Skip to content

Store engine together with Module to mitigate memory increase issue (backport #1982) #1987

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Jan 15, 2024
Merged

Store engine together with Module to mitigate memory increase issue (backport #1982) #1987

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
eb2bcf9
Store engine together with Module to fix memory increase issue
webmaster128 Jan 6, 2024
61c5e53
Add CHANGELOG entry for fixing #1978
webmaster128 Jan 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,13 @@ and this project adheres to

## [Unreleased]

### Fixed

- cosmwasm-vm: Fix memory increase issue (1.3 -> 1.4 regression) by avoiding the
use of a long running Wasmer Engine. ([#1978])

[#1978]: https://github.com/CosmWasm/cosmwasm/issues/1978

## [1.5.1] - 2024-01-10

### Fixed
Expand Down
100 changes: 53 additions & 47 deletions packages/vm/src/cache.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@ use std::io::{Read, Write};
use std::marker::PhantomData;
use std::path::{Path, PathBuf};
use std::sync::Mutex;
use wasmer::{Engine, Store};
use wasmer::{Module, Store};

use crate::backend::{Backend, BackendApi, Querier, Storage};
use crate::capabilities::required_capabilities_from_module;
Expand All @@ -17,7 +17,7 @@ use crate::modules::{CachedModule, FileSystemCache, InMemoryCache, PinnedMemoryC
use crate::parsed_wasm::ParsedWasm;
use crate::size::Size;
use crate::static_analysis::has_ibc_entry_points;
use crate::wasm_backend::{compile, make_compiling_engine, make_runtime_engine};
use crate::wasm_backend::{compile, make_compiling_engine};

const STATE_DIR: &str = "state";
// Things related to the state of the blockchain.
Expand Down Expand Up @@ -88,24 +88,14 @@ pub struct CacheInner {
memory_cache: InMemoryCache,
fs_cache: FileSystemCache,
stats: Stats,
/// A single engine to execute all contracts in this cache instance (usually
/// this means all contracts in the process).
///
/// This engine is headless, i.e. does not contain a Singlepass compiler.
/// It only executes modules compiled with other engines.
///
/// The engine has one memory limit set which is the same for all contracts
/// running with it. If different memory limits would be needed for different
/// contracts at some point, we'd need multiple engines. This is because the tunables
/// that control the limit are attached to the engine.
runtime_engine: Engine,
}

pub struct Cache<A: BackendApi, S: Storage, Q: Querier> {
/// Available capabilities are immutable for the lifetime of the cache,
/// i.e. any number of read-only references is allowed to access it concurrently.
available_capabilities: HashSet<String>,
inner: Mutex<CacheInner>,
instance_memory_limit: Size,
// Those two don't store data but only fix type information
type_api: PhantomData<A>,
type_storage: PhantomData<S>,
Expand Down Expand Up @@ -161,8 +151,8 @@ where
memory_cache: InMemoryCache::new(memory_cache_size),
fs_cache,
stats: Stats::default(),
runtime_engine: make_runtime_engine(Some(instance_memory_limit)),
}),
instance_memory_limit,
type_storage: PhantomData::<S>,
type_api: PhantomData::<A>,
type_querier: PhantomData::<Q>,
Expand Down Expand Up @@ -298,11 +288,12 @@ where
// for a not-so-relevant use case.

// Try to get module from file system cache
if let Some((module, module_size)) = cache.fs_cache.load(checksum, &cache.runtime_engine)? {
if let Some(cached_module) = cache
.fs_cache
.load(checksum, Some(self.instance_memory_limit))?
{
cache.stats.hits_fs_cache = cache.stats.hits_fs_cache.saturating_add(1);
return cache
.pinned_memory_cache
.store(checksum, module, module_size);
return cache.pinned_memory_cache.store(checksum, cached_module);
}

// Re-compile from original Wasm bytecode
Expand All @@ -317,16 +308,16 @@ where
}

// This time we'll hit the file-system cache.
let Some((module, module_size)) = cache.fs_cache.load(checksum, &cache.runtime_engine)?
let Some(cached_module) = cache
.fs_cache
.load(checksum, Some(self.instance_memory_limit))?
else {
return Err(VmError::generic_err(
"Can't load module from file system cache after storing it to file system cache (pin)",
));
};

cache
.pinned_memory_cache
.store(checksum, module, module_size)
cache.pinned_memory_cache.store(checksum, cached_module)
}

