Merge branch 'octeon-drr-config'

Sunil Goutham says:

====================
cn10k: DWRR MTU and weights configuration

On OcteonTx2 DWRR quantum is directly configured into each of
the transmit scheduler queues. And PF/VF drivers were free to
config any value upto 2^24.

On CN10K, HW is modified, the quantum configuration at scheduler
queues is in terms of weight. And SW needs to setup a base DWRR MTU
at NIX_AF_DWRR_RPM_MTU / NIX_AF_DWRR_SDP_MTU. HW will do
'DWRR MTU * weight' to get the quantum.

This patch series addresses this HW change on CN10K silicons,
both admin function and PF/VF drivers are modified.

Also added support to program DWRR MTU via devlink params.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/marvell/octeontx2/af/common.h

@@ -146,10 +146,7 @@ enum nix_scheduler {
 #define TXSCH_RR_QTM_MAX		((1 << 24) - 1)
 #define TXSCH_TL1_DFLT_RR_QTM		TXSCH_RR_QTM_MAX
 #define TXSCH_TL1_DFLT_RR_PRIO		(0x1ull)
-#define MAX_SCHED_WEIGHT		0xFF
-#define DFLT_RR_WEIGHT			71
-#define DFLT_RR_QTM	((DFLT_RR_WEIGHT * TXSCH_RR_QTM_MAX) \
-			 / MAX_SCHED_WEIGHT)
+#define CN10K_MAX_DWRR_WEIGHT          16384 /* Weight is 14bit on CN10K */
 
 /* Min/Max packet sizes, excluding FCS */
 #define	NIC_HW_MIN_FRS			40

drivers/net/ethernet/marvell/octeontx2/af/mbox.h

@@ -1032,8 +1032,12 @@ struct nix_bp_cfg_rsp {
 
 struct nix_hw_info {
 	struct mbox_msghdr hdr;
+	u16 rsvs16;
 	u16 max_mtu;
 	u16 min_mtu;
+	u32 rpm_dwrr_mtu;
+	u32 sdp_dwrr_mtu;
+	u64 rsvd[16]; /* Add reserved fields for future expansion */
 };
 
 struct nix_bandprof_alloc_req {

drivers/net/ethernet/marvell/octeontx2/af/rvu.h

@@ -329,6 +329,7 @@ struct hw_cap {
 	bool	nix_shaping;		 /* Is shaping and coloring supported */
 	bool	nix_tx_link_bp;		 /* Can link backpressure TL queues ? */
 	bool	nix_rx_multicast;	 /* Rx packet replication support */
+	bool	nix_common_dwrr_mtu;	 /* Common DWRR MTU for quantum config */
 	bool	per_pf_mbox_regs; /* PF mbox specified in per PF registers ? */
 	bool	programmable_chans; /* Channels programmable ? */
 	bool	ipolicer;
@@ -706,6 +707,8 @@ int nix_aq_context_read(struct rvu *rvu, struct nix_hw *nix_hw,
 			struct nix_cn10k_aq_enq_rsp *aq_rsp,
 			u16 pcifunc, u8 ctype, u32 qidx);
 int rvu_get_nix_blkaddr(struct rvu *rvu, u16 pcifunc);
+u32 convert_dwrr_mtu_to_bytes(u8 dwrr_mtu);
+u32 convert_bytes_to_dwrr_mtu(u32 bytes);
 
 /* NPC APIs */
 int rvu_npc_init(struct rvu *rvu);

drivers/net/ethernet/marvell/octeontx2/af/rvu_devlink.c

@@ -1364,6 +1364,89 @@ static void rvu_health_reporters_destroy(struct rvu *rvu)
 	rvu_nix_health_reporters_destroy(rvu_dl);
 }
 
