driver/fsl-mc: Add support of MC Flibs

Freescale's Layerscape Management Complex (MC) provide support various
objects like DPRC, DPNI, DPBP and DPIO.
Where:
	DPRC: Place holdes for other MC objectes like DPNI, DPBP, DPIO
	DPBP: Management of buffer pool
	DPIO: Used for used to QBMan portal
	DPNI: Represents standard network interface

These objects are used for DPAA ethernet drivers.

Signed-off-by: J. German Rivera <German.Rivera@freescale.com>
Signed-off-by: Lijun Pan <Lijun.Pan@freescale.com>
Signed-off-by: Stuart Yoder <stuart.yoder@freescale.com>
Signed-off-by: Geoff Thorpe <Geoff.Thorpe@freescale.com>
Signed-off-by: Haiying Wang <Haiying.Wang@freescale.com>
Signed-off-by: Cristian Sovaiala <cristian.sovaiala@freescale.com>
Signed-off-by: pankaj chauhan <pankaj.chauhan@freescale.com>
Signed-off-by: Prabhakar Kushwaha <prabhakar@freescale.com>
Reviewed-by: York Sun <yorksun@freescale.com>

6 files changed, 1320 insertions, 0 deletions

diff --git a/drivers/net/fsl-mc/dpio/Makefile b/drivers/net/fsl-mc/dpio/Makefile
new file mode 100644
index 00000000000..1ccefc0db2a
--- /dev/null
+++ b/drivers/net/fsl-mc/dpio/Makefile
@@ -0,0 +1,9 @@
+#
+# Copyright 2014 Freescale Semiconductor, Inc.
+#
+# SPDX-License-Identifier:	GPL-2.0+
+#
+
+# Layerscape MC DPIO driver
+obj-y += dpio.o	\
+	qbman_portal.o
diff --git a/drivers/net/fsl-mc/dpio/dpio.c b/drivers/net/fsl-mc/dpio/dpio.c
new file mode 100644
index 00000000000..b07eff7e3ca
--- /dev/null
+++ b/drivers/net/fsl-mc/dpio/dpio.c
@@ -0,0 +1,102 @@
+/*
+ * Copyright (C) 2013-2015 Freescale Semiconductor
+ *
+ * SPDX-License-Identifier:	GPL-2.0+
+ */
+
+#include <fsl-mc/fsl_mc_sys.h>
+#include <fsl-mc/fsl_mc_cmd.h>
+#include <fsl-mc/fsl_dpio.h>
+
+int dpio_open(struct fsl_mc_io *mc_io, int dpio_id, uint16_t *token)
+{
+	struct mc_command cmd = { 0 };
+	int err;
+
+	/* prepare command */
+	cmd.header = mc_encode_cmd_header(DPIO_CMDID_OPEN,
+					  MC_CMD_PRI_LOW, 0);
+	DPIO_CMD_OPEN(cmd, dpio_id);
+
+	/* send command to mc*/
+	err = mc_send_command(mc_io, &cmd);
+	if (err)
+		return err;
+
+	/* retrieve response parameters */
+	*token = MC_CMD_HDR_READ_TOKEN(cmd.header);
+
+	return 0;
+}
+
+int dpio_close(struct fsl_mc_io *mc_io, uint16_t token)
+{
+	struct mc_command cmd = { 0 };
+
+	/* prepare command */
+	cmd.header = mc_encode_cmd_header(DPIO_CMDID_CLOSE,
+					  MC_CMD_PRI_HIGH, token);
+
+	/* send command to mc*/
+	return mc_send_command(mc_io, &cmd);
+}
+
+int dpio_enable(struct fsl_mc_io *mc_io, uint16_t token)
+{
+	struct mc_command cmd = { 0 };
+
+	/* prepare command */
+	cmd.header = mc_encode_cmd_header(DPIO_CMDID_ENABLE,
+					  MC_CMD_PRI_LOW, token);
+
+	/* send command to mc*/
+	return mc_send_command(mc_io, &cmd);
+}
+
+int dpio_disable(struct fsl_mc_io *mc_io, uint16_t token)
+{
+	struct mc_command cmd = { 0 };
+
+	/* prepare command */
+	cmd.header = mc_encode_cmd_header(DPIO_CMDID_DISABLE,
+					  MC_CMD_PRI_LOW,
+					  token);
+
+	/* send command to mc*/
+	return mc_send_command(mc_io, &cmd);
+}
+
+int dpio_reset(struct fsl_mc_io *mc_io, uint16_t token)
+{
+	struct mc_command cmd = { 0 };
+
+	/* prepare command */
+	cmd.header = mc_encode_cmd_header(DPIO_CMDID_RESET,
+					  MC_CMD_PRI_LOW, token);
+
+	/* send command to mc*/
+	return mc_send_command(mc_io, &cmd);
+}
+
+int dpio_get_attributes(struct fsl_mc_io *mc_io,
+			uint16_t token,
+			struct dpio_attr *attr)
+{
+	struct mc_command cmd = { 0 };
+	int err;
+
+	/* prepare command */
+	cmd.header = mc_encode_cmd_header(DPIO_CMDID_GET_ATTR,
+					  MC_CMD_PRI_LOW,
+					  token);
+
+	/* send command to mc*/
+	err = mc_send_command(mc_io, &cmd);
+	if (err)
+		return err;
+
+	/* retrieve response parameters */
+	DPIO_RSP_GET_ATTR(cmd, attr);
+
+	return 0;
+}
diff --git a/drivers/net/fsl-mc/dpio/qbman_portal.c b/drivers/net/fsl-mc/dpio/qbman_portal.c
new file mode 100644
index 00000000000..dd2a7deee5f
--- /dev/null
+++ b/drivers/net/fsl-mc/dpio/qbman_portal.c
@@ -0,0 +1,593 @@
+/*
+ * Copyright (C) 2014 Freescale Semiconductor
+ *
+ * SPDX-License-Identifier:	GPL-2.0+
+ */
+
+#include "qbman_portal.h"
+
+/* QBMan portal management command codes */
+#define QBMAN_MC_ACQUIRE       0x30
+#define QBMAN_WQCHAN_CONFIGURE 0x46
+
+/* CINH register offsets */
+#define QBMAN_CINH_SWP_EQAR    0x8c0
+#define QBMAN_CINH_SWP_DCAP    0xac0
+#define QBMAN_CINH_SWP_SDQCR   0xb00
+#define QBMAN_CINH_SWP_RAR     0xcc0
+
+/* CENA register offsets */
+#define QBMAN_CENA_SWP_EQCR(n) (0x000 + ((uint32_t)(n) << 6))
+#define QBMAN_CENA_SWP_DQRR(n) (0x200 + ((uint32_t)(n) << 6))
+#define QBMAN_CENA_SWP_RCR(n)  (0x400 + ((uint32_t)(n) << 6))
+#define QBMAN_CENA_SWP_CR      0x600
+#define QBMAN_CENA_SWP_RR(vb)  (0x700 + ((uint32_t)(vb) >> 1))
+#define QBMAN_CENA_SWP_VDQCR   0x780
+
+/* Reverse mapping of QBMAN_CENA_SWP_DQRR() */
+#define QBMAN_IDX_FROM_DQRR(p) (((unsigned long)p & 0xff) >> 6)
+
+/*******************************/
+/* Pre-defined attribute codes */
+/*******************************/
+
+struct qb_attr_code code_generic_verb = QB_CODE(0, 0, 7);
+struct qb_attr_code code_generic_rslt = QB_CODE(0, 8, 8);
+
+/*************************/
+/* SDQCR attribute codes */
+/*************************/
+
+/* we put these here because at least some of them are required by
+ * qbman_swp_init() */
+struct qb_attr_code code_sdqcr_dct = QB_CODE(0, 24, 2);
+struct qb_attr_code code_sdqcr_fc = QB_CODE(0, 29, 1);
+struct qb_attr_code code_sdqcr_tok = QB_CODE(0, 16, 8);
+#define CODE_SDQCR_DQSRC(n) QB_CODE(0, n, 1)
