We finally test the status quo of remote build trust in a number of
ways. We create a new experimental feature on `nix-daemon` to do so.
PR #3921, which improves the situation with trustless remote building,
will build upon these changes. This code / tests was pull out of there
to make this, so everything is easier to review, and in particular we
test before and after so the new behavior in that PR is readily apparent
from the testsuite diff alone.
Issues:
1. Features gated on disabled experimental settings should warn and be
ignored, not silently succeed.
2. Experimental settings in the same config "batch" (file or env var)
as the enabling of the experimental feature should work.
3. For (2), the order should not matter.
These are analogous to the issues @roberth caught with my changes for
arg handling, but they are instead for config handling.
Co-authored-by: Robert Hensing <roberth@users.noreply.github.com>
This is somewhat hacky fix just for 2.15. I unintentionally hid them
from the manual, when no one wanted to hide them that (including
myself). I also required the experimental feature to be enabled in an
order-dependent way, which is not good.
The simplest fix for this immanent release is just to always show them,
and always allow them to be set.
Effectively undoes some changes from aa663b7e89
How signals should be handled depends on what kind of process Nix
is integrated into. The signal handler thread used by the stand-alone
Nix commands / processes may not work well in the context of other
runtime systems, such as those of Python, Perl, or Haskell.
* Finish converting existing comments for internal API docs
99% of this was just reformatting existing comments. Only two exceptions:
- Expanded upon `BuildResult::status` compat note
- Split up file-level `symbol-table.hh` doc comments to get
per-definition docs
Also fixed a few whitespace goofs, turning leading tabs to spaces and
removing trailing spaces.
Picking up from #8133
* Fix two things from comments
* Use triple-backtick not indent for `dumpPath`
* Convert GNU-style `\`..'` quotes to markdown style in API docs
This will render correctly.
The `write` name is ambiguous and could lead to some funny bugs like
https://github.com/NixOS/nix/pull/8173#issuecomment-1500009480. So
rename it to the more explicit `writeUnbuffered`.
Besides, this method shouldn't be (and isn't) used outside of the class
implementation, so mark it `protected`.
This makes it more symetrical to `BufferedSource` which uses a
`protected readUnbuffered` method.
They are put in the manual separate pages under the new overarching
description of experimental features.
The settings page just lists the valid experimental feature names (so
people know what a valid setting entry looks like), with links to those
pages. It doesn't attempt to describe each experimental feature as that
is too much information for the configuration settings section.
This introduces the SourcePath type from lazy-trees as an abstraction
for accessing files from inputs that may not be materialized in the
real filesystem (e.g. Git repositories). Currently, however, it's just
a wrapper around CanonPath, so it shouldn't change any behaviour. (On
lazy-trees, SourcePath is a <InputAccessor, CanonPath> tuple.)
Instead of constructing a markdown list in C++ (which involved all sorts
of nasty string literals), export some JSON and assemble it with the
manual build system.
Besides following the precedent set with other dumped data, this is a
better separate of content and presentation; if we decide for example we
want to display this information in a different way, or in a different
section of the manual, it will become much easier to do so.
switch statements must now match all enum values or disable the
warning.
Explicit is good. This has helped us find two bugs, after solving
another one by debugging.
From now on, adding to an enum will raise errors where they are
not explicitly handled, which is good for productivity, and helps
us decide the correct behavior in all usages.
Notably still excluded from this though are the cases where the
warning is disabled by local pragmas.
fromTOML.cc did not build despite a top-level pragma, so I've had
to resort to a makefile solution for that.
Prior to this, there was an ad-hoc whitelist in `main.cc`. Now, every
command states its stability.
In a future PR, we will adjust the manual to take advantage of this new
information in the JSON.
(It will be easier to do that once we have some experimental feature
docs to link too; see #5930 and #7798.)
Otherwise, a trace consisting of
frame
frame
frame
non-frame
... would reach the non-frame and print the suggestion, even though
it would have ignored the non-frame anyway.
This resulted in a peculariar situation where --show-trace would have
no apparent effect, as the trace was actually already complete.
