1 /* 2 * Copyright (C) 2014-2018 Vincenzo Maffione, Luigi Rizzo. 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 30 #include <net/netmap.h> 31 #include <sys/selinfo.h> 32 #include <vm/vm.h> 33 #include <vm/pmap.h> /* vtophys ? */ 34 #include <dev/netmap/netmap_kern.h> 35 36 /* Register and unregister. */ 37 static int 38 vtnet_netmap_reg(struct netmap_adapter *na, int state) 39 { 40 struct ifnet *ifp = na->ifp; 41 struct vtnet_softc *sc = ifp->if_softc; 42 43 /* 44 * Trigger a device reinit, asking vtnet_init_locked() to 45 * also enter or exit netmap mode. 46 */ 47 VTNET_CORE_LOCK(sc); 48 ifp->if_drv_flags &= ~IFF_DRV_RUNNING; 49 vtnet_init_locked(sc, state ? VTNET_INIT_NETMAP_ENTER 50 : VTNET_INIT_NETMAP_EXIT); 51 VTNET_CORE_UNLOCK(sc); 52 53 return 0; 54 } 55 56 57 /* Reconcile kernel and user view of the transmit ring. */ 58 static int 59 vtnet_netmap_txsync(struct netmap_kring *kring, int flags) 60 { 61 struct netmap_adapter *na = kring->na; 62 struct ifnet *ifp = na->ifp; 63 struct netmap_ring *ring = kring->ring; 64 u_int ring_nr = kring->ring_id; 65 u_int nm_i; /* index into the netmap ring */ 66 u_int const lim = kring->nkr_num_slots - 1; 67 u_int const head = kring->rhead; 68 69 /* device-specific */ 70 struct vtnet_softc *sc = ifp->if_softc; 71 struct vtnet_txq *txq = &sc->vtnet_txqs[ring_nr]; 72 struct virtqueue *vq = txq->vtntx_vq; 73 int interrupts = !(kring->nr_kflags & NKR_NOINTR); 74 u_int n; 75 76 /* 77 * First part: process new packets to send. 78 */ 79 80 nm_i = kring->nr_hwcur; 81 if (nm_i != head) { /* we have new packets to send */ 82 struct sglist *sg = txq->vtntx_sg; 83 84 for (; nm_i != head; nm_i = nm_next(nm_i, lim)) { 85 /* we use an empty header here */ 86 struct netmap_slot *slot = &ring->slot[nm_i]; 87 u_int len = slot->len; 88 uint64_t paddr; 89 void *addr = PNMB(na, slot, &paddr); 90 int err; 91 92 NM_CHECK_ADDR_LEN(na, addr, len); 93 94 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED); 95 /* Initialize the scatterlist, expose it to the hypervisor, 96 * and kick the hypervisor (if necessary). 97 */ 98 sglist_reset(sg); // cheap 99 err = sglist_append(sg, &txq->vtntx_shrhdr, sc->vtnet_hdr_size); 100 err |= sglist_append_phys(sg, paddr, len); 101 KASSERT(err == 0, ("%s: cannot append to sglist %d", 102 __func__, err)); 103 err = virtqueue_enqueue(vq, /*cookie=*/txq, sg, 104 /*readable=*/sg->sg_nseg, 105 /*writeable=*/0); 106 if (unlikely(err)) { 107 if (err != ENOSPC) 108 nm_prerr("virtqueue_enqueue(%s) failed: %d", 109 kring->name, err); 110 break; 111 } 112 } 113 114 virtqueue_notify(vq); 115 116 /* Update hwcur depending on where we stopped. */ 117 kring->nr_hwcur = nm_i; /* note we migth break early */ 118 } 119 120 /* Free used slots. We only consider our own used buffers, recognized 121 * by the token we passed to virtqueue_enqueue. 122 */ 123 n = 0; 124 for (;;) { 125 void *token = virtqueue_dequeue(vq, NULL); 126 if (token == NULL) 127 break; 128 if (unlikely(token != (void *)txq)) 129 nm_prerr("BUG: TX token mismatch"); 130 else 131 n++; 132 } 133 if (n > 0) { 134 kring->nr_hwtail += n; 135 if (kring->nr_hwtail > lim) 136 kring->nr_hwtail -= lim + 1; 137 } 138 139 if (interrupts && virtqueue_nfree(vq) < 32) 140 virtqueue_postpone_intr(vq, VQ_POSTPONE_LONG); 141 142 return 0; 143 } 144 145 /* 146 * Publish 'num 'netmap receive buffers to the host, starting 147 * from the next available one (rx->vtnrx_nm_refill). 