1 /* 2 * Copyright (c) 2012-2016 VMware, Inc. All rights reserved. 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of EITHER the GNU General Public License 6 * version 2 as published by the Free Software Foundation or the BSD 7 * 2-Clause License. This program is distributed in the hope that it 8 * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED 9 * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. 10 * See the GNU General Public License version 2 for more details at 11 * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html. 12 * 13 * You should have received a copy of the GNU General Public License 14 * along with this program available in the file COPYING in the main 15 * directory of this source tree. 16 * 17 * The BSD 2-Clause License 18 * 19 * Redistribution and use in source and binary forms, with or 20 * without modification, are permitted provided that the following 21 * conditions are met: 22 * 23 * - Redistributions of source code must retain the above 24 * copyright notice, this list of conditions and the following 25 * disclaimer. 26 * 27 * - Redistributions in binary form must reproduce the above 28 * copyright notice, this list of conditions and the following 29 * disclaimer in the documentation and/or other materials 30 * provided with the distribution. 31 * 32 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 33 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 34 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 35 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 36 * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, 37 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 38 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 39 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, 41 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 42 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED 43 * OF THE POSSIBILITY OF SUCH DAMAGE. 44 */ 45 46 #include <asm/page.h> 47 #include <linux/io.h> 48 #include <linux/wait.h> 49 #include <rdma/ib_addr.h> 50 #include <rdma/ib_smi.h> 51 #include <rdma/ib_user_verbs.h> 52 #include <rdma/uverbs_ioctl.h> 53 54 #include "pvrdma.h" 55 56 /** 57 * pvrdma_req_notify_cq - request notification for a completion queue 58 * @ibcq: the completion queue 59 * @notify_flags: notification flags 60 * 61 * @return: 0 for success. 62 */ 63 int pvrdma_req_notify_cq(struct ib_cq *ibcq, 64 enum ib_cq_notify_flags notify_flags) 65 { 66 struct pvrdma_dev *dev = to_vdev(ibcq->device); 67 struct pvrdma_cq *cq = to_vcq(ibcq); 68 u32 val = cq->cq_handle; 69 unsigned long flags; 70 int has_data = 0; 71 72 val |= (notify_flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ? 73 PVRDMA_UAR_CQ_ARM_SOL : PVRDMA_UAR_CQ_ARM; 74 75 spin_lock_irqsave(&cq->cq_lock, flags); 76 77 pvrdma_write_uar_cq(dev, val); 78 79 if (notify_flags & IB_CQ_REPORT_MISSED_EVENTS) { 80 unsigned int head; 81 82 has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx, 83 cq->ibcq.cqe, &head); 84 if (unlikely(has_data == PVRDMA_INVALID_IDX)) 85 dev_err(&dev->pdev->dev, "CQ ring state invalid\n"); 86 } 87 88 spin_unlock_irqrestore(&cq->cq_lock, flags); 89 90 return has_data; 91 } 92 93 /** 94 * pvrdma_create_cq - create completion queue 95 * @ibcq: Allocated CQ 96 * @attr: completion queue attributes 97 * @udata: user data 98 * 99 * @return: 0 on success 100 */ 101 int pvrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, 102 struct ib_udata *udata) 103 { 104 struct ib_device *ibdev = ibcq->device; 105 int entries = attr->cqe; 106 struct pvrdma_dev *dev = to_vdev(ibdev); 107 struct pvrdma_cq *cq = to_vcq(ibcq); 108 int ret; 109 int npages; 110 unsigned long flags; 111 union pvrdma_cmd_req req; 112 union pvrdma_cmd_resp rsp; 113 struct pvrdma_cmd_create_cq *cmd = &req.create_cq; 114 struct pvrdma_cmd_create_cq_resp *resp = &rsp.create_cq_resp; 115 struct pvrdma_create_cq_resp cq_resp = {}; 116 