1 /*
2 * Copyright (c) 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 "../common/zstd_deps.h" /* ZSTD_memcpy, ZSTD_memset, INT_MAX, UINT_MAX */
24 #include "../common/mem.h" /* MEM_STATIC */
25 #include "../common/pool.h" /* threadpool */
26 #include "../common/threading.h" /* mutex */
27 #include "zstd_compress_internal.h" /* MIN, ERROR, ZSTD_*, ZSTD_highbit32 */
28 #include "zstd_ldm.h"
29 #include "zstdmt_compress.h"
30
31 /* Guards code to support resizing the SeqPool.
32 * We will want to resize the SeqPool to save memory in the future.
33 * Until then, comment the code out since it is unused.
34 */
35 #define ZSTD_RESIZE_SEQPOOL 0
36
37 /* ====== Debug ====== */
38 #if defined(DEBUGLEVEL) && (DEBUGLEVEL>=2) \
39 && !defined(_MSC_VER) \
40 && !defined(__MINGW32__)
41
42 # include <stdio.h>
43 # include <unistd.h>
44 # include <sys/times.h>
45
46 # define DEBUG_PRINTHEX(l,p,n) { \
47 unsigned debug_u; \
48 for (debug_u=0; debug_u<(n); debug_u++) \
49 RAWLOG(l, "%02X ", ((const unsigned char*)(p))[debug_u]); \
50 RAWLOG(l, " \n"); \
51 }
52
GetCurrentClockTimeMicroseconds(void)53 static unsigned long long GetCurrentClockTimeMicroseconds(void)
54 {
55 static clock_t _ticksPerSecond = 0;
56 if (_ticksPerSecond <= 0) _ticksPerSecond = sysconf(_SC_CLK_TCK);
57
58 { struct tms junk; clock_t newTicks = (clock_t) times(&junk);
59 return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond);
60 } }
61
62 #define MUTEX_WAIT_TIME_DLEVEL 6
63 #define ZSTD_PTHREAD_MUTEX_LOCK(mutex) { \
64 if (DEBUGLEVEL >= MUTEX_WAIT_TIME_DLEVEL) { \
65 unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds(); \
66 ZSTD_pthread_mutex_lock(mutex); \
67 { unsigned long long const afterTime = GetCurrentClockTimeMicroseconds(); \
68 unsigned long long const elapsedTime = (afterTime-beforeTime); \
69 if (elapsedTime > 1000) { /* or whatever threshold you like; I'm using 1 millisecond here */ \
70 DEBUGLOG(MUTEX_WAIT_TIME_DLEVEL, "Thread took %llu microseconds to acquire mutex %s \n", \
71 elapsedTime, #mutex); \
72 } } \
73 } else { \
74 ZSTD_pthread_mutex_lock(mutex); \
75 } \
76 }
77
78 #else
79
80 # define ZSTD_PTHREAD_MUTEX_LOCK(m) ZSTD_pthread_mutex_lock(m)
81 # define DEBUG_PRINTHEX(l,p,n) {}
82
83 #endif
84
85
86 /* ===== Buffer Pool ===== */
87 /* a single Buffer Pool can be invoked from multiple threads in parallel */
88
89 typedef struct buffer_s {
90 void* start;
91 size_t capacity;
92 } buffer_t;
93
94 static const buffer_t g_nullBuffer = { NULL, 0 };
95
96 typedef struct ZSTDMT_bufferPool_s {
97 ZSTD_pthread_mutex_t poolMutex;
98 size_t bufferSize;
99 unsigned totalBuffers;
100 unsigned nbBuffers;
101 ZSTD_customMem cMem;
102 buffer_t bTable[1]; /* variable size */
103 } ZSTDMT_bufferPool;
104
ZSTDMT_createBufferPool(unsigned nbWorkers,ZSTD_customMem cMem)105 static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned nbWorkers, ZSTD_customMem cMem)
106 {
107 unsigned const maxNbBuffers = 2*nbWorkers + 3;
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);
113 return NULL;
114 }
115 bufPool->bufferSize = 64 KB;
116 bufPool->totalBuffers = maxNbBuffers;
117 bufPool->nbBuffers = 0;
118 bufPool->cMem = cMem;
119 return bufPool;
120 }
121
ZSTDMT_freeBufferPool(ZSTDMT_bufferPool * bufPool)122 static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool)
123 {
124 unsigned u;
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);
130 }
131 ZSTD_pthread_mutex_destroy(&bufPool->poolMutex);
132 ZSTD_customFree(bufPool, bufPool->cMem);
133 }
134
135 /* only works at initialization, not during compression */
ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool * bufPool)136 static size_t ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool* bufPool)
137 {
138 size_t const poolSize = sizeof(*bufPool)
139 + (bufPool->totalBuffers - 1) * sizeof(buffer_t);
140 unsigned u;
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);
146
147 return poolSize + totalBufferSize;
148 }
149
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 */
ZSTDMT_setBufferSize(ZSTDMT_bufferPool * const bufPool,size_t const bSize)154 static void ZSTDMT_setBufferSize(ZSTDMT_bufferPool* const bufPool, size_t const bSize)
155 {
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);
160 }
161
162
ZSTDMT_expandBufferPool(ZSTDMT_bufferPool * srcBufPool,U32 nbWorkers)163 static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, U32 nbWorkers)
164 {
165 unsigned const maxNbBuffers = 2*nbWorkers + 3;
166 if (srcBufPool==NULL) return NULL;
167 if (srcBufPool->totalBuffers >= maxNbBuffers) /* good enough */
168 return srcBufPool;
169 /* need a larger buffer pool */
170 { ZSTD_customMem const cMem = srcBufPool->cMem;
171 size_t const bSize = srcBufPool->bufferSize; /* forward parameters */
172 ZSTDMT_bufferPool* newBufPool;
173 ZSTDMT_freeBufferPool(srcBufPool);
174 newBufPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
175 if (newBufPool==NULL) return newBufPool;
176 ZSTDMT_setBufferSize(newBufPool, bSize);
177 return newBufPool;
178 }
179 }
180
181 /** ZSTDMT_getBuffer() :
182 * assumption : bufPool must be valid
183 * @return : a buffer, with start pointer and size
184 * note: allocation may fail, in this case, start==NULL and size==0 */
ZSTDMT_getBuffer(ZSTDMT_bufferPool * bufPool)185 static buffer_t ZSTDMT_getBuffer(ZSTDMT_bufferPool* bufPool)
186 {
187 size_t const bSize = bufPool->bufferSize;
188 DEBUGLOG(5, "ZSTDMT_getBuffer: bSize = %u", (U32)bufPool->bufferSize);
189 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
190 if (bufPool->nbBuffers) { /* try to use an existing buffer */
191 buffer_t const buf = bufPool->bTable[--(bufPool->nbBuffers)];
192 size_t const availBufferSize = buf.capacity;
193 bufPool->bTable[bufPool->nbBuffers] = g_nullBuffer;
194 if ((availBufferSize >= bSize) & ((availBufferSize>>3) <= bSize)) {
195 /* large enough, but not too much */
196 DEBUGLOG(5, "ZSTDMT_getBuffer: provide buffer %u of size %u",
197 bufPool->nbBuffers, (U32)buf.capacity);
198 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
199 return buf;
200 }
201 /* size conditions not respected : scratch this buffer, create new one */
202 DEBUGLOG(5, "ZSTDMT_getBuffer: existing buffer does not meet size conditions => freeing");
203 ZSTD_customFree(buf.start, bufPool->cMem);
204 }
205 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
206 /* create new buffer */
207 DEBUGLOG(5, "ZSTDMT_getBuffer: create a new buffer");
208 { buffer_t buffer;
209 void* const start = ZSTD_customMalloc(bSize, bufPool->cMem);
210 buffer.start = start; /* note : start can be NULL if malloc fails ! */
211 buffer.capacity = (start==NULL) ? 0 : bSize;
212 if (start==NULL) {
213 DEBUGLOG(5, "ZSTDMT_getBuffer: buffer allocation failure !!");
214 } else {
215 DEBUGLOG(5, "ZSTDMT_getBuffer: created buffer of size %u", (U32)bSize);
216 }
217 return buffer;
218 }
219 }
220
221 #if ZSTD_RESIZE_SEQPOOL
222 /** ZSTDMT_resizeBuffer() :
223 * assumption : bufPool must be valid
224 * @return : a buffer that is at least the buffer pool buffer size.
225 * If a reallocation happens, the data in the input buffer is copied.
