/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


/* ======   Compiler specifics   ====== */
#if defined(_MSC_VER)
#  pragma warning(disable : 4204)   /* disable: C4204: non-constant aggregate initializer */
#endif


/* ======   Dependencies   ====== */
#include "../common/allocations.h" /* ZSTD_customMalloc, ZSTD_customCalloc, ZSTD_customFree */
#include "../common/zstd_deps.h"   /* ZSTD_memcpy, ZSTD_memset, INT_MAX, UINT_MAX */
#include "../common/mem.h"         /* MEM_STATIC */
#include "../common/pool.h"        /* threadpool */
#include "../common/threading.h"   /* mutex */
#include "zstd_compress_internal.h" /* MIN, ERROR, ZSTD_*, ZSTD_highbit32 */
#include "zstd_ldm.h"
#include "zstdmt_compress.h"

/* Guards code to support resizing the SeqPool.
 * We will want to resize the SeqPool to save memory in the future.
 * Until then, comment the code out since it is unused.
 */
#define ZSTD_RESIZE_SEQPOOL 0

/* ======   Debug   ====== */
#if defined(DEBUGLEVEL) && (DEBUGLEVEL>=2) \
    && !defined(_MSC_VER) \
    && !defined(__MINGW32__)

#  include <stdio.h>
#  include <unistd.h>
#  include <sys/times.h>

#  define DEBUG_PRINTHEX(l,p,n)                                       \
    do {                                                              \
        unsigned debug_u;                                             \
        for (debug_u=0; debug_u<(n); debug_u++)                       \
            RAWLOG(l, "%02X ", ((const unsigned char*)(p))[debug_u]); \
        RAWLOG(l, " \n");                                             \
    } while (0)

static unsigned long long GetCurrentClockTimeMicroseconds(void)
{
   static clock_t _ticksPerSecond = 0;
   if (_ticksPerSecond <= 0) _ticksPerSecond = sysconf(_SC_CLK_TCK);

   {   struct tms junk; clock_t newTicks = (clock_t) times(&junk);
       return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond);
}  }

#define MUTEX_WAIT_TIME_DLEVEL 6
#define ZSTD_PTHREAD_MUTEX_LOCK(mutex)                                                  \
    do {                                                                                \
        if (DEBUGLEVEL >= MUTEX_WAIT_TIME_DLEVEL) {                                     \
            unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds();    \
            ZSTD_pthread_mutex_lock(mutex);                                             \
            {   unsigned long long const afterTime = GetCurrentClockTimeMicroseconds(); \
                unsigned long long const elapsedTime = (afterTime-beforeTime);          \
                if (elapsedTime > 1000) {                                               \
                    /* or whatever threshold you like; I'm using 1 millisecond here */  \
                    DEBUGLOG(MUTEX_WAIT_TIME_DLEVEL,                                    \
                        "Thread took %llu microseconds to acquire mutex %s \n",         \
                        elapsedTime, #mutex);                                           \
            }   }                                                                       \
        } else {                                                                        \
            ZSTD_pthread_mutex_lock(mutex);                                             \
        }                                                                               \
    } while (0)

#else

#  define ZSTD_PTHREAD_MUTEX_LOCK(m) ZSTD_pthread_mutex_lock(m)
#  define DEBUG_PRINTHEX(l,p,n) do { } while (0)

#endif


/* =====   Buffer Pool   ===== */
/* a single Buffer Pool can be invoked from multiple threads in parallel */

typedef struct buffer_s {
    void* start;
    size_t capacity;
} Buffer;

static const Buffer g_nullBuffer = { NULL, 0 };

/* The Rust module owns the synchronized pool state.  Keep a C wrapper so
 * custom allocation of the containing ZSTDMT context remains unchanged. */
typedef struct ZSTDMT_RustBufferPool_s ZSTDMT_RustBufferPool;
typedef struct {
    void* start;
    size_t capacity;
} ZSTDMT_RustBuffer;

ZSTDMT_RustBufferPool* ZSTDMT_rust_buffer_pool_create(unsigned maxNbBuffers,
                                                       ZSTD_customMem cMem);
void ZSTDMT_rust_buffer_pool_free(ZSTDMT_RustBufferPool* pool);
size_t ZSTDMT_rust_buffer_pool_sizeof(const ZSTDMT_RustBufferPool* pool);
size_t ZSTDMT_rust_sizeofBufferPool(size_t wrapperSize, size_t rustPoolSize);
void ZSTDMT_rust_buffer_pool_set_size(ZSTDMT_RustBufferPool* pool, size_t bSize);
ZSTDMT_RustBufferPool* ZSTDMT_rust_buffer_pool_expand(ZSTDMT_RustBufferPool* pool,
                                                       unsigned maxNbBuffers);
ZSTDMT_RustBuffer ZSTDMT_rust_buffer_pool_get(ZSTDMT_RustBufferPool* pool);
void ZSTDMT_rust_buffer_pool_release(ZSTDMT_RustBufferPool* pool,
                                     ZSTDMT_RustBuffer buffer);
ZSTDMT_RustBuffer ZSTDMT_rust_buffer_pool_resize(ZSTDMT_RustBufferPool* pool,
                                                  ZSTDMT_RustBuffer buffer);

typedef struct {
    size_t error;
    size_t lastBlockSize;
} ZSTDMT_chunkProcessResult;

typedef struct {
    unsigned firstJob;
    unsigned lastJob;
} ZSTDMT_RustCompressionJobProjection;

typedef struct {
    int status;
    size_t toFlush;
    size_t outputPos;
    size_t dstFlushed;
} ZSTDMT_flushPublicationResult;

typedef void (*ZSTDMT_chunkProgressFn)(void* opaque, size_t cSize, size_t consumed);
typedef size_t (*ZSTDMT_compressionJobStepFn)(void* opaque);
typedef void (*ZSTDMT_compressionJobVoidFn)(void* opaque);
typedef ZSTDMT_chunkProcessResult (*ZSTDMT_compressionJobCompressFn)(
        void* opaque, unsigned lastJob);
typedef void (*ZSTDMT_compressionJobErrorFn)(void* opaque, size_t error);
typedef void (*ZSTDMT_compressionJobFinishFn)(void* opaque, size_t lastBlockSize);

void ZSTDMT_rust_compressionJob(
        const ZSTDMT_RustCompressionJobProjection* projection,
        void* opaque,
        ZSTDMT_compressionJobStepFn acquireResources,
        ZSTDMT_compressionJobVoidFn prepareParameters,
        ZSTDMT_compressionJobVoidFn generateSequences,
        ZSTDMT_compressionJobStepFn beginJob,
        ZSTDMT_compressionJobVoidFn applySequences,
        ZSTDMT_compressionJobStepFn writeFrameHeader,
        ZSTDMT_compressionJobCompressFn compressJob,
        ZSTDMT_compressionJobVoidFn traceJob,
        ZSTDMT_compressionJobErrorFn setError,
        ZSTDMT_compressionJobFinishFn finishJob);

ZSTDMT_chunkProcessResult ZSTDMT_rust_compressJobChunks(
        ZSTD_CCtx* cctx, const void* src, size_t srcSize,
        void* dst, size_t dstCapacity, size_t chunkSize, unsigned lastJob,
        void* progressContext, ZSTDMT_chunkProgressFn progressCallback);

ZSTDMT_flushPublicationResult ZSTDMT_rust_publishJobOutput(
        void* outputDst, size_t outputSize, size_t outputPos,
        const void* jobDst, size_t jobCapacity,
        size_t cSize, size_t dstFlushed);

typedef struct {
    size_t consumed;
    size_t cSize;
    size_t srcSize;
    const void* dstStart;
    size_t dstCapacity;
    size_t dstFlushed;
    unsigned frameChecksumNeeded;
} ZSTDMT_RustFlushJobProjection;
typedef struct {
    unsigned doneJobID;
    unsigned nextJobID;
    unsigned jobIDMask;
    unsigned jobReady;
    unsigned frameEnded;
    size_t inBuffFilled;
} ZSTDMT_RustFlushContextProjection;
typedef struct {
    size_t result;
    size_t outputPos;
    unsigned updateAllJobsCompleted;
    unsigned allJobsCompleted;
} ZSTDMT_RustFlushProducedResult;
typedef void (*ZSTDMT_flushProjectJobFn)(
        void* opaque, unsigned jobID, unsigned blockToFlush,
        ZSTDMT_RustFlushJobProjection* projection);
typedef void (*ZSTDMT_flushChecksumFn)(
        void* opaque, unsigned jobID,
        ZSTDMT_RustFlushJobProjection* projection);
typedef void (*ZSTDMT_flushUpdateJobFn)(
        void* opaque, unsigned jobID, size_t dstFlushed);
typedef void (*ZSTDMT_flushCompleteJobFn)(
        void* opaque, unsigned jobID, size_t srcSize, size_t cSize);
typedef void (*ZSTDMT_flushErrorFn)(void* opaque);
ZSTDMT_RustFlushProducedResult ZSTDMT_rust_flushProduced(
        const ZSTDMT_RustFlushContextProjection* context,
        void* outputDst, size_t outputSize, size_t outputPos,
        unsigned blockToFlush, unsigned end, void* opaque,
        ZSTDMT_flushProjectJobFn projectJob,
        ZSTDMT_flushChecksumFn addFrameChecksum,
        ZSTDMT_flushUpdateJobFn updateJob,
        ZSTDMT_flushCompleteJobFn completeJob,
        ZSTDMT_flushErrorFn onError);

/* The Rust outer scheduler sees only this scalar snapshot.  The MT context,
 * reusable input buffer, worker pool, and all synchronization remain private
 * to this translation unit. */
typedef struct {
    unsigned frameEnded;
    unsigned jobReady;
    void* inBuffStart;
    size_t inBuffCapacity;
    size_t inBuffFilled;
    size_t targetSectionSize;
    int rsyncable;
    U64 rsyncPrimePower;
    U64 rsyncHitMask;
} ZSTDMT_RustCompressStreamContextProjection;
typedef struct {
    void* bufferStart;
    size_t bufferCapacity;
    size_t bufferFilled;
} ZSTDMT_RustStreamInputRangeProjection;
typedef struct {
    const void* src;
    size_t size;
    size_t pos;
} ZSTDMT_RustStreamInputProjection;
typedef struct {
    void* dst;
    size_t size;
    size_t pos;
} ZSTDMT_RustStreamOutputProjection;
typedef struct {
    size_t toLoad;
    int flush;
} ZSTDMT_RustSyncPointProjection;
typedef struct {
    size_t result;
    size_t outputPos;
} ZSTDMT_RustStreamFlushResult;
typedef struct {
    size_t result;
    size_t inputPos;
    size_t outputPos;
} ZSTDMT_RustCompressStreamResult;
typedef int (*ZSTDMT_streamTryGetInputRangeFn)(
        void* opaque, ZSTDMT_RustStreamInputRangeProjection* projection);
typedef int (*ZSTDMT_streamLoadInputFn)(void* opaque, const void* src, size_t size);
typedef size_t (*ZSTDMT_streamCreateJobFn)(void* opaque, size_t srcSize, unsigned end);
typedef ZSTDMT_RustStreamFlushResult (*ZSTDMT_streamFlushProducedFn)(
        void* opaque, void* outputDst, size_t outputSize, size_t outputPos,
        unsigned blockToFlush, unsigned end);
ZSTDMT_RustCompressStreamResult ZSTDMT_rust_compressStreamGeneric(
        const ZSTDMT_RustCompressStreamContextProjection* context,
        const ZSTDMT_RustStreamInputProjection* input,
        const ZSTDMT_RustStreamOutputProjection* output,
        unsigned end, void* opaque,
        ZSTDMT_streamTryGetInputRangeFn tryGetInputRange,
        ZSTDMT_streamLoadInputFn loadInput,
        ZSTDMT_streamCreateJobFn createJob,
        ZSTDMT_streamFlushProducedFn flushProduced);

typedef struct {
    unsigned lastJob;
    size_t srcSize;
    unsigned firstJob;
    void* dstStart;
    size_t dstCapacity;
    size_t consumed;
} ZSTDMT_RustEmptyBlockJobProjection;
typedef struct {
    ZSTDMT_RustBuffer buffer;
    size_t cSize;
    unsigned clearSource;
} ZSTDMT_RustEmptyBlockResult;
typedef ZSTDMT_RustBuffer (*ZSTDMT_bufferGetFn)(void* opaque);
ZSTDMT_RustEmptyBlockResult ZSTDMT_rust_writeLastEmptyBlock(
        const ZSTDMT_RustEmptyBlockJobProjection* projection,
        void* opaque, ZSTDMT_bufferGetFn getBuffer);

typedef struct {
    unsigned doneJobID;
    unsigned nextJobID;
    unsigned jobIDMask;
    unsigned jobReady;
    const void* srcStart;
    size_t srcSize;
    size_t inBuffFilled;
    const void* prefixStart;
    size_t prefixSize;
    size_t targetPrefixSize;
    unsigned endFrame;
    unsigned checksumFlag;
} ZSTDMT_RustCreateJobProjection;
typedef struct {
    const void* srcStart;
    size_t srcSize;
    const void* prefixStart;
    size_t prefixSize;
    const void* nextPrefixStart;
    size_t nextPrefixSize;
    size_t roundBuffPosDelta;
    unsigned jobNumber;
    unsigned firstJob;
    unsigned lastJob;
    unsigned frameChecksumNeeded;
    unsigned clearChecksumFlag;
} ZSTDMT_RustJobInitialization;
typedef struct {
    size_t returnCode;
    unsigned action;
    unsigned jobID;
    unsigned jobNumber;
    unsigned nextJobID;
    unsigned jobReady;
} ZSTDMT_RustCreateJobResult;
enum {
    ZSTDMT_CREATE_JOB_TABLE_FULL = 0,
    ZSTDMT_CREATE_JOB_POST = 1,
    ZSTDMT_CREATE_JOB_EMPTY = 2
};
typedef void (*ZSTDMT_prepareJobFn)(void* opaque, unsigned jobID,
                                   const ZSTDMT_RustJobInitialization* init);
typedef void (*ZSTDMT_writeEmptyJobFn)(void* opaque, unsigned jobID);
typedef int (*ZSTDMT_tryAddJobFn)(void* opaque, unsigned jobID);
ZSTDMT_RustCreateJobResult ZSTDMT_rust_createCompressionJob(
        const ZSTDMT_RustCreateJobProjection* projection,
        void* opaque, ZSTDMT_prepareJobFn prepareJob,
        ZSTDMT_writeEmptyJobFn writeEmptyJob, ZSTDMT_tryAddJobFn tryAddJob);

typedef struct {
    rawSeq* seq;
    size_t pos;
    size_t posInSequence;
    size_t size;
    size_t capacity;
} ZSTDMT_RustRawSeqStore;

typedef char ZSTDMT_rust_raw_seq_layout[
        (sizeof(rawSeq) == 3 * sizeof(U32)) ? 1 : -1];
typedef char ZSTDMT_rust_raw_seq_store_layout[
        (sizeof(ZSTDMT_RustRawSeqStore) == sizeof(RawSeqStore_t)
         && offsetof(ZSTDMT_RustRawSeqStore, pos) == offsetof(RawSeqStore_t, pos)
         && offsetof(ZSTDMT_RustRawSeqStore, capacity) == offsetof(RawSeqStore_t, capacity))
                ? 1 : -1];

