Skip to contentSkip to conversation

pNFS generic device ID cache

PATCH In review [email protected] 3 participants 10 messages Apr 16, 2010 – Apr 22, 2010 3 messages not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover pNFS generic device ID cache

This patch set implements a shared RCU device ID cache servicing multiple mounts of a single layout type per meta data server (struct nfs_client).

0001-pnfs_submit-generic-device-ID-cache.patch 0002-pnfs_submit-fix-multiple-mount-set_pnfs_layoutdriver.patch 0003-pnfs-submit-file-layout-driver-generic-device-ID-cac.patch

These patches apply to the 2.6.34-rc3 pnfs branch below the following patch which is the head of the revert patches - otherwise known as the first patch not included in the pnfs file layout driver submission.

9a75b7356ca22b10903b507383646201cdcc9020

pnfs: filelayout: CB_NOTIFY_DEVICE support

This reverts commit 933b55c15ed2b7fd89d8e0342bd5b2825726201a "SQUASHME pnfs_submit: remove filelayout CB_NOTIFY_DEVICE support"

Testing:

CONFIG_NFS_V4_1 set:

NFSv4.1/pNFS mounts:
Connectathon tests pass against GFS2/pNFS with a single AUTH_SYS mount, a double AUTH_SYS mount, and an AUTH_SYS and AUTH_GSS/KRB5 mount (which creates two superblocks under a struct nfs_client and both share the device id cache).

NFSv4.0 mount;
Connectathon tests pass

Did not test with multiple device ID's. I will create a mulitple device ID test with the pynfs file layout server.

CONFIG_NFS_V4_1 not set:

NFSv4.0 mount: Connectathon tests pass.

-->Andy

#1 lore raw

PATCH 1/3 SQUASHME pnfs_submit: generic device ID cache

From: Andy Adamson <[email protected]>

A shared RCU device ID cache servicing multiple mounts of a single layout type per meta data server (struct nfs_client).

Device IDs of type deviceid4 are required by all layout types, long lived and read at each I/O. They are added to the deviceid cache at first reference by a layout via GETDEVICEINFO and (currently) are only removed at umount.

Reference count the device ID cache for each mounted file system in the initialize_mountpoint layoutdriver_io_operation.

Dereference the device id cache on file system in the uninitialize_mountpoint layoutdriver_io_operation called at umount

---
 fs/nfs/pnfs.c             |  119 +++++++++++++++++++++++++++++++++++++++++++++
 include/linux/nfs4_pnfs.h |   27 ++++++++++
 include/linux/nfs_fs_sb.h |    1 +
 3 files changed, 147 insertions(+), 0 deletions(-)
fs/nfs/pnfs.c +119 -0
index 91572aa..8492aef 100644
@@ -45,6 +45,7 @@
 #include <linux/nfs4.h>
 #include <linux/pnfs_xdr.h>
 #include <linux/nfs4_pnfs.h>
+#include <linux/rculist.h>
 
 #include "internal.h"
 #include "nfs4_fs.h"
@@ -2296,3 +2297,121 @@ struct pnfs_client_operations pnfs_ops = {
 
 EXPORT_SYMBOL(pnfs_unregister_layoutdriver);
 EXPORT_SYMBOL(pnfs_register_layoutdriver);
+
+
+/* Device ID cache. Supports one layout type per struct nfs_client */
+int
+nfs4_alloc_init_deviceid_cache(struct nfs_client *clp,
+			 void (*free_callback)(struct rcu_head *))
+{
+	struct nfs4_deviceid_cache *c;
+
+	c = kzalloc(sizeof(struct nfs4_deviceid_cache), GFP_KERNEL);
+	if (!c)
+		return -ENOMEM;
+	spin_lock(&clp->cl_lock);
+	if (clp->cl_devid_cache != NULL) {
+		kref_get(&clp->cl_devid_cache->dc_kref);
+		spin_unlock(&clp->cl_lock);
+		dprintk("%s [kref [%d]]\n", __func__,
+			atomic_read(&clp->cl_devid_cache->dc_kref.refcount));
+		kfree(c);
+	} else {
+		spin_lock_init(&c->dc_lock);
+		INIT_LIST_HEAD(&c->dc_deviceids);
+		kref_init(&c->dc_kref);
+		c->dc_free_callback = free_callback;
+		c->dc_clp = clp;
+		clp->cl_devid_cache = c;
+		spin_unlock(&clp->cl_lock);
+		dprintk("%s [new]\n", __func__);
+	}
+	return 0;
+}
+EXPORT_SYMBOL(nfs4_alloc_init_deviceid_cache);
+
+void
+nfs4_init_deviceid_node(struct nfs4_deviceid *d)
+{
+	INIT_LIST_HEAD(&d->de_node);
+	INIT_RCU_HEAD(&d->de_rcu);
+}
+EXPORT_SYMBOL(nfs4_init_deviceid_node);
+
+struct nfs4_deviceid *
+nfs4_find_deviceid(struct nfs4_deviceid_cache *c, struct pnfs_deviceid *id)
+{
+	struct nfs4_deviceid *d;
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+		if (!memcmp(&d->de_id, id, NFS4_PNFS_DEVICEID4_SIZE)) {
+			rcu_read_unlock();
+			return d;
+		}
+	}
+	rcu_read_unlock();
+	return NULL;
+}
+EXPORT_SYMBOL(nfs4_find_deviceid);
+
+/*
+ * Add or kref_get a deviceid.
+ * GETDEVICEINFOs for same deviceid can race. If deviceid is found, discard new
+ */
+void
+nfs4_add_deviceid(struct nfs4_deviceid_cache *c, struct nfs4_deviceid *new)
+{
+	struct nfs4_deviceid *d;
+
+	spin_lock(&c->dc_lock);
+	list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+		if (!memcmp(&d->de_id, &new->de_id, NFS4_PNFS_DEVICEID4_SIZE)) {
+			spin_unlock(&c->dc_lock);
+			dprintk("%s [discard]\n", __func__);
+			c->dc_free_callback(&new->de_rcu);
+		}
+	}
+	list_add_rcu(&new->de_node, &c->dc_deviceids);
+	spin_unlock(&c->dc_lock);
+	dprintk("%s [new]\n", __func__);
+}
+EXPORT_SYMBOL(nfs4_add_deviceid);
+
+static int
+nfs4_remove_deviceid(struct nfs4_deviceid_cache *c)
+{
+	struct nfs4_deviceid *d;
+
+	spin_lock(&c->dc_lock);
+	list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+		list_del_rcu(&d->de_node);
+		spin_unlock(&c->dc_lock);
+		synchronize_rcu();
+		c->dc_free_callback(&d->de_rcu);
+		return 1;
+	}
+	spin_unlock(&c->dc_lock);
+	return 0;
+}
+
+static void
+nfs4_free_deviceid_cache(struct kref *kref)
+{
+	struct nfs4_deviceid_cache *cache =
+		container_of(kref, struct nfs4_deviceid_cache, dc_kref);
+	int more = 1;
+
+	while (more)
+		more = nfs4_remove_deviceid(cache);
+	cache->dc_clp->cl_devid_cache = NULL;
+	kfree(cache);
+}
+
+void
+nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *c)
+{
+	dprintk("%s [%d]\n", __func__, atomic_read(&c->dc_kref.refcount));
+	kref_put(&c->dc_kref, nfs4_free_deviceid_cache);
+}
+EXPORT_SYMBOL(nfs4_put_deviceid_cache);
include/linux/nfs4_pnfs.h +27 -0
index 1d521f4..2a88a06 100644
@@ -281,6 +281,33 @@ struct pnfs_devicelist {
 	struct pnfs_deviceid	dev_id[NFS4_PNFS_GETDEVLIST_MAXNUM];
 };
 
+/*
+ * Device ID RCU cache. A device ID is unique per client ID and layout type.
+ */
+struct nfs4_deviceid_cache {
+	spinlock_t		dc_lock;
+	struct list_head	dc_deviceids;
+	struct kref		dc_kref;
+	void			(*dc_free_callback)(struct rcu_head *);
+	struct nfs_client	*dc_clp;
+};
+
+/* Device ID cache node */
+struct nfs4_deviceid {
+	struct list_head	de_node;
+	struct rcu_head		de_rcu;
+	struct pnfs_deviceid	de_id;
+};
+
+extern int nfs4_alloc_init_deviceid_cache(struct nfs_client *,
+				void (*free_callback)(struct rcu_head *));
+extern void nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *);
+extern void nfs4_init_deviceid_node(struct nfs4_deviceid *);
+extern struct nfs4_deviceid *nfs4_find_deviceid(struct nfs4_deviceid_cache *,
+				struct pnfs_deviceid *);
+extern void nfs4_add_deviceid(struct nfs4_deviceid_cache *,
+				struct nfs4_deviceid *);
+
 /* pNFS client callback functions.
  * These operations allow the layout driver to access pNFS client
  * specific information or call pNFS client->server operations.
include/linux/nfs_fs_sb.h +1 -0
index 8522461..ef2e18e 100644
@@ -87,6 +87,7 @@ struct nfs_client {
 	u32			cl_exchange_flags;
 	struct nfs4_session	*cl_session; 	/* sharred session */
 	struct list_head	cl_lo_inodes;	/* Inodes having layouts */
+	struct nfs4_deviceid_cache *cl_devid_cache; /* pNFS deviceid cache */
 #endif /* CONFIG_NFS_V4_1 */
 
 #ifdef CONFIG_NFS_FSCACHE
Signature · 1 line
1.6.6
#2 lore raw

PATCH 2/3 SQUASHME pnfs_submit: fix multiple mount set_pnfs_layoutdriver

From: Andy Adamson <[email protected]>

The same struct nfs_server can enter set_pnfs_layoutdriver for mounts that share a super block. Don't initialize a pnfs mountpoint more than once.

Don't set the pnfs_curr_ld until the pnfs mountpoint initialization succeeds

---
 fs/nfs/pnfs.c |   15 ++++++++++-----
 1 files changed, 10 insertions(+), 5 deletions(-)
fs/nfs/pnfs.c +10 -5
index 8492aef..1d903fd 100644
@@ -215,20 +215,25 @@ set_pnfs_layoutdriver(struct super_block *sb, struct nfs_fh *fh, u32 id)
 	struct pnfs_mount_type *mt;
 	struct nfs_server *server = NFS_SB(sb);
 
+	if (server->pnfs_curr_ld)
+		return;
+
 	if (id > 0 && find_pnfs(id, &mod)) {
-		dprintk("%s: Setting pNFS module\n", __func__);
-		server->pnfs_curr_ld = mod->pnfs_ld_type;
-		mt = server->pnfs_curr_ld->ld_io_ops->initialize_mountpoint(
+		mt = mod->pnfs_ld_type->ld_io_ops->initialize_mountpoint(
 			sb, fh);
 		if (!mt) {
 			printk(KERN_ERR "%s: Error initializing mount point "
 			       "for layout driver %u. ", __func__, id);
 			goto out_err;
 		}
-		/* Layout driver succeeded in initializing mountpoint */
+		/*
+		 * Layout driver succeeded in initializing mountpoint
+		 * and has taken a reference on the nfs_client cl_devid_cache
+		 */
+		server->pnfs_curr_ld = mod->pnfs_ld_type;
 		server->pnfs_mountid = mt;
-		/* Set the rpc_ops */
 		server->nfs_client->rpc_ops = &pnfs_v4_clientops;
+		dprintk("%s: pNFS module for %u set\n", __func__, id);
 		return;
 	}
 
Signature · 1 line
1.6.6
#3 lore raw

PATCH 3/3 SQUASHME pnfs-submit: file layout driver generic device ID cache

From: Andy Adamson <[email protected]>

Replace the per superblock deviceid cache with the generic deviceid cache.

