drivers/infiniband/hw/hns/hns_roce_srq.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2018 Hisilicon Limited.
   4  */
   5
   6 #include <rdma/ib_umem.h>
   7 #include <rdma/hns-abi.h>
   8 #include "hns_roce_device.h"
   9 #include "hns_roce_cmd.h"
  10 #include "hns_roce_hem.h"
  11
  12 void hns_roce_srq_event(struct hns_roce_dev *hr_dev, u32 srqn, int event_type)
  13 {
  14         struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
  15         struct hns_roce_srq *srq;
  16
  17         xa_lock(&srq_table->xa);
  18         srq = xa_load(&srq_table->xa, srqn & (hr_dev->caps.num_srqs - 1));
  19         if (srq)
  20                 atomic_inc(&srq->refcount);
  21         xa_unlock(&srq_table->xa);
  22
  23         if (!srq) {
  24                 dev_warn(hr_dev->dev, "Async event for bogus SRQ %08x\n", srqn);
  25                 return;
  26         }
  27
  28         srq->event(srq, event_type);
  29
  30         if (atomic_dec_and_test(&srq->refcount))
  31                 complete(&srq->free);
  32 }
  33
  34 static void hns_roce_ib_srq_event(struct hns_roce_srq *srq,
  35                                   enum hns_roce_event event_type)
  36 {
  37         struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
  38         struct ib_srq *ibsrq = &srq->ibsrq;
  39         struct ib_event event;
  40
  41         if (ibsrq->event_handler) {
  42                 event.device      = ibsrq->device;
  43                 event.element.srq = ibsrq;
  44                 switch (event_type) {
  45                 case HNS_ROCE_EVENT_TYPE_SRQ_LIMIT_REACH:
  46                         event.event = IB_EVENT_SRQ_LIMIT_REACHED;
  47                         break;
  48                 case HNS_ROCE_EVENT_TYPE_SRQ_CATAS_ERROR:
  49                         event.event = IB_EVENT_SRQ_ERR;
  50                         break;
  51                 default:
  52                         dev_err(hr_dev->dev,
  53                            "hns_roce:Unexpected event type 0x%x on SRQ %06lx\n",
  54                            event_type, srq->srqn);
  55                         return;
  56                 }
  57
  58                 ibsrq->event_handler(&event, ibsrq->srq_context);
  59         }
  60 }
  61
  62 static int hns_roce_hw_create_srq(struct hns_roce_dev *dev,
  63                                   struct hns_roce_cmd_mailbox *mailbox,
  64                                   unsigned long srq_num)
  65 {
  66         return hns_roce_cmd_mbox(dev, mailbox->dma, 0, srq_num, 0,
  67                                  HNS_ROCE_CMD_CREATE_SRQ,
  68                                  HNS_ROCE_CMD_TIMEOUT_MSECS);
  69 }
  70
  71 static int hns_roce_hw_destroy_srq(struct hns_roce_dev *dev,
  72                                    struct hns_roce_cmd_mailbox *mailbox,
  73                                    unsigned long srq_num)
  74 {
  75         return hns_roce_cmd_mbox(dev, 0, mailbox ? mailbox->dma : 0, srq_num,
  76                                  mailbox ? 0 : 1, HNS_ROCE_CMD_DESTROY_SRQ,
  77                                  HNS_ROCE_CMD_TIMEOUT_MSECS);
  78 }
  79
  80 static int hns_roce_srq_alloc(struct hns_roce_dev *hr_dev, u32 pdn, u32 cqn,
  81                               u16 xrcd, struct hns_roce_mtt *hr_mtt,
  82                               u64 db_rec_addr, struct hns_roce_srq *srq)
  83 {
  84         struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
  85         struct hns_roce_cmd_mailbox *mailbox;
  86         dma_addr_t dma_handle_wqe;
  87         dma_addr_t dma_handle_idx;
  88         u64 *mtts_wqe;
  89         u64 *mtts_idx;
  90         int ret;
  91
  92         /* Get the physical address of srq buf */
