freebsd-dev/sys/dev/netmap/ixgbe_netmap.h
Luigi Rizzo 37e3a6d349 Import the current version of netmap, aligned with the one on github.
This commit, long overdue, contains contributions in the last 2 years
from Stefano Garzarella, Giuseppe Lettieri, Vincenzo Maffione, including:
+ fixes on monitor ports
+ the 'ptnet' virtual device driver, and ptnetmap backend, for
  high speed virtual passthrough on VMs (bhyve fixes in an upcoming commit)
+ improved emulated netmap mode
+ more robust error handling
+ removal of stale code
+ various fixes to code and documentation (some mixup between RX and TX
  parameters, and private and public variables)

We also include an additional tool, nmreplay, which is functionally
equivalent to tcpreplay but operating on netmap ports.
2016-10-16 14:13:32 +00:00

508 lines
16 KiB
C

/*
* Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
/*
* $FreeBSD$
*
* netmap support for: ixgbe (both ix and ixv)
*
* This file is meant to be a reference on how to implement
* netmap support for a network driver.
* This file contains code but only static or inline functions used
* by a single driver. To avoid replication of code we just #include
* it near the beginning of the standard driver.
*/
#include <net/netmap.h>
#include <sys/selinfo.h>
/*
* Some drivers may need the following headers. Others
* already include them by default
#include <vm/vm.h>
#include <vm/pmap.h>
*/
#include <dev/netmap/netmap_kern.h>
void ixgbe_netmap_attach(struct adapter *adapter);
/*
* device-specific sysctl variables:
*
* ix_crcstrip: 0: NIC keeps CRC in rx frames (default), 1: NIC strips it.
* During regular operations the CRC is stripped, but on some
* hardware reception of frames not multiple of 64 is slower,
* so using crcstrip=0 helps in benchmarks.
*
* ix_rx_miss, ix_rx_miss_bufs:
* count packets that might be missed due to lost interrupts.
*/
SYSCTL_DECL(_dev_netmap);
static int ix_rx_miss, ix_rx_miss_bufs;
int ix_crcstrip;
SYSCTL_INT(_dev_netmap, OID_AUTO, ix_crcstrip,
CTLFLAG_RW, &ix_crcstrip, 0, "NIC strips CRC on rx frames");
SYSCTL_INT(_dev_netmap, OID_AUTO, ix_rx_miss,
CTLFLAG_RW, &ix_rx_miss, 0, "potentially missed rx intr");
SYSCTL_INT(_dev_netmap, OID_AUTO, ix_rx_miss_bufs,
CTLFLAG_RW, &ix_rx_miss_bufs, 0, "potentially missed rx intr bufs");
static void
set_crcstrip(struct ixgbe_hw *hw, int onoff)
{
/* crc stripping is set in two places:
* IXGBE_HLREG0 (modified on init_locked and hw reset)
* IXGBE_RDRXCTL (set by the original driver in
* ixgbe_setup_hw_rsc() called in init_locked.
* We disable the setting when netmap is compiled in).
* We update the values here, but also in ixgbe.c because
* init_locked sometimes is called outside our control.
*/
uint32_t hl, rxc;
hl = IXGBE_READ_REG(hw, IXGBE_HLREG0);
rxc = IXGBE_READ_REG(hw, IXGBE_RDRXCTL);
if (netmap_verbose)
D("%s read HLREG 0x%x rxc 0x%x",
onoff ? "enter" : "exit", hl, rxc);
/* hw requirements ... */
rxc &= ~IXGBE_RDRXCTL_RSCFRSTSIZE;
rxc |= IXGBE_RDRXCTL_RSCACKC;
if (onoff && !ix_crcstrip) {
/* keep the crc. Fast rx */
hl &= ~IXGBE_HLREG0_RXCRCSTRP;
rxc &= ~IXGBE_RDRXCTL_CRCSTRIP;
} else {
/* reset default mode */
hl |= IXGBE_HLREG0_RXCRCSTRP;
rxc |= IXGBE_RDRXCTL_CRCSTRIP;
}
if (netmap_verbose)
D("%s write HLREG 0x%x rxc 0x%x",
onoff ? "enter" : "exit", hl, rxc);
IXGBE_WRITE_REG(hw, IXGBE_HLREG0, hl);
IXGBE_WRITE_REG(hw, IXGBE_RDRXCTL, rxc);
}
static void
ixgbe_netmap_intr(struct netmap_adapter *na, int onoff)
{
struct ifnet *ifp = na->ifp;
struct adapter *adapter = ifp->if_softc;
IXGBE_CORE_LOCK(adapter);
if (onoff) {
ixgbe_enable_intr(adapter); // XXX maybe ixgbe_stop ?
