octeontx2-pf: TC_MATCHALL ingress ratelimiting offload
authorSunil Goutham <sgoutham@marvell.com>
Tue, 15 Jun 2021 11:34:29 +0000 (17:04 +0530)
committerDavid S. Miller <davem@davemloft.net>
Tue, 15 Jun 2021 18:11:05 +0000 (11:11 -0700)
Add TC_MATCHALL ingress ratelimiting offload support with POLICE
action for entire traffic coming into the interface.

Eg: To ratelimit ingress traffic to 100Mbps

$ ethtool -K eth0 hw-tc-offload on
$ tc qdisc add dev eth0 clsact
$ tc filter add dev eth0 ingress matchall skip_sw \
                action police rate 100Mbit burst 32Kbit

To support this, a leaf level bandwidth profile is allocated and all
RQs' contexts used by this interface are updated to point to it.
And the leaf level bandwidth profile is configured with user specified
rate and burst sizes.

Co-developed-by: Subbaraya Sundeep <sbhatta@marvell.com>
Signed-off-by: Subbaraya Sundeep <sbhatta@marvell.com>
Signed-off-by: Sunil Goutham <sgoutham@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/marvell/octeontx2/nic/cn10k.c
drivers/net/ethernet/marvell/octeontx2/nic/cn10k.h
drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c
drivers/net/ethernet/marvell/octeontx2/nic/otx2_tc.c