  93         mtts_wqe = hns_roce_table_find(hr_dev,
  94                                        &hr_dev->mr_table.mtt_srqwqe_table,
  95                                        srq->mtt.first_seg,
  96                                        &dma_handle_wqe);
  97         if (!mtts_wqe) {
  98                 dev_err(hr_dev->dev, "Failed to find mtt for srq buf.\n");
  99                 return -EINVAL;
 100         }
 101
 102         /* Get physical address of idx que buf */
 103         mtts_idx = hns_roce_table_find(hr_dev, &hr_dev->mr_table.mtt_idx_table,
 104                                        srq->idx_que.mtt.first_seg,
 105                                        &dma_handle_idx);
 106         if (!mtts_idx) {
 107                 dev_err(hr_dev->dev,
 108                         "Failed to find mtt for srq idx queue buf.\n");
 109                 return -EINVAL;
 110         }
 111
 112         ret = hns_roce_bitmap_alloc(&srq_table->bitmap, &srq->srqn);
 113         if (ret) {
 114                 dev_err(hr_dev->dev,
 115                         "Failed to alloc a bit from srq bitmap.\n");
 116                 return -ENOMEM;
 117         }
 118
 119         ret = hns_roce_table_get(hr_dev, &srq_table->table, srq->srqn);
 120         if (ret)
 121                 goto err_out;
 122
 123         ret = xa_err(xa_store(&srq_table->xa, srq->srqn, srq, GFP_KERNEL));
 124         if (ret)
 125                 goto err_put;
 126
 127         mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
 128         if (IS_ERR(mailbox)) {
 129                 ret = PTR_ERR(mailbox);
 130                 goto err_xa;
 131         }
 132
 133         hr_dev->hw->write_srqc(hr_dev, srq, pdn, xrcd, cqn, mailbox->buf,
 134                                mtts_wqe, mtts_idx, dma_handle_wqe,
 135                                dma_handle_idx);
 136
 137         ret = hns_roce_hw_create_srq(hr_dev, mailbox, srq->srqn);
 138         hns_roce_free_cmd_mailbox(hr_dev, mailbox);
 139         if (ret)
 140                 goto err_xa;
 141
 142         atomic_set(&srq->refcount, 1);
 143         init_completion(&srq->free);
 144         return ret;
 145
 146 err_xa:
 147         xa_erase(&srq_table->xa, srq->srqn);
 148
 149 err_put:
 150         hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
 151
 152 err_out:
 153         hns_roce_bitmap_free(&srq_table->bitmap, srq->srqn, BITMAP_NO_RR);
 154         return ret;
 155 }
 156
 157 static void hns_roce_srq_free(struct hns_roce_dev *hr_dev,
 158                               struct hns_roce_srq *srq)
 159 {
 160         struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
 161         int ret;
 162
 163         ret = hns_roce_hw_destroy_srq(hr_dev, NULL, srq->srqn);
 164         if (ret)
 165                 dev_err(hr_dev->dev, "DESTROY_SRQ failed (%d) for SRQN %06lx\n",
 166                         ret, srq->srqn);
 167
 168         xa_erase(&srq_table->xa, srq->srqn);
 169
 170         if (atomic_dec_and_test(&srq->refcount))
 171                 complete(&srq->free);
 172         wait_for_completion(&srq->free);
 173
 174         hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn);
 175         hns_roce_bitmap_free(&srq_table->bitmap, srq->srqn, BITMAP_NO_RR);
 176 }
 177
 178 static int create_user_srq(struct hns_roce_srq *srq, struct ib_udata *udata,
 179                            int srq_buf_size)
 180 {
 181         struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
 182         struct hns_roce_ib_create_srq  ucmd;
 183         struct hns_roce_buf *buf;
 184         int ret;
 185
 186         if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
 187                 return -EFAULT;
 188