} else {
ixgbe_disable_intr(adapter); // XXX maybe ixgbe_stop ?
}
IXGBE_CORE_UNLOCK(adapter);
}
/*
* Register/unregister. We are already under netmap lock.
* Only called on the first register or the last unregister.
*/
static int
ixgbe_netmap_reg(struct netmap_adapter *na, int onoff)
{
struct ifnet *ifp = na->ifp;
struct adapter *adapter = ifp->if_softc;
IXGBE_CORE_LOCK(adapter);
adapter->stop_locked(adapter);
if (!IXGBE_IS_VF(adapter))
set_crcstrip(&adapter->hw, onoff);
/* enable or disable flags and callbacks in na and ifp */
if (onoff) {
nm_set_native_flags(na);
} else {
nm_clear_native_flags(na);
}
adapter->init_locked(adapter); /* also enables intr */
if (!IXGBE_IS_VF(adapter))
set_crcstrip(&adapter->hw, onoff); // XXX why twice ?
IXGBE_CORE_UNLOCK(adapter);
return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1);
}
/*
* Reconcile kernel and user view of the transmit ring.
*
* All information is in the kring.
* Userspace wants to send packets up to the one before kring->rhead,
* kernel knows kring->nr_hwcur is the first unsent packet.
*
* Here we push packets out (as many as possible), and possibly
* reclaim buffers from previously completed transmission.
*
* The caller (netmap) guarantees that there is only one instance
* running at any time. Any interference with other driver
* methods should be handled by the individual drivers.
*/
static int
ixgbe_netmap_txsync(struct netmap_kring *kring, int flags)
{
struct netmap_adapter *na = kring->na;
struct ifnet *ifp = na->ifp;
struct netmap_ring *ring = kring->ring;
u_int nm_i; /* index into the netmap ring */
u_int nic_i; /* index into the NIC ring */
u_int n;
u_int const lim = kring->nkr_num_slots - 1;
u_int const head = kring->rhead;
/*
* interrupts on every tx packet are expensive so request
* them every half ring, or where NS_REPORT is set
*/
u_int report_frequency = kring->nkr_num_slots >> 1;
/* device-specific */
struct adapter *adapter = ifp->if_softc;
struct tx_ring *txr = &adapter->tx_rings[kring->ring_id];
int reclaim_tx;
bus_dmamap_sync(txr->txdma.dma_tag, txr->txdma.dma_map,
BUS_DMASYNC_POSTREAD);
/*
* First part: process new packets to send.
* nm_i is the current index in the netmap ring,
* nic_i is the corresponding index in the NIC ring.
* The two numbers differ because upon a *_init() we reset
* the NIC ring but leave the netmap ring unchanged.
* For the transmit ring, we have
*
* nm_i = kring->nr_hwcur
* nic_i = IXGBE_TDT (not tracked in the driver)
* and
* nm_i == (nic_i + kring->nkr_hwofs) % ring_size
*
* In this driver kring->nkr_hwofs >= 0, but for other
* drivers it might be negative as well.
*/
/*
* If we have packets to send (kring->nr_hwcur != kring->rhead)
* iterate over the netmap ring, fetch length and update
* the corresponding slot in the NIC ring. Some drivers also
* need to update the buffer's physical address in the NIC slot
* even NS_BUF_CHANGED is not set (PNMB computes the addresses).
*
* The netmap_reload_map() calls is especially expensive,
* even when (as in this case) the tag is 0, so do only
* when the buffer has actually changed.
*
* If possible do not set the report/intr bit on all slots,
* but only a few times per ring or when NS_REPORT is set.
*
* Finally, on 10G and faster drivers, it might be useful
* to prefetch the next slot and txr entry.
*/
nm_i = kring->nr_hwcur;
if (nm_i != head) { /* we have new packets to send */
nic_i = netmap_idx_k2n(kring, nm_i);
__builtin_prefetch(&ring->slot[nm_i]);
__builtin_prefetch(&txr->tx_buffers[nic_i]);
for (n = 0; nm_i != head; n++) {
struct netmap_slot *slot = &ring->slot[nm_i];
u_int len = slot->len;
uint64_t paddr;
void *addr = PNMB(na, slot, &paddr);
/* device-specific */
union ixgbe_adv_tx_desc *curr = &txr->tx_base[nic_i];
struct ixgbe_tx_buf *txbuf = &txr->tx_buffers[nic_i];
int flags = (slot->flags & NS_REPORT ||
nic_i == 0 || nic_i == report_frequency) ?
