drivers: fsl-mc: Update qbman driver

Update qbman driver
 - As per latest available qbman driver
 - Use of atomic APIs

Signed-off-by: Prabhakar Kushwaha <prabhakar@freescale.com>
CC: Geoff Thorpe <Geoff.Thorpe@freescale.com>
CC: Haiying Wang <Haiying.Wang@freescale.com>
CC: Roy Pledge <Roy.Pledge@freescale.com>
Reviewed-by: York Sun <yorksun@freescale.com>

3 files changed, 53 insertions, 37 deletions

diff --git a/drivers/net/fsl-mc/dpio/qbman_portal.c b/drivers/net/fsl-mc/dpio/qbman_portal.c
index dd2a7de..5fa8d95 100644
--- a/drivers/net/fsl-mc/dpio/qbman_portal.c
+++ b/drivers/net/fsl-mc/dpio/qbman_portal.c
@@ -64,7 +64,7 @@ enum qbman_sdqcr_fc {
 struct qbman_swp *qbman_swp_init(const struct qbman_swp_desc *d)
 {
 	int ret;
-	struct qbman_swp *p = kmalloc(sizeof(*p), GFP_KERNEL);
+	struct qbman_swp *p = malloc(sizeof(struct qbman_swp));
 
 	if (!p)
 		return NULL;
@@ -77,7 +77,7 @@ struct qbman_swp *qbman_swp_init(const struct qbman_swp_desc *d)
 	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 */
+	atomic_set(&p->vdq.busy, 1);
 	p->vdq.valid_bit = QB_VALID_BIT;
 	p->dqrr.next_idx = 0;
 	p->dqrr.valid_bit = QB_VALID_BIT;
@@ -165,7 +165,6 @@ 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 */
@@ -197,8 +196,7 @@ void qbman_eq_desc_set_response(struct qbman_eq_desc *d,
 {
 	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_64(&code_eq_rsp_lo, (uint64_t *)cl, storage_phys);
 	qb_attr_code_encode(&code_eq_rsp_stash, cl, !!stash);
 }
 
@@ -253,7 +251,6 @@ 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,
@@ -282,8 +279,7 @@ void qbman_pull_desc_set_storage(struct qbman_pull_desc *d,
 	}
 	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));
+	qb_attr_code_encode_64(&code_pull_rsp_lo, (uint64_t *)cl, storage_phys);
 }
 
 void qbman_pull_desc_set_numframes(struct qbman_pull_desc *d, uint8_t numframes)
@@ -316,10 +312,10 @@ 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)
+	if (!atomic_dec_and_test(&s->vdq.busy)) {
+		atomic_inc(&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);
@@ -359,36 +355,44 @@ 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);
+	uint32_t flags;
+	const struct ldpaa_dq *dq;
+	const uint32_t *p;
 
+	dq = qbman_cena_read(&s->sys, QBMAN_CENA_SWP_DQRR(s->dqrr.next_idx));
+	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 the valid-bit isn't of the expected polarity, nothing there. Note,
+	 * in the DQRR reset bug workaround, we shouldn't need to skip these
+	 * check, because we've already determined that a new entry is available
+	 * and we've invalidated the cacheline before reading it, so the
+	 * valid-bit behaviour is repaired and should tell us what we already
+	 * knew from reading PI.
+	 */
 	if ((verb & QB_VALID_BIT) != s->dqrr.valid_bit) {
 		qbman_cena_invalidate_prefetch(&s->sys,
-					       QBMAN_CENA_SWP_DQRR(
-					       s->dqrr.next_idx));
+					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 */
+	s->dqrr.next_idx &= QBMAN_DQRR_SIZE - 1; /* 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;
-	}
+
+	/* If this is the final response to a volatile dequeue command
+	   indicate that the vdq is no longer busy */
+	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) &&
+	    (flags & LDPAA_DQ_STAT_EXPIRED))
+			atomic_inc(&s->vdq.busy);
+
 	qbman_cena_invalidate_prefetch(&s->sys,
 				       QBMAN_CENA_SWP_DQRR(s->dqrr.next_idx));
 	return dq;
@@ -448,8 +452,10 @@ int qbman_dq_entry_has_newtoken(struct qbman_swp *s,
 	 * 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;
+	if (s->vdq.storage == dq) {
+		s->vdq.storage = NULL;
+			atomic_inc(&s->vdq.busy);
+	}
 	return 1;
 }
 
diff --git a/drivers/net/fsl-mc/dpio/qbman_portal.h b/drivers/net/fsl-mc/dpio/qbman_portal.h
index bb67c3b..86e2c3a 100644
--- a/drivers/net/fsl-mc/dpio/qbman_portal.h
+++ b/drivers/net/fsl-mc/dpio/qbman_portal.h
@@ -14,6 +14,10 @@
 /* Management command result codes */
 #define QBMAN_MC_RSLT_OK      0xf0
 
+/* TBD: as of QBMan 4.1, DQRR will be 8 rather than 4! */
+#define QBMAN_DQRR_SIZE 4
+
+
 /* --------------------- */
 /* portal data structure */
 /* --------------------- */
@@ -48,14 +52,13 @@ struct qbman_swp {
 		 * 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;
+		 * command is executing, "vdq" is no longer "busy".
+		 */
+		atomic_t 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
+		 * targeting 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
@@ -127,6 +130,7 @@ static inline uint32_t qb_attr_code_decode(const struct qb_attr_code *code,
 	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)
@@ -136,6 +140,12 @@ static inline void qb_attr_code_encode(const struct qb_attr_code *code,
 		| e32_uint32_t(code->lsoffset, code->width, val);
 }
 
+static inline void qb_attr_code_encode_64(const struct qb_attr_code *code,
+				       uint64_t *cacheline, uint64_t val)
+{
+	cacheline[code->word / 2] = val;
+}
+
 /* ---------------------- */
 /* Descriptors/cachelines */
 /* ---------------------- */
@@ -144,7 +154,7 @@ static inline void qb_attr_code_encode(const struct qb_attr_code *code,
  * 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
+ * holding pre-formatted pieces of hardware 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
diff --git a/drivers/net/fsl-mc/dpio/qbman_private.h b/drivers/net/fsl-mc/dpio/qbman_private.h
index 2d2556b..f1f16b8 100644
--- a/drivers/net/fsl-mc/dpio/qbman_private.h
+++ b/drivers/net/fsl-mc/dpio/qbman_private.h
@@ -9,7 +9,7 @@
 #include <errno.h>
 #include <asm/io.h>
 #include <linux/types.h>
-#include <linux/compat.h>
+#include <asm/atomic.h>
 #include <malloc.h>
 #include <fsl-mc/fsl_qbman_base.h>