diff mbox

[10/10] drivers/crypto/nx: add hardware 842 crypto comp alg

Message ID 1430931066-4536-11-git-send-email-ddstreet@ieee.org (mailing list archive)
State Changes Requested
Delegated to: Herbert Xu
Headers show

Commit Message

Dan Streetman May 6, 2015, 4:51 p.m. UTC
Add crypto compression alg for 842 hardware compression and decompression.

This crypto compression alg is named "nx842" to indicate it uses hardware
to perform the compression and decompression, while the software 842
compression alg is named "sw842".  However, since before this split there
was only one 842 compression alg named "842" which only used hardware,
this is also aliased "842" for backwards compatibility.

This uses only the PowerPC coprocessor hardware for 842 compression.  It
also uses the hardware for decompression, but if the hardware fails it will
fall back to the 842 software decompression library, so that decompression
never fails (for valid 842 compressed buffers).  A header must be used in
most cases, due to the hardware's restrictions on the buffers being
specifically aligned and sized.

Due to the header this driver adds, compressed buffers it creates cannot be
directly passed to the 842 software library for decompression.  However,
compressed buffers created by the software 842 library can be passed to
this driver for hardware 842 decompression (with the exception of buffers
containing the "short data" template, as lib/842/842.h explains).

Signed-off-by: Dan Streetman <ddstreet@ieee.org>
---
 drivers/crypto/nx/Kconfig         |  10 +
 drivers/crypto/nx/Makefile        |   2 +
 drivers/crypto/nx/nx-842-crypto.c | 603 ++++++++++++++++++++++++++++++++++++++
 3 files changed, 615 insertions(+)
 create mode 100644 drivers/crypto/nx/nx-842-crypto.c

Comments

Herbert Xu May 7, 2015, 3:12 a.m. UTC | #1
On Wed, May 06, 2015 at 12:51:06PM -0400, Dan Streetman wrote:
> Add crypto compression alg for 842 hardware compression and decompression.
> 
> This crypto compression alg is named "nx842" to indicate it uses hardware
> to perform the compression and decompression, while the software 842
> compression alg is named "sw842".  However, since before this split there
> was only one 842 compression alg named "842" which only used hardware,
> this is also aliased "842" for backwards compatibility.

This should still be called 842.  You can set the driver name to
nx842 or 842-nx.

Cheers,
Dan Streetman May 7, 2015, 3:06 p.m. UTC | #2
On Wed, May 6, 2015 at 11:12 PM, Herbert Xu <herbert@gondor.apana.org.au> wrote:
> On Wed, May 06, 2015 at 12:51:06PM -0400, Dan Streetman wrote:
>> Add crypto compression alg for 842 hardware compression and decompression.
>>
>> This crypto compression alg is named "nx842" to indicate it uses hardware
>> to perform the compression and decompression, while the software 842
>> compression alg is named "sw842".  However, since before this split there
>> was only one 842 compression alg named "842" which only used hardware,
>> this is also aliased "842" for backwards compatibility.
>
> This should still be called 842.  You can set the driver name to
> nx842 or 842-nx.

ah, ok, will do.

So, I'm wondering about the common NX 842 frontend driver, for the
pSeries and PowerNV platform drivers.  The current setup is:

[ crypto "842-nx" driver ]
           v
[ nx-842 main driver ]
          v
[ nx-842 pSeries driver | nx-842 PowerNV driver ]

The main reason for that is that the HW has specific constraints,
specifically each input and output buffer passed to it for comp or
decomp has to:
-be located at a specific alignment
-have a length of a specific multiple
-have a length between a specific minimum and maximum

The crypto 842-nx has (significant) code in it to handle any alignment
and length input buffers, to match them to what the driver requires.
Would it be better to move that into the crypto code, so that any
crypto compression hw driver can request buffers be specifically
aligned/sized?  I did have to use a header on each compressed buffer
that needed re-alignment or re-sizing, so maybe it's not appropriate
for common crypto compression code.

Since there doesn't seem to be any other hw compression drivers (yet),
maybe it should stay in the 842-nx code, at least for now.  Hopefully
any future compression hw won't have alignment or length multiple
restrictions...

