##// END OF EJS Templates
windows: degrade to py2 behavior when reading a non-symlink as a symlink...
windows: degrade to py2 behavior when reading a non-symlink as a symlink While waiting for the push to hg-committed in WSL to complete, I ran a `phabimport` from Windows and got this traceback: $ hg phabimport 11313 ** Unknown exception encountered with possibly-broken third-party extension "mercurial_keyring" (version N/A) ** which supports versions unknown of Mercurial. ** Please disable "mercurial_keyring" and try your action again. ** If that fixes the bug please report it to https://foss.heptapod.net/mercurial/mercurial_keyring/issues ** Python 3.9.5 (default, May 6 2021, 17:29:31) [MSC v.1928 64 bit (AMD64)] ** Mercurial Distributed SCM (version 5.9rc1+hg32.0e2f5733563d) ** Extensions loaded: absorb, blackbox, evolve 10.3.3, extdiff, fastannotate, fix, mercurial_keyring, mq, phabblocker 20210126, phabricator, rebase, show, strip, topic 0.22.3 Traceback (most recent call last): File "mercurial.lock", line 279, in _trylock File "mercurial.vfs", line 202, in makelock File "mercurial.util", line 2147, in makelock FileExistsError: [WinError 183] Cannot create a file when that file already exists: b'hp-omen:78348' -> b'C:\\Users\\Matt\\hg/.hg/store/lock' During handling of the above exception, another exception occurred: Traceback (most recent call last): File "<string>", line 24, in <module> File "mercurial.dispatch", line 144, in run File "mercurial.dispatch", line 250, in dispatch File "mercurial.dispatch", line 294, in _rundispatch File "mercurial.dispatch", line 470, in _runcatch File "mercurial.dispatch", line 480, in _callcatch File "mercurial.scmutil", line 153, in callcatch File "mercurial.dispatch", line 460, in _runcatchfunc File "mercurial.dispatch", line 1273, in _dispatch File "mercurial.dispatch", line 918, in runcommand File "mercurial.dispatch", line 1285, in _runcommand File "mercurial.dispatch", line 1271, in <lambda> File "mercurial.util", line 1886, in check File "mercurial.util", line 1886, in check File "hgext.mq", line 4239, in mqcommand File "mercurial.util", line 1886, in check File "mercurial.util", line 1886, in check File "hgext.phabricator", line 314, in inner File "hgext.phabricator", line 2222, in phabimport File "hgext.phabricator", line 2123, in readpatch File "hgext.phabricator", line 2199, in _write File "mercurial.localrepo", line 2956, in lock File "mercurial.localrepo", line 2918, in _lock File "mercurial.lock", line 152, in trylock File "mercurial.lock", line 283, in _trylock File "mercurial.lock", line 314, in _readlock File "mercurial.vfs", line 221, in readlock File "mercurial.util", line 2163, in readlock File "mercurial.windows", line 619, in readlink ValueError: not a symbolic link Both exceptions look accurate (the file exists, and the Windows side can't read WSL side symlinks). I didn't try to reproduce this entirely within the Windows side, but we can do better than a cryptic stacktrace. With this change, the same scenario results in this abort: abort: C:\Users\Matt\hg/.hg/store/lock: The file cannot be accessed by the system When both the `push` and `phabimport` are done on the Windows side, it prints a message about waiting for the lock, and successfully applies the patch after the push completes. I'm not sure if there's enough info to be able to convert the abort into the wait scenario. As it stands now, we don't support symlinks on Windows, which requires either a UAC Administrator level process or an opt-in in developer mode, and there are several places where the new symlink on Windows support in py3 was explicitly disabled in order to get tests to pass quicker. Differential Revision: https://phab.mercurial-scm.org/D11333

File last commit:

r48482:78f7f0d4 default
r48680:4162f6b4 stable
Show More
owning.rs
117 lines | 4.8 KiB | application/rls-services+xml | RustLexer
use cpython::PyBytes;
use cpython::Python;
use hg::dirstate_tree::dirstate_map::DirstateMap;
use hg::DirstateError;
use hg::DirstateParents;
/// Keep a `DirstateMap<'on_disk>` next to the `on_disk` buffer that it
/// borrows. This is similar to the owning-ref crate.
///
/// This is similar to [`OwningRef`] which is more limited because it
/// represents exactly one `&T` reference next to the value it borrows, as
/// opposed to a struct that may contain an arbitrary number of references in
/// arbitrarily-nested data structures.
