From this question I understood that it is in charge of Linux NIC's driver code (loaded in kernel space) to allocate at its initialization both RX/TX ring descriptors plus data buffers as well.
NIC's DMA engine will access the NIC driver DMA's allocated memory to deliver incoming/outgoing packets.
From the point of view of incoming packets (i.e. received by the NIC from the wire) it is actually in the Bottom half context (SoftIRQ) that the NIC's driver allocates sk_buff(s) for each received packet meanwhile.
In case tcpdump is running on an interface, then incoming packets are actually copied from data buffers into the PACKET-RX-RING memory shared between the kernel and the tcpdump application (in the latest versions libpcap uses PACKET_MMAP).
Therefore in the above scenario the cloning of sk_buff is not involved but there is actually a copy of data buffers.
Did I get it right ? Thanks.