+enum qbman_sdqcr_dct {
+	qbman_sdqcr_dct_null = 0,
+	qbman_sdqcr_dct_prio_ics,
+	qbman_sdqcr_dct_active_ics,
+	qbman_sdqcr_dct_active
+};
+enum qbman_sdqcr_fc {
+	qbman_sdqcr_fc_one = 0,
+	qbman_sdqcr_fc_up_to_3 = 1
+};
+
+/*********************************/
+/* Portal constructor/destructor */
+/*********************************/
+
+/* Software portals should always be in the power-on state when we initialise,
+ * due to the CCSR-based portal reset functionality that MC has. */
+struct qbman_swp *qbman_swp_init(const struct qbman_swp_desc *d)
+{
+	int ret;
+	struct qbman_swp *p = kmalloc(sizeof(*p), GFP_KERNEL);
+
+	if (!p)
+		return NULL;
+	p->desc = d;
+#ifdef QBMAN_CHECKING
+	p->mc.check = swp_mc_can_start;
+#endif
+	p->mc.valid_bit = QB_VALID_BIT;
+	p->sdq = 0;
+	qb_attr_code_encode(&code_sdqcr_dct, &p->sdq, qbman_sdqcr_dct_prio_ics);
+	qb_attr_code_encode(&code_sdqcr_fc, &p->sdq, qbman_sdqcr_fc_up_to_3);
+	qb_attr_code_encode(&code_sdqcr_tok, &p->sdq, 0xbb);
+	p->vdq.busy = 0; /* TODO: convert to atomic_t */
+	p->vdq.valid_bit = QB_VALID_BIT;
+	p->dqrr.next_idx = 0;
+	p->dqrr.valid_bit = QB_VALID_BIT;
+	ret = qbman_swp_sys_init(&p->sys, d);
+	if (ret) {
+		free(p);
+		printf("qbman_swp_sys_init() failed %d\n", ret);
+		return NULL;
+	}
+	qbman_cinh_write(&p->sys, QBMAN_CINH_SWP_SDQCR, p->sdq);
+	return p;
+}
+
+/***********************/
+/* Management commands */
+/***********************/
+
+/*
+ * Internal code common to all types of management commands.
+ */
+
+void *qbman_swp_mc_start(struct qbman_swp *p)
+{
+	void *ret;
+#ifdef QBMAN_CHECKING
+	BUG_ON(p->mc.check != swp_mc_can_start);
+#endif
+	ret = qbman_cena_write_start(&p->sys, QBMAN_CENA_SWP_CR);
+#ifdef QBMAN_CHECKING
+	if (!ret)
+		p->mc.check = swp_mc_can_submit;
+#endif
+	return ret;
+}
+
+void qbman_swp_mc_submit(struct qbman_swp *p, void *cmd, uint32_t cmd_verb)
+{
+	uint32_t *v = cmd;
+#ifdef QBMAN_CHECKING
+	BUG_ON(!p->mc.check != swp_mc_can_submit);
+#endif
+	lwsync();
+	/* TBD: "|=" is going to hurt performance. Need to move as many fields
+	 * out of word zero, and for those that remain, the "OR" needs to occur
+	 * at the caller side. This debug check helps to catch cases where the
+	 * caller wants to OR but has forgotten to do so. */
+	BUG_ON((*v & cmd_verb) != *v);
+	*v = cmd_verb | p->mc.valid_bit;
+	qbman_cena_write_complete(&p->sys, QBMAN_CENA_SWP_CR, cmd);
+	/* TODO: add prefetch support for GPP */
+#ifdef QBMAN_CHECKING
+	p->mc.check = swp_mc_can_poll;
+#endif
+}
+
+void *qbman_swp_mc_result(struct qbman_swp *p)
+{
+	uint32_t *ret, verb;
+#ifdef QBMAN_CHECKING
+	BUG_ON(p->mc.check != swp_mc_can_poll);
+#endif
+	ret = qbman_cena_read(&p->sys, QBMAN_CENA_SWP_RR(p->mc.valid_bit));
+	/* Remove the valid-bit - command completed iff the rest is non-zero */
+	verb = ret[0] & ~QB_VALID_BIT;
+	if (!verb)
+		return NULL;
+#ifdef QBMAN_CHECKING
+	p->mc.check = swp_mc_can_start;
+#endif
+	p->mc.valid_bit ^= QB_VALID_BIT;
+	return ret;
+}
+
+/***********/
+/* Enqueue */
+/***********/
+
+/* These should be const, eventually */
+static struct qb_attr_code code_eq_cmd = QB_CODE(0, 0, 2);
+static struct qb_attr_code code_eq_orp_en = QB_CODE(0, 2, 1);
+static struct qb_attr_code code_eq_tgt_id = QB_CODE(2, 0, 24);
+/* static struct qb_attr_code code_eq_tag = QB_CODE(3, 0, 32); */
+static struct qb_attr_code code_eq_qd_en = QB_CODE(0, 4, 1);
+static struct qb_attr_code code_eq_qd_bin = QB_CODE(4, 0, 16);
+static struct qb_attr_code code_eq_qd_pri = QB_CODE(4, 16, 4);
+static struct qb_attr_code code_eq_rsp_stash = QB_CODE(5, 16, 1);
+static struct qb_attr_code code_eq_rsp_lo = QB_CODE(6, 0, 32);
+static struct qb_attr_code code_eq_rsp_hi = QB_CODE(7, 0, 32);
+
+enum qbman_eq_cmd_e {
+	/* No enqueue, primarily for plugging ORP gaps for dropped frames */
+	qbman_eq_cmd_empty,
+	/* DMA an enqueue response once complete */
+	qbman_eq_cmd_respond,
+	/* DMA an enqueue response only if the enqueue fails */
+	qbman_eq_cmd_respond_reject
+};
+
+void qbman_eq_desc_clear(struct qbman_eq_desc *d)
+{
+	memset(d, 0, sizeof(*d));
+}
+
+void qbman_eq_desc_set_no_orp(struct qbman_eq_desc *d, int respond_success)
+{
+	uint32_t *cl = qb_cl(d);
+
+	qb_attr_code_encode(&code_eq_orp_en, cl, 0);
+	qb_attr_code_encode(&code_eq_cmd, cl,
+			    respond_success ? qbman_eq_cmd_respond :
+					      qbman_eq_cmd_respond_reject);
+}
+
+void qbman_eq_desc_set_response(struct qbman_eq_desc *d,
+				dma_addr_t storage_phys,
+				int stash)
+{
+	uint32_t *cl = qb_cl(d);
+
+	qb_attr_code_encode(&code_eq_rsp_lo, cl, lower32(storage_phys));
+	qb_attr_code_encode(&code_eq_rsp_hi, cl, upper32(storage_phys));
+	qb_attr_code_encode(&code_eq_rsp_stash, cl, !!stash);
+}
+
+
+void qbman_eq_desc_set_qd(struct qbman_eq_desc *d, uint32_t qdid,
+			  uint32_t qd_bin, uint32_t qd_prio)
+{
+	uint32_t *cl = qb_cl(d);
+
+	qb_attr_code_encode(&code_eq_qd_en, cl, 1);
+	qb_attr_code_encode(&code_eq_tgt_id, cl, qdid);
+	qb_attr_code_encode(&code_eq_qd_bin, cl, qd_bin);
+	qb_attr_code_encode(&code_eq_qd_pri, cl, qd_prio);
+}
+
+#define EQAR_IDX(eqar)     ((eqar) & 0x7)
+#define EQAR_VB(eqar)      ((eqar) & 0x80)
+#define EQAR_SUCCESS(eqar) ((eqar) & 0x100)
+
+int qbman_swp_enqueue(struct qbman_swp *s, const struct qbman_eq_desc *d,
+		      const struct qbman_fd *fd)
+{
+	uint32_t *p;
+	const uint32_t *cl = qb_cl(d);
+	uint32_t eqar = qbman_cinh_read(&s->sys, QBMAN_CINH_SWP_EQAR);
+	debug("EQAR=%08x\n", eqar);
+	if (!EQAR_SUCCESS(eqar))
