2007-05-28 06:59:35

by Nitin Gupta

[permalink] [raw]
Subject: [RFC] LZO de/compression support - take 5

Hi,

This is kernel port of LZO1X-1 compressor and LZO1X decompressor (safe
version only).

--
If we get no perf. problems with this patch, then I beleive it is now
suitable to inclusion in mainline. Further cleanups and optimizations
can surely be done after that. It's still just ~500 LOC.
--

* Changes since 'take 4' (Full Changelog after this):
1) Again applied 1 cleanup (use cpu_to_le16): this was not causing slower perf.
2) Rolled back very minor cleanup: LZO_CHECK_MPOS_NON_DET macro is
back. This is the last of cleanups that can cause different
assembler output than original. _Maybe_ this is cause for that ~9%
slowdown seen by Richard in compressor.

Can anyone do timing measurement in kernel space only. This will
eliminate all possible problems w.r.t usespace testing. I tried doing
the same using get_jiffies_64() across calls to compressor in the
'compress-test' module but this is giving same value when measured
just before and after calls to lzo1x_compress(). I don't know why.

If anyone can simply measure time across lzo1x_compress() and
lzo1x_decompress() calls in this compress-test module, then that will
give us proper perf. figures.

* Changelog vs. original LZO:
1) Used standard/kernel defined data types: (this eliminated _huge_
#ifdef chunks)
lzo_bytep -> unsigned char *
lzo_uint -> size_t
lzo_xint -> size_t
lzo_uint32p -> u32 *
lzo_uintptr_t -> unsigned long
2) Removed everything #ifdef'ed under COPY_DICT (this is not set for
LZO1X, so removed corres. parts).
3) Removed code #ifdef'ed for LZO1Y, LZO1Z, other variants.
4) Reformatted the code to match general kernel style.
5) The only code change: (as suggested by Andrey)
-#if defined(__LITTLE_ENDIAN)
- m_pos = op - 1;
- m_pos -= (*(const unsigned short *)ip) >> 2;
-#else
- m_pos = op - 1;
- m_pos -= (ip[0] >> 2) + (ip[1] << 6);
-#endif

+ m_pos = op - 1 - (cpu_to_le16(*(const u16 *)ip) >> 2);

(Andrey suggested le16_to_cpu for above but I think it should be cpu_to_le16).
*** Need testing on big endian machine ***

Similarly:
-#if defined(__LITTLE_ENDIAN)
- m_pos -= (*(const unsigned short *)ip) >> 2;
-#else
- m_pos -= (ip[0] >> 2) + (ip[1] << 6);
-#endif

+ m_pos -= cpu_to_le16(*(const u16 *)ip) >> 2;


include/linux/lzo1x.h | 66 +++++++++++
lib/Kconfig | 6 +
lib/Makefile | 1 +
lib/lzo1x/Makefile | 3 +
lib/lzo1x/lzo1x_compress.c | 259 ++++++++++++++++++++++++++++++++++++++++++
lib/lzo1x/lzo1x_decompress.c | 238 ++++++++++++++++++++++++++++++++++++++
lib/lzo1x/lzo1x_int.h | 96 ++++++++++++++++
7 files changed, 669 insertions(+), 0 deletions(-)

