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.
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>
53 * device-specific sysctl variables:
55 * ix_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
56 * During regular operations the CRC is stripped, but on some
57 * hardware reception of frames not multiple of 64 is slower,
58 * so using crcstrip=0 helps in benchmarks.
60 * ix_rx_miss, ix_rx_miss_bufs:
61 * count packets that might be missed due to lost interrupts.
63 SYSCTL_DECL(_dev_netmap);
64 static int ix_rx_miss, ix_rx_miss_bufs;
66 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_crcstrip,
67 CTLFLAG_RW, &ix_crcstrip, 0, "strip CRC on rx frames");
68 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_rx_miss,
69 CTLFLAG_RW, &ix_rx_miss, 0, "potentially missed rx intr");
70 SYSCTL_INT(_dev_netmap, OID_AUTO, ix_rx_miss_bufs,
71 CTLFLAG_RW, &ix_rx_miss_bufs, 0, "potentially missed rx intr bufs");
75 set_crcstrip(struct ixgbe_hw *hw, int onoff)
77 /* crc stripping is set in two places:
78 * IXGBE_HLREG0 (modified on init_locked and hw reset)
79 * IXGBE_RDRXCTL (set by the original driver in
80 * ixgbe_setup_hw_rsc() called in init_locked.
81 * We disable the setting when netmap is compiled in).
82 * We update the values here, but also in ixgbe.c because
83 * init_locked sometimes is called outside our control.
87 hl = IXGBE_READ_REG(hw, IXGBE_HLREG0);
88 rxc = IXGBE_READ_REG(hw, IXGBE_RDRXCTL);
90 D("%s read HLREG 0x%x rxc 0x%x",
91 onoff ? "enter" : "exit", hl, rxc);
92 /* hw requirements ... */
93 rxc &= ~IXGBE_RDRXCTL_RSCFRSTSIZE;
94 rxc |= IXGBE_RDRXCTL_RSCACKC;
95 if (onoff && !ix_crcstrip) {
96 /* keep the crc. Fast rx */
97 hl &= ~IXGBE_HLREG0_RXCRCSTRP;
98 rxc &= ~IXGBE_RDRXCTL_CRCSTRIP;
100 /* reset default mode */
101 hl |= IXGBE_HLREG0_RXCRCSTRP;
102 rxc |= IXGBE_RDRXCTL_CRCSTRIP;
105 D("%s write HLREG 0x%x rxc 0x%x",
106 onoff ? "enter" : "exit", hl, rxc);
107 IXGBE_WRITE_REG(hw, IXGBE_HLREG0, hl);
108 IXGBE_WRITE_REG(hw, IXGBE_RDRXCTL, rxc);
113 * Register/unregister. We are already under netmap lock.
114 * Only called on the first register or the last unregister.
117 ixgbe_netmap_reg(struct netmap_adapter *na, int onoff)
119 struct ifnet *ifp = na->ifp;
120 struct adapter *adapter = ifp->if_softc;
122 IXGBE_CORE_LOCK(adapter);
123 ixgbe_disable_intr(adapter); // XXX maybe ixgbe_stop ?
125 /* Tell the stack that the interface is no longer active */
126 ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
128 set_crcstrip(&adapter->hw, onoff);
129 /* enable or disable flags and callbacks in na and ifp */
131 nm_set_native_flags(na);
133 nm_clear_native_flags(na);
135 ixgbe_init_locked(adapter); /* also enables intr */
136 set_crcstrip(&adapter->hw, onoff); // XXX why twice ?
137 IXGBE_CORE_UNLOCK(adapter);
138 return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
143 * Reconcile kernel and user view of the transmit ring.
145 * All information is in the kring.
146 * Userspace wants to send packets up to the one before kring->rhead,
147 * kernel knows kring->nr_hwcur is the first unsent packet.
149 * Here we push packets out (as many as possible), and possibly
150 * reclaim buffers from previously completed transmission.
152 * The caller (netmap) guarantees that there is only one instance
153 * running at any time. Any interference with other driver
154 * methods should be handled by the individual drivers.