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);
typedef void (*ZSTDMT_serialStateLockFn)(void* opaque);
typedef void (*ZSTDMT_serialStateWaitFn)(void* opaque);
typedef struct {
    void* callbackContext;
    unsigned* nextJobID;
    ZSTDMT_serialStateLockFn lock;
    ZSTDMT_serialStateWaitFn wait;
    unsigned jobID;
} ZSTDMT_RustSerialWaitForTurnState;
typedef char ZSTDMT_rust_serial_wait_for_turn_state_layout[
        (offsetof(ZSTDMT_RustSerialWaitForTurnState, callbackContext) == 0
         && offsetof(ZSTDMT_RustSerialWaitForTurnState, nextJobID) == sizeof(void*)
         && offsetof(ZSTDMT_RustSerialWaitForTurnState, lock) == 2 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialWaitForTurnState, wait) == 3 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialWaitForTurnState, jobID) == 4 * sizeof(void*)
         && sizeof(ZSTDMT_RustSerialWaitForTurnState) == 5 * sizeof(void*)) ? 1 : -1];
int ZSTDMT_rust_serialStateWaitForTurn(
        const ZSTDMT_RustSerialWaitForTurnState* state);
typedef void (*ZSTDMT_serialStateBroadcastFn)(void* opaque);
typedef void (*ZSTDMT_serialStateUnlockFn)(void* opaque);
typedef struct {
    void* callbackContext;
    unsigned* nextJobID;
    ZSTDMT_serialStateBroadcastFn broadcast;
    ZSTDMT_serialStateUnlockFn unlock;
} ZSTDMT_RustSerialAdvanceState;
typedef char ZSTDMT_rust_serial_advance_state_layout[
        (offsetof(ZSTDMT_RustSerialAdvanceState, callbackContext) == 0
         && offsetof(ZSTDMT_RustSerialAdvanceState, nextJobID) == sizeof(void*)
         && offsetof(ZSTDMT_RustSerialAdvanceState, broadcast) == 2 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialAdvanceState, unlock) == 3 * sizeof(void*)
         && sizeof(ZSTDMT_RustSerialAdvanceState) == 4 * sizeof(void*)) ? 1 : -1];
void ZSTDMT_rust_serialStateAdvance(const ZSTDMT_RustSerialAdvanceState* state);
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);
typedef struct {
    unsigned skip;
    unsigned nextJobID;
} ZSTDMT_RustSerialEnsureFinishedResult;
ZSTDMT_RustSerialEnsureFinishedResult ZSTDMT_rust_serialStateEnsureFinished(
        unsigned nextJobID, unsigned jobID);
typedef void (*ZSTDMT_serialStateCallbackFn)(void* opaque);
typedef void (*ZSTDMT_serialStateSkipFn)(
        void* opaque, unsigned jobID, size_t cSize);
typedef struct {
    void* callbackContext;
    unsigned* nextJobID;
    ZSTDMT_serialStateCallbackFn lock;
    ZSTDMT_serialStateCallbackFn broadcast;
    ZSTDMT_serialStateCallbackFn ldmLock;
    ZSTDMT_serialStateCallbackFn clearLdmWindow;
    ZSTDMT_serialStateCallbackFn ldmSignal;
    ZSTDMT_serialStateCallbackFn ldmUnlock;
    ZSTDMT_serialStateCallbackFn unlock;
    ZSTDMT_serialStateSkipFn onSkip;
    size_t cSize;
    unsigned jobID;
} ZSTDMT_RustSerialEnsureFinishedState;
typedef char ZSTDMT_rust_serial_ensure_finished_state_layout[
        (offsetof(ZSTDMT_RustSerialEnsureFinishedState, callbackContext) == 0
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, nextJobID) == sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, lock) == 2 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, broadcast) == 3 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, ldmLock) == 4 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, clearLdmWindow) == 5 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, ldmSignal) == 6 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, ldmUnlock) == 7 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, unlock) == 8 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, onSkip) == 9 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, cSize) == 10 * sizeof(void*)
         && offsetof(ZSTDMT_RustSerialEnsureFinishedState, jobID)
                == 10 * sizeof(void*) + sizeof(size_t)
         && sizeof(ZSTDMT_RustSerialEnsureFinishedState) == 12 * sizeof(void*)) ? 1 : -1];
void ZSTDMT_rust_serialStateEnsureFinishedOrchestrated(
        const ZSTDMT_RustSerialEnsureFinishedState* state);
typedef void (*ZSTDMT_waitForJobCompleteFn)(
        void* opaque, unsigned jobID, unsigned doneJobID);
unsigned ZSTDMT_rust_waitForAllJobsCompleted(
        unsigned doneJobID, unsigned nextJobID, unsigned jobIDMask,
        void* opaque, ZSTDMT_waitForJobCompleteFn waitForJob);
typedef void (*ZSTDMT_releaseJobResourceFn)(void* opaque, unsigned jobID);
void ZSTDMT_rust_releaseAllJobResources(
        unsigned jobIDMask, void* opaque, ZSTDMT_releaseJobResourceFn releaseJob);
typedef void (*ZSTDMT_waitForLdmLockFn)(void* opaque);
typedef int (*ZSTDMT_waitForLdmOverlapFn)(
        void* opaque, void* bufferStart, size_t bufferCapacity);
typedef void (*ZSTDMT_waitForLdmWaitFn)(void* opaque);
typedef void (*ZSTDMT_waitForLdmUnlockFn)(void* opaque);
typedef struct {
    void* callbackContext;
    void* bufferStart;
    size_t bufferCapacity;
    int ldmEnabled;
    ZSTDMT_waitForLdmLockFn lock;
    ZSTDMT_waitForLdmOverlapFn overlaps;
    ZSTDMT_waitForLdmWaitFn wait;
    ZSTDMT_waitForLdmUnlockFn unlock;
} ZSTDMT_RustWaitForLdmState;
typedef char ZSTDMT_rust_wait_for_ldm_state_layout[
        (offsetof(ZSTDMT_RustWaitForLdmState, callbackContext) == 0
         && offsetof(ZSTDMT_RustWaitForLdmState, bufferStart) == sizeof(void*)
         && offsetof(ZSTDMT_RustWaitForLdmState, bufferCapacity) == 2 * sizeof(void*)
         && offsetof(ZSTDMT_RustWaitForLdmState, ldmEnabled) == 3 * sizeof(void*)
         && offsetof(ZSTDMT_RustWaitForLdmState, lock) == 4 * sizeof(void*)
         && offsetof(ZSTDMT_RustWaitForLdmState, overlaps) == 5 * sizeof(void*)
         && offsetof(ZSTDMT_RustWaitForLdmState, wait) == 6 * sizeof(void*)
         && offsetof(ZSTDMT_RustWaitForLdmState, unlock) == 7 * sizeof(void*)
         && sizeof(ZSTDMT_RustWaitForLdmState) == 8 * sizeof(void*)) ? 1 : -1];
void ZSTDMT_rust_waitForLdmComplete(
        const ZSTDMT_RustWaitForLdmState* state);

unsigned ZSTDMT_rust_computeTargetJobLog(unsigned windowLog, unsigned chainLog,
                                         int strategy, int enableLdm);
int ZSTDMT_rust_overlapLog(int overlapLog, int strategy);
size_t ZSTDMT_rust_computeOverlapSize(unsigned windowLog, unsigned chainLog,
                                      int strategy, int overlapLog, int enableLdm);
int ZSTDMT_rust_doesOverlapWindow(const void* bufferStart, size_t bufferCapacity,
                                  const void* nextSrc, const void* base,
                                  const void* dictBase, U32 dictLimit, U32 lowLimit);
void ZSTDMT_rust_findSynchronizationPoint(const void* inputSrc, size_t inputSize,
                                           size_t inputPos, size_t targetSectionSize,
                                           const void* inBuffStart, size_t inBuffFilled,
                                           int rsyncable, U64 primePower, U64 hitMask,
                                           size_t* toLoad, int* flush);
U64 ZSTDMT_rust_rollingHashPrimePower(U32 length);
size_t ZSTDMT_rust_nextInputSizeHint(size_t targetSectionSize,
                                     size_t inBuffFilled);
typedef struct {
    unsigned requestedNbWorkers;
    unsigned currentNbWorkers;
    size_t jobSize;
    size_t jobSizeMin;
    size_t jobSizeMax;
    int enableLdm;
    unsigned windowLog;
    unsigned chainLog;
    int strategy;
    int overlapLog;
    int rsyncable;
    size_t roundBuffCapacity;
    unsigned allJobsCompleted;
} ZSTDMT_RustInitCStreamProjection;
typedef size_t (*ZSTDMT_initResizeFn)(void* opaque, unsigned nbWorkers);
typedef void (*ZSTDMT_initDrainFn)(void* opaque);
typedef void (*ZSTDMT_initApplyParametersFn)(void* opaque, size_t jobSize);
typedef size_t (*ZSTDMT_initDictionaryFn)(void* opaque);
typedef void (*ZSTDMT_initSetSizeFn)(void* opaque, size_t size);
typedef void (*ZSTDMT_initSetRsyncFn)(void* opaque, U64 hitMask, U64 primePower);
typedef void (*ZSTDMT_initSetBufferSizeFn)(void* opaque, size_t size);
typedef size_t (*ZSTDMT_initResizeRoundBufferFn)(void* opaque, size_t capacity);
typedef void (*ZSTDMT_initResetStreamFn)(void* opaque);
typedef size_t (*ZSTDMT_initSerialResetFn)(void* opaque, size_t targetSectionSize);
typedef void (*ZSTDMT_serialResetVoidFn)(void* opaque);
typedef void (*ZSTDMT_serialResetSetNbSeqFn)(void* opaque, size_t nbSeq);
typedef int (*ZSTDMT_serialResetResizeFn)(void* opaque);
int ZSTDMT_rust_serialStateReset(
        int enableLdm, int checksumEnabled, size_t maxNbSeq, void* opaque,
        ZSTDMT_serialResetVoidFn resetNextJob,
        ZSTDMT_serialResetVoidFn resetChecksum,
        ZSTDMT_serialResetSetNbSeqFn setNbSeq,
        ZSTDMT_serialResetVoidFn resetWindow,
        ZSTDMT_serialResetResizeFn resizeTables,
        ZSTDMT_serialResetVoidFn zeroTables,
        ZSTDMT_serialResetVoidFn loadDictionary,
        ZSTDMT_serialResetVoidFn copyWindow);
typedef void (*ZSTDMT_serialStateVoidFn)(void* opaque);
typedef int (*ZSTDMT_serialStateInitFn)(void* opaque);
int ZSTDMT_rust_serialStateInit(
        void* opaque, ZSTDMT_serialStateVoidFn zeroState,
        ZSTDMT_serialStateInitFn initMutex,
        ZSTDMT_serialStateInitFn initCond,
        ZSTDMT_serialStateInitFn initLdmMutex,
        ZSTDMT_serialStateInitFn initLdmCond);
void ZSTDMT_rust_serialStateFree(
        void* opaque, ZSTDMT_serialStateVoidFn destroyMutex,
        ZSTDMT_serialStateVoidFn destroyCond,
        ZSTDMT_serialStateVoidFn destroyLdmMutex,
        ZSTDMT_serialStateVoidFn destroyLdmCond,
        ZSTDMT_serialStateVoidFn freeTables);
size_t ZSTDMT_rust_initCStream(
        const ZSTDMT_RustInitCStreamProjection* projection, void* opaque,
        ZSTDMT_initResizeFn resize, ZSTDMT_initDrainFn drain,
        ZSTDMT_initApplyParametersFn applyParameters,
        ZSTDMT_initDictionaryFn prepareDictionary,
        ZSTDMT_initSetSizeFn setTargetPrefixSize,
        ZSTDMT_initSetSizeFn setTargetSectionSize,
        ZSTDMT_initSetRsyncFn setRsync,
        ZSTDMT_initSetBufferSizeFn setBufferSize,
        ZSTDMT_initResizeRoundBufferFn resizeRoundBuffer,
        ZSTDMT_initResetStreamFn resetStream,
        ZSTDMT_initDictionaryFn updateDictionary,
        ZSTDMT_initSerialResetFn serialReset);
typedef struct {
    size_t consumed;
    size_t cSize;
    const void* srcStart;
    size_t srcSize;
    const void* prefixStart;
    size_t prefixSize;
    size_t dstFlushed;
} ZSTDMT_RustJobProjection;
typedef struct {
    const void* start;
    size_t size;
} ZSTDMT_RustInputRange;
typedef struct {
    const void* roundBufferStart;
    size_t roundBufferCapacity;
    size_t roundBufferPos;
    const void* prefixStart;
    size_t prefixSize;
    size_t targetSectionSize;
    const void* inUseStart;
    size_t inUseSize;
} ZSTDMT_RustTryGetInputRangeProjection;
typedef struct {
    unsigned ready;
    unsigned movePrefix;
    void* bufferStart;
    size_t bufferCapacity;
    size_t roundBufferPos;
} ZSTDMT_RustTryGetInputRangeResult;
typedef void (*ZSTDMT_jobProjectionFn)(void* opaque, unsigned jobID,
                                      ZSTDMT_RustJobProjection* projection);
ZSTDMT_RustInputRange ZSTDMT_rust_getInputDataInUse(
        unsigned firstJobID, unsigned lastJobID, unsigned jobIDMask,
        size_t roundBufferCapacity, size_t targetSectionSize,
        void* opaque, ZSTDMT_jobProjectionFn projectJob);
ZSTDMT_RustTryGetInputRangeResult ZSTDMT_rust_tryGetInputRange(
        const ZSTDMT_RustTryGetInputRangeProjection* projection);
size_t ZSTDMT_rust_toFlushNow(unsigned doneJobID, unsigned nextJobID,
                              unsigned jobIDMask, void* opaque,
                              ZSTDMT_jobProjectionFn projectJob);
size_t ZSTDMT_rust_sizeofCCtx(size_t mtctxSize, size_t factorySize,
                              size_t bufferPoolSize, size_t jobsSize,
                              size_t cctxPoolSize, size_t seqPoolSize,
                              size_t cdictSize, size_t roundBuffSize);
ZSTD_frameProgression ZSTDMT_rust_frameProgression(
        unsigned long long consumed, size_t inBuffFilled,
        unsigned long long produced, unsigned currentJobID);
ZSTD_frameProgression ZSTDMT_rust_frameProgressionAddJob(
        ZSTD_frameProgression progression, size_t srcSize, size_t consumed,
        size_t produced, size_t flushed);
ZSTD_frameProgression ZSTDMT_rust_frameProgressionWithJobs(
        unsigned long long consumed, size_t inBuffFilled,
        unsigned long long produced, unsigned currentJobID,
        unsigned doneJobID, unsigned nextJobID, unsigned jobReady,
        unsigned jobIDMask, void* opaque, ZSTDMT_jobProjectionFn projectJob);
typedef int (*ZSTDMT_jobTableInitFn)(void* jobTable, unsigned nbJobs,
                                     size_t jobSize);
typedef void (*ZSTDMT_jobTableDestroyFn)(void* jobTable, unsigned nbJobs,
                                         size_t jobSize);
size_t ZSTDMT_rust_expandJobsTable(
        void** jobTablePtr, unsigned* jobIDMaskPtr, unsigned nbWorkers,
        size_t jobSize, ZSTD_customMem cMem,
        ZSTDMT_jobTableInitFn initSync,
        ZSTDMT_jobTableDestroyFn destroySync);

