1 /* 2 * Copyright (c) 2012, 2013 Intel Corporation. All rights reserved. 3 * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved. 4 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the 10 * OpenIB.org BSD license below: 11 * 12 * Redistribution and use in source and binary forms, with or 13 * without modification, are permitted provided that the following 14 * conditions are met: 15 * 16 * - Redistributions of source code must retain the above 17 * copyright notice, this list of conditions and the following 18 * disclaimer. 19 * 20 * - Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials 23 * provided with the distribution. 24 * 25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 32 * SOFTWARE. 33 */ 34 35 #include <rdma/ib_mad.h> 36 #include <rdma/ib_user_verbs.h> 37 #include <linux/io.h> 38 #include <linux/module.h> 39 #include <linux/utsname.h> 40 #include <linux/rculist.h> 41 #include <linux/mm.h> 42 #include <linux/random.h> 43 44 #include "qib.h" 45 #include "qib_common.h" 46 47 static unsigned int ib_qib_qp_table_size = 256; 48 module_param_named(qp_table_size, ib_qib_qp_table_size, uint, S_IRUGO); 49 MODULE_PARM_DESC(qp_table_size, "QP table size"); 50 51 unsigned int ib_qib_lkey_table_size = 16; 52 module_param_named(lkey_table_size, ib_qib_lkey_table_size, uint, 53 S_IRUGO); 54 MODULE_PARM_DESC(lkey_table_size, 55 "LKEY table size in bits (2^n, 1 <= n <= 23)"); 56 57 static unsigned int ib_qib_max_pds = 0xFFFF; 58 module_param_named(max_pds, ib_qib_max_pds, uint, S_IRUGO); 59 MODULE_PARM_DESC(max_pds, 60 "Maximum number of protection domains to support"); 61 62 static unsigned int ib_qib_max_ahs = 0xFFFF; 63 module_param_named(max_ahs, ib_qib_max_ahs, uint, S_IRUGO); 64 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support"); 65 66 unsigned int ib_qib_max_cqes = 0x2FFFF; 67 module_param_named(max_cqes, ib_qib_max_cqes, uint, S_IRUGO); 68 MODULE_PARM_DESC(max_cqes, 69 "Maximum number of completion queue entries to support"); 70 71 unsigned int ib_qib_max_cqs = 0x1FFFF; 72 module_param_named(max_cqs, ib_qib_max_cqs, uint, S_IRUGO); 73 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support"); 74 75 unsigned int ib_qib_max_qp_wrs = 0x3FFF; 76 module_param_named(max_qp_wrs, ib_qib_max_qp_wrs, uint, S_IRUGO); 77 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support"); 78 79 unsigned int ib_qib_max_qps = 16384; 80 module_param_named(max_qps, ib_qib_max_qps, uint, S_IRUGO); 81 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support"); 82 83 unsigned int ib_qib_max_sges = 0x60; 84 module_param_named(max_sges, ib_qib_max_sges, uint, S_IRUGO); 85 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support"); 86 87 unsigned int ib_qib_max_mcast_grps = 16384; 88 module_param_named(max_mcast_grps, ib_qib_max_mcast_grps, uint, S_IRUGO); 89 MODULE_PARM_DESC(max_mcast_grps, 90 "Maximum number of multicast groups to support"); 91 92 unsigned int ib_qib_max_mcast_qp_attached = 16; 93 module_param_named(max_mcast_qp_attached, ib_qib_max_mcast_qp_attached, 94 uint, S_IRUGO); 95 MODULE_PARM_DESC(max_mcast_qp_attached, 96 "Maximum number of attached QPs to support"); 97 98 unsigned int ib_qib_max_srqs = 1024; 99 module_param_named(max_srqs, ib_qib_max_srqs, uint, S_IRUGO); 100 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support"); 101 102 unsigned int ib_qib_max_srq_sges = 128; 103 module_param_named(max_srq_sges, ib_qib_max_srq_sges, uint, S_IRUGO); 104 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support"); 105 106 unsigned int ib_qib_max_srq_wrs = 0x1FFFF; 107 module_param_named(max_srq_wrs, ib_qib_max_srq_wrs, uint, S_IRUGO); 108 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support"); 109 110 static unsigned int ib_qib_disable_sma; 111 module_param_named(disable_sma, ib_qib_disable_sma, uint, S_IWUSR | S_IRUGO); 112 MODULE_PARM_DESC(disable_sma, "Disable the SMA"); 113 114 /* 115 * Note that it is OK to post send work requests in the SQE and ERR 116 * states; qib_do_send() will process them and generate error 117 * completions as per IB 1.2 C10-96. 118 */ 119 const int ib_qib_state_ops[IB_QPS_ERR + 1] = { 120 [IB_QPS_RESET] = 0, 121 [IB_QPS_INIT] = QIB_POST_RECV_OK, 122 [IB_QPS_RTR] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK, 123 [IB_QPS_RTS] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK | 124 QIB_POST_SEND_OK | QIB_PROCESS_SEND_OK | 125 QIB_PROCESS_NEXT_SEND_OK, 126 [IB_QPS_SQD] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK | 127 QIB_POST_SEND_OK | QIB_PROCESS_SEND_OK, 128 [IB_QPS_SQE] = QIB_POST_RECV_OK | QIB_PROCESS_RECV_OK | 129 QIB_POST_SEND_OK | QIB_FLUSH_SEND, 130 [IB_QPS_ERR] = QIB_POST_RECV_OK | QIB_FLUSH_RECV | 131 QIB_POST_SEND_OK | QIB_FLUSH_SEND, 132 }; 133 134 struct qib_ucontext { 135 struct ib_ucontext ibucontext; 136 }; 137 138 static inline struct qib_ucontext *to_iucontext(struct ib_ucontext 139 *ibucontext) 140 { 141 return container_of(ibucontext, struct qib_ucontext, ibucontext); 142 } 143 144 /* 145 * Translate ib_wr_opcode into ib_wc_opcode. 146 */ 147 const enum ib_wc_opcode ib_qib_wc_opcode[] = { 148 [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE, 149 [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE, 150 [IB_WR_SEND] = IB_WC_SEND, 151 [IB_WR_SEND_WITH_IMM] = IB_WC_SEND, 152 [IB_WR_RDMA_READ] = IB_WC_RDMA_READ, 153 [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP, 154 [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD 155 }; 156 157 /* 158 * System image GUID. 159 */ 160 __be64 ib_qib_sys_image_guid; 161 162 /** 163 * qib_copy_sge - copy data to SGE memory 164 * @ss: the SGE state 165 * @data: the data to copy 166 * @length: the length of the data 167 */ 168 void qib_copy_sge(struct qib_sge_state *ss, void *data, u32 length, int release) 169 { 170 struct qib_sge *sge = &ss->sge; 171 172 while (length) { 173 u32 len = sge->length; 174 175 if (len > length) 176 len = length; 177 if (len > sge->sge_length) 178 len = sge->sge_length; 179 BUG_ON(len == 0); 180 memcpy(sge->vaddr, data, len); 181 sge->vaddr += len; 182 sge->length -= len; 183 sge->sge_length -= len; 184 if (sge->sge_length == 0) { 185 if (release) 186 qib_put_mr(sge->mr); 187 if (--ss->num_sge) 188 *sge = *ss->sg_list++; 189 } else if (sge->length == 0 && sge->mr->lkey) { 190 if (++sge->n >= QIB_SEGSZ) { 191 if (++sge->m >= sge->mr->mapsz) 192 break; 193 sge->n = 0; 194 } 195 sge->vaddr = 196 sge->mr->map[sge->m]->segs[sge->n].vaddr; 197 sge->length = 198 sge->mr->map[sge->m]->segs[sge->n].length; 199 } 200 data += len; 201 length -= len; 202 } 203 } 204 205 /** 206 * qib_skip_sge - skip over SGE memory - XXX almost dup of prev func 207 * @ss: the SGE state 208 * @length: the number of bytes to skip 209 */ 210 void qib_skip_sge(struct qib_sge_state *ss, u32 length, int release) 211 { 212 struct qib_sge *sge = &ss->sge; 213 214 while (length) { 215 u32 len = sge->length; 216 217 if (len > length) 218 len = length; 219 if (len > sge->sge_length) 220 len = sge->sge_length; 221 BUG_ON(len == 0); 222 sge->vaddr += len; 223 sge->length -= len; 224 sge->sge_length -= len; 225 if (sge->sge_length == 0) { 226 if (release) 227 qib_put_mr(sge->mr); 228 if (--ss->num_sge) 229 *sge = *ss->sg_list++; 230 } else if (sge->length == 0 && sge->mr->lkey) { 231 if (++sge->n >= QIB_SEGSZ) { 232 if (++sge->m >= sge->mr->mapsz) 233 break; 234 sge->n = 0; 235 } 236 sge->vaddr = 237 sge->mr->map[sge->m]->segs[sge->n].vaddr; 238 sge->length = 239 sge->mr->map[sge->m]->segs[sge->n].length; 240 } 241 length -= len; 242 } 243 } 244 245 /* 246 * Count the number of DMA descriptors needed to send length bytes of data. 247 * Don't modify the qib_sge_state to get the count. 248 * Return zero if any of the segments is not aligned. 249 */ 250 static u32 qib_count_sge(struct qib_sge_state *ss, u32 length) 251 { 252 struct qib_sge *sg_list = ss->sg_list; 253 struct qib_sge sge = ss->sge; 254 u8 num_sge = ss->num_sge; 255 u32 ndesc = 1; /* count the header */ 256 257 while (length) { 258 u32 len = sge.length; 259 260 if (len > length) 261 len = length; 262 if (len > sge.sge_length) 263 len = sge.sge_length; 264 BUG_ON(len == 0); 265 if (((long) sge.vaddr & (sizeof(u32) - 1)) || 266 (len != length && (len & (sizeof(u32) - 1)))) { 267 ndesc = 0; 268 break; 269 } 270 ndesc++; 271 sge.vaddr += len; 272 sge.length -= len; 273 sge.sge_length -= len; 274 if (sge.sge_length == 0) { 275 if (--num_sge) 276 sge = *sg_list++; 277 } else if (sge.length == 0 && sge.mr->lkey) { 278 if (++sge.n >= QIB_SEGSZ) { 279 if (++sge.m >= sge.mr->mapsz) 280 break; 281 sge.n = 0; 282 } 283 sge.vaddr = 284 sge.mr->map[sge.m]->segs[sge.n].vaddr; 285 sge.length = 286 sge.mr->map[sge.m]->segs[sge.n].length; 287 } 288 length -= len; 289 } 290 return ndesc; 291 } 292 293 /* 294 * Copy from the SGEs to the data buffer. 