226 */
ZSTDMT_resizeBuffer(ZSTDMT_bufferPool * bufPool,buffer_t buffer)227 static buffer_t ZSTDMT_resizeBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buffer)
228 {
229 size_t const bSize = bufPool->bufferSize;
230 if (buffer.capacity < bSize) {
231 void* const start = ZSTD_customMalloc(bSize, bufPool->cMem);
232 buffer_t newBuffer;
233 newBuffer.start = start;
234 newBuffer.capacity = start == NULL ? 0 : bSize;
235 if (start != NULL) {
236 assert(newBuffer.capacity >= buffer.capacity);
237 ZSTD_memcpy(newBuffer.start, buffer.start, buffer.capacity);
238 DEBUGLOG(5, "ZSTDMT_resizeBuffer: created buffer of size %u", (U32)bSize);
239 return newBuffer;
240 }
241 DEBUGLOG(5, "ZSTDMT_resizeBuffer: buffer allocation failure !!");
242 }
243 return buffer;
244 }
245 #endif
246
247 /* store buffer for later re-use, up to pool capacity */
ZSTDMT_releaseBuffer(ZSTDMT_bufferPool * bufPool,buffer_t buf)248 static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, buffer_t buf)
249 {
250 DEBUGLOG(5, "ZSTDMT_releaseBuffer");
251 if (buf.start == NULL) return; /* compatible with release on NULL */
252 ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
253 if (bufPool->nbBuffers < bufPool->totalBuffers) {
254 bufPool->bTable[bufPool->nbBuffers++] = buf; /* stored for later use */
255 DEBUGLOG(5, "ZSTDMT_releaseBuffer: stored buffer of size %u in slot %u",
256 (U32)buf.capacity, (U32)(bufPool->nbBuffers-1));
257 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
258 return;
259 }
260 ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
261 /* Reached bufferPool capacity (should not happen) */
262 DEBUGLOG(5, "ZSTDMT_releaseBuffer: pool capacity reached => freeing ");
263 ZSTD_customFree(buf.start, bufPool->cMem);
264 }
265
266
267 /* ===== Seq Pool Wrapper ====== */
268
269 typedef ZSTDMT_bufferPool ZSTDMT_seqPool;
270
ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool * seqPool)271 static size_t ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool* seqPool)
272 {
273 return ZSTDMT_sizeof_bufferPool(seqPool);
274 }
275
bufferToSeq(buffer_t buffer)276 static rawSeqStore_t bufferToSeq(buffer_t buffer)
277 {
278 rawSeqStore_t seq = kNullRawSeqStore;
279 seq.seq = (rawSeq*)buffer.start;
280 seq.capacity = buffer.capacity / sizeof(rawSeq);
281 return seq;
282 }
283
seqToBuffer(rawSeqStore_t seq)284 static buffer_t seqToBuffer(rawSeqStore_t seq)
285 {
286 buffer_t buffer;
287 buffer.start = seq.seq;
288 buffer.capacity = seq.capacity * sizeof(rawSeq);
289 return buffer;
290 }
291
ZSTDMT_getSeq(ZSTDMT_seqPool * seqPool)292 static rawSeqStore_t ZSTDMT_getSeq(ZSTDMT_seqPool* seqPool)
293 {
294 if (seqPool->bufferSize == 0) {
295 return kNullRawSeqStore;
296 }
297 return bufferToSeq(ZSTDMT_getBuffer(seqPool));
298 }
299
300 #if ZSTD_RESIZE_SEQPOOL
ZSTDMT_resizeSeq(ZSTDMT_seqPool * seqPool,rawSeqStore_t seq)301 static rawSeqStore_t ZSTDMT_resizeSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
302 {
303 return bufferToSeq(ZSTDMT_resizeBuffer(seqPool, seqToBuffer(seq)));
304 }
305 #endif
306
ZSTDMT_releaseSeq(ZSTDMT_seqPool * seqPool,rawSeqStore_t seq)307 static void ZSTDMT_releaseSeq(ZSTDMT_seqPool* seqPool, rawSeqStore_t seq)
308 {
309 ZSTDMT_releaseBuffer(seqPool, seqToBuffer(seq));
310 }
311
ZSTDMT_setNbSeq(ZSTDMT_seqPool * const seqPool,size_t const nbSeq)312 static void ZSTDMT_setNbSeq(ZSTDMT_seqPool* const seqPool, size_t const nbSeq)
313 {
314 ZSTDMT_setBufferSize(seqPool, nbSeq * sizeof(rawSeq));
315 }
316
ZSTDMT_createSeqPool(unsigned nbWorkers,ZSTD_customMem cMem)317 static ZSTDMT_seqPool* ZSTDMT_createSeqPool(unsigned nbWorkers, ZSTD_customMem cMem)
318 {
319 ZSTDMT_seqPool* const seqPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
320 if (seqPool == NULL) return NULL;
321 ZSTDMT_setNbSeq(seqPool, 0);
322 return seqPool;
323 }
324
ZSTDMT_freeSeqPool(ZSTDMT_seqPool * seqPool)325 static void ZSTDMT_freeSeqPool(ZSTDMT_seqPool* seqPool)
326 {
327 ZSTDMT_freeBufferPool(seqPool);
328 }
329
ZSTDMT_expandSeqPool(ZSTDMT_seqPool * pool,U32 nbWorkers)330 static ZSTDMT_seqPool* ZSTDMT_expandSeqPool(ZSTDMT_seqPool* pool, U32 nbWorkers)
331 {
332 return ZSTDMT_expandBufferPool(pool, nbWorkers);
333 }
334
335
336 /* ===== CCtx Pool ===== */
337 /* a single CCtx Pool can be invoked from multiple threads in parallel */
338
339 typedef struct {
340 ZSTD_pthread_mutex_t poolMutex;
341 int totalCCtx;
342 int availCCtx;
343 ZSTD_customMem cMem;
344 ZSTD_CCtx* cctx[1]; /* variable size */
345 } ZSTDMT_CCtxPool;
346
347 /* note : all CCtx borrowed from the pool should be released back to the pool _before_ freeing the pool */
ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool * pool)348 static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
349 {
350 int cid;
351 for (cid=0; cid<pool->totalCCtx; cid++)
352 ZSTD_freeCCtx(pool->cctx[cid]); /* note : compatible with free on NULL */
353 ZSTD_pthread_mutex_destroy(&pool->poolMutex);
354 ZSTD_customFree(pool, pool->cMem);
355 }
356
357 /* ZSTDMT_createCCtxPool() :
358 * implies nbWorkers >= 1 , checked by caller ZSTDMT_createCCtx() */
ZSTDMT_createCCtxPool(int nbWorkers,ZSTD_customMem cMem)359 static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(int nbWorkers,
360 ZSTD_customMem cMem)
361 {
362 ZSTDMT_CCtxPool* const cctxPool = (ZSTDMT_CCtxPool*) ZSTD_customCalloc(
363 sizeof(ZSTDMT_CCtxPool) + (nbWorkers-1)*sizeof(ZSTD_CCtx*), cMem);
364 assert(nbWorkers > 0);
365 if (!cctxPool) return NULL;
366 if (ZSTD_pthread_mutex_init(&cctxPool->poolMutex, NULL)) {
367 ZSTD_customFree(cctxPool, cMem);
368 return NULL;
369 }
370 cctxPool->cMem = cMem;
371 cctxPool->totalCCtx = nbWorkers;
372 cctxPool->availCCtx = 1; /* at least one cctx for single-thread mode */
373 cctxPool->cctx[0] = ZSTD_createCCtx_advanced(cMem);
374 if (!cctxPool->cctx[0]) { ZSTDMT_freeCCtxPool(cctxPool); return NULL; }
375 DEBUGLOG(3, "cctxPool created, with %u workers", nbWorkers);
376 return cctxPool;
377 }
378
ZSTDMT_expandCCtxPool(ZSTDMT_CCtxPool * srcPool,int nbWorkers)379 static ZSTDMT_CCtxPool* ZSTDMT_expandCCtxPool(ZSTDMT_CCtxPool* srcPool,
380 int nbWorkers)
381 {
382 if (srcPool==NULL) return NULL;
383 if (nbWorkers <= srcPool->totalCCtx) return srcPool; /* good enough */
384 /* need a larger cctx pool */
385 { ZSTD_customMem const cMem = srcPool->cMem;
386 ZSTDMT_freeCCtxPool(srcPool);
387 return ZSTDMT_createCCtxPool(nbWorkers, cMem);
388 }
389 }
390
391 /* only works during initialization phase, not during compression */
ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool * cctxPool)392 static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
393 {
394 ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
395 { unsigned const nbWorkers = cctxPool->totalCCtx;
396 size_t const poolSize = sizeof(*cctxPool)
397 + (nbWorkers-1) * sizeof(ZSTD_CCtx*);
398 unsigned u;
399 size_t totalCCtxSize = 0;
400 for (u=0; u<nbWorkers; u++) {
401 totalCCtxSize += ZSTD_sizeof_CCtx(cctxPool->cctx[u]);
402 }
403 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
404 assert(nbWorkers > 0);
405 return poolSize + totalCCtxSize;
406 }
407 }
408
ZSTDMT_getCCtx(ZSTDMT_CCtxPool * cctxPool)409 static ZSTD_CCtx* ZSTDMT_getCCtx(ZSTDMT_CCtxPool* cctxPool)
410 {
411 DEBUGLOG(5, "ZSTDMT_getCCtx");
412 ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
413 if (cctxPool->availCCtx) {
414 cctxPool->availCCtx--;
415 { ZSTD_CCtx* const cctx = cctxPool->cctx[cctxPool->availCCtx];
416 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
417 return cctx;
418 } }
419 ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
420 DEBUGLOG(5, "create one more CCtx");
421 return ZSTD_createCCtx_advanced(cctxPool->cMem); /* note : can be NULL, when creation fails ! */
422 }
423
ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool * pool,ZSTD_CCtx * cctx)424 static void ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool* pool, ZSTD_CCtx* cctx)
425 {
426 if (cctx==NULL) return; /* compatibility with release on NULL */
427 ZSTD_pthread_mutex_lock(&pool->poolMutex);
428 if (pool->availCCtx < pool->totalCCtx)
429 pool->cctx[pool->availCCtx++] = cctx;
430 else {
431 /* pool overflow : should not happen, since totalCCtx==nbWorkers */
432 DEBUGLOG(4, "CCtx pool overflow : free cctx");
433 ZSTD_freeCCtx(cctx);
434 }
435 ZSTD_pthread_mutex_unlock(&pool->poolMutex);
436 }
437
438 /* ==== Serial State ==== */
439
440 typedef struct {
441 void const* start;
442 size_t size;
443 } range_t;
444
445 typedef struct {
446 /* All variables in the struct are protected by mutex. */
447 ZSTD_pthread_mutex_t mutex;
448 ZSTD_pthread_cond_t cond;
449 ZSTD_CCtx_params params;
450 ldmState_t ldmState;
451 XXH64_state_t xxhState;
452 unsigned nextJobID;
453 /* Protects ldmWindow.
454 * Must be acquired after the main mutex when acquiring both.
455 */
456 ZSTD_pthread_mutex_t ldmWindowMutex;
457 ZSTD_pthread_cond_t ldmWindowCond; /* Signaled when ldmWindow is updated */
458 ZSTD_window_t ldmWindow; /* A thread-safe copy of ldmState.window */
459 } serialState_t;
460
461 static int
ZSTDMT_serialState_reset(serialState_t * serialState,ZSTDMT_seqPool * seqPool,ZSTD_CCtx_params params,size_t jobSize,const void * dict,size_t const dictSize,ZSTD_dictContentType_e dictContentType)462 ZSTDMT_serialState_reset(serialState_t* serialState,
463 ZSTDMT_seqPool* seqPool,
464 ZSTD_CCtx_params params,
465 size_t jobSize,
466 const void* dict, size_t const dictSize,
467 ZSTD_dictContentType_e dictContentType)
468 {
469 /* Adjust parameters */
470 if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
471 DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10);
472 ZSTD_ldm_adjustParameters(¶ms.ldmParams, ¶ms.cParams);
473 assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
474 assert(params.ldmParams.hashRateLog < 32);
475 } else {
476 ZSTD_memset(¶ms.ldmParams, 0, sizeof(params.ldmParams));
477 }
478 serialState->nextJobID = 0;
479 if (params.fParams.checksumFlag)
480 XXH64_reset(&serialState->xxhState, 0);
481 if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
482 ZSTD_customMem cMem = params.customMem;
483 unsigned const hashLog = params.ldmParams.hashLog;
484 size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
485 unsigned const bucketLog =
486 params.ldmParams.hashLog - params.ldmParams.bucketSizeLog;
487 unsigned const prevBucketLog =
488 serialState->params.ldmParams.hashLog -
489 serialState->params.ldmParams.bucketSizeLog;
490 size_t const numBuckets = (size_t)1 << bucketLog;
491 /* Size the seq pool tables */
492 ZSTDMT_setNbSeq(seqPool, ZSTD_ldm_getMaxNbSeq(params.ldmParams, jobSize));
493 /* Reset the window */
494 ZSTD_window_init(&serialState->ldmState.window);
495 /* Resize tables and output space if necessary. */
496 if (serialState->ldmState.hashTable == NULL || serialState->params.ldmParams.hashLog < hashLog) {
497 ZSTD_customFree(serialState->ldmState.hashTable, cMem);
498 serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_customMalloc(hashSize, cMem);
499 }
500 if (serialState->ldmState.bucketOffsets == NULL || prevBucketLog < bucketLog) {
501 ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
502 serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_customMalloc(numBuckets, cMem);
503 }
504 if (!serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets)
505 return 1;
506 /* Zero the tables */
507 ZSTD_memset(serialState->ldmState.hashTable, 0, hashSize);
508 ZSTD_memset(serialState->ldmState.bucketOffsets, 0, numBuckets);
509
510 /* Update window state and fill hash table with dict */
511 serialState->ldmState.loadedDictEnd = 0;
512 if (dictSize > 0) {
513 if (dictContentType == ZSTD_dct_rawContent) {
514 BYTE const* const dictEnd = (const BYTE*)dict + dictSize;
515 ZSTD_window_update(&serialState->ldmState.window, dict, dictSize, /* forceNonContiguous */ 0);
516 ZSTD_ldm_fillHashTable(&serialState->ldmState, (const BYTE*)dict, dictEnd, ¶ms.ldmParams);
517 serialState->ldmState.loadedDictEnd = params.forceWindow ? 0 : (U32)(dictEnd - serialState->ldmState.window.base);
518 } else {
519 /* don't even load anything */
520 }
521 }
522
523 /* Initialize serialState's copy of ldmWindow. */
524 serialState->ldmWindow = serialState->ldmState.window;
525 }
526
527 serialState->params = params;
528 serialState->params.jobSize = (U32)jobSize;
529 return 0;
530 }
531
ZSTDMT_serialState_init(serialState_t * serialState)532 static int ZSTDMT_serialState_init(serialState_t* serialState)
533 {
534 int initError = 0;
535 ZSTD_memset(serialState, 0, sizeof(*serialState));
536 initError |= ZSTD_pthread_mutex_init(&serialState->mutex, NULL);
537 initError |= ZSTD_pthread_cond_init(&serialState->cond, NULL);
538 initError |= ZSTD_pthread_mutex_init(&serialState->ldmWindowMutex, NULL);
539 initError |= ZSTD_pthread_cond_init(&serialState->ldmWindowCond, NULL);
540 return initError;
541 }
542
ZSTDMT_serialState_free(serialState_t * serialState)543 static void ZSTDMT_serialState_free(serialState_t* serialState)
544 {
545 ZSTD_customMem cMem = serialState->params.customMem;
546 ZSTD_pthread_mutex_destroy(&serialState->mutex);
547 ZSTD_pthread_cond_destroy(&serialState->cond);
548 ZSTD_pthread_mutex_destroy(&serialState->ldmWindowMutex);
549 ZSTD_pthread_cond_destroy(&serialState->ldmWindowCond);
550 ZSTD_customFree(serialState->ldmState.hashTable, cMem);
551 ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
552 }
553
ZSTDMT_serialState_update(serialState_t * serialState,ZSTD_CCtx * jobCCtx,rawSeqStore_t seqStore,range_t src,unsigned jobID)554 static void ZSTDMT_serialState_update(serialState_t* serialState,
555 ZSTD_CCtx* jobCCtx, rawSeqStore_t seqStore,
556 range_t src, unsigned jobID)
557 {
558 /* Wait for our turn */
559 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
560 while (serialState->nextJobID < jobID) {
561 DEBUGLOG(5, "wait for serialState->cond");
562 ZSTD_pthread_cond_wait(&serialState->cond, &serialState->mutex);
563 }
564 /* A future job may error and skip our job */
565 if (serialState->nextJobID == jobID) {
566 /* It is now our turn, do any processing necessary */
567 if (serialState->params.ldmParams.enableLdm == ZSTD_ps_enable) {
568 size_t error;
569 assert(seqStore.seq != NULL && seqStore.pos == 0 &&
570 seqStore.size == 0 && seqStore.capacity > 0);
571 assert(src.size <= serialState->params.jobSize);
572 ZSTD_window_update(&serialState->ldmState.window, src.start, src.size, /* forceNonContiguous */ 0);
573 error = ZSTD_ldm_generateSequences(
574 &serialState->ldmState, &seqStore,
575 &serialState->params.ldmParams, src.start, src.size);
576 /* We provide a large enough buffer to never fail. */
577 assert(!ZSTD_isError(error)); (void)error;
578 /* Update ldmWindow to match the ldmState.window and signal the main
579 * thread if it is waiting for a buffer.