+/* Devlink Params APIs */
+static int rvu_af_dl_dwrr_mtu_validate(struct devlink *devlink, u32 id,
+				       union devlink_param_value val,
+				       struct netlink_ext_ack *extack)
+{
+	struct rvu_devlink *rvu_dl = devlink_priv(devlink);
+	struct rvu *rvu = rvu_dl->rvu;
+	int dwrr_mtu = val.vu32;
+	struct nix_txsch *txsch;
+	struct nix_hw *nix_hw;
+
+	if (!rvu->hw->cap.nix_common_dwrr_mtu) {
+		NL_SET_ERR_MSG_MOD(extack,
+				   "Setting DWRR_MTU is not supported on this silicon");
+		return -EOPNOTSUPP;
+	}
+
+	if ((dwrr_mtu > 65536 || !is_power_of_2(dwrr_mtu)) &&
+	    (dwrr_mtu != 9728 && dwrr_mtu != 10240)) {
+		NL_SET_ERR_MSG_MOD(extack,
+				   "Invalid, supported MTUs are 0,2,4,8.16,32,64....4K,8K,32K,64K and 9728, 10240");
+		return -EINVAL;
+	}
+
+	nix_hw = get_nix_hw(rvu->hw, BLKADDR_NIX0);
+	if (!nix_hw)
+		return -ENODEV;
+
+	txsch = &nix_hw->txsch[NIX_TXSCH_LVL_SMQ];
+	if (rvu_rsrc_free_count(&txsch->schq) != txsch->schq.max) {
+		NL_SET_ERR_MSG_MOD(extack,
+				   "Changing DWRR MTU is not supported when there are active NIXLFs");
+		NL_SET_ERR_MSG_MOD(extack,
+				   "Makesure none of the PF/VF interfaces are initialized and retry");
+		return -EOPNOTSUPP;
+	}
+
+	return 0;
+}
+
+static int rvu_af_dl_dwrr_mtu_set(struct devlink *devlink, u32 id,
+				  struct devlink_param_gset_ctx *ctx)
+{
+	struct rvu_devlink *rvu_dl = devlink_priv(devlink);
+	struct rvu *rvu = rvu_dl->rvu;
+	u64 dwrr_mtu;
+
+	dwrr_mtu = convert_bytes_to_dwrr_mtu(ctx->val.vu32);
+	rvu_write64(rvu, BLKADDR_NIX0, NIX_AF_DWRR_RPM_MTU, dwrr_mtu);
+
+	return 0;
+}
+
+static int rvu_af_dl_dwrr_mtu_get(struct devlink *devlink, u32 id,
+				  struct devlink_param_gset_ctx *ctx)
+{
+	struct rvu_devlink *rvu_dl = devlink_priv(devlink);
+	struct rvu *rvu = rvu_dl->rvu;
+	u64 dwrr_mtu;
+
+	if (!rvu->hw->cap.nix_common_dwrr_mtu)
+		return -EOPNOTSUPP;
+
+	dwrr_mtu = rvu_read64(rvu, BLKADDR_NIX0, NIX_AF_DWRR_RPM_MTU);
+	ctx->val.vu32 = convert_dwrr_mtu_to_bytes(dwrr_mtu);
+
+	return 0;
+}
+
+enum rvu_af_dl_param_id {
+	RVU_AF_DEVLINK_PARAM_ID_BASE = DEVLINK_PARAM_GENERIC_ID_MAX,
+	RVU_AF_DEVLINK_PARAM_ID_DWRR_MTU,
+};
+
+static const struct devlink_param rvu_af_dl_params[] = {
+	DEVLINK_PARAM_DRIVER(RVU_AF_DEVLINK_PARAM_ID_DWRR_MTU,
+			     "dwrr_mtu", DEVLINK_PARAM_TYPE_U32,
+			     BIT(DEVLINK_PARAM_CMODE_RUNTIME),
+			     rvu_af_dl_dwrr_mtu_get, rvu_af_dl_dwrr_mtu_set,
+			     rvu_af_dl_dwrr_mtu_validate),
+};
+
+/* Devlink switch mode */
 static int rvu_devlink_eswitch_mode_get(struct devlink *devlink, u16 *mode)
 {
 	struct rvu_devlink *rvu_dl = devlink_priv(devlink);
@@ -1438,7 +1521,30 @@ int rvu_register_dl(struct rvu *rvu)
 	rvu_dl->rvu = rvu;
 	rvu->rvu_dl = rvu_dl;
 