Otherwise, when running as root and user namespaces are enabled,
opening the slave fails with EPERM.
Fixes "opening pseudoterminal slave: Permission denied" followed by a
hang (https://hydra.nixos.org/build/213104244), and "error: getting
sandbox mount namespace: No such file or directory" (#8072), which
happens when the child fails very quickly and consequently reading
/proc/<child>/ns fails.
If we conditionally "declare" the argument, as we did before, based upon
weather the feature is enabled, commands like
nix --experimental-features=foo ... --thing-gated-on-foo
won't work, because the experimental feature isn't enabled until *after*
we start parsing.
Instead, allow arguments to also be associated with experimental
features (just as we did for builtins and settings), and then the
command line parser will filter out the experimental ones.
Since the effects of arguments (handler functions) are performed right
away, we get the required behavior: earlier arguments can enable later
arguments enabled!
There is just one catch: we want to keep non-positional
flags...non-positional. So if
nix --experimental-features=foo ... --thing-gated-on-foo
works, then
nix --thing-gated-on-foo --experimental-features=foo ...
should also work.
This is not my favorite long-term solution, but for now this is
implemented by delaying the requirement of needed experimental features
until *after* all the arguments have been parsed.
We hide them in various ways if the experimental feature isn't enabled.
To do this, we had to move the experimental features list out of
libnixstore, because the setting machinary itself depends on it. To do
that, we made a new `ExperimentalFeatureSettings`.
Already, we had classes like `BuiltPathsCommand` and `StorePathsCommand`
which provided alternative `run` virtual functions providing the
implementation with more arguments. This was a very nice and easy way to
make writing command; just fill in the virtual functions and it is
fairly clear what to do.
However, exception to this pattern were `Installable{,s}Command`. These
two classes instead just had a field where the installables would be
stored, and various side-effecting `prepare` and `load` machinery too
fill them in. Command would wish out those fields.
This isn't so clear to use.
What this commit does is make those command classes like the others,
with richer `run` functions.
Not only does this restore the pattern making commands easier to write,
it has a number of other benefits:
- `prepare` and `load` are gone entirely! One command just hands just
hands off to the next.
- `useDefaultInstallables` because `defaultInstallables`. This takes
over `prepare` for the one case that needs it, and provides enough
flexiblity to handle `nix repl`'s idiosyncratic migration.
- We can use `ref` instead of `std::shared_ptr`. The former must be
initialized (so it is like Rust's `Box` rather than `Option<Box>`,
This expresses the invariant that the installable are in fact
initialized much better.
This is possible because since we just have local variables not
fields, we can stop worrying about the not-yet-initialized case.
- Fewer lines of code! (Finally I have a large refactor that makes the
number go down not up...)
- `nix repl` is now implemented in a clearer way.
The last item deserves further mention. `nix repl` is not like the other
installable commands because instead working from once-loaded
installables, it needs to be able to load them again and again.
To properly support this, we make a new superclass
`RawInstallablesCommand`. This class has the argument parsing and
completion logic, but does *not* hand off parsed installables but
instead just the raw string arguments.
This is exactly what `nix repl` needs, and allows us to instead of
having the logic awkwardly split between `prepare`,
`useDefaultInstallables,` and `load`, have everything right next to each
other. I think this will enable future simplifications of that argument
defaulting logic, but I am saving those for a future PR --- best to keep
code motion and more complicated boolean expression rewriting separate
steps.
The "diagnostic ignored `-Woverloaded-virtual`" pragma helps because C++
doesn't like our many `run` methods. In our case, we don't mind the
shadowing it all --- it is *intentional* that the derived class only
provides a `run` method, and doesn't call any of the overridden `run`
methods.
Helps with https://github.com/NixOS/rfcs/pull/134
We make sure the env var paths are actually set (ie. not "") before
sending them to the canonicalization function. If we forget to do so,
the user will end up facing a puzzled failed assertion internal error.
We issue a non-failing warning as a stop-gap measure. We could want to
revisit this to issue a detailed failing error message in the future.
