drivers/infiniband/hw/hns/hns_roce_alloc.c

   1 /*
   2  * Copyright (c) 2016 Hisilicon Limited.
   3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/platform_device.h>
  35 #include <linux/vmalloc.h>
  36 #include "hns_roce_device.h"
  37 #include <rdma/ib_umem.h>
  38
  39 int hns_roce_bitmap_alloc(struct hns_roce_bitmap *bitmap, unsigned long *obj)
  40 {
  41         int ret = 0;
  42
  43         spin_lock(&bitmap->lock);
  44         *obj = find_next_zero_bit(bitmap->table, bitmap->max, bitmap->last);
  45         if (*obj >= bitmap->max) {
  46                 bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
  47                                & bitmap->mask;
  48                 *obj = find_first_zero_bit(bitmap->table, bitmap->max);
  49         }
  50
  51         if (*obj < bitmap->max) {
  52                 set_bit(*obj, bitmap->table);
  53                 bitmap->last = (*obj + 1);
  54                 if (bitmap->last == bitmap->max)
  55                         bitmap->last = 0;
  56                 *obj |= bitmap->top;
  57         } else {
  58                 ret = -1;
  59         }
  60
  61         spin_unlock(&bitmap->lock);
  62
  63         return ret;
  64 }
  65
  66 void hns_roce_bitmap_free(struct hns_roce_bitmap *bitmap, unsigned long obj,
  67                           int rr)
  68 {
  69         hns_roce_bitmap_free_range(bitmap, obj, 1, rr);
  70 }
  71
  72 int hns_roce_bitmap_alloc_range(struct hns_roce_bitmap *bitmap, int cnt,
  73                                 int align, unsigned long *obj)
  74 {
  75         int ret = 0;
  76         int i;
  77
  78         if (likely(cnt == 1 && align == 1))
  79                 return hns_roce_bitmap_alloc(bitmap, obj);
  80
  81         spin_lock(&bitmap->lock);
  82
  83         *obj = bitmap_find_next_zero_area(bitmap->table, bitmap->max,
  84                                           bitmap->last, cnt, align - 1);
  85         if (*obj >= bitmap->max) {
  86                 bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
  87                                & bitmap->mask;
  88                 *obj = bitmap_find_next_zero_area(bitmap->table, bitmap->max, 0,
  89                                                   cnt, align - 1);
  90         }
  91
  92         if (*obj < bitmap->max) {
  93                 for (i = 0; i < cnt; i++)
  94                         set_bit(*obj + i, bitmap->table);
  95
  96                 if (*obj == bitmap->last) {
  97                         bitmap->last = (*obj + cnt);
  98                         if (bitmap->last >= bitmap->max)
  99                                 bitmap->last = 0;
 100                 }
 101                 *obj |= bitmap->top;
 102         } else {
 103                 ret = -1;
 104         }
 105
 106         spin_unlock(&bitmap->lock);
 107
 108         return ret;
 109 }
 110
 111 void hns_roce_bitmap_free_range(struct hns_roce_bitmap *bitmap,
 112                                 unsigned long obj, int cnt,
 113                                 int rr)
 114 {
 115         int i;
 116
 117         obj &= bitmap->max + bitmap->reserved_top - 1;
 118
 119         spin_lock(&bitmap->lock);
 120         for (i = 0; i < cnt; i++)
 121                 clear_bit(obj + i, bitmap->table);
 122
 123         if (!rr)
 124                 bitmap->last = min(bitmap->last, obj);
 125         bitmap->top = (bitmap->top + bitmap->max + bitmap->reserved_top)
 126                        & bitmap->mask;
 127         spin_unlock(&bitmap->lock);
 128 }
 129
 130 int hns_roce_bitmap_init(struct hns_roce_bitmap *bitmap, u32 num, u32 mask,
 131                          u32 reserved_bot, u32 reserved_top)
 132 {
 133         u32 i;
 134
 135         if (num != roundup_pow_of_two(num))
 136                 return -EINVAL;
 137
 138         bitmap->last = 0;