typedef struct ZSTDMT_bufferPool_s {
    ZSTDMT_RustBufferPool* rustPool;
    ZSTD_customMem cMem;
} ZSTDMT_bufferPool;

static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool)
{
    if (!bufPool) return;   /* compatibility with free on NULL */
    ZSTDMT_rust_buffer_pool_free(bufPool->rustPool);
    ZSTD_customFree(bufPool, bufPool->cMem);
}

static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned maxNbBuffers, ZSTD_customMem cMem)
{
    ZSTDMT_bufferPool* const bufPool =
        (ZSTDMT_bufferPool*)ZSTD_customCalloc(sizeof(ZSTDMT_bufferPool), cMem);
    if (bufPool==NULL) return NULL;
    bufPool->rustPool = ZSTDMT_rust_buffer_pool_create(maxNbBuffers, cMem);
    if (bufPool->rustPool == NULL) {
        ZSTD_customFree(bufPool, cMem);
        return NULL;
    }
    bufPool->cMem = cMem;
    return bufPool;
}

/* only works at initialization, not during compression */
static size_t ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool* bufPool)
{
    if (bufPool == NULL) return 0;
    return ZSTDMT_rust_sizeofBufferPool(
        sizeof(*bufPool), ZSTDMT_rust_buffer_pool_sizeof(bufPool->rustPool));
}

/* ZSTDMT_setBufferSize() :
 * all future buffers provided by this buffer pool will have _at least_ this size
 * note : it's better for all buffers to have same size,
 * as they become freely interchangeable, reducing malloc/free usages and memory fragmentation */
static void ZSTDMT_setBufferSize(ZSTDMT_bufferPool* const bufPool, size_t const bSize)
{
    DEBUGLOG(4, "ZSTDMT_setBufferSize: bSize = %u", (U32)bSize);
    ZSTDMT_rust_buffer_pool_set_size(bufPool->rustPool, bSize);
}


static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, unsigned maxNbBuffers)
{
    if (srcBufPool==NULL) return NULL;
    srcBufPool->rustPool = ZSTDMT_rust_buffer_pool_expand(srcBufPool->rustPool, maxNbBuffers);
    if (srcBufPool->rustPool == NULL) {
        ZSTD_customMem const cMem = srcBufPool->cMem;
        ZSTD_customFree(srcBufPool, cMem);
        return NULL;
    }
    return srcBufPool;
}

/** ZSTDMT_getBuffer() :
 *  assumption : bufPool must be valid
 * @return : a buffer, with start pointer and size
 *  note: allocation may fail, in this case, start==NULL and size==0 */
static Buffer ZSTDMT_getBuffer(ZSTDMT_bufferPool* bufPool)
{
    ZSTDMT_RustBuffer const rustBuffer =
        ZSTDMT_rust_buffer_pool_get(bufPool->rustPool);
    Buffer buffer = { rustBuffer.start, rustBuffer.capacity };
    return buffer;
}

static ZSTDMT_RustBuffer ZSTDMT_getBufferForRust(void* opaque)
{
    Buffer const buffer = ZSTDMT_getBuffer((ZSTDMT_bufferPool*)opaque);
    ZSTDMT_RustBuffer const result = { buffer.start, buffer.capacity };
    return result;
}

#if ZSTD_RESIZE_SEQPOOL
/** ZSTDMT_resizeBuffer() :
 * assumption : bufPool must be valid
 * @return : a buffer that is at least the buffer pool buffer size.
 *           If a reallocation happens, the data in the input buffer is copied.
 */
static Buffer ZSTDMT_resizeBuffer(ZSTDMT_bufferPool* bufPool, Buffer buffer)
{
    ZSTDMT_RustBuffer const rustBuffer = { buffer.start, buffer.capacity };
    ZSTDMT_RustBuffer const resized =
        ZSTDMT_rust_buffer_pool_resize(bufPool->rustPool, rustBuffer);
    Buffer const result = { resized.start, resized.capacity };
    return result;
}
#endif

/* store buffer for later re-use, up to pool capacity */
static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, Buffer buf)
{
    ZSTDMT_RustBuffer const rustBuffer = { buf.start, buf.capacity };
    ZSTDMT_rust_buffer_pool_release(bufPool->rustPool, rustBuffer);
}

/* We need 2 output buffers per worker since each dstBuff must be flushed after it is released.
 * The 3 additional buffers are as follows:
 *   1 buffer for input loading
 *   1 buffer for "next input" when submitting current one
 *   1 buffer stuck in queue */
#define BUF_POOL_MAX_NB_BUFFERS(nbWorkers) (2*(nbWorkers) + 3)

/* After a worker releases its rawSeqStore, it is immediately ready for reuse.
 * So we only need one seq buffer per worker. */
#define SEQ_POOL_MAX_NB_BUFFERS(nbWorkers) (nbWorkers)

/* =====   Seq Pool Wrapper   ====== */

typedef ZSTDMT_bufferPool ZSTDMT_seqPool;

static size_t ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool* seqPool)
{
    return ZSTDMT_sizeof_bufferPool(seqPool);
}

static RawSeqStore_t bufferToSeq(Buffer buffer)
{
    ZSTDMT_RustRawSeqStore const rustSeq =
        ZSTDMT_rust_bufferToSeq((ZSTDMT_RustBuffer){ buffer.start, buffer.capacity });
    RawSeqStore_t seq = {
        (rawSeq*)rustSeq.seq,
        rustSeq.pos,
        rustSeq.posInSequence,
        rustSeq.size,
        rustSeq.capacity
    };
    return seq;
}

static Buffer seqToBuffer(RawSeqStore_t seq)
{
    ZSTDMT_RustRawSeqStore const rustSeq = {
        seq.seq,
        seq.pos,
        seq.posInSequence,
        seq.size,
        seq.capacity
    };
    ZSTDMT_RustBuffer const rustBuffer = ZSTDMT_rust_seqToBuffer(rustSeq);
    return (Buffer){ rustBuffer.start, rustBuffer.capacity };
}

static RawSeqStore_t ZSTDMT_getSeq(ZSTDMT_seqPool* seqPool)
{
    return bufferToSeq(ZSTDMT_getBuffer(seqPool));
}

#if ZSTD_RESIZE_SEQPOOL
static RawSeqStore_t ZSTDMT_resizeSeq(ZSTDMT_seqPool* seqPool, RawSeqStore_t seq)
{
  return bufferToSeq(ZSTDMT_resizeBuffer(seqPool, seqToBuffer(seq)));
}
#endif

static void ZSTDMT_releaseSeq(ZSTDMT_seqPool* seqPool, RawSeqStore_t seq)
{
  ZSTDMT_releaseBuffer(seqPool, seqToBuffer(seq));
}

static void ZSTDMT_setNbSeq(ZSTDMT_seqPool* const seqPool, size_t const nbSeq)
{
  ZSTDMT_setBufferSize(seqPool, nbSeq * sizeof(rawSeq));
}

static ZSTDMT_seqPool* ZSTDMT_createSeqPool(unsigned nbWorkers, ZSTD_customMem cMem)
{
    ZSTDMT_seqPool* const seqPool = ZSTDMT_createBufferPool(SEQ_POOL_MAX_NB_BUFFERS(nbWorkers), cMem);
    if (seqPool == NULL) return NULL;
    ZSTDMT_setNbSeq(seqPool, 0);
    return seqPool;
}

static void ZSTDMT_freeSeqPool(ZSTDMT_seqPool* seqPool)
{
    ZSTDMT_freeBufferPool(seqPool);
}

static ZSTDMT_seqPool* ZSTDMT_expandSeqPool(ZSTDMT_seqPool* pool, U32 nbWorkers)
{
    return ZSTDMT_expandBufferPool(pool, SEQ_POOL_MAX_NB_BUFFERS(nbWorkers));
}


/* =====   CCtx Pool   ===== */
/* a single CCtx Pool can be invoked from multiple threads in parallel */

typedef struct ZSTDMT_RustCCtxPool_s ZSTDMT_RustCCtxPool;

ZSTDMT_RustCCtxPool* ZSTDMT_rust_cctx_pool_create(unsigned nbWorkers,
                                                   ZSTD_customMem cMem);
void ZSTDMT_rust_cctx_pool_free(ZSTDMT_RustCCtxPool* pool);
size_t ZSTDMT_rust_cctx_pool_sizeof(const ZSTDMT_RustCCtxPool* pool);
size_t ZSTDMT_rust_sizeofCCtxPool(size_t wrapperSize, size_t rustPoolSize);
ZSTDMT_RustCCtxPool* ZSTDMT_rust_cctx_pool_expand(ZSTDMT_RustCCtxPool* pool,
                                                   unsigned nbWorkers);
ZSTD_CCtx* ZSTDMT_rust_cctx_pool_get(ZSTDMT_RustCCtxPool* pool);
void ZSTDMT_rust_cctx_pool_release(ZSTDMT_RustCCtxPool* pool, ZSTD_CCtx* cctx);

typedef struct {
    ZSTDMT_RustCCtxPool* rustPool;
    int totalCCtx; /* kept for the existing MT parameter diagnostics */
    ZSTD_customMem cMem;
} ZSTDMT_CCtxPool;

/* note : all CCtx borrowed from the pool must be reverted back to the pool _before_ freeing the pool */
static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
{
    if (!pool) return;
    ZSTDMT_rust_cctx_pool_free(pool->rustPool);
    ZSTD_customFree(pool, pool->cMem);
}

