Patchwork [3/4] powerpc/crypto: add 842 hardware compression driver

login
register
mail settings
Submitter Seth Jennings
Date July 19, 2012, 2:42 p.m.
Message ID <1342708961-28587-4-git-send-email-sjenning@linux.vnet.ibm.com>
Download mbox | patch
Permalink /patch/171956/
State Not Applicable
Headers show

Comments

Seth Jennings - July 19, 2012, 2:42 p.m.
This patch adds the driver for interacting with the 842
compression accelerator on IBM Power7+ systems.

The device is a child of the Platform Facilities Option (PFO)
and shows up as a child of the IBM VIO bus.

The compression/decompression API takes the same arguments
as existing compression methods like lzo and deflate.  The 842
hardware operates on 4K hardware pages and the driver breaks up
input on 4K boundaries to submit it to the hardware accelerator.

Signed-off-by: Robert Jennings <rcj@linux.vnet.ibm.com>
Signed-off-by: Seth Jennings <sjenning@linux.vnet.ibm.com>
---
 MAINTAINERS                |    6 +
 drivers/crypto/nx/Kconfig  |    9 +
 drivers/crypto/nx/Makefile |    3 +
 drivers/crypto/nx/nx-842.c | 1615 ++++++++++++++++++++++++++++++++++++++++++++
 include/linux/nx842.h      |   11 +
 5 files changed, 1644 insertions(+)
 create mode 100644 drivers/crypto/nx/nx-842.c
 create mode 100644 include/linux/nx842.h
Michael Ellerman - July 20, 2012, 5:33 a.m.
On Thu, 2012-07-19 at 09:42 -0500, Seth Jennings wrote:
> This patch adds the driver for interacting with the 842
> compression accelerator on IBM Power7+ systems.

...

> +struct nx842_slentry {
> +	unsigned long ptr; /* Absolute address (use virt_to_abs()) */
> /+	unsigned long len;
> +};

These days virt_to_abs() is just __pa() - ie. convert to a real address.

So you should just use __pa().

And you also need to be certain that you only call it on addresses where
that makes sense, ie. not vmalloc etc.

cheers
Seth Jennings - July 20, 2012, 2:01 p.m.
On 07/20/2012 12:33 AM, Michael Ellerman wrote:
> On Thu, 2012-07-19 at 09:42 -0500, Seth Jennings wrote:
>> This patch adds the driver for interacting with the 842
>> compression accelerator on IBM Power7+ systems.
> 
> ...
> 
>> +struct nx842_slentry {
>> +	unsigned long ptr; /* Absolute address (use virt_to_abs()) */
>> /+	unsigned long len;
>> +};
> 
> These days virt_to_abs() is just __pa() - ie. convert to a real address.

Thanks, I'll make that change.

Is it a blocker to the code being pulled in though? I'm
hoping to get this in ASAP for the 3.6 merge window.  As
this isn't a functional defect (I assume __pa() and
virt_to_abs() still achieve the same result), can I get an
OK from you that this isn't a blocker to the code being
accepted?

I'll submit a patch later to make this change.

Thanks again!

Seth
Michael Ellerman - July 30, 2012, 8 a.m.
On Fri, 2012-07-20 at 09:01 -0500, Seth Jennings wrote:
> On 07/20/2012 12:33 AM, Michael Ellerman wrote:
> > On Thu, 2012-07-19 at 09:42 -0500, Seth Jennings wrote:
> >> This patch adds the driver for interacting with the 842
> >> compression accelerator on IBM Power7+ systems.
> > 
> > ...
> > 
> >> +struct nx842_slentry {
> >> +	unsigned long ptr; /* Absolute address (use virt_to_abs()) */
> >> /+	unsigned long len;
> >> +};
> > 
> > These days virt_to_abs() is just __pa() - ie. convert to a real address.
> 
> Thanks, I'll make that change.
> 
> Is it a blocker to the code being pulled in though? I'm
> hoping to get this in ASAP for the 3.6 merge window.  As
> this isn't a functional defect (I assume __pa() and
> virt_to_abs() still achieve the same result), can I get an
> OK from you that this isn't a blocker to the code being
> accepted?

Sorry I missed your reply. No it's not a blocker, just ugly.

I have sent a series to Ben which removes virt_to_abs() entirely, so
we'll want to make sure we fixup the nx driver before that goes in.

cheers

Patch

diff --git a/MAINTAINERS b/MAINTAINERS
index 5dbf8a2..30821ab 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -3387,6 +3387,12 @@  T:	git git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux.git
 S:	Maintained
 F:	arch/ia64/
 
