2 * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
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.
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
29 * netmap support for: ixgbe (both ix and ixv)
31 * This file is meant to be a reference on how to implement
32 * netmap support for a network driver.
33 * This file contains code but only static or inline functions used
34 * by a single driver. To avoid replication of code we just #include
35 * it near the beginning of the standard driver.
39 #include <net/netmap.h>
40 #include <sys/selinfo.h>
42 * Some drivers may need the following headers. Others
43 * already include them by default
49 #include <dev/netmap/netmap_kern.h>
51 void ixgbe_netmap_attach(struct adapter *adapter);
54 * device-specific sysctl variables:
56 * ix_crcstrip: 0: NIC keeps CRC in rx frames (default), 1: NIC strips it.
57 * During regular operations the CRC is stripped, but on some
58 * hardware reception of frames not multiple of 64 is slower,
59 * so using crcstrip=0 helps in benchmarks.
61 * ix_rx_miss, ix_rx_miss_bufs:
62 * count packets that might be missed due to lost interrupts.
64 SYSCTL_DECL(_dev_netmap);
65 static int ix_rx_miss, ix_rx_miss_bufs;
67 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_crcstrip,
68 CTLFLAG_RW, &ix_crcstrip, 0, "NIC strips CRC on rx frames");
69 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_rx_miss,
70 CTLFLAG_RW, &ix_rx_miss, 0, "potentially missed rx intr");
71 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_rx_miss_bufs,
72 CTLFLAG_RW, &ix_rx_miss_bufs, 0, "potentially missed rx intr bufs");
76 set_crcstrip(struct ixgbe_hw *hw, int onoff)
78 /* crc stripping is set in two places:
79 * IXGBE_HLREG0 (modified on init_locked and hw reset)
80 * IXGBE_RDRXCTL (set by the original driver in
81 * ixgbe_setup_hw_rsc() called in init_locked.
82 * We disable the setting when netmap is compiled in).
83 * We update the values here, but also in ixgbe.c because
84 * init_locked sometimes is called outside our control.
88 hl = IXGBE_READ_REG(hw, IXGBE_HLREG0);
89 rxc = IXGBE_READ_REG(hw, IXGBE_RDRXCTL);
91 D("%s read HLREG 0x%x rxc 0x%x",
92 onoff ? "enter" : "exit", hl, rxc);
93 /* hw requirements ... */
94 rxc &= ~IXGBE_RDRXCTL_RSCFRSTSIZE;
95 rxc |= IXGBE_RDRXCTL_RSCACKC;
96 if (onoff && !ix_crcstrip) {
97 /* keep the crc. Fast rx */
98 hl &= ~IXGBE_HLREG0_RXCRCSTRP;
99 rxc &= ~IXGBE_RDRXCTL_CRCSTRIP;
101 /* reset default mode */
102 hl |= IXGBE_HLREG0_RXCRCSTRP;
103 rxc |= IXGBE_RDRXCTL_CRCSTRIP;
106 D("%s write HLREG 0x%x rxc 0x%x",
107 onoff ? "enter" : "exit", hl, rxc);
108 IXGBE_WRITE_REG(hw, IXGBE_HLREG0, hl);
109 IXGBE_WRITE_REG(hw, IXGBE_RDRXCTL, rxc);
113 ixgbe_netmap_intr(struct netmap_adapter *na, int onoff)
115 struct ifnet *ifp = na->ifp;
116 struct adapter *adapter = ifp->if_softc;
118 IXGBE_CORE_LOCK(adapter);
120 ixgbe_enable_intr(adapter); // XXX maybe ixgbe_stop ?
122 ixgbe_disable_intr(adapter); // XXX maybe ixgbe_stop ?
124 IXGBE_CORE_UNLOCK(adapter);
128 * Register/unregister. We are already under netmap lock.
129 * Only called on the first register or the last unregister.
132 ixgbe_netmap_reg(struct netmap_adapter *na, int onoff)
134 struct ifnet *ifp = na->ifp;
135 struct adapter *adapter = ifp->if_softc;
137 IXGBE_CORE_LOCK(adapter);
138 adapter->stop_locked(adapter);
140 if (!IXGBE_IS_VF(adapter))
141 set_crcstrip(&adapter->hw, onoff);
142 /* enable or disable flags and callbacks in na and ifp */
144 nm_set_native_flags(na);
146 nm_clear_native_flags(na);
148 adapter->init_locked(adapter); /* also enables intr */
149 if (!IXGBE_IS_VF(adapter))
150 set_crcstrip(&adapter->hw, onoff); // XXX why twice ?
