2 * Copyright (c) Meta Platforms, Inc. and affiliates.
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 /* ====== Compiler specifics ====== */
14 # pragma warning(disable : 4204) /* disable: C4204: non-constant aggregate initializer */
18 /* ====== Constants ====== */
19 #define ZSTDMT_OVERLAPLOG_DEFAULT 0
22 /* ====== Dependencies ====== */
23 #include "../common/allocations.h" /* ZSTD_customMalloc, ZSTD_customCalloc, ZSTD_customFree */
24 #include "../common/zstd_deps.h" /* ZSTD_memcpy, ZSTD_memset, INT_MAX, UINT_MAX */
25 #include "../common/mem.h" /* MEM_STATIC */
26 #include "../common/pool.h" /* threadpool */
27 #include "../common/threading.h" /* mutex */
28 #include "zstd_compress_internal.h" /* MIN, ERROR, ZSTD_*, ZSTD_highbit32 */
30 #include "zstdmt_compress.h"
32 /* Guards code to support resizing the SeqPool.
33 * We will want to resize the SeqPool to save memory in the future.
34 * Until then, comment the code out since it is unused.
36 #define ZSTD_RESIZE_SEQPOOL 0
38 /* ====== Debug ====== */
39 #if defined(DEBUGLEVEL) && (DEBUGLEVEL>=2) \
40 && !defined(_MSC_VER) \
41 && !defined(__MINGW32__)
45 # include <sys/times.h>
47 # define DEBUG_PRINTHEX(l,p,n) { \
49 for (debug_u=0; debug_u<(n); debug_u++) \
50 RAWLOG(l, "%02X ", ((const unsigned char*)(p))[debug_u]); \
54 static unsigned long long GetCurrentClockTimeMicroseconds(void)
56 static clock_t _ticksPerSecond = 0;
57 if (_ticksPerSecond <= 0) _ticksPerSecond = sysconf(_SC_CLK_TCK);
59 { struct tms junk; clock_t newTicks = (clock_t) times(&junk);
60 return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond);
63 #define MUTEX_WAIT_TIME_DLEVEL 6
64 #define ZSTD_PTHREAD_MUTEX_LOCK(mutex) { \
65 if (DEBUGLEVEL >= MUTEX_WAIT_TIME_DLEVEL) { \
66 unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds(); \
67 ZSTD_pthread_mutex_lock(mutex); \
68 { unsigned long long const afterTime = GetCurrentClockTimeMicroseconds(); \
69 unsigned long long const elapsedTime = (afterTime-beforeTime); \
70 if (elapsedTime > 1000) { /* or whatever threshold you like; I'm using 1 millisecond here */ \
71 DEBUGLOG(MUTEX_WAIT_TIME_DLEVEL, "Thread took %llu microseconds to acquire mutex %s \n", \
72 elapsedTime, #mutex); \
75 ZSTD_pthread_mutex_lock(mutex); \
81 # define ZSTD_PTHREAD_MUTEX_LOCK(m) ZSTD_pthread_mutex_lock(m)
82 # define DEBUG_PRINTHEX(l,p,n) {}
87 /* ===== Buffer Pool ===== */
88 /* a single Buffer Pool can be invoked from multiple threads in parallel */
90 typedef struct buffer_s {
95 static const buffer_t g_nullBuffer = { NULL, 0 };
97 typedef struct ZSTDMT_bufferPool_s {
98 ZSTD_pthread_mutex_t poolMutex;
100 unsigned totalBuffers;
103 buffer_t bTable[1]; /* variable size */
106 static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned maxNbBuffers, ZSTD_customMem cMem)
108 ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)ZSTD_customCalloc(
109 sizeof(ZSTDMT_bufferPool) + (maxNbBuffers-1) * sizeof(buffer_t), cMem);
110 if (bufPool==NULL) return NULL;
111 if (ZSTD_pthread_mutex_init(&bufPool->poolMutex, NULL)) {
112 ZSTD_customFree(bufPool, cMem);
115 bufPool->bufferSize = 64 KB;
116 bufPool->totalBuffers = maxNbBuffers;
117 bufPool->nbBuffers = 0;
118 bufPool->cMem = cMem;
122 static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool)
125 DEBUGLOG(3, "ZSTDMT_freeBufferPool (address:%08X)", (U32)(size_t)bufPool);
126 if (!bufPool) return; /* compatibility with free on NULL */
127 for (u=0; u<bufPool->totalBuffers; u++) {
128 DEBUGLOG(4, "free buffer %2u (address:%08X)", u, (U32)(size_t)bufPool->bTable[u].start);
129 ZSTD_customFree(bufPool->bTable[u].start, bufPool->cMem);
131 ZSTD_pthread_mutex_destroy(&bufPool->poolMutex);
132 ZSTD_customFree(bufPool, bufPool->cMem);
135 /* only works at initialization, not during compression */
136 static size_t ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool* bufPool)
138 size_t const poolSize = sizeof(*bufPool)
139 + (bufPool->totalBuffers - 1) * sizeof(buffer_t);
141 size_t totalBufferSize = 0;
142 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
143 for (u=0; u<bufPool->totalBuffers; u++)
144 totalBufferSize += bufPool->bTable[u].capacity;
145 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
147 return poolSize + totalBufferSize;
150 /* ZSTDMT_setBufferSize() :
151 * all future buffers provided by this buffer pool will have _at least_ this size
152 * note : it's better for all buffers to have same size,
153 * as they become freely interchangeable, reducing malloc/free usages and memory fragmentation */
154 static void ZSTDMT_setBufferSize(ZSTDMT_bufferPool* const bufPool, size_t const bSize)
156 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
157 DEBUGLOG(4, "ZSTDMT_setBufferSize: bSize = %u", (U32)bSize);
158 bufPool->bufferSize = bSize;
159 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
163 static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, unsigned maxNbBuffers)
165 if (srcBufPool==NULL) return NULL;
166 if (srcBufPool->totalBuffers >= maxNbBuffers) /* good enough */
168 /* need a larger buffer pool */
169 { ZSTD_customMem const cMem = srcBufPool->cMem;
170 size_t const bSize = srcBufPool->bufferSize; /* forward parameters */
171 ZSTDMT_bufferPool* newBufPool;
172 ZSTDMT_freeBufferPool(srcBufPool);
173 newBufPool = ZSTDMT_createBufferPool(maxNbBuffers, cMem);
174 if (newBufPool==NULL) return newBufPool;
175 ZSTDMT_setBufferSize(newBufPool, bSize);
180 /** ZSTDMT_getBuffer() :
181 * assumption : bufPool must be valid
182 * @return : a buffer, with start pointer and size
183 * note: allocation may fail, in this case, start==NULL and size==0 */
184 static buffer_t ZSTDMT_getBuffer(ZSTDMT_bufferPool* bufPool)
186 size_t const bSize = bufPool->bufferSize;
187 DEBUGLOG(5, "ZSTDMT_getBuffer: bSize = %u", (U32)bufPool->bufferSize);
188 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
189 if (bufPool->nbBuffers) { /* try to use an existing buffer */
190 buffer_t const buf = bufPool->bTable[--(bufPool->nbBuffers)];
191 size_t const availBufferSize = buf.capacity;
192 bufPool->bTable[bufPool->nbBuffers] = g_nullBuffer;
193 if ((availBufferSize >= bSize) & ((availBufferSize>>3) <= bSize)) {
194 /* large enough, but not too much */
195 DEBUGLOG(5, "ZSTDMT_getBuffer: provide buffer %u of size %u",
196 bufPool->nbBuffers, (U32)buf.capacity);
197 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
200 /* size conditions not respected : scratch this buffer, create new one */
201 DEBUGLOG(5, "ZSTDMT_getBuffer: existing buffer does not meet size conditions => freeing");
202 ZSTD_customFree(buf.start, bufPool->cMem);
204 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
205 /* create new buffer */
206 DEBUGLOG(5, "ZSTDMT_getBuffer: create a new buffer");
208 void* const start = ZSTD_customMalloc(bSize, bufPool->cMem);
209 buffer.start = start; /* note : start can be NULL if malloc fails ! */
210 buffer.capacity = (start==NULL) ? 0 : bSize;
212 DEBUGLOG(5, "ZSTDMT_getBuffer: buffer allocation failure !!");
214 DEBUGLOG(5, "ZSTDMT_getBuffer: created buffer of size %u", (U32)bSize);
220 #if ZSTD_RESIZE_SEQPOOL
221 /** ZSTDMT_resizeBuffer() :
222 * assumption : bufPool must be valid
223 * @return : a buffer that is at least the buffer pool buffer size.
224 * If a reallocation happens, the data in the input buffer is copied.
226 static buffer_t ZSTDMT_resizeBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buffer)
228 size_t const bSize = bufPool->bufferSize;
229 if (buffer.capacity < bSize) {
230 void* const start = ZSTD_customMalloc(bSize, bufPool->cMem);
232 newBuffer.start = start;
233 newBuffer.capacity = start == NULL ? 0 : bSize;
235 assert(newBuffer.capacity >= buffer.capacity);
236 ZSTD_memcpy(newBuffer.start, buffer.start, buffer.capacity);
237 DEBUGLOG(5, "ZSTDMT_resizeBuffer: created buffer of size %u", (U32)bSize);
240 DEBUGLOG(5, "ZSTDMT_resizeBuffer: buffer allocation failure !!");
246 /* store buffer for later re-use, up to pool capacity */
247 static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buf)
249 DEBUGLOG(5, "ZSTDMT_releaseBuffer");
250 if (buf.start == NULL) return; /* compatible with release on NULL */
251 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
252 if (bufPool->nbBuffers < bufPool->totalBuffers) {
253 bufPool->bTable[bufPool->nbBuffers++] = buf; /* stored for later use */
254 DEBUGLOG(5, "ZSTDMT_releaseBuffer: stored buffer of size %u in slot %u",
255 (U32)buf.capacity, (U32)(bufPool->nbBuffers-1));
256 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
259 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
260 /* Reached bufferPool capacity (should not happen) */
261 DEBUGLOG(5, "ZSTDMT_releaseBuffer: pool capacity reached => freeing ");
262 ZSTD_customFree(buf.start, bufPool->cMem);
265 /* We need 2 output buffers per worker since each dstBuff must be flushed after it is released.
