Seems I never replied to this, thanks for the reviews!
On Mon, Dec 07, 2020 at 10:37:32PM -0800, John Fastabend wrote:
> Brendan Jackman wrote:
> > This adds two atomic opcodes, both of which include the BPF_FETCH
> > flag. XCHG without the BPF_FETCH flag would naturally encode
> > atomic_set. This is not supported because it would be of limited
> > value to userspace (it doesn't imply any barriers). CMPXCHG without
> > BPF_FETCH woulud be an atomic compare-and-write. We don't have such
> > an operation in the kernel so it isn't provided to BPF either.
> >
> > There are two significant design decisions made for the CMPXCHG
> > instruction:
> >
> > - To solve the issue that this operation fundamentally has 3
> > operands, but we only have two register fields. Therefore the
> > operand we compare against (the kernel's API calls it 'old') is
> > hard-coded to be R0. x86 has similar design (and A64 doesn't
> > have this problem).
> >
> > A potential alternative might be to encode the other operand's
> > register number in the immediate field.
> >
> > - The kernel's atomic_cmpxchg returns the old value, while the C11
> > userspace APIs return a boolean indicating the comparison
> > result. Which should BPF do? A64 returns the old value. x86 returns
> > the old value in the hard-coded register (and also sets a
> > flag). That means return-old-value is easier to JIT.
>
> Just a nit as it looks like perhaps we get one more spin here. Would
> be nice to be explicit about what the code does here. The above reads
> like it could go either way. Just an extra line "So we use ...' would
> be enough.
Ack, adding the note.
> > Signed-off-by: Brendan Jackman <jackmanb@xxxxxxxxxx>
> > ---
>
> One question below.
>
> > arch/x86/net/bpf_jit_comp.c | 8 ++++++++
> > include/linux/filter.h | 22 ++++++++++++++++++++++
> > include/uapi/linux/bpf.h | 4 +++-
> > kernel/bpf/core.c | 20 ++++++++++++++++++++
> > kernel/bpf/disasm.c | 15 +++++++++++++++
> > kernel/bpf/verifier.c | 19 +++++++++++++++++--
> > tools/include/linux/filter.h | 22 ++++++++++++++++++++++
> > tools/include/uapi/linux/bpf.h | 4 +++-
> > 8 files changed, 110 insertions(+), 4 deletions(-)
> >
>
> [...]
>
> > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > index f8c4e809297d..f5f4460b3e4e 100644
> > --- a/kernel/bpf/verifier.c
> > +++ b/kernel/bpf/verifier.c
> > @@ -3608,11 +3608,14 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
> >
> > static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
> > {
> > + int load_reg;
> > int err;
> >
> > switch (insn->imm) {
> > case BPF_ADD:
> > case BPF_ADD | BPF_FETCH:
> > + case BPF_XCHG:
> > + case BPF_CMPXCHG:
> > break;
> > default:
> > verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
> > @@ -3634,6 +3637,13 @@ static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_i
> > if (err)
> > return err;
> >
> > + if (insn->imm == BPF_CMPXCHG) {
> > + /* Check comparison of R0 with memory location */
> > + err = check_reg_arg(env, BPF_REG_0, SRC_OP);
> > + if (err)
> > + return err;
> > + }
> > +
>
> I need to think a bit more about it, but do we need to update is_reg64()
> at all for these?
I don't think so - this all falls into the same
`if (class == BPF_STX)` case as the existing BPF_STX_XADD instruction.
> > if (is_pointer_value(env, insn->src_reg)) {
> > verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
> > return -EACCES;
> > @@ -3664,8 +3674,13 @@ static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_i
> > if (!(insn->imm & BPF_FETCH))
> > return 0;
> >
> > - /* check and record load of old value into src reg */
> > - err = check_reg_arg(env, insn->src_reg, DST_OP);
> > + if (insn->imm == BPF_CMPXCHG)
> > + load_reg = BPF_REG_0;
> > + else
> > + load_reg = insn->src_reg;
> > +
> > + /* check and record load of old value */
> > + err = check_reg_arg(env, load_reg, DST_OP);
> > if (err)
> > return err;