580 */
581 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
582 serialState->ldmWindow = serialState->ldmState.window;
583 ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
584 ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
585 }
586 if (serialState->params.fParams.checksumFlag && src.size > 0)
587 XXH64_update(&serialState->xxhState, src.start, src.size);
588 }
589 /* Now it is the next jobs turn */
590 serialState->nextJobID++;
591 ZSTD_pthread_cond_broadcast(&serialState->cond);
592 ZSTD_pthread_mutex_unlock(&serialState->mutex);
593
594 if (seqStore.size > 0) {
595 size_t const err = ZSTD_referenceExternalSequences(
596 jobCCtx, seqStore.seq, seqStore.size);
597 assert(serialState->params.ldmParams.enableLdm == ZSTD_ps_enable);
598 assert(!ZSTD_isError(err));
599 (void)err;
600 }
601 }
602
ZSTDMT_serialState_ensureFinished(serialState_t * serialState,unsigned jobID,size_t cSize)603 static void ZSTDMT_serialState_ensureFinished(serialState_t* serialState,
604 unsigned jobID, size_t cSize)
605 {
606 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
607 if (serialState->nextJobID <= jobID) {
608 assert(ZSTD_isError(cSize)); (void)cSize;
609 DEBUGLOG(5, "Skipping past job %u because of error", jobID);
610 serialState->nextJobID = jobID + 1;
611 ZSTD_pthread_cond_broadcast(&serialState->cond);
612
613 ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
614 ZSTD_window_clear(&serialState->ldmWindow);
615 ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
616 ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
617 }
618 ZSTD_pthread_mutex_unlock(&serialState->mutex);
619
620 }
621
622
623 /* ------------------------------------------ */
624 /* ===== Worker thread ===== */
625 /* ------------------------------------------ */
626
627 static const range_t kNullRange = { NULL, 0 };
628
629 typedef struct {
630 size_t consumed; /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx */
631 size_t cSize; /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx, then set0 by mtctx */
632 ZSTD_pthread_mutex_t job_mutex; /* Thread-safe - used by mtctx and worker */
633 ZSTD_pthread_cond_t job_cond; /* Thread-safe - used by mtctx and worker */
634 ZSTDMT_CCtxPool* cctxPool; /* Thread-safe - used by mtctx and (all) workers */
635 ZSTDMT_bufferPool* bufPool; /* Thread-safe - used by mtctx and (all) workers */
636 ZSTDMT_seqPool* seqPool; /* Thread-safe - used by mtctx and (all) workers */
637 serialState_t* serial; /* Thread-safe - used by mtctx and (all) workers */
638 buffer_t dstBuff; /* set by worker (or mtctx), then read by worker & mtctx, then modified by mtctx => no barrier */
639 range_t prefix; /* set by mtctx, then read by worker & mtctx => no barrier */
640 range_t src; /* set by mtctx, then read by worker & mtctx => no barrier */
641 unsigned jobID; /* set by mtctx, then read by worker => no barrier */
642 unsigned firstJob; /* set by mtctx, then read by worker => no barrier */
643 unsigned lastJob; /* set by mtctx, then read by worker => no barrier */
644 ZSTD_CCtx_params params; /* set by mtctx, then read by worker => no barrier */
645 const ZSTD_CDict* cdict; /* set by mtctx, then read by worker => no barrier */
646 unsigned long long fullFrameSize; /* set by mtctx, then read by worker => no barrier */
647 size_t dstFlushed; /* used only by mtctx */
648 unsigned frameChecksumNeeded; /* used only by mtctx */
649 } ZSTDMT_jobDescription;
650
651 #define JOB_ERROR(e) { \
652 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex); \
653 job->cSize = e; \
654 ZSTD_pthread_mutex_unlock(&job->job_mutex); \
655 goto _endJob; \
656 }
657
658 /* ZSTDMT_compressionJob() is a POOL_function type */
ZSTDMT_compressionJob(void * jobDescription)659 static void ZSTDMT_compressionJob(void* jobDescription)
660 {
661 ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)jobDescription;
662 ZSTD_CCtx_params jobParams = job->params; /* do not modify job->params ! copy it, modify the copy */
663 ZSTD_CCtx* const cctx = ZSTDMT_getCCtx(job->cctxPool);
664 rawSeqStore_t rawSeqStore = ZSTDMT_getSeq(job->seqPool);
665 buffer_t dstBuff = job->dstBuff;
666 size_t lastCBlockSize = 0;
667
668 /* resources */
669 if (cctx==NULL) JOB_ERROR(ERROR(memory_allocation));
670 if (dstBuff.start == NULL) { /* streaming job : doesn't provide a dstBuffer */
671 dstBuff = ZSTDMT_getBuffer(job->bufPool);
672 if (dstBuff.start==NULL) JOB_ERROR(ERROR(memory_allocation));
673 job->dstBuff = dstBuff; /* this value can be read in ZSTDMT_flush, when it copies the whole job */
674 }
675 if (jobParams.ldmParams.enableLdm == ZSTD_ps_enable && rawSeqStore.seq == NULL)
676 JOB_ERROR(ERROR(memory_allocation));
677
678 /* Don't compute the checksum for chunks, since we compute it externally,
679 * but write it in the header.
680 */
681 if (job->jobID != 0) jobParams.fParams.checksumFlag = 0;
682 /* Don't run LDM for the chunks, since we handle it externally */
683 jobParams.ldmParams.enableLdm = ZSTD_ps_disable;
684 /* Correct nbWorkers to 0. */
685 jobParams.nbWorkers = 0;
686
687
688 /* init */
689 if (job->cdict) {
690 size_t const initError = ZSTD_compressBegin_advanced_internal(cctx, NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast, job->cdict, &jobParams, job->fullFrameSize);
691 assert(job->firstJob); /* only allowed for first job */
692 if (ZSTD_isError(initError)) JOB_ERROR(initError);
693 } else { /* srcStart points at reloaded section */
694 U64 const pledgedSrcSize = job->firstJob ? job->fullFrameSize : job->src.size;
695 { size_t const forceWindowError = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_forceMaxWindow, !job->firstJob);
696 if (ZSTD_isError(forceWindowError)) JOB_ERROR(forceWindowError);
697 }
698 if (!job->firstJob) {
699 size_t const err = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_deterministicRefPrefix, 0);
700 if (ZSTD_isError(err)) JOB_ERROR(err);
701 }
702 { size_t const initError = ZSTD_compressBegin_advanced_internal(cctx,
703 job->prefix.start, job->prefix.size, ZSTD_dct_rawContent, /* load dictionary in "content-only" mode (no header analysis) */
704 ZSTD_dtlm_fast,
705 NULL, /*cdict*/
706 &jobParams, pledgedSrcSize);
707 if (ZSTD_isError(initError)) JOB_ERROR(initError);
708 } }
709
710 /* Perform serial step as early as possible, but after CCtx initialization */
711 ZSTDMT_serialState_update(job->serial, cctx, rawSeqStore, job->src, job->jobID);
712
713 if (!job->firstJob) { /* flush and overwrite frame header when it's not first job */
714 size_t const hSize = ZSTD_compressContinue(cctx, dstBuff.start, dstBuff.capacity, job->src.start, 0);
715 if (ZSTD_isError(hSize)) JOB_ERROR(hSize);
716 DEBUGLOG(5, "ZSTDMT_compressionJob: flush and overwrite %u bytes of frame header (not first job)", (U32)hSize);
717 ZSTD_invalidateRepCodes(cctx);
718 }
719
720 /* compress */
721 { size_t const chunkSize = 4*ZSTD_BLOCKSIZE_MAX;
722 int const nbChunks = (int)((job->src.size + (chunkSize-1)) / chunkSize);
723 const BYTE* ip = (const BYTE*) job->src.start;
724 BYTE* const ostart = (BYTE*)dstBuff.start;
725 BYTE* op = ostart;
726 BYTE* oend = op + dstBuff.capacity;
727 int chunkNb;
728 if (sizeof(size_t) > sizeof(int)) assert(job->src.size < ((size_t)INT_MAX) * chunkSize); /* check overflow */
729 DEBUGLOG(5, "ZSTDMT_compressionJob: compress %u bytes in %i blocks", (U32)job->src.size, nbChunks);
730 assert(job->cSize == 0);
731 for (chunkNb = 1; chunkNb < nbChunks; chunkNb++) {
732 size_t const cSize = ZSTD_compressContinue(cctx, op, oend-op, ip, chunkSize);
733 if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
734 ip += chunkSize;
735 op += cSize; assert(op < oend);
736 /* stats */
737 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
738 job->cSize += cSize;
739 job->consumed = chunkSize * chunkNb;
740 DEBUGLOG(5, "ZSTDMT_compressionJob: compress new block : cSize==%u bytes (total: %u)",
741 (U32)cSize, (U32)job->cSize);
742 ZSTD_pthread_cond_signal(&job->job_cond); /* warns some more data is ready to be flushed */
743 ZSTD_pthread_mutex_unlock(&job->job_mutex);
744 }
745 /* last block */
746 assert(chunkSize > 0);
747 assert((chunkSize & (chunkSize - 1)) == 0); /* chunkSize must be power of 2 for mask==(chunkSize-1) to work */
748 if ((nbChunks > 0) | job->lastJob /*must output a "last block" flag*/ ) {
749 size_t const lastBlockSize1 = job->src.size & (chunkSize-1);
750 size_t const lastBlockSize = ((lastBlockSize1==0) & (job->src.size>=chunkSize)) ? chunkSize : lastBlockSize1;
751 size_t const cSize = (job->lastJob) ?