>
> Cheers,
> --
> Email: Herbert Xu <herbert@gondor.apana.org.au>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Herbert Xu May 8, 2015, 2:32 a.m. UTC | #3
On Thu, May 07, 2015 at 11:06:06AM -0400, Dan Streetman wrote:
> 
> The crypto 842-nx has (significant) code in it to handle any alignment
> and length input buffers, to match them to what the driver requires.
> Would it be better to move that into the crypto code, so that any
> crypto compression hw driver can request buffers be specifically
> aligned/sized?  I did have to use a header on each compressed buffer
> that needed re-alignment or re-sizing, so maybe it's not appropriate
> for common crypto compression code.

Yes we could certainly move this logic into the crypto layer, as
we do for ciphers and hashes.

But as you say we could make the next guy who writes a comp driver
do this :)

Cheers,
diff mbox

Patch

diff --git a/drivers/crypto/nx/Kconfig b/drivers/crypto/nx/Kconfig
index ee9e259..3e621ad 100644
--- a/drivers/crypto/nx/Kconfig
+++ b/drivers/crypto/nx/Kconfig
@@ -50,4 +50,14 @@  config CRYPTO_DEV_NX_COMPRESS_POWERNV
 	  algorithm.  This supports NX hardware on the PowerNV platform.
 	  If you choose 'M' here, this module will be called nx_compress_powernv.
 
