diff mbox series

[V3,4/5] selftests/powerpc: Add NX-GZIP engine decompress testcase

Message ID 20200413155916.16900-5-rzinsly@linux.ibm.com
State Superseded
Headers show
Series [V3,1/5] selftests/powerpc: Add header files for GZIP engine test | expand

Commit Message

Raphael Moreira Zinsly April 13, 2020, 3:59 p.m. UTC
Include a decompression testcase for the powerpc NX-GZIP
engine.

Signed-off-by: Bulent Abali <abali@us.ibm.com>
Signed-off-by: Raphael Moreira Zinsly <rzinsly@linux.ibm.com>
---
 .../selftests/powerpc/nx-gzip/Makefile        |    7 +-
 .../selftests/powerpc/nx-gzip/gunz_test.c     | 1026 +++++++++++++++++
 2 files changed, 1030 insertions(+), 3 deletions(-)
 create mode 100644 tools/testing/selftests/powerpc/nx-gzip/gunz_test.c
diff mbox series

Patch

diff --git a/tools/testing/selftests/powerpc/nx-gzip/Makefile b/tools/testing/selftests/powerpc/nx-gzip/Makefile
index ab903f63bbbd..82abc19a49a0 100644
--- a/tools/testing/selftests/powerpc/nx-gzip/Makefile
+++ b/tools/testing/selftests/powerpc/nx-gzip/Makefile
@@ -1,9 +1,9 @@ 
 CC = gcc
 CFLAGS = -O3
 INC = ./inc
-SRC = gzfht_test.c
+SRC = gzfht_test.c gunz_test.c
 OBJ = $(SRC:.c=.o)
-TESTS = gzfht_test
+TESTS = gzfht_test gunz_test
 EXTRA_SOURCES = gzip_vas.c
 
 all:	$(TESTS)
@@ -16,6 +16,7 @@  $(TESTS): $(OBJ)
 
 run_tests: $(TESTS)
 	./gzfht_test gzip_vas.c
+	./gunz_test gzip_vas.c.nx.gz
 
 clean:
-	rm -f $(TESTS) *.o *~ *.gz
+	rm -f $(TESTS) *.o *~ *.gz *.gunzip
diff --git a/tools/testing/selftests/powerpc/nx-gzip/gunz_test.c b/tools/testing/selftests/powerpc/nx-gzip/gunz_test.c
new file mode 100644
index 000000000000..94cb79616225
--- /dev/null
+++ b/tools/testing/selftests/powerpc/nx-gzip/gunz_test.c
@@ -0,0 +1,1026 @@ 
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+/* P9 gunzip sample code for demonstrating the P9 NX hardware
+ * interface.  Not intended for productive uses or for performance or
+ * compression ratio measurements.  Note also that /dev/crypto/gzip,
+ * VAS and skiboot support are required
+ *
+ * Copyright 2020 IBM Corp.
+ *
+ * Author: Bulent Abali <abali@us.ibm.com>
+ *
+ * https://github.com/libnxz/power-gzip for zlib api and other utils
+ * Definitions of acronyms used here.  See
+ * P9 NX Gzip Accelerator User's Manual for details:
+ * https://github.com/libnxz/power-gzip/blob/develop/doc/power_nx_gzip_um.pdf
+ *
+ * adler/crc: 32 bit checksums appended to stream tail
+ * ce:       completion extension
+ * cpb:      coprocessor parameter block (metadata)
+ * crb:      coprocessor request block (command)
+ * csb:      coprocessor status block (status)
+ * dht:      dynamic huffman table
+ * dde:      data descriptor element (address, length)
+ * ddl:      list of ddes
+ * dh/fh:    dynamic and fixed huffman types
+ * fc:       coprocessor function code
+ * histlen:  history/dictionary length
+ * history:  sliding window of up to 32KB of data
+ * lzcount:  Deflate LZ symbol counts
+ * rembytecnt: remaining byte count
+ * sfbt:     source final block type; last block's type during decomp
+ * spbc:     source processed byte count
+ * subc:     source unprocessed bit count
+ * tebc:     target ending bit count; valid bits in the last byte
+ * tpbc:     target processed byte count
+ * vas:      virtual accelerator switch; the user mode interface
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <stdint.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/time.h>
+#include <sys/fcntl.h>
+#include <sys/mman.h>
+#include <endian.h>
+#include <bits/endian.h>
+#include <sys/ioctl.h>
+#include <assert.h>
+#include <errno.h>
+#include <signal.h>
+#include "nxu.h"
+#include "nx.h"
+#include "crb.h"
+
+int nx_dbg;
+FILE *nx_gzip_log;
+
+#define NX_MIN(X, Y) (((X) < (Y))?(X):(Y))
+#define NX_MAX(X, Y) (((X) > (Y))?(X):(Y))
+
+#define GETINPC(X) fgetc(X)
+#define FNAME_MAX 1024
+
+/* fifo queue management */
+#define fifo_used_bytes(used) (used)
+#define fifo_free_bytes(used, len) ((len)-(used))
+/* amount of free bytes in the first and last parts */
+#define fifo_free_first_bytes(cur, used, len)  ((((cur)+(used)) <= (len)) \
+						  ? (len)-((cur)+(used)) : 0)
+#define fifo_free_last_bytes(cur, used, len)   ((((cur)+(used)) <= (len)) \
+						  ? (cur) : (len)-(used))
+/* amount of used bytes in the first and last parts */
+#define fifo_used_first_bytes(cur, used, len)  ((((cur)+(used)) <= (len)) \
+						  ? (used) : (len)-(cur))
+#define fifo_used_last_bytes(cur, used, len)   ((((cur)+(used)) <= (len)) \
+						  ? 0 : ((used)+(cur))-(len))
+/* first and last free parts start here */
+#define fifo_free_first_offset(cur, used)      ((cur)+(used))
+#define fifo_free_last_offset(cur, used, len)  \
+					   fifo_used_last_bytes(cur, used, len)
+/* first and last used parts start here */
+#define fifo_used_first_offset(cur)            (cur)
+#define fifo_used_last_offset(cur)             (0)
+
+const int fifo_in_len = 1<<24;
+const int fifo_out_len = 1<<24;
+const int page_sz = 1<<16;
+const int line_sz = 1<<7;
+const int window_max = 1<<15;
+
+/*
+ * Adds an (address, len) pair to the list of ddes (ddl) and updates
+ * the base dde.  ddl[0] is the only dde in a direct dde which
+ * contains a single (addr,len) pair.  For more pairs, ddl[0] becomes
+ * the indirect (base) dde that points to a list of direct ddes.
