In principle, this should avoid deadlocks where two instances of Nix are
holding a shared lock on big-lock and are both waiting to get an
exclusive lock.
However, it seems like `flock(2)` is supposed to do this automatically,
so it's not clear whether this is actually where the problem comes from.
I just had a colleague get confused by the previous phrase for good
reason. "valid" sounds like an *objective* criterion, e.g. and *invalid
signature* would be one that would be trusted by no one, e.g. because it
misformatted or something.
What is actually going is that there might be a signature which is
perfectly valid to *someone else*, but not to the user, because they
don't trust the corresponding public key. This is a *subjective*
criterion, because it depends on the arbitrary and personal choice of
which public keys to trust.
I therefore think "trustworthy" is a better adjective to use. Whether
something is worthy of trust is clearly subjective, and then "trust"
within that word nicely evokes `trusted-public-keys` and friends.
- 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
Implements the approach suggested by feedback on PR #6994, where
tempdir paths are created in the store (now with an exclusive lock).
As part of this work, the currently-broken and unused
`createTempDirInStore` function is updated to create an exclusive lock
on the temp directory in the store.
The GC now makes a non-blocking attempt to lock any store directories
that "look like" the temp directories created by this function, and if
it can't acquire one, ignores the directory.
Without the change any CA deletion triggers linear scan on large
RealisationsRefs table:
sqlite>.eqp full
sqlite> delete from RealisationsRefs where realisationReference IN ( select id from Realisations where outputPath = 1234567890 );
QUERY PLAN
|--SCAN RealisationsRefs
`--LIST SUBQUERY 1
`--SEARCH Realisations USING COVERING INDEX IndexRealisationsRefsOnOutputPath (outputPath=?)
With the change it gets turned into a lookup:
sqlite> CREATE INDEX IndexRealisationsRefsRealisationReference on RealisationsRefs(realisationReference);
sqlite> delete from RealisationsRefs where realisationReference IN ( select id from Realisations where outputPath = 1234567890 );
QUERY PLAN
|--SEARCH RealisationsRefs USING INDEX IndexRealisationsRefsRealisationReference (realisationReference=?)
`--LIST SUBQUERY 1
`--SEARCH Realisations USING COVERING INDEX IndexRealisationsRefsOnOutputPath (outputPath=?)
If the derivation `foo` depends on `bar`, and they both have the same
output path (because they are CA derivations), then this output path
will depend both on the realisation of `foo` and of `bar`, which
themselves depend on each other.
This confuses SQLite which isn’t able to automatically solve this
diamond dependency scheme.
Help it by adding a trigger to delete all the references between the
relevant realisations.
Fix#5320
This ensures that use-sites properly trigger new monomorphisations on
one hand, and on the other hand keeps the main `sqlite.hh` clean and
interface-only. I think that is good practice in general, but in this
situation in particular we do indeed have `sqlite.hh` users that don't
need the `throw_` function.
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.
Rather than having four different but very similar types of hashes, make
only one, with a tag indicating whether it corresponds to a regular of
deferred derivation.
This implies a slight logical change: The original Nix+multiple-outputs
model assumed only one hash-modulo per derivation. Adding
multiple-outputs CA derivations changed this as these have one
hash-modulo per output. This change is now treating each derivation as
having one hash modulo per output.
This obviously means that we internally loose the guaranty that
all the outputs of input-addressed derivations have the same hash
modulo. But it turns out that it doesn’t matter because there’s nothing
in the code taking advantage of that fact (and it probably shouldn’t
anyways).
The upside is that it is now much easier to work with these hashes, and
we can get rid of a lot of useless `std::visit{ overloaded`.
Co-authored-by: John Ericson <John.Ericson@Obsidian.Systems>
1. `DerivationOutput` now as the `std::variant` as a base class. And the
variants are given hierarchical names under `DerivationOutput`.
In 8e0d0689be @matthewbauer and I
didn't know a better idiom, and so we made it a field. But this sort
of "newtype" is anoying for literals downstream.
