This is unnecessary because Darwin builds already cannot create
`set{u,g}id` files due to the minimal sandbox policy, because we can
forbid messing with the top‐level build directory directly in the
sandbox rules, and because Darwin builds can trivially avoid isolation
through temporary directories right now anyway.
This does regress the build directory isolation for builds with
`sandbox = false`, but I can’t imagine that mattering given the
above. The sandbox change prepares us for a world where we close
off shared temporary directories for `sandbox = true` builds and try
harder to achieve proper isolation on Darwin, but probably doesn’t
have a meaningful security impact one way or another for now.
With this change, we get down to 41 byte build directory paths on
Darwin, a ~2–3½ byte improvement over the old status quo. We can
also restore the 0710 permissions on Linux.
Change-Id: I6a6a6964a681c0365241fe7234831db656b76799
The minimum build directory length on Darwin with default settings
when we were still using `/tmp` was 22 bytes. Deriving build directory
names from my local store, the median and mode were 43, the mean was
around 44½, and the maximum was 127.
The switch to `/nix/var/nix/builds` over `/tmp` added a 15 byte
penalty, and the additional `/b` directory added another 2.
Now that we use opaque build directory names, the length is 48, so
we’re still at a ~3½–5 byte penalty over the previous status
quo. This change brings us down to 43, matching the previous median.
Note that these calculations do not take into account the fact that
`/tmp` is a symbolic link to `/private/tmp` on Darwin. Anything
that was canonicalizing paths would have had an additional 8 byte
penalty in the previous status quo that is not applicable here,
so we may already be ahead even without this change. If the more
opaque directory name here is undesirable, then that factor could
potentially help us squeeze by without. Alternatively, in combination
with dropping the `/b` on Darwin we could use `/nix/var/nix/bld` or
similar, but I feel that the paths in general are sufficiently opaque
that it should be okay to go with the shorter option here. Given that
some projects already had to reduce filename lengths to avoid this
limit even before the recent changes, I think it is best to try and
improve on the previous status quo.
Note that `/nix/var/nix/builds` will unfortunately not be cleaned up
on Darwin. However, we don’t clean up the directories inside it on
Darwin anyway, so hopefully that’s okay for now?
Closes: #913
Change-Id: I6a6a6964bffce7194bcddcaefb4c4a37569c7df5
We always use the default temporary directory, because
`createUniqueDir` has an interface nice enough to use directly for
the few bespoke uses.
Change-Id: I6a6a696450b7c0a0bd76655632fb14d7c5e38199
This makes the paths more nondeterministic, but more reliably unique,
and lets us remove the retry loop.
Note that this adds random entropy to the build directory visible
inside derivations on Darwin and unsandboxed Linux. It was already
non‐deterministic in the presence of concurrent builds and similar,
but now we can reliably expect it to be different every time. On the
whole I think that’s a good thing, as it is impossible to ensure
a single consistent build directory and derivation outputs should
not depend on it.
Package reproducibility isn’t great on Darwin to begin with,
though, and the reproducibility bugs this will turn up in packages
will be more urgent to fix than when the build directory was mostly
consistent. A quick survey of my local store shows that many C, C++,
and Rust binaries contain build directory references, likely due to
use of `__FILE__` and its equivalents; non‐binary offenders include:
* Install logs included in the Rust and Cargo bootstrap compilers
* Example errors in the Rust documentation referencing build paths
* Configuration information installed with CPython itself
* Python 2 metadata from resholve’s closure
* Cython metadata
* Generated headers in Facebook libraries referencing source paths
* Generated CMake files in Facebook libraries referencing source paths
I haven’t built that much in this store since the last GC, so this is
probably only a small sample of the problems across the tree. These are
all instances of <https://reproducible-builds.org/docs/build-path/>,
though, and should probably just be treated as general reproducibility
bugs outside of contexts like the Linux sandbox where we can normalize
them away entirely.
I have implemented away build directory paths for C/C++, applied some
additional fixes for non‐`__FILE__`‐related issues in binaries
from ATF and LLVM, and fixed the derivation bug causing the CPython
3 issue, and will work on upstreaming these changes. Rust is working
on the problem upstream, with some temporary workarounds we can
potentially apply in Nixpkgs for now. The rest will require some
distributed effort.
