netmap_kern.h revision 241719
1/* 2 * Copyright (C) 2011-2012 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: head/sys/dev/netmap/netmap_kern.h 241719 2012-10-19 04:13:12Z luigi $ 28 * $Id: netmap_kern.h 11829 2012-09-26 04:06:34Z luigi $ 29 * 30 * The header contains the definitions of constants and function 31 * prototypes used only in kernelspace. 32 */ 33 34#ifndef _NET_NETMAP_KERN_H_ 35#define _NET_NETMAP_KERN_H_ 36 37#define NETMAP_MEM2 // use the new memory allocator 38 39#if defined(__FreeBSD__) 40#define likely(x) __builtin_expect(!!(x), 1) 41#define unlikely(x) __builtin_expect(!!(x), 0) 42 43#define NM_LOCK_T struct mtx 44#define NM_SELINFO_T struct selinfo 45#define MBUF_LEN(m) ((m)->m_pkthdr.len) 46#define NM_SEND_UP(ifp, m) ((ifp)->if_input)(ifp, m) 47#elif defined (linux) 48#define NM_LOCK_T spinlock_t 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#warning apple support is incomplete. 71#define likely(x) __builtin_expect(!!(x), 1) 72#define unlikely(x) __builtin_expect(!!(x), 0) 73#define NM_LOCK_T IOLock * 74#define NM_SELINFO_T struct selinfo 75#define MBUF_LEN(m) ((m)->m_pkthdr.len) 76#define NM_SEND_UP(ifp, m) ((ifp)->if_input)(ifp, m) 77 78#else 79#error unsupported platform 80#endif 81 82#define ND(format, ...) 83#define D(format, ...) \ 84 do { \ 85 struct timeval __xxts; \ 86 microtime(&__xxts); \ 87 printf("%03d.%06d %s [%d] " format "\n", \ 88 (int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec, \ 89 __FUNCTION__, __LINE__, ##__VA_ARGS__); \ 90 } while (0) 91 92/* rate limited, lps indicates how many per second */ 93#define RD(lps, format, ...) \ 94 do { \ 95 static int t0, __cnt; \ 96 if (t0 != time_second) { \ 97 t0 = time_second; \ 98 __cnt = 0; \ 99 } \ 100 if (__cnt++ < lps) \ 101 D(format, ##__VA_ARGS__); \ 102 } while (0) 103 104struct netmap_adapter; 105 106/* 107 * private, kernel view of a ring. Keeps track of the status of 108 * a ring across system calls. 109 * 110 * nr_hwcur index of the next buffer to refill. 111 * It corresponds to ring->cur - ring->reserved 112 * 113 * nr_hwavail the number of slots "owned" by userspace. 114 * nr_hwavail =:= ring->avail + ring->reserved 115 * 116 * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots. 117 * This is so that, on a reset, buffers owned by userspace are not 118 * modified by the kernel. In particular: 119 * RX rings: the next empty buffer (hwcur + hwavail + hwofs) coincides with 120 * the next empty buffer as known by the hardware (next_to_check or so). 121 * TX rings: hwcur + hwofs coincides with next_to_send 122 */ 123struct netmap_kring { 124 struct netmap_ring *ring; 125 u_int nr_hwcur; 126 int nr_hwavail; 127 u_int nr_kflags; /* private driver flags */ 128#define NKR_PENDINTR 0x1 // Pending interrupt. 129 u_int nkr_num_slots; 130 131 int nkr_hwofs; /* offset between NIC and netmap ring */ 132 struct netmap_adapter *na; 133 NM_SELINFO_T si; /* poll/select wait queue */ 134 NM_LOCK_T q_lock; /* used if no device lock available */ 135} __attribute__((__aligned__(64))); 136 137/* 138 * This struct extends the 'struct adapter' (or 139 * equivalent) device descriptor. It contains all fields needed to 140 * support netmap operation. 