Since then we leaned the base class, inherit the constructors trick,
e.g. used in `DerivedPath`. Switching to use that makes this more
ergonomic, and consistent.
2. `store-api.hh` and `derivations.hh` are now independent.
In bcde5456cc I swapped the dependency,
but I now know it is better to just keep on using incomplete types as
much as possible for faster compilation and good separation of
concerns.
This changes was taken from dynamic derivation (#4628). It` somewhat
undoes the refactors I first did for floating CA derivations, as the
benefit of hindsight + requirements of dynamic derivations made me
reconsider some things.
They aren't to consequential, but I figured they might be good to land
first, before the more profound changes @thufschmitt has in the works.
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.
No matter what, we need to resize the buffer to not have any scratch
space after we do the `read`. In the end of file case, `got` will be 0
from it's initial value.
Before, we forgot to resize in the EOF case with the break. Yes, we know
we didn't recieve any data in that case, but we still have the scatch
space to undo.
Co-Authored-By: Will Fancher <Will.Fancher@Obsidian.Systems>
For a typical desktop system (~2K packages) we can easily get 100K
entries in RealisationsRefs. Without indices query for RealisationsRefs
requires linear scan.
RealisationsRefs(referrer)
--------------------------
Inefficiency is seen as a 100% CPU load of nix-daemon for the following
scenario:
$ nix edit -f . bash # add unused environment variable, like FOO="1"
# populate RealisationsRefs, build fresh system
$ nix build -f nixos system --arg config '{ contentAddressedByDefault = true; }'
$ nix edit -f . bash # add unused environment variable, like FOO="2"
$ time nix build -f nixos system --arg config '{ contentAddressedByDefault = true; }'
In this case `bash `will be rebuilt a few times and then rest of CPU
time is spent on scanning RealisationsRefs table (about 5 CPU-minutes
on my machine).
Before the change:
$ time nix build -f nixos system ... # step 4 above
real 34m3,613s
user 0m5,232s
sys 0m0,758s
Of all this time about 29.5 minutes are taken by nix-daemon's CPU time.
After the change:
$ time nix build -f nixos system ... # step 4 above
real 4m50,061s
user 0m5,038s
sys 0m0,677s
Of all this time about 1 minute is taken by nix-daemon's CPU time.
Most of the time is spent polling for non-existent realisations on
cache-nixos.org.
Realisations(outputPath)
------------------------
After running CA system for two weeks I got ~1M entries in Realisations
table. `nix-collect-garbage` became very slow (seemingly 100 path deletions
per second). It happens due to a slow cascading delete from Realisations
triggered by deletion from ValidPaths.
The fix is to add an index on primary key from ValidPaths(id) that
triggers cascading deletions.
Before the change:
$ time nix-collect-garbage -d --max-freed 100G
<interrupted before finish, took too long>
real 23m32.411s
user 17m49.679s
sys 4m50.609s
Most of time was spent in re-scanning Realisations table on each path deletion.
After the change:
$ time nix-collect-garbage -d --max-freed 100G
real 8m43.226s
user 6m16.317s
sys 1m40.188s
Time is spent scanning sqlite indices and in kernel when unlinking directories.
Doing it as a side-effect of calling LocalStore::makeStoreWritable()
is very ugly.
Also, make sure that stopping the progress bar joins the update
thread, otherwise that thread should be unshared as well.
Since 4806f2f6b0, we can't have paths with
references passed to builtins.{path,filterSource}. This prevents many cases
of those functions called on IFD outputs from working. Resolve this by
passing the references found in the original path to the added path.
Rather than having them plain strings scattered through the whole
codebase, create an enum containing all the known experimental features.
This means that
- Nix can now `warn` when an unkwown experimental feature is passed
(making it much nicer to spot typos and spot deprecated features)
- It’s now easy to remove a feature altogether (once the feature isn’t
experimental anymore or is dropped) by just removing the field for the
enum and letting the compiler point us to all the now invalid usages
of it.
