1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Devmem TCP 4 * 5 * Authors: Mina Almasry <almasrymina@google.com> 6 * Willem de Bruijn <willemdebruijn.kernel@gmail.com> 7 * Kaiyuan Zhang <kaiyuanz@google.com 8 */ 9 10 #include <linux/dma-buf.h> 11 #include <linux/genalloc.h> 12 #include <linux/mm.h> 13 #include <linux/netdevice.h> 14 #include <linux/types.h> 15 #include <net/netdev_queues.h> 16 #include <net/netdev_rx_queue.h> 17 #include <net/page_pool/helpers.h> 18 #include <net/page_pool/memory_provider.h> 19 #include <net/sock.h> 20 #include <trace/events/page_pool.h> 21 22 #include "devmem.h" 23 #include "mp_dmabuf_devmem.h" 24 #include "page_pool_priv.h" 25 26 /* Device memory support */ 27 28 static DEFINE_XARRAY_FLAGS(net_devmem_dmabuf_bindings, XA_FLAGS_ALLOC1); 29 30 static const struct memory_provider_ops dmabuf_devmem_ops; 31 32 bool net_is_devmem_iov(struct net_iov *niov) 33 { 34 return niov->type == NET_IOV_DMABUF; 35 } 36 37 static void net_devmem_dmabuf_free_chunk_owner(struct gen_pool *genpool, 38 struct gen_pool_chunk *chunk, 39 void *not_used) 40 { 41 struct dmabuf_genpool_chunk_owner *owner = chunk->owner; 42 43 kvfree(owner->area.niovs); 44 kfree(owner); 45 } 46 47 static dma_addr_t net_devmem_get_dma_addr(const struct net_iov *niov) 48 { 49 struct dmabuf_genpool_chunk_owner *owner; 50 51 owner = net_devmem_iov_to_chunk_owner(niov); 52 return owner->base_dma_addr + 53 ((dma_addr_t)net_iov_idx(niov) << PAGE_SHIFT); 54 } 55 56 void __net_devmem_dmabuf_binding_free(struct work_struct *wq) 57 { 58 struct net_devmem_dmabuf_binding *binding = container_of(wq, typeof(*binding), unbind_w); 59 60 size_t size, avail; 61 62 gen_pool_for_each_chunk(binding->chunk_pool, 63 net_devmem_dmabuf_free_chunk_owner, NULL); 64 65 size = gen_pool_size(binding->chunk_pool); 66 avail = gen_pool_avail(binding->chunk_pool); 67 68 if (!WARN(size != avail, "can't destroy genpool. size=%zu, avail=%zu", 69 size, avail)) 70 gen_pool_destroy(binding->chunk_pool); 71 72 dma_buf_unmap_attachment_unlocked(binding->attachment, binding->sgt, 73 DMA_FROM_DEVICE); 74 dma_buf_detach(binding->dmabuf, binding->attachment); 75 dma_buf_put(binding->dmabuf); 76 xa_destroy(&binding->bound_rxqs); 77 kvfree(binding->tx_vec); 78 kfree(binding); 79 } 80 EXPORT_SYMBOL(__net_devmem_dmabuf_binding_free); 81 82 struct net_iov * 83 net_devmem_alloc_dmabuf(struct net_devmem_dmabuf_binding *binding) 84 { 85 struct dmabuf_genpool_chunk_owner *owner; 86 unsigned long dma_addr; 87 struct net_iov *niov; 88 ssize_t offset; 89 ssize_t index; 90 91 dma_addr = gen_pool_alloc_owner(binding->chunk_pool, PAGE_SIZE, 92 (void **)&owner); 93 if (!dma_addr) 94 return NULL; 95 96 offset = dma_addr - owner->base_dma_addr; 97 index = offset / PAGE_SIZE; 98 niov = &owner->area.niovs[index]; 99 100 niov->pp_magic = 0; 101 niov->pp = NULL; 102 atomic_long_set(&niov->pp_ref_count, 0); 103 104 return niov; 105 } 106 107 void net_devmem_free_dmabuf(struct net_iov *niov) 108 { 109 struct net_devmem_dmabuf_binding *binding = net_devmem_iov_binding(niov); 110 unsigned long dma_addr = net_devmem_get_dma_addr(niov); 111 112 if (WARN_ON(!gen_pool_has_addr(binding->chunk_pool, dma_addr, 113 PAGE_SIZE))) 114 return; 115 116 gen_pool_free(binding->chunk_pool, dma_addr, PAGE_SIZE); 117 } 118 119 void net_devmem_unbind_dmabuf(struct net_devmem_dmabuf_binding *binding) 120 { 121 struct netdev_rx_queue *rxq; 122 unsigned long xa_idx; 123 unsigned int rxq_idx; 124 125 xa_erase(&net_devmem_dmabuf_bindings, binding->id); 126 127 /* Ensure no tx net_devmem_lookup_dmabuf() are in flight after the 128 * erase. 