blob: 9e537b884853e1e582e4f84a65c0781be9ecee5c [file] [log] [blame]
Holger Hildebrandtfa074992020-03-27 15:42:06 +00001/*
2 * Copyright (c) 2016-present, Yann Collet, Facebook, Inc.
3 * All rights reserved.
4 *
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.
9 */
10
11
12/* ====== Compiler specifics ====== */
13#if defined(_MSC_VER)
14# pragma warning(disable : 4204) /* disable: C4204: non-constant aggregate initializer */
15#endif
16
17
18/* ====== Constants ====== */
19#define ZSTDMT_OVERLAPLOG_DEFAULT 0
20
21
22/* ====== Dependencies ====== */
23#include <string.h> /* memcpy, memset */
24#include <limits.h> /* INT_MAX, UINT_MAX */
25#include "mem.h" /* MEM_STATIC */
26#include "pool.h" /* threadpool */
27#include "threading.h" /* mutex */
28#include "zstd_compress_internal.h" /* MIN, ERROR, ZSTD_*, ZSTD_highbit32 */
29#include "zstd_ldm.h"
30#include "zstdmt_compress.h"
31
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.
35 */
36#define ZSTD_RESIZE_SEQPOOL 0
37
38/* ====== Debug ====== */
39#if defined(DEBUGLEVEL) && (DEBUGLEVEL>=2) \
40 && !defined(_MSC_VER) \
41 && !defined(__MINGW32__)
42
43# include <stdio.h>
44# include <unistd.h>
45# include <sys/times.h>
46
47# define DEBUG_PRINTHEX(l,p,n) { \
48 unsigned debug_u; \
49 for (debug_u=0; debug_u<(n); debug_u++) \
50 RAWLOG(l, "%02X ", ((const unsigned char*)(p))[debug_u]); \
51 RAWLOG(l, " \n"); \
52}
53
54static unsigned long long GetCurrentClockTimeMicroseconds(void)
55{
56 static clock_t _ticksPerSecond = 0;
57 if (_ticksPerSecond <= 0) _ticksPerSecond = sysconf(_SC_CLK_TCK);
58
59 { struct tms junk; clock_t newTicks = (clock_t) times(&junk);
60 return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond);
61} }
62
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); \
73 } } \
74 } else { \
75 ZSTD_pthread_mutex_lock(mutex); \
76 } \
77}
78
79#else
80
81# define ZSTD_PTHREAD_MUTEX_LOCK(m) ZSTD_pthread_mutex_lock(m)
82# define DEBUG_PRINTHEX(l,p,n) {}
83
84#endif
85
86
87/* ===== Buffer Pool ===== */
88/* a single Buffer Pool can be invoked from multiple threads in parallel */
89
90typedef struct buffer_s {
91 void* start;
92 size_t capacity;
93} buffer_t;
94
95static const buffer_t g_nullBuffer = { NULL, 0 };
96
97typedef struct ZSTDMT_bufferPool_s {
98 ZSTD_pthread_mutex_t poolMutex;
99 size_t bufferSize;
100 unsigned totalBuffers;
101 unsigned nbBuffers;
102 ZSTD_customMem cMem;
103 buffer_t bTable[1]; /* variable size */
104} ZSTDMT_bufferPool;
105
106static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbWorkers, ZSTD_customMem cMem)
107{
108 unsigned const maxNbBuffers = 2*nbWorkers + 3;
109 ZSTDMT_bufferPool* const bufPool = (ZSTDMT_bufferPool*)ZSTD_calloc(
110 sizeof(ZSTDMT_bufferPool) + (maxNbBuffers-1) * sizeof(buffer_t), cMem);
111 if (bufPool==NULL) return NULL;
112 if (ZSTD_pthread_mutex_init(&bufPool->poolMutex, NULL)) {
113 ZSTD_free(bufPool, cMem);
114 return NULL;
115 }
116 bufPool->bufferSize = 64 KB;
117 bufPool->totalBuffers = maxNbBuffers;
118 bufPool->nbBuffers = 0;
119 bufPool->cMem = cMem;
120 return bufPool;
121}
122
123static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool)
124{
125 unsigned u;
126 DEBUGLOG(3, "ZSTDMT_freeBufferPool (address:%08X)", (U32)(size_t)bufPool);
127 if (!bufPool) return; /* compatibility with free on NULL */
128 for (u=0; u<bufPool->totalBuffers; u++) {
129 DEBUGLOG(4, "free buffer %2u (address:%08X)", u, (U32)(size_t)bufPool->bTable[u].start);
130 ZSTD_free(bufPool->bTable[u].start, bufPool->cMem);
131 }
132 ZSTD_pthread_mutex_destroy(&bufPool->poolMutex);
133 ZSTD_free(bufPool, bufPool->cMem);
134}
135
136/* only works at initialization, not during compression */
137static size_t ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool* bufPool)
138{
139 size_t const poolSize = sizeof(*bufPool)
140 + (bufPool->totalBuffers - 1) * sizeof(buffer_t);
141 unsigned u;
142 size_t totalBufferSize = 0;
143 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
144 for (u=0; u<bufPool->totalBuffers; u++)
145 totalBufferSize += bufPool->bTable[u].capacity;
146 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
147
148 return poolSize + totalBufferSize;
149}
150
151/* ZSTDMT_setBufferSize() :
152 * all future buffers provided by this buffer pool will have _at least_ this size
153 * note : it's better for all buffers to have same size,
154 * as they become freely interchangeable, reducing malloc/free usages and memory fragmentation */
155static void ZSTDMT_setBufferSize(ZSTDMT_bufferPool* const bufPool, size_t const bSize)
156{
157 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
158 DEBUGLOG(4, "ZSTDMT_setBufferSize: bSize = %u", (U32)bSize);
159 bufPool->bufferSize = bSize;
160 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
161}
162
163
164static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, U32 nbWorkers)
165{
166 unsigned const maxNbBuffers = 2*nbWorkers + 3;
167 if (srcBufPool==NULL) return NULL;
168 if (srcBufPool->totalBuffers >= maxNbBuffers) /* good enough */
169 return srcBufPool;
170 /* need a larger buffer pool */
171 { ZSTD_customMem const cMem = srcBufPool->cMem;
172 size_t const bSize = srcBufPool->bufferSize; /* forward parameters */
173 ZSTDMT_bufferPool* newBufPool;
174 ZSTDMT_freeBufferPool(srcBufPool);
175 newBufPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
176 if (newBufPool==NULL) return newBufPool;
177 ZSTDMT_setBufferSize(newBufPool, bSize);
178 return newBufPool;
179 }
180}
181
182/** ZSTDMT_getBuffer() :
183 * assumption : bufPool must be valid
184 * @return : a buffer, with start pointer and size
185 * note: allocation may fail, in this case, start==NULL and size==0 */
186static buffer_t ZSTDMT_getBuffer(ZSTDMT_bufferPool* bufPool)
187{
188 size_t const bSize = bufPool->bufferSize;
189 DEBUGLOG(5, "ZSTDMT_getBuffer: bSize = %u", (U32)bufPool->bufferSize);
190 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
191 if (bufPool->nbBuffers) { /* try to use an existing buffer */
192 buffer_t const buf = bufPool->bTable[--(bufPool->nbBuffers)];
193 size_t const availBufferSize = buf.capacity;
194 bufPool->bTable[bufPool->nbBuffers] = g_nullBuffer;
195 if ((availBufferSize >= bSize) & ((availBufferSize>>3) <= bSize)) {
196 /* large enough, but not too much */
197 DEBUGLOG(5, "ZSTDMT_getBuffer: provide buffer %u of size %u",
198 bufPool->nbBuffers, (U32)buf.capacity);
199 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
200 return buf;
201 }
202 /* size conditions not respected : scratch this buffer, create new one */
203 DEBUGLOG(5, "ZSTDMT_getBuffer: existing buffer does not meet size conditions => freeing");
204 ZSTD_free(buf.start, bufPool->cMem);
205 }
206 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
207 /* create new buffer */
208 DEBUGLOG(5, "ZSTDMT_getBuffer: create a new buffer");
209 { buffer_t buffer;
210 void* const start = ZSTD_malloc(bSize, bufPool->cMem);
211 buffer.start = start; /* note : start can be NULL if malloc fails ! */
212 buffer.capacity = (start==NULL) ? 0 : bSize;
213 if (start==NULL) {
214 DEBUGLOG(5, "ZSTDMT_getBuffer: buffer allocation failure !!");
215 } else {
216 DEBUGLOG(5, "ZSTDMT_getBuffer: created buffer of size %u", (U32)bSize);
217 }
218 return buffer;
219 }
220}
221
222#if ZSTD_RESIZE_SEQPOOL
223/** ZSTDMT_resizeBuffer() :
224 * assumption : bufPool must be valid
225 * @return : a buffer that is at least the buffer pool buffer size.
226 * If a reallocation happens, the data in the input buffer is copied.
227 */
228static buffer_t ZSTDMT_resizeBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buffer)
229{
230 size_t const bSize = bufPool->bufferSize;
231 if (buffer.capacity < bSize) {
232 void* const start = ZSTD_malloc(bSize, bufPool->cMem);
233 buffer_t newBuffer;
234 newBuffer.start = start;
235 newBuffer.capacity = start == NULL ? 0 : bSize;
236 if (start != NULL) {
237 assert(newBuffer.capacity >= buffer.capacity);
238 memcpy(newBuffer.start, buffer.start, buffer.capacity);
239 DEBUGLOG(5, "ZSTDMT_resizeBuffer: created buffer of size %u", (U32)bSize);
240 return newBuffer;
241 }
242 DEBUGLOG(5, "ZSTDMT_resizeBuffer: buffer allocation failure !!");
243 }
244 return buffer;
245}
246#endif
247
248/* store buffer for later re-use, up to pool capacity */
249static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buf)
250{
251 DEBUGLOG(5, "ZSTDMT_releaseBuffer");
252 if (buf.start == NULL) return; /* compatible with release on NULL */
253 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
254 if (bufPool->nbBuffers < bufPool->totalBuffers) {
255 bufPool->bTable[bufPool->nbBuffers++] = buf; /* stored for later use */
256 DEBUGLOG(5, "ZSTDMT_releaseBuffer: stored buffer of size %u in slot %u",
257 (U32)buf.capacity, (U32)(bufPool->nbBuffers-1));
258 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
259 return;
260 }
261 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
262 /* Reached bufferPool capacity (should not happen) */
263 DEBUGLOG(5, "ZSTDMT_releaseBuffer: pool capacity reached => freeing ");
264 ZSTD_free(buf.start, bufPool->cMem);
265}
266
267
268/* ===== Seq Pool Wrapper ====== */
269
270static rawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0};
271
272typedef ZSTDMT_bufferPool ZSTDMT_seqPool;
273
274static size_t ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool* seqPool)
275{
276 return ZSTDMT_sizeof_bufferPool(seqPool);
277}
278
279static rawSeqStore_t bufferToSeq(buffer_t buffer)
280{
281 rawSeqStore_t seq = {NULL, 0, 0, 0};
282 seq.seq = (rawSeq*)buffer.start;
283 seq.capacity = buffer.capacity / sizeof(rawSeq);
284 return seq;
285}
286
287static buffer_t seqToBuffer(rawSeqStore_t seq)
288{
289 buffer_t buffer;
290 buffer.start = seq.seq;
291 buffer.capacity = seq.capacity * sizeof(rawSeq);
292 return buffer;
293}
294
295static rawSeqStore_t ZSTDMT_getSeq(ZSTDMT_seqPool* seqPool)
296{
297 if (seqPool->bufferSize == 0) {
298 return kNullRawSeqStore;
299 }
300 return bufferToSeq(ZSTDMT_getBuffer(seqPool));