IXGBE_TXD_CMD_RS : 0;
/* prefetch for next round */
__builtin_prefetch(&ring->slot[nm_i + 1]);
__builtin_prefetch(&txr->tx_buffers[nic_i + 1]);
NM_CHECK_ADDR_LEN(na, addr, len);
if (slot->flags & NS_BUF_CHANGED) {
/* buffer has changed, reload map */
netmap_reload_map(na, txr->txtag, txbuf->map, addr);
}
slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
/* Fill the slot in the NIC ring. */
/* Use legacy descriptor, they are faster? */
curr->read.buffer_addr = htole64(paddr);
curr->read.olinfo_status = 0;
curr->read.cmd_type_len = htole32(len | flags |
IXGBE_ADVTXD_DCMD_IFCS | IXGBE_TXD_CMD_EOP);
/* make sure changes to the buffer are synced */
bus_dmamap_sync(txr->txtag, txbuf->map,
BUS_DMASYNC_PREWRITE);
nm_i = nm_next(nm_i, lim);
nic_i = nm_next(nic_i, lim);
}
kring->nr_hwcur = head;
/* synchronize the NIC ring */
bus_dmamap_sync(txr->txdma.dma_tag, txr->txdma.dma_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
/* (re)start the tx unit up to slot nic_i (excluded) */
IXGBE_WRITE_REG(&adapter->hw, txr->tail, nic_i);
}
/*
* Second part: reclaim buffers for completed transmissions.
* Because this is expensive (we read a NIC register etc.)
* we only do it in specific cases (see below).
*/
if (flags & NAF_FORCE_RECLAIM) {
reclaim_tx = 1; /* forced reclaim */
} else if (!nm_kr_txempty(kring)) {
reclaim_tx = 0; /* have buffers, no reclaim */
} else {
/*
* No buffers available. Locate previous slot with
* REPORT_STATUS set.
* If the slot has DD set, we can reclaim space,
* otherwise wait for the next interrupt.
* This enables interrupt moderation on the tx
* side though it might reduce throughput.
*/
struct ixgbe_legacy_tx_desc *txd =
(struct ixgbe_legacy_tx_desc *)txr->tx_base;
nic_i = txr->next_to_clean + report_frequency;
if (nic_i > lim)
nic_i -= lim + 1;
// round to the closest with dd set
nic_i = (nic_i < kring->nkr_num_slots / 4 ||
nic_i >= kring->nkr_num_slots*3/4) ?
0 : report_frequency;
reclaim_tx = txd[nic_i].upper.fields.status & IXGBE_TXD_STAT_DD; // XXX cpu_to_le32 ?
}
if (reclaim_tx) {
/*
* Record completed transmissions.
* We (re)use the driver's txr->next_to_clean to keep
* track of the most recently completed transmission.
*
* The datasheet discourages the use of TDH to find
* out the number of sent packets, but we only set
* REPORT_STATUS in a few slots so TDH is the only
* good way.
*/
nic_i = IXGBE_READ_REG(&adapter->hw, IXGBE_IS_VF(adapter) ?
IXGBE_VFTDH(kring->ring_id) : IXGBE_TDH(kring->ring_id));
if (nic_i >= kring->nkr_num_slots) { /* XXX can it happen ? */
D("TDH wrap %d", nic_i);
nic_i -= kring->nkr_num_slots;
}
if (nic_i != txr->next_to_clean) {
/* some tx completed, increment avail */
txr->next_to_clean = nic_i;
kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
}
}
return 0;
}
/*
* Reconcile kernel and user view of the receive ring.
* Same as for the txsync, this routine must be efficient.
* The caller guarantees a single invocations, but races against
* the rest of the driver should be handled here.
*
* On call, kring->rhead is the first packet that userspace wants
* to keep, and kring->rcur is the wakeup point.
* The kernel has previously reported packets up to kring->rtail.
*
* If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
* of whether or not we received an interrupt.
*/
static int
ixgbe_netmap_rxsync(struct netmap_kring *kring, int flags)
{
struct netmap_adapter *na = kring->na;
struct ifnet *ifp = na->ifp;
struct netmap_ring *ring = kring->ring;
u_int nm_i; /* index into the netmap ring */
u_int nic_i; /* index into the NIC ring */
u_int n;
u_int const lim = kring->nkr_num_slots - 1;
u_int const head = kring->rhead;
int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
/* device-specific */
struct adapter *adapter = ifp->if_softc;
struct rx_ring *rxr = &adapter->rx_rings[kring->ring_id];
if (head > lim)
return netmap_ring_reinit(kring);
/* XXX check sync modes */
bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map,
BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
/*
* First part: import newly received packets.
