1 /* 2 * Copyright (C) 2011-2013 Matteo Landi, Luigi Rizzo. All rights reserved. 3 * Copyright (C) 2013 Universita` di Pisa. All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 */ 26 27 /* 28 * $FreeBSD$ 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 WITH_VALE // comment out to disable VALE support 38 39 #if defined(__FreeBSD__) 40 41 #define likely(x) __builtin_expect((long)!!(x), 1L) 42 #define unlikely(x) __builtin_expect((long)!!(x), 0L) 43 44 #define NM_LOCK_T struct mtx 45 #define NMG_LOCK_T struct mtx 46 #define NMG_LOCK_INIT() mtx_init(&netmap_global_lock, \ 47 "netmap global lock", NULL, MTX_DEF) 48 #define NMG_LOCK_DESTROY() mtx_destroy(&netmap_global_lock) 49 #define NMG_LOCK() mtx_lock(&netmap_global_lock) 50 #define NMG_UNLOCK() mtx_unlock(&netmap_global_lock) 51 #define NMG_LOCK_ASSERT() mtx_assert(&netmap_global_lock, MA_OWNED) 52 53 #define NM_SELINFO_T struct selinfo 54 #define MBUF_LEN(m) ((m)->m_pkthdr.len) 55 #define MBUF_IFP(m) ((m)->m_pkthdr.rcvif) 56 #define NM_SEND_UP(ifp, m) ((ifp)->if_input)(ifp, m) 57 58 #define NM_ATOMIC_T volatile int // XXX ? 59 /* atomic operations */ 60 #include <machine/atomic.h> 61 #define NM_ATOMIC_TEST_AND_SET(p) (!atomic_cmpset_acq_int((p), 0, 1)) 62 #define NM_ATOMIC_CLEAR(p) atomic_store_rel_int((p), 0) 63 64 #define prefetch(x) __builtin_prefetch(x) 65 66 MALLOC_DECLARE(M_NETMAP); 67 68 // XXX linux struct, not used in FreeBSD 69 struct net_device_ops { 70 }; 71 struct hrtimer { 72 }; 73 74 #elif defined (linux) 75 76 #define NM_LOCK_T safe_spinlock_t // see bsd_glue.h 77 #define NM_SELINFO_T wait_queue_head_t 78 #define MBUF_LEN(m) ((m)->len) 79 #define MBUF_IFP(m) ((m)->dev) 80 #define NM_SEND_UP(ifp, m) netif_rx(m) 81 82 #define NM_ATOMIC_T volatile long unsigned int 83 84 // XXX a mtx would suffice here too 20130404 gl 85 #define NMG_LOCK_T struct semaphore 86 #define NMG_LOCK_INIT() sema_init(&netmap_global_lock, 1) 87 #define NMG_LOCK_DESTROY() 88 #define NMG_LOCK() down(&netmap_global_lock) 89 #define NMG_UNLOCK() up(&netmap_global_lock) 90 #define NMG_LOCK_ASSERT() // XXX to be completed 91 92 #ifndef DEV_NETMAP 93 #define DEV_NETMAP 94 #endif /* DEV_NETMAP */ 95 96 /* 97 * IFCAP_NETMAP goes into net_device's priv_flags (if_capenable). 98 * This was 16 bits up to linux 2.6.36, so we need a 16 bit value on older 99 * platforms and tolerate the clash with IFF_DYNAMIC and IFF_BRIDGE_PORT. 100 * For the 32-bit value, 0x100000 has no clashes until at least 3.5.1 101 */ 102 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,37) 103 #define IFCAP_NETMAP 0x8000 104 #else 105 #define IFCAP_NETMAP 0x200000 106 #endif 107 108 #elif defined (__APPLE__) 109 110 #warning apple support is incomplete. 111 #define likely(x) __builtin_expect(!!(x), 1) 112 #define unlikely(x) __builtin_expect(!!(x), 0) 113 #define NM_LOCK_T IOLock * 114 #define NM_SELINFO_T struct selinfo 115 #define MBUF_LEN(m) ((m)->m_pkthdr.len) 116 #define NM_SEND_UP(ifp, m) ((ifp)->if_input)(ifp, m) 117 118 #else 119 120 #error unsupported platform 121 122 #endif /* end - platform-specific code */ 123 124 #define ND(format, ...) 125 #define D(format, ...) \ 126 do { \ 127 struct timeval __xxts; \ 128 microtime(&__xxts); \ 129 printf("%03d.%06d %s [%d] " format "\n", \ 130 (int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec, \ 131 __FUNCTION__, __LINE__, ##__VA_ARGS__); \ 132 } while (0) 133 134 /* rate limited, lps indicates how many per second */ 135 #define RD(lps, format, ...) \ 136 do { \ 137 static int t0, __cnt; \ 138 if (t0 != time_second) { \ 139 t0 = time_second; \ 140 __cnt = 0; \ 141 } \ 142 if (__cnt++ < lps) \ 143 D(format, ##__VA_ARGS__); \ 144 } while (0) 145 146 struct netmap_adapter; 147 struct nm_bdg_fwd; 148 struct nm_bridge; 149 struct netmap_priv_d; 150 151 const char *nm_dump_buf(char *p, int len, int lim, char *dst); 152 153 #include "netmap_mbq.h" 154 155 extern NMG_LOCK_T netmap_global_lock; 156 157 /* 158 * private, kernel view of a ring. Keeps track of the status of 159 * a ring across system calls. 160 * 161 * nr_hwcur index of the next buffer to refill. 162 * It corresponds to ring->cur - ring->reserved 163 * 164 * nr_hwavail the number of slots "owned" by userspace. 165 * nr_hwavail =:= ring->avail + ring->reserved 166 * 167 * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots. 168 * This is so that, on a reset, buffers owned by userspace are not 169 * modified by the kernel. In particular: 170 * RX rings: the next empty buffer (hwcur + hwavail + hwofs) coincides with 171 * the next empty buffer as known by the hardware (next_to_check or so). 