/*
* Calculate the checksum flags to be relayed to the peer for the
* packet `mp'.
*/
static uint16_t
xnbo_cksum_to_peer(xnb_t *xnbp, mblk_t *mp)
{
_NOTE(ARGUNUSED(xnbp));
uint16_t r = 0;
uint32_t pflags, csum;
/*
* We might also check for HCK_PARTIALCKSUM here and,
* providing that the partial checksum covers the TCP/UDP
* payload, return NETRXF_data_validated.
*
* It seems that it's probably not worthwhile, as even MAC
* devices which advertise HCKSUM_INET_PARTIAL in their
* capabilities tend to use HCK_FULLCKSUM on the receive side
* - they are actually saying that in the output path the
* caller must use HCK_PARTIALCKSUM.
*
* Then again, if a NIC supports HCK_PARTIALCKSUM in its'
* output path, the host IP stack will use it. If such packets
* are destined for the peer (i.e. looped around) we would
* gain some advantage.
*/
hcksum_retrieve(mp, NULL, NULL, NULL, NULL,
NULL, &csum, &pflags);
/*
* If the MAC driver has asserted that the checksum is
* good, let the peer know.
*/
if (((pflags & HCK_FULLCKSUM) != 0) &&
(((pflags & HCK_FULLCKSUM_OK) != 0) ||
(csum == 0xffff)))
r |= NETRXF_data_validated;
return (r);
}
/*
*---------------------------------------------------------------------------
*
* vmxnet3_tx_prepare_offload --
*
* Build the offload context of a msg.
*
* Results:
* 0 if everything went well.
* +n if n bytes need to be pulled up.
* -1 in case of error (not used).
*
* Side effects:
* None.
*
*---------------------------------------------------------------------------
*/
static int
vmxnet3_tx_prepare_offload(vmxnet3_softc_t *dp,
vmxnet3_offload_t *ol,
mblk_t *mp)
{
int ret = 0;
uint32_t start, stuff, value, flags, lso_flag, mss;
ol->om = VMXNET3_OM_NONE;
ol->hlen = 0;
ol->msscof = 0;
hcksum_retrieve(mp, NULL, NULL, &start, &stuff, NULL, &value, &flags);
mac_lso_get(mp, &mss, &lso_flag);
if (flags || lso_flag) {
struct ether_vlan_header *eth = (void *) mp->b_rptr;
uint8_t ethLen;
if (eth->ether_tpid == htons(ETHERTYPE_VLAN)) {
ethLen = sizeof(struct ether_vlan_header);
} else {
ethLen = sizeof(struct ether_header);
}
VMXNET3_DEBUG(dp, 4, "flags=0x%x, ethLen=%u, start=%u, stuff=%u, value=%u\n",
flags, ethLen, start, stuff, value);
if (lso_flag & HW_LSO) {
mblk_t *mblk = mp;
uint8_t *ip, *tcp;
uint8_t ipLen, tcpLen;
/*
* Copy e1000g's behavior:
* - Do not assume all the headers are in the same mblk.
* - Assume each header is always within one mblk.
* - Assume the ethernet header is in the first mblk.
*/
ip = mblk->b_rptr + ethLen;
if (ip >= mblk->b_wptr) {
mblk = mblk->b_cont;
ip = mblk->b_rptr;
}
ipLen = IPH_HDR_LENGTH((ipha_t *) ip);
tcp = ip + ipLen;
if (tcp >= mblk->b_wptr) {
mblk = mblk->b_cont;
tcp = mblk->b_rptr;
}
tcpLen = TCP_HDR_LENGTH((tcph_t *) tcp);
if (tcp + tcpLen > mblk->b_wptr) { // careful, '>' instead of '>=' here
mblk = mblk->b_cont;
}
ol->om = VMXNET3_OM_TSO;
ol->hlen = ethLen + ipLen + tcpLen;
ol->msscof = mss;
if (mblk != mp) {
ret = ol->hlen;
}
} else if (flags & HCK_PARTIALCKSUM) {
ol->om = VMXNET3_OM_CSUM;
ol->hlen = start + ethLen;
ol->msscof = stuff + ethLen;
}
}
return ret;
}
/*
*---------------------------------------------------------------------------
*
* vmxnet3_tx_prepare_offload --
*
* Build the offload context of a msg.
*
* Results:
* 0 if everything went well.
* +n if n bytes need to be pulled up.
* -1 in case of error (not used).
*
* Side effects:
* None.