752 ZSTD_compressEnd (cctx, op, oend-op, ip, lastBlockSize) :
753 ZSTD_compressContinue(cctx, op, oend-op, ip, lastBlockSize);
754 if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
755 lastCBlockSize = cSize;
756 } }
757 if (!job->firstJob) {
758 /* Double check that we don't have an ext-dict, because then our
759 * repcode invalidation doesn't work.
760 */
761 assert(!ZSTD_window_hasExtDict(cctx->blockState.matchState.window));
762 }
763 ZSTD_CCtx_trace(cctx, 0);
764
765 _endJob:
766 ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize);
767 if (job->prefix.size > 0)
768 DEBUGLOG(5, "Finished with prefix: %zx", (size_t)job->prefix.start);
769 DEBUGLOG(5, "Finished with source: %zx", (size_t)job->src.start);
770 /* release resources */
771 ZSTDMT_releaseSeq(job->seqPool, rawSeqStore);
772 ZSTDMT_releaseCCtx(job->cctxPool, cctx);
773 /* report */
774 ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
775 if (ZSTD_isError(job->cSize)) assert(lastCBlockSize == 0);
776 job->cSize += lastCBlockSize;
777 job->consumed = job->src.size; /* when job->consumed == job->src.size , compression job is presumed completed */
778 ZSTD_pthread_cond_signal(&job->job_cond);
779 ZSTD_pthread_mutex_unlock(&job->job_mutex);
780 }
781
782
783 /* ------------------------------------------ */
784 /* ===== Multi-threaded compression ===== */
785 /* ------------------------------------------ */
786
787 typedef struct {
788 range_t prefix; /* read-only non-owned prefix buffer */
789 buffer_t buffer;
790 size_t filled;
791 } inBuff_t;
792
793 typedef struct {
794 BYTE* buffer; /* The round input buffer. All jobs get references
795 * to pieces of the buffer. ZSTDMT_tryGetInputRange()
796 * handles handing out job input buffers, and makes
797 * sure it doesn't overlap with any pieces still in use.
798 */
799 size_t capacity; /* The capacity of buffer. */
800 size_t pos; /* The position of the current inBuff in the round
801 * buffer. Updated past the end if the inBuff once
802 * the inBuff is sent to the worker thread.
803 * pos <= capacity.
804 */
805 } roundBuff_t;
806
807 static const roundBuff_t kNullRoundBuff = {NULL, 0, 0};
808
809 #define RSYNC_LENGTH 32
810 /* Don't create chunks smaller than the zstd block size.
811 * This stops us from regressing compression ratio too much,
812 * and ensures our output fits in ZSTD_compressBound().
813 *
814 * If this is shrunk < ZSTD_BLOCKSIZELOG_MIN then
815 * ZSTD_COMPRESSBOUND() will need to be updated.
816 */
817 #define RSYNC_MIN_BLOCK_LOG ZSTD_BLOCKSIZELOG_MAX
818 #define RSYNC_MIN_BLOCK_SIZE (1<<RSYNC_MIN_BLOCK_LOG)
819
820 typedef struct {
821 U64 hash;
822 U64 hitMask;
823 U64 primePower;
824 } rsyncState_t;
825
826 struct ZSTDMT_CCtx_s {
827 POOL_ctx* factory;
828 ZSTDMT_jobDescription* jobs;
829 ZSTDMT_bufferPool* bufPool;
830 ZSTDMT_CCtxPool* cctxPool;
831 ZSTDMT_seqPool* seqPool;
832 ZSTD_CCtx_params params;
833 size_t targetSectionSize;
834 size_t targetPrefixSize;
835 int jobReady; /* 1 => one job is already prepared, but pool has shortage of workers. Don't create a new job. */
836 inBuff_t inBuff;
837 roundBuff_t roundBuff;
838 serialState_t serial;
839 rsyncState_t rsync;
840 unsigned jobIDMask;
841 unsigned doneJobID;
842 unsigned nextJobID;
843 unsigned frameEnded;
844 unsigned allJobsCompleted;
845 unsigned long long frameContentSize;
846 unsigned long long consumed;
847 unsigned long long produced;
848 ZSTD_customMem cMem;
849 ZSTD_CDict* cdictLocal;
850 const ZSTD_CDict* cdict;
851 unsigned providedFactory: 1;
852 };
853
ZSTDMT_freeJobsTable(ZSTDMT_jobDescription * jobTable,U32 nbJobs,ZSTD_customMem cMem)854 static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZSTD_customMem cMem)
855 {
856 U32 jobNb;
857 if (jobTable == NULL) return;
858 for (jobNb=0; jobNb<nbJobs; jobNb++) {
859 ZSTD_pthread_mutex_destroy(&jobTable[jobNb].job_mutex);
860 ZSTD_pthread_cond_destroy(&jobTable[jobNb].job_cond);
861 }
862 ZSTD_customFree(jobTable, cMem);
863 }
864
865 /* ZSTDMT_allocJobsTable()
866 * allocate and init a job table.
867 * update *nbJobsPtr to next power of 2 value, as size of table */
ZSTDMT_createJobsTable(U32 * nbJobsPtr,ZSTD_customMem cMem)868 static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
869 {
870 U32 const nbJobsLog2 = ZSTD_highbit32(*nbJobsPtr) + 1;
871 U32 const nbJobs = 1 << nbJobsLog2;
872 U32 jobNb;
873 ZSTDMT_jobDescription* const jobTable = (ZSTDMT_jobDescription*)
874 ZSTD_customCalloc(nbJobs * sizeof(ZSTDMT_jobDescription), cMem);
875 int initError = 0;
876 if (jobTable==NULL) return NULL;
877 *nbJobsPtr = nbJobs;
878 for (jobNb=0; jobNb<nbJobs; jobNb++) {
879 initError |= ZSTD_pthread_mutex_init(&jobTable[jobNb].job_mutex, NULL);
880 initError |= ZSTD_pthread_cond_init(&jobTable[jobNb].job_cond, NULL);
881 }
882 if (initError != 0) {
883 ZSTDMT_freeJobsTable(jobTable, nbJobs, cMem);
884 return NULL;
885 }
886 return jobTable;
887 }
888
ZSTDMT_expandJobsTable(ZSTDMT_CCtx * mtctx,U32 nbWorkers)889 static size_t ZSTDMT_expandJobsTable (ZSTDMT_CCtx* mtctx, U32 nbWorkers) {
890 U32 nbJobs = nbWorkers + 2;
891 if (nbJobs > mtctx->jobIDMask+1) { /* need more job capacity */
892 ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
893 mtctx->jobIDMask = 0;
894 mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, mtctx->cMem);
895 if (mtctx->jobs==NULL) return ERROR(memory_allocation);
896 assert((nbJobs != 0) && ((nbJobs & (nbJobs - 1)) == 0)); /* ensure nbJobs is a power of 2 */
897 mtctx->jobIDMask = nbJobs - 1;
898 }
899 return 0;
900 }
901
902
903 /* ZSTDMT_CCtxParam_setNbWorkers():
904 * Internal use only */
ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params * params,unsigned nbWorkers)905 static size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers)
906 {
907 return ZSTD_CCtxParams_setParameter(params, ZSTD_c_nbWorkers, (int)nbWorkers);
908 }
909
ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers,ZSTD_customMem cMem,ZSTD_threadPool * pool)910 MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool)
911 {
912 ZSTDMT_CCtx* mtctx;
913 U32 nbJobs = nbWorkers + 2;
914 int initError;
915 DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbWorkers = %u)", nbWorkers);
916
917 if (nbWorkers < 1) return NULL;
918 nbWorkers = MIN(nbWorkers , ZSTDMT_NBWORKERS_MAX);
919 if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
920 /* invalid custom allocator */
921 return NULL;
922
923 mtctx = (ZSTDMT_CCtx*) ZSTD_customCalloc(sizeof(ZSTDMT_CCtx), cMem);
924 if (!mtctx) return NULL;
925 ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
926 mtctx->cMem = cMem;
927 mtctx->allJobsCompleted = 1;
928 if (pool != NULL) {
929 mtctx->factory = pool;
930 mtctx->providedFactory = 1;
931 }
932 else {
933 mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem);
934 mtctx->providedFactory = 0;
935 }
936 mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem);
937 assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0); /* ensure nbJobs is a power of 2 */
938 mtctx->jobIDMask = nbJobs - 1;
939 mtctx->bufPool = ZSTDMT_createBufferPool(nbWorkers, cMem);
940 mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem);
941 mtctx->seqPool = ZSTDMT_createSeqPool(nbWorkers, cMem);
942 initError = ZSTDMT_serialState_init(&mtctx->serial);
943 mtctx->roundBuff = kNullRoundBuff;
944 if (!mtctx->factory | !mtctx->jobs | !mtctx->bufPool | !mtctx->cctxPool | !mtctx->seqPool | initError) {
945 ZSTDMT_freeCCtx(mtctx);
946 return NULL;
947 }
948 DEBUGLOG(3, "mt_cctx created, for %u threads", nbWorkers);
949 return mtctx;
950 }
951
ZSTDMT_createCCtx_advanced(unsigned nbWorkers,ZSTD_customMem cMem,ZSTD_threadPool * pool)952 ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool)
953 {
954 #ifdef ZSTD_MULTITHREAD
955 return ZSTDMT_createCCtx_advanced_internal(nbWorkers, cMem, pool);
956 #else
957 (void)nbWorkers;
958 (void)cMem;
959 (void)pool;
960 return NULL;
961 #endif
962 }
963
964
965 /* ZSTDMT_releaseAllJobResources() :
966 * note : ensure all workers are killed first ! */
ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx * mtctx)967 static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx)
968 {
969 unsigned jobID;
970 DEBUGLOG(3, "ZSTDMT_releaseAllJobResources");
971 for (jobID=0; jobID <= mtctx->jobIDMask; jobID++) {
972 /* Copy the mutex/cond out */
973 ZSTD_pthread_mutex_t const mutex = mtctx->jobs[jobID].job_mutex;
974 ZSTD_pthread_cond_t const cond = mtctx->jobs[jobID].job_cond;
975
976 DEBUGLOG(4, "job%02u: release dst address %08X", jobID, (U32)(size_t)mtctx->jobs[jobID].dstBuff.start);
977 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);
978
979 /* Clear the job description, but keep the mutex/cond */
980 ZSTD_memset(&mtctx->jobs[jobID], 0, sizeof(mtctx->jobs[jobID]));
981 mtctx->jobs[jobID].job_mutex = mutex;
982 mtctx->jobs[jobID].job_cond = cond;
983 }
984 mtctx->inBuff.buffer = g_nullBuffer;
985 mtctx->inBuff.filled = 0;
986 mtctx->allJobsCompleted = 1;
987 }
988
ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx * mtctx)989 static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* mtctx)
990 {
991 DEBUGLOG(4, "ZSTDMT_waitForAllJobsCompleted");
992 while (mtctx->doneJobID < mtctx->nextJobID) {
993 unsigned const jobID = mtctx->doneJobID & mtctx->jobIDMask;
994 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
995 while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
996 DEBUGLOG(4, "waiting for jobCompleted signal from job %u", mtctx->doneJobID); /* we want to block when waiting for data to flush */
997 ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
998 }
999 ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
1000 mtctx->doneJobID++;
1001 }
1002 }
1003
ZSTDMT_freeCCtx(ZSTDMT_CCtx * mtctx)1004 size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx)
1005 {
1006 if (mtctx==NULL) return 0; /* compatible with free on NULL */
1007 if (!mtctx->providedFactory)
1008 POOL_free(mtctx->factory); /* stop and free worker threads */
1009 ZSTDMT_releaseAllJobResources(mtctx); /* release job resources into pools first */
1010 ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
1011 ZSTDMT_freeBufferPool(mtctx->bufPool);
1012 ZSTDMT_freeCCtxPool(mtctx->cctxPool);
1013 ZSTDMT_freeSeqPool(mtctx->seqPool);
1014 ZSTDMT_serialState_free(&mtctx->serial);
1015 ZSTD_freeCDict(mtctx->cdictLocal);
1016 if (mtctx->roundBuff.buffer)
1017 ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem);
1018 ZSTD_customFree(mtctx, mtctx->cMem);
1019 return 0;
1020 }
1021
ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx * mtctx)1022 size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx)
1023 {
1024 if (mtctx == NULL) return 0; /* supports sizeof NULL */
1025 return sizeof(*mtctx)
1026 + POOL_sizeof(mtctx->factory)
1027 + ZSTDMT_sizeof_bufferPool(mtctx->bufPool)
1028 + (mtctx->jobIDMask+1) * sizeof(ZSTDMT_jobDescription)
1029 + ZSTDMT_sizeof_CCtxPool(mtctx->cctxPool)
1030 + ZSTDMT_sizeof_seqPool(mtctx->seqPool)
1031 + ZSTD_sizeof_CDict(mtctx->cdictLocal)
1032 + mtctx->roundBuff.capacity;
1033 }
1034
1035
1036 /* ZSTDMT_resize() :
1037 * @return : error code if fails, 0 on success */
ZSTDMT_resize(ZSTDMT_CCtx * mtctx,unsigned nbWorkers)1038 static size_t ZSTDMT_resize(ZSTDMT_CCtx* mtctx, unsigned nbWorkers)
1039 {
1040 if (POOL_resize(mtctx->factory, nbWorkers)) return ERROR(memory_allocation);
1041 FORWARD_IF_ERROR( ZSTDMT_expandJobsTable(mtctx, nbWorkers) , "");
1042 mtctx->bufPool = ZSTDMT_expandBufferPool(mtctx->bufPool, nbWorkers);
1043 if (mtctx->bufPool == NULL) return ERROR(memory_allocation);
1044 mtctx->cctxPool = ZSTDMT_expandCCtxPool(mtctx->cctxPool, nbWorkers);
1045 if (mtctx->cctxPool == NULL) return ERROR(memory_allocation);
1046 mtctx->seqPool = ZSTDMT_expandSeqPool(mtctx->seqPool, nbWorkers);
1047 if (mtctx->seqPool == NULL) return ERROR(memory_allocation);
1048 ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
1049 return 0;
1050 }
1051
1052
1053 /*! ZSTDMT_updateCParams_whileCompressing() :
1054 * Updates a selected set of compression parameters, remaining compatible with currently active frame.