Signed-off-by: Nitin Gupta <[email protected]>
---

b/include/linux/lzo1x.h
new file mode 100755
index 0000000..11a6f23
--- /dev/null
+++ b/include/linux/lzo1x.h
@@ -0,0 +1,66 @@
+/* lzo1x.h -- public interface of the LZO1X compression algorithm
+
+ This file is part of the LZO real-time data compression library.
+
+ Copyright (C) 1996-2005 Markus Franz Xaver Johannes Oberhumer
+ All Rights Reserved.
+
+ The LZO library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU General Public License,
+ version 2, as published by the Free Software Foundation.
+
+ The LZO library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU General Public License for more details.
+
+ You should have received a copy of the GNU General Public License
+ along with the LZO library; see the file COPYING.
+ If not, write to the Free Software Foundation, Inc.,
+ 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
+
+ Markus F.X.J. Oberhumer
+ <[email protected]>
+ http://www.oberhumer.com/opensource/lzo/
+
+
+ This file is modified version of lzo1x.h found in original LZO 2.02
+ code. Some additional changes have also been made to make it work
+ in kernel space.
+
+ Nitin Gupta
+ <[email protected]>
+ */
+
+#ifndef __LZO1X_H
+#define __LZO1X_H
+
+/* LZO return codes */
+#define LZO_E_OK 0
+#define LZO_E_ERROR (-1)
+#define LZO_E_OUT_OF_MEMORY (-2) /* [not used right now] */
+#define LZO_E_NOT_COMPRESSIBLE (-3) /* [not used right now] */
+#define LZO_E_INPUT_OVERRUN (-4)
+#define LZO_E_OUTPUT_OVERRUN (-5)
+#define LZO_E_LOOKBEHIND_OVERRUN (-6)
+#define LZO_E_EOF_NOT_FOUND (-7)
+#define LZO_E_INPUT_NOT_CONSUMED (-8)
+#define LZO_E_NOT_YET_IMPLEMENTED (-9) /* [not used right now] */
+
+/* Size of temp buffer (workmem) required by lzo1x_compress */
+#define LZO1X_WORKMEM_SIZE ((size_t) (16384L * sizeof(unsigned char *)))
+
+/*
+ * This requires 'workmem' of size LZO1X_WORKMEM_SIZE
+ */
+int lzo1x_compress(const unsigned char *src, size_t src_len,
+ unsigned char *dst, size_t *dst_len,
+ void *workmem);
+
+/*
+ * This decompressor will catch all compressed data violations and
+ * return an error code in this case.
+ */
+int lzo1x_decompress(const unsigned char *src, size_t src_len,
+ unsigned char *dst, size_t *dst_len);
+#endif
diff --git a/lib/Kconfig b/lib/Kconfig
index 2e7ae6b..257f377 100644
--- a/lib/Kconfig
+++ b/lib/Kconfig
@@ -64,6 +64,12 @@ config ZLIB_INFLATE
config ZLIB_DEFLATE
tristate

+config LZO1X
+ tristate "LZO1X Compression/Decompression"
+ help
+ Compression: LZO1X-1
+ Decompression: LZO1X (safe)
+
#
# Generic allocator support is selected if needed
#
diff --git a/lib/Makefile b/lib/Makefile
index c8c8e20..4dad99d 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -49,6 +49,7 @@ obj-$(CONFIG_GENERIC_ALLOCATOR) += genalloc.o
obj-$(CONFIG_ZLIB_INFLATE) += zlib_inflate/
obj-$(CONFIG_ZLIB_DEFLATE) += zlib_deflate/
obj-$(CONFIG_REED_SOLOMON) += reed_solomon/
+obj-$(CONFIG_LZO1X) += lzo1x/

