core: add a limited form of the linked-list callsite registry for v0.1.x #2083

hawkw · 2022-04-19T21:24:54Z

Motivation

Currently on v0.1.x, the global callsite registry is implemented as a
Mutex<Vec<&'static dyn Callsite>>. This has a few downsides:

Every time a callsite is registered, the Mutex must be locked. This
can cause a deadlock when a register_callsite implementation calls
into code that also registers a callsite. See tracing, core: fix deadlock in register_callsite #2020 for details.
Registering callsites is relatively slow and forces the program to
synchronize on the callsite registry (and not on individual
callsites). This means that if two threads are both registering
totally different callsites, both threads must wait for the lock.
Although this overhead is amortized over the entire rest of the
program, it does have an impact in short-running applications where
any given callsite may only be hit once or twice.
Memory allocation may occur while the registry is locked. This makes
the use of tracing inside of memory allocators difficult or
impossible.

On the master branch (v0.2.x), PR #988 rewrote the callsite registry
to use an intrusive atomic singly-linked list of Callsites. This
removed the need for locking the callsite registry, and made it possible
to register callsites without ever allocating memory. Because the
callsite registry on v0.2 will never allocate, this also makes it
possible to compile tracing-core for no_std platforms without
requiring liballoc. Unfortunately, however, we cannot use an identical
implementation on v0.1.x, because using the intrusive linked-list
registry for all callsites requires a breaking change to the
Callsite trait (in order to add a way to get the callsite's
linked-list node), which we cannot make on v0.1.x.

Solution

This branch adds a linked-list callsite registry for v0.1.x in a way
that allows us to benefit from some of the advantages of this approach
in a majority of cases. The trick is introducing a new DefaultCallsite
type in tracing-core that implements the Callsite trait. This type
can contain an intrusive list node, and when a callsite is a
DefaultCallsite, we can register it without having to push it to the
Mutex<Vec<...>>. The locked vec still exists, for Callsites that
are not DefaultCallsites, so we can't remove the liballoc
dependency, but in most cases, we can avoid the mutex and allocation.

Naturally, tracing has been updated to use the DefaultCallsite type
from tracing-core, so the Vec will only be used in the following
cases:

User code has a custom implementation of the Callsite trait, which
is not terribly common.
An older version of the tracing crate is in use.

This fixes the deadlock described in #2020 when DefaultCallsites are
used. Additionally, it should reduce the performance impact and memory
usage of the callsite registry.

Furthermore, I've changed the subscriber registry so that a
RwLock<Vec<dyn Dispatch>> is only used when there actually are
multiple subscribers in use. When there's only a global default
subscriber, we can once again avoid locking for the registry of
subscribers. When the std feature is disabled, thread-local scoped
dispatchers are unavailable, so the single global subscriber will always
be used on no_std.

We can make additional changes, such as the ones from #2020, to also
resolve potential deadlocks when non-default callsites are in use, but
since this branch rewrites a lot of the callsite registry code, that
should probably be done in a follow-up.

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

This way, `callsite::register` also does the right thing Signed-off-by: Eliza Weisman <eliza@buoyant.io>

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

hawkw · 2022-04-19T21:25:23Z

@LucioFranco I tagged you on this since you wrote the original linked-list registry PR #988 ages ago, hope that was okay!

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

LucioFranco

LGTM only have some questions but nothing blocking! This seems like a nice work around!

LucioFranco · 2022-04-20T19:26:07Z

tracing-core/src/callsite.rs

+                debug_assert_eq!(
+                    _state,
+                    Self::REGISTERING,
+                    "weird callsite registration state"


lol do we want a better message here?

LucioFranco · 2022-04-20T19:27:19Z

tracing-core/src/callsite.rs

+        let interest = match () {
+            _ if interest.is_never() => Self::INTEREST_NEVER,
+            _ if interest.is_always() => Self::INTEREST_ALWAYS,
+            _ => Self::INTEREST_SOMETIMES,
+        };


wait what is this, does this just replace an if/else if/else?

yeah, it's a couple lines shorter...this was in code that was already there in tracing and was moved to this file. i could change it to an if if you think that's clearer...

nah it just took me a min to figure out what was going on lol. I don't have any strong opinions.

LucioFranco · 2022-04-20T19:33:11Z

tracing-core/src/callsite.rs

+        let this_interest = dispatch.register_callsite(meta);
+        interest = match interest.take() {
+            None => Some(this_interest),
+            Some(that_interest) => Some(that_interest.and(this_interest)),


"that_interest.and(this_interest)" this is poetic

hawkw · 2022-04-21T20:10:18Z

@davidbarsky were you planning to look at this one as well?

davidbarsky · 2022-04-22T15:57:25Z

@hawkw not really, but i'll give it a cursory review and probably stamp it.

