On Thu, May 23, 2019 at 03:28:15PM +0100, Jiong Wang wrote:
>
> > On 23 May 2019, at 03:07, Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote:
> >
> > On Wed, May 22, 2019 at 07:54:57PM +0100, Jiong Wang wrote:
> >> eBPF ISA specification requires high 32-bit cleared when low 32-bit
> >> sub-register is written. This applies to destination register of ALU32 etc.
> >> JIT back-ends must guarantee this semantic when doing code-gen. x86_64 and
> >> AArch64 ISA has the same semantics, so the corresponding JIT back-end
> >> doesn't need to do extra work.
> >>
> >> However, 32-bit arches (arm, x86, nfp etc.) and some other 64-bit arches
> >> (PowerPC, SPARC etc) need to do explicit zero extension to meet this
> >> requirement, otherwise code like the following will fail.
> >>
> >> u64_value = (u64) u32_value
> >> ... other uses of u64_value
> >>
> >> This is because compiler could exploit the semantic described above and
> >> save those zero extensions for extending u32_value to u64_value, these JIT
> >> back-ends are expected to guarantee this through inserting extra zero
> >> extensions which however could be a significant increase on the code size.
> >> Some benchmarks show there could be ~40% sub-register writes out of total
> >> insns, meaning at least ~40% extra code-gen.
> >>
> >> One observation is these extra zero extensions are not always necessary.
> >> Take above code snippet for example, it is possible u32_value will never be
> >> casted into a u64, the value of high 32-bit of u32_value then could be
> >> ignored and extra zero extension could be eliminated.
> >>
> >> This patch implements this idea, insns defining sub-registers will be
> >> marked when the high 32-bit of the defined sub-register matters. For
> >> those unmarked insns, it is safe to eliminate high 32-bit clearnace for
> >> them.
> >>
> >> Algo:
> >> - Split read flags into READ32 and READ64.
> >>
> >> - Record index of insn that does sub-register write. Keep the index inside
> >> reg state and update it during verifier insn walking.
> >>
> >> - A full register read on a sub-register marks its definition insn as
> >> needing zero extension on dst register.
> >>
> >> A new sub-register write overrides the old one.
> >>
> >> - When propagating read64 during path pruning, also mark any insn defining
> >> a sub-register that is read in the pruned path as full-register.
> >>
> >> Reviewed-by: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>
> >> Signed-off-by: Jiong Wang <jiong.wang@xxxxxxxxxxxxx>
> >> ---
> >> include/linux/bpf_verifier.h | 14 +++-
> >> kernel/bpf/verifier.c | 175 +++++++++++++++++++++++++++++++++++++++----
> >> 2 files changed, 173 insertions(+), 16 deletions(-)
> >>
> >> diff --git a/include/linux/bpf_verifier.h b/include/linux/bpf_verifier.h
> >> index 1305ccb..60fb54e 100644
> >> --- a/include/linux/bpf_verifier.h
> >> +++ b/include/linux/bpf_verifier.h
> >> @@ -36,9 +36,11 @@
> >> */
> >> enum bpf_reg_liveness {
> >> REG_LIVE_NONE = 0, /* reg hasn't been read or written this branch */
> >> - REG_LIVE_READ, /* reg was read, so we're sensitive to initial value */
> >> - REG_LIVE_WRITTEN, /* reg was written first, screening off later reads */
> >> - REG_LIVE_DONE = 4, /* liveness won't be updating this register anymore */
> >> + REG_LIVE_READ32 = 0x1, /* reg was read, so we're sensitive to initial value */
> >> + REG_LIVE_READ64 = 0x2, /* likewise, but full 64-bit content matters */
> >> + REG_LIVE_READ = REG_LIVE_READ32 | REG_LIVE_READ64,
> >> + REG_LIVE_WRITTEN = 0x4, /* reg was written first, screening off later reads */
> >> + REG_LIVE_DONE = 0x8, /* liveness won't be updating this register anymore */
> >> };
> >>
> >> struct bpf_reg_state {
> >> @@ -131,6 +133,11 @@ struct bpf_reg_state {
> >> * pointing to bpf_func_state.
> >> */
> >> u32 frameno;
> >> + /* Tracks subreg definition. The stored value is the insn_idx of the
> >> + * writing insn. This is safe because subreg_def is used before any insn
> >> + * patching which only happens after main verification finished.
