On Wed, May 27, 2020 at 10:12 AM Alexander Potapenko <gli...@google.com> wrote: > > On Wed, May 27, 2020 at 6:58 PM Alexei Starovoitov > <alexei.starovoi...@gmail.com> wrote: > > > > On Wed, May 27, 2020 at 8:52 AM Alexander Potapenko <gli...@google.com> > > wrote: > > > > > > This basically means that BPF's output register was uninitialized when > > > ___bpf_prog_run() returned. > > > > > > When I replace the lines initializing registers A and X in > > > net/core/filter.c: > > > > > > - *new_insn++ = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, > > > BPF_REG_A); > > > - *new_insn++ = BPF_ALU32_REG(BPF_XOR, BPF_REG_X, > > > BPF_REG_X); > > > > > > with > > > > > > + *new_insn++ = BPF_MOV32_IMM(BPF_REG_A, 0); > > > + *new_insn++ = BPF_MOV32_IMM(BPF_REG_X, 0); > > > > > > , the bug goes away, therefore I think it's being caused by XORing the > > > initially uninitialized registers with themselves. > > > > > > kernel/bpf/core.c:1408, where the uninitialized value was stored to > > > memory, points to the "ALU(ADD, +)" macro in ___bpf_prog_run(). > > > But the debug info seems to be incorrect here: if I comment this line > > > out and unroll the macro manually, KMSAN points to "ALU(SUB, -)". > > > Most certainly it's actually one of the XOR instruction declarations. > > > > > > Do you think it makes sense to use the UB-proof BPF_MOV32_IMM > > > instructions to initialize the registers? > > > > I think it's better for UBsan to get smarter about xor-ing. > > Could you please elaborate on this? How exactly should KMSAN handle > this situation? > Note that despite the source says "BPF_ALU32_REG(BPF_XOR, BPF_REG_A, > BPF_REG_A);", it doesn't necessarily boil down to an expression like A > = A ^ A. It's more likely that temporary values will be involved, > making it quite hard to figure out whether the two operands are really > the same.
I really don't know who to make it smarter. It's your area of expertise.