Descriptions for commandline flags may not include newlines and should
be rather short for display in a shell. Truncate the description string
of a flag on '\n' or '.' to and add an ellipsis if needed.
XDG Base Directory is a standard for locations for storing various
files. Nix has a few files which seem to fit in the standard, but
currently use a custom location directly in the user's ~, polluting
it:
- ~/.nix-profile
- ~/.nix-defexpr
- ~/.nix-channels
This commit adds a config option (use-xdg-base-directories) to follow
the XDG spec and instead use the following locations:
- $XDG_STATE_HOME/nix/profile
- $XDG_STATE_HOME/nix/defexpr
- $XDG_STATE_HOME/nix/channels
If $XDG_STATE_HOME is not set, it is assumed to be ~/.local/state.
Co-authored-by: Théophane Hufschmitt <7226587+thufschmitt@users.noreply.github.com>
Co-authored-by: Tim Fenney <kodekata@gmail.com>
Co-authored-by: pasqui23 <pasqui23@users.noreply.github.com>
Co-authored-by: Artturin <Artturin@artturin.com>
Co-authored-by: John Ericson <Ericson2314@Yahoo.com>
In unprivileged podman containers, /proc is not fully visible (there
are other filesystems mounted on subdirectories of /proc). Therefore
we can't mount a new /proc in the sandbox that matches the PID
namespace of the sandbox. So this commit automatically disables
sandboxing if /proc is not fully visible.
This didn't work because sandboxing doesn't work in Docker. However,
the sandboxing check is done lazily - after clone(CLONE_NEWNS) fails,
we retry with sandboxing disabled. But at that point, we've already
done UID allocation under the assumption that user namespaces are
enabled.
So let's get rid of the "goto fallback" logic and just detect early
whether user / mount namespaces are enabled.
This commit also gets rid of a compatibility hack for some ancient
Linux kernels (<2.13).
clangStdenv compiles with a single warning:
```
warning: destructor called on non-final 'nix::PosAdapter' that has virtual functions but non-virtual destructor [-Wdelete-non-abstract-non-virtual-dtor]
```
This fixes the warning by making the destructor of PosAdapter virtual,
deffering to the correct destructor from the concrete child classes.
This has no impact in the end, as none of these classes have specific
destructors.
Technicaly, it may be faster not to have this indirection, but as per
the warning, there is only one place where we have to delete abstract
PosAdapter values.
Not worth bikesheding I guess.
Rather than using `/nix/var/nix/{profiles,gcroots}/per-user/`, put the user
profiles and gcroots under `$XDG_DATA_DIR/nix/{profiles,gcroots}`.
This means that the daemon no longer needs to manage these paths itself
(they are fully handled client-side). In particular, it doesn’t have to
`chown` them anymore (removing one need for root).
This does change the layout of the gc-roots created by nix-env, and is
likely to break some stuff, so I’m not sure how to properly handle that.
It appears that on current macOS versions, our use of poll() to detect
client disconnects no longer works. As a workaround, poll() for
POLLRDNORM, since this *will* wake up when the client has
disconnected. The downside is that it also wakes up when input is
available. So just sleep for a bit in that case. This means that on
macOS, a client disconnect may take up to a second to be detected,
but that's better than not being detected at all.
Fixes#7584.
Adds a new boolean structured attribute
`outputChecks.<output>.unsafeDiscardReferences` which disables scanning
an output for runtime references.
__structuredAttrs = true;
outputChecks.out.unsafeDiscardReferences = true;
This is useful when creating filesystem images containing their own embedded Nix
store: they are self-contained blobs of data with no runtime dependencies.
Setting this attribute requires the experimental feature
`discard-references` to be enabled.
This makes the position object used in exceptions abstract, with a
method getSource() to get the source code of the file in which the
error originated. This is needed for lazy trees because source files
don't necessarily exist in the filesystem, and we don't want to make
libutil depend on the InputAccessor type in libfetcher.
When debugging nix expressions the outermost trace tends to be more useful
than the innermost. It is therefore printed last to save developers from
scrolling.
