This fixes a bug in the memfd-related management of a linear memory
where for dynamic memories memfd wasn't informed of the extra room that
the dynamic memory could grow into, only the actual size of linear
memory, which ended up tripping an assert once the memory was grown. The
fix here is pretty simple which is to factor in this extra space when
passing the allocation size to the creation of the `MemFdSlot`.
* Tweak memfd-related features crates
This commit changes the `memfd` feature for the `wasmtime-cli` crate
from an always-on feature to a default-on feature which can be disabled
at compile time. Additionally the `pooling-allocator` feature is also
given similar treatment.
Additionally some documentation was added for the `memfd` feature on the
`wasmtime` crate.
* Don't store `Arc<T>` in `InstanceAllocationRequest`
Instead store `&Arc<T>` to avoid having the clone that lives in
`InstanceAllocationRequest` not actually going anywhere. Otherwise all
instance allocation requires an extra clone to create it for the request
and an extra decrement when the request goes away. Internally clones are
made as necessary when creating instances.
* Enable the pooling allocator by default for `wasmtime-cli`
While perhaps not the most useful option since the CLI doesn't have a
great way to take advantage of this it probably makes sense to at least
match the features of `wasmtime` itself.
* Fix some lints and issues
* More compile fixes
* memfd: Reduce some syscalls in the on-demand case
This tweaks the internal organization of the `MemFdSlot` to avoid some
syscalls in the default case as well as opportunistically in the pooling
case. The two cases added here are:
* A `MemFdSlot` is now created with a specified initial size. For
pooling this is 0 but for the on-demand case this can be non-zero.
* When `instantiate` is called with no prior image and the sizes match
(as will be the case for on-demand allocation) then `mprotect` is
skipped entirely.
* In the `clear_and_remain-ready` case the `mprotect` is skipped if the
heap wasn't grown at all.
This should avoid ever using `mprotect` unnecessarily and makes the
ranges we `mprotect` a bit smaller as well.
* Review comments
* Tweak allow to apply to whole crate
This policy attempts to reuse the same instance slot for subsequent
instantiations of the same module. This is particularly useful when
using a pooling backend such as memfd that benefits from this reuse: for
example, in the memfd case, instantiating the same module into the same
slot allows us to avoid several calls to mmap() because the same
mappings can be reused.
The policy tracks a freelist per "compiled module ID", and when
allocating a slot for an instance, tries these three options in order:
1. A slot from the freelist for this module (i.e., last used for another
instantiation of this particular module), or
3. A slot that was last used by some other module or never before.
The "victim" slot for choice 2 is randomly chosen.
The data structures are carefully designed so that all updates are O(1),
and there is no retry-loop in any of the random selection.
This policy is now the default when the memfd backend is selected via
the `memfd-allocator` feature flag.
* Lazily populate a store's trampoline map
This commit is another installment of "how fast can we make
instantiation". Currently when instantiating a module with many function
imports each function, typically from the host, is inserted into the
store. This insertion process stores the `VMTrampoline` for the host
function in a side table so it can be looked up later if the host
function is called through the `Func` interface. This insertion process,
however, involves a hash map insertion which can be relatively expensive
at the scale of the rest of the instantiation process.
The optimization implemented in this commit is to avoid inserting
trampolines into the store at `Func`-insertion-time (aka instantiation
time) and instead only lazily populate the map of trampolines when
needed. The theory behind this is that almost all `Func` instances that
are called indirectly from the host are actually wasm functions, not
host-defined functions. This means that they already don't need to go
through the map of host trampolines and can instead be looked up from
the module they're defined in. With the assumed rarity of host functions
making `lookup_trampoline` a bit slower seems ok.
The `lookup_trampoline` function will now, on a miss from the wasm
modules and `host_trampolines` map, lazily iterate over the functions
within the store and insert trampolines into the `host_trampolines` map.
This process will eventually reach something which matches the function
provided because it should at least hit the same host function. The
relevant `lookup_trampoline` now sports a new documentation block
explaining all this as well for future readers.
Concretely this commit speeds up instantiation of an empty module with
100 imports and ~80 unique signatures from 10.6us to 6.4us, a 40%
improvement.
* Review comments
* Remove debug assert
Following up on #3696, use the new is-terminal crate to test for a tty
rather than having platform-specific logic in Wasmtime. The is-terminal
crate has a platform-independent API which takes a handle.
This also updates the tree to cap-std 0.24 etc., to avoid depending on
multiple versions of io-lifetimes at once, as enforced by the cargo deny
check.
With the addition of `sock_accept()` in `wasi-0.11.0`, wasmtime can now
implement basic networking for pre-opened sockets.
For Windows `AsHandle` was replaced with `AsRawHandleOrSocket` to cope
with the duality of Handles and Sockets.
For Unix a `wasi_cap_std_sync::net::Socket` enum was created to handle
the {Tcp,Unix}{Listener,Stream} more efficiently in
`WasiCtxBuilder::preopened_socket()`.
The addition of that many `WasiFile` implementors was mainly necessary,
because of the difference in the `num_ready_bytes()` function.