Embed struct nfs4_deviceid into struct nfs4_file_layout_dsaddr, the file layout specific deviceid structure. Provide a free_deviceid_callback.

---
 fs/nfs/client.c            |    1 +
 fs/nfs/nfs4filelayout.c    |   31 +++-----
 fs/nfs/nfs4filelayout.h    |   12 +--
 fs/nfs/nfs4filelayoutdev.c |  190 ++++++++------------------------------------
 4 files changed, 51 insertions(+), 183 deletions(-)
fs/nfs/client.c +1 -0
index e1d1646..82775b7 100644
@@ -38,6 +38,7 @@
 #include <net/ipv6.h>
 #include <linux/nfs_xdr.h>
 #include <linux/sunrpc/bc_xprt.h>
+#include <linux/nfs4_pnfs.h>
 
 #include <asm/system.h>
 
fs/nfs/nfs4filelayout.c +12 -19
index 0530b59..c155586 100644
@@ -76,17 +76,11 @@ filelayout_initialize_mountpoint(struct super_block *sb, struct nfs_fh *fh)
 {
 	struct filelayout_mount_type *fl_mt;
 	struct pnfs_mount_type *mt;
-	int status;
 
 	fl_mt = kmalloc(sizeof(struct filelayout_mount_type), GFP_KERNEL);
 	if (!fl_mt)
 		goto error_ret;
 
-	/* Initialize nfs4 file layout specific device list structure */
-	fl_mt->hlist = kmalloc(sizeof(struct nfs4_pnfs_dev_hlist), GFP_KERNEL);
-	if (!fl_mt->hlist)
-		goto cleanup_fl_mt;
-
 	mt = kmalloc(sizeof(struct pnfs_mount_type), GFP_KERNEL);
 	if (!mt)
 		goto cleanup_fl_mt;
@@ -94,11 +88,11 @@ filelayout_initialize_mountpoint(struct super_block *sb, struct nfs_fh *fh)
 	fl_mt->fl_sb = sb;
 	mt->mountid = (void *)fl_mt;
 
-	status = nfs4_pnfs_devlist_init(fl_mt->hlist);
-	if (status)
+	if (nfs4_alloc_init_deviceid_cache(NFS_SB(sb)->nfs_client,
+					   nfs4_fl_free_deviceid_callback))
 		goto cleanup_mt;
 
-	dprintk("%s: device list has been initialized successfully\n",
+	dprintk("%s: deviceid cache has been initialized successfully\n",
 		__func__);
 	return mt;
 
@@ -106,11 +100,10 @@ cleanup_mt: ;
 	kfree(mt);
 
 cleanup_fl_mt: ;
-	kfree(fl_mt->hlist);
 	kfree(fl_mt);
 
 error_ret: ;
-	printk(KERN_WARNING "%s: device list could not be initialized\n",
+	printk(KERN_WARNING "%s: deviceid cache could not be initialized\n",
 		__func__);
 
 	return NULL;
@@ -123,13 +116,14 @@ filelayout_uninitialize_mountpoint(struct pnfs_mount_type *mountid)
 {
 	struct filelayout_mount_type *fl_mt = NULL;
 
+	dprintk("--> %s\n", __func__);
 	if (mountid) {
-		fl_mt = (struct filelayout_mount_type *)mountid->mountid;
+		struct nfs4_deviceid_cache *cache;
 
-		if (fl_mt != NULL) {
-			nfs4_pnfs_devlist_destroy(fl_mt->hlist);
-			kfree(fl_mt);
-		}
+		fl_mt = (struct filelayout_mount_type *)mountid->mountid;
+		cache = NFS_SB(fl_mt->fl_sb)->nfs_client->cl_devid_cache;
+		nfs4_put_deviceid_cache(cache);
+		kfree(fl_mt);
 		kfree(mountid);
 	}
 	return 0;
@@ -381,8 +375,7 @@ filelayout_check_layout(struct pnfs_layout_type *lo,
 	struct nfs_server *nfss = NFS_SERVER(PNFS_INODE(lo));
 
 	dprintk("--> %s\n", __func__);
-	dsaddr = nfs4_pnfs_device_item_find(FILE_MT(PNFS_INODE(lo))->hlist,
-					     &fl->dev_id);
+	dsaddr = nfs4_pnfs_device_item_find(nfss->nfs_client, &fl->dev_id);
 	if (dsaddr == NULL) {
 		dsaddr = get_device_info(PNFS_INODE(lo), &fl->dev_id);
 		if (dsaddr == NULL) {
@@ -618,7 +611,7 @@ filelayout_commit(struct pnfs_layout_type *layoutid, int sync,
 	stripesz = filelayout_get_stripesize(layoutid);
 	dprintk("%s stripesize %Zd\n", __func__, stripesz);
 
-	dsaddr = nfs4_pnfs_device_item_find(FILE_MT(data->inode)->hlist,
+	dsaddr = nfs4_pnfs_device_item_find(NFS_SERVER(data->inode)->nfs_client,
 					     &nfslay->dev_id);
 	if (dsaddr == NULL) {
 		data->pdata.pnfs_error = -EIO;
fs/nfs/nfs4filelayout.h +4 -8
index 12498a2..2cb05bd 100644
@@ -43,8 +43,7 @@ struct nfs4_pnfs_ds {
 };
 
 struct nfs4_file_layout_dsaddr {
-	struct hlist_node	hash_node;   /* nfs4_pnfs_dev_hlist dev_list */
-	struct pnfs_deviceid	dev_id;
+	struct nfs4_deviceid	deviceid;
 	u32 			stripe_count;
 	u8			*stripe_indices;
 	u32			ds_num;
@@ -86,15 +85,13 @@ struct nfs4_filelayout {
 
 struct filelayout_mount_type {
 	struct super_block *fl_sb;
-	struct nfs4_pnfs_dev_hlist *hlist;
 };
 
 extern struct pnfs_client_operations *pnfs_callback_ops;
 
+extern void nfs4_fl_free_deviceid_callback(struct rcu_head *);
 extern void print_ds(struct nfs4_pnfs_ds *ds);
 char *deviceid_fmt(const struct pnfs_deviceid *dev_id);
-int  nfs4_pnfs_devlist_init(struct nfs4_pnfs_dev_hlist *hlist);
-void nfs4_pnfs_devlist_destroy(struct nfs4_pnfs_dev_hlist *hlist);
 int nfs4_pnfs_dserver_get(struct pnfs_layout_segment *lseg,
 			  loff_t offset,
 			  size_t count,
@@ -102,9 +99,8 @@ int nfs4_pnfs_dserver_get(struct pnfs_layout_segment *lseg,
 u32 filelayout_dserver_get_index(loff_t offset,
 				 struct nfs4_file_layout_dsaddr *di,
 				 struct nfs4_filelayout_segment *layout);
-struct nfs4_file_layout_dsaddr *
-nfs4_pnfs_device_item_find(struct nfs4_pnfs_dev_hlist *hlist,
-			   struct pnfs_deviceid *dev_id);
+extern struct nfs4_file_layout_dsaddr *
+nfs4_pnfs_device_item_find(struct nfs_client *, struct pnfs_deviceid *dev_id);
 struct nfs4_file_layout_dsaddr *
 get_device_info(struct inode *inode, struct pnfs_deviceid *dev_id);
 
fs/nfs/nfs4filelayoutdev.c +34 -156

285 lines not on this page · raw message

Signature · 1 line
1.6.6
#4 lore raw

Re: [PATCH 1/3] SQUASHME pnfs_submit: generic device ID cache

In reply to [email protected]
Quoted 3 lines · On Fri, Apr 16, 2010 at 11:52 AM, <[email protected]> wrote:

From: Andy Adamson <[email protected]>

A shared RCU device ID cache servicing multiple mounts of a single la=

yout type

Quoted 3 lines

per meta data server (struct nfs_client).

Device IDs of type deviceid4 are required by all layout types, long l=

ived and

Quoted 1 line

read at each I/O. =A0They are added to the deviceid cache at first re=

ference by

Quoted 1 line

a layout via GETDEVICEINFO and (currently) are only removed at umount=

=2E

Quoted 5 lines

Reference count the device ID cache for each mounted file system in the initialize_mountpoint layoutdriver_io_operation.

Dereference the device id cache on file system in the uninitialize_mo=

untpoint

Quoted 5 lines

layoutdriver_io_operation called at umount

---
=A0fs/nfs/pnfs.c =A0 =A0 =A0 =A0 =A0 =A0 | =A0119 +++++++++++++++++++=

++++++++++++++++++++++++++

Quoted 17 lines

=A0include/linux/nfs4_pnfs.h | =A0 27 ++++++++++ =A0include/linux/nfs_fs_sb.h | =A0 =A01 + =A03 files changed, 147 insertions(+), 0 deletions(-)

fs/nfs/pnfs.c +0 -0
index 91572aa..8492aef 100644
@@ -45,6 +45,7 @@

=A0#include <linux/nfs4.h>
=A0#include <linux/pnfs_xdr.h>
=A0#include <linux/nfs4_pnfs.h>
+#include <linux/rculist.h>

=A0#include "internal.h"
=A0#include "nfs4_fs.h"

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2302

Quoted 6 lines
fs/nfs/pnfs.c +3 -0
@@ -2296,3 +2297,121 @@ struct pnfs_client_operations pnfs_ops =3D {
 
 =A0EXPORT_SYMBOL(pnfs_unregister_layoutdriver);
 =A0EXPORT_SYMBOL(pnfs_register_layoutdriver);
+
+
+/* Device ID cache. Supports one layout type per struct nfs_client *=

/

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 3 lines
fs/nfs/pnfs.c +3 -0
+int
+nfs4_alloc_init_deviceid_cache(struct nfs_client *clp,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0void (*free_callback=

)(struct rcu_head *))

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 4 lines
fs/nfs/pnfs.c +4 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid_cache *c;
+
+ =A0 =A0 =A0 c =3D kzalloc(sizeof(struct nfs4_deviceid_cache), GFP_K=

ERNEL);

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2314

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+ =A0 =A0 =A0 if (!c)
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 return -ENOMEM;
+ =A0 =A0 =A0 spin_lock(&clp->cl_lock);
+ =A0 =A0 =A0 if (clp->cl_devid_cache !=3D NULL) {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kref_get(&clp->cl_devid_cache->dc_kref)=

;

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2317

Quoted 3 lines
fs/nfs/pnfs.c +3 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&clp->cl_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [kref [%d]]\n", __func__,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 atomic_read(&clp->cl_de=

vid_cache->dc_kref.refcount));

Quoted 15 lines
fs/nfs/pnfs.c +15 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kfree(c);
+ =A0 =A0 =A0 } else {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_lock_init(&c->dc_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 INIT_LIST_HEAD(&c->dc_deviceids);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kref_init(&c->dc_kref);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback =3D free_callback;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_clp =3D clp;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 clp->cl_devid_cache =3D c;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&clp->cl_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [new]\n", __func__);
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 return 0;
+}
+EXPORT_SYMBOL(nfs4_alloc_init_deviceid_cache);
+