The new experimental feature 'cgroups' enables the use of cgroups for
all builds. This allows better containment and enables setting
resource limits and getting some build stats.
- call close explicitly in writeFile to prevent the close exception
from being ignored
- fsync after writing schema file to flush data to disk
- fsync schema file parent to flush metadata to disk
https://github.com/NixOS/nix/issues/7064
This was caused by -L calling setLogFormat() again, which caused the
creation of a new progress bar without destroying the old one. So we
had two progress bars clobbering each other.
We should change 'logger' to be a smart pointer, but I'll do that in a
future PR.
Fixes#6931.
Rather than directly copying the source to its dest, copy it first to a
temporary location, and eventually move that temporary.
That way, the move is at least atomic from the point-of-view of the destination
The recursive copy from the stl doesn’t exactly do what we need because
1. It doesn’t delete things as we go
2. It doesn’t keep the mtime, which change the nars
So re-implement it ourselves. A bit dull, but that way we have what we want
In `nix::rename`, if the call to `rename` fails with `EXDEV` (failure
because the source and the destination are in a different filesystems)
switch to copying and removing the source.
To avoid having to re-implement the copy manually, I switched the
function to use the c++17 `filesystem` library (which has a `copy`
function that should do what we want).
Fix#6262
Defers completion of flake inputs until the whole command line is parsed
so that we know what flakes we need to complete the inputs of.
Previously, `nix build flake --update-input <Tab>` always behaved like
`nix build . --update-input <Tab>`.
Useful because a default `sudo` on darwin doesn't clear `$HOME`, so things like `sudo nix-channel --list`
will surprisingly return the USER'S channels, rather than `root`'s.
Other counterintuitive outcomes can be seen in this PR description:
https://github.com/NixOS/nix/pull/6622
To quote Eelco in #5867:
> Unfortunately we can't do
>
> evalSettings.pureEval.setDefault(false);
>
> because then we have to do the same in main.cc (where
> pureEval is set to true), and that would allow pure-eval
> to be disabled globally from nix.conf.
Instead, a command should specify that it should be impure by
default. Then, `evalSettings.pureEval` will be set to `false;` unless
it's overridden by e.g. a CLI flag.
In that case it's IMHO OK to be (theoretically) able to override
`pure-eval` via `nix.conf` because it doesn't have an effect on commands
where `forceImpureByDefault` returns `false` (i.e. everything where pure
eval actually matters).
Closes#5867
Without the change llvm build fails on this week's gcc-13 snapshot as:
src/libutil/json.cc: In function 'void nix::toJSON(std::ostream&, const char*, const char*)':
src/libutil/json.cc:33:22: error: 'uint16_t' was not declared in this scope
33 | put(hex[(uint16_t(*i) >> 12) & 0xf]);
| ^~~~~~~~
src/libutil/json.cc:5:1: note: 'uint16_t' is defined in header '<cstdint>'; did you forget to '#include <cstdint>'?
4 | #include <cstring>
+++ |+#include <cstdint>
5 |
This solves the error
error: cannot connect to socket at '/nix/var/nix/daemon-socket/socket': Connection refused
on build farm systems that are loaded but operating normally.
I've seen this happen on an M1 mac running a loaded hercules-ci-agent.
Hercules CI uses multiple worker processes, which may connect to
the Nix daemon around the same time. It's not unthinkable that
the Nix daemon listening process isn't scheduled until after 6
workers try to connect, especially on a system under load with
many workers.
Is the increase safe?
The number is the number of connections that the kernel will buffer
while the listening process hasn't `accept`-ed them yet.
It did not - and will not - restrict the total number of daemon
forks that a client can create.
History
The number 5 has remained unchanged since the introduction in
nix-worker with 0130ef88ea in 2006.
Add a new `file` fetcher type, which will fetch a plain file over
http(s), or from the local file.
