Call the OS function to set the main thread's name on program init
Normally, `Thread::spawn` takes care of setting the thread's name, if
one was provided, but since the main thread wasn't created by calling
`Thread::spawn`, we need to call that function in `std::rt::init`.
This is mainly useful for system tools like debuggers and profilers
which might show the thread name to a user. Prior to these changes, gdb
and WinDbg would show all thread names except the main thread's name to
a user. I've validated that this patch resolves the issue for both
debuggers.
Lazily allocate and initialize pthread locks.
Lazily allocate and initialize pthread locks.
This allows {Mutex, Condvar, RwLock}::new() to be const, while still using the platform's native locks for features like priority inheritance and debug tooling. E.g. on macOS, we cannot directly use the (private) APIs that pthread's locks are implemented with, making it impossible for us to use anything other than pthread while still preserving priority inheritance, etc.
This PR doesn't yet make the public APIs const. That's for a separate PR with an FCP.
Tracking issue: https://github.com/rust-lang/rust/issues/93740
The oldest occurence is from 9e224c2bf1,
which is from the pre-1.0 days. In the years since then, std::sys still
hasn't been exported, and the last attempt was met with strong criticism:
https://github.com/rust-lang/rust/pull/97151
Thus, removing the "yet" part makes a lot of sense.
Normally, `Thread::spawn` takes care of setting the thread's name, if
one was provided, but since the main thread wasn't created by calling
`Thread::spawn`, we need to call that function in `std::rt::init`.
This is mainly useful for system tools like debuggers and profilers
which might show the thread name to a user. Prior to these changes, gdb
and WinDbg would show all thread names except the main thread's name to
a user. I've validated that this patch resolves the issue for both
debuggers.
I wrote `state` where I should've used `s`.
This removes the unnecessary `s` variable to prevent that mistake.
Fortunately, this typo didn't affect the correctness of the lock, as the
second half of the condition (!has_writers_waiting) is enough for
correctness, which explains why this mistake didn't show up during
testing.
From reading the source code, it appears like the desired semantic of
std::unix::rand is to always provide some bytes and never block. For
that reason GRND_NONBLOCK is checked before calling getrandom(0), so
that getrandom(0) won't block. If it would block, then the function
falls back to using /dev/urandom, which for the time being doesn't
block. There are some drawbacks to using /dev/urandom, however, and so
getrandom(GRND_INSECURE) was created as a replacement for this exact
circumstance.
getrandom(GRND_INSECURE) is the same as /dev/urandom, except:
- It won't leave a warning in dmesg if used at early boot time, which is
a common occurance (and the reason why I found this issue);
- It won't introduce a tiny delay at early boot on newer kernels when
/dev/urandom tries to opportunistically create jitter entropy;
- It only requires 1 syscall, rather than 3.
Other than that, it returns the same "quality" of randomness as
/dev/urandom, and never blocks.
It's only available on kernels ≥5.6, so we try to use it, cache the
result of that attempt, and fall back to to the previous code if it
didn't work.
Make HashMap fall back to RtlGenRandom if BCryptGenRandom fails
With PR #84096, Rust `std::collections::hash_map::RandomState` changed from using `RtlGenRandom()` ([msdn](https://docs.microsoft.com/en-us/windows/win32/api/ntsecapi/nf-ntsecapi-rtlgenrandom)) to `BCryptGenRandom()` ([msdn](https://docs.microsoft.com/en-us/windows/win32/api/bcrypt/nf-bcrypt-bcryptgenrandom)) as its source of secure randomness after much discussion ([here](https://github.com/rust-random/getrandom/issues/65#issuecomment-753634074), among other places).
