lix-project/lix
20
57
Fork 31
Code Issues 347 Code Review (Gerrit) Projects 3 Releases Wiki Activity

Garbage collection of "too many" paths is unsound #524

New issue
Open
opened 2024-09-14 18:21:06 +00:00 by raito · 4 comments
raito commented 2024-09-14 18:21:06 +00:00
Owner
Copy link

Describe the bug

TL;DR: Garbage collection of large Nix stores can leave the Nix store in an inconsistent state, e.g. referrers without ValidPath rows.

Motivated by https://gerrit.lix.systems/c/lix/+/1916. I have found a curious case of corruption, inconsistencies, idk?

Steps To Reproduce

Way 1: Consult the test in the CL, emit a DELETE to a referenced valid path that bypasses the foreign key constraints (can be done by PRAGMA foreign_keys = OFF; in the session)
Way 2: To discover yet.

Expected behavior

(1) GC should always succeed.
(2) GC/deletion/etc. should never leave the DB in the wrong state.

nix --version output

Happened to me for every version of Nix ≥ 2.18, including Lix ≥ 2.90.

Additional context

The diagnostic may be wrong.

## Describe the bug TL;DR: Garbage collection of large Nix stores can leave the Nix store in an inconsistent state, e.g. referrers without ValidPath rows. Motivated by https://gerrit.lix.systems/c/lix/+/1916. I have found a curious case of corruption, inconsistencies, idk? ## Steps To Reproduce Way 1: Consult the test in the CL, emit a `DELETE` to a **referenced valid path** that bypasses the foreign key constraints (can be done by `PRAGMA foreign_keys = OFF;` in the session) Way 2: To discover yet. ## Expected behavior (1) GC should always succeed. (2) GC/deletion/etc. should never leave the DB in the wrong state. ## `nix --version` output Happened to me for every version of Nix ≥ 2.18, including Lix ≥ 2.90. ## Additional context The diagnostic may be wrong.
raito added the
bug
 label 2024-09-14 18:21:06 +00:00
jade added the
Area/store
 label 2024-09-14 19:17:57 +00:00
jade commented 2024-09-14 19:18:40 +00:00
Owner
Copy link

#505

maybe also part of the observed behaviour

https://git.lix.systems/lix-project/lix/issues/505 maybe also part of the observed behaviour
raito commented 2024-09-14 21:20:09 +00:00
Author
Owner
Copy link

Actually, my root cause analysis seems wrong, FWIW, I expanded why in the Gerrit CL commit message, I will update the initial description.

Actually, my root cause analysis seems wrong, FWIW, I expanded why in the Gerrit CL commit message, I will update the initial description.
raito changed title from Garbage collection of long-chained derivations is unsound to Garbage collection of "too many" paths is unsound 2024-09-14 21:20:23 +00:00
jade commented 2024-09-15 08:34:54 +00:00
Owner
Copy link

I believe that option 1 ought to be impossible, since some trivial auditing reveals that we only call sqlite3_* functions (besides setting some pragmas or pure reads) inside of sqlite.cc inside of some kind of wrapper class that should have set the appropriate foreign keys pragma.

This leads me to wonder: races, somehow? It shouldn't be so, though.

I believe that option 1 ought to be impossible, since some trivial auditing reveals that we only call sqlite3_* functions (besides setting some pragmas or pure reads) inside of sqlite.cc inside of some kind of wrapper class that should have set the appropriate foreign keys pragma. This leads me to wonder: races, somehow? It shouldn't be so, though.
raito commented 2024-09-15 11:26:06 +00:00
Author
Owner
Copy link

@jade Yeah, this is what I wrote in the commit message of the CL as well. SQLite is transactional, though, so races are all serialized in front of the consistency in principle. You cannot add back a (referrer, reference) pair if the referrer does not exist. You cannot delete a valid path's referrer if there's still (referrer, reference) pairs in Refs table, etc.

