Introduce a `CancellationToken` for cancelling specific computations #1007

Veykril · 2025-10-17T17:01:54Z

Introduces a CancellationToken that can be used to cancel a specific database computation opposed to cancelling the whole runtime. This is traced by storing a second cancellation state on ZalsaLocal itself opposed to the runtime cancel flag.

When a thread gets cancelled, it will unwind as usual but with a different payload than pending write cancellation, threads blocked on such a cancelled thread will instead of propagating the cancellation run the computation they are blocked on themselves.

The cancellation state of the database gets reset when we exit the database TLS.

netlify · 2025-10-17T17:02:02Z

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	`2f99827`
🔍 Latest deploy log	https://app.netlify.com/projects/salsa-rs/deploys/695545814804290008c0aa3e

codspeed-hq · 2025-10-17T17:03:18Z

CodSpeed Performance Report

Merging #1007 will degrade performance by 8.36%

_{Comparing Veykril:push-kwpwsmmosonq (2f99827) with master (77c296f)}

Summary

⚡ 1 improvement
❌ 3 (👁 3) regressions
✅ 9 untouched

Benchmarks breakdown

	Benchmark	`BASE`	`HEAD`	Efficiency
👁	`amortized[SupertypeInput]`	2.9 µs	3 µs	-4.61%
👁	`many_tracked_structs`	8.7 µs	9.1 µs	-4.06%
👁	`amortized[InternedInput]`	2 µs	2.2 µs	-8.36%
⚡	`new[Input]`	10.5 µs	10.1 µs	+4.55%

Veykril · 2025-10-17T17:04:49Z

Looks like allocator noise from the new arc alloc

MichaReiser · 2025-10-17T17:08:32Z

Interesting. Is the idea to cancel a single thread rather than all threads? Won't this cancell all threads that currently block on any query executing on the thread being cancelled?

Veykril · 2025-10-17T17:22:35Z

Is the idea to cancel a single thread rather than all threads?

Yes, the thought is that this could allow implementing client side request cancellation for LSP requests for example.

Won't this cancell all threads that currently block on any query executing on the thread being cancelled?

That is the thing I will have to investigate. We unwind with the payload, we don't panic so I believe this does not set the panicking state of the thread actually. So this might actually just work (though I doubt it).

MichaReiser · 2025-10-17T17:27:08Z

That is the thing I will have to investigate. We unwind with the payload, we don't panic so I believe this does not set the panicking state of the thread actually. So this might actually just work (though I doubt it).

Oh, so other threads would reclaim and re-execute the queries then. Interesting.

Veykril · 2025-10-17T17:47:31Z

That would be the ideal I think

Veykril · 2025-10-19T16:52:17Z

I was wrong, unwinding does set the panicking flag after all (which probably makes more sense ...) hmm

src/cancelled.rs

MichaReiser · 2025-10-26T21:49:55Z

Would mind explaining the approach in the pr summary. I'm mainly interested in how it works if other threads are blocked on a query that gets cancelled (including if they participate in a cycle)

Veykril · 2025-12-26T12:44:31Z

.github/workflows/test.yml

      - name: Test with Miri
        run: cargo miri nextest run --no-fail-fast --tests
        env:
-          MIRIFLAGS: -Zmiri-disable-isolation -Zmiri-retag-fields


This is a default now

Copilot

Pull request overview

This PR introduces a CancellationToken that allows cancelling specific database computations instead of the entire runtime. The token is stored on ZalsaLocal and enables fine-grained cancellation control, with special handling during cycle fixpoint iteration where local cancellation is disabled to ensure cycles complete successfully.

Key changes:

Added CancellationToken type with atomic state management for local cancellation
Differentiated Cancelled::Local from Cancelled::PendingWrite to distinguish cancellation types
Modified blocking behavior so threads blocked on locally-cancelled threads recompute instead of propagating cancellation