+		return -EBUSY;
+	p = qbman_cena_write_start(&s->sys,
+				   QBMAN_CENA_SWP_EQCR(EQAR_IDX(eqar)));
+	word_copy(&p[1], &cl[1], 7);
+	word_copy(&p[8], fd, sizeof(*fd) >> 2);
+	lwsync();
+	/* Set the verb byte, have to substitute in the valid-bit */
+	p[0] = cl[0] | EQAR_VB(eqar);
+	qbman_cena_write_complete(&s->sys,
+				  QBMAN_CENA_SWP_EQCR(EQAR_IDX(eqar)),
+				  p);
+	return 0;
+}
+
+/***************************/
+/* Volatile (pull) dequeue */
+/***************************/
+
+/* These should be const, eventually */
+static struct qb_attr_code code_pull_dct = QB_CODE(0, 0, 2);
+static struct qb_attr_code code_pull_dt = QB_CODE(0, 2, 2);
+static struct qb_attr_code code_pull_rls = QB_CODE(0, 4, 1);
+static struct qb_attr_code code_pull_stash = QB_CODE(0, 5, 1);
+static struct qb_attr_code code_pull_numframes = QB_CODE(0, 8, 4);
+static struct qb_attr_code code_pull_token = QB_CODE(0, 16, 8);
+static struct qb_attr_code code_pull_dqsource = QB_CODE(1, 0, 24);
+static struct qb_attr_code code_pull_rsp_lo = QB_CODE(2, 0, 32);
+static struct qb_attr_code code_pull_rsp_hi = QB_CODE(3, 0, 32);
+
+enum qb_pull_dt_e {
+	qb_pull_dt_channel,
+	qb_pull_dt_workqueue,
+	qb_pull_dt_framequeue
+};
+
+void qbman_pull_desc_clear(struct qbman_pull_desc *d)
+{
+	memset(d, 0, sizeof(*d));
+}
+
+void qbman_pull_desc_set_storage(struct qbman_pull_desc *d,
+				 struct ldpaa_dq *storage,
+				 dma_addr_t storage_phys,
+				 int stash)
+{
+	uint32_t *cl = qb_cl(d);
+
+	/* Squiggle the pointer 'storage' into the extra 2 words of the
+	 * descriptor (which aren't copied to the hw command) */
+	*(void **)&cl[4] = storage;
+	if (!storage) {
+		qb_attr_code_encode(&code_pull_rls, cl, 0);
+		return;
+	}
+	qb_attr_code_encode(&code_pull_rls, cl, 1);
+	qb_attr_code_encode(&code_pull_stash, cl, !!stash);
+	qb_attr_code_encode(&code_pull_rsp_lo, cl, lower32(storage_phys));
+	qb_attr_code_encode(&code_pull_rsp_hi, cl, upper32(storage_phys));
+}
+
+void qbman_pull_desc_set_numframes(struct qbman_pull_desc *d, uint8_t numframes)
+{
+	uint32_t *cl = qb_cl(d);
+
+	BUG_ON(!numframes || (numframes > 16));
+	qb_attr_code_encode(&code_pull_numframes, cl,
+			    (uint32_t)(numframes - 1));
+}
+
+void qbman_pull_desc_set_token(struct qbman_pull_desc *d, uint8_t token)
+{
+	uint32_t *cl = qb_cl(d);
+
+	qb_attr_code_encode(&code_pull_token, cl, token);
+}
+
+void qbman_pull_desc_set_fq(struct qbman_pull_desc *d, uint32_t fqid)
+{
+	uint32_t *cl = qb_cl(d);
+
+	qb_attr_code_encode(&code_pull_dct, cl, 1);
+	qb_attr_code_encode(&code_pull_dt, cl, qb_pull_dt_framequeue);
+	qb_attr_code_encode(&code_pull_dqsource, cl, fqid);
+}
+
+int qbman_swp_pull(struct qbman_swp *s, struct qbman_pull_desc *d)
+{
+	uint32_t *p;
+	uint32_t *cl = qb_cl(d);
+
+	/* TODO: convert to atomic_t */
+	if (s->vdq.busy)
+		return -EBUSY;
+	s->vdq.busy = 1;
+	s->vdq.storage = *(void **)&cl[4];
+	s->vdq.token = qb_attr_code_decode(&code_pull_token, cl);
+	p = qbman_cena_write_start(&s->sys, QBMAN_CENA_SWP_VDQCR);
+	word_copy(&p[1], &cl[1], 3);
+	lwsync();
+	/* Set the verb byte, have to substitute in the valid-bit */
+	p[0] = cl[0] | s->vdq.valid_bit;
+	s->vdq.valid_bit ^= QB_VALID_BIT;
+	qbman_cena_write_complete(&s->sys, QBMAN_CENA_SWP_VDQCR, p);
+	return 0;
+}
+
+/****************/
+/* Polling DQRR */
+/****************/
+
+static struct qb_attr_code code_dqrr_verb = QB_CODE(0, 0, 8);
+static struct qb_attr_code code_dqrr_response = QB_CODE(0, 0, 7);
+static struct qb_attr_code code_dqrr_stat = QB_CODE(0, 8, 8);
+
+#define QBMAN_DQRR_RESPONSE_DQ        0x60
+#define QBMAN_DQRR_RESPONSE_FQRN      0x21
+#define QBMAN_DQRR_RESPONSE_FQRNI     0x22
+#define QBMAN_DQRR_RESPONSE_FQPN      0x24
+#define QBMAN_DQRR_RESPONSE_FQDAN     0x25
+#define QBMAN_DQRR_RESPONSE_CDAN      0x26
+#define QBMAN_DQRR_RESPONSE_CSCN_MEM  0x27
+#define QBMAN_DQRR_RESPONSE_CGCU      0x28
+#define QBMAN_DQRR_RESPONSE_BPSCN     0x29
+#define QBMAN_DQRR_RESPONSE_CSCN_WQ   0x2a
+
+
+/* NULL return if there are no unconsumed DQRR entries. Returns a DQRR entry
+ * only once, so repeated calls can return a sequence of DQRR entries, without
+ * requiring they be consumed immediately or in any particular order. */
+const struct ldpaa_dq *qbman_swp_dqrr_next(struct qbman_swp *s)
+{
+	uint32_t verb;
+	uint32_t response_verb;
+	const struct ldpaa_dq *dq = qbman_cena_read(&s->sys,
+					QBMAN_CENA_SWP_DQRR(s->dqrr.next_idx));
+	const uint32_t *p = qb_cl(dq);
+
+	verb = qb_attr_code_decode(&code_dqrr_verb, p);
+	/* If the valid-bit isn't of the expected polarity, nothing there */
+	if ((verb & QB_VALID_BIT) != s->dqrr.valid_bit) {
+		qbman_cena_invalidate_prefetch(&s->sys,
+					       QBMAN_CENA_SWP_DQRR(
+					       s->dqrr.next_idx));
+		return NULL;
+	}
+	/* There's something there. Move "next_idx" attention to the next ring
+	 * entry (and prefetch it) before returning what we found. */
+	s->dqrr.next_idx++;
+	s->dqrr.next_idx &= 3; /* Wrap around at 4 */
+	/* TODO: it's possible to do all this without conditionals, optimise it
+	 * later. */
+	if (!s->dqrr.next_idx)
+		s->dqrr.valid_bit ^= QB_VALID_BIT;
+	/* VDQCR "no longer busy" hook - if VDQCR shows "busy" and this is a
+	 * VDQCR result, mark it as non-busy. */
+	if (s->vdq.busy) {
+		uint32_t flags = ldpaa_dq_flags(dq);
+
+		response_verb = qb_attr_code_decode(&code_dqrr_response, &verb);
+		if ((response_verb == QBMAN_DQRR_RESPONSE_DQ) &&
+		    (flags & LDPAA_DQ_STAT_VOLATILE))
+			s->vdq.busy = 0;
+	}
+	qbman_cena_invalidate_prefetch(&s->sys,
