2 * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
5 * This source code is licensed under both the BSD-style license (found in the
6 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
7 * in the COPYING file in the root directory of this source tree).
8 * You may select, at your option, one of the above-listed licenses.
12 /* ====== Tuning parameters ====== */
13 #define ZSTDMT_NBWORKERS_MAX 200
14 #define ZSTDMT_JOBSIZE_MAX (MEM_32bits() ? (512 MB) : (2 GB)) /* note : limited by `jobSize` type, which is `unsigned` */
15 #define ZSTDMT_OVERLAPLOG_DEFAULT 6
18 /* ====== Compiler specifics ====== */
20 # pragma warning(disable : 4204) /* disable: C4204: non-constant aggregate initializer */
24 /* ====== Dependencies ====== */
25 #include <string.h> /* memcpy, memset */
26 #include <limits.h> /* INT_MAX */
27 #include "pool.h" /* threadpool */
28 #include "threading.h" /* mutex */
29 #include "zstd_compress_internal.h" /* MIN, ERROR, ZSTD_*, ZSTD_highbit32 */
31 #include "zstdmt_compress.h"
33 /* Guards code to support resizing the SeqPool.
34 * We will want to resize the SeqPool to save memory in the future.
35 * Until then, comment the code out since it is unused.
37 #define ZSTD_RESIZE_SEQPOOL 0
39 /* ====== Debug ====== */
40 #if defined(DEBUGLEVEL) && (DEBUGLEVEL>=2) \
41 && !defined(_MSC_VER) \
42 && !defined(__MINGW32__)
46 # include <sys/times.h>
48 # define DEBUG_PRINTHEX(l,p,n) { \
50 for (debug_u=0; debug_u<(n); debug_u++) \
51 RAWLOG(l, "%02X ", ((const unsigned char*)(p))[debug_u]); \
55 static unsigned long long GetCurrentClockTimeMicroseconds(void)
57 static clock_t _ticksPerSecond = 0;
58 if (_ticksPerSecond <= 0) _ticksPerSecond = sysconf(_SC_CLK_TCK);
60 { struct tms junk; clock_t newTicks = (clock_t) times(&junk);
61 return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond); }
64 #define MUTEX_WAIT_TIME_DLEVEL 6
65 #define ZSTD_PTHREAD_MUTEX_LOCK(mutex) { \
66 if (DEBUGLEVEL >= MUTEX_WAIT_TIME_DLEVEL) { \
67 unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds(); \
68 ZSTD_pthread_mutex_lock(mutex); \
69 { unsigned long long const afterTime = GetCurrentClockTimeMicroseconds(); \
70 unsigned long long const elapsedTime = (afterTime-beforeTime); \
71 if (elapsedTime > 1000) { /* or whatever threshold you like; I'm using 1 millisecond here */ \
72 DEBUGLOG(MUTEX_WAIT_TIME_DLEVEL, "Thread took %llu microseconds to acquire mutex %s \n", \
73 elapsedTime, #mutex); \
76 ZSTD_pthread_mutex_lock(mutex); \
82 # define ZSTD_PTHREAD_MUTEX_LOCK(m) ZSTD_pthread_mutex_lock(m)
83 # define DEBUG_PRINTHEX(l,p,n) {}
88 /* ===== Buffer Pool ===== */
89 /* a single Buffer Pool can be invoked from multiple threads in parallel */
91 typedef struct buffer_s {
96 static const buffer_t g_nullBuffer = { NULL, 0 };
98 typedef struct ZSTDMT_bufferPool_s {
99 ZSTD_pthread_mutex_t poolMutex;
101 unsigned totalBuffers;
104 buffer_t bTable[1]; /* variable size */
107 static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbWorkers, ZSTD_customMem cMem)
109 unsigned const maxNbBuffers = 2*nbWorkers + 3;
110 ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)ZSTD_calloc(
111 sizeof(ZSTDMT_bufferPool) + (maxNbBuffers-1) * sizeof(buffer_t), cMem);
112 if (bufPool==NULL) return NULL;
113 if (ZSTD_pthread_mutex_init(&bufPool->poolMutex, NULL)) {
114 ZSTD_free(bufPool, cMem);
117 bufPool->bufferSize = 64 KB;
118 bufPool->totalBuffers = maxNbBuffers;
119 bufPool->nbBuffers = 0;
120 bufPool->cMem = cMem;
124 static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool)
127 DEBUGLOG(3, "ZSTDMT_freeBufferPool (address:%08X)", (U32)(size_t)bufPool);
128 if (!bufPool) return; /* compatibility with free on NULL */
129 for (u=0; u<bufPool->totalBuffers; u++) {
130 DEBUGLOG(4, "free buffer %2u (address:%08X)", u, (U32)(size_t)bufPool->bTable[u].start);
131 ZSTD_free(bufPool->bTable[u].start, bufPool->cMem);
133 ZSTD_pthread_mutex_destroy(&bufPool->poolMutex);
134 ZSTD_free(bufPool, bufPool->cMem);
137 /* only works at initialization, not during compression */
138 static size_t ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool* bufPool)
140 size_t const poolSize = sizeof(*bufPool)
141 + (bufPool->totalBuffers - 1) * sizeof(buffer_t);
143 size_t totalBufferSize = 0;
144 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
145 for (u=0; u<bufPool->totalBuffers; u++)
146 totalBufferSize += bufPool->bTable[u].capacity;
147 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
149 return poolSize + totalBufferSize;
152 /* ZSTDMT_setBufferSize() :
153 * all future buffers provided by this buffer pool will have _at least_ this size
154 * note : it's better for all buffers to have same size,
155 * as they become freely interchangeable, reducing malloc/free usages and memory fragmentation */
156 static void ZSTDMT_setBufferSize(ZSTDMT_bufferPool* const bufPool, size_t const bSize)
158 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
159 DEBUGLOG(4, "ZSTDMT_setBufferSize: bSize = %u", (U32)bSize);
160 bufPool->bufferSize = bSize;
161 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
165 static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, U32 nbWorkers)
167 unsigned const maxNbBuffers = 2*nbWorkers + 3;
168 if (srcBufPool==NULL) return NULL;
169 if (srcBufPool->totalBuffers >= maxNbBuffers) /* good enough */
171 /* need a larger buffer pool */
172 { ZSTD_customMem const cMem = srcBufPool->cMem;
173 size_t const bSize = srcBufPool->bufferSize; /* forward parameters */
174 ZSTDMT_bufferPool* newBufPool;
175 ZSTDMT_freeBufferPool(srcBufPool);
176 newBufPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
177 if (newBufPool==NULL) return newBufPool;
178 ZSTDMT_setBufferSize(newBufPool, bSize);
183 /** ZSTDMT_getBuffer() :
184 * assumption : bufPool must be valid
185 * @return : a buffer, with start pointer and size
186 * note: allocation may fail, in this case, start==NULL and size==0 */
187 static buffer_t ZSTDMT_getBuffer(ZSTDMT_bufferPool* bufPool)
189 size_t const bSize = bufPool->bufferSize;
190 DEBUGLOG(5, "ZSTDMT_getBuffer: bSize = %u", (U32)bufPool->bufferSize);
191 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
192 if (bufPool->nbBuffers) { /* try to use an existing buffer */
193 buffer_t const buf = bufPool->bTable[--(bufPool->nbBuffers)];
194 size_t const availBufferSize = buf.capacity;
195 bufPool->bTable[bufPool->nbBuffers] = g_nullBuffer;
196 if ((availBufferSize >= bSize) & ((availBufferSize>>3) <= bSize)) {
197 /* large enough, but not too much */
198 DEBUGLOG(5, "ZSTDMT_getBuffer: provide buffer %u of size %u",
199 bufPool->nbBuffers, (U32)buf.capacity);
200 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
203 /* size conditions not respected : scratch this buffer, create new one */
204 DEBUGLOG(5, "ZSTDMT_getBuffer: existing buffer does not meet size conditions => freeing");
205 ZSTD_free(buf.start, bufPool->cMem);
207 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
208 /* create new buffer */
209 DEBUGLOG(5, "ZSTDMT_getBuffer: create a new buffer");
211 void* const start = ZSTD_malloc(bSize, bufPool->cMem);
212 buffer.start = start; /* note : start can be NULL if malloc fails ! */
213 buffer.capacity = (start==NULL) ? 0 : bSize;
215 DEBUGLOG(5, "ZSTDMT_getBuffer: buffer allocation failure !!");
217 DEBUGLOG(5, "ZSTDMT_getBuffer: created buffer of size %u", (U32)bSize);
223 #if ZSTD_RESIZE_SEQPOOL
224 /** ZSTDMT_resizeBuffer() :
225 * assumption : bufPool must be valid
226 * @return : a buffer that is at least the buffer pool buffer size.
227 * If a reallocation happens, the data in the input buffer is copied.