 189         srq->umem =
 190                 ib_umem_get(srq->ibsrq.device, ucmd.buf_addr, srq_buf_size, 0);
 191         if (IS_ERR(srq->umem))
 192                 return PTR_ERR(srq->umem);
 193
 194         buf = &srq->buf;
 195         buf->npages = (ib_umem_page_count(srq->umem) +
 196                        (1 << hr_dev->caps.srqwqe_buf_pg_sz) - 1) /
 197                       (1 << hr_dev->caps.srqwqe_buf_pg_sz);
 198         buf->page_shift = PAGE_SHIFT + hr_dev->caps.srqwqe_buf_pg_sz;
 199         ret = hns_roce_mtt_init(hr_dev, buf->npages, buf->page_shift,
 200                                 &srq->mtt);
 201         if (ret)
 202                 goto err_user_buf;
 203
 204         ret = hns_roce_ib_umem_write_mtt(hr_dev, &srq->mtt, srq->umem);
 205         if (ret)
 206                 goto err_user_srq_mtt;
 207
 208         /* config index queue BA */
 209         srq->idx_que.umem = ib_umem_get(srq->ibsrq.device, ucmd.que_addr,
 210                                         srq->idx_que.buf_size, 0);
 211         if (IS_ERR(srq->idx_que.umem)) {
 212                 dev_err(hr_dev->dev, "ib_umem_get error for index queue\n");
 213                 ret = PTR_ERR(srq->idx_que.umem);
 214                 goto err_user_srq_mtt;
 215         }
 216
 217         buf = &srq->idx_que.idx_buf;
 218         buf->npages = DIV_ROUND_UP(ib_umem_page_count(srq->idx_que.umem),
 219                                    1 << hr_dev->caps.idx_buf_pg_sz);
 220         buf->page_shift = PAGE_SHIFT + hr_dev->caps.idx_buf_pg_sz;
 221         ret = hns_roce_mtt_init(hr_dev, buf->npages, buf->page_shift,
 222                                 &srq->idx_que.mtt);
 223         if (ret) {
 224                 dev_err(hr_dev->dev, "hns_roce_mtt_init error for idx que\n");
 225                 goto err_user_idx_mtt;
 226         }
 227
 228         ret = hns_roce_ib_umem_write_mtt(hr_dev, &srq->idx_que.mtt,
 229                                          srq->idx_que.umem);
 230         if (ret) {
 231                 dev_err(hr_dev->dev,
 232                         "hns_roce_ib_umem_write_mtt error for idx que\n");
 233                 goto err_user_idx_buf;
 234         }
 235
 236         return 0;
 237
 238 err_user_idx_buf:
 239         hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 240
 241 err_user_idx_mtt:
 242         ib_umem_release(srq->idx_que.umem);
 243
 244 err_user_srq_mtt:
 245         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 246
 247 err_user_buf:
 248         ib_umem_release(srq->umem);
 249
 250         return ret;
 251 }
 252
 253 static int hns_roce_create_idx_que(struct ib_pd *pd, struct hns_roce_srq *srq,
 254                                    u32 page_shift)
 255 {
 256         struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
 257         struct hns_roce_idx_que *idx_que = &srq->idx_que;
 258
 259         idx_que->bitmap = bitmap_zalloc(srq->wqe_cnt, GFP_KERNEL);
 260         if (!idx_que->bitmap)
 261                 return -ENOMEM;
 262
 263         idx_que->buf_size = srq->idx_que.buf_size;
 264
 265         if (hns_roce_buf_alloc(hr_dev, idx_que->buf_size, (1 << page_shift) * 2,
 266                                &idx_que->idx_buf, page_shift)) {
 267                 bitmap_free(idx_que->bitmap);
 268                 return -ENOMEM;
 269         }
 270
 271         return 0;
 272 }
 273
 274 static int create_kernel_srq(struct hns_roce_srq *srq, int srq_buf_size)
 275 {
 276         struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
 277         u32 page_shift = PAGE_SHIFT + hr_dev->caps.srqwqe_buf_pg_sz;
 278         int ret;
 279
 280         if (hns_roce_buf_alloc(hr_dev, srq_buf_size, (1 << page_shift) * 2,
