1/* 2 * Copyright(c) 2016 - 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/slab.h> 49#include <linux/vmalloc.h> 50#include "cq.h" 51#include "vt.h" 52#include "trace.h" 53 54static struct workqueue_struct *comp_vector_wq; 55 56/** 57 * rvt_cq_enter - add a new entry to the completion queue 58 * @cq: completion queue 59 * @entry: work completion entry to add 60 * @solicited: true if @entry is solicited 61 * 62 * This may be called with qp->s_lock held. 63 * 64 * Return: return true on success, else return 65 * false if cq is full. 66 */ 67bool rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited) 68{ 69 struct ib_uverbs_wc *uqueue = NULL; 70 struct ib_wc *kqueue = NULL; 71 struct rvt_cq_wc *u_wc = NULL; 72 struct rvt_k_cq_wc *k_wc = NULL; 73 unsigned long flags; 74 u32 head; 75 u32 next; 76 u32 tail; 77 78 spin_lock_irqsave(&cq->lock, flags); 79 80 if (cq->ip) { 81 u_wc = cq->queue; 82 uqueue = &u_wc->uqueue[0]; 83 head = RDMA_READ_UAPI_ATOMIC(u_wc->head); 84 tail = RDMA_READ_UAPI_ATOMIC(u_wc->tail); 85 } else { 86 k_wc = cq->kqueue; 87 kqueue = &k_wc->kqueue[0]; 88 head = k_wc->head; 89 tail = k_wc->tail; 90 } 91 92 /* 93 * Note that the head pointer might be writable by 94 * user processes.Take care to verify it is a sane value. 95 */ 96 if (head >= (unsigned)cq->ibcq.cqe) { 97 head = cq->ibcq.cqe; 98 next = 0; 99 } else { 100 next = head + 1; 101 } 102 103 if (unlikely(next == tail || cq->cq_full)) { 104 struct rvt_dev_info *rdi = cq->rdi; 105 106 if (!cq->cq_full) 107 rvt_pr_err_ratelimited(rdi, "CQ is full!\n"); 108 cq->cq_full = true; 109 spin_unlock_irqrestore(&cq->lock, flags); 110 if (cq->ibcq.event_handler) { 111 struct ib_event ev; 112 113 ev.device = cq->ibcq.device; 114 ev.element.cq = &cq->ibcq; 115 ev.event = IB_EVENT_CQ_ERR; 116 cq->ibcq.event_handler(&ev, cq->ibcq.cq_context); 117 } 118 return false; 119 } 120 trace_rvt_cq_enter(cq, entry, head); 121 if (uqueue) { 122 uqueue[head].wr_id = entry->wr_id; 123 uqueue[head].status = entry->status; 124 uqueue[head].opcode = entry->opcode; 125 uqueue[head].vendor_err = entry->vendor_err; 126 uqueue[head].byte_len = entry->byte_len; 127 uqueue[head].ex.imm_data = entry->ex.imm_data; 128 uqueue[head].qp_num = entry->qp->qp_num; 129 uqueue[head].src_qp = entry->src_qp; 130 uqueue[head].wc_flags = entry->wc_flags; 131 uqueue[head].pkey_index = entry->pkey_index; 132 uqueue[head].slid = ib_lid_cpu16(entry->slid); 133 uqueue[head].sl = entry->sl; 134 uqueue[head].dlid_path_bits = entry->dlid_path_bits; 135 uqueue[head].port_num = entry->port_num; 136 /* Make sure entry is written before the head index. */ 137 RDMA_WRITE_UAPI_ATOMIC(u_wc->head, next); 138 } else { 139 kqueue[head] = *entry; 140 k_wc->head = next; 141 } 142 143 if (cq->notify == IB_CQ_NEXT_COMP || 144 (cq->notify == IB_CQ_SOLICITED && 145 (solicited || entry->status != IB_WC_SUCCESS))) { 146 /* 147 * This will cause send_complete() to be called in 148 * another thread. 149 */ 150 cq->notify = RVT_CQ_NONE; 151 cq->triggered++; 152 queue_work_on(cq->comp_vector_cpu, comp_vector_wq, 153 &cq->comptask); 154 } 155 156 spin_unlock_irqrestore(&cq->lock, flags); 157 return true; 158} 159EXPORT_SYMBOL(rvt_cq_enter); 160 161static void send_complete(struct work_struct *work) 162{ 163 struct rvt_cq *cq = container_of(work, struct rvt_cq, comptask); 164 165 /* 166 * The completion handler will most likely rearm the notification 167 * and poll for all pending entries. If a new completion entry 168 * is added while we are in this routine, queue_work() 169 * won't call us again until we return so we check triggered to 170 * see if we need to call the handler again. 