feat(mt): move serial sequence turn policy into Rust

Move MT serial turn/skip control and the ordering of LDM generation,
checksum updates, and turn advancement into the Rust rewrite. Keep the C
serial mutexes, LDM window/hash state, checksum state, and codec callbacks
behind a narrow callback bridge so the synchronization and private layouts
remain unchanged. Add focused tests for skipped predecessors, empty LDM
turns, and LDM-before-checksum ordering.

Test Plan:
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo fmt --manifest-path rust/Cargo.toml -- --check
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo test --manifest-path rust/Cargo.toml serial_turn_ -- --nocapture
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo test --manifest-path rust/Cargo.toml
- ulimit -v 41943040; CARGO_BUILD_JOBS=1 cargo clippy --manifest-path rust/Cargo.toml --all-targets -- -D warnings
- ulimit -v 41943040 make -B -C programs -j1 zstd
- ulimit -v 41943040 make -C tests -j1 test-zstream ZSTREAM_TESTTIME=-T1s
This commit is contained in:
2026-07-19 19:17:41 +02:00
parent bff97833b6
commit 9c298999ac
3 changed files with 237 additions and 37 deletions
+69 -34
View File
@@ -354,6 +354,20 @@ typedef char ZSTDMT_rust_raw_seq_store_layout[
ZSTDMT_RustRawSeqStore ZSTDMT_rust_bufferToSeq(ZSTDMT_RustBuffer buffer);
ZSTDMT_RustBuffer ZSTDMT_rust_seqToBuffer(ZSTDMT_RustRawSeqStore seq);
typedef int (*ZSTDMT_serialWaitForTurnFn)(void* opaque, unsigned jobID);
typedef void (*ZSTDMT_serialGenerateLdmFn)(
void* opaque, ZSTDMT_RustRawSeqStore* seqStore,
const void* src, size_t srcSize);
typedef void (*ZSTDMT_serialUpdateChecksumFn)(
void* opaque, const void* src, size_t srcSize);
typedef void (*ZSTDMT_serialAdvanceFn)(void* opaque);
void ZSTDMT_rust_serialStateGenSequences(
ZSTDMT_RustRawSeqStore* seqStore, const void* src, size_t srcSize,
unsigned jobID, int ldmEnabled, int checksumEnabled, void* opaque,
ZSTDMT_serialWaitForTurnFn waitForTurn,
ZSTDMT_serialGenerateLdmFn generateLdm,
ZSTDMT_serialUpdateChecksumFn updateChecksum,
ZSTDMT_serialAdvanceFn advance);
unsigned ZSTDMT_rust_computeTargetJobLog(unsigned windowLog, unsigned chainLog,
int strategy, int enableLdm);
@@ -862,44 +876,57 @@ static void ZSTDMT_serialState_free(SerialState* serialState)
ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
}
static void
ZSTDMT_serialState_genSequences(SerialState* serialState,
RawSeqStore_t* seqStore,
Range src, unsigned jobID)
/* Rust owns the serial turn/skip decision and operation ordering. The wait
* callback intentionally leaves the main serial mutex locked; the advance
* callback releases it after Rust has performed the current turn's work. */
static int ZSTDMT_serialState_waitForTurn(void* opaque, unsigned jobID)
{
/* Wait for our turn */
SerialState* const serialState = (SerialState*)opaque;
ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
while (serialState->nextJobID < jobID) {
DEBUGLOG(5, "wait for serialState->cond");
ZSTD_pthread_cond_wait(&serialState->cond, &serialState->mutex);
}
/* A future job may error and skip our job */
if (serialState->nextJobID == jobID) {
/* It is now our turn, do any processing necessary */
if (serialState->params.ldmParams.enableLdm == ZSTD_ps_enable) {
size_t error;
DEBUGLOG(6, "ZSTDMT_serialState_genSequences: LDM update");
assert(seqStore->seq != NULL && seqStore->pos == 0 &&
seqStore->size == 0 && seqStore->capacity > 0);
assert(src.size <= serialState->params.jobSize);
ZSTD_window_update(&serialState->ldmState.window, src.start, src.size, /* forceNonContiguous */ 0);
error = ZSTD_ldm_generateSequences(
&serialState->ldmState, seqStore,
&serialState->params.ldmParams, src.start, src.size);
/* We provide a large enough buffer to never fail. */
assert(!ZSTD_isError(error)); (void)error;
/* Update ldmWindow to match the ldmState.window and signal the main
* thread if it is waiting for a buffer.