1055 * New parameters will be applied to next compression job. */
ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx * mtctx,const ZSTD_CCtx_params * cctxParams)1056 void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams)
1057 {
1058 U32 const saved_wlog = mtctx->params.cParams.windowLog; /* Do not modify windowLog while compressing */
1059 int const compressionLevel = cctxParams->compressionLevel;
1060 DEBUGLOG(5, "ZSTDMT_updateCParams_whileCompressing (level:%i)",
1061 compressionLevel);
1062 mtctx->params.compressionLevel = compressionLevel;
1063 { ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
1064 cParams.windowLog = saved_wlog;
1065 mtctx->params.cParams = cParams;
1066 }
1067 }
1068
1069 /* ZSTDMT_getFrameProgression():
1070 * tells how much data has been consumed (input) and produced (output) for current frame.
1071 * able to count progression inside worker threads.
1072 * Note : mutex will be acquired during statistics collection inside workers. */
ZSTDMT_getFrameProgression(ZSTDMT_CCtx * mtctx)1073 ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx)
1074 {
1075 ZSTD_frameProgression fps;
1076 DEBUGLOG(5, "ZSTDMT_getFrameProgression");
1077 fps.ingested = mtctx->consumed + mtctx->inBuff.filled;
1078 fps.consumed = mtctx->consumed;
1079 fps.produced = fps.flushed = mtctx->produced;
1080 fps.currentJobID = mtctx->nextJobID;
1081 fps.nbActiveWorkers = 0;
1082 { unsigned jobNb;
1083 unsigned lastJobNb = mtctx->nextJobID + mtctx->jobReady; assert(mtctx->jobReady <= 1);
1084 DEBUGLOG(6, "ZSTDMT_getFrameProgression: jobs: from %u to <%u (jobReady:%u)",
1085 mtctx->doneJobID, lastJobNb, mtctx->jobReady)
1086 for (jobNb = mtctx->doneJobID ; jobNb < lastJobNb ; jobNb++) {
1087 unsigned const wJobID = jobNb & mtctx->jobIDMask;
1088 ZSTDMT_jobDescription* jobPtr = &mtctx->jobs[wJobID];
1089 ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
1090 { size_t const cResult = jobPtr->cSize;
1091 size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
1092 size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
1093 assert(flushed <= produced);
1094 fps.ingested += jobPtr->src.size;
1095 fps.consumed += jobPtr->consumed;
1096 fps.produced += produced;
1097 fps.flushed += flushed;
1098 fps.nbActiveWorkers += (jobPtr->consumed < jobPtr->src.size);
1099 }
1100 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1101 }
1102 }
1103 return fps;
1104 }
1105
1106
ZSTDMT_toFlushNow(ZSTDMT_CCtx * mtctx)1107 size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx)
1108 {
1109 size_t toFlush;
1110 unsigned const jobID = mtctx->doneJobID;
1111 assert(jobID <= mtctx->nextJobID);
1112 if (jobID == mtctx->nextJobID) return 0; /* no active job => nothing to flush */
1113
1114 /* look into oldest non-fully-flushed job */
1115 { unsigned const wJobID = jobID & mtctx->jobIDMask;
1116 ZSTDMT_jobDescription* const jobPtr = &mtctx->jobs[wJobID];
1117 ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
1118 { size_t const cResult = jobPtr->cSize;
1119 size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
1120 size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
1121 assert(flushed <= produced);
1122 assert(jobPtr->consumed <= jobPtr->src.size);
1123 toFlush = produced - flushed;
1124 /* if toFlush==0, nothing is available to flush.
1125 * However, jobID is expected to still be active:
1126 * if jobID was already completed and fully flushed,
1127 * ZSTDMT_flushProduced() should have already moved onto next job.
1128 * Therefore, some input has not yet been consumed. */
1129 if (toFlush==0) {
1130 assert(jobPtr->consumed < jobPtr->src.size);
1131 }
1132 }
1133 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1134 }
1135
1136 return toFlush;
1137 }
1138
1139
1140 /* ------------------------------------------ */
1141 /* ===== Multi-threaded compression ===== */
1142 /* ------------------------------------------ */
1143
ZSTDMT_computeTargetJobLog(const ZSTD_CCtx_params * params)1144 static unsigned ZSTDMT_computeTargetJobLog(const ZSTD_CCtx_params* params)
1145 {
1146 unsigned jobLog;
1147 if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
1148 /* In Long Range Mode, the windowLog is typically oversized.
1149 * In which case, it's preferable to determine the jobSize
1150 * based on cycleLog instead. */
1151 jobLog = MAX(21, ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy) + 3);
1152 } else {
1153 jobLog = MAX(20, params->cParams.windowLog + 2);
1154 }
1155 return MIN(jobLog, (unsigned)ZSTDMT_JOBLOG_MAX);
1156 }
1157
ZSTDMT_overlapLog_default(ZSTD_strategy strat)1158 static int ZSTDMT_overlapLog_default(ZSTD_strategy strat)
1159 {
1160 switch(strat)
1161 {
1162 case ZSTD_btultra2:
1163 return 9;
1164 case ZSTD_btultra:
1165 case ZSTD_btopt:
1166 return 8;
1167 case ZSTD_btlazy2:
1168 case ZSTD_lazy2:
1169 return 7;
1170 case ZSTD_lazy:
1171 case ZSTD_greedy:
1172 case ZSTD_dfast:
1173 case ZSTD_fast:
1174 default:;
1175 }
1176 return 6;
1177 }
1178
ZSTDMT_overlapLog(int ovlog,ZSTD_strategy strat)1179 static int ZSTDMT_overlapLog(int ovlog, ZSTD_strategy strat)
1180 {
1181 assert(0 <= ovlog && ovlog <= 9);
1182 if (ovlog == 0) return ZSTDMT_overlapLog_default(strat);
1183 return ovlog;
1184 }
1185
ZSTDMT_computeOverlapSize(const ZSTD_CCtx_params * params)1186 static size_t ZSTDMT_computeOverlapSize(const ZSTD_CCtx_params* params)
1187 {
1188 int const overlapRLog = 9 - ZSTDMT_overlapLog(params->overlapLog, params->cParams.strategy);
1189 int ovLog = (overlapRLog >= 8) ? 0 : (params->cParams.windowLog - overlapRLog);
1190 assert(0 <= overlapRLog && overlapRLog <= 8);
1191 if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
1192 /* In Long Range Mode, the windowLog is typically oversized.
1193 * In which case, it's preferable to determine the jobSize
1194 * based on chainLog instead.
