nfp: bpf: refactor offload logic
authorJakub Kicinski <jakub.kicinski@netronome.com>
Fri, 3 Nov 2017 20:56:25 +0000 (13:56 -0700)
committerDavid S. Miller <davem@davemloft.net>
Sun, 5 Nov 2017 13:26:19 +0000 (22:26 +0900)
We currently create a fake cls_bpf offload object when we want
to offload XDP.  Simplify and clarify the code by moving the
TC/XDP specific logic out of common offload code.  This is easy
now that we don't support legacy TC actions.  We only need the
bpf program and state of the skip_sw flag.

Temporarily set @code to NULL in nfp_net_bpf_offload(), compilers
seem to have trouble recognizing it's always initialized.  Next
patches will eliminate that variable.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/netronome/nfp/bpf/main.c
drivers/net/ethernet/netronome/nfp/bpf/main.h
drivers/net/ethernet/netronome/nfp/bpf/offload.c

index 2ff97f12c160e869f00c1eb9c491cb8443ee3d78..9e1286346d425abbc684e90604bd88c7ae4ca964 100644 (file)
@@ -54,28 +54,25 @@ static int
 nfp_bpf_xdp_offload(struct nfp_app *app, struct nfp_net *nn,
                    struct bpf_prog *prog)
 {
-       struct tc_cls_bpf_offload cmd = {
-               .prog = prog,
-       };
+       bool running, xdp_running;
        int ret;
 
        if (!nfp_net_ebpf_capable(nn))
                return -EINVAL;
 
-       if (nn->dp.ctrl & NFP_NET_CFG_CTRL_BPF) {
-               if (!nn->dp.bpf_offload_xdp)
-                       return prog ? -EBUSY : 0;
-               cmd.command = prog ? TC_CLSBPF_REPLACE : TC_CLSBPF_DESTROY;
-       } else {
-               if (!prog)
-                       return 0;
-               cmd.command = TC_CLSBPF_ADD;
-       }
+       running = nn->dp.ctrl & NFP_NET_CFG_CTRL_BPF;
+       xdp_running = running && nn->dp.bpf_offload_xdp;
+
+       if (!prog && !xdp_running)
+               return 0;
+       if (prog && running && !xdp_running)
+               return -EBUSY;
 
-       ret = nfp_net_bpf_offload(nn, &cmd);
+       ret = nfp_net_bpf_offload(nn, prog, running, true);
        /* Stop offload if replace not possible */
-       if (ret && cmd.command == TC_CLSBPF_REPLACE)
+       if (ret && prog)
                nfp_bpf_xdp_offload(app, nn, NULL);
+
        nn->dp.bpf_offload_xdp = prog && !ret;
        return ret;
 }
@@ -96,27 +93,33 @@ static int nfp_bpf_setup_tc_block_cb(enum tc_setup_type type,
 {
        struct tc_cls_bpf_offload *cls_bpf = type_data;
        struct nfp_net *nn = cb_priv;
+       bool skip_sw;
+
+       if (type != TC_SETUP_CLSBPF ||
+           !tc_can_offload(nn->dp.netdev) ||
+           !nfp_net_ebpf_capable(nn) ||
+           cls_bpf->common.protocol != htons(ETH_P_ALL) ||
+           cls_bpf->common.chain_index)
+               return -EOPNOTSUPP;
+       if (nn->dp.bpf_offload_xdp)
+               return -EBUSY;
 
-       if (!tc_can_offload(nn->dp.netdev))
+       /* Only support TC direct action */
+       if (!cls_bpf->exts_integrated ||
+           tcf_exts_has_actions(cls_bpf->exts)) {
+               nn_err(nn, "only direct action with no legacy actions supported\n");
                return -EOPNOTSUPP;
+       }
 
