xref: /illumos-gate/usr/src/uts/common/io/i40e/i40e_transceiver.c (revision ca28c3d8eab8b53ff145fd15cf80cdc2da3fc032)

/*
 * This file and its contents are supplied under the terms of the
 * Common Development and Distribution License ("CDDL"), version 1.0.
 * You may only use this file in accordance with the terms of version
 * 1.0 of the CDDL.
 *
 * A full copy of the text of the CDDL should have accompanied this
 * source.  A copy of the CDDL is also available via the Internet at
 * http://www.illumos.org/license/CDDL.
 */

/*
 * Copyright 2015 OmniTI Computer Consulting, Inc. All rights reserved.
 * Copyright 2019 Joyent, Inc.
 * Copyright 2020 RackTop Systems, Inc.
 */

#include "i40e_sw.h"

/*
 * ---------------------------------------------------------
 * Buffer and Memory Management, Receiving, and Transmitting
 * ---------------------------------------------------------
 *
 * Each physical function (PF), which is what we think of as an instance of the
 * device driver, has a series of associated transmit and receive queue pairs.
 * Effectively, what we think of in MAC as rings. Each of these has their own
 * ring of descriptors which is used as part of doing DMA activity.
 *
 * The transmit ring of descriptors are 16-byte entries which are used to send
 * packets, program filters, etc. The receive ring of descriptors are either
 * 16-byte or 32-bytes each. At the moment, we opt to use the larger descriptor
 * format so that we're in a better position if we ever want to leverage that
 * information later on.
 *
 * However, these rings are just for descriptors, they don't talk or deal with
 * how we actually store the memory that we need for DMA or the associated
 * information that we need for keeping track of message blocks. To correspond
 * to the hardware descriptor ring which is how we communicate with hardware, we
 * introduce a control block which keeps track of our required metadata like DMA
 * mappings.
 *
 * There are two main considerations that dictate how much memory and buffers
 * we end up allocating. Those are:
 *
 *   o The size of the ring (controlled through the driver.conf file)
 *
 *   o The maximum size frame we can receive.
 *
 * The size of the rings currently defaults to 1024 descriptors and is stored in
 * the i40e_t`i40e_rx_ring_size and i40e_t`i40e_tx_ring_size.
 *
 * While the size of the rings is controlled by the driver.conf, the maximum
 * size frame is informed primarily through the use of dladm and the setting of
 * the MTU property on the device. From the MTU, we then go and do some
 * machinations. The first thing we do is we then have to add in space for the
 * Ethernet header, potentially a VLAN header, and the FCS check. This value is
 * what's stored as i40e_t`i40e_frame_max and is derived any time
 * i40e_t`i40e_sdu changes.
 *
 * This size is then rounded up to the nearest 1k chunk, which represents the
 * actual amount of memory that we'll allocate for a single frame.
 *
 * Note, that for RX, we do something that might be unexpected. We always add
 * an extra two bytes to the frame size that we allocate. We then offset the DMA
 * address that we receive a packet into by two bytes. This ensures that the IP
 * header will always be 4 byte aligned because the MAC header is either 14 or
 * 18 bytes in length, depending on the use of 802.1Q tagging, which makes IP's
 * and MAC's lives easier.
 *
 * Both the RX and TX descriptor rings (which are what we use to communicate
 * with hardware) are allocated as a single region of DMA memory which is the
 * size of the descriptor (4 bytes and 2 bytes respectively) times the total
 * number of descriptors for an RX and TX ring.
 *
 * While the RX and TX descriptors are allocated using DMA-based memory, the
 * control blocks for each of them are allocated using normal kernel memory.
 * They aren't special from a DMA perspective. We'll go over the design of both
 * receiving and transmitting separately, as they have slightly different
 * control blocks and different ways that we manage the relationship between
 * control blocks and descriptors.
 *
 * ---------------------------------
 * RX Descriptors and Control Blocks
 * ---------------------------------
 *
 * For every descriptor in the ring that the driver has, we need some associated
 * memory, which means that we need to have the receive specific control block.
 * We have a couple different, but related goals:
 *
 *   o Once we've completed the mc_start GLDv3 endpoint (i40e_m_start), we do
 *     not want to do any additional memory allocations or DMA allocations if
 *     we don't have to.
 *
 *   o We'd like to try and do as much zero-copy as possible, while taking into
 *     account the cost of mapping in DMA resources.
 *
 *   o We'd like to have every receive descriptor available.
 *
 * Now, these rules are a bit in tension with one another. The act of mapping in
 * is an exercise of trying to find the break-even point between page table
 * updates and bcopy. We currently start by using the same metrics that ixgbe
 * used; however, it should be known that this value has effectively been
 * cargo-culted across to yet another driver, sorry.
 *
 * If we receive a packet which is larger than our copy threshold, we'll create
 * a message block out of the DMA memory via desballoc(9F) and send that up to
 * MAC that way. This will cause us to be notified when the message block is
 * then freed because it has been consumed, dropped, or otherwise. Otherwise, if
 * it's less than the threshold, we'll try to use allocb and bcopy it into the
 * block, thus allowing us to immediately reuse the DMA resource. Note, on debug
 * builds, we allow someone to whack the variable i40e_debug_rx_mode to override
 * the behavior and always do a bcopy or a DMA bind.
 *
 * To try and ensure that the device always has blocks that it can receive data
 * into, we maintain two lists of control blocks, a working list and a free
 * list. Each list is sized equal to the number of descriptors in the RX ring.
 * During the GLDv3 mc_start routine, we allocate a number of RX control blocks
 * equal to twice the number of descriptors in the ring and we assign them
 * equally to the free list and to the working list. Each control block also has
 * DMA memory allocated and associated with which it will be used to receive the
 * actual packet data. All of a received frame's data will end up in a single
 * DMA buffer.
 *
 * During operation, we always maintain the invariant that each RX descriptor
 * has an associated RX control block which lives in the working list. If we
 * feel that we should loan up DMA memory to MAC in the form of a message block,
 * we can only do so if we can maintain this invariant. To do that, we swap in
 * one of the buffers from the free list. If none are available, then we resort
 * to using allocb(9F) and bcopy(9F) on the packet instead, regardless of the
 * size.
 *
 * Loaned message blocks come back to use when freemsg(9F) or freeb(9F) is
 * called on the block, at which point we restore the RX control block to the
 * free list and are able to reuse the DMA memory again. While the scheme may
 * seem odd, it importantly keeps us out of trying to do any DMA allocations in
 * the normal path of operation, even though we may still have to allocate
 * message blocks and copy.
 *
 * The following state machine describes the life time of a RX control block. In
 * the diagram we abbrviate the RX ring descriptor entry as rxd and the rx
 * control block entry as rcb.
 *
 *             |                                   |
 *             * ... 1/2 of all initial rcb's  ... *
 *             |                                   |
 *             v                                   v
 *     +------------------+               +------------------+
 *     | rcb on free list |---*---------->| rcb on work list |
 *     +------------------+   .           +------------------+
 *             ^              . moved to          |
 *             |                replace rcb       * . . Frame received,
 *             |                loaned to         |     entry on free list
 *             |                MAC + co.         |     available. rcb's
 *             |                                  |     memory made into mblk_t
 *             * . freemsg(9F)                    |     and sent up to MAC.
 *             |   called on                      |
 *             |   loaned rcb                     |
 *             |   and it is                      v
 *             |   recycled.              +-------------------+
 *             +--------------------<-----| rcb loaned to MAC |
 *                                        +-------------------+
 *
 * Finally, note that every RX control block has a reference count on it. One
 * reference is added as long as the driver has had the GLDv3 mc_start endpoint
 * called. If the GLDv3 mc_stop entry point is called, IP has been unplumbed and
 * no other DLPI consumers remain, then we'll decrement the reference count by
 * one. Whenever we loan up the RX control block and associated buffer to MAC,
 * then we bump the reference count again. Even though the device is stopped,
 * there may still be loaned frames in upper levels that we'll want to account
 * for. Our callback from freemsg(9F)/freeb(9F) will take care of making sure
 * that it is cleaned up.
 *
 * --------------------
 * Managing the RX Ring
 * --------------------
 *
 * The receive ring descriptors are arranged in a circular buffer with a head
 * and tail pointer. There are both the conventional head and tail pointers
 * which are used to partition the ring into two portions, a portion that we,
 * the operating system, manage and a portion that is managed by hardware. When
 * hardware owns a descriptor in the ring, it means that it is waiting for data
 * to be filled in. However, when a portion of the ring is owned by the driver,
 * then that means that the descriptor has been consumed and we need to go take
 * a look at it.
 *
 * The initial head is configured to be zero by writing it as such in the
 * receive queue context in the FPM (function private memory from the host). The
 * initial tail is written to be the last descriptor. This is written to via the
 * PCIe register I40E_QRX_TAIL(). Technically, hardware owns everything between
 * the HEAD and TAIL, inclusive. Note that while we initially program the HEAD,
 * the only values we ever consult ourselves are the TAIL register and our own
 * state tracking. Effectively, we cache the HEAD register and then update it
 * ourselves based on our work.
 *
 * When we iterate over the RX descriptors and thus the received frames, we are
 * either in an interrupt context or we've been asked by MAC to poll on the
 * ring. If we've been asked to poll on the ring, we have a maximum number of
 * bytes of mblk_t's to return. If processing an RX descriptor would cause us to
 * exceed that count, then we do not process it. When in interrupt context, we
 * don't have a strict byte count. However, to ensure liveness, we limit the
 * amount of data based on a configuration value
 * (i40e_t`i40e_rx_limit_per_intr). The number that we've started with for this
 * is based on similar numbers that are used for ixgbe. After some additional
 * time in the field, we'll have a sense as to whether or not it should be
 * changed.
 *
 * When processing, we start at our own HEAD pointer
 * (i40e_rx_data_t`rxd_desc_next), which indicates the descriptor to start
 * processing. Every RX descriptor has what's described as the DD bit. This bit
 * (the LSB of the second 8-byte word), indicates whether or not the descriptor
 * is done.  When we give descriptors to the hardware, this value is always
 * zero. When the hardware has finished a descriptor, it will always be one.
 *
 * The first thing that we check is whether the DD bit indicates that the
 * current HEAD is ready. If it isn't, then we're done. That's the primary
 * invariant of processing a frame. If it's done, then there are a few other
 * things that we want to look at. In the same status word as the DD bit, there
 * are two other important bits:
 *
 *   o End of Packet (EOP)
 *   o Error bits
 *
 * The end of packet indicates that we have reached the last descriptor. Now,
 * you might ask when would there be more than one descriptor. The reason for
 * that might be due to large receive offload (lro) or header splitting
 * functionality, which presently isn't supported in the driver. The error bits
 * in the frame are only valid when EOP is set.
 *
 * If error bits are set on the frame, then we still consume it; however, we
 * will not generate an mblk_t to send up to MAC. If there are no error bits
 * set, then we'll consume the descriptor either using bcopy or DMA binding. See
 * the earlier section 'RX DESCRIPTORS AND CONTROL BLOCKS' for more information
 * on how that selection is made.
 *
 * Regardless of whether we construct an mblk_t or encounter an error, we end up
 * resetting the descriptor. This re-arms the descriptor for hardware and in the
 * process, we may end up assigning it a new receive control bock. After we do
 * this, we always update our HEAD pointer, no matter what.
 *
 * Finally, once we've consumed as much as we will in a given window, we go and
 * update the TAIL register to indicate all the frames we've consumed. We only
 * do a single bulk write for the ring.
 *
 * ---------------------------------
 * TX Descriptors and Control Blocks
 * ---------------------------------
 *
 * While the transmit path is similar in spirit to the receive path, it works
 * differently due to the fact that all data is originated by the operating
 * system and not by the device.
 *
 * Like RX, there is both a descriptor ring that we use to communicate to the
 * driver and which points to the memory used to transmit a frame.  Similarly,
 * there is a corresponding transmit control block, however, the correspondence
 * between descriptors and control blocks is more complex and not necessarily
 * 1-to-1.
 *
 * The driver is asked to process a single frame at a time. That message block
 * may be made up of multiple fragments linked together by the mblk_t`b_cont
 * member. The device has a hard limit of up to 8 buffers being allowed for use
 * for a single non-LSO packet or LSO segment. The number of TX ring entires
 * (and thus TX control blocks) used depends on the fragment sizes and DMA
 * layout, as explained below.
 *
 * We alter our DMA strategy based on a threshold tied to the fragment size.
 * This threshold is configurable via the tx_dma_threshold property. If the
 * fragment is above the threshold, we DMA bind it -- consuming one TCB and
 * potentially several data descriptors. The exact number of descriptors (equal
 * to the number of DMA cookies) depends on page size, MTU size, b_rptr offset
 * into page, b_wptr offset into page, and the physical layout of the dblk's
 * memory (contiguous or not). Essentially, we are at the mercy of the DMA
 * engine and the dblk's memory allocation. Knowing the exact number of
 * descriptors up front is a task best not taken on by the driver itself.
 * Instead, we attempt to DMA bind the fragment and verify the descriptor
 * layout meets hardware constraints. If the proposed DMA bind does not satisfy
 * the hardware constaints, then we discard it and instead copy the entire
 * fragment into the pre-allocated TCB buffer (or buffers if the fragment is
 * larger than the TCB buffer).
 *
 * If the fragment is below or at the threshold, we copy it to the pre-allocated
 * buffer of a TCB. We compress consecutive copy fragments into a single TCB to
 * conserve resources. We are guaranteed that the TCB buffer is made up of only
 * 1 DMA cookie; and therefore consumes only one descriptor on the controller.
 *
 * Furthermore, if the frame requires HW offloads such as LSO, tunneling or
 * filtering, then the TX data descriptors must be preceeded by a single TX
 * context descriptor.  Because there is no DMA transfer associated with the
 * context descriptor, we allocate a control block with a special type which
 * indicates to the TX ring recycle code that there are no associated DMA
 * resources to unbind when the control block is free'd.
 *
 * If we don't have enough space in the ring or TX control blocks available,
 * then we'll return the unprocessed message block to MAC. This will induce flow
 * control and once we recycle enough entries, we'll once again enable sending
 * on the ring.
 *
 * We size the working list as equal to the number of descriptors in the ring.
 * We size the free list as equal to 1.5 times the number of descriptors in the
 * ring. We'll allocate a number of TX control block entries equal to the number
 * of entries in the free list. By default, all entries are placed in the free
 * list. As we come along and try to send something, we'll allocate entries from
 * the free list and add them to the working list, where they'll stay until the
 * hardware indicates that all of the data has been written back to us. The
 * reason that we start with 1.5x is to help facilitate having more than one TX
 * buffer associated with the DMA activity.
 *
 * --------------------
 * Managing the TX Ring
 * --------------------
 *
 * The transmit descriptor ring is driven by us. We maintain our own notion of a
 * HEAD and TAIL register and we update the hardware with updates to the TAIL
 * register. When the hardware is done writing out data, it updates us by
 * writing back to a specific address, not by updating the individual
 * descriptors. That address is a 4-byte region after the main transmit
 * descriptor ring. This is why the descriptor ring has an extra descriptor's
 * worth allocated to it.
 *
 * We maintain our notion of the HEAD in the i40e_trqpair_t`itrq_desc_head and
 * the TAIL in the i40e_trqpair_t`itrq_desc_tail. When we write out frames,
 * we'll update the tail there and in the I40E_QTX_TAIL() register. At various
 * points in time, through both interrupts, and our own internal checks, we'll
 * sync the write-back head portion of the DMA space. Based on the index it
 * reports back, we'll free everything between our current HEAD and the
 * indicated index and update HEAD to the new index.
 *
 * When a frame comes in, we try to use a number of transmit control blocks and
 * we'll transition them from the free list to the work list. They'll get moved
 * to the entry on the work list that corresponds with the transmit descriptor
 * they correspond to. Once we are indicated that the corresponding descriptor
 * has been freed, we'll return it to the list.
 *
 * The transmit control block free list is managed by keeping track of the
 * number of entries in it, i40e_trqpair_t`itrq_tcb_free. We use it as a way to
 * index into the free list and add things to it. In effect, we always push and
 * pop from the tail and protect it with a single lock,
 * i40e_trqpair_t`itrq_tcb_lock. This scheme is somewhat simplistic and may not
 * stand up to further performance testing; however, it does allow us to get off
 * the ground with the device driver.
 *
 * The following image describes where a given transmit control block lives in
 * its lifetime:
 *
 *             |
 *             * ... Initial placement for all tcb's
 *             |
 *             v
 *    +------------------+                       +------------------+
 *    | tcb on free list |---*------------------>| tcb on work list |
 *    +------------------+   .                   +------------------+
 *             ^             . N tcbs allocated[1]         |
 *             |               to send frame               v
 *             |               or fragment on              |
 *             |               wire, mblk from             |
 *             |               MAC associated.             |
 *             |                                           |
 *             +------*-------------------------------<----+
 *                    .
 *                    . Hardware indicates
 *                      entry transmitted.
 *                      tcbs recycled, mblk
 *                      from MAC freed.
 *
 * [1] We allocate N tcbs to transmit a single frame where N can be 1 context
 *     descriptor plus 1 data descriptor, in the non-DMA-bind case.  In the DMA
 *     bind case, N can be 1 context descriptor plus 1 data descriptor per
 *     b_cont in the mblk.  In this case, the mblk is associated with the first
 *     data descriptor and freed as part of freeing that data descriptor.
 *
 * ------------
 * Blocking MAC
 * ------------
 *
 * When performing transmit, we can run out of descriptors and ring entries.
 * When such a case happens, we return the mblk_t to MAC to indicate that we've
 * been blocked. At that point in time, MAC becomes blocked and will not
 * transmit anything out that specific ring until we notify MAC. To indicate
 * that we're in such a situation we set i40e_trqpair_t`itrq_tx_blocked member
 * to B_TRUE.
 *
 * When we recycle TX descriptors then we'll end up signaling MAC by calling
 * mac_tx_ring_update() if we were blocked, letting it know that it's safe to
 * start sending frames out to us again.
 */