///
/// [`OwningRef`]: https://docs.rs/owning_ref/0.4.1/owning_ref/struct.OwningRef.html
pub(super) struct OwningDirstateMap {
/// Owned handle to a bytes buffer with a stable address.
///
/// See <https://docs.rs/owning_ref/0.4.1/owning_ref/trait.StableAddress.html>.
on_disk: PyBytes,
/// Pointer for `Box<DirstateMap<'on_disk>>`, typed-erased because the
/// language cannot represent a lifetime referencing a sibling field.
/// This is not quite a self-referencial struct (moving this struct is not
/// a problem as it doesn’t change the address of the bytes buffer owned
/// by `PyBytes`) but touches similar borrow-checker limitations.
ptr: *mut (),
}
impl OwningDirstateMap {
pub fn new_v1(
py: Python,
on_disk: PyBytes,
) -> Result<(Self, Option<DirstateParents>), DirstateError> {
let bytes: &'_ [u8] = on_disk.data(py);
let (map, parents) = DirstateMap::new_v1(bytes)?;
// Like in `bytes` above, this `'_` lifetime parameter borrows from
// the bytes buffer owned by `on_disk`.
let ptr: *mut DirstateMap<'_> = Box::into_raw(Box::new(map));
// Erase the pointed type entirely in order to erase the lifetime.
let ptr: *mut () = ptr.cast();
Ok((Self { on_disk, ptr }, parents))
}
pub fn new_v2(
py: Python,
on_disk: PyBytes,
data_size: usize,
tree_metadata: PyBytes,
) -> Result<Self, DirstateError> {
let bytes: &'_ [u8] = on_disk.data(py);
let map =
DirstateMap::new_v2(bytes, data_size, tree_metadata.data(py))?;
// Like in `bytes` above, this `'_` lifetime parameter borrows from
// the bytes buffer owned by `on_disk`.
let ptr: *mut DirstateMap<'_> = Box::into_raw(Box::new(map));
// Erase the pointed type entirely in order to erase the lifetime.
let ptr: *mut () = ptr.cast();
Ok(Self { on_disk, ptr })
}
pub fn get_mut<'a>(&'a mut self) -> &'a mut DirstateMap<'a> {
// SAFETY: We cast the type-erased pointer back to the same type it had
// in `new`, except with a different lifetime parameter. This time we
// connect the lifetime to that of `self`. This cast is valid because
// `self` owns the same `PyBytes` whose buffer `DirstateMap`
// references. That buffer has a stable memory address because the byte
// string value of a `PyBytes` is immutable.
let ptr: *mut DirstateMap<'a> = self.ptr.cast();
// SAFETY: we dereference that pointer, connecting the lifetime of the
// new `&mut` to that of `self`. This is valid because the
// raw pointer is to a boxed value, and `self` owns that box.
unsafe { &mut *ptr }
}
pub fn get<'a>(&'a self) -> &'a DirstateMap<'a> {
// SAFETY: same reasoning as in `get_mut` above.
let ptr: *mut DirstateMap<'a> = self.ptr.cast();
unsafe { &*ptr }
}
}
impl Drop for OwningDirstateMap {
fn drop(&mut self) {
// Silence a "field is never read" warning, and demonstrate that this
// value is still alive.
let _ = &self.on_disk;
// SAFETY: this cast is the same as in `get_mut`, and is valid for the
// same reason. `self.on_disk` still exists at this point, drop glue
// will drop it implicitly after this `drop` method returns.
let ptr: *mut DirstateMap<'_> = self.ptr.cast();
// SAFETY: `Box::from_raw` takes ownership of the box away from `self`.
// This is fine because drop glue does nothig for `*mut ()` and we’re
// in `drop`, so `get` and `get_mut` cannot be called again.
unsafe { drop(Box::from_raw(ptr)) }
}
}
fn _static_assert_is_send<T: Send>() {}
fn _static_assert_fields_are_send() {
_static_assert_is_send::<PyBytes>();
_static_assert_is_send::<Box<DirstateMap<'_>>>();
}
// SAFETY: we don’t get this impl implicitly because `*mut (): !Send` because
// thread-safety of raw pointers is unknown in the general case. However this
// particular raw pointer represents a `Box<DirstateMap<'on_disk>>` that we
// own. Since that `Box` and `PyBytes` are both `Send` as shown in above, it
// is sound to mark this struct as `Send` too.
unsafe impl Send for OwningDirstateMap {}