/* ZSTDMT_createCCtxPool() :
 * implies nbWorkers >= 1 , checked by caller ZSTDMT_createCCtx() */
static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(int nbWorkers,
                                              ZSTD_customMem cMem)
{
    ZSTDMT_CCtxPool* const cctxPool =
        (ZSTDMT_CCtxPool*) ZSTD_customCalloc(sizeof(ZSTDMT_CCtxPool), cMem);
    assert(nbWorkers > 0);
    if (!cctxPool) return NULL;
    cctxPool->rustPool = ZSTDMT_rust_cctx_pool_create((unsigned)nbWorkers, cMem);
    if (cctxPool->rustPool == NULL) {
        ZSTD_customFree(cctxPool, cMem);
        return NULL;
    }
    cctxPool->totalCCtx = nbWorkers;
    cctxPool->cMem = cMem;
    DEBUGLOG(3, "cctxPool created, with %u workers", nbWorkers);
    return cctxPool;
}

static ZSTDMT_CCtxPool* ZSTDMT_expandCCtxPool(ZSTDMT_CCtxPool* srcPool,
                                              int nbWorkers)
{
    if (srcPool==NULL) return NULL;
    srcPool->rustPool = ZSTDMT_rust_cctx_pool_expand(srcPool->rustPool,
                                                      (unsigned)nbWorkers);
    if (srcPool->rustPool == NULL) {
        ZSTD_customMem const cMem = srcPool->cMem;
        ZSTD_customFree(srcPool, cMem);
        return NULL;
    }
    srcPool->totalCCtx = nbWorkers;
    return srcPool;
}

/* only works during initialization phase, not during compression */
static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
{
    size_t rustPoolSize;
    if (cctxPool == NULL) return 0;
    rustPoolSize = ZSTDMT_rust_cctx_pool_sizeof(cctxPool->rustPool);
    return ZSTDMT_rust_sizeofCCtxPool(sizeof(*cctxPool), rustPoolSize);
}

static ZSTD_CCtx* ZSTDMT_getCCtx(ZSTDMT_CCtxPool* cctxPool)
{
    DEBUGLOG(5, "ZSTDMT_getCCtx");
    return ZSTDMT_rust_cctx_pool_get(cctxPool->rustPool);
}

static void ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool* pool, ZSTD_CCtx* cctx)
{
    ZSTDMT_rust_cctx_pool_release(pool->rustPool, cctx);
}

/* ====   Serial State   ==== */

typedef struct {
    void const* start;
    size_t size;
} Range;

typedef struct {
    /* All variables in the struct are protected by mutex. */
    ZSTD_pthread_mutex_t mutex;
    ZSTD_pthread_cond_t cond;
    ZSTD_CCtx_params params;
    ldmState_t ldmState;
    XXH64_state_t xxhState;
    unsigned nextJobID;
    /* Protects ldmWindow.
     * Must be acquired after the main mutex when acquiring both.
     */
    ZSTD_pthread_mutex_t ldmWindowMutex;
    ZSTD_pthread_cond_t ldmWindowCond;  /* Signaled when ldmWindow is updated */
    ZSTD_window_t ldmWindow;  /* A thread-safe copy of ldmState.window */
} SerialState;

typedef struct {
    SerialState* serialState;
    ZSTDMT_seqPool* seqPool;
    ZSTD_CCtx_params* params;
    const void* dict;
    size_t dictSize;
    ZSTD_dictContentType_e dictContentType;
    size_t hashSize;
    size_t numBuckets;
    unsigned bucketLog;
    unsigned prevBucketLog;
    ZSTD_customMem cMem;
} ZSTDMT_serialResetContext;

static void ZSTDMT_serialResetNextJob(void* opaque)
{
    ZSTDMT_serialResetContext* const context = (ZSTDMT_serialResetContext*)opaque;
    context->serialState->nextJobID = 0;
}

static void ZSTDMT_serialResetChecksum(void* opaque)
{
    ZSTDMT_serialResetContext* const context = (ZSTDMT_serialResetContext*)opaque;
    XXH64_reset(&context->serialState->xxhState, 0);
}

static void ZSTDMT_serialResetSetNbSeq(void* opaque, size_t nbSeq)
{
    ZSTDMT_serialResetContext* const context = (ZSTDMT_serialResetContext*)opaque;
    ZSTDMT_setNbSeq(context->seqPool, nbSeq);
}

static void ZSTDMT_serialResetWindow(void* opaque)
{
    ZSTDMT_serialResetContext* const context = (ZSTDMT_serialResetContext*)opaque;
    ZSTD_window_init(&context->serialState->ldmState.window);
}

static int ZSTDMT_serialResetResizeTables(void* opaque)
{
    ZSTDMT_serialResetContext* const context = (ZSTDMT_serialResetContext*)opaque;
    SerialState* const serialState = context->serialState;
    ldmParams_t const* const ldmParams = &context->params->ldmParams;

    if (serialState->ldmState.hashTable == NULL ||
            serialState->params.ldmParams.hashLog < ldmParams->hashLog) {
        ZSTD_customFree(serialState->ldmState.hashTable, context->cMem);
        serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_customMalloc(
                context->hashSize, context->cMem);
    }
    if (serialState->ldmState.bucketOffsets == NULL ||
            context->prevBucketLog < context->bucketLog) {
        ZSTD_customFree(serialState->ldmState.bucketOffsets, context->cMem);
        serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_customMalloc(
                context->numBuckets, context->cMem);
    }
    return !serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets;
}

static void ZSTDMT_serialResetZeroTables(void* opaque)
{
    ZSTDMT_serialResetContext* const context = (ZSTDMT_serialResetContext*)opaque;
    SerialState* const serialState = context->serialState;
    ZSTD_memset(serialState->ldmState.hashTable, 0, context->hashSize);
    ZSTD_memset(serialState->ldmState.bucketOffsets, 0, context->numBuckets);
}

static void ZSTDMT_serialResetLoadDictionary(void* opaque)
{
    ZSTDMT_serialResetContext* const context = (ZSTDMT_serialResetContext*)opaque;
    SerialState* const serialState = context->serialState;

    serialState->ldmState.loadedDictEnd = 0;
    if (context->dictSize > 0 &&
            context->dictContentType == ZSTD_dct_rawContent) {
        BYTE const* const dictEnd = (const BYTE*)context->dict + context->dictSize;
        ZSTD_window_update(&serialState->ldmState.window,
                           context->dict, context->dictSize,
                           /* forceNonContiguous */ 0);
        ZSTD_ldm_fillHashTable(&serialState->ldmState,
                               (const BYTE*)context->dict, dictEnd,
                               &context->params->ldmParams);
        serialState->ldmState.loadedDictEnd = context->params->forceWindow
                ? 0 : (U32)(dictEnd - serialState->ldmState.window.base);
    }
}

static void ZSTDMT_serialResetCopyWindow(void* opaque)
{
    ZSTDMT_serialResetContext* const context = (ZSTDMT_serialResetContext*)opaque;
    context->serialState->ldmWindow = context->serialState->ldmState.window;
}

static int
ZSTDMT_serialState_reset(SerialState* serialState,
                         ZSTDMT_seqPool* seqPool,
                         ZSTD_CCtx_params params,
                         size_t jobSize,
                         const void* dict, size_t const dictSize,
                         ZSTD_dictContentType_e dictContentType)
{
    ZSTDMT_serialResetContext context;

    /* Adjust parameters */
    if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
        DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10);
        ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
        assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
        assert(params.ldmParams.hashRateLog < 32);
    } else {
        ZSTD_memset(&params.ldmParams, 0, sizeof(params.ldmParams));
    }

    context.serialState = serialState;
    context.seqPool = seqPool;
    context.params = &params;
    context.dict = dict;
    context.dictSize = dictSize;
    context.dictContentType = dictContentType;
    context.cMem = params.customMem;
    context.hashSize = 0;
    context.numBuckets = 0;
    context.bucketLog = 0;
    context.prevBucketLog = 0;
    if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
        unsigned const hashLog = params.ldmParams.hashLog;
        context.hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
        context.bucketLog = params.ldmParams.hashLog - params.ldmParams.bucketSizeLog;
        context.prevBucketLog =
                serialState->params.ldmParams.hashLog -
                serialState->params.ldmParams.bucketSizeLog;
        context.numBuckets = (size_t)1 << context.bucketLog;
    }

    if (ZSTDMT_rust_serialStateReset(
            params.ldmParams.enableLdm == ZSTD_ps_enable,
            params.fParams.checksumFlag,
            ZSTD_ldm_getMaxNbSeq(params.ldmParams, jobSize),
            &context,
            ZSTDMT_serialResetNextJob,
            ZSTDMT_serialResetChecksum,
            ZSTDMT_serialResetSetNbSeq,
            ZSTDMT_serialResetWindow,
            ZSTDMT_serialResetResizeTables,
            ZSTDMT_serialResetZeroTables,
            ZSTDMT_serialResetLoadDictionary,
            ZSTDMT_serialResetCopyWindow)) {
        return 1;
    }

    serialState->params = params;
    serialState->params.jobSize = (U32)jobSize;
    return 0;
}

static void ZSTDMT_serialState_zero(void* opaque)
{
    ZSTD_memset(opaque, 0, sizeof(SerialState));
}

static int ZSTDMT_serialState_initMutex(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    return ZSTD_pthread_mutex_init(&serialState->mutex, NULL);
}

static int ZSTDMT_serialState_initCond(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    return ZSTD_pthread_cond_init(&serialState->cond, NULL);
}

static int ZSTDMT_serialState_initLdmMutex(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    return ZSTD_pthread_mutex_init(&serialState->ldmWindowMutex, NULL);
}

static int ZSTDMT_serialState_initLdmCond(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    return ZSTD_pthread_cond_init(&serialState->ldmWindowCond, NULL);
}

static void ZSTDMT_serialState_destroyMutex(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_pthread_mutex_destroy(&serialState->mutex);
}

static void ZSTDMT_serialState_destroyCond(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_pthread_cond_destroy(&serialState->cond);
}

static void ZSTDMT_serialState_destroyLdmMutex(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_pthread_mutex_destroy(&serialState->ldmWindowMutex);
}

static void ZSTDMT_serialState_destroyLdmCond(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_pthread_cond_destroy(&serialState->ldmWindowCond);
}

static void ZSTDMT_serialState_freeTables(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_customMem const cMem = serialState->params.customMem;
    ZSTD_customFree(serialState->ldmState.hashTable, cMem);
    ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
}

static int ZSTDMT_serialState_init(SerialState* serialState)
{
    return ZSTDMT_rust_serialStateInit(
            serialState,
            ZSTDMT_serialState_zero,
            ZSTDMT_serialState_initMutex,
            ZSTDMT_serialState_initCond,
            ZSTDMT_serialState_initLdmMutex,
            ZSTDMT_serialState_initLdmCond);
}

static void ZSTDMT_serialState_free(SerialState* serialState)
{
    ZSTDMT_rust_serialStateFree(
            serialState,
            ZSTDMT_serialState_destroyMutex,
            ZSTDMT_serialState_destroyCond,
            ZSTDMT_serialState_destroyLdmMutex,
            ZSTDMT_serialState_destroyLdmCond,
            ZSTDMT_serialState_freeTables);
}

static void ZSTDMT_serialState_lock(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
}

static void ZSTDMT_serialState_wait(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    DEBUGLOG(5, "wait for serialState->cond");
    ZSTD_pthread_cond_wait(&serialState->cond, &serialState->mutex);
}

static void ZSTDMT_serialState_broadcast(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_pthread_cond_broadcast(&serialState->cond);
}

static void ZSTDMT_serialState_unlock(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_pthread_mutex_unlock(&serialState->mutex);
}

/* 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)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTDMT_RustSerialWaitForTurnState state;
    state.callbackContext = serialState;
    state.nextJobID = &serialState->nextJobID;
    state.lock = ZSTDMT_serialState_lock;
    state.wait = ZSTDMT_serialState_wait;
    state.jobID = jobID;
    return ZSTDMT_rust_serialStateWaitForTurn(&state);
}

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;
    ZSTDMT_RustSerialAdvanceState state;
    state.callbackContext = serialState;
    state.nextJobID = &serialState->nextJobID;
    state.broadcast = ZSTDMT_serialState_broadcast;
    state.unlock = ZSTDMT_serialState_unlock;
    ZSTDMT_rust_serialStateAdvance(&state);
}

static void ZSTDMT_serialState_ldmLock(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
}

static void ZSTDMT_serialState_clearLdmWindow(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_rust_windowClear((size_t)(serialState->ldmWindow.nextSrc -
                                   serialState->ldmWindow.base),
                          &serialState->ldmWindow.lowLimit,
                          &serialState->ldmWindow.dictLimit);
}

static void ZSTDMT_serialState_ldmSignal(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
}

static void ZSTDMT_serialState_ldmUnlock(void* opaque)
{
    SerialState* const serialState = (SerialState*)opaque;
    ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
}

static void ZSTDMT_serialState_onSkip(void* opaque, unsigned jobID, size_t cSize)
{
    (void)opaque;
    assert(ZSTD_isError(cSize)); (void)cSize;
    DEBUGLOG(5, "Skipping past job %u because of error", jobID);
}

static void
ZSTDMT_serialState_applySequences(const SerialState* serialState, /* just for an assert() check */
                                  ZSTD_CCtx* jobCCtx,
                                  const RawSeqStore_t* seqStore)
{
    if (seqStore->size > 0) {
        DEBUGLOG(5, "ZSTDMT_serialState_applySequences: uploading %u external sequences", (unsigned)seqStore->size);
        assert(serialState->params.ldmParams.enableLdm == ZSTD_ps_enable); (void)serialState;
        assert(jobCCtx);
        ZSTD_referenceExternalSequences(jobCCtx, seqStore->seq, seqStore->size);
    }
}

static void ZSTDMT_serialState_ensureFinished(SerialState* serialState,
                                              unsigned jobID, size_t cSize)
{
    ZSTDMT_RustSerialEnsureFinishedState state;
    state.callbackContext = serialState;
    state.nextJobID = &serialState->nextJobID;
    state.lock = ZSTDMT_serialState_lock;
    state.broadcast = ZSTDMT_serialState_broadcast;
    state.ldmLock = ZSTDMT_serialState_ldmLock;
    state.clearLdmWindow = ZSTDMT_serialState_clearLdmWindow;
    state.ldmSignal = ZSTDMT_serialState_ldmSignal;
    state.ldmUnlock = ZSTDMT_serialState_ldmUnlock;
    state.unlock = ZSTDMT_serialState_unlock;
    state.onSkip = ZSTDMT_serialState_onSkip;
    state.cSize = cSize;
    state.jobID = jobID;
    ZSTDMT_rust_serialStateEnsureFinishedOrchestrated(&state);

