From dc42add0cf96183e56b3777dde0a1d0500c0884e Mon Sep 17 00:00:00 2001
From: Con Kolivas <kernel@kolivas.org>
Date: Sat, 4 Dec 2010 08:39:52 +1100
Subject: [PATCH] Since 32 bit seems to fall over so easily, get ruthless with
 limits on it, while still maintaining a large window with sliding mmap.

---
 rzip.c   | 20 ++++++++++----------
 stream.c | 13 +++++++++----
 2 files changed, 19 insertions(+), 14 deletions(-)

diff --git a/rzip.c b/rzip.c
index 6e22756..e3281a1 100644
--- a/rzip.c
+++ b/rzip.c
@@ -709,10 +709,7 @@ static void rzip_chunk(struct rzip_state *st, int fd_in, int fd_out, i64 offset,
 {
 	init_sliding_mmap(st, fd_in, offset);
 
-	if (MAXRAM)
-		st->ss = open_stream_out(fd_out, NUM_STREAMS, st->chunk_size);
-	else
-		st->ss = open_stream_out(fd_out, NUM_STREAMS, st->mmap_size);
+	st->ss = open_stream_out(fd_out, NUM_STREAMS, st->chunk_size);
 	if (unlikely(!st->ss))
 		fatal("Failed to open streams in rzip_chunk\n");
 
@@ -770,13 +767,16 @@ void rzip_fd(int fd_in, int fd_out)
 	} else
 		control.st_size = 0;
 
-	/* Optimal use of ram involves no more than 2/3 of it, so if we
-	 * expressly request more with -M or -U, use a sliding mmap */
-	control.max_mmap = control.ramsize / 3 * 2;
-	/* On 32 bits we can have a big window with sliding mmap, but can
-	 * not enable much per mmap/malloc with 3 big buffers required*/
+	/* Optimal use of ram involves no more than 2/3 of it, but 32 bit
+	 * kernels struggle to dish out enough ram */
 	if (BITS32)
-		control.max_mmap = MIN(control.max_mmap, two_gig / 3);
+		control.max_mmap = control.ramsize / 6;
+	else
+		control.max_mmap = control.ramsize / 3 * 2;
+	/* On 32 bits we can have a big window with sliding mmap, but can
+	 * not enable much per mmap/malloc */
+	if (BITS32)
+		control.max_mmap = MIN(control.max_mmap, two_gig / 2);
 	round_to_page(&control.max_mmap);
 
 	/* Set maximum chunk size to proportion of ram according to mode */
diff --git a/stream.c b/stream.c
index 2adbb83..de408d2 100644
--- a/stream.c
+++ b/stream.c
@@ -650,6 +650,7 @@ void *open_stream_out(int f, int n, i64 limit)
 {
 	struct stream_info *sinfo;
 	uchar *testmalloc;
+	i64 testsize;
 	int i;
 
 	sinfo = malloc(sizeof(*sinfo));
@@ -689,7 +690,7 @@ void *open_stream_out(int f, int n, i64 limit)
 
 	/* Serious limits imposed on 32 bit capabilities */
 	if (BITS32)
-		limit = MIN(limit, two_gig / 3);
+		limit = MIN(limit, two_gig / 6);
 
 	sinfo->initial_pos = lseek(f, 0, SEEK_CUR);
 
@@ -703,13 +704,17 @@ void *open_stream_out(int f, int n, i64 limit)
 	 * ram. We need enough for the 2 streams and for the compression
 	 * backend at most, being conservative. */
 retest_malloc:
-	testmalloc = malloc(limit * (n + 1));
+	if (BITS32)
+		testsize = limit * n * 3;
+	else
+		testsize = limit * (n + 1);
+	testmalloc = malloc(testsize);
 	if (!testmalloc) {
 		limit = limit / 10 * 9;
 		goto retest_malloc;
 	}
 	free(testmalloc);
-	print_maxverbose("Succeeded in testing %lld sized malloc for back end compression\n", limit * (n + 1));
+	print_maxverbose("Succeeded in testing %lld sized malloc for back end compression\n", testsize);
 
 	sinfo->bufsize = limit;
 
@@ -765,7 +770,7 @@ void *open_stream_in(int f, int n)
 
 	ucthread = calloc(sizeof(struct uncomp_thread), total_threads);
 	if (unlikely(!ucthread))
-		fatal("Unable to calloc cthread in open_stream_out\n");
+		fatal("Unable to calloc cthread in open_stream_in\n");
 
 	for (i = 0; i < total_threads; i++) {
 		init_sem(&ucthread[i].complete);