+config CRYPTO_DEV_NX_COMPRESS_CRYPTO
+	tristate "Compression acceleration cryptographic interface"
+	select CRYPTO_ALGAPI
+	select 842_DECOMPRESS
+	default y
+	help
+	  Support for PowerPC Nest (NX) accelerators using the cryptographic
+	  API.  If you choose 'M' here, this module will be called
+	  nx_compress_crypto.
+
 endif
diff --git a/drivers/crypto/nx/Makefile b/drivers/crypto/nx/Makefile
index 6619787..868b5e6 100644
--- a/drivers/crypto/nx/Makefile
+++ b/drivers/crypto/nx/Makefile
@@ -13,6 +13,8 @@  nx-crypto-objs := nx.o \
 obj-$(CONFIG_CRYPTO_DEV_NX_COMPRESS) += nx-compress.o
 obj-$(CONFIG_CRYPTO_DEV_NX_COMPRESS_PSERIES) += nx-compress-pseries.o
 obj-$(CONFIG_CRYPTO_DEV_NX_COMPRESS_POWERNV) += nx-compress-powernv.o
+obj-$(CONFIG_CRYPTO_DEV_NX_COMPRESS_CRYPTO) += nx-compress-crypto.o
 nx-compress-objs := nx-842.o
 nx-compress-pseries-objs := nx-842-pseries.o
 nx-compress-powernv-objs := nx-842-powernv.o
+nx-compress-crypto-objs := nx-842-crypto.o
diff --git a/drivers/crypto/nx/nx-842-crypto.c b/drivers/crypto/nx/nx-842-crypto.c
new file mode 100644
index 0000000..42d0da8
--- /dev/null
+++ b/drivers/crypto/nx/nx-842-crypto.c
@@ -0,0 +1,603 @@ 
+/*
+ * Cryptographic API for the NX-842 hardware compression.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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.
+ *
+ * Copyright (C) IBM Corporation, 2011-2015
+ *
+ * Original Authors: Robert Jennings <rcj@linux.vnet.ibm.com>
+ *                   Seth Jennings <sjenning@linux.vnet.ibm.com>
+ *
+ * Rewrite: Dan Streetman <ddstreet@ieee.org>
+ *
+ * This is an interface to the NX-842 compression hardware in PowerPC
+ * processors.  Most of the complexity of this drvier is due to the fact that
+ * the NX-842 compression hardware requires the input and output data buffers
+ * to be specifically aligned, to be a specific multiple in length, and within
+ * specific minimum and maximum lengths.  Those restrictions, provided by the
+ * nx-842 driver via nx842_constraints, mean this driver must use bounce
+ * buffers and headers to correct misaligned in or out buffers, and to split
+ * input buffers that are too large.
+ *
+ * This driver will fall back to software decompression if the hardware
+ * decompression fails, so this driver's decompression should never fail as
+ * long as the provided compressed buffer is valid.  Any compressed buffer
+ * created by this driver will have a header (except ones where the input
+ * perfectly matches the constraints); so users of this driver cannot simply
+ * pass a compressed buffer created by this driver over to the 842 software
+ * decompression library.  Instead, users must use this driver to decompress;
+ * if the hardware fails or is unavailable, the compressed buffer will be
+ * parsed and the header removed, and the raw 842 buffer(s) passed to the 842
+ * software decompression library.
+ *
+ * This does not fall back to software compression, however, since the caller
+ * of this function is specifically requesting hardware compression; if the
+ * hardware compression fails, the caller can fall back to software
+ * compression, and the raw 842 compressed buffer that the software compressor
+ * creates can be passed to this driver for hardware decompression; any
+ * buffer without our specific header magic is assumed to be a raw 842 buffer
+ * and passed directly to the hardware.  Note that the software compression
+ * library will produce a compressed buffer that is incompatible with the
+ * hardware decompressor if the original input buffer length is not a multiple
+ * of 8; if such a compressed buffer is passed to this driver for
+ * decompression, the hardware will reject it and this driver will then pass