151 IXGBE_CORE_UNLOCK(adapter);
152 return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
157 * Reconcile kernel and user view of the transmit ring.
159 * All information is in the kring.
160 * Userspace wants to send packets up to the one before kring->rhead,
161 * kernel knows kring->nr_hwcur is the first unsent packet.
163 * Here we push packets out (as many as possible), and possibly
164 * reclaim buffers from previously completed transmission.
166 * The caller (netmap) guarantees that there is only one instance
167 * running at any time. Any interference with other driver
168 * methods should be handled by the individual drivers.
171 ixgbe_netmap_txsync(struct netmap_kring *kring, int flags)
173 struct netmap_adapter *na = kring->na;
174 struct ifnet *ifp = na->ifp;
175 struct netmap_ring *ring = kring->ring;
176 u_int nm_i; /* index into the netmap ring */
177 u_int nic_i; /* index into the NIC ring */
179 u_int const lim = kring->nkr_num_slots - 1;
180 u_int const head = kring->rhead;
182 * interrupts on every tx packet are expensive so request
183 * them every half ring, or where NS_REPORT is set
185 u_int report_frequency = kring->nkr_num_slots >> 1;
187 /* device-specific */
188 struct adapter *adapter = ifp->if_softc;
189 struct tx_ring *txr = &adapter->tx_rings[kring->ring_id];
192 bus_dmamap_sync(txr->txdma.dma_tag, txr->txdma.dma_map,
193 BUS_DMASYNC_POSTREAD);
196 * First part: process new packets to send.
197 * nm_i is the current index in the netmap ring,
198 * nic_i is the corresponding index in the NIC ring.
199 * The two numbers differ because upon a *_init() we reset
200 * the NIC ring but leave the netmap ring unchanged.
201 * For the transmit ring, we have
203 * nm_i = kring->nr_hwcur
204 * nic_i = IXGBE_TDT (not tracked in the driver)
206 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
208 * In this driver kring->nkr_hwofs >= 0, but for other
209 * drivers it might be negative as well.
213 * If we have packets to send (kring->nr_hwcur != kring->rhead)
214 * iterate over the netmap ring, fetch length and update
215 * the corresponding slot in the NIC ring. Some drivers also
216 * need to update the buffer's physical address in the NIC slot
217 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
219 * The netmap_reload_map() calls is especially expensive,
220 * even when (as in this case) the tag is 0, so do only
221 * when the buffer has actually changed.
223 * If possible do not set the report/intr bit on all slots,
224 * but only a few times per ring or when NS_REPORT is set.
226 * Finally, on 10G and faster drivers, it might be useful
227 * to prefetch the next slot and txr entry.
230 nm_i = kring->nr_hwcur;
231 if (nm_i != head) { /* we have new packets to send */
232 nic_i = netmap_idx_k2n(kring, nm_i);
234 __builtin_prefetch(&ring->slot[nm_i]);
235 __builtin_prefetch(&txr->tx_buffers[nic_i]);
237 for (n = 0; nm_i != head; n++) {
238 struct netmap_slot *slot = &ring->slot[nm_i];
239 u_int len = slot->len;
241 void *addr = PNMB(na, slot, &paddr);
243 /* device-specific */
244 union ixgbe_adv_tx_desc *curr = &txr->tx_base[nic_i];
245 struct ixgbe_tx_buf *txbuf = &txr->tx_buffers[nic_i];
246 int flags = (slot->flags & NS_REPORT ||
247 nic_i == 0 || nic_i == report_frequency) ?