295 */ 296 static void qib_copy_from_sge(void *data, struct qib_sge_state *ss, u32 length) 297 { 298 struct qib_sge *sge = &ss->sge; 299 300 while (length) { 301 u32 len = sge->length; 302 303 if (len > length) 304 len = length; 305 if (len > sge->sge_length) 306 len = sge->sge_length; 307 BUG_ON(len == 0); 308 memcpy(data, sge->vaddr, len); 309 sge->vaddr += len; 310 sge->length -= len; 311 sge->sge_length -= len; 312 if (sge->sge_length == 0) { 313 if (--ss->num_sge) 314 *sge = *ss->sg_list++; 315 } else if (sge->length == 0 && sge->mr->lkey) { 316 if (++sge->n >= QIB_SEGSZ) { 317 if (++sge->m >= sge->mr->mapsz) 318 break; 319 sge->n = 0; 320 } 321 sge->vaddr = 322 sge->mr->map[sge->m]->segs[sge->n].vaddr; 323 sge->length = 324 sge->mr->map[sge->m]->segs[sge->n].length; 325 } 326 data += len; 327 length -= len; 328 } 329 } 330 331 /** 332 * qib_post_one_send - post one RC, UC, or UD send work request 333 * @qp: the QP to post on 334 * @wr: the work request to send 335 */ 336 static int qib_post_one_send(struct qib_qp *qp, struct ib_send_wr *wr, 337 int *scheduled) 338 { 339 struct qib_swqe *wqe; 340 u32 next; 341 int i; 342 int j; 343 int acc; 344 int ret; 345 unsigned long flags; 346 struct qib_lkey_table *rkt; 347 struct qib_pd *pd; 348 349 spin_lock_irqsave(&qp->s_lock, flags); 350 351 /* Check that state is OK to post send. */ 352 if (unlikely(!(ib_qib_state_ops[qp->state] & QIB_POST_SEND_OK))) 353 goto bail_inval; 354 355 /* IB spec says that num_sge == 0 is OK. */ 356 if (wr->num_sge > qp->s_max_sge) 357 goto bail_inval; 358 359 /* 360 * Don't allow RDMA reads or atomic operations on UC or 361 * undefined operations. 362 * Make sure buffer is large enough to hold the result for atomics. 363 */ 364 if (wr->opcode == IB_WR_FAST_REG_MR) { 365 if (qib_fast_reg_mr(qp, wr)) 366 goto bail_inval; 367 } else if (qp->ibqp.qp_type == IB_QPT_UC) { 368 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) 369 goto bail_inval; 370 } else if (qp->ibqp.qp_type != IB_QPT_RC) { 371 /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */ 372 if (wr->opcode != IB_WR_SEND && 373 wr->opcode != IB_WR_SEND_WITH_IMM) 374 goto bail_inval; 375 /* Check UD destination address PD */ 376 if (qp->ibqp.pd != wr->wr.ud.ah->pd) 377 goto bail_inval; 378 } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) 379 goto bail_inval; 380 else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP && 381 (wr->num_sge == 0 || 382 wr->sg_list[0].length < sizeof(u64) || 383 wr->sg_list[0].addr & (sizeof(u64) - 1))) 384 goto bail_inval; 385 else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) 386 goto bail_inval; 387 388 next = qp->s_head + 1; 389 if (next >= qp->s_size) 390 next = 0; 391 if (next == qp->s_last) { 392 ret = -ENOMEM; 393 goto bail; 394 } 395 396 rkt = &to_idev(qp->ibqp.device)->lk_table; 397 pd = to_ipd(qp->ibqp.pd); 398 wqe = get_swqe_ptr(qp, qp->s_head); 399 wqe->wr = *wr; 400 wqe->length = 0; 401 j = 0; 402 if (wr->num_sge) { 403 acc = wr->opcode >= IB_WR_RDMA_READ ? 404 IB_ACCESS_LOCAL_WRITE : 0; 405 for (i = 0; i < wr->num_sge; i++) { 406 u32 length = wr->sg_list[i].length; 407 int ok; 408 409 if (length == 0) 410 continue; 411 ok = qib_lkey_ok(rkt, pd, &wqe->sg_list[j], 412 &wr->sg_list[i], acc); 413 if (!ok) 414 goto bail_inval_free; 415 wqe->length += length; 416 j++; 417 } 418 wqe->wr.num_sge = j; 419 } 420 if (qp->ibqp.qp_type == IB_QPT_UC || 421 qp->ibqp.qp_type == IB_QPT_RC) { 422 if (wqe->length > 0x80000000U) 423 goto bail_inval_free; 424 } else if (wqe->length > (dd_from_ibdev(qp->ibqp.device)->pport + 425 qp->port_num - 1)->ibmtu) 426 goto bail_inval_free; 427 else 428 atomic_inc(&to_iah(wr->wr.ud.ah)->refcount); 429 wqe->ssn = qp->s_ssn++; 430 qp->s_head = next; 431 432 ret = 0; 433 goto bail; 434 435 bail_inval_free: 436 while (j) { 437 struct qib_sge *sge = &wqe->sg_list[--j]; 438 439 qib_put_mr(sge->mr); 440 } 441 bail_inval: 442 ret = -EINVAL; 443 bail: 444 if (!ret && !wr->next && 445 !qib_sdma_empty( 446 dd_from_ibdev(qp->ibqp.device)->pport + qp->port_num - 1)) { 447 qib_schedule_send(qp); 448 *scheduled = 1; 449 } 450 spin_unlock_irqrestore(&qp->s_lock, flags); 451 return ret; 452 } 453 454 /** 455 * qib_post_send - post a send on a QP 456 * @ibqp: the QP to post the send on 457 * @wr: the list of work requests to post 458 * @bad_wr: the first bad WR is put here 459 * 460 * This may be called from interrupt context. 461 */ 462 static int qib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, 463 struct ib_send_wr **bad_wr) 464 { 465 struct qib_qp *qp = to_iqp(ibqp); 466 int err = 0; 467 int scheduled = 0; 468 469 for (; wr; wr = wr->next) { 470 err = qib_post_one_send(qp, wr, &scheduled); 471 if (err) { 472 *bad_wr = wr; 473 goto bail; 474 } 475 } 476 477 /* Try to do the send work in the caller's context. */ 478 if (!scheduled) 479 qib_do_send(&qp->s_work); 480 481 bail: 482 return err; 483 } 484 485 /** 486 * qib_post_receive - post a receive on a QP 487 * @ibqp: the QP to post the receive on 488 * @wr: the WR to post 489 * @bad_wr: the first bad WR is put here 490 * 491 * This may be called from interrupt context. 492 */ 493 static int qib_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr, 494 struct ib_recv_wr **bad_wr) 495 { 496 struct qib_qp *qp = to_iqp(ibqp); 497 struct qib_rwq *wq = qp->r_rq.wq; 498 unsigned long flags; 499 int ret; 500 501 /* Check that state is OK to post receive. */ 502 if (!(ib_qib_state_ops[qp->state] & QIB_POST_RECV_OK) || !wq) { 503 *bad_wr = wr; 504 ret = -EINVAL; 505 goto bail; 506 } 507 508 for (; wr; wr = wr->next) { 509 struct qib_rwqe *wqe; 510 u32 next; 511 int i; 512 513 if ((unsigned) wr->num_sge > qp->r_rq.max_sge) { 514 *bad_wr = wr; 515 ret = -EINVAL; 516 goto bail; 517 } 518 519 spin_lock_irqsave(&qp->r_rq.lock, flags); 520 next = wq->head + 1; 521 if (next >= qp->r_rq.size) 522 next = 0; 523 if (next == wq->tail) { 524 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 525 *bad_wr = wr; 526 ret = -ENOMEM; 527 goto bail; 528 } 529 530 wqe = get_rwqe_ptr(&qp->r_rq, wq->head); 531 wqe->wr_id = wr->wr_id; 532 wqe->num_sge = wr->num_sge; 533 for (i = 0; i < wr->num_sge; i++) 534 wqe->sg_list[i] = wr->sg_list[i]; 535 /* Make sure queue entry is written before the head index. */ 536 smp_wmb(); 537 wq->head = next; 538 spin_unlock_irqrestore(&qp->r_rq.lock, flags); 539 } 540 ret = 0; 541 542 bail: 543 return ret; 544 } 545 546 /** 547 * qib_qp_rcv - processing an incoming packet on a QP 548 * @rcd: the context pointer 549 * @hdr: the packet header 550 * @has_grh: true if the packet has a GRH 551 * @data: the packet data 552 * @tlen: the packet length 553 * @qp: the QP the packet came on 554 * 555 * This is called from qib_ib_rcv() to process an incoming packet 556 * for the given QP. 557 * Called at interrupt level. 558 */ 559 static void qib_qp_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr, 560 int has_grh, void *data, u32 tlen, struct qib_qp *qp) 561 { 562 struct qib_ibport *ibp = &rcd->ppd->ibport_data; 563 564 spin_lock(&qp->r_lock); 565 566 /* Check for valid receive state. */ 567 if (!(ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK)) { 568 ibp->n_pkt_drops++; 569 goto unlock; 570 } 571 572 switch (qp->ibqp.qp_type) { 573 case IB_QPT_SMI: 574 case IB_QPT_GSI: 575 if (ib_qib_disable_sma) 576 break; 577 /* FALLTHROUGH */ 578 case IB_QPT_UD: 579 qib_ud_rcv(ibp, hdr, has_grh, data, tlen, qp); 580 break; 581 582 case IB_QPT_RC: 583 qib_rc_rcv(rcd, hdr, has_grh, data, tlen, qp); 584 break; 585 586 case IB_QPT_UC: 587 qib_uc_rcv(ibp, hdr, has_grh, data, tlen, qp); 588 break; 589 590 default: 591 break; 592 } 593 594 unlock: 595 spin_unlock(&qp->r_lock); 596 } 597 598 /** 599 * qib_ib_rcv - process an incoming packet 600 * @rcd: the context pointer 601 * @rhdr: the header of the packet 602 * @data: the packet payload 603 * @tlen: the packet length 604 * 605 * This is called from qib_kreceive() to process an incoming packet at 606 * interrupt level. Tlen is the length of the header + data + CRC in bytes. 