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2342

Quoted 10 lines
fs/nfs/pnfs.c +10 -0
+void
+nfs4_init_deviceid_node(struct nfs4_deviceid *d)
+{
+ =A0 =A0 =A0 INIT_LIST_HEAD(&d->de_node);
+ =A0 =A0 =A0 INIT_RCU_HEAD(&d->de_rcu);
+}
+EXPORT_SYMBOL(nfs4_init_deviceid_node);
+
+struct nfs4_deviceid *
+nfs4_find_deviceid(struct nfs4_deviceid_cache *c, struct pnfs_device=

id *id)

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2348

Quoted 6 lines
fs/nfs/pnfs.c +6 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 rcu_read_lock();
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 if (!memcmp(&d->de_id, id, NFS4_PNFS_DE=

VICEID4_SIZE)) {

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2360

Quoted 12 lines
fs/nfs/pnfs.c +12 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 rcu_read_unlock();
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 return d;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 }
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 rcu_read_unlock();
+ =A0 =A0 =A0 return NULL;
+}
+EXPORT_SYMBOL(nfs4_find_deviceid);
+
+/*
+ * Add or kref_get a deviceid.
+ * GETDEVICEINFOs for same deviceid can race. If deviceid is found, =

discard new

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 3 lines
fs/nfs/pnfs.c +3 -0
+ */
+void
+nfs4_add_deviceid(struct nfs4_deviceid_cache *c, struct nfs4_devicei=

d *new)

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2369

Quoted 6 lines
fs/nfs/pnfs.c +6 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 spin_lock(&c->dc_lock);
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 if (!memcmp(&d->de_id, &new->de_id, NFS=

4_PNFS_DEVICEID4_SIZE)) {

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2370

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&c->dc_lock=

);

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2371

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [discard]\n=

", __func__);

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2372

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback(&ne=

w->de_rcu);

Quoted 20 lines
fs/nfs/pnfs.c +20 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 }
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 list_add_rcu(&new->de_node, &c->dc_deviceids);
+ =A0 =A0 =A0 spin_unlock(&c->dc_lock);
+ =A0 =A0 =A0 dprintk("%s [new]\n", __func__);
+}
+EXPORT_SYMBOL(nfs4_add_deviceid);
+
+static int
+nfs4_remove_deviceid(struct nfs4_deviceid_cache *c)
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 spin_lock(&c->dc_lock);
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 list_del_rcu(&d->de_node);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&c->dc_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 synchronize_rcu();
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback(&d->de_rcu);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 return 1;

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2402

Quoted 10 lines
fs/nfs/pnfs.c +10 -0
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 spin_unlock(&c->dc_lock);
+ =A0 =A0 =A0 return 0;
+}
+
+static void
+nfs4_free_deviceid_cache(struct kref *kref)
+{
+ =A0 =A0 =A0 struct nfs4_deviceid_cache *cache =3D
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 container_of(kref, struct nfs4_deviceid=

_cache, dc_kref);

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D { · line 2407

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+ =A0 =A0 =A0 int more =3D 1;
+
+ =A0 =A0 =A0 while (more)
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 more =3D nfs4_remove_deviceid(cache);
+ =A0 =A0 =A0 cache->dc_clp->cl_devid_cache =3D NULL;

Need to take the cl_lock around this assignment

spin_lock(&cache->dc_clp->cl_lock);
cache->dc_clp->cl_devid_cache =3D NULL
spin_unlock(&cache->dc_clp->cl_lock);

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 7 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 kfree(cache);
+}
+
+void
+nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *c)
+{
+ =A0 =A0 =A0 dprintk("%s [%d]\n", __func__, atomic_read(&c->dc_kref.=

refcount));

Quoted 9 lines
fs/nfs/pnfs.c +3 -0
+ =A0 =A0 =A0 kref_put(&c->dc_kref, nfs4_free_deviceid_cache);
+}
+EXPORT_SYMBOL(nfs4_put_deviceid_cache);
include/linux/nfs4_pnfs.h +0 -0
index 1d521f4..2a88a06 100644
@@ -281,6 +281,33 @@ struct pnfs_devicelist {

=A0 =A0 =A0 =A0struct pnfs_deviceid =A0 =A0dev_id[NFS4_PNFS_GETDEVLIS=

T_MAXNUM];

Quoted 4 lines

=A0};

+/*
+ * Device ID RCU cache. A device ID is unique per client ID and layo=

ut type.

Quoted 6 lines

+ */
+struct nfs4_deviceid_cache {
+ =A0 =A0 =A0 spinlock_t =A0 =A0 =A0 =A0 =A0 =A0 =A0dc_lock;
+ =A0 =A0 =A0 struct list_head =A0 =A0 =A0 =A0dc_deviceids;
+ =A0 =A0 =A0 struct kref =A0 =A0 =A0 =A0 =A0 =A0 dc_kref;
+ =A0 =A0 =A0 void =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(*dc_free_c=

allback)(struct rcu_head *);

include/linux/nfs4_pnfs.h struct pnfs_devicelist {

Quoted 12 lines

+ =A0 =A0 =A0 struct nfs_client =A0 =A0 =A0 *dc_clp;

include/linux/nfs4_pnfs.h +11 -0
+};
+
+/* Device ID cache node */
+struct nfs4_deviceid {
+ =A0 =A0 =A0 struct list_head =A0 =A0 =A0 =A0de_node;
+ =A0 =A0 =A0 struct rcu_head =A0 =A0 =A0 =A0 de_rcu;
+ =A0 =A0 =A0 struct pnfs_deviceid =A0 =A0de_id;
+};
+
+extern int nfs4_alloc_init_deviceid_cache(struct nfs_client *,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void (*=

free_callback)(struct rcu_head *));

include/linux/nfs4_pnfs.h struct pnfs_devicelist {

Quoted 3 lines
include/linux/nfs4_pnfs.h +3 -0
+extern void nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *);
+extern void nfs4_init_deviceid_node(struct nfs4_deviceid *);
+extern struct nfs4_deviceid *nfs4_find_deviceid(struct nfs4_deviceid=

_cache *,

Quoted 1 line

+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 struct =

pnfs_deviceid *);

Quoted 2 lines

+extern void nfs4_add_deviceid(struct nfs4_deviceid_cache *,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 struct =

nfs4_deviceid *);

Quoted 10 lines

+
=A0/* pNFS client callback functions.
=A0* These operations allow the layout driver to access pNFS client =A0* specific information or call pNFS client->server operations.

include/linux/nfs_fs_sb.h +0 -0
index 8522461..ef2e18e 100644
@@ -87,6 +87,7 @@ struct nfs_client {

=A0 =A0 =A0 =A0u32 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 cl_exchang=

e_flags;

Quoted 1 line

=A0 =A0 =A0 =A0struct nfs4_session =A0 =A0 *cl_session; =A0 =A0/* sha=

rred session */

Quoted 1 line

=A0 =A0 =A0 =A0struct list_head =A0 =A0 =A0 =A0cl_lo_inodes; =A0 /* I=

nodes having layouts */

Quoted 1 line

+ =A0 =A0 =A0 struct nfs4_deviceid_cache *cl_devid_cache; /* pNFS dev=

iceid cache */

Quoted 8 lines

=A0#endif /* CONFIG_NFS_V4_1 */

=A0#ifdef CONFIG_NFS_FSCACHE
--
1.6.6

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" =

in

Quoted 3 lines

the body of a message to [email protected]
More majordomo info at =A0http://vger.kernel.org/majordomo-info.html

A message that is not in the archive (sent off-list or not yet received): u2n89c397151004160904m9e862360xcaf0e187640b0177-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#6 lore raw

Re: [pnfs] [PATCH 1/3] SQUASHME pnfs_submit: generic device ID cache

Quoted 85 lines · William A. (Andy) Adamson
Quoted 84 lines · On Fri, Apr 16, 2010 at 11:52 AM, <[email protected]> wrote:

From: Andy Adamson <[email protected]>

A shared RCU device ID cache servicing multiple mounts of a single layout type per meta data server (struct nfs_client).

Device IDs of type deviceid4 are required by all layout types, long lived and read at each I/O. They are added to the deviceid cache at first reference by a layout via GETDEVICEINFO and (currently) are only removed at umount.

Reference count the device ID cache for each mounted file system in the initialize_mountpoint layoutdriver_io_operation.

Dereference the device id cache on file system in the uninitialize_mountpoint layoutdriver_io_operation called at umount

---
 fs/nfs/pnfs.c             |  119 +++++++++++++++++++++++++++++++++++++++++++++
 include/linux/nfs4_pnfs.h |   27 ++++++++++
 include/linux/nfs_fs_sb.h |    1 +
 3 files changed, 147 insertions(+), 0 deletions(-)
fs/nfs/pnfs.c +47 -0
index 91572aa..8492aef 100644
@@ -45,6 +45,7 @@
 #include <linux/nfs4.h>
 #include <linux/pnfs_xdr.h>
 #include <linux/nfs4_pnfs.h>
+#include <linux/rculist.h>
 
 #include "internal.h"
 #include "nfs4_fs.h"
@@ -2296,3 +2297,121 @@ struct pnfs_client_operations pnfs_ops = {
 
 EXPORT_SYMBOL(pnfs_unregister_layoutdriver);
 EXPORT_SYMBOL(pnfs_register_layoutdriver);
+
+
+/* Device ID cache. Supports one layout type per struct nfs_client */
+int
+nfs4_alloc_init_deviceid_cache(struct nfs_client *clp,
+                        void (*free_callback)(struct rcu_head *))
+{
+       struct nfs4_deviceid_cache *c;
+
+       c = kzalloc(sizeof(struct nfs4_deviceid_cache), GFP_KERNEL);
+       if (!c)
+               return -ENOMEM;
+       spin_lock(&clp->cl_lock);
+       if (clp->cl_devid_cache != NULL) {
+               kref_get(&clp->cl_devid_cache->dc_kref);
+               spin_unlock(&clp->cl_lock);
+               dprintk("%s [kref [%d]]\n", __func__,
+                       atomic_read(&clp->cl_devid_cache->dc_kref.refcount));
+               kfree(c);
+       } else {
+               spin_lock_init(&c->dc_lock);
+               INIT_LIST_HEAD(&c->dc_deviceids);
+               kref_init(&c->dc_kref);
+               c->dc_free_callback = free_callback;
+               c->dc_clp = clp;
+               clp->cl_devid_cache = c;
+               spin_unlock(&clp->cl_lock);
+               dprintk("%s [new]\n", __func__);
+       }
+       return 0;
+}
+EXPORT_SYMBOL(nfs4_alloc_init_deviceid_cache);
+
+void
+nfs4_init_deviceid_node(struct nfs4_deviceid *d)
+{
+       INIT_LIST_HEAD(&d->de_node);
+       INIT_RCU_HEAD(&d->de_rcu);
+}
+EXPORT_SYMBOL(nfs4_init_deviceid_node);
+
+struct nfs4_deviceid *
+nfs4_find_deviceid(struct nfs4_deviceid_cache *c, struct pnfs_deviceid *id)
+{
+       struct nfs4_deviceid *d;
+
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = { · line 2353

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+       rcu_read_lock();
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               if (!memcmp(&d->de_id, id, NFS4_PNFS_DEVICEID4_SIZE)) {
+                       rcu_read_unlock();
+                       return d;
+               }
+       }
+       rcu_read_unlock();

I hope this is worth the added complexity...