248 IXGBE_TXD_CMD_RS : 0;
250 /* prefetch for next round */
251 __builtin_prefetch(&ring->slot[nm_i + 1]);
252 __builtin_prefetch(&txr->tx_buffers[nic_i + 1]);
254 NM_CHECK_ADDR_LEN(na, addr, len);
256 if (slot->flags & NS_BUF_CHANGED) {
257 /* buffer has changed, reload map */
258 netmap_reload_map(na, txr->txtag, txbuf->map, addr);
260 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
262 /* Fill the slot in the NIC ring. */
263 /* Use legacy descriptor, they are faster? */
264 curr->read.buffer_addr = htole64(paddr);
265 curr->read.olinfo_status = 0;
266 curr->read.cmd_type_len = htole32(len | flags |
267 IXGBE_ADVTXD_DCMD_IFCS | IXGBE_TXD_CMD_EOP);
269 /* make sure changes to the buffer are synced */
270 bus_dmamap_sync(txr->txtag, txbuf->map,
271 BUS_DMASYNC_PREWRITE);
273 nm_i = nm_next(nm_i, lim);
274 nic_i = nm_next(nic_i, lim);
276 kring->nr_hwcur = head;
278 /* synchronize the NIC ring */
279 bus_dmamap_sync(txr->txdma.dma_tag, txr->txdma.dma_map,
280 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
282 /* (re)start the tx unit up to slot nic_i (excluded) */
283 IXGBE_WRITE_REG(&adapter->hw, txr->tail, nic_i);
287 * Second part: reclaim buffers for completed transmissions.
288 * Because this is expensive (we read a NIC register etc.)
289 * we only do it in specific cases (see below).
291 if (flags & NAF_FORCE_RECLAIM) {
292 reclaim_tx = 1; /* forced reclaim */
293 } else if (!nm_kr_txempty(kring)) {
294 reclaim_tx = 0; /* have buffers, no reclaim */
297 * No buffers available. Locate previous slot with
299 * If the slot has DD set, we can reclaim space,
300 * otherwise wait for the next interrupt.
301 * This enables interrupt moderation on the tx
302 * side though it might reduce throughput.
304 struct ixgbe_legacy_tx_desc *txd =
305 (struct ixgbe_legacy_tx_desc *)txr->tx_base;
307 nic_i = txr->next_to_clean + report_frequency;
310 // round to the closest with dd set
311 nic_i = (nic_i < kring->nkr_num_slots / 4 ||
312 nic_i >= kring->nkr_num_slots*3/4) ?
313 0 : report_frequency;
314 reclaim_tx = txd[nic_i].upper.fields.status & IXGBE_TXD_STAT_DD; // XXX cpu_to_le32 ?
318 * Record completed transmissions.
319 * We (re)use the driver's txr->next_to_clean to keep
320 * track of the most recently completed transmission.
322 * The datasheet discourages the use of TDH to find
323 * out the number of sent packets, but we only set
324 * REPORT_STATUS in a few slots so TDH is the only
327 nic_i = IXGBE_READ_REG(&adapter->hw, IXGBE_IS_VF(adapter) ?
328 IXGBE_VFTDH(kring->ring_id) : IXGBE_TDH(kring->ring_id));
329 if (nic_i >= kring->nkr_num_slots) { /* XXX can it happen ? */
330 D("TDH wrap %d", nic_i);
331 nic_i -= kring->nkr_num_slots;
333 if (nic_i != txr->next_to_clean) {
334 /* some tx completed, increment avail */
335 txr->next_to_clean = nic_i;
336 kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
345 * Reconcile kernel and user view of the receive ring.
346 * Same as for the txsync, this routine must be efficient.
347 * The caller guarantees a single invocations, but races against
348 * the rest of the driver should be handled here.
350 * On call, kring->rhead is the first packet that userspace wants
351 * to keep, and kring->rcur is the wakeup point.
352 * The kernel has previously reported packets up to kring->rtail.
354 * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
355 * of whether or not we received an interrupt.
358 ixgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
360 struct netmap_adapter *na = kring->na;
361 struct ifnet *ifp = na->ifp;
362 struct netmap_ring *ring = kring->ring;
363 u_int nm_i; /* index into the netmap ring */
364 u_int nic_i; /* index into the NIC ring */
366 u_int const lim = kring->nkr_num_slots - 1;
367 u_int const head = kring->rhead;
368 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
370 /* device-specific */
371 struct adapter *adapter = ifp->if_softc;
372 struct rx_ring *rxr = &adapter->rx_rings[kring->ring_id];
375 return netmap_ring_reinit(kring);
377 /* XXX check sync modes */
378 bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map,
379 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
382 * First part: import newly received packets.