607 */ 608 void qib_ib_rcv(struct qib_ctxtdata *rcd, void *rhdr, void *data, u32 tlen) 609 { 610 struct qib_pportdata *ppd = rcd->ppd; 611 struct qib_ibport *ibp = &ppd->ibport_data; 612 struct qib_ib_header *hdr = rhdr; 613 struct qib_other_headers *ohdr; 614 struct qib_qp *qp; 615 u32 qp_num; 616 int lnh; 617 u8 opcode; 618 u16 lid; 619 620 /* 24 == LRH+BTH+CRC */ 621 if (unlikely(tlen < 24)) 622 goto drop; 623 624 /* Check for a valid destination LID (see ch. 7.11.1). */ 625 lid = be16_to_cpu(hdr->lrh[1]); 626 if (lid < QIB_MULTICAST_LID_BASE) { 627 lid &= ~((1 << ppd->lmc) - 1); 628 if (unlikely(lid != ppd->lid)) 629 goto drop; 630 } 631 632 /* Check for GRH */ 633 lnh = be16_to_cpu(hdr->lrh[0]) & 3; 634 if (lnh == QIB_LRH_BTH) 635 ohdr = &hdr->u.oth; 636 else if (lnh == QIB_LRH_GRH) { 637 u32 vtf; 638 639 ohdr = &hdr->u.l.oth; 640 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR) 641 goto drop; 642 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow); 643 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION) 644 goto drop; 645 } else 646 goto drop; 647 648 opcode = be32_to_cpu(ohdr->bth[0]) >> 24; 649 ibp->opstats[opcode & 0x7f].n_bytes += tlen; 650 ibp->opstats[opcode & 0x7f].n_packets++; 651 652 /* Get the destination QP number. */ 653 qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK; 654 if (qp_num == QIB_MULTICAST_QPN) { 655 struct qib_mcast *mcast; 656 struct qib_mcast_qp *p; 657 658 if (lnh != QIB_LRH_GRH) 659 goto drop; 660 mcast = qib_mcast_find(ibp, &hdr->u.l.grh.dgid); 661 if (mcast == NULL) 662 goto drop; 663 ibp->n_multicast_rcv++; 664 list_for_each_entry_rcu(p, &mcast->qp_list, list) 665 qib_qp_rcv(rcd, hdr, 1, data, tlen, p->qp); 666 /* 667 * Notify qib_multicast_detach() if it is waiting for us 668 * to finish. 669 */ 670 if (atomic_dec_return(&mcast->refcount) <= 1) 671 wake_up(&mcast->wait); 672 } else { 673 if (rcd->lookaside_qp) { 674 if (rcd->lookaside_qpn != qp_num) { 675 if (atomic_dec_and_test( 676 &rcd->lookaside_qp->refcount)) 677 wake_up( 678 &rcd->lookaside_qp->wait); 679 rcd->lookaside_qp = NULL; 680 } 681 } 682 if (!rcd->lookaside_qp) { 683 qp = qib_lookup_qpn(ibp, qp_num); 684 if (!qp) 685 goto drop; 686 rcd->lookaside_qp = qp; 687 rcd->lookaside_qpn = qp_num; 688 } else 689 qp = rcd->lookaside_qp; 690 ibp->n_unicast_rcv++; 691 qib_qp_rcv(rcd, hdr, lnh == QIB_LRH_GRH, data, tlen, qp); 692 } 693 return; 694 695 drop: 696 ibp->n_pkt_drops++; 697 } 698 699 /* 700 * This is called from a timer to check for QPs 701 * which need kernel memory in order to send a packet. 702 */ 703 static void mem_timer(unsigned long data) 704 { 705 struct qib_ibdev *dev = (struct qib_ibdev *) data; 706 struct list_head *list = &dev->memwait; 707 struct qib_qp *qp = NULL; 708 unsigned long flags; 709 710 spin_lock_irqsave(&dev->pending_lock, flags); 711 if (!list_empty(list)) { 712 qp = list_entry(list->next, struct qib_qp, iowait); 713 list_del_init(&qp->iowait); 714 atomic_inc(&qp->refcount); 715 if (!list_empty(list)) 716 mod_timer(&dev->mem_timer, jiffies + 1); 717 } 718 spin_unlock_irqrestore(&dev->pending_lock, flags); 719 720 if (qp) { 721 spin_lock_irqsave(&qp->s_lock, flags); 722 if (qp->s_flags & QIB_S_WAIT_KMEM) { 723 qp->s_flags &= ~QIB_S_WAIT_KMEM; 724 qib_schedule_send(qp); 725 } 726 spin_unlock_irqrestore(&qp->s_lock, flags); 727 if (atomic_dec_and_test(&qp->refcount)) 728 wake_up(&qp->wait); 729 } 730 } 731 732 static void update_sge(struct qib_sge_state *ss, u32 length) 733 { 734 struct qib_sge *sge = &ss->sge; 735 736 sge->vaddr += length; 737 sge->length -= length; 738 sge->sge_length -= length; 739 if (sge->sge_length == 0) { 740 if (--ss->num_sge) 741 *sge = *ss->sg_list++; 742 } else if (sge->length == 0 && sge->mr->lkey) { 743 if (++sge->n >= QIB_SEGSZ) { 744 if (++sge->m >= sge->mr->mapsz) 745 return; 746 sge->n = 0; 747 } 748 sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr; 749 sge->length = sge->mr->map[sge->m]->segs[sge->n].length; 750 } 751 } 752 753 #ifdef __LITTLE_ENDIAN 754 static inline u32 get_upper_bits(u32 data, u32 shift) 755 { 756 return data >> shift; 757 } 758 759 static inline u32 set_upper_bits(u32 data, u32 shift) 760 { 761 return data << shift; 762 } 763 764 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off) 765 { 766 data <<= ((sizeof(u32) - n) * BITS_PER_BYTE); 767 data >>= ((sizeof(u32) - n - off) * BITS_PER_BYTE); 768 return data; 769 } 770 #else 771 static inline u32 get_upper_bits(u32 data, u32 shift) 772 { 773 return data << shift; 774 } 775 776 static inline u32 set_upper_bits(u32 data, u32 shift) 777 { 778 return data >> shift; 779 } 780 781 static inline u32 clear_upper_bytes(u32 data, u32 n, u32 off) 782 { 783 data >>= ((sizeof(u32) - n) * BITS_PER_BYTE); 784 data <<= ((sizeof(u32) - n - off) * BITS_PER_BYTE); 785 return data; 786 } 787 #endif 788 789 static void copy_io(u32 __iomem *piobuf, struct qib_sge_state *ss, 790 u32 length, unsigned flush_wc) 791 { 792 u32 extra = 0; 793 u32 data = 0; 794 u32 last; 795 796 while (1) { 797 u32 len = ss->sge.length; 798 u32 off; 799 800 if (len > length) 801 len = length; 802 if (len > ss->sge.sge_length) 803 len = ss->sge.sge_length; 804 BUG_ON(len == 0); 805 /* If the source address is not aligned, try to align it. */ 806 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1); 807 if (off) { 808 u32 *addr = (u32 *)((unsigned long)ss->sge.vaddr & 809 ~(sizeof(u32) - 1)); 810 u32 v = get_upper_bits(*addr, off * BITS_PER_BYTE); 811 u32 y; 812 813 y = sizeof(u32) - off; 814 if (len > y) 815 len = y; 816 if (len + extra >= sizeof(u32)) { 817 data |= set_upper_bits(v, extra * 818 BITS_PER_BYTE); 819 len = sizeof(u32) - extra; 820 if (len == length) { 821 last = data; 822 break; 823 } 824 __raw_writel(data, piobuf); 825 piobuf++; 826 extra = 0; 827 data = 0; 828 } else { 829 /* Clear unused upper bytes */ 830 data |= clear_upper_bytes(v, len, extra); 831 if (len == length) { 832 last = data; 833 break; 834 } 835 extra += len; 836 } 837 } else if (extra) { 838 /* Source address is aligned. */ 839 u32 *addr = (u32 *) ss->sge.vaddr; 840 int shift = extra * BITS_PER_BYTE; 841 int ushift = 32 - shift; 842 u32 l = len; 843 844 while (l >= sizeof(u32)) { 845 u32 v = *addr; 846 847 data |= set_upper_bits(v, shift); 848 __raw_writel(data, piobuf); 849 data = get_upper_bits(v, ushift); 850 piobuf++; 851 addr++; 852 l -= sizeof(u32); 853 } 854 /* 855 * We still have 'extra' number of bytes leftover. 856 */ 857 if (l) { 858 u32 v = *addr; 859 860 if (l + extra >= sizeof(u32)) { 861 data |= set_upper_bits(v, shift); 862 len -= l + extra - sizeof(u32); 863 if (len == length) { 864 last = data; 865 break; 866 } 867 __raw_writel(data, piobuf); 868 piobuf++; 869 extra = 0; 870 data = 0; 871 } else { 872 /* Clear unused upper bytes */ 873 data |= clear_upper_bytes(v, l, extra); 874 if (len == length) { 875 last = data; 876 break; 877 } 878 extra += l; 879 } 880 } else if (len == length) { 881 last = data; 882 break; 883 } 884 } else if (len == length) { 885 u32 w; 886 887 /* 888 * Need to round up for the last dword in the 889 * packet. 890 */ 891 w = (len + 3) >> 2; 892 qib_pio_copy(piobuf, ss->sge.vaddr, w - 1); 893 piobuf += w - 1; 894 last = ((u32 *) ss->sge.vaddr)[w - 1]; 895 break; 896 } else { 897 u32 w = len >> 2; 898 899 qib_pio_copy(piobuf, ss->sge.vaddr, w); 900 piobuf += w; 901 902 extra = len & (sizeof(u32) - 1); 903 if (extra) { 904 u32 v = ((u32 *) ss->sge.vaddr)[w]; 905 906 /* Clear unused upper bytes */ 907 data = clear_upper_bytes(v, extra, 0); 908 } 909 } 910 update_sge(ss, len); 911 length -= len; 912 } 913 /* Update address before sending packet. */ 914 update_sge(ss, length); 915 if (flush_wc) { 916 /* must flush early everything before trigger word */ 917 qib_flush_wc(); 918 __raw_writel(last, piobuf); 919 /* be sure trigger word is written */ 920 qib_flush_wc(); 921 } else 922 __raw_writel(last, piobuf); 923 } 924 925 static noinline struct qib_verbs_txreq *__get_txreq(struct qib_ibdev *dev, 926 struct qib_qp *qp) 927 { 928 struct qib_verbs_txreq *tx; 929 unsigned long flags; 930 931 spin_lock_irqsave(&qp->s_lock, flags); 932 spin_lock(&dev->pending_lock); 933 934 if (!list_empty(&dev->txreq_free)) { 935 struct list_head *l = dev->txreq_free.next; 936 937 list_del(l); 938 spin_unlock(&dev->pending_lock); 939 spin_unlock_irqrestore(&qp->s_lock, flags); 940 tx = list_entry(l, struct qib_verbs_txreq, txreq.list); 941 } else { 942 if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK && 943 list_empty(&qp->iowait)) { 944 dev->n_txwait++; 945 qp->s_flags |= QIB_S_WAIT_TX; 946 list_add_tail(&qp->iowait, &dev->txwait); 947 } 948 qp->s_flags &= ~QIB_S_BUSY; 949 spin_unlock(&dev->pending_lock); 950 spin_unlock_irqrestore(&qp->s_lock, flags); 951 tx = ERR_PTR(-EBUSY); 952 } 953 return tx; 954 } 955 956 static inline struct qib_verbs_txreq *get_txreq(struct qib_ibdev *dev, 957 struct qib_qp *qp) 958 { 959 struct qib_verbs_txreq *tx; 960 unsigned long flags; 961 962 spin_lock_irqsave(&dev->pending_lock, flags); 963 /* assume the