summaryrefslogtreecommitdiff
path: root/debian
diff options
context:
space:
mode:
Diffstat (limited to 'debian')
-rw-r--r--debian/patches/0001-ext2fs-use-a-seperate-lock-to-protect-nodehash.patch260
-rw-r--r--debian/patches/0002-fatfs-use-a-seperate-lock-to-protect-nodehash.patch304
-rw-r--r--debian/patches/0003-isofs-use-a-seperate-lock-to-protect-node_cache.patch310
-rw-r--r--debian/patches/0004-tmpfs-use-a-seperate-lock-to-protect-all_nodes.patch314
-rw-r--r--debian/patches/0005-libdiskfs-lock-less-reference-counting-of-nodes.patch610
-rw-r--r--debian/patches/series5
6 files changed, 1803 insertions, 0 deletions
diff --git a/debian/patches/0001-ext2fs-use-a-seperate-lock-to-protect-nodehash.patch b/debian/patches/0001-ext2fs-use-a-seperate-lock-to-protect-nodehash.patch
new file mode 100644
index 00000000..640337d1
--- /dev/null
+++ b/debian/patches/0001-ext2fs-use-a-seperate-lock-to-protect-nodehash.patch
@@ -0,0 +1,260 @@
+From 4166312a45357c2ff11b00219dfb83b7475ac4b1 Mon Sep 17 00:00:00 2001
+From: Justus Winter <4winter@informatik.uni-hamburg.de>
+Date: Tue, 13 May 2014 13:09:15 +0200
+Subject: [PATCH hurd 1/5] ext2fs: use a seperate lock to protect nodehash
+
+Previously, ext2fs used diskfs_node_refcnt_lock to serialize access to
+the nodehash.
+
+Use a separate lock to protect nodehash. Adjust the reference
+counting accordingly. Every node in the nodehash carries a light
+reference. When we are asked to give up that light reference, we
+reacquire our lock momentarily to check whether someone else
+reacquired a reference through the nodehash.
+
+* ext2fs/inode.c (nodecache_lock): New lock.
+(diskfs_cached_lookup): Use a separate lock to protect nodehash.
+Adjust the reference counting accordingly.
+(ifind): Likewise.
+(diskfs_node_iterate): Likewise.
+(diskfs_node_norefs): Move the code removing the node from nodehash...
+(diskfs_try_dropping_softrefs): ... here, where we check whether
+someone reacquired a reference, and if so hold on to our light
+reference.
+---
+ ext2fs/inode.c | 128 +++++++++++++++++++++++++++++++++++++++++----------------
+ 1 file changed, 92 insertions(+), 36 deletions(-)
+
+diff --git a/ext2fs/inode.c b/ext2fs/inode.c
+index 6b8b749..7c8d5a8 100644
+--- a/ext2fs/inode.c
++++ b/ext2fs/inode.c
+@@ -46,8 +46,19 @@
+ #define INOHASH(ino) (((unsigned)(ino))%INOHSZ)
+ #endif
+
++/* The nodehash is a cache of nodes.
++
++ Access to nodehash and nodehash_nr_items is protected by
++ nodecache_lock.
++
++ Every node in the nodehash carries a light reference. When we are
++ asked to give up that light reference, we reacquire our lock
++ momentarily to check whether someone else reacquired a reference
++ through the nodehash. */
+ static struct node *nodehash[INOHSZ];
+ static size_t nodehash_nr_items;
++/* nodecache_lock must be acquired before diskfs_node_refcnt_lock. */
++static pthread_rwlock_t nodecache_lock = PTHREAD_RWLOCK_INITIALIZER;
+
+ static error_t read_node (struct node *np);
+
+@@ -62,33 +73,37 @@ inode_init ()
+ nodehash[n] = 0;
+ }
+
++/* Lookup node with inode number INUM. Returns NULL if the node is
++ not found in the node cache. */
++static struct node *
++lookup (ino_t inum)
++{
++ struct node *np;
++ for (np = nodehash[INOHASH(inum)]; np; np = np->dn->hnext)
++ if (np->cache_id == inum)
++ return np;
++ return NULL;
++}
++
+ /* Fetch inode INUM, set *NPP to the node structure;
+ gain one user reference and lock the node. */
+ error_t
+ diskfs_cached_lookup (ino_t inum, struct node **npp)
+ {
+ error_t err;
+- struct node *np;
++ struct node *np, *tmp;
+ struct disknode *dn;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- for (np = nodehash[INOHASH(inum)]; np; np = np->dn->hnext)
+- if (np->cache_id == inum)
+- {
+- np->references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- pthread_mutex_lock (&np->lock);
+- *npp = np;
+- return 0;
+- }
++ pthread_rwlock_rdlock (&nodecache_lock);
++ np = lookup (inum);
++ if (np)
++ goto gotit;
++ pthread_rwlock_unlock (&nodecache_lock);
+
+ /* Format specific data for the new node. */
+ dn = malloc (sizeof (struct disknode));
+ if (! dn)
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- return ENOMEM;
+- }
++ return ENOMEM;
+ dn->dirents = 0;
+ dn->dir_idx = 0;
+ dn->pager = 0;
+@@ -102,14 +117,24 @@ diskfs_cached_lookup (ino_t inum, struct node **npp)
+ pthread_mutex_lock (&np->lock);
+
+ /* Put NP in NODEHASH. */
++ pthread_rwlock_wrlock (&nodecache_lock);
++ tmp = lookup (inum);
++ if (tmp)
++ {
++ /* We lost a race. */
++ diskfs_nput (np);
++ np = tmp;
++ goto gotit;
++ }
++
+ dn->hnext = nodehash[INOHASH(inum)];
+ if (dn->hnext)
+ dn->hnext->dn->hprevp = &dn->hnext;
+ dn->hprevp = &nodehash[INOHASH(inum)];
+ nodehash[INOHASH(inum)] = np;
++ diskfs_nref_light (np);
+ nodehash_nr_items += 1;
+-
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&nodecache_lock);
+
+ /* Get the contents of NP off disk. */
+ err = read_node (np);
+@@ -131,6 +156,13 @@ diskfs_cached_lookup (ino_t inum, struct node **npp)
+ *npp = np;
+ return 0;
+ }
++
++ gotit:
++ diskfs_nref (np);
++ pthread_rwlock_unlock (&nodecache_lock);
++ pthread_mutex_lock (&np->lock);
++ *npp = np;
++ return 0;
+ }
+
+ /* Lookup node INUM (which must have a reference already) and return it
+@@ -140,17 +172,12 @@ ifind (ino_t inum)
+ {
+ struct node *np;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- for (np = nodehash[INOHASH(inum)]; np; np = np->dn->hnext)
+- {
+- if (np->cache_id != inum)
+- continue;
++ pthread_rwlock_rdlock (&nodecache_lock);
++ np = lookup (inum);
++ pthread_rwlock_unlock (&nodecache_lock);
+
+- assert (np->references);
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- return np;
+- }
+- assert (0);
++ assert (np);
++ return np;
+ }
+
+ /* The last reference to a node has gone away; drop
+@@ -158,11 +185,6 @@ ifind (ino_t inum)
+ void
+ diskfs_node_norefs (struct node *np)
+ {
+- *np->dn->hprevp = np->dn->hnext;
+- if (np->dn->hnext)
+- np->dn->hnext->dn->hprevp = np->dn->hprevp;
+- nodehash_nr_items -= 1;
+-
+ if (np->dn->dirents)
+ free (np->dn->dirents);
+ assert (!np->dn->pager);
+@@ -180,6 +202,36 @@ diskfs_node_norefs (struct node *np)
+ void
+ diskfs_try_dropping_softrefs (struct node *np)
+ {
++ pthread_rwlock_wrlock (&nodecache_lock);
++ if (np->dn->hprevp != NULL)
++ {
++ /* Check if someone reacquired a reference through the
++ nodehash. */
++ unsigned int references;
++ pthread_spin_lock (&diskfs_node_refcnt_lock);
++ references = np->references;
++ pthread_spin_unlock (&diskfs_node_refcnt_lock);
++
++ /* An additional reference is acquired by libdiskfs across calls
++ to diskfs_try_dropping_softrefs. */
++ if (references > 1)
++ {
++ /* A reference was reacquired through a hash table lookup.
