During garbage collection we cache several things -- a set of known-dead
paths, a set of known-alive paths, and a map of paths to their derivers.
Currently they use STL maps and sets, which are ordered structures that
typically are backed by binary trees. Since we are putting pseudorandom
paths into these and looking them up by exact key, we don't need the
ordering, and we're paying a nontrivial cost per insertion.
The existing maps require O(n log n) memory and have O(log n) insertion
and lookup time.
We could instead use unordered maps, which are typically backed by
hashmaps. These require O(n) memory and have O(1) insertion and lookup
time.
On my system this appears to result in a dramatic speedup -- prior to
this patch I was able to delete 400k paths out of 9.5 million over the
course of 34.5 hours. After this patch the same result took 89 minutes.
This result should NOT be taken at face value because the two runs
aren't really comparable; in particular the first started when I had 9.5
million store paths and the seconcd started with 7.8 million, so we are
deleting a different set of paths starting from a much cleaner
filesystem. But I do think it's indicative.
Related: https://github.com/NixOS/nix/issues/9581
First the motivation: I recently faced a bug that I assume is coming
from the topoSortPaths function where the GC was trying to delete a
path having some alive referrers. I resolved this by manually deleting
the faulty path referrers using nix-store --query --referrers. I sadly
did not manage to reproduce this bug.
This bug alone is not a big deal. However, this bug is
triggering a cascading failure: invalidatePathChecked is throwing a
PathInUse exception. This exception is not catched and fails the whole GC
run. From there, the machine (a builder machine) was unable to GC its
Nix store, which led to an almost full disk with no way to
automatically delete the dead Nix paths.
Instead, I think we should log the error for the specific store path
we're trying to delete, specifying we can't delete this path because
it still has referrers. Once we're done with logging that, the GC run
should continue to delete the dead store paths it can delete.
We're not replacing `Path` in exposed definitions in many cases, but
just adding alternatives. This will allow us to "top down" change `Path`
to `std::fileysystem::path`, and then we can remove the `Path`-using
utilities which will become unused.
Also add some test files which we forgot to include in the libutil unit
tests `meson.build`.
Co-Authored-By: siddhantCodes <siddhantk232@gmail.com>
On several occasions I've found myself confused when trying to delete
a store path, because I am told it's still alive, but
nix-store --query --roots doesn't show anything. Let's save future
users this confusion by mentioning that a path might be alive due to
having referrers, not just roots.
At this point many features are stripped out, but this works:
- Can run libnix{util,store,expr} unit tests
- Can run some Nix commands
Co-Authored-By volth <volth@volth.com>
Co-Authored-By Brian McKenna <brian@brianmckenna.org>
Most of this is a `catch SysError` -> `catch SystemError` sed. This
is a rather pure-churn change I would like to get out of the way. **The
intersting part is `src/libutil/error.hh`.**
On Unix, we will only throw the `SysError` concrete class, which has
the same constructors that `SystemError` used to have.
On Windows, we will throw `WinError` *and* `SysError`. `WinError`
(which will be created in a later PR), will use a `DWORD` instead of
`int` error value, and `GetLastError()`, which is the Windows equivalent
of the `errno` machinery. Windows will *also* use `SysError` because
Window's "libc" (MSVCRT) implements the POSIX interface, and we use it
too.
As the docs describe, while we *throw* one of the 3 choices above (2
concrete classes or the alias), we should always *catch* `SystemError`.
This ensures no matter how the implementation changes for Windows (e.g.
between `SysError` and `WinError`) the catching logic stays the same
and stays correct.
Co-Authored-By volth <volth@volth.com>
Co-Authored-By Eugene Butler <eugene@eugene4.com>
All OS and IO operations should be moved out, leaving only some misc
portable pure functions.
This is useful to avoid copious CPP when doing things like Windows and
Emscripten ports.
Newly exposed functions to break cycles:
- `restoreSignals`
- `updateWindowSize`
I think it is bad for these reasons when `tests/` contains a mix of
functional and integration tests
- Concepts is harder to understand, the documentation makes a good
unit vs functional vs integration distinction, but when the
integration tests are just two subdirs within `tests/` this is not
clear.
- Source filtering in the `flake.nix` is more complex. We need to
filter out some of the dirs from `tests/`, rather than simply pick
the dirs we want and take all of them. This is a good sign the
structure of what we are trying to do is not matching the structure
of the files.
With this change we have a clean:
```shell-session
$ git show 'HEAD:tests'
tree HEAD:tests
functional/
installer/
nixos/
```
Previously it was not possible to open a local store when its database is on a read-only filesystem. Obviously a store on a read-only filesystem cannot be modified, but it would still be useful to be able to query it.
This change adds a new read-only setting to LocalStore. When set to true, Nix will skip operations that fail when the database is on a read-only filesystem (acquiring big-lock, schema migration, etc), and the store database will be opened in immutable mode.
Co-authored-by: Ben Radford <benradf@users.noreply.github.com>
Co-authored-by: cidkidnix <cidkidnix@protonmail.com>
Co-authored-by: Dylan Green <67574902+cidkidnix@users.noreply.github.com>
Co-authored-by: John Ericson <git@JohnEricson.me>
Co-authored-by: Valentin Gagarin <valentin.gagarin@tweag.io>
If the garbage collector has acquired the global GC lock, but hasn't
created the GC socket yet, then a client attempting to connect would
get ENOENT. Note that this only happens when the GC runs for the first
time on a machine. Subsequently clients will get ECONNREFUSED which
was already handled.
Fixes#7370.
I saw this random failure in https://hydra.nixos.org/build/211811692:
error: opening /proc/15307/fd: No such process
while running nix-collect-garbage in a readfile-context.sh. This is
because we're not handling ESRCH errors reading /proc/<pid>/fd. So
just move the read inside the try/catch where we do handle it.
Previously addTempRoot() acquired the LocalStore state lock and waited
for the garbage collector to reply. If the garbage collector is in the
same process (as it the case with auto-GC), this would deadlock as
soon as the garbage collector thread needs the LocalStore state lock.
So now addTempRoot() uses separate Syncs for the state that it
needs. As long at the auto-GC thread doesn't call addTempRoot() (which
it shouldn't), it shouldn't deadlock.
Fixes#3224.
This also moves the file handle into its own Sync object so we're not
holding the _state while acquiring the file lock. There was no real
deadlock risk here since locking a newly created file cannot block,
but it's still a bit nicer.
We need to close the GC server socket before shutting down the active
GC client connections, otherwise a client may (re)connect and get
ECONNRESET. But also handle ECONNRESET for resilience.
Fixes random failures like
GC socket disconnected
connecting to '/tmp/nix-shell.y07M0H/nix-test/default/var/nix/gc-socket/socket'
sending GC root '/tmp/nix-shell.y07M0H/nix-test/default/store/kb5yzija0f1x5xkqkgclrdzldxj6nnc6-non-blocking'
reading GC root from client: error: unexpected EOF reading a line
1 store paths deleted, 0.00 MiB freed
error: reading from file: Connection reset by peer
in gc-non-blocking.sh.
