net/xdp/xsk_buff_pool.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <net/xsk_buff_pool.h>
   4 #include <net/xdp_sock.h>
   5
   6 #include "xsk_queue.h"
   7
   8 static void xp_addr_unmap(struct xsk_buff_pool *pool)
   9 {
  10         vunmap(pool->addrs);
  11 }
  12
  13 static int xp_addr_map(struct xsk_buff_pool *pool,
  14                        struct page **pages, u32 nr_pages)
  15 {
  16         pool->addrs = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
  17         if (!pool->addrs)
  18                 return -ENOMEM;
  19         return 0;
  20 }
  21
  22 void xp_destroy(struct xsk_buff_pool *pool)
  23 {
  24         if (!pool)
  25                 return;
  26
  27         xp_addr_unmap(pool);
  28         kvfree(pool->heads);
  29         kvfree(pool);
  30 }
  31
  32 struct xsk_buff_pool *xp_create(struct page **pages, u32 nr_pages, u32 chunks,
  33                                 u32 chunk_size, u32 headroom, u64 size,
  34                                 bool unaligned)
  35 {
  36         struct xsk_buff_pool *pool;
  37         struct xdp_buff_xsk *xskb;
  38         int err;
  39         u32 i;
  40
  41         pool = kvzalloc(struct_size(pool, free_heads, chunks), GFP_KERNEL);
  42         if (!pool)
  43                 goto out;
  44
  45         pool->heads = kvcalloc(chunks, sizeof(*pool->heads), GFP_KERNEL);
  46         if (!pool->heads)
  47                 goto out;
  48
  49         pool->chunk_mask = ~((u64)chunk_size - 1);
  50         pool->addrs_cnt = size;
  51         pool->heads_cnt = chunks;
  52         pool->free_heads_cnt = chunks;
  53         pool->headroom = headroom;
  54         pool->chunk_size = chunk_size;
  55         pool->unaligned = unaligned;
  56         pool->frame_len = chunk_size - headroom - XDP_PACKET_HEADROOM;
  57         INIT_LIST_HEAD(&pool->free_list);
  58
  59         for (i = 0; i < pool->free_heads_cnt; i++) {
  60                 xskb = &pool->heads[i];
  61                 xskb->pool = pool;
  62                 xskb->xdp.frame_sz = chunk_size - headroom;
  63                 pool->free_heads[i] = xskb;
  64         }
  65
  66         err = xp_addr_map(pool, pages, nr_pages);
  67         if (!err)
  68                 return pool;
  69
  70 out:
  71         xp_destroy(pool);
  72         return NULL;
  73 }
  74
  75 void xp_set_fq(struct xsk_buff_pool *pool, struct xsk_queue *fq)
  76 {
  77         pool->fq = fq;
  78 }
  79
  80 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
  81 {
  82         u32 i;
  83
  84         for (i = 0; i < pool->heads_cnt; i++)
  85                 pool->heads[i].xdp.rxq = rxq;
  86 }
  87 EXPORT_SYMBOL(xp_set_rxq_info);
  88
  89 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
  90 {
  91         dma_addr_t *dma;
  92         u32 i;
  93
  94         if (pool->dma_pages_cnt == 0)
  95                 return;
  96
  97         for (i = 0; i < pool->dma_pages_cnt; i++) {
  98                 dma = &pool->dma_pages[i];
  99                 if (*dma) {
 100                         dma_unmap_page_attrs(pool->dev, *dma, PAGE_SIZE,
 101                                              DMA_BIDIRECTIONAL, attrs);
 102                         *dma = 0;
 103                 }
 104         }
 105
 106         kvfree(pool->dma_pages);
 107         pool->dma_pages_cnt = 0;
 108         pool->dev = NULL;
 109 }
 110 EXPORT_SYMBOL(xp_dma_unmap);
 111
 112 static void xp_check_dma_contiguity(struct xsk_buff_pool *pool)
 113 {
 114         u32 i;
 115
 116         for (i = 0; i < pool->dma_pages_cnt - 1; i++) {
 117                 if (pool->dma_pages[i] + PAGE_SIZE == pool->dma_pages[i + 1])
 118                         pool->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
 119                 else
 120                         pool->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
 121         }
 122 }
 123
 124 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
 125                unsigned long attrs, struct page **pages, u32 nr_pages)
 126 {
 127         dma_addr_t dma;
 128         u32 i;
 129
 130         pool->dma_pages = kvcalloc(nr_pages, sizeof(*pool->dma_pages),
 131                                    GFP_KERNEL);
 132         if (!pool->dma_pages)
 133                 return -ENOMEM;
 134
 135         pool->dev = dev;
 136         pool->dma_pages_cnt = nr_pages;
 137         pool->dma_need_sync = false;
 138
 139         for (i = 0; i < pool->dma_pages_cnt; i++) {
 140                 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
 141                                          DMA_BIDIRECTIONAL, attrs);
 142                 if (dma_mapping_error(dev, dma)) {
 143                         xp_dma_unmap(pool, attrs);
 144                         return -ENOMEM;
 145                 }
 146                 if (dma_need_sync(dev, dma))
 147                         pool->dma_need_sync = true;
 148                 pool->dma_pages[i] = dma;
 149         }
 150
 151         if (pool->unaligned)
