path: root/arch/tile/include/gxio/mpipe.h

/*
 * Copyright 2012 Tilera Corporation. All Rights Reserved.
 *
 *   This program is free software; you can redistribute it and/or
 *   modify it under the terms of the GNU General Public License
 *   as published by the Free Software Foundation, version 2.
 *
 *   This program is distributed in the hope that it will be useful, but
 *   WITHOUT ANY WARRANTY; without even the implied warranty of
 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
 *   NON INFRINGEMENT.  See the GNU General Public License for
 *   more details.
 */

#ifndef _GXIO_MPIPE_H_
#define _GXIO_MPIPE_H_

/*
 *
 * An API for allocating, configuring, and manipulating mPIPE hardware
 * resources.
 */

#include <gxio/common.h>
#include <gxio/dma_queue.h>

#include <linux/time.h>

#include <arch/mpipe_def.h>
#include <arch/mpipe_shm.h>

#include <hv/drv_mpipe_intf.h>
#include <hv/iorpc.h>

/*
 *
 * The TILE-Gx mPIPE&tm; shim provides Ethernet connectivity, packet
 * classification, and packet load balancing services.  The
 * gxio_mpipe_ API, declared in <gxio/mpipe.h>, allows applications to
 * allocate mPIPE IO channels, configure packet distribution
 * parameters, and send and receive Ethernet packets.  The API is
 * designed to be a minimal wrapper around the mPIPE hardware, making
 * system calls only where necessary to preserve inter-process
 * protection guarantees.
 *
 * The APIs described below allow the programmer to allocate and
 * configure mPIPE resources.  As described below, the mPIPE is a
 * single shared hardware device that provides partitionable resources
 * that are shared between all applications in the system.  The
 * gxio_mpipe_ API allows userspace code to make resource request
 * calls to the hypervisor, which in turns keeps track of the
 * resources in use by all applications, maintains protection
 * guarantees, and resets resources upon application shutdown.
 *
 * We strongly recommend reading the mPIPE section of the IO Device
 * Guide (UG404) before working with this API.  Most functions in the
 * gxio_mpipe_ API are directly analogous to hardware interfaces and
 * the documentation assumes that the reader understands those
 * hardware interfaces.
 *
 * @section mpipe__ingress mPIPE Ingress Hardware Resources
 *
 * The mPIPE ingress hardware provides extensive hardware offload for
 * tasks like packet header parsing, load balancing, and memory
 * management.  This section provides a brief introduction to the
 * hardware components and the gxio_mpipe_ calls used to manage them;
 * see the IO Device Guide for a much more detailed description of the
 * mPIPE's capabilities.
 *
 * When a packet arrives at one of the mPIPE's Ethernet MACs, it is
 * assigned a channel number indicating which MAC received it.  It
 * then proceeds through the following hardware pipeline:
 *
 * @subsection mpipe__classification Classification
 *
 * A set of classification processors run header parsing code on each
 * incoming packet, extracting information including the destination
 * MAC address, VLAN, Ethernet type, and five-tuple hash.  Some of
 * this information is then used to choose which buffer stack will be
 * used to hold the packet, and which bucket will be used by the load
 * balancer to determine which application will receive the packet.
 *
 * The rules by which the buffer stack and bucket are chosen can be
 * configured via the @ref gxio_mpipe_classifier API.  A given app can
 * specify multiple rules, each one specifying a bucket range, and a
 * set of buffer stacks, to be used for packets matching the rule.
 * Each rule can optionally specify a restricted set of channels,
 * VLANs, and/or dMACs, in which it is interested.  By default, a
 * given rule starts out matching all channels associated with the
 * mPIPE context's set of open links; all VLANs; and all dMACs.
 * Subsequent restrictions can then be added.
 *
 * @subsection mpipe__load_balancing Load Balancing
 *
 * The mPIPE load balancer is responsible for choosing the NotifRing
 * to which the packet will be delivered.  This decision is based on
 * the bucket number indicated by the classification program.  In
 * general, the bucket number is based on some number of low bits of
 * the packet's flow hash (applications that aren't interested in flow
 * hashing use a single bucket).  Each load balancer bucket keeps a
 * record of the NotifRing to which packets directed to that bucket
 * are currently being delivered.  Based on the bucket's load
 * balancing mode (@ref gxio_mpipe_bucket_mode_t), the load balancer
 * either forwards the packet to the previously assigned NotifRing or
 * decides to choose a new NotifRing.  If a new NotifRing is required,
 * the load balancer chooses the least loaded ring in the NotifGroup
 * associated with the bucket.
 *
 * The load balancer is a shared resource.  Each application needs to
 * explicitly allocate NotifRings, NotifGroups, and buckets, using
 * gxio_mpipe_alloc_notif_rings(), gxio_mpipe_alloc_notif_groups(),
 * and gxio_mpipe_alloc_buckets().  Then the application needs to
 * configure them using gxio_mpipe_init_notif_ring() and
 * gxio_mpipe_init_notif_group_and_buckets().
 *
 * @subsection mpipe__buffers Buffer Selection and Packet Delivery
 *
 * Once the load balancer has chosen the destination NotifRing, the
 * mPIPE DMA engine pops at least one buffer off of the 'buffer stack'
 * chosen by the classification program and DMAs the packet data into
 * that buffer.  Each buffer stack provides a hardware-accelerated
 * stack of data buffers with the same size.  If the packet data is
 * larger than the buffers provided by the chosen buffer stack, the
 * mPIPE hardware pops off multiple buffers and chains the packet data
 * through a multi-buffer linked list.  Once the packet data is
 * delivered to the buffer(s), the mPIPE hardware writes the
 * ::gxio_mpipe_idesc_t metadata object (calculated by the classifier)
 * into the NotifRing and increments the number of packets delivered
 * to that ring.
 *
 * Applications can push buffers onto a buffer stack by calling
 * gxio_mpipe_push_buffer() or by egressing a packet with the
 * ::gxio_mpipe_edesc_t::hwb bit set, indicating that the egressed
 * buffers should be returned to the stack.
 *
 * Applications can allocate and initialize buffer stacks with the
 * gxio_mpipe_alloc_buffer_stacks() and gxio_mpipe_init_buffer_stack()
 * APIs.
 *
 * The application must also register the memory pages that will hold
 * packets.  This requires calling gxio_mpipe_register_page() for each
 * memory page that will hold packets allocated by the application for
 * a given buffer stack.  Since each buffer stack is limited to 16
 * registered pages, it may be necessary to use huge pages, or even
 * extremely huge pages, to hold all the buffers.
 *
 * @subsection mpipe__iqueue NotifRings
 *
 * Each NotifRing is a region of shared memory, allocated by the
 * application, to which the mPIPE delivers packet descriptors
 * (::gxio_mpipe_idesc_t).  The application can allocate them via
 * gxio_mpipe_alloc_notif_rings().  The application can then either
 * explicitly initialize them with gxio_mpipe_init_notif_ring() and
 * then read from them manually, or can make use of the convenience
 * wrappers provided by @ref gxio_mpipe_wrappers.
 *
 * @section mpipe__egress mPIPE Egress Hardware
 *
 * Applications use eDMA rings to queue packets for egress.  The
 * application can allocate them via gxio_mpipe_alloc_edma_rings().
 * The application can then either explicitly initialize them with
 * gxio_mpipe_init_edma_ring() and then write to them manually, or
 * can make use of the convenience wrappers provided by
 * @ref gxio_mpipe_wrappers.
 *
 * @section gxio__shortcomings Plans for Future API Revisions
 *
 * The API defined here is only an initial version of the mPIPE API.
 * Future plans include:
 *
 * - Higher level wrapper functions to provide common initialization
 * patterns.  This should help users start writing mPIPE programs
 * without having to learn the details of the hardware.
 *
 * - Support for reset and deallocation of resources, including
 * cleanup upon application shutdown.
 *
 * - Support for calling these APIs in the BME.
 *
 * - Support for IO interrupts.
 *
 * - Clearer definitions of thread safety guarantees.
 *
 * @section gxio__mpipe_examples Examples
 *
 * See the following mPIPE example programs for more information about
 * allocating mPIPE resources and using them in real applications:
 *
 * - @ref mpipe/ingress/app.c : Receiving packets.
 *
 * - @ref mpipe/forward/app.c : Forwarding packets.
 *
 * Note that there are several more examples.
 */

/* Flags that can be passed to resource allocation functions. */
enum gxio_mpipe_alloc_flags_e {
	/* Require an allocation to start at a specified resource index. */
	GXIO_MPIPE_ALLOC_FIXED = HV_MPIPE_ALLOC_FIXED,
};

/* Flags that can be passed to memory registration functions. */
enum gxio_mpipe_mem_flags_e {
	/* Do not fill L3 when writing, and invalidate lines upon egress. */
	GXIO_MPIPE_MEM_FLAG_NT_HINT = IORPC_MEM_BUFFER_FLAG_NT_HINT,

