* Remove global state for trap registration
There's a number of changes brought about in this commit, motivated by a
few things. One motivation was to remove an instance of using
`lazy_static!` in an effort to remove global state and encapsulate it
wherever possible. A second motivation came when investigating a
slowly-compiling wasm module (a bit too slowly) where a good chunk of
time was spent in managing trap registrations.
The specific change made here is that `TrapRegistry` is now stored
inside of a `Compiler` instead of inside a global. Additionally traps
are "bulk registered" for a module rather than one-by-one. This form of
bulk-registration allows optimizing the locks used here, where a lock is
only held for a module at-a-time instead of once-per-function.
With these changes the "unregister" logic has also been tweaked a bit
here and there to continue to work. As a nice side effect the `Compiler`
type now has one fewer field that requires actual mutability and has
been updated for multi-threaded compilation, nudging us closer to a
world where we can support multi-threaded compilation. Yay!
In terms of performance improvements, a local wasm test file that
previously took 3 seconds to compile is now 10% faster to compile,
taking ~2.7 seconds now.
* Perform trap resolution after unwinding
This avoids taking locks in signal handlers which feels a bit iffy...
* Remove `TrapRegistration::dummy()`
Avoid an case where you're trying to lookup trap information from a
dummy module for something that happened in a different module.
* Tweak some comments
The intention of the `wasmtime` crate was to disable this verifier by
default, but it looks like cranelift actually has it turned on by
default which was making our documentation incorrect!
This was discovered by seeing a number of timeouts when fuzzing. The
debug verifier is great for fuzzing, however, so fuzzing is updated to
enable this unconditionally, meaning we'll still have timeouts. For
general users though this should make the documentation correct that the
`wasmtime` crate, by default, disables the debug verifier.
Investigating a susprisingly slow-compiling module recently, it turns
out that if you create a wasm module with 40k empty functions (e.g.
`(module (func) (func) (func) ...)`) then it takes **3 seconds** to
compile and drop via the CLI locally on a Linux system. This seems like
an extraordinary amount of time for "doing nothing", and after some
profiling I found that basically all of the time was spent in
`__deregister_frame` calls.
Poking around in the source it looks like libgcc is managing some form
of linked list, and by deregistering in the LIFO order instead of FIFO
order it avoids a quadratic search of all registered functions. Now that
being said it's still pretty bad to do a linear search all the time, and
nothing will be fixed if there are *two* instances both with 40k
functions.
For now though I hope that this will patch over the performance issue
and we can figure out better ways to manage this in the future.
* Reimplement `wasmtime-wasi` on top of `wasmtime`
This commit reimplements the `wasmtime-wasi` crate on top of the
`wasmtime` API crate, instead of being placed on top of the `wasmtime-*`
family of internal crates. The purpose here is to continue to exercise
the API as well as avoid usage of internals wherever possible and
instead use the safe API as much as possible.
The `wasmtime-wasi` crate's API has been updated as part of this PR as
well. The general outline of it is now:
* Each module snapshot has a `WasiCtxBuilder`, `WasiCtx`, and `Wasi`
type.
* The `WasiCtx*` types are reexported from `wasi-common`.
* The `Wasi` type is synthesized by the `wig` crate's procedural macro
* The `Wasi` type exposes one constructor which takes a `Store` and a
`WasiCtx`, and produces a `Wasi`
* Each `Wasi` struct fields for all the exported functions in that wasi
module. They're all public an they all have type `wasmtime::Func`
* The `Wasi` type has a `get_export` method to fetch an struct field by
name.
The intention here is that we can continue to make progress on #727 by
integrating WASI construction into the `Instance::new` experience, but
it requires everything to be part of the same system!
The main oddity required by the `wasmtime-wasi` crate is that it needs
access to the caller's `memory` export, if any. This is currently done
with a bit of a hack and is expected to go away once interface types are
more fully baked in.
* Remove now no-longer-necessary APIs from `wasmtime`
* rustfmt
* Rename to from_abi
* Move `Func` to its own file
* Support `Func` imports with zero shims
This commit extends the `Func` type in the `wasmtime` crate with static
`wrap*` constructors. The goal of these constructors is to create a
`Func` type which has zero shims associated with it, creating as small
of a layer as possible between wasm code and calling imported Rust code.