}


/* ------------------------------------------ */
/* =====          Worker thread         ===== */
/* ------------------------------------------ */

static const Range kNullRange = { NULL, 0 };

typedef struct {
    size_t   consumed;                 /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx */
    size_t   cSize;                    /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx, then set0 by mtctx */
    ZSTD_pthread_mutex_t job_mutex;    /* Thread-safe - used by mtctx and worker */
    ZSTD_pthread_cond_t job_cond;      /* Thread-safe - used by mtctx and worker */
    ZSTDMT_CCtxPool* cctxPool;         /* Thread-safe - used by mtctx and (all) workers */
    ZSTDMT_bufferPool* bufPool;        /* Thread-safe - used by mtctx and (all) workers */
    ZSTDMT_seqPool* seqPool;           /* Thread-safe - used by mtctx and (all) workers */
    SerialState* serial;               /* Thread-safe - used by mtctx and (all) workers */
    Buffer dstBuff;                    /* set by worker (or mtctx), then read by worker & mtctx, then modified by mtctx => no barrier */
    Range prefix;                      /* set by mtctx, then read by worker & mtctx => no barrier */
    Range src;                         /* set by mtctx, then read by worker & mtctx => no barrier */
    unsigned jobID;                    /* set by mtctx, then read by worker => no barrier */
    unsigned firstJob;                 /* set by mtctx, then read by worker => no barrier */
    unsigned lastJob;                  /* set by mtctx, then read by worker => no barrier */
    ZSTD_CCtx_params params;           /* set by mtctx, then read by worker => no barrier */
    const ZSTD_CDict* cdict;           /* set by mtctx, then read by worker => no barrier */
    unsigned long long fullFrameSize;  /* set by mtctx, then read by worker => no barrier */
    size_t   dstFlushed;               /* used only by mtctx */
    unsigned frameChecksumNeeded;      /* used only by mtctx */
} ZSTDMT_jobDescription;

static void ZSTDMT_compressionJobProgress(void* opaque, size_t cSize, size_t consumed)
{
    ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)opaque;
    ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
    job->cSize += cSize;
    job->consumed = consumed;
    DEBUGLOG(5, "ZSTDMT_compressionJob: compress new block : cSize==%u bytes (total: %u)",
                (U32)cSize, (U32)job->cSize);
    ZSTD_pthread_cond_signal(&job->job_cond);   /* warns some more data is ready to be flushed */
    ZSTD_pthread_mutex_unlock(&job->job_mutex);
}

typedef struct {
    ZSTDMT_jobDescription* job;
    ZSTD_CCtx_params jobParams;
    ZSTD_CCtx* cctx;
    RawSeqStore_t rawSeqStore;
    Buffer dstBuff;
} ZSTDMT_compressionJobState;

static size_t ZSTDMT_compressionJobAcquireResources(void* opaque)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;

    state->cctx = ZSTDMT_getCCtx(job->cctxPool);
    state->rawSeqStore = ZSTDMT_getSeq(job->seqPool);
    state->dstBuff = job->dstBuff;
    DEBUGLOG(5, "ZSTDMT_compressionJob: job %u", job->jobID);
    if (state->cctx == NULL) return ERROR(memory_allocation);
    if (state->dstBuff.start == NULL) {
        state->dstBuff = ZSTDMT_getBuffer(job->bufPool);
        if (state->dstBuff.start == NULL) return ERROR(memory_allocation);
        job->dstBuff = state->dstBuff;
    }
    if (state->jobParams.ldmParams.enableLdm == ZSTD_ps_enable &&
        state->rawSeqStore.seq == NULL)
        return ERROR(memory_allocation);
    return 0;
}

static void ZSTDMT_compressionJobPrepareParameters(void* opaque)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;

    /* Don't compute the checksum for chunks, since we compute it externally,
     * but write it in the header. */
    if (job->jobID != 0) state->jobParams.fParams.checksumFlag = 0;
    /* Don't run LDM for the chunks, since we handle it externally. */
    state->jobParams.ldmParams.enableLdm = ZSTD_ps_disable;
    /* Correct nbWorkers to 0. */
    state->jobParams.nbWorkers = 0;
}

static void ZSTDMT_compressionJobGenerateSequences(void* opaque)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;

    /* Perform serial step as early as possible. */
    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)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;

    if (job->cdict) {
        size_t const initError = ZSTD_compressBegin_advanced_internal(
            state->cctx, NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast, job->cdict,
            &state->jobParams, job->fullFrameSize);
        assert(job->firstJob);  /* only allowed for first job */
        return initError;
    }

    {   U64 const pledgedSrcSize = job->firstJob ? job->fullFrameSize : job->src.size;
        size_t const forceWindowError = ZSTD_CCtxParams_setParameter(
            &state->jobParams, ZSTD_c_forceMaxWindow, !job->firstJob);
        if (ZSTD_isError(forceWindowError)) return forceWindowError;
        if (!job->firstJob) {
            size_t const err = ZSTD_CCtxParams_setParameter(
                &state->jobParams, ZSTD_c_deterministicRefPrefix, 0);
            if (ZSTD_isError(err)) return err;
        }
        DEBUGLOG(6, "ZSTDMT_compressionJob: job %u: loading prefix of size %zu", job->jobID, job->prefix.size);
        return ZSTD_compressBegin_advanced_internal(
            state->cctx, job->prefix.start, job->prefix.size,
            ZSTD_dct_rawContent, ZSTD_dtlm_fast, NULL, /*cdict*/
            &state->jobParams, pledgedSrcSize);
    }
}

static void ZSTDMT_compressionJobApplySequences(void* opaque)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;

    /* External Sequences can only be applied after CCtx initialization. */
    ZSTDMT_serialState_applySequences(job->serial, state->cctx,
                                      &state->rawSeqStore);
}

static size_t ZSTDMT_compressionJobWriteFrameHeader(void* opaque)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;
    size_t const hSize = ZSTD_compressContinue_public(
        state->cctx, state->dstBuff.start, state->dstBuff.capacity,
        job->src.start, 0);
    if (ZSTD_isError(hSize)) return hSize;
    DEBUGLOG(5, "ZSTDMT_compressionJob: flush and overwrite %u bytes of frame header (not first job)", (U32)hSize);
    ZSTD_invalidateRepCodes(state->cctx);
    return hSize;
}

static ZSTDMT_chunkProcessResult ZSTDMT_compressionJobCompress(
        void* opaque, unsigned lastJob)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;
    size_t const chunkSize = 4*ZSTD_BLOCKSIZE_MAX;

    assert(lastJob == job->lastJob);
    if (sizeof(size_t) > sizeof(int)) assert(job->src.size < ((size_t)INT_MAX) * chunkSize);   /* check overflow */
    DEBUGLOG(5, "ZSTDMT_compressionJob: compress %u bytes in %zu blocks",
                (U32)job->src.size, (job->src.size + (chunkSize-1)) / chunkSize);
    assert(job->cSize == 0);
    assert(chunkSize > 0);
    assert((chunkSize & (chunkSize - 1)) == 0);  /* chunkSize must be power of 2 for mask==(chunkSize-1) to work */
    return ZSTDMT_rust_compressJobChunks(
        state->cctx, job->src.start, job->src.size,
        state->dstBuff.start, state->dstBuff.capacity, chunkSize, lastJob,
        job, ZSTDMT_compressionJobProgress);
}

static void ZSTDMT_compressionJobTrace(void* opaque)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;

    if (!job->firstJob) {
        /* Double check that we don't have an ext-dict, because then our
         * repcode invalidation doesn't work. */
        assert(!ZSTD_window_hasExtDict(state->cctx->blockState.matchState.window));
    }
    ZSTD_CCtx_trace(state->cctx, 0);
}

static void ZSTDMT_compressionJobSetError(void* opaque, size_t error)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;

    ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
    job->cSize = error;
    ZSTD_pthread_mutex_unlock(&job->job_mutex);
}

static void ZSTDMT_compressionJobFinish(void* opaque, size_t lastCBlockSize)
{
    ZSTDMT_compressionJobState* const state =
        (ZSTDMT_compressionJobState*)opaque;
    ZSTDMT_jobDescription* const job = state->job;

    ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize);
    if (job->prefix.size > 0)
        DEBUGLOG(5, "Finished with prefix: %zx", (size_t)job->prefix.start);
    DEBUGLOG(5, "Finished with source: %zx", (size_t)job->src.start);
    /* release resources */
    ZSTDMT_releaseSeq(job->seqPool, state->rawSeqStore);
    ZSTDMT_releaseCCtx(job->cctxPool, state->cctx);
    /* report */
    ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
    if (ZSTD_isError(job->cSize)) assert(lastCBlockSize == 0);
    job->cSize += lastCBlockSize;
    job->consumed = job->src.size;  /* when job->consumed == job->src.size , compression job is presumed completed */
    ZSTD_pthread_cond_signal(&job->job_cond);
    ZSTD_pthread_mutex_unlock(&job->job_mutex);
}

/* ZSTDMT_compressionJob() is a POOL_function type. Rust owns the stage
 * ordering; these callbacks retain the private job and codec operations. */
static void ZSTDMT_compressionJob(void* jobDescription)
{
    ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)jobDescription;
    ZSTDMT_compressionJobState state;
    ZSTDMT_RustCompressionJobProjection const projection = {
        job->firstJob,
        job->lastJob
    };

    ZSTD_memset(&state, 0, sizeof(state));
    state.job = job;
    state.jobParams = job->params;   /* do not modify job->params ! copy it, modify the copy */
    state.dstBuff = job->dstBuff;

    ZSTDMT_rust_compressionJob(
        &projection, &state,
        ZSTDMT_compressionJobAcquireResources,
        ZSTDMT_compressionJobPrepareParameters,
        ZSTDMT_compressionJobGenerateSequences,
        ZSTDMT_compressionJobBegin,
        ZSTDMT_compressionJobApplySequences,
        ZSTDMT_compressionJobWriteFrameHeader,
        ZSTDMT_compressionJobCompress,
        ZSTDMT_compressionJobTrace,
        ZSTDMT_compressionJobSetError,
        ZSTDMT_compressionJobFinish);
}


/* ------------------------------------------ */
/* =====   Multi-threaded compression   ===== */
/* ------------------------------------------ */

typedef struct {
    Range prefix;         /* read-only non-owned prefix buffer */
    Buffer buffer;
    size_t filled;
} InBuff_t;

typedef struct {
  BYTE* buffer;     /* The round input buffer. All jobs get references
                     * to pieces of the buffer. ZSTDMT_tryGetInputRange()
                     * handles handing out job input buffers, and makes
                     * sure it doesn't overlap with any pieces still in use.
                     */
  size_t capacity;  /* The capacity of buffer. */
  size_t pos;       /* The position of the current inBuff in the round
                     * buffer. Updated past the end if the inBuff once
                     * the inBuff is sent to the worker thread.
                     * pos <= capacity.
                     */
} RoundBuff_t;

static const RoundBuff_t kNullRoundBuff = {NULL, 0, 0};

#define RSYNC_LENGTH 32
/* Don't create chunks smaller than the zstd block size.
 * This stops us from regressing compression ratio too much,
 * and ensures our output fits in ZSTD_compressBound().
 *
 * If this is shrunk < ZSTD_BLOCKSIZELOG_MIN then
 * ZSTD_COMPRESSBOUND() will need to be updated.
 */
#define RSYNC_MIN_BLOCK_LOG ZSTD_BLOCKSIZELOG_MAX
#define RSYNC_MIN_BLOCK_SIZE (1<<RSYNC_MIN_BLOCK_LOG)

typedef struct {
  U64 hash;
  U64 hitMask;
  U64 primePower;
} RSyncState_t;

struct ZSTDMT_CCtx_s {
    POOL_ctx* factory;
    ZSTDMT_jobDescription* jobs;
    ZSTDMT_bufferPool* bufPool;
    ZSTDMT_CCtxPool* cctxPool;
    ZSTDMT_seqPool* seqPool;
    ZSTD_CCtx_params params;
    size_t targetSectionSize;
    size_t targetPrefixSize;
    int jobReady;        /* 1 => one job is already prepared, but pool has shortage of workers. Don't create a new job. */
    InBuff_t inBuff;
    RoundBuff_t roundBuff;
    SerialState serial;
    RSyncState_t rsync;
    unsigned jobIDMask;
    unsigned doneJobID;
    unsigned nextJobID;
    unsigned frameEnded;
    unsigned allJobsCompleted;
    unsigned long long frameContentSize;
    unsigned long long consumed;
    unsigned long long produced;
    ZSTD_customMem cMem;
    ZSTD_CDict* cdictLocal;
    const ZSTD_CDict* cdict;
    unsigned providedFactory: 1;
};

