summaryrefslogtreecommitdiff
path: root/crates/tor-memquota/src
diff options
context:
space:
mode:
Diffstat (limited to 'crates/tor-memquota/src')
-rw-r--r--crates/tor-memquota/src/config.rs324
-rw-r--r--crates/tor-memquota/src/internal_prelude.rs2
-rw-r--r--crates/tor-memquota/src/mtracker.rs2
3 files changed, 303 insertions, 25 deletions
diff --git a/crates/tor-memquota/src/config.rs b/crates/tor-memquota/src/config.rs
index 763316b64..cd71a57b1 100644
--- a/crates/tor-memquota/src/config.rs
+++ b/crates/tor-memquota/src/config.rs
@@ -1,5 +1,10 @@
//! Configuration (private module)
+use std::sync::LazyLock;
+
+use sysinfo::{MemoryRefreshKind, System};
+use tracing::warn;
+
use crate::internal_prelude::*;
/// We want to support at least this many participants with a cache each
@@ -21,13 +26,16 @@ define_derive_deftly! {
impl ConfigBuilder {
$(
- ${when approx_equal($ftype, { Option::<Qty> })}
+ ${when approx_equal($ftype, { Option::<ExplicitOrAuto<Qty>> })}
${fattrs doc}
///
/// (Setter method.)
- pub fn $fname(&mut self, value: usize) -> &mut Self {
- self.$fname = Some(Qty(value));
+ // We use `value: impl Into<ExplicitOrAuto<usize>>` to avoid breaking users who used the
+ // previous `value: usize`. But this isn't 100% foolproof, for example if a user used
+ // `$fname(foo.into())`, which will fail type inference.
+ pub fn $fname(&mut self, value: impl Into<ExplicitOrAuto<usize>>) -> &mut Self {
+ self.$fname = Some(value.into().map(Qty));
self
}
)
@@ -54,17 +62,26 @@ pub struct Config(pub(crate) IfEnabled<ConfigInner>);
pub struct ConfigBuilder {
/// Maximum memory usage tolerated before reclamation starts
///
- /// Setting this to `usize::MAX` disables the memory quota
- /// (and that's the default).
+ /// Setting this to `usize::MAX` disables the memory quota.
+ ///
+ /// The default is "auto",
+ /// which uses a value derived from the total system memory.
+ /// It should not be assumed that the value used for "auto"
+ /// will remain stable across different versions of this library.
///
/// Note that this is not a hard limit.
/// See Approximate in [the overview](crate).
- max: Option<Qty>,
+ max: Option<ExplicitOrAuto<Qty>>,
/// Reclamation will stop when memory use is reduced to below this value
///
- /// Default is 75% of the maximum.
- low_water: Option<Qty>,
+ /// Default is "auto", which uses 75% of the maximum.
+ /// It should not be assumed that the value used for "auto"
+ /// will remain stable across different versions of this library.
+ ///
+ /// If set to an explicit value,
+ /// then `max` must be set to an explicit value as well.
+ low_water: Option<ExplicitOrAuto<Qty>>,
}
/// Configuration, if enabled
@@ -77,7 +94,7 @@ pub struct ConfigBuilder {
pub(crate) struct ConfigInner {
/// Maximum memory usage
///
- /// Guaranteed not to be `MAX`, since we're anbled
+ /// Guaranteed not to be `MAX`, since we're enabled
pub max: Qty,
/// Low water
@@ -98,8 +115,9 @@ impl Config {
///
/// Ad-hoc accessor for testing purposes.
/// (ideally we'd use `visibility` to make fields `pub`, but that doesn't work.)
- #[cfg(feature = "testing")]
- pub fn inner(&self) -> Option<&ConfigInner> {
+ #[cfg(any(test, feature = "testing"))]
+ #[cfg_attr(feature = "testing", visibility::make(pub))]
+ fn inner(&self) -> Option<&ConfigInner> {
self.0.as_ref().into_enabled()
}
}
@@ -107,21 +125,38 @@ impl Config {
impl ConfigBuilder {
/// Builds a new `Config` from a builder
///
- /// Returns an error unless at least `max` has been specified,
- /// or if the fields values are invalid or inconsistent.