+ * See Section 6.4 of the NX-gzip user manual for DDE description.
+ * Addr=NULL, len=0 clears the ddl[0].  Returns the total number of
+ * bytes in ddl.  Caller is responsible for allocting the array of
+ * nx_dde_t *ddl.  If N addresses are required in the scatter-gather
+ * list, the ddl array must have N+1 entries minimum.
+ */
+static inline uint32_t nx_append_dde(struct nx_dde_t *ddl, void *addr,
+					uint32_t len)
+{
+	uint32_t ddecnt;
+	uint32_t bytes;
+
+	if (addr == NULL && len == 0) {
+		clearp_dde(ddl);
+		return 0;
+	}
+
+	NXPRT(fprintf(stderr, "%d: %s addr %p len %x\n", __LINE__, addr,
+			__func__, len));
+
+	/* Number of ddes in the dde list ; == 0 when it is a direct dde */
+	ddecnt = getpnn(ddl, dde_count);
+	bytes = getp32(ddl, ddebc);
+
+	if (ddecnt == 0 && bytes == 0) {
+		/* First dde is unused; make it a direct dde */
+		bytes = len;
+		putp32(ddl, ddebc, bytes);
+		putp64(ddl, ddead, (uint64_t) addr);
+	} else if (ddecnt == 0) {
+		/* Converting direct to indirect dde
+		 * ddl[0] becomes head dde of ddl
+		 * copy direct to indirect first.
+		 */
+		ddl[1] = ddl[0];
+
+		/* Add the new dde next */
+		clear_dde(ddl[2]);
+		put32(ddl[2], ddebc, len);
+		put64(ddl[2], ddead, (uint64_t) addr);
+
+		/* Ddl head points to 2 direct ddes */
+		ddecnt = 2;
+		putpnn(ddl, dde_count, ddecnt);
+		bytes = bytes + len;
+		putp32(ddl, ddebc, bytes);
+		/* Pointer to the first direct dde */
+		putp64(ddl, ddead, (uint64_t) &ddl[1]);
+	} else {
+		/* Append a dde to an existing indirect ddl */
+		++ddecnt;
+		clear_dde(ddl[ddecnt]);
+		put64(ddl[ddecnt], ddead, (uint64_t) addr);
+		put32(ddl[ddecnt], ddebc, len);
+
+		putpnn(ddl, dde_count, ddecnt);
+		bytes = bytes + len;
+		putp32(ddl, ddebc, bytes); /* byte sum of all dde */
+	}
+	return bytes;
+}
+
+/*
+ * Touch specified number of pages represented in number bytes
+ * beginning from the first buffer in a dde list.
+ * Do not touch the pages past buf_sz-th byte's page.
+ *
+ * Set buf_sz = 0 to touch all pages described by the ddep.