+ * it over to the software library for decompression.
+ */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/crypto.h>
+#include <linux/vmalloc.h>
+#include <linux/nx842.h>
+#include <linux/sw842.h>
+#include <linux/ratelimit.h>
+
+/* Use the name "nx842" to register as a crypto compression alg. */
+#define NX842_CRYPTO_NAME	"nx842"
+/* Also alias as "842" to keep backwards compatibility; before the crypto 842
+ * software compression alg (named "sw842") was added, the existing crypto
+ * compression alg was named "842" and used this hardware 842 compression.
+ * This allows users to continue using "842" to access this hardware 842
+ * compression.
+ */
+#define NX842_CRYPTO_ALIAS	"842"
+
+/* The first 5 bits of this magic are 0x1f, which is an invalid 842 5-bit
+ * template (see lib/842/842.h), so this magic number will never appear at
+ * the start of a raw 842 compressed buffer.  That is important, as any buffer
+ * passed to us without this magic is assumed to be a raw 842 compressed
+ * buffer, and passed directly to the hardware to decompress.
+ */
+#define NX842_CRYPTO_MAGIC	(0xf842)
+#define NX842_CRYPTO_GROUP_MAX	(0x20)
+#define NX842_CRYPTO_HEADER_SIZE(g)				\
+	(sizeof(struct nx842_crypto_header) +			\
+	 sizeof(struct nx842_crypto_header_group) * (g))
+#define NX842_CRYPTO_HEADER_MAX_SIZE				\
+	NX842_CRYPTO_HEADER_SIZE(NX842_CRYPTO_GROUP_MAX)
+
+/* bounce buffer size */
+#define BOUNCE_BUFFER_ORDER	(2)
+#define BOUNCE_BUFFER_SIZE					\
+	((unsigned int)(PAGE_SIZE << BOUNCE_BUFFER_ORDER))
+
+/* try longer on comp because we can fallback to sw decomp if hw is busy */
+#define COMP_BUSY_TIMEOUT	(250) /* ms */
+#define DECOMP_BUSY_TIMEOUT	(50) /* ms */
+
+struct nx842_crypto_header_group {
+	__be16 padding;			/* unused bytes at start of group */
+	__be32 compressed_length;	/* compressed bytes in group */
+	__be32 uncompressed_length;	/* bytes after decompression */
+} __packed;
+
+struct nx842_crypto_header {
+	__be16 magic;		/* NX842_CRYPTO_MAGIC */
+	__be16 ignore;		/* decompressed end bytes to ignore */
+	u8 groups;		/* total groups in this header */
+	struct nx842_crypto_header_group group[];
+} __packed;
+
+struct nx842_crypto_param {
+	u8 *in;
+	unsigned int iremain;
+	u8 *out;
+	unsigned int oremain;
+	unsigned int ototal;
+};
+
+static int update_param(struct nx842_crypto_param *p,
+			unsigned int slen, unsigned int dlen)
+{
+	if (p->iremain < slen)
+		return -EOVERFLOW;
+	if (p->oremain < dlen)
+		return -ENOSPC;
+
+	p->in += slen;
+	p->iremain -= slen;
+	p->out += dlen;
+	p->oremain -= dlen;
+	p->ototal += dlen;
+
+	return 0;
+}
+
+struct nx842_crypto_ctx {
+	u8 *wmem;
+	u8 *sbounce, *dbounce;
+
+	struct nx842_crypto_header header;
+	struct nx842_crypto_header_group group[NX842_CRYPTO_GROUP_MAX];
+};
+
+static int nx842_crypto_init(struct crypto_tfm *tfm)
+{
+	struct nx842_crypto_ctx *ctx = crypto_tfm_ctx(tfm);
+
+	ctx->wmem = kmalloc(NX842_MEM_COMPRESS, GFP_KERNEL);
+	ctx->sbounce = (u8 *)__get_free_pages(GFP_KERNEL, BOUNCE_BUFFER_ORDER);
+	ctx->dbounce = (u8 *)__get_free_pages(GFP_KERNEL, BOUNCE_BUFFER_ORDER);
+	if (!ctx->wmem || !ctx->sbounce || !ctx->dbounce) {
+		kfree(ctx->wmem);
+		free_page((unsigned long)ctx->sbounce);
+		free_page((unsigned long)ctx->dbounce);
+		return -ENOMEM;
+	}
+
+	return 0;
+}
+
+static void nx842_crypto_exit(struct crypto_tfm *tfm)
+{
+	struct nx842_crypto_ctx *ctx = crypto_tfm_ctx(tfm);
+
+	kfree(ctx->wmem);
+	free_page((unsigned long)ctx->sbounce);
+	free_page((unsigned long)ctx->dbounce);
+}
+
+static int read_constraints(struct nx842_constraints *c)
+{
+	int ret;
+
+	ret = nx842_constraints(c);
+	if (ret) {
+		pr_err_ratelimited("could not get nx842 constraints : %d\n",
+				   ret);
+		return ret;
+	}
+
