T0b1/wasmtime
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Implement the memory64 proposal in Wasmtime (#3153)

Browse Source 
* Implement the memory64 proposal in Wasmtime

This commit implements the WebAssembly [memory64 proposal][proposal] in
both Wasmtime and Cranelift. In terms of work done Cranelift ended up
needing very little work here since most of it was already prepared for
64-bit memories at one point or another. Most of the work in Wasmtime is
largely refactoring, changing a bunch of `u32` values to something else.

A number of internal and public interfaces are changing as a result of
this commit, for example:

* Acessors on `wasmtime::Memory` that work with pages now all return
  `u64` unconditionally rather than `u32`. This makes it possible to
  accommodate 64-bit memories with this API, but we may also want to
  consider `usize` here at some point since the host can't grow past
  `usize`-limited pages anyway.

* The `wasmtime::Limits` structure is removed in favor of
  minimum/maximum methods on table/memory types.

* Many libcall intrinsics called by jit code now unconditionally take
  `u64` arguments instead of `u32`. Return values are `usize`, however,
  since the return value, if successful, is always bounded by host
  memory while arguments can come from any guest.

* The `heap_addr` clif instruction now takes a 64-bit offset argument
  instead of a 32-bit one. It turns out that the legalization of
  `heap_addr` already worked with 64-bit offsets, so this change was
  fairly trivial to make.

* The runtime implementation of mmap-based linear memories has changed
  to largely work in `usize` quantities in its API and in bytes instead
  of pages. This simplifies various aspects and reflects that
  mmap-memories are always bound by `usize` since that's what the host
  is using to address things, and additionally most calculations care
  about bytes rather than pages except for the very edge where we're
  going to/from wasm.

Overall I've tried to minimize the amount of `as` casts as possible,
using checked `try_from` and checked arithemtic with either error
handling or explicit `unwrap()` calls to tell us about bugs in the
future. Most locations have relatively obvious things to do with various
implications on various hosts, and I think they should all be roughly of
the right shape but time will tell. I mostly relied on the compiler
complaining that various types weren't aligned to figure out
type-casting, and I manually audited some of the more obvious locations.
I suspect we have a number of hidden locations that will panic on 32-bit
hosts if 64-bit modules try to run there, but otherwise I think we
should be generally ok (famous last words). In any case I wouldn't want
to enable this by default naturally until we've fuzzed it for some time.

In terms of the actual underlying implementation, no one should expect
memory64 to be all that fast. Right now it's implemented with
"dynamic" heaps which have a few consequences:

* All memory accesses are bounds-checked. I'm not sure how aggressively
  Cranelift tries to optimize out bounds checks, but I suspect not a ton
  since we haven't stressed this much historically.

* Heaps are always precisely sized. This means that every call to
  `memory.grow` will incur a `memcpy` of memory from the old heap to the
  new. We probably want to at least look into `mremap` on Linux and
  otherwise try to implement schemes where dynamic heaps have some
  reserved pages to grow into to help amortize the cost of
  `memory.grow`.

The memory64 spec test suite is scheduled to now run on CI, but as with
all the other spec test suites it's really not all that comprehensive.
I've tried adding more tests for basic things as I've had to implement
guards for them, but I wouldn't really consider the testing adequate
from just this PR itself. I did try to take care in one test to actually
allocate a 4gb+ heap and then avoid running that in the pooling
allocator or in emulation because otherwise that may fail or take
excessively long.

[proposal]: https://github.com/WebAssembly/memory64/blob/master/proposals/memory64/Overview.md

* Fix some tests

* More test fixes

* Fix wasmtime tests

* Fix doctests

* Revert to 32-bit immediate offsets in `heap_addr`

This commit updates the generation of addresses in wasm code to always
use 32-bit offsets for `heap_addr`, and if the calculated offset is
bigger than 32-bits we emit a manual add with an overflow check.

