ddidderr/zstd-rs
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[lib] Add rsyncable mode

Browse Source 
- Add rsyncable mode to multithreaded mode
- Factor out LDM's hash function for reuse
This commit is contained in:
Nick Terrell
2018-11-14 16:59:57 -08:00
parent c584e84e68
commit b9693d3a49
7 changed files with 218 additions and 80 deletions
Download Patch File Download Diff File Expand all files Collapse all files
lib/compress/zstdmt_compress.c
+112 -17
View File
@@ -471,7 +471,7 @@ static int ZSTDMT_serialState_reset(serialState_t* serialState, ZSTDMT_seqPool*
assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
assert(params.ldmParams.hashEveryLog < 32);
serialState->ldmState.hashPower =
ZSTD_ldm_getHashPower(params.ldmParams.minMatchLength);
ZSTD_rollingHash_primePower(params.ldmParams.minMatchLength);
} else {
memset(&params.ldmParams, 0, sizeof(params.ldmParams));
}
@@ -777,6 +777,14 @@ typedef struct {
static const roundBuff_t kNullRoundBuff = {NULL, 0, 0};
#define RSYNC_LENGTH 32
typedef struct {
U64 hash;
U64 hitMask;
U64 primePower;
} rsyncState_t;
struct ZSTDMT_CCtx_s {
POOL_ctx* factory;
ZSTDMT_jobDescription* jobs;
@@ -790,6 +798,7 @@ struct ZSTDMT_CCtx_s {
inBuff_t inBuff;
roundBuff_t roundBuff;
serialState_t serial;
rsyncState_t rsync;
unsigned singleBlockingThread;
unsigned jobIDMask;
unsigned doneJobID;
@@ -988,6 +997,9 @@ size_t ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params,
DEBUGLOG(4, "ZSTDMT_p_overlapSectionLog : %u", value);
params->overlapSizeLog = (value >= 9) ? 9 : value;
return value;
case ZSTDMT_p_rsyncable :
params->rsyncable = (value == 0 ? 0 : 1);
return value;
default :
return ERROR(parameter_unsupported);
}
@@ -996,15 +1008,7 @@ size_t ZSTDMT_CCtxParam_setMTCtxParameter(ZSTD_CCtx_params* params,
size_t ZSTDMT_setMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, unsigned value)
{
DEBUGLOG(4, "ZSTDMT_setMTCtxParameter");
switch(parameter)
{
case ZSTDMT_p_jobSize :
return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value);
case ZSTDMT_p_overlapSectionLog :
return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value);
default :
return ERROR(parameter_unsupported);
}
return ZSTDMT_CCtxParam_setMTCtxParameter(&mtctx->params, parameter, value);
}
size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter, unsigned* value)
@@ -1016,6 +1020,9 @@ size_t ZSTDMT_getMTCtxParameter(ZSTDMT_CCtx* mtctx, ZSTDMT_parameter parameter,
case ZSTDMT_p_overlapSectionLog:
*value = mtctx->params.overlapSizeLog;
break;
case ZSTDMT_p_rsyncable:
*value = mtctx->params.rsyncable;
break;
default:
return ERROR(parameter_unsupported);
}
@@ -1381,6 +1388,16 @@ size_t ZSTDMT_initCStream_internal(
if (mtctx->targetSectionSize == 0) {
mtctx->targetSectionSize = 1ULL << ZSTDMT_computeTargetJobLog(params);
}
if (params.rsyncable) {
/* Aim for the targetsectionSize as the average job size. */
U32 const jobSizeMB = (U32)(mtctx->targetSectionSize >> 20);
U32 const rsyncBits = ZSTD_highbit32(jobSizeMB) + 20;
assert(jobSizeMB >= 1);
DEBUGLOG(4, "rsyncLog = %u", rsyncBits);
mtctx->rsync.hash = 0;
mtctx->rsync.hitMask = (1ULL << rsyncBits) - 1;
mtctx->rsync.primePower = ZSTD_rollingHash_primePower(RSYNC_LENGTH);
}
if (mtctx->targetSectionSize < mtctx->targetPrefixSize) mtctx->targetSectionSize = mtctx->targetPrefixSize; /* job size must be >= overlap size */
DEBUGLOG(4, "Job Size : %u KB (note : set to %u)", (U32)(mtctx->targetSectionSize>>10), params.jobSize);
DEBUGLOG(4, "inBuff Size : %u KB", (U32)(mtctx->targetSectionSize>>10));
@@ -1818,6 +1835,80 @@ static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx)
return 1;
}
typedef struct {
size_t toLoad; /* The number of bytes to load from the input. */
int flush; /* Boolean declaring if we must flush because we found a synchronization point. */
} syncPoint_t;
/**
* Searches through the input for a synchronization point. If one is found, we
* will instruct the caller to flush, and return the number of bytes to load.