+	/* limit maximum, to always have enough bounce buffer to decompress */
+	if (c->maximum > BOUNCE_BUFFER_SIZE) {
+		c->maximum = BOUNCE_BUFFER_SIZE;
+		pr_info_once("limiting nx842 maximum to %x\n", c->maximum);
+	}
+
+	return 0;
+}
+
+static int nx842_crypto_add_header(struct nx842_crypto_header *hdr, u8 *buf)
+{
+	int s = NX842_CRYPTO_HEADER_SIZE(hdr->groups);
+
+	/* compress should have added space for header */
+	if (s > be16_to_cpu(hdr->group[0].padding)) {
+		pr_err("Internal error: no space for header\n");
+		return -EINVAL;
+	}
+
+	memcpy(buf, hdr, s);
+
+	print_hex_dump_debug("header ", DUMP_PREFIX_OFFSET, 16, 1, buf, s, 0);
+
+	return 0;
+}
+
+static int compress(struct nx842_crypto_ctx *ctx,
+		    struct nx842_crypto_param *p,
+		    struct nx842_crypto_header_group *g,
+		    struct nx842_constraints *c,
+		    u16 *ignore,
+		    unsigned int hdrsize)
+{
+	unsigned int slen = p->iremain, dlen = p->oremain, tmplen;
+	unsigned int adj_slen = slen;
+	u8 *src = p->in, *dst = p->out;
+	int ret, dskip = 0;
+	ktime_t timeout;
+
+	if (p->iremain == 0)
+		return -EOVERFLOW;
+
+	if (p->oremain == 0 || hdrsize + c->minimum > dlen)
+		return -ENOSPC;
+
+	if (slen % c->multiple)
+		adj_slen = round_up(slen, c->multiple);
+	if (slen < c->minimum)
+		adj_slen = c->minimum;
+	if (slen > c->maximum)
+		adj_slen = slen = c->maximum;
+	if (adj_slen > slen || (u64)src % c->alignment) {
+		adj_slen = min(adj_slen, BOUNCE_BUFFER_SIZE);
+		slen = min(slen, BOUNCE_BUFFER_SIZE);
+		if (adj_slen > slen)
+			memset(ctx->sbounce + slen, 0, adj_slen - slen);
+		memcpy(ctx->sbounce, src, slen);
+		src = ctx->sbounce;
+		slen = adj_slen;
+		pr_debug("using comp sbounce buffer, len %x\n", slen);
+	}
+
+	dst += hdrsize;
+	dlen -= hdrsize;
+
+	if ((u64)dst % c->alignment) {
+		dskip = (int)(PTR_ALIGN(dst, c->alignment) - dst);
+		dst += dskip;
+		dlen -= dskip;
+	}
+	if (dlen % c->multiple)
+		dlen = round_down(dlen, c->multiple);
+	if (dlen < c->minimum) {
+nospc:
+		dst = ctx->dbounce;
+		dlen = min(p->oremain, BOUNCE_BUFFER_SIZE);
+		dlen = round_down(dlen, c->multiple);
+		dskip = 0;
+		pr_debug("using comp dbounce buffer, len %x\n", dlen);
+	}
+	if (dlen > c->maximum)
+		dlen = c->maximum;
+
+	tmplen = dlen;
+	timeout = ktime_add_ms(ktime_get(), COMP_BUSY_TIMEOUT);
+	do {
+		dlen = tmplen; /* reset dlen, if we're retrying */
+		ret = nx842_compress(src, slen, dst, &dlen, ctx->wmem);
+		/* possibly we should reduce the slen here, instead of
+		 * retrying with the dbounce buffer?
+		 */
+		if (ret == -ENOSPC && dst != ctx->dbounce)
+			goto nospc;
+	} while (ret == -EBUSY && ktime_before(ktime_get(), timeout));
+	if (ret)
+		return ret;
+
+	dskip += hdrsize;
+
+	if (dst == ctx->dbounce)
+		memcpy(p->out + dskip, dst, dlen);
+
+	g->padding = cpu_to_be16(dskip);
+	g->compressed_length = cpu_to_be32(dlen);
+	g->uncompressed_length = cpu_to_be32(slen);
+
+	if (p->iremain < slen) {
+		*ignore = slen - p->iremain;
+		slen = p->iremain;
+	}
+
+	pr_debug("compress slen %x ignore %x dlen %x padding %x\n",
+		 slen, *ignore, dlen, dskip);
+
+	return update_param(p, slen, dskip + dlen);
+}
+
+static int nx842_crypto_compress(struct crypto_tfm *tfm,
+				 const u8 *src, unsigned int slen,
+				 u8 *dst, unsigned int *dlen)
+{
+	struct nx842_crypto_ctx *ctx = crypto_tfm_ctx(tfm);
+	struct nx842_crypto_header *hdr = &ctx->header;
+	struct nx842_crypto_param p;
+	struct nx842_constraints c;
+	unsigned int groups, hdrsize, h;
+	int ret, n;
+	bool add_header;
+	u16 ignore = 0;
+
+	if (!tfm || !src || !slen || !dst || !dlen)
+		return -EINVAL;
+
+	p.in = (u8 *)src;
+	p.iremain = slen;
+	p.out = dst;
+	p.oremain = *dlen;
+	p.ototal = 0;
+
+	*dlen = 0;
+
+	ret = read_constraints(&c);
+	if (ret)
+		return ret;