301}
302
303#if ZSTD_RESIZE_SEQPOOL
304static rawSeqStore_t ZSTDMT_resizeSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
305{
306 return bufferToSeq(ZSTDMT_resizeBuffer(seqPool, seqToBuffer(seq)));
307}
308#endif
309
310static void ZSTDMT_releaseSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
311{
312 ZSTDMT_releaseBuffer(seqPool, seqToBuffer(seq));
313}
314
315static void ZSTDMT_setNbSeq(ZSTDMT_seqPool* const seqPool, size_t const nbSeq)
316{
317 ZSTDMT_setBufferSize(seqPool, nbSeq * sizeof(rawSeq));
318}
319
320static ZSTDMT_seqPool* ZSTDMT_createSeqPool(unsigned nbWorkers, ZSTD_customMem cMem)
321{
322 ZSTDMT_seqPool* const seqPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
323 if (seqPool == NULL) return NULL;
324 ZSTDMT_setNbSeq(seqPool, 0);
325 return seqPool;
326}
327
328static void ZSTDMT_freeSeqPool(ZSTDMT_seqPool* seqPool)
329{
330 ZSTDMT_freeBufferPool(seqPool);
331}
332
333static ZSTDMT_seqPool* ZSTDMT_expandSeqPool(ZSTDMT_seqPool* pool, U32 nbWorkers)
334{
335 return ZSTDMT_expandBufferPool(pool, nbWorkers);
336}
337
338
339/* ===== CCtx Pool ===== */
340/* a single CCtx Pool can be invoked from multiple threads in parallel */
341
342typedef struct {
343 ZSTD_pthread_mutex_t poolMutex;
344 int totalCCtx;
345 int availCCtx;
346 ZSTD_customMem cMem;
347 ZSTD_CCtx* cctx[1]; /* variable size */
348} ZSTDMT_CCtxPool;
349
350/* note : all CCtx borrowed from the pool should be released back to the pool _before_ freeing the pool */
351static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
352{
353 int cid;
354 for (cid=0; cid<pool->totalCCtx; cid++)
355 ZSTD_freeCCtx(pool->cctx[cid]); /* note : compatible with free on NULL */
356 ZSTD_pthread_mutex_destroy(&pool->poolMutex);
357 ZSTD_free(pool, pool->cMem);
358}
359
360/* ZSTDMT_createCCtxPool() :
361 * implies nbWorkers >= 1 , checked by caller ZSTDMT_createCCtx() */
362static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(int nbWorkers,
363 ZSTD_customMem cMem)
364{
365 ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_calloc(
366 sizeof(ZSTDMT_CCtxPool) + (nbWorkers-1)*sizeof(ZSTD_CCtx*), cMem);
367 assert(nbWorkers > 0);
368 if (!cctxPool) return NULL;
369 if (ZSTD_pthread_mutex_init(&cctxPool->poolMutex, NULL)) {
370 ZSTD_free(cctxPool, cMem);
371 return NULL;
372 }
373 cctxPool->cMem = cMem;
374 cctxPool->totalCCtx = nbWorkers;
375 cctxPool->availCCtx = 1; /* at least one cctx for single-thread mode */
376 cctxPool->cctx[0] = ZSTD_createCCtx_advanced(cMem);
377 if (!cctxPool->cctx[0]) { ZSTDMT_freeCCtxPool(cctxPool); return NULL; }
378 DEBUGLOG(3, "cctxPool created, with %u workers", nbWorkers);
379 return cctxPool;
380}
381
382static ZSTDMT_CCtxPool* ZSTDMT_expandCCtxPool(ZSTDMT_CCtxPool* srcPool,
383 int nbWorkers)
384{
385 if (srcPool==NULL) return NULL;
386 if (nbWorkers <= srcPool->totalCCtx) return srcPool; /* good enough */
387 /* need a larger cctx pool */
388 { ZSTD_customMem const cMem = srcPool->cMem;
389 ZSTDMT_freeCCtxPool(srcPool);
390 return ZSTDMT_createCCtxPool(nbWorkers, cMem);
391 }
392}
393
394/* only works during initialization phase, not during compression */
395static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
396{
397 ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
398 { unsigned const nbWorkers = cctxPool->totalCCtx;
399 size_t const poolSize = sizeof(*cctxPool)
400 + (nbWorkers-1) * sizeof(ZSTD_CCtx*);
401 unsigned u;
402 size_t totalCCtxSize = 0;
403 for (u=0; u<nbWorkers; u++) {
404 totalCCtxSize += ZSTD_sizeof_CCtx(cctxPool->cctx[u]);
405 }
406 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
407 assert(nbWorkers > 0);
408 return poolSize + totalCCtxSize;
409 }
410}
411
412static ZSTD_CCtx* ZSTDMT_getCCtx(ZSTDMT_CCtxPool* cctxPool)
413{
414 DEBUGLOG(5, "ZSTDMT_getCCtx");
415 ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
416 if (cctxPool->availCCtx) {
417 cctxPool->availCCtx--;
418 { ZSTD_CCtx* const cctx = cctxPool->cctx[cctxPool->availCCtx];
419 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
420 return cctx;
421 } }
422 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
423 DEBUGLOG(5, "create one more CCtx");
424 return ZSTD_createCCtx_advanced(cctxPool->cMem); /* note : can be NULL, when creation fails ! */
425}
426
427static void ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool* pool, ZSTD_CCtx* cctx)
428{
429 if (cctx==NULL) return; /* compatibility with release on NULL */
430 ZSTD_pthread_mutex_lock(&pool->poolMutex);
431 if (pool->availCCtx < pool->totalCCtx)
432 pool->cctx[pool->availCCtx++] = cctx;
433 else {
434 /* pool overflow : should not happen, since totalCCtx==nbWorkers */
435 DEBUGLOG(4, "CCtx pool overflow : free cctx");
436 ZSTD_freeCCtx(cctx);
437 }
438 ZSTD_pthread_mutex_unlock(&pool->poolMutex);
439}
440
441/* ==== Serial State ==== */
442
443typedef struct {
444 void const* start;
445 size_t size;
446} range_t;
447
448typedef struct {
449 /* All variables in the struct are protected by mutex. */
450 ZSTD_pthread_mutex_t mutex;
451 ZSTD_pthread_cond_t cond;
452 ZSTD_CCtx_params params;
453 ldmState_t ldmState;
454 XXH64_state_t xxhState;
455 unsigned nextJobID;
456 /* Protects ldmWindow.
457 * Must be acquired after the main mutex when acquiring both.
458 */
459 ZSTD_pthread_mutex_t ldmWindowMutex;
460 ZSTD_pthread_cond_t ldmWindowCond; /* Signaled when ldmWindow is updated */
461 ZSTD_window_t ldmWindow; /* A thread-safe copy of ldmState.window */
462} serialState_t;
463
464static int ZSTDMT_serialState_reset(serialState_t* serialState, ZSTDMT_seqPool* seqPool, ZSTD_CCtx_params params, size_t jobSize)
465{
466 /* Adjust parameters */
467 if (params.ldmParams.enableLdm) {
468 DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10);
469 ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
470 assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
471 assert(params.ldmParams.hashRateLog < 32);
472 serialState->ldmState.hashPower =
473 ZSTD_rollingHash_primePower(params.ldmParams.minMatchLength);
474 } else {
475 memset(&params.ldmParams, 0, sizeof(params.ldmParams));
476 }
477 serialState->nextJobID = 0;
478 if (params.fParams.checksumFlag)
479 XXH64_reset(&serialState->xxhState, 0);
480 if (params.ldmParams.enableLdm) {
481 ZSTD_customMem cMem = params.customMem;
482 unsigned const hashLog = params.ldmParams.hashLog;
483 size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
484 unsigned const bucketLog =
485 params.ldmParams.hashLog - params.ldmParams.bucketSizeLog;
486 size_t const bucketSize = (size_t)1 << bucketLog;
487 unsigned const prevBucketLog =
488 serialState->params.ldmParams.hashLog -
489 serialState->params.ldmParams.bucketSizeLog;
490 /* Size the seq pool tables */
491 ZSTDMT_setNbSeq(seqPool, ZSTD_ldm_getMaxNbSeq(params.ldmParams, jobSize));
492 /* Reset the window */
493 ZSTD_window_clear(&serialState->ldmState.window);
494 serialState->ldmWindow = serialState->ldmState.window;
495 /* Resize tables and output space if necessary. */
496 if (serialState->ldmState.hashTable == NULL || serialState->params.ldmParams.hashLog < hashLog) {
497 ZSTD_free(serialState->ldmState.hashTable, cMem);
498 serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_malloc(hashSize, cMem);
499 }
500 if (serialState->ldmState.bucketOffsets == NULL || prevBucketLog < bucketLog) {
501 ZSTD_free(serialState->ldmState.bucketOffsets, cMem);
502 serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_malloc(bucketSize, cMem);
503 }
504 if (!serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets)
505 return 1;
506 /* Zero the tables */
507 memset(serialState->ldmState.hashTable, 0, hashSize);
508 memset(serialState->ldmState.bucketOffsets, 0, bucketSize);
509 }
510 serialState->params = params;
511 serialState->params.jobSize = (U32)jobSize;
512 return 0;
513}
514
515static int ZSTDMT_serialState_init(serialState_t* serialState)
516{
517 int initError = 0;
518 memset(serialState, 0, sizeof(*serialState));
519 initError |= ZSTD_pthread_mutex_init(&serialState->mutex, NULL);
520 initError |= ZSTD_pthread_cond_init(&serialState->cond, NULL);
521 initError |= ZSTD_pthread_mutex_init(&serialState->ldmWindowMutex, NULL);
522 initError |= ZSTD_pthread_cond_init(&serialState->ldmWindowCond, NULL);
523 return initError;
524}
525
526static void ZSTDMT_serialState_free(serialState_t* serialState)
527{
528 ZSTD_customMem cMem = serialState->params.customMem;
529 ZSTD_pthread_mutex_destroy(&serialState->mutex);
530 ZSTD_pthread_cond_destroy(&serialState->cond);
531 ZSTD_pthread_mutex_destroy(&serialState->ldmWindowMutex);
532 ZSTD_pthread_cond_destroy(&serialState->ldmWindowCond);
533 ZSTD_free(serialState->ldmState.hashTable, cMem);
534 ZSTD_free(serialState->ldmState.bucketOffsets, cMem);
535}
536
537static void ZSTDMT_serialState_update(serialState_t* serialState,
538 ZSTD_CCtx* jobCCtx, rawSeqStore_t seqStore,
539 range_t src, unsigned jobID)
540{
541 /* Wait for our turn */
542 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
543 while (serialState->nextJobID < jobID) {
544 DEBUGLOG(5, "wait for serialState->cond");
545 ZSTD_pthread_cond_wait(&serialState->cond, &serialState->mutex);
546 }
547 /* A future job may error and skip our job */
548 if (serialState->nextJobID == jobID) {
549 /* It is now our turn, do any processing necessary */
550 if (serialState->params.ldmParams.enableLdm) {
551 size_t error;
552 assert(seqStore.seq != NULL && seqStore.pos == 0 &&
553 seqStore.size == 0 && seqStore.capacity > 0);
554 assert(src.size <= serialState->params.jobSize);
555 ZSTD_window_update(&serialState->ldmState.window, src.start, src.size);
556 error = ZSTD_ldm_generateSequences(
557 &serialState->ldmState, &seqStore,
558 &serialState->params.ldmParams, src.start, src.size);
559 /* We provide a large enough buffer to never fail. */
560 assert(!ZSTD_isError(error)); (void)error;
561 /* Update ldmWindow to match the ldmState.window and signal the main
562 * thread if it is waiting for a buffer.