171 */ 172 for (;;) { 173 u8 triggered = cq->triggered; 174 175 /* 176 * IPoIB connected mode assumes the callback is from a 177 * soft IRQ. We simulate this by blocking "bottom halves". 178 * See the implementation for ipoib_cm_handle_tx_wc(), 179 * netif_tx_lock_bh() and netif_tx_lock(). 180 */ 181 local_bh_disable(); 182 cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context); 183 local_bh_enable(); 184 185 if (cq->triggered == triggered) 186 return; 187 } 188} 189 190/** 191 * rvt_create_cq - create a completion queue 192 * @ibcq: Allocated CQ 193 * @attr: creation attributes 194 * @udata: user data for libibverbs.so 195 * 196 * Called by ib_create_cq() in the generic verbs code. 197 * 198 * Return: 0 on success 199 */ 200int rvt_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, 201 struct ib_udata *udata) 202{ 203 struct ib_device *ibdev = ibcq->device; 204 struct rvt_dev_info *rdi = ib_to_rvt(ibdev); 205 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); 206 struct rvt_cq_wc *u_wc = NULL; 207 struct rvt_k_cq_wc *k_wc = NULL; 208 u32 sz; 209 unsigned int entries = attr->cqe; 210 int comp_vector = attr->comp_vector; 211 int err; 212 213 if (attr->flags) 214 return -EINVAL; 215 216 if (entries < 1 || entries > rdi->dparms.props.max_cqe) 217 return -EINVAL; 218 219 if (comp_vector < 0) 220 comp_vector = 0; 221 222 comp_vector = comp_vector % rdi->ibdev.num_comp_vectors; 223 224 /* 225 * Allocate the completion queue entries and head/tail pointers. 226 * This is allocated separately so that it can be resized and 227 * also mapped into user space. 228 * We need to use vmalloc() in order to support mmap and large 229 * numbers of entries. 230 */ 231 if (udata && udata->outlen >= sizeof(__u64)) { 232 sz = sizeof(struct ib_uverbs_wc) * (entries + 1); 233 sz += sizeof(*u_wc); 234 u_wc = vmalloc_user(sz); 235 if (!u_wc) 236 return -ENOMEM; 237 } else { 238 sz = sizeof(struct ib_wc) * (entries + 1); 239 sz += sizeof(*k_wc); 240 k_wc = vzalloc_node(sz, rdi->dparms.node); 241 if (!k_wc) 242 return -ENOMEM; 243 } 244 245 /* 246 * Return the address of the WC as the offset to mmap. 247 * See rvt_mmap() for details. 248 */ 249 if (udata && udata->outlen >= sizeof(__u64)) { 250 cq->ip = rvt_create_mmap_info(rdi, sz, udata, u_wc); 251 if (IS_ERR(cq->ip)) { 252 err = PTR_ERR(cq->ip); 253 goto bail_wc; 254 } 255 256 err = ib_copy_to_udata(udata, &cq->ip->offset, 257 sizeof(cq->ip->offset)); 258 if (err) 259 goto bail_ip; 260 } 261 262 spin_lock_irq(&rdi->n_cqs_lock); 263 if (rdi->n_cqs_allocated == rdi->dparms.props.max_cq) { 264 spin_unlock_irq(&rdi->n_cqs_lock); 265 err = -ENOMEM; 266 goto bail_ip; 267 } 268 269 rdi->n_cqs_allocated++; 270 spin_unlock_irq(&rdi->n_cqs_lock); 271 272 if (cq->ip) { 273 spin_lock_irq(&rdi->pending_lock); 274 list_add(&cq->ip->pending_mmaps, &rdi->pending_mmaps); 275 spin_unlock_irq(&rdi->pending_lock); 276 } 277 278 /* 279 * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe. 280 * The number of entries should be >= the number requested or return 281 * an error. 