 139         bitmap->top = 0;
 140         bitmap->max = num - reserved_top;
 141         bitmap->mask = mask;
 142         bitmap->reserved_top = reserved_top;
 143         spin_lock_init(&bitmap->lock);
 144         bitmap->table = kcalloc(BITS_TO_LONGS(bitmap->max), sizeof(long),
 145                                 GFP_KERNEL);
 146         if (!bitmap->table)
 147                 return -ENOMEM;
 148
 149         for (i = 0; i < reserved_bot; ++i)
 150                 set_bit(i, bitmap->table);
 151
 152         return 0;
 153 }
 154
 155 void hns_roce_bitmap_cleanup(struct hns_roce_bitmap *bitmap)
 156 {
 157         kfree(bitmap->table);
 158 }
 159
 160 void hns_roce_buf_free(struct hns_roce_dev *hr_dev, u32 size,
 161                        struct hns_roce_buf *buf)
 162 {
 163         int i;
 164         struct device *dev = hr_dev->dev;
 165
 166         if (buf->nbufs == 1) {
 167                 dma_free_coherent(dev, size, buf->direct.buf, buf->direct.map);
 168         } else {
 169                 for (i = 0; i < buf->nbufs; ++i)
 170                         if (buf->page_list[i].buf)
 171                                 dma_free_coherent(dev, 1 << buf->page_shift,
 172                                                   buf->page_list[i].buf,
 173                                                   buf->page_list[i].map);
 174                 kfree(buf->page_list);
 175         }
 176 }
 177
 178 int hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size, u32 max_direct,
 179                        struct hns_roce_buf *buf, u32 page_shift)
 180 {
 181         int i = 0;
 182         dma_addr_t t;
 183         struct device *dev = hr_dev->dev;
 184         u32 page_size = 1 << page_shift;
 185         u32 order;
 186
 187         /* SQ/RQ buf lease than one page, SQ + RQ = 8K */
 188         if (size <= max_direct) {
 189                 buf->nbufs = 1;
 190                 /* Npages calculated by page_size */
 191                 order = get_order(size);
 192                 if (order <= page_shift - PAGE_SHIFT)
 193                         order = 0;
 194                 else
 195                         order -= page_shift - PAGE_SHIFT;
 196                 buf->npages = 1 << order;
 197                 buf->page_shift = page_shift;
 198                 /* MTT PA must be recorded in 4k alignment, t is 4k aligned */
 199                 buf->direct.buf = dma_alloc_coherent(dev, size, &t,
 200                                                      GFP_KERNEL);
 201                 if (!buf->direct.buf)
 202                         return -ENOMEM;
 203
 204                 buf->direct.map = t;
 205
 206                 while (t & ((1 << buf->page_shift) - 1)) {
 207                         --buf->page_shift;
 208                         buf->npages *= 2;
 209                 }
 210         } else {
 211                 buf->nbufs = (size + page_size - 1) / page_size;
 212                 buf->npages = buf->nbufs;
 213                 buf->page_shift = page_shift;
 214                 buf->page_list = kcalloc(buf->nbufs, sizeof(*buf->page_list),
 215                                          GFP_KERNEL);
 216
 217                 if (!buf->page_list)
 218                         return -ENOMEM;
 219
 220                 for (i = 0; i < buf->nbufs; ++i) {
 221                         buf->page_list[i].buf = dma_alloc_coherent(dev,
 222                                                                    page_size,
 223                                                                    &t,
 224                                                                    GFP_KERNEL);
 225
 226                         if (!buf->page_list[i].buf)
 227                                 goto err_free;
 228
 229                         buf->page_list[i].map = t;
 230                 }
 231         }
