1 /* 2 * Copyright (c) 2016 Hisilicon Limited. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 */ 32 33 #include <linux/platform_device.h> 34 #include <rdma/ib_umem.h> 35 #include "hns_roce_device.h" 36 #include "hns_roce_cmd.h" 37 #include "hns_roce_hem.h" 38 #include "hns_roce_user.h" 39 #include "hns_roce_common.h" 40 41 static void hns_roce_ib_cq_comp(struct hns_roce_cq *hr_cq) 42 { 43 struct ib_cq *ibcq = &hr_cq->ib_cq; 44 45 ibcq->comp_handler(ibcq, ibcq->cq_context); 46 } 47 48 static void hns_roce_ib_cq_event(struct hns_roce_cq *hr_cq, 49 enum hns_roce_event event_type) 50 { 51 struct hns_roce_dev *hr_dev; 52 struct ib_event event; 53 struct ib_cq *ibcq; 54 55 ibcq = &hr_cq->ib_cq; 56 hr_dev = to_hr_dev(ibcq->device); 57 58 if (event_type != HNS_ROCE_EVENT_TYPE_CQ_ID_INVALID && 59 event_type != HNS_ROCE_EVENT_TYPE_CQ_ACCESS_ERROR && 60 event_type != HNS_ROCE_EVENT_TYPE_CQ_OVERFLOW) { 61 dev_err(&hr_dev->pdev->dev, 62 "hns_roce_ib: Unexpected event type 0x%x on CQ %06lx\n", 63 event_type, hr_cq->cqn); 64 return; 65 } 66 67 if (ibcq->event_handler) { 68 event.device = ibcq->device; 69 event.event = IB_EVENT_CQ_ERR; 70 event.element.cq = ibcq; 71 ibcq->event_handler(&event, ibcq->cq_context); 72 } 73 } 74 75 static int hns_roce_sw2hw_cq(struct hns_roce_dev *dev, 76 struct hns_roce_cmd_mailbox *mailbox, 77 unsigned long cq_num) 78 { 79 return hns_roce_cmd_mbox(dev, mailbox->dma, 0, cq_num, 0, 80 HNS_ROCE_CMD_SW2HW_CQ, HNS_ROCE_CMD_TIME_CLASS_A); 81 } 82 83 static int hns_roce_cq_alloc(struct hns_roce_dev *hr_dev, int nent, 84 struct hns_roce_mtt *hr_mtt, 85 struct hns_roce_uar *hr_uar, 86 struct hns_roce_cq *hr_cq, int vector) 87 { 88 struct hns_roce_cmd_mailbox *mailbox = NULL; 89 struct hns_roce_cq_table *cq_table = NULL; 90 struct device *dev = &hr_dev->pdev->dev; 91 dma_addr_t dma_handle; 92 u64 *mtts = NULL; 93 int ret = 0; 94 95 cq_table = &hr_dev->cq_table; 96 97 /* Get the physical address of cq buf */ 98 mtts = hns_roce_table_find(&hr_dev->mr_table.mtt_table, 99 hr_mtt->first_seg, &dma_handle); 100 if (!mtts) { 101 dev_err(dev, "CQ alloc.Failed to find cq buf addr.\n"); 102 return -EINVAL; 103 } 104 105 if (vector >= hr_dev->caps.num_comp_vectors) { 106 dev_err(dev, "CQ alloc.Invalid vector.\n"); 107 return -EINVAL; 108 } 109 hr_cq->vector = vector; 110 111 ret = hns_roce_bitmap_alloc(&cq_table->bitmap, &hr_cq->cqn); 112 if (ret == -1) { 113 dev_err(dev, "CQ alloc.Failed to alloc index.\n"); 114 return -ENOMEM; 115 } 116 117 /* Get CQC memory HEM(Hardware Entry Memory) table */ 118 ret = hns_roce_table_get(hr_dev, &cq_table->table, hr_cq->cqn); 119 if (ret) { 120 dev_err(dev, "CQ alloc.Failed to get context mem.\n"); 121 goto err_out; 122 } 123 124 /* The cq insert radix tree */ 125 spin_lock_irq(&cq_table->lock); 126 /* Radix_tree: The associated pointer and long integer key value like */ 127 ret = radix_tree_insert(&cq_table->tree, hr_cq->cqn, hr_cq); 128 spin_unlock_irq(&cq_table->lock); 129 if (ret) { 130 dev_err(dev, "CQ alloc.Failed to radix_tree_insert.