	/* L3 cache fills should only populate IO cache ways. */
	GXIO_MPIPE_MEM_FLAG_IO_PIN = IORPC_MEM_BUFFER_FLAG_IO_PIN,
};

/* An ingress packet descriptor.  When a packet arrives, the mPIPE
 * hardware generates this structure and writes it into a NotifRing.
 */
typedef MPIPE_PDESC_t gxio_mpipe_idesc_t;

/* An egress command descriptor.  Applications write this structure
 * into eDMA rings and the hardware performs the indicated operation
 * (normally involving egressing some bytes).  Note that egressing a
 * single packet may involve multiple egress command descriptors.
 */
typedef MPIPE_EDMA_DESC_t gxio_mpipe_edesc_t;

/* Get the "va" field from an "idesc".
 *
 * This is the address at which the ingress hardware copied the first
 * byte of the packet.
 *
 * If the classifier detected a custom header, then this will point to
 * the custom header, and gxio_mpipe_idesc_get_l2_start() will point
 * to the actual L2 header.
 *
 * Note that this value may be misleading if "idesc->be" is set.
 *
 * @param idesc An ingress packet descriptor.
 */
static inline unsigned char *gxio_mpipe_idesc_get_va(gxio_mpipe_idesc_t *idesc)
{
	return (unsigned char *)(long)idesc->va;
}

/* Get the "xfer_size" from an "idesc".
 *
 * This is the actual number of packet bytes transferred into memory
 * by the hardware.
 *
 * Note that this value may be misleading if "idesc->be" is set.
 *
 * @param idesc An ingress packet descriptor.
 *
 * ISSUE: Is this the best name for this?
 * FIXME: Add more docs about chaining, clipping, etc.
 */
static inline unsigned int gxio_mpipe_idesc_get_xfer_size(gxio_mpipe_idesc_t
							  *idesc)
{
	return idesc->l2_size;
}

/* Get the "l2_offset" from an "idesc".
 *
 * Extremely customized classifiers might not support this function.
 *
 * This is the number of bytes between the "va" and the L2 header.
 *
 * The L2 header consists of a destination mac address, a source mac
 * address, and an initial ethertype.  Various initial ethertypes
 * allow encoding extra information in the L2 header, often including
 * a vlan, and/or a new ethertype.
 *
 * Note that the "l2_offset" will be non-zero if (and only if) the
 * classifier processed a custom header for the packet.
 *
 * @param idesc An ingress packet descriptor.
 */
static inline uint8_t gxio_mpipe_idesc_get_l2_offset(gxio_mpipe_idesc_t *idesc)
{
	return (idesc->custom1 >> 32) & 0xFF;
}

/* Get the "l2_start" from an "idesc".
 *
 * This is simply gxio_mpipe_idesc_get_va() plus
 * gxio_mpipe_idesc_get_l2_offset().
 *
 * @param idesc An ingress packet descriptor.
 */
static inline unsigned char *gxio_mpipe_idesc_get_l2_start(gxio_mpipe_idesc_t
							   *idesc)
{
	unsigned char *va = gxio_mpipe_idesc_get_va(idesc);
	return va + gxio_mpipe_idesc_get_l2_offset(idesc);
}

/* Get the "l2_length" from an "idesc".
 *
 * This is simply gxio_mpipe_idesc_get_xfer_size() minus
 * gxio_mpipe_idesc_get_l2_offset().
 *
 * @param idesc An ingress packet descriptor.
 */
static inline unsigned int gxio_mpipe_idesc_get_l2_length(gxio_mpipe_idesc_t
							  *idesc)
{
	unsigned int xfer_size = idesc->l2_size;
	return xfer_size - gxio_mpipe_idesc_get_l2_offset(idesc);
}

/* A context object used to manage mPIPE hardware resources. */
typedef struct {

	/* File descriptor for calling up to Linux (and thus the HV). */
	int fd;

	/* The VA at which configuration registers are mapped. */
	char *mmio_cfg_base;

	/* The VA at which IDMA, EDMA, and buffer manager are mapped. */
	char *mmio_fast_base;

	/* The "initialized" buffer stacks. */
	gxio_mpipe_rules_stacks_t __stacks;

} gxio_mpipe_context_t;

/* This is only used internally, but it's most easily made visible here. */
typedef gxio_mpipe_context_t gxio_mpipe_info_context_t;

/* Initialize an mPIPE context.
 *
 * This function allocates an mPIPE "service domain" and maps the MMIO
 * registers into the caller's VA space.
 *
 * @param context Context object to be initialized.
 * @param mpipe_instance Instance number of mPIPE shim to be controlled via
 *  context.
 */
extern int gxio_mpipe_init(gxio_mpipe_context_t *context,
			   unsigned int mpipe_instance);

/* Destroy an mPIPE context.
 *
 * This function frees the mPIPE "service domain" and unmaps the MMIO
 * registers from the caller's VA space.
 *
 * If a user process exits without calling this routine, the kernel
 * will destroy the mPIPE context as part of process teardown.
 *
 * @param context Context object to be destroyed.
 */
extern int gxio_mpipe_destroy(gxio_mpipe_context_t *context);

/*****************************************************************
 *                         Buffer Stacks                          *
 ******************************************************************/

/* Allocate a set of buffer stacks.
 *
 * The return value is NOT interesting if count is zero.
 *
 * @param context An initialized mPIPE context.
 * @param count Number of stacks required.
 * @param first Index of first stack if ::GXIO_MPIPE_ALLOC_FIXED flag is set,
 *   otherwise ignored.
 * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
 * @return Index of first allocated buffer stack, or
 * ::GXIO_MPIPE_ERR_NO_BUFFER_STACK if allocation failed.
 */
extern int gxio_mpipe_alloc_buffer_stacks(gxio_mpipe_context_t *context,
					  unsigned int count,
					  unsigned int first,
					  unsigned int flags);

/* Enum codes for buffer sizes supported by mPIPE. */
typedef enum {
	/* 128 byte packet data buffer. */
	GXIO_MPIPE_BUFFER_SIZE_128 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_128,
	/* 256 byte packet data buffer. */
	GXIO_MPIPE_BUFFER_SIZE_256 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_256,
	/* 512 byte packet data buffer. */
	GXIO_MPIPE_BUFFER_SIZE_512 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_512,
	/* 1024 byte packet data buffer. */
	GXIO_MPIPE_BUFFER_SIZE_1024 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_1024,
	/* 1664 byte packet data buffer. */
	GXIO_MPIPE_BUFFER_SIZE_1664 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_1664,
	/* 4096 byte packet data buffer. */
	GXIO_MPIPE_BUFFER_SIZE_4096 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_4096,
	/* 10368 byte packet data buffer. */
	GXIO_MPIPE_BUFFER_SIZE_10368 =
		MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_10368,
	/* 16384 byte packet data buffer. */
	GXIO_MPIPE_BUFFER_SIZE_16384 = MPIPE_BSM_INIT_DAT_1__SIZE_VAL_BSZ_16384
} gxio_mpipe_buffer_size_enum_t;

/* Convert a buffer size in bytes into a buffer size enum. */
extern gxio_mpipe_buffer_size_enum_t
gxio_mpipe_buffer_size_to_buffer_size_enum(size_t size);

/* Convert a buffer size enum into a buffer size in bytes. */
extern size_t
gxio_mpipe_buffer_size_enum_to_buffer_size(gxio_mpipe_buffer_size_enum_t
					   buffer_size_enum);

/* Calculate the number of bytes required to store a given number of
 * buffers in the memory registered with a buffer stack via
 * gxio_mpipe_init_buffer_stack().
 */
extern size_t gxio_mpipe_calc_buffer_stack_bytes(unsigned long buffers);

/* Initialize a buffer stack.  This function binds a region of memory
 * to be used by the hardware for storing buffer addresses pushed via
 * gxio_mpipe_push_buffer() or as the result of sending a buffer out
 * the egress with the 'push to stack when done' bit set.  Once this
 * function returns, the memory region's contents may be arbitrarily
 * modified by the hardware at any time and software should not access
 * the memory region again.
 *
 * @param context An initialized mPIPE context.
 * @param stack The buffer stack index.
 * @param buffer_size_enum The size of each buffer in the buffer stack,
 * as an enum.
 * @param mem The address of the buffer stack.  This memory must be
 * physically contiguous and aligned to a 64kB boundary.
 * @param mem_size The size of the buffer stack, in bytes.
 * @param mem_flags ::gxio_mpipe_mem_flags_e memory flags.
 * @return Zero on success, ::GXIO_MPIPE_ERR_INVAL_BUFFER_SIZE if
 * buffer_size_enum is invalid, ::GXIO_MPIPE_ERR_BAD_BUFFER_STACK if
 * stack has not been allocated.
 */
extern int gxio_mpipe_init_buffer_stack(gxio_mpipe_context_t *context,
					unsigned int stack,
					gxio_mpipe_buffer_size_enum_t
					buffer_size_enum, void *mem,
					size_t mem_size,
					unsigned int mem_flags);