282 */ 283 cq->rdi = rdi; 284 if (rdi->driver_f.comp_vect_cpu_lookup) 285 cq->comp_vector_cpu = 286 rdi->driver_f.comp_vect_cpu_lookup(rdi, comp_vector); 287 else 288 cq->comp_vector_cpu = 289 cpumask_first(cpumask_of_node(rdi->dparms.node)); 290 291 cq->ibcq.cqe = entries; 292 cq->notify = RVT_CQ_NONE; 293 spin_lock_init(&cq->lock); 294 INIT_WORK(&cq->comptask, send_complete); 295 if (u_wc) 296 cq->queue = u_wc; 297 else 298 cq->kqueue = k_wc; 299 300 trace_rvt_create_cq(cq, attr); 301 return 0; 302 303bail_ip: 304 kfree(cq->ip); 305bail_wc: 306 vfree(u_wc); 307 vfree(k_wc); 308 return err; 309} 310 311/** 312 * rvt_destroy_cq - destroy a completion queue 313 * @ibcq: the completion queue to destroy. 314 * @udata: user data or NULL for kernel object 315 * 316 * Called by ib_destroy_cq() in the generic verbs code. 317 */ 318int rvt_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata) 319{ 320 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); 321 struct rvt_dev_info *rdi = cq->rdi; 322 323 flush_work(&cq->comptask); 324 spin_lock_irq(&rdi->n_cqs_lock); 325 rdi->n_cqs_allocated--; 326 spin_unlock_irq(&rdi->n_cqs_lock); 327 if (cq->ip) 328 kref_put(&cq->ip->ref, rvt_release_mmap_info); 329 else 330 vfree(cq->kqueue); 331 return 0; 332} 333 334/** 335 * rvt_req_notify_cq - change the notification type for a completion queue 336 * @ibcq: the completion queue 337 * @notify_flags: the type of notification to request 338 * 339 * This may be called from interrupt context. Also called by 340 * ib_req_notify_cq() in the generic verbs code. 341 * 342 * Return: 0 for success. 343 */ 344int rvt_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags) 345{ 346 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); 347 unsigned long flags; 348 int ret = 0; 349 350 spin_lock_irqsave(&cq->lock, flags); 351 /* 352 * Don't change IB_CQ_NEXT_COMP to IB_CQ_SOLICITED but allow 353 * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2). 354 */ 355 if (cq->notify != IB_CQ_NEXT_COMP) 356 cq->notify = notify_flags & IB_CQ_SOLICITED_MASK; 357 358 if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) { 359 if (cq->queue) { 360 if (RDMA_READ_UAPI_ATOMIC(cq->queue->head) != 361 RDMA_READ_UAPI_ATOMIC(cq->queue->tail)) 362 ret = 1; 363 } else { 364 if (cq->kqueue->head != cq->kqueue->tail) 365 ret = 1; 366 } 367 } 368 369 spin_unlock_irqrestore(&cq->lock, flags); 370 371 return ret; 372} 373 374/** 375 * rvt_resize_cq - change the size of the CQ 376 * @ibcq: the completion queue 377 * 378 * Return: 0 for success. 379 */ 380int rvt_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata) 381{ 382 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); 383 u32 head, tail, n; 384 int ret; 385 u32 sz; 386 struct rvt_dev_info *rdi = cq->rdi; 387 struct rvt_cq_wc *u_wc = NULL; 388 struct rvt_cq_wc *old_u_wc = NULL; 389 struct rvt_k_cq_wc *k_wc = NULL; 390 struct rvt_k_cq_wc *old_k_wc = NULL; 391 392 if (cqe < 1 || cqe > rdi->dparms.props.max_cqe) 393 return -EINVAL; 394 395 /* 396 * Need to use vmalloc() if we want to support large #s of entries. 397 */ 398 if (udata && udata->outlen >= sizeof(__u64)) { 399 sz = sizeof(struct ib_uverbs_wc) * (cqe + 1); 400 sz += sizeof(*u_wc); 401 u_wc = vmalloc_user(sz); 402 if (!u_wc) 403 return -ENOMEM; 404 } else { 405 sz = sizeof(struct ib_wc) * (cqe + 1); 406 sz += sizeof(*k_wc); 407 k_wc = vzalloc_node(sz, rdi->dparms.node); 408 if (!k_wc) 409 return -ENOMEM; 410 } 411 /* Check that we can write the offset to mmap. */ 412 if (udata && udata->outlen >= sizeof(__u64)) { 413 __u64 offset = 0; 414 415 ret = ib_copy_to_udata(udata, &offset, sizeof(offset)); 416 if (ret) 417 goto bail_free; 418 } 419 420 spin_lock_irq(&cq->lock); 421 /* 422 * Make sure head and tail are sane since they 423 * might be user writable. 