obj-$(CONFIG_TEXTSEARCH) += textsearch.o
obj-$(CONFIG_TEXTSEARCH_KMP) += ts_kmp.o
diff --git a/lib/lzo1x/Makefile b/lib/lzo1x/Makefile
new file mode 100644
index 0000000..fcd0d3e
--- /dev/null
+++ b/lib/lzo1x/Makefile
@@ -0,0 +1,3 @@
+obj-$(CONFIG_LZO1X) += lzo1x.o
+lzo1x-objs := lzo1x_compress.o lzo1x_decompress.o
+
diff --git a/lib/lzo1x/lzo1x_compress.c b/lib/lzo1x/lzo1x_compress.c
new file mode 100755
index 0000000..5b0e87f
--- /dev/null
+++ b/lib/lzo1x/lzo1x_compress.c
@@ -0,0 +1,259 @@
+/* lzo1x_compress.c -- LZO1X-1 compression
+
+ This file is part of the LZO real-time data compression library.
+
+ Copyright (C) 1996-2005 Markus Franz Xaver Johannes Oberhumer
+ All Rights Reserved.
+
+ The LZO library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU General Public License,
+ version 2, as published by the Free Software Foundation.
+
+ The LZO library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU General Public License for more details.
+
+ You should have received a copy of the GNU General Public License
+ along with the LZO library; see the file COPYING.
+ If not, write to the Free Software Foundation, Inc.,
+ 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
+
+ Markus F.X.J. Oberhumer
+ <[email protected]>
+ http://www.oberhumer.com/opensource/lzo/
+
+
+ This file is derived from lzo1x_1.c and lzo1x_c.ch found in original
+ LZO 2.02 code. Some additional changes have also been made to make
+ it work in kernel space.
+
+ Nitin Gupta
+ <[email protected]>
+ */
+
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/compiler.h>
+#include <linux/lzo1x.h>
+
+#include "lzo1x_int.h"
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("LZO1X Compression");
+
+/* compress a block of data. */
+static noinline unsigned int
+lzo1x_compress_worker(const unsigned char *in, size_t in_len,
+ unsigned char *out, size_t *out_len,
+ void *workmem)
+{
+ register const unsigned char *ip;
+ unsigned char *op;
+ const unsigned char * const in_end = in + in_len;
+ const unsigned char * const ip_end = in + in_len - M2_MAX_LEN - 5;
+ const unsigned char *ii;
+ const unsigned char ** const dict = (const unsigned char **)workmem;
+
+ op = out;
+ ip = in;
+ ii = ip;
+
+ ip += 4;
+ for (;;) {
+ register const unsigned char *m_pos;
+ size_t m_off;
+ size_t m_len;
+ size_t dindex;
+
+ DINDEX1(dindex, ip);
+ m_pos = dict[dindex];
+
+ if (LZO_CHECK_MPOS_NON_DET(m_pos, m_off, in, ip,
+ M4_MAX_OFFSET))
+ goto literal;
+
+ if (m_off <= M2_MAX_OFFSET || m_pos[3] == ip[3])
+ goto try_match;
+
+ DINDEX2(dindex, ip);
+ m_pos = dict[dindex];
+
+ if (LZO_CHECK_MPOS_NON_DET(m_pos, m_off, in, ip,
+ M4_MAX_OFFSET))
+ goto literal;
+
+ if (m_off <= M2_MAX_OFFSET || m_pos[3] == ip[3])
+ goto try_match;
+
+ goto literal;