This ensures any started processes can't write to /nix/store (except
during builds). This partially reverts 01d07b1e, which happened because
of #2646.
The problem was only happening after nix downloads anything, causing
me to suspect the download thread. The problem turns out to be:
"A process can't join a new mount namespace if it is sharing
filesystem-related attributes with another process", in this case this
process is the curl thread.
Ideally, we might kill it before spawning the shell process, but it's
inside a static variable in the getFileTransfer() function. So
instead, stop it from sharing FS state using unshare(). A strategy
such as the one from #5057 (single-threaded chroot helper binary) is
also very much on the table.
Fixes#4337.
This fixes a bug in the garbage collector where if a path
/nix/store/abcd-foo is valid, but we do a
isValidPath("/nix/store/abcd-foo.lock") first, then a negative entry
for /nix/store/abcd is added to pathInfoCache, so /nix/store/abcd-foo
is subsequently considered invalid and deleted.
The garbage collector no longer blocks other processes from
adding/building store paths or adding GC roots. To prevent the
collector from deleting store paths just added by another process,
processes need to connect to the garbage collector via a Unix domain
socket to register new temporary roots.
I had started the trend of doing `std::visit` by value (because a type
error once mislead me into thinking that was the only form that
existed). While the optomizer in principle should be able to deal with
extra coppying or extra indirection once the lambdas inlined, sticking
with by reference is the conventional default. I hope this might even
improve performance.
- This can legitimately happen (for example because of a non-determinism
causing a build-time dependency to be kept or not as a runtime
reference)
- Because of older Nix versions, it can happen that we encounter a
realisation with an (erroneously) empty set of dependencies, in which
case we don’t want to fail, but just warn the user and try to fix it.
Useful when we're using a daemon with a chroot store, e.g.
$ NIX_DAEMON_SOCKET_PATH=/tmp/chroot/nix/var/nix/daemon-socket/socket nix-daemon --store /tmp/chroot
Then the client can now connect with
$ nix build --store unix:///tmp/chroot/nix/var/nix/daemon-socket/socket?root=/tmp/chroot nixpkgs#hello
When adding a path to the local store (via `LocalStore::addToStore`),
ensure that the `ca` field of the provided `ValidPathInfo` does indeed
correspond to the content of the path.
Otherwise any untrusted user (or any binary cache) can add arbitrary
content-addressed paths to the store (as content-addressed paths don’t
need a signature).
I guess the rationale behind the old name wath that
`pathInfoIsTrusted(info)` returns `true` iff we would need to `blindly`
trust the path (because it has no valid signature and `requireSigs` is
set), but I find it to be a really confusing footgun because it's quite
natural to give it the opposite meaning.
Once a build is done, get back to the original derivation, and register
all the newly built outputs for this derivation.
This allows Nix to work properly with derivations that don't have all
their build inputs available − thus allowing garbage collection and
(once it's implemented) binary substitution
Changes:
* The divider lines are gone. These were in practice a bit confusing,
in particular with --show-trace or --keep-going, since then there
were multiple lines, suggesting a start/end which wasn't the case.
* Instead, multi-line error messages are now indented to align with
the prefix (e.g. "error: ").
* The 'description' field is gone since we weren't really using it.
* 'hint' is renamed to 'msg' since it really wasn't a hint.
* The error is now printed *before* the location info.
* The 'name' field is no longer printed since most of the time it
wasn't very useful since it was just the name of the exception (like
EvalError). Ideally in the future this would be a unique, easily
googleable error ID (like rustc).
* "trace:" is now just "…". This assumes error contexts start with
something like "while doing X".