229 static buffer_t ZSTDMT_resizeBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buffer)
231 size_t const bSize = bufPool->bufferSize;
232 if (buffer.capacity < bSize) {
233 void* const start = ZSTD_malloc(bSize, bufPool->cMem);
235 newBuffer.start = start;
236 newBuffer.capacity = start == NULL ? 0 : bSize;
238 assert(newBuffer.capacity >= buffer.capacity);
239 memcpy(newBuffer.start, buffer.start, buffer.capacity);
240 DEBUGLOG(5, "ZSTDMT_resizeBuffer: created buffer of size %u", (U32)bSize);
243 DEBUGLOG(5, "ZSTDMT_resizeBuffer: buffer allocation failure !!");
249 /* store buffer for later re-use, up to pool capacity */
250 static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buf)
252 DEBUGLOG(5, "ZSTDMT_releaseBuffer");
253 if (buf.start == NULL) return; /* compatible with release on NULL */
254 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
255 if (bufPool->nbBuffers < bufPool->totalBuffers) {
256 bufPool->bTable[bufPool->nbBuffers++] = buf; /* stored for later use */
257 DEBUGLOG(5, "ZSTDMT_releaseBuffer: stored buffer of size %u in slot %u",
258 (U32)buf.capacity, (U32)(bufPool->nbBuffers-1));
259 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
262 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
263 /* Reached bufferPool capacity (should not happen) */
264 DEBUGLOG(5, "ZSTDMT_releaseBuffer: pool capacity reached => freeing ");
265 ZSTD_free(buf.start, bufPool->cMem);
269 /* ===== Seq Pool Wrapper ====== */
271 static rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0};
273 typedef ZSTDMT_bufferPool ZSTDMT_seqPool;
275 static size_t ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool* seqPool)
277 return ZSTDMT_sizeof_bufferPool(seqPool);
280 static rawSeqStore_t bufferToSeq(buffer_t buffer)
282 rawSeqStore_t seq = {NULL, 0, 0, 0};
283 seq.seq = (rawSeq*)buffer.start;
284 seq.capacity = buffer.capacity / sizeof(rawSeq);
288 static buffer_t seqToBuffer(rawSeqStore_t seq)
291 buffer.start = seq.seq;
292 buffer.capacity = seq.capacity * sizeof(rawSeq);
296 static rawSeqStore_t ZSTDMT_getSeq(ZSTDMT_seqPool* seqPool)
298 if (seqPool->bufferSize == 0) {
299 return kNullRawSeqStore;
301 return bufferToSeq(ZSTDMT_getBuffer(seqPool));
304 #if ZSTD_RESIZE_SEQPOOL
305 static rawSeqStore_t ZSTDMT_resizeSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
307 return bufferToSeq(ZSTDMT_resizeBuffer(seqPool, seqToBuffer(seq)));
311 static void ZSTDMT_releaseSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
313 ZSTDMT_releaseBuffer(seqPool, seqToBuffer(seq));
316 static void ZSTDMT_setNbSeq(ZSTDMT_seqPool* const seqPool, size_t const nbSeq)
318 ZSTDMT_setBufferSize(seqPool, nbSeq * sizeof(rawSeq));
321 static ZSTDMT_seqPool* ZSTDMT_createSeqPool(unsigned nbWorkers, ZSTD_customMem cMem)
323 ZSTDMT_seqPool* const seqPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
324 if (seqPool == NULL) return NULL;
325 ZSTDMT_setNbSeq(seqPool, 0);
329 static void ZSTDMT_freeSeqPool(ZSTDMT_seqPool* seqPool)
331 ZSTDMT_freeBufferPool(seqPool);
334 static ZSTDMT_seqPool* ZSTDMT_expandSeqPool(ZSTDMT_seqPool* pool, U32 nbWorkers)
336 return ZSTDMT_expandBufferPool(pool, nbWorkers);
340 /* ===== CCtx Pool ===== */
341 /* a single CCtx Pool can be invoked from multiple threads in parallel */
344 ZSTD_pthread_mutex_t poolMutex;
348 ZSTD_CCtx* cctx[1]; /* variable size */
351 /* note : all CCtx borrowed from the pool should be released back to the pool _before_ freeing the pool */
352 static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
355 for (u=0; u<pool->totalCCtx; u++)
356 ZSTD_freeCCtx(pool->cctx[u]); /* note : compatible with free on NULL */
357 ZSTD_pthread_mutex_destroy(&pool->poolMutex);
358 ZSTD_free(pool, pool->cMem);
361 /* ZSTDMT_createCCtxPool() :
362 * implies nbWorkers >= 1 , checked by caller ZSTDMT_createCCtx() */
363 static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(unsigned nbWorkers,
366 ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_calloc(
367 sizeof(ZSTDMT_CCtxPool) + (nbWorkers-1)*sizeof(ZSTD_CCtx*), cMem);
368 assert(nbWorkers > 0);
369 if (!cctxPool) return NULL;
370 if (ZSTD_pthread_mutex_init(&cctxPool->poolMutex, NULL)) {
371 ZSTD_free(cctxPool, cMem);
374 cctxPool->cMem = cMem;
375 cctxPool->totalCCtx = nbWorkers;
376 cctxPool->availCCtx = 1; /* at least one cctx for single-thread mode */
377 cctxPool->cctx[0] = ZSTD_createCCtx_advanced(cMem);
378 if (!cctxPool->cctx[0]) { ZSTDMT_freeCCtxPool(cctxPool); return NULL; }
379 DEBUGLOG(3, "cctxPool created, with %u workers", nbWorkers);
383 static ZSTDMT_CCtxPool* ZSTDMT_expandCCtxPool(ZSTDMT_CCtxPool* srcPool,
386 if (srcPool==NULL) return NULL;
387 if (nbWorkers <= srcPool->totalCCtx) return srcPool; /* good enough */
388 /* need a larger cctx pool */
389 { ZSTD_customMem const cMem = srcPool->cMem;
390 ZSTDMT_freeCCtxPool(srcPool);
391 return ZSTDMT_createCCtxPool(nbWorkers, cMem);
395 /* only works during initialization phase, not during compression */
396 static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
398 ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
399 { unsigned const nbWorkers = cctxPool->totalCCtx;
400 size_t const poolSize = sizeof(*cctxPool)
401 + (nbWorkers-1) * sizeof(ZSTD_CCtx*);
403 size_t totalCCtxSize = 0;
404 for (u=0; u<nbWorkers; u++) {
405 totalCCtxSize += ZSTD_sizeof_CCtx(cctxPool->cctx[u]);
407 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
408 assert(nbWorkers > 0);
409 return poolSize + totalCCtxSize;
413 static ZSTD_CCtx* ZSTDMT_getCCtx(ZSTDMT_CCtxPool* cctxPool)
415 DEBUGLOG(5, "ZSTDMT_getCCtx");
416 ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
417 if (cctxPool->availCCtx) {
418 cctxPool->availCCtx--;
419 { ZSTD_CCtx* const cctx = cctxPool->cctx[cctxPool->availCCtx];
420 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
423 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
424 DEBUGLOG(5, "create one more CCtx");
425 return ZSTD_createCCtx_advanced(cctxPool->cMem); /* note : can be NULL, when creation fails ! */
428 static void ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool* pool, ZSTD_CCtx* cctx)
430 if (cctx==NULL) return; /* compatibility with release on NULL */
431 ZSTD_pthread_mutex_lock(&pool->poolMutex);
432 if (pool->availCCtx < pool->totalCCtx)
433 pool->cctx[pool->availCCtx++] = cctx;
435 /* pool overflow : should not happen, since totalCCtx==nbWorkers */
436 DEBUGLOG(4, "CCtx pool overflow : free cctx");
439 ZSTD_pthread_mutex_unlock(&pool->poolMutex);
442 /* ==== Serial State ==== */
450 /* All variables in the struct are protected by mutex. */
451 ZSTD_pthread_mutex_t mutex;
452 ZSTD_pthread_cond_t cond;
453 ZSTD_CCtx_params params;
455 XXH64_state_t xxhState;
457 /* Protects ldmWindow.
458 * Must be acquired after the main mutex when acquiring both.
460 ZSTD_pthread_mutex_t ldmWindowMutex;
461 ZSTD_pthread_cond_t ldmWindowCond; /* Signaled when ldmWindow is udpated */
462 ZSTD_window_t ldmWindow; /* A thread-safe copy of ldmState.window */
465 static int ZSTDMT_serialState_reset(serialState_t* serialState, ZSTDMT_seqPool* seqPool, ZSTD_CCtx_params params, size_t jobSize)
467 /* Adjust parameters */
468 if (params.ldmParams.enableLdm) {
469 DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10);
470 ZSTD_ldm_adjustParameters(¶ms.ldmParams, ¶ms.cParams);
471 assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
472 assert(params.ldmParams.hashEveryLog < 32);
473 serialState->ldmState.hashPower =
474 ZSTD_ldm_getHashPower(params.ldmParams.minMatchLength);
476 memset(¶ms.ldmParams, 0, sizeof(params.ldmParams));
478 serialState->nextJobID = 0;
479 if (params.fParams.checksumFlag)
480 XXH64_reset(&serialState->xxhState, 0);
481 if (params.ldmParams.enableLdm) {
482 ZSTD_customMem cMem = params.customMem;
483 unsigned const hashLog = params.ldmParams.hashLog;
484 size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
485 unsigned const bucketLog =
486 params.ldmParams.hashLog - params.ldmParams.bucketSizeLog;
487 size_t const bucketSize = (size_t)1 << bucketLog;
488 unsigned const prevBucketLog =
489 serialState->params.ldmParams.hashLog -
490 serialState->params.ldmParams.bucketSizeLog;
491 /* Size the seq pool tables */
492 ZSTDMT_setNbSeq(seqPool, ZSTD_ldm_getMaxNbSeq(params.ldmParams, jobSize));
493 /* Reset the window */
494 ZSTD_window_clear(&serialState->ldmState.window);
495 serialState->ldmWindow = serialState->ldmState.window;
496 /* Resize tables and output space if necessary. */
497 if (serialState->ldmState.hashTable == NULL || serialState->params.ldmParams.hashLog < hashLog) {
498 ZSTD_free(serialState->ldmState.hashTable, cMem);
499 serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_malloc(hashSize, cMem);
501 if (serialState->ldmState.bucketOffsets == NULL || prevBucketLog < bucketLog) {
502 ZSTD_free(serialState->ldmState.bucketOffsets, cMem);
503 serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_malloc(bucketSize, cMem);
505 if (!serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets)
507 /* Zero the tables */
508 memset(serialState->ldmState.hashTable, 0, hashSize);
509 memset(serialState->ldmState.bucketOffsets, 0, bucketSize);
511 serialState->params = params;
512 serialState->params.jobSize = (U32)jobSize;
516 static int ZSTDMT_serialState_init(serialState_t* serialState)
519 memset(serialState, 0, sizeof(*serialState));
520 initError |= ZSTD_pthread_mutex_init(&serialState->mutex, NULL);
521 initError |= ZSTD_pthread_cond_init(&serialState->cond, NULL);
522 initError |= ZSTD_pthread_mutex_init(&serialState->ldmWindowMutex, NULL);
523 initError |= ZSTD_pthread_cond_init(&serialState->ldmWindowCond, NULL);
527 static void ZSTDMT_serialState_free(serialState_t* serialState)
529 ZSTD_customMem cMem = serialState->params.customMem;
530 ZSTD_pthread_mutex_destroy(&serialState->mutex);
531 ZSTD_pthread_cond_destroy(&serialState->cond);
532 ZSTD_pthread_mutex_destroy(&serialState->ldmWindowMutex);
533 ZSTD_pthread_cond_destroy(&serialState->ldmWindowCond);
534 ZSTD_free(serialState->ldmState.hashTable, cMem);
535 ZSTD_free(serialState->ldmState.bucketOffsets, cMem);
538 static void ZSTDMT_serialState_update(serialState_t* serialState,
539 ZSTD_CCtx* jobCCtx, rawSeqStore_t seqStore,
540 range_t src, unsigned jobID)
542 /* Wait for our turn */
543 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
544 while (serialState->nextJobID < jobID) {
545 DEBUGLOG(5, "wait for serialState->cond");
546 ZSTD_pthread_cond_wait(&serialState->cond, &serialState->mutex);
548 /* A future job may error and skip our job */
549 if (serialState->nextJobID == jobID) {
550 /* It is now our turn, do any processing necessary */
551 if (serialState->params.ldmParams.enableLdm) {
553 assert(seqStore.seq != NULL && seqStore.pos == 0 &&
554 seqStore.size == 0 && seqStore.capacity > 0);
555 assert(src.size <= serialState->params.jobSize);
556 ZSTD_window_update(&serialState->ldmState.window, src.start, src.size);
557 error = ZSTD_ldm_generateSequences(
558 &serialState->ldmState, &seqStore,
559 &serialState->params.ldmParams, src.start, src.size);
560 /* We provide a large enough buffer to never fail. */
561 assert(!ZSTD_isError(error)); (void)error;
562 /* Update ldmWindow to match the ldmState.window and signal the main
563 * thread if it is waiting for a buffer.