+
+try_match:
+ if (*(const unsigned short *)m_pos ==
+ *(const unsigned short *)ip) {
+ if (likely(m_pos[2] == ip[2]))
+ goto match;
+ }
+
+ /* a literal */
+literal:
+ dict[dindex] = ip;
+ ++ip;
+ if (unlikely(ip >= ip_end))
+ break;
+ continue;
+
+ /* a match */
+match:
+ dict[dindex] = ip;
+ /* store current literal run */
+ if ((size_t)(ip - ii) > 0) {
+ register size_t t = (size_t)(ip - ii);
+ if (t <= 3)
+ op[-2] |= (unsigned char)(t);
+ else if (t <= 18)
+ *op++ = (unsigned char)(t - 3);
+ else {
+ register size_t tt = t - 18;
+ *op++ = 0;
+ while (tt > 255) {
+ tt -= 255;
+ *op++ = 0;
+ }
+ *op++ = (unsigned char)tt;
+ }
+ do
+ *op++ = *ii++;
+ while (--t > 0);
+ }
+
+ /* code the match */
+ ip += 3;
+ if (m_pos[3] != *ip++ || m_pos[4] != *ip++ ||
+ m_pos[5] != *ip++ || m_pos[6] != *ip++ ||
+ m_pos[7] != *ip++ || m_pos[8] != *ip++) {
+ --ip;
+ m_len = (size_t)(ip - ii);
+
+ if (m_off <= M2_MAX_OFFSET) {
+ m_off -= 1;
+ *op++ = (unsigned char)(((m_len - 1) << 5) |
+ ((m_off & 7) << 2));
+ *op++ = (unsigned char)(m_off >> 3);
+ }
+ else if (m_off <= M3_MAX_OFFSET) {
+ m_off -= 1;
+ *op++ = (unsigned char)(M3_MARKER |
+ (m_len - 2));
+ goto m3_m4_offset;
+ } else {
+ m_off -= 0x4000;
+ *op++ = (unsigned char)(M4_MARKER |
+ ((m_off & 0x4000) >> 11) |
+ (m_len - 2));
+ goto m3_m4_offset;
+ }
+ } else {
+ const unsigned char *end = in_end;
+ const unsigned char *m = m_pos + M2_MAX_LEN + 1;
+ while (ip < end && *m == *ip)
+ m++, ip++;
+ m_len = (size_t)(ip - ii);
+
+ if (m_off <= M3_MAX_OFFSET) {
+ m_off -= 1;
+ if (m_len <= 33)
+ *op++ = (unsigned char)(M3_MARKER |
+ (m_len - 2));
+ else {
+ m_len -= 33;
+ *op++ = M3_MARKER | 0;
+ goto m3_m4_len;
+ }
+ } else {
+ m_off -= 0x4000;
+ if (m_len <= M4_MAX_LEN)
+ *op++ = (unsigned char)(M4_MARKER |
+ ((m_off & 0x4000) >> 11) |
+ (m_len - 2));
+ else {
+ m_len -= M4_MAX_LEN;
+ *op++ = (unsigned char)(M4_MARKER |
+ ((m_off & 0x4000) >> 11));
+m3_m4_len:
+ while (m_len > 255) {
+ m_len -= 255;
+ *op++ = 0;
+ }
+ *op++ = (unsigned char)(m_len);
+ }
+ }
+
+m3_m4_offset:
+ *op++ = (unsigned char)((m_off & 63) << 2);
+ *op++ = (unsigned char)(m_off >> 6);
+ }
+
+ ii = ip;
+ if (unlikely(ip >= ip_end))
+ break;
+ }
+
+ *out_len = (size_t)(op - out);
+ return (size_t)(in_end - ii);
+}
+
+
+/*
+ * This requires buffer (workmem) of size LZO1X_WORKMEM_SIZE
+ * (exported by lzo1x.h).
+ */
+int
+lzo1x_compress(const unsigned char *in, size_t in_len,
+ unsigned char *out, size_t *out_len,
+ void *workmem)
+{
+ unsigned char *op = out;
+ size_t t;
+
+ if (!workmem)
+ return -EINVAL;
+
+ if (unlikely(in_len <= M2_MAX_LEN + 5))
+ t = in_len;
+ else {
+ t = lzo1x_compress_worker(in, in_len, op, out_len, workmem);
+ op += *out_len;
+ }
+
+ if (t > 0) {
