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
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.
40 * Some drivers may need the following headers. Others
41 * already include them by default
50 * device-specific sysctl variables:
52 * ix_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
53 * During regular operations the CRC is stripped, but on some
54 * hardware reception of frames not multiple of 64 is slower,
55 * so using crcstrip=0 helps in benchmarks.
57 * ix_rx_miss, ix_rx_miss_bufs:
58 * count packets that might be missed due to lost interrupts.
60 SYSCTL_DECL(_dev_netmap);
61 static int ix_rx_miss, ix_rx_miss_bufs;
63 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_crcstrip,
64 CTLFLAG_RW, &ix_crcstrip, 0, "strip CRC on rx frames");
65 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_rx_miss,
66 CTLFLAG_RW, &ix_rx_miss, 0, "potentially missed rx intr");
67 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_rx_miss_bufs,
68 CTLFLAG_RW, &ix_rx_miss_bufs, 0, "potentially missed rx intr bufs");
72 set_crcstrip(struct ixgbe_hw *hw, int onoff)
74 /* crc stripping is set in two places:
75 * IXGBE_HLREG0 (modified on init_locked and hw reset)
76 * IXGBE_RDRXCTL (set by the original driver in
77 * ixgbe_setup_hw_rsc() called in init_locked.
78 * We disable the setting when netmap is compiled in).
79 * We update the values here, but also in ixgbe.c because
80 * init_locked sometimes is called outside our control.
84 hl = IXGBE_READ_REG(hw, IXGBE_HLREG0);
85 rxc = IXGBE_READ_REG(hw, IXGBE_RDRXCTL);
87 D("%s read HLREG 0x%x rxc 0x%x",
88 onoff ? "enter" : "exit", hl, rxc);
89 /* hw requirements ... */
90 rxc &= ~IXGBE_RDRXCTL_RSCFRSTSIZE;
91 rxc |= IXGBE_RDRXCTL_RSCACKC;
92 if (onoff && !ix_crcstrip) {
93 /* keep the crc. Fast rx */
94 hl &= ~IXGBE_HLREG0_RXCRCSTRP;
95 rxc &= ~IXGBE_RDRXCTL_CRCSTRIP;
97 /* reset default mode */
98 hl |= IXGBE_HLREG0_RXCRCSTRP;
99 rxc |= IXGBE_RDRXCTL_CRCSTRIP;
102 D("%s write HLREG 0x%x rxc 0x%x",
103 onoff ? "enter" : "exit", hl, rxc);
104 IXGBE_WRITE_REG(hw, IXGBE_HLREG0, hl);
105 IXGBE_WRITE_REG(hw, IXGBE_RDRXCTL, rxc);
110 * Register/unregister. We are already under netmap lock.
111 * Only called on the first register or the last unregister.
114 ixgbe_netmap_reg(struct netmap_adapter *na, int onoff)
116 struct ifnet *ifp = na->ifp;
117 struct adapter *adapter = ifp->if_softc;
119 IXGBE_CORE_LOCK(adapter);
120 adapter->stop_locked(adapter);
122 set_crcstrip(&adapter->hw, onoff);
123 /* enable or disable flags and callbacks in na and ifp */
125 nm_set_native_flags(na);
127 nm_clear_native_flags(na);
129 adapter->init_locked(adapter); /* also enables intr */
130 set_crcstrip(&adapter->hw, onoff); // XXX why twice ?
131 IXGBE_CORE_UNLOCK(adapter);
132 return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
137 * Reconcile kernel and user view of the transmit ring.
139 * All information is in the kring.
140 * Userspace wants to send packets up to the one before kring->rhead,
141 * kernel knows kring->nr_hwcur is the first unsent packet.
143 * Here we push packets out (as many as possible), and possibly
144 * reclaim buffers from previously completed transmission.
146 * The caller (netmap) guarantees that there is only one instance
147 * running at any time. Any interference with other driver
148 * methods should be handled by the individual drivers.