/* Push a buffer onto a previously initialized buffer stack.
 *
 * The size of the buffer being pushed must match the size that was
 * registered with gxio_mpipe_init_buffer_stack().  All packet buffer
 * addresses are 128-byte aligned; the low 7 bits of the specified
 * buffer address will be ignored.
 *
 * @param context An initialized mPIPE context.
 * @param stack The buffer stack index.
 * @param buffer The buffer (the low seven bits are ignored).
 */
static inline void gxio_mpipe_push_buffer(gxio_mpipe_context_t *context,
					  unsigned int stack, void *buffer)
{
	MPIPE_BSM_REGION_ADDR_t offset = { {0} };
	MPIPE_BSM_REGION_VAL_t val = { {0} };

	/*
	 * The mmio_fast_base region starts at the IDMA region, so subtract
	 * off that initial offset.
	 */
	offset.region =
		MPIPE_MMIO_ADDR__REGION_VAL_BSM -
		MPIPE_MMIO_ADDR__REGION_VAL_IDMA;
	offset.stack = stack;

#if __SIZEOF_POINTER__ == 4
	val.va = ((ulong) buffer) >> MPIPE_BSM_REGION_VAL__VA_SHIFT;
#else
	val.va = ((long)buffer) >> MPIPE_BSM_REGION_VAL__VA_SHIFT;
#endif

	__gxio_mmio_write(context->mmio_fast_base + offset.word, val.word);
}

/* Pop a buffer off of a previously initialized buffer stack.
 *
 * @param context An initialized mPIPE context.
 * @param stack The buffer stack index.
 * @return The buffer, or NULL if the stack is empty.
 */
static inline void *gxio_mpipe_pop_buffer(gxio_mpipe_context_t *context,
					  unsigned int stack)
{
	MPIPE_BSM_REGION_ADDR_t offset = { {0} };

	/*
	 * The mmio_fast_base region starts at the IDMA region, so subtract
	 * off that initial offset.
	 */
	offset.region =
		MPIPE_MMIO_ADDR__REGION_VAL_BSM -
		MPIPE_MMIO_ADDR__REGION_VAL_IDMA;
	offset.stack = stack;

	while (1) {
		/*
		 * Case 1: val.c == ..._UNCHAINED, va is non-zero.
		 * Case 2: val.c == ..._INVALID, va is zero.
		 * Case 3: val.c == ..._NOT_RDY, va is zero.
		 */
		MPIPE_BSM_REGION_VAL_t val;
		val.word =
			__gxio_mmio_read(context->mmio_fast_base +
					 offset.word);

		/*
		 * Handle case 1 and 2 by returning the buffer (or NULL).
		 * Handle case 3 by waiting for the prefetch buffer to refill.
		 */
		if (val.c != MPIPE_EDMA_DESC_WORD1__C_VAL_NOT_RDY)
			return (void *)((unsigned long)val.
					va << MPIPE_BSM_REGION_VAL__VA_SHIFT);
	}
}

/*****************************************************************
 *                          NotifRings                            *
 ******************************************************************/

/* Allocate a set of NotifRings.
 *
 * The return value is NOT interesting if count is zero.
 *
 * Note that NotifRings are allocated in chunks, so allocating one at
 * a time is much less efficient than allocating several at once.
 *
 * @param context An initialized mPIPE context.
 * @param count Number of NotifRings required.
 * @param first Index of first NotifRing if ::GXIO_MPIPE_ALLOC_FIXED flag
 *   is set, otherwise ignored.
 * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
 * @return Index of first allocated buffer NotifRing, or
 * ::GXIO_MPIPE_ERR_NO_NOTIF_RING if allocation failed.
 */
extern int gxio_mpipe_alloc_notif_rings(gxio_mpipe_context_t *context,
					unsigned int count, unsigned int first,
					unsigned int flags);

/* Initialize a NotifRing, using the given memory and size.
 *
 * @param context An initialized mPIPE context.
 * @param ring The NotifRing index.
 * @param mem A physically contiguous region of memory to be filled
 * with a ring of ::gxio_mpipe_idesc_t structures.
 * @param mem_size Number of bytes in the ring.  Must be 128, 512,
 * 2048, or 65536 * sizeof(gxio_mpipe_idesc_t).
 * @param mem_flags ::gxio_mpipe_mem_flags_e memory flags.
 *
 * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_NOTIF_RING or
 * ::GXIO_ERR_INVAL_MEMORY_SIZE on failure.
 */
extern int gxio_mpipe_init_notif_ring(gxio_mpipe_context_t *context,
				      unsigned int ring,
				      void *mem, size_t mem_size,
				      unsigned int mem_flags);

/* Configure an interrupt to be sent to a tile on incoming NotifRing
 *  traffic.  Once an interrupt is sent for a particular ring, no more
 *  will be sent until gxio_mica_enable_notif_ring_interrupt() is called.
 *
 * @param context An initialized mPIPE context.
 * @param x X coordinate of interrupt target tile.
 * @param y Y coordinate of interrupt target tile.
 * @param i Index of the IPI register which will receive the interrupt.
 * @param e Specific event which will be set in the target IPI register when
 * the interrupt occurs.
 * @param ring The NotifRing index.
 * @return Zero on success, GXIO_ERR_INVAL if params are out of range.
 */
extern int gxio_mpipe_request_notif_ring_interrupt(gxio_mpipe_context_t
						   *context, int x, int y,
						   int i, int e,
						   unsigned int ring);

/* Enable an interrupt on incoming NotifRing traffic.
 *
 * @param context An initialized mPIPE context.
 * @param ring The NotifRing index.
 * @return Zero on success, GXIO_ERR_INVAL if params are out of range.
 */
extern int gxio_mpipe_enable_notif_ring_interrupt(gxio_mpipe_context_t
						  *context, unsigned int ring);

/* Map all of a client's memory via the given IOTLB.
 * @param context An initialized mPIPE context.
 * @param iotlb IOTLB index.
 * @param pte Page table entry.
 * @param flags Flags.
 * @return Zero on success, or a negative error code.
 */
extern int gxio_mpipe_register_client_memory(gxio_mpipe_context_t *context,
					     unsigned int iotlb, HV_PTE pte,
					     unsigned int flags);

/*****************************************************************
 *                        Notif Groups                            *
 ******************************************************************/

/* Allocate a set of NotifGroups.
 *
 * The return value is NOT interesting if count is zero.
 *
 * @param context An initialized mPIPE context.
 * @param count Number of NotifGroups required.
 * @param first Index of first NotifGroup if ::GXIO_MPIPE_ALLOC_FIXED flag
 *   is set, otherwise ignored.
 * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
 * @return Index of first allocated buffer NotifGroup, or
 * ::GXIO_MPIPE_ERR_NO_NOTIF_GROUP if allocation failed.
 */
extern int gxio_mpipe_alloc_notif_groups(gxio_mpipe_context_t *context,
					 unsigned int count,
					 unsigned int first,
					 unsigned int flags);

/* Add a NotifRing to a NotifGroup.  This only sets a bit in the
 * application's 'group' object; the hardware NotifGroup can be
 * initialized by passing 'group' to gxio_mpipe_init_notif_group() or
 * gxio_mpipe_init_notif_group_and_buckets().
 */
static inline void
gxio_mpipe_notif_group_add_ring(gxio_mpipe_notif_group_bits_t *bits, int ring)
{
	bits->ring_mask[ring / 64] |= (1ull << (ring % 64));
}

/* Set a particular NotifGroup bitmask.  Since the load balancer
 * makes decisions based on both bucket and NotifGroup state, most
 * applications should use gxio_mpipe_init_notif_group_and_buckets()
 * rather than using this function to configure just a NotifGroup.
 */
extern int gxio_mpipe_init_notif_group(gxio_mpipe_context_t *context,
				       unsigned int group,
				       gxio_mpipe_notif_group_bits_t bits);

/*****************************************************************
 *                         Load Balancer                          *
 ******************************************************************/

/* Allocate a set of load balancer buckets.
 *
 * The return value is NOT interesting if count is zero.
 *
 * Note that buckets are allocated in chunks, so allocating one at
 * a time is much less efficient than allocating several at once.
 *
 * Note that the buckets are actually divided into two sub-ranges, of
 * different sizes, and different chunk sizes, and the range you get
 * by default is determined by the size of the request.  Allocations
 * cannot span the two sub-ranges.
 *
 * @param context An initialized mPIPE context.
 * @param count Number of buckets required.
 * @param first Index of first bucket if ::GXIO_MPIPE_ALLOC_FIXED flag is set,
 *   otherwise ignored.
 * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
 * @return Index of first allocated buffer bucket, or
 * ::GXIO_MPIPE_ERR_NO_BUCKET if allocation failed.
 */
extern int gxio_mpipe_alloc_buckets(gxio_mpipe_context_t *context,
				    unsigned int count, unsigned int first,
				    unsigned int flags);