list non empty */ 964 if (likely(!list_empty(&dev->txreq_free))) { 965 struct list_head *l = dev->txreq_free.next; 966 967 list_del(l); 968 spin_unlock_irqrestore(&dev->pending_lock, flags); 969 tx = list_entry(l, struct qib_verbs_txreq, txreq.list); 970 } else { 971 /* call slow path to get the extra lock */ 972 spin_unlock_irqrestore(&dev->pending_lock, flags); 973 tx = __get_txreq(dev, qp); 974 } 975 return tx; 976 } 977 978 void qib_put_txreq(struct qib_verbs_txreq *tx) 979 { 980 struct qib_ibdev *dev; 981 struct qib_qp *qp; 982 unsigned long flags; 983 984 qp = tx->qp; 985 dev = to_idev(qp->ibqp.device); 986 987 if (atomic_dec_and_test(&qp->refcount)) 988 wake_up(&qp->wait); 989 if (tx->mr) { 990 qib_put_mr(tx->mr); 991 tx->mr = NULL; 992 } 993 if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) { 994 tx->txreq.flags &= ~QIB_SDMA_TXREQ_F_FREEBUF; 995 dma_unmap_single(&dd_from_dev(dev)->pcidev->dev, 996 tx->txreq.addr, tx->hdr_dwords << 2, 997 DMA_TO_DEVICE); 998 kfree(tx->align_buf); 999 } 1000 1001 spin_lock_irqsave(&dev->pending_lock, flags); 1002 1003 /* Put struct back on free list */ 1004 list_add(&tx->txreq.list, &dev->txreq_free); 1005 1006 if (!list_empty(&dev->txwait)) { 1007 /* Wake up first QP wanting a free struct */ 1008 qp = list_entry(dev->txwait.next, struct qib_qp, iowait); 1009 list_del_init(&qp->iowait); 1010 atomic_inc(&qp->refcount); 1011 spin_unlock_irqrestore(&dev->pending_lock, flags); 1012 1013 spin_lock_irqsave(&qp->s_lock, flags); 1014 if (qp->s_flags & QIB_S_WAIT_TX) { 1015 qp->s_flags &= ~QIB_S_WAIT_TX; 1016 qib_schedule_send(qp); 1017 } 1018 spin_unlock_irqrestore(&qp->s_lock, flags); 1019 1020 if (atomic_dec_and_test(&qp->refcount)) 1021 wake_up(&qp->wait); 1022 } else 1023 spin_unlock_irqrestore(&dev->pending_lock, flags); 1024 } 1025 1026 /* 1027 * This is called when there are send DMA descriptors that might be 1028 * available. 1029 * 1030 * This is called with ppd->sdma_lock held. 1031 */ 1032 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail) 1033 { 1034 struct qib_qp *qp, *nqp; 1035 struct qib_qp *qps[20]; 1036 struct qib_ibdev *dev; 1037 unsigned i, n; 1038 1039 n = 0; 1040 dev = &ppd->dd->verbs_dev; 1041 spin_lock(&dev->pending_lock); 1042 1043 /* Search wait list for first QP wanting DMA descriptors. */ 1044 list_for_each_entry_safe(qp, nqp, &dev->dmawait, iowait) { 1045 if (qp->port_num != ppd->port) 1046 continue; 1047 if (n == ARRAY_SIZE(qps)) 1048 break; 1049 if (qp->s_tx->txreq.sg_count > avail) 1050 break; 1051 avail -= qp->s_tx->txreq.sg_count; 1052 list_del_init(&qp->iowait); 1053 atomic_inc(&qp->refcount); 1054 qps[n++] = qp; 1055 } 1056 1057 spin_unlock(&dev->pending_lock); 1058 1059 for (i = 0; i < n; i++) { 1060 qp = qps[i]; 1061 spin_lock(&qp->s_lock); 1062 if (qp->s_flags & QIB_S_WAIT_DMA_DESC) { 1063 qp->s_flags &= ~QIB_S_WAIT_DMA_DESC; 1064 qib_schedule_send(qp); 1065 } 1066 spin_unlock(&qp->s_lock); 1067 if (atomic_dec_and_test(&qp->refcount)) 1068 wake_up(&qp->wait); 1069 } 1070 } 1071 1072 /* 1073 * This is called with ppd->sdma_lock held. 1074 */ 1075 static void sdma_complete(struct qib_sdma_txreq *cookie, int status) 1076 { 1077 struct qib_verbs_txreq *tx = 1078 container_of(cookie, struct qib_verbs_txreq, txreq); 1079 struct qib_qp *qp = tx->qp; 1080 1081 spin_lock(&qp->s_lock); 1082 if (tx->wqe) 1083 qib_send_complete(qp, tx->wqe, IB_WC_SUCCESS); 1084 else if (qp->ibqp.qp_type == IB_QPT_RC) { 1085 struct qib_ib_header *hdr; 1086 1087 if (tx->txreq.flags & QIB_SDMA_TXREQ_F_FREEBUF) 1088 hdr = &tx->align_buf->hdr; 1089 else { 1090 struct qib_ibdev *dev = to_idev(qp->ibqp.device); 1091 1092 hdr = &dev->pio_hdrs[tx->hdr_inx].hdr; 1093 } 1094 qib_rc_send_complete(qp, hdr); 1095 } 1096 if (atomic_dec_and_test(&qp->s_dma_busy)) { 1097 if (qp->state == IB_QPS_RESET) 1098 wake_up(&qp->wait_dma); 1099 else if (qp->s_flags & QIB_S_WAIT_DMA) { 1100 qp->s_flags &= ~QIB_S_WAIT_DMA; 1101 qib_schedule_send(qp); 1102 } 1103 } 1104 spin_unlock(&qp->s_lock); 1105 1106 qib_put_txreq(tx); 1107 } 1108 1109 static int wait_kmem(struct qib_ibdev *dev, struct qib_qp *qp) 1110 { 1111 unsigned long flags; 1112 int ret = 0; 1113 1114 spin_lock_irqsave(&qp->s_lock, flags); 1115 if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) { 1116 spin_lock(&dev->pending_lock); 1117 if (list_empty(&qp->iowait)) { 1118 if (list_empty(&dev->memwait)) 1119 mod_timer(&dev->mem_timer, jiffies + 1); 1120 qp->s_flags |= QIB_S_WAIT_KMEM; 1121 list_add_tail(&qp->iowait, &dev->memwait); 1122 } 1123 spin_unlock(&dev->pending_lock); 1124 qp->s_flags &= ~QIB_S_BUSY; 1125 ret = -EBUSY; 1126 } 1127 spin_unlock_irqrestore(&qp->s_lock, flags); 1128 1129 return ret; 1130 } 1131 1132 static int qib_verbs_send_dma(struct qib_qp *qp, struct qib_ib_header *hdr, 1133 u32 hdrwords, struct qib_sge_state *ss, u32 len, 1134 u32 plen, u32 dwords) 1135 { 1136 struct qib_ibdev *dev = to_idev(qp->ibqp.device); 1137 struct qib_devdata *dd = dd_from_dev(dev); 1138 struct qib_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); 1139 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 1140 struct qib_verbs_txreq *tx; 1141 struct qib_pio_header *phdr; 1142 u32 control; 1143 u32 ndesc; 1144 int ret; 1145 1146 tx = qp->s_tx; 1147 if (tx) { 1148 qp->s_tx = NULL; 1149 /* resend previously constructed packet */ 1150 ret = qib_sdma_verbs_send(ppd, tx->ss, tx->dwords, tx); 1151 goto bail; 1152 } 1153 1154 tx = get_txreq(dev, qp); 1155 if (IS_ERR(tx)) 1156 goto bail_tx; 1157 1158 control = dd->f_setpbc_control(ppd, plen, qp->s_srate, 1159 be16_to_cpu(hdr->lrh[0]) >> 12); 1160 tx->qp = qp; 1161 atomic_inc(&qp->refcount); 1162 tx->wqe = qp->s_wqe; 1163 tx->mr = qp->s_rdma_mr; 1164 if (qp->s_rdma_mr) 1165 qp->s_rdma_mr = NULL; 1166 tx->txreq.callback = sdma_complete; 1167 if (dd->flags & QIB_HAS_SDMA_TIMEOUT) 1168 tx->txreq.flags = QIB_SDMA_TXREQ_F_HEADTOHOST; 1169 else 1170 tx->txreq.flags = QIB_SDMA_TXREQ_F_INTREQ; 1171 if (plen + 1 > dd->piosize2kmax_dwords) 1172 tx->txreq.flags |= QIB_SDMA_TXREQ_F_USELARGEBUF; 1173 1174 if (len) { 1175 /* 1176 * Don't try to DMA if it takes more descriptors than 1177 * the queue holds. 1178 */ 1179 ndesc = qib_count_sge(ss, len); 1180 if (ndesc >= ppd->sdma_descq_cnt) 1181 ndesc = 0; 1182 } else 1183 ndesc = 1; 1184 if (ndesc) { 1185 phdr = &dev->pio_hdrs[tx->hdr_inx]; 1186 phdr->pbc[0] = cpu_to_le32(plen); 1187 phdr->pbc[1] = cpu_to_le32(control); 1188 memcpy(&phdr->hdr, hdr, hdrwords << 2); 1189 tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEDESC; 1190 tx->txreq.sg_count = ndesc; 1191 tx->txreq.addr = dev->pio_hdrs_phys + 1192 tx->hdr_inx * sizeof(struct qib_pio_header); 1193 tx->hdr_dwords = hdrwords + 2; /* add PBC length */ 1194 ret = qib_sdma_verbs_send(ppd, ss, dwords, tx); 1195 goto bail; 1196 } 1197 1198 /* Allocate a buffer and copy the header and payload to it. */ 1199 tx->hdr_dwords = plen + 1; 1200 phdr = kmalloc(tx->hdr_dwords << 2, GFP_ATOMIC); 1201 if (!phdr) 1202 goto err_tx; 1203 phdr->pbc[0] = cpu_to_le32(plen); 1204 phdr->pbc[1] = cpu_to_le32(control); 1205 memcpy(&phdr->hdr, hdr, hdrwords << 2); 1206 qib_copy_from_sge((u32 *) &phdr->hdr + hdrwords, ss, len); 1207 1208 tx->txreq.addr = dma_map_single(&dd->pcidev->dev, phdr, 1209 tx->hdr_dwords << 2, DMA_TO_DEVICE); 1210 if (dma_mapping_error(&dd->pcidev->dev, tx->txreq.addr)) 1211 goto map_err; 1212 tx->align_buf = phdr; 1213 tx->txreq.flags |= QIB_SDMA_TXREQ_F_FREEBUF; 1214 tx->txreq.sg_count = 1; 1215 ret = qib_sdma_verbs_send(ppd, NULL, 0, tx); 1216 goto unaligned; 1217 1218 map_err: 1219 kfree(phdr); 1220 err_tx: 1221 qib_put_txreq(tx); 1222 ret = wait_kmem(dev, qp); 1223 unaligned: 1224 ibp->n_unaligned++; 1225 bail: 1226 return ret; 1227 bail_tx: 1228 ret = PTR_ERR(tx); 1229 goto bail; 1230 } 1231 1232 /* 1233 * If we are now in the error state, return zero to flush the 1234 * send work request. 1235 */ 1236 static int no_bufs_available(struct qib_qp *qp) 1237 { 1238 struct qib_ibdev *dev = to_idev(qp->ibqp.device); 1239 struct qib_devdata *dd; 1240 unsigned long flags; 1241 int ret = 0; 1242 1243 /* 1244 * Note that as soon as want_buffer() is called and 1245 * possibly before it returns, qib_ib_piobufavail() 1246 * could be called. Therefore, put QP on the I/O wait list before 1247 * enabling the PIO avail interrupt. 