struct pvrdma_create_cq ucmd; 117 struct pvrdma_ucontext *context = rdma_udata_to_drv_context( 118 udata, struct pvrdma_ucontext, ibucontext); 119 120 BUILD_BUG_ON(sizeof(struct pvrdma_cqe) != 64); 121 122 entries = roundup_pow_of_two(entries); 123 if (entries < 1 || entries > dev->dsr->caps.max_cqe) 124 return -EINVAL; 125 126 if (!atomic_add_unless(&dev->num_cqs, 1, dev->dsr->caps.max_cq)) 127 return -ENOMEM; 128 129 cq->ibcq.cqe = entries; 130 cq->is_kernel = !udata; 131 132 if (!cq->is_kernel) { 133 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) { 134 ret = -EFAULT; 135 goto err_cq; 136 } 137 138 cq->umem = ib_umem_get(udata, ucmd.buf_addr, ucmd.buf_size, 139 IB_ACCESS_LOCAL_WRITE, 1); 140 if (IS_ERR(cq->umem)) { 141 ret = PTR_ERR(cq->umem); 142 goto err_cq; 143 } 144 145 npages = ib_umem_page_count(cq->umem); 146 } else { 147 /* One extra page for shared ring state */ 148 npages = 1 + (entries * sizeof(struct pvrdma_cqe) + 149 PAGE_SIZE - 1) / PAGE_SIZE; 150 151 /* Skip header page. */ 152 cq->offset = PAGE_SIZE; 153 } 154 155 if (npages < 0 || npages > PVRDMA_PAGE_DIR_MAX_PAGES) { 156 dev_warn(&dev->pdev->dev, 157 "overflow pages in completion queue\n"); 158 ret = -EINVAL; 159 goto err_umem; 160 } 161 162 ret = pvrdma_page_dir_init(dev, &cq->pdir, npages, cq->is_kernel); 163 if (ret) { 164 dev_warn(&dev->pdev->dev, 165 "could not allocate page directory\n"); 166 goto err_umem; 167 } 168 169 /* Ring state is always the first page. Set in library for user cq. */ 170 if (cq->is_kernel) 171 cq->ring_state = cq->pdir.pages[0]; 172 else 173 pvrdma_page_dir_insert_umem(&cq->pdir, cq->umem, 0); 174 175 refcount_set(&cq->refcnt, 1); 176 init_completion(&cq->free); 177 spin_lock_init(&cq->cq_lock); 178 179 memset(cmd, 0, sizeof(*cmd)); 180 cmd->hdr.cmd = PVRDMA_CMD_CREATE_CQ; 181 cmd->nchunks = npages; 182 cmd->ctx_handle = context ? context->ctx_handle : 0; 183 cmd->cqe = entries; 184 cmd->pdir_dma = cq->pdir.dir_dma; 185 ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_CQ_RESP); 186 if (ret < 0) { 187 dev_warn(&dev->pdev->dev, 188 "could not create completion queue, error: %d\n", ret); 189 goto err_page_dir; 190 } 191 192 cq->ibcq.cqe = resp->cqe; 193 cq->cq_handle = resp->cq_handle; 194 cq_resp.cqn = resp->cq_handle; 195 spin_lock_irqsave(&dev->cq_tbl_lock, flags); 196 dev->cq_tbl[cq->cq_handle % dev->dsr->caps.max_cq] = cq; 197 spin_unlock_irqrestore(&dev->cq_tbl_lock, flags); 198 199 if (!cq->is_kernel) { 200 cq->uar = &context->uar; 201 202 /* Copy udata back. */ 203 if (ib_copy_to_udata(udata, &cq_resp, sizeof(cq_resp))) { 204 dev_warn(&dev->pdev->dev, 205 "failed to copy back udata\n"); 206 pvrdma_destroy_cq(&cq->ibcq, udata); 207 return -EINVAL; 208 } 209 } 210 211 return 0; 212 213 err_page_dir: 214 pvrdma_page_dir_cleanup(dev, &cq->pdir); 215 err_umem: 216 ib_umem_release(cq->umem); 217 err_cq: 218 atomic_dec(&dev->num_cqs); 219 return ret; 220 } 221 222 static void pvrdma_free_cq(struct pvrdma_dev *dev, struct pvrdma_cq *cq) 223 { 224 if (refcount_dec_and_test(&cq->refcnt)) 225 complete(&cq->free); 226 wait_for_completion(&cq->free); 227 228 ib_umem_release(cq->umem); 229 230 pvrdma_page_dir_cleanup(dev, &cq->pdir); 231 } 232 233 /** 234 * pvrdma_destroy_cq - destroy completion queue 235 * @cq: the completion queue to destroy. 236 * @udata: user data or null for kernel object 237 */ 238 void pvrdma_destroy_cq(struct ib_cq *cq, struct ib_udata *udata) 239 { 240 struct pvrdma_cq *vcq = to_vcq(cq); 241 union pvrdma_cmd_req req; 242 struct pvrdma_cmd_destroy_cq *cmd = &req.destroy_cq; 243 struct pvrdma_dev *dev = to_vdev(cq->device); 244 unsigned long flags; 245 int ret; 246 247 memset(cmd, 0, sizeof(*cmd)); 248 cmd->hdr.cmd = PVRDMA_CMD_DESTROY_CQ; 249 cmd->cq_handle = vcq->cq_handle; 250 251 ret = pvrdma_cmd_post(dev, &req, NULL, 0); 252 if (ret < 0) 253 dev_warn(&dev->pdev->dev, 254 "could not destroy completion