index 9ec0313f13fc5719281d19b625a1f4c19a281ac4..1b08896b46d2ec52be901e3ab366951fc4714949 100644 (file)
@@ -179,3 +179,326 @@ void cn10k_sqe_flush(void *dev, struct otx2_snd_queue *sq, int size, int qidx)
        sq->head++;
        sq->head &= (sq->sqe_cnt - 1);
 }
+
+int cn10k_free_all_ipolicers(struct otx2_nic *pfvf)
+{
+       struct nix_bandprof_free_req *req;
+       int rc;
+
+       if (is_dev_otx2(pfvf->pdev))
+               return 0;
+
+       mutex_lock(&pfvf->mbox.lock);
+
+       req = otx2_mbox_alloc_msg_nix_bandprof_free(&pfvf->mbox);
+       if (!req) {
+               rc =  -ENOMEM;
+               goto out;
+       }
+
+       /* Free all bandwidth profiles allocated */
+       req->free_all = true;
+
+       rc = otx2_sync_mbox_msg(&pfvf->mbox);
+out:
+       mutex_unlock(&pfvf->mbox.lock);
+       return rc;
+}
+
+int cn10k_alloc_leaf_profile(struct otx2_nic *pfvf, u16 *leaf)
+{
+       struct nix_bandprof_alloc_req *req;
+       struct nix_bandprof_alloc_rsp *rsp;
+       int rc;
+
+       req = otx2_mbox_alloc_msg_nix_bandprof_alloc(&pfvf->mbox);
+       if (!req)
+               return  -ENOMEM;
+
+       req->prof_count[BAND_PROF_LEAF_LAYER] = 1;
+
+       rc = otx2_sync_mbox_msg(&pfvf->mbox);
+       if (rc)
+               goto out;
+
+       rsp = (struct  nix_bandprof_alloc_rsp *)
+              otx2_mbox_get_rsp(&pfvf->mbox.mbox, 0, &req->hdr);
+       if (!rsp->prof_count[BAND_PROF_LEAF_LAYER]) {
+               rc = -EIO;
+               goto out;
+       }
+
+       *leaf = rsp->prof_idx[BAND_PROF_LEAF_LAYER][0];
+out:
+       if (rc) {
+               dev_warn(pfvf->dev,
+                        "Failed to allocate ingress bandwidth policer\n");
+       }
+
+       return rc;
+}
+
+int cn10k_alloc_matchall_ipolicer(struct otx2_nic *pfvf)
+{
+       struct otx2_hw *hw = &pfvf->hw;
+       int ret;
+
+       mutex_lock(&pfvf->mbox.lock);
+
+       ret = cn10k_alloc_leaf_profile(pfvf, &hw->matchall_ipolicer);
+
+       mutex_unlock(&pfvf->mbox.lock);
+
+       return ret;
+}
+
+#define POLICER_TIMESTAMP        1  /* 1 second */
+#define MAX_RATE_EXP             22 /* Valid rate exponent range: 0 - 22 */
+
+static void cn10k_get_ingress_burst_cfg(u32 burst, u32 *burst_exp,
+                                       u32 *burst_mantissa)
+{
+       int tmp;
+
+       /* Burst is calculated as
+        * (1+[BURST_MANTISSA]/256)*2^[BURST_EXPONENT]
+        * This is the upper limit on number tokens (bytes) that
+        * can be accumulated in the bucket.
+        */
+       *burst_exp = ilog2(burst);
+       if (burst < 256) {
+               /* No float: can't express mantissa in this case */
+               *burst_mantissa = 0;
+               return;
+       }
+
+       if (*burst_exp > MAX_RATE_EXP)
+               *burst_exp = MAX_RATE_EXP;
+
+       /* Calculate mantissa
+        * Find remaining bytes 'burst - 2^burst_exp'
+        * mantissa = (remaining bytes) / 2^ (burst_exp - 8)
+        */
+       tmp = burst - rounddown_pow_of_two(burst);
+       *burst_mantissa = tmp / (1UL << (*burst_exp - 8));
+}
+
+static void cn10k_get_ingress_rate_cfg(u64 rate, u32 *rate_exp,
+                                      u32 *rate_mantissa, u32 *rdiv)
+{
+       u32 div = 0;
+       u32 exp = 0;
+       u64 tmp;
+
+       /* Figure out mantissa, exponent and divider from given max pkt rate
+        *
+        * To achieve desired rate HW adds
+        * (1+[RATE_MANTISSA]/256)*2^[RATE_EXPONENT] tokens (bytes) at every
+        * policer timeunit * 2^rdiv ie 2 * 2^rdiv usecs, to the token bucket.
+        * Here policer timeunit is 2 usecs and rate is in bits per sec.
+        * Since floating point cannot be used below algorithm uses 1000000
+        * scale factor to support rates upto 100Gbps.
+        */
+       tmp = rate * 32 * 2;
