1 /* 2 * Copyright (C) 2011-2013 Matteo Landi, Luigi Rizzo. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 23 * SUCH DAMAGE. 24 */ 25 26 /* 27 * $FreeBSD$ 28 * 29 * The header contains the definitions of constants and function 30 * prototypes used only in kernelspace. 31 */ 32 33 #ifndef _NET_NETMAP_KERN_H_ 34 #define _NET_NETMAP_KERN_H_ 35 36 #if defined(__FreeBSD__) 37 38 #define likely(x) __builtin_expect(!!(x), 1) 39 #define unlikely(x) __builtin_expect(!!(x), 0) 40 41 #define NM_LOCK_T struct mtx 42 #define NM_SELINFO_T struct selinfo 43 #define MBUF_LEN(m) ((m)->m_pkthdr.len) 44 #define NM_SEND_UP(ifp, m) ((ifp)->if_input)(ifp, m) 45 46 #elif defined (linux) 47 48 #define NM_LOCK_T safe_spinlock_t // see bsd_glue.h 49 #define NM_SELINFO_T wait_queue_head_t 50 #define MBUF_LEN(m) ((m)->len) 51 #define NM_SEND_UP(ifp, m) netif_rx(m) 52 53 #ifndef DEV_NETMAP 54 #define DEV_NETMAP 55 #endif 56 57 /* 58 * IFCAP_NETMAP goes into net_device's priv_flags (if_capenable). 59 * This was 16 bits up to linux 2.6.36, so we need a 16 bit value on older 60 * platforms and tolerate the clash with IFF_DYNAMIC and IFF_BRIDGE_PORT. 61 * For the 32-bit value, 0x100000 has no clashes until at least 3.5.1 62 */ 63 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,37) 64 #define IFCAP_NETMAP 0x8000 65 #else 66 #define IFCAP_NETMAP 0x100000 67 #endif 68 69 #elif defined (__APPLE__) 70 71 #warning apple support is incomplete. 72 #define likely(x) __builtin_expect(!!(x), 1) 73 #define unlikely(x) __builtin_expect(!!(x), 0) 74 #define NM_LOCK_T IOLock * 75 #define NM_SELINFO_T struct selinfo 76 #define MBUF_LEN(m) ((m)->m_pkthdr.len) 77 #define NM_SEND_UP(ifp, m) ((ifp)->if_input)(ifp, m) 78 79 #else 80 81 #error unsupported platform 82 83 #endif /* end - platform-specific code */ 84 85 #define ND(format, ...) 86 #define D(format, ...) \ 87 do { \ 88 struct timeval __xxts; \ 89 microtime(&__xxts); \ 90 printf("%03d.%06d %s [%d] " format "\n", \ 91 (int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec, \ 92 __FUNCTION__, __LINE__, ##__VA_ARGS__); \ 93 } while (0) 94 95 /* rate limited, lps indicates how many per second */ 96 #define RD(lps, format, ...) \ 97 do { \ 98 static int t0, __cnt; \ 99 if (t0 != time_second) { \ 100 t0 = time_second; \ 101 __cnt = 0; \ 102 } \ 103 if (__cnt++ < lps) \ 104 D(format, ##__VA_ARGS__); \ 105 } while (0) 106 107 struct netmap_adapter; 108 109 /* 110 * private, kernel view of a ring. Keeps track of the status of 111 * a ring across system calls. 112 * 113 * nr_hwcur index of the next buffer to refill. 114 * It corresponds to ring->cur - ring->reserved 115 * 116 * nr_hwavail the number of slots "owned" by userspace. 117 * nr_hwavail =:= ring->avail + ring->reserved 118 * 119 * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots. 120 * This is so that, on a reset, buffers owned by userspace are not 121 * modified by the kernel. In particular: 122 * RX rings: the next empty buffer (hwcur + hwavail + hwofs) coincides with 123 * the next empty buffer as known by the hardware (next_to_check or so). 124 * TX rings: hwcur + hwofs coincides with next_to_send 125 * 126 * For received packets, slot->flags is set to nkr_slot_flags 127 * so we can provide a proper initial value (e.g. set NS_FORWARD 128 * when operating in 'transparent' mode). 129 */ 130 struct netmap_kring { 131 struct netmap_ring *ring; 132 u_int nr_hwcur; 133 int nr_hwavail; 134 u_int nr_kflags; /* private driver flags */ 135 #define NKR_PENDINTR 0x1 // Pending interrupt. 