565 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
566 serialState->ldmWindow = serialState->ldmState.window;
567 ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
568 ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
570 if (serialState->params.fParams.checksumFlag && src.size > 0)
571 XXH64_update(&serialState->xxhState, src.start, src.size);
573 /* Now it is the next jobs turn */
574 serialState->nextJobID++;
575 ZSTD_pthread_cond_broadcast(&serialState->cond);
576 ZSTD_pthread_mutex_unlock(&serialState->mutex);
578 if (seqStore.size > 0) {
579 size_t const err = ZSTD_referenceExternalSequences(
580 jobCCtx, seqStore.seq, seqStore.size);
581 assert(serialState->params.ldmParams.enableLdm);
582 assert(!ZSTD_isError(err));
587 static void ZSTDMT_serialState_ensureFinished(serialState_t* serialState,
588 unsigned jobID, size_t cSize)
590 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
591 if (serialState->nextJobID <= jobID) {
592 assert(ZSTD_isError(cSize)); (void)cSize;
593 DEBUGLOG(5, "Skipping past job %u because of error", jobID);
594 serialState->nextJobID = jobID + 1;
595 ZSTD_pthread_cond_broadcast(&serialState->cond);
597 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
598 ZSTD_window_clear(&serialState->ldmWindow);
599 ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
600 ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
602 ZSTD_pthread_mutex_unlock(&serialState->mutex);
607 /* ------------------------------------------ */
608 /* ===== Worker thread ===== */
609 /* ------------------------------------------ */
611 static const range_t kNullRange = { NULL, 0 };
614 size_t consumed; /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx */
615 size_t cSize; /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx, then set0 by mtctx */
616 ZSTD_pthread_mutex_t job_mutex; /* Thread-safe - used by mtctx and worker */
617 ZSTD_pthread_cond_t job_cond; /* Thread-safe - used by mtctx and worker */
618 ZSTDMT_CCtxPool* cctxPool; /* Thread-safe - used by mtctx and (all) workers */
619 ZSTDMT_bufferPool* bufPool; /* Thread-safe - used by mtctx and (all) workers */
620 ZSTDMT_seqPool* seqPool; /* Thread-safe - used by mtctx and (all) workers */
621 serialState_t* serial; /* Thread-safe - used by mtctx and (all) workers */
622 buffer_t dstBuff; /* set by worker (or mtctx), then read by worker & mtctx, then modified by mtctx => no barrier */
623 range_t prefix; /* set by mtctx, then read by worker & mtctx => no barrier */
624 range_t src; /* set by mtctx, then read by worker & mtctx => no barrier */
625 unsigned jobID; /* set by mtctx, then read by worker => no barrier */
626 unsigned firstJob; /* set by mtctx, then read by worker => no barrier */
627 unsigned lastJob; /* set by mtctx, then read by worker => no barrier */
628 ZSTD_CCtx_params params; /* set by mtctx, then read by worker => no barrier */
629 const ZSTD_CDict* cdict; /* set by mtctx, then read by worker => no barrier */
630 unsigned long long fullFrameSize; /* set by mtctx, then read by worker => no barrier */
631 size_t dstFlushed; /* used only by mtctx */
632 unsigned frameChecksumNeeded; /* used only by mtctx */
633 } ZSTDMT_jobDescription;
635 #define JOB_ERROR(e) { \
636 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex); \
638 ZSTD_pthread_mutex_unlock(&job->job_mutex); \
642 /* ZSTDMT_compressionJob() is a POOL_function type */
643 static void ZSTDMT_compressionJob(void* jobDescription)
645 ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)jobDescription;
646 ZSTD_CCtx_params jobParams = job->params; /* do not modify job->params ! copy it, modify the copy */
647 ZSTD_CCtx* const cctx = ZSTDMT_getCCtx(job->cctxPool);
648 rawSeqStore_t rawSeqStore = ZSTDMT_getSeq(job->seqPool);
649 buffer_t dstBuff = job->dstBuff;
650 size_t lastCBlockSize = 0;
653 if (cctx==NULL) JOB_ERROR(ERROR(memory_allocation));
654 if (dstBuff.start == NULL) { /* streaming job : doesn't provide a dstBuffer */
655 dstBuff = ZSTDMT_getBuffer(job->bufPool);
656 if (dstBuff.start==NULL) JOB_ERROR(ERROR(memory_allocation));
657 job->dstBuff = dstBuff; /* this value can be read in ZSTDMT_flush, when it copies the whole job */
659 if (jobParams.ldmParams.enableLdm && rawSeqStore.seq == NULL)
660 JOB_ERROR(ERROR(memory_allocation));
662 /* Don't compute the checksum for chunks, since we compute it externally,
663 * but write it in the header.
665 if (job->jobID != 0) jobParams.fParams.checksumFlag = 0;
666 /* Don't run LDM for the chunks, since we handle it externally */
667 jobParams.ldmParams.enableLdm = 0;
672 size_t const initError = ZSTD_compressBegin_advanced_internal(cctx, NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast, job->cdict, jobParams, job->fullFrameSize);
673 assert(job->firstJob); /* only allowed for first job */
674 if (ZSTD_isError(initError)) JOB_ERROR(initError);
675 } else { /* srcStart points at reloaded section */
676 U64 const pledgedSrcSize = job->firstJob ? job->fullFrameSize : job->src.size;
677 { size_t const forceWindowError = ZSTD_CCtxParam_setParameter(&jobParams, ZSTD_p_forceMaxWindow, !job->firstJob);
678 if (ZSTD_isError(forceWindowError)) JOB_ERROR(forceWindowError);
680 { size_t const initError = ZSTD_compressBegin_advanced_internal(cctx,
681 job->prefix.start, job->prefix.size, ZSTD_dct_rawContent, /* load dictionary in "content-only" mode (no header analysis) */
684 jobParams, pledgedSrcSize);
685 if (ZSTD_isError(initError)) JOB_ERROR(initError);
688 /* Perform serial step as early as possible, but after CCtx initialization */
689 ZSTDMT_serialState_update(job->serial, cctx, rawSeqStore, job->src, job->jobID);
691 if (!job->firstJob) { /* flush and overwrite frame header when it's not first job */
692 size_t const hSize = ZSTD_compressContinue(cctx, dstBuff.start, dstBuff.capacity, job->src.start, 0);
693 if (ZSTD_isError(hSize)) JOB_ERROR(hSize);
694 DEBUGLOG(5, "ZSTDMT_compressionJob: flush and overwrite %u bytes of frame header (not first job)", (U32)hSize);
695 ZSTD_invalidateRepCodes(cctx);
699 { size_t const chunkSize = 4*ZSTD_BLOCKSIZE_MAX;
700 int const nbChunks = (int)((job->src.size + (chunkSize-1)) / chunkSize);
701 const BYTE* ip = (const BYTE*) job->src.start;
702 BYTE* const ostart = (BYTE*)dstBuff.start;
704 BYTE* oend = op + dstBuff.capacity;
706 if (sizeof(size_t) > sizeof(int)) assert(job->src.size < ((size_t)INT_MAX) * chunkSize); /* check overflow */
707 DEBUGLOG(5, "ZSTDMT_compressionJob: compress %u bytes in %i blocks", (U32)job->src.size, nbChunks);
708 assert(job->cSize == 0);
709 for (chunkNb = 1; chunkNb < nbChunks; chunkNb++) {
710 size_t const cSize = ZSTD_compressContinue(cctx, op, oend-op, ip, chunkSize);
711 if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
713 op += cSize; assert(op < oend);
715 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
717 job->consumed = chunkSize * chunkNb;
718 DEBUGLOG(5, "ZSTDMT_compressionJob: compress new block : cSize==%u bytes (total: %u)",
719 (U32)cSize, (U32)job->cSize);
720 ZSTD_pthread_cond_signal(&job->job_cond); /* warns some more data is ready to be flushed */
721 ZSTD_pthread_mutex_unlock(&job->job_mutex);
724 assert(chunkSize > 0);
725 assert((chunkSize & (chunkSize - 1)) == 0); /* chunkSize must be power of 2 for mask==(chunkSize-1) to work */
726 if ((nbChunks > 0) | job->lastJob /*must output a "last block" flag*/ ) {
727 size_t const lastBlockSize1 = job->src.size & (chunkSize-1);
728 size_t const lastBlockSize = ((lastBlockSize1==0) & (job->src.size>=chunkSize)) ? chunkSize : lastBlockSize1;
729 size_t const cSize = (job->lastJob) ?
730 ZSTD_compressEnd (cctx, op, oend-op, ip, lastBlockSize) :
731 ZSTD_compressContinue(cctx, op, oend-op, ip, lastBlockSize);
732 if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
733 lastCBlockSize = cSize;
737 ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize);
738 if (job->prefix.size > 0)
739 DEBUGLOG(5, "Finished with prefix: %zx", (size_t)job->prefix.start);
740 DEBUGLOG(5, "Finished with source: %zx", (size_t)job->src.start);
741 /* release resources */
742 ZSTDMT_releaseSeq(job->seqPool, rawSeqStore);
743 ZSTDMT_releaseCCtx(job->cctxPool, cctx);
745 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
746 if (ZSTD_isError(job->cSize)) assert(lastCBlockSize == 0);
747 job->cSize += lastCBlockSize;
748 job->consumed = job->src.size; /* when job->consumed == job->src.size , compression job is presumed completed */
749 ZSTD_pthread_cond_signal(&job->job_cond);
750 ZSTD_pthread_mutex_unlock(&job->job_mutex);
754 /* ------------------------------------------ */
755 /* ===== Multi-threaded compression ===== */
756 /* ------------------------------------------ */
759 range_t prefix; /* read-only non-owned prefix buffer */
765 BYTE* buffer; /* The round input buffer. All jobs get references
766 * to pieces of the buffer. ZSTDMT_tryGetInputRange()
767 * handles handing out job input buffers, and makes
768 * sure it doesn't overlap with any pieces still in use.
770 size_t capacity; /* The capacity of buffer. */
771 size_t pos; /* The position of the current inBuff in the round
772 * buffer. Updated past the end if the inBuff once
773 * the inBuff is sent to the worker thread.
778 static const roundBuff_t kNullRoundBuff = {NULL, 0, 0};
780 struct ZSTDMT_CCtx_s {
782 ZSTDMT_jobDescription* jobs;
783 ZSTDMT_bufferPool* bufPool;
784 ZSTDMT_CCtxPool* cctxPool;
785 ZSTDMT_seqPool* seqPool;
786 ZSTD_CCtx_params params;
787 size_t targetSectionSize;
788 size_t targetPrefixSize;
789 int jobReady; /* 1 => one job is already prepared, but pool has shortage of workers. Don't create a new job. */
791 roundBuff_t roundBuff;
792 serialState_t serial;
793 unsigned singleBlockingThread;
798 unsigned allJobsCompleted;
799 unsigned long long frameContentSize;
800 unsigned long long consumed;
801 unsigned long long produced;
803 ZSTD_CDict* cdictLocal;
804 const ZSTD_CDict* cdict;
807 static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZSTD_customMem cMem)
810 if (jobTable == NULL) return;
811 for (jobNb=0; jobNb<nbJobs; jobNb++) {
812 ZSTD_pthread_mutex_destroy(&jobTable[jobNb].job_mutex);
813 ZSTD_pthread_cond_destroy(&jobTable[jobNb].job_cond);
815 ZSTD_free(jobTable, cMem);
818 /* ZSTDMT_allocJobsTable()
819 * allocate and init a job table.