563 */
564 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
565 serialState->ldmWindow = serialState->ldmState.window;
566 ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
567 ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
568 }
569 if (serialState->params.fParams.checksumFlag && src.size > 0)
570 XXH64_update(&serialState->xxhState, src.start, src.size);
571 }
572 /* Now it is the next jobs turn */
573 serialState->nextJobID++;
574 ZSTD_pthread_cond_broadcast(&serialState->cond);
575 ZSTD_pthread_mutex_unlock(&serialState->mutex);
576
577 if (seqStore.size > 0) {
578 size_t const err = ZSTD_referenceExternalSequences(
579 jobCCtx, seqStore.seq, seqStore.size);
580 assert(serialState->params.ldmParams.enableLdm);
581 assert(!ZSTD_isError(err));
582 (void)err;
583 }
584}
585
586static void ZSTDMT_serialState_ensureFinished(serialState_t* serialState,
587 unsigned jobID, size_t cSize)
588{
589 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
590 if (serialState->nextJobID <= jobID) {
591 assert(ZSTD_isError(cSize)); (void)cSize;
592 DEBUGLOG(5, "Skipping past job %u because of error", jobID);
593 serialState->nextJobID = jobID + 1;
594 ZSTD_pthread_cond_broadcast(&serialState->cond);
595
596 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
597 ZSTD_window_clear(&serialState->ldmWindow);
598 ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
599 ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
600 }
601 ZSTD_pthread_mutex_unlock(&serialState->mutex);
602
603}
604
605
606/* ------------------------------------------ */
607/* ===== Worker thread ===== */
608/* ------------------------------------------ */
609
610static const range_t kNullRange = { NULL, 0 };
611
612typedef struct {
613 size_t consumed; /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx */
614 size_t cSize; /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx, then set0 by mtctx */
615 ZSTD_pthread_mutex_t job_mutex; /* Thread-safe - used by mtctx and worker */
616 ZSTD_pthread_cond_t job_cond; /* Thread-safe - used by mtctx and worker */
617 ZSTDMT_CCtxPool* cctxPool; /* Thread-safe - used by mtctx and (all) workers */
618 ZSTDMT_bufferPool* bufPool; /* Thread-safe - used by mtctx and (all) workers */
619 ZSTDMT_seqPool* seqPool; /* Thread-safe - used by mtctx and (all) workers */
620 serialState_t* serial; /* Thread-safe - used by mtctx and (all) workers */
621 buffer_t dstBuff; /* set by worker (or mtctx), then read by worker & mtctx, then modified by mtctx => no barrier */
622 range_t prefix; /* set by mtctx, then read by worker & mtctx => no barrier */
623 range_t src; /* set by mtctx, then read by worker & mtctx => no barrier */
624 unsigned jobID; /* set by mtctx, then read by worker => no barrier */
625 unsigned firstJob; /* set by mtctx, then read by worker => no barrier */
626 unsigned lastJob; /* set by mtctx, then read by worker => no barrier */
627 ZSTD_CCtx_params params; /* set by mtctx, then read by worker => no barrier */
628 const ZSTD_CDict* cdict; /* set by mtctx, then read by worker => no barrier */
629 unsigned long long fullFrameSize; /* set by mtctx, then read by worker => no barrier */
630 size_t dstFlushed; /* used only by mtctx */
631 unsigned frameChecksumNeeded; /* used only by mtctx */
632} ZSTDMT_jobDescription;
633
634#define JOB_ERROR(e) { \
635 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex); \
636 job->cSize = e; \
637 ZSTD_pthread_mutex_unlock(&job->job_mutex); \
638 goto _endJob; \
639}
640
641/* ZSTDMT_compressionJob() is a POOL_function type */
642static void ZSTDMT_compressionJob(void* jobDescription)
643{
644 ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)jobDescription;
645 ZSTD_CCtx_params jobParams = job->params; /* do not modify job->params ! copy it, modify the copy */
646 ZSTD_CCtx* const cctx = ZSTDMT_getCCtx(job->cctxPool);
647 rawSeqStore_t rawSeqStore = ZSTDMT_getSeq(job->seqPool);
648 buffer_t dstBuff = job->dstBuff;
649 size_t lastCBlockSize = 0;
650
651 /* resources */
652 if (cctx==NULL) JOB_ERROR(ERROR(memory_allocation));
653 if (dstBuff.start == NULL) { /* streaming job : doesn't provide a dstBuffer */
654 dstBuff = ZSTDMT_getBuffer(job->bufPool);
655 if (dstBuff.start==NULL) JOB_ERROR(ERROR(memory_allocation));
656 job->dstBuff = dstBuff; /* this value can be read in ZSTDMT_flush, when it copies the whole job */
657 }
658 if (jobParams.ldmParams.enableLdm && rawSeqStore.seq == NULL)
659 JOB_ERROR(ERROR(memory_allocation));
660
661 /* Don't compute the checksum for chunks, since we compute it externally,
662 * but write it in the header.
663 */
664 if (job->jobID != 0) jobParams.fParams.checksumFlag = 0;
665 /* Don't run LDM for the chunks, since we handle it externally */
666 jobParams.ldmParams.enableLdm = 0;
667
668
669 /* init */
670 if (job->cdict) {
671 size_t const initError = ZSTD_compressBegin_advanced_internal(cctx, NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast, job->cdict, jobParams, job->fullFrameSize);
672 assert(job->firstJob); /* only allowed for first job */
673 if (ZSTD_isError(initError)) JOB_ERROR(initError);
674 } else { /* srcStart points at reloaded section */
675 U64 const pledgedSrcSize = job->firstJob ? job->fullFrameSize : job->src.size;
676 { size_t const forceWindowError = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_forceMaxWindow, !job->firstJob);
677 if (ZSTD_isError(forceWindowError)) JOB_ERROR(forceWindowError);
678 }
679 { size_t const initError = ZSTD_compressBegin_advanced_internal(cctx,
680 job->prefix.start, job->prefix.size, ZSTD_dct_rawContent, /* load dictionary in "content-only" mode (no header analysis) */
681 ZSTD_dtlm_fast,
682 NULL, /*cdict*/
683 jobParams, pledgedSrcSize);
684 if (ZSTD_isError(initError)) JOB_ERROR(initError);
685 } }
686
687 /* Perform serial step as early as possible, but after CCtx initialization */
688 ZSTDMT_serialState_update(job->serial, cctx, rawSeqStore, job->src, job->jobID);
689
690 if (!job->firstJob) { /* flush and overwrite frame header when it's not first job */
691 size_t const hSize = ZSTD_compressContinue(cctx, dstBuff.start, dstBuff.capacity, job->src.start, 0);
692 if (ZSTD_isError(hSize)) JOB_ERROR(hSize);
693 DEBUGLOG(5, "ZSTDMT_compressionJob: flush and overwrite %u bytes of frame header (not first job)", (U32)hSize);
694 ZSTD_invalidateRepCodes(cctx);
695 }
696
697 /* compress */
698 { size_t const chunkSize = 4*ZSTD_BLOCKSIZE_MAX;
699 int const nbChunks = (int)((job->src.size + (chunkSize-1)) / chunkSize);
700 const BYTE* ip = (const BYTE*) job->src.start;
701 BYTE* const ostart = (BYTE*)dstBuff.start;
702 BYTE* op = ostart;
703 BYTE* oend = op + dstBuff.capacity;
704 int chunkNb;
705 if (sizeof(size_t) > sizeof(int)) assert(job->src.size < ((size_t)INT_MAX) * chunkSize); /* check overflow */
706 DEBUGLOG(5, "ZSTDMT_compressionJob: compress %u bytes in %i blocks", (U32)job->src.size, nbChunks);
707 assert(job->cSize == 0);
708 for (chunkNb = 1; chunkNb < nbChunks; chunkNb++) {
709 size_t const cSize = ZSTD_compressContinue(cctx, op, oend-op, ip, chunkSize);
710 if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
711 ip += chunkSize;
712 op += cSize; assert(op < oend);
713 /* stats */
714 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
715 job->cSize += cSize;
716 job->consumed = chunkSize * chunkNb;
717 DEBUGLOG(5, "ZSTDMT_compressionJob: compress new block : cSize==%u bytes (total: %u)",
718 (U32)cSize, (U32)job->cSize);
719 ZSTD_pthread_cond_signal(&job->job_cond); /* warns some more data is ready to be flushed */
720 ZSTD_pthread_mutex_unlock(&job->job_mutex);
721 }
722 /* last block */
723 assert(chunkSize > 0);
724 assert((chunkSize & (chunkSize - 1)) == 0); /* chunkSize must be power of 2 for mask==(chunkSize-1) to work */
725 if ((nbChunks > 0) | job->lastJob /*must output a "last block" flag*/ ) {
726 size_t const lastBlockSize1 = job->src.size & (chunkSize-1);
727 size_t const lastBlockSize = ((lastBlockSize1==0) & (job->src.size>=chunkSize)) ? chunkSize : lastBlockSize1;
728 size_t const cSize = (job->lastJob) ?
729 ZSTD_compressEnd (cctx, op, oend-op, ip, lastBlockSize) :
730 ZSTD_compressContinue(cctx, op, oend-op, ip, lastBlockSize);
731 if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
732 lastCBlockSize = cSize;
733 } }
734
735_endJob:
736 ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize);
737 if (job->prefix.size > 0)
738 DEBUGLOG(5, "Finished with prefix: %zx", (size_t)job->prefix.start);
739 DEBUGLOG(5, "Finished with source: %zx", (size_t)job->src.start);
740 /* release resources */
741 ZSTDMT_releaseSeq(job->seqPool, rawSeqStore);
742 ZSTDMT_releaseCCtx(job->cctxPool, cctx);
743 /* report */
744 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
745 if (ZSTD_isError(job->cSize)) assert(lastCBlockSize == 0);
746 job->cSize += lastCBlockSize;
747 job->consumed = job->src.size; /* when job->consumed == job->src.size , compression job is presumed completed */
748 ZSTD_pthread_cond_signal(&job->job_cond);
749 ZSTD_pthread_mutex_unlock(&job->job_mutex);
750}
751
752
753/* ------------------------------------------ */
754/* ===== Multi-threaded compression ===== */
755/* ------------------------------------------ */
756
757typedef struct {
758 range_t prefix; /* read-only non-owned prefix buffer */
759 buffer_t buffer;
760 size_t filled;
761} inBuff_t;
762
763typedef struct {
764 BYTE* buffer; /* The round input buffer. All jobs get references
765 * to pieces of the buffer. ZSTDMT_tryGetInputRange()
766 * handles handing out job input buffers, and makes
767 * sure it doesn't overlap with any pieces still in use.
768 */
769 size_t capacity; /* The capacity of buffer. */
770 size_t pos; /* The position of the current inBuff in the round
771 * buffer. Updated past the end if the inBuff once
772 * the inBuff is sent to the worker thread.
773 * pos <= capacity.
774 */
775} roundBuff_t;
776
777static const roundBuff_t kNullRoundBuff = {NULL, 0, 0};
778
779#define RSYNC_LENGTH 32
780
781typedef struct {
782 U64 hash;
783 U64 hitMask;
784 U64 primePower;
785} rsyncState_t;
786
787struct ZSTDMT_CCtx_s {
788 POOL_ctx* factory;
789 ZSTDMT_jobDescription* jobs;
790 ZSTDMT_bufferPool* bufPool;
791 ZSTDMT_CCtxPool* cctxPool;
792 ZSTDMT_seqPool* seqPool;
793 ZSTD_CCtx_params params;
794 size_t targetSectionSize;
795 size_t targetPrefixSize;
796 int jobReady; /* 1 => one job is already prepared, but pool has shortage of workers. Don't create a new job. */
797 inBuff_t inBuff;
798 roundBuff_t roundBuff;
799 serialState_t serial;
800 rsyncState_t rsync;
801 unsigned singleBlockingThread;
802 unsigned jobIDMask;
803 unsigned doneJobID;
804 unsigned nextJobID;
805 unsigned frameEnded;
806 unsigned allJobsCompleted;
807 unsigned long long frameContentSize;
808 unsigned long long consumed;
809 unsigned long long produced;
810 ZSTD_customMem cMem;
811 ZSTD_CDict* cdictLocal;
812 const ZSTD_CDict* cdict;
813};
814
815static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZSTD_customMem cMem)
816{
817 U32 jobNb;
818 if (jobTable == NULL) return;
819 for (jobNb=0; jobNb<nbJobs; jobNb++) {
820 ZSTD_pthread_mutex_destroy(&jobTable[jobNb].job_mutex);
821 ZSTD_pthread_cond_destroy(&jobTable[jobNb].job_cond);
822 }
823 ZSTD_free(jobTable, cMem);
824}
825
826/* ZSTDMT_allocJobsTable()
827 * allocate and init a job table.