 152                 xp_check_dma_contiguity(pool);
 153         return 0;
 154 }
 155 EXPORT_SYMBOL(xp_dma_map);
 156
 157 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
 158                                           u64 addr)
 159 {
 160         return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
 161 }
 162
 163 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
 164 {
 165         *addr = xp_unaligned_extract_addr(*addr);
 166         if (*addr >= pool->addrs_cnt ||
 167             *addr + pool->chunk_size > pool->addrs_cnt ||
 168             xp_addr_crosses_non_contig_pg(pool, *addr))
 169                 return false;
 170         return true;
 171 }
 172
 173 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
 174 {
 175         *addr = xp_aligned_extract_addr(pool, *addr);
 176         return *addr < pool->addrs_cnt;
 177 }
 178
 179 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
 180 {
 181         struct xdp_buff_xsk *xskb;
 182         u64 addr;
 183         bool ok;
 184
 185         if (pool->free_heads_cnt == 0)
 186                 return NULL;
 187
 188         xskb = pool->free_heads[--pool->free_heads_cnt];
 189
 190         for (;;) {
 191                 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
 192                         pool->fq->queue_empty_descs++;
 193                         xp_release(xskb);
 194                         return NULL;
 195                 }
 196
 197                 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
 198                      xp_check_aligned(pool, &addr);
 199                 if (!ok) {
 200                         pool->fq->invalid_descs++;
 201                         xskq_cons_release(pool->fq);
 202                         continue;
 203                 }
 204                 break;
 205         }
 206         xskq_cons_release(pool->fq);
 207
 208         xskb->orig_addr = addr;
 209         xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
 210         if (pool->dma_pages_cnt) {
 211                 xskb->frame_dma = (pool->dma_pages[addr >> PAGE_SHIFT] &
 212                                    ~XSK_NEXT_PG_CONTIG_MASK) +
 213                                   (addr & ~PAGE_MASK);
 214                 xskb->dma = xskb->frame_dma + pool->headroom +
 215                             XDP_PACKET_HEADROOM;
 216         }
 217         return xskb;
 218 }
 219
 220 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
 221 {
 222         struct xdp_buff_xsk *xskb;
 223
 224         if (!pool->free_list_cnt) {
 225                 xskb = __xp_alloc(pool);
 226                 if (!xskb)
 227                         return NULL;
 228         } else {
 229                 pool->free_list_cnt--;
 230                 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
 231                                         free_list_node);
 232                 list_del(&xskb->free_list_node);
 233         }
 234
 235         xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
 236         xskb->xdp.data_meta = xskb->xdp.data;
 237
 238         if (pool->dma_need_sync) {
 239                 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
 240                                                  pool->frame_len,
 241                                                  DMA_BIDIRECTIONAL);
 242         }
 243         return &xskb->xdp;
 244 }
 245 EXPORT_SYMBOL(xp_alloc);
 246
 247 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
 248 {
 249         if (pool->free_list_cnt >= count)
 250                 return true;
 251         return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
 252 }
 253 EXPORT_SYMBOL(xp_can_alloc);
 254
 255 void xp_free(struct xdp_buff_xsk *xskb)
 256 {
 257         xskb->pool->free_list_cnt++;
 258         list_add(&xskb->free_list_node, &xskb->pool->free_list);
 259 }
 260 EXPORT_SYMBOL(xp_free);
 261
 262 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
 263 {
 264         addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 265         return pool->addrs + addr;
 266 }
 267 EXPORT_SYMBOL(xp_raw_get_data);
 268
 269 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
 270 {
 271         addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 272         return (pool->dma_pages[addr >> PAGE_SHIFT] &
 273                 ~XSK_NEXT_PG_CONTIG_MASK) +
 274                 (addr & ~PAGE_MASK);
 275 }
 276 EXPORT_SYMBOL(xp_raw_get_dma);
 277
 278 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
 279 {
 280         dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
 281                                       xskb->pool->frame_len, DMA_BIDIRECTIONAL);
 282 }
 283 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
 284
 285 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
 286                                  size_t size)
 287 {
 288         dma_sync_single_range_for_device(pool->dev, dma, 0,
 289                                          size, DMA_BIDIRECTIONAL);
 290 }
 291 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);