+				       QBMAN_CENA_SWP_DQRR(s->dqrr.next_idx));
+	return dq;
+}
+
+/* Consume DQRR entries previously returned from qbman_swp_dqrr_next(). */
+void qbman_swp_dqrr_consume(struct qbman_swp *s, const struct ldpaa_dq *dq)
+{
+	qbman_cinh_write(&s->sys, QBMAN_CINH_SWP_DCAP, QBMAN_IDX_FROM_DQRR(dq));
+}
+
+/*********************************/
+/* Polling user-provided storage */
+/*********************************/
+
+void qbman_dq_entry_set_oldtoken(struct ldpaa_dq *dq,
+				 unsigned int num_entries,
+				 uint8_t oldtoken)
+{
+	memset(dq, oldtoken, num_entries * sizeof(*dq));
+}
+
+int qbman_dq_entry_has_newtoken(struct qbman_swp *s,
+				const struct ldpaa_dq *dq,
+				uint8_t newtoken)
+{
+	/* To avoid converting the little-endian DQ entry to host-endian prior
+	 * to us knowing whether there is a valid entry or not (and run the
+	 * risk of corrupting the incoming hardware LE write), we detect in
+	 * hardware endianness rather than host. This means we need a different
+	 * "code" depending on whether we are BE or LE in software, which is
+	 * where DQRR_TOK_OFFSET comes in... */
+	static struct qb_attr_code code_dqrr_tok_detect =
+					QB_CODE(0, DQRR_TOK_OFFSET, 8);
+	/* The user trying to poll for a result treats "dq" as const. It is
+	 * however the same address that was provided to us non-const in the
+	 * first place, for directing hardware DMA to. So we can cast away the
+	 * const because it is mutable from our perspective. */
+	uint32_t *p = qb_cl((struct ldpaa_dq *)dq);
+	uint32_t token;
+
+	token = qb_attr_code_decode(&code_dqrr_tok_detect, &p[1]);
+	if (token != newtoken)
+		return 0;
+
+	/* Only now do we convert from hardware to host endianness. Also, as we
+	 * are returning success, the user has promised not to call us again, so
+	 * there's no risk of us converting the endianness twice... */
+	make_le32_n(p, 16);
+
+	/* VDQCR "no longer busy" hook - not quite the same as DQRR, because the
+	 * fact "VDQCR" shows busy doesn't mean that the result we're looking at
+	 * is from the same command. Eg. we may be looking at our 10th dequeue
+	 * result from our first VDQCR command, yet the second dequeue command
+	 * could have been kicked off already, after seeing the 1st result. Ie.
+	 * the result we're looking at is not necessarily proof that we can
+	 * reset "busy".  We instead base the decision on whether the current
+	 * result is sitting at the first 'storage' location of the busy
+	 * command. */
+	if (s->vdq.busy && (s->vdq.storage == dq))
+		s->vdq.busy = 0;
+	return 1;
+}
+
+/********************************/
+/* Categorising dequeue entries */
+/********************************/
+
+static inline int __qbman_dq_entry_is_x(const struct ldpaa_dq *dq, uint32_t x)
+{
+	const uint32_t *p = qb_cl(dq);
+	uint32_t response_verb = qb_attr_code_decode(&code_dqrr_response, p);
+
+	return response_verb == x;
+}
+
+int qbman_dq_entry_is_DQ(const struct ldpaa_dq *dq)
+{
+	return __qbman_dq_entry_is_x(dq, QBMAN_DQRR_RESPONSE_DQ);
+}
+
+/*********************************/
+/* Parsing frame dequeue results */
+/*********************************/
+
+/* These APIs assume qbman_dq_entry_is_DQ() is TRUE */
+
+uint32_t ldpaa_dq_flags(const struct ldpaa_dq *dq)
+{
+	const uint32_t *p = qb_cl(dq);
+
+	return qb_attr_code_decode(&code_dqrr_stat, p);
+}
+
+const struct dpaa_fd *ldpaa_dq_fd(const struct ldpaa_dq *dq)
+{
+	const uint32_t *p = qb_cl(dq);
+
+	return (const struct dpaa_fd *)&p[8];
+}
+
+/******************/
+/* Buffer release */
+/******************/
+
+/* These should be const, eventually */
+/* static struct qb_attr_code code_release_num = QB_CODE(0, 0, 3); */
+static struct qb_attr_code code_release_set_me = QB_CODE(0, 5, 1);
+static struct qb_attr_code code_release_bpid = QB_CODE(0, 16, 16);
+
+void qbman_release_desc_clear(struct qbman_release_desc *d)
+{
+	uint32_t *cl;
+
+	memset(d, 0, sizeof(*d));
+	cl = qb_cl(d);
+	qb_attr_code_encode(&code_release_set_me, cl, 1);
+}
+
+void qbman_release_desc_set_bpid(struct qbman_release_desc *d, uint32_t bpid)
+{
+	uint32_t *cl = qb_cl(d);
+
+	qb_attr_code_encode(&code_release_bpid, cl, bpid);
+}
+
+#define RAR_IDX(rar)     ((rar) & 0x7)
+#define RAR_VB(rar)      ((rar) & 0x80)
+#define RAR_SUCCESS(rar) ((rar) & 0x100)
+
+int qbman_swp_release(struct qbman_swp *s, const struct qbman_release_desc *d,
+		      const uint64_t *buffers, unsigned int num_buffers)
+{
+	uint32_t *p;
+	const uint32_t *cl = qb_cl(d);
+	uint32_t rar = qbman_cinh_read(&s->sys, QBMAN_CINH_SWP_RAR);
+	debug("RAR=%08x\n", rar);
+	if (!RAR_SUCCESS(rar))
+		return -EBUSY;
+	BUG_ON(!num_buffers || (num_buffers > 7));
+	/* Start the release command */
+	p = qbman_cena_write_start(&s->sys,
+				   QBMAN_CENA_SWP_RCR(RAR_IDX(rar)));
+	/* Copy the caller's buffer pointers to the command */
+	u64_to_le32_copy(&p[2], buffers, num_buffers);
+	lwsync();
+	/* Set the verb byte, have to substitute in the valid-bit and the number
+	 * of buffers. */
+	p[0] = cl[0] | RAR_VB(rar) | num_buffers;
+	qbman_cena_write_complete(&s->sys,
+				  QBMAN_CENA_SWP_RCR(RAR_IDX(rar)),
+				  p);
+	return 0;
+}
+
+/*******************/
+/* Buffer acquires */
+/*******************/
+
+/* These should be const, eventually */
+static struct qb_attr_code code_acquire_bpid = QB_CODE(0, 16, 16);
+static struct qb_attr_code code_acquire_num = QB_CODE(1, 0, 3);
+static struct qb_attr_code code_acquire_r_num = QB_CODE(1, 0, 3);
+
+int qbman_swp_acquire(struct qbman_swp *s, uint32_t bpid, uint64_t *buffers,