+IBM Power 842 compression accelerator
+M:	Robert Jennings <rcj@linux.vnet.ibm.com>
+S:	Supported
+F:	drivers/crypto/nx/nx-842.c
+F:	include/linux/nx842.h
+
 IBM Power Linux RAID adapter
 M:	Brian King <brking@us.ibm.com>
 S:	Supported
diff --git a/drivers/crypto/nx/Kconfig b/drivers/crypto/nx/Kconfig
index dedde53..f826166 100644
--- a/drivers/crypto/nx/Kconfig
+++ b/drivers/crypto/nx/Kconfig
@@ -15,3 +15,12 @@  config CRYPTO_DEV_NX_ENCRYPT
 	  Support for Power7+ in-Nest encryption acceleration. This
 	  module supports acceleration for AES and SHA2 algorithms. If you
 	  choose 'M' here, this module will be called nx_crypto.
+
+config CRYPTO_DEV_NX_COMPRESS
+	tristate "Compression acceleration support"
+	depends on PPC64 && IBMVIO
+	default y
+	help
+	  Support for Power7+ in-Nest compression acceleration. This
+	  module supports acceleration for AES and SHA2 algorithms. If you
+	  choose 'M' here, this module will be called nx_compress.
diff --git a/drivers/crypto/nx/Makefile b/drivers/crypto/nx/Makefile
index 7f110e4..bb770ea 100644
--- a/drivers/crypto/nx/Makefile
+++ b/drivers/crypto/nx/Makefile
@@ -9,3 +9,6 @@  nx-crypto-objs := nx.o \
 		  nx-aes-xcbc.o \
 		  nx-sha256.o \
 		  nx-sha512.o
+
+obj-$(CONFIG_CRYPTO_DEV_NX_COMPRESS) += nx-compress.o
+nx-compress-objs := nx-842.o
diff --git a/drivers/crypto/nx/nx-842.c b/drivers/crypto/nx/nx-842.c
new file mode 100644
index 0000000..9da0fb2
--- /dev/null
+++ b/drivers/crypto/nx/nx-842.c
@@ -0,0 +1,1615 @@ 
+/*
+ * Driver for IBM Power 842 compression accelerator
+ *
+ * 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.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
+ *
+ * Copyright (C) IBM Corporation, 2012
+ *
+ * Authors: Robert Jennings <rcj@linux.vnet.ibm.com>
+ *          Seth Jennings <sjenning@linux.vnet.ibm.com>
+ */
+
+#include <linux/module.h>
+#include <asm/vio.h>
+#include <asm/pSeries_reconfig.h>
+#include <linux/slab.h>
+#include <asm/abs_addr.h>
+#include <linux/nx842.h>
+#include <linux/kernel.h>
+
+#include "nx_csbcpb.h" /* struct nx_csbcpb */
+
+#define MODULE_NAME "nx-compress"
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Robert Jennings <rcj@linux.vnet.ibm.com>");
+MODULE_DESCRIPTION("842 H/W Compression driver for IBM Power processors");
+
+#define SHIFT_4K 12
+#define SHIFT_64K 16
+#define SIZE_4K (1UL << SHIFT_4K)
+#define SIZE_64K (1UL << SHIFT_64K)
+
+/* IO buffer must be 128 byte aligned */
+#define IO_BUFFER_ALIGN 128
+
+struct nx842_header {
+	int blocks_nr; /* number of compressed blocks */
+	int offset; /* offset of the first block (from beginning of header) */
+	int sizes[0]; /* size of compressed blocks */
+};
+
+static inline int nx842_header_size(const struct nx842_header *hdr)
+{
+	return sizeof(struct nx842_header) +
+			hdr->blocks_nr * sizeof(hdr->sizes[0]);
+}
+
+/* Macros for fields within nx_csbcpb */
+/* Check the valid bit within the csbcpb valid field */
+#define NX842_CSBCBP_VALID_CHK(x) (x & BIT_MASK(7))
+
+/* CE macros operate on the completion_extension field bits in the csbcpb.
+ * CE0 0=full completion, 1=partial completion
+ * CE1 0=CE0 indicates completion, 1=termination (output may be modified)
+ * CE2 0=processed_bytes is source bytes, 1=processed_bytes is target bytes */
+#define NX842_CSBCPB_CE0(x)	(x & BIT_MASK(7))
+#define NX842_CSBCPB_CE1(x)	(x & BIT_MASK(6))
+#define NX842_CSBCPB_CE2(x)	(x & BIT_MASK(5))
+
+/* The NX unit accepts data only on 4K page boundaries */
+#define NX842_HW_PAGE_SHIFT	SHIFT_4K
+#define NX842_HW_PAGE_SIZE	(ASM_CONST(1) << NX842_HW_PAGE_SHIFT)
+#define NX842_HW_PAGE_MASK	(~(NX842_HW_PAGE_SIZE-1))
+
+enum nx842_status {
+	UNAVAILABLE,
+	AVAILABLE
+};
+
+struct ibm_nx842_counters {
+	atomic64_t comp_complete;
+	atomic64_t comp_failed;
+	atomic64_t decomp_complete;
+	atomic64_t decomp_failed;
+	atomic64_t swdecomp;
+	atomic64_t comp_times[32];
+	atomic64_t decomp_times[32];
+};
+
+static struct nx842_devdata {
+	struct vio_dev *vdev;
+	struct device *dev;
+	struct ibm_nx842_counters *counters;
+	unsigned int max_sg_len;
+	unsigned int max_sync_size;
+	unsigned int max_sync_sg;
+	enum nx842_status status;
+} __rcu *devdata;
+static DEFINE_SPINLOCK(devdata_mutex);
+
+#define NX842_COUNTER_INC(_x) \
+static inline void nx842_inc_##_x( \
+	const struct nx842_devdata *dev) { \
+	if (dev) \
+		atomic64_inc(&dev->counters->_x); \
+}
+NX842_COUNTER_INC(comp_complete);
+NX842_COUNTER_INC(comp_failed);
+NX842_COUNTER_INC(decomp_complete);
+NX842_COUNTER_INC(decomp_failed);
+NX842_COUNTER_INC(swdecomp);
+
+#define NX842_HIST_SLOTS 16
+
+static void ibm_nx842_incr_hist(atomic64_t *times, unsigned int time)
+{
+	int bucket = fls(time);
+
+	if (bucket)
+		bucket = min((NX842_HIST_SLOTS - 1), bucket - 1);
+
+	atomic64_inc(&times[bucket]);
+}
+
+/* NX unit operation flags */
+#define NX842_OP_COMPRESS	0x0
+#define NX842_OP_CRC		0x1
+#define NX842_OP_DECOMPRESS	0x2
+#define NX842_OP_COMPRESS_CRC   (NX842_OP_COMPRESS | NX842_OP_CRC)
+#define NX842_OP_DECOMPRESS_CRC (NX842_OP_DECOMPRESS | NX842_OP_CRC)
+#define NX842_OP_ASYNC		(1<<23)
+#define NX842_OP_NOTIFY		(1<<22)
+#define NX842_OP_NOTIFY_INT(x)	((x & 0xff)<<8)
+
+static unsigned long nx842_get_desired_dma(struct vio_dev *viodev)
+{
+	/* No use of DMA mappings within the driver. */
+	return 0;
+}
+
+struct nx842_slentry {
+	unsigned long ptr; /* Absolute address (use virt_to_abs()) */
+	unsigned long len;
+};
+
+/* pHyp scatterlist entry */
+struct nx842_scatterlist {
+	int entry_nr; /* number of slentries */
+	struct nx842_slentry *entries; /* ptr to array of slentries */
+};
+
+/* Does not include sizeof(entry_nr) in the size */
+static inline unsigned long nx842_get_scatterlist_size(
+				struct nx842_scatterlist *sl)
+{
+	return sl->entry_nr * sizeof(struct nx842_slentry);
+}
+
+static int nx842_build_scatterlist(unsigned long buf, int len,
+			struct nx842_scatterlist *sl)
+{
+	unsigned long nextpage;
+	struct nx842_slentry *entry;
+
+	sl->entry_nr = 0;
+
+	entry = sl->entries;
+	while (len) {
+		entry->ptr = virt_to_abs(buf);
+		nextpage = ALIGN(buf + 1, NX842_HW_PAGE_SIZE);
+		if (nextpage < buf + len) {
+			/* we aren't at the end yet */
+			if (IS_ALIGNED(buf, NX842_HW_PAGE_SIZE))
+				/* we are in the middle (or beginning) */
+				entry->len = NX842_HW_PAGE_SIZE;
+			else
+				/* we are at the beginning */
+				entry->len = nextpage - buf;
+		} else {
+			/* at the end */
+			entry->len = len;
+		}
+
+		len -= entry->len;
+		buf += entry->len;
+		sl->entry_nr++;
+		entry++;
+	}
+
+	return 0;
+}
+
+/*
+ * Working memory for software decompression
+ */
+struct sw842_fifo {
+	union {
+		char f8[256][8];
+		char f4[512][4];
+	};
+	char f2[256][2];
+	unsigned char f84_full;
+	unsigned char f2_full;
+	unsigned char f8_count;
+	unsigned char f2_count;
+	unsigned int f4_count;
+};
+
+/*