*
* nm_i is the index of the next free slot in the netmap ring,
* nic_i is the index of the next received packet in the NIC ring,
* and they may differ in case if_init() has been called while
* in netmap mode. For the receive ring we have
*
* nic_i = rxr->next_to_check;
* nm_i = kring->nr_hwtail (previous)
* and
* nm_i == (nic_i + kring->nkr_hwofs) % ring_size
*
* rxr->next_to_check is set to 0 on a ring reinit
*/
if (netmap_no_pendintr || force_update) {
int crclen = (ix_crcstrip || IXGBE_IS_VF(adapter) ) ? 0 : 4;
uint16_t slot_flags = kring->nkr_slot_flags;
nic_i = rxr->next_to_check; // or also k2n(kring->nr_hwtail)
nm_i = netmap_idx_n2k(kring, nic_i);
for (n = 0; ; n++) {
union ixgbe_adv_rx_desc *curr = &rxr->rx_base[nic_i];
uint32_t staterr = le32toh(curr->wb.upper.status_error);
if ((staterr & IXGBE_RXD_STAT_DD) == 0)
break;
ring->slot[nm_i].len = le16toh(curr->wb.upper.length) - crclen;
ring->slot[nm_i].flags = slot_flags;
bus_dmamap_sync(rxr->ptag,
rxr->rx_buffers[nic_i].pmap, BUS_DMASYNC_POSTREAD);
nm_i = nm_next(nm_i, lim);
nic_i = nm_next(nic_i, lim);
}
if (n) { /* update the state variables */
if (netmap_no_pendintr && !force_update) {
/* diagnostics */
ix_rx_miss ++;
ix_rx_miss_bufs += n;
}
rxr->next_to_check = nic_i;
kring->nr_hwtail = nm_i;
}
kring->nr_kflags &= ~NKR_PENDINTR;
}
/*
* Second part: skip past packets that userspace has released.
* (kring->nr_hwcur to kring->rhead excluded),
* and make the buffers available for reception.
* As usual nm_i is the index in the netmap ring,
* nic_i is the index in the NIC ring, and
* nm_i == (nic_i + kring->nkr_hwofs) % ring_size
*/
nm_i = kring->nr_hwcur;
if (nm_i != head) {
nic_i = netmap_idx_k2n(kring, nm_i);
for (n = 0; nm_i != head; n++) {
struct netmap_slot *slot = &ring->slot[nm_i];
uint64_t paddr;
void *addr = PNMB(na, slot, &paddr);
union ixgbe_adv_rx_desc *curr = &rxr->rx_base[nic_i];
struct ixgbe_rx_buf *rxbuf = &rxr->rx_buffers[nic_i];
if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
goto ring_reset;
if (slot->flags & NS_BUF_CHANGED) {
/* buffer has changed, reload map */
netmap_reload_map(na, rxr->ptag, rxbuf->pmap, addr);
slot->flags &= ~NS_BUF_CHANGED;
}
curr->wb.upper.status_error = 0;
curr->read.pkt_addr = htole64(paddr);
bus_dmamap_sync(rxr->ptag, rxbuf->pmap,
BUS_DMASYNC_PREREAD);
nm_i = nm_next(nm_i, lim);
nic_i = nm_next(nic_i, lim);
}
kring->nr_hwcur = head;
bus_dmamap_sync(rxr->rxdma.dma_tag, rxr->rxdma.dma_map,
BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
/*
* IMPORTANT: we must leave one free slot in the ring,
* so move nic_i back by one unit
*/
nic_i = nm_prev(nic_i, lim);
IXGBE_WRITE_REG(&adapter->hw, rxr->tail, nic_i);
}
return 0;
ring_reset:
return netmap_ring_reinit(kring);
}
/*
* The attach routine, called near the end of ixgbe_attach(),
* fills the parameters for netmap_attach() and calls it.
* It cannot fail, in the worst case (such as no memory)
* netmap mode will be disabled and the driver will only
* operate in standard mode.
*/
void
ixgbe_netmap_attach(struct adapter *adapter)
{
struct netmap_adapter na;
bzero(&na, sizeof(na));
na.ifp = adapter->ifp;
na.na_flags = NAF_BDG_MAYSLEEP;
na.num_tx_desc = adapter->num_tx_desc;
na.num_rx_desc = adapter->num_rx_desc;
na.nm_txsync = ixgbe_netmap_txsync;
na.nm_rxsync = ixgbe_netmap_rxsync;
na.nm_register = ixgbe_netmap_reg;
na.num_tx_rings = na.num_rx_rings = adapter->num_queues;
na.nm_intr = ixgbe_netmap_intr;
netmap_attach(&na);
}
/* end of file */