/* The legal modes for gxio_mpipe_bucket_info_t and
 * gxio_mpipe_init_notif_group_and_buckets().
 *
 * All modes except ::GXIO_MPIPE_BUCKET_ROUND_ROBIN expect that the user
 * will allocate a power-of-two number of buckets and initialize them
 * to the same mode.  The classifier program then uses the appropriate
 * number of low bits from the incoming packet's flow hash to choose a
 * load balancer bucket.  Based on that bucket's load balancing mode,
 * reference count, and currently active NotifRing, the load balancer
 * chooses the NotifRing to which the packet will be delivered.
 */
typedef enum {
	/* All packets for a bucket go to the same NotifRing unless the
	 * NotifRing gets full, in which case packets will be dropped.  If
	 * the bucket reference count ever reaches zero, a new NotifRing may
	 * be chosen.
	 */
	GXIO_MPIPE_BUCKET_DYNAMIC_FLOW_AFFINITY =
		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_DFA,

	/* All packets for a bucket always go to the same NotifRing.
	 */
	GXIO_MPIPE_BUCKET_STATIC_FLOW_AFFINITY =
		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_FIXED,

	/* All packets for a bucket go to the least full NotifRing in the
	 * group, providing load balancing round robin behavior.
	 */
	GXIO_MPIPE_BUCKET_ROUND_ROBIN =
		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_ALWAYS_PICK,

	/* All packets for a bucket go to the same NotifRing unless the
	 * NotifRing gets full, at which point the bucket starts using the
	 * least full NotifRing in the group.  If all NotifRings in the
	 * group are full, packets will be dropped.
	 */
	GXIO_MPIPE_BUCKET_STICKY_FLOW_LOCALITY =
		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_STICKY,

	/* All packets for a bucket go to the same NotifRing unless the
	 * NotifRing gets full, or a random timer fires, at which point the
	 * bucket starts using the least full NotifRing in the group.  If
	 * all NotifRings in the group are full, packets will be dropped.
	 * WARNING: This mode is BROKEN on chips with fewer than 64 tiles.
	 */
	GXIO_MPIPE_BUCKET_PREFER_FLOW_LOCALITY =
		MPIPE_LBL_INIT_DAT_BSTS_TBL__MODE_VAL_STICKY_RAND,

} gxio_mpipe_bucket_mode_t;

/* Copy a set of bucket initialization values into the mPIPE
 * hardware.  Since the load balancer makes decisions based on both
 * bucket and NotifGroup state, most applications should use
 * gxio_mpipe_init_notif_group_and_buckets() rather than using this
 * function to configure a single bucket.
 *
 * @param context An initialized mPIPE context.
 * @param bucket Bucket index to be initialized.
 * @param bucket_info Initial reference count, NotifRing index, and mode.
 * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_BUCKET on failure.
 */
extern int gxio_mpipe_init_bucket(gxio_mpipe_context_t *context,
				  unsigned int bucket,
				  gxio_mpipe_bucket_info_t bucket_info);

/* Initializes a group and range of buckets and range of rings such
 * that the load balancer runs a particular load balancing function.
 *
 * First, the group is initialized with the given rings.
 *
 * Second, each bucket is initialized with the mode and group, and a
 * ring chosen round-robin from the given rings.
 *
 * Normally, the classifier picks a bucket, and then the load balancer
 * picks a ring, based on the bucket's mode, group, and current ring,
 * possibly updating the bucket's ring.
 *
 * @param context An initialized mPIPE context.
 * @param group The group.
 * @param ring The first ring.
 * @param num_rings The number of rings.
 * @param bucket The first bucket.
 * @param num_buckets The number of buckets.
 * @param mode The load balancing mode.
 *
 * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_BUCKET,
 * ::GXIO_MPIPE_ERR_BAD_NOTIF_GROUP, or
 * ::GXIO_MPIPE_ERR_BAD_NOTIF_RING on failure.
 */
extern int gxio_mpipe_init_notif_group_and_buckets(gxio_mpipe_context_t
						   *context,
						   unsigned int group,
						   unsigned int ring,
						   unsigned int num_rings,
						   unsigned int bucket,
						   unsigned int num_buckets,
						   gxio_mpipe_bucket_mode_t
						   mode);

/* Return credits to a NotifRing and/or bucket.
 *
 * @param context An initialized mPIPE context.
 * @param ring The NotifRing index, or -1.
 * @param bucket The bucket, or -1.
 * @param count The number of credits to return.
 */
static inline void gxio_mpipe_credit(gxio_mpipe_context_t *context,
				     int ring, int bucket, unsigned int count)
{
	/* NOTE: Fancy struct initialization would break "C89" header test. */

	MPIPE_IDMA_RELEASE_REGION_ADDR_t offset = { {0} };
	MPIPE_IDMA_RELEASE_REGION_VAL_t val = { {0} };

	/*
	 * The mmio_fast_base region starts at the IDMA region, so subtract
	 * off that initial offset.
	 */
	offset.region =
		MPIPE_MMIO_ADDR__REGION_VAL_IDMA -
		MPIPE_MMIO_ADDR__REGION_VAL_IDMA;
	offset.ring = ring;
	offset.bucket = bucket;
	offset.ring_enable = (ring >= 0);
	offset.bucket_enable = (bucket >= 0);
	val.count = count;

	__gxio_mmio_write(context->mmio_fast_base + offset.word, val.word);
}

/*****************************************************************
 *                         Egress Rings                           *
 ******************************************************************/

/* Allocate a set of eDMA rings.
 *
 * The return value is NOT interesting if count is zero.
 *
 * @param context An initialized mPIPE context.
 * @param count Number of eDMA rings required.
 * @param first Index of first eDMA ring if ::GXIO_MPIPE_ALLOC_FIXED flag
 *   is set, otherwise ignored.
 * @param flags Flag bits from ::gxio_mpipe_alloc_flags_e.
 * @return Index of first allocated buffer eDMA ring, or
 * ::GXIO_MPIPE_ERR_NO_EDMA_RING if allocation failed.
 */
extern int gxio_mpipe_alloc_edma_rings(gxio_mpipe_context_t *context,
				       unsigned int count, unsigned int first,
				       unsigned int flags);

/* Initialize an eDMA ring, using the given memory and size.
 *
 * @param context An initialized mPIPE context.
 * @param ring The eDMA ring index.
 * @param channel The channel to use.  This must be one of the channels
 * associated with the context's set of open links.
 * @param mem A physically contiguous region of memory to be filled
 * with a ring of ::gxio_mpipe_edesc_t structures.
 * @param mem_size Number of bytes in the ring.  Must be 512, 2048,
 * 8192 or 65536, times 16 (i.e. sizeof(gxio_mpipe_edesc_t)).
 * @param mem_flags ::gxio_mpipe_mem_flags_e memory flags.
 *
 * @return 0 on success, ::GXIO_MPIPE_ERR_BAD_EDMA_RING or
 * ::GXIO_ERR_INVAL_MEMORY_SIZE on failure.
 */
extern int gxio_mpipe_init_edma_ring(gxio_mpipe_context_t *context,
				     unsigned int ring, unsigned int channel,
				     void *mem, size_t mem_size,
				     unsigned int mem_flags);

/*****************************************************************
 *                      Classifier Program                        *
 ******************************************************************/