-       switch (type) {
-       case TC_SETUP_CLSBPF:
-               if (!nfp_net_ebpf_capable(nn) ||
-                   cls_bpf->common.protocol != htons(ETH_P_ALL) ||
-                   cls_bpf->common.chain_index)
-                       return -EOPNOTSUPP;
-               if (nn->dp.bpf_offload_xdp)
-                       return -EBUSY;
-
-               /* Only support TC direct action */
-               if (!cls_bpf->exts_integrated ||
-                   tcf_exts_has_actions(cls_bpf->exts)) {
-                       nn_err(nn, "only direct action with no legacy actions supported\n");
-                       return -EOPNOTSUPP;
-               }
-
-               return nfp_net_bpf_offload(nn, cls_bpf);
+       skip_sw = !!(cls_bpf->gen_flags & TCA_CLS_FLAGS_SKIP_SW);
+
+       switch (cls_bpf->command) {
+       case TC_CLSBPF_REPLACE:
+               return nfp_net_bpf_offload(nn, cls_bpf->prog, true, !skip_sw);
+       case TC_CLSBPF_ADD:
+               return nfp_net_bpf_offload(nn, cls_bpf->prog, false, !skip_sw);
+       case TC_CLSBPF_DESTROY:
+               return nfp_net_bpf_offload(nn, NULL, true, !skip_sw);
        default:
                return -EOPNOTSUPP;
        }
index 9f0df6a9786d83648f9ae3ad4e27428fd447e9ff..6dddab95d57a5c5632521ead38c3f5d767c239cf 100644 (file)
@@ -181,8 +181,8 @@ nfp_bpf_jit(struct bpf_prog *filter, void *prog,
 int nfp_prog_verify(struct nfp_prog *nfp_prog, struct bpf_prog *prog);
 
 struct nfp_net;
-struct tc_cls_bpf_offload;
 
-int nfp_net_bpf_offload(struct nfp_net *nn, struct tc_cls_bpf_offload *cls_bpf);
+int nfp_net_bpf_offload(struct nfp_net *nn, struct bpf_prog *prog,
+                       bool old_prog, bool sw_fallback);
 
 #endif
index 268ba1ba82db16bce2ed27141b73708d2504ada8..c09efa1a9649f0abe84f0160733f8c758a859c33 100644 (file)
@@ -52,8 +52,7 @@
 #include "../nfp_net.h"
 
 static int
-nfp_net_bpf_offload_prepare(struct nfp_net *nn,
-                           struct tc_cls_bpf_offload *cls_bpf,
+nfp_net_bpf_offload_prepare(struct nfp_net *nn, struct bpf_prog *prog,
                            struct nfp_bpf_result *res,
                            void **code, dma_addr_t *dma_addr, u16 max_instr)
 {
@@ -73,9 +72,9 @@ nfp_net_bpf_offload_prepare(struct nfp_net *nn,
        done_off = nn_readw(nn, NFP_NET_CFG_BPF_DONE);
 
        stack_size = nn_readb(nn, NFP_NET_CFG_BPF_STACK_SZ) * 64;
-       if (cls_bpf->prog->aux->stack_depth > stack_size) {
+       if (prog->aux->stack_depth > stack_size) {
                nn_info(nn, "stack too large: program %dB > FW stack %dB\n",
-                       cls_bpf->prog->aux->stack_depth, stack_size);
+                       prog->aux->stack_depth, stack_size);
                return -EOPNOTSUPP;
        }
 
@@ -83,8 +82,7 @@ nfp_net_bpf_offload_prepare(struct nfp_net *nn,
        if (!*code)
                return -ENOMEM;
 
-       ret = nfp_bpf_jit(cls_bpf->prog, *code, start_off, done_off,
-                         max_instr, res);
+       ret = nfp_bpf_jit(prog, *code, start_off, done_off, max_instr, res);
        if (ret)
                goto out;
 
@@ -96,13 +94,13 @@ out:
 }
 
 static void
-nfp_net_bpf_load_and_start(struct nfp_net *nn, u32 tc_flags,
+nfp_net_bpf_load_and_start(struct nfp_net *nn, bool sw_fallback,
                           void *code, dma_addr_t dma_addr,
                           unsigned int code_sz, unsigned int n_instr)
 {
        int err;
 
