contracts: Add automated weights for wasm instructions (#7361)

* pallet_contracts: Inline benchmark helper that is only used once

* Move all max_* Schedule items into a new struct

* Limit the number of globals a module can declare

* The current limits are too high for wasmi to even execute

* Limit the amount of parameters any wasm function is allowed to have

* Limit the size the BrTable's immediate value

* Add instruction benchmarks

* Add new benchmarks to the schedule and make use of it

* Add Benchmark Results generated by the bench bot

* Add proc macro that implements `Debug` for `Schedule`

* Add missing imports necessary for no_std build

* Make the WeightDebug macro available for no_std

In this case a dummy implementation is derived in order to not
blow up the code size akin to the RuntimeDebug macro.

* Rework instr_memory_grow benchmark to use only the maximum amount of pages allowed

* Add maximum amount of memory when benching (seal_)call/instantiate

* cargo run --release --features runtime-benchmarks --manifest-path bin/node/cli/Cargo.toml -- benchmark --chain dev --steps 50 --repeat 20 --extrinsic * --execution=wasm --wasm-execution=compiled --heap-pages=4096 --output ./bin/node/runtime/src/weights --header ./HEADER --pallet pallet_contracts

* Added utility benchmark that allows pretty printing of the real schedule

* review: Add missing header to the proc-macro lib.rs

* review: Clarify why #[allow(dead_code)] attribute is there

* review: Fix pwasm-utils line

* review: Fixup rand usage

* review: Fix typo

* review: Imported -> Exported

* cargo run --release --features=runtime-benchmarks --manifest-path=bin/node/cli/Cargo.toml -- benchmark --chain=dev --steps=50 --repeat=20 --pallet=pallet_contracts --extrinsic=* --execution=wasm --wasm-execution=compiled --heap-pages=4096 --output=./frame/contracts/src/weights.rs --template=./.maintain/frame-weight-template.hbs