Out of curiosity, do you have a benchmark comparing the cost of rcu_read_lock (and unlock) vs. spin_lock(&c->dc_lock)?

Quoted 57 lines
Quoted 44 lines
fs/nfs/pnfs.c +44 -0
+       return NULL;
+}
+EXPORT_SYMBOL(nfs4_find_deviceid);
+
+/*
+ * Add or kref_get a deviceid.
+ * GETDEVICEINFOs for same deviceid can race. If deviceid is found, discard new
+ */
+void
+nfs4_add_deviceid(struct nfs4_deviceid_cache *c, struct nfs4_deviceid *new)
+{
+       struct nfs4_deviceid *d;
+
+       spin_lock(&c->dc_lock);
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               if (!memcmp(&d->de_id, &new->de_id, NFS4_PNFS_DEVICEID4_SIZE)) {
+                       spin_unlock(&c->dc_lock);
+                       dprintk("%s [discard]\n", __func__);
+                       c->dc_free_callback(&new->de_rcu);
+               }
+       }
+       list_add_rcu(&new->de_node, &c->dc_deviceids);
+       spin_unlock(&c->dc_lock);
+       dprintk("%s [new]\n", __func__);
+}
+EXPORT_SYMBOL(nfs4_add_deviceid);
+
+static int
+nfs4_remove_deviceid(struct nfs4_deviceid_cache *c)
+{
+       struct nfs4_deviceid *d;
+
+       spin_lock(&c->dc_lock);
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               list_del_rcu(&d->de_node);
+               spin_unlock(&c->dc_lock);
+               synchronize_rcu();
+               c->dc_free_callback(&d->de_rcu);
+               return 1;
+       }
+       spin_unlock(&c->dc_lock);
+       return 0;
+}
+

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = { · line 2407

Quoted 10 lines
fs/nfs/pnfs.c +10 -0
+static void
+nfs4_free_deviceid_cache(struct kref *kref)
+{
+       struct nfs4_deviceid_cache *cache =
+               container_of(kref, struct nfs4_deviceid_cache, dc_kref);
+       int more = 1;
+
+       while (more)
+               more = nfs4_remove_deviceid(cache);
+       cache->dc_clp->cl_devid_cache = NULL;

Need to take the cl_lock around this assignment

Quoted 5 lines

spin_lock(&cache->dc_clp->cl_lock);
cache->dc_clp->cl_devid_cache = NULL
spin_unlock(&cache->dc_clp->cl_lock);

That must happen atomically before kref_put.
It's illegal to have cl_devid_cache be referenced by cache->dc_clp without a reference count backing it up.
Otherwise, if accessed concurrently to this piece of code someone might call kref_get while the refcount is zero.

Normally, you'd first clear the referencing pointer to prevent any new reference to it and only then kref_put it, e.g.:

spin_lock(&cache->dc_clp->cl_lock);
tmp = cache->dc_clp->cl_devid_cache; cache->dc_clp->cl_devid_cache = NULL spin_unlock(&cache->dc_clp->cl_lock); kref_put(&tmp->dc_kref, nfs4_free_deviceid_cache);

Benny

Quoted 72 lines
Quoted 63 lines
fs/nfs/pnfs.c +10 -0
+       kfree(cache);
+}
+
+void
+nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *c)
+{
+       dprintk("%s [%d]\n", __func__, atomic_read(&c->dc_kref.refcount));
+       kref_put(&c->dc_kref, nfs4_free_deviceid_cache);
+}
+EXPORT_SYMBOL(nfs4_put_deviceid_cache);
include/linux/nfs4_pnfs.h +27 -0
index 1d521f4..2a88a06 100644
@@ -281,6 +281,33 @@ struct pnfs_devicelist {
       struct pnfs_deviceid    dev_id[NFS4_PNFS_GETDEVLIST_MAXNUM];
 };
 
+/*
+ * Device ID RCU cache. A device ID is unique per client ID and layout type.
+ */
+struct nfs4_deviceid_cache {
+       spinlock_t              dc_lock;
+       struct list_head        dc_deviceids;
+       struct kref             dc_kref;
+       void                    (*dc_free_callback)(struct rcu_head *);
+       struct nfs_client       *dc_clp;
+};
+
+/* Device ID cache node */
+struct nfs4_deviceid {
+       struct list_head        de_node;
+       struct rcu_head         de_rcu;
+       struct pnfs_deviceid    de_id;
+};
+
+extern int nfs4_alloc_init_deviceid_cache(struct nfs_client *,
+                               void (*free_callback)(struct rcu_head *));
+extern void nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *);
+extern void nfs4_init_deviceid_node(struct nfs4_deviceid *);
+extern struct nfs4_deviceid *nfs4_find_deviceid(struct nfs4_deviceid_cache *,
+                               struct pnfs_deviceid *);
+extern void nfs4_add_deviceid(struct nfs4_deviceid_cache *,
+                               struct nfs4_deviceid *);
+
 /* pNFS client callback functions.
 * These operations allow the layout driver to access pNFS client
 * specific information or call pNFS client->server operations.
include/linux/nfs_fs_sb.h +1 -1
index 8522461..ef2e18e 100644
@@ -87,6 +87,7 @@ struct nfs_client {
       u32                     cl_exchange_flags;
       struct nfs4_session     *cl_session;    /* sharred session */
       struct list_head        cl_lo_inodes;   /* Inodes having layouts */
+       struct nfs4_deviceid_cache *cl_devid_cache; /* pNFS deviceid cache */
 #endif /* CONFIG_NFS_V4_1 */
 
 #ifdef CONFIG_NFS_FSCACHE
--

1.6.6

Quoted 5 lines

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

_______________________________________________
pNFS mailing list
[email protected]
http://linux-nfs.org/cgi-bin/mailman/listinfo/pnfs

#7 lore raw

Re: [pnfs] [PATCH 1/3] SQUASHME pnfs_submit: generic device ID cache

On Wed, Apr 21, 2010 at 1:59 AM, Benny Halevy <[email protected]> wro= te:

Quoted 1 line

On Apr. 16, 2010, 19:04 +0300, "William A. (Andy) Adamson" <androsada=

Quoted 4 lines · [email protected]> wrote:
Quoted 4 lines · On Fri, Apr 16, 2010 at 11:52 AM, =A0<[email protected]> wrote:
Quoted 3 lines

From: Andy Adamson <[email protected]>

A shared RCU device ID cache servicing multiple mounts of a single =

layout type

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

per meta data server (struct nfs_client).

Device IDs of type deviceid4 are required by all layout types, long=

lived and

Quoted 1 line
Quoted 1 line
Quoted 1 line

read at each I/O. =A0They are added to the deviceid cache at first =

reference by

Quoted 1 line
Quoted 1 line
Quoted 1 line

a layout via GETDEVICEINFO and (currently) are only removed at umou=

nt.

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

Reference count the device ID cache for each mounted file system in the initialize_mountpoint layoutdriver_io_operation.

Dereference the device id cache on file system in the uninitialize_=

mountpoint

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

layoutdriver_io_operation called at umount

---
=A0fs/nfs/pnfs.c =A0 =A0 =A0 =A0 =A0 =A0 | =A0119 +++++++++++++++++=

++++++++++++++++++++++++++++

Quoted 14 lines
Quoted 14 lines
Quoted 14 lines

=A0include/linux/nfs4_pnfs.h | =A0 27 ++++++++++ =A0include/linux/nfs_fs_sb.h | =A0 =A01 + =A03 files changed, 147 insertions(+), 0 deletions(-)

fs/nfs/pnfs.c +0 -0
index 91572aa..8492aef 100644
@@ -45,6 +45,7 @@

=A0#include <linux/nfs4.h>
=A0#include <linux/pnfs_xdr.h>
=A0#include <linux/nfs4_pnfs.h>
+#include <linux/rculist.h>

Quoted 3 lines
Quoted 3 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D =

Quoted 3 lines

=A0#include "internal.h"
=A0#include "nfs4_fs.h"

fs/nfs/pnfs.c +0 -0
@@ -2296,3 +2297,121 @@ struct pnfs_client_operations pnfs_ops =3D =

{

Quoted 6 lines
Quoted 6 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2301

Quoted 6 lines
fs/nfs/pnfs.c +3 -0
 =A0EXPORT_SYMBOL(pnfs_unregister_layoutdriver);
 =A0EXPORT_SYMBOL(pnfs_register_layoutdriver);
+
+
+/* Device ID cache. Supports one layout type per struct nfs_client=

*/

Quoted 3 lines
Quoted 3 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 3 lines
fs/nfs/pnfs.c +3 -0
+int
+nfs4_alloc_init_deviceid_cache(struct nfs_client *clp,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0void (*free_callba=

ck)(struct rcu_head *))

Quoted 4 lines
Quoted 4 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 4 lines
fs/nfs/pnfs.c +4 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid_cache *c;
+
+ =A0 =A0 =A0 c =3D kzalloc(sizeof(struct nfs4_deviceid_cache), GFP=

_KERNEL);

Quoted 5 lines
Quoted 5 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2313

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+ =A0 =A0 =A0 if (!c)
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 return -ENOMEM;
+ =A0 =A0 =A0 spin_lock(&clp->cl_lock);
+ =A0 =A0 =A0 if (clp->cl_devid_cache !=3D NULL) {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kref_get(&clp->cl_devid_cache->dc_kre=

f);

Quoted 3 lines
Quoted 3 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2316

Quoted 3 lines
fs/nfs/pnfs.c +3 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&clp->cl_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [kref [%d]]\n", __func__,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 atomic_read(&clp->cl_=

devid_cache->dc_kref.refcount));

Quoted 6 lines
Quoted 6 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2322

Quoted 6 lines
fs/nfs/pnfs.c +6 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kfree(c);
+ =A0 =A0 =A0 } else {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_lock_init(&c->dc_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 INIT_LIST_HEAD(&c->dc_deviceids);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kref_init(&c->dc_kref);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback =3D free_callback=

;

Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
fs/nfs/pnfs.c +11 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_clp =3D clp;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 clp->cl_devid_cache =3D c;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&clp->cl_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [new]\n", __func__);
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 return 0;
+}
+EXPORT_SYMBOL(nfs4_alloc_init_deviceid_cache);
+
+void
+nfs4_init_deviceid_node(struct nfs4_deviceid *d)
Quoted 8 lines
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+{
+ =A0 =A0 =A0 INIT_LIST_HEAD(&d->de_node);
+ =A0 =A0 =A0 INIT_RCU_HEAD(&d->de_rcu);
+}
+EXPORT_SYMBOL(nfs4_init_deviceid_node);
+
+struct nfs4_deviceid *
+nfs4_find_deviceid(struct nfs4_deviceid_cache *c, struct pnfs_devi=

ceid *id)

Quoted 5 lines
Quoted 5 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2346

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 rcu_read_lock();
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_node)=

{

Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2347

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 if (!memcmp(&d->de_id, id, NFS4_PNFS_=

DEVICEID4_SIZE)) {

Quoted 10 lines
Quoted 5 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2352

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 rcu_read_unlock();
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 return d;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 }
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 rcu_read_unlock();

I hope this is worth the added complexity...