/// Unpins a Module, i.e. removes it from the pinned memory cache.
Expand All @@ -350,10 +341,10 @@ where
backend: Backend<A, S, Q>,
options: InstanceOptions,
) -> VmResult<Instance<A, S, Q>> {
let (cached, store) = self.get_module(checksum)?;
let (module, store) = self.get_module(checksum)?;
let instance = Instance::from_module(
store,
&cached.module,
&module,
backend,
options.gas_limit,
options.print_debug,
Expand All @@ -366,36 +357,49 @@ where
/// Returns a module tied to a previously saved Wasm.
/// Depending on availability, this is either generated from a memory cache, file system cache or Wasm code.
/// This is part of `get_instance` but pulled out to reduce the locking time.
fn get_module(&self, checksum: &Checksum) -> VmResult<(CachedModule, Store)> {
fn get_module(&self, checksum: &Checksum) -> VmResult<(Module, Store)> {
let mut cache = self.inner.lock().unwrap();
// Try to get module from the pinned memory cache
if let Some(element) = cache.pinned_memory_cache.load(checksum)? {
cache.stats.hits_pinned_memory_cache =
cache.stats.hits_pinned_memory_cache.saturating_add(1);
let store = Store::new(cache.runtime_engine.clone());
return Ok((element, store));
let CachedModule {
module,
engine,
size_estimate: _,
} = element;
let store = Store::new(engine);
return Ok((module, store));
}

// Get module from memory cache
if let Some(element) = cache.memory_cache.load(checksum)? {
cache.stats.hits_memory_cache = cache.stats.hits_memory_cache.saturating_add(1);
let store = Store::new(cache.runtime_engine.clone());
return Ok((element, store));
let CachedModule {
module,
engine,
size_estimate: _,
} = element;
let store = Store::new(engine);
return Ok((module, store));
}

// Get module from file system cache
if let Some((module, module_size)) = cache.fs_cache.load(checksum, &cache.runtime_engine)? {
if let Some(cached_module) = cache
.fs_cache
.load(checksum, Some(self.instance_memory_limit))?
{
cache.stats.hits_fs_cache = cache.stats.hits_fs_cache.saturating_add(1);

cache
.memory_cache
.store(checksum, module.clone(), module_size)?;
let cached = CachedModule {
cache.memory_cache.store(checksum, cached_module.clone())?;

let CachedModule {
module,
size_estimate: module_size,
};
let store = Store::new(cache.runtime_engine.clone());
return Ok((cached, store));
engine,
size_estimate: _,
} = cached_module;
let store = Store::new(engine);
return Ok((module, store));
}

// Re-compile module from wasm
Expand All @@ -414,21 +418,23 @@ where
}

// This time we'll hit the file-system cache.
let Some((module, module_size)) = cache.fs_cache.load(checksum, &cache.runtime_engine)?
let Some(cached_module) = cache
.fs_cache
.load(checksum, Some(self.instance_memory_limit))?
else {
return Err(VmError::generic_err(
"Can't load module from file system cache after storing it to file system cache (get_module)",
));
};
cache
.memory_cache
.store(checksum, module.clone(), module_size)?;
let cached = CachedModule {
cache.memory_cache.store(checksum, cached_module.clone())?;

let CachedModule {
module,
size_estimate: module_size,
};
let store = Store::new(cache.runtime_engine.clone());
Ok((cached, store))
engine,
size_estimate: _,
} = cached_module;
let store = Store::new(engine);
Ok((module, store))
}
}

Expand Down
23 changes: 20 additions & 3 deletions packages/vm/src/modules/cached_module.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,13 +1,30 @@
use wasmer::Module;
use wasmer::{Engine, Module};

/// Some manual tests on Simon's machine showed that Engine is roughly 3-5 KB big,
/// so give it a constant 10 KiB estimate.
#[inline]
pub fn engine_size_estimate() -> usize {
10 * 1024
}

#[derive(Debug, Clone)]
pub struct CachedModule {
pub module: Module,
/// The runtime engine to run this module. Ideally we could use a single engine
/// for all modules but the memory issue described in <https://github.com/wasmerio/wasmer/issues/4377>
/// requires using one engine per module as a workaround.
pub engine: Engine,
/// The estimated size of this element in memory.
/// Since the cached modules are just [rkyv](https://rkyv.org/) dumps of the Module
/// instances, we use the file size of the module on disk (not the Wasm!)
/// as an estimate for this.
/// Note: Since CosmWasm 1.4 (Wasmer 4), Store/Engine are not cached anymore.
/// The majority of the Module size is the Artifact.
///
/// Between CosmWasm 1.4 (Wasmer 4) and 1.5.2, Store/Engine were not cached. This lead to a
/// memory consumption problem. From 1.5.2 on, Module and Engine are cached and Store is created
/// from Engine on demand.
///
/// The majority of the Module size is the Artifact which is why we use the module filesize as the estimate.
/// Some manual tests on Simon's machine showed that Engine is roughly 3-5 KB big, so give it a constant
/// estimate: [`engine_size_estimate`].
pub size_estimate: usize,
}
59 changes: 40 additions & 19 deletions packages/vm/src/modules/file_system_cache.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,13 +5,18 @@ use std::panic::catch_unwind;
use std::path::{Path, PathBuf};
use thiserror::Error;

use wasmer::{AsEngineRef, DeserializeError, Module, Target};
use wasmer::{DeserializeError, Module, Target};

use crate::checksum::Checksum;
use crate::errors::{VmError, VmResult};

use crate::filesystem::mkdir_p;
use crate::modules::current_wasmer_module_version;
use crate::wasm_backend::make_runtime_engine;
use crate::Size;

use super::cached_module::engine_size_estimate;
use super::CachedModule;