-	return rvu_health_reporters_create(rvu);
+	err = rvu_health_reporters_create(rvu);
+	if (err) {
+		dev_err(rvu->dev,
+			"devlink health reporter creation failed with error %d\n", err);
+		goto err_dl_health;
+	}
+
+	err = devlink_params_register(dl, rvu_af_dl_params,
+				      ARRAY_SIZE(rvu_af_dl_params));
+	if (err) {
+		dev_err(rvu->dev,
+			"devlink params register failed with error %d", err);
+		goto err_dl_health;
+	}
+
+	devlink_params_publish(dl);
+
+	return 0;
+
+err_dl_health:
+	rvu_health_reporters_destroy(rvu);
+	devlink_unregister(dl);
+	devlink_free(dl);
+	return err;
 }
 
 void rvu_unregister_dl(struct rvu *rvu)
@@ -1449,6 +1555,8 @@ void rvu_unregister_dl(struct rvu *rvu)
 	if (!dl)
 		return;
 
+	devlink_params_unregister(dl, rvu_af_dl_params,
+				  ARRAY_SIZE(rvu_af_dl_params));
 	rvu_health_reporters_destroy(rvu);
 	devlink_unregister(dl);
 	devlink_free(dl);

drivers/net/ethernet/marvell/octeontx2/af/rvu_nix.c

@@ -192,6 +192,47 @@ struct nix_hw *get_nix_hw(struct rvu_hwinfo *hw, int blkaddr)
 	return NULL;
 }
 
+u32 convert_dwrr_mtu_to_bytes(u8 dwrr_mtu)
+{
+	dwrr_mtu &= 0x1FULL;
+
+	/* MTU used for DWRR calculation is in power of 2 up until 64K bytes.
+	 * Value of 4 is reserved for MTU value of 9728 bytes.
+	 * Value of 5 is reserved for MTU value of 10240 bytes.
+	 */
+	switch (dwrr_mtu) {
+	case 4:
+		return 9728;
+	case 5:
+		return 10240;
+	default:
+		return BIT_ULL(dwrr_mtu);
+	}
+
+	return 0;
+}
+
+u32 convert_bytes_to_dwrr_mtu(u32 bytes)
+{
+	/* MTU used for DWRR calculation is in power of 2 up until 64K bytes.
+	 * Value of 4 is reserved for MTU value of 9728 bytes.
+	 * Value of 5 is reserved for MTU value of 10240 bytes.
+	 */
+	if (bytes > BIT_ULL(16))
+		return 0;
+
+	switch (bytes) {
+	case 9728:
+		return 4;
+	case 10240:
+		return 5;
+	default:
+		return ilog2(bytes);
+	}
+
+	return 0;
+}
+
 static void nix_rx_sync(struct rvu *rvu, int blkaddr)
 {
 	int err;
@@ -1958,8 +1999,17 @@ static void nix_tl1_default_cfg(struct rvu *rvu, struct nix_hw *nix_hw,
 		return;
 	rvu_write64(rvu, blkaddr, NIX_AF_TL1X_TOPOLOGY(schq),
 		    (TXSCH_TL1_DFLT_RR_PRIO << 1));
-	rvu_write64(rvu, blkaddr, NIX_AF_TL1X_SCHEDULE(schq),
-		    TXSCH_TL1_DFLT_RR_QTM);
+
+	/* On OcteonTx2 the config was in bytes and newer silcons
+	 * it's changed to weight.
+	 */
+	if (!rvu->hw->cap.nix_common_dwrr_mtu)
+		rvu_write64(rvu, blkaddr, NIX_AF_TL1X_SCHEDULE(schq),
+			    TXSCH_TL1_DFLT_RR_QTM);
+	else
+		rvu_write64(rvu, blkaddr, NIX_AF_TL1X_SCHEDULE(schq),
+			    CN10K_MAX_DWRR_WEIGHT);
+
 	rvu_write64(rvu, blkaddr, NIX_AF_TL1X_CIR(schq), 0x00);
 	pfvf_map[schq] = TXSCH_SET_FLAG(pfvf_map[schq], NIX_TXSCHQ_CFG_DONE);
 }
@@ -2667,6 +2717,15 @@ static int nix_setup_txschq(struct rvu *rvu, struct nix_hw *nix_hw, int blkaddr)
 		for (schq = 0; schq < txsch->schq.max; schq++)
 			txsch->pfvf_map[schq] = TXSCH_MAP(0, NIX_TXSCHQ_FREE);
 	}
+
+	/* Setup a default value of 8192 as DWRR MTU */
+	if (rvu->hw->cap.nix_common_dwrr_mtu) {
+		rvu_write64(rvu, blkaddr, NIX_AF_DWRR_RPM_MTU,
+			    convert_bytes_to_dwrr_mtu(8192));
+		rvu_write64(rvu, blkaddr, NIX_AF_DWRR_SDP_MTU,
+			    convert_bytes_to_dwrr_mtu(8192));
+	}
+
 	return 0;
 }
 
