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_sw2hw_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_SW2HW_SRQ,
  68                                  HNS_ROCE_CMD_TIMEOUT_MSECS);
  69 }
  70
  71 static int hns_roce_hw2sw_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_HW2SW_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,
  99                         "SRQ alloc.Failed to find srq buf addr.\n");
 100                 return -EINVAL;
 101         }
 102
 103         /* Get physical address of idx que buf */
 104         mtts_idx = hns_roce_table_find(hr_dev, &hr_dev->mr_table.mtt_idx_table,
 105                                        srq->idx_que.mtt.first_seg,
 106                                        &dma_handle_idx);
 107         if (!mtts_idx) {
 108                 dev_err(hr_dev->dev,
 109                         "SRQ alloc.Failed to find idx que buf addr.\n");
 110                 return -EINVAL;
 111         }
 112
 113         ret = hns_roce_bitmap_alloc(&srq_table->bitmap, &srq->srqn);
 114         if (ret == -1) {
 115                 dev_err(hr_dev->dev, "SRQ alloc.Failed to alloc index.\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_sw2hw_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_hw2sw_srq(hr_dev, NULL, srq->srqn);
 164         if (ret)
 165                 dev_err(hr_dev->dev, "HW2SW_SRQ failed (%d) for CQN %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 = ib_umem_get(udata, ucmd.buf_addr, srq_buf_size, 0, 0);
 190         if (IS_ERR(srq->umem))
 191                 return PTR_ERR(srq->umem);
 192
 193         buf = &srq->buf;
 194         buf->npages = (ib_umem_page_count(srq->umem) +
 195                        (1 << hr_dev->caps.srqwqe_buf_pg_sz) - 1) /
 196                       (1 << hr_dev->caps.srqwqe_buf_pg_sz);
 197         buf->page_shift = PAGE_SHIFT + hr_dev->caps.srqwqe_buf_pg_sz;
 198         ret = hns_roce_mtt_init(hr_dev, buf->npages, buf->page_shift,
 199                                 &srq->mtt);
 200         if (ret)
 201                 goto err_user_buf;
 202
 203         ret = hns_roce_ib_umem_write_mtt(hr_dev, &srq->mtt, srq->umem);
 204         if (ret)
 205                 goto err_user_srq_mtt;
 206
 207         /* config index queue BA */
 208         srq->idx_que.umem = ib_umem_get(udata, ucmd.que_addr,
 209                                         srq->idx_que.buf_size, 0, 0);
 210         if (IS_ERR(srq->idx_que.umem)) {
 211                 dev_err(hr_dev->dev, "ib_umem_get error for index queue\n");
 212                 ret = PTR_ERR(srq->idx_que.umem);
 213                 goto err_user_srq_mtt;
 214         }
 215
 216         buf = &srq->idx_que.idx_buf;
 217         buf->npages = DIV_ROUND_UP(ib_umem_page_count(srq->idx_que.umem),
 218                                    1 << hr_dev->caps.idx_buf_pg_sz);
 219         buf->page_shift = PAGE_SHIFT + hr_dev->caps.idx_buf_pg_sz;
 220         ret = hns_roce_mtt_init(hr_dev, buf->npages, buf->page_shift,
 221                                 &srq->idx_que.mtt);
 222         if (ret) {
 223                 dev_err(hr_dev->dev, "hns_roce_mtt_init error for idx que\n");
 224                 goto err_user_idx_mtt;
 225         }
 226
 227         ret = hns_roce_ib_umem_write_mtt(hr_dev, &srq->idx_que.mtt,
 228                                          srq->idx_que.umem);
 229         if (ret) {
 230                 dev_err(hr_dev->dev,
 231                         "hns_roce_ib_umem_write_mtt error for idx que\n");
 232                 goto err_user_idx_buf;
 233         }
 234
 235         return 0;
 236
 237 err_user_idx_buf:
 238         hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 239
 240 err_user_idx_mtt:
 241         ib_umem_release(srq->idx_que.umem);
 242
 243 err_user_srq_mtt:
 244         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 245
 246 err_user_buf:
 247         ib_umem_release(srq->umem);
 248
 249         return ret;
 250 }
 251
 252 static int hns_roce_create_idx_que(struct ib_pd *pd, struct hns_roce_srq *srq,
 253                                    u32 page_shift)
 254 {
 255         struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