++ It's fine, we didn't touch anything yet. */
++ pthread_rwlock_unlock (&nodecache_lock);
++ return;
++ }
++
++ *np->dn->hprevp = np->dn->hnext;
++ if (np->dn->hnext)
++ np->dn->hnext->dn->hprevp = np->dn->hprevp;
++ np->dn->hnext = NULL;
++ np->dn->hprevp = NULL;
++ nodehash_nr_items -= 1;
++ diskfs_nrele_light (np);
++ }
++ pthread_rwlock_unlock (&nodecache_lock);
++
+ drop_pager_softrefs (np);
+ }
+
+@@ -556,12 +608,12 @@ diskfs_node_iterate (error_t (*fun)(struct node *))
+ size_t num_nodes;
+ struct node *node, **node_list, **p;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_rdlock (&nodecache_lock);
+
+ /* We must copy everything from the hash table into another data structure
+ to avoid running into any problems with the hash-table being modified
+ during processing (normally we delegate access to hash-table with
+- diskfs_node_refcnt_lock, but we can't hold this while locking the
++ nodecache_lock, but we can't hold this while locking the
+ individual node locks). */
+ num_nodes = nodehash_nr_items;
+
+@@ -570,7 +622,7 @@ diskfs_node_iterate (error_t (*fun)(struct node *))
+ node_list = malloc (num_nodes * sizeof (struct node *));
+ if (node_list == NULL)
+ {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&nodecache_lock);
+ ext2_debug ("unable to allocate temporary node table");
+ return ENOMEM;
+ }
+@@ -580,10 +632,14 @@ diskfs_node_iterate (error_t (*fun)(struct node *))
+ for (node = nodehash[n]; node; node = node->dn->hnext)
+ {
+ *p++ = node;
++
++ /* We acquire a hard reference for node, but without using
++ diskfs_nref. We do this so that diskfs_new_hardrefs will not
++ get called. */
+ node->references++;
+ }
+
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&nodecache_lock);
+
+ p = node_list;
+ while (num_nodes-- > 0)
+--
+2.1.4
+
diff --git a/debian/patches/0002-fatfs-use-a-seperate-lock-to-protect-nodehash.patch b/debian/patches/0002-fatfs-use-a-seperate-lock-to-protect-nodehash.patch
new file mode 100644
index 00000000..9dfd9bc8
--- /dev/null
+++ b/debian/patches/0002-fatfs-use-a-seperate-lock-to-protect-nodehash.patch
@@ -0,0 +1,304 @@
+From e17366f09a6003dcc95633ce312aa3753f056e28 Mon Sep 17 00:00:00 2001
+From: Justus Winter <4winter@informatik.uni-hamburg.de>
+Date: Tue, 13 May 2014 15:14:53 +0200
+Subject: [PATCH hurd 2/5] fatfs: use a seperate lock to protect nodehash
+
+Previously, fatfs used diskfs_node_refcnt_lock to serialize access to
+the nodehash.
+
+Use a separate lock to protect nodehash. Adjust the reference
+counting accordingly. Every node in the nodehash carries a light
+reference. When we are asked to give up that light reference, we
+reacquire our lock momentarily to check whether someone else
+reacquired a reference through the nodehash.
+
+* fatfs/inode.c (nodecache_lock): New lock.
+(diskfs_cached_lookup): Use a separate lock to protect nodehash.
+Adjust the reference counting accordingly.
+(ifind): Likewise.
+(diskfs_node_iterate): Likewise.
+(diskfs_node_norefs): Move the code removing the node from nodehash...
+(diskfs_try_dropping_softrefs): ... here, where we check whether
+someone reacquired a reference, and if so hold on to our light
+reference.
+---
+ fatfs/inode.c | 147 ++++++++++++++++++++++++++++++++++++++++------------------
+ 1 file changed, 102 insertions(+), 45 deletions(-)
+
+diff --git a/fatfs/inode.c b/fatfs/inode.c
+index ed6f3f0..1d670f5 100644
+--- a/fatfs/inode.c
++++ b/fatfs/inode.c
+@@ -44,8 +44,19 @@
+ #define INOHASH(ino) (((unsigned)(ino))%INOHSZ)
+ #endif
+
++/* The nodehash is a cache of nodes.
++
++ Access to nodehash and nodehash_nr_items is protected by
++ nodecache_lock.
++
++ Every node in the nodehash carries a light reference. When we are
++ asked to give up that light reference, we reacquire our lock
++ momentarily to check whether someone else reacquired a reference
++ through the nodehash. */
+ static struct node *nodehash[INOHSZ];
+ static size_t nodehash_nr_items;
++/* nodecache_lock must be acquired before diskfs_node_refcnt_lock. */
++static pthread_rwlock_t nodecache_lock = PTHREAD_RWLOCK_INITIALIZER;
+
+ static error_t read_node (struct node *np, vm_address_t buf);
+
+@@ -58,33 +69,38 @@ inode_init ()
+ nodehash[n] = 0;
+ }
+
++/* Lookup node with inode number INUM. Returns NULL if the node is
++ not found in the node cache. */
++static struct node *
++lookup (ino_t inum)
++{
++ struct node *np;
++ for (np = nodehash[INOHASH(inum)]; np; np = np->dn->hnext)
++ if (np->cache_id == inum)
++ return np;
++ return NULL;
++}
++
+ /* Fetch inode INUM, set *NPP to the node structure; gain one user
+ reference and lock the node. */
+ error_t
+ diskfs_cached_lookup (ino64_t inum, struct node **npp)
+ {
+ error_t err;
+- struct node *np;
++ struct node *np, *tmp;
+ struct disknode *dn;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- for (np = nodehash[INOHASH(inum)]; np; np = np->dn->hnext)
+- if (np->cache_id == inum)
+- {
+- np->references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- pthread_mutex_lock (&np->lock);
+- *npp = np;
+- return 0;
+- }
++ pthread_rwlock_rdlock (&nodecache_lock);
++ np = lookup (inum);
++ if (np)
++ goto gotit;
++ pthread_rwlock_unlock (&nodecache_lock);
+
+ /* Format specific data for the new node. */
+ dn = malloc (sizeof (struct disknode));
+ if (! dn)
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- return ENOMEM;
+- }
++ return ENOMEM;
++
+ dn->pager = 0;
+ dn->first = 0;
+ dn->last = 0;
+@@ -102,15 +118,25 @@ diskfs_cached_lookup (ino64_t inum, struct node **npp)
+ pthread_mutex_lock (&np->lock);
+
+ /* Put NP in NODEHASH. */
++ pthread_rwlock_wrlock (&nodecache_lock);
++ tmp = lookup (inum);
++ if (tmp)
++ {
++ /* We lost a race. */
++ diskfs_nput (np);
++ np = tmp;
++ goto gotit;
++ }
++
+ dn->hnext = nodehash[INOHASH(inum)];
+ if (dn->hnext)
+ dn->hnext->dn->hprevp = &dn->hnext;
+ dn->hprevp = &nodehash[INOHASH(inum)];
+ nodehash[INOHASH(inum)] = np;
++ diskfs_nref_light (np);
+ nodehash_nr_items += 1;
++ pthread_rwlock_unlock (&nodecache_lock);
+
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+-
+ /* Get the contents of NP off disk. */
+ err = read_node (np, 0);
+
+@@ -121,6 +147,13 @@ diskfs_cached_lookup (ino64_t inum, struct node **npp)
+ *npp = np;
+ return 0;
+ }
++
++ gotit:
++ diskfs_nref (np);
++ pthread_rwlock_unlock (&nodecache_lock);
++ pthread_mutex_lock (&np->lock);
++ *npp = np;
++ return 0;
+ }
+
+ /* Fetch inode INUM, set *NPP to the node structure;
+@@ -133,24 +166,23 @@ diskfs_cached_lookup_in_dirbuf (int inum, struct node **npp, vm_address_t buf)
+ struct node *np;
+ struct disknode *dn;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_rdlock (&nodecache_lock);
+ for (np = nodehash[INOHASH(inum)]; np; np = np->dn->hnext)
+ if (np->cache_id == inum)
+ {
+- np->references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ diskfs_nref (np);
++ pthread_rwlock_unlock (&nodecache_lock);
+ pthread_mutex_lock (&np->lock);
+ *npp = np;
+ return 0;
+ }
++ pthread_rwlock_unlock (&nodecache_lock);
+
+ /* Format specific data for the new node. */
+ dn = malloc (sizeof (struct disknode));
+ if (! dn)
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- return ENOMEM;
+- }
++ return ENOMEM;
++
+ dn->pager = 0;
+ dn->first = 0;
+ dn->last = 0;
+@@ -168,15 +200,16 @@ diskfs_cached_lookup_in_dirbuf (int inum, struct node **npp, vm_address_t buf)
+ pthread_mutex_lock (&np->lock);
+
+ /* Put NP in NODEHASH. */
++ pthread_rwlock_wrlock (&nodecache_lock);
+ dn->hnext = nodehash[INOHASH(inum)];
+ if (dn->hnext)
+ dn->hnext->dn->hprevp = &dn->hnext;
+ dn->hprevp = &nodehash[INOHASH(inum)];
+ nodehash[INOHASH(inum)] = np;
++ diskfs_nref_light (np);
+ nodehash_nr_items += 1;
++ pthread_rwlock_unlock (&nodecache_lock);
+
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+-
+ /* Get the contents of NP off disk. */
+ err = read_node (np, buf);
+
+@@ -196,17 +229,12 @@ ifind (ino_t inum)
+ {
+ struct node *np;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- for (np = nodehash[INOHASH(inum)]; np; np = np->dn->hnext)
+- {
+- if (np->cache_id != inum)
+- continue;
++ pthread_rwlock_rdlock (&nodecache_lock);
++ np = lookup (inum);
++ pthread_rwlock_unlock (&nodecache_lock);
+
+- assert (np->references);
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- return np;
+- }
+- assert (0);
++ assert (np);
++ return np;
+ }
+
+ /* The last reference to a node has gone away; drop it from the hash
+@@ -216,11 +244,6 @@ diskfs_node_norefs (struct node *np)
+ {
+ struct cluster_chain *last = np->dn->first;
+
+- *np->dn->hprevp = np->dn->hnext;
+- if (np->dn->hnext)
+- np->dn->hnext->dn->hprevp = np->dn->hprevp;
+- nodehash_nr_items -= 1;
+-
+ while (last)
+ {
+ struct cluster_chain *next = last->next;
+@@ -251,6 +274,36 @@ diskfs_node_norefs (struct node *np)
+ void
+ diskfs_try_dropping_softrefs (struct node *np)
+ {
++ pthread_rwlock_wrlock (&nodecache_lock);
++ if (np->dn->hprevp != NULL)
++ {
++ /* Check if someone reacquired a reference through the
++ nodehash. */
++ unsigned int references;
++ pthread_spin_lock (&diskfs_node_refcnt_lock);
++ references = np->references;
++ pthread_spin_unlock (&diskfs_node_refcnt_lock);
++
++ /* An additional reference is acquired by libdiskfs across calls
++ to diskfs_try_dropping_softrefs. */
++ if (references > 1)
++ {
++ /* A reference was reacquired through a hash table lookup.