@@ -2743,6 +2802,7 @@ int rvu_mbox_handler_nix_get_hw_info(struct rvu *rvu, struct msg_req *req,
 				     struct nix_hw_info *rsp)
 {
 	u16 pcifunc = req->hdr.pcifunc;
+	u64 dwrr_mtu;
 	int blkaddr;
 
 	blkaddr = rvu_get_blkaddr(rvu, BLKTYPE_NIX, pcifunc);
@@ -2755,6 +2815,20 @@ int rvu_mbox_handler_nix_get_hw_info(struct rvu *rvu, struct msg_req *req,
 		rvu_get_lmac_link_max_frs(rvu, &rsp->max_mtu);
 
 	rsp->min_mtu = NIC_HW_MIN_FRS;
+
+	if (!rvu->hw->cap.nix_common_dwrr_mtu) {
+		/* Return '1' on OTx2 */
+		rsp->rpm_dwrr_mtu = 1;
+		rsp->sdp_dwrr_mtu = 1;
+		return 0;
+	}
+
+	dwrr_mtu = rvu_read64(rvu, BLKADDR_NIX0, NIX_AF_DWRR_RPM_MTU);
+	rsp->rpm_dwrr_mtu = convert_dwrr_mtu_to_bytes(dwrr_mtu);
+
+	dwrr_mtu = rvu_read64(rvu, BLKADDR_NIX0, NIX_AF_DWRR_SDP_MTU);
+	rsp->sdp_dwrr_mtu = convert_dwrr_mtu_to_bytes(dwrr_mtu);
+
 	return 0;
 }
 
@@ -3647,6 +3721,28 @@ static int nix_aq_init(struct rvu *rvu, struct rvu_block *block)
 	return 0;
 }
 
+static void rvu_nix_setup_capabilities(struct rvu *rvu, int blkaddr)
+{
+	struct rvu_hwinfo *hw = rvu->hw;
+	u64 hw_const;
+
+	hw_const = rvu_read64(rvu, blkaddr, NIX_AF_CONST1);
+
+	/* On OcteonTx2 DWRR quantum is directly configured into each of
+	 * the transmit scheduler queues. And PF/VF drivers were free to
+	 * config any value upto 2^24.
+	 * On CN10K, HW is modified, the quantum configuration at scheduler
+	 * queues is in terms of weight. And SW needs to setup a base DWRR MTU
+	 * at NIX_AF_DWRR_RPM_MTU / NIX_AF_DWRR_SDP_MTU. HW will do
+	 * 'DWRR MTU * weight' to get the quantum.
+	 *
+	 * Check if HW uses a common MTU for all DWRR quantum configs.
+	 * On OcteonTx2 this register field is '0'.
+	 */
+	if (((hw_const >> 56) & 0x10) == 0x10)
+		hw->cap.nix_common_dwrr_mtu = true;
+}
+
 static int rvu_nix_block_init(struct rvu *rvu, struct nix_hw *nix_hw)
 {
 	const struct npc_lt_def_cfg *ltdefs;
@@ -3684,6 +3780,9 @@ static int rvu_nix_block_init(struct rvu *rvu, struct nix_hw *nix_hw)
 	if (err)
 		return err;
 