* contracts: Adapt to new weight structure

* contracts: Fixup runtime WeightInfo

* contracts: Remove unneeded fullpath of WeightInfo type

* Apply suggestions from code review

Co-authored-by: Andrew Jones <ascjones@gmail.com>

* Fix typo in schedule.rs

Co-authored-by: Andrew Jones <ascjones@gmail.com>

* Fix docs in schedule.rs

* Apply suggestions from code review

Co-authored-by: Nikolay Volf <nikvolf@gmail.com>

* Don't publish proc-macro crate until 3.0.0 is ready

* Optimize imports for less repetition

* Break overlong line

Co-authored-by: Parity Benchmarking Bot <admin@parity.io>
Co-authored-by: Andrew Jones <ascjones@gmail.com>
Co-authored-by: Nikolay Volf <nikvolf@gmail.com>
This commit is contained in:
Alexander Theißen
2020-11-09 15:32:14 +01:00
committed by GitHub
parent 9704c204e6
commit 51c67fe881
17 changed files with 3152 additions and 843 deletions
+359 -63
View File
@@ -17,77 +17,180 @@
//! This module contains the cost schedule and supporting code that constructs a
//! sane default schedule from a `WeightInfo` implementation.
use crate::{Trait, WeightInfo};
use crate::{Trait, weights::WeightInfo};
#[cfg(feature = "std")]
use serde::{Serialize, Deserialize};
use pallet_contracts_proc_macro::{ScheduleDebug, WeightDebug};
use frame_support::weights::Weight;
use sp_std::{marker::PhantomData, fmt};
use sp_std::{marker::PhantomData, vec::Vec};
use codec::{Encode, Decode};
use parity_wasm::elements;
use pwasm_utils::rules;
use sp_runtime::RuntimeDebug;
/// How many API calls are executed in a single batch. The reason for increasing the amount
/// of API calls in batches (per benchmark component increase) is so that the linear regression
/// has an easier time determining the contribution of that component.
pub const API_BENCHMARK_BATCH_SIZE: u32 = 100;
/// How many instructions are executed in a single batch. The reasoning is the same
/// as for `API_BENCHMARK_BATCH_SIZE`.
pub const INSTR_BENCHMARK_BATCH_SIZE: u32 = 1_000;
/// Definition of the cost schedule and other parameterizations for wasm vm.
#[cfg_attr(feature = "std", derive(Serialize, Deserialize))]
#[derive(Clone, Encode, Decode, PartialEq, Eq)]
#[cfg_attr(feature = "std", serde(bound(serialize = "", deserialize = "")))]
#[derive(Clone, Encode, Decode, PartialEq, Eq, ScheduleDebug)]
pub struct Schedule<T: Trait> {
/// Version of the schedule.
pub version: u32,
/// The weights for individual wasm instructions.
pub instruction_weights: InstructionWeights,
/// The weights for each imported function a contract is allowed to call.
pub host_fn_weights: HostFnWeights,
/// Whether the `seal_println` function is allowed to be used contracts.
/// MUST only be enabled for `dev` chains, NOT for production chains
pub enable_println: bool,
/// The maximum number of topics supported by an event.
pub max_event_topics: u32,
/// Describes the upper limits on various metrics.
pub limits: Limits,
/// Maximum allowed stack height.
/// The weights for individual wasm instructions.
pub instruction_weights: InstructionWeights<T>,
/// The weights for each imported function a contract is allowed to call.
pub host_fn_weights: HostFnWeights<T>,
}
/// Describes the upper limits on various metrics.
#[cfg_attr(feature = "std", derive(Serialize, Deserialize))]
#[derive(Clone, Encode, Decode, PartialEq, Eq, RuntimeDebug)]
pub struct Limits {
/// The maximum number of topics supported by an event.
pub event_topics: u32,
/// Maximum allowed stack height in number of elements.
///
/// See https://wiki.parity.io/WebAssembly-StackHeight to find out
/// how the stack frame cost is calculated.
pub max_stack_height: u32,
/// how the stack frame cost is calculated. Each element can be of one of the
/// wasm value types. This means the maximum size per element is 64bit.
pub stack_height: u32,
/// Maximum number of globals a module is allowed to declare.
///
/// Globals are not limited through the `stack_height` as locals are. Neither does
/// the linear memory limit `memory_pages` applies to them.
pub globals: u32,
/// Maximum numbers of parameters a function can have.
///
/// Those need to be limited to prevent a potentially exploitable interaction with
/// the stack height instrumentation: The costs of executing the stack height
/// instrumentation for an indirectly called function scales linearly with the amount