* Disable memory64 for spectest fuzzing

* Fix wrong offset being added to heap addr

* More comments!

* Clarify bytes/pages
This commit is contained in:
Alex Crichton
2021-08-12 09:40:20 -05:00
committed by GitHub
parent 76a93dc112
commit e68aa99588
60 changed files with 1356 additions and 635 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
cranelift/wasm/src/code_translator.rs
View File

@@ -2164,10 +2164,6 @@ fn prepare_addr<FE: FuncEnvironment + ?Sized>(
environ: &mut FE,
) -> WasmResult<(MemFlags, Value, Offset32)> {
let addr = state.pop1();
// This function will need updates for 64-bit memories
debug_assert_eq!(builder.func.dfg.value_type(addr), I32);
let offset = u32::try_from(memarg.offset).unwrap();
let heap = state.get_heap(builder.func, memarg.memory, environ)?;
let offset_guard_size: u64 = builder.func.heaps[heap].offset_guard_size.into();
@@ -2176,13 +2172,19 @@ fn prepare_addr<FE: FuncEnvironment + ?Sized>(
// segfaults) to generate traps since that means we don't have to bounds
// check anything explicitly.
//
// If we don't have a guard page of unmapped memory, though, then we can't
// rely on this trapping behavior through segfaults. Instead we need to
// bounds-check the entire memory access here which is everything from
// (1) If we don't have a guard page of unmapped memory, though, then we
// can't rely on this trapping behavior through segfaults. Instead we need
// to bounds-check the entire memory access here which is everything from
// `addr32 + offset` to `addr32 + offset + width` (not inclusive). In this
// scenario our adjusted offset that we're checking is `offset + width`.
// scenario our adjusted offset that we're checking is `memarg.offset +
// access_size`. Note that we do saturating arithmetic here to avoid
// overflow. THe addition here is in the 64-bit space, which means that
// we'll never overflow for 32-bit wasm but for 64-bit this is an issue. If
// our effective offset is u64::MAX though then it's impossible for for
// that to actually be a valid offset because otherwise the wasm linear
// memory would take all of the host memory!
//
// If we have a guard page, however, then we can perform a further
// (2) If we have a guard page, however, then we can perform a further
// optimization of the generated code by only checking multiples of the
// offset-guard size to be more CSE-friendly. Knowing that we have at least
// 1 page of a guard page we're then able to disregard the `width` since we
@@ -2215,32 +2217,104 @@ fn prepare_addr<FE: FuncEnvironment + ?Sized>(
// in-bounds or will hit the guard page, meaning we'll get the desired
// semantics we want.
//
// As one final comment on the bits with the guard size here, another goal
// of this is to hit an optimization in `heap_addr` where if the heap size
// minus the offset is >= 4GB then bounds checks are 100% eliminated. This
// means that with huge guard regions (e.g. our 2GB default) most adjusted
// offsets we're checking here are zero. This means that we'll hit the fast
// path and emit zero conditional traps for bounds checks
// ---
//
// With all that in mind remember that the goal is to bounds check as few
// things as possible. To facilitate this the "fast path" is expected to be
// hit like so:
//
// * For wasm32, wasmtime defaults to 4gb "static" memories with 2gb guard
// regions. This means our `adjusted_offset` is 1 for all offsets <=2gb.
// This hits the optimized case for `heap_addr` on static memories 4gb in
// size in cranelift's legalization of `heap_addr`, eliding the bounds
// check entirely.
//
// * For wasm64 offsets <=2gb will generate a single `heap_addr`
// instruction, but at this time all heaps are "dyanmic" which means that
// a single bounds check is forced. Ideally we'd do better here, but
// that's the current state of affairs.
//
// Basically we assume that most configurations have a guard page and most
// offsets in `memarg` are <=2gb, which means we get the fast path of one
// `heap_addr` instruction plus a hardcoded i32-offset in memory-related