+ * Working memory for crypto API
+ */
+struct nx842_workmem {
+	char bounce[PAGE_SIZE]; /* bounce buffer for decompression input */
+	union {
+		/* hardware working memory */
+		struct {
+			/* scatterlist */
+			char slin[SIZE_4K];
+			char slout[SIZE_4K];
+			/* coprocessor status/parameter block */
+			struct nx_csbcpb csbcpb;
+		};
+		/* software working memory */
+		struct sw842_fifo swfifo; /* software decompression fifo */
+	};
+};
+
+int nx842_get_workmem_size(void)
+{
+	return sizeof(struct nx842_workmem) + NX842_HW_PAGE_SIZE;
+}
+EXPORT_SYMBOL_GPL(nx842_get_workmem_size);
+
+int nx842_get_workmem_size_aligned(void)
+{
+	return sizeof(struct nx842_workmem);
+}
+EXPORT_SYMBOL_GPL(nx842_get_workmem_size_aligned);
+
+static int nx842_validate_result(struct device *dev,
+	struct cop_status_block *csb)
+{
+	/* The csb must be valid after returning from vio_h_cop_sync */
+	if (!NX842_CSBCBP_VALID_CHK(csb->valid)) {
+		dev_err(dev, "%s: cspcbp not valid upon completion.\n",
+				__func__);
+		dev_dbg(dev, "valid:0x%02x cs:0x%02x cc:0x%02x ce:0x%02x\n",
+				csb->valid,
+				csb->crb_seq_number,
+				csb->completion_code,
+				csb->completion_extension);
+		dev_dbg(dev, "processed_bytes:%d address:0x%016lx\n",
+				csb->processed_byte_count,
+				(unsigned long)csb->address);
+		return -EIO;
+	}
+
+	/* Check return values from the hardware in the CSB */
+	switch (csb->completion_code) {
+	case 0:	/* Completed without error */
+		break;
+	case 64: /* Target bytes > Source bytes during compression */
+	case 13: /* Output buffer too small */
+		dev_dbg(dev, "%s: Compression output larger than input\n",
+					__func__);
+		return -ENOSPC;
+	case 66: /* Input data contains an illegal template field */
+	case 67: /* Template indicates data past the end of the input stream */
+		dev_dbg(dev, "%s: Bad data for decompression (code:%d)\n",
+					__func__, csb->completion_code);
+		return -EINVAL;
+	default:
+		dev_dbg(dev, "%s: Unspecified error (code:%d)\n",
+					__func__, csb->completion_code);
+		return -EIO;
+	}
+
+	/* Hardware sanity check */
+	if (!NX842_CSBCPB_CE2(csb->completion_extension)) {
+		dev_err(dev, "%s: No error returned by hardware, but "
+				"data returned is unusable, contact support.\n"
+				"(Additional info: csbcbp->processed bytes "
+				"does not specify processed bytes for the "
+				"target buffer.)\n", __func__);
+		return -EIO;
+	}
+
+	return 0;
+}
+
+/**
+ * nx842_compress - Compress data using the 842 algorithm
+ *
+ * Compression provide by the NX842 coprocessor on IBM Power systems.
+ * The input buffer is compressed and the result is stored in the
+ * provided output buffer.
+ *
+ * Upon return from this function @outlen contains the length of the
+ * compressed data.  If there is an error then @outlen will be 0 and an
+ * error will be specified by the return code from this function.
+ *
+ * @in: Pointer to input buffer, must be page aligned
+ * @inlen: Length of input buffer, must be PAGE_SIZE
+ * @out: Pointer to output buffer
+ * @outlen: Length of output buffer
+ * @wrkmem: ptr to buffer for working memory, size determined by
+ *          nx842_get_workmem_size()
+ *
+ * Returns:
+ *   0		Success, output of length @outlen stored in the buffer at @out
+ *   -ENOMEM	Unable to allocate internal buffers
+ *   -ENOSPC	Output buffer is to small
+ *   -EMSGSIZE	XXX Difficult to describe this limitation
+ *   -EIO	Internal error
+ *   -ENODEV	Hardware unavailable
+ */
+int nx842_compress(const unsigned char *in, unsigned int inlen,
+		       unsigned char *out, unsigned int *outlen, void *wmem)
+{
+	struct nx842_header *hdr;
+	struct nx842_devdata *local_devdata;
+	struct device *dev = NULL;
+	struct nx842_workmem *workmem;
+	struct nx842_scatterlist slin, slout;
+	struct nx_csbcpb *csbcpb;
+	int ret = 0, max_sync_size, i, bytesleft, size, hdrsize;
+	unsigned long inbuf, outbuf, padding;
+	struct vio_pfo_op op = {
+		.done = NULL,
+		.handle = 0,
+		.timeout = 0,
+	};
+	unsigned long start_time = get_tb();
+
+	/*
+	 * Make sure input buffer is 64k page aligned.  This is assumed since
+	 * this driver is designed for page compression only (for now).  This
+	 * is very nice since we can now use direct DDE(s) for the input and
+	 * the alignment is guaranteed.
+	*/
+	inbuf = (unsigned long)in;
+	if (!IS_ALIGNED(inbuf, PAGE_SIZE) || inlen != PAGE_SIZE)
+		return -EINVAL;
+
+	rcu_read_lock();
+	local_devdata = rcu_dereference(devdata);
+	if (!local_devdata || !local_devdata->dev) {
+		rcu_read_unlock();
+		return -ENODEV;
+	}
+	max_sync_size = local_devdata->max_sync_size;
+	dev = local_devdata->dev;
+
+	/* Create the header */
+	hdr = (struct nx842_header *)out;
+	hdr->blocks_nr = PAGE_SIZE / max_sync_size;
+	hdrsize = nx842_header_size(hdr);
+	outbuf = (unsigned long)out + hdrsize;
+	bytesleft = *outlen - hdrsize;
+
+	/* Init scatterlist */
+	workmem = (struct nx842_workmem *)ALIGN((unsigned long)wmem,
+		NX842_HW_PAGE_SIZE);
+	slin.entries = (struct nx842_slentry *)workmem->slin;
+	slout.entries = (struct nx842_slentry *)workmem->slout;
+
+	/* Init operation */
+	op.flags = NX842_OP_COMPRESS;
+	csbcpb = &workmem->csbcpb;
+	memset(csbcpb, 0, sizeof(*csbcpb));
+	op.csbcpb = virt_to_abs(csbcpb);
+	op.out = virt_to_abs(slout.entries);
+
+	for (i = 0; i < hdr->blocks_nr; i++) {
+		/*
+		 * Aligning the output blocks to 128 bytes does waste space,
+		 * but it prevents the need for bounce buffers and memory
+		 * copies.  It also simplifies the code a lot.  In the worst
+		 * case (64k page, 4k max_sync_size), you lose up to
+		 * (128*16)/64k = ~3% the compression factor. For 64k
+		 * max_sync_size, the loss would be at most 128/64k = ~0.2%.
+		 */
+		padding = ALIGN(outbuf, IO_BUFFER_ALIGN) - outbuf;
+		outbuf += padding;
+		bytesleft -= padding;
+		if (i == 0)
+			/* save offset into first block in header */
+			hdr->offset = padding + hdrsize;
+
+		if (bytesleft <= 0) {
+			ret = -ENOSPC;
+			goto unlock;
+		}
+
+		/*
+		 * NOTE: If the default max_sync_size is changed from 4k
+		 * to 64k, remove the "likely" case below, since a
+		 * scatterlist will always be needed.
+		 */
+		if (likely(max_sync_size == NX842_HW_PAGE_SIZE)) {
+			/* Create direct DDE */
+			op.in = virt_to_abs(inbuf);
+			op.inlen = max_sync_size;
+
+		} else {
+			/* Create indirect DDE (scatterlist) */
+			nx842_build_scatterlist(inbuf, max_sync_size, &slin);
+			op.in = virt_to_abs(slin.entries);
+			op.inlen = -nx842_get_scatterlist_size(&slin);
+		}
+
+		/*
+		 * If max_sync_size != NX842_HW_PAGE_SIZE, an indirect
+		 * DDE is required for the outbuf.
+		 * If max_sync_size == NX842_HW_PAGE_SIZE, outbuf must
+		 * also be page aligned (1 in 128/4k=32 chance) in order
+		 * to use a direct DDE.
+		 * This is unlikely, just use an indirect DDE always.
+		 */
+		nx842_build_scatterlist(outbuf,