141 */ 142struct netmap_adapter { 143 /* 144 * On linux we do not have a good way to tell if an interface 145 * is netmap-capable. So we use the following trick: 146 * NA(ifp) points here, and the first entry (which hopefully 147 * always exists and is at least 32 bits) contains a magic 148 * value which we can use to detect that the interface is good. 149 */ 150 uint32_t magic; 151 uint32_t na_flags; /* future place for IFCAP_NETMAP */ 152#define NAF_SKIP_INTR 1 /* use the regular interrupt handler. 153 * useful during initialization 154 */ 155 int refcount; /* number of user-space descriptors using this 156 interface, which is equal to the number of 157 struct netmap_if objs in the mapped region. */ 158 /* 159 * The selwakeup in the interrupt thread can use per-ring 160 * and/or global wait queues. We track how many clients 161 * of each type we have so we can optimize the drivers, 162 * and especially avoid huge contention on the locks. 163 */ 164 int na_single; /* threads attached to a single hw queue */ 165 int na_multi; /* threads attached to multiple hw queues */ 166 167 int separate_locks; /* set if the interface suports different 168 locks for rx, tx and core. */ 169 170 u_int num_rx_rings; /* number of adapter receive rings */ 171 u_int num_tx_rings; /* number of adapter transmit rings */ 172 173 u_int num_tx_desc; /* number of descriptor in each queue */ 174 u_int num_rx_desc; 175 176 /* tx_rings and rx_rings are private but allocated 177 * as a contiguous chunk of memory. Each array has 178 * N+1 entries, for the adapter queues and for the host queue. 179 */ 180 struct netmap_kring *tx_rings; /* array of TX rings. */ 181 struct netmap_kring *rx_rings; /* array of RX rings. */ 182 183 NM_SELINFO_T tx_si, rx_si; /* global wait queues */ 184 185 /* copy of if_qflush and if_transmit pointers, to intercept 186 * packets from the network stack when netmap is active. 187 */ 188 int (*if_transmit)(struct ifnet *, struct mbuf *); 189 190 /* references to the ifnet and device routines, used by 191 * the generic netmap functions. 192 */ 193 struct ifnet *ifp; /* adapter is ifp->if_softc */ 194 195 NM_LOCK_T core_lock; /* used if no device lock available */ 196 197 int (*nm_register)(struct ifnet *, int onoff); 198 void (*nm_lock)(struct ifnet *, int what, u_int ringid); 199 int (*nm_txsync)(struct ifnet *, u_int ring, int lock); 200 int (*nm_rxsync)(struct ifnet *, u_int ring, int lock); 201 202 int bdg_port; 203#ifdef linux 204 struct net_device_ops nm_ndo; 205 int if_refcount; // XXX additions for bridge 206#endif /* linux */ 207}; 208 209/* 210 * The combination of "enable" (ifp->if_capenable & IFCAP_NETMAP) 211 * and refcount gives the status of the interface, namely: 212 * 213 * enable refcount Status 214 * 215 * FALSE 0 normal operation 216 * FALSE != 0 -- (impossible) 217 * TRUE 1 netmap mode 218 * TRUE 0 being deleted. 219 */ 220 221#define NETMAP_DELETING(_na) ( ((_na)->refcount == 0) && \ 222 ( (_na)->ifp->if_capenable & IFCAP_NETMAP) ) 223 224/* 225 * parameters for (*nm_lock)(adapter, what, index) 226 */ 227enum { 228 NETMAP_NO_LOCK = 0, 229 NETMAP_CORE_LOCK, NETMAP_CORE_UNLOCK, 230 NETMAP_TX_LOCK, NETMAP_TX_UNLOCK, 231 NETMAP_RX_LOCK, NETMAP_RX_UNLOCK, 232#ifdef __FreeBSD__ 233#define NETMAP_REG_LOCK NETMAP_CORE_LOCK 234#define NETMAP_REG_UNLOCK NETMAP_CORE_UNLOCK 235#else 236 NETMAP_REG_LOCK, NETMAP_REG_UNLOCK 237#endif 238}; 239 240/* 241 * The following are support routines used by individual drivers to 242 * support netmap operation. 243 * 244 * netmap_attach() initializes a struct netmap_adapter, allocating the 245 * struct netmap_ring's and the struct selinfo. 