xref: /illumos-gate/usr/src/uts/intel/io/vmxnet3s/vmxnet3_tx.c (revision 62dadd65)

/*
 * Copyright (C) 2007 VMware, Inc. All rights reserved.
 *
 * The contents of this file are subject to the terms of the Common
 * Development and Distribution License (the "License") version 1.0
 * and no later version.  You may not use this file except in
 * compliance with the License.
 *
 * You can obtain a copy of the License at
 *         http://www.opensource.org/licenses/cddl1.php
 *
 * See the License for the specific language governing permissions
 * and limitations under the License.
 */

/*
 * Copyright (c) 2012 by Delphix. All rights reserved.
 */

#include <vmxnet3.h>

typedef enum vmxnet3_txstatus {
	VMXNET3_TX_OK,
	VMXNET3_TX_FAILURE,
	VMXNET3_TX_PULLUP,
	VMXNET3_TX_RINGFULL
} vmxnet3_txstatus;

typedef struct vmxnet3_offload_t {
	uint16_t om;
	uint16_t hlen;
	uint16_t msscof;
} vmxnet3_offload_t;

/*
 * vmxnet3_txqueue_init --
 *
 *    Initialize a TxQueue. Currently nothing needs to be done.
 *
 * Results:
 *    DDI_SUCCESS.
 *
 * Side effects:
 *    None.
 */
/* ARGSUSED */
int
vmxnet3_txqueue_init(vmxnet3_softc_t *dp, vmxnet3_txqueue_t *txq)
{
	return (DDI_SUCCESS);
}

/*
 * vmxnet3_txqueue_fini --
 *
 *    Finish a TxQueue by freeing all pending Tx.
 *
 * Results:
 *    DDI_SUCCESS.
 *
 * Side effects:
 *    None.
 */
void
vmxnet3_txqueue_fini(vmxnet3_softc_t *dp, vmxnet3_txqueue_t *txq)
{
	unsigned int i;

	ASSERT(!dp->devEnabled);

	for (i = 0; i < txq->cmdRing.size; i++) {
		mblk_t *mp = txq->metaRing[i].mp;
		if (mp) {
			freemsg(mp);
		}
	}
}

/*
 * 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, lsoflags, 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);
		}

		VMXNET3_DEBUG(dp, 4, "flags=0x%x, ethLen=%u, start=%u, "
		    "stuff=%u, value=%u\n", flags, ethLen, start, stuff, value);

		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, *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);
			/* 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;
			}
		}
	}

	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)
{
	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)) {
					VMXNET3_DEBUG(dp, 2,
					    "overfragmented mp (%u)\n", frags);
					(void) ddi_dma_unbind_handle(
					    dp->txDmaHandle);
					ret = VMXNET3_TX_PULLUP;
					goto error;
				}
				if (cmdRing->avail - frags <= 1) {
					dp->txMustResched = B_TRUE;
					(void) 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);

		(void) 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: M_PROTO mblks could be passed 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);
			atomic_inc_32(&dp->tx_pullup_needed);
			freemsg(mp);
			if (new_mp) {
				mp = new_mp;
			} else {
				atomic_inc_32(&dp->tx_pullup_failed);
				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);
		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);
				atomic_inc_32(&dp->tx_pullup_needed);
				freemsg(mp);
				if (new_mp) {
					mp = new_mp;
					status = vmxnet3_tx_one(dp, txq, &ol,
					    mp);
				} else {
					atomic_inc_32(&dp->tx_pullup_failed);
					continue;
				}
			}
		}
		if (status != VMXNET3_TX_OK && status != VMXNET3_TX_RINGFULL) {
			/* Fatal failure, drop it */
			atomic_inc_32(&dp->tx_error);
			freemsg(mp);
		}
	} while (mps && status != VMXNET3_TX_RINGFULL);

	if (status == VMXNET3_TX_RINGFULL) {
		atomic_inc_32(&dp->tx_ring_full);
		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);
}

/*
 * vmxnet3_tx_complete --
 *
 *    Parse a transmit queue and complete packets.
 *
 * Results:
 *    B_TRUE if Tx must be updated or B_FALSE if no action is required.
 *
 * Side effects:
 *    None.
 */
boolean_t
vmxnet3_tx_complete(vmxnet3_softc_t *dp, vmxnet3_txqueue_t *txq)
{
	vmxnet3_cmdring_t *cmdRing = &txq->cmdRing;
	vmxnet3_compring_t *compRing = &txq->compRing;
	Vmxnet3_GenericDesc *compDesc;
	boolean_t completedTx = B_FALSE;
	boolean_t ret = B_FALSE;

	mutex_enter(&dp->txLock);

	compDesc = VMXNET3_GET_DESC(compRing, compRing->next2comp);
	while (compDesc->tcd.gen == compRing->gen) {
		vmxnet3_metatx_t *sopMetaDesc, *eopMetaDesc;
		uint16_t sopIdx, eopIdx;
		mblk_t *mp;

		eopIdx = compDesc->tcd.txdIdx;
		eopMetaDesc = &txq->metaRing[eopIdx];
		sopIdx = eopMetaDesc->sopIdx;
		sopMetaDesc = &txq->metaRing[sopIdx];

		ASSERT(eopMetaDesc->frags);
		cmdRing->avail += eopMetaDesc->frags;

		ASSERT(sopMetaDesc->mp);
		mp = sopMetaDesc->mp;
		freemsg(mp);

		eopMetaDesc->sopIdx = 0;
		eopMetaDesc->frags = 0;
		sopMetaDesc->mp = NULL;

		completedTx = B_TRUE;

		VMXNET3_DEBUG(dp, 3, "cp 0x%p on [%u;%u]\n", mp, sopIdx,
		    eopIdx);

		VMXNET3_INC_RING_IDX(compRing, compRing->next2comp);
		compDesc = VMXNET3_GET_DESC(compRing, compRing->next2comp);
	}

	if (dp->txMustResched && completedTx) {
		dp->txMustResched = B_FALSE;
		ret = B_TRUE;
	}

	mutex_exit(&dp->txLock);

	return (ret);
}