266 * The 3 additional buffers are as follows:
267 * 1 buffer for input loading
268 * 1 buffer for "next input" when submitting current one
269 * 1 buffer stuck in queue */
270 #define BUF_POOL_MAX_NB_BUFFERS(nbWorkers) (2*(nbWorkers) + 3)
272 /* After a worker releases its rawSeqStore, it is immediately ready for reuse.
273 * So we only need one seq buffer per worker. */
274 #define SEQ_POOL_MAX_NB_BUFFERS(nbWorkers) (nbWorkers)
276 /* ===== Seq Pool Wrapper ====== */
278 typedef ZSTDMT_bufferPool ZSTDMT_seqPool;
280 static size_t ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool* seqPool)
282 return ZSTDMT_sizeof_bufferPool(seqPool);
285 static rawSeqStore_t bufferToSeq(buffer_t buffer)
287 rawSeqStore_t seq = kNullRawSeqStore;
288 seq.seq = (rawSeq*)buffer.start;
289 seq.capacity = buffer.capacity / sizeof(rawSeq);
293 static buffer_t seqToBuffer(rawSeqStore_t seq)
296 buffer.start = seq.seq;
297 buffer.capacity = seq.capacity * sizeof(rawSeq);
301 static rawSeqStore_t ZSTDMT_getSeq(ZSTDMT_seqPool* seqPool)
303 if (seqPool->bufferSize == 0) {
304 return kNullRawSeqStore;
306 return bufferToSeq(ZSTDMT_getBuffer(seqPool));
309 #if ZSTD_RESIZE_SEQPOOL
310 static rawSeqStore_t ZSTDMT_resizeSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
312 return bufferToSeq(ZSTDMT_resizeBuffer(seqPool, seqToBuffer(seq)));
316 static void ZSTDMT_releaseSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
318 ZSTDMT_releaseBuffer(seqPool, seqToBuffer(seq));
321 static void ZSTDMT_setNbSeq(ZSTDMT_seqPool* const seqPool, size_t const nbSeq)
323 ZSTDMT_setBufferSize(seqPool, nbSeq * sizeof(rawSeq));
326 static ZSTDMT_seqPool* ZSTDMT_createSeqPool(unsigned nbWorkers, ZSTD_customMem cMem)
328 ZSTDMT_seqPool* const seqPool = ZSTDMT_createBufferPool(SEQ_POOL_MAX_NB_BUFFERS(nbWorkers), cMem);
329 if (seqPool == NULL) return NULL;
330 ZSTDMT_setNbSeq(seqPool, 0);
334 static void ZSTDMT_freeSeqPool(ZSTDMT_seqPool* seqPool)
336 ZSTDMT_freeBufferPool(seqPool);
339 static ZSTDMT_seqPool* ZSTDMT_expandSeqPool(ZSTDMT_seqPool* pool, U32 nbWorkers)
341 return ZSTDMT_expandBufferPool(pool, SEQ_POOL_MAX_NB_BUFFERS(nbWorkers));
345 /* ===== CCtx Pool ===== */
346 /* a single CCtx Pool can be invoked from multiple threads in parallel */
349 ZSTD_pthread_mutex_t poolMutex;
353 ZSTD_CCtx* cctx[1]; /* variable size */
356 /* note : all CCtx borrowed from the pool should be released back to the pool _before_ freeing the pool */
357 static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
360 for (cid=0; cid<pool->totalCCtx; cid++)
361 ZSTD_freeCCtx(pool->cctx[cid]); /* note : compatible with free on NULL */
362 ZSTD_pthread_mutex_destroy(&pool->poolMutex);
363 ZSTD_customFree(pool, pool->cMem);
366 /* ZSTDMT_createCCtxPool() :
367 * implies nbWorkers >= 1 , checked by caller ZSTDMT_createCCtx() */
368 static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(int nbWorkers,
371 ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_customCalloc(
372 sizeof(ZSTDMT_CCtxPool) + (nbWorkers-1)*sizeof(ZSTD_CCtx*), cMem);
373 assert(nbWorkers > 0);
374 if (!cctxPool) return NULL;
375 if (ZSTD_pthread_mutex_init(&cctxPool->poolMutex, NULL)) {
376 ZSTD_customFree(cctxPool, cMem);
379 cctxPool->cMem = cMem;
380 cctxPool->totalCCtx = nbWorkers;
381 cctxPool->availCCtx = 1; /* at least one cctx for single-thread mode */
382 cctxPool->cctx[0] = ZSTD_createCCtx_advanced(cMem);
383 if (!cctxPool->cctx[0]) { ZSTDMT_freeCCtxPool(cctxPool); return NULL; }
384 DEBUGLOG(3, "cctxPool created, with %u workers", nbWorkers);
388 static ZSTDMT_CCtxPool* ZSTDMT_expandCCtxPool(ZSTDMT_CCtxPool* srcPool,
391 if (srcPool==NULL) return NULL;
392 if (nbWorkers <= srcPool->totalCCtx) return srcPool; /* good enough */
393 /* need a larger cctx pool */
394 { ZSTD_customMem const cMem = srcPool->cMem;
395 ZSTDMT_freeCCtxPool(srcPool);
396 return ZSTDMT_createCCtxPool(nbWorkers, cMem);
400 /* only works during initialization phase, not during compression */
401 static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
403 ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
404 { unsigned const nbWorkers = cctxPool->totalCCtx;
405 size_t const poolSize = sizeof(*cctxPool)
406 + (nbWorkers-1) * sizeof(ZSTD_CCtx*);
408 size_t totalCCtxSize = 0;
409 for (u=0; u<nbWorkers; u++) {
410 totalCCtxSize += ZSTD_sizeof_CCtx(cctxPool->cctx[u]);
412 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
413 assert(nbWorkers > 0);
414 return poolSize + totalCCtxSize;
418 static ZSTD_CCtx* ZSTDMT_getCCtx(ZSTDMT_CCtxPool* cctxPool)
420 DEBUGLOG(5, "ZSTDMT_getCCtx");
421 ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
422 if (cctxPool->availCCtx) {
423 cctxPool->availCCtx--;
424 { ZSTD_CCtx* const cctx = cctxPool->cctx[cctxPool->availCCtx];
425 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
428 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
429 DEBUGLOG(5, "create one more CCtx");
430 return ZSTD_createCCtx_advanced(cctxPool->cMem); /* note : can be NULL, when creation fails ! */
433 static void ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool* pool, ZSTD_CCtx* cctx)
435 if (cctx==NULL) return; /* compatibility with release on NULL */
436 ZSTD_pthread_mutex_lock(&pool->poolMutex);
437 if (pool->availCCtx < pool->totalCCtx)
438 pool->cctx[pool->availCCtx++] = cctx;
440 /* pool overflow : should not happen, since totalCCtx==nbWorkers */
441 DEBUGLOG(4, "CCtx pool overflow : free cctx");
444 ZSTD_pthread_mutex_unlock(&pool->poolMutex);
447 /* ==== Serial State ==== */
455 /* All variables in the struct are protected by mutex. */
456 ZSTD_pthread_mutex_t mutex;
457 ZSTD_pthread_cond_t cond;
458 ZSTD_CCtx_params params;
460 XXH64_state_t xxhState;
462 /* Protects ldmWindow.
463 * Must be acquired after the main mutex when acquiring both.
465 ZSTD_pthread_mutex_t ldmWindowMutex;
466 ZSTD_pthread_cond_t ldmWindowCond; /* Signaled when ldmWindow is updated */
467 ZSTD_window_t ldmWindow; /* A thread-safe copy of ldmState.window */
471 ZSTDMT_serialState_reset(serialState_t* serialState,
472 ZSTDMT_seqPool* seqPool,
473 ZSTD_CCtx_params params,
475 const void* dict, size_t const dictSize,
476 ZSTD_dictContentType_e dictContentType)
478 /* Adjust parameters */
479 if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
480 DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10);
481 ZSTD_ldm_adjustParameters(¶ms.ldmParams, ¶ms.cParams);
482 assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
483 assert(params.ldmParams.hashRateLog < 32);
485 ZSTD_memset(¶ms.ldmParams, 0, sizeof(params.ldmParams));
487 serialState->nextJobID = 0;
488 if (params.fParams.checksumFlag)
489 XXH64_reset(&serialState->xxhState, 0);
490 if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
491 ZSTD_customMem cMem = params.customMem;
492 unsigned const hashLog = params.ldmParams.hashLog;
493 size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
494 unsigned const bucketLog =
495 params.ldmParams.hashLog - params.ldmParams.bucketSizeLog;
496 unsigned const prevBucketLog =
497 serialState->params.ldmParams.hashLog -
498 serialState->params.ldmParams.bucketSizeLog;
499 size_t const numBuckets = (size_t)1 << bucketLog;
500 /* Size the seq pool tables */
501 ZSTDMT_setNbSeq(seqPool, ZSTD_ldm_getMaxNbSeq(params.ldmParams, jobSize));
502 /* Reset the window */
503 ZSTD_window_init(&serialState->ldmState.window);
504 /* Resize tables and output space if necessary. */
505 if (serialState->ldmState.hashTable == NULL || serialState->params.ldmParams.hashLog < hashLog) {
506 ZSTD_customFree(serialState->ldmState.hashTable, cMem);
507 serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_customMalloc(hashSize, cMem);
509 if (serialState->ldmState.bucketOffsets == NULL || prevBucketLog < bucketLog) {
510 ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
511 serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_customMalloc(numBuckets, cMem);
513 if (!serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets)
515 /* Zero the tables */
516 ZSTD_memset(serialState->ldmState.hashTable, 0, hashSize);
517 ZSTD_memset(serialState->ldmState.bucketOffsets, 0, numBuckets);
519 /* Update window state and fill hash table with dict */
520 serialState->ldmState.loadedDictEnd = 0;
522 if (dictContentType == ZSTD_dct_rawContent) {
523 BYTE const* const dictEnd = (const BYTE*)dict + dictSize;
524 ZSTD_window_update(&serialState->ldmState.window, dict, dictSize, /* forceNonContiguous */ 0);
525 ZSTD_ldm_fillHashTable(&serialState->ldmState, (const BYTE*)dict, dictEnd, ¶ms.ldmParams);
526 serialState->ldmState.loadedDictEnd = params.forceWindow ? 0 : (U32)(dictEnd - serialState->ldmState.window.base);
528 /* don't even load anything */
532 /* Initialize serialState's copy of ldmWindow. */
533 serialState->ldmWindow = serialState->ldmState.window;
536 serialState->params = params;
537 serialState->params.jobSize = (U32)jobSize;
541 static int ZSTDMT_serialState_init(serialState_t* serialState)
544 ZSTD_memset(serialState, 0, sizeof(*serialState));
545 initError |= ZSTD_pthread_mutex_init(&serialState->mutex, NULL);
546 initError |= ZSTD_pthread_cond_init(&serialState->cond, NULL);
547 initError |= ZSTD_pthread_mutex_init(&serialState->ldmWindowMutex, NULL);
548 initError |= ZSTD_pthread_cond_init(&serialState->ldmWindowCond, NULL);
552 static void ZSTDMT_serialState_free(serialState_t* serialState)
554 ZSTD_customMem cMem = serialState->params.customMem;
555 ZSTD_pthread_mutex_destroy(&serialState->mutex);
556 ZSTD_pthread_cond_destroy(&serialState->cond);
557 ZSTD_pthread_mutex_destroy(&serialState->ldmWindowMutex);
558 ZSTD_pthread_cond_destroy(&serialState->ldmWindowCond);
559 ZSTD_customFree(serialState->ldmState.hashTable, cMem);
560 ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
563 static void ZSTDMT_serialState_update(serialState_t* serialState,
564 ZSTD_CCtx* jobCCtx, rawSeqStore_t seqStore,
565 range_t src, unsigned jobID)
567 /* Wait for our turn */
568 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
569 while (serialState->nextJobID < jobID) {
570 DEBUGLOG(5, "wait for serialState->cond");
571 ZSTD_pthread_cond_wait(&serialState->cond, &serialState->mutex);
573 /* A future job may error and skip our job */
574 if (serialState->nextJobID == jobID) {
575 /* It is now our turn, do any processing necessary */
576 if (serialState->params.ldmParams.enableLdm == ZSTD_ps_enable) {
578 assert(seqStore.seq != NULL && seqStore.pos == 0 &&
579 seqStore.size == 0 && seqStore.capacity > 0);
580 assert(src.size <= serialState->params.jobSize);
581 ZSTD_window_update(&serialState->ldmState.window, src.start, src.size, /* forceNonContiguous */ 0);
582 error = ZSTD_ldm_generateSequences(
583 &serialState->ldmState, &seqStore,
584 &serialState->params.ldmParams, src.start, src.size);
585 /* We provide a large enough buffer to never fail. */
586 assert(!ZSTD_isError(error)); (void)error;
587 /* Update ldmWindow to match the ldmState.window and signal the main
588 * thread if it is waiting for a buffer.