Because plain `http(s)://` or `file://` urls can already correspond to
`tarball` inputs (if the path ends-up with a know archive extension),
the URL parsing logic is a bit convuluted in that:
- {http,https,file}:// urls will be interpreted as either a tarball or a
file input, depending on the extensions of the path part (so
`https://foo.com/bar` will be a `file` input and
`https://foo.com/bar.tar.gz` as a `tarball` input)
- `file+{something}://` urls will be interpreted as `file` urls (with
the `file+` part removed)
- `tarball+{something}://` urls will be interpreted as `tarball` urls (with
the `tarball+` part removed)
Fix#3785
Co-Authored-By: Tony Olagbaiye <me@fron.io>
Since a26be9f3b8, the same parser is used
to parse the result of sourcehut’s `HEAD` endpoint (coming from [git
dumb protocol]) and the output of `git ls-remote`. However, they are very
slightly different (the former doesn’t specify the current reference
since it’s implied to be `HEAD`).
Unify both, and make the parser a bit more robust and understandable (by
making it more typed and adding tests for it)
[git dumb protocol]: https://git-scm.com/book/en/v2/Git-Internals-Transfer-Protocols#_the_dumb_protocol
this slightly increases the amount of memory used for any given symbol, but this
increase is more than made up for if the symbol is referenced more than once in
the EvalState that holds it. on average every symbol should be referenced at
least twice (once to introduce a binding, once to use it), so we expect no
increase in memory on average.
symbol tables are limited to 2³² entries like position tables, and similar
arguments apply to why overflow is not likely: 2³² symbols would require as many
string instances (at 24 bytes each) and map entries (at 24 bytes or more each,
assuming that the map holds on average at most one item per bucket as the docs
say). a full symbol table would require at least 192GB of memory just for
symbols, which is well out of reach. (an ofborg eval of nixpks today creates
less than a million symbols!)
PosTable deduplicates origin information, so using symbols for paths is no
longer necessary. moving away from path Symbols also reduces the usage of
symbols for things that are not keys in attribute sets, which will become
important in the future when we turn symbols into indices as well.
Pos objects are somewhat wasteful as they duplicate the origin file name and
input type for each object. on files that produce more than one Pos when parsed
this a sizeable waste of memory (one pointer per Pos). the same goes for
ptr<Pos> on 64 bit machines: parsing enough source to require 8 bytes to locate
a position would need at least 8GB of input and 64GB of expression memory. it's
not likely that we'll hit that any time soon, so we can use a uint32_t index to
locate positions instead.
nix show-config --json was serializing experimental features as ints.
nlohmann::json will automatically use these definitions to serialize
and deserialize ExperimentalFeatures.
Strictly, we don't use the from_json instance yet, it's provided for
completeness and hopefully future use.
This was a problem when writing a fetcher that uses e.g. sha256 hashes
for revisions. This doesn't actually do anything new, but allows for
creating such fetchers in the future (perhaps when support for Git's
SHA256 object format gains more popularity).
Saving the cwd fd didn't actually work well -- prior to this commit, the
following would happen:
: ~/w/vc/nix ; doas outputs/out/bin/nix --experimental-features 'nix-command flakes' run nixpkgs#coreutils -- --coreutils-prog=pwd
pwd: couldn't find directory entry in ‘../../../..’ with matching i-node
: ~/w/vc/nix ; doas outputs/out/bin/nix --experimental-features 'nix-command flakes' develop -c pwd
pwd: couldn't find directory entry in ‘../../../..’ with matching i-node
This doesn't work very well (maybe I'm misunderstanding the desired
implementation):
: ~/w/vc/nix ; doas outputs/out/bin/nix --experimental-features 'nix-command flakes' develop -c pwd
pwd: couldn't find directory entry in ‘../../../..’ with matching i-node
I regularly pass around simple scripts by using nix-shell as the script
interpreter, eg. like this:
#!/usr/bin/env nix-shell
#!nix-shell -p dd_rescue coreutils bash -i bash
While this works most of the time, I recently had one occasion where it
would not and the above would result in the following:
$ sudo ./myscript.sh
bash: ./myscript.sh: No such file or directory
Note the "sudo" here, because this error only occurs if we're root.