148 * Return a positive error code on error, and 0 on success. 149 * If we could not publish all of the buffers that's an error, 150 * since the netmap ring and the virtqueue would go out of sync. 151 */ 152 static int 153 vtnet_netmap_kring_refill(struct netmap_kring *kring, u_int num) 154 { 155 struct netmap_adapter *na = kring->na; 156 struct ifnet *ifp = na->ifp; 157 struct netmap_ring *ring = kring->ring; 158 u_int ring_nr = kring->ring_id; 159 u_int const lim = kring->nkr_num_slots - 1; 160 u_int nm_i; 161 162 /* device-specific */ 163 struct vtnet_softc *sc = ifp->if_softc; 164 struct vtnet_rxq *rxq = &sc->vtnet_rxqs[ring_nr]; 165 struct virtqueue *vq = rxq->vtnrx_vq; 166 167 /* use a local sglist, default might be short */ 168 struct sglist_seg ss[2]; 169 struct sglist sg = { ss, 0, 0, 2 }; 170 171 for (nm_i = rxq->vtnrx_nm_refill; num > 0; 172 nm_i = nm_next(nm_i, lim), num--) { 173 struct netmap_slot *slot = &ring->slot[nm_i]; 174 uint64_t paddr; 175 void *addr = PNMB(na, slot, &paddr); 176 int err; 177 178 if (addr == NETMAP_BUF_BASE(na)) { /* bad buf */ 179 if (netmap_ring_reinit(kring)) 180 return EFAULT; 181 } 182 183 slot->flags &= ~NS_BUF_CHANGED; 184 sglist_reset(&sg); 185 err = sglist_append(&sg, &rxq->vtnrx_shrhdr, sc->vtnet_hdr_size); 186 err |= sglist_append_phys(&sg, paddr, NETMAP_BUF_SIZE(na)); 187 KASSERT(err == 0, ("%s: cannot append to sglist %d", 188 __func__, err)); 189 /* writable for the host */ 190 err = virtqueue_enqueue(vq, /*cookie=*/rxq, &sg, 191 /*readable=*/0, /*writeable=*/sg.sg_nseg); 192 if (unlikely(err)) { 193 nm_prerr("virtqueue_enqueue(%s) failed: %d", 194 kring->name, err); 195 break; 196 } 197 } 198 rxq->vtnrx_nm_refill = nm_i; 199 200 return num == 0 ? 0 : ENOSPC; 201 } 202 203 /* 204 * Publish netmap buffers on a RX virtqueue. 205 * Returns -1 if this virtqueue is not being opened in netmap mode. 206 * If the virtqueue is being opened in netmap mode, return 0 on success and 207 * a positive error code on failure. 208 */ 209 static int 210 vtnet_netmap_rxq_populate(struct vtnet_rxq *rxq) 211 { 212 struct netmap_adapter *na = NA(rxq->vtnrx_sc->vtnet_ifp); 213 struct netmap_kring *kring; 214 struct netmap_slot *slot; 215 int error; 216 217 slot = netmap_reset(na, NR_RX, rxq->vtnrx_id, 0); 218 if (slot == NULL) 219 return -1; 220 kring = na->rx_rings[rxq->vtnrx_id]; 221 222 /* Expose all the RX netmap buffers we can. In case of no indirect 223 * buffers, the number of netmap slots in the RX ring matches the 224 * maximum number of 2-elements sglist that the RX virtqueue can 225 * accommodate. We need to start from kring->nr_hwcur, which is 0 226 * on netmap register and may be different from 0 if a virtio 227 * re-init happens while the device is in use by netmap. */ 228 rxq->vtnrx_nm_refill = kring->nr_hwcur; 229 error = vtnet_netmap_kring_refill(kring, na->num_rx_desc - 1); 230 virtqueue_notify(rxq->vtnrx_vq); 231 232 return error; 233 } 234 235 /* Reconcile kernel and user view of the receive ring. */ 236 static int 237 vtnet_netmap_rxsync(struct netmap_kring *kring, int flags) 238 { 239 struct netmap_adapter *na = kring->na; 240 struct ifnet *ifp = na->ifp; 241 struct netmap_ring *ring = kring->ring; 242 u_int ring_nr = kring->ring_id; 243 u_int nm_i; /* index into the netmap ring */ 244 u_int const lim = kring->nkr_num_slots - 1; 245 u_int const head = kring->rhead; 246 int force_update = (flags & NAF_FORCE_READ) || 247 (kring->nr_kflags & NKR_PENDINTR); 248 int interrupts = !