590 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
591 serialState->ldmWindow = serialState->ldmState.window;
592 ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
593 ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
595 if (serialState->params.fParams.checksumFlag && src.size > 0)
596 XXH64_update(&serialState->xxhState, src.start, src.size);
598 /* Now it is the next jobs turn */
599 serialState->nextJobID++;
600 ZSTD_pthread_cond_broadcast(&serialState->cond);
601 ZSTD_pthread_mutex_unlock(&serialState->mutex);
603 if (seqStore.size > 0) {
604 size_t const err = ZSTD_referenceExternalSequences(
605 jobCCtx, seqStore.seq, seqStore.size);
606 assert(serialState->params.ldmParams.enableLdm == ZSTD_ps_enable);
607 assert(!ZSTD_isError(err));
612 static void ZSTDMT_serialState_ensureFinished(serialState_t* serialState,
613 unsigned jobID, size_t cSize)
615 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
616 if (serialState->nextJobID <= jobID) {
617 assert(ZSTD_isError(cSize)); (void)cSize;
618 DEBUGLOG(5, "Skipping past job %u because of error", jobID);
619 serialState->nextJobID = jobID + 1;
620 ZSTD_pthread_cond_broadcast(&serialState->cond);
622 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
623 ZSTD_window_clear(&serialState->ldmWindow);
624 ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
625 ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
627 ZSTD_pthread_mutex_unlock(&serialState->mutex);
632 /* ------------------------------------------ */
633 /* ===== Worker thread ===== */
634 /* ------------------------------------------ */
636 static const range_t kNullRange = { NULL, 0 };
639 size_t consumed; /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx */
640 size_t cSize; /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx, then set0 by mtctx */
641 ZSTD_pthread_mutex_t job_mutex; /* Thread-safe - used by mtctx and worker */
642 ZSTD_pthread_cond_t job_cond; /* Thread-safe - used by mtctx and worker */
643 ZSTDMT_CCtxPool* cctxPool; /* Thread-safe - used by mtctx and (all) workers */
644 ZSTDMT_bufferPool* bufPool; /* Thread-safe - used by mtctx and (all) workers */
645 ZSTDMT_seqPool* seqPool; /* Thread-safe - used by mtctx and (all) workers */
646 serialState_t* serial; /* Thread-safe - used by mtctx and (all) workers */
647 buffer_t dstBuff; /* set by worker (or mtctx), then read by worker & mtctx, then modified by mtctx => no barrier */
648 range_t prefix; /* set by mtctx, then read by worker & mtctx => no barrier */
649 range_t src; /* set by mtctx, then read by worker & mtctx => no barrier */
650 unsigned jobID; /* set by mtctx, then read by worker => no barrier */
651 unsigned firstJob; /* set by mtctx, then read by worker => no barrier */
652 unsigned lastJob; /* set by mtctx, then read by worker => no barrier */
653 ZSTD_CCtx_params params; /* set by mtctx, then read by worker => no barrier */
654 const ZSTD_CDict* cdict; /* set by mtctx, then read by worker => no barrier */
655 unsigned long long fullFrameSize; /* set by mtctx, then read by worker => no barrier */
656 size_t dstFlushed; /* used only by mtctx */
657 unsigned frameChecksumNeeded; /* used only by mtctx */
658 } ZSTDMT_jobDescription;
660 #define JOB_ERROR(e) { \
661 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex); \
663 ZSTD_pthread_mutex_unlock(&job->job_mutex); \
667 /* ZSTDMT_compressionJob() is a POOL_function type */
668 static void ZSTDMT_compressionJob(void* jobDescription)
670 ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)jobDescription;
671 ZSTD_CCtx_params jobParams = job->params; /* do not modify job->params ! copy it, modify the copy */
672 ZSTD_CCtx* const cctx = ZSTDMT_getCCtx(job->cctxPool);
673 rawSeqStore_t rawSeqStore = ZSTDMT_getSeq(job->seqPool);
674 buffer_t dstBuff = job->dstBuff;
675 size_t lastCBlockSize = 0;
678 if (cctx==NULL) JOB_ERROR(ERROR(memory_allocation));
679 if (dstBuff.start == NULL) { /* streaming job : doesn't provide a dstBuffer */
680 dstBuff = ZSTDMT_getBuffer(job->bufPool);
681 if (dstBuff.start==NULL) JOB_ERROR(ERROR(memory_allocation));
682 job->dstBuff = dstBuff; /* this value can be read in ZSTDMT_flush, when it copies the whole job */
684 if (jobParams.ldmParams.enableLdm == ZSTD_ps_enable && rawSeqStore.seq == NULL)
685 JOB_ERROR(ERROR(memory_allocation));
687 /* Don't compute the checksum for chunks, since we compute it externally,
688 * but write it in the header.
690 if (job->jobID != 0) jobParams.fParams.checksumFlag = 0;
691 /* Don't run LDM for the chunks, since we handle it externally */
692 jobParams.ldmParams.enableLdm = ZSTD_ps_disable;
693 /* Correct nbWorkers to 0. */
694 jobParams.nbWorkers = 0;
699 size_t const initError = ZSTD_compressBegin_advanced_internal(cctx, NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast, job->cdict, &jobParams, job->fullFrameSize);
700 assert(job->firstJob); /* only allowed for first job */
701 if (ZSTD_isError(initError)) JOB_ERROR(initError);
702 } else { /* srcStart points at reloaded section */
703 U64 const pledgedSrcSize = job->firstJob ? job->fullFrameSize : job->src.size;
704 { size_t const forceWindowError = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_forceMaxWindow, !job->firstJob);
705 if (ZSTD_isError(forceWindowError)) JOB_ERROR(forceWindowError);
707 if (!job->firstJob) {
708 size_t const err = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_deterministicRefPrefix, 0);
709 if (ZSTD_isError(err)) JOB_ERROR(err);
711 { size_t const initError = ZSTD_compressBegin_advanced_internal(cctx,
712 job->prefix.start, job->prefix.size, ZSTD_dct_rawContent, /* load dictionary in "content-only" mode (no header analysis) */
715 &jobParams, pledgedSrcSize);
716 if (ZSTD_isError(initError)) JOB_ERROR(initError);
719 /* Perform serial step as early as possible, but after CCtx initialization */
720 ZSTDMT_serialState_update(job->serial, cctx, rawSeqStore, job->src, job->jobID);
722 if (!job->firstJob) { /* flush and overwrite frame header when it's not first job */
723 size_t const hSize = ZSTD_compressContinue_public(cctx, dstBuff.start, dstBuff.capacity, job->src.start, 0);
724 if (ZSTD_isError(hSize)) JOB_ERROR(hSize);
725 DEBUGLOG(5, "ZSTDMT_compressionJob: flush and overwrite %u bytes of frame header (not first job)", (U32)hSize);
726 ZSTD_invalidateRepCodes(cctx);
730 { size_t const chunkSize = 4*ZSTD_BLOCKSIZE_MAX;
731 int const nbChunks = (int)((job->src.size + (chunkSize-1)) / chunkSize);
732 const BYTE* ip = (const BYTE*) job->src.start;
733 BYTE* const ostart = (BYTE*)dstBuff.start;
735 BYTE* oend = op + dstBuff.capacity;
737 if (sizeof(size_t) > sizeof(int)) assert(job->src.size < ((size_t)INT_MAX) * chunkSize); /* check overflow */
738 DEBUGLOG(5, "ZSTDMT_compressionJob: compress %u bytes in %i blocks", (U32)job->src.size, nbChunks);
739 assert(job->cSize == 0);
740 for (chunkNb = 1; chunkNb < nbChunks; chunkNb++) {
741 size_t const cSize = ZSTD_compressContinue_public(cctx, op, oend-op, ip, chunkSize);
742 if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
744 op += cSize; assert(op < oend);
746 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
748 job->consumed = chunkSize * chunkNb;
749 DEBUGLOG(5, "ZSTDMT_compressionJob: compress new block : cSize==%u bytes (total: %u)",
750 (U32)cSize, (U32)job->cSize);
751 ZSTD_pthread_cond_signal(&job->job_cond); /* warns some more data is ready to be flushed */
752 ZSTD_pthread_mutex_unlock(&job->job_mutex);
755 assert(chunkSize > 0);
756 assert((chunkSize & (chunkSize - 1)) == 0); /* chunkSize must be power of 2 for mask==(chunkSize-1) to work */
757 if ((nbChunks > 0) | job->lastJob /*must output a "last block" flag*/ ) {
758 size_t const lastBlockSize1 = job->src.size & (chunkSize-1);
759 size_t const lastBlockSize = ((lastBlockSize1==0) & (job->src.size>=chunkSize)) ? chunkSize : lastBlockSize1;
760 size_t const cSize = (job->lastJob) ?