+       if (tmp < 256000000) {
+               while (tmp < 256000000) {
+                       tmp = tmp * 2;
+                       div++;
+               }
+       } else {
+               for (exp = 0; tmp >= 512000000 && exp <= MAX_RATE_EXP; exp++)
+                       tmp = tmp / 2;
+
+               if (exp > MAX_RATE_EXP)
+                       exp = MAX_RATE_EXP;
+       }
+
+       *rate_mantissa = (tmp - 256000000) / 1000000;
+       *rate_exp = exp;
+       *rdiv = div;
+}
+
+int cn10k_map_unmap_rq_policer(struct otx2_nic *pfvf, int rq_idx,
+                              u16 policer, bool map)
+{
+       struct nix_cn10k_aq_enq_req *aq;
+
+       aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
+       if (!aq)
+               return -ENOMEM;
+
+       /* Enable policing and set the bandwidth profile (policer) index */
+       if (map)
+               aq->rq.policer_ena = 1;
+       else
+               aq->rq.policer_ena = 0;
+       aq->rq_mask.policer_ena = 1;
+
+       aq->rq.band_prof_id = policer;
+       aq->rq_mask.band_prof_id = GENMASK(9, 0);
+
+       /* Fill AQ info */
+       aq->qidx = rq_idx;
+       aq->ctype = NIX_AQ_CTYPE_RQ;
+       aq->op = NIX_AQ_INSTOP_WRITE;
+
+       return otx2_sync_mbox_msg(&pfvf->mbox);
+}
+
+int cn10k_free_leaf_profile(struct otx2_nic *pfvf, u16 leaf)
+{
+       struct nix_bandprof_free_req *req;
+
+       req = otx2_mbox_alloc_msg_nix_bandprof_free(&pfvf->mbox);
+       if (!req)
+               return -ENOMEM;
+
+       req->prof_count[BAND_PROF_LEAF_LAYER] = 1;
+       req->prof_idx[BAND_PROF_LEAF_LAYER][0] = leaf;
+
+       return otx2_sync_mbox_msg(&pfvf->mbox);
+}
+
+int cn10k_free_matchall_ipolicer(struct otx2_nic *pfvf)
+{
+       struct otx2_hw *hw = &pfvf->hw;
+       int qidx, rc;
+
+       mutex_lock(&pfvf->mbox.lock);
+
+       /* Remove RQ's policer mapping */
+       for (qidx = 0; qidx < hw->rx_queues; qidx++)
+               cn10k_map_unmap_rq_policer(pfvf, qidx,
+                                          hw->matchall_ipolicer, false);
+
+       rc = cn10k_free_leaf_profile(pfvf, hw->matchall_ipolicer);
+
+       mutex_unlock(&pfvf->mbox.lock);
+       return rc;
+}
+
+int cn10k_set_ipolicer_rate(struct otx2_nic *pfvf, u16 profile,
+                           u32 burst, u64 rate, bool pps)
+{
+       struct nix_cn10k_aq_enq_req *aq;
+       u32 burst_exp, burst_mantissa;
+       u32 rate_exp, rate_mantissa;
+       u32 rdiv;
+
+       /* Get exponent and mantissa values for the desired rate */
+       cn10k_get_ingress_burst_cfg(burst, &burst_exp, &burst_mantissa);
+       cn10k_get_ingress_rate_cfg(rate, &rate_exp, &rate_mantissa, &rdiv);
+
+       /* Init bandwidth profile */
+       aq = otx2_mbox_alloc_msg_nix_cn10k_aq_enq(&pfvf->mbox);
+       if (!aq)
+               return -ENOMEM;
+
+       /* Set initial color mode to blind */
+       aq->prof.icolor = 0x03;
+       aq->prof_mask.icolor = 0x03;
+
+       /* Set rate and burst values */
+       aq->prof.cir_exponent = rate_exp;
+       aq->prof_mask.cir_exponent = 0x1F;
+
+       aq->prof.cir_mantissa = rate_mantissa;
+       aq->prof_mask.cir_mantissa = 0xFF;
+
+       aq->prof.cbs_exponent = burst_exp;
+       aq->prof_mask.cbs_exponent = 0x1F;
+
+       aq->prof.cbs_mantissa = burst_mantissa;
+       aq->prof_mask.cbs_mantissa = 0xFF;
+
+       aq->prof.rdiv = rdiv;
+       aq->prof_mask.rdiv = 0xF;
+
+       if (pps) {
+               /* The amount of decremented tokens is calculated according to
+                * the following equation:
+                * max([ LMODE ? 0 : (packet_length - LXPTR)] +
+                *           ([ADJUST_MANTISSA]/256 - 1) * 2^[ADJUST_EXPONENT],
+                *      1/256)
+                * if LMODE is 1 then rate limiting will be based on