 281                                &srq->buf, page_shift))
 282                 return -ENOMEM;
 283
 284         srq->head = 0;
 285         srq->tail = srq->wqe_cnt - 1;
 286
 287         ret = hns_roce_mtt_init(hr_dev, srq->buf.npages, srq->buf.page_shift,
 288                                 &srq->mtt);
 289         if (ret)
 290                 goto err_kernel_buf;
 291
 292         ret = hns_roce_buf_write_mtt(hr_dev, &srq->mtt, &srq->buf);
 293         if (ret)
 294                 goto err_kernel_srq_mtt;
 295
 296         page_shift = PAGE_SHIFT + hr_dev->caps.idx_buf_pg_sz;
 297         ret = hns_roce_create_idx_que(srq->ibsrq.pd, srq, page_shift);
 298         if (ret) {
 299                 dev_err(hr_dev->dev, "Create idx queue fail(%d)!\n", ret);
 300                 goto err_kernel_srq_mtt;
 301         }
 302
 303         /* Init mtt table for idx_que */
 304         ret = hns_roce_mtt_init(hr_dev, srq->idx_que.idx_buf.npages,
 305                                 srq->idx_que.idx_buf.page_shift,
 306                                 &srq->idx_que.mtt);
 307         if (ret)
 308                 goto err_kernel_create_idx;
 309
 310         /* Write buffer address into the mtt table */
 311         ret = hns_roce_buf_write_mtt(hr_dev, &srq->idx_que.mtt,
 312                                      &srq->idx_que.idx_buf);
 313         if (ret)
 314                 goto err_kernel_idx_buf;
 315
 316         srq->wrid = kvmalloc_array(srq->wqe_cnt, sizeof(u64), GFP_KERNEL);
 317         if (!srq->wrid) {
 318                 ret = -ENOMEM;
 319                 goto err_kernel_idx_buf;
 320         }
 321
 322         return 0;
 323
 324 err_kernel_idx_buf:
 325         hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 326
 327 err_kernel_create_idx:
 328         hns_roce_buf_free(hr_dev, srq->idx_que.buf_size,
 329                           &srq->idx_que.idx_buf);
 330         kfree(srq->idx_que.bitmap);
 331
 332 err_kernel_srq_mtt:
 333         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 334
 335 err_kernel_buf:
 336         hns_roce_buf_free(hr_dev, srq_buf_size, &srq->buf);
 337
 338         return ret;
 339 }
 340
 341 static void destroy_user_srq(struct hns_roce_dev *hr_dev,
 342                              struct hns_roce_srq *srq)
 343 {
 344         hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 345         ib_umem_release(srq->idx_que.umem);
 346         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 347         ib_umem_release(srq->umem);
 348 }
 349
 350 static void destroy_kernel_srq(struct hns_roce_dev *hr_dev,
 351                                struct hns_roce_srq *srq, int srq_buf_size)
 352 {
 353         kvfree(srq->wrid);
 354         hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 355         hns_roce_buf_free(hr_dev, srq->idx_que.buf_size, &srq->idx_que.idx_buf);
 356         kfree(srq->idx_que.bitmap);
 357         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 358         hns_roce_buf_free(hr_dev, srq_buf_size, &srq->buf);
 359 }
 360
 361 int hns_roce_create_srq(struct ib_srq *ib_srq,
 362                         struct ib_srq_init_attr *init_attr,
 363                         struct ib_udata *udata)
 364 {
 365         struct hns_roce_dev *hr_dev = to_hr_dev(ib_srq->device);
 366         struct hns_roce_ib_create_srq_resp resp = {};
 367         struct hns_roce_srq *srq = to_hr_srq(ib_srq);
 368         int srq_desc_size;
 369         int srq_buf_size;
 370         int ret = 0;
 371         u32 cqn;
 372
 373         /* Check the actual SRQ wqe and SRQ sge num */
 374         if (init_attr->attr.max_wr >= hr_dev->caps.max_srq_wrs ||