828 * update *nbJobsPtr to next power of 2 value, as size of table */
829static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
830{
831 U32 const nbJobsLog2 = ZSTD_highbit32(*nbJobsPtr) + 1;
832 U32 const nbJobs = 1 << nbJobsLog2;
833 U32 jobNb;
834 ZSTDMT_jobDescription* const jobTable = (ZSTDMT_jobDescription*)
835 ZSTD_calloc(nbJobs * sizeof(ZSTDMT_jobDescription), cMem);
836 int initError = 0;
837 if (jobTable==NULL) return NULL;
838 *nbJobsPtr = nbJobs;
839 for (jobNb=0; jobNb<nbJobs; jobNb++) {
840 initError |= ZSTD_pthread_mutex_init(&jobTable[jobNb].job_mutex, NULL);
841 initError |= ZSTD_pthread_cond_init(&jobTable[jobNb].job_cond, NULL);
842 }
843 if (initError != 0) {
844 ZSTDMT_freeJobsTable(jobTable, nbJobs, cMem);
845 return NULL;
846 }
847 return jobTable;
848}
849
850static size_t ZSTDMT_expandJobsTable (ZSTDMT_CCtx* mtctx, U32 nbWorkers) {
851 U32 nbJobs = nbWorkers + 2;
852 if (nbJobs > mtctx->jobIDMask+1) { /* need more job capacity */
853 ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
854 mtctx->jobIDMask = 0;
855 mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, mtctx->cMem);
856 if (mtctx->jobs==NULL) return ERROR(memory_allocation);
857 assert((nbJobs != 0) && ((nbJobs & (nbJobs - 1)) == 0)); /* ensure nbJobs is a power of 2 */
858 mtctx->jobIDMask = nbJobs - 1;
859 }
860 return 0;
861}
862
863
864/* ZSTDMT_CCtxParam_setNbWorkers():
865 * Internal use only */
866size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers)
867{
868 return ZSTD_CCtxParams_setParameter(params, ZSTD_c_nbWorkers, (int)nbWorkers);
869}
870
871MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, ZSTD_customMem cMem)
872{
873 ZSTDMT_CCtx* mtctx;
874 U32 nbJobs = nbWorkers + 2;
875 int initError;
876 DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbWorkers = %u)", nbWorkers);
877
878 if (nbWorkers < 1) return NULL;
879 nbWorkers = MIN(nbWorkers , ZSTDMT_NBWORKERS_MAX);
880 if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
881 /* invalid custom allocator */
882 return NULL;
883
884 mtctx = (ZSTDMT_CCtx*) ZSTD_calloc(sizeof(ZSTDMT_CCtx), cMem);
885 if (!mtctx) return NULL;
886 ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
887 mtctx->cMem = cMem;
888 mtctx->allJobsCompleted = 1;
889 mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem);
890 mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem);
891 assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0); /* ensure nbJobs is a power of 2 */
892 mtctx->jobIDMask = nbJobs - 1;
893 mtctx->bufPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
894 mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem);
895 mtctx->seqPool = ZSTDMT_createSeqPool(nbWorkers, cMem);
896 initError = ZSTDMT_serialState_init(&mtctx->serial);
897 mtctx->roundBuff = kNullRoundBuff;
898 if (!mtctx->factory | !mtctx->jobs | !mtctx->bufPool | !mtctx->cctxPool | !mtctx->seqPool | initError) {
899 ZSTDMT_freeCCtx(mtctx);
900 return NULL;
901 }
902 DEBUGLOG(3, "mt_cctx created, for %u threads", nbWorkers);
903 return mtctx;
904}
905
906ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem)
907{
908#ifdef ZSTD_MULTITHREAD
909 return ZSTDMT_createCCtx_advanced_internal(nbWorkers, cMem);
910#else
911 (void)nbWorkers;
912 (void)cMem;
913 return NULL;
914#endif
915}
916
917ZSTDMT_CCtx* ZSTDMT_createCCtx(unsigned nbWorkers)
918{
919 return ZSTDMT_createCCtx_advanced(nbWorkers, ZSTD_defaultCMem);
920}
921
922
923/* ZSTDMT_releaseAllJobResources() :
924 * note : ensure all workers are killed first ! */
925static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx)
926{
927 unsigned jobID;
928 DEBUGLOG(3, "ZSTDMT_releaseAllJobResources");
929 for (jobID=0; jobID <= mtctx->jobIDMask; jobID++) {
930 DEBUGLOG(4, "job%02u: release dst address %08X", jobID, (U32)(size_t)mtctx->jobs[jobID].dstBuff.start);
931 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);
932 mtctx->jobs[jobID].dstBuff = g_nullBuffer;
933 mtctx->jobs[jobID].cSize = 0;
934 }
935 memset(mtctx->jobs, 0, (mtctx->jobIDMask+1)*sizeof(ZSTDMT_jobDescription));
936 mtctx->inBuff.buffer = g_nullBuffer;
937 mtctx->inBuff.filled = 0;
938 mtctx->allJobsCompleted = 1;
939}
940
941static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* mtctx)
942{
943 DEBUGLOG(4, "ZSTDMT_waitForAllJobsCompleted");
944 while (mtctx->doneJobID < mtctx->nextJobID) {
945 unsigned const jobID = mtctx->doneJobID & mtctx->jobIDMask;
946 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
947 while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
948 DEBUGLOG(4, "waiting for jobCompleted signal from job %u", mtctx->doneJobID); /* we want to block when waiting for data to flush */
949 ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
950 }
951 ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
952 mtctx->doneJobID++;
953 }
954}
955
956size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx)
957{
958 if (mtctx==NULL) return 0; /* compatible with free on NULL */
959 POOL_free(mtctx->factory); /* stop and free worker threads */
960 ZSTDMT_releaseAllJobResources(mtctx); /* release job resources into pools first */
961 ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
962 ZSTDMT_freeBufferPool(mtctx->bufPool);
963 ZSTDMT_freeCCtxPool(mtctx->cctxPool);
964 ZSTDMT_freeSeqPool(mtctx->seqPool);
965 ZSTDMT_serialState_free(&mtctx->serial);
966 ZSTD_freeCDict(mtctx->cdictLocal);
967 if (mtctx->roundBuff.buffer)
968 ZSTD_free(mtctx->roundBuff.buffer, mtctx->cMem);
969 ZSTD_free(mtctx, mtctx->cMem);
970 return 0;
971}
972
973size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx)
974{
975 if (mtctx == NULL) return 0; /* supports sizeof NULL */
976 return sizeof(*mtctx)
977 + POOL_sizeof(mtctx->factory)
978 + ZSTDMT_sizeof_bufferPool(mtctx->bufPool)
979 + (mtctx->jobIDMask+1) * sizeof(ZSTDMT_jobDescription)
980 + ZSTDMT_sizeof_CCtxPool(mtctx->cctxPool)
981 + ZSTDMT_sizeof_seqPool(mtctx->seqPool)
982 + ZSTD_sizeof_CDict(mtctx->cdictLocal)
983 + mtctx->roundBuff.capacity;
984}
985
986/* Internal only */
987size_t
988ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params,
989 ZSTDMT_parameter parameter,
990 int value)
991{
992 DEBUGLOG(4, "ZSTDMT_CCtxParam_setMTCtxParameter");
993 switch(parameter)
994 {
995 case ZSTDMT_p_jobSize :
996 DEBUGLOG(4, "ZSTDMT_CCtxParam_setMTCtxParameter : set jobSize to %i", value);
997 return ZSTD_CCtxParams_setParameter(params, ZSTD_c_jobSize, value);
998 case ZSTDMT_p_overlapLog :
999 DEBUGLOG(4, "ZSTDMT_p_overlapLog : %i", value);
1000 return ZSTD_CCtxParams_setParameter(params, ZSTD_c_overlapLog, value);
1001 case ZSTDMT_p_rsyncable :
1002 DEBUGLOG(4, "ZSTD_p_rsyncable : %i", value);
1003 return ZSTD_CCtxParams_setParameter(params, ZSTD_c_rsyncable, value);
1004 default :
1005 return ERROR(parameter_unsupported);
1006 }
1007}
1008
1009size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, int value)
1010{
1011 DEBUGLOG(4, "ZSTDMT_setMTCtxParameter");
1012 return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value);
1013}
1014
1015size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, int* value)
1016{
1017 switch (parameter) {
1018 case ZSTDMT_p_jobSize:
1019 return ZSTD_CCtxParams_getParameter(&mtctx->params, ZSTD_c_jobSize, value);
1020 case ZSTDMT_p_overlapLog:
1021 return ZSTD_CCtxParams_getParameter(&mtctx->params, ZSTD_c_overlapLog, value);
1022 case ZSTDMT_p_rsyncable:
1023 return ZSTD_CCtxParams_getParameter(&mtctx->params, ZSTD_c_rsyncable, value);
1024 default:
1025 return ERROR(parameter_unsupported);
1026 }
1027}
1028
1029/* Sets parameters relevant to the compression job,
1030 * initializing others to default values. */
1031static ZSTD_CCtx_params ZSTDMT_initJobCCtxParams(ZSTD_CCtx_params const params)
1032{
1033 ZSTD_CCtx_params jobParams = params;
1034 /* Clear parameters related to multithreading */
1035 jobParams.forceWindow = 0;
1036 jobParams.nbWorkers = 0;
1037 jobParams.jobSize = 0;
1038 jobParams.overlapLog = 0;
1039 jobParams.rsyncable = 0;
1040 memset(&jobParams.ldmParams, 0, sizeof(ldmParams_t));
1041 memset(&jobParams.customMem, 0, sizeof(ZSTD_customMem));
1042 return jobParams;
1043}
1044
1045
1046/* ZSTDMT_resize() :
1047 * @return : error code if fails, 0 on success */
1048static size_t ZSTDMT_resize(ZSTDMT_CCtx* mtctx, unsigned nbWorkers)
1049{
1050 if (POOL_resize(mtctx->factory, nbWorkers)) return ERROR(memory_allocation);
1051 FORWARD_IF_ERROR( ZSTDMT_expandJobsTable(mtctx, nbWorkers) );
1052 mtctx->bufPool = ZSTDMT_expandBufferPool(mtctx->bufPool, nbWorkers);
1053 if (mtctx->bufPool == NULL) return ERROR(memory_allocation);
1054 mtctx->cctxPool = ZSTDMT_expandCCtxPool(mtctx->cctxPool, nbWorkers);
1055 if (mtctx->cctxPool == NULL) return ERROR(memory_allocation);
1056 mtctx->seqPool = ZSTDMT_expandSeqPool(mtctx->seqPool, nbWorkers);
1057 if (mtctx->seqPool == NULL) return ERROR(memory_allocation);
1058 ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
1059 return 0;
1060}
1061
1062
1063/*! ZSTDMT_updateCParams_whileCompressing() :
1064 * Updates a selected set of compression parameters, remaining compatible with currently active frame.
1065 * New parameters will be applied to next compression job. */
1066void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams)
1067{
1068 U32 const saved_wlog = mtctx->params.cParams.windowLog; /* Do not modify windowLog while compressing */
1069 int const compressionLevel = cctxParams->compressionLevel;
1070 DEBUGLOG(5, "ZSTDMT_updateCParams_whileCompressing (level:%i)",
1071 compressionLevel);
1072 mtctx->params.compressionLevel = compressionLevel;
1073 { ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, 0, 0);
1074 cParams.windowLog = saved_wlog;
1075 mtctx->params.cParams = cParams;
1076 }
1077}
1078
1079/* ZSTDMT_getFrameProgression():
1080 * tells how much data has been consumed (input) and produced (output) for current frame.
1081 * able to count progression inside worker threads.
1082 * Note : mutex will be acquired during statistics collection inside workers. */
1083ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx)
1084{
1085 ZSTD_frameProgression fps;
1086 DEBUGLOG(5, "ZSTDMT_getFrameProgression");
1087 fps.ingested = mtctx->consumed + mtctx->inBuff.filled;
1088 fps.consumed = mtctx->consumed;
1089 fps.produced = fps.flushed = mtctx->produced;
1090 fps.currentJobID = mtctx->nextJobID;
1091 fps.nbActiveWorkers = 0;
1092 { unsigned jobNb;
1093 unsigned lastJobNb = mtctx->nextJobID + mtctx->jobReady; assert(mtctx->jobReady <= 1);
1094 DEBUGLOG(6, "ZSTDMT_getFrameProgression: jobs: from %u to <%u (jobReady:%u)",
1095 mtctx->doneJobID, lastJobNb, mtctx->jobReady)
1096 for (jobNb = mtctx->doneJobID ; jobNb < lastJobNb ; jobNb++) {
1097 unsigned const wJobID = jobNb & mtctx->jobIDMask;
1098 ZSTDMT_jobDescription* jobPtr = &mtctx->jobs[wJobID];
1099 ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
1100 { size_t const cResult = jobPtr->cSize;
1101 size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
1102 size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
1103 assert(flushed <= produced);
1104 fps.ingested += jobPtr->src.size;
1105 fps.consumed += jobPtr->consumed;
1106 fps.produced += produced;
1107 fps.flushed += flushed;
1108 fps.nbActiveWorkers += (jobPtr->consumed < jobPtr->src.size);
1109 }
1110 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1111 }
1112 }
1113 return fps;
1114}
1115
1116
1117size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx)
1118{
1119 size_t toFlush;
1120 unsigned const jobID = mtctx->doneJobID;
1121 assert(jobID <= mtctx->nextJobID);
1122 if (jobID == mtctx->nextJobID) return 0; /* no active job => nothing to flush */
1123
1124 /* look into oldest non-fully-flushed job */
1125 { unsigned const wJobID = jobID & mtctx->jobIDMask;
1126 ZSTDMT_jobDescription* const jobPtr = &mtctx->jobs[wJobID];
1127 ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
1128 { size_t const cResult = jobPtr->cSize;
1129 size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
1130 size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
1131 assert(flushed <= produced);
1132 assert(jobPtr->consumed <= jobPtr->src.size);
1133 toFlush = produced - flushed;
1134 /* if toFlush==0, nothing is available to flush.
1135 * However, jobID is expected to still be active:
1136 * if jobID was already completed and fully flushed,
1137 * ZSTDMT_flushProduced() should have already moved onto next job.
1138 * Therefore, some input has not yet been consumed. */
1139 if (toFlush==0) {
1140 assert(jobPtr->consumed < jobPtr->src.size);
1141 }
1142 }
1143 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1144 }
1145
1146 return toFlush;
1147}
1148
1149
1150/* ------------------------------------------ */
1151/* ===== Multi-threaded compression ===== */
1152/* ------------------------------------------ */
1153
1154static unsigned ZSTDMT_computeTargetJobLog(ZSTD_CCtx_params const params)
1155{
1156 unsigned jobLog;
1157 if (params.ldmParams.enableLdm) {
1158 /* In Long Range Mode, the windowLog is typically oversized.