129 */ 130 synchronize_net(); 131 132 if (binding->list.next) 133 list_del(&binding->list); 134 135 xa_for_each(&binding->bound_rxqs, xa_idx, rxq) { 136 const struct pp_memory_provider_params mp_params = { 137 .mp_priv = binding, 138 .mp_ops = &dmabuf_devmem_ops, 139 }; 140 141 rxq_idx = get_netdev_rx_queue_index(rxq); 142 143 __net_mp_close_rxq(binding->dev, rxq_idx, &mp_params); 144 } 145 146 net_devmem_dmabuf_binding_put(binding); 147 } 148 149 int net_devmem_bind_dmabuf_to_queue(struct net_device *dev, u32 rxq_idx, 150 struct net_devmem_dmabuf_binding *binding, 151 struct netlink_ext_ack *extack) 152 { 153 struct pp_memory_provider_params mp_params = { 154 .mp_priv = binding, 155 .mp_ops = &dmabuf_devmem_ops, 156 }; 157 struct netdev_rx_queue *rxq; 158 u32 xa_idx; 159 int err; 160 161 err = __net_mp_open_rxq(dev, rxq_idx, &mp_params, extack); 162 if (err) 163 return err; 164 165 rxq = __netif_get_rx_queue(dev, rxq_idx); 166 err = xa_alloc(&binding->bound_rxqs, &xa_idx, rxq, xa_limit_32b, 167 GFP_KERNEL); 168 if (err) 169 goto err_close_rxq; 170 171 return 0; 172 173 err_close_rxq: 174 __net_mp_close_rxq(dev, rxq_idx, &mp_params); 175 return err; 176 } 177 178 struct net_devmem_dmabuf_binding * 179 net_devmem_bind_dmabuf(struct net_device *dev, 180 enum dma_data_direction direction, 181 unsigned int dmabuf_fd, struct netlink_ext_ack *extack) 182 { 183 struct net_devmem_dmabuf_binding *binding; 184 static u32 id_alloc_next; 185 struct scatterlist *sg; 186 struct dma_buf *dmabuf; 187 unsigned int sg_idx, i; 188 unsigned long virtual; 189 int err; 190 191 dmabuf = dma_buf_get(dmabuf_fd); 192 if (IS_ERR(dmabuf)) 193 return ERR_CAST(dmabuf); 194 195 binding = kzalloc_node(sizeof(*binding), GFP_KERNEL, 196 dev_to_node(&dev->dev)); 197 if (!binding) { 198 err = -ENOMEM; 199 goto err_put_dmabuf; 200 } 201 202 binding->dev = dev; 203 xa_init_flags(&binding->bound_rxqs, XA_FLAGS_ALLOC); 204 205 refcount_set(&binding->ref, 1); 206 207 mutex_init(&binding->lock); 208 209 binding->dmabuf = dmabuf; 210 211 binding->attachment = dma_buf_attach(binding->dmabuf, dev->dev.parent); 212 if (IS_ERR(binding->attachment)) { 213 err = PTR_ERR(binding->attachment); 214 NL_SET_ERR_MSG(extack, "Failed to bind dmabuf to device"); 215 goto err_free_binding; 216 } 217 218 binding->sgt = dma_buf_map_attachment_unlocked(binding->attachment, 219 direction); 220 if (IS_ERR(binding->sgt)) { 221 err = PTR_ERR(binding->sgt); 222 NL_SET_ERR_MSG(extack, "Failed to map dmabuf attachment"); 223 goto err_detach; 224 } 225 226 if (direction == DMA_TO_DEVICE) { 227 binding->tx_vec = kvmalloc_array(dmabuf->size / PAGE_SIZE, 228 sizeof(struct net_iov *), 229 GFP_KERNEL); 230 if (!binding->tx_vec) { 231 err = -ENOMEM; 232 goto err_unmap; 233 } 234 } 235 236 /* For simplicity we expect to make PAGE_SIZE allocations, but the 237 * binding can be much more flexible than that. We may be able to 238 * allocate MTU sized chunks here. Leave that for future work... 239 */ 240 binding->chunk_pool = gen_pool_create(PAGE_SHIFT, 241 dev_to_node(&dev->dev)); 242 if (!binding->chunk_pool) { 243 err = -ENOMEM; 244 goto err_tx_vec; 245 } 246 247 virtual = 0; 248 for_each_sgtable_dma_sg(binding->sgt, sg, sg_idx) { 249 dma_addr_t dma_addr = sg_dma_address(sg); 250 struct dmabuf_genpool_chunk_owner *owner; 251 size_t len = sg_dma_len(sg); 252 struct net_iov *niov; 253 254 owner = kzalloc_node(sizeof(*owner), GFP_KERNEL, 255 dev_to_node(&dev->dev)); 256 if (!owner) { 257 err = -ENOMEM; 258 goto err_free_chunks; 259 } 260 261 owner->area.base_virtual = virtual; 262 owner->base_dma_addr = dma_addr; 263 owner->area.num_niovs = len / PAGE_SIZE; 264 owner->binding = binding; 265 266 err = gen_pool_add_owner(binding->chunk_pool, dma_addr, 267 dma_addr, len, dev_to_node(&dev->dev), 268 owner); 269 if (err) { 270 kfree(owner); 271 err = -EINVAL; 272 goto err_free_chunks; 273 } 274 275 owner->area.niovs = kvmalloc_array(owner->area.num_niovs, 276 sizeof(*owner->area.niovs), 277 GFP_KERNEL); 278 if (!owner->area.niovs) { 279 err = -ENOMEM; 280 goto err_free_chunks; 281 } 282 283 for (i = 0; i < owner->area.num_niovs; i++) { 284 niov = &owner->area.niovs[i]; 285 niov->type = NET_IOV_DMABUF; 286 niov->owner = &owner->area; 287 page_pool_set_dma_addr_netmem(net_iov_to_netmem(niov), 288 net_devmem_get_dma_addr(niov)); 289 if (direction == DMA_TO_DEVICE) 290 binding->tx_vec[owner->area.base_virtual / PAGE_SIZE + i] = niov; 291 } 292 293 virtual += len; 294 } 295 296 err = xa_alloc_cyclic(&net_devmem_dmabuf_bindings, &binding->id, 297 binding, xa_limit_32b, &id_alloc_next, 298 GFP_KERNEL); 299 if (err < 0) 300 goto err_free_chunks; 301 302 return binding; 303 304 err_free_chunks: 305 gen_pool_for_each_chunk(binding->chunk_pool, 306 net_devmem_dmabuf_free_chunk_owner, NULL); 307 gen_pool_destroy(binding->chunk_pool); 308 err_tx_vec: 309 kvfree(binding->tx_vec); 310 err_unmap: 311 dma_buf_unmap_attachment_unlocked(binding->attachment, binding->sgt, 312 DMA_FROM_DEVICE); 313 err_detach: 314 dma_buf_detach(dmabuf, binding->attachment); 315 err_free_binding: 316 kfree(binding); 317 err_put_dmabuf: 318 dma_buf_put(dmabuf); 319 return ERR_PTR(err); 320 } 321 322 struct net_devmem_dmabuf_binding *net_devmem_lookup_dmabuf(u32 id) 323 { 324 struct net_devmem_dmabuf_binding *binding; 325 326 rcu_read_lock(); 327 binding = xa_load(&net_devmem_dmabuf_bindings, id); 328 if (binding) { 329 if (!net_devmem_dmabuf_binding_get(binding)) 330 binding = NULL; 331 } 332 rcu_read_unlock(); 333 334 return binding; 335 } 336 337 void net_devmem_get_net_iov(struct net_iov *niov) 338 { 339 net_devmem_dmabuf_binding_get(net_devmem_iov_binding(niov)); 340 } 341 342 void net_devmem_put_net_iov(struct net_iov *niov) 343 { 344 net_devmem_dmabuf_binding_put(net_devmem_iov_binding(niov)); 345 } 346 347 struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk, 348 unsigned int dmabuf_id) 349 { 350 struct net_devmem_dmabuf_binding *binding; 351 struct dst_entry *dst = __sk_dst_get(sk); 352 int err = 0; 353 354 binding = net_devmem_lookup_dmabuf(dmabuf_id); 355 if (!binding || !binding->tx_vec) { 356 err = -EINVAL; 357 goto out_err; 358 } 359 360 /* The dma-addrs in this binding are only reachable to the corresponding 361 * net_device. 