172 * TX rings: hwcur + hwofs coincides with next_to_send 173 * 174 * Clients cannot issue concurrent syscall on a ring. The system 175 * detects this and reports an error using two flags, 176 * NKR_WBUSY and NKR_RBUSY 177 * For received packets, slot->flags is set to nkr_slot_flags 178 * so we can provide a proper initial value (e.g. set NS_FORWARD 179 * when operating in 'transparent' mode). 180 * 181 * The following fields are used to implement lock-free copy of packets 182 * from input to output ports in VALE switch: 183 * nkr_hwlease buffer after the last one being copied. 184 * A writer in nm_bdg_flush reserves N buffers 185 * from nr_hwlease, advances it, then does the 186 * copy outside the lock. 187 * In RX rings (used for VALE ports), 188 * nkr_hwcur + nkr_hwavail <= nkr_hwlease < nkr_hwcur+N-1 189 * In TX rings (used for NIC or host stack ports) 190 * nkr_hwcur <= nkr_hwlease < nkr_hwcur+ nkr_hwavail 191 * nkr_leases array of nkr_num_slots where writers can report 192 * completion of their block. NR_NOSLOT (~0) indicates 193 * that the writer has not finished yet 194 * nkr_lease_idx index of next free slot in nr_leases, to be assigned 195 * 196 * The kring is manipulated by txsync/rxsync and generic netmap function. 197 * q_lock is used to arbitrate access to the kring from within the netmap 198 * code, and this and other protections guarantee that there is never 199 * more than 1 concurrent call to txsync or rxsync. So we are free 200 * to manipulate the kring from within txsync/rxsync without any extra 201 * locks. 202 */ 203 struct netmap_kring { 204 struct netmap_ring *ring; 205 uint32_t nr_hwcur; 206 uint32_t nr_hwavail; 207 uint32_t nr_kflags; /* private driver flags */ 208 int32_t nr_hwreserved; 209 #define NKR_PENDINTR 0x1 // Pending interrupt. 210 uint32_t nkr_num_slots; 211 int32_t nkr_hwofs; /* offset between NIC and netmap ring */ 212 213 uint16_t nkr_slot_flags; /* initial value for flags */ 214 struct netmap_adapter *na; 215 struct nm_bdg_fwd *nkr_ft; 216 uint32_t *nkr_leases; 217 #define NR_NOSLOT ((uint32_t)~0) 218 uint32_t nkr_hwlease; 219 uint32_t nkr_lease_idx; 220 221 NM_SELINFO_T si; /* poll/select wait queue */ 222 NM_LOCK_T q_lock; /* protects kring and ring. */ 223 NM_ATOMIC_T nr_busy; /* prevent concurrent syscalls */ 224 225 volatile int nkr_stopped; 226 227 /* support for adapters without native netmap support. 228 * On tx rings we preallocate an array of tx buffers 229 * (same size as the netmap ring), on rx rings we 230 * store incoming packets in a queue. 231 * XXX who writes to the rx queue ? 232 */ 233 struct mbuf **tx_pool; 234 u_int nr_ntc; /* Emulation of a next-to-clean RX ring pointer. */ 235 struct mbq rx_queue; /* A queue for intercepted rx mbufs. */ 236 237 } __attribute__((__aligned__(64))); 238 239 240 /* return the next index, with wraparound */ 241 static inline uint32_t 242 nm_next(uint32_t i, uint32_t lim) 243 { 244 return unlikely (i == lim) ? 0 : i + 1; 245 } 246 247 /* 248 * 249 * Here is the layout for the Rx and Tx rings. 250 251 RxRING TxRING 252 253 +-----------------+ +-----------------+ 254 | | | | 255 |XXX free slot XXX| |XXX free slot XXX| 256 +-----------------+ +-----------------+ 257 | |<-hwcur | |<-hwcur 258 | reserved h | | (ready | 259 +----------- w -+ | to be | 260 cur->| a | | sent) h | 261 | v | +---------- w | 262 | a | cur->| (being a | 263 | i | | prepared) v | 264 | avail l | | a | 265 +-----------------+ + a ------ i + 266 | | ... | v l |<-hwlease 267 | (being | ... | a | ... 268 | prepared) | ... | i | ... 269 +-----------------+ ... | l | ... 270 | |<-hwlease +-----------------+ 271 | | | | 272 | | | | 273 | | | | 274 | | | | 275 +-----------------+ +-----------------+ 276 277 * The cur/avail (user view) and hwcur/hwavail (kernel view) 278 * are used in the normal operation of the card. 279 * 280 * When a ring is the output of a switch port (Rx ring for 281 * a VALE port, Tx ring for the host stack or NIC), slots 282 * are reserved in blocks through 'hwlease' which points 283 * to the next unused slot. 284 * On an Rx ring, hwlease is always after hwavail, 285 * and completions cause avail to advance. 286 * On a Tx ring, hwlease is always between cur and hwavail, 287 * and completions cause cur to advance. 288 * 289 * nm_kr_space() returns the maximum number of slots that 290 * can be assigned. 291 * nm_kr_lease() reserves the required number of buffers, 292 * advances nkr_hwlease and also returns an entry in 293 * a circular array where completions should be reported. 294 */ 295 296 297 298 299 enum txrx { NR_RX = 0, NR_TX = 1 }; 300 301 /* 302 * The "struct netmap_adapter" extends the "struct adapter" 303 * (or equivalent) device descriptor. 304 * It contains all base fields needed to support netmap operation. 