nfsd: add SCSI layout support

This is a simple extension to the block layout driver to use SCSI
persistent reservations for access control and fencing, as well as
SCSI VPD pages for device identification.

For this we need to pass the nfs4_client to the proc_getdeviceinfo method
to generate the reservation key, and add a new fence_client method
to allow for fence actions in the layout driver.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>

Documentation/filesystems/nfs/pnfs-scsi-server.txt

+
+pNFS SCSI layout server user guide
+==================================
+
+This document describes support for pNFS SCSI layouts in the Linux NFS server.
+With pNFS SCSI layouts, the NFS server acts as Metadata Server (MDS) for pNFS,
+which in addition to handling all the metadata access to the NFS export,
+also hands out layouts to the clients so that they can directly access the
+underlying SCSI LUNs that are shared with the client.
+
+To use pNFS SCSI layouts with with the Linux NFS server, the exported file
+system needs to support the pNFS SCSI layouts (currently just XFS), and the
+file system must sit on a SCSI LUN that is accessible to the clients in
+addition to the MDS.  As of now the file system needs to sit directly on the
+exported LUN, striping or concatenation of LUNs on the MDS and clients
+is not supported yet.
+
+On a server built with CONFIG_NFSD_SCSI, the pNFS SCSI volume support is
+automatically enabled if the file system is exported using the "pnfs"
+option and the underlying SCSI device support persistent reservations.
+On the client make sure the kernel has the CONFIG_PNFS_BLOCK option
+enabled, and the file system is mounted using the NFSv4.1 protocol
+version (mount -o vers=4.1).

fs/nfsd/Kconfig

@@ -98,6 +98,19 @@ config NFSD_BLOCKLAYOUT
 
 	  If unsure, say N.
 
+config NFSD_SCSILAYOUT
+	bool "NFSv4.1 server support for pNFS SCSI layouts"
+	depends on NFSD_V4
+	select NFSD_PNFS
+	help
+	  This option enables support for the exporting pNFS SCSI layouts
+	  in the kernel's NFS server. The pNFS SCSI layout enables NFS
+	  clients to directly perform I/O to SCSI devices accesible to both
+	  the server and the clients.  See draft-ietf-nfsv4-scsi-layout for
+	  more details.
+
+	  If unsure, say N.
+
 config NFSD_V4_SECURITY_LABEL
 	bool "Provide Security Label support for NFSv4 server"
 	depends on NFSD_V4 && SECURITY

fs/nfsd/Makefile

@@ -19,3 +19,4 @@ nfsd-$(CONFIG_NFSD_V4)	+= nfs4proc.o nfs4xdr.o nfs4state.o nfs4idmap.o \
 			   nfs4acl.o nfs4callback.o nfs4recover.o
 nfsd-$(CONFIG_NFSD_PNFS) += nfs4layouts.o
 nfsd-$(CONFIG_NFSD_BLOCKLAYOUT) += blocklayout.o blocklayoutxdr.o
+nfsd-$(CONFIG_NFSD_SCSILAYOUT) += blocklayout.o blocklayoutxdr.o

fs/nfsd/blocklayout.c

 /*
- * Copyright (c) 2014 Christoph Hellwig.
+ * Copyright (c) 2014-2016 Christoph Hellwig.
  */
 #include <linux/exportfs.h>
 #include <linux/genhd.h>
 #include <linux/slab.h>
+#include <linux/pr.h>
 
 #include <linux/nfsd/debug.h>
+#include <scsi/scsi_proto.h>
+#include <scsi/scsi_common.h>
 
 #include "blocklayoutxdr.h"
 #include "pnfs.h"
@@ -159,6 +162,7 @@ nfsd4_block_get_device_info_simple(struct super_block *sb,
 