362 */ 363 if (!dst || !dst->dev || dst->dev->ifindex != binding->dev->ifindex) { 364 err = -ENODEV; 365 goto out_err; 366 } 367 368 return binding; 369 370 out_err: 371 if (binding) 372 net_devmem_dmabuf_binding_put(binding); 373 374 return ERR_PTR(err); 375 } 376 377 struct net_iov * 378 net_devmem_get_niov_at(struct net_devmem_dmabuf_binding *binding, 379 size_t virt_addr, size_t *off, size_t *size) 380 { 381 if (virt_addr >= binding->dmabuf->size) 382 return NULL; 383 384 *off = virt_addr % PAGE_SIZE; 385 *size = PAGE_SIZE - *off; 386 387 return binding->tx_vec[virt_addr / PAGE_SIZE]; 388 } 389 390 /*** "Dmabuf devmem memory provider" ***/ 391 392 int mp_dmabuf_devmem_init(struct page_pool *pool) 393 { 394 struct net_devmem_dmabuf_binding *binding = pool->mp_priv; 395 396 if (!binding) 397 return -EINVAL; 398 399 /* dma-buf dma addresses do not need and should not be used with 400 * dma_sync_for_cpu/device. Force disable dma_sync. 401 */ 402 pool->dma_sync = false; 403 pool->dma_sync_for_cpu = false; 404 405 if (pool->p.order != 0) 406 return -E2BIG; 407 408 net_devmem_dmabuf_binding_get(binding); 409 return 0; 410 } 411 412 netmem_ref mp_dmabuf_devmem_alloc_netmems(struct page_pool *pool, gfp_t gfp) 413 { 414 struct net_devmem_dmabuf_binding *binding = pool->mp_priv; 415 struct net_iov *niov; 416 netmem_ref netmem; 417 418 niov = net_devmem_alloc_dmabuf(binding); 419 if (!niov) 420 return 0; 421 422 netmem = net_iov_to_netmem(niov); 423 424 page_pool_set_pp_info(pool, netmem); 425 426 pool->pages_state_hold_cnt++; 427 trace_page_pool_state_hold(pool, netmem, pool->pages_state_hold_cnt); 428 return netmem; 429 } 430 431 void mp_dmabuf_devmem_destroy(struct page_pool *pool) 432 { 433 struct net_devmem_dmabuf_binding *binding = pool->mp_priv; 434 435 net_devmem_dmabuf_binding_put(binding); 436 } 437 438 bool mp_dmabuf_devmem_release_page(struct page_pool *pool, netmem_ref netmem) 439 { 440 long refcount = atomic_long_read(netmem_get_pp_ref_count_ref(netmem)); 441 442 if (WARN_ON_ONCE(!netmem_is_net_iov(netmem))) 443 return false; 444 445 if (WARN_ON_ONCE(refcount != 1)) 446 return false; 447 448 page_pool_clear_pp_info(netmem); 449 450 net_devmem_free_dmabuf(netmem_to_net_iov(netmem)); 451 452 /* We don't want the page pool put_page()ing our net_iovs. */ 453 return false; 454 } 455 456 static int mp_dmabuf_devmem_nl_fill(void *mp_priv, struct sk_buff *rsp, 457 struct netdev_rx_queue *rxq) 458 { 459 const struct net_devmem_dmabuf_binding *binding = mp_priv; 460 int type = rxq ? NETDEV_A_QUEUE_DMABUF : NETDEV_A_PAGE_POOL_DMABUF; 461 462 return nla_put_u32(rsp, type, binding->id); 463 } 464 465 static void mp_dmabuf_devmem_uninstall(void *mp_priv, 466 struct netdev_rx_queue *rxq) 467 { 468 struct net_devmem_dmabuf_binding *binding = mp_priv; 469 struct netdev_rx_queue *bound_rxq; 470 unsigned long xa_idx; 471 472 xa_for_each(&binding->bound_rxqs, xa_idx, bound_rxq) { 473 if (bound_rxq == rxq) { 474 xa_erase(&binding->bound_rxqs, xa_idx); 475 if (xa_empty(&binding->bound_rxqs)) { 476 mutex_lock(&binding->lock); 477 binding->dev = NULL; 478 mutex_unlock(&binding->lock); 479 } 480 break; 481 } 482 } 483 } 484 485 static const struct memory_provider_ops dmabuf_devmem_ops = { 486 .init = mp_dmabuf_devmem_init, 487 .destroy = mp_dmabuf_devmem_destroy, 488 .alloc_netmems = mp_dmabuf_devmem_alloc_netmems, 489 .release_netmem = mp_dmabuf_devmem_release_page, 490 .nl_fill = mp_dmabuf_devmem_nl_fill, 491 .uninstall = mp_dmabuf_devmem_uninstall, 492 }; 493