/* Project only the scalar job state needed by Rust's read-only MT
 * orchestration.  The descriptor layout and its mutex remain private here. */
static void ZSTDMT_projectJob(void* opaque, unsigned jobID,
                              ZSTDMT_RustJobProjection* projection)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_jobDescription* const job = &mtctx->jobs[jobID];
    ZSTD_pthread_mutex_lock(&job->job_mutex);
    projection->consumed = job->consumed;
    projection->cSize = job->cSize;
    projection->dstFlushed = job->dstFlushed;
    ZSTD_pthread_mutex_unlock(&job->job_mutex);
    projection->srcStart = job->src.start;
    projection->srcSize = job->src.size;
    projection->prefixStart = job->prefix.start;
    projection->prefixSize = job->prefix.size;
}

static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZSTD_customMem cMem)
{
    if (jobTable == NULL) return;
    ZSTDMT_job_table_destroy_sync(jobTable, nbJobs, sizeof(*jobTable));
    ZSTDMT_rust_job_table_free(jobTable, nbJobs, sizeof(*jobTable), cMem);
}

int ZSTDMT_job_table_init_sync(void* jobTable, unsigned nbJobs, size_t jobSize)
{
    U32 jobNb;
    int initError = 0;
    BYTE* const table = (BYTE*)jobTable;
    assert(jobSize == sizeof(ZSTDMT_jobDescription));
    if (jobTable == NULL) return 1;
    for (jobNb=0; jobNb<nbJobs; jobNb++) {
        ZSTDMT_jobDescription* const job =
            (ZSTDMT_jobDescription*)(table + jobNb * jobSize);
        initError |= ZSTD_pthread_mutex_init(&job->job_mutex, NULL);
        initError |= ZSTD_pthread_cond_init(&job->job_cond, NULL);
    }
    return initError;
}

void ZSTDMT_job_table_destroy_sync(void* jobTable, unsigned nbJobs, size_t jobSize)
{
    U32 jobNb;
    BYTE* const table = (BYTE*)jobTable;
    if (jobTable == NULL) return;
    assert(jobSize == sizeof(ZSTDMT_jobDescription));
    for (jobNb=0; jobNb<nbJobs; jobNb++) {
        ZSTDMT_jobDescription* const job =
            (ZSTDMT_jobDescription*)(table + jobNb * jobSize);
        ZSTD_pthread_mutex_destroy(&job->job_mutex);
        ZSTD_pthread_cond_destroy(&job->job_cond);
    }
}

/* ZSTDMT_allocJobsTable()
 * allocate and init a job table.
 * update *nbJobsPtr to next power of 2 value, as size of table */
static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
{
    ZSTDMT_jobDescription* const jobTable = (ZSTDMT_jobDescription*)
        ZSTDMT_rust_job_table_create(nbJobsPtr, sizeof(ZSTDMT_jobDescription), cMem);
    if (jobTable == NULL) return NULL;
    if (ZSTDMT_job_table_init_sync(jobTable, *nbJobsPtr, sizeof(*jobTable)) != 0) {
        ZSTDMT_freeJobsTable(jobTable, *nbJobsPtr, cMem);
        return NULL;
    }
    return jobTable;
}

static size_t ZSTDMT_expandJobsTable (ZSTDMT_CCtx* mtctx, U32 nbWorkers) {
    void* jobs = mtctx->jobs;
    U32 jobIDMask = mtctx->jobIDMask;
    size_t const error = ZSTDMT_rust_expandJobsTable(
        &jobs, &jobIDMask, nbWorkers, sizeof(ZSTDMT_jobDescription),
        mtctx->cMem, ZSTDMT_job_table_init_sync, ZSTDMT_job_table_destroy_sync);
    mtctx->jobs = (ZSTDMT_jobDescription*)jobs;
    mtctx->jobIDMask = jobIDMask;
    return error;
}


/* ZSTDMT_CCtxParam_setNbWorkers():
 * Internal use only */
static size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers)
{
    return ZSTD_CCtxParams_setParameter(params, ZSTD_c_nbWorkers, (int)nbWorkers);
}

MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool)
{
    ZSTDMT_CCtx* mtctx;
    U32 nbJobs = nbWorkers + 2;
    int initError;
    DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbWorkers = %u)", nbWorkers);

    if (nbWorkers < 1) return NULL;
    nbWorkers = MIN(nbWorkers , ZSTDMT_NBWORKERS_MAX);
    if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
        /* invalid custom allocator */
        return NULL;

    mtctx = (ZSTDMT_CCtx*) ZSTD_customCalloc(sizeof(ZSTDMT_CCtx), cMem);
    if (!mtctx) return NULL;
    ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
    mtctx->cMem = cMem;
    mtctx->allJobsCompleted = 1;
    if (pool != NULL) {
      mtctx->factory = pool;
      mtctx->providedFactory = 1;
    }
    else {
      mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem);
      mtctx->providedFactory = 0;
    }
    mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem);
    assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0);  /* ensure nbJobs is a power of 2 */
    mtctx->jobIDMask = nbJobs - 1;
    mtctx->bufPool = ZSTDMT_createBufferPool(BUF_POOL_MAX_NB_BUFFERS(nbWorkers), cMem);
    mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem);
    mtctx->seqPool = ZSTDMT_createSeqPool(nbWorkers, cMem);
    initError = ZSTDMT_serialState_init(&mtctx->serial);
    mtctx->roundBuff = kNullRoundBuff;
    if (!mtctx->factory | !mtctx->jobs | !mtctx->bufPool | !mtctx->cctxPool | !mtctx->seqPool | initError) {
        ZSTDMT_freeCCtx(mtctx);
        return NULL;
    }
    DEBUGLOG(3, "mt_cctx created, for %u threads", nbWorkers);
    return mtctx;
}

ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool)
{
#ifdef ZSTD_MULTITHREAD
    return ZSTDMT_createCCtx_advanced_internal(nbWorkers, cMem, pool);
#else
    (void)nbWorkers;
    (void)cMem;
    (void)pool;
    return NULL;
#endif
}


static void ZSTDMT_releaseJobResource(void* opaque, unsigned jobID)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    /* Copy the mutex/cond out */
    ZSTD_pthread_mutex_t const mutex = mtctx->jobs[jobID].job_mutex;
    ZSTD_pthread_cond_t const cond = mtctx->jobs[jobID].job_cond;

    DEBUGLOG(4, "job%02u: release dst address %08X", jobID, (U32)(size_t)mtctx->jobs[jobID].dstBuff.start);
    ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);

    /* Clear the job description, but keep the mutex/cond */
    ZSTD_memset(&mtctx->jobs[jobID], 0, sizeof(mtctx->jobs[jobID]));
    mtctx->jobs[jobID].job_mutex = mutex;
    mtctx->jobs[jobID].job_cond = cond;
}

/* ZSTDMT_releaseAllJobResources() :
 * note : ensure all workers are killed first ! */
static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx)
{
    DEBUGLOG(3, "ZSTDMT_releaseAllJobResources");
    ZSTDMT_rust_releaseAllJobResources(
            mtctx->jobIDMask, mtctx, ZSTDMT_releaseJobResource);
    mtctx->inBuff.buffer = g_nullBuffer;
    mtctx->inBuff.filled = 0;
    mtctx->allJobsCompleted = 1;
}

static void ZSTDMT_waitForJobComplete(
        void* opaque, unsigned jobID, unsigned doneJobID)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_jobDescription* const job = &mtctx->jobs[jobID];
    (void)doneJobID;
    ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
    while (job->consumed < job->src.size) {
        DEBUGLOG(4, "waiting for jobCompleted signal from job %u", doneJobID);
        ZSTD_pthread_cond_wait(&job->job_cond, &job->job_mutex);
    }
    ZSTD_pthread_mutex_unlock(&job->job_mutex);
}

static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* mtctx)
{
    DEBUGLOG(4, "ZSTDMT_waitForAllJobsCompleted");
    mtctx->doneJobID = ZSTDMT_rust_waitForAllJobsCompleted(
            mtctx->doneJobID, mtctx->nextJobID, mtctx->jobIDMask,
            mtctx, ZSTDMT_waitForJobComplete);
}

size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx)
{
    if (mtctx==NULL) return 0;   /* compatible with free on NULL */
    if (!mtctx->providedFactory)
        POOL_free(mtctx->factory);   /* stop and free worker threads */
    ZSTDMT_releaseAllJobResources(mtctx);  /* release job resources into pools first */
    ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
    ZSTDMT_freeBufferPool(mtctx->bufPool);
    ZSTDMT_freeCCtxPool(mtctx->cctxPool);
    ZSTDMT_freeSeqPool(mtctx->seqPool);
    ZSTDMT_serialState_free(&mtctx->serial);
    ZSTD_freeCDict(mtctx->cdictLocal);
    if (mtctx->roundBuff.buffer)
        ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem);
    ZSTD_customFree(mtctx, mtctx->cMem);
    return 0;
}

size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx)
{
    size_t mtctxSize;
    size_t factorySize;
    size_t bufferPoolSize;
    size_t jobsSize;
    size_t cctxPoolSize;
    size_t seqPoolSize;
    size_t cdictSize;
    size_t roundBuffSize;
    if (mtctx == NULL) return 0;   /* supports sizeof NULL */
    mtctxSize = sizeof(*mtctx);
    factorySize = POOL_sizeof(mtctx->factory);
    bufferPoolSize = ZSTDMT_sizeof_bufferPool(mtctx->bufPool);
    jobsSize = (mtctx->jobIDMask+1) * sizeof(ZSTDMT_jobDescription);
    cctxPoolSize = ZSTDMT_sizeof_CCtxPool(mtctx->cctxPool);
    seqPoolSize = ZSTDMT_sizeof_seqPool(mtctx->seqPool);
    cdictSize = ZSTD_sizeof_CDict(mtctx->cdictLocal);
    roundBuffSize = mtctx->roundBuff.capacity;
    return ZSTDMT_rust_sizeofCCtx(mtctxSize, factorySize, bufferPoolSize,
                                  jobsSize, cctxPoolSize, seqPoolSize,
                                  cdictSize, roundBuffSize);
}


/* ZSTDMT_resize() :
 * @return : error code if fails, 0 on success */
static size_t ZSTDMT_resize(ZSTDMT_CCtx* mtctx, unsigned nbWorkers)
{
    if (POOL_resize(mtctx->factory, nbWorkers)) return ERROR(memory_allocation);
    FORWARD_IF_ERROR( ZSTDMT_expandJobsTable(mtctx, nbWorkers) , "");
    mtctx->bufPool = ZSTDMT_expandBufferPool(mtctx->bufPool, BUF_POOL_MAX_NB_BUFFERS(nbWorkers));
    if (mtctx->bufPool == NULL) return ERROR(memory_allocation);
    mtctx->cctxPool = ZSTDMT_expandCCtxPool(mtctx->cctxPool, nbWorkers);
    if (mtctx->cctxPool == NULL) return ERROR(memory_allocation);
    mtctx->seqPool = ZSTDMT_expandSeqPool(mtctx->seqPool, nbWorkers);
    if (mtctx->seqPool == NULL) return ERROR(memory_allocation);
    ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
    return 0;
}


/*! ZSTDMT_updateCParams_whileCompressing() :
 *  Updates a selected set of compression parameters, remaining compatible with currently active frame.
 *  New parameters will be applied to next compression job. */
void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams)
{
    U32 const saved_wlog = mtctx->params.cParams.windowLog;   /* Do not modify windowLog while compressing */
    int const compressionLevel = cctxParams->compressionLevel;
    DEBUGLOG(5, "ZSTDMT_updateCParams_whileCompressing (level:%i)",
                compressionLevel);
    mtctx->params.compressionLevel = compressionLevel;
    {   ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
        cParams.windowLog = saved_wlog;
        mtctx->params.cParams = cParams;
    }
}

/* ZSTDMT_getFrameProgression():
 * tells how much data has been consumed (input) and produced (output) for current frame.
 * able to count progression inside worker threads.
 * Note : mutex will be acquired during statistics collection inside workers. */
ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx)
{
    ZSTD_frameProgression const fps = ZSTDMT_rust_frameProgressionWithJobs(
            mtctx->consumed, mtctx->inBuff.filled, mtctx->produced,
            mtctx->nextJobID, mtctx->doneJobID, mtctx->nextJobID,
            mtctx->jobReady, mtctx->jobIDMask, mtctx,
            ZSTDMT_projectJob);
    DEBUGLOG(5, "ZSTDMT_getFrameProgression");
    return fps;
}


size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx)
{
    assert(mtctx->doneJobID <= mtctx->nextJobID);
    return ZSTDMT_rust_toFlushNow(mtctx->doneJobID, mtctx->nextJobID,
                                   mtctx->jobIDMask, mtctx,
                                   ZSTDMT_projectJob);
}


/* ------------------------------------------ */
/* =====   Multi-threaded compression   ===== */
/* ------------------------------------------ */

typedef struct {
    ZSTDMT_CCtx* mtctx;
    ZSTD_CCtx_params params;
    const void* dict;
    size_t dictSize;
    ZSTD_dictContentType_e dictContentType;
    const ZSTD_CDict* cdict;
    unsigned long long pledgedSrcSize;
} ZSTDMT_initCStreamState;

static size_t ZSTDMT_initCStreamResize(void* opaque, unsigned nbWorkers)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    return ZSTDMT_resize(state->mtctx, nbWorkers);
}

static void ZSTDMT_initCStreamDrain(void* opaque)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    ZSTDMT_CCtx* const mtctx = state->mtctx;
    ZSTDMT_waitForAllJobsCompleted(mtctx);
    ZSTDMT_releaseAllJobResources(mtctx);
    mtctx->allJobsCompleted = 1;
}

static void ZSTDMT_initCStreamApplyParameters(void* opaque, size_t jobSize)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    state->params.jobSize = jobSize;
    state->mtctx->params = state->params;
    state->mtctx->frameContentSize = state->pledgedSrcSize;
}

static size_t ZSTDMT_initCStreamPrepareDictionary(void* opaque)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    ZSTDMT_CCtx* const mtctx = state->mtctx;