820 * update *nbJobsPtr to next power of 2 value, as size of table */
821 static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
823 U32 const nbJobsLog2 = ZSTD_highbit32(*nbJobsPtr) + 1;
824 U32 const nbJobs = 1 << nbJobsLog2;
826 ZSTDMT_jobDescription* const jobTable = (ZSTDMT_jobDescription*)
827 ZSTD_calloc(nbJobs * sizeof(ZSTDMT_jobDescription), cMem);
829 if (jobTable==NULL) return NULL;
831 for (jobNb=0; jobNb<nbJobs; jobNb++) {
832 initError |= ZSTD_pthread_mutex_init(&jobTable[jobNb].job_mutex, NULL);
833 initError |= ZSTD_pthread_cond_init(&jobTable[jobNb].job_cond, NULL);
835 if (initError != 0) {
836 ZSTDMT_freeJobsTable(jobTable, nbJobs, cMem);
842 static size_t ZSTDMT_expandJobsTable (ZSTDMT_CCtx* mtctx, U32 nbWorkers) {
843 U32 nbJobs = nbWorkers + 2;
844 if (nbJobs > mtctx->jobIDMask+1) { /* need more job capacity */
845 ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
846 mtctx->jobIDMask = 0;
847 mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, mtctx->cMem);
848 if (mtctx->jobs==NULL) return ERROR(memory_allocation);
849 assert((nbJobs != 0) && ((nbJobs & (nbJobs - 1)) == 0)); /* ensure nbJobs is a power of 2 */
850 mtctx->jobIDMask = nbJobs - 1;
856 /* ZSTDMT_CCtxParam_setNbWorkers():
857 * Internal use only */
858 size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers)
860 if (nbWorkers > ZSTDMT_NBWORKERS_MAX) nbWorkers = ZSTDMT_NBWORKERS_MAX;
861 params->nbWorkers = nbWorkers;
862 params->overlapSizeLog = ZSTDMT_OVERLAPLOG_DEFAULT;
867 ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem)
870 U32 nbJobs = nbWorkers + 2;
872 DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbWorkers = %u)", nbWorkers);
874 if (nbWorkers < 1) return NULL;
875 nbWorkers = MIN(nbWorkers , ZSTDMT_NBWORKERS_MAX);
876 if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
877 /* invalid custom allocator */
880 mtctx = (ZSTDMT_CCtx*) ZSTD_calloc(sizeof(ZSTDMT_CCtx), cMem);
881 if (!mtctx) return NULL;
882 ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
884 mtctx->allJobsCompleted = 1;
885 mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem);
886 mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem);
887 assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0); /* ensure nbJobs is a power of 2 */
888 mtctx->jobIDMask = nbJobs - 1;
889 mtctx->bufPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
890 mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem);
891 mtctx->seqPool = ZSTDMT_createSeqPool(nbWorkers, cMem);
892 initError = ZSTDMT_serialState_init(&mtctx->serial);
893 mtctx->roundBuff = kNullRoundBuff;
894 if (!mtctx->factory | !mtctx->jobs | !mtctx->bufPool | !mtctx->cctxPool | !mtctx->seqPool | initError) {
895 ZSTDMT_freeCCtx(mtctx);
898 DEBUGLOG(3, "mt_cctx created, for %u threads", nbWorkers);
902 ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers)
904 return ZSTDMT_createCCtx_advanced(nbWorkers, ZSTD_defaultCMem);
908 /* ZSTDMT_releaseAllJobResources() :
909 * note : ensure all workers are killed first ! */
910 static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx)
913 DEBUGLOG(3, "ZSTDMT_releaseAllJobResources");
914 for (jobID=0; jobID <= mtctx->jobIDMask; jobID++) {
915 DEBUGLOG(4, "job%02u: release dst address %08X", jobID, (U32)(size_t)mtctx->jobs[jobID].dstBuff.start);
916 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);
917 mtctx->jobs[jobID].dstBuff = g_nullBuffer;
918 mtctx->jobs[jobID].cSize = 0;
920 memset(mtctx->jobs, 0, (mtctx->jobIDMask+1)*sizeof(ZSTDMT_jobDescription));
921 mtctx->inBuff.buffer = g_nullBuffer;
922 mtctx->inBuff.filled = 0;
923 mtctx->allJobsCompleted = 1;
926 static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* mtctx)
928 DEBUGLOG(4, "ZSTDMT_waitForAllJobsCompleted");
929 while (mtctx->doneJobID < mtctx->nextJobID) {
930 unsigned const jobID = mtctx->doneJobID & mtctx->jobIDMask;
931 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
932 while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
933 DEBUGLOG(4, "waiting for jobCompleted signal from job %u", mtctx->doneJobID); /* we want to block when waiting for data to flush */
934 ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
936 ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
941 size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx)
943 if (mtctx==NULL) return 0; /* compatible with free on NULL */
944 POOL_free(mtctx->factory); /* stop and free worker threads */
945 ZSTDMT_releaseAllJobResources(mtctx); /* release job resources into pools first */
946 ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
947 ZSTDMT_freeBufferPool(mtctx->bufPool);
948 ZSTDMT_freeCCtxPool(mtctx->cctxPool);
949 ZSTDMT_freeSeqPool(mtctx->seqPool);
950 ZSTDMT_serialState_free(&mtctx->serial);
951 ZSTD_freeCDict(mtctx->cdictLocal);
952 if (mtctx->roundBuff.buffer)
953 ZSTD_free(mtctx->roundBuff.buffer, mtctx->cMem);
954 ZSTD_free(mtctx, mtctx->cMem);
958 size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx)
960 if (mtctx == NULL) return 0; /* supports sizeof NULL */
961 return sizeof(*mtctx)
962 + POOL_sizeof(mtctx->factory)
963 + ZSTDMT_sizeof_bufferPool(mtctx->bufPool)
964 + (mtctx->jobIDMask+1) * sizeof(ZSTDMT_jobDescription)
965 + ZSTDMT_sizeof_CCtxPool(mtctx->cctxPool)
966 + ZSTDMT_sizeof_seqPool(mtctx->seqPool)
967 + ZSTD_sizeof_CDict(mtctx->cdictLocal)
968 + mtctx->roundBuff.capacity;
972 size_t ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params,
973 ZSTDMT_parameter parameter, unsigned value) {
974 DEBUGLOG(4, "ZSTDMT_CCtxParam_setMTCtxParameter");
977 case ZSTDMT_p_jobSize :
978 DEBUGLOG(4, "ZSTDMT_CCtxParam_setMTCtxParameter : set jobSize to %u", value);
979 if ( (value > 0) /* value==0 => automatic job size */
980 & (value < ZSTDMT_JOBSIZE_MIN) )
981 value = ZSTDMT_JOBSIZE_MIN;
982 if (value > ZSTDMT_JOBSIZE_MAX)
983 value = ZSTDMT_JOBSIZE_MAX;
984 params->jobSize = value;
986 case ZSTDMT_p_overlapSectionLog :
987 if (value > 9) value = 9;
988 DEBUGLOG(4, "ZSTDMT_p_overlapSectionLog : %u", value);
989 params->overlapSizeLog = (value >= 9) ? 9 : value;
992 return ERROR(parameter_unsupported);
996 size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, unsigned value)
998 DEBUGLOG(4, "ZSTDMT_setMTCtxParameter");
1001 case ZSTDMT_p_jobSize :
1002 return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value);
1003 case ZSTDMT_p_overlapSectionLog :
1004 return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value);
1006 return ERROR(parameter_unsupported);
1010 size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, unsigned* value)
1012 switch (parameter) {
1013 case ZSTDMT_p_jobSize:
1014 *value = mtctx->params.jobSize;
1016 case ZSTDMT_p_overlapSectionLog:
1017 *value = mtctx->params.overlapSizeLog;
1020 return ERROR(parameter_unsupported);
1025 /* Sets parameters relevant to the compression job,
1026 * initializing others to default values. */
1027 static ZSTD_CCtx_params ZSTDMT_initJobCCtxParams(ZSTD_CCtx_params const params)
1029 ZSTD_CCtx_params jobParams;
1030 memset(&jobParams, 0, sizeof(jobParams));
1032 jobParams.cParams = params.cParams;
1033 jobParams.fParams = params.fParams;
1034 jobParams.compressionLevel = params.compressionLevel;
1040 /* ZSTDMT_resize() :
1041 * @return : error code if fails, 0 on success */
1042 static size_t ZSTDMT_resize(ZSTDMT_CCtx* mtctx, unsigned nbWorkers)
1044 if (POOL_resize(mtctx->factory, nbWorkers)) return ERROR(memory_allocation);
1045 CHECK_F( ZSTDMT_expandJobsTable(mtctx, nbWorkers) );
1046 mtctx->bufPool = ZSTDMT_expandBufferPool(mtctx->bufPool, nbWorkers);
1047 if (mtctx->bufPool == NULL) return ERROR(memory_allocation);
1048 mtctx->cctxPool = ZSTDMT_expandCCtxPool(mtctx->cctxPool, nbWorkers);
1049 if (mtctx->cctxPool == NULL) return ERROR(memory_allocation);
1050 mtctx->seqPool = ZSTDMT_expandSeqPool(mtctx->seqPool, nbWorkers);
1051 if (mtctx->seqPool == NULL) return ERROR(memory_allocation);
1052 ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
1057 /*! ZSTDMT_updateCParams_whileCompressing() :
1058 * Updates a selected set of compression parameters, remaining compatible with currently active frame.
1059 * New parameters will be applied to next compression job. */
1060 void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams)
1062 U32 const saved_wlog = mtctx->params.cParams.windowLog; /* Do not modify windowLog while compressing */
1063 int const compressionLevel = cctxParams->compressionLevel;
1064 DEBUGLOG(5, "ZSTDMT_updateCParams_whileCompressing (level:%i)",
1066 mtctx->params.compressionLevel = compressionLevel;
1067 { ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, 0, 0);
1068 cParams.windowLog = saved_wlog;
1069 mtctx->params.cParams = cParams;
1073 /* ZSTDMT_getFrameProgression():
1074 * tells how much data has been consumed (input) and produced (output) for current frame.
1075 * able to count progression inside worker threads.