+ */
+static int nx_touch_pages_dde(struct nx_dde_t *ddep, long buf_sz, long page_sz,
+				int wr)
+{
+	uint32_t indirect_count;
+	uint32_t buf_len;
+	long total;
+	uint64_t buf_addr;
+	struct nx_dde_t *dde_list;
+	int i;
+
+	assert(!!ddep);
+
+	indirect_count = getpnn(ddep, dde_count);
+
+	NXPRT(fprintf(stderr, "%s dde_count %d request len ", __func__,
+			indirect_count));
+	NXPRT(fprintf(stderr, "0x%lx\n", buf_sz));
+
+	if (indirect_count == 0) {
+		/* Direct dde */
+		buf_len = getp32(ddep, ddebc);
+		buf_addr = getp64(ddep, ddead);
+
+		NXPRT(fprintf(stderr, "touch direct ddebc 0x%x ddead %p\n",
+				buf_len, (void *)buf_addr));
+
+		if (buf_sz == 0)
+			nxu_touch_pages((void *)buf_addr, buf_len, page_sz, wr);
+		else
+			nxu_touch_pages((void *)buf_addr, NX_MIN(buf_len,
+					buf_sz), page_sz, wr);
+
+		return ERR_NX_OK;
+	}
+
+	/* Indirect dde */
+	if (indirect_count > MAX_DDE_COUNT)
+		return ERR_NX_EXCESSIVE_DDE;
+
+	/* First address of the list */
+	dde_list = (struct nx_dde_t *) getp64(ddep, ddead);
+
+	if (buf_sz == 0)
+		buf_sz = getp32(ddep, ddebc);
+
+	total = 0;
+	for (i = 0; i < indirect_count; i++) {
+		buf_len = get32(dde_list[i], ddebc);
+		buf_addr = get64(dde_list[i], ddead);
+		total += buf_len;
+
+		NXPRT(fprintf(stderr, "touch loop len 0x%x ddead %p total ",
+				buf_len, (void *)buf_addr));
+		NXPRT(fprintf(stderr, "0x%lx\n", total));
+
+		/* Touching fewer pages than encoded in the ddebc */
+		if (total > buf_sz) {
+			buf_len = NX_MIN(buf_len, total - buf_sz);
+			nxu_touch_pages((void *)buf_addr, buf_len, page_sz, wr);
+			NXPRT(fprintf(stderr, "touch loop break len 0x%x ",
+				      buf_len));
+			NXPRT(fprintf(stderr, "ddead %p\n", (void *)buf_addr));
+			break;
+		}
+		nxu_touch_pages((void *)buf_addr, buf_len, page_sz, wr);
+	}
+	return ERR_NX_OK;
+}
+
+/*
+ * Src and dst buffers are supplied in scatter gather lists.
+ * NX function code and other parameters supplied in cmdp.
+ */
+static int nx_submit_job(struct nx_dde_t *src, struct nx_dde_t *dst,
+			 struct nx_gzip_crb_cpb_t *cmdp, void *handle)
+{
+	int cc;
+	uint64_t csbaddr;
+
+	memset((void *)&cmdp->crb.csb, 0, sizeof(cmdp->crb.csb));
+
+	cmdp->crb.source_dde = *src;
+	cmdp->crb.target_dde = *dst;
+
+	/* Status, output byte count in tpbc */
+	csbaddr = ((uint64_t) &cmdp->crb.csb) & csb_address_mask;
+	put64(cmdp->crb, csb_address, csbaddr);
+
+	/* NX reports input bytes in spbc; cleared */
+	cmdp->cpb.out_spbc_comp_wrap = 0;
+	cmdp->cpb.out_spbc_comp_with_count = 0;
+	cmdp->cpb.out_spbc_decomp = 0;
+
+	/* Clear output */
+	put32(cmdp->cpb, out_crc, INIT_CRC);
+	put32(cmdp->cpb, out_adler, INIT_ADLER);
+
+	/* Submit the crb, the job descriptor, to the accelerator. */
+	return nxu_submit_job(cmdp, handle);
+}
+
+int decompress_file(int argc, char **argv, void *devhandle)
+{
+	FILE *inpf;
+	FILE *outf;
+
+	int c, expect, i, cc, rc = 0;
+	char gzfname[FNAME_MAX];
+
+	/* Queuing, file ops, byte counting */
+	char *fifo_in, *fifo_out;
+	int used_in, cur_in, used_out, cur_out, read_sz, n;
+	int first_free, last_free, first_used, last_used;
+	int first_offset, last_offset;
+	int write_sz, free_space, source_sz;
+	int source_sz_estimate, target_sz_estimate;
+	uint64_t last_comp_ratio; /* 1000 max */
+	uint64_t total_out;
+	int is_final, is_eof;
+
+	/* nx hardware */
+	int sfbt, subc, spbc, tpbc, nx_ce, fc, resuming = 0;
+	int history_len = 0;
+	struct nx_gzip_crb_cpb_t cmd, *cmdp;
+	struct nx_dde_t *ddl_in;
+	struct nx_dde_t dde_in[6] __aligned(128);
+	struct nx_dde_t *ddl_out;
+	struct nx_dde_t dde_out[6] __aligned(128);
+	int pgfault_retries;
+
+	/* when using mmap'ed files */
+	off_t input_file_offset;
+
+	if (argc > 2) {
+		fprintf(stderr, "usage: %s <fname> or stdin\n", argv[0]);
+		fprintf(stderr, "    writes to stdout or <fname>.nx.gunzip\n");
+		return -1;
+	}
+
+	if (argc == 1) {
+		inpf = stdin;
+		outf = stdout;
+	} else if (argc == 2) {
+		char w[1024];
+		char *wp;
+
+		inpf = fopen(argv[1], "r");
+		if (inpf == NULL) {
+			perror(argv[1]);
+			return -1;
+		}
+
+		/* Make a new file name to write to.  Ignoring '.gz' */
+		wp = (NULL != (wp = strrchr(argv[1], '/'))) ? (wp+1) : argv[1];
+		strcpy(w, wp);
+		strcat(w, ".nx.gunzip");
+
+		outf = fopen(w, "w");
+		if (outf == NULL) {
+			perror(w);
+			return -1;
+		}
+	}
+
+	/* Decode the gzip header */
+	c = GETINPC(inpf); expect = 0x1f; /* ID1 */
+	if (c != expect)
+		goto err1;
+
+	c = GETINPC(inpf); expect = 0x8b; /* ID2 */
+	if (c != expect)
+		goto err1;
+
+	c = GETINPC(inpf); expect = 0x08; /* CM */
+	if (c != expect)
+		goto err1;
+
+	int flg = GETINPC(inpf); /* FLG */
+
+	if (flg & 0xE0 || flg & 0x4 || flg == EOF)
+		goto err2;
+
+	fprintf(stderr, "gzHeader FLG %x\n", flg);
+
+	/* Read 6 bytes; ignoring the MTIME, XFL, OS fields in this
+	 * sample code.