1195 * Then, ovLog becomes a fraction of the jobSize, rather than windowSize */
1196 ovLog = MIN(params->cParams.windowLog, ZSTDMT_computeTargetJobLog(params) - 2)
1197 - overlapRLog;
1198 }
1199 assert(0 <= ovLog && ovLog <= ZSTD_WINDOWLOG_MAX);
1200 DEBUGLOG(4, "overlapLog : %i", params->overlapLog);
1201 DEBUGLOG(4, "overlap size : %i", 1 << ovLog);
1202 return (ovLog==0) ? 0 : (size_t)1 << ovLog;
1203 }
1204
1205 /* ====================================== */
1206 /* ======= Streaming API ======= */
1207 /* ====================================== */
1208
ZSTDMT_initCStream_internal(ZSTDMT_CCtx * mtctx,const void * dict,size_t dictSize,ZSTD_dictContentType_e dictContentType,const ZSTD_CDict * cdict,ZSTD_CCtx_params params,unsigned long long pledgedSrcSize)1209 size_t ZSTDMT_initCStream_internal(
1210 ZSTDMT_CCtx* mtctx,
1211 const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
1212 const ZSTD_CDict* cdict, ZSTD_CCtx_params params,
1213 unsigned long long pledgedSrcSize)
1214 {
1215 DEBUGLOG(4, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u, nbWorkers=%u, cctxPool=%u)",
1216 (U32)pledgedSrcSize, params.nbWorkers, mtctx->cctxPool->totalCCtx);
1217
1218 /* params supposed partially fully validated at this point */
1219 assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
1220 assert(!((dict) && (cdict))); /* either dict or cdict, not both */
1221
1222 /* init */
1223 if (params.nbWorkers != mtctx->params.nbWorkers)
1224 FORWARD_IF_ERROR( ZSTDMT_resize(mtctx, params.nbWorkers) , "");
1225
1226 if (params.jobSize != 0 && params.jobSize < ZSTDMT_JOBSIZE_MIN) params.jobSize = ZSTDMT_JOBSIZE_MIN;
1227 if (params.jobSize > (size_t)ZSTDMT_JOBSIZE_MAX) params.jobSize = (size_t)ZSTDMT_JOBSIZE_MAX;
1228
1229 DEBUGLOG(4, "ZSTDMT_initCStream_internal: %u workers", params.nbWorkers);
1230
1231 if (mtctx->allJobsCompleted == 0) { /* previous compression not correctly finished */
1232 ZSTDMT_waitForAllJobsCompleted(mtctx);
1233 ZSTDMT_releaseAllJobResources(mtctx);
1234 mtctx->allJobsCompleted = 1;
1235 }
1236
1237 mtctx->params = params;
1238 mtctx->frameContentSize = pledgedSrcSize;
1239 if (dict) {
1240 ZSTD_freeCDict(mtctx->cdictLocal);
1241 mtctx->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
1242 ZSTD_dlm_byCopy, dictContentType, /* note : a loadPrefix becomes an internal CDict */
1243 params.cParams, mtctx->cMem);
1244 mtctx->cdict = mtctx->cdictLocal;
1245 if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
1246 } else {
1247 ZSTD_freeCDict(mtctx->cdictLocal);
1248 mtctx->cdictLocal = NULL;
1249 mtctx->cdict = cdict;
1250 }
1251
1252 mtctx->targetPrefixSize = ZSTDMT_computeOverlapSize(¶ms);
1253 DEBUGLOG(4, "overlapLog=%i => %u KB", params.overlapLog, (U32)(mtctx->targetPrefixSize>>10));
1254 mtctx->targetSectionSize = params.jobSize;
1255 if (mtctx->targetSectionSize == 0) {
1256 mtctx->targetSectionSize = 1ULL << ZSTDMT_computeTargetJobLog(¶ms);
1257 }
1258 assert(mtctx->targetSectionSize <= (size_t)ZSTDMT_JOBSIZE_MAX);
1259
1260 if (params.rsyncable) {
1261 /* Aim for the targetsectionSize as the average job size. */
1262 U32 const jobSizeKB = (U32)(mtctx->targetSectionSize >> 10);
1263 U32 const rsyncBits = (assert(jobSizeKB >= 1), ZSTD_highbit32(jobSizeKB) + 10);
1264 /* We refuse to create jobs < RSYNC_MIN_BLOCK_SIZE bytes, so make sure our
1265 * expected job size is at least 4x larger. */
1266 assert(rsyncBits >= RSYNC_MIN_BLOCK_LOG + 2);
1267 DEBUGLOG(4, "rsyncLog = %u", rsyncBits);
1268 mtctx->rsync.hash = 0;
1269 mtctx->rsync.hitMask = (1ULL << rsyncBits) - 1;
1270 mtctx->rsync.primePower = ZSTD_rollingHash_primePower(RSYNC_LENGTH);
1271 }
1272 if (mtctx->targetSectionSize < mtctx->targetPrefixSize) mtctx->targetSectionSize = mtctx->targetPrefixSize; /* job size must be >= overlap size */
1273 DEBUGLOG(4, "Job Size : %u KB (note : set to %u)", (U32)(mtctx->targetSectionSize>>10), (U32)params.jobSize);
1274 DEBUGLOG(4, "inBuff Size : %u KB", (U32)(mtctx->targetSectionSize>>10));
1275 ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(mtctx->targetSectionSize));
1276 {
1277 /* If ldm is enabled we need windowSize space. */
1278 size_t const windowSize = mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable ? (1U << mtctx->params.cParams.windowLog) : 0;
1279 /* Two buffers of slack, plus extra space for the overlap
1280 * This is the minimum slack that LDM works with. One extra because
1281 * flush might waste up to targetSectionSize-1 bytes. Another extra
1282 * for the overlap (if > 0), then one to fill which doesn't overlap
1283 * with the LDM window.
1284 */
1285 size_t const nbSlackBuffers = 2 + (mtctx->targetPrefixSize > 0);
1286 size_t const slackSize = mtctx->targetSectionSize * nbSlackBuffers;
1287 /* Compute the total size, and always have enough slack */
1288 size_t const nbWorkers = MAX(mtctx->params.nbWorkers, 1);
1289 size_t const sectionsSize = mtctx->targetSectionSize * nbWorkers;
1290 size_t const capacity = MAX(windowSize, sectionsSize) + slackSize;
1291 if (mtctx->roundBuff.capacity < capacity) {
1292 if (mtctx->roundBuff.buffer)
1293 ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem);
1294 mtctx->roundBuff.buffer = (BYTE*)ZSTD_customMalloc(capacity, mtctx->cMem);
1295 if (mtctx->roundBuff.buffer == NULL) {
1296 mtctx->roundBuff.capacity = 0;
1297 return ERROR(memory_allocation);
1298 }
1299 mtctx->roundBuff.capacity = capacity;
1300 }
1301 }
1302 DEBUGLOG(4, "roundBuff capacity : %u KB", (U32)(mtctx->roundBuff.capacity>>10));
1303 mtctx->roundBuff.pos = 0;
1304 mtctx->inBuff.buffer = g_nullBuffer;
1305 mtctx->inBuff.filled = 0;
1306 mtctx->inBuff.prefix = kNullRange;
1307 mtctx->doneJobID = 0;
1308 mtctx->nextJobID = 0;
1309 mtctx->frameEnded = 0;
1310 mtctx->allJobsCompleted = 0;
1311 mtctx->consumed = 0;
1312 mtctx->produced = 0;
1313 if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params, mtctx->targetSectionSize,
1314 dict, dictSize, dictContentType))
1315 return ERROR(memory_allocation);
1316 return 0;
1317 }
1318
1319
1320 /* ZSTDMT_writeLastEmptyBlock()
1321 * Write a single empty block with an end-of-frame to finish a frame.
1322 * Job must be created from streaming variant.
1323 * This function is always successful if expected conditions are fulfilled.
1324 */
ZSTDMT_writeLastEmptyBlock(ZSTDMT_jobDescription * job)1325 static void ZSTDMT_writeLastEmptyBlock(ZSTDMT_jobDescription* job)
1326 {
1327 assert(job->lastJob == 1);
1328 assert(job->src.size == 0); /* last job is empty -> will be simplified into a last empty block */
1329 assert(job->firstJob == 0); /* cannot be first job, as it also needs to create frame header */
1330 assert(job->dstBuff.start == NULL); /* invoked from streaming variant only (otherwise, dstBuff might be user's output) */
1331 job->dstBuff = ZSTDMT_getBuffer(job->bufPool);
1332 if (job->dstBuff.start == NULL) {
1333 job->cSize = ERROR(memory_allocation);
1334 return;
1335 }
1336 assert(job->dstBuff.capacity >= ZSTD_blockHeaderSize); /* no buffer should ever be that small */
1337 job->src = kNullRange;
1338 job->cSize = ZSTD_writeLastEmptyBlock(job->dstBuff.start, job->dstBuff.capacity);
1339 assert(!ZSTD_isError(job->cSize));
1340 assert(job->consumed == 0);
1341 }
1342
ZSTDMT_createCompressionJob(ZSTDMT_CCtx * mtctx,size_t srcSize,ZSTD_EndDirective endOp)1343 static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* mtctx, size_t srcSize, ZSTD_EndDirective endOp)
1344 {
1345 unsigned const jobID = mtctx->nextJobID & mtctx->jobIDMask;
1346 int const endFrame = (endOp == ZSTD_e_end);
1347
1348 if (mtctx->nextJobID > mtctx->doneJobID + mtctx->jobIDMask) {
1349 DEBUGLOG(5, "ZSTDMT_createCompressionJob: will not create new job : table is full");
1350 assert((mtctx->nextJobID & mtctx->jobIDMask) == (mtctx->doneJobID & mtctx->jobIDMask));
1351 return 0;
1352 }
1353
1354 if (!mtctx->jobReady) {
1355 BYTE const* src = (BYTE const*)mtctx->inBuff.buffer.start;
1356 DEBUGLOG(5, "ZSTDMT_createCompressionJob: preparing job %u to compress %u bytes with %u preload ",
1357 mtctx->nextJobID, (U32)srcSize, (U32)mtctx->inBuff.prefix.size);
1358 mtctx->jobs[jobID].src.start = src;
1359 mtctx->jobs[jobID].src.size = srcSize;
1360 assert(mtctx->inBuff.filled >= srcSize);
1361 mtctx->jobs[jobID].prefix = mtctx->inBuff.prefix;
1362 mtctx->jobs[jobID].consumed = 0;
1363 mtctx->jobs[jobID].cSize = 0;
1364 mtctx->jobs[jobID].params = mtctx->params;
1365 mtctx->jobs[jobID].cdict = mtctx->nextJobID==0 ? mtctx->cdict : NULL;
1366 mtctx->jobs[jobID].fullFrameSize = mtctx->frameContentSize;
1367 mtctx->jobs[jobID].dstBuff = g_nullBuffer;
1368 mtctx->jobs[jobID].cctxPool = mtctx->cctxPool;
1369 mtctx->jobs[jobID].bufPool = mtctx->bufPool;
1370 mtctx->jobs[jobID].seqPool = mtctx->seqPool;
1371 mtctx->jobs[jobID].serial = &mtctx->serial;
1372 mtctx->jobs[jobID].jobID = mtctx->nextJobID;
1373 mtctx->jobs[jobID].firstJob = (mtctx->nextJobID==0);
1374 mtctx->jobs[jobID].lastJob = endFrame;
1375 mtctx->jobs[jobID].frameChecksumNeeded = mtctx->params.fParams.checksumFlag && endFrame && (mtctx->nextJobID>0);
1376 mtctx->jobs[jobID].dstFlushed = 0;
1377
1378 /* Update the round buffer pos and clear the input buffer to be reset */
1379 mtctx->roundBuff.pos += srcSize;
1380 mtctx->inBuff.buffer = g_nullBuffer;
1381 mtctx->inBuff.filled = 0;
1382 /* Set the prefix */
1383 if (!endFrame) {
1384 size_t const newPrefixSize = MIN(srcSize, mtctx->targetPrefixSize);
1385 mtctx->inBuff.prefix.start = src + srcSize - newPrefixSize;
1386 mtctx->inBuff.prefix.size = newPrefixSize;
1387 } else { /* endFrame==1 => no need for another input buffer */
1388 mtctx->inBuff.prefix = kNullRange;
1389 mtctx->frameEnded = endFrame;
1390 if (mtctx->nextJobID == 0) {
1391 /* single job exception : checksum is already calculated directly within worker thread */
1392 mtctx->params.fParams.checksumFlag = 0;
1393 } }
1394
1395 if ( (srcSize == 0)
1396 && (mtctx->nextJobID>0)/*single job must also write frame header*/ ) {
1397 DEBUGLOG(5, "ZSTDMT_createCompressionJob: creating a last empty block to end frame");
1398 assert(endOp == ZSTD_e_end); /* only possible case : need to end the frame with an empty last block */
1399 ZSTDMT_writeLastEmptyBlock(mtctx->jobs + jobID);
1400 mtctx->nextJobID++;
1401 return 0;
1402 }
1403 }
1404
1405 DEBUGLOG(5, "ZSTDMT_createCompressionJob: posting job %u : %u bytes (end:%u, jobNb == %u (mod:%u))",
1406 mtctx->nextJobID,
1407 (U32)mtctx->jobs[jobID].src.size,
1408 mtctx->jobs[jobID].lastJob,
1409 mtctx->nextJobID,
1410 jobID);
1411 if (POOL_tryAdd(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[jobID])) {
1412 mtctx->nextJobID++;
1413 mtctx->jobReady = 0;
1414 } else {
1415 DEBUGLOG(5, "ZSTDMT_createCompressionJob: no worker available for job %u", mtctx->nextJobID);
1416 mtctx->jobReady = 1;
1417 }
1418 return 0;
1419 }
1420
1421
1422 /*! ZSTDMT_flushProduced() :
1423 * flush whatever data has been produced but not yet flushed in current job.