+		      unsigned int num_buffers)
+{
+	uint32_t *p;
+	uint32_t verb, rslt, num;
+
+	BUG_ON(!num_buffers || (num_buffers > 7));
+
+	/* Start the management command */
+	p = qbman_swp_mc_start(s);
+
+	if (!p)
+		return -EBUSY;
+
+	/* Encode the caller-provided attributes */
+	qb_attr_code_encode(&code_acquire_bpid, p, bpid);
+	qb_attr_code_encode(&code_acquire_num, p, num_buffers);
+
+	/* Complete the management command */
+	p = qbman_swp_mc_complete(s, p, p[0] | QBMAN_MC_ACQUIRE);
+
+	/* Decode the outcome */
+	verb = qb_attr_code_decode(&code_generic_verb, p);
+	rslt = qb_attr_code_decode(&code_generic_rslt, p);
+	num = qb_attr_code_decode(&code_acquire_r_num, p);
+	BUG_ON(verb != QBMAN_MC_ACQUIRE);
+
+	/* Determine success or failure */
+	if (unlikely(rslt != QBMAN_MC_RSLT_OK)) {
+		printf("Acquire buffers from BPID 0x%x failed, code=0x%02x\n",
+		       bpid, rslt);
+		return -EIO;
+	}
+	BUG_ON(num > num_buffers);
+	/* Copy the acquired buffers to the caller's array */
+	u64_from_le32_copy(buffers, &p[2], num);
+	return (int)num;
+}
diff --git a/drivers/net/fsl-mc/dpio/qbman_portal.h b/drivers/net/fsl-mc/dpio/qbman_portal.h
new file mode 100644
index 00000000000..bb67c3bd066
--- /dev/null
+++ b/drivers/net/fsl-mc/dpio/qbman_portal.h
@@ -0,0 +1,157 @@
+/*
+ * Copyright (C) 2014 Freescale Semiconductor
+ *
+ * SPDX-License-Identifier:	GPL-2.0+
+ */
+
+#include "qbman_private.h"
+#include <fsl-mc/fsl_qbman_portal.h>
+#include <fsl-mc/fsl_dpaa_fd.h>
+
+/* All QBMan command and result structures use this "valid bit" encoding */
+#define QB_VALID_BIT ((uint32_t)0x80)
+
+/* Management command result codes */
+#define QBMAN_MC_RSLT_OK      0xf0
+
+/* --------------------- */
+/* portal data structure */
+/* --------------------- */
+
+struct qbman_swp {
+	const struct qbman_swp_desc *desc;
+	/* The qbman_sys (ie. arch/OS-specific) support code can put anything it
+	 * needs in here. */
+	struct qbman_swp_sys sys;
+	/* Management commands */
+	struct {
+#ifdef QBMAN_CHECKING
+		enum swp_mc_check {
+			swp_mc_can_start, /* call __qbman_swp_mc_start() */
+			swp_mc_can_submit, /* call __qbman_swp_mc_submit() */
+			swp_mc_can_poll, /* call __qbman_swp_mc_result() */
+		} check;
+#endif
+		uint32_t valid_bit; /* 0x00 or 0x80 */
+	} mc;
+	/* Push dequeues */
+	uint32_t sdq;
+	/* Volatile dequeues */
+	struct {
+		/* VDQCR supports a "1 deep pipeline", meaning that if you know
+		 * the last-submitted command is already executing in the
+		 * hardware (as evidenced by at least 1 valid dequeue result),
+		 * you can write another dequeue command to the register, the
+		 * hardware will start executing it as soon as the
+		 * already-executing command terminates. (This minimises latency
+		 * and stalls.) With that in mind, this "busy" variable refers
+		 * to whether or not a command can be submitted, not whether or
+		 * not a previously-submitted command is still executing. In
+		 * other words, once proof is seen that the previously-submitted
+		 * command is executing, "vdq" is no longer "busy". TODO:
+		 * convert this to "atomic_t" so that it is thread-safe (without
+		 * locking). */
+		int busy;
+		uint32_t valid_bit; /* 0x00 or 0x80 */
+		/* We need to determine when vdq is no longer busy. This depends
+		 * on whether the "busy" (last-submitted) dequeue command is
+		 * targetting DQRR or main-memory, and detected is based on the
+		 * presence of the dequeue command's "token" showing up in
+		 * dequeue entries in DQRR or main-memory (respectively). Debug
+		 * builds will, when submitting vdq commands, verify that the
+		 * dequeue result location is not already equal to the command's
+		 * token value. */
+		struct ldpaa_dq *storage; /* NULL if DQRR */
+		uint32_t token;
+	} vdq;
+	/* DQRR */
+	struct {
+		uint32_t next_idx;
+		uint32_t valid_bit;
+	} dqrr;
+};
+
+/* -------------------------- */
+/* portal management commands */
+/* -------------------------- */
+
+/* Different management commands all use this common base layer of code to issue
+ * commands and poll for results. The first function returns a pointer to where
+ * the caller should fill in their MC command (though they should ignore the
+ * verb byte), the second function commits merges in the caller-supplied command
+ * verb (which should not include the valid-bit) and submits the command to
+ * hardware, and the third function checks for a completed response (returns
+ * non-NULL if only if the response is complete). */
+void *qbman_swp_mc_start(struct qbman_swp *p);
+void qbman_swp_mc_submit(struct qbman_swp *p, void *cmd, uint32_t cmd_verb);
+void *qbman_swp_mc_result(struct qbman_swp *p);
+
+/* Wraps up submit + poll-for-result */
+static inline void *qbman_swp_mc_complete(struct qbman_swp *swp, void *cmd,
+					  uint32_t cmd_verb)
+{
+	int loopvar;
+
+	qbman_swp_mc_submit(swp, cmd, cmd_verb);
+	DBG_POLL_START(loopvar);
+	do {
+		DBG_POLL_CHECK(loopvar);
+		cmd = qbman_swp_mc_result(swp);
+	} while (!cmd);
+	return cmd;
+}
+
+/* ------------ */
+/* qb_attr_code */
+/* ------------ */
+
+/* This struct locates a sub-field within a QBMan portal (CENA) cacheline which
+ * is either serving as a configuration command or a query result. The
+ * representation is inherently little-endian, as the indexing of the words is
+ * itself little-endian in nature and layerscape is little endian for anything
+ * that crosses a word boundary too (64-bit fields are the obvious examples).