+			min(bytesleft, max_sync_size), &slout);
+		/* op.out set before loop */
+		op.outlen = -nx842_get_scatterlist_size(&slout);
+
+		/* Send request to pHyp */
+		ret = vio_h_cop_sync(local_devdata->vdev, &op);
+
+		/* Check for pHyp error */
+		if (ret) {
+			dev_dbg(dev, "%s: vio_h_cop_sync error (ret=%d, hret=%ld)\n",
+				__func__, ret, op.hcall_err);
+			ret = -EIO;
+			goto unlock;
+		}
+
+		/* Check for hardware error */
+		ret = nx842_validate_result(dev, &csbcpb->csb);
+		if (ret && ret != -ENOSPC)
+			goto unlock;
+
+		/* Handle incompressible data */
+		if (unlikely(ret == -ENOSPC)) {
+			if (bytesleft < max_sync_size) {
+				/*
+				 * Not enough space left in the output buffer
+				 * to store uncompressed block
+				 */
+				goto unlock;
+			} else {
+				/* Store incompressible block */
+				memcpy((void *)outbuf, (void *)inbuf,
+					max_sync_size);
+				hdr->sizes[i] = -max_sync_size;
+				outbuf += max_sync_size;
+				bytesleft -= max_sync_size;
+				/* Reset ret, incompressible data handled */
+				ret = 0;
+			}
+		} else {
+			/* Normal case, compression was successful */
+			size = csbcpb->csb.processed_byte_count;
+			dev_dbg(dev, "%s: processed_bytes=%d\n",
+				__func__, size);
+			hdr->sizes[i] = size;
+			outbuf += size;
+			bytesleft -= size;
+		}
+
+		inbuf += max_sync_size;
+	}
+
+	*outlen = (unsigned int)(outbuf - (unsigned long)out);
+
+unlock:
+	if (ret)
+		nx842_inc_comp_failed(local_devdata);
+	else {
+		nx842_inc_comp_complete(local_devdata);
+		ibm_nx842_incr_hist(local_devdata->counters->comp_times,
+			(get_tb() - start_time) / tb_ticks_per_usec);
+	}
+	rcu_read_unlock();
+	return ret;
+}
+EXPORT_SYMBOL_GPL(nx842_compress);
+
+static int sw842_decompress(const unsigned char *, int, unsigned char *, int *,
+			const void *);
+
+/**
+ * nx842_decompress - Decompress data using the 842 algorithm
+ *
+ * Decompression provide by the NX842 coprocessor on IBM Power systems.
+ * The input buffer is decompressed and the result is stored in the
+ * provided output buffer.  The size allocated to the output buffer is
+ * provided by the caller of this function in @outlen.  Upon return from
+ * this function @outlen contains the length of the decompressed data.
+ * If there is an error then @outlen will be 0 and an error will be
+ * specified by the return code from this function.
+ *
+ * @in: Pointer to input buffer, will use bounce buffer if not 128 byte
+ *      aligned
+ * @inlen: Length of input buffer
+ * @out: Pointer to output buffer, must be page aligned
+ * @outlen: Length of output buffer, must be PAGE_SIZE
+ * @wrkmem: ptr to buffer for working memory, size determined by
+ *          nx842_get_workmem_size()
+ *
+ * Returns:
+ *   0		Success, output of length @outlen stored in the buffer at @out
+ *   -ENODEV	Hardware decompression device is unavailable
+ *   -ENOMEM	Unable to allocate internal buffers
+ *   -ENOSPC	Output buffer is to small
+ *   -EINVAL	Bad input data encountered when attempting decompress
+ *   -EIO	Internal error
+ */
+int nx842_decompress(const unsigned char *in, unsigned int inlen,
+			 unsigned char *out, unsigned int *outlen, void *wmem)
+{
+	struct nx842_header *hdr;
+	struct nx842_devdata *local_devdata;
+	struct device *dev = NULL;
+	struct nx842_workmem *workmem;
+	struct nx842_scatterlist slin, slout;
+	struct nx_csbcpb *csbcpb;
+	int ret = 0, i, size, max_sync_size;
+	unsigned long inbuf, outbuf;
+	struct vio_pfo_op op = {
+		.done = NULL,
+		.handle = 0,
+		.timeout = 0,
+	};
+	unsigned long start_time = get_tb();
+
+	/* Ensure page alignment and size */
+	outbuf = (unsigned long)out;
+	if (!IS_ALIGNED(outbuf, PAGE_SIZE) || *outlen != PAGE_SIZE)
+		return -EINVAL;
+
+	rcu_read_lock();
+	local_devdata = rcu_dereference(devdata);
+	if (local_devdata)
+		dev = local_devdata->dev;
+
+	/* Get header */
+	hdr = (struct nx842_header *)in;
+
+	workmem = (struct nx842_workmem *)ALIGN((unsigned long)wmem,
+		NX842_HW_PAGE_SIZE);
+
+	inbuf = (unsigned long)in + hdr->offset;
+	if (likely(!IS_ALIGNED(inbuf, IO_BUFFER_ALIGN))) {
+		/* Copy block(s) into bounce buffer for alignment */
+		memcpy(workmem->bounce, in + hdr->offset, inlen - hdr->offset);
+		inbuf = (unsigned long)workmem->bounce;
+	}
+
+	/* Init scatterlist */
+	slin.entries = (struct nx842_slentry *)workmem->slin;
+	slout.entries = (struct nx842_slentry *)workmem->slout;
+
+	/* Init operation */
+	op.flags = NX842_OP_DECOMPRESS;
+	csbcpb = &workmem->csbcpb;
+	memset(csbcpb, 0, sizeof(*csbcpb));
+	op.csbcpb = virt_to_abs(csbcpb);
+
+	/*
+	 * max_sync_size may have changed since compression,
+	 * so we can't read it from the device info. We need
+	 * to derive it from hdr->blocks_nr.
+	 */
+	max_sync_size = PAGE_SIZE / hdr->blocks_nr;
+
+	for (i = 0; i < hdr->blocks_nr; i++) {
+		/* Skip padding */
+		inbuf = ALIGN(inbuf, IO_BUFFER_ALIGN);
+
+		if (hdr->sizes[i] < 0) {
+			/* Negative sizes indicate uncompressed data blocks */
+			size = abs(hdr->sizes[i]);
+			memcpy((void *)outbuf, (void *)inbuf, size);
+			outbuf += size;
+			inbuf += size;
+			continue;
+		}
+
+		if (!dev)
+			goto sw;
+
+		/*
+		 * The better the compression, the more likely the "likely"
+		 * case becomes.
+		 */
+		if (likely((inbuf & NX842_HW_PAGE_MASK) ==
+			((inbuf + hdr->sizes[i] - 1) & NX842_HW_PAGE_MASK))) {
+			/* Create direct DDE */
+			op.in = virt_to_abs(inbuf);
+			op.inlen = hdr->sizes[i];
+		} else {
+			/* Create indirect DDE (scatterlist) */
+			nx842_build_scatterlist(inbuf, hdr->sizes[i] , &slin);
+			op.in = virt_to_abs(slin.entries);
+			op.inlen = -nx842_get_scatterlist_size(&slin);
+		}
+
+		/*
+		 * NOTE: If the default max_sync_size is changed from 4k
+		 * to 64k, remove the "likely" case below, since a
+		 * scatterlist will always be needed.
+		 */
+		if (likely(max_sync_size == NX842_HW_PAGE_SIZE)) {
+			/* Create direct DDE */
+			op.out = virt_to_abs(outbuf);
+			op.outlen = max_sync_size;
+		} else {
+			/* Create indirect DDE (scatterlist) */
+			nx842_build_scatterlist(outbuf, max_sync_size, &slout);
+			op.out = virt_to_abs(slout.entries);
+			op.outlen = -nx842_get_scatterlist_size(&slout);
+		}
+
+		/* Send request to pHyp */
+		ret = vio_h_cop_sync(local_devdata->vdev, &op);
+
+		/* Check for pHyp error */
+		if (ret) {
+			dev_dbg(dev, "%s: vio_h_cop_sync error (ret=%d, hret=%ld)\n",
+				__func__, ret, op.hcall_err);
+			dev = NULL;
+			goto sw;
+		}
+
+		/* Check for hardware error */
+		ret = nx842_validate_result(dev, &csbcpb->csb);
+		if (ret) {
+			dev = NULL;
+			goto sw;
+		}
+
+		/* HW decompression success */
+		inbuf += hdr->sizes[i];
+		outbuf += csbcpb->csb.processed_byte_count;
+		continue;
+
+sw:
+		/* software decompression */
+		size = max_sync_size;
+		ret = sw842_decompress(
+			(unsigned char *)inbuf, hdr->sizes[i],
+			(unsigned char *)outbuf, &size, wmem);