151 ixgbe_netmap_txsync(struct netmap_kring *kring, int flags)
153 struct netmap_adapter *na = kring->na;
154 struct ifnet *ifp = na->ifp;
155 struct netmap_ring *ring = kring->ring;
156 u_int nm_i; /* index into the netmap ring */
157 u_int nic_i; /* index into the NIC ring */
159 u_int const lim = kring->nkr_num_slots - 1;
160 u_int const head = kring->rhead;
162 * interrupts on every tx packet are expensive so request
163 * them every half ring, or where NS_REPORT is set
165 u_int report_frequency = kring->nkr_num_slots >> 1;
167 /* device-specific */
168 struct adapter *adapter = ifp->if_softc;
169 struct tx_ring *txr = &adapter->tx_rings[kring->ring_id];
172 bus_dmamap_sync(txr->txdma.dma_tag, txr->txdma.dma_map,
173 BUS_DMASYNC_POSTREAD);
176 * First part: process new packets to send.
177 * nm_i is the current index in the netmap ring,
178 * nic_i is the corresponding index in the NIC ring.
179 * The two numbers differ because upon a *_init() we reset
180 * the NIC ring but leave the netmap ring unchanged.
181 * For the transmit ring, we have
183 * nm_i = kring->nr_hwcur
184 * nic_i = IXGBE_TDT (not tracked in the driver)
186 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
188 * In this driver kring->nkr_hwofs >= 0, but for other
189 * drivers it might be negative as well.
193 * If we have packets to send (kring->nr_hwcur != kring->rhead)
194 * iterate over the netmap ring, fetch length and update
195 * the corresponding slot in the NIC ring. Some drivers also
196 * need to update the buffer's physical address in the NIC slot
197 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
199 * The netmap_reload_map() calls is especially expensive,
200 * even when (as in this case) the tag is 0, so do only
201 * when the buffer has actually changed.
203 * If possible do not set the report/intr bit on all slots,
204 * but only a few times per ring or when NS_REPORT is set.
206 * Finally, on 10G and faster drivers, it might be useful
207 * to prefetch the next slot and txr entry.
210 nm_i = kring->nr_hwcur;
211 if (nm_i != head) { /* we have new packets to send */
212 nic_i = netmap_idx_k2n(kring, nm_i);
214 __builtin_prefetch(&ring->slot[nm_i]);
215 __builtin_prefetch(&txr->tx_buffers[nic_i]);
217 for (n = 0; nm_i != head; n++) {
218 struct netmap_slot *slot = &ring->slot[nm_i];
219 u_int len = slot->len;
221 void *addr = PNMB(na, slot, &paddr);
223 /* device-specific */
224 union ixgbe_adv_tx_desc *curr = &txr->tx_base[nic_i];
225 struct ixgbe_tx_buf *txbuf = &txr->tx_buffers[nic_i];
226 int flags = (slot->flags & NS_REPORT ||
227 nic_i == 0 || nic_i == report_frequency) ?
228 IXGBE_TXD_CMD_RS : 0;
230 /* prefetch for next round */
231 __builtin_prefetch(&ring->slot[nm_i + 1]);
232 __builtin_prefetch(&txr->tx_buffers[nic_i + 1]);
234 NM_CHECK_ADDR_LEN(na, addr, len);
236 if (slot->flags & NS_BUF_CHANGED) {
237 /* buffer has changed, reload map */
238 netmap_reload_map(na, txr->txtag, txbuf->map, addr);
240 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
242 /* Fill the slot in the NIC ring. */
243 /* Use legacy descriptor, they are faster? */
244 curr->read.buffer_addr = htole64(paddr);
245 curr->read.olinfo_status = 0;
246 curr->read.cmd_type_len = htole32(len | flags |
247 IXGBE_ADVTXD_DCMD_IFCS | IXGBE_TXD_CMD_EOP);
249 /* make sure changes to the buffer are synced */
250 bus_dmamap_sync(txr->txtag, txbuf->map,
251 BUS_DMASYNC_PREWRITE);
253 nm_i = nm_next(nm_i, lim);
254 nic_i = nm_next(nic_i, lim);
256 kring->nr_hwcur = head;
258 /* synchronize the NIC ring */
259 bus_dmamap_sync(txr->txdma.dma_tag, txr->txdma.dma_map,
260 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
262 /* (re)start the tx unit up to slot nic_i (excluded) */
263 IXGBE_WRITE_REG(&adapter->hw, txr->tail, nic_i);
267 * Second part: reclaim buffers for completed transmissions.