*/
ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
serialState->ldmWindow = serialState->ldmState.window;
ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
}
if (serialState->params.fParams.checksumFlag && src.size > 0)
XXH64_update(&serialState->xxhState, src.start, src.size);
}
/* Now it is the next jobs turn */
return serialState->nextJobID == jobID;
}
static void ZSTDMT_serialState_generateLdm(
void* opaque, ZSTDMT_RustRawSeqStore* seqStore,
const void* src, size_t srcSize)
{
SerialState* const serialState = (SerialState*)opaque;
RawSeqStore_t* const cSeqStore = (RawSeqStore_t*)seqStore;
size_t error;
DEBUGLOG(6, "ZSTDMT_serialState_genSequences: LDM update");
assert(cSeqStore->seq != NULL && cSeqStore->pos == 0 &&
cSeqStore->size == 0 && cSeqStore->capacity > 0);
assert(srcSize <= serialState->params.jobSize);
ZSTD_window_update(&serialState->ldmState.window, src, srcSize,
/* forceNonContiguous */ 0);
error = ZSTD_ldm_generateSequences(
&serialState->ldmState, cSeqStore,
&serialState->params.ldmParams, src, srcSize);
/* We provide a large enough buffer to never fail. */
assert(!ZSTD_isError(error)); (void)error;
/* Update ldmWindow to match the ldmState.window and signal the main
* thread if it is waiting for a buffer. */
ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
serialState->ldmWindow = serialState->ldmState.window;
ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
}
static void ZSTDMT_serialState_updateChecksum(
void* opaque, const void* src, size_t srcSize)
{
SerialState* const serialState = (SerialState*)opaque;
XXH64_update(&serialState->xxhState, src, srcSize);
}
static void ZSTDMT_serialState_advance(void* opaque)
{
SerialState* const serialState = (SerialState*)opaque;
serialState->nextJobID++;
ZSTD_pthread_cond_broadcast(&serialState->cond);
ZSTD_pthread_mutex_unlock(&serialState->mutex);
@@ -1033,8 +1060,16 @@ static void ZSTDMT_compressionJobGenerateSequences(void* opaque)
ZSTDMT_jobDescription* const job = state->job;
/* Perform serial step as early as possible. */
ZSTDMT_serialState_genSequences(job->serial, &state->rawSeqStore,
job->src, job->jobID);
ZSTDMT_rust_serialStateGenSequences(
(ZSTDMT_RustRawSeqStore*)&state->rawSeqStore,
job->src.start, job->src.size, job->jobID,
job->serial->params.ldmParams.enableLdm == ZSTD_ps_enable,
job->serial->params.fParams.checksumFlag,
job->serial,
ZSTDMT_serialState_waitForTurn,
ZSTDMT_serialState_generateLdm,
ZSTDMT_serialState_updateChecksum,
ZSTDMT_serialState_advance);
}
static size_t ZSTDMT_compressionJobBegin(void* opaque)
+6 -3
View File
@@ -97,8 +97,10 @@ zstd ABI:
job-creation decisions, compression-job stage sequencing and error flow,
pending-output decisions through scalar job projections, and frame
progression's job-ring scan, error normalization, and active-worker
accounting. Rust also owns frame-block preparation ordering; C callbacks
retain the private match-state window and workspace operations.
accounting. Rust also owns frame-block preparation ordering and the MT
serial turn/skip policy, including LDM-before-checksum sequencing; C
callbacks retain the private match-state window, workspace operations,
synchronization, LDM state, and checksum state.
- Dictionary support
- `zstd_ddict` owns, loads, copies, and references decode dictionaries.
- Legacy decoding
@@ -149,7 +151,8 @@ shared- and separate-destination decompression scheduling,
single-threaded stream initialization and the buffered/stable stream state
machine, MT stream initialization, MT outer scheduling and flush policy, MT
compression-job stage sequencing and error flow, MT frame-progression job
aggregation, frame-block preparation ordering, public sequence-API
aggregation, frame-block preparation ordering, MT serial turn/skip and
LDM/checksum sequencing, public sequence-API
orchestration, sequence-store and block policy, external-producer invocation
and success-path validation, external-sequence-store reset,
external-sequence/literals block loop, optional-format decompression loops,
+162
View File
@@ -84,6 +84,12 @@ pub type ZSTDMT_compressionJobCompressFn =
pub type ZSTDMT_compressionJobErrorFn = unsafe extern "C" fn(*mut c_void, usize);
pub type ZSTDMT_compressionJobFinishFn = unsafe extern "C" fn(*mut c_void, usize);
pub type ZSTDMT_serialWaitForTurnFn = unsafe extern "C" fn(*mut c_void, c_uint) -> c_int;
pub type ZSTDMT_serialGenerateLdmFn =
unsafe extern "C" fn(*mut c_void, *mut ZstdMtRawSeqStore, *const c_void, usize);
pub type ZSTDMT_serialUpdateChecksumFn = unsafe extern "C" fn(*mut c_void, *const c_void, usize);
pub type ZSTDMT_serialAdvanceFn = unsafe extern "C" fn(*mut c_void);
/// Scalar inputs for the MT streaming initializer. The full parameter
/// object, dictionary handles, pools, buffers, and synchronization remain
/// private to C. Rust owns the order in which the C callbacks are invoked and
@@ -686,6 +692,82 @@ fn compression_job_with<A, P, S, B, Q, H, C, T, E, F>(
finish_job(last_block_size);
}
/// Drive one ordered MT serial-state turn.
///
/// The wait callback returns with the C-owned serial mutex held. Rust then
/// performs the LDM-before-checksum policy only for the current job and always
/// invokes the advance callback exactly once, including when an earlier job
/// has already skipped this turn.