+		if (ret)
+			pr_debug("%s: sw842_decompress failed with %d\n",
+				__func__, ret);
+
+		if (ret) {
+			if (ret != -ENOSPC && ret != -EINVAL &&
+					ret != -EMSGSIZE)
+				ret = -EIO;
+			goto unlock;
+		}
+
+		/* SW decompression success */
+		inbuf += hdr->sizes[i];
+		outbuf += size;
+	}
+
+	*outlen = (unsigned int)(outbuf - (unsigned long)out);
+
+unlock:
+	if (ret)
+		/* decompress fail */
+		nx842_inc_decomp_failed(local_devdata);
+	else {
+		if (!dev)
+			/* software decompress */
+			nx842_inc_swdecomp(local_devdata);
+		nx842_inc_decomp_complete(local_devdata);
+		ibm_nx842_incr_hist(local_devdata->counters->decomp_times,
+			(get_tb() - start_time) / tb_ticks_per_usec);
+	}
+
+	rcu_read_unlock();
+	return ret;
+}
+EXPORT_SYMBOL_GPL(nx842_decompress);
+
+/**
+ * nx842_OF_set_defaults -- Set default (disabled) values for devdata
+ *
+ * @devdata - struct nx842_devdata to update
+ *
+ * Returns:
+ *  0 on success
+ *  -ENOENT if @devdata ptr is NULL
+ */
+static int nx842_OF_set_defaults(struct nx842_devdata *devdata)
+{
+	if (devdata) {
+		devdata->max_sync_size = 0;
+		devdata->max_sync_sg = 0;
+		devdata->max_sg_len = 0;
+		devdata->status = UNAVAILABLE;
+		return 0;
+	} else
+		return -ENOENT;
+}
+
+/**
+ * nx842_OF_upd_status -- Update the device info from OF status prop
+ *
+ * The status property indicates if the accelerator is enabled.  If the
+ * device is in the OF tree it indicates that the hardware is present.
+ * The status field indicates if the device is enabled when the status
+ * is 'okay'.  Otherwise the device driver will be disabled.
+ *
+ * @devdata - struct nx842_devdata to update
+ * @prop - struct property point containing the maxsyncop for the update
+ *
+ * Returns:
+ *  0 - Device is available
+ *  -EINVAL - Device is not available
+ */
+static int nx842_OF_upd_status(struct nx842_devdata *devdata,
+					struct property *prop) {
+	int ret = 0;
+	const char *status = (const char *)prop->value;
+
+	if (!strncmp(status, "okay", (size_t)prop->length)) {
+		devdata->status = AVAILABLE;
+	} else {
+		dev_info(devdata->dev, "%s: status '%s' is not 'okay'\n",
+				__func__, status);
+		devdata->status = UNAVAILABLE;
+	}
+
+	return ret;
+}
+
+/**
+ * nx842_OF_upd_maxsglen -- Update the device info from OF maxsglen prop
+ *
+ * Definition of the 'ibm,max-sg-len' OF property:
+ *  This field indicates the maximum byte length of a scatter list
+ *  for the platform facility. It is a single cell encoded as with encode-int.
+ *
+ * Example:
+ *  # od -x ibm,max-sg-len
+ *  0000000 0000 0ff0
+ *
+ *  In this example, the maximum byte length of a scatter list is
+ *  0x0ff0 (4,080).
+ *
+ * @devdata - struct nx842_devdata to update
+ * @prop - struct property point containing the maxsyncop for the update
+ *
+ * Returns:
+ *  0 on success
+ *  -EINVAL on failure
+ */
+static int nx842_OF_upd_maxsglen(struct nx842_devdata *devdata,
+					struct property *prop) {
+	int ret = 0;
+	const int *maxsglen = prop->value;
+
+	if (prop->length != sizeof(*maxsglen)) {
+		dev_err(devdata->dev, "%s: unexpected format for ibm,max-sg-len property\n", __func__);
+		dev_dbg(devdata->dev, "%s: ibm,max-sg-len is %d bytes long, expected %lu bytes\n", __func__,
+				prop->length, sizeof(*maxsglen));
+		ret = -EINVAL;
+	} else {
+		devdata->max_sg_len = (unsigned int)min(*maxsglen,
+				(int)NX842_HW_PAGE_SIZE);
+	}
+
+	return ret;
+}
+
+/**
+ * nx842_OF_upd_maxsyncop -- Update the device info from OF maxsyncop prop
+ *
+ * Definition of the 'ibm,max-sync-cop' OF property:
+ *  Two series of cells.  The first series of cells represents the maximums
+ *  that can be synchronously compressed. The second series of cells
+ *  represents the maximums that can be synchronously decompressed.
+ *  1. The first cell in each series contains the count of the number of
+ *     data length, scatter list elements pairs that follow – each being
+ *     of the form
+ *    a. One cell data byte length
+ *    b. One cell total number of scatter list elements
+ *
+ * Example:
+ *  # od -x ibm,max-sync-cop
+ *  0000000 0000 0001 0000 1000 0000 01fe 0000 0001
+ *  0000020 0000 1000 0000 01fe
+ *
+ *  In this example, compression supports 0x1000 (4,096) data byte length
+ *  and 0x1fe (510) total scatter list elements.  Decompression supports
+ *  0x1000 (4,096) data byte length and 0x1f3 (510) total scatter list
+ *  elements.
+ *
+ * @devdata - struct nx842_devdata to update
+ * @prop - struct property point containing the maxsyncop for the update
+ *
+ * Returns:
+ *  0 on success
+ *  -EINVAL on failure
+ */
+static int nx842_OF_upd_maxsyncop(struct nx842_devdata *devdata,
+					struct property *prop) {
+	int ret = 0;
+	const struct maxsynccop_t {
+		int comp_elements;
+		int comp_data_limit;
+		int comp_sg_limit;
+		int decomp_elements;
+		int decomp_data_limit;
+		int decomp_sg_limit;
+	} *maxsynccop;
+
+	if (prop->length != sizeof(*maxsynccop)) {
+		dev_err(devdata->dev, "%s: unexpected format for ibm,max-sync-cop property\n", __func__);
+		dev_dbg(devdata->dev, "%s: ibm,max-sync-cop is %d bytes long, expected %lu bytes\n", __func__, prop->length,
+				sizeof(*maxsynccop));
+		ret = -EINVAL;
+		goto out;
+	}
+
+	maxsynccop = (const struct maxsynccop_t *)prop->value;
+
+	/* Use one limit rather than separate limits for compression and
+	 * decompression. Set a maximum for this so as not to exceed the
+	 * size that the header can support and round the value down to
+	 * the hardware page size (4K) */
+	devdata->max_sync_size =
+			(unsigned int)min(maxsynccop->comp_data_limit,
+					maxsynccop->decomp_data_limit);
+
+	devdata->max_sync_size = min_t(unsigned int, devdata->max_sync_size,
+					SIZE_64K);
+
+	if (devdata->max_sync_size < SIZE_4K) {
+		dev_err(devdata->dev, "%s: hardware max data size (%u) is "
+				"less than the driver minimum, unable to use "
+				"the hardware device\n",
+				__func__, devdata->max_sync_size);
+		ret = -EINVAL;
+		goto out;
+	}
+
+	devdata->max_sync_sg = (unsigned int)min(maxsynccop->comp_sg_limit,
+						maxsynccop->decomp_sg_limit);
+	if (devdata->max_sync_sg < 1) {
+		dev_err(devdata->dev, "%s: hardware max sg size (%u) is "
+				"less than the driver minimum, unable to use "
+				"the hardware device\n",
+				__func__, devdata->max_sync_sg);
+		ret = -EINVAL;
+		goto out;
+	}
+
+out:
+	return ret;
+}
+
+/**
+ *
+ * nx842_OF_upd -- Handle OF properties updates for the device.
+ *
+ * Set all properties from the OF tree.  Optionally, a new property
+ * can be provided by the @new_prop pointer to overwrite an existing value.
+ * The device will remain disabled until all values are valid, this function
+ * will return an error for updates unless all values are valid.
+ *
+ * @new_prop: If not NULL, this property is being updated.  If NULL, update