Out of curiosity, do you have a benchmark comparing the cost of rcu_read_lock (and unlock) vs. spin_lock(&c->dc_lock)?

The deviceid cache is read at each I/O. If we use a spin_lock to protect the deviceid cache, this would mean that all I/0 is serialized behind the spin_lock even though the deviceid cache is changed infrequently. The RCU allows readers to "run almost naked" and does not serialize I/O behind reading the deviceid cache.

So I think it is worth it. I have not benchmarked the difference.

Quoted 8 lines
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 8 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 return NULL;
+}
+EXPORT_SYMBOL(nfs4_find_deviceid);
+
+/*
+ * Add or kref_get a deviceid.
+ * GETDEVICEINFOs for same deviceid can race. If deviceid is found=

, discard new

Quoted 3 lines
Quoted 3 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 3 lines
fs/nfs/pnfs.c +3 -0
+ */
+void
+nfs4_add_deviceid(struct nfs4_deviceid_cache *c, struct nfs4_devic=

eid *new)

Quoted 5 lines
Quoted 5 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2367

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 spin_lock(&c->dc_lock);
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_node)=

{

Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2368

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 if (!memcmp(&d->de_id, &new->de_id, N=

=46S4_PNFS_DEVICEID4_SIZE)) {

Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2369

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&c->dc_lo=

ck);

Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2370

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [discard]=

\n", __func__);

Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2371

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback(&=

new->de_rcu);

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 }
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 list_add_rcu(&new->de_node, &c->dc_deviceids);
+ =A0 =A0 =A0 spin_unlock(&c->dc_lock);
+ =A0 =A0 =A0 dprintk("%s [new]\n", __func__);
+}
+EXPORT_SYMBOL(nfs4_add_deviceid);
Quoted 8 lines
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+
+static int
+nfs4_remove_deviceid(struct nfs4_deviceid_cache *c)
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 spin_lock(&c->dc_lock);
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_node)=

{

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 list_del_rcu(&d->de_node);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&c->dc_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 synchronize_rcu();
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback(&d->de_rcu);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 return 1;
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 spin_unlock(&c->dc_lock);
Quoted 8 lines
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+ =A0 =A0 =A0 return 0;
+}
+
+static void
+nfs4_free_deviceid_cache(struct kref *kref)
+{
+ =A0 =A0 =A0 struct nfs4_deviceid_cache *cache =3D
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 container_of(kref, struct nfs4_device=

id_cache, dc_kref);

Quoted 20 lines
Quoted 6 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D = · line 2406

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+ =A0 =A0 =A0 int more =3D 1;
+
+ =A0 =A0 =A0 while (more)
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 more =3D nfs4_remove_deviceid(cache);
+ =A0 =A0 =A0 cache->dc_clp->cl_devid_cache =3D NULL;
Quoted 7 lines

Need to take the cl_lock around this assignment

spin_lock(&cache->dc_clp->cl_lock);
cache->dc_clp->cl_devid_cache =3D NULL
spin_unlock(&cache->dc_clp->cl_lock);

That must happen atomically before kref_put.
It's illegal to have cl_devid_cache be referenced by cache->dc_clp without a reference count backing it up.
Otherwise, if accessed concurrently to this piece of code someone might call kref_get while the refcount is zero.

Quoted 8 lines

Normally, you'd first clear the referencing pointer to prevent any new reference to it and only then kref_put it, e.g.:

spin_lock(&cache->dc_clp->cl_lock);
tmp =3D cache->dc_clp->cl_devid_cache; cache->dc_clp->cl_devid_cache =3D NULL spin_unlock(&cache->dc_clp->cl_lock); kref_put(&tmp->dc_kref, nfs4_free_deviceid_cache);

Good point. Thanks for the review. I'll rethink and resend

-->Andy

Quoted 10 lines

Benny

Quoted 7 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = {

Quoted 7 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 kfree(cache);
+}
+
+void
+nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *c)
+{
+ =A0 =A0 =A0 dprintk("%s [%d]\n", __func__, atomic_read(&c->dc_kre=

f.refcount));

Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
fs/nfs/pnfs.c +3 -0
+ =A0 =A0 =A0 kref_put(&c->dc_kref, nfs4_free_deviceid_cache);
+}
+EXPORT_SYMBOL(nfs4_put_deviceid_cache);
include/linux/nfs4_pnfs.h +0 -0
index 1d521f4..2a88a06 100644
@@ -281,6 +281,33 @@ struct pnfs_devicelist {

=A0 =A0 =A0 =A0struct pnfs_deviceid =A0 =A0dev_id[NFS4_PNFS_GETDEVL=

IST_MAXNUM];

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

=A0};

+/*
+ * Device ID RCU cache. A device ID is unique per client ID and la=

yout type.

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

+ */
+struct nfs4_deviceid_cache {
+ =A0 =A0 =A0 spinlock_t =A0 =A0 =A0 =A0 =A0 =A0 =A0dc_lock;
+ =A0 =A0 =A0 struct list_head =A0 =A0 =A0 =A0dc_deviceids;
+ =A0 =A0 =A0 struct kref =A0 =A0 =A0 =A0 =A0 =A0 dc_kref;
+ =A0 =A0 =A0 void =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(*dc_free=

_callback)(struct rcu_head *);

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

+ =A0 =A0 =A0 struct nfs_client =A0 =A0 =A0 *dc_clp;

include/linux/nfs4_pnfs.h +3 -0
+};
+
+/* Device ID cache node */
Quoted 8 lines
Quoted 8 lines

include/linux/nfs4_pnfs.h struct pnfs_devicelist {

Quoted 8 lines
include/linux/nfs4_pnfs.h +8 -0
+struct nfs4_deviceid {
+ =A0 =A0 =A0 struct list_head =A0 =A0 =A0 =A0de_node;
+ =A0 =A0 =A0 struct rcu_head =A0 =A0 =A0 =A0 de_rcu;
+ =A0 =A0 =A0 struct pnfs_deviceid =A0 =A0de_id;
+};
+
+extern int nfs4_alloc_init_deviceid_cache(struct nfs_client *,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void =

(*free_callback)(struct rcu_head *));

Quoted 3 lines
Quoted 3 lines

include/linux/nfs4_pnfs.h struct pnfs_devicelist {

Quoted 3 lines
include/linux/nfs4_pnfs.h +3 -0
+extern void nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *);
+extern void nfs4_init_deviceid_node(struct nfs4_deviceid *);
+extern struct nfs4_deviceid *nfs4_find_deviceid(struct nfs4_device=

id_cache *,

Quoted 1 line
Quoted 1 line
Quoted 1 line

+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 struc=

t pnfs_deviceid *);

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

+extern void nfs4_add_deviceid(struct nfs4_deviceid_cache *,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 struc=

t nfs4_deviceid *);

Quoted 10 lines
Quoted 10 lines
Quoted 10 lines

+
=A0/* pNFS client callback functions.
=A0* These operations allow the layout driver to access pNFS client =A0* specific information or call pNFS client->server operations.

include/linux/nfs_fs_sb.h +0 -0
index 8522461..ef2e18e 100644
@@ -87,6 +87,7 @@ struct nfs_client {

=A0 =A0 =A0 =A0u32 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 cl_excha=

nge_flags;

Quoted 1 line
Quoted 1 line
Quoted 1 line

=A0 =A0 =A0 =A0struct nfs4_session =A0 =A0 *cl_session; =A0 =A0/* s=

harred session */

Quoted 1 line
Quoted 1 line
Quoted 1 line

=A0 =A0 =A0 =A0struct list_head =A0 =A0 =A0 =A0cl_lo_inodes; =A0 /*=

Inodes having layouts */

Quoted 1 line
Quoted 1 line
Quoted 1 line

+ =A0 =A0 =A0 struct nfs4_deviceid_cache *cl_devid_cache; /* pNFS d=

eviceid cache */

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

=A0#endif /* CONFIG_NFS_V4_1 */

=A0#ifdef CONFIG_NFS_FSCACHE
--
1.6.6

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs=

" in

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

the body of a message to [email protected]
More majordomo info at =A0http://vger.kernel.org/majordomo-info.htm=

l

Quoted 6 lines
Quoted 5 lines
Quoted 1 line

_______________________________________________
pNFS mailing list
[email protected]
http://linux-nfs.org/cgi-bin/mailman/listinfo/pnfs

A message that is not in the archive (sent off-list or not yet received): l2l89c397151004210822j8b43009o3a9e78ceed901fd9-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#9 lore raw

Re: [pnfs] [PATCH 1/3] SQUASHME pnfs_submit: generic device ID cache

Quoted 98 lines · William A. (Andy) Adamson
Quoted 97 lines · Benny Halevy
Quoted 85 lines · William A. (Andy) Adamson
Quoted 84 lines · On Fri, Apr 16, 2010 at 11:52 AM, <[email protected]> wrote:

From: Andy Adamson <[email protected]>

A shared RCU device ID cache servicing multiple mounts of a single layout type per meta data server (struct nfs_client).

Device IDs of type deviceid4 are required by all layout types, long lived and read at each I/O. They are added to the deviceid cache at first reference by a layout via GETDEVICEINFO and (currently) are only removed at umount.

Reference count the device ID cache for each mounted file system in the initialize_mountpoint layoutdriver_io_operation.

Dereference the device id cache on file system in the uninitialize_mountpoint layoutdriver_io_operation called at umount

---
 fs/nfs/pnfs.c             |  119 +++++++++++++++++++++++++++++++++++++++++++++
 include/linux/nfs4_pnfs.h |   27 ++++++++++
 include/linux/nfs_fs_sb.h |    1 +
 3 files changed, 147 insertions(+), 0 deletions(-)
fs/nfs/pnfs.c +47 -0
index 91572aa..8492aef 100644
@@ -45,6 +45,7 @@
 #include <linux/nfs4.h>
 #include <linux/pnfs_xdr.h>
 #include <linux/nfs4_pnfs.h>
+#include <linux/rculist.h>
 
 #include "internal.h"
 #include "nfs4_fs.h"
@@ -2296,3 +2297,121 @@ struct pnfs_client_operations pnfs_ops = {
 
 EXPORT_SYMBOL(pnfs_unregister_layoutdriver);
 EXPORT_SYMBOL(pnfs_register_layoutdriver);
+
+
+/* Device ID cache. Supports one layout type per struct nfs_client */
+int
+nfs4_alloc_init_deviceid_cache(struct nfs_client *clp,
+                        void (*free_callback)(struct rcu_head *))
+{
+       struct nfs4_deviceid_cache *c;
+
+       c = kzalloc(sizeof(struct nfs4_deviceid_cache), GFP_KERNEL);
+       if (!c)
+               return -ENOMEM;
+       spin_lock(&clp->cl_lock);
+       if (clp->cl_devid_cache != NULL) {
+               kref_get(&clp->cl_devid_cache->dc_kref);
+               spin_unlock(&clp->cl_lock);
+               dprintk("%s [kref [%d]]\n", __func__,
+                       atomic_read(&clp->cl_devid_cache->dc_kref.refcount));
+               kfree(c);
+       } else {
+               spin_lock_init(&c->dc_lock);
+               INIT_LIST_HEAD(&c->dc_deviceids);
+               kref_init(&c->dc_kref);
+               c->dc_free_callback = free_callback;
+               c->dc_clp = clp;
+               clp->cl_devid_cache = c;
+               spin_unlock(&clp->cl_lock);
+               dprintk("%s [new]\n", __func__);
+       }
+       return 0;
+}
+EXPORT_SYMBOL(nfs4_alloc_init_deviceid_cache);
+
+void
+nfs4_init_deviceid_node(struct nfs4_deviceid *d)
+{
+       INIT_LIST_HEAD(&d->de_node);
+       INIT_RCU_HEAD(&d->de_rcu);
+}
+EXPORT_SYMBOL(nfs4_init_deviceid_node);
+
+struct nfs4_deviceid *
+nfs4_find_deviceid(struct nfs4_deviceid_cache *c, struct pnfs_deviceid *id)
+{
+       struct nfs4_deviceid *d;
+
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = { · line 2353

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+       rcu_read_lock();
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               if (!memcmp(&d->de_id, id, NFS4_PNFS_DEVICEID4_SIZE)) {
+                       rcu_read_unlock();
+                       return d;
+               }
+       }
+       rcu_read_unlock();

I hope this is worth the added complexity...