1076 * Note : mutex will be acquired during statistics collection inside workers. */
1077 ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx)
1079 ZSTD_frameProgression fps;
1080 DEBUGLOG(5, "ZSTDMT_getFrameProgression");
1081 fps.ingested = mtctx->consumed + mtctx->inBuff.filled;
1082 fps.consumed = mtctx->consumed;
1083 fps.produced = fps.flushed = mtctx->produced;
1084 fps.currentJobID = mtctx->nextJobID;
1085 fps.nbActiveWorkers = 0;
1087 unsigned lastJobNb = mtctx->nextJobID + mtctx->jobReady; assert(mtctx->jobReady <= 1);
1088 DEBUGLOG(6, "ZSTDMT_getFrameProgression: jobs: from %u to <%u (jobReady:%u)",
1089 mtctx->doneJobID, lastJobNb, mtctx->jobReady)
1090 for (jobNb = mtctx->doneJobID ; jobNb < lastJobNb ; jobNb++) {
1091 unsigned const wJobID = jobNb & mtctx->jobIDMask;
1092 ZSTDMT_jobDescription* jobPtr = &mtctx->jobs[wJobID];
1093 ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
1094 { size_t const cResult = jobPtr->cSize;
1095 size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
1096 size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
1097 assert(flushed <= produced);
1098 fps.ingested += jobPtr->src.size;
1099 fps.consumed += jobPtr->consumed;
1100 fps.produced += produced;
1101 fps.flushed += flushed;
1102 fps.nbActiveWorkers += (jobPtr->consumed < jobPtr->src.size);
1104 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1111 size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx)
1114 unsigned const jobID = mtctx->doneJobID;
1115 assert(jobID <= mtctx->nextJobID);
1116 if (jobID == mtctx->nextJobID) return 0; /* no active job => nothing to flush */
1118 /* look into oldest non-fully-flushed job */
1119 { unsigned const wJobID = jobID & mtctx->jobIDMask;
1120 ZSTDMT_jobDescription* const jobPtr = &mtctx->jobs[wJobID];
1121 ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
1122 { size_t const cResult = jobPtr->cSize;
1123 size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
1124 size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
1125 assert(flushed <= produced);
1126 toFlush = produced - flushed;
1127 if (toFlush==0 && (jobPtr->consumed >= jobPtr->src.size)) {
1128 /* doneJobID is not-fully-flushed, but toFlush==0 : doneJobID should be compressing some more data */
1129 assert(jobPtr->consumed < jobPtr->src.size);
1132 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1139 /* ------------------------------------------ */
1140 /* ===== Multi-threaded compression ===== */
1141 /* ------------------------------------------ */
1143 static size_t ZSTDMT_computeTargetJobLog(ZSTD_CCtx_params const params)
1145 if (params.ldmParams.enableLdm)
1146 return MAX(21, params.cParams.chainLog + 4);
1147 return MAX(20, params.cParams.windowLog + 2);
1150 static size_t ZSTDMT_computeOverlapLog(ZSTD_CCtx_params const params)
1152 unsigned const overlapRLog = (params.overlapSizeLog>9) ? 0 : 9-params.overlapSizeLog;
1153 if (params.ldmParams.enableLdm)
1154 return (MIN(params.cParams.windowLog, ZSTDMT_computeTargetJobLog(params) - 2) - overlapRLog);
1155 return overlapRLog >= 9 ? 0 : (params.cParams.windowLog - overlapRLog);
1158 static unsigned ZSTDMT_computeNbJobs(ZSTD_CCtx_params params, size_t srcSize, unsigned nbWorkers) {
1159 assert(nbWorkers>0);
1160 { size_t const jobSizeTarget = (size_t)1 << ZSTDMT_computeTargetJobLog(params);
1161 size_t const jobMaxSize = jobSizeTarget << 2;
1162 size_t const passSizeMax = jobMaxSize * nbWorkers;
1163 unsigned const multiplier = (unsigned)(srcSize / passSizeMax) + 1;
1164 unsigned const nbJobsLarge = multiplier * nbWorkers;
1165 unsigned const nbJobsMax = (unsigned)(srcSize / jobSizeTarget) + 1;
1166 unsigned const nbJobsSmall = MIN(nbJobsMax, nbWorkers);
1167 return (multiplier>1) ? nbJobsLarge : nbJobsSmall;
1170 /* ZSTDMT_compress_advanced_internal() :
1171 * This is a blocking function : it will only give back control to caller after finishing its compression job.
1173 static size_t ZSTDMT_compress_advanced_internal(
1175 void* dst, size_t dstCapacity,
1176 const void* src, size_t srcSize,
1177 const ZSTD_CDict* cdict,
1178 ZSTD_CCtx_params params)
1180 ZSTD_CCtx_params const jobParams = ZSTDMT_initJobCCtxParams(params);
1181 size_t const overlapSize = (size_t)1 << ZSTDMT_computeOverlapLog(params);
1182 unsigned const nbJobs = ZSTDMT_computeNbJobs(params, srcSize, params.nbWorkers);
1183 size_t const proposedJobSize = (srcSize + (nbJobs-1)) / nbJobs;
1184 size_t const avgJobSize = (((proposedJobSize-1) & 0x1FFFF) < 0x7FFF) ? proposedJobSize + 0xFFFF : proposedJobSize; /* avoid too small last block */
1185 const char* const srcStart = (const char*)src;
1186 size_t remainingSrcSize = srcSize;
1187 unsigned const compressWithinDst = (dstCapacity >= ZSTD_compressBound(srcSize)) ? nbJobs : (unsigned)(dstCapacity / ZSTD_compressBound(avgJobSize)); /* presumes avgJobSize >= 256 KB, which should be the case */
1188 size_t frameStartPos = 0, dstBufferPos = 0;
1189 assert(jobParams.nbWorkers == 0);
1190 assert(mtctx->cctxPool->totalCCtx == params.nbWorkers);
1192 params.jobSize = (U32)avgJobSize;
1193 DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: nbJobs=%2u (rawSize=%u bytes; fixedSize=%u) ",
1194 nbJobs, (U32)proposedJobSize, (U32)avgJobSize);
1196 if ((nbJobs==1) | (params.nbWorkers<=1)) { /* fallback to single-thread mode : this is a blocking invocation anyway */
1197 ZSTD_CCtx* const cctx = mtctx->cctxPool->cctx[0];
1198 DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: fallback to single-thread mode");
1199 if (cdict) return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, jobParams.fParams);
1200 return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, NULL, 0, jobParams);
1203 assert(avgJobSize >= 256 KB); /* condition for ZSTD_compressBound(A) + ZSTD_compressBound(B) <= ZSTD_compressBound(A+B), required to compress directly into Dst (no additional buffer) */
1204 ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(avgJobSize) );
1205 if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params, avgJobSize))
1206 return ERROR(memory_allocation);
1208 CHECK_F( ZSTDMT_expandJobsTable(mtctx, nbJobs) ); /* only expands if necessary */
1211 for (u=0; u<nbJobs; u++) {
1212 size_t const jobSize = MIN(remainingSrcSize, avgJobSize);
1213 size_t const dstBufferCapacity = ZSTD_compressBound(jobSize);
1214 buffer_t const dstAsBuffer = { (char*)dst + dstBufferPos, dstBufferCapacity };
1215 buffer_t const dstBuffer = u < compressWithinDst ? dstAsBuffer : g_nullBuffer;
1216 size_t dictSize = u ? overlapSize : 0;
1218 mtctx->jobs[u].prefix.start = srcStart + frameStartPos - dictSize;
1219 mtctx->jobs[u].prefix.size = dictSize;
1220 mtctx->jobs[u].src.start = srcStart + frameStartPos;
1221 mtctx->jobs[u].src.size = jobSize; assert(jobSize > 0); /* avoid job.src.size == 0 */
1222 mtctx->jobs[u].consumed = 0;
1223 mtctx->jobs[u].cSize = 0;
1224 mtctx->jobs[u].cdict = (u==0) ? cdict : NULL;
1225 mtctx->jobs[u].fullFrameSize = srcSize;
1226 mtctx->jobs[u].params = jobParams;
1227 /* do not calculate checksum within sections, but write it in header for first section */
1228 mtctx->jobs[u].dstBuff = dstBuffer;
1229 mtctx->jobs[u].cctxPool = mtctx->cctxPool;
1230 mtctx->jobs[u].bufPool = mtctx->bufPool;
1231 mtctx->jobs[u].seqPool = mtctx->seqPool;
1232 mtctx->jobs[u].serial = &mtctx->serial;
1233 mtctx->jobs[u].jobID = u;
1234 mtctx->jobs[u].firstJob = (u==0);
1235 mtctx->jobs[u].lastJob = (u==nbJobs-1);
1237 DEBUGLOG(5, "ZSTDMT_compress_advanced_internal: posting job %u (%u bytes)", u, (U32)jobSize);
1238 DEBUG_PRINTHEX(6, mtctx->jobs[u].prefix.start, 12);
1239 POOL_add(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[u]);
1241 frameStartPos += jobSize;
1242 dstBufferPos += dstBufferCapacity;
1243 remainingSrcSize -= jobSize;
1246 /* collect result */
1247 { size_t error = 0, dstPos = 0;
1249 for (jobID=0; jobID<nbJobs; jobID++) {
1250 DEBUGLOG(5, "waiting for job %u ", jobID);
1251 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
1252 while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
1253 DEBUGLOG(5, "waiting for jobCompleted signal from job %u", jobID);
1254 ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
1256 ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
1257 DEBUGLOG(5, "ready to write job %u ", jobID);
1259 { size_t const cSize = mtctx->jobs[jobID].cSize;
1260 if (ZSTD_isError(cSize)) error = cSize;
1261 if ((!error) && (dstPos + cSize > dstCapacity)) error = ERROR(dstSize_tooSmall);
1262 if (jobID) { /* note : job 0 is written directly at dst, which is correct position */
1264 memmove((char*)dst + dstPos, mtctx->jobs[jobID].dstBuff.start, cSize); /* may overlap when job compressed within dst */
1265 if (jobID >= compressWithinDst) { /* job compressed into its own buffer, which must be released */
1266 DEBUGLOG(5, "releasing buffer %u>=%u", jobID, compressWithinDst);
