[PATCH 05/14] pnfs: wave 2: change how lsegs are removed from layout list

[Fred Isaman <iisaman@netapp.com>]

This is to prepare the way for sensible io draining.  Instead of just
removing the lseg from the list, we instead clear the VALID flag
(preventing new io from grabbing references to the lseg) and remove
the reference holding it in the list.  Thus the lseg will be removed
once any io in progress completes and any references still held are
dropped.

Signed-off-by: Fred Isaman <iisaman@netapp.com>
---
 fs/nfs/inode.c |    2 +-
 fs/nfs/pnfs.c  |  121 ++++++++++++++++++++++++++++++++++++-------------------
 fs/nfs/pnfs.h  |    8 +++-
 3 files changed, 87 insertions(+), 44 deletions(-)

diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
index e67e31c..43a69da 100644
--- a/fs/nfs/inode.c
+++ b/fs/nfs/inode.c
@@ -1410,9 +1410,9 @@ static int nfs_update_inode(struct inode *inode, struct nfs_fattr *fattr)
  */
 void nfs4_evict_inode(struct inode *inode)
 {
+	pnfs_destroy_layout(NFS_I(inode));
 	truncate_inode_pages(&inode->i_data, 0);
 	end_writeback(inode);
-	pnfs_destroy_layout(NFS_I(inode));
 	/* If we are holding a delegation, return it! */
 	nfs_inode_return_delegation_noreclaim(inode);
 	/* First call standard NFS clear_inode() code */
diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
index 7e4ad2a..49591e4 100644
--- a/fs/nfs/pnfs.c
+++ b/fs/nfs/pnfs.c
@@ -211,69 +211,101 @@ static void
 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
 {
 	INIT_LIST_HEAD(&lseg->fi_list);
-	kref_init(&lseg->kref);
+	atomic_set(&lseg->pls_refcount, 1);
+	smp_mb();
+	set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
 	lseg->layout = lo;
 }
 
-/* Called without i_lock held, as the free_lseg call may sleep */
-static void
-destroy_lseg(struct kref *kref)
+static void free_lseg(struct pnfs_layout_segment *lseg)
 {
-	struct pnfs_layout_segment *lseg =
-		container_of(kref, struct pnfs_layout_segment, kref);
 	struct inode *ino = lseg->layout->inode;
 
-	dprintk("--> %s\n", __func__);
+	BUG_ON(atomic_read(&lseg->pls_refcount) != 0);
 	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
 	/* Matched by get_layout_hdr in pnfs_insert_layout */
 	put_layout_hdr(ino);
 }
 
+/* The use of tmp_list is necessary because pnfs_curr_ld->free_lseg
+ * could sleep, so must be called outside of the lock.
+ */
 static void
-put_lseg(struct pnfs_layout_segment *lseg)
+put_lseg_locked(struct pnfs_layout_segment *lseg,
+		struct list_head *tmp_list)
 {
-	if (!lseg)
-		return;
+	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
+		atomic_read(&lseg->pls_refcount),
+		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
+	if (atomic_dec_and_test(&lseg->pls_refcount)) {
+		struct inode *ino = lseg->layout->inode;
+
+		BUG_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
+		list_del(&lseg->fi_list);
+		if (list_empty(&lseg->layout->segs)) {
+			struct nfs_client *clp;
+
+			clp = NFS_SERVER(ino)->nfs_client;
+			spin_lock(&clp->cl_lock);
+			/* List does not take a reference, so no need for put here */
+			list_del_init(&lseg->layout->layouts);
+			spin_unlock(&clp->cl_lock);
+		}
+		list_add(&lseg->fi_list, tmp_list);
+	}
+}
 
-	dprintk("%s: lseg %p ref %d\n", __func__, lseg,
-		atomic_read(&lseg->kref.refcount));
-	kref_put(&lseg->kref, destroy_lseg);
+static bool
+should_free_lseg(u32 lseg_iomode, u32 recall_iomode)
+{
+	return (recall_iomode == IOMODE_ANY ||
+		lseg_iomode == recall_iomode);
 }
 
-static void
-pnfs_clear_lseg_list(struct pnfs_layout_hdr *lo, struct list_head *tmp_list)
+static void mark_lseg_invalid(struct pnfs_layout_segment *lseg,
+			      struct list_head *tmp_list)
+{
+	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
+		/* Remove the reference keeping the lseg in the
+		 * list.  It will now be removed when all
+		 * outstanding io is finished.
+		 */
+		put_lseg_locked(lseg, tmp_list);
+	}
+}
+
+/* Returns false if no lsegs match, true otherwise */
+static bool
+mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
+			    struct list_head *tmp_list,
+			    u32 iomode)
 {
 	struct pnfs_layout_segment *lseg, *next;
-	struct nfs_client *clp;
+	bool rv = false;
 
 	dprintk("%s:Begin lo %p\n", __func__, lo);
 