This is achieved by creating an `extern "C"` shim function which matches
the ABI of what Cranelift will generate, and then the host function is
passed directly into an `InstanceHandle` to get called later. This also
enables enough inlining opportunities that LLVM will be able to see all
functions and inline everything to the point where your function is
called immediately from wasm, no questions asked.
Whenever we enter wasm code we need to verify that the correct number
and the correct types of arguments were passed in, lest we misinterpret
bits!
Closes#52
Print out an explicit `rerun-if-changed` annotation in the `build.rs`
for the `wasmtime-cli` crate to avoid rebuilding it constantly as files
like tests change which don't need to cause a rebuild.
Crates used exclusively at build time, such as proc-macro crates and
their dependencies, don't benefit substantially from optimization; they
take far longer to optimize than time saved when running them during the
build. No machine code from these crates will appear in the final
compiled binary.
Use the new profile-overrides mechanism in Rust 1.41
(https://doc.rust-lang.org/cargo/reference/profiles.html#overrides) to
build such crates with opt-level 0.
On a 4-thread laptop, this brings build time from 6m27s to 4m21s, and
CPU time spent from 23m43s to 16m23s.
* Remove another thread local in `instance.rs`
This commit removes another usage of `thread_local!` in the continued
effort to centralize all thread-local state per-call (or basically state
needed for traps) in one location. This removal is targeted at the
support for custom signal handlers on instances, removing the previous
stack of instances with instead a linked list of instances.
The `with_signals_on` method is no longer necessary (since it was always
called anyway) and is inferred from the first `vmctx` argument of the
entrypoints into wasm. These functions establish a linked list of
instances on the stack, if needed, to handle signals when they happen.
This involved some refactoring where some C++ glue was moved into Rust,
so now Rust handles a bit more of the signal handling logic.
* Update some inline docs about `HandleTrap`
During creation of an `InstanceHandle` if a link error occurred (such as
an element segment doesn't fit) then the instance itself would be leaked
by accident. This commit fixes the issue by ensuring that an
`InstanceHandle` is created very quickly so if any initialization later
fails it will be cleaned up through normal destructors.
We've been getting some errors on Linux which seem like they might be
related to a pinned `wheel` dependency. Apparently I originally added
this dependency to CI and I have no idea why I wrote down a `==`
dependency for it, so let's try not pinning and see what happens.
* Improve panics/traps from imported functions
This commit performs a few refactorings and fixes a bug as well. The
changes here are:
* The `thread_local!` in the `wasmtime` crate for trap information is
removed. The thread local in the `wasmtime_runtime` crate is now
leveraged to transmit trap information.
* Panics in user-provided functions are now caught explicitly to be
carried across JIT code manually. Getting Rust panics unwinding
through JIT code is pretty likely to be super tricky and difficult to
do, so in the meantime we can get by with catching panics and resuming
the panic once we've resumed in Rust code.
* Various take/record trap apis have all been removed in favor of
working directly with `Trap` objects, where the internal trap object
has been expanded slightly to encompass user-provided errors as well.
This borrows a bit #839 and otherwise will...
Closes#848
* Rename `r#return` to `ret`
* Support parsing the text format in `wasmtime` crate
This commit adds support to the `wasmtime::Module` type to parse the
text format. This is often quite convenient to support in testing or
tinkering with the runtime. Additionally the `wat` parser is pretty
lightweight and easy to add to builds, so it's relatively easy for us to
support as well!
The exact manner that this is now supported comes with a few updates to
the existing API:
* A new optional feature of the `wasmtime` crate, `wat`, has been added.
This is enabled by default.
* The `Module::new` API now takes `impl AsRef<[u8]>` instead of just
`&[u8]`, and when the `wat` feature is enabled it will attempt to
interpret it either as a wasm binary or as the text format. Note that
this check is quite cheap since you just check the first byte.
* A `Module::from_file` API was added as a convenience to parse a file
from disk, allowing error messages for `*.wat` files on disk to be a
bit nicer.
* APIs like `Module::new_unchecked` and `Module::validate` remain
unchanged, they require the binary format to be called.
The intention here is to make this as convenient as possible for new
developers of the `wasmtime` crate. By changing the default behavior
though this has ramifications such as, for example, supporting the text
format implicitly through the C API now.
* Handle review comments
* Update more tests to avoid usage of `wat` crate
* Go back to unchecked for now in wasm_module_new
Looks like C# tests rely on this?