+ */
+struct qb_attr_code {
+	unsigned int word; /* which uint32_t[] array member encodes the field */
+	unsigned int lsoffset; /* encoding offset from ls-bit */
+	unsigned int width; /* encoding width. (bool must be 1.) */
+};
+
+/* Macros to define codes */
+#define QB_CODE(a, b, c) { a, b, c}
+
+/* decode a field from a cacheline */
+static inline uint32_t qb_attr_code_decode(const struct qb_attr_code *code,
+				      const uint32_t *cacheline)
+{
+	return d32_uint32_t(code->lsoffset, code->width, cacheline[code->word]);
+}
+
+/* encode a field to a cacheline */
+static inline void qb_attr_code_encode(const struct qb_attr_code *code,
+				       uint32_t *cacheline, uint32_t val)
+{
+	cacheline[code->word] =
+		r32_uint32_t(code->lsoffset, code->width, cacheline[code->word])
+		| e32_uint32_t(code->lsoffset, code->width, val);
+}
+
+/* ---------------------- */
+/* Descriptors/cachelines */
+/* ---------------------- */
+
+/* To avoid needless dynamic allocation, the driver API often gives the caller
+ * a "descriptor" type that the caller can instantiate however they like.
+ * Ultimately though, it is just a cacheline of binary storage (or something
+ * smaller when it is known that the descriptor doesn't need all 64 bytes) for
+ * holding pre-formatted pieces of harware commands. The performance-critical
+ * code can then copy these descriptors directly into hardware command
+ * registers more efficiently than trying to construct/format commands
+ * on-the-fly. The API user sees the descriptor as an array of 32-bit words in
+ * order for the compiler to know its size, but the internal details are not
+ * exposed. The following macro is used within the driver for converting *any*
+ * descriptor pointer to a usable array pointer. The use of a macro (instead of
+ * an inline) is necessary to work with different descriptor types and to work
+ * correctly with const and non-const inputs (and similarly-qualified outputs).
+ */
+#define qb_cl(d) (&(d)->dont_manipulate_directly[0])
diff --git a/drivers/net/fsl-mc/dpio/qbman_private.h b/drivers/net/fsl-mc/dpio/qbman_private.h
new file mode 100644
index 00000000000..2d2556b7556
--- /dev/null
+++ b/drivers/net/fsl-mc/dpio/qbman_private.h
@@ -0,0 +1,169 @@
+/*
+ * Copyright (C) 2014 Freescale Semiconductor
+ *
+ * SPDX-License-Identifier:	GPL-2.0+
+ */
+
+/* Perform extra checking */
+#include <common.h>
+#include <errno.h>
+#include <asm/io.h>
+#include <linux/types.h>
+#include <linux/compat.h>
+#include <malloc.h>
+#include <fsl-mc/fsl_qbman_base.h>
+
+#define QBMAN_CHECKING
+
+/* Any time there is a register interface which we poll on, this provides a
+ * "break after x iterations" scheme for it. It's handy for debugging, eg.
+ * where you don't want millions of lines of log output from a polling loop
+ * that won't, because such things tend to drown out the earlier log output
+ * that might explain what caused the problem. (NB: put ";" after each macro!)
+ * TODO: we should probably remove this once we're done sanitising the
+ * simulator...
+ */
+#define DBG_POLL_START(loopvar) (loopvar = 10)
+#define DBG_POLL_CHECK(loopvar) \
+	do {if (!(loopvar--)) BUG_ON(NULL == "DBG_POLL_CHECK"); } while (0)
+
+/* For CCSR or portal-CINH registers that contain fields at arbitrary offsets
+ * and widths, these macro-generated encode/decode/isolate/remove inlines can
+ * be used.
+ *
+ * Eg. to "d"ecode a 14-bit field out of a register (into a "uint16_t" type),
+ * where the field is located 3 bits "up" from the least-significant bit of the
+ * register (ie. the field location within the 32-bit register corresponds to a
+ * mask of 0x0001fff8), you would do;
+ *                uint16_t field = d32_uint16_t(3, 14, reg_value);
+ *
+ * Or to "e"ncode a 1-bit boolean value (input type is "int", zero is FALSE,
+ * non-zero is TRUE, so must convert all non-zero inputs to 1, hence the "!!"