/*
 * We set our DMA alignment requests based on the smallest supported page size
 * of the corresponding platform.
 */
#if	defined(__sparc)
#define	I40E_DMA_ALIGNMENT 0x2000ull
#elif defined(__x86)
#define	I40E_DMA_ALIGNMENT 0x1000ull
#else
#error	"unknown architecture for i40e"
#endif

/*
 * This structure is used to maintain information and flags related to
 * transmitting a frame.  These fields are ultimately used to construct the
 * TX data descriptor(s) and, if necessary, the TX context descriptor.
 */
typedef struct i40e_tx_context {
	enum i40e_tx_desc_cmd_bits	itc_data_cmdflags;
	uint32_t			itc_data_offsets;
	enum i40e_tx_ctx_desc_cmd_bits	itc_ctx_cmdflags;
	uint32_t			itc_ctx_tsolen;
	uint32_t			itc_ctx_mss;
} i40e_tx_context_t;

/*
 * Toggles on debug builds which can be used to override our RX behaviour based
 * on thresholds.
 */
#ifdef	DEBUG
typedef enum {
	I40E_DEBUG_RX_DEFAULT	= 0,
	I40E_DEBUG_RX_BCOPY	= 1,
	I40E_DEBUG_RX_DMABIND	= 2
} i40e_debug_rx_t;

i40e_debug_rx_t i40e_debug_rx_mode = I40E_DEBUG_RX_DEFAULT;
#endif	/* DEBUG */

/*
 * Notes on the following pair of DMA attributes. The first attribute,
 * i40e_static_dma_attr, is designed to be used for both the descriptor rings
 * and the static buffers that we associate with control blocks. For this
 * reason, we force an SGL length of one. While technically the driver supports
 * a larger SGL (5 on RX and 8 on TX), we opt to only use one to simplify our
 * management here. In addition, when the Intel common code wants to allocate
 * memory via the i40e_allocate_virt_mem osdep function, we have it leverage
 * the static dma attr.
 *
 * The latter two sets of attributes, are what we use when we're binding a
 * bunch of mblk_t fragments to go out the door. Note that the main difference
 * here is that we're allowed a larger SGL length.  For non-LSO TX, we
 * restrict the SGL length to match the number of TX buffers available to the
 * PF (8).  For the LSO case we can go much larger, with the caveat that each
 * MSS-sized chunk (segment) must not span more than 8 data descriptors and
 * hence must not span more than 8 cookies.
 *
 * Note, we default to setting ourselves to be DMA capable here. However,
 * because we could have multiple instances which have different FMA error
 * checking capabilities, or end up on different buses, we make these static
 * and const and copy them into the i40e_t for the given device with the actual
 * values that reflect the actual capabilities.
 */
static const ddi_dma_attr_t i40e_g_static_dma_attr = {
	DMA_ATTR_V0,			/* version number */
	0x0000000000000000ull,		/* low address */
	0xFFFFFFFFFFFFFFFFull,		/* high address */
	0x00000000FFFFFFFFull,		/* dma counter max */
	I40E_DMA_ALIGNMENT,		/* alignment */
	0x00000FFF,			/* burst sizes */
	0x00000001,			/* minimum transfer size */
	0x00000000FFFFFFFFull,		/* maximum transfer size */
	0xFFFFFFFFFFFFFFFFull,		/* maximum segment size */
	1,				/* scatter/gather list length */
	0x00000001,			/* granularity */
	DDI_DMA_FLAGERR			/* DMA flags */
};

static const ddi_dma_attr_t i40e_g_txbind_dma_attr = {
	DMA_ATTR_V0,			/* version number */
	0x0000000000000000ull,		/* low address */
	0xFFFFFFFFFFFFFFFFull,		/* high address */
	I40E_MAX_TX_BUFSZ - 1,		/* dma counter max */
	I40E_DMA_ALIGNMENT,		/* alignment */
	0x00000FFF,			/* burst sizes */
	0x00000001,			/* minimum transfer size */
	0x00000000FFFFFFFFull,		/* maximum transfer size */
	0xFFFFFFFFFFFFFFFFull,		/* maximum segment size	 */
	I40E_TX_MAX_COOKIE,		/* scatter/gather list length */
	0x00000001,			/* granularity */
	DDI_DMA_FLAGERR			/* DMA flags */
};

static const ddi_dma_attr_t i40e_g_txbind_lso_dma_attr = {
	DMA_ATTR_V0,			/* version number */
	0x0000000000000000ull,		/* low address */
	0xFFFFFFFFFFFFFFFFull,		/* high address */
	I40E_MAX_TX_BUFSZ - 1,		/* dma counter max */
	I40E_DMA_ALIGNMENT,		/* alignment */
	0x00000FFF,			/* burst sizes */
	0x00000001,			/* minimum transfer size */
	0x00000000FFFFFFFFull,		/* maximum transfer size */
	0xFFFFFFFFFFFFFFFFull,		/* maximum segment size	 */
	I40E_TX_LSO_MAX_COOKIE,		/* scatter/gather list length */
	0x00000001,			/* granularity */
	DDI_DMA_FLAGERR			/* DMA flags */
};

/*
 * Next, we have the attributes for these structures. The descriptor rings are
 * all strictly little endian, while the data buffers are just arrays of bytes
 * representing frames. Because of this, we purposefully simplify the driver
 * programming life by programming the descriptor ring as little endian, while
 * for the buffer data we keep it as unstructured.
 *
 * Note, that to keep the Intel common code operating in a reasonable way, when
 * we allocate DMA memory for it, we do not use byte swapping and thus use the
 * standard i40e_buf_acc_attr.
 */
static const ddi_device_acc_attr_t i40e_g_desc_acc_attr = {
	DDI_DEVICE_ATTR_V0,
	DDI_STRUCTURE_LE_ACC,
	DDI_STRICTORDER_ACC
};

static const ddi_device_acc_attr_t i40e_g_buf_acc_attr = {
	DDI_DEVICE_ATTR_V0,
	DDI_NEVERSWAP_ACC,
	DDI_STRICTORDER_ACC
};

/*
 * The next two functions are designed to be type-safe versions of macros that
 * are used to increment and decrement a descriptor index in the loop. Note,
 * these are marked inline to try and keep the data path hot and they were
 * effectively inlined in their previous life as macros.
 */
static inline int
i40e_next_desc(int base, int count, int size)
{
	int out;

	ASSERT(base >= 0);
	ASSERT(count > 0);
	ASSERT(size > 0);

	if (base + count < size) {
		out = base + count;
	} else {
		out = base + count - size;
	}

	ASSERT(out >= 0 && out < size);
	return (out);
}

static inline int
i40e_prev_desc(int base, int count, int size)
{
	int out;

	ASSERT(base >= 0);
	ASSERT(count > 0);
	ASSERT(size > 0);

	if (base >= count) {
		out = base - count;
	} else {
		out = base - count + size;
	}

	ASSERT(out >= 0 && out < size);
	return (out);
}

/*
 * Free DMA memory that is represented by a i40e_dma_buffer_t.
 */
static void
i40e_free_dma_buffer(i40e_dma_buffer_t *dmap)
{
	if (dmap->dmab_dma_address != 0) {
		VERIFY(dmap->dmab_dma_handle != NULL);
		(void) ddi_dma_unbind_handle(dmap->dmab_dma_handle);
		dmap->dmab_dma_address = 0;
		dmap->dmab_size = 0;
	}

	if (dmap->dmab_acc_handle != NULL) {
		ddi_dma_mem_free(&dmap->dmab_acc_handle);
		dmap->dmab_acc_handle = NULL;
		dmap->dmab_address = NULL;
	}

	if (dmap->dmab_dma_handle != NULL) {
		ddi_dma_free_handle(&dmap->dmab_dma_handle);
		dmap->dmab_dma_handle = NULL;
	}

	/*
	 * These should only be set if we have valid handles allocated and
	 * therefore should always be NULLed out due to the above code. This
	 * is here to catch us acting sloppy.
	 */
	ASSERT(dmap->dmab_dma_address == 0);
	ASSERT(dmap->dmab_address == NULL);
	ASSERT(dmap->dmab_size == 0);
	dmap->dmab_len = 0;
}

/*
 * Allocate size bytes of DMA memory based on the passed in attributes. This
 * fills in the information in dmap and is designed for all of our single cookie
 * allocations.
 */
static boolean_t
i40e_alloc_dma_buffer(i40e_t *i40e, i40e_dma_buffer_t *dmap,
    ddi_dma_attr_t *attrsp, ddi_device_acc_attr_t *accp, boolean_t stream,
    boolean_t zero, size_t size)
{
	int ret;
	uint_t flags;
	size_t len;
	ddi_dma_cookie_t cookie;
	uint_t ncookies;

	if (stream == B_TRUE)
		flags = DDI_DMA_STREAMING;
	else
		flags = DDI_DMA_CONSISTENT;

	/*
	 * Step one: Allocate the DMA handle
	 */
	ret = ddi_dma_alloc_handle(i40e->i40e_dip, attrsp, DDI_DMA_DONTWAIT,
	    NULL, &dmap->dmab_dma_handle);
	if (ret != DDI_SUCCESS) {
		i40e_error(i40e, "failed to allocate dma handle for I/O "
		    "buffers: %d", ret);
		dmap->dmab_dma_handle = NULL;
		return (B_FALSE);
	}

	/*
	 * Step two: Allocate the DMA memory
	 */
	ret = ddi_dma_mem_alloc(dmap->dmab_dma_handle, size, accp, flags,
	    DDI_DMA_DONTWAIT, NULL, &dmap->dmab_address, &len,
	    &dmap->dmab_acc_handle);
	if (ret != DDI_SUCCESS) {
		i40e_error(i40e, "failed to allocate %ld bytes of DMA for I/O "
		    "buffers", size);
		dmap->dmab_address = NULL;
		dmap->dmab_acc_handle = NULL;
		i40e_free_dma_buffer(dmap);
		return (B_FALSE);
	}

	/*
	 * Step three: Optionally zero
	 */
	if (zero == B_TRUE)
		bzero(dmap->dmab_address, len);

	/*
	 * Step four: Bind the memory
	 */
	ret = ddi_dma_addr_bind_handle(dmap->dmab_dma_handle, NULL,
	    dmap->dmab_address, len, DDI_DMA_RDWR | flags, DDI_DMA_DONTWAIT,
	    NULL, &cookie, &ncookies);
	if (ret != DDI_DMA_MAPPED) {
		i40e_error(i40e, "failed to allocate %ld bytes of DMA for I/O "
		    "buffers: %d", size, ret);
		i40e_free_dma_buffer(dmap);
		return (B_FALSE);
	}

	VERIFY(ncookies == 1);
	dmap->dmab_dma_address = cookie.dmac_laddress;
	dmap->dmab_size = len;
	dmap->dmab_len = 0;
	return (B_TRUE);
}