+                * PPS otherwise bps.
+                * The aim of the ADJUST value is to specify a token cost per
+                * packet in contrary to the packet length that specifies a
+                * cost per byte. To rate limit based on PPS adjust mantissa
+                * is set as 384 and exponent as 1 so that number of tokens
+                * decremented becomes 1 i.e, 1 token per packeet.
+                */
+               aq->prof.adjust_exponent = 1;
+               aq->prof_mask.adjust_exponent = 0x1F;
+
+               aq->prof.adjust_mantissa = 384;
+               aq->prof_mask.adjust_mantissa = 0x1FF;
+
+               aq->prof.lmode = 0x1;
+               aq->prof_mask.lmode = 0x1;
+       }
+
+       /* Two rate three color marker
+        * With PEIR/EIR set to zero, color will be either green or red
+        */
+       aq->prof.meter_algo = 2;
+       aq->prof_mask.meter_algo = 0x3;
+
+       aq->prof.rc_action = NIX_RX_BAND_PROF_ACTIONRESULT_DROP;
+       aq->prof_mask.rc_action = 0x3;
+
+       aq->prof.yc_action = NIX_RX_BAND_PROF_ACTIONRESULT_PASS;
+       aq->prof_mask.yc_action = 0x3;
+
+       aq->prof.gc_action = NIX_RX_BAND_PROF_ACTIONRESULT_PASS;
+       aq->prof_mask.gc_action = 0x3;
+
+       /* Setting exponent value as 24 and mantissa as 0 configures
+        * the bucket with zero values making bucket unused. Peak
+        * information rate and Excess information rate buckets are
+        * unused here.
+        */
+       aq->prof.peir_exponent = 24;
+       aq->prof_mask.peir_exponent = 0x1F;
+
+       aq->prof.peir_mantissa = 0;
+       aq->prof_mask.peir_mantissa = 0xFF;
+
+       aq->prof.pebs_exponent = 24;
+       aq->prof_mask.pebs_exponent = 0x1F;
+
+       aq->prof.pebs_mantissa = 0;
+       aq->prof_mask.pebs_mantissa = 0xFF;
+
+       /* Fill AQ info */
+       aq->qidx = profile;
+       aq->ctype = NIX_AQ_CTYPE_BANDPROF;
+       aq->op = NIX_AQ_INSTOP_WRITE;
+
+       return otx2_sync_mbox_msg(&pfvf->mbox);
+}
+
+int cn10k_set_matchall_ipolicer_rate(struct otx2_nic *pfvf,
+                                    u32 burst, u64 rate)
+{
+       struct otx2_hw *hw = &pfvf->hw;
+       int qidx, rc;
+
+       mutex_lock(&pfvf->mbox.lock);
+
+       rc = cn10k_set_ipolicer_rate(pfvf, hw->matchall_ipolicer, burst,
+                                    rate, false);
+       if (rc)
+               goto out;
+
+       for (qidx = 0; qidx < hw->rx_queues; qidx++) {
+               rc = cn10k_map_unmap_rq_policer(pfvf, qidx,
+                                               hw->matchall_ipolicer, true);
+               if (rc)
+                       break;
+       }
+
+out:
+       mutex_unlock(&pfvf->mbox.lock);
+       return rc;
+}
index e0bc595cbb7893c11bc998c7d447f9ebede4bf09..71292a4cf1f3cc037af2c3ef56dec716506f8ce0 100644 (file)
@@ -14,4 +14,15 @@ 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);
 int cn10k_pf_lmtst_init(struct otx2_nic *pf);
 int cn10k_vf_lmtst_init(struct otx2_nic *vf);
+int cn10k_free_all_ipolicers(struct otx2_nic *pfvf);
+int cn10k_alloc_matchall_ipolicer(struct otx2_nic *pfvf);
+int cn10k_free_matchall_ipolicer(struct otx2_nic *pfvf);
+int cn10k_set_matchall_ipolicer_rate(struct otx2_nic *pfvf,
+                                    u32 burst, u64 rate);
+int cn10k_map_unmap_rq_policer(struct otx2_nic *pfvf, int rq_idx,
+                              u16 policer, bool map);
+int cn10k_alloc_leaf_profile(struct otx2_nic *pfvf, u16 *leaf);
+int cn10k_set_ipolicer_rate(struct otx2_nic *pfvf, u16 profile,
+                           u32 burst, u64 rate, bool pps);
+int cn10k_free_leaf_profile(struct otx2_nic *pfvf, u16 leaf);
 #endif /* CN10K_H */
index e5616d466236df34b0dc3e4b95bfae5257baeb32..e0a3e28f556837407373e6007fcf644cd8ef3679 100644 (file)
@@ -180,6 +180,7 @@ struct otx2_hw {
 