1267 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);
1269 mtctx->jobs[jobID].dstBuff = g_nullBuffer;
1270 mtctx->jobs[jobID].cSize = 0;
1273 } /* for (jobID=0; jobID<nbJobs; jobID++) */
1275 DEBUGLOG(4, "checksumFlag : %u ", params.fParams.checksumFlag);
1276 if (params.fParams.checksumFlag) {
1277 U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
1278 if (dstPos + 4 > dstCapacity) {
1279 error = ERROR(dstSize_tooSmall);
1281 DEBUGLOG(4, "writing checksum : %08X \n", checksum);
1282 MEM_writeLE32((char*)dst + dstPos, checksum);
1286 if (!error) DEBUGLOG(4, "compressed size : %u ", (U32)dstPos);
1287 return error ? error : dstPos;
1291 size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx,
1292 void* dst, size_t dstCapacity,
1293 const void* src, size_t srcSize,
1294 const ZSTD_CDict* cdict,
1295 ZSTD_parameters params,
1296 unsigned overlapLog)
1298 ZSTD_CCtx_params cctxParams = mtctx->params;
1299 cctxParams.cParams = params.cParams;
1300 cctxParams.fParams = params.fParams;
1301 cctxParams.overlapSizeLog = overlapLog;
1302 return ZSTDMT_compress_advanced_internal(mtctx,
1309 size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
1310 void* dst, size_t dstCapacity,
1311 const void* src, size_t srcSize,
1312 int compressionLevel)
1314 U32 const overlapLog = (compressionLevel >= ZSTD_maxCLevel()) ? 9 : ZSTDMT_OVERLAPLOG_DEFAULT;
1315 ZSTD_parameters params = ZSTD_getParams(compressionLevel, srcSize, 0);
1316 params.fParams.contentSizeFlag = 1;
1317 return ZSTDMT_compress_advanced(mtctx, dst, dstCapacity, src, srcSize, NULL, params, overlapLog);
1321 /* ====================================== */
1322 /* ======= Streaming API ======= */
1323 /* ====================================== */
1325 size_t ZSTDMT_initCStream_internal(
1327 const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
1328 const ZSTD_CDict* cdict, ZSTD_CCtx_params params,
1329 unsigned long long pledgedSrcSize)
1331 DEBUGLOG(4, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u, nbWorkers=%u, cctxPool=%u)",
1332 (U32)pledgedSrcSize, params.nbWorkers, mtctx->cctxPool->totalCCtx);
1334 /* params supposed partially fully validated at this point */
1335 assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
1336 assert(!((dict) && (cdict))); /* either dict or cdict, not both */
1339 if (params.nbWorkers != mtctx->params.nbWorkers)
1340 CHECK_F( ZSTDMT_resize(mtctx, params.nbWorkers) );
1342 if (params.jobSize > 0 && params.jobSize < ZSTDMT_JOBSIZE_MIN) params.jobSize = ZSTDMT_JOBSIZE_MIN;
1343 if (params.jobSize > ZSTDMT_JOBSIZE_MAX) params.jobSize = ZSTDMT_JOBSIZE_MAX;
1345 mtctx->singleBlockingThread = (pledgedSrcSize <= ZSTDMT_JOBSIZE_MIN); /* do not trigger multi-threading when srcSize is too small */
1346 if (mtctx->singleBlockingThread) {
1347 ZSTD_CCtx_params const singleThreadParams = ZSTDMT_initJobCCtxParams(params);
1348 DEBUGLOG(5, "ZSTDMT_initCStream_internal: switch to single blocking thread mode");
1349 assert(singleThreadParams.nbWorkers == 0);
1350 return ZSTD_initCStream_internal(mtctx->cctxPool->cctx[0],
1351 dict, dictSize, cdict,
1352 singleThreadParams, pledgedSrcSize);
1355 DEBUGLOG(4, "ZSTDMT_initCStream_internal: %u workers", params.nbWorkers);
1357 if (mtctx->allJobsCompleted == 0) { /* previous compression not correctly finished */
1358 ZSTDMT_waitForAllJobsCompleted(mtctx);
1359 ZSTDMT_releaseAllJobResources(mtctx);
1360 mtctx->allJobsCompleted = 1;
1363 mtctx->params = params;
1364 mtctx->frameContentSize = pledgedSrcSize;
1366 ZSTD_freeCDict(mtctx->cdictLocal);
1367 mtctx->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
1368 ZSTD_dlm_byCopy, dictContentType, /* note : a loadPrefix becomes an internal CDict */
1369 params.cParams, mtctx->cMem);
1370 mtctx->cdict = mtctx->cdictLocal;
1371 if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
1373 ZSTD_freeCDict(mtctx->cdictLocal);
1374 mtctx->cdictLocal = NULL;
1375 mtctx->cdict = cdict;
1378 mtctx->targetPrefixSize = (size_t)1 << ZSTDMT_computeOverlapLog(params);
1379 DEBUGLOG(4, "overlapLog=%u => %u KB", params.overlapSizeLog, (U32)(mtctx->targetPrefixSize>>10));
1380 mtctx->targetSectionSize = params.jobSize;
1381 if (mtctx->targetSectionSize == 0) {
1382 mtctx->targetSectionSize = 1ULL << ZSTDMT_computeTargetJobLog(params);
1384 if (mtctx->targetSectionSize < mtctx->targetPrefixSize) mtctx->targetSectionSize = mtctx->targetPrefixSize; /* job size must be >= overlap size */
1385 DEBUGLOG(4, "Job Size : %u KB (note : set to %u)", (U32)(mtctx->targetSectionSize>>10), params.jobSize);
1386 DEBUGLOG(4, "inBuff Size : %u KB", (U32)(mtctx->targetSectionSize>>10));
1387 ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(mtctx->targetSectionSize));
1389 /* If ldm is enabled we need windowSize space. */
1390 size_t const windowSize = mtctx->params.ldmParams.enableLdm ? (1U << mtctx->params.cParams.windowLog) : 0;
1391 /* Two buffers of slack, plus extra space for the overlap
1392 * This is the minimum slack that LDM works with. One extra because
1393 * flush might waste up to targetSectionSize-1 bytes. Another extra
1394 * for the overlap (if > 0), then one to fill which doesn't overlap
1395 * with the LDM window.
1397 size_t const nbSlackBuffers = 2 + (mtctx->targetPrefixSize > 0);
1398 size_t const slackSize = mtctx->targetSectionSize * nbSlackBuffers;
1399 /* Compute the total size, and always have enough slack */
1400 size_t const nbWorkers = MAX(mtctx->params.nbWorkers, 1);
1401 size_t const sectionsSize = mtctx->targetSectionSize * nbWorkers;
1402 size_t const capacity = MAX(windowSize, sectionsSize) + slackSize;
1403 if (mtctx->roundBuff.capacity < capacity) {
1404 if (mtctx->roundBuff.buffer)
1405 ZSTD_free(mtctx->roundBuff.buffer, mtctx->cMem);
1406 mtctx->roundBuff.buffer = (BYTE*)ZSTD_malloc(capacity, mtctx->cMem);
1407 if (mtctx->roundBuff.buffer == NULL) {
1408 mtctx->roundBuff.capacity = 0;
1409 return ERROR(memory_allocation);
1411 mtctx->roundBuff.capacity = capacity;
1414 DEBUGLOG(4, "roundBuff capacity : %u KB", (U32)(mtctx->roundBuff.capacity>>10));
1415 mtctx->roundBuff.pos = 0;
1416 mtctx->inBuff.buffer = g_nullBuffer;
1417 mtctx->inBuff.filled = 0;
1418 mtctx->inBuff.prefix = kNullRange;
1419 mtctx->doneJobID = 0;
1420 mtctx->nextJobID = 0;
1421 mtctx->frameEnded = 0;
1422 mtctx->allJobsCompleted = 0;
1423 mtctx->consumed = 0;
1424 mtctx->produced = 0;
1425 if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params, mtctx->targetSectionSize))
1426 return ERROR(memory_allocation);
1430 size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx,
1431 const void* dict, size_t dictSize,
1432 ZSTD_parameters params,
1433 unsigned long long pledgedSrcSize)
1435 ZSTD_CCtx_params cctxParams = mtctx->params; /* retrieve sticky params */
1436 DEBUGLOG(4, "ZSTDMT_initCStream_advanced (pledgedSrcSize=%u)", (U32)pledgedSrcSize);
1437 cctxParams.cParams = params.cParams;
1438 cctxParams.fParams = params.fParams;
1439 return ZSTDMT_initCStream_internal(mtctx, dict, dictSize, ZSTD_dct_auto, NULL,
1440 cctxParams, pledgedSrcSize);
1443 size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx,
1444 const ZSTD_CDict* cdict,
1445 ZSTD_frameParameters fParams,
1446 unsigned long long pledgedSrcSize)
1448 ZSTD_CCtx_params cctxParams = mtctx->params;
1449 if (cdict==NULL) return ERROR(dictionary_wrong); /* method incompatible with NULL cdict */
1450 cctxParams.cParams = ZSTD_getCParamsFromCDict(cdict);
1451 cctxParams.fParams = fParams;
1452 return ZSTDMT_initCStream_internal(mtctx, NULL, 0 /*dictSize*/, ZSTD_dct_auto, cdict,
1453 cctxParams, pledgedSrcSize);
1457 /* ZSTDMT_resetCStream() :
1458 * pledgedSrcSize can be zero == unknown (for the time being)
1459 * prefer using ZSTD_CONTENTSIZE_UNKNOWN,
1460 * as `0` might mean "empty" in the future */
1461 size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* mtctx, unsigned long long pledgedSrcSize)
1463 if (!pledgedSrcSize) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;
1464 return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dct_auto, 0, mtctx->params,
1468 size_t ZSTDMT_initCStream(ZSTDMT_CCtx* mtctx, int compressionLevel) {
1469 ZSTD_parameters const params = ZSTD_getParams(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0);
1470 ZSTD_CCtx_params cctxParams = mtctx->params; /* retrieve sticky params */
1471 DEBUGLOG(4, "ZSTDMT_initCStream (cLevel=%i)", compressionLevel);
1472 cctxParams.cParams = params.cParams;
1473 cctxParams.fParams = params.fParams;
1474 return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dct_auto, NULL, cctxParams, ZSTD_CONTENTSIZE_UNKNOWN);
1478 /* ZSTDMT_writeLastEmptyBlock()
1479 * Write a single empty block with an end-of-frame to finish a frame.
1480 * Job must be created from streaming variant.
1481 * This function is always successfull if expected conditions are fulfilled.
