to_jiffies(timeoff); 1273 ppd->led_override_vals[1] = msecs_to_jiffies(timeon); 1274 1275 /* Arbitrarily start from LED on phase */ 1276 ppd->led_override_phase = 1; 1277 1278 /* 1279 * If the timer has not already been started, do so. Use a "quick" 1280 * timeout so the handler will be called soon to look at our request. 1281 */ 1282 if (!timer_pending(&ppd->led_override_timer)) { 1283 timer_setup(&ppd->led_override_timer, run_led_override, 0); 1284 ppd->led_override_timer.expires = jiffies + 1; 1285 add_timer(&ppd->led_override_timer); 1286 atomic_set(&ppd->led_override_timer_active, 1); 1287 /* Ensure the atomic_set is visible to all CPUs */ 1288 smp_wmb(); 1289 } 1290 } 1291 1292 /** 1293 * hfi1_reset_device - reset the chip if possible 1294 * @unit: the device to reset 1295 * 1296 * Whether or not reset is successful, we attempt to re-initialize the chip 1297 * (that is, much like a driver unload/reload). We clear the INITTED flag 1298 * so that the various entry points will fail until we reinitialize. For 1299 * now, we only allow this if no user contexts are open that use chip resources 1300 */ 1301 int hfi1_reset_device(int unit) 1302 { 1303 int ret; 1304 struct hfi1_devdata *dd = hfi1_lookup(unit); 1305 struct hfi1_pportdata *ppd; 1306 int pidx; 1307 1308 if (!dd) { 1309 ret = -ENODEV; 1310 goto bail; 1311 } 1312 1313 dd_dev_info(dd, "Reset on unit %u requested\n", unit); 1314 1315 if (!dd->kregbase1 || !(dd->flags & HFI1_PRESENT)) { 1316 dd_dev_info(dd, 1317 "Invalid unit number %u or not initialized or not present\n", 1318 unit); 1319 ret = -ENXIO; 1320 goto bail; 1321 } 1322 1323 /* If there are any user/vnic contexts, we cannot reset */ 1324 mutex_lock(&hfi1_mutex); 1325 if (dd->rcd) 1326 if (hfi1_stats.sps_ctxts) { 1327 mutex_unlock(&hfi1_mutex); 1328 ret = -EBUSY; 1329 goto bail; 1330 } 1331 mutex_unlock(&hfi1_mutex); 1332 1333 for (pidx = 0; pidx < dd->num_pports; ++pidx) { 1334 ppd = dd->pport + pidx; 1335 1336 shutdown_led_override(ppd); 1337 } 1338 if (dd->flags & HFI1_HAS_SEND_DMA) 1339 sdma_exit(dd); 1340 1341 hfi1_reset_cpu_counters(dd); 1342 1343 ret = hfi1_init(dd, 1); 1344 1345 if (ret) 1346 dd_dev_err(dd, 1347 "Reinitialize unit %u after reset failed with %d\n", 1348 unit, ret); 1349 else 1350 dd_dev_info(dd, "Reinitialized unit %u after resetting\n", 1351 unit); 1352 1353 bail: 1354 return ret; 1355 } 1356 1357 static inline void hfi1_setup_ib_header(struct hfi1_packet *packet) 1358 { 1359 packet->hdr = (struct hfi1_ib_message_header *) 1360 hfi1_get_msgheader(packet->rcd, 1361 packet->rhf_addr); 1362 packet->hlen = (u8 *)packet->rhf_addr - (u8 *)packet->hdr; 1363 } 1364 1365 static int hfi1_bypass_ingress_pkt_check(struct hfi1_packet *packet) 1366 { 1367 struct hfi1_pportdata *ppd = packet->rcd->ppd; 1368 1369 /* slid and dlid cannot be 0 */ 1370 if ((!packet->slid) || (!packet->dlid)) 1371 return -EINVAL; 1372 1373 /* Compare port lid with incoming packet dlid */ 1374 if ((!