-	assert_spin_locked(&lo->inode->i_lock);
-	list_for_each_entry_safe(lseg, next, &lo->segs, fi_list) {
-		dprintk("%s: freeing lseg %p\n", __func__, lseg);
-		list_move(&lseg->fi_list, tmp_list);
-	}
-	clp = NFS_SERVER(lo->inode)->nfs_client;
-	spin_lock(&clp->cl_lock);
-	/* List does not take a reference, so no need for put here */
-	list_del_init(&lo->layouts);
-	spin_unlock(&clp->cl_lock);
-
+	list_for_each_entry_safe(lseg, next, &lo->segs, fi_list)
+		if (should_free_lseg(lseg->range.iomode, iomode)) {
+			dprintk("%s: freeing lseg %p iomode %d "
+				"offset %llu length %llu\n", __func__,
+				lseg, lseg->range.iomode, lseg->range.offset,
+				lseg->range.length);
+			mark_lseg_invalid(lseg, tmp_list);
+			rv = true;
+		}
 	dprintk("%s:Return\n", __func__);
+	return rv;
 }
 
 static void
-pnfs_free_lseg_list(struct list_head *tmp_list)
+pnfs_free_lseg_list(struct list_head *free_me)
 {
-	struct pnfs_layout_segment *lseg;
+	struct pnfs_layout_segment *lseg, *tmp;
 
-	while (!list_empty(tmp_list)) {
-		lseg = list_entry(tmp_list->next, struct pnfs_layout_segment,
-				fi_list);
-		dprintk("%s calling put_lseg on %p\n", __func__, lseg);
-		list_del(&lseg->fi_list);
-		put_lseg(lseg);
-	}
+	list_for_each_entry_safe(lseg, tmp, free_me, fi_list)
+		free_lseg(lseg);
+	INIT_LIST_HEAD(free_me);
 }
 
 void
@@ -285,7 +317,8 @@ pnfs_destroy_layout(struct nfs_inode *nfsi)
 	spin_lock(&nfsi->vfs_inode.i_lock);
 	lo = nfsi->layout;
 	if (lo) {
-		pnfs_clear_lseg_list(lo, &tmp_list);
+		set_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags);
+		mark_matching_lsegs_invalid(lo, &tmp_list, IOMODE_ANY);
 		/* Matched by refcount set to 1 in alloc_init_layout_hdr */
 		put_layout_hdr_locked(lo);
 	}
@@ -477,9 +510,12 @@ pnfs_find_alloc_layout(struct inode *ino)
 	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
 
 	assert_spin_locked(&ino->i_lock);
-	if (nfsi->layout)
-		return nfsi->layout;
-
+	if (nfsi->layout) {
+		if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
+			return NULL;
+		else
+			return nfsi->layout;
+	}
 	spin_unlock(&ino->i_lock);
 	new = alloc_init_layout_hdr(ino);
 	spin_lock(&ino->i_lock);
@@ -520,7 +556,8 @@ pnfs_has_layout(struct pnfs_layout_hdr *lo, u32 iomode)
 
 	assert_spin_locked(&lo->inode->i_lock);
 	list_for_each_entry(lseg, &lo->segs, fi_list) {
-		if (is_matching_lseg(lseg, iomode)) {
+		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
+		    is_matching_lseg(lseg, iomode)) {
 			ret = lseg;
 			break;
 		}
@@ -529,7 +566,7 @@ pnfs_has_layout(struct pnfs_layout_hdr *lo, u32 iomode)
 	}
 
 	dprintk("%s:Return lseg %p ref %d\n",
-		__func__, ret, ret ? atomic_read(&ret->kref.refcount) : 0);
+		__func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
 	return ret;
 }
 
diff --git a/fs/nfs/pnfs.h b/fs/nfs/pnfs.h
index 89eedf41..96b23bc 100644
--- a/fs/nfs/pnfs.h
+++ b/fs/nfs/pnfs.h
@@ -30,10 +30,15 @@
 #ifndef FS_NFS_PNFS_H
 #define FS_NFS_PNFS_H
 
+enum {
+	NFS_LSEG_VALID = 0,	/* cleared when lseg is recalled/returned */
+};
+
 struct pnfs_layout_segment {
 	struct list_head fi_list;
 	struct pnfs_layout_range range;
-	struct kref kref;
+	atomic_t pls_refcount;
+	unsigned long pls_flags;
 	struct pnfs_layout_hdr *layout;
 };
 
@@ -44,6 +49,7 @@ struct pnfs_layout_segment {
 enum {
 	NFS_LAYOUT_RO_FAILED = 0,	/* get ro layout failed stop trying */
 	NFS_LAYOUT_RW_FAILED,		/* get rw layout failed stop trying */
+	NFS_LAYOUT_DESTROYED,		/* no new use of layout allowed */
 };
 
 /* Per-layout driver specific registration structure */
-- 
1.7.2.1