/*
 * This function is called once the last pending rcb has been freed by the upper
 * levels of the system.
 */
static void
i40e_free_rx_data(i40e_rx_data_t *rxd)
{
	VERIFY(rxd->rxd_rcb_pending == 0);

	if (rxd->rxd_rcb_area != NULL) {
		kmem_free(rxd->rxd_rcb_area,
		    sizeof (i40e_rx_control_block_t) *
		    (rxd->rxd_free_list_size + rxd->rxd_ring_size));
		rxd->rxd_rcb_area = NULL;
	}

	if (rxd->rxd_free_list != NULL) {
		kmem_free(rxd->rxd_free_list,
		    sizeof (i40e_rx_control_block_t *) *
		    rxd->rxd_free_list_size);
		rxd->rxd_free_list = NULL;
	}

	if (rxd->rxd_work_list != NULL) {
		kmem_free(rxd->rxd_work_list,
		    sizeof (i40e_rx_control_block_t *) *
		    rxd->rxd_ring_size);
		rxd->rxd_work_list = NULL;
	}

	kmem_free(rxd, sizeof (i40e_rx_data_t));
}

static boolean_t
i40e_alloc_rx_data(i40e_t *i40e, i40e_trqpair_t *itrq)
{
	i40e_rx_data_t *rxd;

	rxd = kmem_zalloc(sizeof (i40e_rx_data_t), KM_NOSLEEP);
	if (rxd == NULL)
		return (B_FALSE);
	itrq->itrq_rxdata = rxd;
	rxd->rxd_i40e = i40e;

	rxd->rxd_ring_size = i40e->i40e_rx_ring_size;
	rxd->rxd_free_list_size = i40e->i40e_rx_ring_size;

	rxd->rxd_rcb_free = rxd->rxd_free_list_size;

	rxd->rxd_work_list = kmem_zalloc(sizeof (i40e_rx_control_block_t *) *
	    rxd->rxd_ring_size, KM_NOSLEEP);
	if (rxd->rxd_work_list == NULL) {
		i40e_error(i40e, "failed to allocate RX work list for a ring "
		    "of %d entries for ring %d", rxd->rxd_ring_size,
		    itrq->itrq_index);
		goto cleanup;
	}

	rxd->rxd_free_list = kmem_zalloc(sizeof (i40e_rx_control_block_t *) *
	    rxd->rxd_free_list_size, KM_NOSLEEP);
	if (rxd->rxd_free_list == NULL) {
		i40e_error(i40e, "failed to allocate a %d entry RX free list "
		    "for ring %d", rxd->rxd_free_list_size, itrq->itrq_index);
		goto cleanup;
	}

	rxd->rxd_rcb_area = kmem_zalloc(sizeof (i40e_rx_control_block_t) *
	    (rxd->rxd_free_list_size + rxd->rxd_ring_size), KM_NOSLEEP);
	if (rxd->rxd_rcb_area == NULL) {
		i40e_error(i40e, "failed to allocate a %d entry rcb area for "
		    "ring %d", rxd->rxd_ring_size + rxd->rxd_free_list_size,
		    itrq->itrq_index);
		goto cleanup;
	}

	return (B_TRUE);

cleanup:
	i40e_free_rx_data(rxd);
	itrq->itrq_rxdata = NULL;
	return (B_FALSE);
}

/*
 * Free all of the memory that we've allocated for DMA. Note that we may have
 * buffers that we've loaned up to the OS which are still outstanding. We'll
 * always free up the descriptor ring, because we no longer need that. For each
 * rcb, we'll iterate over it and if we send the reference count to zero, then
 * we'll free the message block and DMA related resources. However, if we don't
 * take the last one, then we'll go ahead and keep track that we'll have pending
 * data and clean it up when we get there.
 */
static void
i40e_free_rx_dma(i40e_rx_data_t *rxd, boolean_t failed_init)
{
	uint32_t i, count, ref;

	i40e_rx_control_block_t *rcb;
	i40e_t *i40e = rxd->rxd_i40e;

	i40e_free_dma_buffer(&rxd->rxd_desc_area);
	rxd->rxd_desc_ring = NULL;
	rxd->rxd_desc_next = 0;

	mutex_enter(&i40e->i40e_rx_pending_lock);

	rcb = rxd->rxd_rcb_area;
	count = rxd->rxd_ring_size + rxd->rxd_free_list_size;

	for (i = 0; i < count; i++, rcb++) {
		VERIFY(rcb != NULL);

		/*
		 * If we're cleaning up from a failed creation attempt, then an
		 * entry may never have been assembled which would mean that
		 * it's reference count is zero. If we find that, we leave it
		 * be, because nothing else should be modifying it at this
		 * point. We're not at the point that any more references can be
		 * added, just removed.
		 */
		if (failed_init == B_TRUE && rcb->rcb_ref == 0)
			continue;

		ref = atomic_dec_32_nv(&rcb->rcb_ref);
		if (ref == 0) {
			freemsg(rcb->rcb_mp);
			rcb->rcb_mp = NULL;
			i40e_free_dma_buffer(&rcb->rcb_dma);
		} else {
			atomic_inc_32(&rxd->rxd_rcb_pending);
			atomic_inc_32(&i40e->i40e_rx_pending);
		}
	}
	mutex_exit(&i40e->i40e_rx_pending_lock);
}

/*
 * Initialize the DMA memory for the descriptor ring and for each frame in the
 * control block list.
 */
static boolean_t
i40e_alloc_rx_dma(i40e_rx_data_t *rxd)
{
	int i, count;
	size_t dmasz;
	i40e_rx_control_block_t *rcb;
	i40e_t *i40e = rxd->rxd_i40e;

	/*
	 * First allocate the RX descriptor ring.
	 */
	dmasz = sizeof (i40e_rx_desc_t) * rxd->rxd_ring_size;
	VERIFY(dmasz > 0);
	if (i40e_alloc_dma_buffer(i40e, &rxd->rxd_desc_area,
	    &i40e->i40e_static_dma_attr, &i40e->i40e_desc_acc_attr, B_FALSE,
	    B_TRUE, dmasz) == B_FALSE) {
		i40e_error(i40e, "failed to allocate DMA resources "
		    "for RX descriptor ring");
		return (B_FALSE);
	}
	rxd->rxd_desc_ring =
	    (i40e_rx_desc_t *)(uintptr_t)rxd->rxd_desc_area.dmab_address;
	rxd->rxd_desc_next = 0;

	count = rxd->rxd_ring_size + rxd->rxd_free_list_size;
	rcb = rxd->rxd_rcb_area;

	dmasz = i40e->i40e_rx_buf_size;
	VERIFY(dmasz > 0);
	for (i = 0; i < count; i++, rcb++) {
		i40e_dma_buffer_t *dmap;
		VERIFY(rcb != NULL);

		if (i < rxd->rxd_ring_size) {
			rxd->rxd_work_list[i] = rcb;
		} else {
			rxd->rxd_free_list[i - rxd->rxd_ring_size] = rcb;
		}

		dmap = &rcb->rcb_dma;
		if (i40e_alloc_dma_buffer(i40e, dmap,
		    &i40e->i40e_static_dma_attr, &i40e->i40e_buf_acc_attr,
		    B_TRUE, B_FALSE, dmasz) == B_FALSE) {
			i40e_error(i40e, "failed to allocate RX dma buffer");
			return (B_FALSE);
		}

		/*
		 * Initialize the control block and offset the DMA address. See
		 * the note in the big theory statement that explains how this
		 * helps IP deal with alignment. Note, we don't worry about
		 * whether or not we successfully get an mblk_t from desballoc,
		 * it's a common case that we have to handle later on in the
		 * system.
		 */
		dmap->dmab_size -= I40E_BUF_IPHDR_ALIGNMENT;
		dmap->dmab_address += I40E_BUF_IPHDR_ALIGNMENT;
		dmap->dmab_dma_address += I40E_BUF_IPHDR_ALIGNMENT;

		rcb->rcb_ref = 1;
		rcb->rcb_rxd = rxd;
		rcb->rcb_free_rtn.free_func = i40e_rx_recycle;
		rcb->rcb_free_rtn.free_arg = (caddr_t)rcb;
		rcb->rcb_mp = desballoc((unsigned char *)dmap->dmab_address,
		    dmap->dmab_size, 0, &rcb->rcb_free_rtn);
	}

	return (B_TRUE);
}

static void
i40e_free_tx_dma(i40e_trqpair_t *itrq)
{
	size_t fsz;

	if (itrq->itrq_tcb_area != NULL) {
		uint32_t i;
		i40e_tx_control_block_t *tcb = itrq->itrq_tcb_area;

		for (i = 0; i < itrq->itrq_tx_free_list_size; i++, tcb++) {
			i40e_free_dma_buffer(&tcb->tcb_dma);
			if (tcb->tcb_dma_handle != NULL) {
				ddi_dma_free_handle(&tcb->tcb_dma_handle);
				tcb->tcb_dma_handle = NULL;
			}
			if (tcb->tcb_lso_dma_handle != NULL) {
				ddi_dma_free_handle(&tcb->tcb_lso_dma_handle);
				tcb->tcb_lso_dma_handle = NULL;
			}
		}

		fsz = sizeof (i40e_tx_control_block_t) *
		    itrq->itrq_tx_free_list_size;
		kmem_free(itrq->itrq_tcb_area, fsz);
		itrq->itrq_tcb_area = NULL;
	}

	if (itrq->itrq_tcb_free_list != NULL) {
		fsz = sizeof (i40e_tx_control_block_t *) *
		    itrq->itrq_tx_free_list_size;
		kmem_free(itrq->itrq_tcb_free_list, fsz);
		itrq->itrq_tcb_free_list = NULL;
	}

	if (itrq->itrq_tcb_work_list != NULL) {
		fsz = sizeof (i40e_tx_control_block_t *) *
		    itrq->itrq_tx_ring_size;
		kmem_free(itrq->itrq_tcb_work_list, fsz);
		itrq->itrq_tcb_work_list = NULL;
	}

	i40e_free_dma_buffer(&itrq->itrq_desc_area);
	itrq->itrq_desc_ring = NULL;

}

static boolean_t
i40e_alloc_tx_dma(i40e_trqpair_t *itrq)
{
	int i, ret;
	size_t dmasz;
	i40e_tx_control_block_t *tcb;
	i40e_t *i40e = itrq->itrq_i40e;

	itrq->itrq_tx_ring_size = i40e->i40e_tx_ring_size;
	itrq->itrq_tx_free_list_size = i40e->i40e_tx_ring_size +
	    (i40e->i40e_tx_ring_size >> 1);

	/*
	 * Allocate an additional TX descriptor for the writeback head.
	 */
	dmasz = sizeof (i40e_tx_desc_t) * itrq->itrq_tx_ring_size;
	dmasz += sizeof (i40e_tx_desc_t);

	VERIFY(dmasz > 0);
	if (i40e_alloc_dma_buffer(i40e, &itrq->itrq_desc_area,
	    &i40e->i40e_static_dma_attr, &i40e->i40e_desc_acc_attr,
	    B_FALSE, B_TRUE, dmasz) == B_FALSE) {
		i40e_error(i40e, "failed to allocate DMA resources for TX "
		    "descriptor ring");
		return (B_FALSE);
	}
	itrq->itrq_desc_ring =
	    (i40e_tx_desc_t *)(uintptr_t)itrq->itrq_desc_area.dmab_address;
	itrq->itrq_desc_wbhead = (uint32_t *)(itrq->itrq_desc_ring +
	    itrq->itrq_tx_ring_size);
	itrq->itrq_desc_head = 0;
	itrq->itrq_desc_tail = 0;
	itrq->itrq_desc_free = itrq->itrq_tx_ring_size;

	itrq->itrq_tcb_work_list = kmem_zalloc(itrq->itrq_tx_ring_size *
	    sizeof (i40e_tx_control_block_t *), KM_NOSLEEP);
	if (itrq->itrq_tcb_work_list == NULL) {
		i40e_error(i40e, "failed to allocate a %d entry TX work list "
		    "for ring %d", itrq->itrq_tx_ring_size, itrq->itrq_index);
		goto cleanup;
	}

	itrq->itrq_tcb_free_list = kmem_zalloc(itrq->itrq_tx_free_list_size *
	    sizeof (i40e_tx_control_block_t *), KM_SLEEP);
	if (itrq->itrq_tcb_free_list == NULL) {
		i40e_error(i40e, "failed to allocate a %d entry TX free list "
		    "for ring %d", itrq->itrq_tx_free_list_size,
		    itrq->itrq_index);
		goto cleanup;
	}

	/*
	 * We allocate enough TX control blocks to cover the free list.
	 */
	itrq->itrq_tcb_area = kmem_zalloc(sizeof (i40e_tx_control_block_t) *
	    itrq->itrq_tx_free_list_size, KM_NOSLEEP);
	if (itrq->itrq_tcb_area == NULL) {
		i40e_error(i40e, "failed to allocate a %d entry tcb area for "
		    "ring %d", itrq->itrq_tx_free_list_size, itrq->itrq_index);
		goto cleanup;
	}

	/*
	 * For each tcb, allocate DMA memory.
	 */
	dmasz = i40e->i40e_tx_buf_size;
	VERIFY(dmasz > 0);
	tcb = itrq->itrq_tcb_area;
	for (i = 0; i < itrq->itrq_tx_free_list_size; i++, tcb++) {
		VERIFY(tcb != NULL);

		/*
		 * Allocate both a DMA buffer which we'll use for when we copy
		 * packets for transmission and allocate a DMA handle which
		 * we'll use when we bind data.
		 */
		ret = ddi_dma_alloc_handle(i40e->i40e_dip,
		    &i40e->i40e_txbind_dma_attr, DDI_DMA_DONTWAIT, NULL,
		    &tcb->tcb_dma_handle);
		if (ret != DDI_SUCCESS) {
			i40e_error(i40e, "failed to allocate DMA handle for TX "
			    "data binding on ring %d: %d", itrq->itrq_index,
			    ret);
			tcb->tcb_dma_handle = NULL;
			goto cleanup;
		}

		ret = ddi_dma_alloc_handle(i40e->i40e_dip,
		    &i40e->i40e_txbind_lso_dma_attr, DDI_DMA_DONTWAIT, NULL,
		    &tcb->tcb_lso_dma_handle);
		if (ret != DDI_SUCCESS) {
			i40e_error(i40e, "failed to allocate DMA handle for TX "
			    "LSO data binding on ring %d: %d", itrq->itrq_index,
			    ret);
			tcb->tcb_lso_dma_handle = NULL;
			goto cleanup;
		}

		if (i40e_alloc_dma_buffer(i40e, &tcb->tcb_dma,
		    &i40e->i40e_static_dma_attr, &i40e->i40e_buf_acc_attr,
		    B_TRUE, B_FALSE, dmasz) == B_FALSE) {
			i40e_error(i40e, "failed to allocate %ld bytes of "
			    "DMA for TX data binding on ring %d", dmasz,
			    itrq->itrq_index);
			goto cleanup;
		}

		itrq->itrq_tcb_free_list[i] = tcb;
	}

	itrq->itrq_tcb_free = itrq->itrq_tx_free_list_size;

	return (B_TRUE);

cleanup:
	i40e_free_tx_dma(itrq);
	return (B_FALSE);
}

/*
 * Free all memory associated with a ring. Note, this is done as part of
 * the GLDv3 ring stop routine.
 */
void
i40e_free_ring_mem(i40e_trqpair_t *itrq, boolean_t failed_init)
{
	i40e_t *i40e = itrq->itrq_i40e;
	i40e_rx_data_t *rxd = itrq->itrq_rxdata;