++ It's fine, we didn't touch anything yet. */
++ pthread_rwlock_unlock (&nodecache_lock);
++ return;
++ }
++
++ *np->dn->hprevp = np->dn->hnext;
++ if (np->dn->hnext)
++ np->dn->hnext->dn->hprevp = np->dn->hprevp;
++ np->dn->hnext = NULL;
++ np->dn->hprevp = NULL;
++ nodehash_nr_items -= 1;
++ diskfs_nrele_light (np);
++ }
++ pthread_rwlock_unlock (&nodecache_lock);
++
+ drop_pager_softrefs (np);
+ }
+
+@@ -554,12 +607,12 @@ diskfs_node_iterate (error_t (*fun)(struct node *))
+ size_t num_nodes;
+ struct node *node, **node_list, **p;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_rdlock (&nodecache_lock);
+
+ /* We must copy everything from the hash table into another data structure
+ to avoid running into any problems with the hash-table being modified
+ during processing (normally we delegate access to hash-table with
+- diskfs_node_refcnt_lock, but we can't hold this while locking the
++ nodecache_lock, but we can't hold this while locking the
+ individual node locks). */
+
+ num_nodes = nodehash_nr_items;
+@@ -570,10 +623,14 @@ diskfs_node_iterate (error_t (*fun)(struct node *))
+ for (node = nodehash[n]; node; node = node->dn->hnext)
+ {
+ *p++ = node;
+- node->references++;
++
++ /* We acquire a hard reference for node, but without using
++ diskfs_nref. We do this so that diskfs_new_hardrefs will not
++ get called. */
++ node->references++;
+ }
+
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&nodecache_lock);
+
+ p = node_list;
+ while (num_nodes-- > 0)
+--
+2.1.4
+
diff --git a/debian/patches/0003-isofs-use-a-seperate-lock-to-protect-node_cache.patch b/debian/patches/0003-isofs-use-a-seperate-lock-to-protect-node_cache.patch
new file mode 100644
index 00000000..fc3fc03b
--- /dev/null
+++ b/debian/patches/0003-isofs-use-a-seperate-lock-to-protect-node_cache.patch
@@ -0,0 +1,310 @@
+From 912151f13737af551d3d3dd2e45faef3bd7c78cb Mon Sep 17 00:00:00 2001
+From: Justus Winter <4winter@informatik.uni-hamburg.de>
+Date: Tue, 13 May 2014 15:16:31 +0200
+Subject: [PATCH hurd 3/5] isofs: use a seperate lock to protect node_cache
+
+Previously, isofs used diskfs_node_refcnt_lock to serialize access to
+the node_cache.
+
+Use a separate lock to protect node_cache. Adjust the reference
+counting accordingly. Every node in the node_cache carries a light
+reference. When we are asked to give up that light reference, we
+reacquire our lock momentarily to check whether someone else
+reacquired a reference through the node_cache.
+
+* isofs/inode.c (nodecache_lock): New lock.
+(inode_cache_find): Use a separate lock to protect node_cache.
+Adjust the reference counting accordingly.
+(diskfs_cached_lookup): Likewise.
+(load_inode): Likewise.
+(cache_inode): Update comment accordingly.
+(diskfs_node_iterate): Likewise.
+(diskfs_node_norefs): Move the code removing the node from node_cache...
+(diskfs_try_dropping_softrefs): ... here, where we check whether
+someone reacquired a reference, and if so hold on to our light
+reference.
+---
+ isofs/inode.c | 146 +++++++++++++++++++++++++++++++++++++++++-----------------
+ 1 file changed, 105 insertions(+), 41 deletions(-)
+
+diff --git a/isofs/inode.c b/isofs/inode.c
+index 247d8ac..37bf1ac 100644
+--- a/isofs/inode.c
++++ b/isofs/inode.c
+@@ -48,35 +48,53 @@ struct node_cache
+ struct node *np; /* if live */
+ };
+
++/* The node_cache is a cache of nodes.
++
++ Access to node_cache, node_cache_size, and node_cache_alloced is
++ protected by nodecache_lock.
++
++ Every node in the node_cache carries a light reference. When we
++ are asked to give up that light reference, we reacquire our lock
++ momentarily to check whether someone else reacquired a reference
++ through the node_cache. */
+ static int node_cache_size = 0;
+ static int node_cache_alloced = 0;
+ struct node_cache *node_cache = 0;
++/* nodecache_lock must be acquired before diskfs_node_refcnt_lock. */
++static pthread_rwlock_t nodecache_lock = PTHREAD_RWLOCK_INITIALIZER;
+
+ /* Forward */
+ static error_t read_disknode (struct node *,
+ struct dirrect *, struct rrip_lookup *);
+
+
++/* Lookup node with id ID. Returns NULL if the node is not found in
++ the node cache. */
++static struct node *
++lookup (off_t id)
++{
++ int i;
++ for (i = 0; i < node_cache_size; i++)
++ if (node_cache[i].id == id
++ && node_cache[i].np)
++ return node_cache[i].np;
++ return NULL;
++}
++
+ /* See if node with identifier ID is in the cache. If so, return it,
+- with one additional reference. diskfs_node_refcnt_lock must be held
++ with one additional reference. nodecache_lock must be held
+ on entry to the call, and will be released iff the node was found
+ in the cache. */
+ void
+ inode_cache_find (off_t id, struct node **npp)
+ {
+- int i;
+-
+- for (i = 0; i < node_cache_size; i++)
+- if (node_cache[i].id == id
+- && node_cache[i].np)
+- {
+- *npp = node_cache[i].np;
+- (*npp)->references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- pthread_mutex_lock (&(*npp)->lock);
+- return;
+- }
+- *npp = 0;
++ *npp = lookup (id);
++ if (*npp)
++ {
++ diskfs_nref (*npp);
++ pthread_rwlock_unlock (&nodecache_lock);
++ pthread_mutex_lock (&(*npp)->lock);
++ }
+ }
+
+
+@@ -92,7 +110,7 @@ use_file_start_id (struct dirrect *record, struct rrip_lookup *rr)
+ }
+
+ /* Enter NP into the cache. The directory entry we used is DR, the
+- cached Rock-Ridge info RR. diskfs_node_refcnt_lock must be held. */
++ cached Rock-Ridge info RR. nodecache_lock must be held. */
+ void
+ cache_inode (struct node *np, struct dirrect *record,
+ struct rrip_lookup *rr)
+@@ -137,6 +155,7 @@ cache_inode (struct node *np, struct dirrect *record,
+ c->id = id;
+ c->dr = record;
+ c->file_start = np->dn->file_start;
++ diskfs_nref_light (np);
+ c->np = np;
+
+ /* PLUS 1 so that we don't store zero cache ID's (not allowed by diskfs) */
+@@ -155,7 +174,7 @@ diskfs_cached_lookup (ino_t id, struct node **npp)
+ to avoid presenting zero cache ID's. */
+ id--;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_rdlock (&nodecache_lock);
+ assert (id < node_cache_size);
+
+ np = node_cache[id].np;
+@@ -166,6 +185,8 @@ diskfs_cached_lookup (ino_t id, struct node **npp)
+ struct rrip_lookup rr;
+ struct disknode *dn;
+
++ pthread_rwlock_unlock (&nodecache_lock);
++
+ rrip_lookup (node_cache[id].dr, &rr, 1);
+
+ /* We should never cache the wrong directory entry */
+@@ -174,7 +195,7 @@ diskfs_cached_lookup (ino_t id, struct node **npp)
+ dn = malloc (sizeof (struct disknode));
+ if (!dn)
+ {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&nodecache_lock);