+ const unsigned char *ii = in + in_len - t;
+
+ if (op == out && t <= 238)
+ *op++ = (unsigned char)(17 + t);
+ else if (t <= 3)
+ op[-2] |= (unsigned char)t;
+ else if (t <= 18)
+ *op++ = (unsigned char)(t - 3);
+ else {
+ size_t tt = t - 18;
+ *op++ = 0;
+ while (tt > 255) {
+ tt -= 255;
+ *op++ = 0;
+ }
+ *op++ = (unsigned char)tt;
+ }
+ do
+ *op++ = *ii++;
+ while (--t > 0);
+ }
+ *op++ = M4_MARKER | 1;
+ *op++ = 0;
+ *op++ = 0;
+
+ *out_len = (size_t)(op - out);
+ return LZO_E_OK;
+}
+
+EXPORT_SYMBOL(lzo1x_compress);
diff --git a/lib/lzo1x/lzo1x_decompress.c b/lib/lzo1x/lzo1x_decompress.c
new file mode 100755
index 0000000..75ce294
--- /dev/null
+++ b/lib/lzo1x/lzo1x_decompress.c
@@ -0,0 +1,238 @@
+/* lzo1x_decompress.c -- LZO1X decompression
+
+ This file is part of the LZO real-time data compression library.
+
+ Copyright (C) 1996-2005 Markus Franz Xaver Johannes Oberhumer
+ All Rights Reserved.
+
+ The LZO library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU General Public License,
+ version 2, as published by the Free Software Foundation.
+
+ The LZO library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU General Public License for more details.
+
+ You should have received a copy of the GNU General Public License
+ along with the LZO library; see the file COPYING.
+ If not, write to the Free Software Foundation, Inc.,
+ 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
+
+ Markus F.X.J. Oberhumer
+ <[email protected]>
+ http://www.oberhumer.com/opensource/lzo/
+
+
+ This file is derived from lzo1x_d1.c and lzo1x_d.ch found in original
+ LZO 2.02 code. Some additional changes have also been made to make
+ it work in kernel space.
+
+ Nitin Gupta
+ <[email protected]>
+ */
+
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <asm/byteorder.h>
+#include <linux/lzo1x.h>
+
+#include "lzo1x_int.h"
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("LZO1X Decompression");
+
+int
+lzo1x_decompress(const unsigned char *in, size_t in_len,
+ unsigned char *out, size_t *out_len)
+{
+ register size_t t;
+ register unsigned char *op = out;
+ register const unsigned char *ip = in, *m_pos;
+ const unsigned char * const ip_end = in + in_len;
+ unsigned char * const op_end = out + *out_len;
+ *out_len = 0;
+
+ if (*ip > 17) {
+ t = *ip++ - 17;
+ if (t < 4)
+ goto match_next;
+ NEED_OP(t);
+ NEED_IP(t + 1);
+ do
+ *op++ = *ip++;
+ while (--t > 0);
+ goto first_literal_run;
+ }
+
+ while (TEST_IP) {
+ t = *ip++;
+ if (t >= 16)
+ goto match;
+ /* a literal run */
+ if (t == 0) {
+ NEED_IP(1);
+ while (*ip == 0) {
+ t += 255;
+ ip++;
+ NEED_IP(1);
+ }
+ t += 15 + *ip++;
+ }
+ /* copy literals */
+ NEED_OP(t + 3);
+ NEED_IP(t + 4);
+ COPY4(op, ip);
+ op += 4; ip += 4;
+ if (--t > 0) {
+ if (t >= 4) {