+ *  all properties from the current values in the OF tree.
+ *
+ * Returns:
+ *  0 - Success
+ *  -ENOMEM - Could not allocate memory for new devdata structure
+ *  -EINVAL - property value not found, new_prop is not a recognized
+ *	property for the device or property value is not valid.
+ *  -ENODEV - Device is not available
+ */
+static int nx842_OF_upd(struct property *new_prop)
+{
+	struct nx842_devdata *old_devdata = NULL;
+	struct nx842_devdata *new_devdata = NULL;
+	struct device_node *of_node = NULL;
+	struct property *status = NULL;
+	struct property *maxsglen = NULL;
+	struct property *maxsyncop = NULL;
+	int ret = 0;
+	unsigned long flags;
+
+	spin_lock_irqsave(&devdata_mutex, flags);
+	old_devdata = rcu_dereference_check(devdata,
+			lockdep_is_held(&devdata_mutex));
+	if (old_devdata)
+		of_node = old_devdata->dev->of_node;
+
+	if (!old_devdata || !of_node) {
+		pr_err("%s: device is not available\n", __func__);
+		spin_unlock_irqrestore(&devdata_mutex, flags);
+		return -ENODEV;
+	}
+
+	new_devdata = kzalloc(sizeof(*new_devdata), GFP_NOFS);
+	if (!new_devdata) {
+		dev_err(old_devdata->dev, "%s: Could not allocate memory for device data\n", __func__);
+		ret = -ENOMEM;
+		goto error_out;
+	}
+
+	memcpy(new_devdata, old_devdata, sizeof(*old_devdata));
+	new_devdata->counters = old_devdata->counters;
+
+	/* Set ptrs for existing properties */
+	status = of_find_property(of_node, "status", NULL);
+	maxsglen = of_find_property(of_node, "ibm,max-sg-len", NULL);
+	maxsyncop = of_find_property(of_node, "ibm,max-sync-cop", NULL);
+	if (!status || !maxsglen || !maxsyncop) {
+		dev_err(old_devdata->dev, "%s: Could not locate device properties\n", __func__);
+		ret = -EINVAL;
+		goto error_out;
+	}
+
+	/* Set ptr to new property if provided */
+	if (new_prop) {
+		/* Single property */
+		if (!strncmp(new_prop->name, "status", new_prop->length)) {
+			status = new_prop;
+
+		} else if (!strncmp(new_prop->name, "ibm,max-sg-len",
+					new_prop->length)) {
+			maxsglen = new_prop;
+
+		} else if (!strncmp(new_prop->name, "ibm,max-sync-cop",
+					new_prop->length)) {
+			maxsyncop = new_prop;
+
+		} else {
+			/*
+			 * Skip the update, the property being updated
+			 * has no impact.
+			 */
+			goto out;
+		}
+	}
+
+	/* Perform property updates */
+	ret = nx842_OF_upd_status(new_devdata, status);
+	if (ret)
+		goto error_out;
+
+	ret = nx842_OF_upd_maxsglen(new_devdata, maxsglen);
+	if (ret)
+		goto error_out;
+
+	ret = nx842_OF_upd_maxsyncop(new_devdata, maxsyncop);
+	if (ret)
+		goto error_out;
+
+out:
+	dev_info(old_devdata->dev, "%s: max_sync_size new:%u old:%u\n",
+			__func__, new_devdata->max_sync_size,
+			old_devdata->max_sync_size);
+	dev_info(old_devdata->dev, "%s: max_sync_sg new:%u old:%u\n",
+			__func__, new_devdata->max_sync_sg,
+			old_devdata->max_sync_sg);
+	dev_info(old_devdata->dev, "%s: max_sg_len new:%u old:%u\n",
+			__func__, new_devdata->max_sg_len,
+			old_devdata->max_sg_len);
+
+	rcu_assign_pointer(devdata, new_devdata);
+	spin_unlock_irqrestore(&devdata_mutex, flags);
+	synchronize_rcu();
+	dev_set_drvdata(new_devdata->dev, new_devdata);
+	kfree(old_devdata);
+	return 0;
+
+error_out:
+	if (new_devdata) {
+		dev_info(old_devdata->dev, "%s: device disabled\n", __func__);
+		nx842_OF_set_defaults(new_devdata);
+		rcu_assign_pointer(devdata, new_devdata);
+		spin_unlock_irqrestore(&devdata_mutex, flags);
+		synchronize_rcu();
+		dev_set_drvdata(new_devdata->dev, new_devdata);
+		kfree(old_devdata);
+	} else {
+		dev_err(old_devdata->dev, "%s: could not update driver from hardware\n", __func__);
+		spin_unlock_irqrestore(&devdata_mutex, flags);
+	}
+
+	if (!ret)
+		ret = -EINVAL;
+	return ret;
+}
+
+/**
+ * nx842_OF_notifier - Process updates to OF properties for the device
+ *
+ * @np: notifier block
+ * @action: notifier action
+ * @update: struct pSeries_reconfig_prop_update pointer if action is
+ *	PSERIES_UPDATE_PROPERTY
+ *
+ * Returns:
+ *	NOTIFY_OK on success
+ *	NOTIFY_BAD encoded with error number on failure, use
+ *		notifier_to_errno() to decode this value
+ */
+static int nx842_OF_notifier(struct notifier_block *np,
+					unsigned long action,
+					void *update)
+{
+	struct pSeries_reconfig_prop_update *upd;
+	struct nx842_devdata *local_devdata;
+	struct device_node *node = NULL;
+
+	upd = (struct pSeries_reconfig_prop_update *)update;
+
+	rcu_read_lock();
+	local_devdata = rcu_dereference(devdata);
+	if (local_devdata)
+		node = local_devdata->dev->of_node;
+
+	if (local_devdata &&
+			action == PSERIES_UPDATE_PROPERTY &&
+			!strcmp(upd->node->name, node->name)) {
+		rcu_read_unlock();
+		nx842_OF_upd(upd->property);
+	} else
+		rcu_read_unlock();
+
+	return NOTIFY_OK;
+}
+
+static struct notifier_block nx842_of_nb = {
+	.notifier_call = nx842_OF_notifier,
+};
+
+#define nx842_counter_read(_name)					\
+static ssize_t nx842_##_name##_show(struct device *dev,		\
+		struct device_attribute *attr,				\
+		char *buf) {						\
+	struct nx842_devdata *local_devdata;			\
+	int p = 0;							\
+	rcu_read_lock();						\
+	local_devdata = rcu_dereference(devdata);			\
+	if (local_devdata)						\
+		p = snprintf(buf, PAGE_SIZE, "%ld\n",			\
+		       atomic64_read(&local_devdata->counters->_name));	\
+	rcu_read_unlock();						\
+	return p;							\
+}
+
+#define NX842DEV_COUNTER_ATTR_RO(_name)					\
+	nx842_counter_read(_name);					\
+	static struct device_attribute dev_attr_##_name = __ATTR(_name,	\
+						0444,			\
+						nx842_##_name##_show,\
+						NULL);
+
+NX842DEV_COUNTER_ATTR_RO(comp_complete);
+NX842DEV_COUNTER_ATTR_RO(comp_failed);
+NX842DEV_COUNTER_ATTR_RO(decomp_complete);
+NX842DEV_COUNTER_ATTR_RO(decomp_failed);
+NX842DEV_COUNTER_ATTR_RO(swdecomp);
+
+static ssize_t nx842_timehist_show(struct device *,
+		struct device_attribute *, char *);
+
+static struct device_attribute dev_attr_comp_times = __ATTR(comp_times, 0444,
+		nx842_timehist_show, NULL);
+static struct device_attribute dev_attr_decomp_times = __ATTR(decomp_times,
+		0444, nx842_timehist_show, NULL);
+
+static ssize_t nx842_timehist_show(struct device *dev,
+		struct device_attribute *attr, char *buf) {
+	char *p = buf;
+	struct nx842_devdata *local_devdata;
+	atomic64_t *times;
+	int bytes_remain = PAGE_SIZE;
+	int bytes;
+	int i;
+
+	rcu_read_lock();
+	local_devdata = rcu_dereference(devdata);
+	if (!local_devdata) {
+		rcu_read_unlock();
+		return 0;
+	}
+
+	if (attr == &dev_attr_comp_times)
+		times = local_devdata->counters->comp_times;
+	else if (attr == &dev_attr_decomp_times)
+		times = local_devdata->counters->decomp_times;
+	else {
+		rcu_read_unlock();
+		return 0;
+	}
+
+	for (i = 0; i < (NX842_HIST_SLOTS - 2); i++) {
+		bytes = snprintf(p, bytes_remain, "%u-%uus:\t%ld\n",
+			       i ? (2<<(i-1)) : 0, (2<<i)-1,
+			       atomic64_read(&times[i]));
+		bytes_remain -= bytes;