157 ixgbe_netmap_txsync(struct netmap_kring *kring, int flags)
159 struct netmap_adapter *na = kring->na;
160 struct ifnet *ifp = na->ifp;
161 struct netmap_ring *ring = kring->ring;
162 u_int nm_i; /* index into the netmap ring */
163 u_int nic_i; /* index into the NIC ring */
165 u_int const lim = kring->nkr_num_slots - 1;
166 u_int const head = kring->rhead;
168 * interrupts on every tx packet are expensive so request
169 * them every half ring, or where NS_REPORT is set
171 u_int report_frequency = kring->nkr_num_slots >> 1;
173 /* device-specific */
174 struct adapter *adapter = ifp->if_softc;
175 struct tx_ring *txr = &adapter->tx_rings[kring->ring_id];
178 bus_dmamap_sync(txr->txdma.dma_tag, txr->txdma.dma_map,
179 BUS_DMASYNC_POSTREAD);
182 * First part: process new packets to send.
183 * nm_i is the current index in the netmap ring,
184 * nic_i is the corresponding index in the NIC ring.
185 * The two numbers differ because upon a *_init() we reset
186 * the NIC ring but leave the netmap ring unchanged.
187 * For the transmit ring, we have
189 * nm_i = kring->nr_hwcur
190 * nic_i = IXGBE_TDT (not tracked in the driver)
192 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
194 * In this driver kring->nkr_hwofs >= 0, but for other
195 * drivers it might be negative as well.
199 * If we have packets to send (kring->nr_hwcur != kring->rhead)
200 * iterate over the netmap ring, fetch length and update
201 * the corresponding slot in the NIC ring. Some drivers also
202 * need to update the buffer's physical address in the NIC slot
203 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
205 * The netmap_reload_map() calls is especially expensive,
206 * even when (as in this case) the tag is 0, so do only
207 * when the buffer has actually changed.
209 * If possible do not set the report/intr bit on all slots,
210 * but only a few times per ring or when NS_REPORT is set.
212 * Finally, on 10G and faster drivers, it might be useful
213 * to prefetch the next slot and txr entry.
216 nm_i = kring->nr_hwcur;
217 if (nm_i != head) { /* we have new packets to send */
218 nic_i = netmap_idx_k2n(kring, nm_i);
220 __builtin_prefetch(&ring->slot[nm_i]);
221 __builtin_prefetch(&txr->tx_buffers[nic_i]);
223 for (n = 0; nm_i != head; n++) {
224 struct netmap_slot *slot = &ring->slot[nm_i];
225 u_int len = slot->len;
227 void *addr = PNMB(na, slot, &paddr);
229 /* device-specific */
230 union ixgbe_adv_tx_desc *curr = &txr->tx_base[nic_i];
231 struct ixgbe_tx_buf *txbuf = &txr->tx_buffers[nic_i];
232 int flags = (slot->flags & NS_REPORT ||
233 nic_i == 0 || nic_i == report_frequency) ?
234 IXGBE_TXD_CMD_RS : 0;
236 /* prefetch for next round */
237 __builtin_prefetch(&ring->slot[nm_i + 1]);
238 __builtin_prefetch(&txr->tx_buffers[nic_i + 1]);
240 NM_CHECK_ADDR_LEN(na, addr, len);
242 if (slot->flags & NS_BUF_CHANGED) {
243 /* buffer has changed, reload map */
244 netmap_reload_map(na, txr->txtag, txbuf->map, addr);
246 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
248 /* Fill the slot in the NIC ring. */
249 /* Use legacy descriptor, they are faster? */
250 curr->read.buffer_addr = htole64(paddr);
251 curr->read.olinfo_status = 0;
252 curr->read.cmd_type_len = htole32(len | flags |
253 IXGBE_ADVTXD_DCMD_IFCS | IXGBE_TXD_CMD_EOP);
255 /* make sure changes to the buffer are synced */
256 bus_dmamap_sync(txr->txtag, txbuf->map,
257 BUS_DMASYNC_PREWRITE);
259 nm_i = nm_next(nm_i, lim);
260 nic_i = nm_next(nic_i, lim);
262 kring->nr_hwcur = head;
264 /* synchronize the NIC ring */
265 bus_dmamap_sync(txr->txdma.dma_tag, txr->txdma.dma_map,
266 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
268 /* (re)start the tx unit up to slot nic_i (excluded) */
269 IXGBE_WRITE_REG(&adapter->hw, IXGBE_TDT(txr->me), nic_i);
273 * Second part: reclaim buffers for completed transmissions.