1424 * move to next job if current one is fully flushed.
1425 * `output` : `pos` will be updated with amount of data flushed .
1426 * `blockToFlush` : if >0, the function will block and wait if there is no data available to flush .
1427 * @return : amount of data remaining within internal buffer, 0 if no more, 1 if unknown but > 0, or an error code */
ZSTDMT_flushProduced(ZSTDMT_CCtx * mtctx,ZSTD_outBuffer * output,unsigned blockToFlush,ZSTD_EndDirective end)1428 static size_t ZSTDMT_flushProduced(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, unsigned blockToFlush, ZSTD_EndDirective end)
1429 {
1430 unsigned const wJobID = mtctx->doneJobID & mtctx->jobIDMask;
1431 DEBUGLOG(5, "ZSTDMT_flushProduced (blocking:%u , job %u <= %u)",
1432 blockToFlush, mtctx->doneJobID, mtctx->nextJobID);
1433 assert(output->size >= output->pos);
1434
1435 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
1436 if ( blockToFlush
1437 && (mtctx->doneJobID < mtctx->nextJobID) ) {
1438 assert(mtctx->jobs[wJobID].dstFlushed <= mtctx->jobs[wJobID].cSize);
1439 while (mtctx->jobs[wJobID].dstFlushed == mtctx->jobs[wJobID].cSize) { /* nothing to flush */
1440 if (mtctx->jobs[wJobID].consumed == mtctx->jobs[wJobID].src.size) {
1441 DEBUGLOG(5, "job %u is completely consumed (%u == %u) => don't wait for cond, there will be none",
1442 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].consumed, (U32)mtctx->jobs[wJobID].src.size);
1443 break;
1444 }
1445 DEBUGLOG(5, "waiting for something to flush from job %u (currently flushed: %u bytes)",
1446 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
1447 ZSTD_pthread_cond_wait(&mtctx->jobs[wJobID].job_cond, &mtctx->jobs[wJobID].job_mutex); /* block when nothing to flush but some to come */
1448 } }
1449
1450 /* try to flush something */
1451 { size_t cSize = mtctx->jobs[wJobID].cSize; /* shared */
1452 size_t const srcConsumed = mtctx->jobs[wJobID].consumed; /* shared */
1453 size_t const srcSize = mtctx->jobs[wJobID].src.size; /* read-only, could be done after mutex lock, but no-declaration-after-statement */
1454 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1455 if (ZSTD_isError(cSize)) {
1456 DEBUGLOG(5, "ZSTDMT_flushProduced: job %u : compression error detected : %s",
1457 mtctx->doneJobID, ZSTD_getErrorName(cSize));
1458 ZSTDMT_waitForAllJobsCompleted(mtctx);
1459 ZSTDMT_releaseAllJobResources(mtctx);
1460 return cSize;
1461 }
1462 /* add frame checksum if necessary (can only happen once) */
1463 assert(srcConsumed <= srcSize);
1464 if ( (srcConsumed == srcSize) /* job completed -> worker no longer active */
1465 && mtctx->jobs[wJobID].frameChecksumNeeded ) {
1466 U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
1467 DEBUGLOG(4, "ZSTDMT_flushProduced: writing checksum : %08X \n", checksum);
1468 MEM_writeLE32((char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].cSize, checksum);
1469 cSize += 4;
1470 mtctx->jobs[wJobID].cSize += 4; /* can write this shared value, as worker is no longer active */
1471 mtctx->jobs[wJobID].frameChecksumNeeded = 0;
1472 }
1473
1474 if (cSize > 0) { /* compression is ongoing or completed */
1475 size_t const toFlush = MIN(cSize - mtctx->jobs[wJobID].dstFlushed, output->size - output->pos);
1476 DEBUGLOG(5, "ZSTDMT_flushProduced: Flushing %u bytes from job %u (completion:%u/%u, generated:%u)",
1477 (U32)toFlush, mtctx->doneJobID, (U32)srcConsumed, (U32)srcSize, (U32)cSize);
1478 assert(mtctx->doneJobID < mtctx->nextJobID);
1479 assert(cSize >= mtctx->jobs[wJobID].dstFlushed);
1480 assert(mtctx->jobs[wJobID].dstBuff.start != NULL);
1481 if (toFlush > 0) {
1482 ZSTD_memcpy((char*)output->dst + output->pos,
1483 (const char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].dstFlushed,
1484 toFlush);
1485 }
1486 output->pos += toFlush;
1487 mtctx->jobs[wJobID].dstFlushed += toFlush; /* can write : this value is only used by mtctx */
1488
1489 if ( (srcConsumed == srcSize) /* job is completed */
1490 && (mtctx->jobs[wJobID].dstFlushed == cSize) ) { /* output buffer fully flushed => free this job position */
1491 DEBUGLOG(5, "Job %u completed (%u bytes), moving to next one",
1492 mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
1493 ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[wJobID].dstBuff);
1494 DEBUGLOG(5, "dstBuffer released");
1495 mtctx->jobs[wJobID].dstBuff = g_nullBuffer;
1496 mtctx->jobs[wJobID].cSize = 0; /* ensure this job slot is considered "not started" in future check */
1497 mtctx->consumed += srcSize;
1498 mtctx->produced += cSize;
1499 mtctx->doneJobID++;
1500 } }
1501
1502 /* return value : how many bytes left in buffer ; fake it to 1 when unknown but >0 */
1503 if (cSize > mtctx->jobs[wJobID].dstFlushed) return (cSize - mtctx->jobs[wJobID].dstFlushed);
1504 if (srcSize > srcConsumed) return 1; /* current job not completely compressed */
1505 }
1506 if (mtctx->doneJobID < mtctx->nextJobID) return 1; /* some more jobs ongoing */
1507 if (mtctx->jobReady) return 1; /* one job is ready to push, just not yet in the list */
1508 if (mtctx->inBuff.filled > 0) return 1; /* input is not empty, and still needs to be converted into a job */
1509 mtctx->allJobsCompleted = mtctx->frameEnded; /* all jobs are entirely flushed => if this one is last one, frame is completed */
1510 if (end == ZSTD_e_end) return !mtctx->frameEnded; /* for ZSTD_e_end, question becomes : is frame completed ? instead of : are internal buffers fully flushed ? */
1511 return 0; /* internal buffers fully flushed */
1512 }
1513
1514 /**
1515 * Returns the range of data used by the earliest job that is not yet complete.
1516 * If the data of the first job is broken up into two segments, we cover both
1517 * sections.
1518 */
ZSTDMT_getInputDataInUse(ZSTDMT_CCtx * mtctx)1519 static range_t ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx)
1520 {
1521 unsigned const firstJobID = mtctx->doneJobID;
1522 unsigned const lastJobID = mtctx->nextJobID;
1523 unsigned jobID;
1524
1525 for (jobID = firstJobID; jobID < lastJobID; ++jobID) {
1526 unsigned const wJobID = jobID & mtctx->jobIDMask;
1527 size_t consumed;
1528
1529 ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
1530 consumed = mtctx->jobs[wJobID].consumed;
1531 ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
1532
1533 if (consumed < mtctx->jobs[wJobID].src.size) {
1534 range_t range = mtctx->jobs[wJobID].prefix;
1535 if (range.size == 0) {
1536 /* Empty prefix */
1537 range = mtctx->jobs[wJobID].src;
1538 }
1539 /* Job source in multiple segments not supported yet */
1540 assert(range.start <= mtctx->jobs[wJobID].src.start);
1541 return range;
1542 }
1543 }
1544 return kNullRange;
1545 }
1546
1547 /**
1548 * Returns non-zero iff buffer and range overlap.
1549 */
ZSTDMT_isOverlapped(buffer_t buffer,range_t range)1550 static int ZSTDMT_isOverlapped(buffer_t buffer, range_t range)
1551 {
1552 BYTE const* const bufferStart = (BYTE const*)buffer.start;
1553 BYTE const* const rangeStart = (BYTE const*)range.start;
1554
1555 if (rangeStart == NULL || bufferStart == NULL)
1556 return 0;
1557
1558 {
1559 BYTE const* const bufferEnd = bufferStart + buffer.capacity;
1560 BYTE const* const rangeEnd = rangeStart + range.size;
1561
1562 /* Empty ranges cannot overlap */
1563 if (bufferStart == bufferEnd || rangeStart == rangeEnd)
1564 return 0;
1565
1566 return bufferStart < rangeEnd && rangeStart < bufferEnd;
1567 }
1568 }
1569
ZSTDMT_doesOverlapWindow(buffer_t buffer,ZSTD_window_t window)1570 static int ZSTDMT_doesOverlapWindow(buffer_t buffer, ZSTD_window_t window)
1571 {
1572 range_t extDict;
1573 range_t prefix;
1574
1575 DEBUGLOG(5, "ZSTDMT_doesOverlapWindow");
1576 extDict.start = window.dictBase + window.lowLimit;
1577 extDict.size = window.dictLimit - window.lowLimit;
1578
1579 prefix.start = window.base + window.dictLimit;
1580 prefix.size = window.nextSrc - (window.base + window.dictLimit);
1581 DEBUGLOG(5, "extDict [0x%zx, 0x%zx)",
1582 (size_t)extDict.start,
1583 (size_t)extDict.start + extDict.size);
1584 DEBUGLOG(5, "prefix [0x%zx, 0x%zx)",
1585 (size_t)prefix.start,
1586 (size_t)prefix.start + prefix.size);
1587
1588 return ZSTDMT_isOverlapped(buffer, extDict)
1589 || ZSTDMT_isOverlapped(buffer, prefix);
1590 }
1591
ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx * mtctx,buffer_t buffer)1592 static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, buffer_t buffer)
1593 {
1594 if (mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable) {
1595 ZSTD_pthread_mutex_t* mutex = &mtctx->serial.ldmWindowMutex;
1596 DEBUGLOG(5, "ZSTDMT_waitForLdmComplete");
1597 DEBUGLOG(5, "source [0x%zx, 0x%zx)",
1598 (size_t)buffer.start,
1599 (size_t)buffer.start + buffer.capacity);
1600 ZSTD_PTHREAD_MUTEX_LOCK(mutex);
1601 while (ZSTDMT_doesOverlapWindow(buffer, mtctx->serial.ldmWindow)) {
1602 DEBUGLOG(5, "Waiting for LDM to finish...");
1603 ZSTD_pthread_cond_wait(&mtctx->serial.ldmWindowCond, mutex);
1604 }
1605 DEBUGLOG(6, "Done waiting for LDM to finish");
1606 ZSTD_pthread_mutex_unlock(mutex);
1607 }
1608 }
1609
1610 /**
1611 * Attempts to set the inBuff to the next section to fill.
1612 * If any part of the new section is still in use we give up.
1613 * Returns non-zero if the buffer is filled.
1614 */
ZSTDMT_tryGetInputRange(ZSTDMT_CCtx * mtctx)1615 static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx)
1616 {
1617 range_t const inUse = ZSTDMT_getInputDataInUse(mtctx);
1618 size_t const spaceLeft = mtctx->roundBuff.capacity - mtctx->roundBuff.pos;
1619 size_t const target = mtctx->targetSectionSize;
1620 buffer_t buffer;
1621
1622 DEBUGLOG(5, "ZSTDMT_tryGetInputRange");
1623 assert(mtctx->inBuff.buffer.start == NULL);
1624 assert(mtctx->roundBuff.capacity >= target);
1625
1626 if (spaceLeft < target) {
1627 /* ZSTD_invalidateRepCodes() doesn't work for extDict variants.
1628 * Simply copy the prefix to the beginning in that case.