	/*
	 * In some cases i40e_alloc_rx_data() may have failed
	 * and in that case there is no rxd to free.
	 */
	if (rxd == NULL)
		return;

	/*
	 * Clean up our RX data. We have to free DMA resources first and
	 * then if we have no more pending RCB's, then we'll go ahead
	 * and clean things up. Note, we can't set the stopped flag on
	 * the RX data until after we've done the first pass of the
	 * pending resources. Otherwise we might race with
	 * i40e_rx_recycle on determining who should free the
	 * i40e_rx_data_t above.
	 */
	i40e_free_rx_dma(rxd, failed_init);

	mutex_enter(&i40e->i40e_rx_pending_lock);
	rxd->rxd_shutdown = B_TRUE;
	if (rxd->rxd_rcb_pending == 0) {
		i40e_free_rx_data(rxd);
		itrq->itrq_rxdata = NULL;
	}
	mutex_exit(&i40e->i40e_rx_pending_lock);

	i40e_free_tx_dma(itrq);
}

/*
 * Allocate all of the resources associated with a ring.
 * Note this is done as part of the GLDv3 ring start routine.
 * This takes care of both DMA and non-DMA related resources.
 */
boolean_t
i40e_alloc_ring_mem(i40e_trqpair_t *itrq)
{
	if (!i40e_alloc_rx_data(itrq->itrq_i40e, itrq))
		goto free;

	if (!i40e_alloc_rx_dma(itrq->itrq_rxdata))
		goto free;

	if (!i40e_alloc_tx_dma(itrq))
		goto free;

	return (B_TRUE);

free:
	i40e_free_ring_mem(itrq, B_TRUE);
	return (B_FALSE);
}


/*
 * Because every instance of i40e may have different support for FMA
 * capabilities, we copy the DMA attributes into the i40e_t and set them that
 * way and use them for determining attributes.
 */
void
i40e_init_dma_attrs(i40e_t *i40e, boolean_t fma)
{
	bcopy(&i40e_g_static_dma_attr, &i40e->i40e_static_dma_attr,
	    sizeof (ddi_dma_attr_t));
	bcopy(&i40e_g_txbind_dma_attr, &i40e->i40e_txbind_dma_attr,
	    sizeof (ddi_dma_attr_t));
	bcopy(&i40e_g_txbind_lso_dma_attr, &i40e->i40e_txbind_lso_dma_attr,
	    sizeof (ddi_dma_attr_t));
	bcopy(&i40e_g_desc_acc_attr, &i40e->i40e_desc_acc_attr,
	    sizeof (ddi_device_acc_attr_t));
	bcopy(&i40e_g_buf_acc_attr, &i40e->i40e_buf_acc_attr,
	    sizeof (ddi_device_acc_attr_t));

	if (fma == B_TRUE) {
		i40e->i40e_static_dma_attr.dma_attr_flags |= DDI_DMA_FLAGERR;
		i40e->i40e_txbind_dma_attr.dma_attr_flags |= DDI_DMA_FLAGERR;
		i40e->i40e_txbind_lso_dma_attr.dma_attr_flags |=
		    DDI_DMA_FLAGERR;
	} else {
		i40e->i40e_static_dma_attr.dma_attr_flags &= ~DDI_DMA_FLAGERR;
		i40e->i40e_txbind_dma_attr.dma_attr_flags &= ~DDI_DMA_FLAGERR;
		i40e->i40e_txbind_lso_dma_attr.dma_attr_flags &=
		    ~DDI_DMA_FLAGERR;
	}
}

static void
i40e_rcb_free(i40e_rx_data_t *rxd, i40e_rx_control_block_t *rcb)
{
	mutex_enter(&rxd->rxd_free_lock);
	ASSERT(rxd->rxd_rcb_free < rxd->rxd_free_list_size);
	ASSERT(rxd->rxd_free_list[rxd->rxd_rcb_free] == NULL);
	rxd->rxd_free_list[rxd->rxd_rcb_free] = rcb;
	rxd->rxd_rcb_free++;
	mutex_exit(&rxd->rxd_free_lock);
}

static i40e_rx_control_block_t *
i40e_rcb_alloc(i40e_rx_data_t *rxd)
{
	i40e_rx_control_block_t *rcb;

	mutex_enter(&rxd->rxd_free_lock);
	if (rxd->rxd_rcb_free == 0) {
		mutex_exit(&rxd->rxd_free_lock);
		return (NULL);
	}
	rxd->rxd_rcb_free--;
	rcb = rxd->rxd_free_list[rxd->rxd_rcb_free];
	VERIFY(rcb != NULL);
	rxd->rxd_free_list[rxd->rxd_rcb_free] = NULL;
	mutex_exit(&rxd->rxd_free_lock);

	return (rcb);
}

/*
 * This is the callback that we get from the OS when freemsg(9F) has been called
 * on a loaned descriptor. In addition, if we take the last reference count
 * here, then we have to tear down all of the RX data.
 */
void
i40e_rx_recycle(caddr_t arg)
{
	uint32_t ref;
	i40e_rx_control_block_t *rcb;
	i40e_rx_data_t *rxd;
	i40e_t *i40e;

	/* LINTED: E_BAD_PTR_CAST_ALIGN */
	rcb = (i40e_rx_control_block_t *)arg;
	rxd = rcb->rcb_rxd;
	i40e = rxd->rxd_i40e;

	/*
	 * It's possible for this to be called with a reference count of zero.
	 * That will happen when we're doing the freemsg after taking the last
	 * reference because we're tearing down everything and this rcb is not
	 * outstanding.
	 */
	if (rcb->rcb_ref == 0)
		return;

	/*
	 * Don't worry about failure of desballoc here. It'll only become fatal
	 * if we're trying to use it and we can't in i40e_rx_bind().
	 */
	rcb->rcb_mp = desballoc((unsigned char *)rcb->rcb_dma.dmab_address,
	    rcb->rcb_dma.dmab_size, 0, &rcb->rcb_free_rtn);
	i40e_rcb_free(rxd, rcb);

	/*
	 * It's possible that the rcb was being used while we are shutting down
	 * the device. In that case, we'll take the final reference from the
	 * device here.
	 */
	ref = atomic_dec_32_nv(&rcb->rcb_ref);
	if (ref == 0) {
		freemsg(rcb->rcb_mp);
		rcb->rcb_mp = NULL;
		i40e_free_dma_buffer(&rcb->rcb_dma);

		mutex_enter(&i40e->i40e_rx_pending_lock);
		atomic_dec_32(&rxd->rxd_rcb_pending);
		atomic_dec_32(&i40e->i40e_rx_pending);

		/*
		 * If this was the last block and it's been indicated that we've
		 * passed the shutdown point, we should clean up.
		 */
		if (rxd->rxd_shutdown == B_TRUE && rxd->rxd_rcb_pending == 0) {
			i40e_free_rx_data(rxd);
			cv_broadcast(&i40e->i40e_rx_pending_cv);
		}

		mutex_exit(&i40e->i40e_rx_pending_lock);
	}
}

static mblk_t *
i40e_rx_bind(i40e_trqpair_t *itrq, i40e_rx_data_t *rxd, uint32_t index,
    uint32_t plen)
{
	mblk_t *mp;
	i40e_t *i40e = rxd->rxd_i40e;
	i40e_rx_control_block_t *rcb, *rep_rcb;

	ASSERT(MUTEX_HELD(&itrq->itrq_rx_lock));

	if ((rep_rcb = i40e_rcb_alloc(rxd)) == NULL) {
		itrq->itrq_rxstat.irxs_rx_bind_norcb.value.ui64++;
		return (NULL);
	}

	rcb = rxd->rxd_work_list[index];

	/*
	 * Check to make sure we have a mblk_t. If we don't, this is our last
	 * chance to try and get one.
	 */
	if (rcb->rcb_mp == NULL) {
		rcb->rcb_mp =
		    desballoc((unsigned char *)rcb->rcb_dma.dmab_address,
		    rcb->rcb_dma.dmab_size, 0, &rcb->rcb_free_rtn);
		if (rcb->rcb_mp == NULL) {
			itrq->itrq_rxstat.irxs_rx_bind_nomp.value.ui64++;
			i40e_rcb_free(rxd, rcb);
			return (NULL);
		}
	}

	I40E_DMA_SYNC(&rcb->rcb_dma, DDI_DMA_SYNC_FORKERNEL);

	if (i40e_check_dma_handle(rcb->rcb_dma.dmab_dma_handle) != DDI_FM_OK) {
		ddi_fm_service_impact(i40e->i40e_dip, DDI_SERVICE_DEGRADED);
		atomic_or_32(&i40e->i40e_state, I40E_ERROR);
		i40e_rcb_free(rxd, rcb);
		return (NULL);
	}

	/*
	 * Note, we've already accounted for the I40E_BUF_IPHDR_ALIGNMENT.
	 */
	mp = rcb->rcb_mp;
	atomic_inc_32(&rcb->rcb_ref);
	mp->b_wptr = mp->b_rptr + plen;
	mp->b_next = mp->b_cont = NULL;

	rxd->rxd_work_list[index] = rep_rcb;
	return (mp);
}

/*
 * We're going to allocate a new message block for this frame and attempt to
 * receive it. See the big theory statement for more information on when we copy
 * versus bind.
 */
static mblk_t *
i40e_rx_copy(i40e_trqpair_t *itrq, i40e_rx_data_t *rxd, uint32_t index,
    uint32_t plen)
{
	i40e_t *i40e = rxd->rxd_i40e;
	i40e_rx_control_block_t *rcb;
	mblk_t *mp;

	ASSERT(index < rxd->rxd_ring_size);
	rcb = rxd->rxd_work_list[index];

	I40E_DMA_SYNC(&rcb->rcb_dma, DDI_DMA_SYNC_FORKERNEL);

	if (i40e_check_dma_handle(rcb->rcb_dma.dmab_dma_handle) != DDI_FM_OK) {
		ddi_fm_service_impact(i40e->i40e_dip, DDI_SERVICE_DEGRADED);
		atomic_or_32(&i40e->i40e_state, I40E_ERROR);
		return (NULL);
	}

	mp = allocb(plen + I40E_BUF_IPHDR_ALIGNMENT, 0);
	if (mp == NULL) {
		itrq->itrq_rxstat.irxs_rx_copy_nomem.value.ui64++;
		return (NULL);
	}

	mp->b_rptr += I40E_BUF_IPHDR_ALIGNMENT;
	bcopy(rcb->rcb_dma.dmab_address, mp->b_rptr, plen);
	mp->b_wptr = mp->b_rptr + plen;

	return (mp);
}

/*
 * Determine if the device has enabled any checksum flags for us. The level of
 * checksum computed will depend on the type packet that we have, which is
 * contained in ptype. For example, the checksum logic it does will vary
 * depending on whether or not the packet is considered tunneled, whether it
 * recognizes the L4 type, etc. Section 8.3.4.3 summarizes which checksums are
 * valid.
 *
 * While there are additional checksums that we could recognize here, we'll need
 * to get some additional GLDv3 enhancements to be able to properly describe
 * them.
 */
static void
i40e_rx_hcksum(i40e_trqpair_t *itrq, mblk_t *mp, uint64_t status, uint32_t err,
    uint32_t ptype)
{
	uint32_t cksum;
	struct i40e_rx_ptype_decoded pinfo;

	ASSERT(ptype <= 255);
	pinfo = decode_rx_desc_ptype(ptype);

	cksum = 0;

	/*
	 * If the ptype isn't something that we know in the driver, then we
	 * shouldn't even consider moving forward.
	 */
	if (pinfo.known == 0) {
		itrq->itrq_rxstat.irxs_hck_unknown.value.ui64++;
		return;
	}

	/*
	 * If hardware didn't set the L3L4P bit on the frame, then there is no
	 * checksum offload to consider.
	 */
	if ((status & (1 << I40E_RX_DESC_STATUS_L3L4P_SHIFT)) == 0) {
		itrq->itrq_rxstat.irxs_hck_nol3l4p.value.ui64++;
		return;
	}

	/*
	 * The device tells us that IPv6 checksums where a Destination Options
	 * Header or a Routing header shouldn't be trusted. Discard all
	 * checksums in this case.
	 */
	if (pinfo.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
	    pinfo.outer_ip_ver == I40E_RX_PTYPE_OUTER_IPV6 &&
	    (status & (1 << I40E_RX_DESC_STATUS_IPV6EXADD_SHIFT))) {
		itrq->itrq_rxstat.irxs_hck_v6skip.value.ui64++;
		return;
	}

	/*
	 * The hardware denotes three kinds of possible errors. Two are reserved
	 * for inner and outer IP checksum errors (IPE and EIPE) and the latter
	 * is for L4 checksum errors (L4E). If there is only one IP header, then
	 * the only thing that we care about is IPE. Note that since we don't
	 * support inner checksums, we will ignore IPE being set on tunneled
	 * packets and only care about EIPE.
	 */
	if (pinfo.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
	    pinfo.outer_ip_ver == I40E_RX_PTYPE_OUTER_IPV4) {
		if (pinfo.tunnel_type == I40E_RX_PTYPE_OUTER_NONE) {
			if ((err & (1 << I40E_RX_DESC_ERROR_IPE_SHIFT)) != 0) {
				itrq->itrq_rxstat.irxs_hck_iperr.value.ui64++;
			} else {
				itrq->itrq_rxstat.irxs_hck_v4hdrok.value.ui64++;
				cksum |= HCK_IPV4_HDRCKSUM_OK;
			}
		} else {
			if ((err & (1 << I40E_RX_DESC_ERROR_EIPE_SHIFT)) != 0) {
				itrq->itrq_rxstat.irxs_hck_eiperr.value.ui64++;
			} else {
				itrq->itrq_rxstat.irxs_hck_v4hdrok.value.ui64++;
				cksum |= HCK_IPV4_HDRCKSUM_OK;
			}
		}
	}