A known issue is Windows now busy polling on sockets, because except
for `stdin`, nothing is querying the status of windows handles/sockets.
Another know issue on Windows, is that there is no crate providing
support for `fcntl(fd, F_GETFL, 0)` on a socket.
Signed-off-by: Harald Hoyer <harald@profian.com>
Testing so far with recent Wasmtime has not been able to show the need
for avoiding the process-wide mmap lock in real-world use-cases. As
such, the technique of using an anonymous file and ftruncate() to extend
it seems unnecessary; instead, memfd can always use anonymous zeroed
memory for heap backing where the CoW image is not present, and
mprotect() to extend the heap limit by changing page protections.
As first suggested by Jan on the Zulip here [1], a cheap and effective
way to obtain copy-on-write semantics of a "backing image" for a Wasm
memory is to mmap a file with `MAP_PRIVATE`. The `memfd` mechanism
provided by the Linux kernel allows us to create anonymous,
in-memory-only files that we can use for this mapping, so we can
construct the image contents on-the-fly then effectively create a CoW
overlay. Furthermore, and importantly, `madvise(MADV_DONTNEED, ...)`
will discard the CoW overlay, returning the mapping to its original
state.
By itself this is almost enough for a very fast
instantiation-termination loop of the same image over and over,
without changing the address space mapping at all (which is
expensive). The only missing bit is how to implement
heap *growth*. But here memfds can help us again: if we create another
anonymous file and map it where the extended parts of the heap would
go, we can take advantage of the fact that a `mmap()` mapping can
be *larger than the file itself*, with accesses beyond the end
generating a `SIGBUS`, and the fact that we can cheaply resize the
file with `ftruncate`, even after a mapping exists. So we can map the
"heap extension" file once with the maximum memory-slot size and grow
the memfd itself as `memory.grow` operations occur.
The above CoW technique and heap-growth technique together allow us a
fastpath of `madvise()` and `ftruncate()` only when we re-instantiate
the same module over and over, as long as we can reuse the same
slot. This fastpath avoids all whole-process address-space locks in
the Linux kernel, which should mean it is highly scalable. It also
avoids the cost of copying data on read, as the `uffd` heap backend
does when servicing pagefaults; the kernel's own optimized CoW
logic (same as used by all file mmaps) is used instead.
[1] https://bytecodealliance.zulipchat.com/#narrow/stream/206238-general/topic/Copy.20on.20write.20based.20instance.20reuse/near/266657772
* Don't copy `VMBuiltinFunctionsArray` into each `VMContext`
This is another PR along the lines of "let's squeeze all possible
performance we can out of instantiation". Before this PR we would copy,
by value, the contents of `VMBuiltinFunctionsArray` into each
`VMContext` allocated. This array of function pointers is modestly-sized
but growing over time as we add various intrinsics. Additionally it's
the exact same for all `VMContext` allocations.
This PR attempts to speed up instantiation slightly by instead storing
an indirection to the function array. This means that calling a builtin
intrinsic is a tad bit slower since it requires two loads instead of one
(one to get the base pointer, another to get the actual address).
Otherwise though `VMContext` initialization is now simply setting one
pointer instead of doing a `memcpy` from one location to another.
With some macro-magic this commit also replaces the previous
implementation with one that's more `const`-friendly which also gets us
compile-time type-checks of libcalls as well as compile-time
verification that all libcalls are defined.
Overall, as with #3739, the win is very modest here. Locally I measured
a speedup from 1.9us to 1.7us taken to instantiate an empty module with
one function. While small at these scales it's still a 10% improvement!
* Review comments
This commit updates the allocation of a `VMExternRefActivationsTable`
structure to perform zero malloc memory allocations. Previously it would
allocate a page-size of `chunk` plus some space in hash sets for future
insertions. The main trick here implemented is that after the first gc
during the slow path the fast chunk allocation is allocated and
configured.
The motivation for this PR is that given our recent work to further
refine and optimize the instantiation process this allocation started to
show up in a nontrivial fashion. Most modules today never touch this
table anyway as almost none of them use reference types, so the time
spent allocation and deallocating the table per-store was largely wasted
time.
Concretely on a microbenchmark this PR speeds up instantiation of a
module with one function by 30%, decreasing the instantiation cost from
1.8us to 1.2us. Overall a pretty minor win but when the instantiation
times we're measuring start being in the single-digit microseconds this
win ends up getting magnified!
This will make it generate `table.set`s that come from `global.get`s and
`global.get`s that come from `table.set`s. Ultimately, it should give us much
more fuzzing coverage of `externref` globals, their barriers, and passing
`externref`s into and out of Wasm to get or set globals.
When we GC, we assert the invariant that all `externref`s we find on the stack
have a corresponding entry in the `VMExternRefActivationsTable`. However, we
also might be in code that is in the process of fixing up this invariant and
adding an entry to the table, but the table's bump chunk is full, and so we do a
GC and then add the entry into the table. This will ultimately maintain our
desired invariant, but there is a moment in time when we are doing the GC where
the invariant is relaxed which is okay because the reference will be in the
table before we return to Wasm or do anything else. This isn't a possible UAF,
in other words. To make it so that the assertion won't trip, we explicitly
insert the reference into the table *before* we GC, so that the invariant is not
relaxed across a possibly-GCing operation (even though it would be safe in this
particular case).
