nixpkgs/pkgs/top-level/release-attrpaths-superset.nix

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

195 lines
6.3 KiB
Nix
Raw Normal View History

pkgs/top-level/release-attrpaths-superset.nix: init This file walks the entire nixpkgs tree and emits a superset of all release attrnames in only 44 seconds on a 3ghz CPU, using 5 gbytes of memory. By comparison, on the same CPU the `nix-env` hack used by ofborg on every PR submission requires 41 *minutes* and peaks at 60 gbytes, even with checkMeta turned off. Full details below. This is: - 46x faster (or 2.1% of the elapsed time) - 12.5x less memory (or 8.0% of the peak memory usage) In order to replace the ofborg check, this list of attrnames must then be post-filtered for platform-relevance. However, crucially, the post-filtering can be done *in parallel* on multiple cores by splitting the attrname list in to chunks. Generating the list of attrnames cannot be parallelized because it is a single-threaded cppnix task. This PR also adds `recurseForDerivations` where necessary within nixpkgs in order to make this possible -- it screens out various non-tryEval-catchable failures and infinite recursions. Before undraftifying, I will add an invocation of this command to the CI tests, to ensure that the work performed here is not immediately undone. My next PR will then add an additional CI check confirming that the emitted attrpaths are in fact a superset of the release attrpaths calculated by the slow-memory-hog ofborg method. I have manually confirmed that this is the case at the tip commit of this PR, but we need CI to make sure this remains true until ofborg switches to this more-efficient method of calculation; at that point the superset-check can be dropped. According to GNU Time, Command being timed: "nix-instantiate --eval --strict --json pkgs/top-level/release-attrpaths-superset.nix -A names" User time (seconds): 44.88 System time (seconds): 8.09 Percent of CPU this job got: 99% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:53.20 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 4823028 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 3611240 Voluntary context switches: 113 Involuntary context switches: 949 Swaps: 0 File system inputs: 1480 File system outputs: 5944 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 Compared to release-outpaths.nix: Command being timed: "nix-env -qaP --no-name --out-path --arg checkMeta false --argstr path /git/work/pr/release-outpaths -f pkgs/top-level/release-outpaths.nix" User time (seconds): 2120.67 System time (seconds): 337.80 Percent of CPU this job got: 98% Elapsed (wall clock) time (h:mm:ss or m:ss): 41:37.91 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 60171768 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 2 Minor (reclaiming a frame) page faults: 230608113 Voluntary context switches: 8876 Involuntary context switches: 22275 Swaps: 0 File system inputs: 62624 File system outputs: 72 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0
2023-11-23 02:52:27 +01:00
# This expression will, as efficiently as possible, dump a
# *superset* of all attrpaths of derivations which might be
# part of a release on *any* platform.
#
# Both this expression and what ofborg uses (release-outpaths.nix)
# are essentially single-threaded (under the current cppnix
# implementation).
#
# This expression runs much, much, much faster and uses much, much
# less memory than the ofborg script by skipping the
# platform-relevance checks. The ofborg outpaths.nix script takes
# half an hour on a 3ghz core and peaks at 60gbytes of memory; this
# expression runs on the same machine in 44 seconds with peak memory
# usage of 5gbytes.
#
# Once you have the list of attrnames you can split it up into
# $NUM_CORES batches and run the platform checks separately for each
# batch, in parallel.
#
# To dump the attrnames:
#
# nix-instantiate --eval --strict --json pkgs/top-level/release-attrpaths-superset.nix -A names
#
{ lib ? import (path + "/lib")
, trace ? false
, enableWarnings ? true
, checkMeta ? true
, path ? ./../..
}:
let
# No release package attrpath may have any of these attrnames as
# its initial component.
#
# If you can find a way to remove any of these entries without
# causing CI to fail, please do so.
#
excluded-toplevel-attrs = {
# spliced packagesets
__splicedPackages = true;
pkgsBuildBuild = true;
pkgsBuildHost = true;
pkgsBuildTarget = true;
pkgsHostHost = true;
pkgsHostTarget = true;
pkgsTargetTarget = true;
buildPackages = true;
targetPackages = true;
# cross packagesets
pkgsLLVM = true;
pkgsMusl = true;
pkgsStatic = true;
pkgsCross = true;
pkgsi686Linux = true;
pkgsLinux = true;
pkgsExtraHardening = true;
pkgs/top-level/release-attrpaths-superset.nix: init This file walks the entire nixpkgs tree and emits a superset of all release attrnames in only 44 seconds on a 3ghz CPU, using 5 gbytes of memory. By comparison, on the same CPU the `nix-env` hack used by ofborg on every PR submission requires 41 *minutes* and peaks at 60 gbytes, even with checkMeta turned off. Full details below. This is: - 46x faster (or 2.1% of the elapsed time) - 12.5x less memory (or 8.0% of the peak memory usage) In order to replace the ofborg check, this list of attrnames must then be post-filtered for platform-relevance. However, crucially, the post-filtering can be done *in parallel* on multiple cores by splitting the attrname list in to chunks. Generating the list of attrnames cannot be parallelized because it is a single-threaded cppnix task. This PR also adds `recurseForDerivations` where necessary within nixpkgs in order to make this possible -- it screens out various non-tryEval-catchable failures and infinite recursions. Before undraftifying, I will add an invocation of this command to the CI tests, to ensure that the work performed here is not