 static __be32
 nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
+		struct nfs4_client *clp,
 		struct nfsd4_getdeviceinfo *gdp)
 {
 	if (sb->s_bdev != sb->s_bdev->bd_contains)
@@ -200,3 +204,205 @@ const struct nfsd4_layout_ops bl_layout_ops = {
 	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
 };
 #endif /* CONFIG_NFSD_BLOCKLAYOUT */
+
+#ifdef CONFIG_NFSD_SCSILAYOUT
+static int nfsd4_scsi_identify_device(struct block_device *bdev,
+		struct pnfs_block_volume *b)
+{
+	struct request_queue *q = bdev->bd_disk->queue;
+	struct request *rq;
+	size_t bufflen = 252, len, id_len;
+	u8 *buf, *d, type, assoc;
+	int error;
+
+	buf = kzalloc(bufflen, GFP_KERNEL);
+	if (!buf)
+		return -ENOMEM;
+
+	rq = blk_get_request(q, READ, GFP_KERNEL);
+	if (IS_ERR(rq)) {
+		error = -ENOMEM;
+		goto out_free_buf;
+	}
+	blk_rq_set_block_pc(rq);
+
+	error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL);
+	if (error)
+		goto out_put_request;
+
+	rq->cmd[0] = INQUIRY;
+	rq->cmd[1] = 1;
+	rq->cmd[2] = 0x83;
+	rq->cmd[3] = bufflen >> 8;
+	rq->cmd[4] = bufflen & 0xff;
+	rq->cmd_len = COMMAND_SIZE(INQUIRY);
+
+	error = blk_execute_rq(rq->q, NULL, rq, 1);
+	if (error) {
+		pr_err("pNFS: INQUIRY 0x83 failed with: %x\n",
+			rq->errors);
+		goto out_put_request;
+	}
+
+	len = (buf[2] << 8) + buf[3] + 4;
+	if (len > bufflen) {
+		pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n",
+			len);
+		goto out_put_request;
+	}
+
+	d = buf + 4;
+	for (d = buf + 4; d < buf + len; d += id_len + 4) {
+		id_len = d[3];
+		type = d[1] & 0xf;
+		assoc = (d[1] >> 4) & 0x3;
+
+		/*
+		 * We only care about a EUI-64 and NAA designator types
+		 * with LU association.
+		 */
+		if (assoc != 0x00)
+			continue;
+		if (type != 0x02 && type != 0x03)
+			continue;
+		if (id_len != 8 && id_len != 12 && id_len != 16)
+			continue;
+
+		b->scsi.code_set = PS_CODE_SET_BINARY;
+		b->scsi.designator_type = type == 0x02 ?
+			PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA;
+		b->scsi.designator_len = id_len;
+		memcpy(b->scsi.designator, d + 4, id_len);
+
+		/*
+		 * If we found a 8 or 12 byte descriptor continue on to
+		 * see if a 16 byte one is available.  If we find a
+		 * 16 byte descriptor we're done.
+		 */
+		if (id_len == 16)
+			break;
+	}
+
+out_put_request:
+	blk_put_request(rq);
+out_free_buf:
+	kfree(buf);
+	return error;
+}
+
+#define NFSD_MDS_PR_KEY		0x0100000000000000
+
+/*
+ * We use the client ID as a unique key for the reservations.
+ * This allows us to easily fence a client when recalls fail.
+ */
+static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
+{
+	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
+}
+
+static int
+nfsd4_block_get_device_info_scsi(struct super_block *sb,
+		struct nfs4_client *clp,
+		struct nfsd4_getdeviceinfo *gdp)
+{
+	struct pnfs_block_deviceaddr *dev;
+	struct pnfs_block_volume *b;
+	const struct pr_ops *ops;
+	int error;
+
+	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
+		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
+	if (!dev)
+		return -ENOMEM;
+	gdp->gd_device = dev;
+
+	dev->nr_volumes = 1;
+	b = &dev->volumes[0];
+
+	b->type = PNFS_BLOCK_VOLUME_SCSI;
+	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
+
+	error = nfsd4_scsi_identify_device(sb->s_bdev, b);
+	if (error)
+		return error;
+
+	ops = sb->s_bdev->bd_disk->fops->pr_ops;
+	if (!ops) {
+		pr_err("pNFS: device %s does not support PRs.\n",
+			sb->s_id);
+		return -EINVAL;
+	}
+
+	error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
+	if (error) {
+		pr_err("pNFS: failed to register key for device %s.\n",
+			sb->s_id);
+		return -EINVAL;
+	}
+
+	error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
+			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
+	if (error) {
+		pr_err("pNFS: failed to reserve device %s.\n",
+			sb->s_id);
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+static __be32
+nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
+		struct nfs4_client *clp,
+		struct nfsd4_getdeviceinfo *gdp)
+{
+	if (sb->s_bdev != sb->s_bdev->bd_contains)
+		return nfserr_inval;
+	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
+}
+static __be32
+nfsd4_scsi_proc_layoutcommit(struct inode *inode,
+		struct nfsd4_layoutcommit *lcp)
+{
+	struct iomap *iomaps;
+	int nr_iomaps;
+
+	nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout,
+			lcp->lc_up_len, &iomaps, 1 << inode->i_blkbits);
+	if (nr_iomaps < 0)
+		return nfserrno(nr_iomaps);
+
+	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
+}
+
+static void
+nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls)
+{
+	struct nfs4_client *clp = ls->ls_stid.sc_client;
+	struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev;
+
+	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
+			nfsd4_scsi_pr_key(clp), 0, true);
+}
+
+const struct nfsd4_layout_ops scsi_layout_ops = {
+	/*
+	 * Pretend that we send notification to the client.  This is a blatant
+	 * lie to force recent Linux clients to cache our device IDs.
+	 * We rarely ever change the device ID, so the harm of leaking deviceids
+	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
+	 * in this regard, but I filed errata 4119 for this a while ago, and
+	 * hopefully the Linux client will eventually start caching deviceids
+	 * without this again.
+	 */
+	.notify_types		=
+			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
+	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
+	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
+	.proc_layoutget		= nfsd4_block_proc_layoutget,
+	.encode_layoutget	= nfsd4_block_encode_layoutget,
+	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
+	.fence_client		= nfsd4_scsi_fence_client,
+};
+#endif /* CONFIG_NFSD_SCSILAYOUT */