+		p += bytes;
+	}
+	/* The last bucket holds everything over
+	 * 2<<(NX842_HIST_SLOTS - 2) us */
+	bytes = snprintf(p, bytes_remain, "%uus - :\t%ld\n",
+			2<<(NX842_HIST_SLOTS - 2),
+			atomic64_read(&times[(NX842_HIST_SLOTS - 1)]));
+	p += bytes;
+
+	rcu_read_unlock();
+	return p - buf;
+}
+
+static struct attribute *nx842_sysfs_entries[] = {
+	&dev_attr_comp_complete.attr,
+	&dev_attr_comp_failed.attr,
+	&dev_attr_decomp_complete.attr,
+	&dev_attr_decomp_failed.attr,
+	&dev_attr_swdecomp.attr,
+	&dev_attr_comp_times.attr,
+	&dev_attr_decomp_times.attr,
+	NULL,
+};
+
+static struct attribute_group nx842_attribute_group = {
+	.name = NULL,		/* put in device directory */
+	.attrs = nx842_sysfs_entries,
+};
+
+static int __init nx842_probe(struct vio_dev *viodev,
+				  const struct vio_device_id *id)
+{
+	struct nx842_devdata *old_devdata, *new_devdata = NULL;
+	unsigned long flags;
+	int ret = 0;
+
+	spin_lock_irqsave(&devdata_mutex, flags);
+	old_devdata = rcu_dereference_check(devdata,
+			lockdep_is_held(&devdata_mutex));
+
+	if (old_devdata && old_devdata->vdev != NULL) {
+		dev_err(&viodev->dev, "%s: Attempt to register more than one instance of the hardware\n", __func__);
+		ret = -1;
+		goto error_unlock;
+	}
+
+	dev_set_drvdata(&viodev->dev, NULL);
+
+	new_devdata = kzalloc(sizeof(*new_devdata), GFP_NOFS);
+	if (!new_devdata) {
+		dev_err(&viodev->dev, "%s: Could not allocate memory for device data\n", __func__);
+		ret = -ENOMEM;
+		goto error_unlock;
+	}
+
+	new_devdata->counters = kzalloc(sizeof(*new_devdata->counters),
+			GFP_NOFS);
+	if (!new_devdata->counters) {
+		dev_err(&viodev->dev, "%s: Could not allocate memory for performance counters\n", __func__);
+		ret = -ENOMEM;
+		goto error_unlock;
+	}
+
+	new_devdata->vdev = viodev;
+	new_devdata->dev = &viodev->dev;
+	nx842_OF_set_defaults(new_devdata);
+
+	rcu_assign_pointer(devdata, new_devdata);
+	spin_unlock_irqrestore(&devdata_mutex, flags);
+	synchronize_rcu();
+	kfree(old_devdata);
+
+	pSeries_reconfig_notifier_register(&nx842_of_nb);
+
+	ret = nx842_OF_upd(NULL);
+	if (ret && ret != -ENODEV) {
+		dev_err(&viodev->dev, "could not parse device tree. %d\n", ret);
+		ret = -1;
+		goto error;
+	}
+
+	rcu_read_lock();
+	if (dev_set_drvdata(&viodev->dev, rcu_dereference(devdata))) {
+		rcu_read_unlock();
+		dev_err(&viodev->dev, "failed to set driver data for device\n");
+		ret = -1;
+		goto error;
+	}
+	rcu_read_unlock();
+
+	if (sysfs_create_group(&viodev->dev.kobj, &nx842_attribute_group)) {
+		dev_err(&viodev->dev, "could not create sysfs device attributes\n");
+		ret = -1;
+		goto error;
+	}
+
+	return 0;
+
+error_unlock:
+	spin_unlock_irqrestore(&devdata_mutex, flags);
+	if (new_devdata)
+		kfree(new_devdata->counters);
+	kfree(new_devdata);
+error:
+	return ret;
+}
+
+static int __exit nx842_remove(struct vio_dev *viodev)
+{
+	struct nx842_devdata *old_devdata;
+	unsigned long flags;
+
+	pr_info("Removing IBM Power 842 compression device\n");
+	sysfs_remove_group(&viodev->dev.kobj, &nx842_attribute_group);
+
+	spin_lock_irqsave(&devdata_mutex, flags);
+	old_devdata = rcu_dereference_check(devdata,
+			lockdep_is_held(&devdata_mutex));
+	pSeries_reconfig_notifier_unregister(&nx842_of_nb);
+	rcu_assign_pointer(devdata, NULL);
+	spin_unlock_irqrestore(&devdata_mutex, flags);
+	synchronize_rcu();
+	dev_set_drvdata(&viodev->dev, NULL);
+	if (old_devdata)
+		kfree(old_devdata->counters);
+	kfree(old_devdata);
+	return 0;
+}
+
+static struct vio_device_id nx842_driver_ids[] = {
+	{"ibm,compression-v1", "ibm,compression"},
+	{"", ""},
+};
+
+static struct vio_driver nx842_driver = {
+	.name = MODULE_NAME,
+	.probe = nx842_probe,
+	.remove = nx842_remove,
+	.get_desired_dma = nx842_get_desired_dma,
+	.id_table = nx842_driver_ids,
+};
+
+static int __init nx842_init(void)
+{
+	struct nx842_devdata *new_devdata;
+	pr_info("Registering IBM Power 842 compression driver\n");
+
+	RCU_INIT_POINTER(devdata, NULL);
+	new_devdata = kzalloc(sizeof(*new_devdata), GFP_KERNEL);
+	if (!new_devdata) {
+		pr_err("Could not allocate memory for device data\n");
+		return -ENOMEM;
+	}
+	new_devdata->status = UNAVAILABLE;
+	RCU_INIT_POINTER(devdata, new_devdata);
+
+	return vio_register_driver(&nx842_driver);
+}
+
+module_init(nx842_init);
+
+static void __exit nx842_exit(void)
+{
+	struct nx842_devdata *old_devdata;
+	unsigned long flags;
+
+	pr_info("Exiting IBM Power 842 compression driver\n");
+	spin_lock_irqsave(&devdata_mutex, flags);
+	old_devdata = rcu_dereference_check(devdata,
+			lockdep_is_held(&devdata_mutex));
+	rcu_assign_pointer(devdata, NULL);
+	spin_unlock_irqrestore(&devdata_mutex, flags);
+	synchronize_rcu();
+	if (old_devdata)
+		dev_set_drvdata(old_devdata->dev, NULL);
+	kfree(old_devdata);
+	vio_unregister_driver(&nx842_driver);
+}
+
+module_exit(nx842_exit);
+
+/*********************************
+ * 842 software decompressor
+*********************************/
+typedef int (*sw842_template_op)(const char **, int *, unsigned char **,
+						struct sw842_fifo *);
+
+static int sw842_data8(const char **, int *, unsigned char **,
+						struct sw842_fifo *);
+static int sw842_data4(const char **, int *, unsigned char **,
+						struct sw842_fifo *);
+static int sw842_data2(const char **, int *, unsigned char **,
+						struct sw842_fifo *);
+static int sw842_ptr8(const char **, int *, unsigned char **,
+						struct sw842_fifo *);
+static int sw842_ptr4(const char **, int *, unsigned char **,
+						struct sw842_fifo *);
+static int sw842_ptr2(const char **, int *, unsigned char **,
+						struct sw842_fifo *);
+
+/* special templates */
+#define SW842_TMPL_REPEAT 0x1B
+#define SW842_TMPL_ZEROS 0x1C
+#define SW842_TMPL_EOF 0x1E
+
+static sw842_template_op sw842_tmpl_ops[26][4] = {
+	{ sw842_data8, NULL}, /* 0 (00000) */
+	{ sw842_data4, sw842_data2, sw842_ptr2,  NULL},
+	{ sw842_data4, sw842_ptr2,  sw842_data2, NULL},
+	{ sw842_data4, sw842_ptr2,  sw842_ptr2,  NULL},
+	{ sw842_data4, sw842_ptr4,  NULL},
+	{ sw842_data2, sw842_ptr2,  sw842_data4, NULL},
+	{ sw842_data2, sw842_ptr2,  sw842_data2, sw842_ptr2},
+	{ sw842_data2, sw842_ptr2,  sw842_ptr2,  sw842_data2},
+	{ sw842_data2, sw842_ptr2,  sw842_ptr2,  sw842_ptr2,},
+	{ sw842_data2, sw842_ptr2,  sw842_ptr4,  NULL},
+	{ sw842_ptr2,  sw842_data2, sw842_data4, NULL}, /* 10 (01010) */
+	{ sw842_ptr2,  sw842_data4, sw842_ptr2,  NULL},
+	{ sw842_ptr2,  sw842_data2, sw842_ptr2,  sw842_data2},
+	{ sw842_ptr2,  sw842_data2, sw842_ptr2,  sw842_ptr2},
+	{ sw842_ptr2,  sw842_data2, sw842_ptr4,  NULL},
+	{ sw842_ptr2,  sw842_ptr2,  sw842_data4, NULL},
+	{ sw842_ptr2,  sw842_ptr2,  sw842_data2, sw842_ptr2},
+	{ sw842_ptr2,  sw842_ptr2,  sw842_ptr2,  sw842_data2},
+	{ sw842_ptr2,  sw842_ptr2,  sw842_ptr2,  sw842_ptr2},
+	{ sw842_ptr2,  sw842_ptr2,  sw842_ptr4,  NULL},