/// Bump this version whenever the module system changes in a way
/// that old stored modules would be corrupt when loaded in the new system.
Expand Down Expand Up @@ -130,24 +135,29 @@ impl FileSystemCache {
path
}

/// Loads a serialized module from the file system and returns a module (i.e. artifact + store),
/// along with the size of the serialized module.
/// Loads a serialized module from the file system and returns a Module + Engine,
/// along with a size estimation for the pair.
pub fn load(
&self,
checksum: &Checksum,
engine: &impl AsEngineRef,
) -> VmResult<Option<(Module, usize)>> {
memory_limit: Option<Size>,
) -> VmResult<Option<CachedModule>> {
let file_path = self.module_file(checksum);

let engine = make_runtime_engine(memory_limit);
let result = if self.unchecked_modules {
unsafe { Module::deserialize_from_file_unchecked(engine, &file_path) }
unsafe { Module::deserialize_from_file_unchecked(&engine, &file_path) }
} else {
unsafe { Module::deserialize_from_file(engine, &file_path) }
unsafe { Module::deserialize_from_file(&engine, &file_path) }
};
match result {
Ok(module) => {
let module_size = module_size(&file_path)?;
Ok(Some((module, module_size)))
Ok(Some(CachedModule {
module,
engine,
size_estimate: module_size + engine_size_estimate(),
}))
}
Err(DeserializeError::Io(err)) => match err.kind() {
io::ErrorKind::NotFound => Ok(None),
Expand Down Expand Up @@ -229,7 +239,7 @@ mod tests {
use super::*;
use crate::{
size::Size,
wasm_backend::{compile, make_compiling_engine, make_runtime_engine},
wasm_backend::{compile, make_compiling_engine},
};
use tempfile::TempDir;
use wasmer::{imports, Instance as WasmerInstance, Store};
Expand All @@ -256,8 +266,7 @@ mod tests {
let checksum = Checksum::generate(&wasm);

// Module does not exist
let runtime_engine = make_runtime_engine(TESTING_MEMORY_LIMIT);
let cached = cache.load(&checksum, &runtime_engine).unwrap();
let cached = cache.load(&checksum, TESTING_MEMORY_LIMIT).unwrap();
assert!(cached.is_none());

// Store module
Expand All @@ -266,14 +275,21 @@ mod tests {
cache.store(&checksum, &module).unwrap();

// Load module
let cached = cache.load(&checksum, &runtime_engine).unwrap();
let cached = cache.load(&checksum, TESTING_MEMORY_LIMIT).unwrap();
assert!(cached.is_some());

// Check the returned module is functional.
// This is not really testing the cache API but better safe than sorry.
{
let (cached_module, module_size) = cached.unwrap();
assert_eq!(module_size, module.serialize().unwrap().len());
let CachedModule {
module: cached_module,
engine: runtime_engine,
size_estimate,
} = cached.unwrap();
assert_eq!(
size_estimate,
module.serialize().unwrap().len() + 10240 /* engine size estimate */
);
let import_object = imports! {};
let mut store = Store::new(runtime_engine);
let instance = WasmerInstance::new(&mut store, &cached_module, &import_object).unwrap();
Expand Down Expand Up @@ -318,20 +334,25 @@ mod tests {
let checksum = Checksum::generate(&wasm);

// Store module
let engine1 = make_compiling_engine(TESTING_MEMORY_LIMIT);
let module = compile(&engine1, &wasm).unwrap();
let compiling_engine = make_compiling_engine(TESTING_MEMORY_LIMIT);
let module = compile(&compiling_engine, &wasm).unwrap();
cache.store(&checksum, &module).unwrap();

// It's there
let engine2 = make_runtime_engine(TESTING_MEMORY_LIMIT);
assert!(cache.load(&checksum, &engine2).unwrap().is_some());
assert!(cache
.load(&checksum, TESTING_MEMORY_LIMIT)
.unwrap()
.is_some());

// Remove module
let existed = cache.remove(&checksum).unwrap();
assert!(existed);

// it's gone now
assert!(cache.load(&checksum, &engine2).unwrap().is_none());
assert!(cache
.load(&checksum, TESTING_MEMORY_LIMIT)
.unwrap()
.is_none());

// Remove again
let existed = cache.remove(&checksum).unwrap();
Expand Down
Loading