    ZSTD_freeCDict(mtctx->cdictLocal);
    if (state->dict) {
        mtctx->cdictLocal = ZSTD_createCDict_advanced(
            state->dict, state->dictSize, ZSTD_dlm_byCopy,
            state->dictContentType, state->params.cParams, mtctx->cMem);
        mtctx->cdict = mtctx->cdictLocal;
        if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
    } else {
        mtctx->cdictLocal = NULL;
        mtctx->cdict = state->cdict;
    }
    return 0;
}

static void ZSTDMT_initCStreamSetTargetPrefixSize(void* opaque, size_t size)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    state->mtctx->targetPrefixSize = size;
    DEBUGLOG(4, "overlapLog=%i => %u KB", state->params.overlapLog, (U32)(size >> 10));
}

static void ZSTDMT_initCStreamSetTargetSectionSize(void* opaque, size_t size)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    state->mtctx->targetSectionSize = size;
}

static void ZSTDMT_initCStreamSetRsync(void* opaque, U64 hitMask, U64 primePower)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    state->mtctx->rsync.hash = 0;
    state->mtctx->rsync.hitMask = hitMask;
    state->mtctx->rsync.primePower = primePower;
    DEBUGLOG(4, "rsyncLog = %u", ZSTD_highbit32((U32)(hitMask + 1)));
}

static void ZSTDMT_initCStreamSetBufferSize(void* opaque, size_t size)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    DEBUGLOG(4, "Job Size : %u KB (note : set to %u)",
             (U32)(state->mtctx->targetSectionSize >> 10),
             (U32)state->params.jobSize);
    DEBUGLOG(4, "inBuff Size : %u KB", (U32)(state->mtctx->targetSectionSize >> 10));
    ZSTDMT_setBufferSize(state->mtctx->bufPool, size);
}

static size_t ZSTDMT_initCStreamResizeRoundBuffer(void* opaque, size_t capacity)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    ZSTDMT_CCtx* const mtctx = state->mtctx;
    if (mtctx->roundBuff.capacity < capacity) {
        if (mtctx->roundBuff.buffer)
            ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem);
        mtctx->roundBuff.buffer = (BYTE*)ZSTD_customMalloc(capacity, mtctx->cMem);
        if (mtctx->roundBuff.buffer == NULL) {
            mtctx->roundBuff.capacity = 0;
            return ERROR(memory_allocation);
        }
        mtctx->roundBuff.capacity = capacity;
    }
    return 0;
}

static void ZSTDMT_initCStreamResetStream(void* opaque)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    ZSTDMT_CCtx* const mtctx = state->mtctx;
    DEBUGLOG(4, "roundBuff capacity : %u KB", (U32)(mtctx->roundBuff.capacity >> 10));
    mtctx->roundBuff.pos = 0;
    mtctx->inBuff.buffer = g_nullBuffer;
    mtctx->inBuff.filled = 0;
    mtctx->inBuff.prefix = kNullRange;
    mtctx->doneJobID = 0;
    mtctx->nextJobID = 0;
    mtctx->frameEnded = 0;
    mtctx->allJobsCompleted = 0;
    mtctx->consumed = 0;
    mtctx->produced = 0;
}

static size_t ZSTDMT_initCStreamUpdateDictionary(void* opaque)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    ZSTDMT_CCtx* const mtctx = state->mtctx;

    ZSTD_freeCDict(mtctx->cdictLocal);
    mtctx->cdictLocal = NULL;
    mtctx->cdict = NULL;
    if (state->dict) {
        if (state->dictContentType == ZSTD_dct_rawContent) {
            mtctx->inBuff.prefix.start = (const BYTE*)state->dict;
            mtctx->inBuff.prefix.size = state->dictSize;
        } else {
            /* note : a loadPrefix becomes an internal CDict */
            mtctx->cdictLocal = ZSTD_createCDict_advanced(
                state->dict, state->dictSize, ZSTD_dlm_byRef,
                state->dictContentType, state->params.cParams, mtctx->cMem);
            mtctx->cdict = mtctx->cdictLocal;
            if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
        }
    } else {
        mtctx->cdict = state->cdict;
    }
    return 0;
}

static size_t ZSTDMT_initCStreamSerialReset(void* opaque, size_t targetSectionSize)
{
    ZSTDMT_initCStreamState* const state = (ZSTDMT_initCStreamState*)opaque;
    if (ZSTDMT_serialState_reset(
            &state->mtctx->serial, state->mtctx->seqPool, state->params,
            targetSectionSize, state->dict, state->dictSize,
            state->dictContentType))
        return ERROR(memory_allocation);
    return 0;
}

/* ====================================== */
/* =======      Streaming API     ======= */
/* ====================================== */

size_t ZSTDMT_initCStream_internal(
        ZSTDMT_CCtx* mtctx,
        const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
        const ZSTD_CDict* cdict, ZSTD_CCtx_params params,
        unsigned long long pledgedSrcSize)
{
    ZSTDMT_initCStreamState state;
    ZSTDMT_RustInitCStreamProjection projection;

    DEBUGLOG(4, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u, nbWorkers=%u, cctxPool=%u)",
                (U32)pledgedSrcSize, params.nbWorkers, mtctx->cctxPool->totalCCtx);

    /* params supposed partially fully validated at this point */
    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */

    state = (ZSTDMT_initCStreamState){
        mtctx, params, dict, dictSize, dictContentType, cdict, pledgedSrcSize
    };
    projection = (ZSTDMT_RustInitCStreamProjection){
        (unsigned)params.nbWorkers,
        (unsigned)mtctx->params.nbWorkers,
        params.jobSize,
        ZSTDMT_JOBSIZE_MIN,
        (size_t)ZSTDMT_JOBSIZE_MAX,
        params.ldmParams.enableLdm,
        params.cParams.windowLog,
        params.cParams.chainLog,
        params.cParams.strategy,
        params.overlapLog,
        params.rsyncable,
        mtctx->roundBuff.capacity,
        mtctx->allJobsCompleted
    };

    return ZSTDMT_rust_initCStream(
        &projection, &state,
        ZSTDMT_initCStreamResize,
        ZSTDMT_initCStreamDrain,
        ZSTDMT_initCStreamApplyParameters,
        ZSTDMT_initCStreamPrepareDictionary,
        ZSTDMT_initCStreamSetTargetPrefixSize,
        ZSTDMT_initCStreamSetTargetSectionSize,
        ZSTDMT_initCStreamSetRsync,
        ZSTDMT_initCStreamSetBufferSize,
        ZSTDMT_initCStreamResizeRoundBuffer,
        ZSTDMT_initCStreamResetStream,
        ZSTDMT_initCStreamUpdateDictionary,
        ZSTDMT_initCStreamSerialReset);
}


/* ZSTDMT_writeLastEmptyBlock()
 * Write a single empty block with an end-of-frame to finish a frame.
 * Job must be created from streaming variant.
 * This function is always successful if expected conditions are fulfilled.
 */
static void ZSTDMT_writeLastEmptyBlock(ZSTDMT_jobDescription* job)
{
    ZSTDMT_RustEmptyBlockJobProjection projection;
    ZSTDMT_RustEmptyBlockResult result;

    assert(job->lastJob == 1);
    assert(job->src.size == 0);   /* last job is empty -> will be simplified into a last empty block */
    assert(job->firstJob == 0);   /* cannot be first job, as it also needs to create frame header */
    assert(job->dstBuff.start == NULL);   /* invoked from streaming variant only (otherwise, dstBuff might be user's output) */
    assert(job->consumed == 0);

    projection = (ZSTDMT_RustEmptyBlockJobProjection){
        job->lastJob,
        job->src.size,
        job->firstJob,
        job->dstBuff.start,
        job->dstBuff.capacity,
        job->consumed
    };
    result = ZSTDMT_rust_writeLastEmptyBlock(
            &projection, job->bufPool, ZSTDMT_getBufferForRust);

    job->dstBuff = (Buffer){ result.buffer.start, result.buffer.capacity };
    if (job->dstBuff.start == NULL) {
        assert(!result.clearSource);
        job->cSize = result.cSize;
        return;
    }
    assert(result.clearSource);
    assert(job->dstBuff.capacity >= ZSTD_blockHeaderSize);   /* no buffer should ever be that small */
    if (result.clearSource) job->src = kNullRange;
    job->cSize = result.cSize;
    assert(!ZSTD_isError(job->cSize));
}

static void ZSTDMT_prepareCompressionJob(
        void* opaque, unsigned jobID,
        const ZSTDMT_RustJobInitialization* initialization)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_jobDescription* const job = &mtctx->jobs[jobID];
    BYTE const* const src = (const BYTE*)initialization->srcStart;

    DEBUGLOG(5, "ZSTDMT_createCompressionJob: preparing job %u to compress %u bytes with %u preload ",
                initialization->jobNumber, (U32)initialization->srcSize,
                (U32)initialization->prefixSize);
    assert(mtctx->inBuff.filled >= initialization->srcSize);
    job->src = (Range){ src, initialization->srcSize };
    job->prefix = (Range){
        (const BYTE*)initialization->prefixStart, initialization->prefixSize
    };
    job->consumed = 0;
    job->cSize = 0;
    job->params = mtctx->params;
    job->cdict = initialization->firstJob ? mtctx->cdict : NULL;
    job->fullFrameSize = mtctx->frameContentSize;
    job->dstBuff = g_nullBuffer;
    job->cctxPool = mtctx->cctxPool;
    job->bufPool = mtctx->bufPool;
    job->seqPool = mtctx->seqPool;
    job->serial = &mtctx->serial;
    job->jobID = initialization->jobNumber;
    job->firstJob = initialization->firstJob;
    job->lastJob = initialization->lastJob;
    job->frameChecksumNeeded = initialization->frameChecksumNeeded;
    job->dstFlushed = 0;

    /* Update the round buffer position and clear the input buffer to be reset. */
    mtctx->roundBuff.pos += initialization->roundBuffPosDelta;
    mtctx->inBuff.buffer = g_nullBuffer;
    mtctx->inBuff.filled = 0;
    mtctx->inBuff.prefix = (Range){
        (const BYTE*)initialization->nextPrefixStart,
        initialization->nextPrefixSize
    };
    if (initialization->lastJob) {
        mtctx->frameEnded = 1;
        if (initialization->clearChecksumFlag)
            mtctx->params.fParams.checksumFlag = 0;
    }
}

static void ZSTDMT_writeEmptyCompressionJob(void* opaque, unsigned jobID)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_writeLastEmptyBlock(&mtctx->jobs[jobID]);
}

static int ZSTDMT_tryAddCompressionJob(void* opaque, unsigned jobID)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    return POOL_tryAdd(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[jobID]);
}

static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* mtctx, size_t srcSize, ZSTD_EndDirective endOp)
{
    ZSTDMT_RustCreateJobProjection const projection = {
        mtctx->doneJobID,
        mtctx->nextJobID,
        mtctx->jobIDMask,
        mtctx->jobReady,
        mtctx->inBuff.buffer.start,
        srcSize,
        mtctx->inBuff.filled,
        mtctx->inBuff.prefix.start,
        mtctx->inBuff.prefix.size,
        mtctx->targetPrefixSize,
        (unsigned)(endOp == ZSTD_e_end),
        (unsigned)mtctx->params.fParams.checksumFlag
    };
    ZSTDMT_RustCreateJobResult const result = ZSTDMT_rust_createCompressionJob(
        &projection, mtctx, ZSTDMT_prepareCompressionJob,
        ZSTDMT_writeEmptyCompressionJob, ZSTDMT_tryAddCompressionJob);

    if (result.action == ZSTDMT_CREATE_JOB_TABLE_FULL) {
        DEBUGLOG(5, "ZSTDMT_createCompressionJob: will not create new job : table is full");
        assert((mtctx->nextJobID & mtctx->jobIDMask) == (mtctx->doneJobID & mtctx->jobIDMask));
        return result.returnCode;
    }

    mtctx->nextJobID = result.nextJobID;
    mtctx->jobReady = result.jobReady;
    if (result.action == ZSTDMT_CREATE_JOB_EMPTY) {
        DEBUGLOG(5, "ZSTDMT_createCompressionJob: creating a last empty block to end frame");
        assert(endOp == ZSTD_e_end);  /* only possible case : need to end the frame with an empty last block */
        return result.returnCode;
    }

    DEBUGLOG(5, "ZSTDMT_createCompressionJob: posting job %u : %u bytes  (end:%u, jobNb == %u (mod:%u))",
                result.jobNumber,
                (U32)mtctx->jobs[result.jobID].src.size,
                mtctx->jobs[result.jobID].lastJob,
                result.jobNumber,
                result.jobID);
    if (result.jobReady)
        DEBUGLOG(5, "ZSTDMT_createCompressionJob: no worker available for job %u", result.jobNumber);
    return result.returnCode;
}

static size_t ZSTDMT_streamCreateJob(void* opaque, size_t srcSize, unsigned end)
{
    return ZSTDMT_createCompressionJob((ZSTDMT_CCtx*)opaque, srcSize,
                                       (ZSTD_EndDirective)end);
}