/*
 *
 * Functions for loading or configuring the mPIPE classifier program.
 *
 * The mPIPE classification processors all run a special "classifier"
 * program which, for each incoming packet, parses the packet headers,
 * encodes some packet metadata in the "idesc", and either drops the
 * packet, or picks a notif ring to handle the packet, and a buffer
 * stack to contain the packet, usually based on the channel, VLAN,
 * dMAC, flow hash, and packet size, under the guidance of the "rules"
 * API described below.
 *
 * @section gxio_mpipe_classifier_default Default Classifier
 *
 * The MDE provides a simple "default" classifier program.  It is
 * shipped as source in "$TILERA_ROOT/src/sys/mpipe/classifier.c",
 * which serves as its official documentation.  It is shipped as a
 * binary program in "$TILERA_ROOT/tile/boot/classifier", which is
 * automatically included in bootroms created by "tile-monitor", and
 * is automatically loaded by the hypervisor at boot time.
 *
 * The L2 analysis handles LLC packets, SNAP packets, and "VLAN
 * wrappers" (keeping the outer VLAN).
 *
 * The L3 analysis handles IPv4 and IPv6, dropping packets with bad
 * IPv4 header checksums, requesting computation of a TCP/UDP checksum
 * if appropriate, and hashing the dest and src IP addresses, plus the
 * ports for TCP/UDP packets, into the flow hash.  No special analysis
 * is done for "fragmented" packets or "tunneling" protocols.  Thus,
 * the first fragment of a fragmented TCP/UDP packet is hashed using
 * src/dest IP address and ports and all subsequent fragments are only
 * hashed according to src/dest IP address.
 *
 * The L3 analysis handles other packets too, hashing the dMAC
 * smac into a flow hash.
 *
 * The channel, VLAN, and dMAC used to pick a "rule" (see the
 * "rules" APIs below), which in turn is used to pick a buffer stack
 * (based on the packet size) and a bucket (based on the flow hash).
 *
 * To receive traffic matching a particular (channel/VLAN/dMAC
 * pattern, an application should allocate its own buffer stacks and
 * load balancer buckets, and map traffic to those stacks and buckets,
 * as decribed by the "rules" API below.
 *
 * Various packet metadata is encoded in the idesc.  The flow hash is
 * four bytes at 0x0C.  The VLAN is two bytes at 0x10.  The ethtype is
 * two bytes at 0x12.  The l3 start is one byte at 0x14.  The l4 start
 * is one byte at 0x15 for IPv4 and IPv6 packets, and otherwise zero.
 * The protocol is one byte at 0x16 for IPv4 and IPv6 packets, and
 * otherwise zero.
 *
 * @section gxio_mpipe_classifier_custom Custom Classifiers.
 *
 * A custom classifier may be created using "tile-mpipe-cc" with a
 * customized version of the default classifier sources.
 *
 * The custom classifier may be included in bootroms using the
 * "--classifier" option to "tile-monitor", or loaded dynamically
 * using gxio_mpipe_classifier_load_from_file().
 *
 * Be aware that "extreme" customizations may break the assumptions of
 * the "rules" APIs described below, but simple customizations, such
 * as adding new packet metadata, should be fine.
 */

/* A set of classifier rules, plus a context. */
typedef struct {

	/* The context. */
	gxio_mpipe_context_t *context;

	/* The actual rules. */
	gxio_mpipe_rules_list_t list;

} gxio_mpipe_rules_t;

/* Initialize a classifier program rules list.
 *
 * This function can be called on a previously initialized rules list
 * to discard any previously added rules.
 *
 * @param rules Rules list to initialize.
 * @param context An initialized mPIPE context.
 */
extern void gxio_mpipe_rules_init(gxio_mpipe_rules_t *rules,
				  gxio_mpipe_context_t *context);

/* Begin a new rule on the indicated rules list.
 *
 * Note that an empty rule matches all packets, but an empty rule list
 * matches no packets.
 *
 * @param rules Rules list to which new rule is appended.
 * @param bucket First load balancer bucket to which packets will be
 * delivered.
 * @param num_buckets Number of buckets (must be a power of two) across
 * which packets will be distributed based on the "flow hash".
 * @param stacks Either NULL, to assign each packet to the smallest
 * initialized buffer stack which does not induce chaining (and to
 * drop packets which exceed the largest initialized buffer stack
 * buffer size), or an array, with each entry indicating which buffer
 * stack should be used for packets up to that size (with 255
 * indicating that those packets should be dropped).
 * @return 0 on success, or a negative error code on failure.
 */
extern int gxio_mpipe_rules_begin(gxio_mpipe_rules_t *rules,
				  unsigned int bucket,
				  unsigned int num_buckets,
				  gxio_mpipe_rules_stacks_t *stacks);

/* Set the headroom of the current rule.
 *
 * @param rules Rules list whose current rule will be modified.
 * @param headroom The headroom.
 * @return 0 on success, or a negative error code on failure.
 */
extern int gxio_mpipe_rules_set_headroom(gxio_mpipe_rules_t *rules,
					 uint8_t headroom);

/* Indicate that packets from a particular channel can be delivered
 * to the buckets and buffer stacks associated with the current rule.
 *
 * Channels added must be associated with links opened by the mPIPE context
 * used in gxio_mpipe_rules_init().  A rule with no channels is equivalent
 * to a rule naming all such associated channels.
 *
 * @param rules Rules list whose current rule will be modified.
 * @param channel The channel to add.
 * @return 0 on success, or a negative error code on failure.
 */
extern int gxio_mpipe_rules_add_channel(gxio_mpipe_rules_t *rules,
					unsigned int channel);

/* Commit rules.
 *
 * The rules are sent to the hypervisor, where they are combined with
 * the rules from other apps, and used to program the hardware classifier.
 *
 * Note that if this function returns an error, then the rules will NOT
 * have been committed, even if the error is due to interactions with
 * rules from another app.
 *
 * @param rules Rules list to commit.
 * @return 0 on success, or a negative error code on failure.
 */
extern int gxio_mpipe_rules_commit(gxio_mpipe_rules_t *rules);

/*****************************************************************
 *                     Ingress Queue Wrapper                      *
 ******************************************************************/

/*
 *
 * Convenience functions for receiving packets from a NotifRing and
 * sending packets via an eDMA ring.
 *
 * The mpipe ingress and egress hardware uses shared memory packet
 * descriptors to describe packets that have arrived on ingress or
 * are destined for egress.  These descriptors are stored in shared
 * memory ring buffers and written or read by hardware as necessary.
 * The gxio library provides wrapper functions that manage the head and
 * tail pointers for these rings, allowing the user to easily read or
 * write packet descriptors.
 *
 * The initialization interface for ingress and egress rings is quite
 * similar.  For example, to create an ingress queue, the user passes
 * a ::gxio_mpipe_iqueue_t state object, a ring number from
 * gxio_mpipe_alloc_notif_rings(), and the address of memory to hold a
 * ring buffer to the gxio_mpipe_iqueue_init() function.  The function
 * returns success when the state object has been initialized and the
 * hardware configured to deliver packets to the specified ring
 * buffer.  Similarly, gxio_mpipe_equeue_init() takes a
 * ::gxio_mpipe_equeue_t state object, a ring number from
 * gxio_mpipe_alloc_edma_rings(), and a shared memory buffer.
 *
 * @section gxio_mpipe_iqueue Working with Ingress Queues
 *
 * Once initialized, the gxio_mpipe_iqueue_t API provides two flows
 * for getting the ::gxio_mpipe_idesc_t packet descriptor associated
 * with incoming packets.  The simplest is to call
 * gxio_mpipe_iqueue_get() or gxio_mpipe_iqueue_try_get().  These
 * functions copy the oldest packet descriptor out of the NotifRing and
 * into a descriptor provided by the caller.  They also immediately
 * inform the hardware that a descriptor has been processed.
 *
 * For applications with stringent performance requirements, higher
 * efficiency can be achieved by avoiding the packet descriptor copy
 * and processing multiple descriptors at once.  The
 * gxio_mpipe_iqueue_peek() and gxio_mpipe_iqueue_try_peek() functions
 * allow such optimizations.  These functions provide a pointer to the
 * next valid ingress descriptor in the NotifRing's shared memory ring
 * buffer, and a count of how many contiguous descriptors are ready to
 * be processed.  The application can then process any number of those
 * descriptors in place, calling gxio_mpipe_iqueue_consume() to inform
 * the hardware after each one has been processed.
 *
 * @section gxio_mpipe_equeue Working with Egress Queues
 *
 * Similarly, the egress queue API provides a high-performance
 * interface plus a simple wrapper for use in posting
 * ::gxio_mpipe_edesc_t egress packet descriptors.  The simple
 * version, gxio_mpipe_equeue_put(), allows the programmer to wait for
 * an eDMA ring slot to become available and write a single descriptor
 * into the ring.
 *
 * Alternatively, you can reserve slots in the eDMA ring using
 * gxio_mpipe_equeue_reserve() or gxio_mpipe_equeue_try_reserve(), and
 * then fill in each slot using gxio_mpipe_equeue_put_at().  This
 * capability can be used to amortize the cost of reserving slots
 * across several packets.  It also allows gather operations to be
 * performed on a shared equeue, by ensuring that the edescs for all
 * the fragments are all contiguous in the eDMA ring.
 *
 * The gxio_mpipe_equeue_reserve() and gxio_mpipe_equeue_try_reserve()
 * functions return a 63-bit "completion slot", which is actually a
 * sequence number, the low bits of which indicate the ring buffer
 * index and the high bits the number of times the application has
 * gone around the egress ring buffer.  The extra bits allow an
 * application to check for egress completion by calling
 * gxio_mpipe_equeue_is_complete() to see whether a particular 'slot'
 * number has finished.  Given the maximum packet rates of the Gx
 * processor, the 63-bit slot number will never wrap.
 *
 * In practice, most applications use the ::gxio_mpipe_edesc_t::hwb
 * bit to indicate that the buffers containing egress packet data
 * should be pushed onto a buffer stack when egress is complete.  Such
 * applications generally do not need to know when an egress operation
 * completes (since there is no need to free a buffer post-egress),
 * and thus can use the optimized gxio_mpipe_equeue_reserve_fast() or
 * gxio_mpipe_equeue_try_reserve_fast() functions, which return a 24
 * bit "slot", instead of a 63-bit "completion slot".
 *
 * Once a slot has been "reserved", it MUST be filled.  If the
 * application reserves a slot and then decides that it does not
 * actually need it, it can set the ::gxio_mpipe_edesc_t::ns (no send)
 * bit on the descriptor passed to gxio_mpipe_equeue_put_at() to
 * indicate that no data should be sent.  This technique can also be
 * used to drop an incoming packet, instead of forwarding it, since
 * any buffer will still be pushed onto the buffer stack when the
 * egress descriptor is processed.
 */