1483 static void ZSTDMT_writeLastEmptyBlock(ZSTDMT_jobDescription* job)
1485 assert(job->lastJob == 1);
1486 assert(job->src.size == 0); /* last job is empty -> will be simplified into a last empty block */
1487 assert(job->firstJob == 0); /* cannot be first job, as it also needs to create frame header */
1488 assert(job->dstBuff.start == NULL); /* invoked from streaming variant only (otherwise, dstBuff might be user's output) */
1489 job->dstBuff = ZSTDMT_getBuffer(job->bufPool);
1490 if (job->dstBuff.start == NULL) {
1491 job->cSize = ERROR(memory_allocation);
1494 assert(job->dstBuff.capacity >= ZSTD_blockHeaderSize); /* no buffer should ever be that small */
1495 job->src = kNullRange;
1496 job->cSize = ZSTD_writeLastEmptyBlock(job->dstBuff.start, job->dstBuff.capacity);
1497 assert(!ZSTD_isError(job->cSize));
1498 assert(job->consumed == 0);
1501 static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* mtctx, size_t srcSize, ZSTD_EndDirective endOp)
1503 unsigned const jobID = mtctx->nextJobID & mtctx->jobIDMask;
1504 int const endFrame = (endOp == ZSTD_e_end);
1506 if (mtctx->nextJobID > mtctx->doneJobID + mtctx->jobIDMask) {
1507 DEBUGLOG(5, "ZSTDMT_createCompressionJob: will not create new job : table is full");
1508 assert((mtctx->nextJobID & mtctx->jobIDMask) == (mtctx->doneJobID & mtctx->jobIDMask));
1512 if (!mtctx->jobReady) {
1513 BYTE const* src = (BYTE const*)mtctx->inBuff.buffer.start;
1514 DEBUGLOG(5, "ZSTDMT_createCompressionJob: preparing job %u to compress %u bytes with %u preload ",
1515 mtctx->nextJobID, (U32)srcSize, (U32)mtctx->inBuff.prefix.size);
1516 mtctx->jobs[jobID].src.start = src;
1517 mtctx->jobs[jobID].src.size = srcSize;
1518 assert(mtctx->inBuff.filled >= srcSize);
1519 mtctx->jobs[jobID].prefix = mtctx->inBuff.prefix;
1520 mtctx->jobs[jobID].consumed = 0;
1521 mtctx->jobs[jobID].cSize = 0;
1522 mtctx->jobs[jobID].params = mtctx->params;
1523 mtctx->jobs[jobID].cdict = mtctx->nextJobID==0 ? mtctx->cdict : NULL;
1524 mtctx->jobs[jobID].fullFrameSize = mtctx->frameContentSize;
1525 mtctx->jobs[jobID].dstBuff = g_nullBuffer;
1526 mtctx->jobs[jobID].cctxPool = mtctx->cctxPool;
1527 mtctx->jobs[jobID].bufPool = mtctx->bufPool;
1528 mtctx->jobs[jobID].seqPool = mtctx->seqPool;
1529 mtctx->jobs[jobID].serial = &mtctx->serial;
1530 mtctx->jobs[jobID].jobID = mtctx->nextJobID;
1531 mtctx->jobs[jobID].firstJob = (mtctx->nextJobID==0);
1532 mtctx->jobs[jobID].lastJob = endFrame;
1533 mtctx->jobs[jobID].frameChecksumNeeded = mtctx->params.fParams.checksumFlag && endFrame && (mtctx->nextJobID>0);
1534 mtctx->jobs[jobID].dstFlushed = 0;
1536 /* Update the round buffer pos and clear the input buffer to be reset */
1537 mtctx->roundBuff.pos += srcSize;
1538 mtctx->inBuff.buffer = g_nullBuffer;
1539 mtctx->inBuff.filled = 0;
1540 /* Set the prefix */
1542 size_t const newPrefixSize = MIN(srcSize, mtctx->targetPrefixSize);
1543 mtctx->inBuff.prefix.start = src + srcSize - newPrefixSize;
1544 mtctx->inBuff.prefix.size = newPrefixSize;
1545 } else { /* endFrame==1 => no need for another input buffer */
1546 mtctx->inBuff.prefix = kNullRange;
1547 mtctx->frameEnded = endFrame;
1548 if (mtctx->nextJobID == 0) {
1549 /* single job exception : checksum is already calculated directly within worker thread */
1550 mtctx->params.fParams.checksumFlag = 0;
1554 && (mtctx->nextJobID>0)/*single job must also write frame header*/ ) {
1555 DEBUGLOG(5, "ZSTDMT_createCompressionJob: creating a last empty block to end frame");
1556 assert(endOp == ZSTD_e_end); /* only possible case : need to end the frame with an empty last block */
1557 ZSTDMT_writeLastEmptyBlock(mtctx->jobs + jobID);
1563 DEBUGLOG(5, "ZSTDMT_createCompressionJob: posting job %u : %u bytes (end:%u, jobNb == %u (mod:%u))",
1565 (U32)mtctx->jobs[jobID].src.size,
1566 mtctx->jobs[jobID].lastJob,
1569 if (POOL_tryAdd(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[jobID])) {
1571 mtctx->jobReady = 0;
1573 DEBUGLOG(5, "ZSTDMT_createCompressionJob: no worker available for job %u", mtctx->nextJobID);
1574 mtctx->jobReady = 1;
1580 /*! ZSTDMT_flushProduced() :
1581 * flush whatever data has been produced but not yet flushed in current job.
1582 * move to next job if current one is fully flushed.
1583 * `output` : `pos` will be updated with amount of data flushed .
1584 * `blockToFlush` : if >0, the function will block and wait if there is no data available to flush .
1585 * @return : amount of data remaining within internal buffer, 0 if no more, 1 if unknown but > 0, or an error code */
1586 static size_t ZSTDMT_flushProduced(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, unsigned blockToFlush, ZSTD_EndDirective end)
1588 unsigned const wJobID = mtctx->doneJobID & mtctx->jobIDMask;
1589 DEBUGLOG(5, "ZSTDMT_flushProduced (blocking:%u , job %u <= %u)",
1590 blockToFlush, mtctx->doneJobID, mtctx->nextJobID);
1591 assert(output->size >= output->pos);
1593 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
1595 && (mtctx->doneJobID < mtctx->nextJobID) ) {
1596 assert(mtctx->jobs[wJobID].dstFlushed <= mtctx->jobs[wJobID].cSize);
1597 while (mtctx->jobs[wJobID].dstFlushed == mtctx->jobs[wJobID].cSize) { /* nothing to flush */
1598 if (mtctx->jobs[wJobID].consumed == mtctx->jobs[wJobID].src.size) {
1599 DEBUGLOG(5, "job %u is completely consumed (%u == %u) => don't wait for cond, there will be none",
1600 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].consumed, (U32)mtctx->jobs[wJobID].src.size);
1603 DEBUGLOG(5, "waiting for something to flush from job %u (currently flushed: %u bytes)",
1604 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
1605 ZSTD_pthread_cond_wait(&mtctx->jobs[wJobID].job_cond, &mtctx->jobs[wJobID].job_mutex); /* block when nothing to flush but some to come */
1608 /* try to flush something */
1609 { size_t cSize = mtctx->jobs[wJobID].cSize; /* shared */
1610 size_t const srcConsumed = mtctx->jobs[wJobID].consumed; /* shared */
1611 size_t const srcSize = mtctx->jobs[wJobID].src.size; /* read-only, could be done after mutex lock, but no-declaration-after-statement */
1612 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1613 if (ZSTD_isError(cSize)) {
1614 DEBUGLOG(5, "ZSTDMT_flushProduced: job %u : compression error detected : %s",
1615 mtctx->doneJobID, ZSTD_getErrorName(cSize));
1616 ZSTDMT_waitForAllJobsCompleted(mtctx);
1617 ZSTDMT_releaseAllJobResources(mtctx);
1620 /* add frame checksum if necessary (can only happen once) */
1621 assert(srcConsumed <= srcSize);
1622 if ( (srcConsumed == srcSize) /* job completed -> worker no longer active */
1623 && mtctx->jobs[wJobID].frameChecksumNeeded ) {
1624 U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
1625 DEBUGLOG(4, "ZSTDMT_flushProduced: writing checksum : %08X \n", checksum);
1626 MEM_writeLE32((char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].cSize, checksum);
1628 mtctx->jobs[wJobID].cSize += 4; /* can write this shared value, as worker is no longer active */
1629 mtctx->jobs[wJobID].frameChecksumNeeded = 0;
1632 if (cSize > 0) { /* compression is ongoing or completed */
1633 size_t const toFlush = MIN(cSize - mtctx->jobs[wJobID].dstFlushed, output->size - output->pos);
1634 DEBUGLOG(5, "ZSTDMT_flushProduced: Flushing %u bytes from job %u (completion:%u/%u, generated:%u)",
1635 (U32)toFlush, mtctx->doneJobID, (U32)srcConsumed, (U32)srcSize, (U32)cSize);
1636 assert(mtctx->doneJobID < mtctx->nextJobID);
1637 assert(cSize >= mtctx->jobs[wJobID].dstFlushed);
1638 assert(mtctx->jobs[wJobID].dstBuff.start != NULL);
1639 memcpy((char*)output->dst + output->pos,
1640 (const char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].dstFlushed,
1642 output->pos += toFlush;
1643 mtctx->jobs[wJobID].dstFlushed += toFlush; /* can write : this value is only used by mtctx */
1645 if ( (srcConsumed == srcSize) /* job is completed */
1646 && (mtctx->jobs[wJobID].dstFlushed == cSize) ) { /* output buffer fully flushed => free this job position */
1647 DEBUGLOG(5, "Job %u completed (%u bytes), moving to next one",
1648 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
1649 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[wJobID].dstBuff);
1650 DEBUGLOG(5, "dstBuffer released");
1651 mtctx->jobs[wJobID].dstBuff = g_nullBuffer;
1652 mtctx->jobs[wJobID].cSize = 0; /* ensure this job slot is considered "not started" in future check */
1653 mtctx->consumed += srcSize;
1654 mtctx->produced += cSize;
1658 /* return value : how many bytes left in buffer ; fake it to 1 when unknown but >0 */
1659 if (cSize > mtctx->jobs[wJobID].dstFlushed) return (cSize - mtctx->jobs[wJobID].dstFlushed);
1660 if (srcSize > srcConsumed) return 1; /* current job not completely compressed */
1662 if (mtctx->doneJobID < mtctx->nextJobID) return 1; /* some more jobs ongoing */
1663 if (mtctx->jobReady) return 1; /* one job is ready to push, just not yet in the list */
1664 if (mtctx->inBuff.filled > 0) return 1; /* input is not empty, and still needs to be converted into a job */
1665 mtctx->allJobsCompleted = mtctx->frameEnded; /* all jobs are entirely flushed => if this one is last one, frame is completed */
1666 if (end == ZSTD_e_end) return !mtctx->frameEnded; /* for ZSTD_e_end, question becomes : is frame completed ? instead of : are internal buffers fully flushed ? */
1667 return 0; /* internal buffers fully flushed */
1671 * Returns the range of data used by the earliest job that is not yet complete.
1672 * If the data of the first job is broken up into two segments, we cover both
1675 static range_t ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx)
1677 unsigned const firstJobID = mtctx->doneJobID;
1678 unsigned const lastJobID = mtctx->nextJobID;
1681 for (jobID = firstJobID; jobID < lastJobID; ++jobID) {
1682 unsigned const wJobID = jobID & mtctx->jobIDMask;
1685 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
1686 consumed = mtctx->jobs[wJobID].consumed;
1687 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1689 if (consumed < mtctx->jobs[wJobID].src.size) {
1690 range_t range = mtctx->jobs[wJobID].prefix;
1691 if (range.size == 0) {
1693 range = mtctx->jobs[wJobID].src;
1695 /* Job source in multiple segments not supported yet */
1696 assert(range.start <= mtctx->jobs[wJobID].src.start);
1704 * Returns non-zero iff buffer and range overlap.