/// of parameters of this function. Because the stack height instrumentation itself is
/// is not weight metered its costs must be static (via this limit) and included in
/// the costs of the instructions that cause them (call, call_indirect).
pub parameters: u32,
/// Maximum number of memory pages allowed for a contract.
pub max_memory_pages: u32,
pub memory_pages: u32,
/// Maximum allowed size of a declared table.
pub max_table_size: u32,
/// Maximum number of elements allowed in a table.
///
/// Currently, the only type of element that is allowed in a table is funcref.
pub table_size: u32,
/// The maximum length of a subject used for PRNG generation.
pub max_subject_len: u32,
/// Maximum number of elements that can appear as immediate value to the br_table instruction.
pub br_table_size: u32,
/// The maximum length of a subject in bytes used for PRNG generation.
pub subject_len: u32,
/// The maximum length of a contract code in bytes. This limit applies to the uninstrumented
/// and pristine form of the code as supplied to `put_code`.
pub max_code_size: u32,
pub code_size: u32,
}
/// Describes the weight for all categories of supported wasm instructions.
///
/// There there is one field for each wasm instruction that describes the weight to
/// execute one instruction of that name. There are a few execptions:
///
/// 1. If there is a i64 and a i32 variant of an instruction we use the weight
/// of the former for both.
/// 2. The following instructions are free of charge because they merely structure the
/// wasm module and cannot be spammed without making the module invalid (and rejected):
/// End, Unreachable, Return, Else
/// 3. The following instructions cannot be benchmarked because they are removed by any
/// real world execution engine as a preprocessing step and therefore don't yield a
/// meaningful benchmark result. However, in contrast to the instructions mentioned
/// in 2. they can be spammed. We price them with the same weight as the "default"
/// instruction (i64.const): Block, Loop, Nop
/// 4. We price both i64.const and drop as InstructionWeights.i64const / 2. The reason
/// for that is that we cannot benchmark either of them on its own but we need their
/// individual values to derive (by subtraction) the weight of all other instructions
/// that use them as supporting instructions. Supporting means mainly pushing arguments
/// and dropping return values in order to maintain a valid module.
#[cfg_attr(feature = "std", derive(Serialize, Deserialize))]
#[derive(Clone, Encode, Decode, PartialEq, Eq, WeightDebug)]
pub struct InstructionWeights<T: Trait> {
pub i64const: u32,
pub i64load: u32,
pub i64store: u32,
pub select: u32,
pub r#if: u32,
pub br: u32,
pub br_if: u32,
pub br_table: u32,
pub br_table_per_entry: u32,
pub call: u32,
pub call_indirect: u32,
pub call_indirect_per_param: u32,
pub local_get: u32,
pub local_set: u32,
pub local_tee: u32,
pub global_get: u32,
pub global_set: u32,
pub memory_current: u32,
pub memory_grow: u32,
pub i64clz: u32,
pub i64ctz: u32,
pub i64popcnt: u32,
pub i64eqz: u32,
pub i64extendsi32: u32,
pub i64extendui32: u32,
pub i32wrapi64: u32,
pub i64eq: u32,
pub i64ne: u32,
pub i64lts: u32,
pub i64ltu: u32,
pub i64gts: u32,
pub i64gtu: u32,
pub i64les: u32,
pub i64leu: u32,
pub i64ges: u32,
pub i64geu: u32,
pub i64add: u32,
pub i64sub: u32,
pub i64mul: u32,
pub i64divs: u32,
pub i64divu: u32,
pub i64rems: u32,
pub i64remu: u32,
pub i64and: u32,
pub i64or: u32,
pub i64xor: u32,
pub i64shl: u32,
pub i64shrs: u32,
pub i64shru: u32,
pub i64rotl: u32,
pub i64rotr: u32,
/// The type parameter is used in the default implementation.
pub _phantom: PhantomData<T>,
}
/// Describes the weight for all categories of supported wasm instructions.
#[cfg_attr(feature = "std", derive(Serialize, Deserialize))]
#[derive(Clone, Encode, Decode, PartialEq, Eq)]
pub struct InstructionWeights {
/// Weight of a growing memory by single page.
pub grow_mem: Weight,
/// Weight of a regular operation.
pub regular: Weight,
}
/// Describes the weight for each imported function that a contract is allowed to call.
#[cfg_attr(feature = "std", derive(Serialize, Deserialize))]
#[derive(Clone, Encode, Decode, PartialEq, Eq)]
pub struct HostFnWeights {
#[derive(Clone, Encode, Decode, PartialEq, Eq, WeightDebug)]