	/*
	 * We only have meaningful L4 checksums in the case of IP->L4 and
	 * IP->IP->L4. There is not outer L4 checksum data available in any
	 * other case. Further, we don't bother reporting the valid checksum in
	 * the case of IP->IP->L4 set.
	 */
	if (pinfo.outer_ip == I40E_RX_PTYPE_OUTER_IP &&
	    pinfo.tunnel_type == I40E_RX_PTYPE_TUNNEL_NONE &&
	    (pinfo.inner_prot == I40E_RX_PTYPE_INNER_PROT_UDP ||
	    pinfo.inner_prot == I40E_RX_PTYPE_INNER_PROT_TCP ||
	    pinfo.inner_prot == I40E_RX_PTYPE_INNER_PROT_ICMP ||
	    pinfo.inner_prot == I40E_RX_PTYPE_INNER_PROT_SCTP)) {
		ASSERT(pinfo.payload_layer == I40E_RX_PTYPE_PAYLOAD_LAYER_PAY4);
		if ((err & (1 << I40E_RX_DESC_ERROR_L4E_SHIFT)) != 0) {
			itrq->itrq_rxstat.irxs_hck_l4err.value.ui64++;
		} else {
			itrq->itrq_rxstat.irxs_hck_l4hdrok.value.ui64++;
			cksum |= HCK_FULLCKSUM_OK;
		}
	}

	if (cksum != 0) {
		itrq->itrq_rxstat.irxs_hck_set.value.ui64++;
		mac_hcksum_set(mp, 0, 0, 0, 0, cksum);
	} else {
		itrq->itrq_rxstat.irxs_hck_miss.value.ui64++;
	}
}

mblk_t *
i40e_ring_rx(i40e_trqpair_t *itrq, int poll_bytes)
{
	i40e_t *i40e;
	i40e_hw_t *hw;
	i40e_rx_data_t *rxd;
	uint32_t cur_head;
	i40e_rx_desc_t *cur_desc;
	i40e_rx_control_block_t *rcb;
	uint64_t rx_bytes, rx_frames;
	uint64_t stword;
	mblk_t *mp, *mp_head, **mp_tail;

	ASSERT(MUTEX_HELD(&itrq->itrq_rx_lock));
	rxd = itrq->itrq_rxdata;
	i40e = itrq->itrq_i40e;
	hw = &i40e->i40e_hw_space;

	if (!(i40e->i40e_state & I40E_STARTED) ||
	    (i40e->i40e_state & I40E_OVERTEMP) ||
	    (i40e->i40e_state & I40E_SUSPENDED) ||
	    (i40e->i40e_state & I40E_ERROR))
		return (NULL);

	/*
	 * Before we do anything else, we have to make sure that all of the DMA
	 * buffers are synced up and then check to make sure that they're
	 * actually good from an FM perspective.
	 */
	I40E_DMA_SYNC(&rxd->rxd_desc_area, DDI_DMA_SYNC_FORKERNEL);
	if (i40e_check_dma_handle(rxd->rxd_desc_area.dmab_dma_handle) !=
	    DDI_FM_OK) {
		ddi_fm_service_impact(i40e->i40e_dip, DDI_SERVICE_DEGRADED);
		atomic_or_32(&i40e->i40e_state, I40E_ERROR);
		return (NULL);
	}

	/*
	 * Prepare our stats. We do a limited amount of processing in both
	 * polling and interrupt context. The limit in interrupt context is
	 * based on frames, in polling context based on bytes.
	 */
	rx_bytes = rx_frames = 0;
	mp_head = NULL;
	mp_tail = &mp_head;

	/*
	 * At this point, the descriptor ring is available to check. We'll try
	 * and process until we either run out of poll_bytes or descriptors.
	 */
	cur_head = rxd->rxd_desc_next;
	cur_desc = &rxd->rxd_desc_ring[cur_head];
	stword = LE64_TO_CPU(cur_desc->wb.qword1.status_error_len);

	/*
	 * Note, the primary invariant of this loop should be that cur_head,
	 * cur_desc, and stword always point to the currently processed
	 * descriptor. When we leave the loop, it should point to a descriptor
	 * that HAS NOT been processed. Meaning, that if we haven't consumed the
	 * frame, the descriptor should not be advanced.
	 */
	while ((stword & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)) != 0) {
		uint32_t error, eop, plen, ptype;

		/*
		 * The DD, PLEN, and EOP bits are the only ones that are valid
		 * in every frame. The error information is only valid when EOP
		 * is set in the same frame.
		 *
		 * At this time, because we don't do any LRO or header
		 * splitting. We expect that every frame should have EOP set in
		 * it. When later functionality comes in, we'll want to
		 * re-evaluate this.
		 */
		eop = stword & (1 << I40E_RX_DESC_STATUS_EOF_SHIFT);
		VERIFY(eop != 0);

		error = (stword & I40E_RXD_QW1_ERROR_MASK) >>
		    I40E_RXD_QW1_ERROR_SHIFT;
		if (error & I40E_RX_ERR_BITS) {
			itrq->itrq_rxstat.irxs_rx_desc_error.value.ui64++;
			goto discard;
		}

		plen = (stword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
		    I40E_RXD_QW1_LENGTH_PBUF_SHIFT;

		ptype = (stword & I40E_RXD_QW1_PTYPE_MASK) >>
		    I40E_RXD_QW1_PTYPE_SHIFT;

		/*
		 * This packet contains valid data. We should check to see if
		 * we're actually going to consume it based on its length (to
		 * ensure that we don't overshoot our quota). We determine
		 * whether to bcopy or bind the DMA resources based on the size
		 * of the frame. However, if on debug, we allow it to be
		 * overridden for testing purposes.
		 *
		 * We should be smarter about this and do DMA binding for
		 * larger frames, but for now, it's really more important that
		 * we actually just get something simple working.
		 */

		/*
		 * Ensure we don't exceed our polling quota by reading this
		 * frame. Note we only bump bytes now, we bump frames later.
		 */
		if ((poll_bytes != I40E_POLL_NULL) &&
		    (rx_bytes + plen) > poll_bytes)
			break;
		rx_bytes += plen;

		mp = NULL;
		if (plen >= i40e->i40e_rx_dma_min)
			mp = i40e_rx_bind(itrq, rxd, cur_head, plen);
		if (mp == NULL)
			mp = i40e_rx_copy(itrq, rxd, cur_head, plen);

		if (mp != NULL) {
			if (i40e->i40e_rx_hcksum_enable)
				i40e_rx_hcksum(itrq, mp, stword, error, ptype);
			*mp_tail = mp;
			mp_tail = &mp->b_next;
		}

		/*
		 * Now we need to prepare this frame for use again. See the
		 * discussion in the big theory statements.
		 *
		 * However, right now we're doing the simple version of this.
		 * Normally what we'd do would depend on whether or not we were
		 * doing DMA binding or bcopying. But because we're always doing
		 * bcopying, we can just always use the current index as a key
		 * for what to do and reassign the buffer based on the ring.
		 */
discard:
		rcb = rxd->rxd_work_list[cur_head];
		cur_desc->read.pkt_addr =
		    CPU_TO_LE64((uintptr_t)rcb->rcb_dma.dmab_dma_address);
		cur_desc->read.hdr_addr = 0;

		/*
		 * Finally, update our loop invariants.
		 */
		cur_head = i40e_next_desc(cur_head, 1, rxd->rxd_ring_size);
		cur_desc = &rxd->rxd_desc_ring[cur_head];
		stword = LE64_TO_CPU(cur_desc->wb.qword1.status_error_len);

		/*
		 * To help provide liveness, we limit the amount of data that
		 * we'll end up counting. Note that in these cases, an interrupt
		 * is not dissimilar from a polling request.
		 */
		rx_frames++;
		if (rx_frames > i40e->i40e_rx_limit_per_intr) {
			itrq->itrq_rxstat.irxs_rx_intr_limit.value.ui64++;
			break;
		}
	}

	/*
	 * As we've modified the ring, we need to make sure that we sync the
	 * descriptor ring for the device. Next, we update the hardware and
	 * update our notion of where the head for us to read from hardware is
	 * next.
	 */
	I40E_DMA_SYNC(&rxd->rxd_desc_area, DDI_DMA_SYNC_FORDEV);
	if (i40e_check_dma_handle(rxd->rxd_desc_area.dmab_dma_handle) !=
	    DDI_FM_OK) {
		ddi_fm_service_impact(i40e->i40e_dip, DDI_SERVICE_DEGRADED);
		atomic_or_32(&i40e->i40e_state, I40E_ERROR);
	}

	if (rx_frames != 0) {
		uint32_t tail;
		ddi_acc_handle_t rh = i40e->i40e_osdep_space.ios_reg_handle;
		rxd->rxd_desc_next = cur_head;
		tail = i40e_prev_desc(cur_head, 1, rxd->rxd_ring_size);

		I40E_WRITE_REG(hw, I40E_QRX_TAIL(itrq->itrq_index), tail);
		if (i40e_check_acc_handle(rh) != DDI_FM_OK) {
			ddi_fm_service_impact(i40e->i40e_dip,
			    DDI_SERVICE_DEGRADED);
			atomic_or_32(&i40e->i40e_state, I40E_ERROR);
		}

		itrq->itrq_rxstat.irxs_bytes.value.ui64 += rx_bytes;
		itrq->itrq_rxstat.irxs_packets.value.ui64 += rx_frames;
	}

#ifdef DEBUG
	if (rx_frames == 0) {
		ASSERT(rx_bytes == 0);
	}
#endif

	return (mp_head);
}

/*
 * This function is called by the GLDv3 when it wants to poll on a ring. The
 * only primary difference from when we call this during an interrupt is that we
 * have a limit on the number of bytes that we should consume.
 */
mblk_t *
i40e_ring_rx_poll(void *arg, int poll_bytes)
{
	i40e_trqpair_t *itrq = arg;
	mblk_t *mp;

	ASSERT(poll_bytes > 0);
	if (poll_bytes == 0)
		return (NULL);

	mutex_enter(&itrq->itrq_rx_lock);
	mp = i40e_ring_rx(itrq, poll_bytes);
	mutex_exit(&itrq->itrq_rx_lock);

	return (mp);
}

/*
 * Attempt to put togther the information we'll need to feed into a descriptor
 * to properly program the hardware for checksum offload as well as the
 * generally required flags.
 *
 * The i40e_tx_context_t`itc_data_cmdflags contains the set of flags we need to
 * 'or' into the descriptor based on the checksum flags for this mblk_t and the
 * actual information we care about.
 *
 * If the mblk requires LSO then we'll also gather the information that will be
 * used to construct the Transmit Context Descriptor.
 */
static int
i40e_tx_context(i40e_t *i40e, i40e_trqpair_t *itrq, mblk_t *mp,
    mac_ether_offload_info_t *meo, i40e_tx_context_t *tctx)
{
	uint32_t chkflags, start, mss, lsoflags;
	i40e_txq_stat_t *txs = &itrq->itrq_txstat;

	bzero(tctx, sizeof (i40e_tx_context_t));

	if (i40e->i40e_tx_hcksum_enable != B_TRUE)
		return (0);

	mac_hcksum_get(mp, &start, NULL, NULL, NULL, &chkflags);
	mac_lso_get(mp, &mss, &lsoflags);

	if (chkflags == 0 && lsoflags == 0)
		return (0);

	/*
	 * Have we been asked to checksum an IPv4 header. If so, verify that we
	 * have sufficient information and then set the proper fields in the
	 * command structure.
	 */
	if (chkflags & HCK_IPV4_HDRCKSUM) {
		if ((meo->meoi_flags & MEOI_L2INFO_SET) == 0) {
			txs->itxs_hck_nol2info.value.ui64++;
			return (-1);
		}
		if ((meo->meoi_flags & MEOI_L3INFO_SET) == 0) {
			txs->itxs_hck_nol3info.value.ui64++;
			return (-1);
		}
		if (meo->meoi_l3proto != ETHERTYPE_IP) {
			txs->itxs_hck_badl3.value.ui64++;
			return (-1);
		}
		tctx->itc_data_cmdflags |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
		tctx->itc_data_offsets |= (meo->meoi_l2hlen >> 1) <<
		    I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
		tctx->itc_data_offsets |= (meo->meoi_l3hlen >> 2) <<
		    I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
	}