+	 */
+	for (i = 0; i < 6; i++) {
+		char tmp[10];
+
+		tmp[i] = GETINPC(inpf);
+		if (tmp[i] == EOF)
+			goto err3;
+		fprintf(stderr, "%02x ", tmp[i]);
+		if (i == 5)
+			fprintf(stderr, "\n");
+	}
+	fprintf(stderr, "gzHeader MTIME, XFL, OS ignored\n");
+
+	/* FNAME */
+	if (flg & 0x8) {
+		int k = 0;
+
+		do {
+			c = GETINPC(inpf);
+			if (c == EOF || k >= FNAME_MAX)
+				goto err3;
+			gzfname[k++] = c;
+		} while (c);
+		fprintf(stderr, "gzHeader FNAME: %s\n", gzfname);
+	}
+
+	/* FHCRC */
+	if (flg & 0x2) {
+		c = GETINPC(inpf);
+		if (c == EOF)
+			goto err3;
+		c = GETINPC(inpf);
+		if (c == EOF)
+			goto err3;
+		fprintf(stderr, "gzHeader FHCRC: ignored\n");
+	}
+
+	used_in = cur_in = used_out = cur_out = 0;
+	is_final = is_eof = 0;
+
+	/* Allocate one page larger to prevent page faults due to NX
+	 * overfetching.
+	 * Either do this (char*)(uintptr_t)aligned_alloc or use
+	 * -std=c11 flag to make the int-to-pointer warning go away.
+	 */
+	assert((fifo_in  = (char *)(uintptr_t)aligned_alloc(line_sz,
+				   fifo_in_len + page_sz)) != NULL);
+	assert((fifo_out = (char *)(uintptr_t)aligned_alloc(line_sz,
+				   fifo_out_len + page_sz + line_sz)) != NULL);
+	/* Leave unused space due to history rounding rules */
+	fifo_out = fifo_out + line_sz;
+	nxu_touch_pages(fifo_out, fifo_out_len, page_sz, 1);
+
+	ddl_in  = &dde_in[0];
+	ddl_out = &dde_out[0];
+	cmdp = &cmd;
+	memset(&cmdp->crb, 0, sizeof(cmdp->crb));
+
+read_state:
+
+	/* Read from .gz file */
+
+	NXPRT(fprintf(stderr, "read_state:\n"));
+
+	if (is_eof != 0)
+		goto write_state;
+
+	/* We read in to fifo_in in two steps: first: read in to from
+	 * cur_in to the end of the buffer.  last: if free space wrapped
+	 * around, read from fifo_in offset 0 to offset cur_in.
+	 */
+
+	/* Reset fifo head to reduce unnecessary wrap arounds */
+	cur_in = (used_in == 0) ? 0 : cur_in;
+
+	/* Free space total is reduced by a gap */
+	free_space = NX_MAX(0, fifo_free_bytes(used_in, fifo_in_len)
+			    - line_sz);
+
+	/* Free space may wrap around as first and last */
+	first_free = fifo_free_first_bytes(cur_in, used_in, fifo_in_len);
+	last_free  = fifo_free_last_bytes(cur_in, used_in, fifo_in_len);
+
+	/* Start offsets of the free memory */
+	first_offset = fifo_free_first_offset(cur_in, used_in);
+	last_offset  = fifo_free_last_offset(cur_in, used_in, fifo_in_len);
+
+	/* Reduce read_sz because of the line_sz gap */
+	read_sz = NX_MIN(free_space, first_free);
+	n = 0;
+	if (read_sz > 0) {
+		/* Read in to offset cur_in + used_in */
+		n = fread(fifo_in + first_offset, 1, read_sz, inpf);
+		used_in = used_in + n;
+		free_space = free_space - n;
+		assert(n <= read_sz);
+		if (n != read_sz) {
+			/* Either EOF or error; exit the read loop */
+			is_eof = 1;
+			goto write_state;
+		}
+	}
+
+	/* If free space wrapped around */
+	if (last_free > 0) {
+		/* Reduce read_sz because of the line_sz gap */
+		read_sz = NX_MIN(free_space, last_free);
+		n = 0;
+		if (read_sz > 0) {
+			n = fread(fifo_in + last_offset, 1, read_sz, inpf);
+			used_in = used_in + n;       /* Increase used space */
+			free_space = free_space - n; /* Decrease free space */
+			assert(n <= read_sz);
+			if (n != read_sz) {
+				/* Either EOF or error; exit the read loop */
+				is_eof = 1;
+				goto write_state;
+			}
+		}
+	}
+
+	/* At this point we have used_in bytes in fifo_in with the
+	 * data head starting at cur_in and possibly wrapping around.