268 * Because this is expensive (we read a NIC register etc.)
269 * we only do it in specific cases (see below).
271 if (flags & NAF_FORCE_RECLAIM) {
272 reclaim_tx = 1; /* forced reclaim */
273 } else if (!nm_kr_txempty(kring)) {
274 reclaim_tx = 0; /* have buffers, no reclaim */
277 * No buffers available. Locate previous slot with
279 * If the slot has DD set, we can reclaim space,
280 * otherwise wait for the next interrupt.
281 * This enables interrupt moderation on the tx
282 * side though it might reduce throughput.
284 struct ixgbe_legacy_tx_desc *txd =
285 (struct ixgbe_legacy_tx_desc *)txr->tx_base;
287 nic_i = txr->next_to_clean + report_frequency;
290 // round to the closest with dd set
291 nic_i = (nic_i < kring->nkr_num_slots / 4 ||
292 nic_i >= kring->nkr_num_slots*3/4) ?
293 0 : report_frequency;
294 reclaim_tx = txd[nic_i].upper.fields.status & IXGBE_TXD_STAT_DD; // XXX cpu_to_le32 ?
298 * Record completed transmissions.
299 * We (re)use the driver's txr->next_to_clean to keep
300 * track of the most recently completed transmission.
302 * The datasheet discourages the use of TDH to find
303 * out the number of sent packets, but we only set
304 * REPORT_STATUS in a few slots so TDH is the only
307 nic_i = IXGBE_READ_REG(&adapter->hw, IXGBE_TDH(kring->ring_id));
308 if (nic_i >= kring->nkr_num_slots) { /* XXX can it happen ? */
309 D("TDH wrap %d", nic_i);
310 nic_i -= kring->nkr_num_slots;
312 if (nic_i != txr->next_to_clean) {
313 /* some tx completed, increment avail */
314 txr->next_to_clean = nic_i;
315 kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
319 nm_txsync_finalize(kring);
326 * Reconcile kernel and user view of the receive ring.
327 * Same as for the txsync, this routine must be efficient.
328 * The caller guarantees a single invocations, but races against
329 * the rest of the driver should be handled here.
331 * On call, kring->rhead is the first packet that userspace wants
332 * to keep, and kring->rcur is the wakeup point.
333 * The kernel has previously reported packets up to kring->rtail.
335 * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
336 * of whether or not we received an interrupt.
339 ixgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
341 struct netmap_adapter *na = kring->na;
342 struct ifnet *ifp = na->ifp;
343 struct netmap_ring *ring = kring->ring;
344 u_int nm_i; /* index into the netmap ring */
345 u_int nic_i; /* index into the NIC ring */
347 u_int const lim = kring->nkr_num_slots - 1;
348 u_int const head = nm_rxsync_prologue(kring);
349 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
351 /* device-specific */
352 struct adapter *adapter = ifp->if_softc;
353 struct rx_ring *rxr = &adapter->rx_rings[kring->ring_id];
356 return netmap_ring_reinit(kring);
358 /* XXX check sync modes */
359 bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map,
360 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
363 * First part: import newly received packets.