+	/* Setup capabilities of the NIX block */
+	rvu_nix_setup_capabilities(rvu, blkaddr);
+
 	/* Initialize admin queue */
 	err = nix_aq_init(rvu, block);
 	if (err)

drivers/net/ethernet/marvell/octeontx2/af/rvu_reg.h

@@ -269,6 +269,8 @@
 #define NIX_AF_DEBUG_NPC_RESP_DATAX(a)          (0x680 | (a) << 3)
 #define NIX_AF_SMQX_CFG(a)                      (0x700 | (a) << 16)
 #define NIX_AF_SQM_DBG_CTL_STATUS               (0x750)
+#define NIX_AF_DWRR_SDP_MTU                     (0x790)
+#define NIX_AF_DWRR_RPM_MTU                     (0x7A0)
 #define NIX_AF_PSE_CHANNEL_LEVEL                (0x800)
 #define NIX_AF_PSE_SHAPER_CFG                   (0x810)
 #define NIX_AF_TX_EXPR_CREDIT			(0x830)

drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c

@@ -92,8 +92,7 @@ int cn10k_sq_aq_init(void *dev, u16 qidx, u16 sqb_aura)
 	aq->sq.ena = 1;
 	/* Only one SMQ is allocated, map all SQ's to that SMQ  */
 	aq->sq.smq = pfvf->hw.txschq_list[NIX_TXSCH_LVL_SMQ][0];
-	/* FIXME: set based on NIX_AF_DWRR_RPM_MTU*/
-	aq->sq.smq_rr_weight = pfvf->netdev->mtu;
+	aq->sq.smq_rr_weight = mtu_to_dwrr_weight(pfvf, pfvf->max_frs);
 	aq->sq.default_chan = pfvf->hw.tx_chan_base;
 	aq->sq.sqe_stype = NIX_STYPE_STF; /* Cache SQB */
 	aq->sq.sqb_aura = sqb_aura;

drivers/net/ethernet/marvell/octeontx2/nic/cn10k.h

@@ -9,6 +9,20 @@
 
 #include "otx2_common.h"
 
+static inline int mtu_to_dwrr_weight(struct otx2_nic *pfvf, int mtu)
+{
+	u32 weight;
+
+	/* On OTx2, since AF returns DWRR_MTU as '1', this logic
+	 * will work on those silicons as well.
+	 */
+	weight = mtu / pfvf->hw.dwrr_mtu;
+	if (mtu % pfvf->hw.dwrr_mtu)
+		weight += 1;
+
+	return weight;
+}
+
 void cn10k_refill_pool_ptrs(void *dev, struct otx2_cq_queue *cq);
 void cn10k_sqe_flush(void *dev, struct otx2_snd_queue *sq, int size, int qidx);
 int cn10k_sq_aq_init(void *dev, u16 qidx, u16 sqb_aura);

drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c

@@ -596,6 +596,9 @@ int otx2_txschq_config(struct otx2_nic *pfvf, int lvl)
 	struct otx2_hw *hw = &pfvf->hw;
 	struct nix_txschq_config *req;
 	u64 schq, parent;
+	u64 dwrr_val;
+
+	dwrr_val = mtu_to_dwrr_weight(pfvf, pfvf->max_frs);
 
