Thanks for your reply. With your solution a), I guess 'single mempool' mean a mempool fit in one memseg (continuous memory). What about a mempool across many memsegs ? I'm afraid it's still not safe. Just like this one: -------------- MEMPOOL ---------------- mempool @0x7ff9e4833d00 flags=10 pool=0x7ff9fbfffe00 phys_addr=0xc4fc33d00 nb_mem_chunks=91 size=524288 populated_size=524288 header_size=64 elt_size=2432 trailer_size=0 total_obj_size=2496 private_data_size=64 avg bytes/object=2496.233643 Zone 0: name:, phys:0xc4fdb7f40, len:0x34000, virt:0x7ff9e49b7f40, socket_id:0, flags:0 Zone 1: name:, phys:0xc4fc33d00, len:0x182100, virt:0x7ff9e4833d00, socket_id:0, flags:0 Zone 2: name:, phys:0xb22000080, len:0x16ffff40, virt:0x7ffa3a800080, socket_id:0, flags:0 Zone 3: name:, phys:0xc199ffe00, len:0x800180, virt:0x7ff9fbfffe00, socket_id:0, flags:0 Zone 4: name:, phys:0xc29c00080, len:0x77fff40, virt:0x7ff9e5800080, socket_id:0, flags:0 Zone 5: name:, phys:0xc22c00080, len:0x67fff40, virt:0x7ff9ed200080, socket_id:0, flags:0 Zone 6: name:, phys:0xc1dc00080, len:0x3bfff40, virt:0x7ff9f4800080, socket_id:0, flags:0 Zone 7: name:, phys:0xc1bc00080, len:0x1bfff40, virt:0x7ff9f8600080, socket_id:0, flags:0 Zone 8: name:, phys:0xbf4600080, len:0xffff40, virt:0x7ffa1ea00080, socket_id:0, flags:0 Zone 9: name:, phys:0xc0e000080, len:0xdfff40, virt:0x7ffa06400080, socket_id:0, flags:0 Zone 10: name:, phys:0xbe0600080, len:0xdfff40, virt:0x7ffa32000080, socket_id:0, flags:0 Zone 11: name:, phys:0xc18000080, len:0xbfff40, virt:0x7ff9fd000080, socket_id:0, flags:0 Zone 12: name:, phys:0x65000080, len:0xbfff40, virt:0x7ffa54e00080, socket_id:0, flags:0 Zone 13: name:, phys:0xc12a00080, len:0x7fff40, virt:0x7ffa02200080, socket_id:0, flags:0 Zone 14: name:, phys:0xc0d600080, len:0x7fff40, virt:0x7ffa07400080, socket_id:0, flags:0 Zone 15: name:, phys:0xc06600080, len:0x7fff40, virt:0x7ffa0de00080, socket_id:0, flags:0 ... 在2018年03月13 22时57分, "Ferruh Yigit"写道: On 3/9/2018 12:14 PM, cys wrote: > Commit 8451269e6d7ba7501723fe2efd0 said "remove continuous memory restriction"; > http://dpdk.org/browse/dpdk/commit/lib/librte_eal/linuxapp/kni/kni_net.c?id=8451269e6d7ba7501723fe2efd05745010295bac > For chained mbufs(nb_segs > 1), function va2pa use the offset of previous mbuf > to calculate physical address of next mbuf. > So anywhere guarante that all mbufs have the same offset (buf_addr - buf_physaddr) ? > Or have I misunderstood chained mbufs? Hi, Your description is correct, KNI chained mbufs is broken if chained mbufs are from different mempools. Two commits seems involved, in time order: [1] d89a58dfe90b ("kni: support chained mbufs") [2] 8451269e6d7b ("kni: remove continuous memory restriction") With current implementation, kernel needs to know physical address of the mbuf to be able to access it. For chained mbufs, first mbuf is OK but for rest kernel side gets the virtual address of the mbuf and this only works if all chained mbufs are from same mempool. I don't have any good solution indeed, but it is possible to: a) If you are using chained mbufs, keep old limitation of using singe mempool b) Serialize chained mbufs for KNI in userspace &}+bץry^4מm5j_z(ا\d^qyik^ץr{{^&ݹ睵Vvd_"13wu.ׯvdyx1_4מ5@Mp&[R(9jZrا,jjEWZuEj[E~&~k&4ץrnwۍo'hNߚ)zW(DL-/DR۞ć8sm5^nMKur+s