fs/nfsd/blocklayoutxdr.c

 /*
- * Copyright (c) 2014 Christoph Hellwig.
+ * Copyright (c) 2014-2016 Christoph Hellwig.
  */
 #include <linux/sunrpc/svc.h>
 #include <linux/exportfs.h>
@@ -53,6 +53,18 @@ nfsd4_block_encode_volume(struct xdr_stream *xdr, struct pnfs_block_volume *b)
 		p = xdr_encode_hyper(p, b->simple.offset);
 		p = xdr_encode_opaque(p, b->simple.sig, b->simple.sig_len);
 		break;
+	case PNFS_BLOCK_VOLUME_SCSI:
+		len = 4 + 4 + 4 + 4 + b->scsi.designator_len + 8;
+		p = xdr_reserve_space(xdr, len);
+		if (!p)
+			return -ETOOSMALL;
+
+		*p++ = cpu_to_be32(b->type);
+		*p++ = cpu_to_be32(b->scsi.code_set);
+		*p++ = cpu_to_be32(b->scsi.designator_type);
+		p = xdr_encode_opaque(p, b->scsi.designator, b->scsi.designator_len);
+		p = xdr_encode_hyper(p, b->scsi.pr_key);
+		break;
 	default:
 		return -ENOTSUPP;
 	}
@@ -155,3 +167,54 @@ nfsd4_block_decode_layoutupdate(__be32 *p, u32 len, struct iomap **iomapp,
 	kfree(iomaps);
 	return -EINVAL;
 }
+
+int
+nfsd4_scsi_decode_layoutupdate(__be32 *p, u32 len, struct iomap **iomapp,
+		u32 block_size)
+{
+	struct iomap *iomaps;
+	u32 nr_iomaps, expected, i;
+
+	if (len < sizeof(u32)) {
+		dprintk("%s: extent array too small: %u\n", __func__, len);
+		return -EINVAL;
+	}
+
+	nr_iomaps = be32_to_cpup(p++);
+	expected = sizeof(__be32) + nr_iomaps * PNFS_SCSI_RANGE_SIZE;
+	if (len != expected) {
+		dprintk("%s: extent array size mismatch: %u/%u\n",
+			__func__, len, expected);
+		return -EINVAL;
+	}
+
+	iomaps = kcalloc(nr_iomaps, sizeof(*iomaps), GFP_KERNEL);
+	if (!iomaps) {
+		dprintk("%s: failed to allocate extent array\n", __func__);
+		return -ENOMEM;
+	}
+
+	for (i = 0; i < nr_iomaps; i++) {
+		u64 val;
+
+		p = xdr_decode_hyper(p, &val);
+		if (val & (block_size - 1)) {
+			dprintk("%s: unaligned offset 0x%llx\n", __func__, val);
+			goto fail;
+		}
+		iomaps[i].offset = val;
+
+		p = xdr_decode_hyper(p, &val);
+		if (val & (block_size - 1)) {
+			dprintk("%s: unaligned length 0x%llx\n", __func__, val);
+			goto fail;
+		}
+		iomaps[i].length = val;
+	}
+
+	*iomapp = iomaps;
+	return nr_iomaps;
+fail:
+	kfree(iomaps);
+	return -EINVAL;
+}