Change-Id: I6a6a69645b4915c56c0fdef904aa81684e4136c6
This does change the behaviour when the global temporary directory
does not exist, but other uses of the global temporary directory are
already broken in that circumstance, and it should be fixed centrally
if the use case is considered desirable. The logic was not present
before the recent churn around build directories – it was added now
that Lix is taking ownership of the build directory in the store –
so this should not be a meaningful regression.
Change-Id: I6a6a6964e345ea6803226c5ad759e836de7cb0ed
There is now no risk of race conditions on a system with a functioning
entropy source, and the bespoke prefixes are either redundant to the
default or unnecessary.
Change-Id: I6a6a69641211c6bb979ea48ad30aecb1a53d03f0
Relax the constraints on keeping the exact same filename format to
provide a more robust source of entropy with a simpler interface
(as previously suggested by eldritch horrors). Using 128 bits of
OS‐provided entropy ensures global uniqueness and allows us to
skip any thought of gracefully handling the case where these files
already exist.
My microbenchmark that repeatedly constructed paths like this and
printed them out showed that this takes about 1.23× the time of
the previous implementation, both taking on the order of a couple
microseconds for one iteration. Since everything that uses it is doing
things more expensive than printing to standard output, the actual
performance delta is likely to be lost in the noise. If it somehow
becomes a bottleneck, it can be optimized without sacrificing the
guarantees by reading from the system RNG only to seed a thread‐local
CSPRNG like [ChaCha8Rand], but I think that’s very unlikely.
We also tweak the recommended way of creating a temporary file inside
a directory in anticipation of later changes, and rename the `suffix`
parameter to `prefix` (it’s a prefix to the random characters and
a suffix to the root, but this way is more consistent).
[ChaCha8Rand]: https://c2sp.org/chacha8rand
Change-Id: I6a6a69648502c746d13d8c3bd2768cbbf1b90466
base32Encode now takes std::span<std::byte>, with a base32EncodeStr
convenience wrapper which takes std::string_view.
Co-authored-by: Qyriad <qyriad@qyriad.me>
Change-Id: I6a6a6964f799dc84ecbfb55c7ca03a064cff71d9
The prospective callers of this should probably be doing something
smarter or more abstracted to begin with, but this is useful as an
incremental improvement for call sites with existing `makeTempPath`
logic in the face of filename length limits.
Change-Id: I6a6a69644292f5bbf984a1df90192e06c6022b53
We always use the default temporary directory, because
`createUniqueDir` has an interface nice enough to use directly for
the few bespoke uses.
Change-Id: I6a6a6964c15c31bb3e131fbe1db1837987a6d6dc
This makes the paths more nondeterministic, but more reliably unique,
and lets us remove the retry loop.
Note that this adds random entropy to the build directory visible
inside derivations on Darwin and unsandboxed Linux. It was already
non‐deterministic in the presence of concurrent builds and similar,
but now we can reliably expect it to be different every time. On the
whole I think that’s a good thing, as it is impossible to ensure
a single consistent build directory and derivation outputs should
not depend on it.
Package reproducibility isn’t great on Darwin to begin with,
though, and the reproducibility bugs this will turn up in packages
will be more urgent to fix than when the build directory was mostly
consistent. A quick survey of my local store shows that many C, C++,
and Rust binaries contain build directory references, likely due to
use of `__FILE__` and its equivalents; non‐binary offenders include:
* Install logs included in the Rust and Cargo bootstrap compilers
* Example errors in the Rust documentation referencing build paths
* Configuration information installed with CPython itself
* Python 2 metadata from resholve’s closure
* Cython metadata
* Generated headers in Facebook libraries referencing source paths
* Generated CMake files in Facebook libraries referencing source paths
I haven’t built that much in this store since the last GC, so this is
probably only a small sample of the problems across the tree. These are
all instances of <https://reproducible-builds.org/docs/build-path/>,
though, and should probably just be treated as general reproducibility
bugs outside of contexts like the Linux sandbox where we can normalize
them away entirely.
I have implemented away build directory paths for C/C++, applied some
additional fixes for non‐`__FILE__`‐related issues in binaries
from ATF and LLVM, and fixed the derivation bug causing the CPython
3 issue, and will work on upstreaming these changes. Rust is working
on the problem upstream, with some temporary workarounds we can
potentially apply in Nixpkgs for now. The rest will require some
distributed effort.