1706 static int ZSTDMT_isOverlapped(buffer_t buffer, range_t range)
1708 BYTE const* const bufferStart = (BYTE const*)buffer.start;
1709 BYTE const* const bufferEnd = bufferStart + buffer.capacity;
1710 BYTE const* const rangeStart = (BYTE const*)range.start;
1711 BYTE const* const rangeEnd = rangeStart + range.size;
1713 if (rangeStart == NULL || bufferStart == NULL)
1715 /* Empty ranges cannot overlap */
1716 if (bufferStart == bufferEnd || rangeStart == rangeEnd)
1719 return bufferStart < rangeEnd && rangeStart < bufferEnd;
1722 static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window)
1727 DEBUGLOG(5, "ZSTDMT_doesOverlapWindow");
1728 extDict.start = window.dictBase + window.lowLimit;
1729 extDict.size = window.dictLimit - window.lowLimit;
1731 prefix.start = window.base + window.dictLimit;
1732 prefix.size = window.nextSrc - (window.base + window.dictLimit);
1733 DEBUGLOG(5, "extDict [0x%zx, 0x%zx)",
1734 (size_t)extDict.start,
1735 (size_t)extDict.start + extDict.size);
1736 DEBUGLOG(5, "prefix [0x%zx, 0x%zx)",
1737 (size_t)prefix.start,
1738 (size_t)prefix.start + prefix.size);
1740 return ZSTDMT_isOverlapped(buffer, extDict)
1741 || ZSTDMT_isOverlapped(buffer, prefix);
1744 static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, buffer_t buffer)
1746 if (mtctx->params.ldmParams.enableLdm) {
1747 ZSTD_pthread_mutex_t* mutex = &mtctx->serial.ldmWindowMutex;
1748 DEBUGLOG(5, "ZSTDMT_waitForLdmComplete");
1749 DEBUGLOG(5, "source [0x%zx, 0x%zx)",
1750 (size_t)buffer.start,
1751 (size_t)buffer.start + buffer.capacity);
1752 ZSTD_PTHREAD_MUTEX_LOCK(mutex);
1753 while (ZSTDMT_doesOverlapWindow(buffer, mtctx->serial.ldmWindow)) {
1754 DEBUGLOG(5, "Waiting for LDM to finish...");
1755 ZSTD_pthread_cond_wait(&mtctx->serial.ldmWindowCond, mutex);
1757 DEBUGLOG(6, "Done waiting for LDM to finish");
1758 ZSTD_pthread_mutex_unlock(mutex);
1763 * Attempts to set the inBuff to the next section to fill.
1764 * If any part of the new section is still in use we give up.
1765 * Returns non-zero if the buffer is filled.
1767 static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx)
1769 range_t const inUse = ZSTDMT_getInputDataInUse(mtctx);
1770 size_t const spaceLeft = mtctx->roundBuff.capacity - mtctx->roundBuff.pos;
1771 size_t const target = mtctx->targetSectionSize;
1774 DEBUGLOG(5, "ZSTDMT_tryGetInputRange");
1775 assert(mtctx->inBuff.buffer.start == NULL);
1776 assert(mtctx->roundBuff.capacity >= target);
1778 if (spaceLeft < target) {
1779 /* ZSTD_invalidateRepCodes() doesn't work for extDict variants.
1780 * Simply copy the prefix to the beginning in that case.
1782 BYTE* const start = (BYTE*)mtctx->roundBuff.buffer;
1783 size_t const prefixSize = mtctx->inBuff.prefix.size;
1785 buffer.start = start;
1786 buffer.capacity = prefixSize;
1787 if (ZSTDMT_isOverlapped(buffer, inUse)) {
1788 DEBUGLOG(5, "Waiting for buffer...");
1791 ZSTDMT_waitForLdmComplete(mtctx, buffer);
1792 memmove(start, mtctx->inBuff.prefix.start, prefixSize);
1793 mtctx->inBuff.prefix.start = start;
1794 mtctx->roundBuff.pos = prefixSize;
1796 buffer.start = mtctx->roundBuff.buffer + mtctx->roundBuff.pos;
1797 buffer.capacity = target;
1799 if (ZSTDMT_isOverlapped(buffer, inUse)) {
1800 DEBUGLOG(5, "Waiting for buffer...");
1803 assert(!ZSTDMT_isOverlapped(buffer, mtctx->inBuff.prefix));
1805 ZSTDMT_waitForLdmComplete(mtctx, buffer);
1807 DEBUGLOG(5, "Using prefix range [%zx, %zx)",
1808 (size_t)mtctx->inBuff.prefix.start,
1809 (size_t)mtctx->inBuff.prefix.start + mtctx->inBuff.prefix.size);
1810 DEBUGLOG(5, "Using source range [%zx, %zx)",
1811 (size_t)buffer.start,
1812 (size_t)buffer.start + buffer.capacity);
1815 mtctx->inBuff.buffer = buffer;
1816 mtctx->inBuff.filled = 0;
1817 assert(mtctx->roundBuff.pos + buffer.capacity <= mtctx->roundBuff.capacity);
1822 /** ZSTDMT_compressStream_generic() :
1823 * internal use only - exposed to be invoked from zstd_compress.c
1824 * assumption : output and input are valid (pos <= size)
1825 * @return : minimum amount of data remaining to flush, 0 if none */
1826 size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
1827 ZSTD_outBuffer* output,
1828 ZSTD_inBuffer* input,
1829 ZSTD_EndDirective endOp)
1831 unsigned forwardInputProgress = 0;
1832 DEBUGLOG(5, "ZSTDMT_compressStream_generic (endOp=%u, srcSize=%u)",
1833 (U32)endOp, (U32)(input->size - input->pos));
1834 assert(output->pos <= output->size);
1835 assert(input->pos <= input->size);
1837 if (mtctx->singleBlockingThread) { /* delegate to single-thread (synchronous) */
1838 return ZSTD_compressStream_generic(mtctx->cctxPool->cctx[0], output, input, endOp);
1841 if ((mtctx->frameEnded) && (endOp==ZSTD_e_continue)) {
1842 /* current frame being ended. Only flush/end are allowed */
1843 return ERROR(stage_wrong);
1846 /* single-pass shortcut (note : synchronous-mode) */
1847 if ( (mtctx->nextJobID == 0) /* just started */
1848 && (mtctx->inBuff.filled == 0) /* nothing buffered */
1849 && (!mtctx->jobReady) /* no job already created */
1850 && (endOp == ZSTD_e_end) /* end order */
1851 && (output->size - output->pos >= ZSTD_compressBound(input->size - input->pos)) ) { /* enough space in dst */
1852 size_t const cSize = ZSTDMT_compress_advanced_internal(mtctx,
1853 (char*)output->dst + output->pos, output->size - output->pos,
1854 (const char*)input->src + input->pos, input->size - input->pos,
1855 mtctx->cdict, mtctx->params);
1856 if (ZSTD_isError(cSize)) return cSize;
1857 input->pos = input->size;
1858 output->pos += cSize;
1859 mtctx->allJobsCompleted = 1;
1860 mtctx->frameEnded = 1;
1864 /* fill input buffer */
1865 if ( (!mtctx->jobReady)
1866 && (input->size > input->pos) ) { /* support NULL input */
1867 if (mtctx->inBuff.buffer.start == NULL) {
1868 assert(mtctx->inBuff.filled == 0); /* Can't fill an empty buffer */
1869 if (!ZSTDMT_tryGetInputRange(mtctx)) {
1870 /* It is only possible for this operation to fail if there are
1871 * still compression jobs ongoing.
1873 DEBUGLOG(5, "ZSTDMT_tryGetInputRange failed");
1874 assert(mtctx->doneJobID != mtctx->nextJobID);
1876 DEBUGLOG(5, "ZSTDMT_tryGetInputRange completed successfully : mtctx->inBuff.buffer.start = %p", mtctx->inBuff.buffer.start);
1878 if (mtctx->inBuff.buffer.start != NULL) {
1879 size_t const toLoad = MIN(input->size - input->pos, mtctx->targetSectionSize - mtctx->inBuff.filled);
1880 assert(mtctx->inBuff.buffer.capacity >= mtctx->targetSectionSize);
1881 DEBUGLOG(5, "ZSTDMT_compressStream_generic: adding %u bytes on top of %u to buffer of size %u",
1882 (U32)toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize);
1883 memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, toLoad);
1884 input->pos += toLoad;
1885 mtctx->inBuff.filled += toLoad;
1886 forwardInputProgress = toLoad>0;
1888 if ((input->pos < input->size) && (endOp == ZSTD_e_end))
1889 endOp = ZSTD_e_flush; /* can't end now : not all input consumed */
1892 if ( (mtctx->jobReady)
1893 || (mtctx->inBuff.filled >= mtctx->targetSectionSize) /* filled enough : let's compress */
1894 || ((endOp != ZSTD_e_continue) && (mtctx->inBuff.filled > 0)) /* something to flush : let's go */
1895 || ((endOp == ZSTD_e_end) && (!mtctx->frameEnded)) ) { /* must finish the frame with a zero-size block */
1896 size_t const jobSize = mtctx->inBuff.filled;
1897 assert(mtctx->inBuff.filled <= mtctx->targetSectionSize);
1898 CHECK_F( ZSTDMT_createCompressionJob(mtctx, jobSize, endOp) );
1901 /* check for potential compressed data ready to be flushed */
1902 { size_t const remainingToFlush = ZSTDMT_flushProduced(mtctx, output, !forwardInputProgress, endOp); /* block if there was no forward input progress */
1903 if (input->pos < input->size) return MAX(remainingToFlush, 1); /* input not consumed : do not end flush yet */
1904 DEBUGLOG(5, "end of ZSTDMT_compressStream_generic: remainingToFlush = %u", (U32)remainingToFlush);
1905 return remainingToFlush;
1910 size_t ZSTDMT_compressStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
1912 CHECK_F( ZSTDMT_compressStream_generic(mtctx, output, input, ZSTD_e_continue) );
1914 /* recommended next input size : fill current input buffer */
1915 return mtctx->targetSectionSize - mtctx->inBuff.filled; /* note : could be zero when input buffer is fully filled and no more availability to create new job */
1919 static size_t ZSTDMT_flushStream_internal(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_EndDirective endFrame)
1921 size_t const srcSize = mtctx->inBuff.filled;
1922 DEBUGLOG(5, "ZSTDMT_flushStream_internal");
1924 if ( mtctx->jobReady /* one job ready for a worker to pick up */
1925 || (srcSize > 0) /* still some data within input buffer */
1926 || ((endFrame==ZSTD_e_end) && !mtctx->frameEnded)) { /* need a last 0-size block to end frame */
1927 DEBUGLOG(5, "ZSTDMT_flushStream_internal : create a new job (%u bytes, end:%u)",
1928 (U32)srcSize, (U32)endFrame);
1929 CHECK_F( ZSTDMT_createCompressionJob(mtctx, srcSize, endFrame) );
1932 /* check if there is any data available to flush */
1933 return ZSTDMT_flushProduced(mtctx, output, 1 /* blockToFlush */, endFrame);
1937 size_t ZSTDMT_flushStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output)
1939 DEBUGLOG(5, "ZSTDMT_flushStream");
1940 if (mtctx->singleBlockingThread)
1941 return ZSTD_flushStream(mtctx->cctxPool->cctx[0], output);
1942 return ZSTDMT_flushStream_internal(mtctx, output, ZSTD_e_flush);
1945 size_t ZSTDMT_endStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output)
1947 DEBUGLOG(4, "ZSTDMT_endStream");
1948 if (mtctx->singleBlockingThread)
1949 return ZSTD_endStream(mtctx->cctxPool->cctx[0], output);
1950 return ZSTDMT_flushStream_internal(mtctx, output, ZSTD_e_end);