	/*
	 * We've been asked to provide an L4 header, first, set up the IP
	 * information in the descriptor if we haven't already before moving
	 * onto seeing if we have enough information for the L4 checksum
	 * offload.
	 */
	if (chkflags & HCK_PARTIALCKSUM) {
		if ((meo->meoi_flags & MEOI_L4INFO_SET) == 0) {
			txs->itxs_hck_nol4info.value.ui64++;
			return (-1);
		}

		if (!(chkflags & HCK_IPV4_HDRCKSUM)) {
			if ((meo->meoi_flags & MEOI_L2INFO_SET) == 0) {
				txs->itxs_hck_nol2info.value.ui64++;
				return (-1);
			}
			if ((meo->meoi_flags & MEOI_L3INFO_SET) == 0) {
				txs->itxs_hck_nol3info.value.ui64++;
				return (-1);
			}

			if (meo->meoi_l3proto == ETHERTYPE_IP) {
				tctx->itc_data_cmdflags |=
				    I40E_TX_DESC_CMD_IIPT_IPV4;
			} else if (meo->meoi_l3proto == ETHERTYPE_IPV6) {
				tctx->itc_data_cmdflags |=
				    I40E_TX_DESC_CMD_IIPT_IPV6;
			} else {
				txs->itxs_hck_badl3.value.ui64++;
				return (-1);
			}
			tctx->itc_data_offsets |= (meo->meoi_l2hlen >> 1) <<
			    I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
			tctx->itc_data_offsets |= (meo->meoi_l3hlen >> 2) <<
			    I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
		}

		switch (meo->meoi_l4proto) {
		case IPPROTO_TCP:
			tctx->itc_data_cmdflags |=
			    I40E_TX_DESC_CMD_L4T_EOFT_TCP;
			break;
		case IPPROTO_UDP:
			tctx->itc_data_cmdflags |=
			    I40E_TX_DESC_CMD_L4T_EOFT_UDP;
			break;
		case IPPROTO_SCTP:
			tctx->itc_data_cmdflags |=
			    I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
			break;
		default:
			txs->itxs_hck_badl4.value.ui64++;
			return (-1);
		}

		tctx->itc_data_offsets |= (meo->meoi_l4hlen >> 2) <<
		    I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
	}

	if (lsoflags & HW_LSO) {
		/*
		 * LSO requires that checksum offloads are enabled.  If for
		 * some reason they're not we bail out with an error.
		 *
		 * Fulfilling this requirement also ensures that the L4 info was
		 * parsed by meoi, which is also necessary for LSO.
		 */
		if ((meo->meoi_l3proto == ETHERTYPE_IP &&
		    (chkflags & HCK_IPV4_HDRCKSUM) == 0) ||
		    (chkflags & HCK_PARTIALCKSUM) == 0) {
			txs->itxs_lso_nohck.value.ui64++;
			return (-1);
		}

		tctx->itc_ctx_cmdflags |= I40E_TX_CTX_DESC_TSO;
		tctx->itc_ctx_mss = mss;
		tctx->itc_ctx_tsolen = msgsize(mp) -
		    (meo->meoi_l2hlen + meo->meoi_l3hlen + meo->meoi_l4hlen);
	}

	return (0);
}

static void
i40e_tcb_free(i40e_trqpair_t *itrq, i40e_tx_control_block_t *tcb)
{
	ASSERT(tcb != NULL);

	mutex_enter(&itrq->itrq_tcb_lock);
	ASSERT(itrq->itrq_tcb_free < itrq->itrq_tx_free_list_size);
	itrq->itrq_tcb_free_list[itrq->itrq_tcb_free] = tcb;
	itrq->itrq_tcb_free++;
	mutex_exit(&itrq->itrq_tcb_lock);
}

static i40e_tx_control_block_t *
i40e_tcb_alloc(i40e_trqpair_t *itrq)
{
	i40e_tx_control_block_t *ret;

	mutex_enter(&itrq->itrq_tcb_lock);
	if (itrq->itrq_tcb_free == 0) {
		mutex_exit(&itrq->itrq_tcb_lock);
		return (NULL);
	}

	itrq->itrq_tcb_free--;
	ret = itrq->itrq_tcb_free_list[itrq->itrq_tcb_free];
	itrq->itrq_tcb_free_list[itrq->itrq_tcb_free] = NULL;
	mutex_exit(&itrq->itrq_tcb_lock);

	ASSERT(ret != NULL);
	return (ret);
}

/*
 * This should be used to free any DMA resources, associated mblk_t's, etc. It's
 * used as part of recycling the message blocks when we have either an interrupt
 * or other activity that indicates that we need to take a look.
 */
static void
i40e_tcb_reset(i40e_tx_control_block_t *tcb)
{
	switch (tcb->tcb_type) {
	case I40E_TX_COPY:
		tcb->tcb_dma.dmab_len = 0;
		break;
	case I40E_TX_DMA:
		if (tcb->tcb_used_lso == B_TRUE && tcb->tcb_bind_ncookies > 0)
			(void) ddi_dma_unbind_handle(tcb->tcb_lso_dma_handle);
		else if (tcb->tcb_bind_ncookies > 0)
			(void) ddi_dma_unbind_handle(tcb->tcb_dma_handle);
		if (tcb->tcb_bind_info != NULL) {
			kmem_free(tcb->tcb_bind_info,
			    tcb->tcb_bind_ncookies *
			    sizeof (struct i40e_dma_bind_info));
		}
		tcb->tcb_bind_info = NULL;
		tcb->tcb_bind_ncookies = 0;
		tcb->tcb_used_lso = B_FALSE;
		break;
	case I40E_TX_DESC:
		break;
	case I40E_TX_NONE:
		/* Cast to pacify lint */
		panic("trying to free tcb %p with bad type none", (void *)tcb);
	default:
		panic("unknown i40e tcb type: %d", tcb->tcb_type);
	}

	tcb->tcb_type = I40E_TX_NONE;
	if (tcb->tcb_mp != NULL) {
		freemsg(tcb->tcb_mp);
		tcb->tcb_mp = NULL;
	}
	tcb->tcb_next = NULL;
}

/*
 * This is called as part of shutting down to clean up all outstanding
 * descriptors. Similar to recycle, except we don't re-arm anything and instead
 * just return control blocks to the free list.
 */
void
i40e_tx_cleanup_ring(i40e_trqpair_t *itrq)
{
	uint32_t index;

	ASSERT(MUTEX_HELD(&itrq->itrq_tx_lock));
	ASSERT(itrq->itrq_desc_free <= itrq->itrq_tx_ring_size);

	/*
	 * Because we should have shut down the chip at this point, it should be
	 * safe to just clean up all the entries between our head and tail.
	 */
#ifdef	DEBUG
	index = I40E_READ_REG(&itrq->itrq_i40e->i40e_hw_space,
	    I40E_QTX_ENA(itrq->itrq_index));
	VERIFY0(index & (I40E_QTX_ENA_QENA_REQ_MASK |
	    I40E_QTX_ENA_QENA_STAT_MASK));
#endif

	index = itrq->itrq_desc_head;
	while (itrq->itrq_desc_free < itrq->itrq_tx_ring_size) {
		i40e_tx_control_block_t *tcb;

		tcb = itrq->itrq_tcb_work_list[index];
		if (tcb != NULL) {
			itrq->itrq_tcb_work_list[index] = NULL;
			i40e_tcb_reset(tcb);
			i40e_tcb_free(itrq, tcb);
		}

		bzero(&itrq->itrq_desc_ring[index], sizeof (i40e_tx_desc_t));
		index = i40e_next_desc(index, 1, itrq->itrq_tx_ring_size);
		itrq->itrq_desc_free++;
	}

	ASSERT(index == itrq->itrq_desc_tail);
	itrq->itrq_desc_head = index;
}

/*
 * We're here either by hook or by crook. We need to see if there are transmit
 * descriptors available for us to go and clean up and return to the hardware.
 * We may also be blocked, and if so, we should make sure that we let it know
 * we're good to go.
 */
void
i40e_tx_recycle_ring(i40e_trqpair_t *itrq)
{
	uint32_t wbhead, toclean, count;
	i40e_tx_control_block_t *tcbhead;
	i40e_t *i40e = itrq->itrq_i40e;
	uint_t desc_per_tcb, i;

	mutex_enter(&itrq->itrq_tx_lock);

	ASSERT(itrq->itrq_desc_free <= itrq->itrq_tx_ring_size);
	if (itrq->itrq_desc_free == itrq->itrq_tx_ring_size) {
		if (itrq->itrq_tx_blocked == B_TRUE) {
			itrq->itrq_tx_blocked = B_FALSE;
			mac_tx_ring_update(i40e->i40e_mac_hdl,
			    itrq->itrq_mactxring);
			itrq->itrq_txstat.itxs_num_unblocked.value.ui64++;
		}
		mutex_exit(&itrq->itrq_tx_lock);
		return;
	}

	/*
	 * Now we need to try and see if there's anything available. The driver
	 * will write to the head location and it guarantees that it does not
	 * use relaxed ordering.
	 */
	VERIFY0(ddi_dma_sync(itrq->itrq_desc_area.dmab_dma_handle,
	    (uintptr_t)itrq->itrq_desc_wbhead,
	    sizeof (uint32_t), DDI_DMA_SYNC_FORKERNEL));

	if (i40e_check_dma_handle(itrq->itrq_desc_area.dmab_dma_handle) !=
	    DDI_FM_OK) {
		mutex_exit(&itrq->itrq_tx_lock);
		ddi_fm_service_impact(i40e->i40e_dip, DDI_SERVICE_DEGRADED);
		atomic_or_32(&i40e->i40e_state, I40E_ERROR);
		return;
	}

	wbhead = *itrq->itrq_desc_wbhead;
	toclean = itrq->itrq_desc_head;
	count = 0;
	tcbhead = NULL;

	while (toclean != wbhead) {
		i40e_tx_control_block_t *tcb;

		tcb = itrq->itrq_tcb_work_list[toclean];
		itrq->itrq_tcb_work_list[toclean] = NULL;
		ASSERT(tcb != NULL);
		tcb->tcb_next = tcbhead;
		tcbhead = tcb;

		/*
		 * In the DMA bind case, there may not necessarily be a 1:1
		 * mapping between tcb's and descriptors.  If the tcb type
		 * indicates a DMA binding then check the number of DMA
		 * cookies to determine how many entries to clean in the
		 * descriptor ring.
		 */
		if (tcb->tcb_type == I40E_TX_DMA)
			desc_per_tcb = tcb->tcb_bind_ncookies;
		else
			desc_per_tcb = 1;

		for (i = 0; i < desc_per_tcb; i++) {
			/*
			 * We zero this out for sanity purposes.
			 */
			bzero(&itrq->itrq_desc_ring[toclean],
			    sizeof (i40e_tx_desc_t));
			toclean = i40e_next_desc(toclean, 1,
			    itrq->itrq_tx_ring_size);
			count++;
		}
	}

	itrq->itrq_desc_head = wbhead;
	itrq->itrq_desc_free += count;
	itrq->itrq_txstat.itxs_recycled.value.ui64 += count;
	ASSERT(itrq->itrq_desc_free <= itrq->itrq_tx_ring_size);

	if (itrq->itrq_tx_blocked == B_TRUE &&
	    itrq->itrq_desc_free > i40e->i40e_tx_block_thresh) {
		itrq->itrq_tx_blocked = B_FALSE;

		mac_tx_ring_update(i40e->i40e_mac_hdl, itrq->itrq_mactxring);
		itrq->itrq_txstat.itxs_num_unblocked.value.ui64++;
	}

	mutex_exit(&itrq->itrq_tx_lock);

	/*
	 * Now clean up the tcb.
	 */
	while (tcbhead != NULL) {
		i40e_tx_control_block_t *tcb = tcbhead;

		tcbhead = tcb->tcb_next;
		i40e_tcb_reset(tcb);
		i40e_tcb_free(itrq, tcb);
	}

	DTRACE_PROBE2(i40e__recycle, i40e_trqpair_t *, itrq, uint32_t, count);
}

static void
i40e_tx_copy_fragment(i40e_tx_control_block_t *tcb, const mblk_t *mp,
    const size_t off, const size_t len)
{
	const void *soff = mp->b_rptr + off;
	void *doff = tcb->tcb_dma.dmab_address + tcb->tcb_dma.dmab_len;

	ASSERT3U(len, >, 0);
	ASSERT3P(soff, >=, mp->b_rptr);
	ASSERT3P(soff, <=, mp->b_wptr);
	ASSERT3U(len, <=, MBLKL(mp));
	ASSERT3U((uintptr_t)soff + len, <=, (uintptr_t)mp->b_wptr);
	ASSERT3U(tcb->tcb_dma.dmab_size - tcb->tcb_dma.dmab_len, >=, len);
	bcopy(soff, doff, len);
	tcb->tcb_type = I40E_TX_COPY;
	tcb->tcb_dma.dmab_len += len;
	I40E_DMA_SYNC(&tcb->tcb_dma, DDI_DMA_SYNC_FORDEV);
}

static i40e_tx_control_block_t *
i40e_tx_bind_fragment(i40e_trqpair_t *itrq, const mblk_t *mp,
    size_t off, boolean_t use_lso)
{
	ddi_dma_handle_t dma_handle;
	ddi_dma_cookie_t dma_cookie;
	uint_t i = 0, ncookies = 0, dmaflags;
	i40e_tx_control_block_t *tcb;
	i40e_txq_stat_t *txs = &itrq->itrq_txstat;

	if ((tcb = i40e_tcb_alloc(itrq)) == NULL) {
		txs->itxs_err_notcb.value.ui64++;
		return (NULL);
	}
	tcb->tcb_type = I40E_TX_DMA;

	if (use_lso == B_TRUE)
		dma_handle = tcb->tcb_lso_dma_handle;
	else
		dma_handle = tcb->tcb_dma_handle;

	dmaflags = DDI_DMA_WRITE | DDI_DMA_STREAMING;
	if (ddi_dma_addr_bind_handle(dma_handle, NULL,
	    (caddr_t)(mp->b_rptr + off), MBLKL(mp) - off, dmaflags,
	    DDI_DMA_DONTWAIT, NULL, &dma_cookie, &ncookies) != DDI_DMA_MAPPED) {
		txs->itxs_bind_fails.value.ui64++;
		goto bffail;
	}

	tcb->tcb_bind_ncookies = ncookies;
	tcb->tcb_used_lso = use_lso;

	tcb->tcb_bind_info =
	    kmem_zalloc(ncookies * sizeof (struct i40e_dma_bind_info),
	    KM_NOSLEEP);
	if (tcb->tcb_bind_info == NULL)
		goto bffail;

	while (i < ncookies) {
		if (i > 0)
			ddi_dma_nextcookie(dma_handle, &dma_cookie);

		tcb->tcb_bind_info[i].dbi_paddr =
		    (caddr_t)dma_cookie.dmac_laddress;
		tcb->tcb_bind_info[i++].dbi_len = dma_cookie.dmac_size;
	}

	return (tcb);

bffail:
	i40e_tcb_reset(tcb);
	i40e_tcb_free(itrq, tcb);
	return (NULL);
}

static void
i40e_tx_set_data_desc(i40e_trqpair_t *itrq, i40e_tx_context_t *tctx,
    caddr_t buff, size_t len, boolean_t last_desc)
{
	i40e_tx_desc_t *txdesc;
	int cmd;

	ASSERT(MUTEX_HELD(&itrq->itrq_tx_lock));
	itrq->itrq_desc_free--;
	txdesc = &itrq->itrq_desc_ring[itrq->itrq_desc_tail];
	itrq->itrq_desc_tail = i40e_next_desc(itrq->itrq_desc_tail, 1,
	    itrq->itrq_tx_ring_size);

	cmd = I40E_TX_DESC_CMD_ICRC | tctx->itc_data_cmdflags;

	/*
	 * The last data descriptor needs the EOP bit set, so that the HW knows
	 * that we're ready to send.  Additionally, we set the RS (Report
	 * Status) bit, so that we are notified when the transmit engine has
	 * completed DMA'ing all of the data descriptors and data buffers
	 * associated with this frame.
	 */
	if (last_desc == B_TRUE) {
		cmd |= I40E_TX_DESC_CMD_EOP;
		cmd |= I40E_TX_DESC_CMD_RS;
	}