(kring->nr_kflags & NKR_NOINTR); 249 250 /* device-specific */ 251 struct vtnet_softc *sc = ifp->if_softc; 252 struct vtnet_rxq *rxq = &sc->vtnet_rxqs[ring_nr]; 253 struct virtqueue *vq = rxq->vtnrx_vq; 254 255 /* 256 * First part: import newly received packets. 257 * Only accept our own buffers (matching the token). We should only get 258 * matching buffers. The hwtail should never overrun hwcur, because 259 * we publish only N-1 receive buffers (and non N). 260 * In any case we must not leave this routine with the interrupts 261 * disabled, pending packets in the VQ and hwtail == (hwcur - 1), 262 * otherwise the pending packets could stall. 263 */ 264 if (netmap_no_pendintr || force_update) { 265 uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim); 266 void *token; 267 268 vtnet_rxq_disable_intr(rxq); 269 270 nm_i = kring->nr_hwtail; 271 for (;;) { 272 int len; 273 token = virtqueue_dequeue(vq, &len); 274 if (token == NULL) { 275 /* 276 * Enable the interrupts again and double-check 277 * for more work. We can go on until we win the 278 * race condition, since we are not replenishing 279 * in the meanwhile, and thus we will process at 280 * most N-1 slots. 281 */ 282 if (interrupts && vtnet_rxq_enable_intr(rxq)) { 283 vtnet_rxq_disable_intr(rxq); 284 continue; 285 } 286 break; 287 } 288 if (unlikely(token != (void *)rxq)) { 289 nm_prerr("BUG: RX token mismatch"); 290 } else { 291 if (nm_i == hwtail_lim) { 292 KASSERT(false, ("hwtail would " 293 "overrun hwcur")); 294 } 295 296 /* Skip the virtio-net header. */ 297 len -= sc->vtnet_hdr_size; 298 if (unlikely(len < 0)) { 299 nm_prlim(1, "Truncated virtio-net-header, " 300 "missing %d bytes", -len); 301 len = 0; 302 } 303 ring->slot[nm_i].len = len; 304 ring->slot[nm_i].flags = 0; 305 nm_i = nm_next(nm_i, lim); 306 } 307 } 308 kring->nr_hwtail = nm_i; 309 kring->nr_kflags &= ~NKR_PENDINTR; 310 } 311 312 /* 313 * Second part: skip past packets that userspace has released. 314 */ 315 nm_i = kring->nr_hwcur; /* netmap ring index */ 316 if (nm_i != head) { 317 int released; 318 int error; 319 320 released = head - nm_i; 321 if (released < 0) 322 released += kring->nkr_num_slots; 323 error = vtnet_netmap_kring_refill(kring, released); 324 if (error) { 325 nm_prerr("Failed to replenish RX VQ with %u sgs", 326 released); 327 return error; 328 } 329 kring->nr_hwcur = head; 330 virtqueue_notify(vq); 331 } 332 333 nm_prdis("h %d c %d t %d hwcur %d hwtail %d", kring->rhead, 334 kring->rcur, kring->rtail, kring->nr_hwcur, kring->nr_hwtail); 335 336 return 0; 337 } 338 339 340 /* Enable/disable interrupts on all virtqueues. */ 341 static void 342 vtnet_netmap_intr(struct netmap_adapter *na, int state) 343 { 344 struct vtnet_softc *sc = na->ifp->if_softc; 345 int i; 346 347 for (i = 0; i < sc->vtnet_max_vq_pairs; i++) { 