\n"); 131 goto err_put; 132 } 133 134 /* Allocate mailbox memory */ 135 mailbox = hns_roce_alloc_cmd_mailbox(hr_dev); 136 if (IS_ERR(mailbox)) { 137 ret = PTR_ERR(mailbox); 138 goto err_radix; 139 } 140 141 hr_dev->hw->write_cqc(hr_dev, hr_cq, mailbox->buf, mtts, dma_handle, 142 nent, vector); 143 144 /* Send mailbox to hw */ 145 ret = hns_roce_sw2hw_cq(hr_dev, mailbox, hr_cq->cqn); 146 hns_roce_free_cmd_mailbox(hr_dev, mailbox); 147 if (ret) { 148 dev_err(dev, "CQ alloc.Failed to cmd mailbox.\n"); 149 goto err_radix; 150 } 151 152 hr_cq->cons_index = 0; 153 hr_cq->uar = hr_uar; 154 155 atomic_set(&hr_cq->refcount, 1); 156 init_completion(&hr_cq->free); 157 158 return 0; 159 160 err_radix: 161 spin_lock_irq(&cq_table->lock); 162 radix_tree_delete(&cq_table->tree, hr_cq->cqn); 163 spin_unlock_irq(&cq_table->lock); 164 165 err_put: 166 hns_roce_table_put(hr_dev, &cq_table->table, hr_cq->cqn); 167 168 err_out: 169 hns_roce_bitmap_free(&cq_table->bitmap, hr_cq->cqn); 170 return ret; 171 } 172 173 static int hns_roce_hw2sw_cq(struct hns_roce_dev *dev, 174 struct hns_roce_cmd_mailbox *mailbox, 175 unsigned long cq_num) 176 { 177 return hns_roce_cmd_mbox(dev, 0, mailbox ? mailbox->dma : 0, cq_num, 178 mailbox ? 0 : 1, HNS_ROCE_CMD_HW2SW_CQ, 179 HNS_ROCE_CMD_TIME_CLASS_A); 180 } 181 182 static void hns_roce_free_cq(struct hns_roce_dev *hr_dev, 183 struct hns_roce_cq *hr_cq) 184 { 185 struct hns_roce_cq_table *cq_table = &hr_dev->cq_table; 186 struct device *dev = &hr_dev->pdev->dev; 187 int ret; 188 189 ret = hns_roce_hw2sw_cq(hr_dev, NULL, hr_cq->cqn); 190 if (ret) 191 dev_err(dev, "HW2SW_CQ failed (%d) for CQN %06lx\n", ret, 192 hr_cq->cqn); 193 194 /* Waiting interrupt process procedure carried out */ 195 synchronize_irq(hr_dev->eq_table.eq[hr_cq->vector].irq); 196 197 /* wait for all interrupt processed */ 198 if (atomic_dec_and_test(&hr_cq->refcount)) 199 complete(&hr_cq->free); 200 wait_for_completion(&hr_cq->free); 201 202 spin_lock_irq(&cq_table->lock); 203 radix_tree_delete(&cq_table->tree, hr_cq->cqn); 204 spin_unlock_irq(&cq_table->lock); 205 206 hns_roce_table_put(hr_dev, &cq_table->table, hr_cq->cqn); 207 hns_roce_bitmap_free(&cq_table->bitmap, hr_cq->cqn); 208 } 209 210 static int hns_roce_ib_get_cq_umem(struct hns_roce_dev *hr_dev, 211 struct ib_ucontext *context, 212 struct hns_roce_cq_buf *buf, 213 struct ib_umem **umem, u64 buf_addr, int cqe) 214 { 215 int ret; 216 217 *umem = ib_umem_get(context, buf_addr, cqe * hr_dev->caps.cq_entry_sz, 218 IB_ACCESS_LOCAL_WRITE, 1); 219 if (IS_ERR(*umem)) 220 return PTR_ERR(*umem); 221 222 ret = hns_roce_mtt_init(hr_dev, ib_umem_page_count(*umem), 223 ilog2((unsigned int)(*umem)->page_size), 224 &buf->hr_mtt); 225 if (ret) 226 goto err_buf; 227 228 ret = hns_roce_ib_umem_write_mtt(hr_dev, &buf->hr_mtt, *umem); 229 if (ret) 230 goto err_mtt; 