761 ZSTD_compressEnd_public(cctx, op, oend-op, ip, lastBlockSize) :
762 ZSTD_compressContinue_public(cctx, op, oend-op, ip, lastBlockSize);
763 if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
764 lastCBlockSize = cSize;
766 if (!job->firstJob) {
767 /* Double check that we don't have an ext-dict, because then our
768 * repcode invalidation doesn't work.
770 assert(!ZSTD_window_hasExtDict(cctx->blockState.matchState.window));
772 ZSTD_CCtx_trace(cctx, 0);
775 ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize);
776 if (job->prefix.size > 0)
777 DEBUGLOG(5, "Finished with prefix: %zx", (size_t)job->prefix.start);
778 DEBUGLOG(5, "Finished with source: %zx", (size_t)job->src.start);
779 /* release resources */
780 ZSTDMT_releaseSeq(job->seqPool, rawSeqStore);
781 ZSTDMT_releaseCCtx(job->cctxPool, cctx);
783 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
784 if (ZSTD_isError(job->cSize)) assert(lastCBlockSize == 0);
785 job->cSize += lastCBlockSize;
786 job->consumed = job->src.size; /* when job->consumed == job->src.size , compression job is presumed completed */
787 ZSTD_pthread_cond_signal(&job->job_cond);
788 ZSTD_pthread_mutex_unlock(&job->job_mutex);
792 /* ------------------------------------------ */
793 /* ===== Multi-threaded compression ===== */
794 /* ------------------------------------------ */
797 range_t prefix; /* read-only non-owned prefix buffer */
803 BYTE* buffer; /* The round input buffer. All jobs get references
804 * to pieces of the buffer. ZSTDMT_tryGetInputRange()
805 * handles handing out job input buffers, and makes
806 * sure it doesn't overlap with any pieces still in use.
808 size_t capacity; /* The capacity of buffer. */
809 size_t pos; /* The position of the current inBuff in the round
810 * buffer. Updated past the end if the inBuff once
811 * the inBuff is sent to the worker thread.
816 static const roundBuff_t kNullRoundBuff = {NULL, 0, 0};
818 #define RSYNC_LENGTH 32
819 /* Don't create chunks smaller than the zstd block size.
820 * This stops us from regressing compression ratio too much,
821 * and ensures our output fits in ZSTD_compressBound().
823 * If this is shrunk < ZSTD_BLOCKSIZELOG_MIN then
824 * ZSTD_COMPRESSBOUND() will need to be updated.
826 #define RSYNC_MIN_BLOCK_LOG ZSTD_BLOCKSIZELOG_MAX
827 #define RSYNC_MIN_BLOCK_SIZE (1<<RSYNC_MIN_BLOCK_LOG)
835 struct ZSTDMT_CCtx_s {
837 ZSTDMT_jobDescription* jobs;
838 ZSTDMT_bufferPool* bufPool;
839 ZSTDMT_CCtxPool* cctxPool;
840 ZSTDMT_seqPool* seqPool;
841 ZSTD_CCtx_params params;
842 size_t targetSectionSize;
843 size_t targetPrefixSize;
844 int jobReady; /* 1 => one job is already prepared, but pool has shortage of workers. Don't create a new job. */
846 roundBuff_t roundBuff;
847 serialState_t serial;
853 unsigned allJobsCompleted;
854 unsigned long long frameContentSize;
855 unsigned long long consumed;
856 unsigned long long produced;
858 ZSTD_CDict* cdictLocal;
859 const ZSTD_CDict* cdict;
860 unsigned providedFactory: 1;
863 static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZSTD_customMem cMem)
866 if (jobTable == NULL) return;
867 for (jobNb=0; jobNb<nbJobs; jobNb++) {
868 ZSTD_pthread_mutex_destroy(&jobTable[jobNb].job_mutex);
869 ZSTD_pthread_cond_destroy(&jobTable[jobNb].job_cond);
871 ZSTD_customFree(jobTable, cMem);
874 /* ZSTDMT_allocJobsTable()
875 * allocate and init a job table.
876 * update *nbJobsPtr to next power of 2 value, as size of table */
877 static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
879 U32 const nbJobsLog2 = ZSTD_highbit32(*nbJobsPtr) + 1;
880 U32 const nbJobs = 1 << nbJobsLog2;
882 ZSTDMT_jobDescription* const jobTable = (ZSTDMT_jobDescription*)
883 ZSTD_customCalloc(nbJobs * sizeof(ZSTDMT_jobDescription), cMem);
885 if (jobTable==NULL) return NULL;
887 for (jobNb=0; jobNb<nbJobs; jobNb++) {
888 initError |= ZSTD_pthread_mutex_init(&jobTable[jobNb].job_mutex, NULL);
889 initError |= ZSTD_pthread_cond_init(&jobTable[jobNb].job_cond, NULL);
891 if (initError != 0) {
892 ZSTDMT_freeJobsTable(jobTable, nbJobs, cMem);
898 static size_t ZSTDMT_expandJobsTable (ZSTDMT_CCtx* mtctx, U32 nbWorkers) {
899 U32 nbJobs = nbWorkers + 2;
900 if (nbJobs > mtctx->jobIDMask+1) { /* need more job capacity */
901 ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
902 mtctx->jobIDMask = 0;
903 mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, mtctx->cMem);
904 if (mtctx->jobs==NULL) return ERROR(memory_allocation);
905 assert((nbJobs != 0) && ((nbJobs & (nbJobs - 1)) == 0)); /* ensure nbJobs is a power of 2 */
906 mtctx->jobIDMask = nbJobs - 1;
912 /* ZSTDMT_CCtxParam_setNbWorkers():
913 * Internal use only */
914 static size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers)
916 return ZSTD_CCtxParams_setParameter(params, ZSTD_c_nbWorkers, (int)nbWorkers);
919 MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool)
922 U32 nbJobs = nbWorkers + 2;
924 DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbWorkers = %u)", nbWorkers);
926 if (nbWorkers < 1) return NULL;
927 nbWorkers = MIN(nbWorkers , ZSTDMT_NBWORKERS_MAX);
928 if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
929 /* invalid custom allocator */
932 mtctx = (ZSTDMT_CCtx*) ZSTD_customCalloc(sizeof(ZSTDMT_CCtx), cMem);
933 if (!mtctx) return NULL;
934 ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
936 mtctx->allJobsCompleted = 1;
938 mtctx->factory = pool;
939 mtctx->providedFactory = 1;
942 mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem);
943 mtctx->providedFactory = 0;
945 mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem);
946 assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0); /* ensure nbJobs is a power of 2 */
947 mtctx->jobIDMask = nbJobs - 1;
948 mtctx->bufPool = ZSTDMT_createBufferPool(BUF_POOL_MAX_NB_BUFFERS(nbWorkers), cMem);
949 mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem);
950 mtctx->seqPool = ZSTDMT_createSeqPool(nbWorkers, cMem);
951 initError = ZSTDMT_serialState_init(&mtctx->serial);
952 mtctx->roundBuff = kNullRoundBuff;
953 if (!mtctx->factory | !mtctx->jobs | !mtctx->bufPool | !mtctx->cctxPool | !mtctx->seqPool | initError) {
954 ZSTDMT_freeCCtx(mtctx);
957 DEBUGLOG(3, "mt_cctx created, for %u threads", nbWorkers);
961 ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool)
963 #ifdef ZSTD_MULTITHREAD
964 return ZSTDMT_createCCtx_advanced_internal(nbWorkers, cMem, pool);
974 /* ZSTDMT_releaseAllJobResources() :
975 * note : ensure all workers are killed first ! */
976 static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx)
979 DEBUGLOG(3, "ZSTDMT_releaseAllJobResources");
980 for (jobID=0; jobID <= mtctx->jobIDMask; jobID++) {
981 /* Copy the mutex/cond out */
982 ZSTD_pthread_mutex_t const mutex = mtctx->jobs[jobID].job_mutex;
983 ZSTD_pthread_cond_t const cond = mtctx->jobs[jobID].job_cond;
985 DEBUGLOG(4, "job%02u: release dst address %08X", jobID, (U32)(size_t)mtctx->jobs[jobID].dstBuff.start);
986 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);
988 /* Clear the job description, but keep the mutex/cond */
989 ZSTD_memset(&mtctx->jobs[jobID], 0, sizeof(mtctx->jobs[jobID]));
990 mtctx->jobs[jobID].job_mutex = mutex;
991 mtctx->jobs[jobID].job_cond = cond;
993 mtctx->inBuff.buffer = g_nullBuffer;
994 mtctx->inBuff.filled = 0;
995 mtctx->allJobsCompleted = 1;
998 static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* mtctx)
1000 DEBUGLOG(4, "ZSTDMT_waitForAllJobsCompleted");
1001 while (mtctx->doneJobID < mtctx->nextJobID) {
1002 unsigned const jobID = mtctx->doneJobID & mtctx->jobIDMask;
1003 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
1004 while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
1005 DEBUGLOG(4, "waiting for jobCompleted signal from job %u", mtctx->doneJobID); /* we want to block when waiting for data to flush */
1006 ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
1008 ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
1013 size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx)
1015 if (mtctx==NULL) return 0; /* compatible with free on NULL */
1016 if (!mtctx->providedFactory)
1017 POOL_free(mtctx->factory); /* stop and free worker threads */
1018 ZSTDMT_releaseAllJobResources(mtctx); /* release job resources into pools first */
1019 ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
1020 ZSTDMT_freeBufferPool(mtctx->bufPool);
1021 ZSTDMT_freeCCtxPool(mtctx->cctxPool);
1022 ZSTDMT_freeSeqPool(mtctx->seqPool);
1023 ZSTDMT_serialState_free(&mtctx->serial);
1024 ZSTD_freeCDict(mtctx->cdictLocal);
1025 if (mtctx->roundBuff.buffer)
1026 ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem);
1027 ZSTD_customFree(mtctx, mtctx->cMem);
1031 size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx)
1033 if (mtctx == NULL) return 0; /* supports sizeof NULL */
1034 return sizeof(*mtctx)
1035 + POOL_sizeof(mtctx->factory)
1036 + ZSTDMT_sizeof_bufferPool(mtctx->bufPool)
1037 + (mtctx->jobIDMask+1) * sizeof(ZSTDMT_jobDescription)
1038 + ZSTDMT_sizeof_CCtxPool(mtctx->cctxPool)
1039 + ZSTDMT_sizeof_seqPool(mtctx->seqPool)
1040 + ZSTD_sizeof_CDict(mtctx->cdictLocal)
1041 + mtctx->roundBuff.capacity;
1045 /* ZSTDMT_resize() :
1046 * @return : error code if fails, 0 on success */
1047 static size_t ZSTDMT_resize(ZSTDMT_CCtx* mtctx, unsigned nbWorkers)
1049 if (POOL_resize(mtctx->factory, nbWorkers)) return ERROR(memory_allocation);
1050 FORWARD_IF_ERROR( ZSTDMT_expandJobsTable(mtctx, nbWorkers) , "");
1051 mtctx->bufPool = ZSTDMT_expandBufferPool(mtctx->bufPool, BUF_POOL_MAX_NB_BUFFERS(nbWorkers));
1052 if (mtctx->bufPool == NULL) return ERROR(memory_allocation);
1053 mtctx->cctxPool = ZSTDMT_expandCCtxPool(mtctx->cctxPool, nbWorkers);
1054 if (mtctx->cctxPool == NULL) return ERROR(memory_allocation);
1055 mtctx->seqPool = ZSTDMT_expandSeqPool(mtctx->seqPool, nbWorkers);
1056 if (mtctx->seqPool == NULL) return ERROR(memory_allocation);
1057 ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
1062 /*! ZSTDMT_updateCParams_whileCompressing() :
1063 * Updates a selected set of compression parameters, remaining compatible with currently active frame.