+ release_rrip (&rr);
+ return ENOMEM;
+ }
+@@ -185,16 +206,26 @@ diskfs_cached_lookup (ino_t id, struct node **npp)
+ if (!np)
+ {
+ free (dn);
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&nodecache_lock);
+ release_rrip (&rr);
+ return ENOMEM;
+ }
+ np->cache_id = id + 1; /* see above for rationale for increment */
+ pthread_mutex_lock (&np->lock);
++
++ pthread_rwlock_wrlock (&nodecache_lock);
++ if (c->np != NULL)
++ {
++ /* We lost a race. */
++ diskfs_nput (np);
++ np = c->np;
++ goto gotit;
++ }
+ c->np = np;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ diskfs_nref_light (np);
++ pthread_rwlock_unlock (&nodecache_lock);
+
+- err = read_disknode (np, node_cache[id].dr, &rr);
++ err = read_disknode (np, dn->dr, &rr);
+ if (!err)
+ *npp = np;
+
+@@ -203,9 +234,9 @@ diskfs_cached_lookup (ino_t id, struct node **npp)
+ return err;
+ }
+
+-
+- np->references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ gotit:
++ diskfs_nref (np);
++ pthread_rwlock_unlock (&nodecache_lock);
+ pthread_mutex_lock (&np->lock);
+ *npp = np;
+ return 0;
+@@ -307,7 +338,8 @@ load_inode (struct node **npp, struct dirrect *record,
+ error_t err;
+ off_t file_start;
+ struct disknode *dn;
+- struct node *np;
++ struct node *np, *tmp;
++ off_t id;
+
+ err = calculate_file_start (record, &file_start, rr);
+ if (err)
+@@ -315,27 +347,23 @@ load_inode (struct node **npp, struct dirrect *record,
+ if (rr->valid & VALID_CL)
+ record = rr->realdirent;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+-
+ /* First check the cache */
+ if (use_file_start_id (record, rr))
+- inode_cache_find (file_start << store->log2_block_size, npp);
++ id = file_start << store->log2_block_size;
+ else
+- inode_cache_find ((off_t) ((void *) record - (void *) disk_image), npp);
++ id = (off_t) ((void *) record - (void *) disk_image);
+
++ pthread_rwlock_rdlock (&nodecache_lock);
++ inode_cache_find (id, npp);
+ if (*npp)
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- return 0;
+- }
++ return 0;
++ pthread_rwlock_unlock (&nodecache_lock);
+
+ /* Create a new node */
+ dn = malloc (sizeof (struct disknode));
+ if (!dn)
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- return ENOMEM;
+- }
++ return ENOMEM;
++
+ dn->fileinfo = 0;
+ dn->dr = record;
+ dn->file_start = file_start;
+@@ -344,14 +372,25 @@ load_inode (struct node **npp, struct dirrect *record,
+ if (!np)
+ {
+ free (dn);
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+ return ENOMEM;
+ }
+
+ pthread_mutex_lock (&np->lock);
+
++ pthread_rwlock_wrlock (&nodecache_lock);
++ tmp = lookup (id);
++ if (tmp)
++ {
++ /* We lost a race. */
++ diskfs_nput (np);
++ diskfs_nref (tmp);
++ *npp = tmp;
++ pthread_rwlock_unlock (&nodecache_lock);
++ return 0;
++ }
++
+ cache_inode (np, record, rr);
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&nodecache_lock);
+
+ err = read_disknode (np, record, rr);
+ *npp = np;
+@@ -505,9 +544,6 @@ error_t (*diskfs_read_symlink_hook) (struct node *, char *)
+ void
+ diskfs_node_norefs (struct node *np)
+ {
+- assert (node_cache[np->cache_id - 1].np == np);
+- node_cache[np->cache_id - 1].np = 0;
+-
+ if (np->dn->translator)
+ free (np->dn->translator);
+
+@@ -521,6 +557,34 @@ diskfs_node_norefs (struct node *np)
+ void
+ diskfs_try_dropping_softrefs (struct node *np)
+ {
++ pthread_rwlock_wrlock (&nodecache_lock);
++ if (np->cache_id != 0)
++ {
++ assert (node_cache[np->cache_id - 1].np == np);
++
++ /* Check if someone reacquired a reference through the
++ node_cache. */
++ unsigned int references;
++ pthread_spin_lock (&diskfs_node_refcnt_lock);
++ references = np->references;
++ pthread_spin_unlock (&diskfs_node_refcnt_lock);
++
++ /* An additional reference is acquired by libdiskfs across calls
++ to diskfs_try_dropping_softrefs. */
++ if (references > 1)
++ {
++ /* A reference was reacquired through a hash table lookup.
++ It's fine, we didn't touch anything yet. */
++ pthread_rwlock_unlock (&nodecache_lock);
++ return;
++ }
++
++ node_cache[np->cache_id - 1].np = 0;
++ np->cache_id = 0;
++ diskfs_nrele_light (np);
++ }
++ pthread_rwlock_unlock (&nodecache_lock);
++
+ drop_pager_softrefs (np);
+ }
+
+--
+2.1.4
+
diff --git a/debian/patches/0004-tmpfs-use-a-seperate-lock-to-protect-all_nodes.patch b/debian/patches/0004-tmpfs-use-a-seperate-lock-to-protect-all_nodes.patch
new file mode 100644
index 00000000..81a00ed1
--- /dev/null
+++ b/debian/patches/0004-tmpfs-use-a-seperate-lock-to-protect-all_nodes.patch
@@ -0,0 +1,314 @@
+From ba3560bd19188297a2bc51447b2006ced7d5c615 Mon Sep 17 00:00:00 2001
+From: Justus Winter <4winter@informatik.uni-hamburg.de>
+Date: Tue, 13 May 2014 15:35:42 +0200
+Subject: [PATCH hurd 4/5] tmpfs: use a seperate lock to protect all_nodes
+
+Previously, tmpfs used diskfs_node_refcnt_lock to serialize access to
+the all_nodes and some other related global state related to memory
+consumption.
+
+Use a separate lock to protect all_nodes, and atomic operations to
+access the state related to memory consumption. Adjust the reference
+counting accordingly. Every node in the all_nodes carries a light
+reference. When we are asked to give up that light reference, we
+reacquire our lock momentarily to check whether someone else
+reacquired a reference through the all_nodes.
+
+* tmpfs/tmpfs.h (num_files, tmpfs_space_used): Use atomic operations
+for these variables.
+(adjust_used): Use atomic operations.
+(get_used): New convenience function to atomically retrieve
+tmpfs_space_used.
+* tmpfs/node.c (all_nodes_lock): New lock.
+(diskfs_alloc_node): Use a separate lock to protect all_nodes.
+Adjust the reference counting accordingly.
+(diskfs_free_node): Likewise.
+(diskfs_cached_lookup):Likewise.
+(diskfs_node_iterate): Likewise.
+(diskfs_node_norefs): Do not remove the node from all_nodes. This
+actually looks like a mistake, I do not know why they did that here as
+well as in diskfs_free_node.
+(diskfs_try_dropping_softrefs): Check whether someone reacquired a
+reference, and if so hold on to our light reference.
+(diskfs_grow): Use atomic operations.
+* tmpfs/tmpfs.c (diskfs_set_statfs): Likewise.
+---
+ tmpfs/node.c | 107 ++++++++++++++++++++++++++++++++++++++++++----------------
+ tmpfs/tmpfs.c | 6 ++--
+ tmpfs/tmpfs.h | 20 +++++++----
+ 3 files changed, 94 insertions(+), 39 deletions(-)
+
+diff --git a/tmpfs/node.c b/tmpfs/node.c
+index 02d7a60..428b6d9 100644
+--- a/tmpfs/node.c
++++ b/tmpfs/node.c
+@@ -29,8 +29,19 @@
+ unsigned int num_files;
+ static unsigned int gen;
+
++/* all_nodes is a list of all nodes.