Change-Id: I6a6a69648f74d85c6fca86cc52f38fd957e4f9ad
This does change the behaviour when the global temporary directory
does not exist, but other uses of the global temporary directory are
already broken in that circumstance, and it should be fixed centrally
if the use case is considered desirable. The logic was not present
before the recent churn around build directories – it was added now
that Lix is taking ownership of the build directory in the store –
so this should not be a meaningful regression.
Change-Id: I6a6a69648054ae201b3ce36d11e49c93793fdb0e
There is now no risk of race conditions on a system with a functioning
entropy source, and the bespoke prefixes are either redundant to the
default or unnecessary.
Change-Id: I6a6a69648a3b8060333e97269ea8b72499614559
Relax the constraints on keeping the exact same filename format to
provide a more robust source of entropy with a simpler interface
(as previously suggested by eldritch horrors). Using 128 bits of
OS‐provided entropy ensures global uniqueness and allows us to
skip any thought of gracefully handling the case where these files
already exist.
My microbenchmark that repeatedly constructed paths like this and
printed them out showed that this takes about 1.23× the time of
the previous implementation, both taking on the order of a couple
microseconds for one iteration. Since everything that uses it is doing
things more expensive than printing to standard output, the actual
performance delta is likely to be lost in the noise. If it somehow
becomes a bottleneck, it can be optimized without sacrificing the
guarantees by reading from the system RNG only to seed a thread‐local
CSPRNG like [ChaCha8Rand], but I think that’s very unlikely.
We also tweak the recommended way of creating a temporary file inside
a directory in anticipation of later changes, and rename the `suffix`
parameter to `prefix` (it’s a prefix to the random characters and
a suffix to the root, but this way is more consistent).
[ChaCha8Rand]: https://c2sp.org/chacha8rand
Change-Id: I5bd7badf1392243f485935c4a016c1f833cb16d3
The prospective callers of this should probably be doing something
smarter or more abstracted to begin with, but this is useful as an
incremental improvement for call sites with existing `makeTempPath`
logic in the face of filename length limits.
Change-Id: I6a6a6964374f47abbf0ec10aa8d945c4e50a43af
See [my comment] on the Nix PR to restore the previous behaviour
for why I believe we should remove this for the next release. The PR
should still be backported to stable releases to avoid making breaking
changes to their semantics.
[my comment]: <https://github.com/NixOS/nix/pull/13741#issuecomment-3180851635>
Fixing this across supported Lix versions is required for Nixpkgs to
update toml11, which is a blocker for the CMake 4 update.
Change-Id: I6a6a69642e6b6cb13a9fccc0778e9158b53102d5
This version changes the handling of TOML timestamps, and throws an
error on out‐of‐range integer literals rather than the previous
saturating behaviour, as required by [the TOML v1.0.0 specification]:
> Arbitrary 64-bit signed integers (from −2^63 to 2^63−1) should be
> accepted and handled losslessly. If an integer cannot be represented
> losslessly, an error must be thrown.
[the TOML v1.0.0 specification]: <https://toml.io/en/v1.0.0#integer>
The only known use of this is a questionable Nixpkgs test that I have
proposed [a fix] for.
[a fix]: <https://github.com/NixOS/nixpkgs/pull/433710>
Bumping this ahead of Nixpkgs ensures we can test these cases on
HEAD in advance. I presume that the next Lix major version will be
released after 25.05 goes out of support, so it should be fine to
drop support for the old version of toml11.
The co‐authors of this commit are the contributors to the vendored
package definition from Nixpkgs.
Co-authored-by: Anderson Torres <torres.anderson.85@protonmail.com>
Co-authored-by: Artturin <Artturin@artturin.com>
Co-authored-by: Silvan Mosberger <silvan.mosberger@moduscreate.com>
Change-Id: I6a6a69644a188b6e09eee5c9cf91ddd3c81d24ee
This behaviour is bad and will be fixed, but adding a test for the
present state documents the change.
Change-Id: I6a6a6964b88dc929269ee136804857b3852dfafe
JSON has only one numeric type, with vague semantics. [RFC 8259] says:
> This specification allows implementations to set limits on the range
> and precision of numbers accepted. Since software that implements
> IEEE 754 binary64 (double precision) numbers [IEEE754] is generally
> available and widely used, good interoperability can be achieved by
> implementations that expect no more precision or range than these
> provide, in the sense that implementations will approximate JSON
> numbers within the expected precision. A JSON number such as 1E400
> or 3.141592653589793238462643383279 may indicate potential
> interoperability problems, since it suggests that the software that
> created it expects receiving software to have greater capabilities
> for numeric magnitude and precision than is widely available.