Unfortunately, after that PR landed, Mozilla Firefox started experiencing fairly-rare crashes during startup while attempting to initialize the `env_logger` crate. ([docs for env_logger](https://docs.rs/env_logger/latest/env_logger/)) The root issue is that on some machines, `BCryptGenRandom()` will fail with an `Access is denied. (os error 5)` error message. ([Bugzilla issue 1754490](https://bugzilla.mozilla.org/show_bug.cgi?id=1754490)) (Discussion in issue #94098)
Note that this is happening upon startup of Firefox's unsandboxed Main Process, so this behavior is different and separate from previous issues ([like this](https://bugzilla.mozilla.org/show_bug.cgi?id=1746254)) where BCrypt DLLs were blocked by process sandboxing. In the case of sandboxing, we knew we were doing something abnormal and expected that we'd have to resort to abnormal measures to make it work.
However, in this case we are in a regular unsandboxed process just trying to initialize `env_logger` and getting a panic. We suspect that this may be caused by a virus scanner or some other security software blocking the loading of the BCrypt DLLs, but we're not completely sure as we haven't been able to replicate locally.
It is also possible that Firefox is not the only software affected by this; we just may be one of the pieces of Rust software that has the telemetry and crash reporting necessary to catch it.
I have read some of the historical discussion around using `BCryptGenRandom()` in Rust code, and I respect the decision that was made and agree that it was a good course of action, so I'm not trying to open a discussion about a return to `RtlGenRandom()`. Instead, I'd like to suggest that perhaps we use `RtlGenRandom()` as a "fallback RNG" in the case that BCrypt doesn't work.
This pull request implements this fallback behavior. I believe this would improve the robustness of this essential data structure within the standard library, and I see only 2 potential drawbacks:
1. Slight added overhead: It should be quite minimal though. The first call to `sys::rand::hashmap_random_keys()` will incur a bit of initialization overhead, and every call after will incur roughly 2 non-atomic global reads and 2 easily predictable branches. Both should be negligible compared to the actual cost of generating secure random numbers
2. `RtlGenRandom()` is deprecated by Microsoft: Technically true, but as mentioned in [this comment on GoLang](https://github.com/golang/go/issues/33542#issuecomment-626124873), this API is ubiquitous in Windows software and actually removing it would break lots of things. Also, Firefox uses it already in [our C++ code](https://searchfox.org/mozilla-central/rev/5f88c1d6977e03e22d3420d0cdf8ad0113c2eb31/mfbt/RandomNum.cpp#25), and [Chromium uses it in their code as well](https://source.chromium.org/chromium/chromium/src/+/main:base/rand_util_win.cc) (which transitively means that Microsoft uses it in their own web browser, Edge). If there did come a time when Microsoft truly removes this API, it should be easy enough for Rust to simply remove the fallback in the code I've added here
Issue #84096 changed the hashmap RNG to use BCryptGenRandom instead of
RtlGenRandom on Windows.
Mozilla Firefox started experiencing random failures in
env_logger::Builder::new() (Issue #94098) during initialization of their
unsandboxed main process with an "Access Denied" error message from
BCryptGenRandom(), which is used by the HashMap contained in
env_logger::Builder
The root cause appears to be a virus scanner or other software interfering
with BCrypt DLLs loading.
This change adds a fallback option if BCryptGenRandom is unusable for
whatever reason. It will fallback to RtlGenRandom in this case.
Fixes#94098
Windows: Make stdin pipes synchronous
Stdin pipes do not need to be used asynchronously within the standard library. This is a first step in making pipes mostly synchronous.
r? `@m-ou-se`
std: directly use pthread in UNIX parker implementation
`Mutex` and `Condvar` are being replaced by more efficient implementations, which need thread parking themselves (see #93740). Therefore we should use the `pthread` synchronization primitives directly. Also, we can avoid allocating the mutex and condition variable because the `Parker` struct is being placed in an `Arc` anyways.
This basically is just a copy of the current `Mutex` and `Condvar` code, which will however be removed (again, see #93740). An alternative implementation could be to use dedicated private `OsMutex` and `OsCondvar` types, but all the other platforms supported by std actually have their own thread parking primitives.
I used `Pin` to guarantee a stable address for the `Parker` struct, while the current implementation does not, rather using extra unsafe declaration. Since the thread struct is shared anyways, I assumed this would not add too much clutter while being clearer.