1159 * In which case, it's preferable to determine the jobSize
1160 * based on chainLog instead. */
1161 jobLog = MAX(21, params.cParams.chainLog + 4);
1162 } else {
1163 jobLog = MAX(20, params.cParams.windowLog + 2);
1164 }
1165 return MIN(jobLog, (unsigned)ZSTDMT_JOBLOG_MAX);
1166}
1167
1168static int ZSTDMT_overlapLog_default(ZSTD_strategy strat)
1169{
1170 switch(strat)
1171 {
1172 case ZSTD_btultra2:
1173 return 9;
1174 case ZSTD_btultra:
1175 case ZSTD_btopt:
1176 return 8;
1177 case ZSTD_btlazy2:
1178 case ZSTD_lazy2:
1179 return 7;
1180 case ZSTD_lazy:
1181 case ZSTD_greedy:
1182 case ZSTD_dfast:
1183 case ZSTD_fast:
1184 default:;
1185 }
1186 return 6;
1187}
1188
1189static int ZSTDMT_overlapLog(int ovlog, ZSTD_strategy strat)
1190{
1191 assert(0 <= ovlog && ovlog <= 9);
1192 if (ovlog == 0) return ZSTDMT_overlapLog_default(strat);
1193 return ovlog;
1194}
1195
1196static size_t ZSTDMT_computeOverlapSize(ZSTD_CCtx_params const params)
1197{
1198 int const overlapRLog = 9 - ZSTDMT_overlapLog(params.overlapLog, params.cParams.strategy);
1199 int ovLog = (overlapRLog >= 8) ? 0 : (params.cParams.windowLog - overlapRLog);
1200 assert(0 <= overlapRLog && overlapRLog <= 8);
1201 if (params.ldmParams.enableLdm) {
1202 /* In Long Range Mode, the windowLog is typically oversized.
1203 * In which case, it's preferable to determine the jobSize
1204 * based on chainLog instead.
1205 * Then, ovLog becomes a fraction of the jobSize, rather than windowSize */
1206 ovLog = MIN(params.cParams.windowLog, ZSTDMT_computeTargetJobLog(params) - 2)
1207 - overlapRLog;
1208 }
1209 assert(0 <= ovLog && ovLog <= ZSTD_WINDOWLOG_MAX);
1210 DEBUGLOG(4, "overlapLog : %i", params.overlapLog);
1211 DEBUGLOG(4, "overlap size : %i", 1 << ovLog);
1212 return (ovLog==0) ? 0 : (size_t)1 << ovLog;
1213}
1214
1215static unsigned
1216ZSTDMT_computeNbJobs(ZSTD_CCtx_params params, size_t srcSize, unsigned nbWorkers)
1217{
1218 assert(nbWorkers>0);
1219 { size_t const jobSizeTarget = (size_t)1 << ZSTDMT_computeTargetJobLog(params);
1220 size_t const jobMaxSize = jobSizeTarget << 2;
1221 size_t const passSizeMax = jobMaxSize * nbWorkers;
1222 unsigned const multiplier = (unsigned)(srcSize / passSizeMax) + 1;
1223 unsigned const nbJobsLarge = multiplier * nbWorkers;
1224 unsigned const nbJobsMax = (unsigned)(srcSize / jobSizeTarget) + 1;
1225 unsigned const nbJobsSmall = MIN(nbJobsMax, nbWorkers);
1226 return (multiplier>1) ? nbJobsLarge : nbJobsSmall;
1227} }
1228
1229/* ZSTDMT_compress_advanced_internal() :
1230 * This is a blocking function : it will only give back control to caller after finishing its compression job.
1231 */
1232static size_t ZSTDMT_compress_advanced_internal(
1233 ZSTDMT_CCtx* mtctx,
1234 void* dst, size_t dstCapacity,
1235 const void* src, size_t srcSize,
1236 const ZSTD_CDict* cdict,
1237 ZSTD_CCtx_params params)
1238{
1239 ZSTD_CCtx_params const jobParams = ZSTDMT_initJobCCtxParams(params);
1240 size_t const overlapSize = ZSTDMT_computeOverlapSize(params);
1241 unsigned const nbJobs = ZSTDMT_computeNbJobs(params, srcSize, params.nbWorkers);
1242 size_t const proposedJobSize = (srcSize + (nbJobs-1)) / nbJobs;
1243 size_t const avgJobSize = (((proposedJobSize-1) & 0x1FFFF) < 0x7FFF) ? proposedJobSize + 0xFFFF : proposedJobSize; /* avoid too small last block */
1244 const char* const srcStart = (const char*)src;
1245 size_t remainingSrcSize = srcSize;
1246 unsigned const compressWithinDst = (dstCapacity >= ZSTD_compressBound(srcSize)) ? nbJobs : (unsigned)(dstCapacity / ZSTD_compressBound(avgJobSize)); /* presumes avgJobSize >= 256 KB, which should be the case */
1247 size_t frameStartPos = 0, dstBufferPos = 0;
1248 assert(jobParams.nbWorkers == 0);
1249 assert(mtctx->cctxPool->totalCCtx == params.nbWorkers);
1250
1251 params.jobSize = (U32)avgJobSize;
1252 DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: nbJobs=%2u (rawSize=%u bytes; fixedSize=%u) ",
1253 nbJobs, (U32)proposedJobSize, (U32)avgJobSize);
1254
1255 if ((nbJobs==1) | (params.nbWorkers<=1)) { /* fallback to single-thread mode : this is a blocking invocation anyway */
1256 ZSTD_CCtx* const cctx = mtctx->cctxPool->cctx[0];
1257 DEBUGLOG(4, "ZSTDMT_compress_advanced_internal: fallback to single-thread mode");
1258 if (cdict) return ZSTD_compress_usingCDict_advanced(cctx, dst, dstCapacity, src, srcSize, cdict, jobParams.fParams);
1259 return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, NULL, 0, jobParams);
1260 }
1261
1262 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) */
1263 ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(avgJobSize) );
1264 if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params, avgJobSize))
1265 return ERROR(memory_allocation);
1266
1267 FORWARD_IF_ERROR( ZSTDMT_expandJobsTable(mtctx, nbJobs) ); /* only expands if necessary */
1268
1269 { unsigned u;
1270 for (u=0; u<nbJobs; u++) {
1271 size_t const jobSize = MIN(remainingSrcSize, avgJobSize);
1272 size_t const dstBufferCapacity = ZSTD_compressBound(jobSize);
1273 buffer_t const dstAsBuffer = { (char*)dst + dstBufferPos, dstBufferCapacity };
1274 buffer_t const dstBuffer = u < compressWithinDst ? dstAsBuffer : g_nullBuffer;
1275 size_t dictSize = u ? overlapSize : 0;
1276
1277 mtctx->jobs[u].prefix.start = srcStart + frameStartPos - dictSize;
1278 mtctx->jobs[u].prefix.size = dictSize;
1279 mtctx->jobs[u].src.start = srcStart + frameStartPos;
1280 mtctx->jobs[u].src.size = jobSize; assert(jobSize > 0); /* avoid job.src.size == 0 */
1281 mtctx->jobs[u].consumed = 0;
1282 mtctx->jobs[u].cSize = 0;
1283 mtctx->jobs[u].cdict = (u==0) ? cdict : NULL;
1284 mtctx->jobs[u].fullFrameSize = srcSize;
1285 mtctx->jobs[u].params = jobParams;
1286 /* do not calculate checksum within sections, but write it in header for first section */
1287 mtctx->jobs[u].dstBuff = dstBuffer;
1288 mtctx->jobs[u].cctxPool = mtctx->cctxPool;
1289 mtctx->jobs[u].bufPool = mtctx->bufPool;
1290 mtctx->jobs[u].seqPool = mtctx->seqPool;
1291 mtctx->jobs[u].serial = &mtctx->serial;
1292 mtctx->jobs[u].jobID = u;
1293 mtctx->jobs[u].firstJob = (u==0);
1294 mtctx->jobs[u].lastJob = (u==nbJobs-1);
1295
1296 DEBUGLOG(5, "ZSTDMT_compress_advanced_internal: posting job %u (%u bytes)", u, (U32)jobSize);
1297 DEBUG_PRINTHEX(6, mtctx->jobs[u].prefix.start, 12);
1298 POOL_add(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[u]);
1299
1300 frameStartPos += jobSize;
1301 dstBufferPos += dstBufferCapacity;
1302 remainingSrcSize -= jobSize;
1303 } }
1304
1305 /* collect result */
1306 { size_t error = 0, dstPos = 0;
1307 unsigned jobID;
1308 for (jobID=0; jobID<nbJobs; jobID++) {
1309 DEBUGLOG(5, "waiting for job %u ", jobID);
1310 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
1311 while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
1312 DEBUGLOG(5, "waiting for jobCompleted signal from job %u", jobID);
1313 ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
1314 }
1315 ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
1316 DEBUGLOG(5, "ready to write job %u ", jobID);
1317
1318 { size_t const cSize = mtctx->jobs[jobID].cSize;
1319 if (ZSTD_isError(cSize)) error = cSize;
1320 if ((!error) && (dstPos + cSize > dstCapacity)) error = ERROR(dstSize_tooSmall);
1321 if (jobID) { /* note : job 0 is written directly at dst, which is correct position */
1322 if (!error)
1323 memmove((char*)dst + dstPos, mtctx->jobs[jobID].dstBuff.start, cSize); /* may overlap when job compressed within dst */
1324 if (jobID >= compressWithinDst) { /* job compressed into its own buffer, which must be released */
1325 DEBUGLOG(5, "releasing buffer %u>=%u", jobID, compressWithinDst);
1326 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);
1327 } }
1328 mtctx->jobs[jobID].dstBuff = g_nullBuffer;
1329 mtctx->jobs[jobID].cSize = 0;
1330 dstPos += cSize ;
1331 }
1332 } /* for (jobID=0; jobID<nbJobs; jobID++) */
1333
1334 DEBUGLOG(4, "checksumFlag : %u ", params.fParams.checksumFlag);
1335 if (params.fParams.checksumFlag) {
1336 U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
1337 if (dstPos + 4 > dstCapacity) {
1338 error = ERROR(dstSize_tooSmall);
1339 } else {
1340 DEBUGLOG(4, "writing checksum : %08X \n", checksum);
1341 MEM_writeLE32((char*)dst + dstPos, checksum);
1342 dstPos += 4;
1343 } }
1344
1345 if (!error) DEBUGLOG(4, "compressed size : %u ", (U32)dstPos);
1346 return error ? error : dstPos;
1347 }
1348}
1349
1350size_t ZSTDMT_compress_advanced(ZSTDMT_CCtx* mtctx,
1351 void* dst, size_t dstCapacity,
1352 const void* src, size_t srcSize,
1353 const ZSTD_CDict* cdict,
1354 ZSTD_parameters params,
1355 int overlapLog)
1356{
1357 ZSTD_CCtx_params cctxParams = mtctx->params;
1358 cctxParams.cParams = params.cParams;
1359 cctxParams.fParams = params.fParams;
1360 assert(ZSTD_OVERLAPLOG_MIN <= overlapLog && overlapLog <= ZSTD_OVERLAPLOG_MAX);
1361 cctxParams.overlapLog = overlapLog;
1362 return ZSTDMT_compress_advanced_internal(mtctx,
1363 dst, dstCapacity,
1364 src, srcSize,
1365 cdict, cctxParams);
1366}
1367
1368
1369size_t ZSTDMT_compressCCtx(ZSTDMT_CCtx* mtctx,
1370 void* dst, size_t dstCapacity,
1371 const void* src, size_t srcSize,
1372 int compressionLevel)
1373{
1374 ZSTD_parameters params = ZSTD_getParams(compressionLevel, srcSize, 0);
1375 int const overlapLog = ZSTDMT_overlapLog_default(params.cParams.strategy);
1376 params.fParams.contentSizeFlag = 1;
1377 return ZSTDMT_compress_advanced(mtctx, dst, dstCapacity, src, srcSize, NULL, params, overlapLog);
1378}
1379
1380
1381/* ====================================== */
1382/* ======= Streaming API ======= */
1383/* ====================================== */
1384
1385size_t ZSTDMT_initCStream_internal(
1386 ZSTDMT_CCtx* mtctx,
1387 const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
1388 const ZSTD_CDict* cdict, ZSTD_CCtx_params params,
1389 unsigned long long pledgedSrcSize)
1390{
1391 DEBUGLOG(4, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u, nbWorkers=%u, cctxPool=%u)",
1392 (U32)pledgedSrcSize, params.nbWorkers, mtctx->cctxPool->totalCCtx);
1393
1394 /* params supposed partially fully validated at this point */
1395 assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
1396 assert(!((dict) && (cdict))); /* either dict or cdict, not both */
1397
1398 /* init */
1399 if (params.nbWorkers != mtctx->params.nbWorkers)
1400 FORWARD_IF_ERROR( ZSTDMT_resize(mtctx, params.nbWorkers) );
1401
1402 if (params.jobSize != 0 && params.jobSize < ZSTDMT_JOBSIZE_MIN) params.jobSize = ZSTDMT_JOBSIZE_MIN;
1403 if (params.jobSize > (size_t)ZSTDMT_JOBSIZE_MAX) params.jobSize = (size_t)ZSTDMT_JOBSIZE_MAX;
1404
1405 mtctx->singleBlockingThread = (pledgedSrcSize <= ZSTDMT_JOBSIZE_MIN); /* do not trigger multi-threading when srcSize is too small */
1406 if (mtctx->singleBlockingThread) {
1407 ZSTD_CCtx_params const singleThreadParams = ZSTDMT_initJobCCtxParams(params);
1408 DEBUGLOG(5, "ZSTDMT_initCStream_internal: switch to single blocking thread mode");
1409 assert(singleThreadParams.nbWorkers == 0);
1410 return ZSTD_initCStream_internal(mtctx->cctxPool->cctx[0],