305 * There are in fact different types of netmap adapters 306 * (native, generic, VALE switch...) so a netmap_adapter is 307 * just the first field in the derived type. 308 */ 309 struct netmap_adapter { 310 /* 311 * On linux we do not have a good way to tell if an interface 312 * is netmap-capable. So we always use the following trick: 313 * NA(ifp) points here, and the first entry (which hopefully 314 * always exists and is at least 32 bits) contains a magic 315 * value which we can use to detect that the interface is good. 316 */ 317 uint32_t magic; 318 uint32_t na_flags; /* enabled, and other flags */ 319 #define NAF_SKIP_INTR 1 /* use the regular interrupt handler. 320 * useful during initialization 321 */ 322 #define NAF_SW_ONLY 2 /* forward packets only to sw adapter */ 323 #define NAF_BDG_MAYSLEEP 4 /* the bridge is allowed to sleep when 324 * forwarding packets coming from this 325 * interface 326 */ 327 #define NAF_MEM_OWNER 8 /* the adapter is responsible for the 328 * deallocation of the memory allocator 329 */ 330 #define NAF_NATIVE_ON 16 /* the adapter is native and the attached 331 * interface is in netmap mode 332 */ 333 #define NAF_NETMAP_ON 32 /* netmap is active (either native or 334 * emulated. Where possible (e.g. FreeBSD) 335 * IFCAP_NETMAP also mirrors this flag. 336 */ 337 int active_fds; /* number of user-space descriptors using this 338 interface, which is equal to the number of 339 struct netmap_if objs in the mapped region. */ 340 341 u_int num_rx_rings; /* number of adapter receive rings */ 342 u_int num_tx_rings; /* number of adapter transmit rings */ 343 344 u_int num_tx_desc; /* number of descriptor in each queue */ 345 u_int num_rx_desc; 346 347 /* tx_rings and rx_rings are private but allocated 348 * as a contiguous chunk of memory. Each array has 349 * N+1 entries, for the adapter queues and for the host queue. 350 */ 351 struct netmap_kring *tx_rings; /* array of TX rings. */ 352 struct netmap_kring *rx_rings; /* array of RX rings. */ 353 void *tailroom; /* space below the rings array */ 354 /* (used for leases) */ 355 356 357 NM_SELINFO_T tx_si, rx_si; /* global wait queues */ 358 359 /* copy of if_qflush and if_transmit pointers, to intercept 360 * packets from the network stack when netmap is active. 361 */ 362 int (*if_transmit)(struct ifnet *, struct mbuf *); 363 364 /* references to the ifnet and device routines, used by 365 * the generic netmap functions. 366 */ 367 struct ifnet *ifp; /* adapter is ifp->if_softc */ 368 369 /* private cleanup */ 370 void (*nm_dtor)(struct netmap_adapter *); 371 372 int (*nm_register)(struct netmap_adapter *, int onoff); 373 374 int (*nm_txsync)(struct netmap_adapter *, u_int ring, int flags); 375 int (*nm_rxsync)(struct netmap_adapter *, u_int ring, int flags); 376 #define NAF_FORCE_READ 1 377 #define NAF_FORCE_RECLAIM 2 378 /* return configuration information */ 379 int (*nm_config)(struct netmap_adapter *, 380 u_int *txr, u_int *txd, u_int *rxr, u_int *rxd); 381 int (*nm_krings_create)(struct netmap_adapter *); 382 void (*nm_krings_delete)(struct netmap_adapter *); 383 int (*nm_notify)(struct netmap_adapter *, 384 u_int ring, enum txrx, int flags); 385 #define NAF_GLOBAL_NOTIFY 4 386 #define NAF_DISABLE_NOTIFY 8 387 388 /* standard refcount to control the lifetime of the adapter 389 * (it should be equal to the lifetime of the corresponding ifp) 390 */ 391 int na_refcount; 392 393 /* memory allocator (opaque) 394 * We also cache a pointer to the lut_entry for translating 395 * buffer addresses, and the total number of buffers. 396 */ 397 struct netmap_mem_d *nm_mem; 398 struct lut_entry *na_lut; 399 uint32_t na_lut_objtotal; /* max buffer index */ 400 401 /* used internally. If non-null, the interface cannot be bound 402 * from userspace 403 */ 404 void *na_private; 405 }; 406 407 /* 408 * If the NIC is owned by the kernel 409 * (i.e., bridge), neither another bridge nor user can use it; 410 * if the NIC is owned by a user, only users can share it. 411 * Evaluation must be done under NMG_LOCK(). 412 */ 413 #define NETMAP_OWNED_BY_KERN(na) (na->na_private) 414 #define NETMAP_OWNED_BY_ANY(na) \ 415 (NETMAP_OWNED_BY_KERN(na) || (na->active_fds > 0)) 416 417 418 /* 419 * derived netmap adapters for various types of ports 420 */ 421 struct netmap_vp_adapter { /* VALE software port */ 422 struct netmap_adapter up; 423 424 /* 425 * Bridge support: 426 * 427 * bdg_port is the port number used in the bridge; 428 * na_bdg points to the bridge this NA is attached to. 429 */ 430 int bdg_port; 431 struct nm_bridge *na_bdg; 432 int retry; 433 434 u_int offset; /* Offset of ethernet header for each packet. */ 435 }; 436 437 struct netmap_hw_adapter { /* physical device */ 438 struct netmap_adapter up; 439 440 struct net_device_ops