1064 * New parameters will be applied to next compression job. */
1065 void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams)
1067 U32 const saved_wlog = mtctx->params.cParams.windowLog; /* Do not modify windowLog while compressing */
1068 int const compressionLevel = cctxParams->compressionLevel;
1069 DEBUGLOG(5, "ZSTDMT_updateCParams_whileCompressing (level:%i)",
1071 mtctx->params.compressionLevel = compressionLevel;
1072 { ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
1073 cParams.windowLog = saved_wlog;
1074 mtctx->params.cParams = cParams;
1078 /* ZSTDMT_getFrameProgression():
1079 * tells how much data has been consumed (input) and produced (output) for current frame.
1080 * able to count progression inside worker threads.
1081 * Note : mutex will be acquired during statistics collection inside workers. */
1082 ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx)
1084 ZSTD_frameProgression fps;
1085 DEBUGLOG(5, "ZSTDMT_getFrameProgression");
1086 fps.ingested = mtctx->consumed + mtctx->inBuff.filled;
1087 fps.consumed = mtctx->consumed;
1088 fps.produced = fps.flushed = mtctx->produced;
1089 fps.currentJobID = mtctx->nextJobID;
1090 fps.nbActiveWorkers = 0;
1092 unsigned lastJobNb = mtctx->nextJobID + mtctx->jobReady; assert(mtctx->jobReady <= 1);
1093 DEBUGLOG(6, "ZSTDMT_getFrameProgression: jobs: from %u to <%u (jobReady:%u)",
1094 mtctx->doneJobID, lastJobNb, mtctx->jobReady)
1095 for (jobNb = mtctx->doneJobID ; jobNb < lastJobNb ; jobNb++) {
1096 unsigned const wJobID = jobNb & mtctx->jobIDMask;
1097 ZSTDMT_jobDescription* jobPtr = &mtctx->jobs[wJobID];
1098 ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
1099 { size_t const cResult = jobPtr->cSize;
1100 size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
1101 size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
1102 assert(flushed <= produced);
1103 fps.ingested += jobPtr->src.size;
1104 fps.consumed += jobPtr->consumed;
1105 fps.produced += produced;
1106 fps.flushed += flushed;
1107 fps.nbActiveWorkers += (jobPtr->consumed < jobPtr->src.size);
1109 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1116 size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx)
1119 unsigned const jobID = mtctx->doneJobID;
1120 assert(jobID <= mtctx->nextJobID);
1121 if (jobID == mtctx->nextJobID) return 0; /* no active job => nothing to flush */
1123 /* look into oldest non-fully-flushed job */
1124 { unsigned const wJobID = jobID & mtctx->jobIDMask;
1125 ZSTDMT_jobDescription* const jobPtr = &mtctx->jobs[wJobID];
1126 ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
1127 { size_t const cResult = jobPtr->cSize;
1128 size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
1129 size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
1130 assert(flushed <= produced);
1131 assert(jobPtr->consumed <= jobPtr->src.size);
1132 toFlush = produced - flushed;
1133 /* if toFlush==0, nothing is available to flush.
1134 * However, jobID is expected to still be active:
1135 * if jobID was already completed and fully flushed,
1136 * ZSTDMT_flushProduced() should have already moved onto next job.
1137 * Therefore, some input has not yet been consumed. */
1139 assert(jobPtr->consumed < jobPtr->src.size);
1142 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1149 /* ------------------------------------------ */
1150 /* ===== Multi-threaded compression ===== */
1151 /* ------------------------------------------ */
1153 static unsigned ZSTDMT_computeTargetJobLog(const ZSTD_CCtx_params* params)
1156 if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
1157 /* In Long Range Mode, the windowLog is typically oversized.
1158 * In which case, it's preferable to determine the jobSize
1159 * based on cycleLog instead. */
1160 jobLog = MAX(21, ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy) + 3);
1162 jobLog = MAX(20, params->cParams.windowLog + 2);
1164 return MIN(jobLog, (unsigned)ZSTDMT_JOBLOG_MAX);
1167 static int ZSTDMT_overlapLog_default(ZSTD_strategy strat)
1188 static int ZSTDMT_overlapLog(int ovlog, ZSTD_strategy strat)
1190 assert(0 <= ovlog && ovlog <= 9);
1191 if (ovlog == 0) return ZSTDMT_overlapLog_default(strat);
1195 static size_t ZSTDMT_computeOverlapSize(const ZSTD_CCtx_params* params)
1197 int const overlapRLog = 9 - ZSTDMT_overlapLog(params->overlapLog, params->cParams.strategy);
1198 int ovLog = (overlapRLog >= 8) ? 0 : (params->cParams.windowLog - overlapRLog);
1199 assert(0 <= overlapRLog && overlapRLog <= 8);
1200 if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
1201 /* In Long Range Mode, the windowLog is typically oversized.
1202 * In which case, it's preferable to determine the jobSize
1203 * based on chainLog instead.