1248 */ 1249 spin_lock_irqsave(&qp->s_lock, flags); 1250 if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) { 1251 spin_lock(&dev->pending_lock); 1252 if (list_empty(&qp->iowait)) { 1253 dev->n_piowait++; 1254 qp->s_flags |= QIB_S_WAIT_PIO; 1255 list_add_tail(&qp->iowait, &dev->piowait); 1256 dd = dd_from_dev(dev); 1257 dd->f_wantpiobuf_intr(dd, 1); 1258 } 1259 spin_unlock(&dev->pending_lock); 1260 qp->s_flags &= ~QIB_S_BUSY; 1261 ret = -EBUSY; 1262 } 1263 spin_unlock_irqrestore(&qp->s_lock, flags); 1264 return ret; 1265 } 1266 1267 static int qib_verbs_send_pio(struct qib_qp *qp, struct qib_ib_header *ibhdr, 1268 u32 hdrwords, struct qib_sge_state *ss, u32 len, 1269 u32 plen, u32 dwords) 1270 { 1271 struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device); 1272 struct qib_pportdata *ppd = dd->pport + qp->port_num - 1; 1273 u32 *hdr = (u32 *) ibhdr; 1274 u32 __iomem *piobuf_orig; 1275 u32 __iomem *piobuf; 1276 u64 pbc; 1277 unsigned long flags; 1278 unsigned flush_wc; 1279 u32 control; 1280 u32 pbufn; 1281 1282 control = dd->f_setpbc_control(ppd, plen, qp->s_srate, 1283 be16_to_cpu(ibhdr->lrh[0]) >> 12); 1284 pbc = ((u64) control << 32) | plen; 1285 piobuf = dd->f_getsendbuf(ppd, pbc, &pbufn); 1286 if (unlikely(piobuf == NULL)) 1287 return no_bufs_available(qp); 1288 1289 /* 1290 * Write the pbc. 1291 * We have to flush after the PBC for correctness on some cpus 1292 * or WC buffer can be written out of order. 1293 */ 1294 writeq(pbc, piobuf); 1295 piobuf_orig = piobuf; 1296 piobuf += 2; 1297 1298 flush_wc = dd->flags & QIB_PIO_FLUSH_WC; 1299 if (len == 0) { 1300 /* 1301 * If there is just the header portion, must flush before 1302 * writing last word of header for correctness, and after 1303 * the last header word (trigger word). 1304 */ 1305 if (flush_wc) { 1306 qib_flush_wc(); 1307 qib_pio_copy(piobuf, hdr, hdrwords - 1); 1308 qib_flush_wc(); 1309 __raw_writel(hdr[hdrwords - 1], piobuf + hdrwords - 1); 1310 qib_flush_wc(); 1311 } else 1312 qib_pio_copy(piobuf, hdr, hdrwords); 1313 goto done; 1314 } 1315 1316 if (flush_wc) 1317 qib_flush_wc(); 1318 qib_pio_copy(piobuf, hdr, hdrwords); 1319 piobuf += hdrwords; 1320 1321 /* The common case is aligned and contained in one segment. */ 1322 if (likely(ss->num_sge == 1 && len <= ss->sge.length && 1323 !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) { 1324 u32 *addr = (u32 *) ss->sge.vaddr; 1325 1326 /* Update address before sending packet. */ 1327 update_sge(ss, len); 1328 if (flush_wc) { 1329 qib_pio_copy(piobuf, addr, dwords - 1); 1330 /* must flush early everything before trigger word */ 1331 qib_flush_wc(); 1332 __raw_writel(addr[dwords - 1], piobuf + dwords - 1); 1333 /* be sure trigger word is written */ 1334 qib_flush_wc(); 1335 } else 1336 qib_pio_copy(piobuf, addr, dwords); 1337 goto done; 1338 } 1339 copy_io(piobuf, ss, len, flush_wc); 1340 done: 1341 if (dd->flags & QIB_USE_SPCL_TRIG) { 1342 u32 spcl_off = (pbufn >= dd->piobcnt2k) ? 2047 : 1023; 1343 qib_flush_wc(); 1344 __raw_writel(0xaebecede, piobuf_orig + spcl_off); 1345 } 1346 qib_sendbuf_done(dd, pbufn); 1347 if (qp->s_rdma_mr) { 1348 qib_put_mr(qp->s_rdma_mr); 1349 qp->s_rdma_mr = NULL; 1350 } 1351 if (qp->s_wqe) { 1352 spin_lock_irqsave(&qp->s_lock, flags); 1353 qib_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS); 1354 spin_unlock_irqrestore(&qp->s_lock, flags); 1355 } else if (qp->ibqp.qp_type == IB_QPT_RC) { 1356 spin_lock_irqsave(&qp->s_lock, flags); 1357 qib_rc_send_complete(qp, ibhdr); 1358 spin_unlock_irqrestore(&qp->s_lock, flags); 1359 } 1360 return 0; 1361 } 1362 1363 /** 1364 * qib_verbs_send - send a packet 1365 * @qp: the QP to send on 1366 * @hdr: the packet header 1367 * @hdrwords: the number of 32-bit words in the header 1368 * @ss: the SGE to send 1369 * @len: the length of the packet in bytes 1370 * 1371 * Return zero if packet is sent or queued OK. 1372 * Return non-zero and clear qp->s_flags QIB_S_BUSY otherwise. 1373 */ 1374 int qib_verbs_send(struct qib_qp *qp, struct qib_ib_header *hdr, 1375 u32 hdrwords, struct qib_sge_state *ss, u32 len) 1376 { 1377 struct qib_devdata *dd = dd_from_ibdev(qp->ibqp.device); 1378 u32 plen; 1379 int ret; 1380 u32 dwords = (len + 3) >> 2; 1381 1382 /* 1383 * Calculate the send buffer trigger address. 1384 * The +1 counts for the pbc control dword following the pbc length. 1385 */ 1386 plen = hdrwords + dwords + 1; 1387 1388 /* 1389 * VL15 packets (IB_QPT_SMI) will always use PIO, so we 1390 * can defer SDMA restart until link goes ACTIVE without 1391 * worrying about just how we got there. 1392 */ 1393 if (qp->ibqp.qp_type == IB_QPT_SMI || 1394 !(dd->flags & QIB_HAS_SEND_DMA)) 1395 ret = qib_verbs_send_pio(qp, hdr, hdrwords, ss, len, 1396 plen, dwords); 1397 else 1398 ret = qib_verbs_send_dma(qp, hdr, hdrwords, ss, len, 1399 plen, dwords); 1400 1401 return ret; 1402 } 1403 1404 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords, 1405 u64 *rwords, u64 *spkts, u64 *rpkts, 1406 u64 *xmit_wait) 1407 { 1408 int ret; 1409 struct qib_devdata *dd = ppd->dd; 1410 1411 if (!(dd->flags & QIB_PRESENT)) { 1412 /* no hardware, freeze, etc. */ 1413 ret = -EINVAL; 1414 goto bail; 1415 } 1416 *swords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDSEND); 1417 *rwords = dd->f_portcntr(ppd, QIBPORTCNTR_WORDRCV); 1418 *spkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTSEND); 1419 *rpkts = dd->f_portcntr(ppd, QIBPORTCNTR_PKTRCV); 1420 *xmit_wait = dd->f_portcntr(ppd, QIBPORTCNTR_SENDSTALL); 1421 1422 ret = 0; 1423 1424 bail: 1425 return ret; 1426 } 1427 1428 /** 1429 * qib_get_counters - get various chip counters 1430 * @dd: the qlogic_ib device 1431 * @cntrs: counters are placed here 1432 * 1433 * Return the counters needed by recv_pma_get_portcounters(). 1434 */ 1435 int qib_get_counters(struct qib_pportdata *ppd, 1436 struct qib_verbs_counters *cntrs) 1437 { 1438 int ret; 1439 1440 if (!(ppd->dd->flags & QIB_PRESENT)) { 1441 /* no hardware, freeze, etc. */ 1442 ret = -EINVAL; 1443 goto bail; 1444 } 1445 cntrs->symbol_error_counter = 1446 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBSYMBOLERR); 1447 cntrs->link_error_recovery_counter = 1448 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKERRRECOV); 1449 /* 1450 * The link downed counter counts when the other side downs the 1451 * connection. We add in the number of times we downed the link 1452 * due to local link integrity errors to compensate. 1453 */ 1454 cntrs->link_downed_counter = 1455 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_IBLINKDOWN); 1456 cntrs->port_rcv_errors = 1457 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXDROPPKT) + 1458 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVOVFL) + 1459 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERR_RLEN) + 1460 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_INVALIDRLEN) + 1461 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLINK) + 1462 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRICRC) + 1463 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRVCRC) + 1464 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_ERRLPCRC) + 1465 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_BADFORMAT); 1466 cntrs->port_rcv_errors += 1467 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXLOCALPHYERR); 1468 cntrs->port_rcv_errors += 1469 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RXVLERR); 1470 cntrs->port_rcv_remphys_errors = 1471 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_RCVEBP); 1472 cntrs->port_xmit_discards = 1473 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_UNSUPVL); 1474 cntrs->port_xmit_data = ppd->dd->f_portcntr(ppd, 1475 QIBPORTCNTR_WORDSEND); 1476 cntrs->port_rcv_data = ppd->dd->f_portcntr(ppd, 1477 QIBPORTCNTR_WORDRCV); 1478 cntrs->port_xmit_packets = ppd->dd->f_portcntr(ppd, 1479 QIBPORTCNTR_PKTSEND); 1480 cntrs->port_rcv_packets = ppd->dd->f_portcntr(ppd, 1481 QIBPORTCNTR_PKTRCV); 1482 cntrs->local_link_integrity_errors = 1483 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_LLI); 1484 cntrs->excessive_buffer_overrun_errors = 1485 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_EXCESSBUFOVFL); 1486 cntrs->vl15_dropped = 1487 ppd->dd->f_portcntr(ppd, QIBPORTCNTR_VL15PKTDROP); 1488 1489 ret = 0; 1490 1491 bail: 1492 return ret; 1493 } 1494 1495 /** 1496 * qib_ib_piobufavail - callback when a PIO buffer is available 1497 * @dd: the device pointer 1498 * 1499 * This is called from qib_intr() at interrupt level when a PIO buffer is 1500 * available after qib_verbs_send() returned an error that no buffers were 1501 * available. Disable the interrupt if there are no more QPs waiting. 1502 */ 1503 void qib_ib_piobufavail(struct qib_devdata *dd) 1504 { 1505 struct qib_ibdev *dev = &dd->verbs_dev; 1506 struct list_head *list; 1507 struct qib_qp *qps[5]; 1508 struct qib_qp *qp; 1509 unsigned long flags; 1510 unsigned i, n; 1511 1512 list = &dev->piowait; 1513 n = 0; 1514 1515 /* 1516 * Note: checking that the piowait list is empty and clearing 1517 * the buffer available interrupt needs to be atomic or we 1518 * could end up with QPs on the wait list with the interrupt 1519 * disabled. 