@jamesmunns

# 0.1.27 (June 7, 2022) This release of `tracing-core` introduces a new `DefaultCallsite` type, which can be used by instrumentation crates rather than implementing their own callsite types. Using `DefaultCallsite` may offer reduced overhead from callsite registration. ### Added - `DefaultCallsite`, a pre-written `Callsite` implementation for use in instrumentation crates ([#2083]) - `ValueSet::len` and `Record::len` methods returning the number of fields in a `ValueSet` or `Record` ([#2152]) ### Changed - Replaced `lazy_static` dependency with `once_cell` ([#2147]) ### Documented - Added documentation to the `callsite` module ([#2088], [#2149]) Thanks to new contributors @jamesmunns and @james7132 for contributing to this release! [#2083]: #2083 [#2152]: #2152 [#2147]: #2147 [#2088]: #2088 [#2149]: #2149

@jamesmunns

# 0.1.27 (June 7, 2022) This release of `tracing-core` introduces a new `DefaultCallsite` type, which can be used by instrumentation crates rather than implementing their own callsite types. Using `DefaultCallsite` may offer reduced overhead from callsite registration. ### Added - `DefaultCallsite`, a pre-written `Callsite` implementation for use in instrumentation crates ([#2083]) - `ValueSet::len` and `Record::len` methods returning the number of fields in a `ValueSet` or `Record` ([#2152]) ### Changed - Replaced `lazy_static` dependency with `once_cell` ([#2147]) ### Documented - Added documentation to the `callsite` module ([#2088], [#2149]) Thanks to new contributors @jamesmunns and @james7132 for contributing to this release! [#2083]: #2083 [#2152]: #2152 [#2147]: #2147 [#2088]: #2088 [#2149]: #2149

# 0.1.35 (June 7, 2022) This release reduces the overhead of callsite registration by using new `tracing-core` APIs. ### Added - Use `DefaultCallsite` to reduce callsite registration overhead ([#2083]) ### Changed - `tracing-core`: updated to [0.1.27][core-0.1.27] [core-0.1.27]: https://github.com/tokio-rs/tracing/releases/tag/tracing-core-0.1.27 [#2088]: #2083

# 0.1.35 (June 8, 2022) This release reduces the overhead of callsite registration by using new `tracing-core` APIs. ### Added - Use `DefaultCallsite` to reduce callsite registration overhead ([#2083]) ### Changed - `tracing-core`: updated to [0.1.27][core-0.1.27] [core-0.1.27]: https://github.com/tokio-rs/tracing/releases/tag/tracing-core-0.1.27 [#2088]: #2083

When the interest cache is configured as part of initing the `LogTracer` before the global dispatcher is set, the main thread fails to log. The problem is that the interests cache registers a dummy callsite. With tokio-rs#2083 the callsite registration changed to use `dispatcher::get_default`. This call internally sets the thread local `default` dispatcher to `Dispatch::none` when there isn't yet any dispatcher set. This `default` dispatcher isn't overwritten when setting the global dispatcher. The solution is to not store the `Dispatch::none` as `default` dispatcher.

When the interest cache is configured as part of initing the `LogTracer` before the global dispatcher is set, the main thread fails to log. The problem is that the interests cache registers a dummy callsite. With #2083 the callsite registration changed to use `dispatcher::get_default`. This call internally sets the thread local `default` dispatcher to `Dispatch::none` when there isn't yet any dispatcher set. This `default` dispatcher isn't overwritten when setting the global dispatcher. The solution is to not store the `Dispatch::none` as `default` dispatcher.

When the interest cache is configured as part of initing the `LogTracer` before the global dispatcher is set, the main thread fails to log. The problem is that the interests cache registers a dummy callsite. With tokio-rs#2083 the callsite registration changed to use `dispatcher::get_default`. This call internally sets the thread local `default` dispatcher to `Dispatch::none` when there isn't yet any dispatcher set. This `default` dispatcher isn't overwritten when setting the global dispatcher. The solution is to not store the `Dispatch::none` as `default` dispatcher.

## Motivation A deadlock exists when a collector's `register_callsite` method calls into code that also contains tracing instrumentation and triggers a second `register_callsite` call for the same callsite. This is because the current implementation of the `MacroCallsite` type holds a `core::sync::Once` which it uses to ensure that it is only added to the callsite registry a single time. This deadlock was fixed in v0.1.x in PR #2083, but the issue still exists on v0.2.x. ## Solution This branch forward-ports the solution from #2083. Rather than using a `core::sync::Once`, we now track the callsite's registration state directly in `MacroCallsite`. If a callsite has started registering, but has not yet completed, subsequent `register` calls will just immediately receive an `Interest::sometimes` until the registration has completed, rather than waiting to attempt their own registration. I've also forward-ported the tests for this that were added in #2083.