-       nn->dp.bpf_offload_skip_sw = !!(tc_flags & TCA_CLS_FLAGS_SKIP_SW);
+       nn->dp.bpf_offload_skip_sw = !sw_fallback;
 
        nn_writew(nn, NFP_NET_CFG_BPF_SIZE, n_instr);
        nn_writeq(nn, NFP_NET_CFG_BPF_ADDR, dma_addr);
@@ -134,7 +132,8 @@ static int nfp_net_bpf_stop(struct nfp_net *nn)
        return nfp_net_reconfig(nn, NFP_NET_CFG_UPDATE_GEN);
 }
 
-int nfp_net_bpf_offload(struct nfp_net *nn, struct tc_cls_bpf_offload *cls_bpf)
+int nfp_net_bpf_offload(struct nfp_net *nn, struct bpf_prog *prog,
+                       bool old_prog, bool sw_fallback)
 {
        struct nfp_bpf_result res;
        dma_addr_t dma_addr;
@@ -142,49 +141,37 @@ int nfp_net_bpf_offload(struct nfp_net *nn, struct tc_cls_bpf_offload *cls_bpf)
        void *code;
        int err;
 
+       /* There is nothing stopping us from implementing seamless
+        * replace but the simple method of loading I adopted in
+        * the firmware does not handle atomic replace (i.e. we have to
+        * stop the BPF offload and re-enable it).  Leaking-in a few
+        * frames which didn't have BPF applied in the hardware should
+        * be fine if software fallback is available, though.
+        */
+       if (prog && old_prog && nn->dp.bpf_offload_skip_sw)
+               return -EBUSY;
+
+       /* Something else is loaded, different program type? */
+       if (!old_prog && nn->dp.ctrl & NFP_NET_CFG_CTRL_BPF)
+               return -EBUSY;
+
        max_instr = nn_readw(nn, NFP_NET_CFG_BPF_MAX_LEN);
+       code = NULL;
 
-       switch (cls_bpf->command) {
-       case TC_CLSBPF_REPLACE:
-               /* There is nothing stopping us from implementing seamless
-                * replace but the simple method of loading I adopted in
-                * the firmware does not handle atomic replace (i.e. we have to
-                * stop the BPF offload and re-enable it).  Leaking-in a few
-                * frames which didn't have BPF applied in the hardware should
-                * be fine if software fallback is available, though.
-                */
-               if (nn->dp.bpf_offload_skip_sw)
-                       return -EBUSY;
-
-               err = nfp_net_bpf_offload_prepare(nn, cls_bpf, &res, &code,
+       if (prog) {
+               err = nfp_net_bpf_offload_prepare(nn, prog, &res, &code,
                                                  &dma_addr, max_instr);
                if (err)
                        return err;
+       }
 
+       if (old_prog)
                nfp_net_bpf_stop(nn);
-               nfp_net_bpf_load_and_start(nn, cls_bpf->gen_flags, code,
-                                          dma_addr, max_instr * sizeof(u64),
-                                          res.n_instr);
-               return 0;
-
-       case TC_CLSBPF_ADD:
-               if (nn->dp.ctrl & NFP_NET_CFG_CTRL_BPF)
-                       return -EBUSY;
-
-               err = nfp_net_bpf_offload_prepare(nn, cls_bpf, &res, &code,
-                                                 &dma_addr, max_instr);
-               if (err)
-                       return err;
 
-               nfp_net_bpf_load_and_start(nn, cls_bpf->gen_flags, code,
+       if (prog)
+               nfp_net_bpf_load_and_start(nn, sw_fallback, code,
                                           dma_addr, max_instr * sizeof(u64),
                                           res.n_instr);
-               return 0;
 
-       case TC_CLSBPF_DESTROY:
-               return nfp_net_bpf_stop(nn);
-
-       default:
-               return -EOPNOTSUPP;
-       }
+       return 0;
 }