*
*---------------------------------------------------------------------------
*/
static int
vmxnet3_tx_prepare_offload(vmxnet3_softc_t *dp,
vmxnet3_offload_t *ol,
mblk_t *mp)
{
int ret = 0;
uint32_t start, stuff, value, flags;
#if defined(OPEN_SOLARIS) || defined(SOL11)
uint32_t lso_flag, mss;
#endif
ol->om = VMXNET3_OM_NONE;
ol->hlen = 0;
ol->msscof = 0;
hcksum_retrieve(mp, NULL, NULL, &start, &stuff, NULL, &value, &flags);
#if defined(OPEN_SOLARIS) || defined(SOL11)
mac_lso_get(mp, &mss, &lso_flag);
if (flags || lso_flag) {
#else
if (flags) {
#endif
struct ether_vlan_header *eth = (void *) mp->b_rptr;
uint8_t ethLen;
if (eth->ether_tpid == htons(ETHERTYPE_VLAN)) {
ethLen = sizeof(struct ether_vlan_header);
} else {
ethLen = sizeof(struct ether_header);
}
VMXNET3_DEBUG(dp, 4, "flags=0x%x, ethLen=%u, start=%u, stuff=%u, value=%u\n",
flags, ethLen, start, stuff, value);
#if defined(OPEN_SOLARIS) || defined(SOL11)
if (lso_flag & HW_LSO) {
#else
if (flags & HCK_PARTIALCKSUM) {
ol->om = VMXNET3_OM_CSUM;
ol->hlen = start + ethLen;
ol->msscof = stuff + ethLen;
}
if (flags & HW_LSO) {
#endif
mblk_t *mblk = mp;
uint8_t *ip, *tcp;
uint8_t ipLen, tcpLen;
/*
* Copy e1000g's behavior:
* - Do not assume all the headers are in the same mblk.
* - Assume each header is always within one mblk.
* - Assume the ethernet header is in the first mblk.
*/
ip = mblk->b_rptr + ethLen;
if (ip >= mblk->b_wptr) {
mblk = mblk->b_cont;
ip = mblk->b_rptr;
}
ipLen = IPH_HDR_LENGTH((ipha_t *) ip);
tcp = ip + ipLen;
if (tcp >= mblk->b_wptr) {
mblk = mblk->b_cont;
tcp = mblk->b_rptr;
}
tcpLen = TCP_HDR_LENGTH((tcph_t *) tcp);
if (tcp + tcpLen > mblk->b_wptr) { // careful, '>' instead of '>=' here
mblk = mblk->b_cont;
}
ol->om = VMXNET3_OM_TSO;
ol->hlen = ethLen + ipLen + tcpLen;
#if defined(OPEN_SOLARIS) || defined(SOL11)
ol->msscof = mss;
#else
/* OpenSolaris fills 'value' with the MSS but Solaris doesn't. */
ol->msscof = DB_LSOMSS(mp);
#endif
if (mblk != mp) {
ret = ol->hlen;
}
}
#if defined(OPEN_SOLARIS) || defined(SOL11)
else if (flags & HCK_PARTIALCKSUM) {
ol->om = VMXNET3_OM_CSUM;
ol->hlen = start + ethLen;
ol->msscof = stuff + ethLen;
}
#endif
}
return ret;
}
/*
*---------------------------------------------------------------------------
*
* vmxnet3_tx_one --
*
* Map a msg into the Tx command ring of a vmxnet3 device.
*
* Results:
* VMXNET3_TX_OK if everything went well.
* VMXNET3_TX_RINGFULL if the ring is nearly full.
* VMXNET3_TX_PULLUP if the msg is overfragmented.
* VMXNET3_TX_FAILURE if there was a DMA or offload error.
*
* Side effects:
* The ring is filled if VMXNET3_TX_OK is returned.