+ /// Returns an error if the fields values are invalid or inconsistent.
pub fn build(&self) -> Result<Config, ConfigBuildError> {
- let max = self.max.unwrap_or(Qty::MAX);
+ // both options default to "auto"
+ let max = self.max.unwrap_or(ExplicitOrAuto::Auto);
+ let low_water = self.low_water.unwrap_or(ExplicitOrAuto::Auto);
- if max == Qty::MAX {
- if self.low_water.is_some() {
+ // `MAX` indicates "disabled".
+ // TODO: Should we add a new "enabled" config option instead of using a sentinel value?
+ // But this would be a breaking change. Or maybe we should always enable the memquota
+ // machinery even if the user chooses an unreasonably large value, and not give users a way
+ // to disable it.
+ if max == ExplicitOrAuto::Explicit(Qty::MAX) {
+ // If it should be disabled, but the user provided an explicit value for `low_water`.
+ if matches!(low_water, ExplicitOrAuto::Explicit(_)) {
return Err(ConfigBuildError::Inconsistent {
fields: vec!["max".into(), "low_water".into()],
- problem: "low_water supplied, but max omitted".into(),
+ problem: "low_water supplied, but max indicates that we should disable the memory quota".into(),
});
};
return Ok(Config(IfEnabled::Noop));
}
+ // We don't want the user to set "auto" for `max`, but an explicit value for `low_water`.
+ // Otherwise this config is prone to breaking since a `max` of "auto" may change as system
+ // memory is removed (either physically or if running in a VM/container).
+ if matches!(max, ExplicitOrAuto::Auto) && matches!(low_water, ExplicitOrAuto::Explicit(_)) {
+ return Err(ConfigBuildError::Inconsistent {
+ fields: vec!["max".into(), "low_water".into()],
+ problem: "max is \"auto\", but low_water is set to an explicit quantity".into(),
+ });
+ }
+
let enabled = EnabledToken::new_if_compiled_in()
//
.ok_or_else(|| ConfigBuildError::NoCompileTimeSupport {
@@ -129,10 +164,18 @@ impl ConfigBuilder {
problem: "cargo feature `memquota` disabled (in tor-memquota crate)".into(),
})?;
- let low_water = self.low_water.unwrap_or_else(
- //
- || Qty((*max as f32 * 0.75) as _),
- );
+ // The general logic is taken from c-tor (see `compute_real_max_mem_in_queues`).
+ // NOTE: Relays have an additional lower bound for explicitly given values (64 MiB),
+ // but we have no way of knowing whether we are a relay or not here.
+ let max = match max {
+ ExplicitOrAuto::Explicit(x) => x,
+ ExplicitOrAuto::Auto => compute_max_from_total_system_mem(total_available_memory()),
+ };
+
+ let low_water = match low_water {
+ ExplicitOrAuto::Explicit(x) => x,
+ ExplicitOrAuto::Auto => Qty((*max as f32 * 0.75) as _),
+ };
let config = ConfigInner { max, low_water };
@@ -160,6 +203,136 @@ impl ConfigBuilder {
}
}
+/// Determine a max given the system's total available memory.
+///
+/// This is used when `max` is configured as "auto".
+/// It takes a `Result` so that we can handle the case where the total memory isn't available.
+fn compute_max_from_total_system_mem(mem: Result<usize, MemQueryError>) -> Qty {
+ const MIB: usize = 1024 * 1024;
+ const GIB: usize = 1024 * 1024 * 1024;
+
+ let mem = match mem {
+ Ok(x) => x,
+ Err(e) => {
+ warn!("Unable to get the total available memory. Using a constant max instead: {e}");
+
+ // Can't get the total available memory,
+ // so we return a max depending on whether the architecture is 32-bit or 64-bit.