+ * operator) into a register at bit location 0x00080000 (19 bits "in" from the
+ * LS bit), do;
+ *                reg_value |= e32_int(19, 1, !!field);
+ *
+ * If you wish to read-modify-write a register, such that you leave the 14-bit
+ * field as-is but have all other fields set to zero, then "i"solate the 14-bit
+ * value using;
+ *                reg_value = i32_uint16_t(3, 14, reg_value);
+ *
+ * Alternatively, you could "r"emove the 1-bit boolean field (setting it to
+ * zero) but leaving all other fields as-is;
+ *                reg_val = r32_int(19, 1, reg_value);
+ *
+ */
+#define MAKE_MASK32(width) (width == 32 ? 0xffffffff : \
+				 (uint32_t)((1 << width) - 1))
+#define DECLARE_CODEC32(t) \
+static inline uint32_t e32_##t(uint32_t lsoffset, uint32_t width, t val) \
+{ \
+	BUG_ON(width > (sizeof(t) * 8)); \
+	return ((uint32_t)val & MAKE_MASK32(width)) << lsoffset; \
+} \
+static inline t d32_##t(uint32_t lsoffset, uint32_t width, uint32_t val) \
+{ \
+	BUG_ON(width > (sizeof(t) * 8)); \
+	return (t)((val >> lsoffset) & MAKE_MASK32(width)); \
+} \
+static inline uint32_t i32_##t(uint32_t lsoffset, uint32_t width, \
+				uint32_t val) \
+{ \
+	BUG_ON(width > (sizeof(t) * 8)); \
+	return e32_##t(lsoffset, width, d32_##t(lsoffset, width, val)); \
+} \
+static inline uint32_t r32_##t(uint32_t lsoffset, uint32_t width, \
+				uint32_t val) \
+{ \
+	BUG_ON(width > (sizeof(t) * 8)); \
+	return ~(MAKE_MASK32(width) << lsoffset) & val; \
+}
+DECLARE_CODEC32(uint32_t)
+DECLARE_CODEC32(uint16_t)
+DECLARE_CODEC32(uint8_t)
+DECLARE_CODEC32(int)
+
+	/*********************/
+	/* Debugging assists */
+	/*********************/
+
+static inline void __hexdump(unsigned long start, unsigned long end,
+			unsigned long p, size_t sz, const unsigned char *c)
+{
+	while (start < end) {
+		unsigned int pos = 0;
+		char buf[64];
+		int nl = 0;
+
+		pos += sprintf(buf + pos, "%08lx: ", start);
+		do {
+			if ((start < p) || (start >= (p + sz)))
+				pos += sprintf(buf + pos, "..");
+			else
+				pos += sprintf(buf + pos, "%02x", *(c++));
+			if (!(++start & 15)) {
+				buf[pos++] = '\n';
+				nl = 1;
+			} else {
+				nl = 0;
+				if (!(start & 1))
+					buf[pos++] = ' ';
+				if (!(start & 3))
+					buf[pos++] = ' ';
+			}
+		} while (start & 15);
+		if (!nl)
+			buf[pos++] = '\n';
+		buf[pos] = '\0';
+		debug("%s", buf);
+	}
+}
+static inline void hexdump(const void *ptr, size_t sz)
+{
+	unsigned long p = (unsigned long)ptr;
+	unsigned long start = p & ~(unsigned long)15;
+	unsigned long end = (p + sz + 15) & ~(unsigned long)15;
+	const unsigned char *c = ptr;
+
+	__hexdump(start, end, p, sz, c);
+}
+
+#if defined(__BIG_ENDIAN)
+#define DQRR_TOK_OFFSET 0
+#else
+#define DQRR_TOK_OFFSET 24
+#endif
+
+/* Similarly-named functions */
+#define upper32(a) upper_32_bits(a)
+#define lower32(a) lower_32_bits(a)
+
+	/****************/
+	/* arch assists */
+	/****************/
+
+static inline void dcbz(void *ptr)
+{
+	uint32_t *p = ptr;
+	BUG_ON((unsigned long)ptr & 63);
+	p[0] = 0;
+	p[1] = 0;
+	p[2] = 0;
+	p[3] = 0;
+	p[4] = 0;
+	p[5] = 0;
+	p[6] = 0;
+	p[7] = 0;
+	p[8] = 0;
+	p[9] = 0;
+	p[10] = 0;
+	p[11] = 0;
+	p[12] = 0;
+	p[13] = 0;
+	p[14] = 0;
+	p[15] = 0;
+}
+
+#define lwsync()
+
+#include "qbman_sys.h"
diff --git a/drivers/net/fsl-mc/dpio/qbman_sys.h b/drivers/net/fsl-mc/dpio/qbman_sys.h
new file mode 100644
index 00000000000..235d641bd4c
--- /dev/null
+++ b/drivers/net/fsl-mc/dpio/qbman_sys.h
@@ -0,0 +1,290 @@
+/*
+ * Copyright (C) 2014 Freescale Semiconductor
+ *
+ * SPDX-License-Identifier:	GPL-2.0+
+ */
+
+/* qbman_sys_decl.h and qbman_sys.h are the two platform-specific files in the
+ * driver. They are only included via qbman_private.h, which is itself a
+ * platform-independent file and is included by all the other driver source.
+ *
+ * qbman_sys_decl.h is included prior to all other declarations and logic, and
+ * it exists to provide compatibility with any linux interfaces our
+ * single-source driver code is dependent on (eg. kmalloc). Ie. this file
+ * provides linux compatibility.
+ *
+ * This qbman_sys.h header, on the other hand, is included *after* any common
+ * and platform-neutral declarations and logic in qbman_private.h, and exists to
+ * implement any platform-specific logic of the qbman driver itself. Ie. it is
+ * *not* to provide linux compatibility.
+ */
+
+/* Trace the 3 different classes of read/write access to QBMan. #undef as
+ * required. */
+#undef QBMAN_CCSR_TRACE
+#undef QBMAN_CINH_TRACE
+#undef QBMAN_CENA_TRACE
+
+/* Temporarily define this to get around the fact that cache enabled mapping is
+ * not working right now. Will remove this after uboot could map the cache
+ * enabled portal memory.
+ */
+#define QBMAN_CINH_ONLY
+
+static inline void word_copy(void *d, const void *s, unsigned int cnt)
+{
+	uint32_t *dd = d;
+	const uint32_t *ss = s;
+
+	while (cnt--)
+		*(dd++) = *(ss++);
+}
+
+/* Currently, the CENA support code expects each 32-bit word to be written in
+ * host order, and these are converted to hardware (little-endian) order on
+ * command submission. However, 64-bit quantities are must be written (and read)
+ * as two 32-bit words with the least-significant word first, irrespective of
+ * host endianness. */
+static inline void u64_to_le32_copy(void *d, const uint64_t *s,
+					unsigned int cnt)
+{
+	uint32_t *dd = d;
+	const uint32_t *ss = (const uint32_t *)s;
+
+	while (cnt--) {
+		/* TBD: the toolchain was choking on the use of 64-bit types up
+		 * until recently so this works entirely with 32-bit variables.
+		 * When 64-bit types become usable again, investigate better
+		 * ways of doing this. */
+#if defined(__BIG_ENDIAN)
+		*(dd++) = ss[1];
+		*(dd++) = ss[0];
+		ss += 2;
+#else
+		*(dd++) = *(ss++);
+		*(dd++) = *(ss++);
+#endif
+	}
+}
+static inline void u64_from_le32_copy(uint64_t *d, const void *s,
+					unsigned int cnt)
+{
+	const uint32_t *ss = s;
+	uint32_t *dd = (uint32_t *)d;
+
+	while (cnt--) {
+#if defined(__BIG_ENDIAN)
+		dd[1] = *(ss++);
+		dd[0] = *(ss++);
+		dd += 2;
+#else
+		*(dd++) = *(ss++);
+		*(dd++) = *(ss++);
+#endif
+	}
+}