*
*---------------------------------------------------------------------------
*/
static vmxnet3_txstatus
vmxnet3_tx_one(vmxnet3_softc_t *dp,
vmxnet3_txqueue_t *txq,
vmxnet3_offload_t *ol,
mblk_t *mp,
boolean_t retry)
{
int ret = VMXNET3_TX_OK;
unsigned int frags = 0, totLen = 0;
vmxnet3_cmdring_t *cmdRing = &txq->cmdRing;
Vmxnet3_TxQueueCtrl *txqCtrl = txq->sharedCtrl;
Vmxnet3_GenericDesc *txDesc;
uint16_t sopIdx, eopIdx;
uint8_t sopGen, curGen;
mblk_t *mblk;
mutex_enter(&dp->txLock);
sopIdx = eopIdx = cmdRing->next2fill;
sopGen = cmdRing->gen;
curGen = !cmdRing->gen;
for (mblk = mp; mblk != NULL; mblk = mblk->b_cont) {
unsigned int len = MBLKL(mblk);
ddi_dma_cookie_t cookie;
uint_t cookieCount;
if (len) {
totLen += len;
} else {
continue;
}
if (ddi_dma_addr_bind_handle(dp->txDmaHandle, NULL,
(caddr_t) mblk->b_rptr, len,
DDI_DMA_RDWR | DDI_DMA_STREAMING,
DDI_DMA_DONTWAIT, NULL,
&cookie, &cookieCount) != DDI_DMA_MAPPED) {
VMXNET3_WARN(dp, "ddi_dma_addr_bind_handle() failed\n");
ret = VMXNET3_TX_FAILURE;
goto error;
}
ASSERT(cookieCount);
do {
uint64_t addr = cookie.dmac_laddress;
size_t len = cookie.dmac_size;
do {
uint32_t dw2, dw3;
size_t chunkLen;
ASSERT(!txq->metaRing[eopIdx].mp);
ASSERT(cmdRing->avail - frags);
if (frags >= cmdRing->size - 1 ||
(ol->om != VMXNET3_OM_TSO && frags >= VMXNET3_MAX_TXD_PER_PKT)) {
if (retry) {
VMXNET3_DEBUG(dp, 2, "overfragmented, frags=%u ring=%hu om=%hu\n",
frags, cmdRing->size, ol->om);
}
ddi_dma_unbind_handle(dp->txDmaHandle);
ret = VMXNET3_TX_PULLUP;
goto error;
}
if (cmdRing->avail - frags <= 1) {
dp->txMustResched = B_TRUE;
ddi_dma_unbind_handle(dp->txDmaHandle);
ret = VMXNET3_TX_RINGFULL;
goto error;
}
if (len > VMXNET3_MAX_TX_BUF_SIZE) {
chunkLen = VMXNET3_MAX_TX_BUF_SIZE;
} else {
chunkLen = len;
}
frags++;
eopIdx = cmdRing->next2fill;
txDesc = VMXNET3_GET_DESC(cmdRing, eopIdx);
ASSERT(txDesc->txd.gen != cmdRing->gen);
// txd.addr
txDesc->txd.addr = addr;
// txd.dw2
dw2 = chunkLen == VMXNET3_MAX_TX_BUF_SIZE ? 0 : chunkLen;
dw2 |= curGen << VMXNET3_TXD_GEN_SHIFT;
txDesc->dword[2] = dw2;
ASSERT(txDesc->txd.len == len || txDesc->txd.len == 0);
// txd.dw3
dw3 = 0;
txDesc->dword[3] = dw3;
VMXNET3_INC_RING_IDX(cmdRing, cmdRing->next2fill);
curGen = cmdRing->gen;
addr += chunkLen;
len -= chunkLen;
} while (len);
if (--cookieCount) {
ddi_dma_nextcookie(dp->txDmaHandle, &cookie);
}
} while (cookieCount);
ddi_dma_unbind_handle(dp->txDmaHandle);
}
/* Update the EOP descriptor */
txDesc = VMXNET3_GET_DESC(cmdRing, eopIdx);
txDesc->dword[3] |= VMXNET3_TXD_CQ | VMXNET3_TXD_EOP;
/* Update the SOP descriptor. Must be done last */
txDesc = VMXNET3_GET_DESC(cmdRing, sopIdx);
if (ol->om == VMXNET3_OM_TSO &&
txDesc->txd.len != 0 &&
txDesc->txd.len < ol->hlen) {
ret = VMXNET3_TX_FAILURE;
goto error;
}
txDesc->txd.om = ol->om;
txDesc->txd.hlen = ol->hlen;
txDesc->txd.msscof = ol->msscof;
membar_producer();
txDesc->txd.gen = sopGen;
/* Update the meta ring & metadata */
txq->metaRing[sopIdx].mp = mp;
txq->metaRing[eopIdx].sopIdx = sopIdx;
txq->metaRing[eopIdx].frags = frags;
cmdRing->avail -= frags;
if (ol->om == VMXNET3_OM_TSO) {
txqCtrl->txNumDeferred +=
(totLen - ol->hlen + ol->msscof - 1) / ol->msscof;
} else {
txqCtrl->txNumDeferred++;
}
VMXNET3_DEBUG(dp, 3, "tx 0x%p on [%u;%u]\n", mp, sopIdx, eopIdx);
goto done;
error:
/* Reverse the generation bits */
while (sopIdx != cmdRing->next2fill) {
VMXNET3_DEC_RING_IDX(cmdRing, cmdRing->next2fill);
txDesc = VMXNET3_GET_DESC(cmdRing, cmdRing->next2fill);
txDesc->txd.gen = !cmdRing->gen;
}
done:
mutex_exit(&dp->txLock);
return ret;
}
/*
*---------------------------------------------------------------------------
*
* vmxnet3_tx --
*
* Send packets on a vmxnet3 device.