@jade Yeah, this is what I wrote in the commit message of the CL as well. SQLite is transactional, though, so races are all serialized in front of the consistency in principle. You cannot add back a (referrer, reference) pair if the referrer does not exist. You cannot delete a valid path's referrer if there's still (referrer, reference) pairs in Refs table, etc.
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
release-2.91
sb/raito/phantom-referrers-gc
repl-overlay-example
reuse-compliance
release-2.90
git-series/activity-cleanup
sb/tom-hubrecht/primops.cc-v2
sb/bacchanalia/cache-strlen%topic=substr-perf
sb/arcuru/commit-msg
sb/qyriad/beta-1
sb/qyriad/2.90-beta.0
rbt/pre-commit-update
sb/rbt/test-branch
sb/rbt/justfile
sb/rbt/repl-overlays
sb/fuck/what
sb/rbt/pre-commit
sb/rbt/makeflags
sb/pennae/io-rewrite
sb/qyriad/maint-meson
sb/qyriad/maint/meson
sb/puck/meson-c-api
sb/pennae/parser-rewrite
sb/jade/burn-it-all
dependabot/github_actions/zeebe-io/backport-action-2.4.1
dependabot/github_actions/cachix/install-nix-action-25
dependabot/github_actions/actions/labeler-5
dependabot/github_actions/cachix/cachix-action-14
2.91.1
2.91.0
2.90.0
2.90.0-rc1
2.90-beta.1
2.90-beta.0
Labels
Clear labels   
Area/build-packaging

regarding the building or packaging of Lix itself

  
Area/cli

The command line interface of Lix

  
Area/evaluator

libexpr and anything else related to the evaluation of Nix expressions

  
Area/fetching

Fetching sources or other materials from the World Wide Web in Lix

  
Area/flakes

   
Area/language

language design, features, extensions, etc

  
Area/profiles

profiles: nix-env, nix profile, and installer flavors all alike

  
Area/protocol

The protocol, used for communicating between the client and the daemon

  
Area/releng

Release engineering

  
Area/remote-builds

remote builds

  
Area/repl

nix repl

  
Area/store

libstore and anything else relating to the Nix store

  
bug

it's a bug

  
crash 💥

Lix crashes (possibly bad error handling)

  
Cross Compilation

   
devx

development experience issues

  
docs

requires or is about documentation changes

  
Downstream Dependents

regarding an external tool's usage with Lix (e.g. nix-eval-jobs)

  
E/easy

bug is relatively simple and could simply be fixed with some minor effort

  
E/hard

this issue should probably be solved with close involvement of an expert

  
E/help wanted

maintainers may not get to this soon, but we want it fixed

  
E/reproducible

This bug is reproducible and obvious.

  
E/requires rearchitecture

this bug cannot be fixed properly without fixing a known architectural flaw

  
imported

Issue imported from upstream nix (please do not remove this label, it will cause dupes if import is rerun)

  
Needs Langver

We can't get to this until we implement language versioning

  
OS/Linux

Affects specifically Lix on Linux

  
OS/macOS

Affects specifically Lix on macOS

  
performance

make lix go faster

  
regression

This is a regression since some previous version of Lix

  
release-blocker

Blocks public release

  
RFD

Request for discussion

  
stability

Fixing this makes the software more stable

  
Status
blocked
Blocked on something

  
Status
invalid
not our bug, duplicate, similar

  
Status
postponed
closed but should maybe look at later (generally with S-wontfix)

  
Status
wontfix
We chose not to change this (see issue comments for more detail).

  
testing

Tests that should be added

  
testing/flakey

Tests that are unreliable/flakey (not to be confused with flakes that test)

  
ux

makes the user experience better :)

No labels
Area/build-packaging
Area/cli
Area/evaluator
Area/fetching
Area/flakes
Area/language
Area/profiles
Area/protocol
Area/releng
Area/remote-builds
Area/repl
Area/store
bug
crash 💥
Cross Compilation
devx
docs
Downstream Dependents
E/easy
E/hard
E/help wanted
E/reproducible
E/requires rearchitecture
imported
Needs Langver
OS/Linux
OS/macOS
performance
regression
release-blocker
RFD
stability
Status
blocked
Status
invalid
Status
postponed
Status
wontfix
testing
testing/flakey
ux
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: lix-project/lix#524
No description provided.
Delete branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?