231 232 return 0; 233 234 err_mtt: 235 hns_roce_mtt_cleanup(hr_dev, &buf->hr_mtt); 236 237 err_buf: 238 ib_umem_release(*umem); 239 return ret; 240 } 241 242 static int hns_roce_ib_alloc_cq_buf(struct hns_roce_dev *hr_dev, 243 struct hns_roce_cq_buf *buf, u32 nent) 244 { 245 int ret; 246 247 ret = hns_roce_buf_alloc(hr_dev, nent * hr_dev->caps.cq_entry_sz, 248 PAGE_SIZE * 2, &buf->hr_buf); 249 if (ret) 250 goto out; 251 252 ret = hns_roce_mtt_init(hr_dev, buf->hr_buf.npages, 253 buf->hr_buf.page_shift, &buf->hr_mtt); 254 if (ret) 255 goto err_buf; 256 257 ret = hns_roce_buf_write_mtt(hr_dev, &buf->hr_mtt, &buf->hr_buf); 258 if (ret) 259 goto err_mtt; 260 261 return 0; 262 263 err_mtt: 264 hns_roce_mtt_cleanup(hr_dev, &buf->hr_mtt); 265 266 err_buf: 267 hns_roce_buf_free(hr_dev, nent * hr_dev->caps.cq_entry_sz, 268 &buf->hr_buf); 269 out: 270 return ret; 271 } 272 273 static void hns_roce_ib_free_cq_buf(struct hns_roce_dev *hr_dev, 274 struct hns_roce_cq_buf *buf, int cqe) 275 { 276 hns_roce_buf_free(hr_dev, (cqe + 1) * hr_dev->caps.cq_entry_sz, 277 &buf->hr_buf); 278 } 279 280 struct ib_cq *hns_roce_ib_create_cq(struct ib_device *ib_dev, 281 const struct ib_cq_init_attr *attr, 282 struct ib_ucontext *context, 283 struct ib_udata *udata) 284 { 285 struct hns_roce_dev *hr_dev = to_hr_dev(ib_dev); 286 struct device *dev = &hr_dev->pdev->dev; 287 struct hns_roce_ib_create_cq ucmd; 288 struct hns_roce_cq *hr_cq = NULL; 289 struct hns_roce_uar *uar = NULL; 290 int vector = attr->comp_vector; 291 int cq_entries = attr->cqe; 292 int ret = 0; 293 294 if (cq_entries < 1 || cq_entries > hr_dev->caps.max_cqes) { 295 dev_err(dev, "Creat CQ failed. entries=%d, max=%d\n", 296 cq_entries, hr_dev->caps.max_cqes); 297 return ERR_PTR(-EINVAL); 298 } 299 300 hr_cq = kmalloc(sizeof(*hr_cq), GFP_KERNEL); 301 if (!hr_cq) 302 return ERR_PTR(-ENOMEM); 303 304 /* In v1 engine, parameter verification */ 305 if (cq_entries < HNS_ROCE_MIN_CQE_NUM) 306 cq_entries = HNS_ROCE_MIN_CQE_NUM; 307 308 cq_entries = roundup_pow_of_two((unsigned int)cq_entries); 309 hr_cq->ib_cq.cqe = cq_entries - 1; 310 spin_lock_init(&hr_cq->lock); 311 312 if (context) { 313 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) { 314 dev_err(dev, "Failed to copy_from_udata.\n"); 315 ret = -EFAULT; 316 goto err_cq; 317 } 318 319 /* Get user space address, write it into mtt table */ 320 ret = hns_roce_ib_get_cq_umem(hr_dev, context, &hr_cq->hr_buf, 321 &hr_cq->umem, ucmd.buf_addr, 322 cq_entries); 323 if (ret) { 324 dev_err(dev, "Failed to get_cq_umem.\n"); 325 goto err_cq; 326 } 327 328 /* Get user space parameters */ 329 uar = &to_hr_ucontext(context)->uar; 330 } else { 331 /* Init mmt table and write buff address to mtt table */ 332 ret = hns_roce_ib_alloc_cq_buf(hr_dev, &hr_cq->hr_buf, 333 cq_entries); 334 if (ret) { 335 dev_err(dev, "Failed to alloc_cq_buf.