Quoted 4 lines
Quoted 2 lines

Out of curiosity, do you have a benchmark comparing the cost of rcu_read_lock (and unlock) vs. spin_lock(&c->dc_lock)?

The deviceid cache is read at each I/O. If we use a spin_lock to

Yeah, I see where this goes...
In the objects layout driver we get a reference on the device structure at alloc_lseg time and keep a pointer to it throughout the lseg's life time. This saves the deviceid lookup on every I/O.

Benny

Quoted 167 lines

protect the deviceid cache, this would mean that all I/0 is serialized behind the spin_lock even though the deviceid cache is changed infrequently. The RCU allows readers to "run almost naked" and does not serialize I/O behind reading the deviceid cache.

So I think it is worth it. I have not benchmarked the difference.

Quoted 70 lines
Quoted 58 lines
Quoted 45 lines
fs/nfs/pnfs.c +44 -0
+       return NULL;
+}
+EXPORT_SYMBOL(nfs4_find_deviceid);
+
+/*
+ * Add or kref_get a deviceid.
+ * GETDEVICEINFOs for same deviceid can race. If deviceid is found, discard new
+ */
+void
+nfs4_add_deviceid(struct nfs4_deviceid_cache *c, struct nfs4_deviceid *new)
+{
+       struct nfs4_deviceid *d;
+
+       spin_lock(&c->dc_lock);
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               if (!memcmp(&d->de_id, &new->de_id, NFS4_PNFS_DEVICEID4_SIZE)) {
+                       spin_unlock(&c->dc_lock);
+                       dprintk("%s [discard]\n", __func__);
+                       c->dc_free_callback(&new->de_rcu);
+               }
+       }
+       list_add_rcu(&new->de_node, &c->dc_deviceids);
+       spin_unlock(&c->dc_lock);
+       dprintk("%s [new]\n", __func__);
+}
+EXPORT_SYMBOL(nfs4_add_deviceid);
+
+static int
+nfs4_remove_deviceid(struct nfs4_deviceid_cache *c)
+{
+       struct nfs4_deviceid *d;
+
+       spin_lock(&c->dc_lock);
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               list_del_rcu(&d->de_node);
+               spin_unlock(&c->dc_lock);
+               synchronize_rcu();
+               c->dc_free_callback(&d->de_rcu);
+               return 1;
+       }
+       spin_unlock(&c->dc_lock);
+       return 0;
+}
+

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = { · line 2407

Quoted 10 lines
fs/nfs/pnfs.c +10 -0
+static void
+nfs4_free_deviceid_cache(struct kref *kref)
+{
+       struct nfs4_deviceid_cache *cache =
+               container_of(kref, struct nfs4_deviceid_cache, dc_kref);
+       int more = 1;
+
+       while (more)
+               more = nfs4_remove_deviceid(cache);
+       cache->dc_clp->cl_devid_cache = NULL;

Need to take the cl_lock around this assignment

Quoted 5 lines

spin_lock(&cache->dc_clp->cl_lock);
cache->dc_clp->cl_devid_cache = NULL
spin_unlock(&cache->dc_clp->cl_lock);

That must happen atomically before kref_put.
It's illegal to have cl_devid_cache be referenced by cache->dc_clp without a reference count backing it up.
Otherwise, if accessed concurrently to this piece of code someone might call kref_get while the refcount is zero.

Quoted 8 lines

Normally, you'd first clear the referencing pointer to prevent any new reference to it and only then kref_put it, e.g.:

spin_lock(&cache->dc_clp->cl_lock);
tmp = cache->dc_clp->cl_devid_cache; cache->dc_clp->cl_devid_cache = NULL spin_unlock(&cache->dc_clp->cl_lock); kref_put(&tmp->dc_kref, nfs4_free_deviceid_cache);

Good point. Thanks for the review. I'll rethink and resend

-->Andy

Quoted 76 lines

Benny

Quoted 72 lines
Quoted 63 lines
fs/nfs/pnfs.c +10 -0
+       kfree(cache);
+}
+
+void
+nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *c)
+{
+       dprintk("%s [%d]\n", __func__, atomic_read(&c->dc_kref.refcount));
+       kref_put(&c->dc_kref, nfs4_free_deviceid_cache);
+}
+EXPORT_SYMBOL(nfs4_put_deviceid_cache);
include/linux/nfs4_pnfs.h +27 -0
index 1d521f4..2a88a06 100644
@@ -281,6 +281,33 @@ struct pnfs_devicelist {
       struct pnfs_deviceid    dev_id[NFS4_PNFS_GETDEVLIST_MAXNUM];
 };
 
+/*
+ * Device ID RCU cache. A device ID is unique per client ID and layout type.
+ */
+struct nfs4_deviceid_cache {
+       spinlock_t              dc_lock;
+       struct list_head        dc_deviceids;
+       struct kref             dc_kref;
+       void                    (*dc_free_callback)(struct rcu_head *);
+       struct nfs_client       *dc_clp;
+};
+
+/* Device ID cache node */
+struct nfs4_deviceid {
+       struct list_head        de_node;
+       struct rcu_head         de_rcu;
+       struct pnfs_deviceid    de_id;
+};
+
+extern int nfs4_alloc_init_deviceid_cache(struct nfs_client *,
+                               void (*free_callback)(struct rcu_head *));
+extern void nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *);
+extern void nfs4_init_deviceid_node(struct nfs4_deviceid *);
+extern struct nfs4_deviceid *nfs4_find_deviceid(struct nfs4_deviceid_cache *,
+                               struct pnfs_deviceid *);
+extern void nfs4_add_deviceid(struct nfs4_deviceid_cache *,
+                               struct nfs4_deviceid *);
+
 /* pNFS client callback functions.
 * These operations allow the layout driver to access pNFS client
 * specific information or call pNFS client->server operations.
include/linux/nfs_fs_sb.h +1 -1
index 8522461..ef2e18e 100644
@@ -87,6 +87,7 @@ struct nfs_client {
       u32                     cl_exchange_flags;
       struct nfs4_session     *cl_session;    /* sharred session */
       struct list_head        cl_lo_inodes;   /* Inodes having layouts */
+       struct nfs4_deviceid_cache *cl_devid_cache; /* pNFS deviceid cache */
 #endif /* CONFIG_NFS_V4_1 */
 
 #ifdef CONFIG_NFS_FSCACHE
--

1.6.6

Quoted 5 lines

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

_______________________________________________
pNFS mailing list
[email protected]
http://linux-nfs.org/cgi-bin/mailman/listinfo/pnfs

Signature · 9 lines
Benny Halevy
Software Architect
Panasas, Inc.
[email protected]
Tel/Fax: +972-3-647-8340
Mobile: +972-54-802-8340

Panasas: The Leader in Parallel Storage
www.panasas.com
#10 lore raw

Re: [pnfs] [PATCH 1/3] SQUASHME pnfs_submit: generic device ID cache

On Thu, Apr 22, 2010 at 7:20 AM, Benny Halevy <[email protected]> wro= te:

Quoted 1 line

On Apr. 21, 2010, 18:22 +0300, "William A. (Andy) Adamson" <androsada=

Quoted 1 line · [email protected]> wrote:
Quoted 1 line

On Wed, Apr 21, 2010 at 1:59 AM, Benny Halevy <[email protected]> =

Quoted 1 line · wrote:
Quoted 1 line
Quoted 1 line

On Apr. 16, 2010, 19:04 +0300, "William A. (Andy) Adamson" <androsa=

Quoted 4 lines · [email protected]> wrote:
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines · On Fri, Apr 16, 2010 at 11:52 AM, =A0<[email protected]> wrote:
Quoted 3 lines

From: Andy Adamson <[email protected]>

A shared RCU device ID cache servicing multiple mounts of a singl=

e layout type

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

per meta data server (struct nfs_client).

Device IDs of type deviceid4 are required by all layout types, lo=

ng lived and

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

read at each I/O. =A0They are added to the deviceid cache at firs=

t reference by

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

a layout via GETDEVICEINFO and (currently) are only removed at um=

ount.

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

Reference count the device ID cache for each mounted file system in the initialize_mountpoint layoutdriver_io_operation.

Dereference the device id cache on file system in the uninitializ=

e_mountpoint

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

layoutdriver_io_operation called at umount

---
=A0fs/nfs/pnfs.c =A0 =A0 =A0 =A0 =A0 =A0 | =A0119 +++++++++++++++=

++++++++++++++++++++++++++++++

Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
Quoted 14 lines
Quoted 14 lines

=A0include/linux/nfs4_pnfs.h | =A0 27 ++++++++++ =A0include/linux/nfs_fs_sb.h | =A0 =A01 + =A03 files changed, 147 insertions(+), 0 deletions(-)

fs/nfs/pnfs.c +0 -0
index 91572aa..8492aef 100644
@@ -45,6 +45,7 @@

=A0#include <linux/nfs4.h>
=A0#include <linux/pnfs_xdr.h>
=A0#include <linux/nfs4_pnfs.h>
+#include <linux/rculist.h>

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D=

Quoted 3 lines

=A0#include "internal.h"
=A0#include "nfs4_fs.h"

fs/nfs/pnfs.c +0 -0
@@ -2296,3 +2297,121 @@ struct pnfs_client_operations pnfs_ops =3D=

{

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2301

Quoted 6 lines
fs/nfs/pnfs.c +3 -0
 =A0EXPORT_SYMBOL(pnfs_unregister_layoutdriver);
 =A0EXPORT_SYMBOL(pnfs_register_layoutdriver);
+
+
+/* Device ID cache. Supports one layout type per struct nfs_clie=

nt */

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2304

Quoted 3 lines
fs/nfs/pnfs.c +3 -0
+int
+nfs4_alloc_init_deviceid_cache(struct nfs_client *clp,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0void (*free_call=

back)(struct rcu_head *))

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2308

Quoted 4 lines
fs/nfs/pnfs.c +4 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid_cache *c;
+
+ =A0 =A0 =A0 c =3D kzalloc(sizeof(struct nfs4_deviceid_cache), G=

=46P_KERNEL);

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2313

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+ =A0 =A0 =A0 if (!c)
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 return -ENOMEM;
+ =A0 =A0 =A0 spin_lock(&clp->cl_lock);
+ =A0 =A0 =A0 if (clp->cl_devid_cache !=3D NULL) {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kref_get(&clp->cl_devid_cache->dc_k=

ref);

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2315

Quoted 2 lines
fs/nfs/pnfs.c +2 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&clp->cl_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [kref [%d]]\n", __func_=

_,

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2316

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 atomic_read(&clp->c=

l_devid_cache->dc_kref.refcount));

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2322

Quoted 6 lines
fs/nfs/pnfs.c +6 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kfree(c);
+ =A0 =A0 =A0 } else {
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_lock_init(&c->dc_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 INIT_LIST_HEAD(&c->dc_deviceids);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 kref_init(&c->dc_kref);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback =3D free_callba=

ck;

Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
fs/nfs/pnfs.c +11 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_clp =3D clp;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 clp->cl_devid_cache =3D c;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&clp->cl_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [new]\n", __func__);
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 return 0;
+}
+EXPORT_SYMBOL(nfs4_alloc_init_deviceid_cache);
+
+void
+nfs4_init_deviceid_node(struct nfs4_deviceid *d)
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2341

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+{
+ =A0 =A0 =A0 INIT_LIST_HEAD(&d->de_node);
+ =A0 =A0 =A0 INIT_RCU_HEAD(&d->de_rcu);
+}
+EXPORT_SYMBOL(nfs4_init_deviceid_node);
+
+struct nfs4_deviceid *
+nfs4_find_deviceid(struct nfs4_deviceid_cache *c, struct pnfs_de=

viceid *id)

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2346

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 rcu_read_lock();
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_nod=

e) {

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2347

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 if (!memcmp(&d->de_id, id, NFS4_PNF=

S_DEVICEID4_SIZE)) {

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 5 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2352

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 rcu_read_unlock();
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 return d;
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 }
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 rcu_read_unlock();

I hope this is worth the added complexity...