1204 * Then, ovLog becomes a fraction of the jobSize, rather than windowSize */
1205 ovLog = MIN(params->cParams.windowLog, ZSTDMT_computeTargetJobLog(params) - 2)
1208 assert(0 <= ovLog && ovLog <= ZSTD_WINDOWLOG_MAX);
1209 DEBUGLOG(4, "overlapLog : %i", params->overlapLog);
1210 DEBUGLOG(4, "overlap size : %i", 1 << ovLog);
1211 return (ovLog==0) ? 0 : (size_t)1 << ovLog;
1214 /* ====================================== */
1215 /* ======= Streaming API ======= */
1216 /* ====================================== */
1218 size_t ZSTDMT_initCStream_internal(
1220 const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
1221 const ZSTD_CDict* cdict, ZSTD_CCtx_params params,
1222 unsigned long long pledgedSrcSize)
1224 DEBUGLOG(4, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u, nbWorkers=%u, cctxPool=%u)",
1225 (U32)pledgedSrcSize, params.nbWorkers, mtctx->cctxPool->totalCCtx);
1227 /* params supposed partially fully validated at this point */
1228 assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
1229 assert(!((dict) && (cdict))); /* either dict or cdict, not both */
1232 if (params.nbWorkers != mtctx->params.nbWorkers)
1233 FORWARD_IF_ERROR( ZSTDMT_resize(mtctx, params.nbWorkers) , "");
1235 if (params.jobSize != 0 && params.jobSize < ZSTDMT_JOBSIZE_MIN) params.jobSize = ZSTDMT_JOBSIZE_MIN;
1236 if (params.jobSize > (size_t)ZSTDMT_JOBSIZE_MAX) params.jobSize = (size_t)ZSTDMT_JOBSIZE_MAX;
1238 DEBUGLOG(4, "ZSTDMT_initCStream_internal: %u workers", params.nbWorkers);
1240 if (mtctx->allJobsCompleted == 0) { /* previous compression not correctly finished */
1241 ZSTDMT_waitForAllJobsCompleted(mtctx);
1242 ZSTDMT_releaseAllJobResources(mtctx);
1243 mtctx->allJobsCompleted = 1;
1246 mtctx->params = params;
1247 mtctx->frameContentSize = pledgedSrcSize;
1249 ZSTD_freeCDict(mtctx->cdictLocal);
1250 mtctx->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
1251 ZSTD_dlm_byCopy, dictContentType, /* note : a loadPrefix becomes an internal CDict */
1252 params.cParams, mtctx->cMem);
1253 mtctx->cdict = mtctx->cdictLocal;
1254 if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
1256 ZSTD_freeCDict(mtctx->cdictLocal);
1257 mtctx->cdictLocal = NULL;
1258 mtctx->cdict = cdict;
1261 mtctx->targetPrefixSize = ZSTDMT_computeOverlapSize(¶ms);
1262 DEBUGLOG(4, "overlapLog=%i => %u KB", params.overlapLog, (U32)(mtctx->targetPrefixSize>>10));
1263 mtctx->targetSectionSize = params.jobSize;
1264 if (mtctx->targetSectionSize == 0) {
1265 mtctx->targetSectionSize = 1ULL << ZSTDMT_computeTargetJobLog(¶ms);
1267 assert(mtctx->targetSectionSize <= (size_t)ZSTDMT_JOBSIZE_MAX);
1269 if (params.rsyncable) {
1270 /* Aim for the targetsectionSize as the average job size. */
1271 U32 const jobSizeKB = (U32)(mtctx->targetSectionSize >> 10);
1272 U32 const rsyncBits = (assert(jobSizeKB >= 1), ZSTD_highbit32(jobSizeKB) + 10);
1273 /* We refuse to create jobs < RSYNC_MIN_BLOCK_SIZE bytes, so make sure our
1274 * expected job size is at least 4x larger. */
1275 assert(rsyncBits >= RSYNC_MIN_BLOCK_LOG + 2);
1276 DEBUGLOG(4, "rsyncLog = %u", rsyncBits);
1277 mtctx->rsync.hash = 0;
1278 mtctx->rsync.hitMask = (1ULL << rsyncBits) - 1;
1279 mtctx->rsync.primePower = ZSTD_rollingHash_primePower(RSYNC_LENGTH);
1281 if (mtctx->targetSectionSize < mtctx->targetPrefixSize) mtctx->targetSectionSize = mtctx->targetPrefixSize; /* job size must be >= overlap size */
1282 DEBUGLOG(4, "Job Size : %u KB (note : set to %u)", (U32)(mtctx->targetSectionSize>>10), (U32)params.jobSize);
1283 DEBUGLOG(4, "inBuff Size : %u KB", (U32)(mtctx->targetSectionSize>>10));
1284 ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(mtctx->targetSectionSize));
1286 /* If ldm is enabled we need windowSize space. */
1287 size_t const windowSize = mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable ? (1U << mtctx->params.cParams.windowLog) : 0;
1288 /* Two buffers of slack, plus extra space for the overlap
1289 * This is the minimum slack that LDM works with. One extra because
1290 * flush might waste up to targetSectionSize-1 bytes. Another extra
1291 * for the overlap (if > 0), then one to fill which doesn't overlap
1292 * with the LDM window.
1294 size_t const nbSlackBuffers = 2 + (mtctx->targetPrefixSize > 0);
1295 size_t const slackSize = mtctx->targetSectionSize * nbSlackBuffers;
1296 /* Compute the total size, and always have enough slack */
1297 size_t const nbWorkers = MAX(mtctx->params.nbWorkers, 1);
1298 size_t const sectionsSize = mtctx->targetSectionSize * nbWorkers;
1299 size_t const capacity = MAX(windowSize, sectionsSize) + slackSize;
1300 if (mtctx->roundBuff.capacity < capacity) {
1301 if (mtctx->roundBuff.buffer)
1302 ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem);
1303 mtctx->roundBuff.buffer = (BYTE*)ZSTD_customMalloc(capacity, mtctx->cMem);
1304 if (mtctx->roundBuff.buffer == NULL) {
1305 mtctx->roundBuff.capacity = 0;
1306 return ERROR(memory_allocation);
1308 mtctx->roundBuff.capacity = capacity;
1311 DEBUGLOG(4, "roundBuff capacity : %u KB", (U32)(mtctx->roundBuff.capacity>>10));
1312 mtctx->roundBuff.pos = 0;
1313 mtctx->inBuff.buffer = g_nullBuffer;
1314 mtctx->inBuff.filled = 0;
1315 mtctx->inBuff.prefix = kNullRange;
1316 mtctx->doneJobID = 0;
1317 mtctx->nextJobID = 0;
1318 mtctx->frameEnded = 0;
1319 mtctx->allJobsCompleted = 0;
1320 mtctx->consumed = 0;
1321 mtctx->produced = 0;
1322 if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params, mtctx->targetSectionSize,
1323 dict, dictSize, dictContentType))
1324 return ERROR(memory_allocation);
1329 /* ZSTDMT_writeLastEmptyBlock()
1330 * Write a single empty block with an end-of-frame to finish a frame.
1331 * Job must be created from streaming variant.
1332 * This function is always successful if expected conditions are fulfilled.
1334 static void ZSTDMT_writeLastEmptyBlock(ZSTDMT_jobDescription* job)
1336 assert(job->lastJob == 1);
1337 assert(job->src.size == 0); /* last job is empty -> will be simplified into a last empty block */
1338 assert(job->firstJob == 0); /* cannot be first job, as it also needs to create frame header */
1339 assert(job->dstBuff.start == NULL); /* invoked from streaming variant only (otherwise, dstBuff might be user's output) */
1340 job->dstBuff = ZSTDMT_getBuffer(job->bufPool);
1341 if (job->dstBuff.start == NULL) {
1342 job->cSize = ERROR(memory_allocation);
1345 assert(job->dstBuff.capacity >= ZSTD_blockHeaderSize); /* no buffer should ever be that small */
1346 job->src = kNullRange;
1347 job->cSize = ZSTD_writeLastEmptyBlock(job->dstBuff.start, job->dstBuff.capacity);
1348 assert(!ZSTD_isError(job->cSize));
1349 assert(job->consumed == 0);
1352 static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* mtctx, size_t srcSize, ZSTD_EndDirective endOp)
1354 unsigned const jobID = mtctx->nextJobID & mtctx->jobIDMask;
1355 int const endFrame = (endOp == ZSTD_e_end);
1357 if (mtctx->nextJobID > mtctx->doneJobID + mtctx->jobIDMask) {
1358 DEBUGLOG(5, "ZSTDMT_createCompressionJob: will not create new job : table is full");
1359 assert((mtctx->nextJobID & mtctx->jobIDMask) == (mtctx->doneJobID & mtctx->jobIDMask));
1363 if (!mtctx->jobReady) {
1364 BYTE const* src = (BYTE const*)mtctx->inBuff.buffer.start;
1365 DEBUGLOG(5, "ZSTDMT_createCompressionJob: preparing job %u to compress %u bytes with %u preload ",
1366 mtctx->nextJobID, (U32)srcSize, (U32)mtctx->inBuff.prefix.size);
1367 mtctx->jobs[jobID].src.start = src;
1368 mtctx->jobs[jobID].src.size = srcSize;
1369 assert(mtctx->inBuff.filled >= srcSize);
1370 mtctx->jobs[jobID].prefix = mtctx->inBuff.prefix;
1371 mtctx->jobs[jobID].consumed = 0;
1372 mtctx->jobs[jobID].cSize = 0;
1373 mtctx->jobs[jobID].params = mtctx->params;
1374 mtctx->jobs[jobID].cdict = mtctx->nextJobID==0 ? mtctx->cdict : NULL;
1375 mtctx->jobs[jobID].fullFrameSize = mtctx->frameContentSize;
1376 mtctx->jobs[jobID].dstBuff = g_nullBuffer;
1377 mtctx->jobs[jobID].cctxPool = mtctx->cctxPool;
1378 mtctx->jobs[jobID].bufPool = mtctx->bufPool;
1379 mtctx->jobs[jobID].seqPool = mtctx->seqPool;
1380 mtctx->jobs[jobID].serial = &mtctx->serial;
1381 mtctx->jobs[jobID].jobID = mtctx->nextJobID;
1382 mtctx->jobs[jobID].firstJob = (mtctx->nextJobID==0);
1383 mtctx->jobs[jobID].lastJob = endFrame;
1384 mtctx->jobs[jobID].frameChecksumNeeded = mtctx->params.fParams.checksumFlag && endFrame && (mtctx->nextJobID>0);
1385 mtctx->jobs[jobID].dstFlushed = 0;
1387 /* Update the round buffer pos and clear the input buffer to be reset */
1388 mtctx->roundBuff.pos += srcSize;
1389 mtctx->inBuff.buffer = g_nullBuffer;
1390 mtctx->inBuff.filled = 0;
1391 /* Set the prefix */
1393 size_t const newPrefixSize = MIN(srcSize, mtctx->targetPrefixSize);
1394 mtctx->inBuff.prefix.start = src + srcSize - newPrefixSize;
1395 mtctx->inBuff.prefix.size = newPrefixSize;
1396 } else { /* endFrame==1 => no need for another input buffer */
1397 mtctx->inBuff.prefix = kNullRange;
1398 mtctx->frameEnded = endFrame;
1399 if (mtctx->nextJobID == 0) {
1400 /* single job exception : checksum is already calculated directly within worker thread */
1401 mtctx->params.fParams.checksumFlag = 0;
1405 && (mtctx->nextJobID>0)/*single job must also write frame header*/ ) {
1406 DEBUGLOG(5, "ZSTDMT_createCompressionJob: creating a last empty block to end frame");
1407 assert(endOp == ZSTD_e_end); /* only possible case : need to end the frame with an empty last block */
1408 ZSTDMT_writeLastEmptyBlock(mtctx->jobs + jobID);
1414 DEBUGLOG(5, "ZSTDMT_createCompressionJob: posting job %u : %u bytes (end:%u, jobNb == %u (mod:%u))",
1416 (U32)mtctx->jobs[jobID].src.size,
1417 mtctx->jobs[jobID].lastJob,
1420 if (POOL_tryAdd(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[jobID])) {
1422 mtctx->jobReady = 0;
1424 DEBUGLOG(5, "ZSTDMT_createCompressionJob: no worker available for job %u", mtctx->nextJobID);
1425 mtctx->jobReady = 1;
1431 /*! ZSTDMT_flushProduced() :
1432 * flush whatever data has been produced but not yet flushed in current job.
1433 * move to next job if current one is fully flushed.
1434 * `output` : `pos` will be updated with amount of data flushed .
