Fix an issue that caused premature fallback to software

Backport the first patch from https://github.com/libnxz/power-gzip/pull/174.

libnxz.spec

@@ -4,7 +4,7 @@
 
 Name:		libnxz
 Version:	0.63
-Release:	3%{?dist}
+Release:	4%{?dist}
 Summary:	Zlib implementation for POWER processors
 License:	ASL 2.0 or GPLv2+
 Url:		https://github.com/libnxz/power-gzip
@@ -15,6 +15,8 @@ Source0:	%{url}/archive/v%{version}/%{name}-%{version}.tar.gz
 Patch0: pr150.patch
 # https://github.com/libnxz/power-gzip/pull/155
 Patch1: pr155.patch
+# https://github.com/libnxz/power-gzip/pull/174
+Patch2: pr174.patch
 
 # Be explicit about the soname in order to avoid unintentional changes.
 %global soname libnxz.so.0
@@ -85,6 +87,9 @@ fi
 %{_libdir}/libnxz.la
 
 %changelog
+* Mon Jun 20 2022 Tulio Magno Quites Machado Filho <tuliom@linux.ibm.com> - 0.63-4
+- Fix an issue that caused premature fallback to software.
+
 * Tue Jun 14 2022 Jakub Čajka <jcajka@redhat.com> - 0.63-3
 - Move udev rule to the powerpc-utils-core
 

pr174.patch

@@ -0,0 +1,105 @@
+From f8b15699d654c8ace6819287dcdd3c8e493b7f6c Mon Sep 17 00:00:00 2001
+From: Tulio Magno Quites Machado Filho <tuliom@linux.ibm.com>
+Date: Fri, 3 Jun 2022 13:47:34 -0300
+Subject: [PATCH] Reduce the scope for SW usage on length threshold
+
+When compressing/decompressing large amounts of data, the first call to
+deflate()/inflate() may not have data enough to pass the threshold,
+causing libnxz to fallback to software prematurely.
+We can only take that decision if the software indicates that all the
+input has been made available, i.e. when flush == Z_FINISH.
+
+Signed-off-by: Tulio Magno Quites Machado Filho <tuliom@linux.ibm.com>
+---
+ lib/nx_deflate.c |  3 ++-
+ lib/nx_inflate.c |  3 ++-
+ lib/nx_zlib.h    | 27 +++++++++++++++++++--------
+ 3 files changed, 23 insertions(+), 10 deletions(-)
+
+diff --git a/lib/nx_deflate.c b/lib/nx_deflate.c
+index f2d5222..05b3580 100644
+--- a/lib/nx_deflate.c
++++ b/lib/nx_deflate.c
+@@ -1534,7 +1534,8 @@ int nx_deflate(z_streamp strm, int flush)
+ 		return Z_STREAM_ERROR;
+ 
+ 	/* check for sw deflate first */
+-	if( (has_nx_state(strm)) && s->switchable && (0 == use_nx_deflate(strm))){
++	if (has_nx_state(strm) && s->switchable
++	    && (0 == use_nx_deflate(strm, flush))) {
+ 		/* Use software zlib, switch the sw and hw state */
+ 		s = (nx_streamp) strm->state;
+ 		s->switchable = 0; /* decided to use sw zlib and not switchable */
+diff --git a/lib/nx_inflate.c b/lib/nx_inflate.c
+index 6dac0ad..d59d5db 100644
+--- a/lib/nx_inflate.c
++++ b/lib/nx_inflate.c
+@@ -289,7 +289,8 @@ int nx_inflate(z_streamp strm, int flush)
+ 	if (s == NULL) return Z_STREAM_ERROR;
+ 
+ 	/* check for sw deflate first*/
+-	if(has_nx_state(strm) && s->switchable && (0 == use_nx_inflate(strm))){
++	if (has_nx_state(strm) && s->switchable
++	    && (0 == use_nx_inflate(strm, flush))) {
+ 		/*Use software zlib, switch the sw and hw state*/
+ 		s = (nx_streamp) strm->state;
+ 		s->switchable = 0; /* decided to use sw zlib and not switchable */
+diff --git a/lib/nx_zlib.h b/lib/nx_zlib.h
+index 7550c65..2c0bdb5 100644
+--- a/lib/nx_zlib.h
++++ b/lib/nx_zlib.h
+@@ -341,7 +341,7 @@ static inline int has_nx_state(z_streamp strm)
+ 	return (nx_state->magic1 == MAGIC1);
+ }
+ 
+-static inline int use_nx_inflate(z_streamp strm)
++static inline int use_nx_inflate(z_streamp strm, int flush)
+ {
+ 	uint64_t rnd;
+ 	assert(strm != NULL);
+@@ -349,8 +349,13 @@ static inline int use_nx_inflate(z_streamp strm)
+ 	if(nx_config.mode.inflate == GZIP_NX) return 1;
+ 	if(nx_config.mode.inflate == GZIP_SW) return 0;
+ 
+-	/* #1 Threshold */
+-	if(strm->avail_in <= DECOMPRESS_THRESHOLD) return 0;
++	/* #2 Length threshold
++	   Even when decompressing a large amount of data, the first call to
++	   inflate() may not have enough input. So, avoid switching to software
++	   decompression prematurely unless there is a guarantee that all the
++	   input has been provided, i.e. when using Z_FINISH. */
++	if(flush == Z_FINISH && strm->avail_in <= DECOMPRESS_THRESHOLD)
++		return 0;
+ 
+ 	if(nx_config.mode.inflate == GZIP_AUTO) return 1;
+ 
+@@ -363,15 +368,21 @@ static inline int use_nx_inflate(z_streamp strm)
+ 	}
+ }
+ 
+-static inline int use_nx_deflate(z_streamp strm)
++static inline int use_nx_deflate(z_streamp strm, int flush)
+ {
+ 	assert(strm != NULL);
+ 
+-        if(nx_config.mode.deflate == GZIP_NX) return 1;
+-        if(nx_config.mode.deflate == GZIP_SW) return 0;
++	if(nx_config.mode.deflate == GZIP_NX) return 1;
++	if(nx_config.mode.deflate == GZIP_SW) return 0;
++
++	/* #2 Length threshold
++	   Even when compressing a large amount of data, the first call to
++	   deflate() may not have enough input. So, avoid switching to software
++	   compression prematurely unless there is a guarantee that all the
++	   input has been provided, i.e. when using Z_FINISH. */
++	if(flush == Z_FINISH && strm->avail_in <= COMPRESS_THRESHOLD)
++		return 0;
+ 
+-	/* #1 Threshold */
+-	if(strm->avail_in <= COMPRESS_THRESHOLD) return 0;
+ 	return 1;
+ }
+ 
+-- 
+2.35.3
+