Example before:
error: --- AssertionError ---------------------------------------------------------------------------------------- nix
at: (7:7) in file: /home/eelco/Dev/nixpkgs/pkgs/applications/misc/hello/default.nix
6|
7| x = assert false; 1;
| ^
8|
assertion 'false' failed
----------------------------------------------------- show-trace -----------------------------------------------------
trace: while evaluating the attribute 'x' of the derivation 'hello-2.10'
at: (192:11) in file: /home/eelco/Dev/nixpkgs/pkgs/stdenv/generic/make-derivation.nix
191| // (lib.optionalAttrs (!(attrs ? name) && attrs ? pname && attrs ? version)) {
192| name = "${attrs.pname}-${attrs.version}";
| ^
193| } // (lib.optionalAttrs (stdenv.hostPlatform != stdenv.buildPlatform && !dontAddHostSuffix && (attrs ? name || (attrs ? pname && attrs ? version)))) {
Example after:
error: assertion 'false' failed
at: (7:7) in file: /home/eelco/Dev/nixpkgs/pkgs/applications/misc/hello/default.nix
6|
7| x = assert false; 1;
| ^
8|
… while evaluating the attribute 'x' of the derivation 'hello-2.10'
at: (192:11) in file: /home/eelco/Dev/nixpkgs/pkgs/stdenv/generic/make-derivation.nix
191| // (lib.optionalAttrs (!(attrs ? name) && attrs ? pname && attrs ? version)) {
192| name = "${attrs.pname}-${attrs.version}";
| ^
193| } // (lib.optionalAttrs (stdenv.hostPlatform != stdenv.buildPlatform && !dontAddHostSuffix && (attrs ? name || (attrs ? pname && attrs ? version)))) {
With the `ca-derivation` experimental features, non-ca derivations used
to have their output paths returned as unknown as long as they weren't
built (because of a mistake in the code that systematically erased the
previous value)
Thanks @regnat and @edolstra for catching this and comming up with the
solution.
They way I had generalized those is wrong, because local settings for
non-local stores is confusing default. And due to the nature of C++
inheritance, fixing the defaults is more annoying than it should be.
Additionally, I thought we might just drop the check in the substitution
logic since `Store::addToStore` is now streaming, but @regnat rightfully
pointed out that as it downloads dependencies first, that would still be
too late, and also waste effort on possibly unneeded/unwanted
dependencies.
The simple and correct thing to do is just make a store method for the
boolean logic, keeping all the setting and key stuff the way it was
before. That new method is both used by `LocalStore::addToStore` and the
substitution goal check. Perhaps we might eventually make it fancier,
e.g. sending the ValidPathInfo to remote stores for them to validate,
but this is good enough for now.
We embrace virtual the rest of the way, and get rid of the
`assert(false)` 0-param constructors.
We also list config base classes first, so the constructor order is
always:
1. all the configs
2. all the stores
Each in the same order
PRs #4370 and #4348 had a bad interaction in that the second broke the fist
one in a not trivial way.
The issue was that since #4348 the logic for detecting whether a
derivation output is already built requires some logic that was specific
to the `LocalStore`.
It happens though that most of this logic could be upstreamed to any `Store`,
which is what this commit does.
This ignore was here because `queryPartialDrvOutputMap` was used both
1. as a cache to avoid having to re-read the derivation (when gc-ing for
example), and
2. as the source of truth for ca realisations
The use-case 2. required it to be able to work even when the derivation
wasn't there anymore (see https://github.com/NixOS/nix/issues/4138).
However, this use-case is now handled by `queryRealisation`, meaning
that we can safely error out if the derivation isn't there anymore
Rather than storing the derivation outputs as `drvPath!outputName` internally,
store them as `drvHashModulo!outputName` (or `outputHash!outputName` for
fixed-output derivations).
This makes the storage slightly more opaque, but enables an earlier
cutoff in cases where a fixed-output dependency changes (but keeps the
same output hash) − same as what we already do for input-addressed
derivations.
Add a new table for tracking the derivation output mappings.
We used to hijack the `DerivationOutputs` table for that, but (despite its
name), it isn't a really good fit:
- Its entries depend on the drv being a valid path, making it play badly with
garbage collection and preventing us to copy a drv output without copying
the whole drv closure too;
- It dosen't guaranty that the output path exists;
By using a different table, we can experiment with a different schema better
suited for tracking the output mappings of CA derivations.
(incidentally, this also fixes#4138)