+	 */
+
+write_state:
+
+	/* Write decompressed data to output file */
+
+	NXPRT(fprintf(stderr, "write_state:\n"));
+
+	if (used_out == 0)
+		goto decomp_state;
+
+	/* If fifo_out has data waiting, write it out to the file to
+	 * make free target space for the accelerator used bytes in
+	 * the first and last parts of fifo_out.
+	 */
+
+	first_used = fifo_used_first_bytes(cur_out, used_out, fifo_out_len);
+	last_used  = fifo_used_last_bytes(cur_out, used_out, fifo_out_len);
+
+	write_sz = first_used;
+
+	n = 0;
+	if (write_sz > 0) {
+		n = fwrite(fifo_out + cur_out, 1, write_sz, outf);
+		used_out = used_out - n;
+		/* Move head of the fifo */
+		cur_out = (cur_out + n) % fifo_out_len;
+		assert(n <= write_sz);
+		if (n != write_sz) {
+			fprintf(stderr, "error: write\n");
+			rc = -1;
+			goto err5;
+		}
+	}
+
+	if (last_used > 0) { /* If more data available in the last part */
+		write_sz = last_used; /* Keep it here for later */
+		n = 0;
+		if (write_sz > 0) {
+			n = fwrite(fifo_out, 1, write_sz, outf);
+			used_out = used_out - n;
+			cur_out = (cur_out + n) % fifo_out_len;
+			assert(n <= write_sz);
+			if (n != write_sz) {
+				fprintf(stderr, "error: write\n");
+				rc = -1;
+				goto err5;
+			}
+		}
+	}
+
+decomp_state:
+
+	/* NX decompresses input data */
+
+	NXPRT(fprintf(stderr, "decomp_state:\n"));
+
+	if (is_final)
+		goto finish_state;
+
+	/* Address/len lists */
+	clearp_dde(ddl_in);
+	clearp_dde(ddl_out);
+
+	/* FC, CRC, HistLen, Table 6-6 */
+	if (resuming) {
+		/* Resuming a partially decompressed input.
+		 * The key to resume is supplying the 32KB
+		 * dictionary (history) to NX, which is basically
+		 * the last 32KB of output produced.
+		 */
+		fc = GZIP_FC_DECOMPRESS_RESUME;
+
+		cmdp->cpb.in_crc   = cmdp->cpb.out_crc;
+		cmdp->cpb.in_adler = cmdp->cpb.out_adler;
+
+		/* Round up the history size to quadword.  Section 2.10 */
+		history_len = (history_len + 15) / 16;
+		putnn(cmdp->cpb, in_histlen, history_len);
+		history_len = history_len * 16; /* bytes */
+
+		if (history_len > 0) {
+			/* Chain in the history buffer to the DDE list */
+			if (cur_out >= history_len) {
+				nx_append_dde(ddl_in, fifo_out
+					      + (cur_out - history_len),
+					      history_len);
+			} else {
+				nx_append_dde(ddl_in, fifo_out
+					      + ((fifo_out_len + cur_out)
+					      - history_len),
+					      history_len - cur_out);
+				/* Up to 32KB history wraps around fifo_out */
+				nx_append_dde(ddl_in, fifo_out, cur_out);
+			}
+
+		}
+	} else {
+		/* First decompress job */
+		fc = GZIP_FC_DECOMPRESS;
+
+		history_len = 0;
+		/* Writing 0 clears out subc as well */
+		cmdp->cpb.in_histlen = 0;
+		total_out = 0;
+
+		put32(cmdp->cpb, in_crc, INIT_CRC);
+		put32(cmdp->cpb, in_adler, INIT_ADLER);
+		put32(cmdp->cpb, out_crc, INIT_CRC);
+		put32(cmdp->cpb, out_adler, INIT_ADLER);
+
+		/* Assuming 10% compression ratio initially; use the
+		 * most recently measured compression ratio as a
+		 * heuristic to estimate the input and output
+		 * sizes.  If we give too much input, the target buffer
+		 * overflows and NX cycles are wasted, and then we
+		 * must retry with smaller input size.  1000 is 100%.