1435 * `blockToFlush` : if >0, the function will block and wait if there is no data available to flush .
1436 * @return : amount of data remaining within internal buffer, 0 if no more, 1 if unknown but > 0, or an error code */
1437 static size_t ZSTDMT_flushProduced(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, unsigned blockToFlush, ZSTD_EndDirective end)
1439 unsigned const wJobID = mtctx->doneJobID & mtctx->jobIDMask;
1440 DEBUGLOG(5, "ZSTDMT_flushProduced (blocking:%u , job %u <= %u)",
1441 blockToFlush, mtctx->doneJobID, mtctx->nextJobID);
1442 assert(output->size >= output->pos);
1444 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
1446 && (mtctx->doneJobID < mtctx->nextJobID) ) {
1447 assert(mtctx->jobs[wJobID].dstFlushed <= mtctx->jobs[wJobID].cSize);
1448 while (mtctx->jobs[wJobID].dstFlushed == mtctx->jobs[wJobID].cSize) { /* nothing to flush */
1449 if (mtctx->jobs[wJobID].consumed == mtctx->jobs[wJobID].src.size) {
1450 DEBUGLOG(5, "job %u is completely consumed (%u == %u) => don't wait for cond, there will be none",
1451 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].consumed, (U32)mtctx->jobs[wJobID].src.size);
1454 DEBUGLOG(5, "waiting for something to flush from job %u (currently flushed: %u bytes)",
1455 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
1456 ZSTD_pthread_cond_wait(&mtctx->jobs[wJobID].job_cond, &mtctx->jobs[wJobID].job_mutex); /* block when nothing to flush but some to come */
1459 /* try to flush something */
1460 { size_t cSize = mtctx->jobs[wJobID].cSize; /* shared */
1461 size_t const srcConsumed = mtctx->jobs[wJobID].consumed; /* shared */
1462 size_t const srcSize = mtctx->jobs[wJobID].src.size; /* read-only, could be done after mutex lock, but no-declaration-after-statement */
1463 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1464 if (ZSTD_isError(cSize)) {
1465 DEBUGLOG(5, "ZSTDMT_flushProduced: job %u : compression error detected : %s",
1466 mtctx->doneJobID, ZSTD_getErrorName(cSize));
1467 ZSTDMT_waitForAllJobsCompleted(mtctx);
1468 ZSTDMT_releaseAllJobResources(mtctx);
1471 /* add frame checksum if necessary (can only happen once) */
1472 assert(srcConsumed <= srcSize);
1473 if ( (srcConsumed == srcSize) /* job completed -> worker no longer active */
1474 && mtctx->jobs[wJobID].frameChecksumNeeded ) {
1475 U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
1476 DEBUGLOG(4, "ZSTDMT_flushProduced: writing checksum : %08X \n", checksum);
1477 MEM_writeLE32((char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].cSize, checksum);
1479 mtctx->jobs[wJobID].cSize += 4; /* can write this shared value, as worker is no longer active */
1480 mtctx->jobs[wJobID].frameChecksumNeeded = 0;
1483 if (cSize > 0) { /* compression is ongoing or completed */
1484 size_t const toFlush = MIN(cSize - mtctx->jobs[wJobID].dstFlushed, output->size - output->pos);
1485 DEBUGLOG(5, "ZSTDMT_flushProduced: Flushing %u bytes from job %u (completion:%u/%u, generated:%u)",
1486 (U32)toFlush, mtctx->doneJobID, (U32)srcConsumed, (U32)srcSize, (U32)cSize);
1487 assert(mtctx->doneJobID < mtctx->nextJobID);
1488 assert(cSize >= mtctx->jobs[wJobID].dstFlushed);
1489 assert(mtctx->jobs[wJobID].dstBuff.start != NULL);
1491 ZSTD_memcpy((char*)output->dst + output->pos,
1492 (const char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].dstFlushed,
1495 output->pos += toFlush;
1496 mtctx->jobs[wJobID].dstFlushed += toFlush; /* can write : this value is only used by mtctx */
1498 if ( (srcConsumed == srcSize) /* job is completed */
1499 && (mtctx->jobs[wJobID].dstFlushed == cSize) ) { /* output buffer fully flushed => free this job position */
1500 DEBUGLOG(5, "Job %u completed (%u bytes), moving to next one",
1501 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
1502 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[wJobID].dstBuff);
1503 DEBUGLOG(5, "dstBuffer released");
1504 mtctx->jobs[wJobID].dstBuff = g_nullBuffer;
1505 mtctx->jobs[wJobID].cSize = 0; /* ensure this job slot is considered "not started" in future check */
1506 mtctx->consumed += srcSize;
1507 mtctx->produced += cSize;
1511 /* return value : how many bytes left in buffer ; fake it to 1 when unknown but >0 */
1512 if (cSize > mtctx->jobs[wJobID].dstFlushed) return (cSize - mtctx->jobs[wJobID].dstFlushed);
1513 if (srcSize > srcConsumed) return 1; /* current job not completely compressed */
1515 if (mtctx->doneJobID < mtctx->nextJobID) return 1; /* some more jobs ongoing */
1516 if (mtctx->jobReady) return 1; /* one job is ready to push, just not yet in the list */
1517 if (mtctx->inBuff.filled > 0) return 1; /* input is not empty, and still needs to be converted into a job */
1518 mtctx->allJobsCompleted = mtctx->frameEnded; /* all jobs are entirely flushed => if this one is last one, frame is completed */
1519 if (end == ZSTD_e_end) return !mtctx->frameEnded; /* for ZSTD_e_end, question becomes : is frame completed ? instead of : are internal buffers fully flushed ? */
1520 return 0; /* internal buffers fully flushed */
1524 * Returns the range of data used by the earliest job that is not yet complete.
1525 * If the data of the first job is broken up into two segments, we cover both
1528 static range_t ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx)
1530 unsigned const firstJobID = mtctx->doneJobID;
1531 unsigned const lastJobID = mtctx->nextJobID;
1534 for (jobID = firstJobID; jobID < lastJobID; ++jobID) {
1535 unsigned const wJobID = jobID & mtctx->jobIDMask;
1538 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
1539 consumed = mtctx->jobs[wJobID].consumed;
1540 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1542 if (consumed < mtctx->jobs[wJobID].src.size) {
1543 range_t range = mtctx->jobs[wJobID].prefix;
1544 if (range.size == 0) {
1546 range = mtctx->jobs[wJobID].src;
1548 /* Job source in multiple segments not supported yet */
1549 assert(range.start <= mtctx->jobs[wJobID].src.start);
1557 * Returns non-zero iff buffer and range overlap.
1559 static int ZSTDMT_isOverlapped(buffer_t buffer, range_t range)
1561 BYTE const* const bufferStart = (BYTE const*)buffer.start;
1562 BYTE const* const rangeStart = (BYTE const*)range.start;
1564 if (rangeStart == NULL || bufferStart == NULL)
1568 BYTE const* const bufferEnd = bufferStart + buffer.capacity;
1569 BYTE const* const rangeEnd = rangeStart + range.size;
1571 /* Empty ranges cannot overlap */
1572 if (bufferStart == bufferEnd || rangeStart == rangeEnd)
1575 return bufferStart < rangeEnd && rangeStart < bufferEnd;
1579 static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window)
1584 DEBUGLOG(5, "ZSTDMT_doesOverlapWindow");
1585 extDict.start = window.dictBase + window.lowLimit;
1586 extDict.size = window.dictLimit - window.lowLimit;
1588 prefix.start = window.base + window.dictLimit;
1589 prefix.size = window.nextSrc - (window.base + window.dictLimit);
1590 DEBUGLOG(5, "extDict [0x%zx, 0x%zx)",
1591 (size_t)extDict.start,
1592 (size_t)extDict.start + extDict.size);
1593 DEBUGLOG(5, "prefix [0x%zx, 0x%zx)",
1594 (size_t)prefix.start,
1595 (size_t)prefix.start + prefix.size);
1597 return ZSTDMT_isOverlapped(buffer, extDict)
1598 || ZSTDMT_isOverlapped(buffer, prefix);
1601 static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, buffer_t buffer)
1603 if (mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable) {
1604 ZSTD_pthread_mutex_t* mutex = &mtctx->serial.ldmWindowMutex;
1605 DEBUGLOG(5, "ZSTDMT_waitForLdmComplete");
1606 DEBUGLOG(5, "source [0x%zx, 0x%zx)",
1607 (size_t)buffer.start,
1608 (size_t)buffer.start + buffer.capacity);
1609 ZSTD_PTHREAD_MUTEX_LOCK(mutex);
1610 while (ZSTDMT_doesOverlapWindow(buffer, mtctx->serial.ldmWindow)) {
1611 DEBUGLOG(5, "Waiting for LDM to finish...");
1612 ZSTD_pthread_cond_wait(&mtctx->serial.ldmWindowCond, mutex);
1614 DEBUGLOG(6, "Done waiting for LDM to finish");
1615 ZSTD_pthread_mutex_unlock(mutex);
1620 * Attempts to set the inBuff to the next section to fill.
1621 * If any part of the new section is still in use we give up.
1622 * Returns non-zero if the buffer is filled.
1624 static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx)
1626 range_t const inUse = ZSTDMT_getInputDataInUse(mtctx);
1627 size_t const spaceLeft = mtctx->roundBuff.capacity - mtctx->roundBuff.pos;
1628 size_t const target = mtctx->targetSectionSize;
1631 DEBUGLOG(5, "ZSTDMT_tryGetInputRange");
1632 assert(mtctx->inBuff.buffer.start == NULL);
1633 assert(mtctx->roundBuff.capacity >= target);
1635 if (spaceLeft < target) {
1636 /* ZSTD_invalidateRepCodes() doesn't work for extDict variants.
1637 * Simply copy the prefix to the beginning in that case.