pub struct HostFnWeights<T: Trait> {
/// Weight of calling `seal_caller`.
pub caller: Weight,
@@ -222,21 +325,11 @@ pub struct HostFnWeights {
/// Weight per byte hashed by `seal_hash_blake2_128`.
pub hash_blake2_128_per_byte: Weight,
}
/// We need to implement Debug manually because the automatic derive enforces T
/// to also implement Debug.
impl<T: Trait> fmt::Debug for Schedule<T> {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
f.debug_struct("Schedule").finish()
}
/// The type parameter is used in the default implementation.
pub _phantom: PhantomData<T>
}
/// 500 (2 instructions per nano second on 2GHZ) * 1000x slowdown through wasmi
/// This is a wild guess and should be viewed as a rough estimation.
/// Proper benchmarks are needed before this value and its derivatives can be used in production.
const WASM_INSTRUCTION_COST: Weight = 500_000;
macro_rules! replace_token {
($_in:tt $replacement:tt) => { $replacement };
}
@@ -259,6 +352,25 @@ macro_rules! cost_batched_args {
}
}
macro_rules! cost_instr_no_params_with_batch_size {
($name:ident, $batch_size:expr) => {
(cost_args!($name, 1) / Weight::from($batch_size)) as u32
}
}
macro_rules! cost_instr_with_batch_size {
($name:ident, $num_params:expr, $batch_size:expr) => {
cost_instr_no_params_with_batch_size!($name, $batch_size)
.saturating_sub((cost_instr_no_params_with_batch_size!(instr_i64const, $batch_size) / 2).saturating_mul($num_params))
}
}
macro_rules! cost_instr {
($name:ident, $num_params:expr) => {
cost_instr_with_batch_size!($name, $num_params, INSTR_BENCHMARK_BATCH_SIZE)
}
}
macro_rules! cost_byte_args {
($name:ident, $( $arg: expr ),+) => {
cost_args!($name, $( $arg ),+) / 1024
@@ -297,12 +409,97 @@ macro_rules! cost_byte_batched {
impl<T: Trait> Default for Schedule<T> {
fn default() -> Self {
let instruction_weights = InstructionWeights {
grow_mem: WASM_INSTRUCTION_COST,
regular: WASM_INSTRUCTION_COST,
};
Self {
version: 0,
enable_println: false,
limits: Default::default(),
instruction_weights: Default::default(),
host_fn_weights: Default::default(),
}
}
}
let host_fn_weights = HostFnWeights {
impl Default for Limits {
fn default() -> Self {
Self {
event_topics: 4,
// 512 * sizeof(i64) will give us a 4k stack.
stack_height: 512,
globals: 256,
parameters: 128,
memory_pages: 16,
// 4k function pointers (This is in count not bytes).
table_size: 4096,
br_table_size: 256,
subject_len: 32,
code_size: 512 * 1024,
}
}
}
impl<T: Trait> Default for InstructionWeights<T> {
fn default() -> Self {
let max_pages = Limits::default().memory_pages;
Self {
i64const: cost_instr!(instr_i64const, 1),
i64load: cost_instr!(instr_i64load, 2),
i64store: cost_instr!(instr_i64store, 2),
select: cost_instr!(instr_select, 4),
r#if: cost_instr!(instr_if, 3),
br: cost_instr!(instr_br, 2),
br_if: cost_instr!(instr_br_if, 5),
br_table: cost_instr!(instr_br_table, 3),
br_table_per_entry: cost_instr!(instr_br_table_per_entry, 0),
call: cost_instr!(instr_call, 2),
call_indirect: cost_instr!(instr_call_indirect, 3),
call_indirect_per_param: cost_instr!(instr_call_indirect_per_param, 1),
local_get: cost_instr!(instr_local_get, 1),
local_set: cost_instr!(instr_local_set, 1),
local_tee: cost_instr!(instr_local_tee, 2),
global_get: cost_instr!(instr_global_get, 1),
global_set: cost_instr!(instr_global_set, 1),
memory_current: cost_instr!(instr_memory_current, 1),
memory_grow: cost_instr_with_batch_size!(instr_memory_grow, 1, max_pages),
i64clz: cost_instr!(instr_i64clz, 2),
i64ctz: cost_instr!(instr_i64ctz, 2),
i64popcnt: cost_instr!(instr_i64popcnt, 2),
i64eqz: cost_instr!(instr_i64eqz, 2),
i64extendsi32: cost_instr!(instr_i64extendsi32, 2),
i64extendui32: cost_instr!(instr_i64extendui32, 2),
i32wrapi64: cost_instr!(instr_i32wrapi64, 2),
i64eq: cost_instr!(instr_i64eq, 3),
i64ne: cost_instr!(instr_i64ne, 3),
i64lts: cost_instr!(instr_i64lts, 3),
i64ltu: cost_instr!(instr_i64ltu, 3),
i64gts: cost_instr!(instr_i64gts, 3),
i64gtu: cost_instr!(instr_i64gtu, 3),
i64les: cost_instr!(instr_i64les, 3),
i64leu: cost_instr!(instr_i64leu, 3),
i64ges: cost_instr!(instr_i64ges, 3),
i64geu: cost_instr!(instr_i64geu, 3),
i64add: cost_instr!(instr_i64add, 3),
i64sub: cost_instr!(instr_i64sub, 3),
i64mul: cost_instr!(instr_i64mul, 3),
i64divs: cost_instr!(instr_i64divs, 3),
i64divu: cost_instr!(instr_i64divu, 3),