/* A convenient interface to a NotifRing, for use by a single thread.
 */
typedef struct {

	/* The context. */
	gxio_mpipe_context_t *context;

	/* The actual NotifRing. */
	gxio_mpipe_idesc_t *idescs;

	/* The number of entries. */
	unsigned long num_entries;

	/* The number of entries minus one. */
	unsigned long mask_num_entries;

	/* The log2() of the number of entries. */
	unsigned long log2_num_entries;

	/* The next entry. */
	unsigned int head;

	/* The NotifRing id. */
	unsigned int ring;

#ifdef __BIG_ENDIAN__
	/* The number of byteswapped entries. */
	unsigned int swapped;
#endif

} gxio_mpipe_iqueue_t;

/* Initialize an "iqueue".
 *
 * Takes the iqueue plus the same args as gxio_mpipe_init_notif_ring().
 */
extern int gxio_mpipe_iqueue_init(gxio_mpipe_iqueue_t *iqueue,
				  gxio_mpipe_context_t *context,
				  unsigned int ring,
				  void *mem, size_t mem_size,
				  unsigned int mem_flags);

/* Advance over some old entries in an iqueue.
 *
 * Please see the documentation for gxio_mpipe_iqueue_consume().
 *
 * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
 * @param count The number of entries to advance over.
 */
static inline void gxio_mpipe_iqueue_advance(gxio_mpipe_iqueue_t *iqueue,
					     int count)
{
	/* Advance with proper wrap. */
	int head = iqueue->head + count;
	iqueue->head =
		(head & iqueue->mask_num_entries) +
		(head >> iqueue->log2_num_entries);

#ifdef __BIG_ENDIAN__
	/* HACK: Track swapped entries. */
	iqueue->swapped -= count;
#endif
}

/* Release the ring and bucket for an old entry in an iqueue.
 *
 * Releasing the ring allows more packets to be delivered to the ring.
 *
 * Releasing the bucket allows flows using the bucket to be moved to a
 * new ring when using GXIO_MPIPE_BUCKET_DYNAMIC_FLOW_AFFINITY.
 *
 * This function is shorthand for "gxio_mpipe_credit(iqueue->context,
 * iqueue->ring, idesc->bucket_id, 1)", and it may be more convenient
 * to make that underlying call, using those values, instead of
 * tracking the entire "idesc".
 *
 * If packet processing is deferred, optimal performance requires that
 * the releasing be deferred as well.
 *
 * Please see the documentation for gxio_mpipe_iqueue_consume().
 *
 * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
 * @param idesc The descriptor which was processed.
 */
static inline void gxio_mpipe_iqueue_release(gxio_mpipe_iqueue_t *iqueue,
					     gxio_mpipe_idesc_t *idesc)
{
	gxio_mpipe_credit(iqueue->context, iqueue->ring, idesc->bucket_id, 1);
}

/* Consume a packet from an "iqueue".
 *
 * After processing packets peeked at via gxio_mpipe_iqueue_peek()
 * or gxio_mpipe_iqueue_try_peek(), you must call this function, or
 * gxio_mpipe_iqueue_advance() plus gxio_mpipe_iqueue_release(), to
 * advance over those entries, and release their rings and buckets.
 *
 * You may call this function as each packet is processed, or you can
 * wait until several packets have been processed.
 *
 * Note that if you are using a single bucket, and you are handling
 * batches of N packets, then you can replace several calls to this
 * function with calls to "gxio_mpipe_iqueue_advance(iqueue, N)" and
 * "gxio_mpipe_credit(iqueue->context, iqueue->ring, bucket, N)".
 *
 * Note that if your classifier sets "idesc->nr", then you should
 * explicitly call "gxio_mpipe_iqueue_advance(iqueue, idesc)" plus
 * "gxio_mpipe_credit(iqueue->context, iqueue->ring, -1, 1)", to
 * avoid incorrectly crediting the (unused) bucket.
 *
 * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
 * @param idesc The descriptor which was processed.
 */
static inline void gxio_mpipe_iqueue_consume(gxio_mpipe_iqueue_t *iqueue,
					     gxio_mpipe_idesc_t *idesc)
{
	gxio_mpipe_iqueue_advance(iqueue, 1);
	gxio_mpipe_iqueue_release(iqueue, idesc);
}

/* Peek at the next packet(s) in an "iqueue", without waiting.
 *
 * If no packets are available, fills idesc_ref with NULL, and then
 * returns ::GXIO_MPIPE_ERR_IQUEUE_EMPTY.  Otherwise, fills idesc_ref
 * with the address of the next valid packet descriptor, and returns
 * the maximum number of valid descriptors which can be processed.
 * You may process fewer descriptors if desired.
 *
 * Call gxio_mpipe_iqueue_consume() on each packet once it has been
 * processed (or dropped), to allow more packets to be delivered.
 *
 * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
 * @param idesc_ref A pointer to a packet descriptor pointer.
 * @return The (positive) number of packets which can be processed,
 * or ::GXIO_MPIPE_ERR_IQUEUE_EMPTY if no packets are available.
 */
static inline int gxio_mpipe_iqueue_try_peek(gxio_mpipe_iqueue_t *iqueue,
					     gxio_mpipe_idesc_t **idesc_ref)
{
	gxio_mpipe_idesc_t *next;

	uint64_t head = iqueue->head;
	uint64_t tail = __gxio_mmio_read(iqueue->idescs);

	/* Available entries. */
	uint64_t avail =
		(tail >= head) ? (tail - head) : (iqueue->num_entries - head);

	if (avail == 0) {
		*idesc_ref = NULL;
		return GXIO_MPIPE_ERR_IQUEUE_EMPTY;
	}

	next = &iqueue->idescs[head];

	/* ISSUE: Is this helpful? */
	__insn_prefetch(next);

#ifdef __BIG_ENDIAN__
	/* HACK: Swap new entries directly in memory. */
	{
		int i, j;
		for (i = iqueue->swapped; i < avail; i++) {
			for (j = 0; j < 8; j++)
				next[i].words[j] =
					__builtin_bswap64(next[i].words[j]);
		}
		iqueue->swapped = avail;
	}
#endif

	*idesc_ref = next;

	return avail;
}

/* Drop a packet by pushing its buffer (if appropriate).
 *
 * NOTE: The caller must still call gxio_mpipe_iqueue_consume() if idesc
 * came from gxio_mpipe_iqueue_try_peek() or gxio_mpipe_iqueue_peek().
 *
 * @param iqueue An ingress queue initialized via gxio_mpipe_iqueue_init().
 * @param idesc A packet descriptor.
 */
static inline void gxio_mpipe_iqueue_drop(gxio_mpipe_iqueue_t *iqueue,
					  gxio_mpipe_idesc_t *idesc)
{
	/* FIXME: Handle "chaining" properly. */

	if (!idesc->be) {
		unsigned char *va = gxio_mpipe_idesc_get_va(idesc);
		gxio_mpipe_push_buffer(iqueue->context, idesc->stack_idx, va);
	}
}

/*****************************************************************
 *                      Egress Queue Wrapper                      *
 ******************************************************************/

/* A convenient, thread-safe interface to an eDMA ring. */
typedef struct {

	/* State object for tracking head and tail pointers. */
	__gxio_dma_queue_t dma_queue;

	/* The ring entries. */
	gxio_mpipe_edesc_t *edescs;

	/* The number of entries minus one. */
	unsigned long mask_num_entries;

	/* The log2() of the number of entries. */
	unsigned long log2_num_entries;

} gxio_mpipe_equeue_t;

/* Initialize an "equeue".
 *
 * Takes the equeue plus the same args as gxio_mpipe_init_edma_ring().
 */
extern int gxio_mpipe_equeue_init(gxio_mpipe_equeue_t *equeue,
				  gxio_mpipe_context_t *context,
				  unsigned int edma_ring_id,
				  unsigned int channel,
				  void *mem, unsigned int mem_size,
				  unsigned int mem_flags);

/* Reserve completion slots for edescs.
 *
 * Use gxio_mpipe_equeue_put_at() to actually populate the slots.
 *
 * This function is slower than gxio_mpipe_equeue_reserve_fast(), but
 * returns a full 64 bit completion slot, which can be used with
 * gxio_mpipe_equeue_is_complete().
 *
 * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
 * @param num Number of slots to reserve (must be non-zero).
 * @return The first reserved completion slot, or a negative error code.
 */
static inline int64_t gxio_mpipe_equeue_reserve(gxio_mpipe_equeue_t *equeue,
						unsigned int num)
{
	return __gxio_dma_queue_reserve_aux(&equeue->dma_queue, num, true);
}

