On Tue, Apr 18, 2017 at 5:48 PM, William Tu via iovisor-dev <[email protected]> wrote: > Hi, > > I have a struct like below: > struct bpf_flow_keys { > long src; > }; > > I'm initializing an array of it using structures like below: > struct bpf_flow_keys flow_mask_array[] = { > {.src = 5555}, > {.src = 4444}, > }; > instead of initialize it one-by-one > struct bpf_flow_keys flow_mask_array[2]; > flow_mask_array[0].src = 5555; > flow_mask_array[1].src = 4444; > > However, the former fails the verifier due to > bpf_load_program() err=13 > 0: (bf) r6 = r1 > 1: (30) r0 = *(u8 *)skb[23] > 2: (63) *(u32 *)(r10 -4) = r0 > 3: (18) r1 = 0x223f6000 > 5: (79) r2 = *(u64 *)(r1 +8) > R1 invalid mem access 'map_ptr' > I guess it tries to store "4444, 5555" at some memory location (3: r1 > = 0x223f6000), and loads to stack (5: r2). > > The latter case works OK. Is this a limitation of verifier/bpf loader? > The C code, failed BPF, and passed BPF: > https://gist.github.com/williamtu/070a29e2fa8a15e512f26ea4ea9cffeb
yeah, that's the same issue that you mentioned in the other thread. We need to teach llvm to emit direct stores with imm values instead of copy from global rodata. _______________________________________________ iovisor-dev mailing list [email protected] https://lists.iovisor.org/mailman/listinfo/iovisor-dev