i64rems: cost_instr!(instr_i64rems, 3),
i64remu: cost_instr!(instr_i64remu, 3),
i64and: cost_instr!(instr_i64and, 3),
i64or: cost_instr!(instr_i64or, 3),
i64xor: cost_instr!(instr_i64xor, 3),
i64shl: cost_instr!(instr_i64shl, 3),
i64shrs: cost_instr!(instr_i64shrs, 3),
i64shru: cost_instr!(instr_i64shru, 3),
i64rotl: cost_instr!(instr_i64rotl, 3),
i64rotr: cost_instr!(instr_i64rotr, 3),
_phantom: PhantomData,
}
}
}
impl<T: Trait> Default for HostFnWeights<T> {
fn default() -> Self {
Self {
caller: cost_batched!(seal_caller),
address: cost_batched!(seal_address),
gas_left: cost_batched!(seal_gas_left),
@@ -348,20 +545,119 @@ impl<T: Trait> Default for Schedule<T> {
hash_blake2_256_per_byte: cost_byte_batched!(seal_hash_blake2_256_per_kb),
hash_blake2_128: cost_batched!(seal_hash_blake2_128),
hash_blake2_128_per_byte: cost_byte_batched!(seal_hash_blake2_128_per_kb),
};
Self {
version: 0,
instruction_weights,
host_fn_weights,
enable_println: false,
max_event_topics: 4,
max_stack_height: 64 * 1024,
max_memory_pages: 16,
max_table_size: 16 * 1024,
max_subject_len: 32,
max_code_size: 512 * 1024,
_phantom: PhantomData,
}
}
}
struct ScheduleRules<'a, T: Trait> {
schedule: &'a Schedule<T>,
params: Vec<u32>,
}
impl<T: Trait> Schedule<T> {
pub fn rules(&self, module: &elements::Module) -> impl rules::Rules + '_ {
ScheduleRules {
schedule: &self,
params: module
.type_section()
.iter()
.flat_map(|section| section.types())
.map(|func| {
let elements::Type::Function(func) = func;
func.params().len() as u32
})
.collect()
}
}
}
impl<'a, T: Trait> rules::Rules for ScheduleRules<'a, T> {
fn instruction_cost(&self, instruction: &elements::Instruction) -> Option<u32> {
use parity_wasm::elements::Instruction::*;
let w = &self.schedule.instruction_weights;
let max_params = self.schedule.limits.parameters;
let weight = match *instruction {
End | Unreachable | Return | Else => 0,
I32Const(_) | I64Const(_) | Block(_) | Loop(_) | Nop | Drop => w.i64const,
I32Load(_, _) | I32Load8S(_, _) | I32Load8U(_, _) | I32Load16S(_, _) |
I32Load16U(_, _) | I64Load(_, _) | I64Load8S(_, _) | I64Load8U(_, _) |
I64Load16S(_, _) | I64Load16U(_, _) | I64Load32S(_, _) | I64Load32U(_, _)
=> w.i64load,
I32Store(_, _) | I32Store8(_, _) | I32Store16(_, _) | I64Store(_, _) |
I64Store8(_, _) | I64Store16(_, _) | I64Store32(_, _) => w.i64store,
Select => w.select,
If(_) => w.r#if,
Br(_) => w.br,
BrIf(_) => w.br_if,
Call(_) => w.call,
GetLocal(_) => w.local_get,
SetLocal(_) => w.local_set,
TeeLocal(_) => w.local_tee,
GetGlobal(_) => w.global_get,
SetGlobal(_) => w.global_set,
CurrentMemory(_) => w.memory_current,
GrowMemory(_) => w.memory_grow,
CallIndirect(idx, _) => *self.params.get(idx as usize).unwrap_or(&max_params),
BrTable(ref data) =>
w.br_table.saturating_add(
w.br_table_per_entry.saturating_mul(data.table.len() as u32)
),
I32Clz | I64Clz => w.i64clz,
I32Ctz | I64Ctz => w.i64ctz,
I32Popcnt | I64Popcnt => w.i64popcnt,
I64ExtendSI32 => w.i64extendsi32,
I64ExtendUI32 => w.i64extendui32,
I32WrapI64 => w.i32wrapi64,
I32Eq | I64Eq => w.i64eq,
I32Ne | I64Ne => w.i64ne,
I32LtS | I64LtS => w.i64lts,
I32LtU | I64LtU => w.i64ltu,
I32GtS | I64GtS => w.i64gts,
I32GtU | I64GtU => w.i64gtu,
I32LeS | I64LeS => w.i64les,
I32LeU | I64LeU => w.i64leu,
I32GeS | I64GeS => w.i64ges,
I32GeU | I64GeU => w.i64geu,
I32Add | I64Add => w.i64add,
I32Sub | I64Sub => w.i64sub,
I32Mul | I64Mul => w.i64mul,
I32DivS | I64DivS => w.i64divs,
I32DivU | I64DivU => w.i64divu,
I32RemS | I64RemS => w.i64rems,
I32RemU | I64RemU => w.i64remu,
I32And | I64And => w.i64and,
I32Or | I64Or => w.i64or,
I32Xor | I64Xor => w.i64xor,
I32Shl | I64Shl => w.i64shl,
I32ShrS | I64ShrS => w.i64shrs,
I32ShrU | I64ShrU => w.i64shru,
I32Rotl | I64Rotl => w.i64rotl,
I32Rotr | I64Rotr => w.i64rotr,
// Returning None makes the gas instrumentation fail which we intend for
// unsupported or unknown instructions.
_ => return None,
};
Some(weight)
}
fn memory_grow_cost(&self) -> Option<rules::MemoryGrowCost> {
// We benchmarked the memory.grow instruction with the maximum allowed pages.
// The cost for growing is therefore already included in the instruction cost.
None
}
}
#[cfg(test)]
mod test {
use crate::tests::Test;
use super::*;
#[test]
fn print_test_schedule() {
let schedule = Schedule::<Test>::default();
println!("{:#?}", schedule);
}
}