#[inline]
fn serial_state_gen_sequences_with<W, L, C, A>(
seq_store: &mut ZstdMtRawSeqStore,
src: *const c_void,
src_size: usize,
job_id: c_uint,
ldm_enabled: bool,
checksum_enabled: bool,
mut wait_for_turn: W,
mut generate_ldm: L,
mut update_checksum: C,
mut advance: A,
) where
W: FnMut(c_uint) -> bool,
L: FnMut(&mut ZstdMtRawSeqStore, *const c_void, usize),
C: FnMut(*const c_void, usize),
A: FnMut(),
{
if wait_for_turn(job_id) {
if ldm_enabled {
generate_ldm(seq_store, src, src_size);
}
if checksum_enabled && src_size != 0 {
update_checksum(src, src_size);
}
}
advance();
}
/// Own MT serial turn/skip and LDM/checksum ordering while C retains the
/// mutexes, LDM window/hash state, checksum state, and raw sequence storage.
#[no_mangle]
pub unsafe extern "C" fn ZSTDMT_rust_serialStateGenSequences(
seq_store: *mut ZstdMtRawSeqStore,
src: *const c_void,
src_size: usize,
job_id: c_uint,
ldm_enabled: c_int,
checksum_enabled: c_int,
opaque: *mut c_void,
wait_for_turn: Option<ZSTDMT_serialWaitForTurnFn>,
generate_ldm: Option<ZSTDMT_serialGenerateLdmFn>,
update_checksum: Option<ZSTDMT_serialUpdateChecksumFn>,
advance: Option<ZSTDMT_serialAdvanceFn>,
) {
let (Some(wait_for_turn), Some(generate_ldm), Some(update_checksum), Some(advance)) =
(wait_for_turn, generate_ldm, update_checksum, advance)
else {
return;
};
if seq_store.is_null() {
return;
}
unsafe {
serial_state_gen_sequences_with(
&mut *seq_store,
src,
src_size,
job_id,
ldm_enabled != 0,
checksum_enabled != 0,
|job_id| wait_for_turn(opaque, job_id) != 0,
|seq_store, src, src_size| generate_ldm(opaque, seq_store, src, src_size),
|src, src_size| update_checksum(opaque, src, src_size),
|| advance(opaque),
);
}
}
/// C ABI entry point for the worker-job orchestration. C supplies callbacks
/// that keep the private descriptor, pools, mutexes, and codec operations on
/// the C side of this narrow projection.
@@ -3156,6 +3238,86 @@ mod tests {
assert_eq!(state.finished, vec![7]);
}
#[test]
fn serial_turn_runs_ldm_before_checksum_and_advances_once() {
let events = Rc::new(RefCell::new(Vec::new()));
let wait_events = Rc::clone(&events);
let ldm_events = Rc::clone(&events);
let checksum_events = Rc::clone(&events);
let advance_events = Rc::clone(&events);
let mut seq_store = ZstdMtRawSeqStore::default();
serial_state_gen_sequences_with(
&mut seq_store,
ptr::null(),
8,
4,
true,
true,
move |job_id| {
assert_eq!(job_id, 4);
wait_events.borrow_mut().push("wait");
true
},
move |_seq_store, _src, src_size| {
assert_eq!(src_size, 8);
ldm_events.borrow_mut().push("ldm");
},
move |_src, src_size| {
assert_eq!(src_size, 8);
checksum_events.borrow_mut().push("checksum");
},
move || advance_events.borrow_mut().push("advance"),
);
assert_eq!(&*events.borrow(), &["wait", "ldm", "checksum", "advance"]);
}
#[test]
fn serial_turn_skips_failed_predecessor_and_keeps_empty_ldm_turn() {
let events = Rc::new(RefCell::new(Vec::new()));
let wait_events = Rc::clone(&events);
let advance_events = Rc::clone(&events);
let mut seq_store = ZstdMtRawSeqStore::default();
serial_state_gen_sequences_with(
&mut seq_store,
ptr::null(),
0,
9,
true,
true,
move |_job_id| {
wait_events.borrow_mut().push("wait");
false
},
|_seq_store, _src, _src_size| panic!("skipped jobs must not generate LDM sequences"),
|_src, _src_size| panic!("skipped jobs must not update the checksum"),
move || advance_events.borrow_mut().push("advance"),
);
assert_eq!(&*events.borrow(), &["wait", "advance"]);
let empty_events = Rc::new(RefCell::new(Vec::new()));
let empty_ldm_events = Rc::clone(&empty_events);
let empty_advance_events = Rc::clone(&empty_events);
serial_state_gen_sequences_with(
&mut seq_store,
ptr::null(),
0,
10,
true,
true,
|_job_id| true,
move |_seq_store, _src, src_size| {
assert_eq!(src_size, 0);
empty_ldm_events.borrow_mut().push("ldm");
},
|_src, _src_size| panic!("empty input must not update the checksum"),
move || empty_advance_events.borrow_mut().push("advance"),
);
assert_eq!(&*empty_events.borrow(), &["ldm", "advance"]);
}
#[test]
fn compression_job_stops_on_non_first_chunk_error_and_cleans_up() {
let state = Rc::new(RefCell::new(MockCompressionJob::default()));