+ do {
+ COPY4(op, ip);
+ op += 4; ip += 4; t -= 4;
+ } while (t >= 4);
+ if (t > 0)
+ do
+ *op++ = *ip++;
+ while (--t > 0);
+ }
+ else
+ do
+ *op++ = *ip++;
+ while (--t > 0);
+ }
+
+first_literal_run:
+ t = *ip++;
+ if (t >= 16)
+ goto match;
+ m_pos = op - (1 + M2_MAX_OFFSET);
+ m_pos -= t >> 2;
+ m_pos -= *ip++ << 2;
+ TEST_LB(m_pos);
+ NEED_OP(3);
+ *op++ = *m_pos++;
+ *op++ = *m_pos++;
+ *op++ = *m_pos;
+ goto match_done;
+
+ /* handle matches */
+ do {
+match:
+ if (t >= 64) { /* a M2 match */
+ m_pos = op - 1;
+ m_pos -= (t >> 2) & 7;
+ m_pos -= *ip++ << 3;
+ t = (t >> 5) - 1;
+ TEST_LB(m_pos);
+ NEED_OP(t + 3 - 1);
+ goto copy_match;
+ } else if (t >= 32) { /* a M3 match */
+ t &= 31;
+ if (t == 0) {
+ NEED_IP(1);
+ while (*ip == 0) {
+ t += 255;
+ ip++;
+ NEED_IP(1);
+ }
+ t += 31 + *ip++;
+ }
+ m_pos = op - 1 - (cpu_to_le16(
+ *(const unsigned short *)ip) >> 2);
+ ip += 2;
+ } else if (t >= 16) { /* a M4 match */
+ m_pos = op;
+ m_pos -= (t & 8) << 11;
+ t &= 7;
+ if (t == 0) {
+ NEED_IP(1);
+ while (*ip == 0) {
+ t += 255;
+ ip++;
+ NEED_IP(1);
+ }
+ t += 7 + *ip++;
+ }
+ m_pos -= cpu_to_le16(
+ *(const unsigned short *)ip) >> 2;
+ ip += 2;
+ if (m_pos == op)
+ goto eof_found;
+ m_pos -= 0x4000;
+ } else { /* a M1 match */
+ m_pos = op - 1;
+ m_pos -= t >> 2;
+ m_pos -= *ip++ << 2;
+ TEST_LB(m_pos);
+ NEED_OP(2);
+ *op++ = *m_pos++;
+ *op++ = *m_pos;
+ goto match_done;
+ }
+
+ /* copy match */
+ TEST_LB(m_pos);
+ NEED_OP(t + 3 - 1);
+
+ if (t >= 2 * 4 - (3 - 1) && (op - m_pos) >= 4) {
+ COPY4(op, m_pos);
+ op += 4; m_pos += 4; t -= 4 - (3 - 1);
+ do {
+ COPY4(op, m_pos);
+ op += 4; m_pos += 4; t -= 4;
+ } while (t >= 4);
+ if (t > 0)
+ do *op++ = *m_pos++;
+ while (--t > 0);
+ } else {
+copy_match:
+ *op++ = *m_pos++;
+ *op++ = *m_pos++;
+ do
+ *op++ = *m_pos++;
+ while (--t > 0);
+ }
+
+match_done:
+ t = ip[-2] & 3;
+ if (t == 0)
+ break;
+
+ /* copy literals */
+match_next:
+ NEED_OP(t);
+ NEED_IP(t + 1);
+ *op++ = *ip++;
+ if (t > 1) {
+ *op++ = *ip++;
+ if (t > 2)
+ *op++ = *ip++;
+ }
+ t = *ip++;
+ } while (TEST_IP);
+ }
+
+ /* no EOF code was found */
+ *out_len = (size_t)(op - out);
+ return LZO_E_EOF_NOT_FOUND;
+
+eof_found:
+ *out_len = (size_t)(op - out);
+ return (ip == ip_end ? LZO_E_OK :
+ (ip < ip_end ? LZO_E_INPUT_NOT_CONSUMED :
+ LZO_E_INPUT_OVERRUN));
+
+input_overrun:
+ *out_len = (size_t)(op - out);
+ return LZO_E_INPUT_OVERRUN;
+
+output_overrun:
+ *out_len = (size_t)(op - out);
+ return LZO_E_OUTPUT_OVERRUN;
+
+lookbehind_overrun:
+ *out_len = (size_t)(op - out);
+ return LZO_E_LOOKBEHIND_OVERRUN;
+}
+
+EXPORT_SYMBOL(lzo1x_decompress);
diff --git a/lib/lzo1x/lzo1x_int.h b/lib/lzo1x/lzo1x_int.h
new file mode 100755
index 0000000..6c7850c
--- /dev/null