nm_ndo; // XXX linux only 441 }; 442 443 struct netmap_generic_adapter { /* non-native device */ 444 struct netmap_hw_adapter up; 445 446 /* Pointer to a previously used netmap adapter. */ 447 struct netmap_adapter *prev; 448 449 /* generic netmap adapters support: 450 * a net_device_ops struct overrides ndo_select_queue(), 451 * save_if_input saves the if_input hook (FreeBSD), 452 * mit_timer and mit_pending implement rx interrupt mitigation, 453 */ 454 struct net_device_ops generic_ndo; 455 void (*save_if_input)(struct ifnet *, struct mbuf *); 456 457 struct hrtimer mit_timer; 458 int mit_pending; 459 }; 460 461 #ifdef WITH_VALE 462 463 /* bridge wrapper for non VALE ports. It is used to connect real devices to the bridge. 464 * 465 * The real device must already have its own netmap adapter (hwna). The 466 * bridge wrapper and the hwna adapter share the same set of netmap rings and 467 * buffers, but they have two separate sets of krings descriptors, with tx/rx 468 * meanings swapped: 469 * 470 * netmap 471 * bwrap krings rings krings hwna 472 * +------+ +------+ +-----+ +------+ +------+ 473 * |tx_rings->| |\ /| |----| |<-tx_rings| 474 * | | +------+ \ / +-----+ +------+ | | 475 * | | X | | 476 * | | / \ | | 477 * | | +------+/ \+-----+ +------+ | | 478 * |rx_rings->| | | |----| |<-rx_rings| 479 * | | +------+ +-----+ +------+ | | 480 * +------+ +------+ 481 * 482 * - packets coming from the bridge go to the brwap rx rings, which are also the 483 * hwna tx rings. The bwrap notify callback will then complete the hwna tx 484 * (see netmap_bwrap_notify). 485 * - packets coming from the outside go to the hwna rx rings, which are also the 486 * bwrap tx rings. The (overwritten) hwna notify method will then complete 487 * the bridge tx (see netmap_bwrap_intr_notify). 488 * 489 * The bridge wrapper may optionally connect the hwna 'host' rings to the 490 * bridge. This is done by using a second port in the bridge and connecting it 491 * to the 'host' netmap_vp_adapter contained in the netmap_bwrap_adapter. 492 * The brwap host adapter cross-links the hwna host rings in the same way as shown above. 493 * 494 * - packets coming from the bridge and directed to host stack are handled by the 495 * bwrap host notify callback (see netmap_bwrap_host_notify) 496 * - packets coming from the host stack are still handled by the overwritten 497 * hwna notify callback (netmap_bwrap_intr_notify), but are diverted to the 498 * host adapter depending on the ring number. 499 * 500 */ 501 struct netmap_bwrap_adapter { 502 struct netmap_vp_adapter up; 503 struct netmap_vp_adapter host; /* for host rings */ 504 struct netmap_adapter *hwna; /* the underlying device */ 505 506 /* backup of the hwna notify callback */ 507 int (*save_notify)(struct netmap_adapter *, 508 u_int ring, enum txrx, int flags); 509 /* When we attach a physical interface to the bridge, we 510 * allow the controlling process to terminate, so we need 511 * a place to store the netmap_priv_d data structure. 512 * This is only done when physical interfaces are attached to a bridge. 513 */ 514 struct netmap_priv_d *na_kpriv; 515 }; 516 517 518 /* 519 * Available space in the ring. Only used in VALE code 520 */ 521 static inline uint32_t 522 nm_kr_space(struct netmap_kring *k, int is_rx) 523 { 524 int space; 525 526 if (is_rx) { 527 int busy = k->nkr_hwlease - k->nr_hwcur + k->nr_hwreserved; 528 if (busy < 0) 529 busy += k->nkr_num_slots; 530 space = k->nkr_num_slots - 1 - busy; 531 } else { 532 space = k->nr_hwcur + k->nr_hwavail - k->nkr_hwlease; 533 if (space < 0) 534 space += k->nkr_num_slots; 535 } 536 #if 0 537 // sanity check 538 if (k->nkr_hwlease >= k->nkr_num_slots || 539 k->nr_hwcur >= k->nkr_num_slots || 540 k->nr_hwavail >= k->nkr_num_slots || 541 busy < 0 || 542 busy >= k->nkr_num_slots) { 543 D("invalid kring, cur %d avail %d lease %d lease_idx %d lim %d", k->nr_hwcur, k->nr_hwavail, k->nkr_hwlease, 544 k->nkr_lease_idx, k->nkr_num_slots); 545 } 546 #endif 547 return space; 548 } 549 550 551 552 553 /* make a lease on the kring for N positions. return the 554 * lease index 555 */ 556 static inline uint32_t 557 nm_kr_lease(struct netmap_kring *k, u_int n, int is_rx) 558 { 559 uint32_t lim = k->nkr_num_slots - 1; 560 uint32_t lease_idx = k->nkr_lease_idx; 561 562 k->nkr_leases[lease_idx] = NR_NOSLOT; 563 k->nkr_lease_idx = nm_next(lease_idx, lim); 564 565 if (n > nm_kr_space(k, is_rx)) { 566 D("invalid request for %d slots", n); 567 panic("x"); 568 } 569 /* XXX verify that there are n slots */ 570 k->nkr_hwlease += n; 571 if (k->nkr_hwlease > lim) 572 k->nkr_hwlease -= lim + 1; 573 574 if (k->nkr_hwlease >= k->nkr_num_slots || 575 k->nr_hwcur >= k->nkr_num_slots || 576 k->nr_hwavail >= k->nkr_num_slots || 577 k->nkr_lease_idx >= k->nkr_num_slots) { 578 D("invalid kring %s, cur %d avail %d lease %d lease_idx %d lim %d", 579 k->na->ifp->if_xname, 580 k->nr_hwcur, k->nr_hwavail, k->nkr_hwlease, 581 k->nkr_lease_idx, k->nkr_num_slots); 582 } 583 return lease_idx; 584 } 585 586 #endif /* WITH_VALE */ 587 588 /* return update position */ 589 static inline uint32_t 590 nm_kr_rxpos(struct netmap_kring *k) 591 { 592 uint32_t pos = k->nr_hwcur + k->nr_hwavail; 593 if (pos >= k->nkr_num_slots) 594 pos -= k->nkr_num_slots; 595 #if 0 596 if (pos >= k->nkr_num_slots || 597 k->nkr_hwlease >= k->nkr_num_slots || 598 k->nr_hwcur >= k->nkr_num_slots || 599 k->nr_hwavail >= k->nkr_num_slots || 600 k->nkr_lease_idx >= k->nkr_num_slots) { 601 D("invalid kring, cur %d avail %d lease %d lease_idx %d lim %d", k->nr_hwcur, k->nr_hwavail, k->nkr_hwlease, 602 k->nkr_lease_idx, k->nkr_num_slots); 603 } 604 #endif 605 return pos; 606 } 607 608 609 /* 610 * protect against multiple threads using the same ring. 611 * also check that the ring has not been stopped. 612 * We only care for 0 or !=0 as a return code. 613 */ 614 #define NM_KR_BUSY 1 615 #define NM_KR_STOPPED 2 616 617 static __inline void nm_kr_put(struct netmap_kring *kr) 618 { 619 NM_ATOMIC_CLEAR(&kr->nr_busy); 620 } 621 622 static __inline int nm_kr_tryget(struct netmap_kring *kr) 623 { 624 /* check a first time without taking the lock 625 * to avoid starvation for nm_kr_get() 626 */ 627 if (unlikely(kr->nkr_stopped)) { 628 ND("ring %p stopped (%d)", kr, kr->nkr_stopped); 629 return NM_KR_STOPPED; 630 } 631 if (unlikely(NM_ATOMIC_TEST_AND_SET(&kr->nr_busy))) 632 return NM_KR_BUSY; 633 /* check a second time with lock held */ 634 if (unlikely(kr->nkr_stopped)) { 635 ND("ring %p stopped (%d)", kr, kr->nkr_stopped); 636 nm_kr_put(kr); 637 return NM_KR_STOPPED; 638 } 639 return 0; 640 } 641 642 643 /* 644 * The following are support routines used by individual drivers to 645 * support netmap operation. 646 * 647 * netmap_attach() initializes a struct netmap_adapter, allocating the 648 * struct netmap_ring's and the struct selinfo. 649 * 650 * netmap_detach() frees the memory allocated by netmap_attach(). 651 * 652 * netmap_transmit() replaces the if_transmit routine of the interface, 653 * and is used to intercept packets coming from the stack. 654 * 655 * netmap_load_map/netmap_reload_map are helper routines to set/reset 656 * the dmamap for a packet buffer 657 * 658 * netmap_reset() is a helper routine to be called in the driver 659 * when reinitializing a ring. 660 */ 661 int netmap_attach(struct netmap_adapter *); 662 int netmap_attach_common(struct netmap_adapter *); 663 void netmap_detach_common(struct netmap_adapter *na); 664 void netmap_detach(struct ifnet *); 665 int netmap_transmit(struct ifnet *, struct mbuf *); 666 struct netmap_slot *netmap_reset(struct netmap_adapter *na, 667 enum txrx tx, u_int n, u_int new_cur); 668 int netmap_ring_reinit(struct netmap_kring *); 669 670 /* set/clear native flags. XXX maybe also if_transmit ? */ 671 static inline void 672 nm_set_native_flags(struct netmap_adapter *na) 673 { 674 struct ifnet *ifp = na->ifp; 675 676 na->na_flags |= (NAF_NATIVE_ON | NAF_NETMAP_ON); 677 #ifdef IFCAP_NETMAP /* or FreeBSD ? */ 678 ifp->if_capenable |= IFCAP_NETMAP; 679 #endif 680 #ifdef __FreeBSD__ 681 na->if_transmit = ifp->if_transmit; 682 ifp->if_transmit = netmap_transmit; 683 #else 684 na->if_transmit = (void *)ifp->netdev_ops; 685 ifp->netdev_ops = &((struct netmap_hw_adapter *)na)->nm_ndo; 686 #endif 687 } 688 689 static inline void 690 nm_clear_native_flags(struct netmap_adapter *na) 691 { 692 struct ifnet *ifp = na->ifp; 693 694 #ifdef __FreeBSD__ 695 ifp->if_transmit = na->if_transmit; 696 #else 697 ifp->netdev_ops = (void *)na->if_transmit; 698 #endif 699 na->na_flags &= ~(NAF_NATIVE_ON | NAF_NETMAP_ON); 700 #ifdef IFCAP_NETMAP /* or FreeBSD ? */ 701 ifp->if_capenable &= ~IFCAP_NETMAP; 702 #endif 703 } 704 705 /* 706 * validates parameters in the ring/kring, returns a value for cur, 707 * and the 'new_slots' value in the argument. 708 * If any error, returns cur > lim to force a reinit. 709 */ 710 u_int nm_txsync_prologue(struct netmap_kring *, u_int *); 711 712 /* 713 * validates parameters in the ring/kring, returns a value for cur, 714 * and the 'reserved' value in the argument. 715 * If any error, returns cur > lim to force a reinit. 