1520 */ 1521 spin_lock_irqsave(&dev->pending_lock, flags); 1522 while (!list_empty(list)) { 1523 if (n == ARRAY_SIZE(qps)) 1524 goto full; 1525 qp = list_entry(list->next, struct qib_qp, iowait); 1526 list_del_init(&qp->iowait); 1527 atomic_inc(&qp->refcount); 1528 qps[n++] = qp; 1529 } 1530 dd->f_wantpiobuf_intr(dd, 0); 1531 full: 1532 spin_unlock_irqrestore(&dev->pending_lock, flags); 1533 1534 for (i = 0; i < n; i++) { 1535 qp = qps[i]; 1536 1537 spin_lock_irqsave(&qp->s_lock, flags); 1538 if (qp->s_flags & QIB_S_WAIT_PIO) { 1539 qp->s_flags &= ~QIB_S_WAIT_PIO; 1540 qib_schedule_send(qp); 1541 } 1542 spin_unlock_irqrestore(&qp->s_lock, flags); 1543 1544 /* Notify qib_destroy_qp() if it is waiting. */ 1545 if (atomic_dec_and_test(&qp->refcount)) 1546 wake_up(&qp->wait); 1547 } 1548 } 1549 1550 static int qib_query_device(struct ib_device *ibdev, 1551 struct ib_device_attr *props) 1552 { 1553 struct qib_devdata *dd = dd_from_ibdev(ibdev); 1554 struct qib_ibdev *dev = to_idev(ibdev); 1555 1556 memset(props, 0, sizeof(*props)); 1557 1558 props->device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR | 1559 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT | 1560 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN | 1561 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE; 1562 props->page_size_cap = PAGE_SIZE; 1563 props->vendor_id = 1564 QIB_SRC_OUI_1 << 16 | QIB_SRC_OUI_2 << 8 | QIB_SRC_OUI_3; 1565 props->vendor_part_id = dd->deviceid; 1566 props->hw_ver = dd->minrev; 1567 props->sys_image_guid = ib_qib_sys_image_guid; 1568 props->max_mr_size = ~0ULL; 1569 props->max_qp = ib_qib_max_qps; 1570 props->max_qp_wr = ib_qib_max_qp_wrs; 1571 props->max_sge = ib_qib_max_sges; 1572 props->max_cq = ib_qib_max_cqs; 1573 props->max_ah = ib_qib_max_ahs; 1574 props->max_cqe = ib_qib_max_cqes; 1575 props->max_mr = dev->lk_table.max; 1576 props->max_fmr = dev->lk_table.max; 1577 props->max_map_per_fmr = 32767; 1578 props->max_pd = ib_qib_max_pds; 1579 props->max_qp_rd_atom = QIB_MAX_RDMA_ATOMIC; 1580 props->max_qp_init_rd_atom = 255; 1581 /* props->max_res_rd_atom */ 1582 props->max_srq = ib_qib_max_srqs; 1583 props->max_srq_wr = ib_qib_max_srq_wrs; 1584 props->max_srq_sge = ib_qib_max_srq_sges; 1585 /* props->local_ca_ack_delay */ 1586 props->atomic_cap = IB_ATOMIC_GLOB; 1587 props->max_pkeys = qib_get_npkeys(dd); 1588 props->max_mcast_grp = ib_qib_max_mcast_grps; 1589 props->max_mcast_qp_attach = ib_qib_max_mcast_qp_attached; 1590 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach * 1591 props->max_mcast_grp; 1592 1593 return 0; 1594 } 1595 1596 static int qib_query_port(struct ib_device *ibdev, u8 port, 1597 struct ib_port_attr *props) 1598 { 1599 struct qib_devdata *dd = dd_from_ibdev(ibdev); 1600 struct qib_ibport *ibp = to_iport(ibdev, port); 1601 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 1602 enum ib_mtu mtu; 1603 u16 lid = ppd->lid; 1604 1605 memset(props, 0, sizeof(*props)); 1606 props->lid = lid ? lid : be16_to_cpu(IB_LID_PERMISSIVE); 1607 props->lmc = ppd->lmc; 1608 props->sm_lid = ibp->sm_lid; 1609 props->sm_sl = ibp->sm_sl; 1610 props->state = dd->f_iblink_state(ppd->lastibcstat); 1611 props->phys_state = dd->f_ibphys_portstate(ppd->lastibcstat); 1612 props->port_cap_flags = ibp->port_cap_flags; 1613 props->gid_tbl_len = QIB_GUIDS_PER_PORT; 1614 props->max_msg_sz = 0x80000000; 1615 props->pkey_tbl_len = qib_get_npkeys(dd); 1616 props->bad_pkey_cntr = ibp->pkey_violations; 1617 props->qkey_viol_cntr = ibp->qkey_violations; 1618 props->active_width = ppd->link_width_active; 1619 /* See rate_show() */ 1620 props->active_speed = ppd->link_speed_active; 1621 props->max_vl_num = qib_num_vls(ppd->vls_supported); 1622 props->init_type_reply = 0; 1623 1624 props->max_mtu = qib_ibmtu ? qib_ibmtu : IB_MTU_4096; 1625 switch (ppd->ibmtu) { 1626 case 4096: 1627 mtu = IB_MTU_4096; 1628 break; 1629 case 2048: 1630 mtu = IB_MTU_2048; 1631 break; 1632 case 1024: 1633 mtu = IB_MTU_1024; 1634 break; 1635 case 512: 1636 mtu = IB_MTU_512; 1637 break; 1638 case 256: 1639 mtu = IB_MTU_256; 1640 break; 1641 default: 1642 mtu = IB_MTU_2048; 1643 } 1644 props->active_mtu = mtu; 1645 props->subnet_timeout = ibp->subnet_timeout; 1646 1647 return 0; 1648 } 1649 1650 static int qib_modify_device(struct ib_device *device, 1651 int device_modify_mask, 1652 struct ib_device_modify *device_modify) 1653 { 1654 struct qib_devdata *dd = dd_from_ibdev(device); 1655 unsigned i; 1656 int ret; 1657 1658 if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID | 1659 IB_DEVICE_MODIFY_NODE_DESC)) { 1660 ret = -EOPNOTSUPP; 1661 goto bail; 1662 } 1663 1664 if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) { 1665 memcpy(device->node_desc, device_modify->node_desc, 64); 1666 for (i = 0; i < dd->num_pports; i++) { 1667 struct qib_ibport *ibp = &dd->pport[i].ibport_data; 1668 1669 qib_node_desc_chg(ibp); 1670 } 1671 } 1672 1673 if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) { 1674 ib_qib_sys_image_guid = 1675 cpu_to_be64(device_modify->sys_image_guid); 1676 for (i = 0; i < dd->num_pports; i++) { 1677 struct qib_ibport *ibp = &dd->pport[i].ibport_data; 1678 1679 qib_sys_guid_chg(ibp); 1680 } 1681 } 1682 1683 ret = 0; 1684 1685 bail: 1686 return ret; 1687 } 1688 1689 static int qib_modify_port(struct ib_device *ibdev, u8 port, 1690 int port_modify_mask, struct ib_port_modify *props) 1691 { 1692 struct qib_ibport *ibp = to_iport(ibdev, port); 1693 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 1694 1695 ibp->port_cap_flags |= props->set_port_cap_mask; 1696 ibp->port_cap_flags &= ~props->clr_port_cap_mask; 1697 if (props->set_port_cap_mask || props->clr_port_cap_mask) 1698 qib_cap_mask_chg(ibp); 1699 if (port_modify_mask & IB_PORT_SHUTDOWN) 1700 qib_set_linkstate(ppd, QIB_IB_LINKDOWN); 1701 if (port_modify_mask & IB_PORT_RESET_QKEY_CNTR) 1702 ibp->qkey_violations = 0; 1703 return 0; 1704 } 1705 1706 static int qib_query_gid(struct ib_device *ibdev, u8 port, 1707 int index, union ib_gid *gid) 1708 { 1709 struct qib_devdata *dd = dd_from_ibdev(ibdev); 1710 int ret = 0; 1711 1712 if (!port || port > dd->num_pports) 1713 ret = -EINVAL; 1714 else { 1715 struct qib_ibport *ibp = to_iport(ibdev, port); 1716 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 1717 1718 gid->global.subnet_prefix = ibp->gid_prefix; 1719 if (index == 0) 1720 gid->global.interface_id = ppd->guid; 1721 else if (index < QIB_GUIDS_PER_PORT) 1722 gid->global.interface_id = ibp->guids[index - 1]; 1723 else 1724 ret = -EINVAL; 1725 } 1726 1727 return ret; 1728 } 1729 1730 static struct ib_pd *qib_alloc_pd(struct ib_device *ibdev, 1731 struct ib_ucontext *context, 1732 struct ib_udata *udata) 1733 { 1734 struct qib_ibdev *dev = to_idev(ibdev); 1735 struct qib_pd *pd; 1736 struct ib_pd *ret; 1737 1738 /* 1739 * This is actually totally arbitrary. Some correctness tests 1740 * assume there's a maximum number of PDs that can be allocated. 1741 * We don't actually have this limit, but we fail the test if 1742 * we allow allocations of more than we report for this value. 1743 */ 1744 1745 pd = kmalloc(sizeof *pd, GFP_KERNEL); 1746 if (!pd) { 1747 ret = ERR_PTR(-ENOMEM); 1748 goto bail; 1749 } 1750 1751 spin_lock(&dev->n_pds_lock); 1752 if (dev->n_pds_allocated == ib_qib_max_pds) { 1753 spin_unlock(&dev->n_pds_lock); 1754 kfree(pd); 1755 ret = ERR_PTR(-ENOMEM); 1756 goto bail; 1757 } 1758 1759 dev->n_pds_allocated++; 1760 spin_unlock(&dev->n_pds_lock); 1761 1762 /* ib_alloc_pd() will initialize pd->ibpd. */ 1763 pd->user = udata != NULL; 1764 1765 ret = &pd->ibpd; 1766 1767 bail: 1768 return ret; 1769 } 1770 1771 static int qib_dealloc_pd(struct ib_pd *ibpd) 1772 { 1773 struct qib_pd *pd = to_ipd(ibpd); 1774 struct qib_ibdev *dev = to_idev(ibpd->device); 1775 1776 spin_lock(&dev->n_pds_lock); 1777 dev->n_pds_allocated--; 1778 spin_unlock(&dev->n_pds_lock); 1779 1780 kfree(pd); 1781 1782 return 0; 1783 } 1784 1785 int qib_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr) 1786 { 1787 /* A multicast address requires a GRH (see ch. 8.4.1). */ 1788 if (ah_attr->dlid >= QIB_MULTICAST_LID_BASE && 1789 ah_attr->dlid != QIB_PERMISSIVE_LID && 1790 !