136 u_int nkr_num_slots; 137 138 uint16_t nkr_slot_flags; /* initial value for flags */ 139 int nkr_hwofs; /* offset between NIC and netmap ring */ 140 struct netmap_adapter *na; 141 NM_SELINFO_T si; /* poll/select wait queue */ 142 NM_LOCK_T q_lock; /* used if no device lock available */ 143 } __attribute__((__aligned__(64))); 144 145 /* 146 * This struct extends the 'struct adapter' (or 147 * equivalent) device descriptor. It contains all fields needed to 148 * support netmap operation. 149 */ 150 struct netmap_adapter { 151 /* 152 * On linux we do not have a good way to tell if an interface 153 * is netmap-capable. So we use the following trick: 154 * NA(ifp) points here, and the first entry (which hopefully 155 * always exists and is at least 32 bits) contains a magic 156 * value which we can use to detect that the interface is good. 157 */ 158 uint32_t magic; 159 uint32_t na_flags; /* future place for IFCAP_NETMAP */ 160 #define NAF_SKIP_INTR 1 /* use the regular interrupt handler. 161 * useful during initialization 162 */ 163 int refcount; /* number of user-space descriptors using this 164 interface, which is equal to the number of 165 struct netmap_if objs in the mapped region. */ 166 /* 167 * The selwakeup in the interrupt thread can use per-ring 168 * and/or global wait queues. We track how many clients 169 * of each type we have so we can optimize the drivers, 170 * and especially avoid huge contention on the locks. 171 */ 172 int na_single; /* threads attached to a single hw queue */ 173 int na_multi; /* threads attached to multiple hw queues */ 174 175 int separate_locks; /* set if the interface suports different 176 locks for rx, tx and core. */ 177 178 u_int num_rx_rings; /* number of adapter receive rings */ 179 u_int num_tx_rings; /* number of adapter transmit rings */ 180 181 u_int num_tx_desc; /* number of descriptor in each queue */ 182 u_int num_rx_desc; 183 184 /* tx_rings and rx_rings are private but allocated 185 * as a contiguous chunk of memory. Each array has 186 * N+1 entries, for the adapter queues and for the host queue. 187 */ 188 struct netmap_kring *tx_rings; /* array of TX rings. */ 189 struct netmap_kring *rx_rings; /* array of RX rings. */ 190 191 NM_SELINFO_T tx_si, rx_si; /* global wait queues */ 192 193 /* copy of if_qflush and if_transmit pointers, to intercept 194 * packets from the network stack when netmap is active. 195 */ 196 int (*if_transmit)(struct ifnet *, struct mbuf *); 197 198 /* references to the ifnet and device routines, used by 199 * the generic netmap functions. 200 */ 201 struct ifnet *ifp; /* adapter is ifp->if_softc */ 202 203 NM_LOCK_T core_lock; /* used if no device lock available */ 204 205 int (*nm_register)(struct ifnet *, int onoff); 206 void (*nm_lock)(struct ifnet *, int what, u_int ringid); 207 int (*nm_txsync)(struct ifnet *, u_int ring, int lock); 208 int (*nm_rxsync)(struct ifnet *, u_int ring, int lock); 209 /* return configuration information */ 210 int (*nm_config)(struct ifnet *, u_int *txr, u_int *txd, 211 u_int *rxr, u_int *rxd); 212 213 int bdg_port; 214 #ifdef linux 215 struct net_device_ops nm_ndo; 216 int if_refcount; // XXX additions for bridge 217 #endif /* linux */ 218 }; 219 220 /* 221 * The combination of "enable" (ifp->if_capenable & IFCAP_NETMAP) 222 * and refcount gives the status of the interface, namely: 223 * 224 * enable refcount Status 225 * 226 * FALSE 0 normal operation 227 * FALSE != 0 -- (impossible) 228 * TRUE 1 netmap mode 229 * TRUE 0 being deleted. 