* Otherwise, we will load as many bytes as possible and instruct the caller
* to continue as normal.
*/
static syncPoint_t findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input) {
BYTE const* const istart = (BYTE const*)input.src + input.pos;
U64 const primePower = mtctx->rsync.primePower;
U64 const hitMask = mtctx->rsync.hitMask;
syncPoint_t syncPoint;
U64 hash;
BYTE const* prev;
size_t pos;
syncPoint.toLoad = MIN(input.size - input.pos, mtctx->targetSectionSize - mtctx->inBuff.filled);
syncPoint.flush = 0;
if (!mtctx->params.rsyncable)
/* Rsync is disabled. */
return syncPoint;
if (mtctx->inBuff.filled + syncPoint.toLoad < RSYNC_LENGTH)
/* Not enough to compute the hash.
* We will miss any synchronization points in this RSYNC_LENGTH byte
* window. However, since it depends only in the internal buffers, if the
* state is already synchronized, we will remain synchronized.
* Additionally, the probability that we miss a synchronization point is
* low: RSYNC_LENGTH / targetSectionSize.
*/
return syncPoint;
/* Initialize the loop variables. */
if (mtctx->inBuff.filled >= RSYNC_LENGTH) {
/* We have enough bytes buffered to initialize the hash.
* Start scanning at the beginning of the input.
*/
pos = 0;
prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
} else {
/* We don't have enough bytes buffered to initialize the hash, but
* we know we have at least RSYNC_LENGTH bytes total.
* Start scanning after the first RSYNC_LENGTH bytes less the bytes
* already buffered.
*/
pos = RSYNC_LENGTH - mtctx->inBuff.filled;
prev = (BYTE const*)mtctx->inBuff.buffer.start - pos;
hash = ZSTD_rollingHash_compute(mtctx->inBuff.buffer.start, mtctx->inBuff.filled);
hash = ZSTD_rollingHash_append(hash, istart, pos);
}
/* Starting with the hash of the previous RSYNC_LENGTH bytes, roll
* through the input. If we hit a synchronization point, then cut the
* job off, and tell the compressor to flush the job. Otherwise, load
* all the bytes and continue as normal.
* If we go too long without a synchronization point (targetSectionSize)
* then a block will be emitted anyways, but this is okay, since if we
* are already synchronized we will remain synchronized.
*/
for (; pos < syncPoint.toLoad; ++pos) {
BYTE const toRemove = pos < RSYNC_LENGTH ? prev[pos] : istart[pos - RSYNC_LENGTH];
/* if (pos >= RSYNC_LENGTH) assert(ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash); */
hash = ZSTD_rollingHash_rotate(hash, toRemove, istart[pos], primePower);
if ((hash & hitMask) == hitMask) {
syncPoint.toLoad = pos + 1;
syncPoint.flush = 1;
break;
}
}
return syncPoint;
}
/** ZSTDMT_compressStream_generic() :
* internal use only - exposed to be invoked from zstd_compress.c
@@ -1844,7 +1935,8 @@ size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
}
/* single-pass shortcut (note : synchronous-mode) */
if ( (mtctx->nextJobID == 0) /* just started */
if ( (!mtctx->params.rsyncable) /* rsyncable mode is disabled */
&& (mtctx->nextJobID == 0) /* just started */
&& (mtctx->inBuff.filled == 0) /* nothing buffered */
&& (!mtctx->jobReady) /* no job already created */
&& (endOp == ZSTD_e_end) /* end order */
@@ -1876,14 +1968,17 @@ size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
DEBUGLOG(5, "ZSTDMT_tryGetInputRange completed successfully : mtctx->inBuff.buffer.start = %p", mtctx->inBuff.buffer.start);
}
if (mtctx->inBuff.buffer.start != NULL) {
size_t const toLoad = MIN(input->size - input->pos, mtctx->targetSectionSize - mtctx->inBuff.filled);
syncPoint_t const syncPoint = findSynchronizationPoint(mtctx, *input);
if (syncPoint.flush && endOp == ZSTD_e_continue) {
endOp = ZSTD_e_flush;
}
assert(mtctx->inBuff.buffer.capacity >= mtctx->targetSectionSize);
DEBUGLOG(5, "ZSTDMT_compressStream_generic: adding %u bytes on top of %u to buffer of size %u",
(U32)toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize);
memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, toLoad);
input->pos += toLoad;
mtctx->inBuff.filled += toLoad;
forwardInputProgress = toLoad>0;
(U32)syncPoint.toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize);
memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, syncPoint.toLoad);
input->pos += syncPoint.toLoad;
mtctx->inBuff.filled += syncPoint.toLoad;
forwardInputProgress = syncPoint.toLoad>0;
}
if ((input->pos < input->size) && (endOp == ZSTD_e_end))
endOp = ZSTD_e_flush; /* can't end now : not all input consumed */