+
+/* Convert a host-native 32bit value into little endian */
+#if defined(__BIG_ENDIAN)
+static inline uint32_t make_le32(uint32_t val)
+{
+	return ((val & 0xff) << 24) | ((val & 0xff00) << 8) |
+		((val & 0xff0000) >> 8) | ((val & 0xff000000) >> 24);
+}
+#else
+#define make_le32(val) (val)
+#endif
+static inline void make_le32_n(uint32_t *val, unsigned int num)
+{
+	while (num--) {
+		*val = make_le32(*val);
+		val++;
+	}
+}
+
+	/******************/
+	/* Portal access  */
+	/******************/
+struct qbman_swp_sys {
+	/* On GPP, the sys support for qbman_swp is here. The CENA region isi
+	 * not an mmap() of the real portal registers, but an allocated
+	 * place-holder, because the actual writes/reads to/from the portal are
+	 * marshalled from these allocated areas using QBMan's "MC access
+	 * registers". CINH accesses are atomic so there's no need for a
+	 * place-holder. */
+	void *cena;
+	void __iomem *addr_cena;
+	void __iomem *addr_cinh;
+};
+
+/* P_OFFSET is (ACCESS_CMD,0,12) - offset within the portal
+ * C is (ACCESS_CMD,12,1) - is inhibited? (0==CENA, 1==CINH)
+ * SWP_IDX is (ACCESS_CMD,16,10) - Software portal index
+ * P is (ACCESS_CMD,28,1) - (0==special portal, 1==any portal)
+ * T is (ACCESS_CMD,29,1) - Command type (0==READ, 1==WRITE)
+ * E is (ACCESS_CMD,31,1) - Command execute (1 to issue, poll for 0==complete)
+ */
+
+static inline void qbman_cinh_write(struct qbman_swp_sys *s, uint32_t offset,
+				    uint32_t val)
+{
+	__raw_writel(val, s->addr_cinh + offset);
+#ifdef QBMAN_CINH_TRACE
+	pr_info("qbman_cinh_write(%p:0x%03x) 0x%08x\n",
+		s->addr_cinh, offset, val);
+#endif
+}
+
+static inline uint32_t qbman_cinh_read(struct qbman_swp_sys *s, uint32_t offset)
+{
+	uint32_t reg = __raw_readl(s->addr_cinh + offset);
+
+#ifdef QBMAN_CINH_TRACE
+	pr_info("qbman_cinh_read(%p:0x%03x) 0x%08x\n",
+		s->addr_cinh, offset, reg);
+#endif
+	return reg;
+}
+
+static inline void *qbman_cena_write_start(struct qbman_swp_sys *s,
+						uint32_t offset)
+{
+	void *shadow = s->cena + offset;
+
+#ifdef QBMAN_CENA_TRACE
+	pr_info("qbman_cena_write_start(%p:0x%03x) %p\n",
+		s->addr_cena, offset, shadow);
+#endif
+	BUG_ON(offset & 63);
+	dcbz(shadow);
+	return shadow;
+}
+
+static inline void qbman_cena_write_complete(struct qbman_swp_sys *s,
+						uint32_t offset, void *cmd)
+{
+	const uint32_t *shadow = cmd;
+	int loop;
+
+#ifdef QBMAN_CENA_TRACE
+	pr_info("qbman_cena_write_complete(%p:0x%03x) %p\n",
+		s->addr_cena, offset, shadow);
+	hexdump(cmd, 64);
+#endif
+	for (loop = 15; loop >= 0; loop--)
+#ifdef QBMAN_CINH_ONLY
+		__raw_writel(shadow[loop], s->addr_cinh +
+					 offset + loop * 4);
+#else
+		__raw_writel(shadow[loop], s->addr_cena +
+					 offset + loop * 4);
+#endif
+}
+
+static inline void *qbman_cena_read(struct qbman_swp_sys *s, uint32_t offset)
+{
+	uint32_t *shadow = s->cena + offset;
+	unsigned int loop;
+
+#ifdef QBMAN_CENA_TRACE
+	pr_info("qbman_cena_read(%p:0x%03x) %p\n",
+		s->addr_cena, offset, shadow);
+#endif
+
+	for (loop = 0; loop < 16; loop++)
+#ifdef QBMAN_CINH_ONLY
+		shadow[loop] = __raw_readl(s->addr_cinh + offset
+					+ loop * 4);
+#else
+		shadow[loop] = __raw_readl(s->addr_cena + offset
+					+ loop * 4);
+#endif
+#ifdef QBMAN_CENA_TRACE
+	hexdump(shadow, 64);
+#endif
+	return shadow;
+}
+
+static inline void qbman_cena_invalidate_prefetch(struct qbman_swp_sys *s,
+						  uint32_t offset)
+{
+}
+
+	/******************/
+	/* Portal support */
+	/******************/
+
+/* The SWP_CFG portal register is special, in that it is used by the
+ * platform-specific code rather than the platform-independent code in
+ * qbman_portal.c. So use of it is declared locally here. */
+#define QBMAN_CINH_SWP_CFG   0xd00
+
+/* For MC portal use, we always configure with
+ * DQRR_MF is (SWP_CFG,20,3) - DQRR max fill (<- 0x4)
+ * EST is (SWP_CFG,16,3) - EQCR_CI stashing threshold (<- 0x0)
+ * RPM is (SWP_CFG,12,2) - RCR production notification mode (<- 0x3)
+ * DCM is (SWP_CFG,10,2) - DQRR consumption notification mode (<- 0x2)
+ * EPM is (SWP_CFG,8,2) - EQCR production notification mode (<- 0x3)
+ * SD is (SWP_CFG,5,1) - memory stashing drop enable (<- FALSE)
+ * SP is (SWP_CFG,4,1) - memory stashing priority (<- TRUE)
+ * SE is (SWP_CFG,3,1) - memory stashing enable (<- 0x0)
+ * DP is (SWP_CFG,2,1) - dequeue stashing priority (<- TRUE)
+ * DE is (SWP_CFG,1,1) - dequeue stashing enable (<- 0x0)
+ * EP is (SWP_CFG,0,1) - EQCR_CI stashing priority (<- FALSE)
+ */
+static inline uint32_t qbman_set_swp_cfg(uint8_t max_fill, uint8_t wn,
+					uint8_t est, uint8_t rpm, uint8_t dcm,
+					uint8_t epm, int sd, int sp, int se,
+					int dp, int de, int ep)
+{
+	uint32_t reg;
+
+	reg = e32_uint8_t(20, 3, max_fill) | e32_uint8_t(16, 3, est) |
+		e32_uint8_t(12, 2, rpm) | e32_uint8_t(10, 2, dcm) |
+		e32_uint8_t(8, 2, epm) | e32_int(5, 1, sd) |
+		e32_int(4, 1, sp) | e32_int(3, 1, se) | e32_int(2, 1, dp) |
+		e32_int(1, 1, de) | e32_int(0, 1, ep) |	e32_uint8_t(14, 1, wn);
+	return reg;
+}
+
+static inline int qbman_swp_sys_init(struct qbman_swp_sys *s,
+				     const struct qbman_swp_desc *d)
+{
+	uint32_t reg;
+
+	s->addr_cena = d->cena_bar;
+	s->addr_cinh = d->cinh_bar;
+	s->cena = (void *)valloc(CONFIG_SYS_PAGE_SIZE);
+	memset((void *)s->cena, 0x00, CONFIG_SYS_PAGE_SIZE);
+	if (!s->cena) {
+		printf("Could not allocate page for cena shadow\n");
+		return -1;
+	}
+
+#ifdef QBMAN_CHECKING
+	/* We should never be asked to initialise for a portal that isn't in
+	 * the power-on state. (Ie. don't forget to reset portals when they are
+	 * decommissioned!)
+	 */
+	reg = qbman_cinh_read(s, QBMAN_CINH_SWP_CFG);
+	BUG_ON(reg);
+#endif
+#ifdef QBMAN_CINH_ONLY
+	reg = qbman_set_swp_cfg(4, 1, 0, 3, 2, 3, 0, 1, 0, 1, 0, 0);
+#else
+	reg = qbman_set_swp_cfg(4, 0, 0, 3, 2, 3, 0, 1, 0, 1, 0, 0);
+#endif
+	qbman_cinh_write(s, QBMAN_CINH_SWP_CFG, reg);
+	reg = qbman_cinh_read(s, QBMAN_CINH_SWP_CFG);
+	if (!reg) {
+		printf("The portal is not enabled!\n");
+		free(s->cena);
+		return -1;
+	}
+	return 0;
+}
+
+static inline void qbman_swp_sys_finish(struct qbman_swp_sys *s)
+{
+	free((void *)s->cena);
+}