230 */ 231 232 #define NETMAP_DELETING(_na) ( ((_na)->refcount == 0) && \ 233 ( (_na)->ifp->if_capenable & IFCAP_NETMAP) ) 234 235 /* 236 * parameters for (*nm_lock)(adapter, what, index) 237 */ 238 enum { 239 NETMAP_NO_LOCK = 0, 240 NETMAP_CORE_LOCK, NETMAP_CORE_UNLOCK, 241 NETMAP_TX_LOCK, NETMAP_TX_UNLOCK, 242 NETMAP_RX_LOCK, NETMAP_RX_UNLOCK, 243 #ifdef __FreeBSD__ 244 #define NETMAP_REG_LOCK NETMAP_CORE_LOCK 245 #define NETMAP_REG_UNLOCK NETMAP_CORE_UNLOCK 246 #else 247 NETMAP_REG_LOCK, NETMAP_REG_UNLOCK 248 #endif 249 }; 250 251 /* 252 * The following are support routines used by individual drivers to 253 * support netmap operation. 254 * 255 * netmap_attach() initializes a struct netmap_adapter, allocating the 256 * struct netmap_ring's and the struct selinfo. 257 * 258 * netmap_detach() frees the memory allocated by netmap_attach(). 259 * 260 * netmap_start() replaces the if_transmit routine of the interface, 261 * and is used to intercept packets coming from the stack. 262 * 263 * netmap_load_map/netmap_reload_map are helper routines to set/reset 264 * the dmamap for a packet buffer 265 * 266 * netmap_reset() is a helper routine to be called in the driver 267 * when reinitializing a ring. 268 */ 269 int netmap_attach(struct netmap_adapter *, int); 270 void netmap_detach(struct ifnet *); 271 int netmap_start(struct ifnet *, struct mbuf *); 272 enum txrx { NR_RX = 0, NR_TX = 1 }; 273 struct netmap_slot *netmap_reset(struct netmap_adapter *na, 274 enum txrx tx, int n, u_int new_cur); 275 int netmap_ring_reinit(struct netmap_kring *); 276 277 extern u_int netmap_buf_size; 278 #define NETMAP_BUF_SIZE netmap_buf_size 279 extern int netmap_mitigate; 280 extern int netmap_no_pendintr; 281 extern u_int netmap_total_buffers; 282 extern char *netmap_buffer_base; 283 extern int netmap_verbose; // XXX debugging 284 enum { /* verbose flags */ 285 NM_VERB_ON = 1, /* generic verbose */ 286 NM_VERB_HOST = 0x2, /* verbose host stack */ 287 NM_VERB_RXSYNC = 0x10, /* verbose on rxsync/txsync */ 288 NM_VERB_TXSYNC = 0x20, 289 NM_VERB_RXINTR = 0x100, /* verbose on rx/tx intr (driver) */ 290 NM_VERB_TXINTR = 0x200, 291 NM_VERB_NIC_RXSYNC = 0x1000, /* verbose on rx/tx intr (driver) */ 292 NM_VERB_NIC_TXSYNC = 0x2000, 293 }; 294 295 /* 296 * NA returns a pointer to the struct netmap adapter from the ifp, 297 * WNA is used to write it. 298 */ 299 #ifndef WNA 300 #define WNA(_ifp) (_ifp)->if_pspare[0] 301 #endif 302 #define NA(_ifp) ((struct netmap_adapter *)WNA(_ifp)) 303 304 /* 305 * Macros to determine if an interface is netmap capable or netmap enabled. 306 * See the magic field in struct netmap_adapter. 307 */ 308 #ifdef __FreeBSD__ 309 /* 310 * on FreeBSD just use if_capabilities and if_capenable. 311 */ 312 #define NETMAP_CAPABLE(ifp) (NA(ifp) && \ 313 (ifp)->if_capabilities & IFCAP_NETMAP ) 314 315 #define NETMAP_SET_CAPABLE(ifp) \ 316 (ifp)->if_capabilities |= IFCAP_NETMAP 317 318 #else /* linux */ 319 320 /* 321 * on linux: 322 * we check if NA(ifp) is set and its first element has a related 323 * magic value. The capenable is within the struct netmap_adapter. 324 */ 325 #define NETMAP_MAGIC 0x52697a7a 326 327 #define NETMAP_CAPABLE(ifp) (NA(ifp) && \ 328 ((uint32_t)(uintptr_t)NA(ifp) ^ NA(ifp)->magic) == NETMAP_MAGIC ) 329 330 #define NETMAP_SET_CAPABLE(ifp) \ 331 NA(ifp)->magic = ((uint32_t)(uintptr_t)NA(ifp)) ^ NETMAP_MAGIC 332 333 #endif /* linux */ 334 335 #ifdef __FreeBSD__ 336 /* Callback invoked by the dma machinery after a successfull dmamap_load */ 337 static void netmap_dmamap_cb(__unused void *arg, 338 __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error) 339 { 340 } 341 342 /* bus_dmamap_load wrapper: call aforementioned function if map != NULL. 343 * XXX can we do it without a callback ? 