 	req = otx2_mbox_alloc_msg_nix_txschq_cfg(&pfvf->mbox);
 	if (!req)
@@ -621,21 +624,21 @@ int otx2_txschq_config(struct otx2_nic *pfvf, int lvl)
 		req->num_regs++;
 		/* Set DWRR quantum */
 		req->reg[2] = NIX_AF_MDQX_SCHEDULE(schq);
-		req->regval[2] =  DFLT_RR_QTM;
+		req->regval[2] =  dwrr_val;
 	} else if (lvl == NIX_TXSCH_LVL_TL4) {
 		parent =  hw->txschq_list[NIX_TXSCH_LVL_TL3][0];
 		req->reg[0] = NIX_AF_TL4X_PARENT(schq);
 		req->regval[0] = parent << 16;
 		req->num_regs++;
 		req->reg[1] = NIX_AF_TL4X_SCHEDULE(schq);
-		req->regval[1] = DFLT_RR_QTM;
+		req->regval[1] = dwrr_val;
 	} else if (lvl == NIX_TXSCH_LVL_TL3) {
 		parent = hw->txschq_list[NIX_TXSCH_LVL_TL2][0];
 		req->reg[0] = NIX_AF_TL3X_PARENT(schq);
 		req->regval[0] = parent << 16;
 		req->num_regs++;
 		req->reg[1] = NIX_AF_TL3X_SCHEDULE(schq);
-		req->regval[1] = DFLT_RR_QTM;
+		req->regval[1] = dwrr_val;
 	} else if (lvl == NIX_TXSCH_LVL_TL2) {
 		parent =  hw->txschq_list[NIX_TXSCH_LVL_TL1][0];
 		req->reg[0] = NIX_AF_TL2X_PARENT(schq);
@@ -643,7 +646,7 @@ int otx2_txschq_config(struct otx2_nic *pfvf, int lvl)
 
 		req->num_regs++;
 		req->reg[1] = NIX_AF_TL2X_SCHEDULE(schq);
-		req->regval[1] = TXSCH_TL1_DFLT_RR_PRIO << 24 | DFLT_RR_QTM;
+		req->regval[1] = TXSCH_TL1_DFLT_RR_PRIO << 24 | dwrr_val;
 
 		req->num_regs++;
 		req->reg[2] = NIX_AF_TL3_TL2X_LINKX_CFG(schq,
@@ -656,7 +659,10 @@ int otx2_txschq_config(struct otx2_nic *pfvf, int lvl)
 		 * For VF this is always ignored.
 		 */
 
-		/* Set DWRR quantum */
+		/* On CN10K, if RR_WEIGHT is greater than 16384, HW will
+		 * clip it to 16384, so configuring a 24bit max value
+		 * will work on both OTx2 and CN10K.
+		 */
 		req->reg[0] = NIX_AF_TL1X_SCHEDULE(schq);
 		req->regval[0] = TXSCH_TL1_DFLT_RR_QTM;
 
@@ -803,7 +809,7 @@ int otx2_sq_aq_init(void *dev, u16 qidx, u16 sqb_aura)
 	aq->sq.ena = 1;
 	/* Only one SMQ is allocated, map all SQ's to that SMQ  */
 	aq->sq.smq = pfvf->hw.txschq_list[NIX_TXSCH_LVL_SMQ][0];
-	aq->sq.smq_rr_quantum = DFLT_RR_QTM;
+	aq->sq.smq_rr_quantum = mtu_to_dwrr_weight(pfvf, pfvf->max_frs);
 	aq->sq.default_chan = pfvf->hw.tx_chan_base;
 	aq->sq.sqe_stype = NIX_STYPE_STF; /* Cache SQB */
 	aq->sq.sqb_aura = sqb_aura;
@@ -1668,6 +1674,11 @@ u16 otx2_get_max_mtu(struct otx2_nic *pfvf)
 		 * SMQ errors
 		 */
 		max_mtu = rsp->max_mtu - 8 - OTX2_ETH_HLEN;
+
+		/* Also save DWRR MTU, needed for DWRR weight calculation */
+		pfvf->hw.dwrr_mtu = rsp->rpm_dwrr_mtu;
+		if (!pfvf->hw.dwrr_mtu)
+			pfvf->hw.dwrr_mtu = 1;
 	}
 
 out:

drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h

@@ -181,6 +181,7 @@ struct otx2_hw {
 	/* NIX */
 	u16		txschq_list[NIX_TXSCH_LVL_CNT][MAX_TXSCHQ_PER_FUNC];
 	u16			matchall_ipolicer;
+	u32			dwrr_mtu;
 
 	/* HW settings, coalescing etc */
 	u16			rx_chan_base;