> >> + */
> >> + s32 subreg_def;
> >> enum bpf_reg_liveness live;
> >> };
> >>
> >> @@ -232,6 +239,7 @@ struct bpf_insn_aux_data {
> >> int ctx_field_size; /* the ctx field size for load insn, maybe 0 */
> >> int sanitize_stack_off; /* stack slot to be cleared */
> >> bool seen; /* this insn was processed by the verifier */
> >> + bool zext_dst; /* this insn zero extends dst reg */
> >> u8 alu_state; /* used in combination with alu_limit */
> >> unsigned int orig_idx; /* original instruction index */
> >> };
> >> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> >> index 95f93544..0efccf8 100644
> >> --- a/kernel/bpf/verifier.c
> >> +++ b/kernel/bpf/verifier.c
> >> @@ -981,6 +981,7 @@ static void mark_reg_not_init(struct bpf_verifier_env *env,
> >> __mark_reg_not_init(regs + regno);
> >> }
> >>
> >> +#define DEF_NOT_SUBREG (-1)
> >> static void init_reg_state(struct bpf_verifier_env *env,
> >> struct bpf_func_state *state)
> >> {
> >> @@ -991,6 +992,7 @@ static void init_reg_state(struct bpf_verifier_env *env,
> >> mark_reg_not_init(env, regs, i);
> >> regs[i].live = REG_LIVE_NONE;
> >> regs[i].parent = NULL;
> >> + regs[i].subreg_def = DEF_NOT_SUBREG;
> >
> > shouldn't it be moved into __mark_reg_unknown ?
>
> I had pondered this in early version, but thought it is not correct.
> Like “live”, “subreg_def" is initialised inside init_reg_state, they are
> not touched by __mark_reg_unknown, instead, they are updated mostly
> updated by check_reg_arg which is strictly called at every reg read/write
> sites which are places where register def and use happened and are the
> places we want to update “subreg_def”.
>
> From my understanding, __mark_reg_unknown could be called when verifier
> is not sure about the value of a register, but the call site doesn’t care
> about width at the moment. For example, for a narrow loads from stack,
> __mark_reg_unknown could be triggered, the dst_reg then will be marked as
> DEF_NOT_SUBREG, but it is a subreg define.
>
> Make sense?
yes.
> > Also what about my old suggestion to use DEF_NOT_SUBREG==0
> > to be on safer side when we zero things out?
>
> Will change DEF_NOT_SUBREG to 0 if no problem shown up.
>
> TBH, I thought about this in initial version, but later used -1
> to avoid index + 1 and some other code change issues IIRC.
>
> if we use DEF_NOT_SUBREG == 0, then when we zero things out, we will
> mark one register def as a full register, it then won’t be zero extended
> in any case.
>
> if we use DEF_NOT_SUBREG == -1, then when we zero things out, one register
> will be marked as sub-register def and defined by insn 0, which normally
> is a context register move.
'normally is a context register' is a dangerous assumption to make.
that's exactly the reason I prefer DEF_NOT_SUBREG == 0
>
> >
> >> }
> >>
> >> /* frame pointer */
> >> @@ -1136,7 +1138,7 @@ static int check_subprogs(struct bpf_verifier_env *env)
> >> */
> >> static int mark_reg_read(struct bpf_verifier_env *env,
> >> const struct bpf_reg_state *state,
> >> - struct bpf_reg_state *parent)
> >> + struct bpf_reg_state *parent, u8 flag)
> >> {
> >> bool writes = parent == state->parent; /* Observe write marks */
> >> int cnt = 0;
> >> @@ -1151,17 +1153,26 @@ static int mark_reg_read(struct bpf_verifier_env *env,
> >> parent->var_off.value, parent->off);
> >> return -EFAULT;
> >> }
> >> - if (parent->live & REG_LIVE_READ)
> >> + /* The first condition is more likely to be true than the
> >> + * second, checked it first.
> >> + */
> >> + if ((parent->live & REG_LIVE_READ) == flag ||
> >> + parent->live & REG_LIVE_READ64)
> >> /* The parentage chain never changes and
> >> * this parent was already marked as LIVE_READ.
> >> * There is no need to keep walking the chain again and
> >> * keep re-marking all parents as LIVE_READ.
> >> * This case happens when the same register is read
> >> * multiple times without writes into it in-between.
> >> + * Also, if parent has the stronger REG_LIVE_READ64 set,
> >> + * then no need to set the weak REG_LIVE_READ32.