246 * 247 * netmap_detach() frees the memory allocated by netmap_attach(). 248 * 249 * netmap_start() replaces the if_transmit routine of the interface, 250 * and is used to intercept packets coming from the stack. 251 * 252 * netmap_load_map/netmap_reload_map are helper routines to set/reset 253 * the dmamap for a packet buffer 254 * 255 * netmap_reset() is a helper routine to be called in the driver 256 * when reinitializing a ring. 257 */ 258int netmap_attach(struct netmap_adapter *, int); 259void netmap_detach(struct ifnet *); 260int netmap_start(struct ifnet *, struct mbuf *); 261enum txrx { NR_RX = 0, NR_TX = 1 }; 262struct netmap_slot *netmap_reset(struct netmap_adapter *na, 263 enum txrx tx, int n, u_int new_cur); 264int netmap_ring_reinit(struct netmap_kring *); 265 266extern u_int netmap_buf_size; 267#define NETMAP_BUF_SIZE netmap_buf_size 268extern int netmap_mitigate; 269extern int netmap_no_pendintr; 270extern u_int netmap_total_buffers; 271extern char *netmap_buffer_base; 272extern int netmap_verbose; // XXX debugging 273enum { /* verbose flags */ 274 NM_VERB_ON = 1, /* generic verbose */ 275 NM_VERB_HOST = 0x2, /* verbose host stack */ 276 NM_VERB_RXSYNC = 0x10, /* verbose on rxsync/txsync */ 277 NM_VERB_TXSYNC = 0x20, 278 NM_VERB_RXINTR = 0x100, /* verbose on rx/tx intr (driver) */ 279 NM_VERB_TXINTR = 0x200, 280 NM_VERB_NIC_RXSYNC = 0x1000, /* verbose on rx/tx intr (driver) */ 281 NM_VERB_NIC_TXSYNC = 0x2000, 282}; 283 284/* 285 * NA returns a pointer to the struct netmap adapter from the ifp, 286 * WNA is used to write it. 287 */ 288#ifndef WNA 289#define WNA(_ifp) (_ifp)->if_pspare[0] 290#endif 291#define NA(_ifp) ((struct netmap_adapter *)WNA(_ifp)) 292 293/* 294 * Macros to determine if an interface is netmap capable or netmap enabled. 295 * See the magic field in struct netmap_adapter. 296 */ 297#ifdef __FreeBSD__ 298/* 299 * on FreeBSD just use if_capabilities and if_capenable. 300 */ 301#define NETMAP_CAPABLE(ifp) (NA(ifp) && \ 302 (ifp)->if_capabilities & IFCAP_NETMAP ) 303 304#define NETMAP_SET_CAPABLE(ifp) \ 305 (ifp)->if_capabilities |= IFCAP_NETMAP 306 307#else /* linux */ 308 309/* 310 * on linux: 311 * we check if NA(ifp) is set and its first element has a related 312 * magic value. The capenable is within the struct netmap_adapter. 313 */ 314#define NETMAP_MAGIC 0x52697a7a 315 316#define NETMAP_CAPABLE(ifp) (NA(ifp) && \ 317 ((uint32_t)(uintptr_t)NA(ifp) ^ NA(ifp)->magic) == NETMAP_MAGIC ) 318 319#define NETMAP_SET_CAPABLE(ifp) \ 320 NA(ifp)->magic = ((uint32_t)(uintptr_t)NA(ifp)) ^ NETMAP_MAGIC 321 322#endif /* linux */ 323 324#ifdef __FreeBSD__ 325/* Callback invoked by the dma machinery after a successfull dmamap_load */ 326static void netmap_dmamap_cb(__unused void *arg, 327 __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error) 328{ 329} 330 331/* bus_dmamap_load wrapper: call aforementioned function if map != NULL. 332 * XXX can we do it without a callback ? 