The reason for the latter is because running Nix as root means that we
can directly access the store, which makes sure we use a filesystem
namespace to make the store writable. XXX - REWORD!
So when stracing the process, I stumbled on the following sequence:
openat(AT_FDCWD, "/proc/self/ns/mnt", O_RDONLY) = 3
unshare(CLONE_NEWNS) = 0
... later ...
getcwd("/the/real/cwd", 4096) = 14
setns(3, CLONE_NEWNS) = 0
getcwd("/", 4096) = 2
In the whole strace output there are no calls to chdir() whatsoever, so
I decided to look into the kernel source to see what else could change
directories and found this[1]:
/* Update the pwd and root */
set_fs_pwd(fs, &root);
set_fs_root(fs, &root);
The set_fs_pwd() call is roughly equivalent to a chdir() syscall and
this is called when the setns() syscall is invoked[2].
[1]: b14ffae378/fs/namespace.c (L4659)
[2]: b14ffae378/kernel/nsproxy.c (L346)
Impure derivations are derivations that can produce a different result
every time they're built. Example:
stdenv.mkDerivation {
name = "impure";
__impure = true; # marks this derivation as impure
outputHashAlgo = "sha256";
outputHashMode = "recursive";
buildCommand = "date > $out";
};
Some important characteristics:
* This requires the 'impure-derivations' experimental feature.
* Impure derivations are not "cached". Thus, running "nix-build" on
the example above multiple times will cause a rebuild every time.
* They are implemented similar to CA derivations, i.e. the output is
moved to a content-addressed path in the store. The difference is
that we don't register a realisation in the Nix database.
* Pure derivations are not allowed to depend on impure derivations. In
the future fixed-output derivations will be allowed to depend on
impure derivations, thus forming an "impurity barrier" in the
dependency graph.
* When sandboxing is enabled, impure derivations can access the
network in the same way as fixed-output derivations. In relaxed
sandboxing mode, they can access the local filesystem.
The return value of BaseError::addTrace(...) is never used and
error-prone as subclasses calling it will return a BaseError instead of
the subclass.
This commit changes its return value to be void.
When importing e.g. a local `nixpkgs` in a flake to test a change like
{
inputs.nixpkgs.url = path:/home/ma27/Projects/nixpkgs;
outputs = /* ... */
}
then the input is missing a `lastModified`-field that's e.g. used in
`nixpkgs.lib.nixosSystem`. Due to the missing `lastMoified`-field, the
mtime is set to 19700101:
result -> /nix/store/b7dg1lmmsill2rsgyv2w7b6cnmixkvc1-nixos-system-nixos-22.05.19700101.dirty
With this change, the `path`-fetcher now sets a `lastModified` attribute
to the `mtime` just like it's the case in the `tarball`-fetcher already.
When building NixOS systems with `nixpkgs` being a `path`-input and this
patch, the output-path now looks like this:
result -> /nix/store/ld2qf9c1s98dxmiwcaq5vn9k5ylzrm1s-nixos-system-nixos-22.05.20220217.dirty
Before the change on a system with `auto-optimise-store = true`:
$ nix store gc --verbose --max 1
deleted all the paths instead of one path (we requested 1 byte limit).
It happens because every file in `auto-optimise-store = true` has at
least 2 links: file itself and a link in /nix/store/.links/ directory.
The change conservatively assumes that any file that has one (as before)
or two links (assume auto-potimise mode) will free space.
Co-authored-by: Sandro <sandro.jaeckel@gmail.com>
no need for function<> with c++17 deduction. this saves allocations and virtual
calls, but has the same semantics otherwise. not going through function has the
side effect of giving compilers more insight into the cleanup code, so we need a
few local warning disables.
reduces peak hep memory use on eval of our test system from 264.4MB to 242.3MB,
possibly also a slight performance boost.
theoretically memory use could be cut down by another eight bytes per Pos on
average by turning it into a tuple containing an index into a global base
position table with row and column offsets, but that doesn't seem worth the
effort at this point.
No real need for keeping a separate header for such a simple class.