> >> */
> >> break;
> >> /* ... then we depend on parent's value */
> >> - parent->live |= REG_LIVE_READ;
> >> + parent->live |= flag;
> >> + /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
> >> + if (flag == REG_LIVE_READ64)
> >> + parent->live &= ~REG_LIVE_READ32;
> >> state = parent;
> >> parent = state->parent;
> >> writes = true;
> >> @@ -1173,12 +1184,111 @@ static int mark_reg_read(struct bpf_verifier_env *env,
> >> return 0;
> >> }
> >>
> >> +/* This function is supposed to be used by the following 32-bit optimization
> >> + * code only. It returns TRUE if the source or destination register operates
> >> + * on 64-bit, otherwise return FALSE.
> >> + */
> >> +static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
> >> + u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
> >> +{
> >> + u8 code, class, op;
> >> +
> >> + code = insn->code;
> >> + class = BPF_CLASS(code);
> >> + op = BPF_OP(code);
> >> + if (class == BPF_JMP) {
> >> + /* BPF_EXIT for "main" will reach here. Return TRUE
> >> + * conservatively.
> >> + */
> >> + if (op == BPF_EXIT)
> >> + return true;
> >> + if (op == BPF_CALL) {
> >> + /* BPF to BPF call will reach here because of marking
> >> + * caller saved clobber with DST_OP_NO_MARK for which we
> >> + * don't care the register def because they are anyway
> >> + * marked as NOT_INIT already.
> >> + */
> >> + if (insn->src_reg == BPF_PSEUDO_CALL)
> >> + return false;
> >> + /* Helper call will reach here because of arg type
> >> + * check, conservatively return TRUE.
> >> + */
> >> + if (t == SRC_OP)
> >> + return true;
> >> +
> >> + return false;
> >> + }
> >> + }
> >> +
> >> + if (class == BPF_ALU64 || class == BPF_JMP ||
> >> + /* BPF_END always use BPF_ALU class. */
> >> + (class == BPF_ALU && op == BPF_END && insn->imm == 64))
> >> + return true;
> >> +
> >> + if (class == BPF_ALU || class == BPF_JMP32)
> >> + return false;
> >> +
> >> + if (class == BPF_LDX) {
> >> + if (t != SRC_OP)
> >> + return BPF_SIZE(code) == BPF_DW;
> >> + /* LDX source must be ptr. */
> >> + return true;
> >> + }
> >> +
> >> + if (class == BPF_STX) {
> >> + if (reg->type != SCALAR_VALUE)
> >> + return true;
> >> + return BPF_SIZE(code) == BPF_DW;
> >> + }
> >> +
> >> + if (class == BPF_LD) {
> >> + u8 mode = BPF_MODE(code);
> >> +
> >> + /* LD_IMM64 */
> >> + if (mode == BPF_IMM)
> >> + return true;
> >> +
> >> + /* Both LD_IND and LD_ABS return 32-bit data. */
> >> + if (t != SRC_OP)
> >> + return false;
> >> +
> >> + /* Implicit ctx ptr. */
> >> + if (regno == BPF_REG_6)
> >> + return true;
> >
> > compiler will optimize above 'if' away.
> > What's the point of the above stmt?
>
> I implemented this function insn class by class to make sure I had thorough
> checks for each insn class and won’t miss some corner case. I think the code
> is more clear for maintain. And was making each insn class do full return
> after finishing their own checks, tried to avoid fall through, this is why
> these “redundant” checks are here.
>
> And for this BPF_REG_6 check, BPF_LD insn looked quite special to me, so was
> implementing it following the comments above check_ld_abs case by case, this
> check was for the following comment:
>
> * Implicit input:
> * ctx == skb == R6 == CTX
>
> It is true these checks could be merged, checks inside BPF_LD could be
> even further compressed into single line along with the class == BPF_LD
> check but I am thinking that would be a little bit over compressed,
> will just drop the following BPF_ST check and this BPF_REG_6 check, I
> am thinking avoid fall through might make the code more clear, so will
> stop at letting BPF_LD fall through to default return, new code then
> will look like the following:
>
> if (class == BPF_LD) {
> u8 mode = BPF_MODE(code);
>
> /* LD_IMM64 */
> if (mode == BPF_IMM)
> return true;
>
> /* Both LD_IND and LD_ABS return 32-bit data. */
> if (t != SRC_OP)
> return false;
>
> /* SRC is implicit ctx ptr which is 64-bit, or explicit source
> * which could be any width.
> */
> return true;
> }
>
> /* Conservatively return true for all the other cases. */
> return true;
>
> Does this looks good?
well, it made me realize that we're probably doing it wrong,
since after calling check_reg_arg() we need to re-parse insn encoding.
How about we change check_reg_arg()'s enum reg_arg_type instead?
The caller has much more context and no need to parse insn opcode again.