+	{ sw842_ptr4,  sw842_data4, NULL}, /* 20 (10100) */
+	{ sw842_ptr4,  sw842_data2, sw842_ptr2,  NULL},
+	{ sw842_ptr4,  sw842_ptr2,  sw842_data2, NULL},
+	{ sw842_ptr4,  sw842_ptr2,  sw842_ptr2,  NULL},
+	{ sw842_ptr4,  sw842_ptr4,  NULL},
+	{ sw842_ptr8,  NULL}
+};
+
+/* Software decompress helpers */
+
+static uint8_t sw842_get_byte(const char *buf, int bit)
+{
+	uint8_t tmpl;
+	uint16_t tmp;
+	tmp = htons(*(uint16_t *)(buf));
+	tmp = (uint16_t)(tmp << bit);
+	tmp = ntohs(tmp);
+	memcpy(&tmpl, &tmp, 1);
+	return tmpl;
+}
+
+static uint8_t sw842_get_template(const char **buf, int *bit)
+{
+	uint8_t byte;
+	byte = sw842_get_byte(*buf, *bit);
+	byte = byte >> 3;
+	byte &= 0x1F;
+	*buf += (*bit + 5) / 8;
+	*bit = (*bit + 5) % 8;
+	return byte;
+}
+
+/* repeat_count happens to be 5-bit too (like the template) */
+static uint8_t sw842_get_repeat_count(const char **buf, int *bit)
+{
+	uint8_t byte;
+	byte = sw842_get_byte(*buf, *bit);
+	byte = byte >> 2;
+	byte &= 0x3F;
+	*buf += (*bit + 6) / 8;
+	*bit = (*bit + 6) % 8;
+	return byte;
+}
+
+static uint8_t sw842_get_ptr2(const char **buf, int *bit)
+{
+	uint8_t ptr;
+	ptr = sw842_get_byte(*buf, *bit);
+	(*buf)++;
+	return ptr;
+}
+
+static uint16_t sw842_get_ptr4(const char **buf, int *bit,
+		struct sw842_fifo *fifo)
+{
+	uint16_t ptr;
+	ptr = htons(*(uint16_t *)(*buf));
+	ptr = (uint16_t)(ptr << *bit);
+	ptr = ptr >> 7;
+	ptr &= 0x01FF;
+	*buf += (*bit + 9) / 8;
+	*bit = (*bit + 9) % 8;
+	return ptr;
+}
+
+static uint8_t sw842_get_ptr8(const char **buf, int *bit,
+		struct sw842_fifo *fifo)
+{
+	return sw842_get_ptr2(buf, bit);
+}
+
+/* Software decompress template ops */
+
+static int sw842_data8(const char **inbuf, int *inbit,
+		unsigned char **outbuf, struct sw842_fifo *fifo)
+{
+	int ret;
+
+	ret = sw842_data4(inbuf, inbit, outbuf, fifo);
+	if (ret)
+		return ret;
+	ret = sw842_data4(inbuf, inbit, outbuf, fifo);
+	return ret;
+}
+
+static int sw842_data4(const char **inbuf, int *inbit,
+		unsigned char **outbuf, struct sw842_fifo *fifo)
+{
+	int ret;
+
+	ret = sw842_data2(inbuf, inbit, outbuf, fifo);
+	if (ret)
+		return ret;
+	ret = sw842_data2(inbuf, inbit, outbuf, fifo);
+	return ret;
+}
+
+static int sw842_data2(const char **inbuf, int *inbit,
+		unsigned char **outbuf, struct sw842_fifo *fifo)
+{
+	**outbuf = sw842_get_byte(*inbuf, *inbit);
+	(*inbuf)++;
+	(*outbuf)++;
+	**outbuf = sw842_get_byte(*inbuf, *inbit);
+	(*inbuf)++;
+	(*outbuf)++;
+	return 0;
+}
+
+static int sw842_ptr8(const char **inbuf, int *inbit,
+		unsigned char **outbuf, struct sw842_fifo *fifo)
+{
+	uint8_t ptr;
+	ptr = sw842_get_ptr8(inbuf, inbit, fifo);
+	if (!fifo->f84_full && (ptr >= fifo->f8_count))
+		return 1;
+	memcpy(*outbuf, fifo->f8[ptr], 8);
+	*outbuf += 8;
+	return 0;
+}
+
+static int sw842_ptr4(const char **inbuf, int *inbit,
+		unsigned char **outbuf, struct sw842_fifo *fifo)
+{
+	uint16_t ptr;
+	ptr = sw842_get_ptr4(inbuf, inbit, fifo);
+	if (!fifo->f84_full && (ptr >= fifo->f4_count))
+		return 1;
+	memcpy(*outbuf, fifo->f4[ptr], 4);
+	*outbuf += 4;
+	return 0;
+}
+
+static int sw842_ptr2(const char **inbuf, int *inbit,
+		unsigned char **outbuf, struct sw842_fifo *fifo)
+{
+	uint8_t ptr;
+	ptr = sw842_get_ptr2(inbuf, inbit);
+	if (!fifo->f2_full && (ptr >= fifo->f2_count))
+		return 1;
+	memcpy(*outbuf, fifo->f2[ptr], 2);
+	*outbuf += 2;
+	return 0;
+}
+
+static void sw842_copy_to_fifo(const char *buf, struct sw842_fifo *fifo)
+{
+	unsigned char initial_f2count = fifo->f2_count;
+
+	memcpy(fifo->f8[fifo->f8_count], buf, 8);
+	fifo->f4_count += 2;
+	fifo->f8_count += 1;
+
+	if (!fifo->f84_full && fifo->f4_count >= 512) {
+		fifo->f84_full = 1;
+		fifo->f4_count /= 512;
+	}
+
+	memcpy(fifo->f2[fifo->f2_count++], buf, 2);
+	memcpy(fifo->f2[fifo->f2_count++], buf + 2, 2);
+	memcpy(fifo->f2[fifo->f2_count++], buf + 4, 2);
+	memcpy(fifo->f2[fifo->f2_count++], buf + 6, 2);
+	if (fifo->f2_count < initial_f2count)
+		fifo->f2_full = 1;
+}
+
+static int sw842_decompress(const unsigned char *src, int srclen,
+			unsigned char *dst, int *destlen,
+			const void *wrkmem)
+{
+	uint8_t tmpl;
+	const char *inbuf;
+	int inbit = 0;
+	unsigned char *outbuf, *outbuf_end, *origbuf, *prevbuf;
+	const char *inbuf_end;
+	sw842_template_op op;
+	int opindex;
+	int i, repeat_count;
+	struct sw842_fifo *fifo;
+	int ret = 0;
+
+	fifo = &((struct nx842_workmem *)(wrkmem))->swfifo;
+	memset(fifo, 0, sizeof(*fifo));
+
+	origbuf = NULL;
+	inbuf = src;
+	inbuf_end = src + srclen;
+	outbuf = dst;
+	outbuf_end = dst + *destlen;
+
+	while ((tmpl = sw842_get_template(&inbuf, &inbit)) != SW842_TMPL_EOF) {
+		if (inbuf >= inbuf_end) {
+			ret = -EINVAL;
+			goto out;
+		}
+
+		opindex = 0;
+		prevbuf = origbuf;
+		origbuf = outbuf;
+		switch (tmpl) {
+		case SW842_TMPL_REPEAT:
+			if (prevbuf == NULL) {
+				ret = -EINVAL;
+				goto out;
+			}
+
+			repeat_count = sw842_get_repeat_count(&inbuf,
+								&inbit) + 1;
+
+			/* Did the repeat count advance past the end of input */
+			if (inbuf > inbuf_end) {
+				ret = -EINVAL;
+				goto out;
+			}
+
+			for (i = 0; i < repeat_count; i++) {
+				/* Would this overflow the output buffer */
+				if ((outbuf + 8) > outbuf_end) {
+					ret = -ENOSPC;
+					goto out;
+				}
+
+				memcpy(outbuf, prevbuf, 8);
+				sw842_copy_to_fifo(outbuf, fifo);
+				outbuf += 8;
+			}
+			break;
+
+		case SW842_TMPL_ZEROS:
+			/* Would this overflow the output buffer */
+			if ((outbuf + 8) > outbuf_end) {
+				ret = -ENOSPC;
+				goto out;
+			}
+
+			memset(outbuf, 0, 8);
+			sw842_copy_to_fifo(outbuf, fifo);
+			outbuf += 8;
+			break;
+
+		default:
+			if (tmpl > 25) {
+				ret = -EINVAL;
+				goto out;
+			}
+
+			/* Does this go past the end of the input buffer */
+			if ((inbuf + 2) > inbuf_end) {
+				ret = -EINVAL;
+				goto out;
+			}
+
+			/* Would this overflow the output buffer */
+			if ((outbuf + 8) > outbuf_end) {
+				ret = -ENOSPC;
+				goto out;
+			}
+
+			while (opindex < 4 &&
+				(op = sw842_tmpl_ops[tmpl][opindex++])
+					!= NULL) {
+				ret = (*op)(&inbuf, &inbit, &outbuf, fifo);
+				if (ret) {
+					ret = -EINVAL;
+					goto out;
+				}
+				sw842_copy_to_fifo(origbuf, fifo);
+			}
+		}
+	}
+
+out:
+	if (!ret)
+		*destlen = (unsigned int)(outbuf - dst);
+	else
+		*destlen = 0;
+
+	return ret;
+}
diff --git a/include/linux/nx842.h b/include/linux/nx842.h
new file mode 100644
index 0000000..a4d324c
--- /dev/null
+++ b/include/linux/nx842.h
@@ -0,0 +1,11 @@ 
+#ifndef __NX842_H__
+#define __NX842_H__
+
+int nx842_get_workmem_size(void);
+int nx842_get_workmem_size_aligned(void);
+int nx842_compress(const unsigned char *in, unsigned int in_len,
+		unsigned char *out, unsigned int *out_len, void *wrkmem);
+int nx842_decompress(const unsigned char *in, unsigned int in_len,
+		unsigned char *out, unsigned int *out_len, void *wrkmem);
+
+#endif