This requires changing a bit `OrSuggestions<T>::operator*` to not throw
an `Error` to prevent a cyclic dependency. But since this error is only
thrown on programmer error, we can replace the whole method by a direct
call to `std::get` which will raise its own assertion if needs be.
Allows completing `nix build ~/flake#<Tab>`.
We can implement expansion for `~user` later if needed.
Not using wordexp(3) since that expands way too much.
Starts progress on #5729.
The idea is that we should not have these default methods throwing
"unimplemented". This is a small step in that direction.
I kept `addTempRoot` because it is a no-op, rather than failure. Also,
as a practical matter, it is called all over the place, while doing
other tasks, so the downcasting would be annoying.
Maybe in the future I could move the "real" `addTempRoot` to `GcStore`,
and the existing usecases use a `tryAddTempRoot` wrapper to downcast or
do nothing, but I wasn't sure whether that was a good idea so with a
bias to less churn I didn't do it yet.
To avoid that JSON messages are parsed twice in case of
remote builds with `ssh-ng://`, I split up the original
`handleJSONLogMessage` into three parts:
* `parseJSONMessage(const std::string&)` checks if it's a message in the
form of `@nix {...}` and tries to parse it (and prints an error if the
parsing fails).
* `handleJSONLogMessage(nlohmann::json&, ...)` reads the fields from the
message and passes them to the logger.
* `handleJSONLogMessage(const std::string&, ...)` behaves as before, but
uses the two functions mentioned above as implementation.
In case of `ssh-ng://`-logs the first two methods are invoked manually.
This changes the representation of the interrupt callback list to
be safe to use during interrupt handling.
Holding a lock while executing arbitrary functions is something to
avoid in general, because of the risk of deadlock.
Such a deadlock occurs in https://github.com/NixOS/nix/issues/3294
where ~CurlDownloader tries to deregister its interrupt callback.
This happens during what seems to be a triggerInterrupt() by the
daemon connection's MonitorFdHup thread. This bit I can not confirm
based on the stack trace though; it's based on reading the code,
so no absolute certainty, but a smoking gun nonetheless.
we'll retain the old coerceToString interface that returns a string, but callers
that don't need the returned value to outlive the Value it came from can save
copies by using the new interface instead. for values that weren't stringy we'll
pass a new buffer argument that'll be used for storage and shouldn't be
inspected.
when given a string yacc will copy the entire input to a newly allocated
location so that it can add a second terminating NUL byte. since the
parser is a very internal thing to EvalState we can ensure that having
two terminating NUL bytes is always possible without copying, and have
the parser itself merely check that the expected NULs are present.
# before
Benchmark 1: nix search --offline nixpkgs hello
Time (mean ± σ): 572.4 ms ± 2.3 ms [User: 563.4 ms, System: 8.6 ms]
Range (min … max): 566.9 ms … 579.1 ms 50 runs
Benchmark 2: nix eval -f ../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix
Time (mean ± σ): 381.7 ms ± 1.0 ms [User: 348.3 ms, System: 33.1 ms]
Range (min … max): 380.2 ms … 387.7 ms 50 runs
Benchmark 3: nix eval --raw --impure --expr 'with import <nixpkgs/nixos> {}; system'
Time (mean ± σ): 2.936 s ± 0.005 s [User: 2.715 s, System: 0.221 s]
Range (min … max): 2.923 s … 2.946 s 50 runs
# after
Benchmark 1: nix search --offline nixpkgs hello
Time (mean ± σ): 571.7 ms ± 2.4 ms [User: 563.3 ms, System: 8.0 ms]
Range (min … max): 566.7 ms … 579.7 ms 50 runs
Benchmark 2: nix eval -f ../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix
Time (mean ± σ): 376.6 ms ± 1.0 ms [User: 345.8 ms, System: 30.5 ms]
Range (min … max): 374.5 ms … 379.1 ms 50 runs
Benchmark 3: nix eval --raw --impure --expr 'with import <nixpkgs/nixos> {}; system'
Time (mean ± σ): 2.922 s ± 0.006 s [User: 2.707 s, System: 0.215 s]
Range (min … max): 2.906 s … 2.934 s 50 runs
there's a couple places that can be easily converted from using strings to using
string_views instead. gives a slight (~1%) boost to system eval.