+ return Qty({
+ cfg_if::cfg_if! {
+ if #[cfg(target_pointer_width = "64")] {
+ 8 * GIB
+ } else {
+ 1 * GIB
+ }
+ }
+ });
+ }
+ };
+
+ let mem = if mem >= 8 * GIB {
+ // From c-tor:
+ //
+ // > The idea behind this value is that the amount of RAM is more than enough
+ // > for a single relay and should allow the relay operator to run two relays
+ // > if they have additional bandwidth available.
+ (mem as f64 * 0.40) as usize
+ } else {
+ (mem as f64 * 0.75) as usize
+ };
+
+ // The (min, max) range to clamp `mem` to.
+ let clamp = {
+ cfg_if::cfg_if! {
+ if #[cfg(target_pointer_width = "64")] {
+ (256 * MIB, 8 * GIB)
+ } else {
+ (256 * MIB, 2 * GIB)
+ }
+ }
+ };
+
+ let mem = mem.clamp(clamp.0, clamp.1);
+
+ Qty(mem)
+}
+
+/// The total available memory in bytes.
+///
+/// This is generally the amount of system RAM,
+/// but we may also take into account other OS-specific limits such as cgroups.
+///
+/// Returns `None` if we were unable to get the total available memory.
+/// But see internal comments for details.
+fn total_available_memory() -> Result<usize, MemQueryError> {
+ // The sysinfo crate says we should use only one `System` per application.
+ // But we're a library, so it's probably best to just make this global and reuse it.
+ // In reality getting the system memory probably shouldn't require persistent state,
+ // but since the internals of the sysinfo crate are opaque to us,
+ // we'll just follow their documentation and cache the `System`.
+ //
+ // NOTE: The sysinfo crate in practice gets more information than we ask for.
+ // For example `System::new()` will always query the `_SC_PAGESIZE` and `_SC_CLK_TCK`
+ // on Linux even though we only refresh the memory info below
+ // (see https://github.com/GuillaumeGomez/sysinfo/blob/fc31b411eea7b9983176399dc5be162786dec95b/src/unix/linux/system.rs#L152).
+ // This means that miri will fail to run on tests that build the config, even if the config uses
+ // explicit values.
+ static SYSTEM: LazyLock<Mutex<System>> = LazyLock::new(|| Mutex::new(System::new()));
+ let mut system = SYSTEM.lock().unwrap_or_else(|mut e| {
+ // The sysinfo crate has some internal panics which would poison this mutex.
+ // But we can easily reset it, rather than panicking ourselves if it's poisoned.
+ **e.get_mut() = System::new();
+ SYSTEM.clear_poison();
+ e.into_inner()
+ });
+
+ system.refresh_memory_specifics(MemoryRefreshKind::nothing().with_ram());
+
+ // It might be possible for 32-bit systems to return >usize::MAX due to PAE (I haven't looked
+ // into this), so we just saturate the value and don't consider this an error.
+ let mem = to_usize_saturating(system.total_memory());
+
+ // The sysinfo crate doesn't report errors, so the best we can do is guess that a value of 0
+ // implies that it was unable to get the total memory.
+ //
+ // We also need to return early to prevent a panic below.
+ if mem == 0 {
+ return Err(MemQueryError::Unavailable);
+ }
+
+ // Note: The docs for the sysinfo crate say:
+ //
+ // > You need to have run refresh_memory at least once before calling this method.
+ //
+ // But as implemented, it also panics if `sys.mem_total == 0` (for example if the refresh
+ // silently failed).
+ let Some(cgroups) = system.cgroup_limits() else {
+ // There is no cgroup (or we're a non-Linux platform).
+ return Ok(mem);
+ };
+
+ // The `cgroup_limits()` surprisingly doesn't actually return the unaltered cgroups limits.
+ // It also adjusts them depending on the total memory.