/* The Rust flush state machine receives only this synchronized scalar view.
 * The descriptor, condition variable, serial checksum state, and buffer pool
 * remain private to C. */
static void ZSTDMT_projectFlushJob(void* opaque, unsigned jobID,
                                   unsigned blockToFlush,
                                   ZSTDMT_RustFlushJobProjection* projection)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_jobDescription* const job = &mtctx->jobs[jobID];

    ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
    if (blockToFlush && (mtctx->doneJobID < mtctx->nextJobID)) {
        assert(job->dstFlushed <= job->cSize);
        while (job->dstFlushed == job->cSize) {  /* nothing to flush */
            if (job->consumed == job->src.size) break;
            ZSTD_pthread_cond_wait(&job->job_cond, &job->job_mutex);
        }
    }
    *projection = (ZSTDMT_RustFlushJobProjection){
        job->consumed,
        job->cSize,
        job->src.size,
        job->dstBuff.start,
        job->dstBuff.capacity,
        job->dstFlushed,
        job->frameChecksumNeeded
    };
    ZSTD_pthread_mutex_unlock(&job->job_mutex);
}

static void ZSTDMT_addFrameChecksum(void* opaque, unsigned jobID,
                                    ZSTDMT_RustFlushJobProjection* projection)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_jobDescription* const job = &mtctx->jobs[jobID];
    U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);

    assert(projection->frameChecksumNeeded);
    assert(projection->consumed == projection->srcSize);
    assert(projection->cSize == job->cSize);
    MEM_writeLE32((char*)job->dstBuff.start + job->cSize, checksum);
    job->cSize += 4;
    job->frameChecksumNeeded = 0;
    projection->cSize = job->cSize;
    projection->frameChecksumNeeded = 0;
}

static void ZSTDMT_updateFlushJob(void* opaque, unsigned jobID, size_t dstFlushed)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_jobDescription* const job = &mtctx->jobs[jobID];

    assert(dstFlushed <= job->cSize);
    job->dstFlushed = dstFlushed;
}

static void ZSTDMT_completeFlushJob(void* opaque, unsigned jobID,
                                    size_t srcSize, size_t cSize)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_jobDescription* const job = &mtctx->jobs[jobID];

    assert(jobID == (mtctx->doneJobID & mtctx->jobIDMask));
    assert(job->src.size == srcSize);
    assert(job->cSize == cSize);
    assert(job->dstFlushed == cSize);
    ZSTDMT_releaseBuffer(mtctx->bufPool, job->dstBuff);
    job->dstBuff = g_nullBuffer;
    job->cSize = 0;   /* ensure this job slot is considered "not started" in future check */
    mtctx->consumed += srcSize;
    mtctx->produced += cSize;
    mtctx->doneJobID++;
}

static void ZSTDMT_flushError(void* opaque)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTDMT_waitForAllJobsCompleted(mtctx);
    ZSTDMT_releaseAllJobResources(mtctx);
}

/*! ZSTDMT_flushProduced() :
 *  flush whatever data has been produced but not yet flushed in current job.
 *  move to next job if current one is fully flushed.
 * `output` : `pos` will be updated with amount of data flushed .
 * `blockToFlush` : if >0, the function will block and wait if there is no data available to flush .
 * @return : amount of data remaining within internal buffer, 0 if no more, 1 if unknown but > 0, or an error code */
static size_t ZSTDMT_flushProduced(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, unsigned blockToFlush, ZSTD_EndDirective end)
{
    ZSTDMT_RustFlushContextProjection const context = {
        mtctx->doneJobID,
        mtctx->nextJobID,
        mtctx->jobIDMask,
        mtctx->jobReady,
        mtctx->frameEnded,
        mtctx->inBuff.filled
    };
    ZSTDMT_RustFlushProducedResult const result = ZSTDMT_rust_flushProduced(
            &context,
            output->dst, output->size, output->pos,
            blockToFlush, (unsigned)end, mtctx,
            ZSTDMT_projectFlushJob,
            ZSTDMT_addFrameChecksum,
            ZSTDMT_updateFlushJob,
            ZSTDMT_completeFlushJob,
            ZSTDMT_flushError);

    assert(output->size >= output->pos);
    output->pos = result.outputPos;
    if (result.updateAllJobsCompleted)
        mtctx->allJobsCompleted = result.allJobsCompleted;
    return result.result;
}

static ZSTDMT_RustStreamFlushResult ZSTDMT_streamFlushProduced(
        void* opaque, void* outputDst, size_t outputSize, size_t outputPos,
        unsigned blockToFlush, unsigned end)
{
    ZSTD_outBuffer output = { outputDst, outputSize, outputPos };
    size_t const result = ZSTDMT_flushProduced(
            (ZSTDMT_CCtx*)opaque, &output, blockToFlush, (ZSTD_EndDirective)end);
    return (ZSTDMT_RustStreamFlushResult){ result, output.pos };
}

/**
 * Returns the range of data used by the earliest job that is not yet complete.
 * If the data of the first job is broken up into two segments, we cover both
 * sections.
 */
static Range ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx)
{
    ZSTDMT_RustInputRange const range = ZSTDMT_rust_getInputDataInUse(
            mtctx->doneJobID, mtctx->nextJobID, mtctx->jobIDMask,
            mtctx->roundBuff.capacity, mtctx->targetSectionSize,
            mtctx, ZSTDMT_projectJob);
    return (Range){ range.start, range.size };
}

static int ZSTDMT_doesOverlapWindow(Buffer buffer, ZSTD_window_t window)
{
    DEBUGLOG(5, "ZSTDMT_doesOverlapWindow");
    DEBUGLOG(5, "extDict [0x%zx, 0x%zx)",
                (size_t)window.dictBase + window.lowLimit,
                (size_t)window.dictBase + window.dictLimit);
    DEBUGLOG(5, "prefix  [0x%zx, 0x%zx)",
                (size_t)window.base + window.dictLimit,
                (size_t)window.nextSrc);

    return ZSTDMT_rust_doesOverlapWindow(buffer.start, buffer.capacity,
                                         window.nextSrc, window.base,
                                         window.dictBase, window.dictLimit,
                                         window.lowLimit);
}

static void ZSTDMT_waitForLdmLock(void* opaque)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->serial.ldmWindowMutex);
}

static int ZSTDMT_waitForLdmOverlap(
        void* opaque, void* bufferStart, size_t bufferCapacity)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    Buffer const buffer = { bufferStart, bufferCapacity };
    return ZSTDMT_doesOverlapWindow(buffer, mtctx->serial.ldmWindow);
}

static void ZSTDMT_waitForLdmWait(void* opaque)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    DEBUGLOG(5, "Waiting for LDM to finish...");
    ZSTD_pthread_cond_wait(&mtctx->serial.ldmWindowCond,
                           &mtctx->serial.ldmWindowMutex);
}

static void ZSTDMT_waitForLdmUnlock(void* opaque)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    DEBUGLOG(6, "Done waiting for LDM to finish");
    ZSTD_pthread_mutex_unlock(&mtctx->serial.ldmWindowMutex);
}

static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, Buffer buffer)
{
    int const ldmEnabled = mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable;
    ZSTDMT_RustWaitForLdmState state;
    if (ldmEnabled) {
        DEBUGLOG(5, "ZSTDMT_waitForLdmComplete");
        DEBUGLOG(5, "source  [0x%zx, 0x%zx)",
                    (size_t)buffer.start,
                    (size_t)buffer.start + buffer.capacity);
    }
    state.callbackContext = mtctx;
    state.bufferStart = buffer.start;
    state.bufferCapacity = buffer.capacity;
    state.ldmEnabled = ldmEnabled;
    state.lock = ZSTDMT_waitForLdmLock;
    state.overlaps = ZSTDMT_waitForLdmOverlap;
    state.wait = ZSTDMT_waitForLdmWait;
    state.unlock = ZSTDMT_waitForLdmUnlock;
    ZSTDMT_rust_waitForLdmComplete(&state);
}

/**
 * Attempts to set the inBuff to the next section to fill.
 * If any part of the new section is still in use we give up.
 * Returns non-zero if the buffer is filled.
 */
static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx)
{
    Range const inUse = ZSTDMT_getInputDataInUse(mtctx);
    ZSTDMT_RustTryGetInputRangeProjection const projection = {
        mtctx->roundBuff.buffer,
        mtctx->roundBuff.capacity,
        mtctx->roundBuff.pos,
        mtctx->inBuff.prefix.start,
        mtctx->inBuff.prefix.size,
        mtctx->targetSectionSize,
        inUse.start,
        inUse.size
    };
    ZSTDMT_RustTryGetInputRangeResult const result =
        ZSTDMT_rust_tryGetInputRange(&projection);
    Buffer buffer;

    DEBUGLOG(5, "ZSTDMT_tryGetInputRange");
    assert(mtctx->inBuff.buffer.start == NULL);

    if (!result.ready) {
        DEBUGLOG(5, "Waiting for buffer...");
        return 0;
    }

    if (result.movePrefix) {
        /* ZSTD_invalidateRepCodes() doesn't work for extDict variants.
         * Simply copy the prefix to the beginning in that case.
         */
        BYTE* const start = (BYTE*)mtctx->roundBuff.buffer;
        size_t const prefixSize = mtctx->inBuff.prefix.size;

        buffer.start = start;
        buffer.capacity = prefixSize;
        ZSTDMT_waitForLdmComplete(mtctx, buffer);
        ZSTD_memmove(start, mtctx->inBuff.prefix.start, prefixSize);
        mtctx->inBuff.prefix.start = start;
        mtctx->roundBuff.pos = result.roundBufferPos;
    }

    buffer.start = (BYTE*)result.bufferStart;
    buffer.capacity = result.bufferCapacity;

    ZSTDMT_waitForLdmComplete(mtctx, buffer);

    DEBUGLOG(5, "Using prefix range [%zx, %zx)",
                (size_t)mtctx->inBuff.prefix.start,
                (size_t)mtctx->inBuff.prefix.start + mtctx->inBuff.prefix.size);
    DEBUGLOG(5, "Using source range [%zx, %zx)",
                (size_t)buffer.start,
                (size_t)buffer.start + buffer.capacity);


    mtctx->inBuff.buffer = buffer;
    mtctx->inBuff.filled = 0;
    assert(mtctx->roundBuff.pos + buffer.capacity <= mtctx->roundBuff.capacity);
    return 1;
}

/**
 * Searches through the input for a synchronization point. If one is found, we
 * will instruct the caller to flush, and return the number of bytes to load.
 * Otherwise, we will load as many bytes as possible and instruct the caller
 * to continue as normal.
 */
/* Adapter callback for the C-owned reusable input range. */
static int ZSTDMT_streamTryGetInputRange(
        void* opaque, ZSTDMT_RustStreamInputRangeProjection* projection)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    int const ready = ZSTDMT_tryGetInputRange(mtctx);
    *projection = (ZSTDMT_RustStreamInputRangeProjection){
        mtctx->inBuff.buffer.start,
        mtctx->inBuff.buffer.capacity,
        mtctx->inBuff.filled
    };
    return ready;
}

static int ZSTDMT_streamLoadInput(void* opaque, const void* src, size_t size)
{
    ZSTDMT_CCtx* const mtctx = (ZSTDMT_CCtx*)opaque;
    assert(mtctx->inBuff.buffer.start != NULL);
    assert(mtctx->inBuff.filled <= mtctx->inBuff.buffer.capacity);
    assert(size <= mtctx->inBuff.buffer.capacity - mtctx->inBuff.filled);
    ZSTD_memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, src, size);
    mtctx->inBuff.filled += size;
    return 1;
}

size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx)
{
    return ZSTDMT_rust_nextInputSizeHint(mtctx->targetSectionSize,
                                         mtctx->inBuff.filled);
}

/** ZSTDMT_compressStream_generic() :
 *  internal use only - exposed to be invoked from zstd_compress.c
 *  assumption : output and input are valid (pos <= size)
 * @return : minimum amount of data remaining to flush, 0 if none */
size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
                                     ZSTD_outBuffer* output,
                                     ZSTD_inBuffer* input,
                                     ZSTD_EndDirective endOp)
{
    ZSTDMT_RustCompressStreamContextProjection const context = {
        mtctx->frameEnded,
        (unsigned)mtctx->jobReady,
        mtctx->inBuff.buffer.start,
        mtctx->inBuff.buffer.capacity,
        mtctx->inBuff.filled,
        mtctx->targetSectionSize,
        mtctx->params.rsyncable,
        mtctx->rsync.primePower,
        mtctx->rsync.hitMask
    };
    ZSTDMT_RustStreamInputProjection const inputProjection = {
        input->src,
        input->size,
        input->pos
    };
    ZSTDMT_RustStreamOutputProjection const outputProjection = {
        output->dst,
        output->size,
        output->pos
    };
    ZSTDMT_RustCompressStreamResult const result =
        ZSTDMT_rust_compressStreamGeneric(
            &context, &inputProjection, &outputProjection, (unsigned)endOp,
            mtctx, ZSTDMT_streamTryGetInputRange, ZSTDMT_streamLoadInput,
            ZSTDMT_streamCreateJob, ZSTDMT_streamFlushProduced);

    DEBUGLOG(5, "ZSTDMT_compressStream_generic (endOp=%u, srcSize=%u)",
                (U32)endOp, (U32)(input->size - input->pos));
    assert(output->pos <= output->size);
    assert(input->pos  <= input->size);
    input->pos = result.inputPos;
    output->pos = result.outputPos;
    DEBUGLOG(5, "end of ZSTDMT_compressStream_generic: remainingToFlush = %u", (U32)result.result);
    return result.result;
}