++
++ Access to all_nodes and all_nodes_nr_items is protected by
++ all_nodes_lock.
++
++ Every node in all_nodes carries a light reference. When we are
++ asked to give up that light reference, we reacquire our lock
++ momentarily to check whether someone else reacquired a
++ reference. */
+ struct node *all_nodes;
+ static size_t all_nodes_nr_items;
++/* all_nodes_lock must be acquired before diskfs_node_refcnt_lock. */
++pthread_rwlock_t all_nodes_lock = PTHREAD_RWLOCK_INITIALIZER;
+
+ error_t
+ diskfs_alloc_node (struct node *dp, mode_t mode, struct node **npp)
+@@ -40,18 +51,17 @@ diskfs_alloc_node (struct node *dp, mode_t mode, struct node **npp)
+ dn = calloc (1, sizeof *dn);
+ if (dn == 0)
+ return ENOSPC;
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- if (round_page (tmpfs_space_used + sizeof *dn) / vm_page_size
++
++ if (round_page (get_used () + sizeof *dn) / vm_page_size
+ > tmpfs_page_limit)
+ {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&all_nodes_lock);
+ free (dn);
+ return ENOSPC;
+ }
+ dn->gen = gen++;
+- ++num_files;
+- tmpfs_space_used += sizeof *dn;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ __atomic_add_fetch (&num_files, 1, __ATOMIC_RELAXED);
++ adjust_used (sizeof *dn);
+
+ dn->type = IFTODT (mode & S_IFMT);
+ return diskfs_cached_lookup ((ino_t) (uintptr_t) dn, npp);
+@@ -75,15 +85,19 @@ diskfs_free_node (struct node *np, mode_t mode)
+ free (np->dn->u.lnk);
+ break;
+ }
++
++ pthread_rwlock_wrlock (&all_nodes_lock);
+ *np->dn->hprevp = np->dn->hnext;
+ if (np->dn->hnext != 0)
+ np->dn->hnext->dn->hprevp = np->dn->hprevp;
+ all_nodes_nr_items -= 1;
++ pthread_rwlock_unlock (&all_nodes_lock);
++
+ free (np->dn);
+ np->dn = 0;
+
+- --num_files;
+- tmpfs_space_used -= sizeof *np->dn;
++ __atomic_sub_fetch (&num_files, 1, __ATOMIC_RELAXED);
++ adjust_used (-sizeof *np->dn);
+ }
+
+ void
+@@ -117,14 +131,6 @@ diskfs_node_norefs (struct node *np)
+ np->dn->u.chr = np->dn_stat.st_rdev;
+ break;
+ }
+-
+- /* Remove this node from the cache list rooted at `all_nodes'. */
+- *np->dn->hprevp = np->dn->hnext;
+- if (np->dn->hnext != 0)
+- np->dn->hnext->dn->hprevp = np->dn->hprevp;
+- all_nodes_nr_items -= 1;
+- np->dn->hnext = 0;
+- np->dn->hprevp = 0;
+ }
+
+ free (np);
+@@ -167,30 +173,34 @@ diskfs_cached_lookup (ino_t inum, struct node **npp)
+
+ assert (npp);
+
++ pthread_rwlock_rdlock (&all_nodes_lock);
+ if (dn->hprevp != 0) /* There is already a node. */
+- {
+- np = *dn->hprevp;
+- assert (np->dn == dn);
+- assert (*dn->hprevp == np);
+-
+- diskfs_nref (np);
+- }
++ goto gotit;
+ else
+ /* Create the new node. */
+ {
+ struct stat *st;
++ pthread_rwlock_unlock (&all_nodes_lock);
+
+ np = diskfs_make_node (dn);
+ np->cache_id = (ino_t) (uintptr_t) dn;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_wrlock (&all_nodes_lock);
++ if (dn->hprevp != NULL)
++ {
++ /* We lost a race. */
++ diskfs_nrele (np);
++ goto gotit;
++ }
++
+ dn->hnext = all_nodes;
+ if (dn->hnext)
+ dn->hnext->dn->hprevp = &dn->hnext;
+ dn->hprevp = &all_nodes;
+ all_nodes = np;
+ all_nodes_nr_items += 1;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ diskfs_nref_light (np);
++ pthread_rwlock_unlock (&all_nodes_lock);
+
+ st = &np->dn_stat;
+ memset (st, 0, sizeof *st);
+@@ -220,6 +230,16 @@ diskfs_cached_lookup (ino_t inum, struct node **npp)
+ pthread_mutex_lock (&np->lock);
+ *npp = np;
+ return 0;
++
++ gotit:
++ np = *dn->hprevp;
++ assert (np->dn == dn);
++ assert (*dn->hprevp == np);
++ diskfs_nref (np);
++ pthread_rwlock_unlock (&all_nodes_lock);
++ pthread_mutex_lock (&np->lock);
++ *npp = np;
++ return 0;
+ }
+
+ error_t
+@@ -229,12 +249,12 @@ diskfs_node_iterate (error_t (*fun) (struct node *))
+ size_t num_nodes;
+ struct node *node, **node_list, **p;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_rdlock (&all_nodes_lock);
+
+ /* We must copy everything from the hash table into another data structure
+ to avoid running into any problems with the hash-table being modified
+ during processing (normally we delegate access to hash-table with
+- diskfs_node_refcnt_lock, but we can't hold this while locking the
++ all_nodes_lock, but we can't hold this while locking the
+ individual node locks). */
+
+ num_nodes = all_nodes_nr_items;
+@@ -243,10 +263,14 @@ diskfs_node_iterate (error_t (*fun) (struct node *))
+ for (node = all_nodes; node != 0; node = node->dn->hnext)
+ {
+ *p++ = node;
++
++ /* We acquire a hard reference for node, but without using
++ diskfs_nref. We do this so that diskfs_new_hardrefs will not
++ get called. */
+ node->references++;
+ }
+
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ pthread_rwlock_unlock (&all_nodes_lock);
+
+ p = node_list;
+ while (num_nodes-- > 0)
+@@ -272,6 +296,31 @@ diskfs_node_iterate (error_t (*fun) (struct node *))
+ void
+ diskfs_try_dropping_softrefs (struct node *np)
+ {
++ pthread_rwlock_wrlock (&all_nodes_lock);
++ if (np->cache_id != 0)
++ {
++ /* Check if someone reacquired a reference. */
++ unsigned int references;
++ pthread_spin_lock (&diskfs_node_refcnt_lock);
++ references = np->references;
++ pthread_spin_unlock (&diskfs_node_refcnt_lock);
++
++ /* An additional reference is acquired by libdiskfs across calls
++ to diskfs_try_dropping_softrefs. */
++ if (references > 1)
++ {
++ /* A reference was reacquired. It's fine, we didn't touch
++ anything yet. */
++ pthread_rwlock_unlock (&all_nodes_lock);
++ return;
++ }
++
++ /* Just let go of the weak reference. The node will be removed
++ from all_nodes in diskfs_free_node. */
++ np->cache_id = 0;
++ diskfs_nrele_light (np);
++ }
++ pthread_rwlock_unlock (&all_nodes_lock);
+ }
+
+ /* The user must define this funcction. Node NP has some light
+@@ -447,7 +496,7 @@ diskfs_grow (struct node *np, off_t size, struct protid *cred)
+
+ off_t set_size = size;
+ size = round_page (size);
+- if (round_page (tmpfs_space_used + size - np->allocsize)
++ if (round_page (get_used () + size - np->allocsize)
+ / vm_page_size > tmpfs_page_limit)
+ return ENOSPC;
+
+diff --git a/tmpfs/tmpfs.c b/tmpfs/tmpfs.c
+index 1b5b374..fd1c9aa 100644
+--- a/tmpfs/tmpfs.c
++++ b/tmpfs/tmpfs.c
+@@ -67,10 +67,8 @@ diskfs_set_statfs (struct statfs *st)
+ st->f_bsize = vm_page_size;
+ st->f_blocks = tmpfs_page_limit;
+
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- st->f_files = num_files;
+- pages = round_page (tmpfs_space_used) / vm_page_size;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ st->f_files = __atomic_load_n (&num_files, __ATOMIC_RELAXED);
++ pages = round_page (get_used ()) / vm_page_size;
+
+ st->f_bfree = pages < tmpfs_page_limit ? tmpfs_page_limit - pages : 0;
+ st->f_bavail = st->f_bfree;
+diff --git a/tmpfs/tmpfs.h b/tmpfs/tmpfs.h
+index b3c636d..ad47200 100644
+--- a/tmpfs/tmpfs.h
++++ b/tmpfs/tmpfs.h
+@@ -69,17 +69,25 @@ struct tmpfs_dirent
+ char name[0];
+ };
+
+-extern unsigned int num_files;
+-extern off_t tmpfs_page_limit, tmpfs_space_used;
+-
++extern off_t tmpfs_page_limit;
+ extern mach_port_t default_pager;
+
++/* These two must be accessed using atomic operations. */
++extern unsigned int num_files;
++extern off_t tmpfs_space_used;
++
++/* Convenience function to adjust tmpfs_space_used. */
+ static inline void
+ adjust_used (off_t change)
+ {
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- tmpfs_space_used += change;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ __atomic_add_fetch (&num_files, change, __ATOMIC_RELAXED);
++}
++
++/* Convenience function to get tmpfs_space_used. */
++static inline off_t
++get_used (void)
++{
++ return __atomic_load_n (&num_files, __ATOMIC_RELAXED);
+ }
+
+ #endif
+--
+2.1.4
+
diff --git a/debian/patches/0005-libdiskfs-lock-less-reference-counting-of-nodes.patch b/debian/patches/0005-libdiskfs-lock-less-reference-counting-of-nodes.patch
new file mode 100644
index 00000000..8d5b9a5c
--- /dev/null
+++ b/debian/patches/0005-libdiskfs-lock-less-reference-counting-of-nodes.patch
@@ -0,0 +1,610 @@
+From e7f6157094adb09c027ef607ceb86d95442fb357 Mon Sep 17 00:00:00 2001
+From: Justus Winter <4winter@informatik.uni-hamburg.de>
+Date: Wed, 14 May 2014 11:19:35 +0200
+Subject: [PATCH hurd 5/5] libdiskfs: lock-less reference counting of nodes
+
+* libdiskfs/diskfs.h (struct node): Use refcounts_t for reference counting.