1411 dict, dictSize, cdict,
1412 singleThreadParams, pledgedSrcSize);
1413 }
1414
1415 DEBUGLOG(4, "ZSTDMT_initCStream_internal: %u workers", params.nbWorkers);
1416
1417 if (mtctx->allJobsCompleted == 0) { /* previous compression not correctly finished */
1418 ZSTDMT_waitForAllJobsCompleted(mtctx);
1419 ZSTDMT_releaseAllJobResources(mtctx);
1420 mtctx->allJobsCompleted = 1;
1421 }
1422
1423 mtctx->params = params;
1424 mtctx->frameContentSize = pledgedSrcSize;
1425 if (dict) {
1426 ZSTD_freeCDict(mtctx->cdictLocal);
1427 mtctx->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
1428 ZSTD_dlm_byCopy, dictContentType, /* note : a loadPrefix becomes an internal CDict */
1429 params.cParams, mtctx->cMem);
1430 mtctx->cdict = mtctx->cdictLocal;
1431 if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
1432 } else {
1433 ZSTD_freeCDict(mtctx->cdictLocal);
1434 mtctx->cdictLocal = NULL;
1435 mtctx->cdict = cdict;
1436 }
1437
1438 mtctx->targetPrefixSize = ZSTDMT_computeOverlapSize(params);
1439 DEBUGLOG(4, "overlapLog=%i => %u KB", params.overlapLog, (U32)(mtctx->targetPrefixSize>>10));
1440 mtctx->targetSectionSize = params.jobSize;
1441 if (mtctx->targetSectionSize == 0) {
1442 mtctx->targetSectionSize = 1ULL << ZSTDMT_computeTargetJobLog(params);
1443 }
1444 assert(mtctx->targetSectionSize <= (size_t)ZSTDMT_JOBSIZE_MAX);
1445
1446 if (params.rsyncable) {
1447 /* Aim for the targetsectionSize as the average job size. */
1448 U32 const jobSizeMB = (U32)(mtctx->targetSectionSize >> 20);
1449 U32 const rsyncBits = ZSTD_highbit32(jobSizeMB) + 20;
1450 assert(jobSizeMB >= 1);
1451 DEBUGLOG(4, "rsyncLog = %u", rsyncBits);
1452 mtctx->rsync.hash = 0;
1453 mtctx->rsync.hitMask = (1ULL << rsyncBits) - 1;
1454 mtctx->rsync.primePower = ZSTD_rollingHash_primePower(RSYNC_LENGTH);
1455 }
1456 if (mtctx->targetSectionSize < mtctx->targetPrefixSize) mtctx->targetSectionSize = mtctx->targetPrefixSize; /* job size must be >= overlap size */
1457 DEBUGLOG(4, "Job Size : %u KB (note : set to %u)", (U32)(mtctx->targetSectionSize>>10), (U32)params.jobSize);
1458 DEBUGLOG(4, "inBuff Size : %u KB", (U32)(mtctx->targetSectionSize>>10));
1459 ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(mtctx->targetSectionSize));
1460 {
1461 /* If ldm is enabled we need windowSize space. */
1462 size_t const windowSize = mtctx->params.ldmParams.enableLdm ? (1U << mtctx->params.cParams.windowLog) : 0;
1463 /* Two buffers of slack, plus extra space for the overlap
1464 * This is the minimum slack that LDM works with. One extra because
1465 * flush might waste up to targetSectionSize-1 bytes. Another extra
1466 * for the overlap (if > 0), then one to fill which doesn't overlap
1467 * with the LDM window.
1468 */
1469 size_t const nbSlackBuffers = 2 + (mtctx->targetPrefixSize > 0);
1470 size_t const slackSize = mtctx->targetSectionSize * nbSlackBuffers;
1471 /* Compute the total size, and always have enough slack */
1472 size_t const nbWorkers = MAX(mtctx->params.nbWorkers, 1);
1473 size_t const sectionsSize = mtctx->targetSectionSize * nbWorkers;
1474 size_t const capacity = MAX(windowSize, sectionsSize) + slackSize;
1475 if (mtctx->roundBuff.capacity < capacity) {
1476 if (mtctx->roundBuff.buffer)
1477 ZSTD_free(mtctx->roundBuff.buffer, mtctx->cMem);
1478 mtctx->roundBuff.buffer = (BYTE*)ZSTD_malloc(capacity, mtctx->cMem);
1479 if (mtctx->roundBuff.buffer == NULL) {
1480 mtctx->roundBuff.capacity = 0;
1481 return ERROR(memory_allocation);
1482 }
1483 mtctx->roundBuff.capacity = capacity;
1484 }
1485 }
1486 DEBUGLOG(4, "roundBuff capacity : %u KB", (U32)(mtctx->roundBuff.capacity>>10));
1487 mtctx->roundBuff.pos = 0;
1488 mtctx->inBuff.buffer = g_nullBuffer;
1489 mtctx->inBuff.filled = 0;
1490 mtctx->inBuff.prefix = kNullRange;
1491 mtctx->doneJobID = 0;
1492 mtctx->nextJobID = 0;
1493 mtctx->frameEnded = 0;
1494 mtctx->allJobsCompleted = 0;
1495 mtctx->consumed = 0;
1496 mtctx->produced = 0;
1497 if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params, mtctx->targetSectionSize))
1498 return ERROR(memory_allocation);
1499 return 0;
1500}
1501
1502size_t ZSTDMT_initCStream_advanced(ZSTDMT_CCtx* mtctx,
1503 const void* dict, size_t dictSize,
1504 ZSTD_parameters params,
1505 unsigned long long pledgedSrcSize)
1506{
1507 ZSTD_CCtx_params cctxParams = mtctx->params; /* retrieve sticky params */
1508 DEBUGLOG(4, "ZSTDMT_initCStream_advanced (pledgedSrcSize=%u)", (U32)pledgedSrcSize);
1509 cctxParams.cParams = params.cParams;
1510 cctxParams.fParams = params.fParams;
1511 return ZSTDMT_initCStream_internal(mtctx, dict, dictSize, ZSTD_dct_auto, NULL,
1512 cctxParams, pledgedSrcSize);
1513}
1514
1515size_t ZSTDMT_initCStream_usingCDict(ZSTDMT_CCtx* mtctx,
1516 const ZSTD_CDict* cdict,
1517 ZSTD_frameParameters fParams,
1518 unsigned long long pledgedSrcSize)
1519{
1520 ZSTD_CCtx_params cctxParams = mtctx->params;
1521 if (cdict==NULL) return ERROR(dictionary_wrong); /* method incompatible with NULL cdict */
1522 cctxParams.cParams = ZSTD_getCParamsFromCDict(cdict);
1523 cctxParams.fParams = fParams;
1524 return ZSTDMT_initCStream_internal(mtctx, NULL, 0 /*dictSize*/, ZSTD_dct_auto, cdict,
1525 cctxParams, pledgedSrcSize);
1526}
1527
1528
1529/* ZSTDMT_resetCStream() :
1530 * pledgedSrcSize can be zero == unknown (for the time being)
1531 * prefer using ZSTD_CONTENTSIZE_UNKNOWN,
1532 * as `0` might mean "empty" in the future */
1533size_t ZSTDMT_resetCStream(ZSTDMT_CCtx* mtctx, unsigned long long pledgedSrcSize)
1534{
1535 if (!pledgedSrcSize) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;
1536 return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dct_auto, 0, mtctx->params,
1537 pledgedSrcSize);
1538}
1539
1540size_t ZSTDMT_initCStream(ZSTDMT_CCtx* mtctx, int compressionLevel) {
1541 ZSTD_parameters const params = ZSTD_getParams(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0);
1542 ZSTD_CCtx_params cctxParams = mtctx->params; /* retrieve sticky params */
1543 DEBUGLOG(4, "ZSTDMT_initCStream (cLevel=%i)", compressionLevel);
1544 cctxParams.cParams = params.cParams;
1545 cctxParams.fParams = params.fParams;
1546 return ZSTDMT_initCStream_internal(mtctx, NULL, 0, ZSTD_dct_auto, NULL, cctxParams, ZSTD_CONTENTSIZE_UNKNOWN);
1547}
1548
1549
1550/* ZSTDMT_writeLastEmptyBlock()
1551 * Write a single empty block with an end-of-frame to finish a frame.
1552 * Job must be created from streaming variant.
1553 * This function is always successful if expected conditions are fulfilled.
1554 */
1555static void ZSTDMT_writeLastEmptyBlock(ZSTDMT_jobDescription* job)
1556{
1557 assert(job->lastJob == 1);
1558 assert(job->src.size == 0); /* last job is empty -> will be simplified into a last empty block */
1559 assert(job->firstJob == 0); /* cannot be first job, as it also needs to create frame header */
1560 assert(job->dstBuff.start == NULL); /* invoked from streaming variant only (otherwise, dstBuff might be user's output) */
1561 job->dstBuff = ZSTDMT_getBuffer(job->bufPool);
1562 if (job->dstBuff.start == NULL) {
1563 job->cSize = ERROR(memory_allocation);
1564 return;
1565 }
1566 assert(job->dstBuff.capacity >= ZSTD_blockHeaderSize); /* no buffer should ever be that small */
1567 job->src = kNullRange;
1568 job->cSize = ZSTD_writeLastEmptyBlock(job->dstBuff.start, job->dstBuff.capacity);
1569 assert(!ZSTD_isError(job->cSize));
1570 assert(job->consumed == 0);
1571}
1572
1573static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* mtctx, size_t srcSize, ZSTD_EndDirective endOp)
1574{
1575 unsigned const jobID = mtctx->nextJobID & mtctx->jobIDMask;
1576 int const endFrame = (endOp == ZSTD_e_end);
1577
1578 if (mtctx->nextJobID > mtctx->doneJobID + mtctx->jobIDMask) {
1579 DEBUGLOG(5, "ZSTDMT_createCompressionJob: will not create new job : table is full");
1580 assert((mtctx->nextJobID & mtctx->jobIDMask) == (mtctx->doneJobID & mtctx->jobIDMask));
1581 return 0;
1582 }
1583
1584 if (!mtctx->jobReady) {
1585 BYTE const* src = (BYTE const*)mtctx->inBuff.buffer.start;
1586 DEBUGLOG(5, "ZSTDMT_createCompressionJob: preparing job %u to compress %u bytes with %u preload ",
1587 mtctx->nextJobID, (U32)srcSize, (U32)mtctx->inBuff.prefix.size);
1588 mtctx->jobs[jobID].src.start = src;
1589 mtctx->jobs[jobID].src.size = srcSize;
1590 assert(mtctx->inBuff.filled >= srcSize);
1591 mtctx->jobs[jobID].prefix = mtctx->inBuff.prefix;
1592 mtctx->jobs[jobID].consumed = 0;
1593 mtctx->jobs[jobID].cSize = 0;
1594 mtctx->jobs[jobID].params = mtctx->params;
1595 mtctx->jobs[jobID].cdict = mtctx->nextJobID==0 ? mtctx->cdict : NULL;
1596 mtctx->jobs[jobID].fullFrameSize = mtctx->frameContentSize;
1597 mtctx->jobs[jobID].dstBuff = g_nullBuffer;
1598 mtctx->jobs[jobID].cctxPool = mtctx->cctxPool;
1599 mtctx->jobs[jobID].bufPool = mtctx->bufPool;
1600 mtctx->jobs[jobID].seqPool = mtctx->seqPool;
1601 mtctx->jobs[jobID].serial = &mtctx->serial;
1602 mtctx->jobs[jobID].jobID = mtctx->nextJobID;
1603 mtctx->jobs[jobID].firstJob = (mtctx->nextJobID==0);
1604 mtctx->jobs[jobID].lastJob = endFrame;
1605 mtctx->jobs[jobID].frameChecksumNeeded = mtctx->params.fParams.checksumFlag && endFrame && (mtctx->nextJobID>0);
1606 mtctx->jobs[jobID].dstFlushed = 0;
1607
1608 /* Update the round buffer pos and clear the input buffer to be reset */
1609 mtctx->roundBuff.pos += srcSize;
1610 mtctx->inBuff.buffer = g_nullBuffer;
1611 mtctx->inBuff.filled = 0;
1612 /* Set the prefix */
1613 if (!endFrame) {
1614 size_t const newPrefixSize = MIN(srcSize, mtctx->targetPrefixSize);
1615 mtctx->inBuff.prefix.start = src + srcSize - newPrefixSize;
1616 mtctx->inBuff.prefix.size = newPrefixSize;
1617 } else { /* endFrame==1 => no need for another input buffer */
1618 mtctx->inBuff.prefix = kNullRange;
1619 mtctx->frameEnded = endFrame;
1620 if (mtctx->nextJobID == 0) {
1621 /* single job exception : checksum is already calculated directly within worker thread */
1622 mtctx->params.fParams.checksumFlag = 0;
1623 } }
1624
1625 if ( (srcSize == 0)
1626 && (mtctx->nextJobID>0)/*single job must also write frame header*/ ) {
1627 DEBUGLOG(5, "ZSTDMT_createCompressionJob: creating a last empty block to end frame");
1628 assert(endOp == ZSTD_e_end); /* only possible case : need to end the frame with an empty last block */
1629 ZSTDMT_writeLastEmptyBlock(mtctx->jobs + jobID);
1630 mtctx->nextJobID++;
1631 return 0;
1632 }
1633 }
1634
1635 DEBUGLOG(5, "ZSTDMT_createCompressionJob: posting job %u : %u bytes (end:%u, jobNb == %u (mod:%u))",
1636 mtctx->nextJobID,
1637 (U32)mtctx->jobs[jobID].src.size,
1638 mtctx->jobs[jobID].lastJob,
1639 mtctx->nextJobID,
1640 jobID);
1641 if (POOL_tryAdd(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[jobID])) {
1642 mtctx->nextJobID++;
1643 mtctx->jobReady = 0;
1644 } else {
1645 DEBUGLOG(5, "ZSTDMT_createCompressionJob: no worker available for job %u", mtctx->nextJobID);
1646 mtctx->jobReady = 1;
1647 }
1648 return 0;
1649}
1650
1651
1652/*! ZSTDMT_flushProduced() :
1653 * flush whatever data has been produced but not yet flushed in current job.