// instructions.
let adjusted_offset = if offset_guard_size == 0 {
u64::from(offset) + u64::from(access_size)
// Why saturating? see (1) above
memarg.offset.saturating_add(u64::from(access_size))
} else {
// Why is there rounding here? see (2) above
assert!(access_size < 1024);
cmp::max(u64::from(offset) / offset_guard_size * offset_guard_size, 1)
cmp::max(memarg.offset / offset_guard_size * offset_guard_size, 1)
};
debug_assert!(adjusted_offset > 0); // want to bounds check at least 1 byte
let check_size = u32::try_from(adjusted_offset).unwrap_or(u32::MAX);
let base = builder
.ins()
.heap_addr(environ.pointer_type(), heap, addr, check_size);
// Native load/store instructions take a signed `Offset32` immediate, so adjust the base
// pointer if necessary.
let (addr, offset) = if offset > i32::MAX as u32 {
// Offset doesn't fit in the load/store instruction.
let adj = builder.ins().iadd_imm(base, i64::from(i32::MAX) + 1);
(adj, (offset - (i32::MAX as u32 + 1)) as i32)
} else {
(base, offset as i32)
debug_assert!(adjusted_offset > 0); // want to bounds check at least 1 byte
let (addr, offset) = match u32::try_from(adjusted_offset) {
// If our adjusted offset fits within a u32, then we can place the
// entire offset into the offset of the `heap_addr` instruction. After
// the `heap_addr` instruction, though, we need to factor the the offset
// into the returned address. This is either an immediate to later
// memory instructions if the offset further fits within `i32`, or a
// manual add instruction otherwise.
//
// Note that native instructions take a signed offset hence the switch
// to i32. Note also the lack of overflow checking in the offset
// addition, which should be ok since if `heap_addr` passed we're
// guaranteed that this won't overflow.
Ok(adjusted_offset) => {
let base = builder
.ins()
.heap_addr(environ.pointer_type(), heap, addr, adjusted_offset);
match i32::try_from(memarg.offset) {
Ok(val) => (base, val),
Err(_) => {
let adj = builder.ins().iadd_imm(base, memarg.offset as i64);
(adj, 0)
}
}
}
// If the adjusted offset doesn't fit within a u32, then we can't pass
// the adjust sized to `heap_addr` raw.
//
// One reasonable question you might ask is "why not?". There's no
// fundamental reason why `heap_addr` *must* take a 32-bit offset. The
// reason this isn't done, though, is that blindly changing the offset
// to a 64-bit offset increases the size of the `InstructionData` enum
// in cranelift by 8 bytes (16 to 24). This can have significant
// performance implications so the conclusion when this was written was
// that we shouldn't do that.
//
// Without the ability to put the whole offset into the `heap_addr`
// instruction we need to fold the offset into the address itself with
// an unsigned addition. In doing so though we need to check for
// overflow because that would mean the address is out-of-bounds (wasm
// bounds checks happen on the effective 33 or 65 bit address once the
// offset is factored in).
//
// Once we have the effective address, offset already folded in, then
// `heap_addr` is used to verify that the address is indeed in-bounds.
// The access size of the `heap_addr` is what we were passed in from
// above.
//
// Note that this is generating what's likely to be at least two
// branches, one for the overflow and one for the bounds check itself.
// For now though that should hopefully be ok since 4gb+ offsets are
// relatively odd/rare. In the future if needed we can look into
// optimizing this more.
Err(_) => {
let index_type = builder.func.heaps[heap].index_type;
let offset = builder.ins().iconst(index_type, memarg.offset as i64);
let (addr, overflow) = builder.ins().iadd_ifcout(addr, offset);
builder.ins().trapif(
environ.unsigned_add_overflow_condition(),
overflow,
ir::TrapCode::HeapOutOfBounds,
);
let base = builder
.ins()
.heap_addr(environ.pointer_type(), heap, addr, access_size);
(base, 0)
}
};
// Note that we don't set `is_aligned` here, even if the load instruction's