Reviewed changes

Copilot reviewed 22 out of 22 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
`src/zalsa_local.rs`	Implements `CancellationToken` with atomic state management and integrates it into `ZalsaLocal`
`src/zalsa.rs`	Updates cancellation checks to handle both local and pending-write cancellation
`src/database.rs`	Adds public `cancellation_token()` API method
`src/cancelled.rs`	Adds `Cancelled::Local` variant for local cancellation
`src/attach.rs`	Resets cancellation state when exiting database TLS
`src/storage.rs`	Ensures proper cleanup of `zalsa_local` when converting to handle
`src/runtime.rs`	Adds `WaitResult::Cancelled`, renames `revision_canceled` to `revision_cancelled`
`src/function/sync.rs`	Reorders `ClaimResult` variants, adds `zalsa_local` parameter, handles cancelled release
`src/function/fetch.rs`	Returns `None` on local cancellation to trigger recomputation
`src/function/memo.rs`	Converts local cancellation to panic in fixpoint, fixes comment typo
`src/function/execute.rs`	Disables local cancellation during fixpoint iteration
`src/function/maybe_changed_after.rs`	Updates blocking to discard return value
`src/interned.rs`	Adds `#[allow(dead_code)]` for feature-gated field
`tests/cancellation_token.rs`	Basic cancellation token test
`tests/parallel/cancellation_token_*.rs`	Comprehensive parallel cancellation tests
`tests/dataflow.rs`	Refactors `Box<[usize]>` to `BTreeSet<usize>`
`tests/interned-revisions.rs`	Adjusts iteration counts for miri/non-miri
`.github/workflows/test.yml`	Removes `-Zmiri-retag-fields` flag
`tests/parallel/main.rs`	Adds new test module declarations

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

src/zalsa_local.rs

tests/parallel/cancellation_token_cycle_nested.rs

tests/interned-revisions.rs

MichaReiser

It probably makes sense to have this in Salsa but is an implementation outside Salsa considerably more work?

MichaReiser · 2025-12-31T14:30:26Z

src/function/memo.rs

+                            // We cannot handle local cancellations in fixpoints
+                            // so we treat it as a general cancellation / panic.
+                            //
+                            // We shouldn't hit this though as we disable local cancellation
+                            // in cycles.


block_on_heads isn't used by fixpoint anymore. It's only used by the FallbackImmediate cycle recovery strategy

Is FallbackImmediate even necessary anymore? iirc you've removed the requirement of fixpoint having to converge right?

Whether it's still necessary depends on r-a. We aren't using it in ty

iirc you've removed the requirement of fixpoint having to converge right?

No, fixpoint queries still have to converge. I only changed how we run nested queries.

Ah i mixed it up with #991

src/function/sync.rs

tests/dataflow.rs

src/attach.rs

MichaReiser · 2025-12-31T14:48:14Z

src/attach.rs

+                if let Some(attached) = self.state {
+                    if let Some(prev) = attached.database.replace(self.prev) {
+                        // SAFETY: `prev` is a valid pointer to a database.
+                        unsafe { prev.as_ref().zalsa_local().uncancel() };


What if prev was cancelled too? Which is possible because CancellationToken is Clone.

Should we instead simply consider that db as "dead". Meaning, you have to call Clone to get a new Db afterwards (What you do is probably the opposite, you clone the Db before passing it to a request handler, and the db is later dropped)

If prev was cancelled too then the query that initially attached it will reset the state once it exits its scope. We always put the databases back that we replaced in attached when the scope ends due to this.

Okay, I think I got confused by the variable being named prev. I assumed it's DbGuard.prev when it is the database stored in attached.database (which isn't prev but the current db up to when we started dropping).

I think there's still a risk that we uncancel if we have something like a -> b -> a and a gets cancelled. I understand that we call a.uncancel() when we drop the guard for a -> b -> a (the last db).

Should we make this a method on attached instead, saying that it resets the database state (e.g. .attached.exit`). This also avoids duplicating the same logic.

Right, I believe this should be fine though. This could basically cause cancellations to be .. well cancelled. But only if you interleave a bunch of different databases (which ideally we wouldn't even allow imo, but is kind of a necessary evil given the lack of speculative execution).

Veykril · 2025-12-31T15:20:33Z

I am not sure how we can make this work as an implementation outside of salsa, given it needs to special case around fixpoint executions. And to me this feels very much like something salsa should offer natively anyways.

MichaReiser · 2025-12-31T16:29:55Z

src/attach.rs

+                if let Some(attached) = self.state {
+                    if let Some(prev) = attached.database.replace(self.prev) {
+                        // SAFETY: `prev` is a valid pointer to a database.
+                        unsafe { prev.as_ref().zalsa_local().uncancel() };


Okay, I think I got confused by the variable being named prev. I assumed it's DbGuard.prev when it is the database stored in attached.database (which isn't prev but the current db up to when we started dropping).

I think there's still a risk that we uncancel if we have something like a -> b -> a and a gets cancelled. I understand that we call a.uncancel() when we drop the guard for a -> b -> a (the last db).

Should we make this a method on attached instead, saying that it resets the database state (e.g. .attached.exit`). This also avoids duplicating the same logic.