365 * nm_i is the index of the next free slot in the netmap ring,
366 * nic_i is the index of the next received packet in the NIC ring,
367 * and they may differ in case if_init() has been called while
368 * in netmap mode. For the receive ring we have
370 * nic_i = rxr->next_to_check;
371 * nm_i = kring->nr_hwtail (previous)
373 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
375 * rxr->next_to_check is set to 0 on a ring reinit
377 if (netmap_no_pendintr || force_update) {
378 int crclen = (ix_crcstrip) ? 0 : 4;
379 uint16_t slot_flags = kring->nkr_slot_flags;
381 nic_i = rxr->next_to_check; // or also k2n(kring->nr_hwtail)
382 nm_i = netmap_idx_n2k(kring, nic_i);
385 union ixgbe_adv_rx_desc *curr = &rxr->rx_base[nic_i];
386 uint32_t staterr = le32toh(curr->wb.upper.status_error);
388 if ((staterr & IXGBE_RXD_STAT_DD) == 0)
390 ring->slot[nm_i].len = le16toh(curr->wb.upper.length) - crclen;
391 ring->slot[nm_i].flags = slot_flags;
392 bus_dmamap_sync(rxr->ptag,
393 rxr->rx_buffers[nic_i].pmap, BUS_DMASYNC_POSTREAD);
394 nm_i = nm_next(nm_i, lim);
395 nic_i = nm_next(nic_i, lim);
397 if (n) { /* update the state variables */
398 if (netmap_no_pendintr && !force_update) {
401 ix_rx_miss_bufs += n;
403 rxr->next_to_check = nic_i;
404 kring->nr_hwtail = nm_i;
406 kring->nr_kflags &= ~NKR_PENDINTR;
410 * Second part: skip past packets that userspace has released.
411 * (kring->nr_hwcur to kring->rhead excluded),
412 * and make the buffers available for reception.
413 * As usual nm_i is the index in the netmap ring,
414 * nic_i is the index in the NIC ring, and
415 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
417 nm_i = kring->nr_hwcur;
419 nic_i = netmap_idx_k2n(kring, nm_i);
420 for (n = 0; nm_i != head; n++) {
421 struct netmap_slot *slot = &ring->slot[nm_i];
423 void *addr = PNMB(na, slot, &paddr);
425 union ixgbe_adv_rx_desc *curr = &rxr->rx_base[nic_i];
426 struct ixgbe_rx_buf *rxbuf = &rxr->rx_buffers[nic_i];
428 if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
431 if (slot->flags & NS_BUF_CHANGED) {
432 /* buffer has changed, reload map */
433 netmap_reload_map(na, rxr->ptag, rxbuf->pmap, addr);
434 slot->flags &= ~NS_BUF_CHANGED;
436 curr->wb.upper.status_error = 0;
437 curr->read.pkt_addr = htole64(paddr);
438 bus_dmamap_sync(rxr->ptag, rxbuf->pmap,
439 BUS_DMASYNC_PREREAD);
440 nm_i = nm_next(nm_i, lim);
441 nic_i = nm_next(nic_i, lim);
443 kring->nr_hwcur = head;
445 bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map,
446 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
448 * IMPORTANT: we must leave one free slot in the ring,
449 * so move nic_i back by one unit
451 nic_i = nm_prev(nic_i, lim);
452 IXGBE_WRITE_REG(&adapter->hw, rxr->tail, nic_i);
455 /* tell userspace that there might be new packets */
456 nm_rxsync_finalize(kring);
461 return netmap_ring_reinit(kring);
466 * The attach routine, called near the end of ixgbe_attach(),
467 * fills the parameters for netmap_attach() and calls it.
468 * It cannot fail, in the worst case (such as no memory)
469 * netmap mode will be disabled and the driver will only
470 * operate in standard mode.
473 ixgbe_netmap_attach(struct adapter *adapter)
475 struct netmap_adapter na;
477 bzero(&na, sizeof(na));
479 na.ifp = adapter->ifp;
480 na.na_flags = NAF_BDG_MAYSLEEP;
481 na.num_tx_desc = adapter->num_tx_desc;
482 na.num_rx_desc = adapter->num_rx_desc;
483 na.nm_txsync = ixgbe_netmap_txsync;
484 na.nm_rxsync = ixgbe_netmap_rxsync;
485 na.nm_register = ixgbe_netmap_reg;
486 na.num_tx_rings = na.num_rx_rings = adapter->num_queues;
490 #endif /* DEV_NETMAP */