1654 * move to next job if current one is fully flushed.
1655 * `output` : `pos` will be updated with amount of data flushed .
1656 * `blockToFlush` : if >0, the function will block and wait if there is no data available to flush .
1657 * @return : amount of data remaining within internal buffer, 0 if no more, 1 if unknown but > 0, or an error code */
1658static size_t ZSTDMT_flushProduced(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, unsigned blockToFlush, ZSTD_EndDirective end)
1659{
1660 unsigned const wJobID = mtctx->doneJobID & mtctx->jobIDMask;
1661 DEBUGLOG(5, "ZSTDMT_flushProduced (blocking:%u , job %u <= %u)",
1662 blockToFlush, mtctx->doneJobID, mtctx->nextJobID);
1663 assert(output->size >= output->pos);
1664
1665 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
1666 if ( blockToFlush
1667 && (mtctx->doneJobID < mtctx->nextJobID) ) {
1668 assert(mtctx->jobs[wJobID].dstFlushed <= mtctx->jobs[wJobID].cSize);
1669 while (mtctx->jobs[wJobID].dstFlushed == mtctx->jobs[wJobID].cSize) { /* nothing to flush */
1670 if (mtctx->jobs[wJobID].consumed == mtctx->jobs[wJobID].src.size) {
1671 DEBUGLOG(5, "job %u is completely consumed (%u == %u) => don't wait for cond, there will be none",
1672 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].consumed, (U32)mtctx->jobs[wJobID].src.size);
1673 break;
1674 }
1675 DEBUGLOG(5, "waiting for something to flush from job %u (currently flushed: %u bytes)",
1676 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
1677 ZSTD_pthread_cond_wait(&mtctx->jobs[wJobID].job_cond, &mtctx->jobs[wJobID].job_mutex); /* block when nothing to flush but some to come */
1678 } }
1679
1680 /* try to flush something */
1681 { size_t cSize = mtctx->jobs[wJobID].cSize; /* shared */
1682 size_t const srcConsumed = mtctx->jobs[wJobID].consumed; /* shared */
1683 size_t const srcSize = mtctx->jobs[wJobID].src.size; /* read-only, could be done after mutex lock, but no-declaration-after-statement */
1684 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1685 if (ZSTD_isError(cSize)) {
1686 DEBUGLOG(5, "ZSTDMT_flushProduced: job %u : compression error detected : %s",
1687 mtctx->doneJobID, ZSTD_getErrorName(cSize));
1688 ZSTDMT_waitForAllJobsCompleted(mtctx);
1689 ZSTDMT_releaseAllJobResources(mtctx);
1690 return cSize;
1691 }
1692 /* add frame checksum if necessary (can only happen once) */
1693 assert(srcConsumed <= srcSize);
1694 if ( (srcConsumed == srcSize) /* job completed -> worker no longer active */
1695 && mtctx->jobs[wJobID].frameChecksumNeeded ) {
1696 U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
1697 DEBUGLOG(4, "ZSTDMT_flushProduced: writing checksum : %08X \n", checksum);
1698 MEM_writeLE32((char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].cSize, checksum);
1699 cSize += 4;
1700 mtctx->jobs[wJobID].cSize += 4; /* can write this shared value, as worker is no longer active */
1701 mtctx->jobs[wJobID].frameChecksumNeeded = 0;
1702 }
1703
1704 if (cSize > 0) { /* compression is ongoing or completed */
1705 size_t const toFlush = MIN(cSize - mtctx->jobs[wJobID].dstFlushed, output->size - output->pos);
1706 DEBUGLOG(5, "ZSTDMT_flushProduced: Flushing %u bytes from job %u (completion:%u/%u, generated:%u)",
1707 (U32)toFlush, mtctx->doneJobID, (U32)srcConsumed, (U32)srcSize, (U32)cSize);
1708 assert(mtctx->doneJobID < mtctx->nextJobID);
1709 assert(cSize >= mtctx->jobs[wJobID].dstFlushed);
1710 assert(mtctx->jobs[wJobID].dstBuff.start != NULL);
1711 memcpy((char*)output->dst + output->pos,
1712 (const char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].dstFlushed,
1713 toFlush);
1714 output->pos += toFlush;
1715 mtctx->jobs[wJobID].dstFlushed += toFlush; /* can write : this value is only used by mtctx */
1716
1717 if ( (srcConsumed == srcSize) /* job is completed */
1718 && (mtctx->jobs[wJobID].dstFlushed == cSize) ) { /* output buffer fully flushed => free this job position */
1719 DEBUGLOG(5, "Job %u completed (%u bytes), moving to next one",
1720 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
1721 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[wJobID].dstBuff);
1722 DEBUGLOG(5, "dstBuffer released");
1723 mtctx->jobs[wJobID].dstBuff = g_nullBuffer;
1724 mtctx->jobs[wJobID].cSize = 0; /* ensure this job slot is considered "not started" in future check */
1725 mtctx->consumed += srcSize;
1726 mtctx->produced += cSize;
1727 mtctx->doneJobID++;
1728 } }
1729
1730 /* return value : how many bytes left in buffer ; fake it to 1 when unknown but >0 */
1731 if (cSize > mtctx->jobs[wJobID].dstFlushed) return (cSize - mtctx->jobs[wJobID].dstFlushed);
1732 if (srcSize > srcConsumed) return 1; /* current job not completely compressed */
1733 }
1734 if (mtctx->doneJobID < mtctx->nextJobID) return 1; /* some more jobs ongoing */
1735 if (mtctx->jobReady) return 1; /* one job is ready to push, just not yet in the list */
1736 if (mtctx->inBuff.filled > 0) return 1; /* input is not empty, and still needs to be converted into a job */
1737 mtctx->allJobsCompleted = mtctx->frameEnded; /* all jobs are entirely flushed => if this one is last one, frame is completed */
1738 if (end == ZSTD_e_end) return !mtctx->frameEnded; /* for ZSTD_e_end, question becomes : is frame completed ? instead of : are internal buffers fully flushed ? */
1739 return 0; /* internal buffers fully flushed */
1740}
1741
1742/**
1743 * Returns the range of data used by the earliest job that is not yet complete.
1744 * If the data of the first job is broken up into two segments, we cover both
1745 * sections.
1746 */
1747static range_t ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx)
1748{
1749 unsigned const firstJobID = mtctx->doneJobID;
1750 unsigned const lastJobID = mtctx->nextJobID;
1751 unsigned jobID;
1752
1753 for (jobID = firstJobID; jobID < lastJobID; ++jobID) {
1754 unsigned const wJobID = jobID & mtctx->jobIDMask;
1755 size_t consumed;
1756
1757 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
1758 consumed = mtctx->jobs[wJobID].consumed;
1759 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1760
1761 if (consumed < mtctx->jobs[wJobID].src.size) {
1762 range_t range = mtctx->jobs[wJobID].prefix;
1763 if (range.size == 0) {
1764 /* Empty prefix */
1765 range = mtctx->jobs[wJobID].src;
1766 }
1767 /* Job source in multiple segments not supported yet */
1768 assert(range.start <= mtctx->jobs[wJobID].src.start);
1769 return range;
1770 }
1771 }
1772 return kNullRange;
1773}
1774
1775/**
1776 * Returns non-zero iff buffer and range overlap.
1777 */
1778static int ZSTDMT_isOverlapped(buffer_t buffer, range_t range)
1779{
1780 BYTE const* const bufferStart = (BYTE const*)buffer.start;
1781 BYTE const* const bufferEnd = bufferStart + buffer.capacity;
1782 BYTE const* const rangeStart = (BYTE const*)range.start;
1783 BYTE const* const rangeEnd = rangeStart + range.size;
1784
1785 if (rangeStart == NULL || bufferStart == NULL)
1786 return 0;
1787 /* Empty ranges cannot overlap */
1788 if (bufferStart == bufferEnd || rangeStart == rangeEnd)
1789 return 0;
1790
1791 return bufferStart < rangeEnd && rangeStart < bufferEnd;
1792}
1793
1794static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window)
1795{
1796 range_t extDict;
1797 range_t prefix;
1798
1799 DEBUGLOG(5, "ZSTDMT_doesOverlapWindow");
1800 extDict.start = window.dictBase + window.lowLimit;
1801 extDict.size = window.dictLimit - window.lowLimit;
1802
1803 prefix.start = window.base + window.dictLimit;
1804 prefix.size = window.nextSrc - (window.base + window.dictLimit);
1805 DEBUGLOG(5, "extDict [0x%zx, 0x%zx)",
1806 (size_t)extDict.start,
1807 (size_t)extDict.start + extDict.size);
1808 DEBUGLOG(5, "prefix [0x%zx, 0x%zx)",
1809 (size_t)prefix.start,
1810 (size_t)prefix.start + prefix.size);
1811
1812 return ZSTDMT_isOverlapped(buffer, extDict)
1813 || ZSTDMT_isOverlapped(buffer, prefix);
1814}
1815
1816static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, buffer_t buffer)
1817{
1818 if (mtctx->params.ldmParams.enableLdm) {
1819 ZSTD_pthread_mutex_t* mutex = &mtctx->serial.ldmWindowMutex;
1820 DEBUGLOG(5, "ZSTDMT_waitForLdmComplete");
1821 DEBUGLOG(5, "source [0x%zx, 0x%zx)",
1822 (size_t)buffer.start,
1823 (size_t)buffer.start + buffer.capacity);
1824 ZSTD_PTHREAD_MUTEX_LOCK(mutex);
1825 while (ZSTDMT_doesOverlapWindow(buffer, mtctx->serial.ldmWindow)) {
1826 DEBUGLOG(5, "Waiting for LDM to finish...");
1827 ZSTD_pthread_cond_wait(&mtctx->serial.ldmWindowCond, mutex);
1828 }
1829 DEBUGLOG(6, "Done waiting for LDM to finish");
1830 ZSTD_pthread_mutex_unlock(mutex);
1831 }
1832}
1833
1834/**
1835 * Attempts to set the inBuff to the next section to fill.
1836 * If any part of the new section is still in use we give up.
1837 * Returns non-zero if the buffer is filled.
1838 */
1839static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx)
1840{
1841 range_t const inUse = ZSTDMT_getInputDataInUse(mtctx);
1842 size_t const spaceLeft = mtctx->roundBuff.capacity - mtctx->roundBuff.pos;
1843 size_t const target = mtctx->targetSectionSize;
1844 buffer_t buffer;
1845
1846 DEBUGLOG(5, "ZSTDMT_tryGetInputRange");
1847 assert(mtctx->inBuff.buffer.start == NULL);
1848 assert(mtctx->roundBuff.capacity >= target);
1849
1850 if (spaceLeft < target) {
1851 /* ZSTD_invalidateRepCodes() doesn't work for extDict variants.
1852 * Simply copy the prefix to the beginning in that case.