+
+	groups = min_t(unsigned int, NX842_CRYPTO_GROUP_MAX,
+		       DIV_ROUND_UP(p.iremain, c.maximum));
+	hdrsize = NX842_CRYPTO_HEADER_SIZE(groups);
+
+	/* skip adding header if the buffers meet all constraints */
+	add_header = (p.iremain % c.multiple	||
+		      p.iremain < c.minimum	||
+		      p.iremain > c.maximum	||
+		      (u64)p.in % c.alignment	||
+		      p.oremain % c.multiple	||
+		      p.oremain < c.minimum	||
+		      p.oremain > c.maximum	||
+		      (u64)p.out % c.alignment);
+
+	hdr->magic = cpu_to_be16(NX842_CRYPTO_MAGIC);
+	hdr->groups = 0;
+	hdr->ignore = 0;
+
+	while (p.iremain > 0) {
+		n = hdr->groups++;
+		if (hdr->groups > NX842_CRYPTO_GROUP_MAX)
+			return -ENOSPC;
+
+		/* header goes before first group */
+		h = !n && add_header ? hdrsize : 0;
+
+		if (ignore)
+			pr_warn("interal error, ignore is set %x\n", ignore);
+
+		ret = compress(ctx, &p, &hdr->group[n], &c, &ignore, h);
+		if (ret)
+			return ret;
+	}
+
+	if (!add_header && hdr->groups > 1) {
+		pr_err("Internal error: No header but multiple groups\n");
+		return -EINVAL;
+	}
+
+	/* ignore indicates the input stream needed to be padded */
+	hdr->ignore = cpu_to_be16(ignore);
+	if (ignore)
+		pr_debug("marked %d bytes as ignore\n", ignore);
+
+	if (add_header)
+		ret = nx842_crypto_add_header(hdr, dst);
+	if (ret)
+		return ret;
+
+	*dlen = p.ototal;
+
+	pr_debug("compress total slen %x dlen %x\n", slen, *dlen);
+
+	return 0;
+}
+
+static int decompress(struct nx842_crypto_ctx *ctx,
+		      struct nx842_crypto_param *p,
+		      struct nx842_crypto_header_group *g,
+		      struct nx842_constraints *c,
+		      u16 ignore,
+		      bool usehw)
+{
+	unsigned int slen = be32_to_cpu(g->compressed_length);
+	unsigned int required_len = be32_to_cpu(g->uncompressed_length);
+	unsigned int dlen = p->oremain, tmplen;
+	unsigned int adj_slen = slen;
+	u8 *src = p->in, *dst = p->out;
+	u16 padding = be16_to_cpu(g->padding);
+	int ret, spadding = 0, dpadding = 0;
+	ktime_t timeout;
+
+	if (!slen || !required_len)
+		return -EINVAL;
+
+	if (p->iremain <= 0 || padding + slen > p->iremain)
+		return -EOVERFLOW;
+
+	if (p->oremain <= 0 || required_len - ignore > p->oremain)
+		return -ENOSPC;
+
+	src += padding;
+
+	if (!usehw)
+		goto usesw;
+
+	if (slen % c->multiple)
+		adj_slen = round_up(slen, c->multiple);
+	if (slen < c->minimum)
+		adj_slen = c->minimum;
+	if (slen > c->maximum)
+		goto usesw;
+	if (slen < adj_slen || (u64)src % c->alignment) {
+		/* we can append padding bytes because the 842 format defines
+		 * an "end" template (see lib/842/842_decompress.c) and will
+		 * ignore any bytes following it.
+		 */
+		if (slen < adj_slen)
+			memset(ctx->sbounce + slen, 0, adj_slen - slen);
+		memcpy(ctx->sbounce, src, slen);
+		src = ctx->sbounce;
+		spadding = adj_slen - slen;
+		slen = adj_slen;
+		pr_debug("using decomp sbounce buffer, len %x\n", slen);
+	}
+
+	if (dlen % c->multiple)
+		dlen = round_down(dlen, c->multiple);
+	if (dlen < required_len || (u64)dst % c->alignment) {
+		dst = ctx->dbounce;
+		dlen = min(required_len, BOUNCE_BUFFER_SIZE);
+		pr_debug("using decomp dbounce buffer, len %x\n", dlen);
+	}
+	if (dlen < c->minimum)
+		goto usesw;
+	if (dlen > c->maximum)
+		dlen = c->maximum;
+
+	tmplen = dlen;
+	timeout = ktime_add_ms(ktime_get(), DECOMP_BUSY_TIMEOUT);
+	do {
+		dlen = tmplen; /* reset dlen, if we're retrying */
+		ret = nx842_decompress(src, slen, dst, &dlen, ctx->wmem);
+	} while (ret == -EBUSY && ktime_before(ktime_get(), timeout));
+	if (ret) {
+usesw:
+		/* reset everything, sw doesn't have constraints */
+		src = p->in + padding;
+		slen = be32_to_cpu(g->compressed_length);
+		spadding = 0;
+		dst = p->out;
+		dlen = p->oremain;
+		dpadding = 0;