274 * Because this is expensive (we read a NIC register etc.)
275 * we only do it in specific cases (see below).
277 if (flags & NAF_FORCE_RECLAIM) {
278 reclaim_tx = 1; /* forced reclaim */
279 } else if (!nm_kr_txempty(kring)) {
280 reclaim_tx = 0; /* have buffers, no reclaim */
283 * No buffers available. Locate previous slot with
285 * If the slot has DD set, we can reclaim space,
286 * otherwise wait for the next interrupt.
287 * This enables interrupt moderation on the tx
288 * side though it might reduce throughput.
290 struct ixgbe_legacy_tx_desc *txd =
291 (struct ixgbe_legacy_tx_desc *)txr->tx_base;
293 nic_i = txr->next_to_clean + report_frequency;
296 // round to the closest with dd set
297 nic_i = (nic_i < kring->nkr_num_slots / 4 ||
298 nic_i >= kring->nkr_num_slots*3/4) ?
299 0 : report_frequency;
300 reclaim_tx = txd[nic_i].upper.fields.status & IXGBE_TXD_STAT_DD; // XXX cpu_to_le32 ?
304 * Record completed transmissions.
305 * We (re)use the driver's txr->next_to_clean to keep
306 * track of the most recently completed transmission.
308 * The datasheet discourages the use of TDH to find
309 * out the number of sent packets, but we only set
310 * REPORT_STATUS in a few slots so TDH is the only
313 nic_i = IXGBE_READ_REG(&adapter->hw, IXGBE_TDH(kring->ring_id));
314 if (nic_i >= kring->nkr_num_slots) { /* XXX can it happen ? */
315 D("TDH wrap %d", nic_i);
316 nic_i -= kring->nkr_num_slots;
318 if (nic_i != txr->next_to_clean) {
319 /* some tx completed, increment avail */
320 txr->next_to_clean = nic_i;
321 kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
325 nm_txsync_finalize(kring);
332 * Reconcile kernel and user view of the receive ring.
333 * Same as for the txsync, this routine must be efficient.
334 * The caller guarantees a single invocations, but races against
335 * the rest of the driver should be handled here.
337 * On call, kring->rhead is the first packet that userspace wants
338 * to keep, and kring->rcur is the wakeup point.
339 * The kernel has previously reported packets up to kring->rtail.
341 * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
342 * of whether or not we received an interrupt.
345 ixgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
347 struct netmap_adapter *na = kring->na;
348 struct ifnet *ifp = na->ifp;
349 struct netmap_ring *ring = kring->ring;
350 u_int nm_i; /* index into the netmap ring */
351 u_int nic_i; /* index into the NIC ring */
353 u_int const lim = kring->nkr_num_slots - 1;
354 u_int const head = nm_rxsync_prologue(kring);
355 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
357 /* device-specific */
358 struct adapter *adapter = ifp->if_softc;
359 struct rx_ring *rxr = &adapter->rx_rings[kring->ring_id];
362 return netmap_ring_reinit(kring);
364 /* XXX check sync modes */
365 bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map,
366 BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
369 * First part: import newly received packets.