+(diskfs_node_refcnt_lock): Remove.
+(diskfs_node_norefs,diskfs_drop_node): Change comments accordingly.
+* libdiskfs/init-init.c: Likewise.
+* libdiskfs/node-drop.c: Likewise.
+* libdiskfs/node-make.c: Likewise.
+* libdiskfs/node-nput.c: Likewise.
+* libdiskfs/node-nputl.c: Likewise.
+* libdiskfs/node-nref.c: Likewise.
+* libdiskfs/node-nrefl.c: Likewise.
+* libdiskfs/node-nrele.c: Likewise.
+* libdiskfs/node-nrelel.c: Likewise.
+* ext2fs/inode.c: Likewise.
+* fatfs/inode.c: Likewise.
+* isofs/inode.c: Likewise.
+* tmpfs/node.c: Likewise.
+* doc/hurd.texi: Likewise.
+---
+ doc/hurd.texi | 11 ++---------
+ ext2fs/inode.c | 15 +++++---------
+ fatfs/inode.c | 27 ++++++++-----------------
+ isofs/inode.c | 13 ++++---------
+ libdiskfs/diskfs.h | 15 ++++++--------
+ libdiskfs/init-init.c | 2 --
+ libdiskfs/node-drop.c | 11 +++++------
+ libdiskfs/node-make.c | 3 +--
+ libdiskfs/node-nput.c | 52 +++++++++++++++++++------------------------------
+ libdiskfs/node-nputl.c | 12 ++++--------
+ libdiskfs/node-nref.c | 9 +++------
+ libdiskfs/node-nrefl.c | 4 +---
+ libdiskfs/node-nrele.c | 48 ++++++++++++++++++++++-----------------------
+ libdiskfs/node-nrelel.c | 9 +++------
+ tmpfs/node.c | 15 +++++---------
+ 15 files changed, 90 insertions(+), 156 deletions(-)
+
+diff --git a/doc/hurd.texi b/doc/hurd.texi
+index 7e7b5ee..2f36bdc 100644
+--- a/doc/hurd.texi
++++ b/doc/hurd.texi
+@@ -3780,10 +3780,6 @@ new thread and (eventually) get rid of the old one; the old thread won't
+ do any more syncs, regardless.
+ @end deftypefun
+
+-@deftypevar spin_lock_t diskfs_node_refcnt_lock
+-Pager reference count lock.
+-@end deftypevar
+-
+ @deftypevar int diskfs_readonly
+ Set to zero if the filesystem is currently writable.
+ @end deftypevar
+@@ -3818,9 +3814,7 @@ Every file or directory is a diskfs @dfn{node}. The following functions
+ help your diskfs callbacks manage nodes and their references:
+
+ @deftypefun void diskfs_drop_node (@w{struct node *@var{np}})
+-Node @var{np} now has no more references; clean all state. The
+-@var{diskfs_node_refcnt_lock} must be held, and will be released upon
+-return. @var{np} must be locked.
++Node @var{np} now has no more references; clean all state.
+ @end deftypefun
+
+ @deftypefun void diskfs_node_update (@w{struct node *@var{np}}, @w{int @var{wait}})
+@@ -4236,14 +4230,13 @@ without real users.
+ @deftypefun void diskfs_try_dropping_softrefs (@w{struct node *@var{np}})
+ Node @var{np} has some light references, but has just lost its last hard
+ references. Take steps so that if any light references can be freed,
+-they are. Both @var{diskfs_node_refcnt_lock} and @var{np} are locked.
++they are. @var{np} is locked.
+ This function will be called after @code{diskfs_lost_hardrefs}.
+ @end deftypefun
+
+ @deftypefun void diskfs_node_norefs (@w{struct node *@var{np}})
+ Node @var{np} has no more references; free local state, including
+ @code{*@var{np}} if it shouldn't be retained.
+-@var{diskfs_node_refcnt_lock} is held.
+ @end deftypefun
+
+ @deftypefun error_t diskfs_set_hypermetadata (@w{int @var{wait}}, @w{int @var{clean}})
+diff --git a/ext2fs/inode.c b/ext2fs/inode.c
+index 7c8d5a8..2a0c3cf 100644
+--- a/ext2fs/inode.c
++++ b/ext2fs/inode.c
+@@ -57,7 +57,6 @@
+ through the nodehash. */
+ static struct node *nodehash[INOHSZ];
+ static size_t nodehash_nr_items;
+-/* nodecache_lock must be acquired before diskfs_node_refcnt_lock. */
+ static pthread_rwlock_t nodecache_lock = PTHREAD_RWLOCK_INITIALIZER;
+
+ static error_t read_node (struct node *np);
+@@ -207,14 +206,10 @@ diskfs_try_dropping_softrefs (struct node *np)
+ {
+ /* Check if someone reacquired a reference through the
+ nodehash. */
+- unsigned int references;
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- references = np->references;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+-
+- /* An additional reference is acquired by libdiskfs across calls
+- to diskfs_try_dropping_softrefs. */
+- if (references > 1)
++ struct references result;
++ refcounts_references (&np->refcounts, &result);
++
++ if (result.hard > 0)
+ {
+ /* A reference was reacquired through a hash table lookup.
+ It's fine, we didn't touch anything yet. */
+@@ -636,7 +631,7 @@ diskfs_node_iterate (error_t (*fun)(struct node *))
+ /* We acquire a hard reference for node, but without using
+ diskfs_nref. We do this so that diskfs_new_hardrefs will not
+ get called. */
+- node->references++;
++ refcounts_ref (&node->refcounts, NULL);
+ }
+
+ pthread_rwlock_unlock (&nodecache_lock);
+diff --git a/fatfs/inode.c b/fatfs/inode.c
+index 1d670f5..f228618 100644
+--- a/fatfs/inode.c
++++ b/fatfs/inode.c
+@@ -55,7 +55,6 @@
+ through the nodehash. */
+ static struct node *nodehash[INOHSZ];
+ static size_t nodehash_nr_items;
+-/* nodecache_lock must be acquired before diskfs_node_refcnt_lock. */
+ static pthread_rwlock_t nodecache_lock = PTHREAD_RWLOCK_INITIALIZER;
+
+ static error_t read_node (struct node *np, vm_address_t buf);
+@@ -254,14 +253,8 @@ diskfs_node_norefs (struct node *np)
+ if (np->dn->translator)
+ free (np->dn->translator);
+
+- /* It is safe to unlock diskfs_node_refcnt_lock here for a while because
+- all references to the node have been deleted. */
+ if (np->dn->dirnode)
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- diskfs_nrele (np->dn->dirnode);
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- }
++ diskfs_nrele (np->dn->dirnode);
+
+ assert (!np->dn->pager);
+
+@@ -279,14 +272,10 @@ diskfs_try_dropping_softrefs (struct node *np)
+ {
+ /* Check if someone reacquired a reference through the
+ nodehash. */
+- unsigned int references;
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- references = np->references;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+-
+- /* An additional reference is acquired by libdiskfs across calls
+- to diskfs_try_dropping_softrefs. */
+- if (references > 1)
++ struct references result;
++ refcounts_references (&np->refcounts, &result);
++
++ if (result.hard > 0)
+ {
+ /* A reference was reacquired through a hash table lookup.