716 */ 717 u_int nm_rxsync_prologue(struct netmap_kring *, u_int *); 718 719 /* 720 * update kring and ring at the end of txsync 721 */ 722 static inline void 723 nm_txsync_finalize(struct netmap_kring *kring, u_int cur) 724 { 725 /* recompute hwreserved */ 726 kring->nr_hwreserved = cur - kring->nr_hwcur; 727 if (kring->nr_hwreserved < 0) 728 kring->nr_hwreserved += kring->nkr_num_slots; 729 730 /* update avail and reserved to what the kernel knows */ 731 kring->ring->avail = kring->nr_hwavail; 732 kring->ring->reserved = kring->nr_hwreserved; 733 } 734 735 /* check/fix address and len in tx rings */ 736 #if 1 /* debug version */ 737 #define NM_CHECK_ADDR_LEN(_a, _l) do { \ 738 if (_a == netmap_buffer_base || _l > NETMAP_BUF_SIZE) { \ 739 RD(5, "bad addr/len ring %d slot %d idx %d len %d", \ 740 ring_nr, nm_i, slot->buf_idx, len); \ 741 if (_l > NETMAP_BUF_SIZE) \ 742 _l = NETMAP_BUF_SIZE; \ 743 } } while (0) 744 #else /* no debug version */ 745 #define NM_CHECK_ADDR_LEN(_a, _l) do { \ 746 if (_l > NETMAP_BUF_SIZE) \ 747 _l = NETMAP_BUF_SIZE; \ 748 } while (0) 749 #endif 750 751 752 /*---------------------------------------------------------------*/ 753 /* 754 * Support routines to be used with the VALE switch 755 */ 756 int netmap_update_config(struct netmap_adapter *na); 757 int netmap_krings_create(struct netmap_adapter *na, u_int ntx, u_int nrx, u_int tailroom); 758 void netmap_krings_delete(struct netmap_adapter *na); 759 760 struct netmap_if * 761 netmap_do_regif(struct netmap_priv_d *priv, struct netmap_adapter *na, 762 uint16_t ringid, int *err); 763 764 765 766 u_int nm_bound_var(u_int *v, u_int dflt, u_int lo, u_int hi, const char *msg); 767 int netmap_get_na(struct nmreq *nmr, struct netmap_adapter **na, int create); 768 int netmap_get_hw_na(struct ifnet *ifp, struct netmap_adapter **na); 769 770 #ifdef WITH_VALE 771 /* 772 * The following bridge-related interfaces are used by other kernel modules 773 * In the version that only supports unicast or broadcast, the lookup 774 * function can return 0 .. NM_BDG_MAXPORTS-1 for regular ports, 775 * NM_BDG_MAXPORTS for broadcast, NM_BDG_MAXPORTS+1 for unknown. 776 * XXX in practice "unknown" might be handled same as broadcast. 777 */ 778 typedef u_int (*bdg_lookup_fn_t)(char *buf, u_int len, 779 uint8_t *ring_nr, struct netmap_vp_adapter *); 780 u_int netmap_bdg_learning(char *, u_int, uint8_t *, 781 struct netmap_vp_adapter *); 782 783 #define NM_BDG_MAXPORTS 254 /* up to 254 */ 784 #define NM_BDG_BROADCAST NM_BDG_MAXPORTS 785 #define NM_BDG_NOPORT (NM_BDG_MAXPORTS+1) 786 787 #define NM_NAME "vale" /* prefix for bridge port name */ 788 789 790 /* these are redefined in case of no VALE support */ 791 int netmap_get_bdg_na(struct nmreq *nmr, struct netmap_adapter **na, int create); 792 void netmap_init_bridges(void); 793 int netmap_bdg_ctl(struct nmreq *nmr, bdg_lookup_fn_t func); 794 795 #else /* !WITH_VALE */ 796 #define netmap_get_bdg_na(_1, _2, _3) 0 797 #define netmap_init_bridges(_1) 798 #define netmap_bdg_ctl(_1, _2) EINVAL 799 #endif /* !WITH_VALE */ 800 801 /* Various prototypes */ 802 int netmap_poll(struct cdev *dev, int events, struct thread *td); 803 804 805 int netmap_init(void); 806 void netmap_fini(void); 807 int netmap_get_memory(struct netmap_priv_d* p); 808 void netmap_dtor(void *data); 809 int netmap_dtor_locked(struct netmap_priv_d *priv); 810 811 int netmap_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td); 812 813 /* netmap_adapter creation/destruction */ 814 #define NM_IFPNAME(ifp) ((ifp) ? (ifp)->if_xname : "zombie") 815 #define NM_DEBUG_PUTGET 1 816 817 #ifdef NM_DEBUG_PUTGET 818 819 #define NM_DBG(f) __##f 820 821 void __netmap_adapter_get(struct netmap_adapter *na); 822 823 #define netmap_adapter_get(na) \ 824 do { \ 825 struct netmap_adapter *__na = na; \ 826 D("getting %p:%s (%d)", __na, NM_IFPNAME(__na->ifp), __na->na_refcount); \ 827 __netmap_adapter_get(__na); \ 828 } while (0) 829 830 int __netmap_adapter_put(struct netmap_adapter *na); 831 832 #define netmap_adapter_put(na) \ 833 do { \ 834 struct netmap_adapter *__na = na; \ 835 D("putting %p:%s (%d)", __na, NM_IFPNAME(__na->ifp), __na->na_refcount); \ 836 __netmap_adapter_put(__na); \ 837 } while (0) 838 839 #else /* !NM_DEBUG_PUTGET */ 840 841 #define NM_DBG(f) f 842 void netmap_adapter_get(struct netmap_adapter *na); 843 int netmap_adapter_put(struct netmap_adapter *na); 844 845 #endif /* !NM_DEBUG_PUTGET */ 846 847 848 extern u_int netmap_buf_size; 849 #define NETMAP_BUF_SIZE netmap_buf_size // XXX remove 850 extern int netmap_mitigate; 851 extern int netmap_no_pendintr; 852 extern u_int netmap_total_buffers; 853 extern char *netmap_buffer_base; 854 extern int netmap_verbose; // XXX debugging 855 enum { /* verbose flags */ 856 NM_VERB_ON = 1, /* generic verbose */ 857 NM_VERB_HOST = 0x2, /* verbose host stack */ 858 NM_VERB_RXSYNC = 0x10, /* verbose on rxsync/txsync */ 859 NM_VERB_TXSYNC = 0x20, 860 NM_VERB_RXINTR = 0x100, /* verbose on rx/tx intr (driver) */ 861 NM_VERB_TXINTR = 0x200, 862 NM_VERB_NIC_RXSYNC = 0x1000, /* verbose on rx/tx intr (driver) */ 863 NM_VERB_NIC_TXSYNC = 0x2000, 864 }; 865 866 extern int netmap_txsync_retry; 867 extern int netmap_generic_mit; 868 extern int netmap_generic_ringsize; 869 870 /* 871 * NA returns a pointer to the struct netmap adapter from the ifp, 872 * WNA is used to write it. 873 */ 874 #ifndef WNA 875 #define WNA(_ifp) (_ifp)->if_pspare[0] 876 #endif 877 #define NA(_ifp) ((struct netmap_adapter *)WNA(_ifp)) 878 879 /* 880 * Macros to determine if an interface is netmap capable or netmap enabled. 