nfp: bpf: remember maps by ID
authorJakub Kicinski <jakub.kicinski@netronome.com>
Thu, 26 Jul 2018 02:53:34 +0000 (19:53 -0700)
committerDaniel Borkmann <daniel@iogearbox.net>
Fri, 27 Jul 2018 05:14:35 +0000 (07:14 +0200)
Record perf maps by map ID, not raw kernel pointer.  This helps
with debug messages, because printing pointers to logs is frowned
upon, and makes debug easier for the users, as map ID is something
they should be more familiar with.  Note that perf maps are offload
neutral, therefore IDs won't be orphaned.

While at it use a rate limited print helper for the error message.

Reported-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
drivers/net/ethernet/netronome/nfp/bpf/cmsg.c
drivers/net/ethernet/netronome/nfp/bpf/jit.c
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 1946291bf4fdd8d810138181fb639dd8af6410e4..2572a4b91c7c85e067183df29d98f6a8c59e5d57 100644 (file)
@@ -43,8 +43,6 @@
 #include "fw.h"
 #include "main.h"
 
-#define cmsg_warn(bpf, msg...) nn_dp_warn(&(bpf)->app->ctrl->dp, msg)
-
 #define NFP_BPF_TAG_ALLOC_SPAN (U16_MAX / 4)
 
 static bool nfp_bpf_all_tags_busy(struct nfp_app_bpf *bpf)
index 1d9e3683540439da2ee7d6e57fec4080542ed7ba..3c22d27de9da723896d2a5a2f21e1bbb54a985d4 100644 (file)
@@ -3883,6 +3883,7 @@ static int nfp_bpf_replace_map_ptrs(struct nfp_prog *nfp_prog)
        struct nfp_insn_meta *meta1, *meta2;
        struct nfp_bpf_map *nfp_map;
        struct bpf_map *map;
+       u32 id;
 
        nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) {
                if (meta1->skip || meta2->skip)
@@ -3894,11 +3895,14 @@ static int nfp_bpf_replace_map_ptrs(struct nfp_prog *nfp_prog)
 
                map = (void *)(unsigned long)((u32)meta1->insn.imm |
                                              (u64)meta2->insn.imm << 32);
-               if (bpf_map_offload_neutral(map))
-                       continue;
-               nfp_map = map_to_offmap(map)->dev_priv;
+               if (bpf_map_offload_neutral(map)) {
+                       id = map->id;
+               } else {
+                       nfp_map = map_to_offmap(map)->dev_priv;
+                       id = nfp_map->tid;
+               }
 
-               meta1->insn.imm = nfp_map->tid;
+               meta1->insn.imm = id;
                meta2->insn.imm = 0;
        }
 
index 192e88981fb23d58b79dbd2fe2847bd10f58f0b6..cce1d2945a320ba799d8aefce267ee7db4dccc0a 100644 (file)
@@ -45,8 +45,8 @@
 
 const struct rhashtable_params nfp_bpf_maps_neutral_params = {
        .nelem_hint             = 4,
-       .key_len                = FIELD_SIZEOF(struct nfp_bpf_neutral_map, ptr),
-       .key_offset             = offsetof(struct nfp_bpf_neutral_map, ptr),
+       .key_len                = FIELD_SIZEOF(struct bpf_map, id),
+       .key_offset             = offsetof(struct nfp_bpf_neutral_map, map_id),
        .head_offset            = offsetof(struct nfp_bpf_neutral_map, l),
        .automatic_shrinking    = true,
 };
index 017e0ae5e7364c5939f85840ec92cc59e21c870b..57573bfa8c0314dec11223c91636dca9de070e86 100644 (file)
@@ -47,6 +47,8 @@
 #include "../nfp_asm.h"
 #include "fw.h"
 
+#define cmsg_warn(bpf, msg...) nn_dp_warn(&(bpf)->app->ctrl->dp, msg)
+
 /* For relocation logic use up-most byte of branch instruction as scratch
  * area.  Remember to clear this before sending instructions to HW!
  */
@@ -221,6 +223,7 @@ struct nfp_bpf_map {
 struct nfp_bpf_neutral_map {
        struct rhash_head l;
        struct bpf_map *ptr;
+       u32 map_id;
        u32 count;
 };
 
index 293dda84818f29c8e92fd029bf54599d3336b10f..b1fbb3babc7f3846c62055c44b5b4fdf3aa6dd32 100644 (file)
@@ -67,7 +67,7 @@ nfp_map_ptr_record(struct nfp_app_bpf *bpf, struct nfp_prog *nfp_prog,
        ASSERT_RTNL();
 
        /* Reuse path - other offloaded program is already tracking this map. */
-       record = rhashtable_lookup_fast(&bpf->maps_neutral, &map,
+       record = rhashtable_lookup_fast(&bpf->maps_neutral, &map->id,
                                        nfp_bpf_maps_neutral_params);
        if (record) {
                nfp_prog->map_records[nfp_prog->map_records_cnt++] = record;
@@ -89,6 +89,7 @@ nfp_map_ptr_record(struct nfp_app_bpf *bpf, struct nfp_prog *nfp_prog,
        }
 
        record->ptr = map;
+       record->map_id = map->id;
        record->count = 1;
 
        err = rhashtable_insert_fast(&bpf->maps_neutral, &record->l,
@@ -457,15 +458,17 @@ int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data,
                         unsigned int len)
 {
        struct cmsg_bpf_event *cbe = (void *)data;
-       u32 pkt_size, data_size;
-       struct bpf_map *map;
+       struct nfp_bpf_neutral_map *record;
+       u32 pkt_size, data_size, map_id;
+       u64 map_id_full;
 
        if (len < sizeof(struct cmsg_bpf_event))
                return -EINVAL;
 
        pkt_size = be32_to_cpu(cbe->pkt_size);
        data_size = be32_to_cpu(cbe->data_size);
-       map = (void *)(unsigned long)be64_to_cpu(cbe->map_ptr);
+       map_id_full = be64_to_cpu(cbe->map_ptr);
+       map_id = map_id_full;
 
        if (len < sizeof(struct cmsg_bpf_event) + pkt_size + data_size)
                return -EINVAL;
@@ -473,15 +476,16 @@ int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data,
                return -EINVAL;
 
        rcu_read_lock();
-       if (!rhashtable_lookup_fast(&bpf->maps_neutral, &map,
-                                   nfp_bpf_maps_neutral_params)) {
+       record = rhashtable_lookup_fast(&bpf->maps_neutral, &map_id,
+                                       nfp_bpf_maps_neutral_params);
+       if (!record || map_id_full > U32_MAX) {
                rcu_read_unlock();
-               pr_warn("perf event: dest map pointer %px not recognized, dropping event\n",
-                       map);
+               cmsg_warn(bpf, "perf event: map id %lld (0x%llx) not recognized, dropping event\n",
+                         map_id_full, map_id_full);
                return -EINVAL;
        }
 
-       bpf_event_output(map, be32_to_cpu(cbe->cpu_id),
+       bpf_event_output(record->ptr, be32_to_cpu(cbe->cpu_id),
                         &cbe->data[round_up(pkt_size, 4)], data_size,
                         cbe->data, pkt_size, nfp_bpf_perf_event_copy);
        rcu_read_unlock();