348 struct vtnet_rxq *rxq = &sc->vtnet_rxqs[i]; 349 struct vtnet_txq *txq = &sc->vtnet_txqs[i]; 350 struct virtqueue *txvq = txq->vtntx_vq; 351 352 if (state) { 353 vtnet_rxq_enable_intr(rxq); 354 virtqueue_enable_intr(txvq); 355 } else { 356 vtnet_rxq_disable_intr(rxq); 357 virtqueue_disable_intr(txvq); 358 } 359 } 360 } 361 362 static int 363 vtnet_netmap_tx_slots(struct vtnet_softc *sc) 364 { 365 int div; 366 367 /* We need to prepend a virtio-net header to each netmap buffer to be 368 * transmitted, therefore calling virtqueue_enqueue() passing sglist 369 * with 2 elements. 370 * TX virtqueues use indirect descriptors if the feature was negotiated 371 * with the host, and if sc->vtnet_tx_nsegs > 1. With indirect 372 * descriptors, a single virtio descriptor is sufficient to reference 373 * each TX sglist. Without them, we need two separate virtio descriptors 374 * for each TX sglist. We therefore compute the number of netmap TX 375 * slots according to these assumptions. 376 */ 377 if ((sc->vtnet_flags & VTNET_FLAG_INDIRECT) && sc->vtnet_tx_nsegs > 1) 378 div = 1; 379 else 380 div = 2; 381 382 return virtqueue_size(sc->vtnet_txqs[0].vtntx_vq) / div; 383 } 384 385 static int 386 vtnet_netmap_rx_slots(struct vtnet_softc *sc) 387 { 388 int div; 389 390 /* We need to prepend a virtio-net header to each netmap buffer to be 391 * received, therefore calling virtqueue_enqueue() passing sglist 392 * with 2 elements. 393 * RX virtqueues use indirect descriptors if the feature was negotiated 394 * with the host, and if sc->vtnet_rx_nsegs > 1. With indirect 395 * descriptors, a single virtio descriptor is sufficient to reference 396 * each RX sglist. Without them, we need two separate virtio descriptors 397 * for each RX sglist. We therefore compute the number of netmap RX 398 * slots according to these assumptions. 399 */ 400 if ((sc->vtnet_flags & VTNET_FLAG_INDIRECT) && sc->vtnet_rx_nsegs > 1) 401 div = 1; 402 else 403 div = 2; 404 405 return virtqueue_size(sc->vtnet_rxqs[0].vtnrx_vq) / div; 406 } 407 408 static int 409 vtnet_netmap_config(struct netmap_adapter *na, struct nm_config_info *info) 410 { 411 struct vtnet_softc *sc = na->ifp->if_softc; 412 413 info->num_tx_rings = sc->vtnet_act_vq_pairs; 414 info->num_rx_rings = sc->vtnet_act_vq_pairs; 415 info->num_tx_descs = vtnet_netmap_tx_slots(sc); 416 info->num_rx_descs = vtnet_netmap_rx_slots(sc); 417 info->rx_buf_maxsize = NETMAP_BUF_SIZE(na); 418 419 return 0; 420 } 421 422 static void 423 vtnet_netmap_attach(struct vtnet_softc *sc) 424 { 425 struct netmap_adapter na; 426 427 bzero(&na, sizeof(na)); 428 429 na.ifp = sc->vtnet_ifp; 430 na.na_flags = 0; 431 na.num_tx_desc = vtnet_netmap_tx_slots(sc); 432 na.num_rx_desc = vtnet_netmap_rx_slots(sc); 433 na.num_tx_rings = na.num_rx_rings = sc->vtnet_max_vq_pairs; 434 na.rx_buf_maxsize = 0; 435 na.nm_register = vtnet_netmap_reg; 436 na.nm_txsync = vtnet_netmap_txsync; 437 na.nm_rxsync = vtnet_netmap_rxsync; 438 na.nm_intr = vtnet_netmap_intr; 439 na.nm_config = vtnet_netmap_config; 440 441 netmap_attach(&na); 442 443 nm_prinf("vtnet attached txq=%d, txd=%d rxq=%d, rxd=%d", 444 na.num_tx_rings, na.num_tx_desc, 445 na.num_tx_rings, na.num_rx_desc); 446 } 447 /* end of file */ 448