 256         struct hns_roce_idx_que *idx_que = &srq->idx_que;
 257
 258         idx_que->bitmap = bitmap_zalloc(srq->max, GFP_KERNEL);
 259         if (!idx_que->bitmap)
 260                 return -ENOMEM;
 261
 262         idx_que->buf_size = srq->idx_que.buf_size;
 263
 264         if (hns_roce_buf_alloc(hr_dev, idx_que->buf_size, (1 << page_shift) * 2,
 265                                &idx_que->idx_buf, page_shift)) {
 266                 bitmap_free(idx_que->bitmap);
 267                 return -ENOMEM;
 268         }
 269
 270         return 0;
 271 }
 272
 273 static int create_kernel_srq(struct hns_roce_srq *srq, int srq_buf_size)
 274 {
 275         struct hns_roce_dev *hr_dev = to_hr_dev(srq->ibsrq.device);
 276         u32 page_shift = PAGE_SHIFT + hr_dev->caps.srqwqe_buf_pg_sz;
 277         int ret;
 278
 279         if (hns_roce_buf_alloc(hr_dev, srq_buf_size, (1 << page_shift) * 2,
 280                                &srq->buf, page_shift))
 281                 return -ENOMEM;
 282
 283         srq->head = 0;
 284         srq->tail = srq->max - 1;
 285
 286         ret = hns_roce_mtt_init(hr_dev, srq->buf.npages, srq->buf.page_shift,
 287                                 &srq->mtt);
 288         if (ret)
 289                 goto err_kernel_buf;
 290
 291         ret = hns_roce_buf_write_mtt(hr_dev, &srq->mtt, &srq->buf);
 292         if (ret)
 293                 goto err_kernel_srq_mtt;
 294
 295         page_shift = PAGE_SHIFT + hr_dev->caps.idx_buf_pg_sz;
 296         ret = hns_roce_create_idx_que(srq->ibsrq.pd, srq, page_shift);
 297         if (ret) {
 298                 dev_err(hr_dev->dev, "Create idx queue fail(%d)!\n", ret);
 299                 goto err_kernel_srq_mtt;
 300         }
 301
 302         /* Init mtt table for idx_que */
 303         ret = hns_roce_mtt_init(hr_dev, srq->idx_que.idx_buf.npages,
 304                                 srq->idx_que.idx_buf.page_shift,
 305                                 &srq->idx_que.mtt);
 306         if (ret)
 307                 goto err_kernel_create_idx;
 308
 309         /* Write buffer address into the mtt table */
 310         ret = hns_roce_buf_write_mtt(hr_dev, &srq->idx_que.mtt,
 311                                      &srq->idx_que.idx_buf);
 312         if (ret)
 313                 goto err_kernel_idx_buf;
 314
 315         srq->wrid = kvmalloc_array(srq->max, sizeof(u64), GFP_KERNEL);
 316         if (!srq->wrid) {
 317                 ret = -ENOMEM;
 318                 goto err_kernel_idx_buf;
 319         }
 320
 321         return 0;
 322
 323 err_kernel_idx_buf:
 324         hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 325
 326 err_kernel_create_idx:
 327         hns_roce_buf_free(hr_dev, srq->idx_que.buf_size,
 328                           &srq->idx_que.idx_buf);
 329         kfree(srq->idx_que.bitmap);
 330
 331 err_kernel_srq_mtt:
 332         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 333
 334 err_kernel_buf:
 335         hns_roce_buf_free(hr_dev, srq_buf_size, &srq->buf);
 336
 337         return ret;
 338 }
 339
 340 static void destroy_user_srq(struct hns_roce_dev *hr_dev,
 341                              struct hns_roce_srq *srq)
 342 {
 343         hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 344         ib_umem_release(srq->idx_que.umem);
 345         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 346         ib_umem_release(srq->umem);
 347 }
 348
 349 static void destroy_kernel_srq(struct hns_roce_dev *hr_dev,
 350                                struct hns_roce_srq *srq, int srq_buf_size)
 351 {
 352         kvfree(srq->wrid);
 353         hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 354         hns_roce_buf_free(hr_dev, srq->idx_que.buf_size, &srq->idx_que.idx_buf);
 355         kfree(srq->idx_que.bitmap);
 356         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 357         hns_roce_buf_free(hr_dev, srq_buf_size, &srq->buf);
 358 }
 359
 360 int hns_roce_create_srq(struct ib_srq *ib_srq,
 361                         struct ib_srq_init_attr *srq_init_attr,
 362                         struct ib_udata *udata)