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

Out of curiosity, do you have a benchmark comparing the cost of rcu_read_lock (and unlock) vs. spin_lock(&c->dc_lock)?

The deviceid cache is read at each I/O. If we use a spin_lock to

Yeah, I see where this goes...
In the objects layout driver we get a reference on the device structu=

re

Quoted 1 line

at alloc_lseg time and keep a pointer to it throughout the lseg's lif=

e time.

Quoted 1 line

This saves the deviceid lookup on every I/O.

Perhaps that is a better way to go. How many deviceid's is 'normal' for the object driver?

-->Andy

Quoted 4 lines

Benny

Quoted 1 line

protect the deviceid cache, this would mean that all I/0 is serializ=

ed

Quoted 8 lines
Quoted 8 lines

behind the spin_lock even though the deviceid cache is changed infrequently. The RCU allows readers to "run almost naked" and does not serialize I/O behind reading the deviceid cache.

So I think it is worth it. I have not benchmarked the difference.

Quoted 1 line
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2359

Quoted 7 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 return NULL;
+}
+EXPORT_SYMBOL(nfs4_find_deviceid);
+
+/*
+ * Add or kref_get a deviceid.
+ * GETDEVICEINFOs for same deviceid can race. If deviceid is fou=

nd, discard new

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2362

Quoted 3 lines
fs/nfs/pnfs.c +3 -0
+ */
+void
+nfs4_add_deviceid(struct nfs4_deviceid_cache *c, struct nfs4_dev=

iceid *new)

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2367

Quoted 5 lines
fs/nfs/pnfs.c +5 -0
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 spin_lock(&c->dc_lock);
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_nod=

e) {

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2368

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 if (!memcmp(&d->de_id, &new->de_id,=

NFS4_PNFS_DEVICEID4_SIZE)) {

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2369

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&c->dc_=

lock);

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2370

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 dprintk("%s [discar=

d]\n", __func__);

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2371

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback=

(&new->de_rcu);

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 }
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 list_add_rcu(&new->de_node, &c->dc_deviceids);
+ =A0 =A0 =A0 spin_unlock(&c->dc_lock);
+ =A0 =A0 =A0 dprintk("%s [new]\n", __func__);
+}
+EXPORT_SYMBOL(nfs4_add_deviceid);
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2386

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+
+static int
+nfs4_remove_deviceid(struct nfs4_deviceid_cache *c)
+{
+ =A0 =A0 =A0 struct nfs4_deviceid *d;
+
+ =A0 =A0 =A0 spin_lock(&c->dc_lock);
+ =A0 =A0 =A0 list_for_each_entry_rcu(d, &c->dc_deviceids, de_nod=

e) {

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 list_del_rcu(&d->de_node);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 spin_unlock(&c->dc_lock);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 synchronize_rcu();
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 c->dc_free_callback(&d->de_rcu);
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 return 1;
+ =A0 =A0 =A0 }
+ =A0 =A0 =A0 spin_unlock(&c->dc_lock);
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2401

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+ =A0 =A0 =A0 return 0;
+}
+
+static void
+nfs4_free_deviceid_cache(struct kref *kref)
+{
+ =A0 =A0 =A0 struct nfs4_deviceid_cache *cache =3D
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 container_of(kref, struct nfs4_devi=

ceid_cache, dc_kref);

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2405

Quoted 4 lines
fs/nfs/pnfs.c +4 -0
+ =A0 =A0 =A0 int more =3D 1;
+
+ =A0 =A0 =A0 while (more)
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 more =3D nfs4_remove_deviceid(cache=

);

Quoted 32 lines
Quoted 32 lines
Quoted 16 lines
Quoted 9 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2406

Quoted 1 line
fs/nfs/pnfs.c +1 -0
+ =A0 =A0 =A0 cache->dc_clp->cl_devid_cache =3D NULL;

Need to take the cl_lock around this assignment

spin_lock(&cache->dc_clp->cl_lock);
cache->dc_clp->cl_devid_cache =3D NULL
spin_unlock(&cache->dc_clp->cl_lock);

That must happen atomically before kref_put.
It's illegal to have cl_devid_cache be referenced by cache->dc_clp without a reference count backing it up.
Otherwise, if accessed concurrently to this piece of code someone might call kref_get while the refcount is zero.

Quoted 8 lines

Normally, you'd first clear the referencing pointer to prevent any new reference to it and only then kref_put it, e.g.:

spin_lock(&cache->dc_clp->cl_lock);
tmp =3D cache->dc_clp->cl_devid_cache; cache->dc_clp->cl_devid_cache =3D NULL spin_unlock(&cache->dc_clp->cl_lock); kref_put(&tmp->dc_kref, nfs4_free_deviceid_cache);

Good point. Thanks for the review. I'll rethink and resend

-->Andy

Quoted 3 lines

Benny

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops =3D= · line 2413

Quoted 7 lines
fs/nfs/pnfs.c +7 -0
+ =A0 =A0 =A0 kfree(cache);
+}
+
+void
+nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *c)
+{
+ =A0 =A0 =A0 dprintk("%s [%d]\n", __func__, atomic_read(&c->dc_k=

ref.refcount));

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

include/linux/nfs4_pnfs.=

Quoted 4 lines
fs/nfs/pnfs.c +3 -0
+ =A0 =A0 =A0 kref_put(&c->dc_kref, nfs4_free_deviceid_cache);
+}
+EXPORT_SYMBOL(nfs4_put_deviceid_cache);
include/linux/nfs4_pnfs.= renamed from include/linux/nfs4_pnfs.h +0 -0

h

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

index 1d521f4..2a88a06 100644

include/linux/nfs4_pnfs.h +0 -0
@@ -281,6 +281,33 @@ struct pnfs_devicelist {

=A0 =A0 =A0 =A0struct pnfs_deviceid =A0 =A0dev_id[NFS4_PNFS_GETDE=

VLIST_MAXNUM];

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

=A0};

+/*
+ * Device ID RCU cache. A device ID is unique per client ID and =

layout type.

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

+ */
+struct nfs4_deviceid_cache {
+ =A0 =A0 =A0 spinlock_t =A0 =A0 =A0 =A0 =A0 =A0 =A0dc_lock;
+ =A0 =A0 =A0 struct list_head =A0 =A0 =A0 =A0dc_deviceids;
+ =A0 =A0 =A0 struct kref =A0 =A0 =A0 =A0 =A0 =A0 dc_kref;
+ =A0 =A0 =A0 void =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(*dc_fr=

ee_callback)(struct rcu_head *);

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

+ =A0 =A0 =A0 struct nfs_client =A0 =A0 =A0 *dc_clp;

include/linux/nfs4_pnfs.h +3 -0
+};
+
+/* Device ID cache node */
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

include/linux/nfs4_pnfs.h

Quoted 8 lines
include/linux/nfs4_pnfs.h +8 -0
+struct nfs4_deviceid {
+ =A0 =A0 =A0 struct list_head =A0 =A0 =A0 =A0de_node;
+ =A0 =A0 =A0 struct rcu_head =A0 =A0 =A0 =A0 de_rcu;
+ =A0 =A0 =A0 struct pnfs_deviceid =A0 =A0de_id;
+};
+
+extern int nfs4_alloc_init_deviceid_cache(struct nfs_client *,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 voi=

d (*free_callback)(struct rcu_head *));

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

+extern void nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *=

);

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

include/linux/nfs4_pnfs.h

Quoted 2 lines
include/linux/nfs4_pnfs.h +2 -0
+extern void nfs4_init_deviceid_node(struct nfs4_deviceid *);
+extern struct nfs4_deviceid *nfs4_find_deviceid(struct nfs4_devi=

ceid_cache *,

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 str=

uct pnfs_deviceid *);

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

+extern void nfs4_add_deviceid(struct nfs4_deviceid_cache *,
+ =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 str=

uct nfs4_deviceid *);

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

+
=A0/* pNFS client callback functions.
=A0* These operations allow the layout driver to access pNFS clie=

nt

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

include/linux/nfs_fs_sb.=

Quoted 2 lines

=A0* specific information or call pNFS client->server operations.

include/linux/nfs_fs_sb.= renamed from include/linux/nfs_fs_sb.h +0 -0

h

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

index 8522461..ef2e18e 100644

include/linux/nfs_fs_sb.h +0 -0
@@ -87,6 +87,7 @@ struct nfs_client {

=A0 =A0 =A0 =A0u32 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 cl_exc=

hange_flags;

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

=A0 =A0 =A0 =A0struct nfs4_session =A0 =A0 *cl_session; =A0 =A0/*=

sharred session */

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

=A0 =A0 =A0 =A0struct list_head =A0 =A0 =A0 =A0cl_lo_inodes; =A0 =

/* Inodes having layouts */

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

+ =A0 =A0 =A0 struct nfs4_deviceid_cache *cl_devid_cache; /* pNFS=

deviceid cache */

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

=A0#endif /* CONFIG_NFS_V4_1 */

=A0#ifdef CONFIG_NFS_FSCACHE
--
1.6.6

--
To unsubscribe from this list: send the line "unsubscribe linux-n=

fs" in

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

the body of a message to [email protected]
More majordomo info at =A0http://vger.kernel.org/majordomo-info.h=

tml

Quoted 19 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Quoted 1 line

_______________________________________________
pNFS mailing list
[email protected]
http://linux-nfs.org/cgi-bin/mailman/listinfo/pnfs

--
Benny Halevy
Software Architect
Panasas, Inc.
[email protected]
Tel/Fax: +972-3-647-8340
Mobile: +972-54-802-8340

Panasas: The Leader in Parallel Storage
www.panasas.com

A message that is not in the archive (sent off-list or not yet received): v2h89c397151004220847v3a31c493s4089d0cd53cf3e19-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#12 lore raw

Re: [pnfs] [PATCH 1/3] SQUASHME pnfs_submit: generic device ID cache

Quoted 105 lines · William A. (Andy) Adamson
Quoted 104 lines · Benny Halevy
Quoted 102 lines · William A. (Andy) Adamson
Quoted 97 lines · Benny Halevy
Quoted 85 lines · William A. (Andy) Adamson
Quoted 84 lines · On Fri, Apr 16, 2010 at 11:52 AM, <[email protected]> wrote:

From: Andy Adamson <[email protected]>

A shared RCU device ID cache servicing multiple mounts of a single layout type per meta data server (struct nfs_client).