424 */ 425 if (u_wc) { 426 old_u_wc = cq->queue; 427 head = RDMA_READ_UAPI_ATOMIC(old_u_wc->head); 428 tail = RDMA_READ_UAPI_ATOMIC(old_u_wc->tail); 429 } else { 430 old_k_wc = cq->kqueue; 431 head = old_k_wc->head; 432 tail = old_k_wc->tail; 433 } 434 435 if (head > (u32)cq->ibcq.cqe) 436 head = (u32)cq->ibcq.cqe; 437 if (tail > (u32)cq->ibcq.cqe) 438 tail = (u32)cq->ibcq.cqe; 439 if (head < tail) 440 n = cq->ibcq.cqe + 1 + head - tail; 441 else 442 n = head - tail; 443 if (unlikely((u32)cqe < n)) { 444 ret = -EINVAL; 445 goto bail_unlock; 446 } 447 for (n = 0; tail != head; n++) { 448 if (u_wc) 449 u_wc->uqueue[n] = old_u_wc->uqueue[tail]; 450 else 451 k_wc->kqueue[n] = old_k_wc->kqueue[tail]; 452 if (tail == (u32)cq->ibcq.cqe) 453 tail = 0; 454 else 455 tail++; 456 } 457 cq->ibcq.cqe = cqe; 458 if (u_wc) { 459 RDMA_WRITE_UAPI_ATOMIC(u_wc->head, n); 460 RDMA_WRITE_UAPI_ATOMIC(u_wc->tail, 0); 461 cq->queue = u_wc; 462 } else { 463 k_wc->head = n; 464 k_wc->tail = 0; 465 cq->kqueue = k_wc; 466 } 467 spin_unlock_irq(&cq->lock); 468 469 if (u_wc) 470 vfree(old_u_wc); 471 else 472 vfree(old_k_wc); 473 474 if (cq->ip) { 475 struct rvt_mmap_info *ip = cq->ip; 476 477 rvt_update_mmap_info(rdi, ip, sz, u_wc); 478 479 /* 480 * Return the offset to mmap. 481 * See rvt_mmap() for details. 482 */ 483 if (udata && udata->outlen >= sizeof(__u64)) { 484 ret = ib_copy_to_udata(udata, &ip->offset, 485 sizeof(ip->offset)); 486 if (ret) 487 return ret; 488 } 489 490 spin_lock_irq(&rdi->pending_lock); 491 if (list_empty(&ip->pending_mmaps)) 492 list_add(&ip->pending_mmaps, &rdi->pending_mmaps); 493 spin_unlock_irq(&rdi->pending_lock); 494 } 495 496 return 0; 497 498bail_unlock: 499 spin_unlock_irq(&cq->lock); 500bail_free: 501 vfree(u_wc); 502 vfree(k_wc); 503 504 return ret; 505} 506 507/** 508 * rvt_poll_cq - poll for work completion entries 509 * @ibcq: the completion queue to poll 510 * @num_entries: the maximum number of entries to return 511 * @entry: pointer to array where work completions are placed 512 * 513 * This may be called from interrupt context. Also called by ib_poll_cq() 514 * in the generic verbs code. 515 * 516 * Return: the number of completion entries polled. 517 */ 518int rvt_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry) 519{ 520 struct rvt_cq *cq = ibcq_to_rvtcq(ibcq); 521 struct rvt_k_cq_wc *wc; 522 unsigned long flags; 523 int npolled; 524 u32 tail; 525 526 /* The kernel can only poll a kernel completion queue */ 527 if (cq->ip) 528 return -EINVAL; 529 530 spin_lock_irqsave(&cq->lock, flags); 531 532 wc = cq->kqueue; 533 tail = wc->tail; 534 if (tail > (u32)cq->ibcq.cqe) 535 tail = (u32)cq->ibcq.cqe; 536 for (npolled = 0; npolled < num_entries; ++npolled, ++entry) { 537 if (tail == wc->head) 538 break; 539 /* The kernel doesn't need a RMB since it has the lock. */ 540 trace_rvt_cq_poll(cq, &wc->kqueue[tail], npolled); 541 *entry = wc->kqueue[tail]; 542 if (tail >= cq->ibcq.cqe) 543 tail = 0; 544 else 545 tail++; 546 } 547 wc->tail = tail; 548 549 spin_unlock_irqrestore(&cq->lock, flags); 550 551 return npolled; 552} 553 554/** 555 * rvt_driver_cq_init - Init cq resources on behalf of driver 556 * 557 * Return: 0 on success 558 */ 559int rvt_driver_cq_init(void) 560{ 561 comp_vector_wq = alloc_workqueue("%s", WQ_HIGHPRI | WQ_CPU_INTENSIVE, 562 0, "rdmavt_cq"); 563 if (!comp_vector_wq) 564 return -ENOMEM; 565 566 return 0; 567} 568 569/** 570 * rvt_cq_exit - tear down cq reources 571 */ 572void rvt_cq_exit(void) 573{ 574 destroy_workqueue(comp_vector_wq); 575 comp_vector_wq = NULL; 576} 577