+ It's fine, we didn't touch anything yet. */
+@@ -392,7 +381,7 @@ read_node (struct node *np, vm_address_t buf)
+ /* Files in fatfs depend on the directory that hold the file. */
+ np->dn->dirnode = dp;
+ if (dp)
+- dp->references++;
++ refcounts_ref (&dp->refcounts, NULL);
+
+ pthread_rwlock_rdlock (&np->dn->dirent_lock);
+
+@@ -627,7 +616,7 @@ diskfs_node_iterate (error_t (*fun)(struct node *))
+ /* We acquire a hard reference for node, but without using
+ diskfs_nref. We do this so that diskfs_new_hardrefs will not
+ get called. */
+- node->references++;
++ refcounts_ref (&node->refcounts, NULL);
+ }
+
+ pthread_rwlock_unlock (&nodecache_lock);
+@@ -838,7 +827,7 @@ diskfs_alloc_node (struct node *dir, mode_t mode, struct node **node)
+
+ /* FIXME: We know that readnode couldn't put this in. */
+ np->dn->dirnode = dir;
+- dir->references++;
++ refcounts_ref (&dir->refcounts, NULL);
+
+ *node = np;
+ return 0;
+diff --git a/isofs/inode.c b/isofs/inode.c
+index 37bf1ac..340bc9c 100644
+--- a/isofs/inode.c
++++ b/isofs/inode.c
+@@ -60,7 +60,6 @@ struct node_cache
+ static int node_cache_size = 0;
+ static int node_cache_alloced = 0;
+ struct node_cache *node_cache = 0;
+-/* nodecache_lock must be acquired before diskfs_node_refcnt_lock. */
+ static pthread_rwlock_t nodecache_lock = PTHREAD_RWLOCK_INITIALIZER;
+
+ /* Forward */
+@@ -564,14 +563,10 @@ diskfs_try_dropping_softrefs (struct node *np)
+
+ /* Check if someone reacquired a reference through the
+ node_cache. */
+- unsigned int references;
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- references = np->references;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+-
+- /* An additional reference is acquired by libdiskfs across calls
+- to diskfs_try_dropping_softrefs. */
+- if (references > 1)
++ struct references result;
++ refcounts_references (&np->refcounts, &result);
++
++ if (result.hard > 0)
+ {
+ /* A reference was reacquired through a hash table lookup.
+ It's fine, we didn't touch anything yet. */
+diff --git a/libdiskfs/diskfs.h b/libdiskfs/diskfs.h
+index 2818225..535fb39 100644
+--- a/libdiskfs/diskfs.h
++++ b/libdiskfs/diskfs.h
+@@ -96,8 +96,7 @@ struct node
+
+ pthread_mutex_t lock;
+
+- int references; /* hard references */
+- int light_references; /* light references */
++ refcounts_t refcounts;
+
+ mach_port_t sockaddr; /* address for S_IFSOCK shortcut */
+
+@@ -198,8 +197,6 @@ extern volatile struct mapped_time_value *diskfs_mtime;
+ be done by format independent code. */
+ extern int diskfs_synchronous;
+
+-extern pthread_spinlock_t diskfs_node_refcnt_lock;
+-
+ extern int pager_port_type;
+
+ /* Whether the filesystem is currently writable or not. */
+@@ -448,14 +445,15 @@ error_t diskfs_alloc_node (struct node *dp, mode_t mode, struct node **np);
+ void diskfs_free_node (struct node *np, mode_t mode);
+
+ /* Node NP has no more references; free local state, including *NP
+- if it isn't to be retained. diskfs_node_refcnt_lock is held. */
++ if it isn't to be retained. */
+ void diskfs_node_norefs (struct node *np);
+
+ /* The user must define this function. Node NP has some light
+ references, but has just lost its last hard references. Take steps
+ so that if any light references can be freed, they are. NP is locked
+ as is the pager refcount lock. This function will be called after
+- diskfs_lost_hardrefs. */
++ diskfs_lost_hardrefs. An additional light reference is acquired by
++ libdiskfs across calls to this function. */
+ void diskfs_try_dropping_softrefs (struct node *np);
+
+ /* The user must define this funcction. Node NP has some light
+@@ -611,9 +609,8 @@ void diskfs_spawn_first_thread (ports_demuxer_type demuxer);
+ diskfs_init_completed once it has a valid proc and auth port. */
+ void diskfs_start_bootstrap ();
+
+-/* Node NP now has no more references; clean all state. The
+- _diskfs_node_refcnt_lock must be held, and will be released
+- upon return. NP must be locked. */
++/* Node NP now has no more references; clean all state. NP must be
++ locked. */
+ void diskfs_drop_node (struct node *np);
+
+ /* Set on disk fields from NP->dn_stat; update ctime, atime, and mtime
+diff --git a/libdiskfs/init-init.c b/libdiskfs/init-init.c
+index ffb99e0..357960b 100644
+--- a/libdiskfs/init-init.c
++++ b/libdiskfs/init-init.c
+@@ -41,8 +41,6 @@ int _diskfs_noatime;
+
+ struct hurd_port _diskfs_exec_portcell;
+
+-pthread_spinlock_t diskfs_node_refcnt_lock = PTHREAD_SPINLOCK_INITIALIZER;
+-
+ pthread_spinlock_t _diskfs_control_lock = PTHREAD_SPINLOCK_INITIALIZER;
+ int _diskfs_ncontrol_ports;
+
+diff --git a/libdiskfs/node-drop.c b/libdiskfs/node-drop.c
+index 83eb590..455031b 100644
+--- a/libdiskfs/node-drop.c
++++ b/libdiskfs/node-drop.c
+@@ -31,9 +31,8 @@ free_modreqs (struct modreq *mr)
+ }
+
+
+-/* Node NP now has no more references; clean all state. The
+- diskfs_node_refcnt_lock must be held, and will be released
+- upon return. NP must be locked. */
++/* Node NP now has no more references; clean all state. NP must be
++ locked. */
+ void
+ diskfs_drop_node (struct node *np)
+ {
+@@ -60,8 +59,7 @@ diskfs_drop_node (struct node *np)
+ and an nput. The next time through, this routine
+ will notice that the size is zero, and not have to
+ do anything. */
+- np->references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ refcounts_unsafe_ref (&np->refcounts, NULL);
+ diskfs_truncate (np, 0);
+
+ /* Force allocsize to zero; if truncate consistently fails this
+@@ -93,6 +91,7 @@ diskfs_drop_node (struct node *np)
+
+ assert (!np->sockaddr);
+
++ pthread_mutex_unlock(&np->lock);
++ pthread_mutex_destroy(&np->lock);
+ diskfs_node_norefs (np);
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+ }
+diff --git a/libdiskfs/node-make.c b/libdiskfs/node-make.c
+index ff0cc0d..c7ca3b0 100644
+--- a/libdiskfs/node-make.c
++++ b/libdiskfs/node-make.c
+@@ -29,8 +29,7 @@ init_node (struct node *np, struct disknode *dn)
+ np->dn_stat_dirty = 0;
+
+ pthread_mutex_init (&np->lock, NULL);
+- np->references = 1;
+- np->light_references = 0;
++ refcounts_init (&np->refcounts, 1, 0);
+ np->owner = 0;
+ np->sockaddr = MACH_PORT_NULL;
+
+diff --git a/libdiskfs/node-nput.c b/libdiskfs/node-nput.c
+index 5043ad1..d23c103 100644
+--- a/libdiskfs/node-nput.c
++++ b/libdiskfs/node-nput.c
+@@ -26,56 +26,44 @@
+ void
+ diskfs_nput (struct node *np)
+ {
+- int tried_drop_softrefs = 0;
++ struct references result;
+
+- loop:
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- assert (np->references);
+- np->references--;
+- if (np->references + np->light_references == 0)
+- diskfs_drop_node (np);
+- else if (np->references == 0 && !tried_drop_softrefs)
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ /* While we call the diskfs_try_dropping_softrefs, we need to hold
++ one reference. We use a weak reference for this purpose, which
++ we acquire by demoting our hard reference to a weak one. */
++ refcounts_demote (&np->refcounts, &result);
+
++ if (result.hard == 0)
++ {
+ /* This is our cue that something akin to "last process closes file"
+ in the POSIX.1 sense happened, so make sure any pending node time
+ updates now happen in a timely fashion. */
+ diskfs_set_node_times (np);
+-
+ diskfs_lost_hardrefs (np);
+ if (!np->dn_stat.st_nlink)
+ {
+- /* There are no links. If there are soft references that
+- can be dropped, we can't let them postpone deallocation.
