fix(Layer): carry gate until eviction is complete (#7838)

the gate was accidentially being dropped before the final blocking
phase, possibly explaining the resident physical size global problems
during deletions.

it could had caused more harm as well, but the path is not actively
being tested because cplane no longer puts locationconfigs with higher
generation number during normal operation which prompted the last wave
of fixes.

Cc: #7341.
This commit is contained in:
Joonas Koivunen
2024-05-22 18:13:45 +03:00
committed by GitHub
parent e015b2bf3e
commit 62aac6c8ad

View File

@@ -12,7 +12,7 @@ use std::time::{Duration, SystemTime};
use tracing::Instrument;
use utils::id::TimelineId;
use utils::lsn::Lsn;
use utils::sync::heavier_once_cell;
use utils::sync::{gate, heavier_once_cell};
use crate::config::PageServerConf;
use crate::context::{DownloadBehavior, RequestContext};
@@ -1333,7 +1333,7 @@ impl LayerInner {
is_good_to_continue(&rx.borrow_and_update())?;
let Ok(_gate) = timeline.gate.enter() else {
let Ok(gate) = timeline.gate.enter() else {
return Err(EvictionCancelled::TimelineGone);
};
@@ -1421,7 +1421,7 @@ impl LayerInner {
Self::spawn_blocking(move || {
let _span = span.entered();
let res = self.evict_blocking(&timeline, &permit);
let res = self.evict_blocking(&timeline, &gate, &permit);
let waiters = self.inner.initializer_count();
@@ -1447,6 +1447,7 @@ impl LayerInner {
fn evict_blocking(
&self,
timeline: &Timeline,
_gate: &gate::GateGuard,
_permit: &heavier_once_cell::InitPermit,
) -> Result<(), EvictionCancelled> {
// now accesses to `self.inner.get_or_init*` wait on the semaphore or the `_permit`