 232
 233         return 0;
 234
 235 err_free:
 236         hns_roce_buf_free(hr_dev, size, buf);
 237         return -ENOMEM;
 238 }
 239
 240 int hns_roce_get_kmem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
 241                            int buf_cnt, int start, struct hns_roce_buf *buf)
 242 {
 243         int i, end;
 244         int total;
 245
 246         end = start + buf_cnt;
 247         if (end > buf->npages) {
 248                 dev_err(hr_dev->dev,
 249                         "invalid kmem region,offset %d,buf_cnt %d,total %d!\n",
 250                         start, buf_cnt, buf->npages);
 251                 return -EINVAL;
 252         }
 253
 254         total = 0;
 255         for (i = start; i < end; i++)
 256                 if (buf->nbufs == 1)
 257                         bufs[total++] = buf->direct.map +
 258                                         ((dma_addr_t)i << buf->page_shift);
 259                 else
 260                         bufs[total++] = buf->page_list[i].map;
 261
 262         return total;
 263 }
 264
 265 int hns_roce_get_umem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
 266                            int buf_cnt, int start, struct ib_umem *umem,
 267                            int page_shift)
 268 {
 269         struct ib_block_iter biter;
 270         int total = 0;
 271         int idx = 0;
 272         u64 addr;
 273
 274         if (page_shift < PAGE_SHIFT) {
 275                 dev_err(hr_dev->dev, "invalid page shift %d!\n", page_shift);
 276                 return -EINVAL;
 277         }
 278
 279         /* convert system page cnt to hw page cnt */
 280         rdma_for_each_block(umem->sg_head.sgl, &biter, umem->nmap,
 281                             1 << page_shift) {
 282                 addr = rdma_block_iter_dma_address(&biter);
 283                 if (idx >= start) {
 284                         bufs[total++] = addr;
 285                         if (total >= buf_cnt)
 286                                 goto done;
 287                 }
 288                 idx++;
 289         }
 290
 291 done:
 292         return total;
 293 }
 294
 295 void hns_roce_init_buf_region(struct hns_roce_buf_region *region, int hopnum,
 296                               int offset, int buf_cnt)
 297 {
 298         if (hopnum == HNS_ROCE_HOP_NUM_0)
 299                 region->hopnum = 0;
 300         else
 301                 region->hopnum = hopnum;
 302
 303         region->offset = offset;
 304         region->count = buf_cnt;
 305 }
 306
 307 void hns_roce_free_buf_list(dma_addr_t **bufs, int region_cnt)
 308 {
 309         int i;
 310
 311         for (i = 0; i < region_cnt; i++) {
 312                 kfree(bufs[i]);
 313                 bufs[i] = NULL;
 314         }
 315 }
 316
 317 int hns_roce_alloc_buf_list(struct hns_roce_buf_region *regions,
 318                             dma_addr_t **bufs, int region_cnt)
 319 {
 320         struct hns_roce_buf_region *r;
 321         int i;
 322
 323         for (i = 0; i < region_cnt; i++) {
 324                 r = &regions[i];
 325                 bufs[i] = kcalloc(r->count, sizeof(dma_addr_t), GFP_KERNEL);
 326                 if (!bufs[i])
 327                         goto err_alloc;
 328         }
 329
 330         return 0;
 331
 332 err_alloc:
 333         hns_roce_free_buf_list(bufs, i);
 334
 335         return -ENOMEM;
 336 }
 337
 338 void hns_roce_cleanup_bitmap(struct hns_roce_dev *hr_dev)
 339 {
 340         if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_SRQ)
 341                 hns_roce_cleanup_srq_table(hr_dev);
 342         hns_roce_cleanup_qp_table(hr_dev);
 343         hns_roce_cleanup_cq_table(hr_dev);
 344         hns_roce_cleanup_mr_table(hr_dev);
 345         hns_roce_cleanup_pd_table(hr_dev);
 346         hns_roce_cleanup_uar_table(hr_dev);
 347 }