+ // Since this is all undocumented, we'll also do the same calculation here.
+ let mem = std::cmp::min(mem, to_usize_saturating(cgroups.total_memory));
+
+ Ok(mem)
+}
+
+/// An error when we are unable to obtain the system's total available memory.
+#[derive(Clone, Debug, thiserror::Error)]
+enum MemQueryError {
+ /// The total available memory is unavailable.
+ #[error("total available memory is unavailable")]
+ Unavailable,
+}
+
+/// Convert a `u64` to a `usize`, saturating if the value would overflow.
+fn to_usize_saturating(x: u64) -> usize {
+ // this will be optimized to a no-op on 64-bit systems
+ x.try_into().unwrap_or(usize::MAX)
+}
+
#[cfg(test)]
mod test {
// @@ begin test lint list maintained by maint/add_warning @@
@@ -180,6 +353,9 @@ mod test {
use serde_json::json;
#[test]
+ // A value of "auto" depends on the system memory,
+ // which typically results in libc calls or syscall that aren't supported by miri.
+ #[cfg_attr(miri, ignore)]
fn configs() {
let chk_ok_raw = |j, c| {
let b: ConfigBuilder = serde_json::from_value(j).unwrap();
@@ -215,17 +391,119 @@ mod test {
chk_err(j, "UNSUPPORTED");
};
+ let chk_builds = |j| {
+ cfg_if::cfg_if! {
+ if #[cfg(feature = "memquota")] {
+ let b: ConfigBuilder = serde_json::from_value(j).unwrap();
+ b.build().unwrap();
+ } else {
+ chk_err(j, "UNSUPPORTED");
+ }
+ }
+ };
+
chk_ok(json! {{ "max": "8 MiB" }}, 8, 6);
+ chk_ok(json! {{ "max": "8 MiB", "low_water": "auto" }}, 8, 6);
chk_ok(json! {{ "max": "8 MiB", "low_water": "4 MiB" }}, 8, 4);
- chk_ok_raw(json! {{ }}, Config(IfEnabled::Noop));
+
+ // We don't know what the exact values will be since they are derived from the system
+ // memory.
+ chk_builds(json! {{ }});
+ chk_builds(json! {{ "max": "auto" }});
+ chk_builds(json! {{ "low_water": "auto" }});
+ chk_builds(json! {{ "max": "auto", "low_water": "auto" }});
chk_err(
json! {{ "low_water": "4 MiB" }},
- "low_water supplied, but max omitted",
+ "max is \"auto\", but low_water is set to an explicit quantity",
);
chk_err(
json! {{ "max": "8 MiB", "low_water": "8 MiB" }},
"inconsistent: low_water / max",
);
+
+ // `usize::MAX` is a special value.
+ chk_err(
+ json! {{ "max": usize::MAX.to_string(), "low_water": "8 MiB" }},
+ "low_water supplied, but max indicates that we should disable the memory quota",
+ );
+ chk_builds(json! {{ "max": (usize::MAX - 1).to_string(), "low_water": "8 MiB" }});
+
+ // check that the builder works as expected
+ #[cfg(feature = "memquota")]
+ {
+ let mut b = Config::builder();
+ b.max(ExplicitOrAuto::Explicit(100_000_000));
+ if let Some(inner) = b.build().unwrap().inner() {
+ assert_eq!(inner.max, Qty(100_000_000));
+ }
+
+ let mut b = Config::builder();
+ b.max(100_000_000);
+ if let Some(inner) = b.build().unwrap().inner() {
+ assert_eq!(inner.max, Qty(100_000_000));
+ }
+
+ let mut b = ConfigBuilder::default();
+ b.max(ExplicitOrAuto::Auto);
+ b.build().unwrap();
+ }
+ }
+
+ /// Test the logic that computes the `max` when configured as "auto".
+ #[test]
+ // We do some `1 * X` operations below for readability.