881 * See the magic field in struct netmap_adapter. 882 */ 883 #ifdef __FreeBSD__ 884 /* 885 * on FreeBSD just use if_capabilities and if_capenable. 886 */ 887 #define NETMAP_CAPABLE(ifp) (NA(ifp) && \ 888 (ifp)->if_capabilities & IFCAP_NETMAP ) 889 890 #define NETMAP_SET_CAPABLE(ifp) \ 891 (ifp)->if_capabilities |= IFCAP_NETMAP 892 893 #else /* linux */ 894 895 /* 896 * on linux: 897 * we check if NA(ifp) is set and its first element has a related 898 * magic value. The capenable is within the struct netmap_adapter. 899 */ 900 #define NETMAP_MAGIC 0x52697a7a 901 902 #define NETMAP_CAPABLE(ifp) (NA(ifp) && \ 903 ((uint32_t)(uintptr_t)NA(ifp) ^ NA(ifp)->magic) == NETMAP_MAGIC ) 904 905 #define NETMAP_SET_CAPABLE(ifp) \ 906 NA(ifp)->magic = ((uint32_t)(uintptr_t)NA(ifp)) ^ NETMAP_MAGIC 907 908 #endif /* linux */ 909 910 #ifdef __FreeBSD__ 911 912 /* Callback invoked by the dma machinery after a successfull dmamap_load */ 913 static void netmap_dmamap_cb(__unused void *arg, 914 __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error) 915 { 916 } 917 918 /* bus_dmamap_load wrapper: call aforementioned function if map != NULL. 919 * XXX can we do it without a callback ? 920 */ 921 static inline void 922 netmap_load_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 923 { 924 if (map) 925 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE, 926 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); 927 } 928 929 /* update the map when a buffer changes. */ 930 static inline void 931 netmap_reload_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 932 { 933 if (map) { 934 bus_dmamap_unload(tag, map); 935 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE, 936 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); 937 } 938 } 939 940 #else /* linux */ 941 942 /* 943 * XXX How do we redefine these functions: 944 * 945 * on linux we need 946 * dma_map_single(&pdev->dev, virt_addr, len, direction) 947 * dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction 948 * The len can be implicit (on netmap it is NETMAP_BUF_SIZE) 949 * unfortunately the direction is not, so we need to change 950 * something to have a cross API 951 */ 952 #define netmap_load_map(_t, _m, _b) 953 #define netmap_reload_map(_t, _m, _b) 954 #if 0 955 struct e1000_buffer *buffer_info = &tx_ring->buffer_info[l]; 956 /* set time_stamp *before* dma to help avoid a possible race */ 957 buffer_info->time_stamp = jiffies; 958 buffer_info->mapped_as_page = false; 959 buffer_info->length = len; 960 //buffer_info->next_to_watch = l; 961 /* reload dma map */ 962 dma_unmap_single(&adapter->pdev->dev, buffer_info->dma, 963 NETMAP_BUF_SIZE, DMA_TO_DEVICE); 964 buffer_info->dma = dma_map_single(&adapter->pdev->dev, 965 addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE); 966 967 if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) { 968 D("dma mapping error"); 969 /* goto dma_error; See e1000_put_txbuf() */ 970 /* XXX reset */ 971 } 972 tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX 973 974 #endif 975 976 /* 977 * The bus_dmamap_sync() can be one of wmb() or rmb() depending on direction. 978 */ 979 #define bus_dmamap_sync(_a, _b, _c) 980 981 #endif /* linux */ 982 983 984 /* 985 * functions to map NIC to KRING indexes (n2k) and vice versa (k2n) 986 */ 987 static inline int 988 netmap_idx_n2k(struct netmap_kring *kr, int idx) 989 { 990 int n = kr->nkr_num_slots; 991 idx += kr->nkr_hwofs; 992 if (idx < 0) 993 return idx + n; 994 else if (idx < n) 995 return idx; 996 else 997 return idx - n; 998 } 999 1000 1001 static inline int 1002 netmap_idx_k2n(struct netmap_kring *kr, int idx) 1003 { 1004 int n = kr->nkr_num_slots; 1005 idx -= kr->nkr_hwofs; 1006 if (idx < 0) 1007 return idx + n; 1008 else if (idx < n) 1009 return idx; 1010 else 1011 return idx - n; 1012 } 1013 1014 1015 /* Entries of the look-up table. */ 1016 struct lut_entry { 1017 void *vaddr; /* virtual address. */ 1018 vm_paddr_t paddr; /* physical address. */ 1019 }; 1020 1021 struct netmap_obj_pool; 1022 extern struct lut_entry *netmap_buffer_lut; 1023 #define NMB_VA(i) (netmap_buffer_lut[i].vaddr) 1024 #define NMB_PA(i) (netmap_buffer_lut[i].paddr) 