344 */ 345 static inline void 346 netmap_load_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 347 { 348 if (map) 349 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE, 350 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); 351 } 352 353 /* update the map when a buffer changes. */ 354 static inline void 355 netmap_reload_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 356 { 357 if (map) { 358 bus_dmamap_unload(tag, map); 359 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE, 360 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); 361 } 362 } 363 #else /* linux */ 364 365 /* 366 * XXX How do we redefine these functions: 367 * 368 * on linux we need 369 * dma_map_single(&pdev->dev, virt_addr, len, direction) 370 * dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction 371 * The len can be implicit (on netmap it is NETMAP_BUF_SIZE) 372 * unfortunately the direction is not, so we need to change 373 * something to have a cross API 374 */ 375 #define netmap_load_map(_t, _m, _b) 376 #define netmap_reload_map(_t, _m, _b) 377 #if 0 378 struct e1000_buffer *buffer_info = &tx_ring->buffer_info[l]; 379 /* set time_stamp *before* dma to help avoid a possible race */ 380 buffer_info->time_stamp = jiffies; 381 buffer_info->mapped_as_page = false; 382 buffer_info->length = len; 383 //buffer_info->next_to_watch = l; 384 /* reload dma map */ 385 dma_unmap_single(&adapter->pdev->dev, buffer_info->dma, 386 NETMAP_BUF_SIZE, DMA_TO_DEVICE); 387 buffer_info->dma = dma_map_single(&adapter->pdev->dev, 388 addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE); 389 390 if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) { 391 D("dma mapping error"); 392 /* goto dma_error; See e1000_put_txbuf() */ 393 /* XXX reset */ 394 } 395 tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX 396 397 #endif 398 399 /* 400 * The bus_dmamap_sync() can be one of wmb() or rmb() depending on direction. 401 */ 402 #define bus_dmamap_sync(_a, _b, _c) 403 404 #endif /* linux */ 405 406 /* 407 * functions to map NIC to KRING indexes (n2k) and vice versa (k2n) 408 */ 409 static inline int 410 netmap_idx_n2k(struct netmap_kring *kr, int idx) 411 { 412 int n = kr->nkr_num_slots; 413 idx += kr->nkr_hwofs; 414 if (idx < 0) 415 return idx + n; 416 else if (idx < n) 417 return idx; 418 else 419 return idx - n; 420 } 421 422 423 static inline int 424 netmap_idx_k2n(struct netmap_kring *kr, int idx) 425 { 426 int n = kr->nkr_num_slots; 427 idx -= kr->nkr_hwofs; 428 if (idx < 0) 429 return idx + n; 430 else if (idx < n) 431 return idx; 432 else 433 return idx - n; 434 } 435 436 437 /* Entries of the look-up table. */ 438 struct lut_entry { 439 void *vaddr; /* virtual address. */ 440 vm_paddr_t paddr; /* phisical address. */ 441 }; 442 443 struct netmap_obj_pool; 444 extern struct lut_entry *netmap_buffer_lut; 445 #define NMB_VA(i) (netmap_buffer_lut[i].vaddr) 446 #define NMB_PA(i) (netmap_buffer_lut[i].paddr) 447 448 /* 449 * NMB return the virtual address of a buffer (buffer 0 on bad index) 450 * PNMB also fills the physical address 451 */ 452 static inline void * 453 NMB(struct netmap_slot *slot) 454 { 455 uint32_t i = slot->buf_idx; 456 return (unlikely(i >= netmap_total_buffers)) ? NMB_VA(0) : NMB_VA(i); 457 } 458 459 static inline void * 460 PNMB(struct netmap_slot *slot, uint64_t *pp) 461 { 462 uint32_t i = slot->buf_idx; 463 void *ret = (i >= netmap_total_buffers) ? NMB_VA(0) : NMB_VA(i); 464 465 *pp = (i >= netmap_total_buffers) ? NMB_PA(0) : NMB_PA(i); 466 return ret; 467 } 468 469 /* default functions to handle rx/tx interrupts */ 470 int netmap_rx_irq(struct ifnet *, int, int *); 471 #define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL) 472 473 474 extern int netmap_copy; 475 #endif /* _NET_NETMAP_KERN_H_ */ 476