* Reel in unsafety around `InstanceHandle`
This commit is an attempt, or at least is targeted at being a start, at
reeling in the unsafety around the `InstanceHandle` type. Currently this
type represents a sort of moral `Rc<Instance>` but is a bit more
specialized since the underlying memory is allocated through mmap.
Additionally, though, `InstanceHandle` exposes a fundamental flaw in its
safety by safetly allowing mutable access so long as you have `&mut
InstanceHandle`. This type, however, is trivially created by simply
cloning a `InstanceHandle` to get an owned reference. This means that
`&mut InstanceHandle` does not actually provide any guarantees about
uniqueness, so there's no more safety than `&InstanceHandle` itself.
This commit removes all `&mut self` APIs from `InstanceHandle`,
additionally removing some where `&self` was `unsafe` and `&mut self`
was safe (since it was trivial to subvert this "safety"). In doing so
interior mutability patterns are now used much more extensively through
structures such as `Table` and `Memory`. Additionally a number of
methods were refactored to be a bit clearer and use helper functions
where possible.
This is a relatively large commit unfortunately, but it snowballed very
quickly into touching quite a few places. My hope though is that this
will prevent developers working on wasmtime internals as well as
developers still yet to migrate to the `wasmtime` crate from falling
into trivial unsafe traps by accidentally using `&mut` when they can't.
All existing users relying on `&mut` will need to migrate to some form
of interior mutability, such as using `RefCell` or `Cell`.
This commit also additionally marks `InstanceHandle::new` as an `unsafe`
function. The rationale for this is that the `&mut`-safety is only the
beginning for the safety of `InstanceHandle`. In general the wasmtime
internals are extremely unsafe and haven't been audited for appropriate
usage of `unsafe`. Until that's done it's hoped that we can warn users
with this `unsafe` constructor and otherwise push users to the
`wasmtime` crate which we know is safe.
* Fix windows build
* Wrap up mutable memory state in one structure
Rather than having separate fields
* Use `Cell::set`, not `Cell::replace`, where possible
* Add a helper function for offsets from VMContext
* Fix a typo from merging
* rustfmt
* Use try_from, not as
* Tweak style of some setters
* Improve handling of strings for backtraces
Largely avoid storing strings at all in the `wasmtime-*` internal
crates, and instead only store strings in a separate global cache
specific to the `wasmtime` crate itself. This global cache is inserted
and removed from dynamically as modules are created and deallocated, and
the global cache is consulted whenever a `Trap` is created to
symbolicate any wasm frames.
This also avoids the need to thread `module_name` through the jit crates
and back, and additionally removes the need for `ModuleSyncString`.
* Run rustfmt
This commit deletes the old C implementation of the original
`wasi_unstable` module, instead only leaving around our single
`wasmtime-wasi` crate as the implementation for both
`wasi_snapshot_preview1` and `wasi_unstable`.
This hasn't been discussed (AFAIK) up until now, so this is also a
proposal! Some thoughts in favor of this deletion I would have are:
* This has been off-by-default for ages
* We don't build or test any of this on CI
* Published binaries with `wasmtime` do not have this possibility
enabled
* Future refactorings to the `wasmtime-wasi` crate will either need to
work around how the C implementation is different or bring it up to
speed.
This is motivated by the last bullet point where I was working on
getting `wasmtime-wasi` working purely as an implementation detail on
top of the `wasmtime` crate itself, but quickly ran into a case where
the CLI would need to multiplex all sorts of wasi implementations. In
any case I'm curious what others think, is this too soon? Is there
something remaining blocking this? (etc)
* Remove the unused EnsureDarwinMachPorts function
When compiling the C++ shims for longjmp/setjmp/signal handling we don't
use the `USE_APPLE_MACH_PORTS` directive, so this function was entirely
unused anyway. This looks to be a holdover from when this was originally
copied from C++, but no need for keeping around this now-legacy
initialization.
* Remove the `wasmtime_init_finish` function
This looks like it's perhaps largely historical cruft at this point now
I think? The function, with the removal of the mach ports from the
previous commit, only reads the initializtion state of the signal
handlers. If the signal handlers failed to get installed, though, it
simply returns early rather than erroring out anyway. In any case a
follow-up commit will refactor `wasmtime_init_eager` to handle this as
well.