Device IDs of type deviceid4 are required by all layout types, long lived and read at each I/O. They are added to the deviceid cache at first reference by a layout via GETDEVICEINFO and (currently) are only removed at umount.

Reference count the device ID cache for each mounted file system in the initialize_mountpoint layoutdriver_io_operation.

Dereference the device id cache on file system in the uninitialize_mountpoint layoutdriver_io_operation called at umount

---
 fs/nfs/pnfs.c             |  119 +++++++++++++++++++++++++++++++++++++++++++++
 include/linux/nfs4_pnfs.h |   27 ++++++++++
 include/linux/nfs_fs_sb.h |    1 +
 3 files changed, 147 insertions(+), 0 deletions(-)
fs/nfs/pnfs.c +47 -0
index 91572aa..8492aef 100644
@@ -45,6 +45,7 @@
 #include <linux/nfs4.h>
 #include <linux/pnfs_xdr.h>
 #include <linux/nfs4_pnfs.h>
+#include <linux/rculist.h>
 
 #include "internal.h"
 #include "nfs4_fs.h"
@@ -2296,3 +2297,121 @@ struct pnfs_client_operations pnfs_ops = {
 
 EXPORT_SYMBOL(pnfs_unregister_layoutdriver);
 EXPORT_SYMBOL(pnfs_register_layoutdriver);
+
+
+/* Device ID cache. Supports one layout type per struct nfs_client */
+int
+nfs4_alloc_init_deviceid_cache(struct nfs_client *clp,
+                        void (*free_callback)(struct rcu_head *))
+{
+       struct nfs4_deviceid_cache *c;
+
+       c = kzalloc(sizeof(struct nfs4_deviceid_cache), GFP_KERNEL);
+       if (!c)
+               return -ENOMEM;
+       spin_lock(&clp->cl_lock);
+       if (clp->cl_devid_cache != NULL) {
+               kref_get(&clp->cl_devid_cache->dc_kref);
+               spin_unlock(&clp->cl_lock);
+               dprintk("%s [kref [%d]]\n", __func__,
+                       atomic_read(&clp->cl_devid_cache->dc_kref.refcount));
+               kfree(c);
+       } else {
+               spin_lock_init(&c->dc_lock);
+               INIT_LIST_HEAD(&c->dc_deviceids);
+               kref_init(&c->dc_kref);
+               c->dc_free_callback = free_callback;
+               c->dc_clp = clp;
+               clp->cl_devid_cache = c;
+               spin_unlock(&clp->cl_lock);
+               dprintk("%s [new]\n", __func__);
+       }
+       return 0;
+}
+EXPORT_SYMBOL(nfs4_alloc_init_deviceid_cache);
+
+void
+nfs4_init_deviceid_node(struct nfs4_deviceid *d)
+{
+       INIT_LIST_HEAD(&d->de_node);
+       INIT_RCU_HEAD(&d->de_rcu);
+}
+EXPORT_SYMBOL(nfs4_init_deviceid_node);
+
+struct nfs4_deviceid *
+nfs4_find_deviceid(struct nfs4_deviceid_cache *c, struct pnfs_deviceid *id)
+{
+       struct nfs4_deviceid *d;
+
Quoted 8 lines

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = { · line 2353

Quoted 8 lines
fs/nfs/pnfs.c +8 -0
+       rcu_read_lock();
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               if (!memcmp(&d->de_id, id, NFS4_PNFS_DEVICEID4_SIZE)) {
+                       rcu_read_unlock();
+                       return d;
+               }
+       }
+       rcu_read_unlock();

I hope this is worth the added complexity...

Quoted 2 lines

Out of curiosity, do you have a benchmark comparing the cost of rcu_read_lock (and unlock) vs. spin_lock(&c->dc_lock)?

The deviceid cache is read at each I/O. If we use a spin_lock to

Quoted 8 lines
Quoted 4 lines

Yeah, I see where this goes...
In the objects layout driver we get a reference on the device structure at alloc_lseg time and keep a pointer to it throughout the lseg's life time. This saves the deviceid lookup on every I/O.

Perhaps that is a better way to go. How many deviceid's is 'normal' for the object driver?

I'd say that order of 10's to a few 100's is normal. 1000 and up is a big installation.

Benny

Quoted 186 lines

-->Andy

Quoted 183 lines

Benny

Quoted 87 lines

protect the deviceid cache, this would mean that all I/0 is serialized behind the spin_lock even though the deviceid cache is changed infrequently. The RCU allows readers to "run almost naked" and does not serialize I/O behind reading the deviceid cache.

So I think it is worth it. I have not benchmarked the difference.

Quoted 70 lines
Quoted 58 lines
Quoted 45 lines
fs/nfs/pnfs.c +44 -0
+       return NULL;
+}
+EXPORT_SYMBOL(nfs4_find_deviceid);
+
+/*
+ * Add or kref_get a deviceid.
+ * GETDEVICEINFOs for same deviceid can race. If deviceid is found, discard new
+ */
+void
+nfs4_add_deviceid(struct nfs4_deviceid_cache *c, struct nfs4_deviceid *new)
+{
+       struct nfs4_deviceid *d;
+
+       spin_lock(&c->dc_lock);
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               if (!memcmp(&d->de_id, &new->de_id, NFS4_PNFS_DEVICEID4_SIZE)) {
+                       spin_unlock(&c->dc_lock);
+                       dprintk("%s [discard]\n", __func__);
+                       c->dc_free_callback(&new->de_rcu);
+               }
+       }
+       list_add_rcu(&new->de_node, &c->dc_deviceids);
+       spin_unlock(&c->dc_lock);
+       dprintk("%s [new]\n", __func__);
+}
+EXPORT_SYMBOL(nfs4_add_deviceid);
+
+static int
+nfs4_remove_deviceid(struct nfs4_deviceid_cache *c)
+{
+       struct nfs4_deviceid *d;
+
+       spin_lock(&c->dc_lock);
+       list_for_each_entry_rcu(d, &c->dc_deviceids, de_node) {
+               list_del_rcu(&d->de_node);
+               spin_unlock(&c->dc_lock);
+               synchronize_rcu();
+               c->dc_free_callback(&d->de_rcu);
+               return 1;
+       }
+       spin_unlock(&c->dc_lock);
+       return 0;
+}
+

fs/nfs/pnfs.c struct pnfs_client_operations pnfs_ops = { · line 2407

Quoted 10 lines
fs/nfs/pnfs.c +10 -0
+static void
+nfs4_free_deviceid_cache(struct kref *kref)
+{
+       struct nfs4_deviceid_cache *cache =
+               container_of(kref, struct nfs4_deviceid_cache, dc_kref);
+       int more = 1;
+
+       while (more)
+               more = nfs4_remove_deviceid(cache);
+       cache->dc_clp->cl_devid_cache = NULL;

Need to take the cl_lock around this assignment

Quoted 5 lines

spin_lock(&cache->dc_clp->cl_lock);
cache->dc_clp->cl_devid_cache = NULL
spin_unlock(&cache->dc_clp->cl_lock);

That must happen atomically before kref_put.
It's illegal to have cl_devid_cache be referenced by cache->dc_clp without a reference count backing it up.
Otherwise, if accessed concurrently to this piece of code someone might call kref_get while the refcount is zero.

Quoted 8 lines

Normally, you'd first clear the referencing pointer to prevent any new reference to it and only then kref_put it, e.g.:

spin_lock(&cache->dc_clp->cl_lock);
tmp = cache->dc_clp->cl_devid_cache; cache->dc_clp->cl_devid_cache = NULL spin_unlock(&cache->dc_clp->cl_lock); kref_put(&tmp->dc_kref, nfs4_free_deviceid_cache);

Quoted 80 lines

Good point. Thanks for the review. I'll rethink and resend

-->Andy

Quoted 76 lines

Benny

Quoted 72 lines
Quoted 63 lines
fs/nfs/pnfs.c +10 -0
+       kfree(cache);
+}
+
+void
+nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *c)
+{
+       dprintk("%s [%d]\n", __func__, atomic_read(&c->dc_kref.refcount));
+       kref_put(&c->dc_kref, nfs4_free_deviceid_cache);
+}
+EXPORT_SYMBOL(nfs4_put_deviceid_cache);
include/linux/nfs4_pnfs.h +27 -0
index 1d521f4..2a88a06 100644
@@ -281,6 +281,33 @@ struct pnfs_devicelist {
       struct pnfs_deviceid    dev_id[NFS4_PNFS_GETDEVLIST_MAXNUM];
 };
 
+/*
+ * Device ID RCU cache. A device ID is unique per client ID and layout type.
+ */
+struct nfs4_deviceid_cache {
+       spinlock_t              dc_lock;
+       struct list_head        dc_deviceids;
+       struct kref             dc_kref;
+       void                    (*dc_free_callback)(struct rcu_head *);
+       struct nfs_client       *dc_clp;
+};
+
+/* Device ID cache node */
+struct nfs4_deviceid {
+       struct list_head        de_node;
+       struct rcu_head         de_rcu;
+       struct pnfs_deviceid    de_id;
+};
+
+extern int nfs4_alloc_init_deviceid_cache(struct nfs_client *,
+                               void (*free_callback)(struct rcu_head *));
+extern void nfs4_put_deviceid_cache(struct nfs4_deviceid_cache *);
+extern void nfs4_init_deviceid_node(struct nfs4_deviceid *);
+extern struct nfs4_deviceid *nfs4_find_deviceid(struct nfs4_deviceid_cache *,
+                               struct pnfs_deviceid *);
+extern void nfs4_add_deviceid(struct nfs4_deviceid_cache *,
+                               struct nfs4_deviceid *);
+
 /* pNFS client callback functions.
 * These operations allow the layout driver to access pNFS client
 * specific information or call pNFS client->server operations.
include/linux/nfs_fs_sb.h +1 -1
index 8522461..ef2e18e 100644
@@ -87,6 +87,7 @@ struct nfs_client {
       u32                     cl_exchange_flags;
       struct nfs4_session     *cl_session;    /* sharred session */
       struct list_head        cl_lo_inodes;   /* Inodes having layouts */
+       struct nfs4_deviceid_cache *cl_devid_cache; /* pNFS deviceid cache */
 #endif /* CONFIG_NFS_V4_1 */
 
 #ifdef CONFIG_NFS_FSCACHE
--

1.6.6

Quoted 5 lines

--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

_______________________________________________
pNFS mailing list
[email protected]
http://linux-nfs.org/cgi-bin/mailman/listinfo/pnfs

--
Benny Halevy
Software Architect
Panasas, Inc.
[email protected]
Tel/Fax: +972-3-647-8340
Mobile: +972-54-802-8340

Panasas: The Leader in Parallel Storage
www.panasas.com