/* Reserve completion slots for edescs, if possible.
 *
 * Use gxio_mpipe_equeue_put_at() to actually populate the slots.
 *
 * This function is slower than gxio_mpipe_equeue_try_reserve_fast(),
 * but returns a full 64 bit completion slot, which can be used with
 * gxio_mpipe_equeue_is_complete().
 *
 * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
 * @param num Number of slots to reserve (must be non-zero).
 * @return The first reserved completion slot, or a negative error code.
 */
static inline int64_t gxio_mpipe_equeue_try_reserve(gxio_mpipe_equeue_t
						    *equeue, unsigned int num)
{
	return __gxio_dma_queue_reserve_aux(&equeue->dma_queue, num, false);
}

/* Reserve slots for edescs.
 *
 * Use gxio_mpipe_equeue_put_at() to actually populate the slots.
 *
 * This function is faster than gxio_mpipe_equeue_reserve(), but
 * returns a 24 bit slot (instead of a 64 bit completion slot), which
 * thus cannot be used with gxio_mpipe_equeue_is_complete().
 *
 * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
 * @param num Number of slots to reserve (should be non-zero).
 * @return The first reserved slot, or a negative error code.
 */
static inline int64_t gxio_mpipe_equeue_reserve_fast(gxio_mpipe_equeue_t
						     *equeue, unsigned int num)
{
	return __gxio_dma_queue_reserve(&equeue->dma_queue, num, true, false);
}

/* Reserve slots for edescs, if possible.
 *
 * Use gxio_mpipe_equeue_put_at() to actually populate the slots.
 *
 * This function is faster than gxio_mpipe_equeue_try_reserve(), but
 * returns a 24 bit slot (instead of a 64 bit completion slot), which
 * thus cannot be used with gxio_mpipe_equeue_is_complete().
 *
 * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
 * @param num Number of slots to reserve (should be non-zero).
 * @return The first reserved slot, or a negative error code.
 */
static inline int64_t gxio_mpipe_equeue_try_reserve_fast(gxio_mpipe_equeue_t
							 *equeue,
							 unsigned int num)
{
	return __gxio_dma_queue_reserve(&equeue->dma_queue, num, false, false);
}

/*
 * HACK: This helper function tricks gcc 4.6 into avoiding saving
 * a copy of "edesc->words[0]" on the stack for no obvious reason.
 */

static inline void gxio_mpipe_equeue_put_at_aux(gxio_mpipe_equeue_t *equeue,
						uint_reg_t ew[2],
						unsigned long slot)
{
	unsigned long edma_slot = slot & equeue->mask_num_entries;
	gxio_mpipe_edesc_t *edesc_p = &equeue->edescs[edma_slot];

	/*
	 * ISSUE: Could set eDMA ring to be on generation 1 at start, which
	 * would avoid the negation here, perhaps allowing "__insn_bfins()".
	 */
	ew[0] |= !((slot >> equeue->log2_num_entries) & 1);

	/*
	 * NOTE: We use "__gxio_mpipe_write()", plus the fact that the eDMA
	 * queue alignment restrictions ensure that these two words are on
	 * the same cacheline, to force proper ordering between the stores.
	 */
	__gxio_mmio_write64(&edesc_p->words[1], ew[1]);
	__gxio_mmio_write64(&edesc_p->words[0], ew[0]);
}

/* Post an edesc to a given slot in an equeue.
 *
 * This function copies the supplied edesc into entry "slot mod N" in
 * the underlying ring, setting the "gen" bit to the appropriate value
 * based on "(slot mod N*2)", where "N" is the size of the ring.  Note
 * that the higher bits of slot are unused, and thus, this function
 * can handle "slots" as well as "completion slots".
 *
 * Normally this function is used to fill in slots reserved by
 * gxio_mpipe_equeue_try_reserve(), gxio_mpipe_equeue_reserve(),
 * gxio_mpipe_equeue_try_reserve_fast(), or
 * gxio_mpipe_equeue_reserve_fast(),
 *
 * This function can also be used without "reserving" slots, if the
 * application KNOWS that the ring can never overflow, for example, by
 * pushing fewer buffers into the buffer stacks than there are total
 * slots in the equeue, but this is NOT recommended.
 *
 * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
 * @param edesc The egress descriptor to be posted.
 * @param slot An egress slot (only the low bits are actually used).
 */
static inline void gxio_mpipe_equeue_put_at(gxio_mpipe_equeue_t *equeue,
					    gxio_mpipe_edesc_t edesc,
					    unsigned long slot)
{
	gxio_mpipe_equeue_put_at_aux(equeue, edesc.words, slot);
}

/* Post an edesc to the next slot in an equeue.
 *
 * This is a convenience wrapper around
 * gxio_mpipe_equeue_reserve_fast() and gxio_mpipe_equeue_put_at().
 *
 * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
 * @param edesc The egress descriptor to be posted.
 * @return 0 on success.
 */
static inline int gxio_mpipe_equeue_put(gxio_mpipe_equeue_t *equeue,
					gxio_mpipe_edesc_t edesc)
{
	int64_t slot = gxio_mpipe_equeue_reserve_fast(equeue, 1);
	if (slot < 0)
		return (int)slot;

	gxio_mpipe_equeue_put_at(equeue, edesc, slot);

	return 0;
}

/* Ask the mPIPE hardware to egress outstanding packets immediately.
 *
 * This call is not necessary, but may slightly reduce overall latency.
 *
 * Technically, you should flush all gxio_mpipe_equeue_put_at() writes
 * to memory before calling this function, to ensure the descriptors
 * are visible in memory before the mPIPE hardware actually looks for
 * them.  But this should be very rare, and the only side effect would
 * be increased latency, so it is up to the caller to decide whether
 * or not to flush memory.
 *
 * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
 */
static inline void gxio_mpipe_equeue_flush(gxio_mpipe_equeue_t *equeue)
{
	/* Use "ring_idx = 0" and "count = 0" to "wake up" the eDMA ring. */
	MPIPE_EDMA_POST_REGION_VAL_t val = { {0} };
	/* Flush the write buffers. */
	__insn_flushwb();
	__gxio_mmio_write(equeue->dma_queue.post_region_addr, val.word);
}

/* Determine if a given edesc has been completed.
 *
 * Note that this function requires a "completion slot", and thus may
 * NOT be used with a "slot" from gxio_mpipe_equeue_reserve_fast() or
 * gxio_mpipe_equeue_try_reserve_fast().
 *
 * @param equeue An egress queue initialized via gxio_mpipe_equeue_init().
 * @param completion_slot The completion slot used by the edesc.
 * @param update If true, and the desc does not appear to have completed
 * yet, then update any software cache of the hardware completion counter,
 * and check again.  This should normally be true.
 * @return True iff the given edesc has been completed.
 */
static inline int gxio_mpipe_equeue_is_complete(gxio_mpipe_equeue_t *equeue,
						int64_t completion_slot,
						int update)
{
	return __gxio_dma_queue_is_complete(&equeue->dma_queue,
					    completion_slot, update);
}

/*****************************************************************
 *                        Link Management                         *
 ******************************************************************/