1639 BYTE* const start = (BYTE*)mtctx->roundBuff.buffer;
1640 size_t const prefixSize = mtctx->inBuff.prefix.size;
1642 buffer.start = start;
1643 buffer.capacity = prefixSize;
1644 if (ZSTDMT_isOverlapped(buffer, inUse)) {
1645 DEBUGLOG(5, "Waiting for buffer...");
1648 ZSTDMT_waitForLdmComplete(mtctx, buffer);
1649 ZSTD_memmove(start, mtctx->inBuff.prefix.start, prefixSize);
1650 mtctx->inBuff.prefix.start = start;
1651 mtctx->roundBuff.pos = prefixSize;
1653 buffer.start = mtctx->roundBuff.buffer + mtctx->roundBuff.pos;
1654 buffer.capacity = target;
1656 if (ZSTDMT_isOverlapped(buffer, inUse)) {
1657 DEBUGLOG(5, "Waiting for buffer...");
1660 assert(!ZSTDMT_isOverlapped(buffer, mtctx->inBuff.prefix));
1662 ZSTDMT_waitForLdmComplete(mtctx, buffer);
1664 DEBUGLOG(5, "Using prefix range [%zx, %zx)",
1665 (size_t)mtctx->inBuff.prefix.start,
1666 (size_t)mtctx->inBuff.prefix.start + mtctx->inBuff.prefix.size);
1667 DEBUGLOG(5, "Using source range [%zx, %zx)",
1668 (size_t)buffer.start,
1669 (size_t)buffer.start + buffer.capacity);
1672 mtctx->inBuff.buffer = buffer;
1673 mtctx->inBuff.filled = 0;
1674 assert(mtctx->roundBuff.pos + buffer.capacity <= mtctx->roundBuff.capacity);
1679 size_t toLoad; /* The number of bytes to load from the input. */
1680 int flush; /* Boolean declaring if we must flush because we found a synchronization point. */
1684 * Searches through the input for a synchronization point. If one is found, we
1685 * will instruct the caller to flush, and return the number of bytes to load.
1686 * Otherwise, we will load as many bytes as possible and instruct the caller
1687 * to continue as normal.
1690 findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
1692 BYTE const* const istart = (BYTE const*)input.src + input.pos;
1693 U64 const primePower = mtctx->rsync.primePower;
1694 U64 const hitMask = mtctx->rsync.hitMask;
1696 syncPoint_t syncPoint;
1701 syncPoint.toLoad = MIN(input.size - input.pos, mtctx->targetSectionSize - mtctx->inBuff.filled);
1702 syncPoint.flush = 0;
1703 if (!mtctx->params.rsyncable)
1704 /* Rsync is disabled. */
1706 if (mtctx->inBuff.filled + input.size - input.pos < RSYNC_MIN_BLOCK_SIZE)
1707 /* We don't emit synchronization points if it would produce too small blocks.
1708 * We don't have enough input to find a synchronization point, so don't look.
1711 if (mtctx->inBuff.filled + syncPoint.toLoad < RSYNC_LENGTH)
1712 /* Not enough to compute the hash.
1713 * We will miss any synchronization points in this RSYNC_LENGTH byte
1714 * window. However, since it depends only in the internal buffers, if the
1715 * state is already synchronized, we will remain synchronized.
1716 * Additionally, the probability that we miss a synchronization point is
1717 * low: RSYNC_LENGTH / targetSectionSize.
1720 /* Initialize the loop variables. */
1721 if (mtctx->inBuff.filled < RSYNC_MIN_BLOCK_SIZE) {
1722 /* We don't need to scan the first RSYNC_MIN_BLOCK_SIZE positions
1723 * because they can't possibly be a sync point. So we can start
1724 * part way through the input buffer.
1726 pos = RSYNC_MIN_BLOCK_SIZE - mtctx->inBuff.filled;
1727 if (pos >= RSYNC_LENGTH) {
1728 prev = istart + pos - RSYNC_LENGTH;
1729 hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
1731 assert(mtctx->inBuff.filled >= RSYNC_LENGTH);
1732 prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
1733 hash = ZSTD_rollingHash_compute(prev + pos, (RSYNC_LENGTH - pos));
1734 hash = ZSTD_rollingHash_append(hash, istart, pos);
1737 /* We have enough bytes buffered to initialize the hash,
1738 * and have processed enough bytes to find a sync point.
1739 * Start scanning at the beginning of the input.
1741 assert(mtctx->inBuff.filled >= RSYNC_MIN_BLOCK_SIZE);
1742 assert(RSYNC_MIN_BLOCK_SIZE >= RSYNC_LENGTH);
1744 prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
1745 hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
1746 if ((hash & hitMask) == hitMask) {
1747 /* We're already at a sync point so don't load any more until
1748 * we're able to flush this sync point.
1749 * This likely happened because the job table was full so we
1750 * couldn't add our job.
1752 syncPoint.toLoad = 0;
1753 syncPoint.flush = 1;
1757 /* Starting with the hash of the previous RSYNC_LENGTH bytes, roll
1758 * through the input. If we hit a synchronization point, then cut the
1759 * job off, and tell the compressor to flush the job. Otherwise, load
1760 * all the bytes and continue as normal.
1761 * If we go too long without a synchronization point (targetSectionSize)
1762 * then a block will be emitted anyways, but this is okay, since if we
1763 * are already synchronized we will remain synchronized.
1765 assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
1766 for (; pos < syncPoint.toLoad; ++pos) {
1767 BYTE const toRemove = pos < RSYNC_LENGTH ? prev[pos] : istart[pos - RSYNC_LENGTH];
1768 /* This assert is very expensive, and Debian compiles with asserts enabled.
1769 * So disable it for now. We can get similar coverage by checking it at the
1770 * beginning & end of the loop.
1771 * assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
1773 hash = ZSTD_rollingHash_rotate(hash, toRemove, istart[pos], primePower);
1774 assert(mtctx->inBuff.filled + pos >= RSYNC_MIN_BLOCK_SIZE);
1775 if ((hash & hitMask) == hitMask) {
1776 syncPoint.toLoad = pos + 1;
1777 syncPoint.flush = 1;
1778 ++pos; /* for assert */
1782 assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
1786 size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx)
1788 size_t hintInSize = mtctx->targetSectionSize - mtctx->inBuff.filled;
1789 if (hintInSize==0) hintInSize = mtctx->targetSectionSize;
1793 /** ZSTDMT_compressStream_generic() :
1794 * internal use only - exposed to be invoked from zstd_compress.c
1795 * assumption : output and input are valid (pos <= size)
1796 * @return : minimum amount of data remaining to flush, 0 if none */
1797 size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
1798 ZSTD_outBuffer* output,
1799 ZSTD_inBuffer* input,
1800 ZSTD_EndDirective endOp)
1802 unsigned forwardInputProgress = 0;
1803 DEBUGLOG(5, "ZSTDMT_compressStream_generic (endOp=%u, srcSize=%u)",
1804 (U32)endOp, (U32)(input->size - input->pos));
1805 assert(output->pos <= output->size);
1806 assert(input->pos <= input->size);
1808 if ((mtctx->frameEnded) && (endOp==ZSTD_e_continue)) {
1809 /* current frame being ended. Only flush/end are allowed */
1810 return ERROR(stage_wrong);
1813 /* fill input buffer */
1814 if ( (!mtctx->jobReady)
1815 && (input->size > input->pos) ) { /* support NULL input */
1816 if (mtctx->inBuff.buffer.start == NULL) {
1817 assert(mtctx->inBuff.filled == 0); /* Can't fill an empty buffer */
1818 if (!ZSTDMT_tryGetInputRange(mtctx)) {
1819 /* It is only possible for this operation to fail if there are
1820 * still compression jobs ongoing.
1822 DEBUGLOG(5, "ZSTDMT_tryGetInputRange failed");
1823 assert(mtctx->doneJobID != mtctx->nextJobID);
1825 DEBUGLOG(5, "ZSTDMT_tryGetInputRange completed successfully : mtctx->inBuff.buffer.start = %p", mtctx->inBuff.buffer.start);
1827 if (mtctx->inBuff.buffer.start != NULL) {
1828 syncPoint_t const syncPoint = findSynchronizationPoint(mtctx, *input);
1829 if (syncPoint.flush && endOp == ZSTD_e_continue) {
1830 endOp = ZSTD_e_flush;
1832 assert(mtctx->inBuff.buffer.capacity >= mtctx->targetSectionSize);
1833 DEBUGLOG(5, "ZSTDMT_compressStream_generic: adding %u bytes on top of %u to buffer of size %u",
1834 (U32)syncPoint.toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize);
1835 ZSTD_memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, syncPoint.toLoad);
1836 input->pos += syncPoint.toLoad;
1837 mtctx->inBuff.filled += syncPoint.toLoad;
1838 forwardInputProgress = syncPoint.toLoad>0;
1841 if ((input->pos < input->size) && (endOp == ZSTD_e_end)) {
1842 /* Can't end yet because the input is not fully consumed.
1843 * We are in one of these cases:
1844 * - mtctx->inBuff is NULL & empty: we couldn't get an input buffer so don't create a new job.
1845 * - We filled the input buffer: flush this job but don't end the frame.
1846 * - We hit a synchronization point: flush this job but don't end the frame.
1848 assert(mtctx->inBuff.filled == 0 || mtctx->inBuff.filled == mtctx->targetSectionSize || mtctx->params.rsyncable);
1849 endOp = ZSTD_e_flush;
1852 if ( (mtctx->jobReady)
1853 || (mtctx->inBuff.filled >= mtctx->targetSectionSize) /* filled enough : let's compress */
1854 || ((endOp != ZSTD_e_continue) && (mtctx->inBuff.filled > 0)) /* something to flush : let's go */
1855 || ((endOp == ZSTD_e_end) && (!mtctx->frameEnded)) ) { /* must finish the frame with a zero-size block */
1856 size_t const jobSize = mtctx->inBuff.filled;
1857 assert(mtctx->inBuff.filled <= mtctx->targetSectionSize);
1858 FORWARD_IF_ERROR( ZSTDMT_createCompressionJob(mtctx, jobSize, endOp) , "");
1861 /* check for potential compressed data ready to be flushed */
1862 { size_t const remainingToFlush = ZSTDMT_flushProduced(mtctx, output, !forwardInputProgress, endOp); /* block if there was no forward input progress */
1863 if (input->pos < input->size) return MAX(remainingToFlush, 1); /* input not consumed : do not end flush yet */
1864 DEBUGLOG(5, "end of ZSTDMT_compressStream_generic: remainingToFlush = %u", (U32)remainingToFlush);
1865 return remainingToFlush;