lib/lzo: implement run-length encoding

Patch series "lib/lzo: run-length encoding support", v5.

Following on from the previous lzo-rle patchset:

  https://lkml.org/lkml/2018/11/30/972

This patchset contains only the RLE patches, and should be applied on
top of the non-RLE patches ( https://lkml.org/lkml/2019/2/5/366 ).

Previously, some questions were raised around the RLE patches.  I've
done some additional benchmarking to answer these questions.  In short:

 - RLE offers significant additional performance (data-dependent)

 - I didn't measure any regressions that were clearly outside the noise

One concern with this patchset was around performance - specifically,
measuring RLE impact separately from Matt Sealey's patches (CTZ & fast
copy).  I have done some additional benchmarking which I hope clarifies
the benefits of each part of the patchset.

Firstly, I've captured some memory via /dev/fmem from a Chromebook with
many tabs open which is starting to swap, and then split this into 4178
4k pages.  I've excluded the all-zero pages (as zram does), and also the
no-zero pages (which won't tell us anything about RLE performance).
This should give a realistic test dataset for zram.  What I found was
that the data is VERY bimodal: 44% of pages in this dataset contain 5%
or fewer zeros, and 44% contain over 90% zeros (30% if you include the
no-zero pages).  This supports the idea of special-casing zeros in zram.

Next, I've benchmarked four variants of lzo on these pages (on 64-bit
Arm at max frequency): baseline LZO; baseline + Matt Sealey's patches
(aka MS); baseline + RLE only; baseline + MS + RLE.  Numbers are for
weighted roundtrip throughput (the weighting reflects that zram does
more compression than decompression).

  https://drive.google.com/file/d/1VLtLjRVxgUNuWFOxaGPwJYhl_hMQXpHe/view?usp=sharing

Matt's patches help in all cases for Arm (and no effect on Intel), as
expected.

RLE also behaves as expected: with few zeros present, it makes no
difference; above ~75%, it gives a good improvement (50 - 300 MB/s on
top of the benefit from Matt's patches).

Best performance is seen with both MS and RLE patches.

Finally, I have benchmarked the same dataset on an x86-64 device.  Here,
the MS patches make no difference (as expected); RLE helps, similarly as
on Arm.  There were no definite regressions; allowing for observational
error, 0.1% (3/4178) of cases had a regression > 1 standard deviation,
of which the largest was 4.6% (1.2 standard deviations).  I think this
is probably within the noise.

  https://drive.google.com/file/d/1xCUVwmiGD0heEMx5gcVEmLBI4eLaageV/view?usp=sharing

One point to note is that the graphs show RLE appears to help very
slightly with no zeros present! This is because the extra code causes
the clang optimiser to change code layout in a way that happens to have
a significant benefit.  Taking baseline LZO and adding a do-nothing line
like "__builtin_prefetch(out_len);" immediately before the "goto next"
has the same effect.  So this is a real, but basically spurious effect -
it's small enough not to upset the overall findings.

This patch (of 3):

When using zram, we frequently encounter long runs of zero bytes.  This
adds a special case which identifies runs of zeros and encodes them
using run-length encoding.

This is faster for both compression and decompresion.  For high-entropy
data which doesn't hit this case, impact is minimal.

Compression ratio is within a few percent in all cases.

This modifies the bitstream in a way which is backwards compatible
(i.e., we can decompress old bitstreams, but old versions of lzo cannot
decompress new bitstreams).

Link: http://lkml.kernel.org/r/20190205155944.16007-2-dave.rodgman@arm.com
Signed-off-by: Dave Rodgman <dave.rodgman@arm.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Markus F.X.J. Oberhumer <markus@oberhumer.com>
Cc: Matt Sealey <matt.sealey@arm.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Nitin Gupta <nitingupta910@gmail.com>
Cc: Richard Purdie <rpurdie@openedhand.com>
Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
Cc: Sonny Rao <sonnyrao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 28 insertions, 7 deletions

diff --git a/Documentation/lzo.txt b/Documentation/lzo.txt
index 6fa6a93d0949..306c60344ca7 100644
--- a/Documentation/lzo.txt
+++ b/Documentation/lzo.txt
@@ -78,16 +78,30 @@ Description
      is an implementation design choice independent on the algorithm or
      encoding.
 
+Versions
+
+0: Original version
+1: LZO-RLE
+
+Version 1 of LZO implements an extension to encode runs of zeros using run
+length encoding. This improves speed for data with many zeros, which is a
+common case for zram. This modifies the bitstream in a backwards compatible way
+(v1 can correctly decompress v0 compressed data, but v0 cannot read v1 data).
+
 Byte sequences
 ==============
 
   First byte encoding::
 
-      0..17   : follow regular instruction encoding, see below. It is worth
-                noting that codes 16 and 17 will represent a block copy from
-                the dictionary which is empty, and that they will always be
+      0..16   : follow regular instruction encoding, see below. It is worth
+                noting that code 16 will represent a block copy from the
+                dictionary which is empty, and that it will always be
                 invalid at this place.
 
+      17      : bitstream version. If the first byte is 17, the next byte
+                gives the bitstream version. If the first byte is not 17,
+                the bitstream version is 0.
+
       18..21  : copy 0..3 literals
                 state = (byte - 17) = 0..3  [ copy <state> literals ]
                 skip byte
@@ -140,6 +154,11 @@ Byte sequences
            state = S (copy S literals after this block)
            End of stream is reached if distance == 16384
 
+        In version 1, this instruction is also used to encode a run of zeros if
+        distance = 0xbfff, i.e. H = 1 and the D bits are all 1.
+           In this case, it is followed by a fourth byte, X.
+           run length = ((X << 3) | (0 0 0 0 0 L L L)) + 4.
+
       0 0 1 L L L L L  (32..63)
            Copy of small block within 16kB distance (preferably less than 34B)
            length = 2 + (L ?: 31 + (zero_bytes * 255) + non_zero_byte)
@@ -165,7 +184,9 @@ Authors
 =======
 
   This document was written by Willy Tarreau <w@1wt.eu> on 2014/07/19 during an
-  analysis of the decompression code available in Linux 3.16-rc5. The code is
-  tricky, it is possible that this document contains mistakes or that a few
-  corner cases were overlooked. In any case, please report any doubt, fix, or
-  proposed updates to the author(s) so that the document can be updated.
+  analysis of the decompression code available in Linux 3.16-rc5, and updated
+  by Dave Rodgman <dave.rodgman@arm.com> on 2018/10/30 to introduce run-length
+  encoding. The code is tricky, it is possible that this document contains
+  mistakes or that a few corner cases were overlooked. In any case, please
+  report any doubt, fix, or proposed updates to the author(s) so that the
+  document can be updated.