Something like:
enum reg_arg_type {
SRC_OP64,
DST_OP64,
DST_OP_NO_MARK, // probably no need to split this one ?
SRC_OP32,
DST_OP32,
};
> >
> >> +
> >> + /* Explicit source could be any width. */
> >> + return true;
> >> + }
> >> +
> >> + if (class == BPF_ST)
> >> + /* The only source register for BPF_ST is a ptr. */
> >> + return true;
> >
> > ditto
>
> >
> >> +
> >> + /* Conservatively return true at default. */
> >> + return true;
> >> +}
> >> +
> >> +static void mark_insn_zext(struct bpf_verifier_env *env,
> >> + struct bpf_reg_state *reg)
> >> +{
> >> + s32 def_idx = reg->subreg_def;
> >> +
> >> + if (def_idx == DEF_NOT_SUBREG)
> >> + return;
> >> +
> >> + env->insn_aux_data[def_idx].zext_dst = true;
> >> + /* The dst will be zero extended, so won't be sub-register anymore. */
> >> + reg->subreg_def = DEF_NOT_SUBREG;
> >> +}
> >> +
> >> static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
> >> enum reg_arg_type t)
> >> {
> >> struct bpf_verifier_state *vstate = env->cur_state;
> >> struct bpf_func_state *state = vstate->frame[vstate->curframe];
> >> + struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
> >> struct bpf_reg_state *reg, *regs = state->regs;
> >> + bool rw64;
> >>
> >> if (regno >= MAX_BPF_REG) {
> >> verbose(env, "R%d is invalid\n", regno);
> >> @@ -1186,6 +1296,7 @@ static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
> >> }
> >>
> >> reg = ®s[regno];
> >> + rw64 = is_reg64(env, insn, regno, reg, t);
> >> if (t == SRC_OP) {
> >> /* check whether register used as source operand can be read */
> >> if (reg->type == NOT_INIT) {
> >> @@ -1196,7 +1307,11 @@ static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
> >> if (regno == BPF_REG_FP)
> >> return 0;
> >>
> >> - return mark_reg_read(env, reg, reg->parent);
> >> + if (rw64)
> >> + mark_insn_zext(env, reg);
> >> +
> >> + return mark_reg_read(env, reg, reg->parent,
> >> + rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
> >> } else {
> >> /* check whether register used as dest operand can be written to */
> >> if (regno == BPF_REG_FP) {
> >> @@ -1204,6 +1319,7 @@ static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
> >> return -EACCES;
> >> }
> >> reg->live |= REG_LIVE_WRITTEN;
> >> + reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx;
> >> if (t == DST_OP)
> >> mark_reg_unknown(env, regs, regno);
> >> }
> >> @@ -1383,7 +1499,8 @@ static int check_stack_read(struct bpf_verifier_env *env,
> >> state->regs[value_regno].live |= REG_LIVE_WRITTEN;
> >> }
> >> mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
> >> - reg_state->stack[spi].spilled_ptr.parent);
> >> + reg_state->stack[spi].spilled_ptr.parent,
> >> + REG_LIVE_READ64);
> >> return 0;
> >> } else {
> >> int zeros = 0;
> >> @@ -1400,7 +1517,9 @@ static int check_stack_read(struct bpf_verifier_env *env,
> >> return -EACCES;
> >> }
> >> mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
> >> - reg_state->stack[spi].spilled_ptr.parent);
> >> + reg_state->stack[spi].spilled_ptr.parent,
> >> + size == BPF_REG_SIZE
> >> + ? REG_LIVE_READ64 : REG_LIVE_READ32);
> >
> > I don't think that's correct.
> > It's a parial read from a stack slot that doesn't contain a pointer.
> > It can be of any size and any offset.
> > Hence 4-byte read can read upper bits too.
>
> That’s true. But I think this won’t cause correctness issue. Because
> for stack read, we call mark_reg_read mostly to set the read bit in
> “live”, and it is used by stacksafe to know whether one stack slot
> is ever used or not. Either REG_LIVE_READ64 or REG_LIVE_READ32 will
> let the slot marked as “read”, and let stacksafe know it is used.
yes. that's technically correct, but very fragile.
> For stack content, if it comes from eBPF register A, then it must have
> been stored by STX to stack which should have triggered zero extension
> on the original A if it is a sub-register. So the content is correct.
>
> Here, the issue is stack slot read is reusing mark_reg_read, the last
> parameter for stack slot read just need to let mark_reg_read know this
> slot has been read, perhaps just pass REG_LIVE_READ64 for all stack
> read?
yes. please.