fs/nfsd/blocklayoutxdr.h

@@ -15,6 +15,11 @@ struct pnfs_block_extent {
 	enum pnfs_block_extent_state	es;
 };
 
+struct pnfs_block_range {
+	u64				foff;
+	u64				len;
+};
+
 /*
  * Random upper cap for the uuid length to avoid unbounded allocation.
  * Not actually limited by the protocol.
@@ -29,6 +34,13 @@ struct pnfs_block_volume {
 			u32		sig_len;
 			u8		sig[PNFS_BLOCK_UUID_LEN];
 		} simple;
+		struct {
+			enum scsi_code_set		code_set;
+			enum scsi_designator_type	designator_type;
+			int				designator_len;
+			u8				designator[256];
+			u64				pr_key;
+		} scsi;
 	};
 };
 
@@ -43,5 +55,7 @@ __be32 nfsd4_block_encode_layoutget(struct xdr_stream *xdr,
 		struct nfsd4_layoutget *lgp);
 int nfsd4_block_decode_layoutupdate(__be32 *p, u32 len, struct iomap **iomapp,
 		u32 block_size);
+int nfsd4_scsi_decode_layoutupdate(__be32 *p, u32 len, struct iomap **iomapp,
+		u32 block_size);
 
 #endif /* _NFSD_BLOCKLAYOUTXDR_H */

fs/nfsd/nfs4layouts.c

 /*
  * Copyright (c) 2014 Christoph Hellwig.
  */
+#include <linux/blkdev.h>
 #include <linux/kmod.h>
 #include <linux/file.h>
 #include <linux/jhash.h>
@@ -29,6 +30,9 @@ const struct nfsd4_layout_ops *nfsd4_layout_ops[LAYOUT_TYPE_MAX] =  {
 #ifdef CONFIG_NFSD_BLOCKLAYOUT
 	[LAYOUT_BLOCK_VOLUME]	= &bl_layout_ops,
 #endif
+#ifdef CONFIG_NFSD_SCSILAYOUT
+	[LAYOUT_SCSI]		= &scsi_layout_ops,
+#endif
 };
 
 /* pNFS device ID to export fsid mapping */
@@ -123,12 +127,24 @@ void nfsd4_setup_layout_type(struct svc_export *exp)
 	if (!(exp->ex_flags & NFSEXP_PNFS))
 		return;
 
+	/*
+	 * Check if the file system supports exporting a block-like layout.
+	 * If the block device supports reservations prefer the SCSI layout,
+	 * otherwise advertise the block layout.
+	 */
 #ifdef CONFIG_NFSD_BLOCKLAYOUT
 	if (sb->s_export_op->get_uuid &&
 	    sb->s_export_op->map_blocks &&
 	    sb->s_export_op->commit_blocks)
 		exp->ex_layout_type = LAYOUT_BLOCK_VOLUME;
 #endif
+#ifdef CONFIG_NFSD_SCSILAYOUT
+	/* overwrite block layout selection if needed */
+	if (sb->s_export_op->map_blocks &&
+	    sb->s_export_op->commit_blocks &&
+	    sb->s_bdev && sb->s_bdev->bd_disk->fops->pr_ops)
+		exp->ex_layout_type = LAYOUT_SCSI;
+#endif
 }
 
 static void
@@ -594,8 +610,6 @@ nfsd4_cb_layout_fail(struct nfs4_layout_stateid *ls)
 
 	rpc_ntop((struct sockaddr *)&clp->cl_addr, addr_str, sizeof(addr_str));
 