*
* Results:
* NULL in case of success or failure.
* The mps to be retransmitted later if the ring is full.
*
* Side effects:
* None.
*
*---------------------------------------------------------------------------
*/
mblk_t *
vmxnet3_tx(void *data, mblk_t *mps)
{
vmxnet3_softc_t *dp = data;
vmxnet3_txqueue_t *txq = &dp->txQueue;
vmxnet3_cmdring_t *cmdRing = &txq->cmdRing;
Vmxnet3_TxQueueCtrl *txqCtrl = txq->sharedCtrl;
vmxnet3_txstatus status = VMXNET3_TX_OK;
mblk_t *mp;
ASSERT(mps != NULL);
do {
vmxnet3_offload_t ol;
int pullup;
mp = mps;
mps = mp->b_next;
mp->b_next = NULL;
if (DB_TYPE(mp) != M_DATA) {
/*
* PR #315560: Solaris might pass M_PROTO mblks for some reason.
* Drop them because we don't understand them and because their
* contents are not Ethernet frames anyway.
*/
ASSERT(B_FALSE);
freemsg(mp);
continue;
}
/*
* Prepare the offload while we're still handling the original
* message -- msgpullup() discards the metadata afterwards.
*/
pullup = vmxnet3_tx_prepare_offload(dp, &ol, mp);
if (pullup) {
mblk_t *new_mp = msgpullup(mp, pullup);
freemsg(mp);
if (new_mp) {
mp = new_mp;
} else {
continue;
}
}
/*
* Try to map the message in the Tx ring.
* This call might fail for non-fatal reasons.
*/
status = vmxnet3_tx_one(dp, txq, &ol, mp, B_FALSE);
if (status == VMXNET3_TX_PULLUP) {
/*
* Try one more time after flattening
* the message with msgpullup().
*/
if (mp->b_cont != NULL) {
mblk_t *new_mp = msgpullup(mp, -1);
freemsg(mp);
if (new_mp) {
mp = new_mp;
status = vmxnet3_tx_one(dp, txq, &ol, mp, B_TRUE);
} else {
continue;
}
}
}
if (status != VMXNET3_TX_OK && status != VMXNET3_TX_RINGFULL) {
/* Fatal failure, drop it */
freemsg(mp);
}
} while (mps && status != VMXNET3_TX_RINGFULL);
if (status == VMXNET3_TX_RINGFULL) {
mp->b_next = mps;
mps = mp;
} else {
ASSERT(!mps);
}
/* Notify the device */
mutex_enter(&dp->txLock);
if (txqCtrl->txNumDeferred >= txqCtrl->txThreshold) {
txqCtrl->txNumDeferred = 0;
VMXNET3_BAR0_PUT32(dp, VMXNET3_REG_TXPROD, cmdRing->next2fill);
}
mutex_exit(&dp->txLock);
return mps;
}
/*
* DL_UINTDATA_REQ
*/
void
proto_unitdata_req(dld_str_t *dsp, mblk_t *mp)
{
queue_t *q = dsp->ds_wq;
dl_unitdata_req_t *dlp = (dl_unitdata_req_t *)mp->b_rptr;
off_t off;
size_t len, size;
const uint8_t *addr;
uint16_t sap;
uint_t addr_length;
mblk_t *bp, *payload;
uint32_t start, stuff, end, value, flags;
t_uscalar_t dl_err;
uint_t max_sdu;
if (MBLKL(mp) < sizeof (dl_unitdata_req_t) || mp->b_cont == NULL) {
dlerrorack(q, mp, DL_UNITDATA_REQ, DL_BADPRIM, 0);
return;
}
mutex_enter(&dsp->ds_lock);
if (dsp->ds_dlstate != DL_IDLE) {
mutex_exit(&dsp->ds_lock);
dlerrorack(q, mp, DL_UNITDATA_REQ, DL_OUTSTATE, 0);
return;
}
DLD_DATATHR_INC(dsp);
mutex_exit(&dsp->ds_lock);
addr_length = dsp->ds_mip->mi_addr_length;
off = dlp->dl_dest_addr_offset;
len = dlp->dl_dest_addr_length;
if (!MBLKIN(mp, off, len) || !IS_P2ALIGNED(off, sizeof (uint16_t))) {
dl_err = DL_BADPRIM;
goto failed;
}
if (len != addr_length + sizeof (uint16_t)) {
dl_err = DL_BADADDR;
goto failed;
}
addr = mp->b_rptr + off;
sap = *(uint16_t *)(mp->b_rptr + off + addr_length);
/*
* Check the length of the packet and the block types.