+- So attempt to drop them. But that's a user-supplied
+- routine, which might result in further recursive calls to
+- the ref-counting system. So we have to reacquire our
+- reference around the call to forestall disaster. */
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- np->references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+-
+ if (np->sockaddr != MACH_PORT_NULL)
+ {
+ mach_port_deallocate (mach_task_self (), np->sockaddr);
+ np->sockaddr = MACH_PORT_NULL;
+ }
+
++ /* There are no links. If there are soft references that
++ can be dropped, we can't let them postpone deallocation.
++ So attempt to drop them. But that's a user-supplied
++ routine, which might result in further recursive calls to
++ the ref-counting system. This is not a problem, as we
++ hold a weak reference ourselves. */
+ diskfs_try_dropping_softrefs (np);
+-
+- /* But there's no value in looping forever in this
+- routine; only try to drop soft refs once. */
+- tried_drop_softrefs = 1;
+-
+- /* Now we can drop the reference back... */
+- goto loop;
+ }
+ pthread_mutex_unlock (&np->lock);
+ }
++
++ /* Finally get rid of our reference. */
++ refcounts_deref_weak (&np->refcounts, &result);
++
++ if (result.hard == 0 && result.weak == 0)
++ diskfs_drop_node (np);
+ else
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- pthread_mutex_unlock (&np->lock);
+- }
++ pthread_mutex_unlock (&np->lock);
+ }
+diff --git a/libdiskfs/node-nputl.c b/libdiskfs/node-nputl.c
+index 1959665..8dac16e 100644
+--- a/libdiskfs/node-nputl.c
++++ b/libdiskfs/node-nputl.c
+@@ -25,14 +25,10 @@
+ void
+ diskfs_nput_light (struct node *np)
+ {
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- assert (np->light_references);
+- np->light_references--;
+- if (np->references + np->light_references == 0)
++ struct references result;
++ refcounts_deref_weak (&np->refcounts, &result);
++ if (result.hard == 0 && result.weak == 0)
+ diskfs_drop_node (np);
+ else
+- {
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- pthread_mutex_unlock (&np->lock);
+- }
++ pthread_mutex_unlock (&np->lock);
+ }
+diff --git a/libdiskfs/node-nref.c b/libdiskfs/node-nref.c
+index 13cea05..766a69c 100644
+--- a/libdiskfs/node-nref.c
++++ b/libdiskfs/node-nref.c
+@@ -26,12 +26,9 @@
+ void
+ diskfs_nref (struct node *np)
+ {
+- int new_hardref;
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- np->references++;
+- new_hardref = (np->references == 1);
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+- if (new_hardref)
++ struct references result;
++ refcounts_ref (&np->refcounts, &result);
++ if (result.hard == 1)
+ {
+ pthread_mutex_lock (&np->lock);
+ diskfs_new_hardrefs (np);
+diff --git a/libdiskfs/node-nrefl.c b/libdiskfs/node-nrefl.c
+index 9692247..f7a823d 100644
+--- a/libdiskfs/node-nrefl.c
++++ b/libdiskfs/node-nrefl.c
+@@ -24,7 +24,5 @@
+ void
+ diskfs_nref_light (struct node *np)
+ {
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- np->light_references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++ refcounts_ref_weak (&np->refcounts, NULL);
+ }
+diff --git a/libdiskfs/node-nrele.c b/libdiskfs/node-nrele.c
+index cc68089..d962846 100644
+--- a/libdiskfs/node-nrele.c
++++ b/libdiskfs/node-nrele.c
+@@ -28,38 +28,36 @@
+ void
+ diskfs_nrele (struct node *np)
+ {
+- int tried_drop_softrefs = 0;
++ struct references result;
+
+- loop:
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- assert (np->references);
+- np->references--;
+- if (np->references + np->light_references == 0)
+- {
+- pthread_mutex_lock (&np->lock);
+- diskfs_drop_node (np);
+- }
+- else if (np->references == 0)
++ /* While we call the diskfs_try_dropping_softrefs, we need to hold
++ one reference. We use a weak reference for this purpose, which
++ we acquire by demoting our hard reference to a weak one. */
++ refcounts_demote (&np->refcounts, &result);
++
++ if (result.hard == 0)
+ {
+ pthread_mutex_lock (&np->lock);
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+ diskfs_lost_hardrefs (np);
+- if (!np->dn_stat.st_nlink && !tried_drop_softrefs)
++ if (!np->dn_stat.st_nlink)
+ {
+- /* Same issue here as in nput; see that for explanation */
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- np->references++;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+-
++ /* There are no links. If there are soft references that
++ can be dropped, we can't let them postpone deallocation.
++ So attempt to drop them. But that's a user-supplied
++ routine, which might result in further recursive calls to
++ the ref-counting system. This is not a problem, as we
++ hold a weak reference ourselves. */
+ diskfs_try_dropping_softrefs (np);
+- tried_drop_softrefs = 1;
+-
+- /* Now we can drop the reference back... */
+- pthread_mutex_unlock (&np->lock);
+- goto loop;
+ }
+ pthread_mutex_unlock (&np->lock);
+ }
+- else
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
++
++ /* Finally get rid of our reference. */
++ refcounts_deref_weak (&np->refcounts, &result);
++
++ if (result.hard == 0 && result.weak == 0)
++ {
++ pthread_mutex_lock (&np->lock);
++ diskfs_drop_node (np);
++ }
+ }
+diff --git a/libdiskfs/node-nrelel.c b/libdiskfs/node-nrelel.c
+index ee53b22..dc4f920 100644
+--- a/libdiskfs/node-nrelel.c
++++ b/libdiskfs/node-nrelel.c
+@@ -26,14 +26,11 @@
+ void
+ diskfs_nrele_light (struct node *np)
+ {
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- assert (np->light_references);
+- np->light_references--;
+- if (np->references + np->light_references == 0)
++ struct references result;
++ refcounts_deref_weak (&np->refcounts, &result);
++ if (result.hard == 0 && result.weak == 0)
+ {
+ pthread_mutex_lock (&np->lock);
+ diskfs_drop_node (np);
+ }
+- else
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+ }
+diff --git a/tmpfs/node.c b/tmpfs/node.c
+index 428b6d9..2a4489c 100644
+--- a/tmpfs/node.c
++++ b/tmpfs/node.c
+@@ -40,7 +40,6 @@ static unsigned int gen;
+ reference. */
+ struct node *all_nodes;
+ static size_t all_nodes_nr_items;
+-/* all_nodes_lock must be acquired before diskfs_node_refcnt_lock. */
+ pthread_rwlock_t all_nodes_lock = PTHREAD_RWLOCK_INITIALIZER;
+
+ error_t
+@@ -267,7 +266,7 @@ diskfs_node_iterate (error_t (*fun) (struct node *))
+ /* We acquire a hard reference for node, but without using
+ diskfs_nref. We do this so that diskfs_new_hardrefs will not
+ get called. */
+- node->references++;
++ refcounts_ref (&node->refcounts, NULL);
+ }
+
+ pthread_rwlock_unlock (&all_nodes_lock);
+@@ -300,14 +299,10 @@ diskfs_try_dropping_softrefs (struct node *np)
+ if (np->cache_id != 0)
+ {
+ /* Check if someone reacquired a reference. */
+- unsigned int references;
+- pthread_spin_lock (&diskfs_node_refcnt_lock);
+- references = np->references;
+- pthread_spin_unlock (&diskfs_node_refcnt_lock);
+-
+- /* An additional reference is acquired by libdiskfs across calls
+- to diskfs_try_dropping_softrefs. */
+- if (references > 1)
++ struct references result;
++ refcounts_references (&np->refcounts, &result);
++
++ if (result.hard > 0)
+ {
+ /* A reference was reacquired. It's fine, we didn't touch
+ anything yet. */
+--
+2.1.4
+
diff --git a/debian/patches/series b/debian/patches/series
index 79656c5d..e644f762 100644
--- a/debian/patches/series
+++ b/debian/patches/series
@@ -45,3 +45,8 @@ merge-me-0002-startup-faster-reboots.patch
thomas_term.patch
ajoin.patch
proc_disable_new_task_notifications.patch
+0001-ext2fs-use-a-seperate-lock-to-protect-nodehash.patch
+0002-fatfs-use-a-seperate-lock-to-protect-nodehash.patch
+0003-isofs-use-a-seperate-lock-to-protect-node_cache.patch
+0004-tmpfs-use-a-seperate-lock-to-protect-all_nodes.patch
+0005-libdiskfs-lock-less-reference-counting-of-nodes.patch