 375             init_attr->attr.max_sge > hr_dev->caps.max_srq_sges)
 376                 return -EINVAL;
 377
 378         mutex_init(&srq->mutex);
 379         spin_lock_init(&srq->lock);
 380
 381         srq->wqe_cnt = roundup_pow_of_two(init_attr->attr.max_wr + 1);
 382         srq->max_gs = init_attr->attr.max_sge;
 383
 384         srq_desc_size = roundup_pow_of_two(max(16, 16 * srq->max_gs));
 385
 386         srq->wqe_shift = ilog2(srq_desc_size);
 387
 388         srq_buf_size = srq->wqe_cnt * srq_desc_size;
 389
 390         srq->idx_que.entry_sz = HNS_ROCE_IDX_QUE_ENTRY_SZ;
 391         srq->idx_que.buf_size = srq->wqe_cnt * srq->idx_que.entry_sz;
 392         srq->mtt.mtt_type = MTT_TYPE_SRQWQE;
 393         srq->idx_que.mtt.mtt_type = MTT_TYPE_IDX;
 394
 395         if (udata) {
 396                 ret = create_user_srq(srq, udata, srq_buf_size);
 397                 if (ret) {
 398                         dev_err(hr_dev->dev, "Create user srq failed\n");
 399                         goto err_srq;
 400                 }
 401         } else {
 402                 ret = create_kernel_srq(srq, srq_buf_size);
 403                 if (ret) {
 404                         dev_err(hr_dev->dev, "Create kernel srq failed\n");
 405                         goto err_srq;
 406                 }
 407         }
 408
 409         cqn = ib_srq_has_cq(init_attr->srq_type) ?
 410               to_hr_cq(init_attr->ext.cq)->cqn : 0;
 411
 412         srq->db_reg_l = hr_dev->reg_base + SRQ_DB_REG;
 413
 414         ret = hns_roce_srq_alloc(hr_dev, to_hr_pd(ib_srq->pd)->pdn, cqn, 0,
 415                                  &srq->mtt, 0, srq);
 416         if (ret)
 417                 goto err_wrid;
 418
 419         srq->event = hns_roce_ib_srq_event;
 420         resp.srqn = srq->srqn;
 421
 422         if (udata) {
 423                 if (ib_copy_to_udata(udata, &resp,
 424                                      min(udata->outlen, sizeof(resp)))) {
 425                         ret = -EFAULT;
 426                         goto err_srqc_alloc;
 427                 }
 428         }
 429
 430         return 0;
 431
 432 err_srqc_alloc:
 433         hns_roce_srq_free(hr_dev, srq);
 434
 435 err_wrid:
 436         if (udata)
 437                 destroy_user_srq(hr_dev, srq);
 438         else
 439                 destroy_kernel_srq(hr_dev, srq, srq_buf_size);
 440
 441 err_srq:
 442         return ret;
 443 }
 444
 445 void hns_roce_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
 446 {
 447         struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
 448         struct hns_roce_srq *srq = to_hr_srq(ibsrq);
 449
 450         hns_roce_srq_free(hr_dev, srq);
 451         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 452
 453         if (udata) {
 454                 hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 455         } else {
 456                 kvfree(srq->wrid);
 457                 hns_roce_buf_free(hr_dev, srq->wqe_cnt << srq->wqe_shift,
 458                                   &srq->buf);
 459         }
 460         ib_umem_release(srq->idx_que.umem);
 461         ib_umem_release(srq->umem);
 462 }
 463
 464 int hns_roce_init_srq_table(struct hns_roce_dev *hr_dev)
 465 {
 466         struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
 467
 468         xa_init(&srq_table->xa);
 469
 470         return hns_roce_bitmap_init(&srq_table->bitmap, hr_dev->caps.num_srqs,
 471                                     hr_dev->caps.num_srqs - 1,
 472                                     hr_dev->caps.reserved_srqs, 0);
 473 }
 474
 475 void hns_roce_cleanup_srq_table(struct hns_roce_dev *hr_dev)
 476 {
 477         hns_roce_bitmap_cleanup(&hr_dev->srq_table.bitmap);
 478 }