* Lazily load types into `Func`
This commit changes the construction of a `Func` to lazily load the type
information if required instead of always loading the type information
at `Func`-construction time. The main purpose of this change is to
accelerate instantiation of instances which have many imports. Currently
in the fast way of doing this the instantiation loop looks like:
let mut store = Store::new(&engine, ...);
let instance = instance_pre.instantiate(&mut store);
In this situation the `instance_pre` will typically load host-defined
functions (defined via `Linker` APIs) into the `Store` as individual
`Func` items and then perform the instantiation process. The operation
of loading a `HostFunc` into a `Store` however currently involves two
expensive operations:
* First a read-only lock is taken on the `RwLock` around engine
signatures.
* Next a clone of the wasm type is made to pull it out of the engine
signature registry.
Neither of these is actually necessary for imported functions. The
`FuncType` for imported functions is never used since all comparisons
happen with the intern'd indices instead. The only time a `FuncType` is
used typically is for exported functions when using `Func::typed` or
similar APIs which need type information.
This commit makes this path faster by storing `Option<FuncType>` instead
of `FuncType` within a `Func`. This means that it starts out as `None`
and is only filled in on-demand as necessary. This means that when
instantiating a module with many imports no clones/locks are done.
On a simple microbenchmark where a module with 100 imports is
instantiated this PR improves instantiation time by ~35%. Due to the
rwlock used here and the general inefficiency of pthreads rwlocks the
effect is even more profound when many threads are performing the same
instantiation process. On x86_64 with 8 threads performing instantiation
this PR improves instantiation time by 80% and on arm64 it improves by
97% (wow read-contended glibc rwlocks on arm64 are slow).
Note that much of the improvement here is also from memory
allocatoins/deallocations no longer being performed because dropping
functions within a store no longer requires deallocating the `FuncType`
if it's not present.
A downside of this PR is that `Func::ty` is now unconditionally taking
an rwlock if the type hasn't already been filled in. (it uses the
engine). If this is an issue in the future though we can investigate at
that time using somthing like a `Once` to lazily fill in even when
mutable access to the store isn't available.
* Review comments
WASI doesn't have an `isatty` ioctl or syscall, so wasi-libc's `isatty`
implementation uses the file descriptor type and rights to determine if
the file descriptor is likely to be a tty. The real fix here will be to
add an `isatty` call to WASI. But for now, have Wasmtime set the
filetype and rights for file descriptors so that wasi-libc's `isatty`
works as expected.
`ptr::cast` has the advantage of being unable to silently cast
`*const T` to `*mut T`. This turned up several places that were
performing such casts, which this PR also fixes.
I, after-the-fact, now recall that we document fuel and other
interruption schemes in the `Config::async_support` documentation so
I've re-worded that section to mention epoch-based interruption as well.
This PR introduces a new way of performing cooperative timeslicing that
is intended to replace the "fuel" mechanism. The tradeoff is that this
mechanism interrupts with less precision: not at deterministic points
where fuel runs out, but rather when the Engine enters a new epoch. The
generated code instrumentation is substantially faster, however, because
it does not need to do as much work as when tracking fuel; it only loads
the global "epoch counter" and does a compare-and-branch at backedges
and function prologues.
This change has been measured as ~twice as fast as fuel-based
timeslicing for some workloads, especially control-flow-intensive
workloads such as the SpiderMonkey JS interpreter on Wasm/WASI.
The intended interface is that the embedder of the `Engine` performs an
`engine.increment_epoch()` call periodically, e.g. once per millisecond.
An async invocation of a Wasm guest on a `Store` can specify a number of
epoch-ticks that are allowed before an async yield back to the
executor's event loop. (The initial amount and automatic "refills" are
configured on the `Store`, just as for fuel.) This call does only
signal-safe work (it increments an `AtomicU64`) so could be invoked from
a periodic signal, or from a thread that wakes up once per period.
* Provide helpers for demangling function names
* Profile trampolines in vtune too
* get rid of mapping
* avoid code duplication with jitdump_linux
* maintain previous default display name for wasm functions
* no dash, grrr
* Remove unused profiling error type
Before this PR, each profiler (perf/vtune, at the moment) had to have a
demangler for each of the programming languages that could have been
compiled to wasm and fed into wasmtime. With this, wasmtime now
demangles names before even forwarding them to the underlying profiler,
which makes for a unified representation in profilers, and avoids
incorrect demangling in profilers.
This is unconditionally used on aarch64 and otherwise trying to manage
the precise clause for making it conditional vs unconditional probably
isn't worth it.
Closes#3672
I forgot in the recent refactoring to add back in fuel support to the
`table_ops` fuzzer. This commit re-adds the previously existent logic to
always use fuel to cancel execution of the table_ops fuzzer.