(hfi1_is_16B_mcast(packet->dlid))) && 1375 (packet->dlid != 1376 opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B))) { 1377 if ((packet->dlid & ~((1 << ppd->lmc) - 1)) != ppd->lid) 1378 return -EINVAL; 1379 } 1380 1381 /* No multicast packets with SC15 */ 1382 if ((hfi1_is_16B_mcast(packet->dlid)) && (packet->sc == 0xF)) 1383 return -EINVAL; 1384 1385 /* Packets with permissive DLID always on SC15 */ 1386 if ((packet->dlid == opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 1387 16B)) && 1388 (packet->sc != 0xF)) 1389 return -EINVAL; 1390 1391 return 0; 1392 } 1393 1394 static int hfi1_setup_9B_packet(struct hfi1_packet *packet) 1395 { 1396 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd); 1397 struct ib_header *hdr; 1398 u8 lnh; 1399 1400 hfi1_setup_ib_header(packet); 1401 hdr = packet->hdr; 1402 1403 lnh = ib_get_lnh(hdr); 1404 if (lnh == HFI1_LRH_BTH) { 1405 packet->ohdr = &hdr->u.oth; 1406 packet->grh = NULL; 1407 } else if (lnh == HFI1_LRH_GRH) { 1408 u32 vtf; 1409 1410 packet->ohdr = &hdr->u.l.oth; 1411 packet->grh = &hdr->u.l.grh; 1412 if (packet->grh->next_hdr != IB_GRH_NEXT_HDR) 1413 goto drop; 1414 vtf = be32_to_cpu(packet->grh->version_tclass_flow); 1415 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION) 1416 goto drop; 1417 } else { 1418 goto drop; 1419 } 1420 1421 /* Query commonly used fields from packet header */ 1422 packet->payload = packet->ebuf; 1423 packet->opcode = ib_bth_get_opcode(packet->ohdr); 1424 packet->slid = ib_get_slid(hdr); 1425 packet->dlid = ib_get_dlid(hdr); 1426 if (unlikely((packet->dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) && 1427 (packet->dlid != be16_to_cpu(IB_LID_PERMISSIVE)))) 1428 packet->dlid += opa_get_mcast_base(OPA_MCAST_NR) - 1429 be16_to_cpu(IB_MULTICAST_LID_BASE); 1430 packet->sl = ib_get_sl(hdr); 1431 packet->sc = hfi1_9B_get_sc5(hdr, packet->rhf); 1432 packet->pad = ib_bth_get_pad(packet->ohdr); 1433 packet->extra_byte = 0; 1434 packet->pkey = ib_bth_get_pkey(packet->ohdr); 1435 packet->migrated = ib_bth_is_migration(packet->ohdr); 1436 1437 return 0; 1438 drop: 1439 ibp->rvp.n_pkt_drops++; 1440 return -EINVAL; 1441 } 1442 1443 static int hfi1_setup_bypass_packet(struct hfi1_packet *packet) 1444 { 1445 /* 1446 * Bypass packets have a different header/payload split 1447 * compared to an IB packet. 1448 * Current split is set such that 16 bytes of the actual 1449 * header is in the header buffer and the remining is in 1450 * the eager buffer. We chose 16 since hfi1 driver only 1451 * supports 16B bypass packets and we will be able to 1452 * receive the entire LRH with such a split. 