/*
 *
 * Functions for manipulating and sensing the state and configuration
 * of physical network links.
 *
 * @section gxio_mpipe_link_perm Link Permissions
 *
 * Opening a link (with gxio_mpipe_link_open()) requests a set of link
 * permissions, which control what may be done with the link, and potentially
 * what permissions may be granted to other processes.
 *
 * Data permission allows the process to receive packets from the link by
 * specifying the link's channel number in mPIPE packet distribution rules,
 * and to send packets to the link by using the link's channel number as
 * the target for an eDMA ring.
 *
 * Stats permission allows the process to retrieve link attributes (such as
 * the speeds it is capable of running at, or whether it is currently up), and
 * to read and write certain statistics-related registers in the link's MAC.
 *
 * Control permission allows the process to retrieve and modify link attributes
 * (so that it may, for example, bring the link up and take it down), and
 * read and write many registers in the link's MAC and PHY.
 *
 * Any permission may be requested as shared, which allows other processes
 * to also request shared permission, or exclusive, which prevents other
 * processes from requesting it.  In keeping with GXIO's typical usage in
 * an embedded environment, the defaults for all permissions are shared.
 *
 * Permissions are granted on a first-come, first-served basis, so if two
 * applications request an exclusive permission on the same link, the one
 * to run first will win.  Note, however, that some system components, like
 * the kernel Ethernet driver, may get an opportunity to open links before
 * any applications run.
 *
 * @section gxio_mpipe_link_names Link Names
 *
 * Link names are of the form gbe<em>number</em> (for Gigabit Ethernet),
 * xgbe<em>number</em> (for 10 Gigabit Ethernet), loop<em>number</em> (for
 * internal mPIPE loopback), or ilk<em>number</em>/<em>channel</em>
 * (for Interlaken links); for instance, gbe0, xgbe1, loop3, and
 * ilk0/12 are all possible link names.  The correspondence between
 * the link name and an mPIPE instance number or mPIPE channel number is
 * system-dependent; all links will not exist on all systems, and the set
 * of numbers used for a particular link type may not start at zero and may
 * not be contiguous.  Use gxio_mpipe_link_enumerate() to retrieve the set of
 * links which exist on a system, and always use gxio_mpipe_link_instance()
 * to determine which mPIPE controls a particular link.
 *
 * Note that in some cases, links may share hardware, such as PHYs, or
 * internal mPIPE buffers; in these cases, only one of the links may be
 * opened at a time.  This is especially common with xgbe and gbe ports,
 * since each xgbe port uses 4 SERDES lanes, each of which may also be
 * configured as one gbe port.
 *
 * @section gxio_mpipe_link_states Link States
 *
 * The mPIPE link management model revolves around three different states,
 * which are maintained for each link:
 *
 * 1. The <em>current</em> link state: is the link up now, and if so, at
 *    what speed?
 *
 * 2. The <em>desired</em> link state: what do we want the link state to be?
 *    The system is always working to make this state the current state;
 *    thus, if the desired state is up, and the link is down, we'll be
 *    constantly trying to bring it up, automatically.
 *
 * 3. The <em>possible</em> link state: what speeds are valid for this
 *    particular link?  Or, in other words, what are the capabilities of
 *    the link hardware?
 *
 * These link states are not, strictly speaking, related to application
 * state; they may be manipulated at any time, whether or not the link
 * is currently being used for data transfer.  However, for convenience,
 * gxio_mpipe_link_open() and gxio_mpipe_link_close() (or application exit)
 * can affect the link state.  These implicit link management operations
 * may be modified or disabled by the use of link open flags.
 *
 * From an application, you can use gxio_mpipe_link_get_attr()
 * and gxio_mpipe_link_set_attr() to manipulate the link states.
 * gxio_mpipe_link_get_attr() with ::GXIO_MPIPE_LINK_POSSIBLE_STATE
 * gets you the possible link state.  gxio_mpipe_link_get_attr() with
 * ::GXIO_MPIPE_LINK_CURRENT_STATE gets you the current link state.
 * Finally, gxio_mpipe_link_set_attr() and gxio_mpipe_link_get_attr()
 * with ::GXIO_MPIPE_LINK_DESIRED_STATE allow you to modify or retrieve
 * the desired link state.
 *
 * If you want to manage a link from a part of your application which isn't
 * involved in packet processing, you can use the ::GXIO_MPIPE_LINK_NO_DATA
 * flags on a gxio_mpipe_link_open() call.  This opens the link, but does
 * not request data permission, so it does not conflict with any exclusive
 * permissions which may be held by other processes.  You can then can use
 * gxio_mpipe_link_get_attr() and gxio_mpipe_link_set_attr() on this link
 * object to bring up or take down the link.
 *
 * Some links support link state bits which support various loopback
 * modes. ::GXIO_MPIPE_LINK_LOOP_MAC tests datapaths within the Tile
 * Processor itself; ::GXIO_MPIPE_LINK_LOOP_PHY tests the datapath between
 * the Tile Processor and the external physical layer interface chip; and
 * ::GXIO_MPIPE_LINK_LOOP_EXT tests the entire network datapath with the
 * aid of an external loopback connector.  In addition to enabling hardware
 * testing, such configuration can be useful for software testing, as well.
 *
 * When LOOP_MAC or LOOP_PHY is enabled, packets transmitted on a channel
 * will be received by that channel, instead of being emitted on the
 * physical link, and packets received on the physical link will be ignored.
 * Other than that, all standard GXIO operations work as you might expect.
 * Note that loopback operation requires that the link be brought up using
 * one or more of the GXIO_MPIPE_LINK_SPEED_xxx link state bits.
 *
 * Those familiar with previous versions of the MDE on TILEPro hardware
 * will notice significant similarities between the NetIO link management
 * model and the mPIPE link management model.  However, the NetIO model
 * was developed in stages, and some of its features -- for instance,
 * the default setting of certain flags -- were shaped by the need to be
 * compatible with previous versions of NetIO.  Since the features provided
 * by the mPIPE hardware and the mPIPE GXIO library are significantly
 * different than those provided by NetIO, in some cases, we have made
 * different choices in the mPIPE link management API.  Thus, please read
 * this documentation carefully before assuming that mPIPE link management
 * operations are exactly equivalent to their NetIO counterparts.
 */

/* An object used to manage mPIPE link state and resources. */
typedef struct {
	/* The overall mPIPE context. */
	gxio_mpipe_context_t *context;

	/* The channel number used by this link. */
	uint8_t channel;

	/* The MAC index used by this link. */
	uint8_t mac;
} gxio_mpipe_link_t;

/* Retrieve one of this system's legal link names, and its MAC address.
 *
 * @param index Link name index.  If a system supports N legal link names,
 *  then indices between 0 and N - 1, inclusive, each correspond to one of
 *  those names.  Thus, to retrieve all of a system's legal link names,
 *  call this function in a loop, starting with an index of zero, and
 *  incrementing it once per iteration until -1 is returned.
 * @param link_name Pointer to the buffer which will receive the retrieved
 *  link name.  The buffer should contain space for at least
 *  ::GXIO_MPIPE_LINK_NAME_LEN bytes; the returned name, including the
 *  terminating null byte, will be no longer than that.
 * @param link_name Pointer to the buffer which will receive the retrieved
 *  MAC address.  The buffer should contain space for at least 6 bytes.
 * @return Zero if a link name was successfully retrieved; -1 if one was
 *  not.
 */
extern int gxio_mpipe_link_enumerate_mac(int index, char *link_name,
					 uint8_t *mac_addr);

/* Open an mPIPE link.
 *
 *  A link must be opened before it may be used to send or receive packets,
 *  and before its state may be examined or changed.  Depending up on the
 *  link's intended use, one or more link permissions may be requested via
 *  the flags parameter; see @ref gxio_mpipe_link_perm.  In addition, flags
 *  may request that the link's state be modified at open time.  See @ref
 *  gxio_mpipe_link_states and @ref gxio_mpipe_link_open_flags for more detail.
 *
 * @param link A link state object, which will be initialized if this
 *  function completes successfully.
 * @param context An initialized mPIPE context.
 * @param link_name Name of the link.
 * @param flags Zero or more @ref gxio_mpipe_link_open_flags, ORed together.
 * @return 0 if the link was successfully opened, or a negative error code.
 *
 */
extern int gxio_mpipe_link_open(gxio_mpipe_link_t *link,
				gxio_mpipe_context_t *context,
				const char *link_name, unsigned int flags);

/* Close an mPIPE link.
 *
 *  Closing a link makes it available for use by other processes.  Once
 *  a link has been closed, packets may no longer be sent on or received
 *  from the link, and its state may not be examined or changed.
 *
 * @param link A link state object, which will no longer be initialized
 *  if this function completes successfully.
 * @return 0 if the link was successfully closed, or a negative error code.
 *
 */
extern int gxio_mpipe_link_close(gxio_mpipe_link_t *link);

/* Return a link's channel number.
 *
 * @param link A properly initialized link state object.
 * @return The channel number for the link.
 */
static inline int gxio_mpipe_link_channel(gxio_mpipe_link_t *link)
{
	return link->channel;
}

///////////////////////////////////////////////////////////////////
//                             Timestamp                         //
///////////////////////////////////////////////////////////////////

/* Get the timestamp of mPIPE when this routine is called.
 *
 * @param context An initialized mPIPE context.
 * @param ts A timespec structure to store the current clock.
 * @return If the call was successful, zero; otherwise, a negative error
 *  code.
 */
extern int gxio_mpipe_get_timestamp(gxio_mpipe_context_t *context,
				    struct timespec *ts);

/* Set the timestamp of mPIPE.
 *
 * @param context An initialized mPIPE context.
 * @param ts A timespec structure to store the requested clock.
 * @return If the call was successful, zero; otherwise, a negative error
 *  code.
 */
extern int gxio_mpipe_set_timestamp(gxio_mpipe_context_t *context,
				    const struct timespec *ts);

/* Adjust the timestamp of mPIPE.
 *
 * @param context An initialized mPIPE context.
 * @param delta A signed time offset to adjust, in nanoseconds.
 * The absolute value of this parameter must be less than or
 * equal to 1000000000.
 * @return If the call was successful, zero; otherwise, a negative error
 *  code.
 */
extern int gxio_mpipe_adjust_timestamp(gxio_mpipe_context_t *context,
				       int64_t delta);

#endif /* !_GXIO_MPIPE_H_ */