*/
size = 0;
payload = mp->b_cont;
for (bp = payload; bp != NULL; bp = bp->b_cont) {
if (DB_TYPE(bp) != M_DATA)
goto baddata;
size += MBLKL(bp);
}
mac_sdu_get(dsp->ds_mh, NULL, &max_sdu);
if (size > max_sdu)
goto baddata;
/*
* Build a packet header.
*/
if ((bp = dls_header(dsp, addr, sap, dlp->dl_priority.dl_max,
&payload)) == NULL) {
dl_err = DL_BADADDR;
goto failed;
}
/*
* We no longer need the M_PROTO header, so free it.
*/
freeb(mp);
/*
* Transfer the checksum offload information if it is present.
*/
hcksum_retrieve(payload, NULL, NULL, &start, &stuff, &end, &value,
&flags);
(void) hcksum_assoc(bp, NULL, NULL, start, stuff, end, value, flags, 0);
/*
* Link the payload onto the new header.
*/
ASSERT(bp->b_cont == NULL);
bp->b_cont = payload;
/*
* No lock can be held across modules and putnext()'s,
* which can happen here with the call from DLD_TX().
*/
if (DLD_TX(dsp, bp, 0, 0) != NULL) {
/* flow-controlled */
DLD_SETQFULL(dsp);
}
DLD_DATATHR_DCR(dsp);
return;
failed:
dlerrorack(q, mp, DL_UNITDATA_REQ, dl_err, 0);
DLD_DATATHR_DCR(dsp);
return;
baddata:
dluderrorind(q, mp, (void *)addr, len, DL_BADDATA, 0);
DLD_DATATHR_DCR(dsp);
}
/* ARGSUSED */
static int
vmxnet3s_tx_prepare_offload(vmxnet3s_softc_t *dp, vmxnet3s_offload_t *ol,
mblk_t *mp, int *to_copy)
{
int ret = 0;
uint32_t start;
uint32_t stuff;
uint32_t value;
uint32_t flags;
uint32_t lsoflags;
uint32_t mss;
ol->om = VMXNET3_OM_NONE;
ol->hlen = 0;
ol->msscof = 0;
hcksum_retrieve(mp, NULL, NULL, &start, &stuff, NULL, &value, &flags);
mac_lso_get(mp, &mss, &lsoflags);
if (lsoflags & HW_LSO)
flags |= HW_LSO;
if (flags) {
struct ether_vlan_header *eth = (void *) mp->b_rptr;
uint8_t ethlen;
if (eth->ether_tpid == htons(ETHERTYPE_VLAN))
ethlen = sizeof (struct ether_vlan_header);
else
ethlen = sizeof (struct ether_header);
if (flags & HCK_PARTIALCKSUM) {
ol->om = VMXNET3_OM_CSUM;
ol->hlen = start + ethlen;
ol->msscof = stuff + ethlen;
}
if (flags & HW_LSO) {
mblk_t *mblk = mp;
uint8_t *ip;
uint8_t *tcp;
uint8_t iplen;
uint8_t tcplen;
/*
* Copy e1000g's behavior:
* - Do not assume all the headers are in the same mblk.
* - Assume each header is always within one mblk.
* - Assume the ethernet header is in the first mblk.
*/
ip = mblk->b_rptr + ethlen;
if (ip >= mblk->b_wptr) {
mblk = mblk->b_cont;
ip = mblk->b_rptr;
}
iplen = IPH_HDR_LENGTH((ipha_t *)ip);
tcp = ip + iplen;
if (tcp >= mblk->b_wptr) {
mblk = mblk->b_cont;
tcp = mblk->b_rptr;
}
tcplen = TCP_HDR_LENGTH((tcph_t *)tcp);
/* Careful, '>' instead of '>=' here */
if (tcp + tcplen > mblk->b_wptr)
mblk = mblk->b_cont;
ol->om = VMXNET3_OM_TSO;
ol->hlen = ethlen + iplen + tcplen;
ol->msscof = DB_LSOMSS(mp);
if (mblk != mp)
ret = ol->hlen;
else
*to_copy = ol->hlen;
}
}
return (ret);
}
