summaryrefslogtreecommitdiffstats
path: root/src/common/FastCDC.cc
diff options
context:
space:
mode:
authorDaniel Baumann <daniel.baumann@progress-linux.org>2024-04-07 18:45:59 +0000
committerDaniel Baumann <daniel.baumann@progress-linux.org>2024-04-07 18:45:59 +0000
commit19fcec84d8d7d21e796c7624e521b60d28ee21ed (patch)
tree42d26aa27d1e3f7c0b8bd3fd14e7d7082f5008dc /src/common/FastCDC.cc
parentInitial commit. (diff)
downloadceph-19fcec84d8d7d21e796c7624e521b60d28ee21ed.tar.xz
ceph-19fcec84d8d7d21e796c7624e521b60d28ee21ed.zip
Adding upstream version 16.2.11+ds.upstream/16.2.11+dsupstream
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'src/common/FastCDC.cc')
-rw-r--r--src/common/FastCDC.cc176
1 files changed, 176 insertions, 0 deletions
diff --git a/src/common/FastCDC.cc b/src/common/FastCDC.cc
new file mode 100644
index 000000000..941fc873c
--- /dev/null
+++ b/src/common/FastCDC.cc
@@ -0,0 +1,176 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include <random>
+
+#include "FastCDC.h"
+
+
+// Unlike FastCDC described in the paper, if we are close to the
+// target, use the target mask. If we are very small or very large,
+// use an adjusted mask--like the paper. This tries to keep more
+// cut points using the same mask, and fewer using the small or large
+// masks.
+
+// How many more/fewer bits to set in the small/large masks.
+//
+// This is the "normalization level" or "NC level" in the FastCDC
+// paper.
+#define TARGET_WINDOW_MASK_BITS 2
+
+// How big the 'target window' is (in which we use the target mask).
+//
+// In the FastCDC paper, this is always 0: there is not "target
+// window," and either small_mask (maskS) or large_mask (maskL) is
+// used--never target_mask (maskA).
+#define TARGET_WINDOW_BITS 1
+
+// How many bits larger/smaller than target for hard limits on chunk
+// size.
+//
+// We assume the min and max sizes are always this many bits
+// larger/smaller than the target. (Note that the FastCDC paper 8KB
+// example has a min of 2KB (2 bits smaller) and max of 64 KB (3 bits
+// larger), although it is not clear why they chose those values.)
+#define SIZE_WINDOW_BITS 2
+
+void FastCDC::_setup(int target, int size_window_bits)
+{
+ target_bits = target;
+
+ if (!size_window_bits) {
+ size_window_bits = SIZE_WINDOW_BITS;
+ }
+ min_bits = target - size_window_bits;
+ max_bits = target + size_window_bits;
+
+ std::mt19937_64 engine;
+
+ // prefill table
+ for (unsigned i = 0; i < 256; ++i) {
+ table[i] = engine();
+ }
+
+ // set mask
+ int did = 0;
+ uint64_t m = 0;
+ while (did < target_bits + TARGET_WINDOW_MASK_BITS) {
+ uint64_t bit = 1ull << (engine() & 63);
+ if (m & bit) {
+ continue; // this bit is already set
+ }
+ m |= bit;
+ ++did;
+ if (did == target_bits - TARGET_WINDOW_MASK_BITS) {
+ large_mask = m;
+ } else if (did == target_bits) {
+ target_mask = m;
+ } else if (did == target_bits + TARGET_WINDOW_MASK_BITS) {
+ small_mask = m;
+ }
+ }
+}
+
+static inline bool _scan(
+ // these are our cursor/postion...
+ bufferlist::buffers_t::const_iterator *p,
+ const char **pp, const char **pe,
+ size_t& pos,
+ size_t max, // how much to read
+ uint64_t& fp, // fingerprint
+ uint64_t mask, const uint64_t *table)
+{
+ while (pos < max) {
+ if (*pp == *pe) {
+ ++(*p);
+ *pp = (*p)->c_str();
+ *pe = *pp + (*p)->length();
+ }
+ const char *te = std::min(*pe, *pp + max - pos);
+ for (; *pp < te; ++(*pp), ++pos) {
+ if ((fp & mask) == mask) {
+ return false;
+ }
+ fp = (fp << 1) ^ table[*(unsigned char*)*pp];
+ }
+ if (pos >= max) {
+ return true;
+ }
+ }
+ return true;
+}
+
+void FastCDC::calc_chunks(
+ const bufferlist& bl,
+ std::vector<std::pair<uint64_t, uint64_t>> *chunks) const
+{
+ if (bl.length() == 0) {
+ return;
+ }
+ auto p = bl.buffers().begin();
+ const char *pp = p->c_str();
+ const char *pe = pp + p->length();
+
+ size_t pos = 0;
+ size_t len = bl.length();
+ while (pos < len) {
+ size_t cstart = pos;
+ uint64_t fp = 0;
+
+ // are we left with a min-sized (or smaller) chunk?
+ if (len - pos <= (1ul << min_bits)) {
+ chunks->push_back(std::pair<uint64_t,uint64_t>(pos, len - pos));
+ break;
+ }
+
+ // skip forward to the min chunk size cut point (minus the window, so
+ // we can initialize the rolling fingerprint).
+ size_t skip = (1 << min_bits) - window;
+ pos += skip;
+ while (skip) {
+ size_t s = std::min<size_t>(pe - pp, skip);
+ skip -= s;
+ pp += s;
+ if (pp == pe) {
+ ++p;
+ pp = p->c_str();
+ pe = pp + p->length();
+ }
+ }
+
+ // first fill the window
+ size_t max = pos + window;
+ while (pos < max) {
+ if (pp == pe) {
+ ++p;
+ pp = p->c_str();
+ pe = pp + p->length();
+ }
+ const char *te = std::min(pe, pp + (max - pos));
+ for (; pp < te; ++pp, ++pos) {
+ fp = (fp << 1) ^ table[*(unsigned char*)pp];
+ }
+ }
+ ceph_assert(pos < len);
+
+ // find an end marker
+ if (
+ // for the first "small" region
+ _scan(&p, &pp, &pe, pos,
+ std::min(len, cstart + (1 << (target_bits - TARGET_WINDOW_BITS))),
+ fp, small_mask, table) &&
+ // for the middle range (close to our target)
+ (TARGET_WINDOW_BITS == 0 ||
+ _scan(&p, &pp, &pe, pos,
+ std::min(len, cstart + (1 << (target_bits + TARGET_WINDOW_BITS))),
+ fp, target_mask, table)) &&
+ // we're past target, use large_mask!
+ _scan(&p, &pp, &pe, pos,
+ std::min(len,
+ cstart + (1 << max_bits)),
+ fp, large_mask, table))
+ ;
+
+ chunks->push_back(std::pair<uint64_t,uint64_t>(cstart, pos - cstart));
+ }
+}