On Tue, Jan 14, 2025 at 06:17:22PM +0100, Alexander Lobakin wrote:

From: Dan Carpenter <redacted>
Date: Tue, 14 Jan 2025 13:45:04 +0300

quoted

[ I tried to send this email yesterday but apparently gmail blocked
  it for security reasons?  So weird. - dan ]

On Mon, Jan 13, 2025 at 01:32:11PM +0100, Alexander Lobakin wrote:

quoted

From: Dan Carpenter <redacted>
Date: Mon, 13 Jan 2025 09:18:39 +0300

quoted

The "sizeof(struct cmsg_bpf_event) + pkt_size + data_size" math could
potentially have an integer wrapping bug on 32bit systems.  Check for

Not in practice I suppose? Do we need to fix "never" bugs?

No, this is from static analysis.  We don't need to fix never bugs.

This is called from nfp_bpf_ctrl_msg_rx() and nfp_bpf_ctrl_msg_rx_raw()
and I assumed that since pkt_size and data_size come from skb->data on
the rx path then they couldn't be trusted.

skbs are always valid and skb->len could never cross INT_MAX to provoke
an overflow.

True but unrelated.  I think you are looking at the wrong code...

drivers/net/ethernet/netronome/nfp/bpf/offload.c
   445  int nfp_bpf_event_output(struct nfp_app_bpf *bpf, const void *data,
                                                                      ^^^^
This code comes from the network so it cannot be trusted.

   446                           unsigned int len)
   447  {
   448          struct cmsg_bpf_event *cbe = (void *)data;
                                       ^^^^^^^^^^^^^^^^^^^
It is cast to a struct here.

   449          struct nfp_bpf_neutral_map *record;
   450          u32 pkt_size, data_size, map_id;
   451          u64 map_id_full;
   452  
   453          if (len < sizeof(struct cmsg_bpf_event))
   454                  return -EINVAL;
   455  
   456          pkt_size = be32_to_cpu(cbe->pkt_size);
   457          data_size = be32_to_cpu(cbe->data_size);

pkt_size and data_size are u32 values which are controlled from
over the network.

   458          map_id_full = be64_to_cpu(cbe->map_ptr);
   459          map_id = map_id_full;
   460  
   461          if (len < sizeof(struct cmsg_bpf_event) + pkt_size + data_size)
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
On a 32bit system, then this math can overflow.  The pkt_size and
data_size are too high and we should return -EINVAL but this check
doesn't work because of integer wrapping.

   462                  return -EINVAL;
   63          if (cbe->hdr.ver != NFP_CCM_ABI_VERSION)
   464                  return -EINVAL;
   465  
   466          rcu_read_lock();
   467          record = rhashtable_lookup(&bpf->maps_neutral, &map_id,
   468                                     nfp_bpf_maps_neutral_params);
   469          if (!record || map_id_full > U32_MAX) {
   470                  rcu_read_unlock();
   471                  cmsg_warn(bpf, "perf event: map id %lld (0x%llx) not recognized, dropping event\n",
   472                            map_id_full, map_id_full);
   473                  return -EINVAL;
   474          }
   475  
   476          bpf_event_output(record->ptr, be32_to_cpu(cbe->cpu_id),
   477                           &cbe->data[round_up(pkt_size, 4)], data_size,
                                            ^^^^^^^^^^^^^^^^^^^^^
Here we are way out of bounds with regards to the cbe->data[] array.

regards,
dan carpenter

   478                           cbe->data, pkt_size, nfp_bpf_perf_event_copy);
   479          rcu_read_unlock();
   480  
   481          return 0;
   482  }