…1.x (tokio-rs#2083) ## Motivation Currently on `v0.1.x`, the global callsite registry is implemented as a `Mutex<Vec<&'static dyn Callsite>>`. This has a few downsides: * Every time a callsite is registered, the `Mutex` must be locked. This can cause a deadlock when a `register_callsite` implementation calls into code that _also_ registers a callsite. See tokio-rs#2020 for details. * Registering callsites is relatively slow and forces the program to synchronize on the callsite registry (and not on *individual* callsites). This means that if two threads are both registering totally different callsites, both threads must wait for the lock. Although this overhead is amortized over the entire rest of the program, it does have an impact in short-running applications where any given callsite may only be hit once or twice. * Memory allocation may occur while the registry is locked. This makes the use of `tracing` inside of memory allocators difficult or impossible. On the `master` branch (v0.2.x), PR tokio-rs#988 rewrote the callsite registry to use an intrusive atomic singly-linked list of `Callsite`s. This removed the need for locking the callsite registry, and made it possible to register callsites without ever allocating memory. Because the callsite registry on v0.2 will *never* allocate, this also makes it possible to compile `tracing-core` for `no_std` platforms without requiring `liballoc`. Unfortunately, however, we cannot use an identical implementation on v0.1.x, because using the intrusive linked-list registry for *all* callsites requires a breaking change to the `Callsite` trait (in order to add a way to get the callsite's linked-list node), which we cannot make on v0.1.x. ## Solution This branch adds a linked-list callsite registry for v0.1.x in a way that allows us to benefit from *some* of the advantages of this approach in a majority of cases. The trick is introducing a new `DefaultCallsite` type in `tracing-core` that implements the `Callsite` trait. This type can contain an intrusive list node, and *when a callsite is a `DefaultCallsite`*, we can register it without having to push it to the `Mutex<Vec<...>>`. The locked vec still _exists_, for `Callsite`s that are *not* `DefaultCallsite`s, so we can't remove the `liballoc` dependency, but in most cases, we can avoid the mutex and allocation. Naturally, `tracing` has been updated to use the `DefaultCallsite` type from `tracing-core`, so the `Vec` will only be used in the following cases: * User code has a custom implementation of the `Callsite` trait, which is [not terribly common][1]. * An older version of the `tracing` crate is in use. This fixes the deadlock described in tokio-rs#2020 when `DefaultCallsite`s are used. Additionally, it should reduce the performance impact and memory usage of the callsite registry. Furthermore, I've changed the subscriber registry so that a `RwLock<Vec<dyn Dispatch>>` is only used when there actually are multiple subscribers in use. When there's only a global default subscriber, we can once again avoid locking for the registry of subscribers. When the `std` feature is disabled, thread-local scoped dispatchers are unavailable, so the single global subscriber will always be used on `no_std`. We can make additional changes, such as the ones from tokio-rs#2020, to _also_ resolve potential deadlocks when non-default callsites are in use, but since this branch rewrites a lot of the callsite registry code, that should probably be done in a follow-up. [1]: https://cs.github.com/?scopeName=All+repos&scope=&q=%28%2Fimpl+.*Callsite%2F+AND+language%3Arust%29+NOT+%28path%3A%2Ftracing%2F**+OR+path%3A%2Ftracing-*%2F**+OR+path%3A%2Ftokio-trace*%2F**%29%29 Signed-off-by: Eliza Weisman <eliza@buoyant.io>

@jamesmunns

# 0.1.27 (June 7, 2022) This release of `tracing-core` introduces a new `DefaultCallsite` type, which can be used by instrumentation crates rather than implementing their own callsite types. Using `DefaultCallsite` may offer reduced overhead from callsite registration. ### Added - `DefaultCallsite`, a pre-written `Callsite` implementation for use in instrumentation crates ([tokio-rs#2083]) - `ValueSet::len` and `Record::len` methods returning the number of fields in a `ValueSet` or `Record` ([tokio-rs#2152]) ### Changed - Replaced `lazy_static` dependency with `once_cell` ([tokio-rs#2147]) ### Documented - Added documentation to the `callsite` module ([tokio-rs#2088], [tokio-rs#2149]) Thanks to new contributors @jamesmunns and @james7132 for contributing to this release! [tokio-rs#2083]: tokio-rs#2083 [tokio-rs#2152]: tokio-rs#2152 [tokio-rs#2147]: tokio-rs#2147 [tokio-rs#2088]: tokio-rs#2088 [tokio-rs#2149]: tokio-rs#2149