371 * nm_i is the index of the next free slot in the netmap ring,
372 * nic_i is the index of the next received packet in the NIC ring,
373 * and they may differ in case if_init() has been called while
374 * in netmap mode. For the receive ring we have
376 * nic_i = rxr->next_to_check;
377 * nm_i = kring->nr_hwtail (previous)
379 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
381 * rxr->next_to_check is set to 0 on a ring reinit
383 if (netmap_no_pendintr || force_update) {
384 int crclen = ix_crcstrip ? 0 : 4;
385 uint16_t slot_flags = kring->nkr_slot_flags;
387 nic_i = rxr->next_to_check; // or also k2n(kring->nr_hwtail)
388 nm_i = netmap_idx_n2k(kring, nic_i);
391 union ixgbe_adv_rx_desc *curr = &rxr->rx_base[nic_i];
392 uint32_t staterr = le32toh(curr->wb.upper.status_error);
394 if ((staterr & IXGBE_RXD_STAT_DD) == 0)
396 ring->slot[nm_i].len = le16toh(curr->wb.upper.length) - crclen;
397 ring->slot[nm_i].flags = slot_flags;
398 bus_dmamap_sync(rxr->ptag,
399 rxr->rx_buffers[nic_i].pmap, BUS_DMASYNC_POSTREAD);
400 nm_i = nm_next(nm_i, lim);
401 nic_i = nm_next(nic_i, lim);
403 if (n) { /* update the state variables */
404 if (netmap_no_pendintr && !force_update) {
407 ix_rx_miss_bufs += n;
409 rxr->next_to_check = nic_i;
410 kring->nr_hwtail = nm_i;
412 kring->nr_kflags &= ~NKR_PENDINTR;
416 * Second part: skip past packets that userspace has released.
417 * (kring->nr_hwcur to kring->rhead excluded),
418 * and make the buffers available for reception.
419 * As usual nm_i is the index in the netmap ring,
420 * nic_i is the index in the NIC ring, and
421 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
423 nm_i = kring->nr_hwcur;
425 nic_i = netmap_idx_k2n(kring, nm_i);
426 for (n = 0; nm_i != head; n++) {
427 struct netmap_slot *slot = &ring->slot[nm_i];
429 void *addr = PNMB(na, slot, &paddr);
431 union ixgbe_adv_rx_desc *curr = &rxr->rx_base[nic_i];
432 struct ixgbe_rx_buf *rxbuf = &rxr->rx_buffers[nic_i];
434 if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
437 if (slot->flags & NS_BUF_CHANGED) {
438 /* buffer has changed, reload map */
439 netmap_reload_map(na, rxr->ptag, rxbuf->pmap, addr);
440 slot->flags &= ~NS_BUF_CHANGED;
442 curr->wb.upper.status_error = 0;
443 curr->read.pkt_addr = htole64(paddr);
444 bus_dmamap_sync(rxr->ptag, rxbuf->pmap,
445 BUS_DMASYNC_PREREAD);
446 nm_i = nm_next(nm_i, lim);
447 nic_i = nm_next(nic_i, lim);
449 kring->nr_hwcur = head;
451 bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map,
452 BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
454 * IMPORTANT: we must leave one free slot in the ring,
455 * so move nic_i back by one unit
457 nic_i = nm_prev(nic_i, lim);
458 IXGBE_WRITE_REG(&adapter->hw, IXGBE_RDT(rxr->me), nic_i);
461 /* tell userspace that there might be new packets */
462 nm_rxsync_finalize(kring);
467 return netmap_ring_reinit(kring);
472 * The attach routine, called near the end of ixgbe_attach(),
473 * fills the parameters for netmap_attach() and calls it.
474 * It cannot fail, in the worst case (such as no memory)
475 * netmap mode will be disabled and the driver will only
476 * operate in standard mode.
479 ixgbe_netmap_attach(struct adapter *adapter)
481 struct netmap_adapter na;
483 bzero(&na, sizeof(na));
485 na.ifp = adapter->ifp;
486 na.na_flags = NAF_BDG_MAYSLEEP;
487 na.num_tx_desc = adapter->num_tx_desc;
488 na.num_rx_desc = adapter->num_rx_desc;
489 na.nm_txsync = ixgbe_netmap_txsync;
490 na.nm_rxsync = ixgbe_netmap_rxsync;
491 na.nm_register = ixgbe_netmap_reg;
492 na.num_tx_rings = na.num_rx_rings = adapter->num_queues;