\n"); 336 goto err_cq; 337 } 338 339 uar = &hr_dev->priv_uar; 340 hr_cq->cq_db_l = hr_dev->reg_base + ROCEE_DB_OTHERS_L_0_REG + 341 0x1000 * uar->index; 342 } 343 344 /* Allocate cq index, fill cq_context */ 345 ret = hns_roce_cq_alloc(hr_dev, cq_entries, &hr_cq->hr_buf.hr_mtt, uar, 346 hr_cq, vector); 347 if (ret) { 348 dev_err(dev, "Creat CQ .Failed to cq_alloc.\n"); 349 goto err_mtt; 350 } 351 352 /* Get created cq handler and carry out event */ 353 hr_cq->comp = hns_roce_ib_cq_comp; 354 hr_cq->event = hns_roce_ib_cq_event; 355 hr_cq->cq_depth = cq_entries; 356 357 if (context) { 358 if (ib_copy_to_udata(udata, &hr_cq->cqn, sizeof(u64))) { 359 ret = -EFAULT; 360 goto err_cqc; 361 } 362 } 363 364 return &hr_cq->ib_cq; 365 366 err_cqc: 367 hns_roce_free_cq(hr_dev, hr_cq); 368 369 err_mtt: 370 hns_roce_mtt_cleanup(hr_dev, &hr_cq->hr_buf.hr_mtt); 371 if (context) 372 ib_umem_release(hr_cq->umem); 373 else 374 hns_roce_ib_free_cq_buf(hr_dev, &hr_cq->hr_buf, 375 hr_cq->ib_cq.cqe); 376 377 err_cq: 378 kfree(hr_cq); 379 return ERR_PTR(ret); 380 } 381 382 int hns_roce_ib_destroy_cq(struct ib_cq *ib_cq) 383 { 384 struct hns_roce_dev *hr_dev = to_hr_dev(ib_cq->device); 385 struct hns_roce_cq *hr_cq = to_hr_cq(ib_cq); 386 387 hns_roce_free_cq(hr_dev, hr_cq); 388 hns_roce_mtt_cleanup(hr_dev, &hr_cq->hr_buf.hr_mtt); 389 390 if (ib_cq->uobject) 391 ib_umem_release(hr_cq->umem); 392 else 393 /* Free the buff of stored cq */ 394 hns_roce_ib_free_cq_buf(hr_dev, &hr_cq->hr_buf, ib_cq->cqe); 395 396 kfree(hr_cq); 397 398 return 0; 399 } 400 401 void hns_roce_cq_completion(struct hns_roce_dev *hr_dev, u32 cqn) 402 { 403 struct device *dev = &hr_dev->pdev->dev; 404 struct hns_roce_cq *cq; 405 406 cq = radix_tree_lookup(&hr_dev->cq_table.tree, 407 cqn & (hr_dev->caps.num_cqs - 1)); 408 if (!cq) { 409 dev_warn(dev, "Completion event for bogus CQ 0x%08x\n", cqn); 410 return; 411 } 412 413 cq->comp(cq); 414 } 415 416 void hns_roce_cq_event(struct hns_roce_dev *hr_dev, u32 cqn, int event_type) 417 { 418 struct hns_roce_cq_table *cq_table = &hr_dev->cq_table; 419 struct device *dev = &hr_dev->pdev->dev; 420 struct hns_roce_cq *cq; 421 422 cq = radix_tree_lookup(&cq_table->tree, 423 cqn & (hr_dev->caps.num_cqs - 1)); 424 if (cq) 425 atomic_inc(&cq->refcount); 426 427 if (!cq) { 428 dev_warn(dev, "Async event for bogus CQ %08x\n", cqn); 429 return; 430 } 431 432 cq->event(cq, (enum hns_roce_event)event_type); 433 434 if (atomic_dec_and_test(&cq->refcount)) 435 complete(&cq->free); 436 } 437 438 int hns_roce_init_cq_table(struct hns_roce_dev *hr_dev) 439 { 440 struct hns_roce_cq_table *cq_table = &hr_dev->cq_table; 441 442 spin_lock_init(&cq_table->lock); 443 INIT_RADIX_TREE(&cq_table->tree, GFP_ATOMIC); 444 445 return hns_roce_bitmap_init(&cq_table->bitmap, hr_dev->caps.num_cqs, 446 hr_dev->caps.num_cqs - 1, 447 hr_dev->caps.reserved_cqs, 0); 448 } 449 450 void hns_roce_cleanup_cq_table(struct hns_roce_dev *hr_dev) 451 { 452 hns_roce_bitmap_cleanup(&hr_dev->cq_table.bitmap); 453 } 454