333 */ 334static inline void 335netmap_load_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 336{ 337 if (map) 338 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE, 339 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); 340} 341 342/* update the map when a buffer changes. */ 343static inline void 344netmap_reload_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 345{ 346 if (map) { 347 bus_dmamap_unload(tag, map); 348 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE, 349 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); 350 } 351} 352#else /* linux */ 353 354/* 355 * XXX How do we redefine these functions: 356 * 357 * on linux we need 358 * dma_map_single(&pdev->dev, virt_addr, len, direction) 359 * dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction 360 * The len can be implicit (on netmap it is NETMAP_BUF_SIZE) 361 * unfortunately the direction is not, so we need to change 362 * something to have a cross API 363 */ 364#define netmap_load_map(_t, _m, _b) 365#define netmap_reload_map(_t, _m, _b) 366#if 0 367 struct e1000_buffer *buffer_info = &tx_ring->buffer_info[l]; 368 /* set time_stamp *before* dma to help avoid a possible race */ 369 buffer_info->time_stamp = jiffies; 370 buffer_info->mapped_as_page = false; 371 buffer_info->length = len; 372 //buffer_info->next_to_watch = l; 373 /* reload dma map */ 374 dma_unmap_single(&adapter->pdev->dev, buffer_info->dma, 375 NETMAP_BUF_SIZE, DMA_TO_DEVICE); 376 buffer_info->dma = dma_map_single(&adapter->pdev->dev, 377 addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE); 378 379 if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) { 380 D("dma mapping error"); 381 /* goto dma_error; See e1000_put_txbuf() */ 382 /* XXX reset */ 383 } 384 tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX 385 386#endif 387 388/* 389 * The bus_dmamap_sync() can be one of wmb() or rmb() depending on direction. 390 */ 391#define bus_dmamap_sync(_a, _b, _c) 392 393#endif /* linux */ 394 395/* 396 * functions to map NIC to KRING indexes (n2k) and vice versa (k2n) 397 */ 398static inline int 399netmap_idx_n2k(struct netmap_kring *kr, int idx) 400{ 401 int n = kr->nkr_num_slots; 402 idx += kr->nkr_hwofs; 403 if (idx < 0) 404 return idx + n; 405 else if (idx < n) 406 return idx; 407 else 408 return idx - n; 409} 410 411 412static inline int 413netmap_idx_k2n(struct netmap_kring *kr, int idx) 414{ 415 int n = kr->nkr_num_slots; 416 idx -= kr->nkr_hwofs; 417 if (idx < 0) 418 return idx + n; 419 else if (idx < n) 420 return idx; 421 else 422 return idx - n; 423} 424 425 426#ifdef NETMAP_MEM2 427/* Entries of the look-up table. */ 428struct lut_entry { 429 void *vaddr; /* virtual address. */ 430 vm_paddr_t paddr; /* phisical address. */ 431}; 432 433struct netmap_obj_pool; 434extern struct lut_entry *netmap_buffer_lut; 435#define NMB_VA(i) (netmap_buffer_lut[i].vaddr) 436#define NMB_PA(i) (netmap_buffer_lut[i].paddr) 437#else /* NETMAP_MEM1 */ 438#define NMB_VA(i) (netmap_buffer_base + (i * NETMAP_BUF_SIZE) ) 439#endif /* NETMAP_MEM2 */ 440 441/* 442 * NMB return the virtual address of a buffer (buffer 0 on bad index) 443 * PNMB also fills the physical address 444 */ 445static inline void * 446NMB(struct netmap_slot *slot) 447{ 448 uint32_t i = slot->buf_idx; 449 return (unlikely(i >= netmap_total_buffers)) ? NMB_VA(0) : NMB_VA(i); 450} 451 452static inline void * 453PNMB(struct netmap_slot *slot, uint64_t *pp) 454{ 455 uint32_t i = slot->buf_idx; 456 void *ret = (i >= netmap_total_buffers) ? NMB_VA(0) : NMB_VA(i); 457#ifdef NETMAP_MEM2 458 *pp = (i >= netmap_total_buffers) ? NMB_PA(0) : NMB_PA(i); 459#else 460 *pp = vtophys(ret); 461#endif 462 return ret; 463} 464 465/* default functions to handle rx/tx interrupts */ 466int netmap_rx_irq(struct ifnet *, int, int *); 467#define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL) 468 469extern int netmap_copy; 470#endif /* _NET_NETMAP_KERN_H_ */ 471