+++ b/lib/lzo1x/lzo1x_int.h
@@ -0,0 +1,96 @@
+/* lzo1x_int.h -- to be used internally by LZO de/compression algorithms
+
+ This file is part of the LZO real-time data compression library.
+
+ Copyright (C) 1996-2005 Markus Franz Xaver Johannes Oberhumer
+ All Rights Reserved.
+
+ The LZO library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU General Public License,
+ version 2, as published by the Free Software Foundation.
+
+ The LZO library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU General Public License for more details.
+
+ You should have received a copy of the GNU General Public License
+ along with the LZO library; see the file COPYING.
+ If not, write to the Free Software Foundation, Inc.,
+ 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
+
+ Markus F.X.J. Oberhumer
+ <[email protected]>
+ http://www.oberhumer.com/opensource/lzo/
+
+
+ This file was derived from several header files found in original
+ LZO 2.02 code. Some additional changes have also been made to make
+ it work in kernel space.
+
+ Nitin Gupta
+ <[email protected]>
+ */
+
+#ifndef __LZO1X_INT_H
+#define __LZO1X_INT_H
+
+#include <linux/types.h>
+
+#define D_BITS 14
+#define D_SIZE (1u << D_BITS)
+#define D_MASK (D_SIZE - 1)
+#define D_HIGH ((D_MASK >> 1) + 1)
+
+#define PTR(a) ((unsigned long)(a))
+#define PTR_LT(a,b) (PTR(a) < PTR(b))
+#define PTR_GE(a,b) (PTR(a) >= PTR(b))
+#define PTR_DIFF(a,b) (PTR(a) - PTR(b))
+
+#define LZO_CHECK_MPOS_NON_DET(m_pos,m_off,in,ip,max_offset) \
+ (m_pos = ip - (size_t)PTR_DIFF(ip, m_pos), \
+ PTR_LT(m_pos, in) || \
+ (m_off = (size_t)PTR_DIFF(ip, m_pos)) <= 0 || \
+ m_off > max_offset)
+
+#define DX2(p,s1,s2) \
+ (((((size_t)((p)[2]) << (s2)) ^ (p)[1]) << (s1)) ^ (p)[0])
+#define DX3(p,s1,s2,s3) \
+ ((DX2((p) + 1, s2, s3) << (s1)) ^ (p)[0])
+#define DINDEX1(d,p) \
+ d = ((size_t)(0x21 * DX3(p, 5, 5, 6)) >> 5) & D_MASK
+#define DINDEX2(d,p) \
+ d = (d & (D_MASK & 0x7ff)) ^ (D_HIGH | 0x1f)
+
+#define COPY4(dst,src) *(u32 *)(dst) = *(u32 *)(src)
+
+/* LZO1X Specific constants */
+#define M1_MAX_OFFSET 0x0400
+#define M2_MAX_OFFSET 0x0800
+#define M3_MAX_OFFSET 0x4000
+#define M4_MAX_OFFSET 0xbfff
+
+#define M1_MIN_LEN 2
+#define M1_MAX_LEN 2
+#define M2_MIN_LEN 3
+#define M2_MAX_LEN 8
+#define M3_MIN_LEN 3
+#define M3_MAX_LEN 33
+#define M4_MIN_LEN 3
+#define M4_MAX_LEN 9
+
+#define M1_MARKER 0
+#define M2_MARKER 64
+#define M3_MARKER 32
+#define M4_MARKER 16
+
+/* Bounds checking */
+#define TEST_IP (ip < ip_end)
+#define NEED_IP(x) \
+ if ((size_t)(ip_end - ip) < (size_t)(x)) goto input_overrun
+#define NEED_OP(x) \
+ if ((size_t)(op_end - op) < (size_t)(x)) goto output_overrun
+#define TEST_LB(m_pos) \
+ if (m_pos < out || m_pos >= op) goto lookbehind_overrun
+
+#endif