MichaReiser · 2025-12-31T16:30:49Z

Curious to see how you use this in r-a.

Veykril · 2026-01-01T14:46:42Z

Basically like rust-lang/rust-analyzer#21380. Right now rust-analyzer will run LSP requests to completion even if they were cancelled (unless a user types/changes an input of course)

Veykril force-pushed the push-kwpwsmmosonq branch 3 times, most recently from db715b7 to 47b7a7d Compare October 19, 2025 16:45

Veykril force-pushed the push-kwpwsmmosonq branch 12 times, most recently from f97346e to de32cab Compare October 26, 2025 15:52

Veykril marked this pull request as ready for review October 26, 2025 15:52

Veykril commented Oct 26, 2025

View reviewed changes

src/cancelled.rs Outdated Show resolved Hide resolved

Veykril force-pushed the push-kwpwsmmosonq branch 3 times, most recently from ba0f832 to 2318de4 Compare October 26, 2025 16:32

Veykril requested a review from MichaReiser October 26, 2025 17:14

Veykril requested a review from MichaReiser December 26, 2025 12:13

Veykril force-pushed the push-kwpwsmmosonq branch from f26560a to 2415a00 Compare December 26, 2025 12:30

Veykril commented Dec 26, 2025

View reviewed changes

Veykril force-pushed the push-kwpwsmmosonq branch 6 times, most recently from 6e14f3d to 7d927c2 Compare December 26, 2025 18:43

Veykril requested a review from Copilot December 31, 2025 14:00

Copilot started reviewing on behalf of Veykril December 31, 2025 14:00 View session

Copilot AI reviewed Dec 31, 2025

View reviewed changes

src/zalsa_local.rs Outdated Show resolved Hide resolved

tests/parallel/cancellation_token_cycle_nested.rs Outdated Show resolved Hide resolved

tests/interned-revisions.rs Outdated Show resolved Hide resolved

Veykril force-pushed the push-kwpwsmmosonq branch from 7d927c2 to 8a82499 Compare December 31, 2025 14:18

Veykril and others added 4 commits December 31, 2025 15:20

Introduce a CancellationToken for cancelling specific computations

5be84e6

Address reviews

796d08b

Disable local cancellation while within a cycle computation

2141ac9

Add cycle test

c07e297

Veykril force-pushed the push-kwpwsmmosonq branch from 8a82499 to b2e2844 Compare December 31, 2025 14:20

MichaReiser reviewed Dec 31, 2025

View reviewed changes

Veykril added 2 commits December 31, 2025 16:42

Fix set_cancellation_disabled implementation

48ad331

Expose zalsa_local from ClaimGuard

08d8981

Veykril force-pushed the push-kwpwsmmosonq branch 2 times, most recently from 15c4e98 to c5b15ae Compare December 31, 2025 15:45

Document attached states

2f99827

Veykril force-pushed the push-kwpwsmmosonq branch from c5b15ae to 2f99827 Compare December 31, 2025 15:47

MichaReiser approved these changes Dec 31, 2025

View reviewed changes

Veykril mentioned this pull request Jan 1, 2026

Correctly implement client side request cancellation support rust-lang/rust-analyzer#21380

Draft

Introduce a CancellationToken for cancelling specific computations #1007

Are you sure you want to change the base?

Introduce a CancellationToken for cancelling specific computations #1007

Uh oh!

Conversation

Veykril commented Oct 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

netlify bot commented Oct 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ Deploy Preview for salsa-rs canceled.

Uh oh!

codspeed-hq bot commented Oct 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merging #1007 will degrade performance by 8.36%

Summary

Benchmarks breakdown

Uh oh!

Veykril commented Oct 17, 2025

Uh oh!

MichaReiser commented Oct 17, 2025

Uh oh!

Veykril commented Oct 17, 2025

Uh oh!

MichaReiser commented Oct 17, 2025

Uh oh!

Veykril commented Oct 17, 2025

Uh oh!

Veykril commented Oct 19, 2025

Uh oh!

Uh oh!

MichaReiser commented Oct 26, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

MichaReiser left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Veykril commented Dec 31, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

MichaReiser commented Dec 31, 2025

Uh oh!

Veykril commented Jan 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Introduce a `CancellationToken` for cancelling specific computations #1007

Introduce a `CancellationToken` for cancelling specific computations #1007

Veykril commented Oct 17, 2025 •

edited

Loading

netlify bot commented Oct 17, 2025 •

edited

Loading

codspeed-hq bot commented Oct 17, 2025 •

edited

Loading

Veykril commented Dec 31, 2025 •

edited

Loading