immediately undone. My next PR will then add an additional CI check confirming that the emitted attrpaths are in fact a superset of the release attrpaths calculated by the slow-memory-hog ofborg method. I have manually confirmed that this is the case at the tip commit of this PR, but we need CI to make sure this remains true until ofborg switches to this more-efficient method of calculation; at that point the superset-check can be dropped. According to GNU Time, Command being timed: "nix-instantiate --eval --strict --json pkgs/top-level/release-attrpaths-superset.nix -A names" User time (seconds): 44.88 System time (seconds): 8.09 Percent of CPU this job got: 99% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:53.20 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 4823028 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 3611240 Voluntary context switches: 113 Involuntary context switches: 949 Swaps: 0 File system inputs: 1480 File system outputs: 5944 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 Compared to release-outpaths.nix: Command being timed: "nix-env -qaP --no-name --out-path --arg checkMeta false --argstr path /git/work/pr/release-outpaths -f pkgs/top-level/release-outpaths.nix" User time (seconds): 2120.67 System time (seconds): 337.80 Percent of CPU this job got: 98% Elapsed (wall clock) time (h:mm:ss or m:ss): 41:37.91 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 60171768 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 2 Minor (reclaiming a frame) page faults: 230608113 Voluntary context switches: 8876 Involuntary context switches: 22275 Swaps: 0 File system inputs: 62624 File system outputs: 72 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0
2023-11-23 02:52:27 +01:00
};
# No release package attrname may have any of these at a component
# anywhere in its attrpath. These are the names of gigantic
# top-level attrsets that have leaked into so many sub-packagesets
# that it's easier to simply exclude them entirely.
#
# If you can find a way to remove any of these entries without
# causing CI to fail, please do so.
#
excluded-attrnames-at-any-depth = {
lib = true;
override = true;
__functor = true;
__functionArgs = true;
newScope = true;
scope = true;
pkgs = true;
buildHaskellPackages = true;
buildPackages = true;
generateOptparseApplicativeCompletions = true;
callPackage = true;
mkDerivation = true;
overrideDerivation = true;
overrideScope = true;
overrideScope' = true;
# Special case: lib/types.nix leaks into a lot of nixos-related
# derivations, and does not eval deeply.
type = true;
};
# __attrsFailEvaluation is a temporary workaround to get top-level
# eval to succeed (under builtins.tryEval) for the entire
# packageset, without deep invasve changes into individual
# packages.
#
# Now that CI has been added, ensuring that top-level eval will
# not be broken by any new commits, you should not add any new
# occurrences of __attrsFailEvaluation, and should remove them
# wherever you are able to (doing so will likely require deep
# adjustments within packages). Once all of the uses of
# __attrsFailEvaluation are removed, it will be deleted from the
# routine below. In the meantime,
#
# The intended semantics are that an attrpath rooted at pkgs is
# part of the (unfiltered) release jobset iff all of the following
# are true:
#
# 1. The first component of the attrpath is not in
# `excluded-toplevel-attrs`
#
# 2. No attrname in the attrpath belongs to the list of forbidden
# attrnames `excluded-attrnames-at-any-depth`
#
# 3. The attrpath leads to a value for which lib.isDerivation is true
#
# 4. No proper prefix of the attrpath has __attrsFailEvaluation=true
#
# 5. Any proper prefix of the attrpath at which lib.isDerivation
# is true also has __recurseIntoDerivationForReleaseJobs=true.
#
# The last condition is unfortunately necessary because there are
# Hydra release jobnames which have proper prefixes which are
# attrnames of derivations (!). We should probably restructure
# the job tree so that this is not the case.
#
justAttrNames = path: value:
let
attempt =
if lib.isDerivation value &&
# in some places we have *derivations* with jobsets as subattributes, ugh
!(value.__recurseIntoDerivationForReleaseJobs or false) then
[ path ]
# Even wackier case: we have meta.broken==true jobs with
# !meta.broken jobs as subattributes with license=unfree, and
# check-meta.nix won't throw an "unfree" failure because the
# enclosing derivation is marked broken. Yeah. Bonkers.
# We should just forbid jobsets enclosed by derivations.
else if lib.isDerivation value &&
!value.meta.available then []
else if !(lib.isAttrs value) then []
else if (value.__attrsFailEvaluation or false) then []
else lib.pipe value [
(builtins.mapAttrs
(name: value:
if excluded-attrnames-at-any-depth.${name} or false then [] else
(justAttrNames (path ++ [name]) value)))
builtins.attrValues
builtins.concatLists
];
seq = builtins.deepSeq attempt attempt;
tried = builtins.tryEval seq;
result =
if tried.success
then tried.value
else if enableWarnings && path != [ "AAAAAASomeThingsFailToEvaluate" ]
then lib.warn "tryEval failed at: ${lib.concatStringsSep "." path}" []
else [];
in
if !trace
then result
else lib.trace "** ${lib.concatStringsSep "." path}" result;
unfiltered = import ./release-outpaths.nix {
inherit checkMeta;
attrNamesOnly = true;
inherit path;
};
filtered = lib.pipe unfiltered [
(pkgs: builtins.removeAttrs pkgs (builtins.attrNames excluded-toplevel-attrs))
];
paths =
[
# I am not entirely sure why these three packages end up in
# the Hydra jobset. But they do, and they don't meet the
# criteria above, so at the moment they are special-cased.
[ "pkgsLLVM" "stdenv" ]
[ "pkgsStatic" "stdenv" ]
[ "pkgsMusl" "stdenv" ]
] ++ justAttrNames [] filtered;
names =
map (path: (lib.concatStringsSep "." path)) paths;
in
{
inherit paths names;
}