 363 {
 364         struct hns_roce_dev *hr_dev = to_hr_dev(ib_srq->device);
 365         struct hns_roce_ib_create_srq_resp resp = {};
 366         struct hns_roce_srq *srq = to_hr_srq(ib_srq);
 367         int srq_desc_size;
 368         int srq_buf_size;
 369         int ret = 0;
 370         u32 cqn;
 371
 372         /* Check the actual SRQ wqe and SRQ sge num */
 373         if (srq_init_attr->attr.max_wr >= hr_dev->caps.max_srq_wrs ||
 374             srq_init_attr->attr.max_sge > hr_dev->caps.max_srq_sges)
 375                 return -EINVAL;
 376
 377         mutex_init(&srq->mutex);
 378         spin_lock_init(&srq->lock);
 379
 380         srq->max = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
 381         srq->max_gs = srq_init_attr->attr.max_sge;
 382
 383         srq_desc_size = max(16, 16 * srq->max_gs);
 384
 385         srq->wqe_shift = ilog2(srq_desc_size);
 386
 387         srq_buf_size = srq->max * srq_desc_size;
 388
 389         srq->idx_que.entry_sz = HNS_ROCE_IDX_QUE_ENTRY_SZ;
 390         srq->idx_que.buf_size = srq->max * srq->idx_que.entry_sz;
 391         srq->mtt.mtt_type = MTT_TYPE_SRQWQE;
 392         srq->idx_que.mtt.mtt_type = MTT_TYPE_IDX;
 393
 394         if (udata) {
 395                 ret = create_user_srq(srq, udata, srq_buf_size);
 396                 if (ret) {
 397                         dev_err(hr_dev->dev, "Create user srq failed\n");
 398                         goto err_srq;
 399                 }
 400         } else {
 401                 ret = create_kernel_srq(srq, srq_buf_size);
 402                 if (ret) {
 403                         dev_err(hr_dev->dev, "Create kernel srq failed\n");
 404                         goto err_srq;
 405                 }
 406         }
 407
 408         cqn = ib_srq_has_cq(srq_init_attr->srq_type) ?
 409               to_hr_cq(srq_init_attr->ext.cq)->cqn : 0;
 410
 411         srq->db_reg_l = hr_dev->reg_base + SRQ_DB_REG;
 412
 413         ret = hns_roce_srq_alloc(hr_dev, to_hr_pd(ib_srq->pd)->pdn, cqn, 0,
 414                                  &srq->mtt, 0, srq);
 415         if (ret)
 416                 goto err_wrid;
 417
 418         srq->event = hns_roce_ib_srq_event;
 419         resp.srqn = srq->srqn;
 420
 421         if (udata) {
 422                 if (ib_copy_to_udata(udata, &resp,
 423                                      min(udata->outlen, sizeof(resp)))) {
 424                         ret = -EFAULT;
 425                         goto err_srqc_alloc;
 426                 }
 427         }
 428
 429         return 0;
 430
 431 err_srqc_alloc:
 432         hns_roce_srq_free(hr_dev, srq);
 433
 434 err_wrid:
 435         if (udata)
 436                 destroy_user_srq(hr_dev, srq);
 437         else
 438                 destroy_kernel_srq(hr_dev, srq, srq_buf_size);
 439
 440 err_srq:
 441         return ret;
 442 }
 443
 444 void hns_roce_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
 445 {
 446         struct hns_roce_dev *hr_dev = to_hr_dev(ibsrq->device);
 447         struct hns_roce_srq *srq = to_hr_srq(ibsrq);
 448
 449         hns_roce_srq_free(hr_dev, srq);
 450         hns_roce_mtt_cleanup(hr_dev, &srq->mtt);
 451
 452         if (udata) {
 453                 hns_roce_mtt_cleanup(hr_dev, &srq->idx_que.mtt);
 454         } else {
 455                 kvfree(srq->wrid);
 456                 hns_roce_buf_free(hr_dev, srq->max << srq->wqe_shift,
 457                                   &srq->buf);
 458         }
 459         ib_umem_release(srq->idx_que.umem);
 460         ib_umem_release(srq->umem);
 461 }
 462
 463 int hns_roce_init_srq_table(struct hns_roce_dev *hr_dev)
 464 {
 465         struct hns_roce_srq_table *srq_table = &hr_dev->srq_table;
 466
 467         xa_init(&srq_table->xa);
 468
 469         return hns_roce_bitmap_init(&srq_table->bitmap, hr_dev->caps.num_srqs,
 470                                     hr_dev->caps.num_srqs - 1,
 471                                     hr_dev->caps.reserved_srqs, 0);
 472 }
 473
 474 void hns_roce_cleanup_srq_table(struct hns_roce_dev *hr_dev)
 475 {
 476         hns_roce_bitmap_cleanup(&hr_dev->srq_table.bitmap);
 477 }