Attachments:
(No filename) (20.68 kB)
patch_lzo_2.6.22-rc3.bz2 (4.79 kB)
Download all attachments

2007-05-28 09:44:48

by Satyam Sharma

[permalink] [raw]
Subject: Re: [RFC] LZO de/compression support - take 5

On 5/28/07, Nitin Gupta <[email protected]> wrote:
> [...]
> Can anyone do timing measurement in kernel space only. This will
> eliminate all possible problems w.r.t usespace testing. I tried doing
> the same using get_jiffies_64() across calls to compressor in the
> 'compress-test' module but this is giving same value when measured
> just before and after calls to lzo1x_compress(). I don't know why.

Jiffies wouldn't really offer the kind of precision to serve your purpose.

> If anyone can simply measure time across lzo1x_compress() and
> lzo1x_decompress() calls in this compress-test module, then that will
> give us proper perf. figures.

I wish I had time for this myself (I'll see if I can do this tonight) ... but
anyhow, the idea is quite simply this:

cycles_t start, end;
unsigned long long diff;

start = get_cycles();
/*
* lzo1x_compress() or
* lzo1x_decompress() or whatever
* on large enough input size.
*/
end = get_cycles();

diff = end - start;

That is all there is to it.

Note that get_cycles() (on x86 boxes, at least) might not be usable
or trustworthy on SMP machines unless you ensure that your code
(including the start and end timing measurements) executes on the
same particular CPU.

You could also use something like this when pin-pointing the exact
cause of the performance loss you experienced when the code was
fully cleaned up initially (before you started rolling back some stuff).

Satyam

2007-05-28 10:50:13

by Michael-Luke Jones

[permalink] [raw]
Subject: Re: [RFC] LZO de/compression support - take 5

On 28 May 2007, at 07:59, Nitin Gupta wrote:

> If we get no perf. problems with this patch, then I beleive it is now
> suitable to inclusion in mainline. Further cleanups and optimizations
> can surely be done after that. It's still just ~500 LOC.

Before LZO code is sent to Linus, its selection in Kconfig should be
made orthogonal to the current zlib selection code.

This means:
1) Options in lib/Kconfig hidden (selectable by drivers as required)
2) Decompression and Compression support separated, as read-only
filesystems only need to build in decompression support.

Thanks,

Michael-Luke

2007-05-28 12:09:30

by Nitin Gupta

[permalink] [raw]
Subject: Re: [RFC] LZO de/compression support - take 5

On 5/28/07, Michael-Luke Jones <[email protected]> wrote:
> On 28 May 2007, at 07:59, Nitin Gupta wrote:
>
> > If we get no perf. problems with this patch, then I beleive it is now
> > suitable to inclusion in mainline. Further cleanups and optimizations
> > can surely be done after that. It's still just ~500 LOC.
>
> Before LZO code is sent to Linus, its selection in Kconfig should be
> made orthogonal to the current zlib selection code.
>
> This means:
> 1) Options in lib/Kconfig hidden (selectable by drivers as required)

LZO as hidden option has no practical sense. Although LZO should be
auto-selected when some dependent project is selected (e.g. reieser4)
- there should be separate patch for this. Mixing such changes with
'core' LZO patch will just add side noise.

> 2) Decompression and Compression support separated, as read-only
> filesystems only need to build in decompression support.
>

Ok, I will do this. I wonder if some difference in opinion in such
things can actually cause 10+ extra RFCs?

Cheers,
Nitin

2007-05-28 12:16:18

by Michael-Luke Jones

[permalink] [raw]
Subject: Re: [RFC] LZO de/compression support - take 5

On 28 May 2007, at 13:09, Nitin Gupta wrote:

>> This means:
>> 1) Options in lib/Kconfig hidden (selectable by drivers as required)
>
> LZO as hidden option has no practical sense. Although LZO should be
> auto-selected when some dependent project is selected (e.g. reieser4)
> - there should be separate patch for this. Mixing such changes with
> 'core' LZO patch will just add side noise.

No, LZO as a visible option makes no practical sense. Why would
anyone want to build LZO into a kernel when there are no in-kernel
users of the code?

In fact, all of the library code should probably go this way...

Michael-Luke Jones

2007-05-28 13:47:54

by Pekka Enberg

[permalink] [raw]
Subject: Re: [RFC] LZO de/compression support - take 5

On 5/28/07, Michael-Luke Jones <[email protected]> wrote:
> No, LZO as a visible option makes no practical sense. Why would
> anyone want to build LZO into a kernel when there are no in-kernel
> users of the code?

Agreed. It should be a auto-selected hidden config option.