384 * nm_i is the index of the next free slot in the netmap ring,
385 * nic_i is the index of the next received packet in the NIC ring,
386 * and they may differ in case if_init() has been called while
387 * in netmap mode. For the receive ring we have
389 * nic_i = rxr->next_to_check;
390 * nm_i = kring->nr_hwtail (previous)
392 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
394 * rxr->next_to_check is set to 0 on a ring reinit
396 if (netmap_no_pendintr || force_update) {
397 int crclen = (ix_crcstrip || IXGBE_IS_VF(adapter) ) ? 0 : 4;
398 uint16_t slot_flags = kring->nkr_slot_flags;
400 nic_i = rxr->next_to_check; // or also k2n(kring->nr_hwtail)
401 nm_i = netmap_idx_n2k(kring, nic_i);
404 union ixgbe_adv_rx_desc *curr = &rxr->rx_base[nic_i];
405 uint32_t staterr = le32toh(curr->wb.upper.status_error);
407 if ((staterr & IXGBE_RXD_STAT_DD) == 0)
409 ring->slot[nm_i].len = le16toh(curr->wb.upper.length) - crclen;
410 ring->slot[nm_i].flags = slot_flags;
411 bus_dmamap_sync(rxr->ptag,
412 rxr->rx_buffers[nic_i].pmap, BUS_DMASYNC_POSTREAD);
413 nm_i = nm_next(nm_i, lim);
414 nic_i = nm_next(nic_i, lim);
416 if (n) { /* update the state variables */
417 if (netmap_no_pendintr && !force_update) {
420 ix_rx_miss_bufs += n;
422 rxr->next_to_check = nic_i;
423 kring->nr_hwtail = nm_i;
425 kring->nr_kflags &= ~NKR_PENDINTR;
429 * Second part: skip past packets that userspace has released.
430 * (kring->nr_hwcur to kring->rhead excluded),
431 * and make the buffers available for reception.
432 * As usual nm_i is the index in the netmap ring,
433 * nic_i is the index in the NIC ring, and
434 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
436 nm_i = kring->nr_hwcur;
438 nic_i = netmap_idx_k2n(kring, nm_i);
439 for (n = 0; nm_i != head; n++) {
440 struct netmap_slot *slot = &ring->slot[nm_i];
442 void *addr = PNMB(na, slot, &paddr);
444 union ixgbe_adv_rx_desc *curr = &rxr->rx_base[nic_i];
445 struct ixgbe_rx_buf *rxbuf = &rxr->rx_buffers[nic_i];
447 if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
450 if (slot->flags & NS_BUF_CHANGED) {
451 /* buffer has changed, reload map */
452 netmap_reload_map(na, rxr->ptag, rxbuf->pmap, addr);
453 slot->flags &= ~NS_BUF_CHANGED;
455 curr->wb.upper.status_error = 0;
456 curr->read.pkt_addr = htole64(paddr);
457 bus_dmamap_sync(rxr->ptag, rxbuf->pmap,
458 BUS_DMASYNC_PREREAD);
459 nm_i = nm_next(nm_i, lim);
460 nic_i = nm_next(nic_i, lim);
462 kring->nr_hwcur = head;
464 bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map,
465 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
467 * IMPORTANT: we must leave one free slot in the ring,
468 * so move nic_i back by one unit
470 nic_i = nm_prev(nic_i, lim);
471 IXGBE_WRITE_REG(&adapter->hw, rxr->tail, nic_i);
477 return netmap_ring_reinit(kring);
482 * The attach routine, called near the end of ixgbe_attach(),
483 * fills the parameters for netmap_attach() and calls it.
484 * It cannot fail, in the worst case (such as no memory)
485 * netmap mode will be disabled and the driver will only
486 * operate in standard mode.
489 ixgbe_netmap_attach(struct adapter *adapter)
491 struct netmap_adapter na;
493 bzero(&na, sizeof(na));
495 na.ifp = adapter->ifp;
496 na.na_flags = NAF_BDG_MAYSLEEP;
497 na.num_tx_desc = adapter->num_tx_desc;
498 na.num_rx_desc = adapter->num_rx_desc;
499 na.nm_txsync = ixgbe_netmap_txsync;
500 na.nm_rxsync = ixgbe_netmap_rxsync;
501 na.nm_register = ixgbe_netmap_reg;
502 na.num_tx_rings = na.num_rx_rings = adapter->num_queues;
503 na.nm_intr = ixgbe_netmap_intr;