1453 */ 1454 1455 struct hfi1_ctxtdata *rcd = packet->rcd; 1456 struct hfi1_pportdata *ppd = rcd->ppd; 1457 struct hfi1_ibport *ibp = &ppd->ibport_data; 1458 u8 l4; 1459 1460 packet->hdr = (struct hfi1_16b_header *) 1461 hfi1_get_16B_header(packet->rcd, 1462 packet->rhf_addr); 1463 l4 = hfi1_16B_get_l4(packet->hdr); 1464 if (l4 == OPA_16B_L4_IB_LOCAL) { 1465 packet->ohdr = packet->ebuf; 1466 packet->grh = NULL; 1467 packet->opcode = ib_bth_get_opcode(packet->ohdr); 1468 packet->pad = hfi1_16B_bth_get_pad(packet->ohdr); 1469 /* hdr_len_by_opcode already has an IB LRH factored in */ 1470 packet->hlen = hdr_len_by_opcode[packet->opcode] + 1471 (LRH_16B_BYTES - LRH_9B_BYTES); 1472 packet->migrated = opa_bth_is_migration(packet->ohdr); 1473 } else if (l4 == OPA_16B_L4_IB_GLOBAL) { 1474 u32 vtf; 1475 u8 grh_len = sizeof(struct ib_grh); 1476 1477 packet->ohdr = packet->ebuf + grh_len; 1478 packet->grh = packet->ebuf; 1479 packet->opcode = ib_bth_get_opcode(packet->ohdr); 1480 packet->pad = hfi1_16B_bth_get_pad(packet->ohdr); 1481 /* hdr_len_by_opcode already has an IB LRH factored in */ 1482 packet->hlen = hdr_len_by_opcode[packet->opcode] + 1483 (LRH_16B_BYTES - LRH_9B_BYTES) + grh_len; 1484 packet->migrated = opa_bth_is_migration(packet->ohdr); 1485 1486 if (packet->grh->next_hdr != IB_GRH_NEXT_HDR) 1487 goto drop; 1488 vtf = be32_to_cpu(packet->grh->version_tclass_flow); 1489 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION) 1490 goto drop; 1491 } else if (l4 == OPA_16B_L4_FM) { 1492 packet->mgmt = packet->ebuf; 1493 packet->ohdr = NULL; 1494 packet->grh = NULL; 1495 packet->opcode = IB_OPCODE_UD_SEND_ONLY; 1496 packet->pad = OPA_16B_L4_FM_PAD; 1497 packet->hlen = OPA_16B_L4_FM_HLEN; 1498 packet->migrated = false; 1499 } else { 1500 goto drop; 1501 } 1502 1503 /* Query commonly used fields from packet header */ 1504 packet->payload = packet->ebuf + packet->hlen - LRH_16B_BYTES; 1505 packet->slid = hfi1_16B_get_slid(packet->hdr); 1506 packet->dlid = hfi1_16B_get_dlid(packet->hdr); 1507 if (unlikely(hfi1_is_16B_mcast(packet->dlid))) 1508 packet->dlid += opa_get_mcast_base(OPA_MCAST_NR) - 1509 opa_get_lid(opa_get_mcast_base(OPA_MCAST_NR), 1510 16B); 1511 packet->sc = hfi1_16B_get_sc(packet->hdr); 1512 packet->sl = ibp->sc_to_sl[packet->sc]; 1513 packet->extra_byte = SIZE_OF_LT; 1514 packet->pkey = hfi1_16B_get_pkey(packet->hdr); 1515 1516 if (hfi1_bypass_ingress_pkt_check(packet)) 1517 goto drop; 1518 1519 return 0; 1520 drop: 1521 hfi1_cdbg(PKT, "%s: packet dropped\n", __func__); 1522 ibp->rvp.n_pkt_drops++; 1523 return -EINVAL; 1524 } 1525 1526 static void show_eflags_errs(struct hfi1_packet *packet) 1527 { 1528 struct hfi1_ctxtdata *rcd = packet->rcd; 1529 u32 rte = rhf_rcv_type_err(packet->rhf); 1530 1531 dd_dev_err(rcd->dd, 1532 "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s] rte 0x%x\n", 1533 rcd->ctxt, packet->rhf, 1534 packet->rhf & RHF_K_HDR_LEN_ERR ? "k_hdr_len " : "", 1535 packet->rhf & RHF_DC_UNC_ERR ? "dc_unc " : "", 1536 packet->rhf & RHF_DC_ERR ? "dc " : "", 1537 packet->rhf & RHF_TID_ERR ? "tid " : "", 1538 packet->rhf & RHF_LEN_ERR ? "len " : "", 1539 packet->rhf & RHF_ECC_ERR ? "ecc " : "", 1540 packet->rhf & RHF_ICRC_ERR ? "icrc " : "", 1541 rte); 1542 } 1543 1544 void handle_eflags(struct hfi1_packet *packet) 1545 { 1546 struct hfi1_ctxtdata *rcd = packet->rcd; 1547 1548 rcv_hdrerr(rcd, rcd->ppd, packet); 1549 if (rhf_err_flags(packet->rhf)) 1550 show_eflags_errs(packet); 