+		 */
+		last_comp_ratio = 100UL;
+	}
+	cmdp->crb.gzip_fc = 0;
+	putnn(cmdp->crb, gzip_fc, fc);
+
+	/*
+	 * NX source buffers
+	 */
+	first_used = fifo_used_first_bytes(cur_in, used_in, fifo_in_len);
+	last_used = fifo_used_last_bytes(cur_in, used_in, fifo_in_len);
+
+	if (first_used > 0)
+		nx_append_dde(ddl_in, fifo_in + cur_in, first_used);
+
+	if (last_used > 0)
+		nx_append_dde(ddl_in, fifo_in, last_used);
+
+	/*
+	 * NX target buffers
+	 */
+	first_free = fifo_free_first_bytes(cur_out, used_out, fifo_out_len);
+	last_free = fifo_free_last_bytes(cur_out, used_out, fifo_out_len);
+
+	/* Reduce output free space amount not to overwrite the history */
+	int target_max = NX_MAX(0, fifo_free_bytes(used_out, fifo_out_len)
+				- (1<<16));
+
+	NXPRT(fprintf(stderr, "target_max %d (0x%x)\n", target_max,
+		      target_max));
+
+	first_free = NX_MIN(target_max, first_free);
+	if (first_free > 0) {
+		first_offset = fifo_free_first_offset(cur_out, used_out);
+		nx_append_dde(ddl_out, fifo_out + first_offset, first_free);
+	}
+
+	if (last_free > 0) {
+		last_free = NX_MIN(target_max - first_free, last_free);
+		if (last_free > 0) {
+			last_offset = fifo_free_last_offset(cur_out, used_out,
+							    fifo_out_len);
+			nx_append_dde(ddl_out, fifo_out + last_offset,
+				      last_free);
+		}
+	}
+
+	/* Target buffer size is used to limit the source data size
+	 * based on previous measurements of compression ratio.
+	 */
+
+	/* source_sz includes history */
+	source_sz = getp32(ddl_in, ddebc);
+	assert(source_sz > history_len);
+	source_sz = source_sz - history_len;
+
+	/* Estimating how much source is needed to 3/4 fill a
+	 * target_max size target buffer.  If we overshoot, then NX
+	 * must repeat the job with smaller input and we waste
+	 * bandwidth.  If we undershoot then we use more NX calls than
+	 * necessary.
+	 */
+
+	source_sz_estimate = ((uint64_t)target_max * last_comp_ratio * 3UL)
+				/ 4000;
+
+	if (source_sz_estimate < source_sz) {
+		/* Target might be small, therefore limiting the
+		 * source data.
+		 */
+		source_sz = source_sz_estimate;
+		target_sz_estimate = target_max;
+	} else {
+		/* Source file might be small, therefore limiting target
+		 * touch pages to a smaller value to save processor cycles.
+		 */
+		target_sz_estimate = ((uint64_t)source_sz * 1000UL)
+					/ (last_comp_ratio + 1);
+		target_sz_estimate = NX_MIN(2 * target_sz_estimate,
+					    target_max);
+	}
+
+	source_sz = source_sz + history_len;
+
+	/* Some NX condition codes require submitting the NX job again.
+	 * Kernel doesn't handle NX page faults. Expects user code to
+	 * touch pages.
+	 */
+	pgfault_retries = NX_MAX_FAULTS;
+
+restart_nx:
+
+	putp32(ddl_in, ddebc, source_sz);
+
+	/* Fault in pages */
+	nxu_touch_pages(cmdp, sizeof(struct nx_gzip_crb_cpb_t), page_sz, 1);
+	nx_touch_pages_dde(ddl_in, 0, page_sz, 0);
+	nx_touch_pages_dde(ddl_out, target_sz_estimate, page_sz, 1);
+
+	/* Send job to NX */
+	cc = nx_submit_job(ddl_in, ddl_out, cmdp, devhandle);
+
+	switch (cc) {
+
+	case ERR_NX_TRANSLATION:
+
+		/* We touched the pages ahead of time.  In the most common case
+		 * we shouldn't be here.  But may be some pages were paged out.
+		 * Kernel should have placed the faulting address to fsaddr.
+		 */
+		NXPRT(fprintf(stderr, "ERR_NX_TRANSLATION %p\n",
+			      (void *)cmdp->crb.csb.fsaddr));
+
+		if (pgfault_retries == NX_MAX_FAULTS) {
+			/* Try once with exact number of pages */
+			--pgfault_retries;
+			goto restart_nx;
+		} else if (pgfault_retries > 0) {
+			/* If still faulting try fewer input pages
+			 * assuming memory outage
+			 */
+			if (source_sz > page_sz)
+				source_sz = NX_MAX(source_sz / 2, page_sz);
+			--pgfault_retries;
+			goto restart_nx;
+		} else {
+			fprintf(stderr, "cannot make progress; too many ");
+			fprintf(stderr, "page fault retries cc= %d\n", cc);
+			rc = -1;
+			goto err5;
+		}
+
+	case ERR_NX_DATA_LENGTH:
+
+		NXPRT(fprintf(stderr, "ERR_NX_DATA_LENGTH; "));
+		NXPRT(fprintf(stderr, "stream may have trailing data\n"));
+
+		/* Not an error in the most common case; it just says
+		 * there is trailing data that we must examine.
+		 *
+		 * CC=3 CE(1)=0 CE(0)=1 indicates partial completion
+		 * Fig.6-7 and Table 6-8.