+		if (dlen < required_len) { /* have ignore bytes */
+			dst = ctx->dbounce;
+			dlen = BOUNCE_BUFFER_SIZE;
+		}
+		pr_info_ratelimited("using software 842 decompression\n");
+		ret = sw842_decompress(src, slen, dst, &dlen);
+	}
+	if (ret)
+		return ret;
+
+	slen -= spadding;
+
+	dlen -= ignore;
+	if (ignore)
+		pr_debug("ignoring last %x bytes\n", ignore);
+
+	if (dst == ctx->dbounce)
+		memcpy(p->out, dst, dlen);
+
+	pr_debug("decompress slen %x padding %x dlen %x ignore %x\n",
+		 slen, padding, dlen, ignore);
+
+	return update_param(p, slen + padding, dlen);
+}
+
+static int nx842_crypto_decompress(struct crypto_tfm *tfm,
+				   const u8 *src, unsigned int slen,
+				   u8 *dst, unsigned int *dlen)
+{
+	struct nx842_crypto_ctx *ctx = crypto_tfm_ctx(tfm);
+	struct nx842_crypto_header *hdr;
+	struct nx842_crypto_param p;
+	struct nx842_constraints c;
+	int n, ret, hdr_len;
+	u16 ignore = 0;
+	bool usehw = true;
+
+	if (!tfm || !src || !slen || !dst || !dlen)
+		return -EINVAL;
+
+	p.in = (u8 *)src;
+	p.iremain = slen;
+	p.out = dst;
+	p.oremain = *dlen;
+	p.ototal = 0;
+
+	*dlen = 0;
+
+	if (read_constraints(&c))
+		usehw = false;
+
+	hdr = (struct nx842_crypto_header *)src;
+
+	/* If it doesn't start with our header magic number, assume it's a raw
+	 * 842 compressed buffer and pass it directly to the hardware driver
+	 */
+	if (be16_to_cpu(hdr->magic) != NX842_CRYPTO_MAGIC) {
+		struct nx842_crypto_header_group g = {
+			.padding =		0,
+			.compressed_length =	cpu_to_be32(p.iremain),
+			.uncompressed_length =	cpu_to_be32(p.oremain),
+		};
+
+		ret = decompress(ctx, &p, &g, &c, 0, usehw);
+		if (ret)
+			return ret;
+
+		*dlen = p.ototal;
+
+		return 0;
+	}
+
+	if (!hdr->groups) {
+		pr_err("header has no groups\n");
+		return -EINVAL;
+	}
+	if (hdr->groups > NX842_CRYPTO_GROUP_MAX) {
+		pr_err("header has too many groups %x, max %x\n",
+		       hdr->groups, NX842_CRYPTO_GROUP_MAX);
+		return -EINVAL;
+	}
+
+	hdr_len = NX842_CRYPTO_HEADER_SIZE(hdr->groups);
+	if (hdr_len > slen)
+		return -EOVERFLOW;
+
+	memcpy(&ctx->header, src, hdr_len);
+	hdr = &ctx->header;
+
+	for (n = 0; n < hdr->groups; n++) {
+		/* ignore applies to last group */
+		if (n + 1 == hdr->groups)
+			ignore = be16_to_cpu(hdr->ignore);
+
+		ret = decompress(ctx, &p, &hdr->group[n], &c, ignore, usehw);
+		if (ret)
+			return ret;
+	}
+
+	*dlen = p.ototal;
+
+	pr_debug("decompress total slen %x dlen %x\n", slen, *dlen);
+
+	return 0;
+}
+
+static struct crypto_alg algs[2] = { {
+	.cra_name		= NX842_CRYPTO_NAME,
+	.cra_flags		= CRYPTO_ALG_TYPE_COMPRESS,
+	.cra_ctxsize		= sizeof(struct nx842_crypto_ctx),
+	.cra_module		= THIS_MODULE,
+	.cra_init		= nx842_crypto_init,
+	.cra_exit		= nx842_crypto_exit,
+	.cra_u			= { .compress = {
+	.coa_compress		= nx842_crypto_compress,
+	.coa_decompress		= nx842_crypto_decompress } }
+}, {
+	.cra_name		= NX842_CRYPTO_ALIAS,
+	.cra_flags		= CRYPTO_ALG_TYPE_COMPRESS,
+	.cra_ctxsize		= sizeof(struct nx842_crypto_ctx),
+	.cra_module		= THIS_MODULE,
+	.cra_init		= nx842_crypto_init,
+	.cra_exit		= nx842_crypto_exit,
+	.cra_u			= { .compress = {
+	.coa_compress		= nx842_crypto_compress,
+	.coa_decompress		= nx842_crypto_decompress } }
+} };
+
+static int __init nx842_crypto_mod_init(void)
+{
+	return crypto_register_algs(algs, ARRAY_SIZE(algs));
+}
+module_init(nx842_crypto_mod_init);
+
+static void __exit nx842_crypto_mod_exit(void)
+{
+	crypto_unregister_algs(algs, ARRAY_SIZE(algs));
+}
+module_exit(nx842_crypto_mod_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("IBM PowerPC Nest (NX) 842 Hardware Compression Interface");
+MODULE_ALIAS_CRYPTO(NX842_CRYPTO_NAME);
+MODULE_ALIAS_CRYPTO(NX842_CRYPTO_ALIAS);
+MODULE_AUTHOR("Dan Streetman <ddstreet@ieee.org>");