	/*
	 * Per the X710 manual, section 8.4.2.1.1, the buffer size
	 * must be a value from 1 to 16K minus 1, inclusive.
	 */
	ASSERT3U(len, >=, 1);
	ASSERT3U(len, <=, I40E_MAX_TX_BUFSZ);

	txdesc->buffer_addr = CPU_TO_LE64((uintptr_t)buff);
	txdesc->cmd_type_offset_bsz =
	    LE_64(((uint64_t)I40E_TX_DESC_DTYPE_DATA |
	    ((uint64_t)tctx->itc_data_offsets << I40E_TXD_QW1_OFFSET_SHIFT) |
	    ((uint64_t)cmd << I40E_TXD_QW1_CMD_SHIFT) |
	    ((uint64_t)len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT)));
}

/*
 * Place 'tcb' on the tail of the list represented by 'head'/'tail'.
 */
static inline void
tcb_list_append(i40e_tx_control_block_t **head, i40e_tx_control_block_t **tail,
    i40e_tx_control_block_t *tcb)
{
	if (*head == NULL) {
		*head = tcb;
		*tail = *head;
	} else {
		ASSERT3P(*tail, !=, NULL);
		ASSERT3P((*tail)->tcb_next, ==, NULL);
		(*tail)->tcb_next = tcb;
		*tail = tcb;
	}
}

/*
 * This function takes a single packet, possibly consisting of
 * multiple mblks, and creates a TCB chain to send to the controller.
 * This TCB chain may span up to a maximum of 8 descriptors. A copy
 * TCB consumes one descriptor; whereas a DMA TCB may consume 1 or
 * more, depending on several factors. For each fragment (invidual
 * mblk making up the packet), we determine if its size dictates a
 * copy to the TCB buffer or a DMA bind of the dblk buffer. We keep a
 * count of descriptors used; when that count reaches the max we force
 * all remaining fragments into a single TCB buffer. We have a
 * guarantee that the TCB buffer is always larger than the MTU -- so
 * there is always enough room. Consecutive fragments below the DMA
 * threshold are copied into a single TCB. In the event of an error
 * this function returns NULL but leaves 'mp' alone.
 */
static i40e_tx_control_block_t *
i40e_non_lso_chain(i40e_trqpair_t *itrq, mblk_t *mp, uint_t *ndesc)
{
	const mblk_t *nmp = mp;
	uint_t needed_desc = 0;
	boolean_t force_copy = B_FALSE;
	i40e_tx_control_block_t *tcb = NULL, *tcbhead = NULL, *tcbtail = NULL;
	i40e_t *i40e = itrq->itrq_i40e;
	i40e_txq_stat_t *txs = &itrq->itrq_txstat;

	/* TCB buffer is always larger than MTU. */
	ASSERT3U(msgsize(mp), <, i40e->i40e_tx_buf_size);

	while (nmp != NULL) {
		const size_t nmp_len = MBLKL(nmp);

		/* Ignore zero-length mblks. */
		if (nmp_len == 0) {
			nmp = nmp->b_cont;
			continue;
		}

		if (nmp_len < i40e->i40e_tx_dma_min || force_copy) {
			/* Compress consecutive copies into one TCB. */
			if (tcb != NULL && tcb->tcb_type == I40E_TX_COPY) {
				i40e_tx_copy_fragment(tcb, nmp, 0, nmp_len);
				nmp = nmp->b_cont;
				continue;
			}

			if ((tcb = i40e_tcb_alloc(itrq)) == NULL) {
				txs->itxs_err_notcb.value.ui64++;
				goto fail;
			}

			/*
			 * TCB DMA buffer is guaranteed to be one
			 * cookie by i40e_alloc_dma_buffer().
			 */
			i40e_tx_copy_fragment(tcb, nmp, 0, nmp_len);
			needed_desc++;
			tcb_list_append(&tcbhead, &tcbtail, tcb);
		} else {
			uint_t total_desc;

			tcb = i40e_tx_bind_fragment(itrq, nmp, 0, B_FALSE);
			if (tcb == NULL) {
				i40e_error(i40e, "dma bind failed!");
				goto fail;
			}

			/*
			 * If the new total exceeds the max or we've
			 * reached the limit and there's data left,
			 * then give up binding and copy the rest into
			 * the pre-allocated TCB buffer.
			 */
			total_desc = needed_desc + tcb->tcb_bind_ncookies;
			if ((total_desc > I40E_TX_MAX_COOKIE) ||
			    (total_desc == I40E_TX_MAX_COOKIE &&
			    nmp->b_cont != NULL)) {
				i40e_tcb_reset(tcb);
				i40e_tcb_free(itrq, tcb);

				if (tcbtail != NULL &&
				    tcbtail->tcb_type == I40E_TX_COPY) {
					tcb = tcbtail;
				} else {
					tcb = NULL;
				}

				force_copy = B_TRUE;
				txs->itxs_force_copy.value.ui64++;
				continue;
			}

			needed_desc += tcb->tcb_bind_ncookies;
			tcb_list_append(&tcbhead, &tcbtail, tcb);
		}

		nmp = nmp->b_cont;
	}

	ASSERT3P(nmp, ==, NULL);
	ASSERT3U(needed_desc, <=, I40E_TX_MAX_COOKIE);
	ASSERT3P(tcbhead, !=, NULL);
	*ndesc += needed_desc;
	return (tcbhead);

fail:
	tcb = tcbhead;
	while (tcb != NULL) {
		i40e_tx_control_block_t *next = tcb->tcb_next;

		ASSERT(tcb->tcb_type == I40E_TX_DMA ||
		    tcb->tcb_type == I40E_TX_COPY);

		tcb->tcb_mp = NULL;
		i40e_tcb_reset(tcb);
		i40e_tcb_free(itrq, tcb);
		tcb = next;
	}

	return (NULL);
}

/*
 * Section 8.4.1 of the 700-series programming guide states that a
 * segment may span up to 8 data descriptors; including both header
 * and payload data. However, empirical evidence shows that the
 * controller freezes the Tx queue when presented with a segment of 8
 * descriptors. Or, at least, when the first segment contains 8
 * descriptors. One explanation is that the controller counts the
 * context descriptor against the first segment, even though the
 * programming guide makes no mention of such a constraint. In any
 * case, we limit TSO segments to 7 descriptors to prevent Tx queue
 * freezes. We still allow non-TSO segments to utilize all 8
 * descriptors as they have not demonstrated the faulty behavior.
 */
uint_t i40e_lso_num_descs = 7;

#define	I40E_TCB_LEFT(tcb)				\
	((tcb)->tcb_dma.dmab_size - (tcb)->tcb_dma.dmab_len)

/*
 * This function is similar in spirit to i40e_non_lso_chain(), but
 * much more complicated in reality. Like the previous function, it
 * takes a packet (an LSO packet) as input and returns a chain of
 * TCBs. The complication comes with the fact that we are no longer
 * trying to fit the entire packet into 8 descriptors, but rather we
 * must fit each MSS-size segment of the LSO packet into 8 descriptors.
 * Except it's really 7 descriptors, see i40e_lso_num_descs.
 *
 * Your first inclination might be to verify that a given segment
 * spans no more than 7 mblks; but it's actually much more subtle than
 * that. First, let's describe what the hardware expects, and then we
 * can expound on the software side of things.
 *
 * For an LSO packet the hardware expects the following:
 *
 *	o Each MSS-sized segment must span no more than 7 descriptors.
 *
 *	o The header size does not count towards the segment size.
 *
 *	o If header and payload share the first descriptor, then the
 *	  controller will count the descriptor twice.
 *
 * The most important thing to keep in mind is that the hardware does
 * not view the segments in terms of mblks, like we do. The hardware
 * only sees descriptors. It will iterate each descriptor in turn,
 * keeping a tally of bytes seen and descriptors visited. If the byte
 * count hasn't reached MSS by the time the descriptor count reaches
 * 7, then the controller freezes the queue and we are stuck.
 * Furthermore, the hardware picks up its tally where it left off. So
 * if it reached MSS in the middle of a descriptor, it will start
 * tallying the next segment in the middle of that descriptor. The
 * hardware's view is entirely removed from the mblk chain or even the
 * descriptor layout. Consider these facts:
 *
 *	o The MSS will vary dpeneding on MTU and other factors.
 *
 *	o The dblk allocation will sit at various offsets within a
 *	  memory page.
 *
 *	o The page size itself could vary in the future (i.e. not
 *	  always 4K).
 *
 *	o Just because a dblk is virtually contiguous doesn't mean
 *	  it's physically contiguous. The number of cookies
 *	  (descriptors) required by a DMA bind of a single dblk is at
 *	  the mercy of the page size and physical layout.
 *
 *	o The descriptors will most often NOT start/end on a MSS
 *	  boundary. Thus the hardware will often start counting the
 *	  MSS mid descriptor and finish mid descriptor.
 *
 * The upshot of all this is that the driver must learn to think like
 * the controller; and verify that none of the constraints are broken.
 * It does this by tallying up the segment just like the hardware
 * would. This is handled by the two variables 'segsz' and 'segdesc'.
 * After each attempt to bind a dblk, we check the constaints. If
 * violated, we undo the DMA and force a copy until MSS is met. We
 * have a guarantee that the TCB buffer is larger than MTU; thus
 * ensuring we can always meet the MSS with a single copy buffer. We
 * also copy consecutive non-DMA fragments into the same TCB buffer.
 */
static i40e_tx_control_block_t *
i40e_lso_chain(i40e_trqpair_t *itrq, const mblk_t *mp,
    const mac_ether_offload_info_t *meo, const i40e_tx_context_t *tctx,
    uint_t *ndesc)
{
	size_t mp_len = MBLKL(mp);
	/*
	 * The cpoff (copy offset) variable tracks the offset inside
	 * the current mp. There are cases where the entire mp is not
	 * fully copied in one go: such as the header copy followed by
	 * a non-DMA mblk, or a TCB buffer that only has enough space
	 * to copy part of the current mp.
	 */
	size_t cpoff = 0;
	/*
	 * The segsz and segdesc variables track the controller's view
	 * of the segment. The needed_desc variable tracks the total
	 * number of data descriptors used by the driver.
	 */
	size_t segsz = 0;
	uint_t segdesc = 0;
	uint_t needed_desc = 0;
	size_t hdrcopied = 0;
	const size_t hdrlen =
	    meo->meoi_l2hlen + meo->meoi_l3hlen + meo->meoi_l4hlen;
	const size_t mss = tctx->itc_ctx_mss;
	boolean_t force_copy = B_FALSE;
	i40e_tx_control_block_t *tcb = NULL, *tcbhead = NULL, *tcbtail = NULL;
	i40e_t *i40e = itrq->itrq_i40e;
	i40e_txq_stat_t *txs = &itrq->itrq_txstat;

	/*
	 * We always copy the header in order to avoid more
	 * complicated code dealing with various edge cases.
	 */
	if ((tcb = i40e_tcb_alloc(itrq)) == NULL) {
		txs->itxs_err_notcb.value.ui64++;
		goto fail;
	}

	needed_desc++;
	tcb_list_append(&tcbhead, &tcbtail, tcb);

	while (hdrcopied < hdrlen) {
		const size_t tocopy = MIN(hdrlen - hdrcopied, mp_len);
		i40e_tx_copy_fragment(tcb, mp, 0, tocopy);
		hdrcopied += tocopy;
		cpoff += tocopy;
		if (tocopy == mp_len) {
			/*
			 * This is a bit of defensive programming. We
			 * should never have a chain too short to
			 * satisfy the headers -- but just in case.
			 */
			if ((mp = mp->b_cont) == NULL) {
				txs->itxs_tx_short.value.ui64++;
				goto fail;
			}

			while ((mp_len = MBLKL(mp)) == 0) {
				if ((mp = mp->b_cont) == NULL) {
					txs->itxs_tx_short.value.ui64++;
					goto fail;
				}
			}
			cpoff = 0;
		}
	}
	ASSERT3U(hdrcopied, ==, hdrlen);

	/*
	 * A single descriptor containing both header and data is
	 * counted twice by the controller.
	 */
	if (mp_len < i40e->i40e_tx_dma_min) {
		segdesc = 2;
	} else {
		segdesc = 1;
	}

	while (mp != NULL) {
		mp_len = MBLKL(mp);
force_copy:
		/* Ignore zero-length mblks. */
		if (mp_len == 0) {
			mp = mp->b_cont;
			cpoff = 0;
			continue;
		}

		/*
		 * We copy into the preallocated TCB buffer when the
		 * current fragment is less than the DMA threshold OR
		 * when the DMA bind can't meet the controller's
		 * segment descriptor limit.
		 */
		if (mp_len < i40e->i40e_tx_dma_min || force_copy) {
			size_t tocopy;

			/*
			 * Our objective here is to compress
			 * consecutive copies into one TCB (until it
			 * is full). If there is no current TCB, or if
			 * it is a DMA TCB, then allocate a new one.
			 */
			if (tcb == NULL ||
			    (tcb != NULL && tcb->tcb_type != I40E_TX_COPY)) {
				if ((tcb = i40e_tcb_alloc(itrq)) == NULL) {
					txs->itxs_err_notcb.value.ui64++;
					goto fail;
				}

				/*
				 * The TCB DMA buffer is guaranteed to
				 * be one cookie by i40e_alloc_dma_buffer().
				 */
				needed_desc++;
				segdesc++;
				ASSERT3U(segdesc, <=, i40e_lso_num_descs);
				tcb_list_append(&tcbhead, &tcbtail, tcb);
			} else if (segdesc == 0) {
				/*
				 * We are copying into an existing TCB
				 * but we just crossed the MSS
				 * boundary. Make sure to increment
				 * segdesc to track the descriptor
				 * count as the hardware would.
				 */
				segdesc++;
			}

			tocopy = MIN(I40E_TCB_LEFT(tcb), mp_len - cpoff);
			i40e_tx_copy_fragment(tcb, mp, cpoff, tocopy);
			cpoff += tocopy;
			segsz += tocopy;

			/* We have consumed the current mp. */
			if (cpoff == mp_len) {
				mp = mp->b_cont;
				cpoff = 0;
			}

			/* We have consumed the current TCB buffer. */
			if (I40E_TCB_LEFT(tcb) == 0) {
				tcb = NULL;
			}

			/*
			 * We have met MSS with this copy; restart the
			 * counters.
			 */
			if (segsz >= mss) {
				segsz = segsz % mss;
				segdesc = segsz == 0 ? 0 : 1;
				force_copy = B_FALSE;
			}