+		 */
+		nx_ce = get_csb_ce_ms3b(cmdp->crb.csb);
+
+		if (!csb_ce_termination(nx_ce) &&
+		    csb_ce_partial_completion(nx_ce)) {
+			/* Check CPB for more information
+			 * spbc and tpbc are valid
+			 */
+			sfbt = getnn(cmdp->cpb, out_sfbt); /* Table 6-4 */
+			subc = getnn(cmdp->cpb, out_subc); /* Table 6-4 */
+			spbc = get32(cmdp->cpb, out_spbc_decomp);
+			tpbc = get32(cmdp->crb.csb, tpbc);
+			assert(target_max >= tpbc);
+
+			goto ok_cc3; /* not an error */
+		} else {
+			/* History length error when CE(1)=1 CE(0)=0. */
+			rc = -1;
+			fprintf(stderr, "history length error cc= %d\n", cc);
+			goto err5;
+		}
+
+	case ERR_NX_TARGET_SPACE:
+
+		/* Target buffer not large enough; retry smaller input
+		 * data; give at least 1 byte.  SPBC/TPBC are not valid.
+		 */
+		assert(source_sz > history_len);
+		source_sz = ((source_sz - history_len + 2) / 2) + history_len;
+		NXPRT(fprintf(stderr, "ERR_NX_TARGET_SPACE; retry with "));
+		NXPRT(fprintf(stderr, "smaller input data src %d hist %d\n",
+			      source_sz, history_len));
+		goto restart_nx;
+
+	case ERR_NX_OK:
+
+		/* This should not happen for gzip formatted data;
+		 * we need trailing crc and isize
+		 */
+		fprintf(stderr, "ERR_NX_OK\n");
+		spbc = get32(cmdp->cpb, out_spbc_decomp);
+		tpbc = get32(cmdp->crb.csb, tpbc);
+		assert(target_max >= tpbc);
+		assert(spbc >= history_len);
+		source_sz = spbc - history_len;
+		goto offsets_state;
+
+	default:
+		fprintf(stderr, "error: cc= %d\n", cc);
+		rc = -1;
+		goto err5;
+	}
+
+ok_cc3:
+
+	NXPRT(fprintf(stderr, "cc3: sfbt: %x\n", sfbt));
+
+	assert(spbc > history_len);
+	source_sz = spbc - history_len;
+
+	/* Table 6-4: Source Final Block Type (SFBT) describes the
+	 * last processed deflate block and clues the software how to
+	 * resume the next job.  SUBC indicates how many input bits NX
+	 * consumed but did not process.  SPBC indicates how many
+	 * bytes of source were given to the accelerator including
+	 * history bytes.
+	 */
+
+	switch (sfbt) {
+		int dhtlen;
+
+	case 0x0: /* Deflate final EOB received */
+
+		/* Calculating the checksum start position. */
+
+		source_sz = source_sz - subc / 8;
+		is_final = 1;
+		break;
+
+		/* Resume decompression cases are below. Basically
+		 * indicates where NX has suspended and how to resume
+		 * the input stream.
+		 */
+
+	case 0x8: /* Within a literal block; use rembytecount */
+	case 0x9: /* Within a literal block; use rembytecount; bfinal=1 */
+
+		/* Supply the partially processed source byte again */
+		source_sz = source_sz - ((subc + 7) / 8);
+
+		/* SUBC LS 3bits: number of bits in the first source byte need
+		 * to be processed.
+		 * 000 means all 8 bits;  Table 6-3
+		 * Clear subc, histlen, sfbt, rembytecnt, dhtlen
+		 */
+		cmdp->cpb.in_subc = 0;
+		cmdp->cpb.in_sfbt = 0;
+		putnn(cmdp->cpb, in_subc, subc % 8);
+		putnn(cmdp->cpb, in_sfbt, sfbt);
+		putnn(cmdp->cpb, in_rembytecnt, getnn(cmdp->cpb,
+						      out_rembytecnt));
+		break;
+
+	case 0xA: /* Within a FH block; */
+	case 0xB: /* Within a FH block; bfinal=1 */
+
+		source_sz = source_sz - ((subc + 7) / 8);
+
+		/* Clear subc, histlen, sfbt, rembytecnt, dhtlen */
+		cmdp->cpb.in_subc = 0;
+		cmdp->cpb.in_sfbt = 0;
+		putnn(cmdp->cpb, in_subc, subc % 8);
+		putnn(cmdp->cpb, in_sfbt, sfbt);
+		break;
+
+	case 0xC: /* Within a DH block; */
+	case 0xD: /* Within a DH block; bfinal=1 */
+
+		source_sz = source_sz - ((subc + 7) / 8);
+
+		/* Clear subc, histlen, sfbt, rembytecnt, dhtlen */
+		cmdp->cpb.in_subc = 0;
+		cmdp->cpb.in_sfbt = 0;
+		putnn(cmdp->cpb, in_subc, subc % 8);
+		putnn(cmdp->cpb, in_sfbt, sfbt);
+
+		dhtlen = getnn(cmdp->cpb, out_dhtlen);
+		putnn(cmdp->cpb, in_dhtlen, dhtlen);
+		assert(dhtlen >= 42);
+
+		/* Round up to a qword */
+		dhtlen = (dhtlen + 127) / 128;
+
+		while (dhtlen > 0) { /* Copy dht from cpb.out to cpb.in */
+			--dhtlen;
+			cmdp->cpb.in_dht[dhtlen] = cmdp->cpb.out_dht[dhtlen];
+		}
+		break;
+
+	case 0xE: /* Within a block header; bfinal=0; */
+		     /* Also given if source data exactly ends (SUBC=0) with
+		      * EOB code with BFINAL=0.  Means the next byte will
+		      * contain a block header.
