End the GC stall on issue #5537: answer the collector's demand signal, and stop it paying for work that finds nothing - #5609

Merged

shai-almog merged 24 commits into

masterfrom

gc-stall-5537

Aug 29, 2026

Commits on Aug 28, 2026

Answer the collector's demand signal instead of idling through it (issue #5537)

shai-almog
and
claude
committed
1210c8a
Stop the collector paying for work that finds nothing (issue #5537)

shai-almog
and
claude
committed
b6ed99f
Fix a deadlock the latency work exposed, and close the bulk-copy barrier gap (issue #5537)

shai-almog
and
claude
committed
18a2def
Address the two review findings, and stop asserting the runner (issue #5537)

shai-almog
and
claude
committed
35b6d38
Retry a failed allocation by looping, not by recursing (issue #5537)

shai-almog
and
claude
committed
5321f73
Bind the new native for JavaScript, and stop gating "a collection is owed" on who asked (issue #5537)

shai-almog
and
claude
committed
80d819f
Fix three more review findings, one of which was in the measurement itself (issue #5537)

shai-almog
and
claude
committed
dcd5e55
Never time out the GC handshake (issue #5537)

shai-almog
and
claude
committed
176d741
Invalidate the JavaAPI cache on a deleted source too (issue #5537)

shai-almog
and
claude
committed
7890725
Back off between failed allocations, and make that path testable (issue #5537)

shai-almog
and
claude
committed
0acda2f
Stop the simulated-failure budget from refilling itself (issue #5537)

shai-almog
and
claude
committed
1c99355
Take the SATB mutex once per chunk on bulk copies, not once per reference
shai-almog
committed
9da3fef
Correct the duty-cycle figures the collector was inflating
shai-almog
committed
e5c870f
Stop the duty figure inflating itself, and shut the bulk SATB path down cleanly
shai-almog
committed
76126e2
Register the whole bulk barrier, not each range inside it
shai-almog
committed
a4fa589
Never trace a new SATB discovery with the barrier already down
shai-almog
committed
d14a4b9
Record the trial clear and what satbReopens is for (issue #5537)
shai-almog
committed
efa53ab
Count the same threads in the duty numerator that the denominator counts
shai-almog
committed
1fdd221
Make an empty catch the only way out of SATB termination
shai-almog
committed
926dc96
Bound the lost-request window, and divide duty by real thread-time
shai-almog
committed
06d7afc
Make the bulk barrier's fast-path check safe across the trial clear
shai-almog
committed
0b474db
Handshake bulk copies with mark STARTUP, and make thread-time exact
shai-almog
committed
2249bed
Hold the bulk registration across the copy, not just the logging
shai-almog
committed
becc8db
Correct the cost claim for the spanning bulk registration
shai-almog
committed
483f93b

Navigation Menu

Uh oh!

End the GC stall on issue #5537: answer the collector's demand signal, and stop it paying for work that finds nothing - #5609

End the GC stall on issue #5537: answer the collector's demand signal, and stop it paying for work that finds nothing#5609
shai-almog merged 24 commits into
masterfrom
gc-stall-5537

Commits on Aug 28, 2026

Answer the collector's demand signal instead of idling through it (issue #5537)

Stop the collector paying for work that finds nothing (issue #5537)

Fix a deadlock the latency work exposed, and close the bulk-copy barrier gap (issue #5537)

Address the two review findings, and stop asserting the runner (issue #5537)

Retry a failed allocation by looping, not by recursing (issue #5537)

Bind the new native for JavaScript, and stop gating "a collection is owed" on who asked (issue #5537)

Fix three more review findings, one of which was in the measurement itself (issue #5537)

Never time out the GC handshake (issue #5537)

Invalidate the JavaAPI cache on a deleted source too (issue #5537)

Back off between failed allocations, and make that path testable (issue #5537)

Stop the simulated-failure budget from refilling itself (issue #5537)

Take the SATB mutex once per chunk on bulk copies, not once per reference

Correct the duty-cycle figures the collector was inflating

Stop the duty figure inflating itself, and shut the bulk SATB path down cleanly

Register the whole bulk barrier, not each range inside it

Never trace a new SATB discovery with the barrier already down

Record the trial clear and what satbReopens is for (issue #5537)

Count the same threads in the duty numerator that the denominator counts

Make an empty catch the only way out of SATB termination

Bound the lost-request window, and divide duty by real thread-time

Make the bulk barrier's fast-path check safe across the trial clear

Handshake bulk copies with mark STARTUP, and make thread-time exact

Hold the bulk registration across the copy, not just the logging

Correct the cost claim for the spanning bulk registration