(ah_attr->ah_flags & IB_AH_GRH)) 1791 goto bail; 1792 if ((ah_attr->ah_flags & IB_AH_GRH) && 1793 ah_attr->grh.sgid_index >= QIB_GUIDS_PER_PORT) 1794 goto bail; 1795 if (ah_attr->dlid == 0) 1796 goto bail; 1797 if (ah_attr->port_num < 1 || 1798 ah_attr->port_num > ibdev->phys_port_cnt) 1799 goto bail; 1800 if (ah_attr->static_rate != IB_RATE_PORT_CURRENT && 1801 ib_rate_to_mult(ah_attr->static_rate) < 0) 1802 goto bail; 1803 if (ah_attr->sl > 15) 1804 goto bail; 1805 return 0; 1806 bail: 1807 return -EINVAL; 1808 } 1809 1810 /** 1811 * qib_create_ah - create an address handle 1812 * @pd: the protection domain 1813 * @ah_attr: the attributes of the AH 1814 * 1815 * This may be called from interrupt context. 1816 */ 1817 static struct ib_ah *qib_create_ah(struct ib_pd *pd, 1818 struct ib_ah_attr *ah_attr) 1819 { 1820 struct qib_ah *ah; 1821 struct ib_ah *ret; 1822 struct qib_ibdev *dev = to_idev(pd->device); 1823 unsigned long flags; 1824 1825 if (qib_check_ah(pd->device, ah_attr)) { 1826 ret = ERR_PTR(-EINVAL); 1827 goto bail; 1828 } 1829 1830 ah = kmalloc(sizeof *ah, GFP_ATOMIC); 1831 if (!ah) { 1832 ret = ERR_PTR(-ENOMEM); 1833 goto bail; 1834 } 1835 1836 spin_lock_irqsave(&dev->n_ahs_lock, flags); 1837 if (dev->n_ahs_allocated == ib_qib_max_ahs) { 1838 spin_unlock_irqrestore(&dev->n_ahs_lock, flags); 1839 kfree(ah); 1840 ret = ERR_PTR(-ENOMEM); 1841 goto bail; 1842 } 1843 1844 dev->n_ahs_allocated++; 1845 spin_unlock_irqrestore(&dev->n_ahs_lock, flags); 1846 1847 /* ib_create_ah() will initialize ah->ibah. */ 1848 ah->attr = *ah_attr; 1849 atomic_set(&ah->refcount, 0); 1850 1851 ret = &ah->ibah; 1852 1853 bail: 1854 return ret; 1855 } 1856 1857 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid) 1858 { 1859 struct ib_ah_attr attr; 1860 struct ib_ah *ah = ERR_PTR(-EINVAL); 1861 struct qib_qp *qp0; 1862 1863 memset(&attr, 0, sizeof attr); 1864 attr.dlid = dlid; 1865 attr.port_num = ppd_from_ibp(ibp)->port; 1866 rcu_read_lock(); 1867 qp0 = rcu_dereference(ibp->qp0); 1868 if (qp0) 1869 ah = ib_create_ah(qp0->ibqp.pd, &attr); 1870 rcu_read_unlock(); 1871 return ah; 1872 } 1873 1874 /** 1875 * qib_destroy_ah - destroy an address handle 1876 * @ibah: the AH to destroy 1877 * 1878 * This may be called from interrupt context. 1879 */ 1880 static int qib_destroy_ah(struct ib_ah *ibah) 1881 { 1882 struct qib_ibdev *dev = to_idev(ibah->device); 1883 struct qib_ah *ah = to_iah(ibah); 1884 unsigned long flags; 1885 1886 if (atomic_read(&ah->refcount) != 0) 1887 return -EBUSY; 1888 1889 spin_lock_irqsave(&dev->n_ahs_lock, flags); 1890 dev->n_ahs_allocated--; 1891 spin_unlock_irqrestore(&dev->n_ahs_lock, flags); 1892 1893 kfree(ah); 1894 1895 return 0; 1896 } 1897 1898 static int qib_modify_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr) 1899 { 1900 struct qib_ah *ah = to_iah(ibah); 1901 1902 if (qib_check_ah(ibah->device, ah_attr)) 1903 return -EINVAL; 1904 1905 ah->attr = *ah_attr; 1906 1907 return 0; 1908 } 1909 1910 static int qib_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr) 1911 { 1912 struct qib_ah *ah = to_iah(ibah); 1913 1914 *ah_attr = ah->attr; 1915 1916 return 0; 1917 } 1918 1919 /** 1920 * qib_get_npkeys - return the size of the PKEY table for context 0 1921 * @dd: the qlogic_ib device 1922 */ 1923 unsigned qib_get_npkeys(struct qib_devdata *dd) 1924 { 1925 return ARRAY_SIZE(dd->rcd[0]->pkeys); 1926 } 1927 1928 /* 1929 * Return the indexed PKEY from the port PKEY table. 1930 * No need to validate rcd[ctxt]; the port is setup if we are here. 1931 */ 1932 unsigned qib_get_pkey(struct qib_ibport *ibp, unsigned index) 1933 { 1934 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 1935 struct qib_devdata *dd = ppd->dd; 1936 unsigned ctxt = ppd->hw_pidx; 1937 unsigned ret; 1938 1939 /* dd->rcd null if mini_init or some init failures */ 1940 if (!dd->rcd || index >= ARRAY_SIZE(dd->rcd[ctxt]->pkeys)) 1941 ret = 0; 1942 else 1943 ret = dd->rcd[ctxt]->pkeys[index]; 1944 1945 return ret; 1946 } 1947 1948 static int qib_query_pkey(struct ib_device *ibdev, u8 port, u16 index, 1949 u16 *pkey) 1950 { 1951 struct qib_devdata *dd = dd_from_ibdev(ibdev); 1952 int ret; 1953 1954 if (index >= qib_get_npkeys(dd)) { 1955 ret = -EINVAL; 1956 goto bail; 1957 } 1958 1959 *pkey = qib_get_pkey(to_iport(ibdev, port), index); 1960 ret = 0; 1961 1962 bail: 1963 return ret; 1964 } 1965 1966 /** 1967 * qib_alloc_ucontext - allocate a ucontest 1968 * @ibdev: the infiniband device 1969 * @udata: not used by the QLogic_IB driver 1970 */ 1971 1972 static struct ib_ucontext *qib_alloc_ucontext(struct ib_device *ibdev, 1973 struct ib_udata *udata) 1974 { 1975 struct qib_ucontext *context; 1976 struct ib_ucontext *ret; 1977 1978 context = kmalloc(sizeof *context, GFP_KERNEL); 1979 if (!context) { 1980 ret = ERR_PTR(-ENOMEM); 1981 goto bail; 1982 } 1983 1984 ret = &context->ibucontext; 1985 1986 bail: 1987 return ret; 1988 } 1989 1990 static int qib_dealloc_ucontext(struct ib_ucontext *context) 1991 { 1992 kfree(to_iucontext(context)); 1993 return 0; 1994 } 1995 1996 static void init_ibport(struct qib_pportdata *ppd) 1997 { 1998 struct qib_verbs_counters cntrs; 1999 struct qib_ibport *ibp = &ppd->ibport_data; 2000 2001 spin_lock_init(&ibp->lock); 2002 /* Set the prefix to the default value (see ch. 4.1.1) */ 2003 ibp->gid_prefix = IB_DEFAULT_GID_PREFIX; 2004 ibp->sm_lid = be16_to_cpu(IB_LID_PERMISSIVE); 2005 ibp->port_cap_flags = IB_PORT_SYS_IMAGE_GUID_SUP | 2006 IB_PORT_CLIENT_REG_SUP | IB_PORT_SL_MAP_SUP | 2007 IB_PORT_TRAP_SUP | IB_PORT_AUTO_MIGR_SUP | 2008 IB_PORT_DR_NOTICE_SUP | IB_PORT_CAP_MASK_NOTICE_SUP | 2009 IB_PORT_OTHER_LOCAL_CHANGES_SUP; 2010 if (ppd->dd->flags & QIB_HAS_LINK_LATENCY) 2011 ibp->port_cap_flags |= IB_PORT_LINK_LATENCY_SUP; 2012 ibp->pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA; 2013 ibp->pma_counter_select[1] = IB_PMA_PORT_RCV_DATA; 2014 ibp->pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS; 2015 ibp->pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS; 2016 ibp->pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT; 2017 2018 /* Snapshot current HW counters to "clear" them. */ 2019 qib_get_counters(ppd, &cntrs); 2020 ibp->z_symbol_error_counter = cntrs.symbol_error_counter; 2021 ibp->z_link_error_recovery_counter = 2022 cntrs.link_error_recovery_counter; 2023 ibp->z_link_downed_counter = cntrs.link_downed_counter; 2024 ibp->z_port_rcv_errors = cntrs.port_rcv_errors; 2025 ibp->z_port_rcv_remphys_errors = cntrs.port_rcv_remphys_errors; 2026 ibp->z_port_xmit_discards = cntrs.port_xmit_discards; 2027 ibp->z_port_xmit_data = cntrs.port_xmit_data; 2028 ibp->z_port_rcv_data = cntrs.port_rcv_data; 2029 ibp->z_port_xmit_packets = cntrs.port_xmit_packets; 2030 ibp->z_port_rcv_packets = cntrs.port_rcv_packets; 2031 ibp->z_local_link_integrity_errors = 2032 cntrs.local_link_integrity_errors; 2033 ibp->z_excessive_buffer_overrun_errors = 2034 cntrs.excessive_buffer_overrun_errors; 2035 ibp->z_vl15_dropped = cntrs.vl15_dropped; 2036 RCU_INIT_POINTER(ibp->qp0, NULL); 2037 RCU_INIT_POINTER(ibp->qp1, NULL); 2038 } 2039 2040 /** 2041 * qib_register_ib_device - register our device with the infiniband core 2042 * @dd: the device data structure 2043 * Return the allocated qib_ibdev pointer or NULL on error. 2044 */ 2045 int qib_register_ib_device(struct qib_devdata *dd) 2046 { 2047 struct qib_ibdev *dev = &dd->verbs_dev; 2048 struct ib_device *ibdev = &dev->ibdev; 2049 struct qib_pportdata *ppd = dd->pport; 2050 unsigned i, lk_tab_size; 2051 int ret; 2052 2053 dev->qp_table_size = ib_qib_qp_table_size; 2054 get_random_bytes(&dev->qp_rnd, sizeof(dev->qp_rnd)); 2055 dev->qp_table = kmalloc(dev->qp_table_size * sizeof *dev->qp_table, 2056 GFP_KERNEL); 2057 if (!dev->qp_table) { 2058 ret = -ENOMEM; 2059 goto err_qpt; 2060 } 2061 for (i = 0; i < dev->qp_table_size; i++) 2062 RCU_INIT_POINTER(dev->qp_table[i], NULL); 2063 2064 for (i = 0; i < dd->num_pports; i++) 2065 init_ibport(ppd + i); 2066 2067 /* Only need to initialize non-zero fields. */ 2068 spin_lock_init(&dev->qpt_lock); 2069 spin_lock_init(&dev->n_pds_lock); 2070 spin_lock_init(&dev->n_ahs_lock); 2071 spin_lock_init(&dev->n_cqs_lock); 2072 spin_lock_init(&dev->n_qps_lock); 2073 spin_lock_init(&dev->n_srqs_lock); 2074 spin_lock_init(&dev->n_mcast_grps_lock); 2075 init_timer(&dev->mem_timer); 2076 dev->mem_timer.function = mem_timer; 2077 dev->mem_timer.data = (unsigned long) dev; 2078 2079 qib_init_qpn_table(dd, &dev->qpn_table); 2080 2081 /* 2082 * The top ib_qib_lkey_table_size bits are used to index the 2083 * table. The lower 8 bits can be owned by the user (copied from 2084 * the LKEY). The remaining bits act as a generation number or tag. 2085 */ 2086 spin_lock_init(&dev->lk_table.lock); 2087 dev->lk_table.max = 1 << ib_qib_lkey_table_size; 2088 lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table); 2089 dev->lk_table.table = (struct qib_mregion __rcu **) 2090 __get_free_pages(GFP_KERNEL, get_order(lk_tab_size)); 2091 if (dev->lk_table.table == NULL) { 2092 ret = -ENOMEM; 2093 goto err_lk; 2094 } 2095 RCU_INIT_POINTER(dev->dma_mr, NULL); 2096 for (i = 0; i < dev->lk_table.max; i++) 2097 RCU_INIT_POINTER(dev->lk_table.table[i], NULL); 2098 INIT_LIST_HEAD(&dev->pending_mmaps); 2099 spin_lock_init(&dev->pending_lock); 2100 dev->mmap_offset = PAGE_SIZE; 2101 spin_lock_init(&dev->mmap_offset_lock); 2102 INIT_LIST_HEAD(&dev->piowait); 2103 INIT_LIST_HEAD(&dev->dmawait); 2104 INIT_LIST_HEAD(&dev->txwait); 2105 INIT_LIST_HEAD(&dev->memwait); 2106 INIT_LIST_HEAD(&dev->txreq_free); 2107 2108 if (ppd->sdma_descq_cnt) { 2109 dev->pio_hdrs = dma_alloc_coherent(&dd->pcidev->dev, 2110 ppd->sdma_descq_cnt * 2111 sizeof(struct qib_pio_header), 2112 &dev->pio_hdrs_phys, 2113 GFP_KERNEL); 2114 if (!dev->pio_hdrs) { 2115 ret = -ENOMEM; 2116 goto err_hdrs; 2117 } 2118 } 2119 2120 for (i = 0; i < ppd->sdma_descq_cnt; i++) { 2121 struct qib_verbs_txreq *tx; 2122 2123 tx = kzalloc(sizeof *tx, GFP_KERNEL); 2124 if (!tx) { 2125 ret = -ENOMEM; 2126 goto err_tx; 2127 } 2128 tx->hdr_inx = i; 2129 list_add(&tx->txreq.list, &dev->txreq_free); 2130 } 2131 2132 /* 2133 * The system image GUID is supposed to be the same for all 2134 * IB HCAs in a single system but since there can be other 2135 * device types in the system, we can't be sure this is unique. 2136 */ 2137 if (!ib_qib_sys_image_guid) 2138 ib_qib_sys_image_guid = ppd->guid; 2139 2140 strlcpy(ibdev->name, "qib%d", IB_DEVICE_NAME_MAX); 2141 ibdev->owner = THIS_MODULE; 2142 ibdev->node_guid = ppd->guid; 2143 ibdev->uverbs_abi_ver = QIB_UVERBS_ABI_VERSION; 2144 ibdev->uverbs_cmd_mask = 2145 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) | 2146 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) | 2147 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) | 2148 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) | 2149 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) | 2150 (1ull << IB_USER_VERBS_CMD_CREATE_AH) | 2151 (1ull << IB_USER_VERBS_CMD_MODIFY_AH) | 2152 (1ull << IB_USER_VERBS_CMD_QUERY_AH) | 2153 (1ull << IB_USER_VERBS_CMD_DESTROY_AH) | 2154 (1ull << IB_USER_VERBS_CMD_REG_MR) | 2155 (1ull << IB_USER_VERBS_CMD_DEREG_MR) | 2156 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) | 2157 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) | 2158 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) | 2159 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) | 2160 (1ull << IB_USER_VERBS_CMD_POLL_CQ) | 2161 (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) | 2162 (1ull << IB_USER_VERBS_CMD_CREATE_QP) | 2163 (1ull << IB_USER_VERBS_CMD_QUERY_QP) | 2164 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) | 2165 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) | 2166 (1ull << IB_USER_VERBS_CMD_POST_SEND) | 2167 (1ull << IB_USER_VERBS_CMD_POST_RECV) | 2168 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) | 2169 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) | 2170 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) | 2171 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) | 2172 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) | 2173 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) | 2174 (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV); 2175 ibdev->node_type = RDMA_NODE_IB_CA; 2176 ibdev->phys_port_cnt = dd->num_pports; 2177 ibdev->num_comp_vectors = 1; 2178 ibdev->dma_device = &dd->pcidev->dev; 2179 ibdev->query_device = qib_query_device; 2180 ibdev->modify_device = qib_modify_device; 2181 ibdev->query_port = qib_query_port; 2182 ibdev->modify_port = qib_modify_port; 2183 ibdev->query_pkey = qib_query_pkey; 2184 ibdev->query_gid = qib_query_gid; 2185 ibdev->alloc_ucontext = qib_alloc_ucontext; 2186 ibdev->dealloc_ucontext = qib_dealloc_ucontext; 2187 ibdev->alloc_pd = qib_alloc_pd; 2188 ibdev->dealloc_pd = qib_dealloc_pd; 2189 ibdev->create_ah = qib_create_ah; 2190 ibdev->destroy_ah = qib_destroy_ah; 2191 ibdev->modify_ah = qib_modify_ah; 2192 ibdev->query_ah = qib_query_ah; 2193 ibdev->create_srq = qib_create_srq; 2194 ibdev->modify_srq = qib_modify_srq; 2195 ibdev->query_srq = qib_query_srq; 2196 ibdev->destroy_srq = qib_destroy_srq; 2197 ibdev->create_qp = qib_create_qp; 2198 ibdev->modify_qp = qib_modify_qp; 2199 ibdev->query_qp = qib_query_qp; 2200 ibdev->destroy_qp = qib_destroy_qp; 2201 ibdev->post_send = qib_post_send; 2202 ibdev->post_recv = qib_post_receive; 2203 ibdev->post_srq_recv = qib_post_srq_receive; 2204 ibdev->create_cq = qib_create_cq; 2205 ibdev->destroy_cq = qib_destroy_cq; 2206 ibdev->resize_cq = qib_resize_cq; 2207 ibdev->poll_cq = qib_poll_cq; 2208 ibdev->req_notify_cq = qib_req_notify_cq; 2209 ibdev->get_dma_mr = qib_get_dma_mr; 2210 ibdev->reg_phys_mr = qib_reg_phys_mr; 2211 ibdev->reg_user_mr = qib_reg_user_mr; 2212 ibdev->dereg_mr = qib_dereg_mr; 2213 ibdev->alloc_fast_reg_mr = qib_alloc_fast_reg_mr; 2214 ibdev->alloc_fast_reg_page_list = qib_alloc_fast_reg_page_list; 2215 ibdev->free_fast_reg_page_list = qib_free_fast_reg_page_list; 2216 ibdev->alloc_fmr = qib_alloc_fmr; 2217 ibdev->map_phys_fmr = qib_map_phys_fmr; 2218 ibdev->unmap_fmr = qib_unmap_fmr; 2219 ibdev->dealloc_fmr = qib_dealloc_fmr; 2220 ibdev->attach_mcast = qib_multicast_attach; 2221 ibdev->detach_mcast = qib_multicast_detach; 2222 ibdev->process_mad = qib_process_mad; 2223 ibdev->mmap = qib_mmap; 2224 ibdev->dma_ops = &qib_dma_mapping_ops; 2225 2226 snprintf(ibdev->node_desc, sizeof(ibdev->node_desc), 2227 "Intel Infiniband HCA %s", init_utsname()->nodename); 2228 2229 ret = ib_register_device(ibdev, qib_create_port_files); 2230 if (ret) 2231 goto err_reg; 2232 2233 ret = qib_create_agents(dev); 2234 if (ret) 2235 goto err_agents; 2236 2237 ret = qib_verbs_register_sysfs(dd); 2238 if (ret) 2239 goto err_class; 2240 2241 goto bail; 2242 2243 err_class: 2244 qib_free_agents(dev); 2245 err_agents: 2246 ib_unregister_device(ibdev); 2247 err_reg: 2248 err_tx: 2249 while (!list_empty(&dev->txreq_free)) { 2250 struct list_head *l = dev->txreq_free.next; 2251 struct qib_verbs_txreq *tx; 2252 2253 list_del(l); 2254 tx = list_entry(l, struct qib_verbs_txreq, txreq.list); 2255 kfree(tx); 2256 } 2257 if (ppd->sdma_descq_cnt) 2258 dma_free_coherent(&dd->pcidev->dev, 2259 ppd->sdma_descq_cnt * 2260 sizeof(struct qib_pio_header), 2261 dev->pio_hdrs, dev->pio_hdrs_phys); 2262 err_hdrs: 2263 free_pages((unsigned long) dev->lk_table.table, get_order(lk_tab_size)); 2264 err_lk: 2265 kfree(dev->qp_table); 2266 err_qpt: 2267 qib_dev_err(dd, "cannot register verbs: %d!\n", -ret); 2268 bail: 2269 return ret; 2270 } 2271 2272 void qib_unregister_ib_device(struct qib_devdata *dd) 2273 { 2274 struct qib_ibdev *dev = &dd->verbs_dev; 2275 struct ib_device *ibdev = &dev->ibdev; 2276 u32 qps_inuse; 2277 unsigned lk_tab_size; 2278 2279 qib_verbs_unregister_sysfs(dd); 2280 2281 qib_free_agents(dev); 2282 2283 ib_unregister_device(ibdev); 2284 2285 if (!list_empty(&dev->piowait)) 2286 qib_dev_err(dd, "piowait list not empty!\n"); 2287 if (!list_empty(&dev->dmawait)) 2288 qib_dev_err(dd, "dmawait list not empty!\n"); 2289 if (!list_empty(&dev->txwait)) 2290 qib_dev_err(dd, "txwait list not empty!\n"); 2291 if (!list_empty(&dev->memwait)) 2292 qib_dev_err(dd, "memwait list not empty!\n"); 2293 if (dev->dma_mr) 2294 qib_dev_err(dd, "DMA MR not NULL!\n"); 2295 2296 qps_inuse = qib_free_all_qps(dd); 2297 if (qps_inuse) 2298 qib_dev_err(dd, "QP memory leak! %u still in use\n", 2299 qps_inuse); 2300 2301 del_timer_sync(&dev->mem_timer); 2302 qib_free_qpn_table(&dev->qpn_table); 2303 while (!list_empty(&dev->txreq_free)) { 2304 struct list_head *l = dev->txreq_free.next; 2305 struct qib_verbs_txreq *tx; 2306 2307 list_del(l); 2308 tx = list_entry(l, struct qib_verbs_txreq, txreq.list); 2309 kfree(tx); 2310 } 2311 if (dd->pport->sdma_descq_cnt) 2312 dma_free_coherent(&dd->pcidev->dev, 2313 dd->pport->sdma_descq_cnt * 2314 sizeof(struct qib_pio_header), 2315 dev->pio_hdrs, dev->pio_hdrs_phys); 2316 lk_tab_size = dev->lk_table.max * sizeof(*dev->lk_table.table); 2317 free_pages((unsigned long) dev->lk_table.table, 2318 get_order(lk_tab_size)); 2319 kfree(dev->qp_table); 2320 } 2321 2322 /* 2323 * This must be called with s_lock held. 2324 */ 2325 void qib_schedule_send(struct qib_qp *qp) 2326 { 2327 if (qib_send_ok(qp)) { 2328 struct qib_ibport *ibp = 2329 to_iport(qp->ibqp.device, qp->port_num); 2330 struct qib_pportdata *ppd = ppd_from_ibp(ibp); 2331 2332 queue_work(ppd->qib_wq, &qp->s_work); 2333 } 2334 } 2335