queue, error: %d\n", 255 ret); 256 257 /* free cq's resources */ 258 spin_lock_irqsave(&dev->cq_tbl_lock, flags); 259 dev->cq_tbl[vcq->cq_handle] = NULL; 260 spin_unlock_irqrestore(&dev->cq_tbl_lock, flags); 261 262 pvrdma_free_cq(dev, vcq); 263 atomic_dec(&dev->num_cqs); 264 } 265 266 static inline struct pvrdma_cqe *get_cqe(struct pvrdma_cq *cq, int i) 267 { 268 return (struct pvrdma_cqe *)pvrdma_page_dir_get_ptr( 269 &cq->pdir, 270 cq->offset + 271 sizeof(struct pvrdma_cqe) * i); 272 } 273 274 void _pvrdma_flush_cqe(struct pvrdma_qp *qp, struct pvrdma_cq *cq) 275 { 276 unsigned int head; 277 int has_data; 278 279 if (!cq->is_kernel) 280 return; 281 282 /* Lock held */ 283 has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx, 284 cq->ibcq.cqe, &head); 285 if (unlikely(has_data > 0)) { 286 int items; 287 int curr; 288 int tail = pvrdma_idx(&cq->ring_state->rx.prod_tail, 289 cq->ibcq.cqe); 290 struct pvrdma_cqe *cqe; 291 struct pvrdma_cqe *curr_cqe; 292 293 items = (tail > head) ? (tail - head) : 294 (cq->ibcq.cqe - head + tail); 295 curr = --tail; 296 while (items-- > 0) { 297 if (curr < 0) 298 curr = cq->ibcq.cqe - 1; 299 if (tail < 0) 300 tail = cq->ibcq.cqe - 1; 301 curr_cqe = get_cqe(cq, curr); 302 if ((curr_cqe->qp & 0xFFFF) != qp->qp_handle) { 303 if (curr != tail) { 304 cqe = get_cqe(cq, tail); 305 *cqe = *curr_cqe; 306 } 307 tail--; 308 } else { 309 pvrdma_idx_ring_inc( 310 &cq->ring_state->rx.cons_head, 311 cq->ibcq.cqe); 312 } 313 curr--; 314 } 315 } 316 } 317 318 static int pvrdma_poll_one(struct pvrdma_cq *cq, struct pvrdma_qp **cur_qp, 319 struct ib_wc *wc) 320 { 321 struct pvrdma_dev *dev = to_vdev(cq->ibcq.device); 322 int has_data; 323 unsigned int head; 324 bool tried = false; 325 struct pvrdma_cqe *cqe; 326 327 retry: 328 has_data = pvrdma_idx_ring_has_data(&cq->ring_state->rx, 329 cq->ibcq.cqe, &head); 330 if (has_data == 0) { 331 if (tried) 332 return -EAGAIN; 333 334 pvrdma_write_uar_cq(dev, cq->cq_handle | PVRDMA_UAR_CQ_POLL); 335 336 tried = true; 337 goto retry; 338 } else if (has_data == PVRDMA_INVALID_IDX) { 339 dev_err(&dev->pdev->dev, "CQ ring state invalid\n"); 340 return -EAGAIN; 341 } 342 343 cqe = get_cqe(cq, head); 344 345 /* Ensure cqe is valid. */ 346 rmb(); 347 if (dev->qp_tbl[cqe->qp & 0xffff]) 348 *cur_qp = (struct pvrdma_qp *)dev->qp_tbl[cqe->qp & 0xffff]; 349 else 350 return -EAGAIN; 351 352 wc->opcode = pvrdma_wc_opcode_to_ib(cqe->opcode); 353 wc->status = pvrdma_wc_status_to_ib(cqe->status); 354 wc->wr_id = cqe->wr_id; 355 wc->qp = &(*cur_qp)->ibqp; 356 wc->byte_len = cqe->byte_len; 357 wc->ex.imm_data = cqe->imm_data; 358 wc->src_qp = cqe->src_qp; 359 wc->wc_flags = pvrdma_wc_flags_to_ib(cqe->wc_flags); 360 wc->pkey_index = cqe->pkey_index; 361 wc->slid = cqe->slid; 362 wc->sl = cqe->sl; 363 wc->dlid_path_bits = cqe->dlid_path_bits; 364 wc->port_num = cqe->port_num; 365 wc->vendor_err = cqe->vendor_err; 366 wc->network_hdr_type = cqe->network_hdr_type; 367 368 /* Update shared ring state */ 369 pvrdma_idx_ring_inc(&cq->ring_state->rx.cons_head, cq->ibcq.cqe); 370 371 return 0; 372 } 373 374 /** 375 * pvrdma_poll_cq - poll for work completion queue entries 376 * @ibcq: completion queue 377 * @num_entries: the maximum number of entries 378 * @entry: pointer to work completion array 379 * 380 * @return: number of polled completion entries 381 */ 382 int pvrdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc) 383 { 384 struct pvrdma_cq *cq = to_vcq(ibcq); 385 struct pvrdma_qp *cur_qp = NULL; 386 unsigned long flags; 387 int npolled; 388 389 if (num_entries < 1 || wc == NULL) 390 return 0; 391 392 spin_lock_irqsave(&cq->cq_lock, flags); 393 for (npolled = 0; npolled < num_entries; ++npolled) { 394 if (pvrdma_poll_one(cq, &cur_qp, wc + npolled)) 395 break; 396 } 397 398 spin_unlock_irqrestore(&cq->cq_lock, flags); 399 400 /* Ensure we do not return errors from poll_cq */ 401 return npolled; 402 } 403