1629 */
1630 BYTE* const start = (BYTE*)mtctx->roundBuff.buffer;
1631 size_t const prefixSize = mtctx->inBuff.prefix.size;
1632
1633 buffer.start = start;
1634 buffer.capacity = prefixSize;
1635 if (ZSTDMT_isOverlapped(buffer, inUse)) {
1636 DEBUGLOG(5, "Waiting for buffer...");
1637 return 0;
1638 }
1639 ZSTDMT_waitForLdmComplete(mtctx, buffer);
1640 ZSTD_memmove(start, mtctx->inBuff.prefix.start, prefixSize);
1641 mtctx->inBuff.prefix.start = start;
1642 mtctx->roundBuff.pos = prefixSize;
1643 }
1644 buffer.start = mtctx->roundBuff.buffer + mtctx->roundBuff.pos;
1645 buffer.capacity = target;
1646
1647 if (ZSTDMT_isOverlapped(buffer, inUse)) {
1648 DEBUGLOG(5, "Waiting for buffer...");
1649 return 0;
1650 }
1651 assert(!ZSTDMT_isOverlapped(buffer, mtctx->inBuff.prefix));
1652
1653 ZSTDMT_waitForLdmComplete(mtctx, buffer);
1654
1655 DEBUGLOG(5, "Using prefix range [%zx, %zx)",
1656 (size_t)mtctx->inBuff.prefix.start,
1657 (size_t)mtctx->inBuff.prefix.start + mtctx->inBuff.prefix.size);
1658 DEBUGLOG(5, "Using source range [%zx, %zx)",
1659 (size_t)buffer.start,
1660 (size_t)buffer.start + buffer.capacity);
1661
1662
1663 mtctx->inBuff.buffer = buffer;
1664 mtctx->inBuff.filled = 0;
1665 assert(mtctx->roundBuff.pos + buffer.capacity <= mtctx->roundBuff.capacity);
1666 return 1;
1667 }
1668
1669 typedef struct {
1670 size_t toLoad; /* The number of bytes to load from the input. */
1671 int flush; /* Boolean declaring if we must flush because we found a synchronization point. */
1672 } syncPoint_t;
1673
1674 /**
1675 * Searches through the input for a synchronization point. If one is found, we
1676 * will instruct the caller to flush, and return the number of bytes to load.
1677 * Otherwise, we will load as many bytes as possible and instruct the caller
1678 * to continue as normal.
1679 */
1680 static syncPoint_t
findSynchronizationPoint(ZSTDMT_CCtx const * mtctx,ZSTD_inBuffer const input)1681 findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
1682 {
1683 BYTE const* const istart = (BYTE const*)input.src + input.pos;
1684 U64 const primePower = mtctx->rsync.primePower;
1685 U64 const hitMask = mtctx->rsync.hitMask;
1686
1687 syncPoint_t syncPoint;
1688 U64 hash;
1689 BYTE const* prev;
1690 size_t pos;
1691
1692 syncPoint.toLoad = MIN(input.size - input.pos, mtctx->targetSectionSize - mtctx->inBuff.filled);
1693 syncPoint.flush = 0;
1694 if (!mtctx->params.rsyncable)
1695 /* Rsync is disabled. */
1696 return syncPoint;
1697 if (mtctx->inBuff.filled + input.size - input.pos < RSYNC_MIN_BLOCK_SIZE)
1698 /* We don't emit synchronization points if it would produce too small blocks.
1699 * We don't have enough input to find a synchronization point, so don't look.
1700 */
1701 return syncPoint;
1702 if (mtctx->inBuff.filled + syncPoint.toLoad < RSYNC_LENGTH)
1703 /* Not enough to compute the hash.
1704 * We will miss any synchronization points in this RSYNC_LENGTH byte
1705 * window. However, since it depends only in the internal buffers, if the
1706 * state is already synchronized, we will remain synchronized.
1707 * Additionally, the probability that we miss a synchronization point is
1708 * low: RSYNC_LENGTH / targetSectionSize.
1709 */
1710 return syncPoint;
1711 /* Initialize the loop variables. */
1712 if (mtctx->inBuff.filled < RSYNC_MIN_BLOCK_SIZE) {
1713 /* We don't need to scan the first RSYNC_MIN_BLOCK_SIZE positions
1714 * because they can't possibly be a sync point. So we can start
1715 * part way through the input buffer.
1716 */
1717 pos = RSYNC_MIN_BLOCK_SIZE - mtctx->inBuff.filled;
1718 if (pos >= RSYNC_LENGTH) {
1719 prev = istart + pos - RSYNC_LENGTH;
1720 hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
1721 } else {
1722 assert(mtctx->inBuff.filled >= RSYNC_LENGTH);
1723 prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
1724 hash = ZSTD_rollingHash_compute(prev + pos, (RSYNC_LENGTH - pos));
1725 hash = ZSTD_rollingHash_append(hash, istart, pos);
1726 }
1727 } else {
1728 /* We have enough bytes buffered to initialize the hash,
1729 * and are have processed enough bytes to find a sync point.
1730 * Start scanning at the beginning of the input.
1731 */
1732 assert(mtctx->inBuff.filled >= RSYNC_MIN_BLOCK_SIZE);
1733 assert(RSYNC_MIN_BLOCK_SIZE >= RSYNC_LENGTH);
1734 pos = 0;
1735 prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
1736 hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
1737 if ((hash & hitMask) == hitMask) {
1738 /* We're already at a sync point so don't load any more until
1739 * we're able to flush this sync point.
1740 * This likely happened because the job table was full so we
1741 * couldn't add our job.
1742 */
1743 syncPoint.toLoad = 0;
1744 syncPoint.flush = 1;
1745 return syncPoint;
1746 }
1747 }
1748 /* Starting with the hash of the previous RSYNC_LENGTH bytes, roll
1749 * through the input. If we hit a synchronization point, then cut the
1750 * job off, and tell the compressor to flush the job. Otherwise, load
1751 * all the bytes and continue as normal.
1752 * If we go too long without a synchronization point (targetSectionSize)
1753 * then a block will be emitted anyways, but this is okay, since if we
1754 * are already synchronized we will remain synchronized.
1755 */
1756 for (; pos < syncPoint.toLoad; ++pos) {
1757 BYTE const toRemove = pos < RSYNC_LENGTH ? prev[pos] : istart[pos - RSYNC_LENGTH];
1758 assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
1759 hash = ZSTD_rollingHash_rotate(hash, toRemove, istart[pos], primePower);
1760 assert(mtctx->inBuff.filled + pos >= RSYNC_MIN_BLOCK_SIZE);
1761 if ((hash & hitMask) == hitMask) {
1762 syncPoint.toLoad = pos + 1;
1763 syncPoint.flush = 1;
1764 break;
1765 }
1766 }
1767 return syncPoint;
1768 }
1769
ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx * mtctx)1770 size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx)
1771 {
1772 size_t hintInSize = mtctx->targetSectionSize - mtctx->inBuff.filled;
1773 if (hintInSize==0) hintInSize = mtctx->targetSectionSize;
1774 return hintInSize;
1775 }
1776
1777 /** ZSTDMT_compressStream_generic() :
1778 * internal use only - exposed to be invoked from zstd_compress.c
1779 * assumption : output and input are valid (pos <= size)
1780 * @return : minimum amount of data remaining to flush, 0 if none */
ZSTDMT_compressStream_generic(ZSTDMT_CCtx * mtctx,ZSTD_outBuffer * output,ZSTD_inBuffer * input,ZSTD_EndDirective endOp)1781 size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
1782 ZSTD_outBuffer* output,
1783 ZSTD_inBuffer* input,
1784 ZSTD_EndDirective endOp)
1785 {
1786 unsigned forwardInputProgress = 0;
1787 DEBUGLOG(5, "ZSTDMT_compressStream_generic (endOp=%u, srcSize=%u)",
1788 (U32)endOp, (U32)(input->size - input->pos));
1789 assert(output->pos <= output->size);
1790 assert(input->pos <= input->size);
1791
1792 if ((mtctx->frameEnded) && (endOp==ZSTD_e_continue)) {
1793 /* current frame being ended. Only flush/end are allowed */
1794 return ERROR(stage_wrong);
1795 }
1796
1797 /* fill input buffer */
1798 if ( (!mtctx->jobReady)
1799 && (input->size > input->pos) ) { /* support NULL input */
1800 if (mtctx->inBuff.buffer.start == NULL) {
1801 assert(mtctx->inBuff.filled == 0); /* Can't fill an empty buffer */
1802 if (!ZSTDMT_tryGetInputRange(mtctx)) {
1803 /* It is only possible for this operation to fail if there are
1804 * still compression jobs ongoing.
1805 */
1806 DEBUGLOG(5, "ZSTDMT_tryGetInputRange failed");
1807 assert(mtctx->doneJobID != mtctx->nextJobID);
1808 } else
1809 DEBUGLOG(5, "ZSTDMT_tryGetInputRange completed successfully : mtctx->inBuff.buffer.start = %p", mtctx->inBuff.buffer.start);
1810 }
1811 if (mtctx->inBuff.buffer.start != NULL) {
1812 syncPoint_t const syncPoint = findSynchronizationPoint(mtctx, *input);
1813 if (syncPoint.flush && endOp == ZSTD_e_continue) {
1814 endOp = ZSTD_e_flush;
1815 }
1816 assert(mtctx->inBuff.buffer.capacity >= mtctx->targetSectionSize);
1817 DEBUGLOG(5, "ZSTDMT_compressStream_generic: adding %u bytes on top of %u to buffer of size %u",
1818 (U32)syncPoint.toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize);
1819 ZSTD_memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, syncPoint.toLoad);
1820 input->pos += syncPoint.toLoad;
1821 mtctx->inBuff.filled += syncPoint.toLoad;
1822 forwardInputProgress = syncPoint.toLoad>0;
1823 }
1824 }
1825 if ((input->pos < input->size) && (endOp == ZSTD_e_end)) {
1826 /* Can't end yet because the input is not fully consumed.
1827 * We are in one of these cases:
1828 * - mtctx->inBuff is NULL & empty: we couldn't get an input buffer so don't create a new job.
1829 * - We filled the input buffer: flush this job but don't end the frame.
1830 * - We hit a synchronization point: flush this job but don't end the frame.
1831 */
1832 assert(mtctx->inBuff.filled == 0 || mtctx->inBuff.filled == mtctx->targetSectionSize || mtctx->params.rsyncable);
1833 endOp = ZSTD_e_flush;
1834 }
1835
1836 if ( (mtctx->jobReady)
1837 || (mtctx->inBuff.filled >= mtctx->targetSectionSize) /* filled enough : let's compress */
1838 || ((endOp != ZSTD_e_continue) && (mtctx->inBuff.filled > 0)) /* something to flush : let's go */
1839 || ((endOp == ZSTD_e_end) && (!mtctx->frameEnded)) ) { /* must finish the frame with a zero-size block */
1840 size_t const jobSize = mtctx->inBuff.filled;
1841 assert(mtctx->inBuff.filled <= mtctx->targetSectionSize);
1842 FORWARD_IF_ERROR( ZSTDMT_createCompressionJob(mtctx, jobSize, endOp) , "");
1843 }
1844
1845 /* check for potential compressed data ready to be flushed */
1846 { size_t const remainingToFlush = ZSTDMT_flushProduced(mtctx, output, !forwardInputProgress, endOp); /* block if there was no forward input progress */
1847 if (input->pos < input->size) return MAX(remainingToFlush, 1); /* input not consumed : do not end flush yet */
1848 DEBUGLOG(5, "end of ZSTDMT_compressStream_generic: remainingToFlush = %u", (U32)remainingToFlush);
1849 return remainingToFlush;
1850 }
1851 }
1852