# 0.1.35 (June 8, 2022) This release reduces the overhead of callsite registration by using new `tracing-core` APIs. ### Added - Use `DefaultCallsite` to reduce callsite registration overhead ([tokio-rs#2083]) ### Changed - `tracing-core`: updated to [0.1.27][core-0.1.27] [core-0.1.27]: https://github.com/tokio-rs/tracing/releases/tag/tracing-core-0.1.27 [tokio-rs#2088]: tokio-rs#2083

hawkw added 8 commits April 18, 2022 11:08

tracing: test reproducing register_callsite deadlock

2209bca

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

crazy idea that might even work...

dff2fa4

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

cleanup

88fbab0

avoid hitting the lock in more cases

e4ee130

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

fix no-std code

969b618

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

update everything to use the new callsites

3cd2f15

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

downcasty trick

94bc0c2

This way, `callsite::register` also does the right thing Signed-off-by: Eliza Weisman <eliza@buoyant.io>

docs

3e6fadd

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

hawkw added kind/feature New feature or request crate/core Related to the `tracing-core` crate kind/perf Performance improvements labels Apr 19, 2022

hawkw requested review from davidbarsky and LucioFranco April 19, 2022 21:24

hawkw requested a review from a team as a code owner April 19, 2022 21:24

hawkw self-assigned this Apr 19, 2022

hawkw added 5 commits April 19, 2022 14:29

Merge branch 'v0.1.x' into eliza/linked-list-registry

251f013

fix lifetimes in log support code

a9893a6

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

fix max_level_hint on nostd

e04f82e

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

fix deadlock on no_std when hitting once cell

b6c70b0

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

cleanup warnings

18cc711

Signed-off-by: Eliza Weisman <eliza@buoyant.io>

hawkw force-pushed the eliza/linked-list-registry branch from 4a30546 to 18cc711 Compare April 20, 2022 17:07

LucioFranco approved these changes Apr 20, 2022

View reviewed changes

Merge branch 'v0.1.x' into eliza/linked-list-registry

f40ae75

davidbarsky approved these changes Apr 22, 2022

View reviewed changes

hawkw enabled auto-merge (squash) April 22, 2022 16:10

hawkw merged commit 45a5df1 into v0.1.x Apr 22, 2022

hawkw deleted the eliza/linked-list-registry branch April 22, 2022 16:11

This was referenced Apr 26, 2022

backport recent commits from master #2096

Merged

tracing, core: fix deadlock in register_callsite #2020

Closed

hawkw mentioned this pull request Jun 7, 2022

core: prepare to release v0.1.27 #2155

Merged

hawkw mentioned this pull request Jun 7, 2022

tracing: prepare to releases v0.1.35 #2156

Merged

bkchr mentioned this pull request Jun 17, 2022

Fix logging with interest cache #2165

Open

Geal mentioned this pull request Jun 27, 2022

update dependencies apollographql/router#1311

Merged

hawkw mentioned this pull request Jun 24, 2023

tracing: fix recursive register_callsite deadlock #2634

Merged

davidbarsky mentioned this pull request Sep 26, 2023

chore: backport roughly a year's worth of changes #2728

Merged

kaffarell mentioned this pull request Jun 25, 2024

fix Name / parent variant of event!() #2983

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

core: add a limited form of the linked-list callsite registry for v0.1.x #2083

core: add a limited form of the linked-list callsite registry for v0.1.x #2083

hawkw commented Apr 19, 2022

hawkw commented Apr 19, 2022

LucioFranco left a comment

LucioFranco Apr 20, 2022

LucioFranco Apr 20, 2022

hawkw Apr 20, 2022

LucioFranco Apr 20, 2022

LucioFranco Apr 20, 2022

hawkw commented Apr 21, 2022

davidbarsky commented Apr 22, 2022

core: add a limited form of the linked-list callsite registry for v0.1.x #2083

core: add a limited form of the linked-list callsite registry for v0.1.x #2083

Conversation

hawkw commented Apr 19, 2022

Motivation

Solution

hawkw commented Apr 19, 2022

LucioFranco left a comment

Choose a reason for hiding this comment

LucioFranco Apr 20, 2022

Choose a reason for hiding this comment

LucioFranco Apr 20, 2022

Choose a reason for hiding this comment

hawkw Apr 20, 2022

Choose a reason for hiding this comment

LucioFranco Apr 20, 2022

Choose a reason for hiding this comment

LucioFranco Apr 20, 2022

Choose a reason for hiding this comment

hawkw commented Apr 21, 2022

davidbarsky commented Apr 22, 2022