-	trace_layout_recall_fail(&ls->ls_stid.sc_stateid);
-
 	printk(KERN_WARNING
 		"nfsd: client %s failed to respond to layout recall. "
 		"  Fencing..\n", addr_str);
@@ -630,6 +644,7 @@ nfsd4_cb_layout_done(struct nfsd4_callback *cb, struct rpc_task *task)
 		container_of(cb, struct nfs4_layout_stateid, ls_recall);
 	struct nfsd_net *nn;
 	ktime_t now, cutoff;
+	const struct nfsd4_layout_ops *ops;
 	LIST_HEAD(reaplist);
 
 
@@ -665,6 +680,12 @@ nfsd4_cb_layout_done(struct nfsd4_callback *cb, struct rpc_task *task)
 		/*
 		 * Unknown error or non-responding client, we'll need to fence.
 		 */
+		trace_layout_recall_fail(&ls->ls_stid.sc_stateid);
+
+		ops = nfsd4_layout_ops[ls->ls_layout_type];
+		if (ops->fence_client)
+			ops->fence_client(ls);
+		else
 			nfsd4_cb_layout_fail(ls);
 		return -1;
 	}

fs/nfsd/nfs4proc.c

@@ -1268,8 +1268,10 @@ nfsd4_getdeviceinfo(struct svc_rqst *rqstp,
 		goto out;
 
 	nfserr = nfs_ok;
-	if (gdp->gd_maxcount != 0)
-		nfserr = ops->proc_getdeviceinfo(exp->ex_path.mnt->mnt_sb, gdp);
+	if (gdp->gd_maxcount != 0) {
+		nfserr = ops->proc_getdeviceinfo(exp->ex_path.mnt->mnt_sb,
+					cstate->session->se_client, gdp);
+	}
 
 	gdp->gd_notify_types &= ops->notify_types;
 out:

fs/nfsd/pnfs.h

@@ -21,6 +21,7 @@ struct nfsd4_layout_ops {
 	u32		notify_types;
 
 	__be32 (*proc_getdeviceinfo)(struct super_block *sb,
+			struct nfs4_client *clp,
 			struct nfsd4_getdeviceinfo *gdevp);
 	__be32 (*encode_getdeviceinfo)(struct xdr_stream *xdr,
 			struct nfsd4_getdeviceinfo *gdevp);
@@ -32,12 +33,17 @@ struct nfsd4_layout_ops {
 
 	__be32 (*proc_layoutcommit)(struct inode *inode,
 			struct nfsd4_layoutcommit *lcp);
+
+	void (*fence_client)(struct nfs4_layout_stateid *ls);
 };
 
 extern const struct nfsd4_layout_ops *nfsd4_layout_ops[];
 #ifdef CONFIG_NFSD_BLOCKLAYOUT
 extern const struct nfsd4_layout_ops bl_layout_ops;
 #endif
+#ifdef CONFIG_NFSD_SCSILAYOUT
+extern const struct nfsd4_layout_ops scsi_layout_ops;
+#endif
 
 __be32 nfsd4_preprocess_layout_stateid(struct svc_rqst *rqstp,
 		struct nfsd4_compound_state *cstate, stateid_t *stateid,

fs/xfs/Makefile

@@ -122,3 +122,4 @@ xfs-$(CONFIG_XFS_POSIX_ACL)	+= xfs_acl.o
 xfs-$(CONFIG_SYSCTL)		+= xfs_sysctl.o
 xfs-$(CONFIG_COMPAT)		+= xfs_ioctl32.o
 xfs-$(CONFIG_NFSD_BLOCKLAYOUT)	+= xfs_pnfs.o
+xfs-$(CONFIG_NFSD_SCSILAYOUT)	+= xfs_pnfs.o

fs/xfs/xfs_pnfs.h

 #ifndef _XFS_PNFS_H
 #define _XFS_PNFS_H 1
 
-#ifdef CONFIG_NFSD_BLOCKLAYOUT
+#if defined(CONFIG_NFSD_BLOCKLAYOUT) || defined(CONFIG_NFSD_SCSILAYOUT)
 int xfs_fs_get_uuid(struct super_block *sb, u8 *buf, u32 *len, u64 *offset);
 int xfs_fs_map_blocks(struct inode *inode, loff_t offset, u64 length,
 		struct iomap *iomap, bool write, u32 *device_generation);