			/*
			 * We are at the controller's descriptor
			 * limit; we must copy into the current TCB
			 * until MSS is reached. The TCB buffer is
			 * always bigger than the MTU so we know it is
			 * big enough to meet the MSS.
			 */
			if (segdesc == i40e_lso_num_descs) {
				force_copy = B_TRUE;
			}
		} else {
			uint_t tsegdesc = segdesc;
			size_t tsegsz = segsz;

			ASSERT(force_copy == B_FALSE);
			ASSERT3U(tsegdesc, <, i40e_lso_num_descs);

			tcb = i40e_tx_bind_fragment(itrq, mp, cpoff, B_TRUE);
			if (tcb == NULL) {
				i40e_error(i40e, "dma bind failed!");
				goto fail;
			}

			for (uint_t i = 0; i < tcb->tcb_bind_ncookies; i++) {
				struct i40e_dma_bind_info dbi =
				    tcb->tcb_bind_info[i];

				tsegsz += dbi.dbi_len;
				tsegdesc++;
				ASSERT3U(tsegdesc, <=, i40e_lso_num_descs);

				/*
				 * We've met the MSS with this portion
				 * of the DMA.
				 */
				if (tsegsz >= mss) {
					tsegsz = tsegsz % mss;
					tsegdesc = tsegsz == 0 ? 0 : 1;
				}

				/*
				 * We've reached max descriptors but
				 * have not met the MSS. Undo the bind
				 * and instead copy.
				 */
				if (tsegdesc == i40e_lso_num_descs) {
					i40e_tcb_reset(tcb);
					i40e_tcb_free(itrq, tcb);

					if (tcbtail != NULL &&
					    I40E_TCB_LEFT(tcb) > 0 &&
					    tcbtail->tcb_type == I40E_TX_COPY) {
						tcb = tcbtail;
					} else {
						tcb = NULL;
					}

					/*
					 * Remember, we are still on
					 * the same mp.
					 */
					force_copy = B_TRUE;
					txs->itxs_tso_force_copy.value.ui64++;
					goto force_copy;
				}
			}

			ASSERT3U(tsegdesc, <=, i40e_lso_num_descs);
			ASSERT3U(tsegsz, <, mss);

			/*
			 * We've made if through the loop without
			 * breaking the segment descriptor contract
			 * with the controller -- replace the segment
			 * tracking values with the temporary ones.
			 */
			segdesc = tsegdesc;
			segsz = tsegsz;
			needed_desc += tcb->tcb_bind_ncookies;
			cpoff = 0;
			tcb_list_append(&tcbhead, &tcbtail, tcb);
			mp = mp->b_cont;
		}
	}

	ASSERT3P(mp, ==, NULL);
	ASSERT3P(tcbhead, !=, NULL);
	*ndesc += needed_desc;
	return (tcbhead);

fail:
	tcb = tcbhead;
	while (tcb != NULL) {
		i40e_tx_control_block_t *next = tcb->tcb_next;

		ASSERT(tcb->tcb_type == I40E_TX_DMA ||
		    tcb->tcb_type == I40E_TX_COPY);

		tcb->tcb_mp = NULL;
		i40e_tcb_reset(tcb);
		i40e_tcb_free(itrq, tcb);
		tcb = next;
	}

	return (NULL);
}

/*
 * Keep track of activity through the transmit data path.
 *
 * We need to ensure we don't try and transmit when a trqpair has been
 * stopped, nor do we want to stop a trqpair whilst transmitting.
 */
static boolean_t
i40e_ring_tx_enter(i40e_trqpair_t *itrq)
{
	boolean_t allow;

	mutex_enter(&itrq->itrq_tx_lock);
	allow = !itrq->itrq_tx_quiesce;
	if (allow)
		itrq->itrq_tx_active++;
	mutex_exit(&itrq->itrq_tx_lock);

	return (allow);
}

static void
i40e_ring_tx_exit_nolock(i40e_trqpair_t *itrq)
{
	ASSERT(MUTEX_HELD(&itrq->itrq_tx_lock));

	itrq->itrq_tx_active--;
	if (itrq->itrq_tx_quiesce)
		cv_signal(&itrq->itrq_tx_cv);
}

static void
i40e_ring_tx_exit(i40e_trqpair_t *itrq)
{
	mutex_enter(&itrq->itrq_tx_lock);
	i40e_ring_tx_exit_nolock(itrq);
	mutex_exit(&itrq->itrq_tx_lock);
}


/*
 * Tell the transmit path to quiesce and wait until there is no
 * more activity.
 * Will return B_TRUE if the transmit path is already quiesced, B_FALSE
 * otherwise.
 */
boolean_t
i40e_ring_tx_quiesce(i40e_trqpair_t *itrq)
{
	mutex_enter(&itrq->itrq_tx_lock);
	if (itrq->itrq_tx_quiesce) {
		/*
		 * When itrq_tx_quiesce is set, then the ring has already
		 * been shutdown.
		 */
		mutex_exit(&itrq->itrq_tx_lock);
		return (B_TRUE);
	}

	/*
	 * Tell any threads in transmit path this trqpair is quiesced and
	 * wait until they've all exited the critical code path.
	 */
	itrq->itrq_tx_quiesce = B_TRUE;
	while (itrq->itrq_tx_active > 0)
		cv_wait(&itrq->itrq_tx_cv, &itrq->itrq_tx_lock);

	mutex_exit(&itrq->itrq_tx_lock);

	return (B_FALSE);
}

/*
 * We've been asked to send a message block on the wire. We'll only have a
 * single chain. There will not be any b_next pointers; however, there may be
 * multiple b_cont blocks. The number of b_cont blocks may exceed the
 * controller's Tx descriptor limit.
 *
 * We may do one of three things with any given mblk_t chain:
 *
 *   1) Drop it
 *   2) Transmit it
 *   3) Return it
 *
 * If we return it to MAC, then MAC will flow control on our behalf. In other
 * words, it won't send us anything until we tell it that it's okay to send us
 * something.
 */
mblk_t *
i40e_ring_tx(void *arg, mblk_t *mp)
{
	size_t msglen;
	i40e_tx_control_block_t *tcb_ctx = NULL, *tcb = NULL, *tcbhead = NULL;
	i40e_tx_context_desc_t *ctxdesc;
	mac_ether_offload_info_t meo;
	i40e_tx_context_t tctx;
	int type;
	uint_t needed_desc = 0;
	boolean_t do_ctx_desc = B_FALSE, use_lso = B_FALSE;

	i40e_trqpair_t *itrq = arg;
	i40e_t *i40e = itrq->itrq_i40e;
	i40e_hw_t *hw = &i40e->i40e_hw_space;
	i40e_txq_stat_t *txs = &itrq->itrq_txstat;

	ASSERT(mp->b_next == NULL);

	if (!(i40e->i40e_state & I40E_STARTED) ||
	    (i40e->i40e_state & I40E_OVERTEMP) ||
	    (i40e->i40e_state & I40E_SUSPENDED) ||
	    (i40e->i40e_state & I40E_ERROR) ||
	    (i40e->i40e_link_state != LINK_STATE_UP) ||
	    !i40e_ring_tx_enter(itrq)) {
		freemsg(mp);
		return (NULL);
	}

	/*
	 * Parse packet headers for use by any requested offloads.  That offload
	 * logic will later determine if the results here were adequate.
	 */
	mac_ether_offload_info(mp, &meo);

	/*
	 * Figure out the relevant context about this frame that we might need
	 * for enabling checksum, LSO, etc. This also fills in information that
	 * we might set around the packet type, etc.
	 */
	if (i40e_tx_context(i40e, itrq, mp, &meo, &tctx) < 0) {
		freemsg(mp);
		itrq->itrq_txstat.itxs_err_context.value.ui64++;
		i40e_ring_tx_exit(itrq);
		return (NULL);
	}
	if (tctx.itc_ctx_cmdflags & I40E_TX_CTX_DESC_TSO) {
		use_lso = B_TRUE;
		do_ctx_desc = B_TRUE;
	}

	/*
	 * For the primordial driver we can punt on doing any recycling right
	 * now; however, longer term we need to probably do some more pro-active
	 * recycling to cut back on stalls in the TX path.
	 */

	msglen = msgsize(mp);

	if (do_ctx_desc) {
		/*
		 * If we're doing tunneling or LSO, then we'll need a TX
		 * context descriptor in addition to one or more TX data
		 * descriptors.  Since there's no data DMA block or handle
		 * associated with the context descriptor, we create a special
		 * control block that behaves effectively like a NOP.
		 */
		if ((tcb_ctx = i40e_tcb_alloc(itrq)) == NULL) {
			txs->itxs_err_notcb.value.ui64++;
			goto txfail;
		}
		tcb_ctx->tcb_type = I40E_TX_DESC;
		needed_desc++;
	}

	if (!use_lso) {
		tcbhead = i40e_non_lso_chain(itrq, mp, &needed_desc);
	} else {
		tcbhead = i40e_lso_chain(itrq, mp, &meo, &tctx, &needed_desc);
	}

	if (tcbhead == NULL)
		goto txfail;

	tcbhead->tcb_mp = mp;

	/*
	 * The second condition ensures that 'itrq_desc_tail' never
	 * equals 'itrq_desc_head'. This enforces the rule found in
	 * the second bullet point of section 8.4.3.1.5 of the XL710
	 * PG, which declares the TAIL pointer in I40E_QTX_TAIL should
	 * never overlap with the head. This means that we only ever
	 * have 'itrq_tx_ring_size - 1' total available descriptors.
	 */
	mutex_enter(&itrq->itrq_tx_lock);
	if (itrq->itrq_desc_free < i40e->i40e_tx_block_thresh ||
	    (itrq->itrq_desc_free - 1) < needed_desc) {
		txs->itxs_err_nodescs.value.ui64++;
		mutex_exit(&itrq->itrq_tx_lock);
		goto txfail;
	}

	if (do_ctx_desc) {
		/*
		 * If we're enabling any offloads for this frame, then we'll
		 * need to build up a transmit context descriptor, first.  The
		 * context descriptor needs to be placed in the TX ring before
		 * the data descriptor(s).  See section 8.4.2, table 8-16
		 */
		uint_t tail = itrq->itrq_desc_tail;
		itrq->itrq_desc_free--;
		ctxdesc = (i40e_tx_context_desc_t *)&itrq->itrq_desc_ring[tail];
		itrq->itrq_tcb_work_list[tail] = tcb_ctx;
		itrq->itrq_desc_tail = i40e_next_desc(tail, 1,
		    itrq->itrq_tx_ring_size);

		/* QW0 */
		type = I40E_TX_DESC_DTYPE_CONTEXT;
		ctxdesc->tunneling_params = 0;
		ctxdesc->l2tag2 = 0;

		/* QW1 */
		ctxdesc->type_cmd_tso_mss = CPU_TO_LE64((uint64_t)type);
		if (tctx.itc_ctx_cmdflags & I40E_TX_CTX_DESC_TSO) {
			ctxdesc->type_cmd_tso_mss |= CPU_TO_LE64((uint64_t)
			    ((uint64_t)tctx.itc_ctx_cmdflags <<
			    I40E_TXD_CTX_QW1_CMD_SHIFT) |
			    ((uint64_t)tctx.itc_ctx_tsolen <<
			    I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
			    ((uint64_t)tctx.itc_ctx_mss <<
			    I40E_TXD_CTX_QW1_MSS_SHIFT));
		}
	}

	tcb = tcbhead;
	while (tcb != NULL) {

		itrq->itrq_tcb_work_list[itrq->itrq_desc_tail] = tcb;
		if (tcb->tcb_type == I40E_TX_COPY) {
			boolean_t last_desc = (tcb->tcb_next == NULL);

			i40e_tx_set_data_desc(itrq, &tctx,
			    (caddr_t)tcb->tcb_dma.dmab_dma_address,
			    tcb->tcb_dma.dmab_len, last_desc);
		} else {
			boolean_t last_desc = B_FALSE;
			ASSERT3S(tcb->tcb_type, ==, I40E_TX_DMA);

			for (uint_t c = 0; c < tcb->tcb_bind_ncookies; c++) {
				last_desc = (c == tcb->tcb_bind_ncookies - 1) &&
				    (tcb->tcb_next == NULL);

				i40e_tx_set_data_desc(itrq, &tctx,
				    tcb->tcb_bind_info[c].dbi_paddr,
				    tcb->tcb_bind_info[c].dbi_len,
				    last_desc);
			}
		}

		tcb = tcb->tcb_next;
	}

	/*
	 * Now, finally, sync the DMA data and alert hardware.
	 */
	I40E_DMA_SYNC(&itrq->itrq_desc_area, DDI_DMA_SYNC_FORDEV);

	I40E_WRITE_REG(hw, I40E_QTX_TAIL(itrq->itrq_index),
	    itrq->itrq_desc_tail);

	if (i40e_check_acc_handle(i40e->i40e_osdep_space.ios_reg_handle) !=
	    DDI_FM_OK) {
		/*
		 * Note, we can't really go through and clean this up very well,
		 * because the memory has been given to the device, so just
		 * indicate it's been transmitted.
		 */
		ddi_fm_service_impact(i40e->i40e_dip, DDI_SERVICE_DEGRADED);
		atomic_or_32(&i40e->i40e_state, I40E_ERROR);
	}

	txs->itxs_bytes.value.ui64 += msglen;
	txs->itxs_packets.value.ui64++;
	txs->itxs_descriptors.value.ui64 += needed_desc;

	i40e_ring_tx_exit_nolock(itrq);

	mutex_exit(&itrq->itrq_tx_lock);

	return (NULL);

txfail:
	/*
	 * We ran out of resources. Return it to MAC and indicate that we'll
	 * need to signal MAC. If there are allocated tcb's, return them now.
	 * Make sure to reset their message block's, since we'll return them
	 * back to MAC.
	 */
	if (tcb_ctx != NULL) {
		tcb_ctx->tcb_mp = NULL;
		i40e_tcb_reset(tcb_ctx);
		i40e_tcb_free(itrq, tcb_ctx);
	}

	tcb = tcbhead;
	while (tcb != NULL) {
		i40e_tx_control_block_t *next = tcb->tcb_next;

		ASSERT(tcb->tcb_type == I40E_TX_DMA ||
		    tcb->tcb_type == I40E_TX_COPY);

		tcb->tcb_mp = NULL;
		i40e_tcb_reset(tcb);
		i40e_tcb_free(itrq, tcb);
		tcb = next;
	}

	mutex_enter(&itrq->itrq_tx_lock);
	i40e_ring_tx_exit_nolock(itrq);
	itrq->itrq_tx_blocked = B_TRUE;
	mutex_exit(&itrq->itrq_tx_lock);

	return (mp);
}