+ #[allow(clippy::identity_op)]
+ fn auto_max() {
+ #[allow(unused)]
+ fn check_helper(val: Qty, expected_32: Qty, expected_64: Qty) {
+ assert_eq!(val, {
+ cfg_if::cfg_if! {
+ if #[cfg(target_pointer_width = "64")] {
+ expected_64
+ } else if #[cfg(target_pointer_width = "32")] {
+ expected_32
+ } else {
+ panic!("Unsupported architecture :(");
+ }
+ }
+ });
+ }
+
+ check_helper(
+ compute_max_from_total_system_mem(Err(MemQueryError::Unavailable)),
+ /* 32-bit */ Qty(1 * 1024 * 1024 * 1024),
+ /* 64-bit */ Qty(8 * 1024 * 1024 * 1024),
+ );
+ check_helper(
+ compute_max_from_total_system_mem(Ok(8 * 1024 * 1024 * 1024)),
+ /* 32-bit */ Qty(2 * 1024 * 1024 * 1024),
+ /* 64-bit */ Qty(3435973836),
+ );
+ check_helper(
+ compute_max_from_total_system_mem(Ok(7 * 1024 * 1024 * 1024)),
+ /* 32-bit */ Qty(2 * 1024 * 1024 * 1024),
+ /* 64-bit */ Qty(5637144576),
+ );
+ check_helper(
+ compute_max_from_total_system_mem(Ok(1 * 1024 * 1024 * 1024)),
+ /* 32-bit */ Qty(805306368),
+ /* 64-bit */ Qty(805306368),
+ );
+ check_helper(
+ compute_max_from_total_system_mem(Ok(7 * 1024)),
+ /* 32-bit */ Qty(256 * 1024 * 1024),
+ /* 64-bit */ Qty(256 * 1024 * 1024),
+ );
+ check_helper(
+ compute_max_from_total_system_mem(Ok(0)),
+ /* 32-bit */ Qty(256 * 1024 * 1024),
+ /* 64-bit */ Qty(256 * 1024 * 1024),
+ );
+ check_helper(
+ compute_max_from_total_system_mem(Ok(usize::MAX)),
+ /* 32-bit */ Qty(2 * 1024 * 1024 * 1024),
+ /* 64-bit */ Qty(8 * 1024 * 1024 * 1024),
+ );
}
}
diff --git a/crates/tor-memquota/src/internal_prelude.rs b/crates/tor-memquota/src/internal_prelude.rs
index 5bddc6466..c4c358885 100644
--- a/crates/tor-memquota/src/internal_prelude.rs
+++ b/crates/tor-memquota/src/internal_prelude.rs
@@ -55,7 +55,7 @@ pub(crate) use {
tor_async_utils::mpsc_channel_no_memquota,
tor_async_utils::stream_peek::StreamUnobtrusivePeeker,
tor_basic_utils::ByteQty as Qty,
- tor_config::{ConfigBuildError, ReconfigureError},
+ tor_config::{ConfigBuildError, ExplicitOrAuto, ReconfigureError},
tor_error::{error_report, internal, into_internal, trace_report, Bug, ErrorKind, HasKind},
tor_log_ratelim::log_ratelim,
tor_rtcompat::{CoarseInstant, CoarseTimeProvider, DynTimeProvider},
diff --git a/crates/tor-memquota/src/mtracker.rs b/crates/tor-memquota/src/mtracker.rs
index b3569376e..e2a8f053c 100644
--- a/crates/tor-memquota/src/mtracker.rs
+++ b/crates/tor-memquota/src/mtracker.rs
@@ -568,7 +568,7 @@ impl MemoryQuotaTracker {
let for_task = Arc::downgrade(&tracker);
runtime.spawn(reclaim::task(for_task, reclaim_rx, enabled))?;
- info!(%max, %low_water, "memory quota tracking initialised");
+ info!(%max, %low_water, "Memory quota tracking initialised");
Ok(tracker)
}