1551 } 1552 1553 /* 1554 * The following functions are called by the interrupt handler. They are type 1555 * specific handlers for each packet type. 1556 */ 1557 static void process_receive_ib(struct hfi1_packet *packet) 1558 { 1559 if (hfi1_setup_9B_packet(packet)) 1560 return; 1561 1562 if (unlikely(hfi1_dbg_should_fault_rx(packet))) 1563 return; 1564 1565 trace_hfi1_rcvhdr(packet); 1566 1567 if (unlikely(rhf_err_flags(packet->rhf))) { 1568 handle_eflags(packet); 1569 return; 1570 } 1571 1572 hfi1_ib_rcv(packet); 1573 } 1574 1575 static inline bool hfi1_is_vnic_packet(struct hfi1_packet *packet) 1576 { 1577 /* Packet received in VNIC context via RSM */ 1578 if (packet->rcd->is_vnic) 1579 return true; 1580 1581 if ((hfi1_16B_get_l2(packet->ebuf) == OPA_16B_L2_TYPE) && 1582 (hfi1_16B_get_l4(packet->ebuf) == OPA_16B_L4_ETHR)) 1583 return true; 1584 1585 return false; 1586 } 1587 1588 static void process_receive_bypass(struct hfi1_packet *packet) 1589 { 1590 struct hfi1_devdata *dd = packet->rcd->dd; 1591 1592 if (hfi1_is_vnic_packet(packet)) { 1593 hfi1_vnic_bypass_rcv(packet); 1594 return; 1595 } 1596 1597 if (hfi1_setup_bypass_packet(packet)) 1598 return; 1599 1600 trace_hfi1_rcvhdr(packet); 1601 1602 if (unlikely(rhf_err_flags(packet->rhf))) { 1603 handle_eflags(packet); 1604 return; 1605 } 1606 1607 if (hfi1_16B_get_l2(packet->hdr) == 0x2) { 1608 hfi1_16B_rcv(packet); 1609 } else { 1610 dd_dev_err(dd, 1611 "Bypass packets other than 16B are not supported in normal operation. Dropping\n"); 1612 incr_cntr64(&dd->sw_rcv_bypass_packet_errors); 1613 if (!(dd->err_info_rcvport.status_and_code & 1614 OPA_EI_STATUS_SMASK)) { 1615 u64 *flits = packet->ebuf; 1616 1617 if (flits && !(packet->rhf & RHF_LEN_ERR)) { 1618 dd->err_info_rcvport.packet_flit1 = flits[0]; 1619 dd->err_info_rcvport.packet_flit2 = 1620 packet->tlen > sizeof(flits[0]) ? 1621 flits[1] : 0; 1622 } 1623 dd->err_info_rcvport.status_and_code |= 1624 (OPA_EI_STATUS_SMASK | BAD_L2_ERR); 1625 } 1626 } 1627 } 1628 1629 static void process_receive_error(struct hfi1_packet *packet) 1630 { 1631 /* KHdrHCRCErr -- KDETH packet with a bad HCRC */ 1632 if (unlikely( 1633 hfi1_dbg_fault_suppress_err(&packet->rcd->dd->verbs_dev) && 1634 (rhf_rcv_type_err(packet->rhf) == RHF_RCV_TYPE_ERROR || 1635 packet->rhf & RHF_DC_ERR))) 1636 return; 1637 1638 hfi1_setup_ib_header(packet); 1639 handle_eflags(packet); 1640 1641 if (unlikely(rhf_err_flags(packet->rhf))) 1642 dd_dev_err(packet->rcd->dd, 1643 "Unhandled error packet received. Dropping.\n"); 1644 } 1645 1646 static void kdeth_process_expected(struct hfi1_packet *packet) 1647 { 1648 hfi1_setup_9B_packet(packet); 1649 if (unlikely(hfi1_dbg_should_fault_rx(packet))) 1650 return; 1651 1652 if (unlikely(rhf_err_flags(packet->rhf))) { 1653 struct hfi1_ctxtdata *rcd = packet->rcd; 1654 1655 if (hfi1_handle_kdeth_eflags(rcd, rcd->ppd, packet)) 1656 return; 1657 } 1658 1659 hfi1_kdeth_expected_rcv(packet); 1660 } 1661 1662 static void kdeth_process_eager(struct hfi1_packet *packet) 1663 { 1664 hfi1_setup_9B_packet(packet); 1665 if (unlikely(hfi1_dbg_should_fault_rx(packet))) 1666 return; 1667 1668 trace_hfi1_rcvhdr(packet); 1669 if (unlikely(rhf_err_flags(packet->rhf))) { 1670 struct hfi1_ctxtdata *rcd = packet->rcd; 1671 1672 show_eflags_errs(packet); 1673 if (hfi1_handle_kdeth_eflags(rcd, rcd->ppd, packet)) 1674 return; 1675 } 1676 1677 hfi1_kdeth_eager_rcv(packet); 1678 } 1679 1680 static void process_receive_invalid(struct hfi1_packet *packet) 1681 { 1682 dd_dev_err(packet->rcd->dd, "Invalid packet type %d. Dropping\n", 1683 rhf_rcv_type(packet->rhf)); 1684 } 1685 1686 #define HFI1_RCVHDR_DUMP_MAX 5 1687 1688 void seqfile_dump_rcd(struct seq_file *s, struct hfi1_ctxtdata *rcd) 1689 { 1690 struct hfi1_packet packet; 1691 struct ps_mdata mdata; 1692 int i; 1693 1694 seq_printf(s, "Rcd %u: RcvHdr cnt %u entsize %u %s ctrl 0x%08llx status 0x%08llx, head %llu tail %llu sw head %u\n", 1695 rcd->ctxt, get_hdrq_cnt(rcd), get_hdrqentsize(rcd), 1696 get_dma_rtail_setting(rcd) ? 1697 "dma_rtail" : "nodma_rtail", 1698 read_kctxt_csr(rcd->dd, rcd->ctxt, RCV_CTXT_CTRL), 1699 read_kctxt_csr(rcd->dd, rcd->ctxt, RCV_CTXT_STATUS), 1700 read_uctxt_csr(rcd->dd, rcd->ctxt, RCV_HDR_HEAD) & 1701 RCV_HDR_HEAD_HEAD_MASK, 1702 read_uctxt_csr(rcd->dd, rcd->ctxt, RCV_HDR_TAIL), 1703 rcd->head); 1704 1705 init_packet(rcd, &packet); 1706 init_ps_mdata(&mdata, &packet); 1707 1708 for (i = 0; i < HFI1_RCVHDR_DUMP_MAX; i++) { 1709 __le32 *rhf_addr = (__le32 *)rcd->rcvhdrq + mdata.ps_head + 1710 rcd->rhf_offset; 1711 struct ib_header *hdr; 1712 u64 rhf = rhf_to_cpu(rhf_addr); 1713 u32 etype = rhf_rcv_type(rhf), qpn; 1714 u8 opcode; 1715 u32 psn; 1716 u8 lnh; 1717 1718 if (ps_done(&mdata, rhf, rcd)) 1719 break; 1720 1721 if (ps_skip(&mdata, rhf, rcd)) 1722 goto next; 1723 1724 if (etype > RHF_RCV_TYPE_IB) 1725 goto next; 1726 1727 packet.hdr = hfi1_get_msgheader(rcd, rhf_addr); 1728 hdr = packet.hdr; 1729 1730 lnh = be16_to_cpu(hdr->lrh[0]) & 3; 1731 1732 if (lnh == HFI1_LRH_BTH) 1733 packet.ohdr = &hdr->u.oth; 1734 else if (lnh == HFI1_LRH_GRH) 1735 packet.ohdr = &hdr->u.l.oth; 1736 else 1737 goto next; /* just in case */ 1738 1739 opcode = (be32_to_cpu(packet.ohdr->bth[0]) >> 24); 1740 qpn = be32_to_cpu(packet.ohdr->bth[1]) & RVT_QPN_MASK; 1741 psn = mask_psn(be32_to_cpu(packet.ohdr->bth[2])); 1742 1743 seq_printf(s, "\tEnt %u: opcode 0x%x, qpn 0x%x, psn 0x%x\n", 1744 mdata.ps_head, opcode, qpn, psn); 1745 next: 1746 update_ps_mdata(&mdata, rcd); 1747 } 1748 } 1749 1750 const rhf_rcv_function_ptr normal_rhf_rcv_functions[] = { 1751 [RHF_RCV_TYPE_EXPECTED] = kdeth_process_expected, 1752 [RHF_RCV_TYPE_EAGER] = kdeth_process_eager, 1753 [RHF_RCV_TYPE_IB] = process_receive_ib, 1754 [RHF_RCV_TYPE_ERROR] = process_receive_error, 1755 [RHF_RCV_TYPE_BYPASS] = process_receive_bypass, 1756 [RHF_RCV_TYPE_INVALID5] = process_receive_invalid, 1757 [RHF_RCV_TYPE_INVALID6] = process_receive_invalid, 1758 [RHF_RCV_TYPE_INVALID7] = process_receive_invalid, 1759 }; 1760