        /* NIX */
        u16             txschq_list[NIX_TXSCH_LVL_CNT][MAX_TXSCHQ_PER_FUNC];
+       u16                     matchall_ipolicer;
 
        /* HW settings, coalescing etc */
        u16                     rx_chan_base;
@@ -327,6 +328,7 @@ struct otx2_nic {
 #define OTX2_FLAG_TX_PAUSE_ENABLED             BIT_ULL(10)
 #define OTX2_FLAG_TC_FLOWER_SUPPORT            BIT_ULL(11)
 #define OTX2_FLAG_TC_MATCHALL_EGRESS_ENABLED   BIT_ULL(12)
+#define OTX2_FLAG_TC_MATCHALL_INGRESS_ENABLED  BIT_ULL(13)
        u64                     flags;
 
        struct otx2_qset        qset;
index 65f505b07b5d9b95451990ee6e8caa554074d351..59912f73417b8f31dfc03d0d62ac2a108ecaf261 100644 (file)
@@ -1461,6 +1461,9 @@ static void otx2_free_hw_resources(struct otx2_nic *pf)
 
        otx2_free_cq_res(pf);
 
+       /* Free all ingress bandwidth profiles allocated */
+       cn10k_free_all_ipolicers(pf);
+
        mutex_lock(&mbox->lock);
        /* Reset NIX LF */
        free_req = otx2_mbox_alloc_msg_nix_lf_free(mbox);
index 26712c091c6365171efbf7848233a8d3ee924aed..5767fa4ef2050f651380133df7a808f27ee1f450 100644 (file)
@@ -15,6 +15,7 @@
 #include <net/tc_act/tc_vlan.h>
 #include <net/ipv6.h>
 
+#include "cn10k.h"
 #include "otx2_common.h"
 
 /* Egress rate limiting definitions */
@@ -675,6 +676,87 @@ static int otx2_setup_tc_cls_flower(struct otx2_nic *nic,
        }
 }
 
+static int otx2_tc_ingress_matchall_install(struct otx2_nic *nic,
+                                           struct tc_cls_matchall_offload *cls)
+{
+       struct netlink_ext_ack *extack = cls->common.extack;
+       struct flow_action *actions = &cls->rule->action;
+       struct flow_action_entry *entry;
+       u64 rate;
+       int err;
+
+       err = otx2_tc_validate_flow(nic, actions, extack);
+       if (err)
+               return err;
+
+       if (nic->flags & OTX2_FLAG_TC_MATCHALL_INGRESS_ENABLED) {
+               NL_SET_ERR_MSG_MOD(extack,
+                                  "Only one ingress MATCHALL ratelimitter can be offloaded");
+               return -ENOMEM;
+       }
+
+       entry = &cls->rule->action.entries[0];
+       switch (entry->id) {
+       case FLOW_ACTION_POLICE:
+               /* Ingress ratelimiting is not supported on OcteonTx2 */
+               if (is_dev_otx2(nic->pdev)) {
+                       NL_SET_ERR_MSG_MOD(extack,
+                                          "Ingress policing not supported on this platform");
+                       return -EOPNOTSUPP;
+               }
+
+               err = cn10k_alloc_matchall_ipolicer(nic);
+               if (err)
+                       return err;
+
+               /* Convert to bits per second */
+               rate = entry->police.rate_bytes_ps * 8;
+               err = cn10k_set_matchall_ipolicer_rate(nic, entry->police.burst, rate);
+               if (err)
+                       return err;
+               nic->flags |= OTX2_FLAG_TC_MATCHALL_INGRESS_ENABLED;
+               break;
+       default:
+               NL_SET_ERR_MSG_MOD(extack,
+                                  "Only police action supported with Ingress MATCHALL offload");
+               return -EOPNOTSUPP;
+       }
+
+       return 0;
+}
+
+static int otx2_tc_ingress_matchall_delete(struct otx2_nic *nic,
+                                          struct tc_cls_matchall_offload *cls)
+{
+       struct netlink_ext_ack *extack = cls->common.extack;
+       int err;
+
+       if (nic->flags & OTX2_FLAG_INTF_DOWN) {
+               NL_SET_ERR_MSG_MOD(extack, "Interface not initialized");
+               return -EINVAL;
+       }
+
+       err = cn10k_free_matchall_ipolicer(nic);
+       nic->flags &= ~OTX2_FLAG_TC_MATCHALL_INGRESS_ENABLED;
+       return err;
+}
+
+static int otx2_setup_tc_ingress_matchall(struct otx2_nic *nic,
+                                         struct tc_cls_matchall_offload *cls_matchall)
+{
+       switch (cls_matchall->command) {
+       case TC_CLSMATCHALL_REPLACE:
+               return otx2_tc_ingress_matchall_install(nic, cls_matchall);
+       case TC_CLSMATCHALL_DESTROY:
+               return otx2_tc_ingress_matchall_delete(nic, cls_matchall);
+       case TC_CLSMATCHALL_STATS:
+       default:
+               break;
+       }
+
+       return -EOPNOTSUPP;
+}
+
 static int otx2_setup_tc_block_ingress_cb(enum tc_setup_type type,
                                          void *type_data, void *cb_priv)
 {
@@ -686,6 +768,8 @@ static int otx2_setup_tc_block_ingress_cb(enum tc_setup_type type,
        switch (type) {
        case TC_SETUP_CLSFLOWER:
                return otx2_setup_tc_cls_flower(nic, type_data);
+       case TC_SETUP_CLSMATCHALL:
+               return otx2_setup_tc_ingress_matchall(nic, type_data);
        default:
                break;
        }