1025 1026 /* 1027 * NMB return the virtual address of a buffer (buffer 0 on bad index) 1028 * PNMB also fills the physical address 1029 */ 1030 static inline void * 1031 NMB(struct netmap_slot *slot) 1032 { 1033 uint32_t i = slot->buf_idx; 1034 return (unlikely(i >= netmap_total_buffers)) ? NMB_VA(0) : NMB_VA(i); 1035 } 1036 1037 static inline void * 1038 PNMB(struct netmap_slot *slot, uint64_t *pp) 1039 { 1040 uint32_t i = slot->buf_idx; 1041 void *ret = (i >= netmap_total_buffers) ? NMB_VA(0) : NMB_VA(i); 1042 1043 *pp = (i >= netmap_total_buffers) ? NMB_PA(0) : NMB_PA(i); 1044 return ret; 1045 } 1046 1047 /* Generic version of NMB, which uses device-specific memory. */ 1048 static inline void * 1049 BDG_NMB(struct netmap_adapter *na, struct netmap_slot *slot) 1050 { 1051 struct lut_entry *lut = na->na_lut; 1052 uint32_t i = slot->buf_idx; 1053 return (unlikely(i >= na->na_lut_objtotal)) ? 1054 lut[0].vaddr : lut[i].vaddr; 1055 } 1056 1057 /* default functions to handle rx/tx interrupts */ 1058 int netmap_rx_irq(struct ifnet *, u_int, u_int *); 1059 #define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL) 1060 void netmap_common_irq(struct ifnet *, u_int, u_int *work_done); 1061 1062 1063 void netmap_txsync_to_host(struct netmap_adapter *na); 1064 void netmap_disable_all_rings(struct ifnet *); 1065 void netmap_enable_all_rings(struct ifnet *); 1066 void netmap_disable_ring(struct netmap_kring *kr); 1067 1068 1069 /* Structure associated to each thread which registered an interface. 1070 * 1071 * The first 4 fields of this structure are written by NIOCREGIF and 1072 * read by poll() and NIOC?XSYNC. 1073 * There is low contention among writers (actually, a correct user program 1074 * should have no contention among writers) and among writers and readers, 1075 * so we use a single global lock to protect the structure initialization. 1076 * Since initialization involves the allocation of memory, we reuse the memory 1077 * allocator lock. 1078 * Read access to the structure is lock free. Readers must check that 1079 * np_nifp is not NULL before using the other fields. 1080 * If np_nifp is NULL initialization has not been performed, so they should 1081 * return an error to userlevel. 1082 * 1083 * The ref_done field is used to regulate access to the refcount in the 1084 * memory allocator. The refcount must be incremented at most once for 1085 * each open("/dev/netmap"). The increment is performed by the first 1086 * function that calls netmap_get_memory() (currently called by 1087 * mmap(), NIOCGINFO and NIOCREGIF). 1088 * If the refcount is incremented, it is then decremented when the 1089 * private structure is destroyed. 1090 */ 1091 struct netmap_priv_d { 1092 struct netmap_if * volatile np_nifp; /* netmap if descriptor. */ 1093 1094 struct netmap_adapter *np_na; 1095 int np_ringid; /* from the ioctl */ 1096 u_int np_qfirst, np_qlast; /* range of rings to scan */ 1097 uint16_t np_txpoll; 1098 1099 struct netmap_mem_d *np_mref; /* use with NMG_LOCK held */ 1100 /* np_refcount is only used on FreeBSD */ 1101 int np_refcount; /* use with NMG_LOCK held */ 1102 }; 1103 1104 1105 /* 1106 * generic netmap emulation for devices that do not have 1107 * native netmap support. 1108 * XXX generic_netmap_register() is only exported to implement 1109 * nma_is_generic(). 1110 */ 1111 int generic_netmap_register(struct netmap_adapter *na, int enable); 1112 int generic_netmap_attach(struct ifnet *ifp); 1113 1114 int netmap_catch_rx(struct netmap_adapter *na, int intercept); 1115 void generic_rx_handler(struct ifnet *ifp, struct mbuf *m);; 1116 void netmap_catch_packet_steering(struct netmap_generic_adapter *na, int enable); 1117 int generic_xmit_frame(struct ifnet *ifp, struct mbuf *m, void *addr, u_int len, u_int ring_nr); 1118 int generic_find_num_desc(struct ifnet *ifp, u_int *tx, u_int *rx); 1119 void generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq); 1120 1121 static __inline int 1122 nma_is_generic(struct netmap_adapter *na) 1123 { 1124 return na->nm_register == generic_netmap_register; 1125 } 1126 1127 /* 1128 * netmap_mitigation API. This is used by the generic adapter 1129 * to reduce the number of interrupt requests/selwakeup 1130 * to clients on incoming packets. 1131 */ 1132 void netmap_mitigation_init(struct netmap_generic_adapter *na); 1133 void netmap_mitigation_start(struct netmap_generic_adapter *na); 1134 void netmap_mitigation_restart(struct netmap_generic_adapter *na); 1135 int netmap_mitigation_active(struct netmap_generic_adapter *na); 1136 void netmap_mitigation_cleanup(struct netmap_generic_adapter *na); 1137 1138 // int generic_timer_handler(struct hrtimer *t); 1139 1140 #endif /* _NET_NETMAP_KERN_H_ */ 1141