+		      */
+	case 0xF: /* within a block header with BFINAL=1. */
+
+		source_sz = source_sz - ((subc + 7) / 8);
+
+		/* Clear subc, histlen, sfbt, rembytecnt, dhtlen */
+		cmdp->cpb.in_subc = 0;
+		cmdp->cpb.in_sfbt = 0;
+		putnn(cmdp->cpb, in_subc, subc % 8);
+		putnn(cmdp->cpb, in_sfbt, sfbt);
+
+		/* Engine did not process any data */
+		if (is_eof && (source_sz == 0))
+			is_final = 1;
+	}
+
+offsets_state:
+
+	/* Adjust the source and target buffer offsets and lengths  */
+
+	NXPRT(fprintf(stderr, "offsets_state:\n"));
+
+	/* Delete input data from fifo_in */
+	used_in = used_in - source_sz;
+	cur_in = (cur_in + source_sz) % fifo_in_len;
+	input_file_offset = input_file_offset + source_sz;
+
+	/* Add output data to fifo_out */
+	used_out = used_out + tpbc;
+
+	assert(used_out <= fifo_out_len);
+
+	total_out = total_out + tpbc;
+
+	/* Deflate history is 32KB max.  No need to supply more
+	 * than 32KB on a resume.
+	 */
+	history_len = (total_out > window_max) ? window_max : total_out;
+
+	/* To estimate expected expansion in the next NX job; 500 means 50%.
+	 * Deflate best case is around 1 to 1000.
+	 */
+	last_comp_ratio = (1000UL * ((uint64_t)source_sz + 1))
+			  / ((uint64_t)tpbc + 1);
+	last_comp_ratio = NX_MAX(NX_MIN(1000UL, last_comp_ratio), 1);
+	NXPRT(fprintf(stderr, "comp_ratio %ld source_sz %d spbc %d tpbc %d\n",
+		      last_comp_ratio, source_sz, spbc, tpbc));
+
+	resuming = 1;
+
+finish_state:
+
+	NXPRT(fprintf(stderr, "finish_state:\n"));
+
+	if (is_final) {
+		if (used_out)
+			goto write_state; /* More data to write out */
+		else if (used_in < 8) {
+			/* Need at least 8 more bytes containing gzip crc
+			 * and isize.
+			 */
+			rc = -1;
+			goto err4;
+		} else {
+			/* Compare checksums and exit */
+			int i;
+			unsigned char tail[8];
+			uint32_t cksum, isize;
+
+			for (i = 0; i < 8; i++)
+				tail[i] = fifo_in[(cur_in + i) % fifo_in_len];
+			fprintf(stderr, "computed checksum %08x isize %08x\n",
+				cmdp->cpb.out_crc, (uint32_t) (total_out
+				% (1ULL<<32)));
+			cksum = ((uint32_t) tail[0] | (uint32_t) tail[1]<<8
+				 | (uint32_t) tail[2]<<16
+				 | (uint32_t) tail[3]<<24);
+			isize = ((uint32_t) tail[4] | (uint32_t) tail[5]<<8
+				 | (uint32_t) tail[6]<<16
+				 | (uint32_t) tail[7]<<24);
+			fprintf(stderr, "stored   checksum %08x isize %08x\n",
+				cksum, isize);
+
+			if (cksum == cmdp->cpb.out_crc && isize == (uint32_t)
+			    (total_out % (1ULL<<32))) {
+				rc = 0;	goto ok1;
+			} else {
+				rc = -1; goto err4;
+			}
+		}
+	} else
+		goto read_state;
+
+	return -1;
+
+err1:
+	fprintf(stderr, "error: not a gzip file, expect %x, read %x\n",
+		expect, c);
+	return -1;
+
+err2:
+	fprintf(stderr, "error: the FLG byte is wrong or not being handled\n");
+	return -1;
+
+err3:
+	fprintf(stderr, "error: gzip header\n");
+	return -1;
+
+err4:
+	fprintf(stderr, "error: checksum missing or mismatch\n");
+
+err5:
+ok1:
+	fprintf(stderr, "decomp is complete: fclose\n");
+	fclose(outf);
+
+	return rc;
+}
+
+
+int main(int argc, char **argv)
+{
+	int rc;
+	struct sigaction act;
+	void *handle;
+
+	nx_dbg = 0;
+	nx_gzip_log = NULL;
+	act.sa_handler = 0;
+	act.sa_sigaction = nxu_sigsegv_handler;
+	act.sa_flags = SA_SIGINFO;
+	act.sa_restorer = 0;
+	sigemptyset(&act.sa_mask);
+	sigaction(SIGSEGV, &act, NULL);
+
+	handle = nx_function_begin(NX_FUNC_COMP_GZIP, 0);
+	if (!handle) {
+		fprintf(stderr, "Unable to init NX, errno %d\n", errno);
+		exit(-1);
+	}
+
+	rc = decompress_file(argc, argv, handle);
+
+	nx_function_end(handle);
+
+	return rc;
+}