# before
nix eval --raw --impure --expr 'with import <nixpkgs/nixos> {}; system'
Time (mean ± σ): 2.946 s ± 0.026 s [User: 2.655 s, System: 0.209 s]
Range (min … max): 2.905 s … 2.995 s 20 runs
# after
Time (mean ± σ): 2.928 s ± 0.024 s [User: 2.638 s, System: 0.211 s]
Range (min … max): 2.893 s … 2.970 s 20 runs
this avoids one copy from `s` into `str`, and possibly another copy needed to
construct `s` at the call site. lexical_cast is also more efficient in general.
There already existed a smoke test for the link content length,
but it appears that there exists some corruptions pernicious enough
to replace the file content with zeros, and keeping the same length.
--repair-path now goes as far as checking the content of the link,
making it true to its name and actually repairing the path for such
coruption cases.
we don't have to create an ostream sentry object for every character of a JSON
string we write. format a bunch of characters and flush them to the stream all
at once instead.
this doesn't affect small numbers of string characters, but larger numbers of
total JSON string characters written gain a lot. at 1MB of total string written
we gain almost 30%, at 16MB it's almost a factor of 3x. large numbers of JSON
string characters do occur naturally in a nixos system evaluation to generate
documentation (though this is now somewhat mitigated by caching the largest part
of nixos option docs).
benchmarked with
hyperfine 'nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) {e})"' --warmup 1 -L e 1,4,256,4096,65536
before:
Benchmark 1: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 1)"
Time (mean ± σ): 12.5 ms ± 0.2 ms [User: 9.2 ms, System: 4.0 ms]
Range (min … max): 11.9 ms … 13.1 ms 223 runs
Benchmark 2: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 4)"
Time (mean ± σ): 12.5 ms ± 0.2 ms [User: 9.3 ms, System: 3.8 ms]
Range (min … max): 11.9 ms … 13.2 ms 220 runs
Benchmark 3: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 256)"
Time (mean ± σ): 13.2 ms ± 0.3 ms [User: 9.8 ms, System: 4.0 ms]
Range (min … max): 12.6 ms … 14.3 ms 205 runs
Benchmark 4: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 4096)"
Time (mean ± σ): 24.0 ms ± 0.4 ms [User: 19.4 ms, System: 5.2 ms]
Range (min … max): 22.7 ms … 25.8 ms 119 runs
Benchmark 5: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 65536)"
Time (mean ± σ): 196.0 ms ± 3.7 ms [User: 171.2 ms, System: 25.8 ms]
Range (min … max): 190.6 ms … 201.5 ms 14 runs
after:
Benchmark 1: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 1)"
Time (mean ± σ): 12.4 ms ± 0.3 ms [User: 9.1 ms, System: 4.0 ms]
Range (min … max): 11.7 ms … 13.3 ms 204 runs
Benchmark 2: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 4)"
Time (mean ± σ): 12.4 ms ± 0.2 ms [User: 9.2 ms, System: 3.9 ms]
Range (min … max): 11.8 ms … 13.0 ms 214 runs
Benchmark 3: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 256)"
Time (mean ± σ): 12.6 ms ± 0.2 ms [User: 9.5 ms, System: 3.8 ms]
Range (min … max): 12.1 ms … 13.3 ms 209 runs
Benchmark 4: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 4096)"
Time (mean ± σ): 15.9 ms ± 0.2 ms [User: 11.4 ms, System: 5.1 ms]
Range (min … max): 15.2 ms … 16.4 ms 171 runs
Benchmark 5: nix eval --raw --expr "let s = __concatStringsSep \"\" (__genList (_: \"c\") 256); in __toJSON (__genList (_: s) 65536)"
Time (mean ± σ): 69.0 ms ± 0.9 ms [User: 44.3 ms, System: 25.3 ms]
Range (min … max): 67.2 ms … 70.9 ms 42 runs