1853 */
1854 BYTE* const start = (BYTE*)mtctx->roundBuff.buffer;
1855 size_t const prefixSize = mtctx->inBuff.prefix.size;
1856
1857 buffer.start = start;
1858 buffer.capacity = prefixSize;
1859 if (ZSTDMT_isOverlapped(buffer, inUse)) {
1860 DEBUGLOG(5, "Waiting for buffer...");
1861 return 0;
1862 }
1863 ZSTDMT_waitForLdmComplete(mtctx, buffer);
1864 memmove(start, mtctx->inBuff.prefix.start, prefixSize);
1865 mtctx->inBuff.prefix.start = start;
1866 mtctx->roundBuff.pos = prefixSize;
1867 }
1868 buffer.start = mtctx->roundBuff.buffer + mtctx->roundBuff.pos;
1869 buffer.capacity = target;
1870
1871 if (ZSTDMT_isOverlapped(buffer, inUse)) {
1872 DEBUGLOG(5, "Waiting for buffer...");
1873 return 0;
1874 }
1875 assert(!ZSTDMT_isOverlapped(buffer, mtctx->inBuff.prefix));
1876
1877 ZSTDMT_waitForLdmComplete(mtctx, buffer);
1878
1879 DEBUGLOG(5, "Using prefix range [%zx, %zx)",
1880 (size_t)mtctx->inBuff.prefix.start,
1881 (size_t)mtctx->inBuff.prefix.start + mtctx->inBuff.prefix.size);
1882 DEBUGLOG(5, "Using source range [%zx, %zx)",
1883 (size_t)buffer.start,
1884 (size_t)buffer.start + buffer.capacity);
1885
1886
1887 mtctx->inBuff.buffer = buffer;
1888 mtctx->inBuff.filled = 0;
1889 assert(mtctx->roundBuff.pos + buffer.capacity <= mtctx->roundBuff.capacity);
1890 return 1;
1891}
1892
1893typedef struct {
1894 size_t toLoad; /* The number of bytes to load from the input. */
1895 int flush; /* Boolean declaring if we must flush because we found a synchronization point. */
1896} syncPoint_t;
1897
1898/**
1899 * Searches through the input for a synchronization point. If one is found, we
1900 * will instruct the caller to flush, and return the number of bytes to load.
1901 * Otherwise, we will load as many bytes as possible and instruct the caller
1902 * to continue as normal.
1903 */
1904static syncPoint_t
1905findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
1906{
1907 BYTE const* const istart = (BYTE const*)input.src + input.pos;
1908 U64 const primePower = mtctx->rsync.primePower;
1909 U64 const hitMask = mtctx->rsync.hitMask;
1910
1911 syncPoint_t syncPoint;
1912 U64 hash;
1913 BYTE const* prev;
1914 size_t pos;
1915
1916 syncPoint.toLoad = MIN(input.size - input.pos, mtctx->targetSectionSize - mtctx->inBuff.filled);
1917 syncPoint.flush = 0;
1918 if (!mtctx->params.rsyncable)
1919 /* Rsync is disabled. */
1920 return syncPoint;
1921 if (mtctx->inBuff.filled + syncPoint.toLoad < RSYNC_LENGTH)
1922 /* Not enough to compute the hash.
1923 * We will miss any synchronization points in this RSYNC_LENGTH byte
1924 * window. However, since it depends only in the internal buffers, if the
1925 * state is already synchronized, we will remain synchronized.
1926 * Additionally, the probability that we miss a synchronization point is
1927 * low: RSYNC_LENGTH / targetSectionSize.
1928 */
1929 return syncPoint;
1930 /* Initialize the loop variables. */
1931 if (mtctx->inBuff.filled >= RSYNC_LENGTH) {
1932 /* We have enough bytes buffered to initialize the hash.
1933 * Start scanning at the beginning of the input.
1934 */
1935 pos = 0;
1936 prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
1937 hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
1938 } else {
1939 /* We don't have enough bytes buffered to initialize the hash, but
1940 * we know we have at least RSYNC_LENGTH bytes total.
1941 * Start scanning after the first RSYNC_LENGTH bytes less the bytes
1942 * already buffered.
1943 */
1944 pos = RSYNC_LENGTH - mtctx->inBuff.filled;
1945 prev = (BYTE const*)mtctx->inBuff.buffer.start - pos;
1946 hash = ZSTD_rollingHash_compute(mtctx->inBuff.buffer.start, mtctx->inBuff.filled);
1947 hash = ZSTD_rollingHash_append(hash, istart, pos);
1948 }
1949 /* Starting with the hash of the previous RSYNC_LENGTH bytes, roll
1950 * through the input. If we hit a synchronization point, then cut the
1951 * job off, and tell the compressor to flush the job. Otherwise, load
1952 * all the bytes and continue as normal.
1953 * If we go too long without a synchronization point (targetSectionSize)
1954 * then a block will be emitted anyways, but this is okay, since if we
1955 * are already synchronized we will remain synchronized.
1956 */
1957 for (; pos < syncPoint.toLoad; ++pos) {
1958 BYTE const toRemove = pos < RSYNC_LENGTH ? prev[pos] : istart[pos - RSYNC_LENGTH];
1959 /* if (pos >= RSYNC_LENGTH) assert(ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash); */
1960 hash = ZSTD_rollingHash_rotate(hash, toRemove, istart[pos], primePower);
1961 if ((hash & hitMask) == hitMask) {
1962 syncPoint.toLoad = pos + 1;
1963 syncPoint.flush = 1;
1964 break;
1965 }
1966 }
1967 return syncPoint;
1968}
1969
1970size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx)
1971{
1972 size_t hintInSize = mtctx->targetSectionSize - mtctx->inBuff.filled;
1973 if (hintInSize==0) hintInSize = mtctx->targetSectionSize;
1974 return hintInSize;
1975}
1976
1977/** ZSTDMT_compressStream_generic() :
1978 * internal use only - exposed to be invoked from zstd_compress.c
1979 * assumption : output and input are valid (pos <= size)
1980 * @return : minimum amount of data remaining to flush, 0 if none */
1981size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
1982 ZSTD_outBuffer* output,
1983 ZSTD_inBuffer* input,
1984 ZSTD_EndDirective endOp)
1985{
1986 unsigned forwardInputProgress = 0;
1987 DEBUGLOG(5, "ZSTDMT_compressStream_generic (endOp=%u, srcSize=%u)",
1988 (U32)endOp, (U32)(input->size - input->pos));
1989 assert(output->pos <= output->size);
1990 assert(input->pos <= input->size);
1991
1992 if (mtctx->singleBlockingThread) { /* delegate to single-thread (synchronous) */
1993 return ZSTD_compressStream2(mtctx->cctxPool->cctx[0], output, input, endOp);
1994 }
1995
1996 if ((mtctx->frameEnded) && (endOp==ZSTD_e_continue)) {
1997 /* current frame being ended. Only flush/end are allowed */
1998 return ERROR(stage_wrong);
1999 }
2000
2001 /* single-pass shortcut (note : synchronous-mode) */
2002 if ( (!mtctx->params.rsyncable) /* rsyncable mode is disabled */
2003 && (mtctx->nextJobID == 0) /* just started */
2004 && (mtctx->inBuff.filled == 0) /* nothing buffered */
2005 && (!mtctx->jobReady) /* no job already created */
2006 && (endOp == ZSTD_e_end) /* end order */
2007 && (output->size - output->pos >= ZSTD_compressBound(input->size - input->pos)) ) { /* enough space in dst */
2008 size_t const cSize = ZSTDMT_compress_advanced_internal(mtctx,
2009 (char*)output->dst + output->pos, output->size - output->pos,
2010 (const char*)input->src + input->pos, input->size - input->pos,
2011 mtctx->cdict, mtctx->params);
2012 if (ZSTD_isError(cSize)) return cSize;
2013 input->pos = input->size;
2014 output->pos += cSize;
2015 mtctx->allJobsCompleted = 1;
2016 mtctx->frameEnded = 1;
2017 return 0;
2018 }
2019
2020 /* fill input buffer */
2021 if ( (!mtctx->jobReady)
2022 && (input->size > input->pos) ) { /* support NULL input */
2023 if (mtctx->inBuff.buffer.start == NULL) {
2024 assert(mtctx->inBuff.filled == 0); /* Can't fill an empty buffer */
2025 if (!ZSTDMT_tryGetInputRange(mtctx)) {
2026 /* It is only possible for this operation to fail if there are
2027 * still compression jobs ongoing.
2028 */
2029 DEBUGLOG(5, "ZSTDMT_tryGetInputRange failed");
2030 assert(mtctx->doneJobID != mtctx->nextJobID);
2031 } else
2032 DEBUGLOG(5, "ZSTDMT_tryGetInputRange completed successfully : mtctx->inBuff.buffer.start = %p", mtctx->inBuff.buffer.start);
2033 }
2034 if (mtctx->inBuff.buffer.start != NULL) {
2035 syncPoint_t const syncPoint = findSynchronizationPoint(mtctx, *input);
2036 if (syncPoint.flush && endOp == ZSTD_e_continue) {
2037 endOp = ZSTD_e_flush;
2038 }
2039 assert(mtctx->inBuff.buffer.capacity >= mtctx->targetSectionSize);
2040 DEBUGLOG(5, "ZSTDMT_compressStream_generic: adding %u bytes on top of %u to buffer of size %u",
2041 (U32)syncPoint.toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize);
2042 memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, syncPoint.toLoad);
2043 input->pos += syncPoint.toLoad;
2044 mtctx->inBuff.filled += syncPoint.toLoad;
2045 forwardInputProgress = syncPoint.toLoad>0;
2046 }
2047 if ((input->pos < input->size) && (endOp == ZSTD_e_end))
2048 endOp = ZSTD_e_flush; /* can't end now : not all input consumed */
2049 }
2050
2051 if ( (mtctx->jobReady)
2052 || (mtctx->inBuff.filled >= mtctx->targetSectionSize) /* filled enough : let's compress */
2053 || ((endOp != ZSTD_e_continue) && (mtctx->inBuff.filled > 0)) /* something to flush : let's go */
2054 || ((endOp == ZSTD_e_end) && (!mtctx->frameEnded)) ) { /* must finish the frame with a zero-size block */
2055 size_t const jobSize = mtctx->inBuff.filled;
2056 assert(mtctx->inBuff.filled <= mtctx->targetSectionSize);
2057 FORWARD_IF_ERROR( ZSTDMT_createCompressionJob(mtctx, jobSize, endOp) );
2058 }
2059
2060 /* check for potential compressed data ready to be flushed */
2061 { size_t const remainingToFlush = ZSTDMT_flushProduced(mtctx, output, !forwardInputProgress, endOp); /* block if there was no forward input progress */
2062 if (input->pos < input->size) return MAX(remainingToFlush, 1); /* input not consumed : do not end flush yet */
2063 DEBUGLOG(5, "end of ZSTDMT_compressStream_generic: remainingToFlush = %u", (U32)remainingToFlush);
2064 return remainingToFlush;
2065 }
2066}
2067
2068
2069size_t ZSTDMT_compressStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
2070{
2071 FORWARD_IF_ERROR( ZSTDMT_compressStream_generic(mtctx, output, input, ZSTD_e_continue) );
2072
2073 /* recommended next input size : fill current input buffer */
2074 return mtctx->targetSectionSize - mtctx->inBuff.filled; /* note : could be zero when input buffer is fully filled and no more availability to create new job */
2075}
2076
2077
2078static size_t ZSTDMT_flushStream_internal(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, ZSTD_EndDirective endFrame)
2079{
2080 size_t const srcSize = mtctx->inBuff.filled;
2081 DEBUGLOG(5, "ZSTDMT_flushStream_internal");
2082
2083 if ( mtctx->jobReady /* one job ready for a worker to pick up */
2084 || (srcSize > 0) /* still some data within input buffer */
2085 || ((endFrame==ZSTD_e_end) && !mtctx->frameEnded)) { /* need a last 0-size block to end frame */
2086 DEBUGLOG(5, "ZSTDMT_flushStream_internal : create a new job (%u bytes, end:%u)",
2087 (U32)srcSize, (U32)endFrame);
2088 FORWARD_IF_ERROR( ZSTDMT_createCompressionJob(mtctx, srcSize, endFrame) );
2089 }
2090
2091 /* check if there is any data available to flush */
2092 return ZSTDMT_flushProduced(mtctx, output, 1 /* blockToFlush */, endFrame);
2093}
2094
2095
2096size_t ZSTDMT_flushStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output)
2097{
2098 DEBUGLOG(5, "ZSTDMT_flushStream");
2099 if (mtctx->singleBlockingThread)
2100 return ZSTD_flushStream(mtctx->cctxPool->cctx[0], output);
2101 return ZSTDMT_flushStream_internal(mtctx, output, ZSTD_e_flush);
2102}
2103
2104size_t ZSTDMT_endStream(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output)
2105{
2106 DEBUGLOG(4, "ZSTDMT_endStream");
2107 if (mtctx->singleBlockingThread)
2108 return ZSTD_endStream(mtctx->cctxPool->cctx[0], output);
2109 return ZSTDMT_flushStream_internal(mtctx, output, ZSTD_e_end);
2110}