>
> Note that when such software is used, numbers that are integers and
> are in the range [-(2**53)+1, (2**53)-1] are interoperable in the
> sense that implementations will agree exactly on their numeric
> values.
[RFC 8259]: <https://www.rfc-editor.org/rfc/rfc8259.html#section-6>
Floating‐point numbers are annoying to deal with in Nix, so it
optimistically parses integer‐looking literals as Nix‐language
integers where possible. Nixpkgs relies on this behaviour, as it backs
its `lib.toInt` family of functions with `builtins.fromJSON` in lieu
of a real integer‐parsing built‐in, and treats floating‐point
outputs as an error. Therefore, dealing with integer‐looking
JSON number literals that are outside the interoperable range is
unavoidable.
However, this raises the question of how literals that look
like integers, but exceed the range of a Nix‐language integer,
should be handled. The JSON library we use attempts to represent
integer‐looking literals as a unsigned or signed C++ integer type
before falling back to floating‐point numbers.
This means that we were parsing literals below −2⁶³ as
floating‐point numbers, while rejecting ones above (2⁶³ + 1)
with an error. This was done to avoid the C++ undefined behaviour in
the previous code path, but is hard to justify. This change causes
them to both be parsed as floating‐point numbers.
The alternative would be to reject the signed case too. However,
I believe that is less consistent with how JSON handles its single
numeric type, less interoperable with JSON documents found in the wild,
and that it is preferable to avoid the function being needlessly
partial in this case. It does mean that round‐tripping is less
lossless than before, but extreme floating‐point values already
caused these kinds of issues, and from JSON’s point of view that’s
exactly what these literals represent. Only numeric values outside the
RFC’s suggested interoperable range will have any round‐tripping
issues, and we continue to guarantee the behaviour on all values that
can be represented as Nix‐language integers.
Change-Id: I6a6a696412383e8a2cc160397716cb7f7bc7a2d4
`mesonCheckPhase` of course runs `preCheck` and `postCheck` hooks,
not `preInstallCheck`, so this was not doing the right thing. Really
sorry for breaking CI :(
Fixes: e6da29ad6b
Change-Id: I6a6a69642a242e2c8bfe10ac61d3f6756d666548
This gets us `--print-errorlogs` and `--timeout-multiplier=0` for free,
and also fixes the concatenation logic with `__structuredAttrs`.
Change-Id: I6a6a69643ea7224cb24508505c554143f83ae243
We’re already allowing `/tmp` anyway, so this should be harmless,
and it fixes a regression in the default configuration caused by
moving the build directories out of `temp-dir`. (For instance, that
broke the Lix `guessOrInventPath.sockets` test.)
Note that removing `/tmp` breaks quite a few builds, so although it may
be a good idea in general it would require work on the Nixpkgs side.
Fixes: 749afbbe99
Change-Id: I6a6a69645f429bc50d4cb24283feda3d3091f534
Using `AllowDaemon::Disallow` here broke `ssh-ng://` remote builds in
multi‐user setups where the remote builder user does not have write
access to the store, now that the automatic store selection logic
has changed. Switch to the default behaviour for this path to fix that.
This causes `ssh-ng://` builds to use the daemon by default on the
remote end, even as `root`. I think this is desirable, as the previous
change already made `ssh://` behave this way, and the pitfalls of
local stores apply to remote builds too. For instance, there were
persistent `ulimit` issues on the NixOS Hydra macOS builders that were
resolved by forcing use of the daemon, and I believe the Linux builders
also go through the daemon these days due to using non‐`root` SSH
users. I believe that the `root` vs. non‐`root` difference is just
as confusing for remote builds as it is for local ones.
`ssh-ng://root@builder?remote-store=local` can be used to revert back
to the previous default if necessary.
Closes: #884
Fixes: 9a59106c17
Change-Id: I6a6a696410f46cd3f2f5a94073ea924ad45dc99c
This allows other functions to parameterize over it themselves. An
enum class is used to avoid API misuse.
Change-Id: I6a6a6964d2b5ad47ae5ea9eb11af9b6373ce2141