* Pare down `wasmtime_init_eager`
Similar to previous commits it looks like this function may have accrued
some debt over time, nowadays it doesn't really do much other than
capture a backtrace and install signal handlers. The `lazy_static` state
isn't really that necessary and we can rely on the `Once` primitive in
the standard library for one-time initialization.
This also updates the code to unconditionally panic if signal handlers
fail to get installed, which I think is the behavior that we'll want for
now and we can enhance it over time if necessary, but I don't think we
have a use case where it's currently necessary.
* Reduce number of thread locals in trap handling
This commit refactors the trap handling portion of wasmtime with a few
goals in mind. I've been reading around a bit lately and feel that we
have a bit too few globals and thread locals floating around rather than
handles attached to contexts. I'm hoping that we can reduce the number
of thread locals and globals, and this commit is the start of reducing
this number.
The changes applied in this commit remove the set of thread locals in
the `traphandlers` module in favor of one thread local that's managed in
a sort of stack discipline. This way each call to `wasmtime_call*` sets
up its own stack local state that can be managed and read on that stack
frame.
Additionally the C++ glue code around `setjmp` and `longjmp` has all
been refactored to avoid going back and forth between Rust and C++. Now
we'll simply enter C++, go straight into `setjmp`/the call, and then
traps will enter Rust only once to both learn if the trap should be
acted upon and record information about the trap.
Overall the hope here is that context passing between `wasmtime_call*`
and the trap handling function will be a bit easier. For example I hope
to remove the global `get_trap_registry()` function next in favor of
storing a handle to a registry inside each instance, and the `*mut
VMContext` can be used to reach the `InstanceHandle` underneath, and
this trap registry.
* Update crates/runtime/src/traphandlers.rs
Co-Authored-By: Sergei Pepyakin <s.pepyakin@gmail.com>
Co-authored-by: Sergei Pepyakin <s.pepyakin@gmail.com>
This commit removes the `signature_cache` field from the `Store` type
and performs a few internal changes which are aimed to be a bit forward
looking towards #777, making `Store` threadsafe.
The changes made here are:
* The `SignatureRegistry` internal type now contains the reverse map
that `signature_cache` was serving to do. This is populated on calls
to `register` automatically and is accompanied by a `lookup` method as
well.
* The `register_wasmtime_signature` and `lookup_wasmtime_signature`
methods were removed from `Store` and now instead work by using the
`Compiler::signatures` field.
* The `SignatureRegistry` type was updated to have interior mutability.
The global `Compiler` type is highly likely to get shared across many
threads through `Store`, so it needs some form of lock somewhere for
mutation of the registry of signatures and this commit opts to put it
inside `SignatureRegistry` which will eventually allow for the removal
of most `&mut self` method on `Compiler`.
* Auto-generate shims for old `wasi_unstable` module
This commit is effectively just doing what #707 already did, but
applying it to the `snapshot_0` module as well. The end result is the
same, where we cut down on all the boilerplate in `snapshot_0` and bring
it in line with the main `wasi_snapshot_preview1` implementation. The
goal here is to make it easier to change the two in tandem since they're
both doing the same thing.
* Migrate `wasi_common::hostcalls` to a macro
This commit migrates the `hostcalls` module to being auto-generated by a
macro rather than duplicating a handwritten signature for each wasi
syscall.
* Auto-generate snapshot_0's `hostcalls` module
Similar to the previous commit, but for `snapshot_0`
* Delete the `wasi-common-cbindgen` crate
This is no longer needed with the hostcalls macro now, we can easily
fold the definition of the cbindgen macro into the same crate.
* Rustfmt
* Fix windows build errors
* Rustfmt
* Remove now no-longer-necessary code
* rustfmt
* Replaces `load_file` with `load_bytes` in macro
Loading an `AsRef<[u8]>` object is just more flexible than a filestring.
In the end you can just do `std::fs::read(&str)?` as the argument to get
the same behavior, but the reverse is a lot harder.
* updates markdown rust macro test example with new macro syntax
* Adds the `load_file` method back to rust macro
`load_file` was removed preferring `load_bytes`, but then later readded
with the `load_bytes` method as backend
Alex Crichton
Yury Delendik
Nathan Froyd
Nick Fitzgerald
Peter Huene
Josh Triplett
Sergei Pepyakin
Andrew Brown
Jakub Konka
Jakub Konka
Julian Popescu