[TCP]: high_seq parameter removed (all callers use tp->high_seq)
[pandora-kernel.git] / fs / dcache.c
index b5f6139..678d39d 100644 (file)
@@ -21,7 +21,6 @@
 #include <linux/fsnotify.h>
 #include <linux/slab.h>
 #include <linux/init.h>
-#include <linux/smp_lock.h>
 #include <linux/hash.h>
 #include <linux/cache.h>
 #include <linux/module.h>
@@ -121,6 +120,28 @@ static void dentry_iput(struct dentry * dentry)
        }
 }
 
+/**
+ * d_kill - kill dentry and return parent
+ * @dentry: dentry to kill
+ *
+ * Called with dcache_lock and d_lock, releases both.  The dentry must
+ * already be unhashed and removed from the LRU.
+ *
+ * If this is the root of the dentry tree, return NULL.
+ */
+static struct dentry *d_kill(struct dentry *dentry)
+{
+       struct dentry *parent;
+
+       list_del(&dentry->d_u.d_child);
+       dentry_stat.nr_dentry--;        /* For d_free, below */
+       /*drops the locks, at that point nobody can reach this dentry */
+       dentry_iput(dentry);
+       parent = dentry->d_parent;
+       d_free(dentry);
+       return dentry == parent ? NULL : parent;
+}
+
 /* 
  * This is dput
  *
@@ -189,28 +210,17 @@ repeat:
 
 unhash_it:
        __d_drop(dentry);
-
-kill_it: {
-               struct dentry *parent;
-
-               /* If dentry was on d_lru list
-                * delete it from there
-                */
-               if (!list_empty(&dentry->d_lru)) {
-                       list_del(&dentry->d_lru);
-                       dentry_stat.nr_unused--;
-               }
-               list_del(&dentry->d_u.d_child);
-               dentry_stat.nr_dentry--;        /* For d_free, below */
-               /*drops the locks, at that point nobody can reach this dentry */
-               dentry_iput(dentry);
-               parent = dentry->d_parent;
-               d_free(dentry);
-               if (dentry == parent)
-                       return;
-               dentry = parent;
-               goto repeat;
+kill_it:
+       /* If dentry was on d_lru list
+        * delete it from there
+        */
+       if (!list_empty(&dentry->d_lru)) {
+               list_del(&dentry->d_lru);
+               dentry_stat.nr_unused--;
        }
+       dentry = d_kill(dentry);
+       if (dentry)
+               goto repeat;
 }
 
 /**
@@ -371,22 +381,40 @@ restart:
  * Throw away a dentry - free the inode, dput the parent.  This requires that
  * the LRU list has already been removed.
  *
+ * If prune_parents is true, try to prune ancestors as well.
+ *
  * Called with dcache_lock, drops it and then regains.
  * Called with dentry->d_lock held, drops it.
  */
-static void prune_one_dentry(struct dentry * dentry)
+static void prune_one_dentry(struct dentry * dentry, int prune_parents)
 {
-       struct dentry * parent;
-
        __d_drop(dentry);
-       list_del(&dentry->d_u.d_child);
-       dentry_stat.nr_dentry--;        /* For d_free, below */
-       dentry_iput(dentry);
-       parent = dentry->d_parent;
-       d_free(dentry);
-       if (parent != dentry)
-               dput(parent);
+       dentry = d_kill(dentry);
+       if (!prune_parents) {
+               dput(dentry);
+               spin_lock(&dcache_lock);
+               return;
+       }
+
+       /*
+        * Prune ancestors.  Locking is simpler than in dput(),
+        * because dcache_lock needs to be taken anyway.
+        */
        spin_lock(&dcache_lock);
+       while (dentry) {
+               if (!atomic_dec_and_lock(&dentry->d_count, &dentry->d_lock))
+                       return;
+
+               if (dentry->d_op && dentry->d_op->d_delete)
+                       dentry->d_op->d_delete(dentry);
+               if (!list_empty(&dentry->d_lru)) {
+                       list_del(&dentry->d_lru);
+                       dentry_stat.nr_unused--;
+               }
+               __d_drop(dentry);
+               dentry = d_kill(dentry);
+               spin_lock(&dcache_lock);
+       }
 }
 
 /**
@@ -394,6 +422,7 @@ static void prune_one_dentry(struct dentry * dentry)
  * @count: number of entries to try and free
  * @sb: if given, ignore dentries for other superblocks
  *         which are being unmounted.
+ * @prune_parents: if true, try to prune ancestors as well in one go
  *
  * Shrink the dcache. This is done when we need
  * more memory, or simply when we need to unmount
@@ -404,7 +433,7 @@ static void prune_one_dentry(struct dentry * dentry)
  * all the dentries are in use.
  */
  
-static void prune_dcache(int count, struct super_block *sb)
+static void prune_dcache(int count, struct super_block *sb, int prune_parents)
 {
        spin_lock(&dcache_lock);
        for (; count ; count--) {
@@ -464,7 +493,7 @@ static void prune_dcache(int count, struct super_block *sb)
                 * without taking the s_umount lock (I already hold it).
                 */
                if (sb && dentry->d_sb == sb) {
-                       prune_one_dentry(dentry);
+                       prune_one_dentry(dentry, prune_parents);
                        continue;
                }
                /*
@@ -479,7 +508,7 @@ static void prune_dcache(int count, struct super_block *sb)
                s_umount = &dentry->d_sb->s_umount;
                if (down_read_trylock(s_umount)) {
                        if (dentry->d_sb->s_root != NULL) {
-                               prune_one_dentry(dentry);
+                               prune_one_dentry(dentry, prune_parents);
                                up_read(s_umount);
                                continue;
                        }
@@ -550,7 +579,7 @@ repeat:
                        spin_unlock(&dentry->d_lock);
                        continue;
                }
-               prune_one_dentry(dentry);
+               prune_one_dentry(dentry, 1);
                cond_resched_lock(&dcache_lock);
                goto repeat;
        }
@@ -829,7 +858,7 @@ void shrink_dcache_parent(struct dentry * parent)
        int found;
 
        while ((found = select_parent(parent)) != 0)
-               prune_dcache(found, parent->d_sb);
+               prune_dcache(found, parent->d_sb, 1);
 }
 
 /*
@@ -849,11 +878,16 @@ static int shrink_dcache_memory(int nr, gfp_t gfp_mask)
        if (nr) {
                if (!(gfp_mask & __GFP_FS))
                        return -1;
-               prune_dcache(nr, NULL);
+               prune_dcache(nr, NULL, 1);
        }
        return (dentry_stat.nr_unused / 100) * sysctl_vfs_cache_pressure;
 }
 
+static struct shrinker dcache_shrinker = {
+       .shrink = shrink_dcache_memory,
+       .seeks = DEFAULT_SEEKS,
+};
+
 /**
  * d_alloc     -       allocate a dcache entry
  * @parent: parent of entry to allocate
@@ -1739,41 +1773,45 @@ shouldnt_be_hashed:
  * @rootmnt: vfsmnt to which the root dentry belongs
  * @buffer: buffer to return value in
  * @buflen: buffer length
- * @fail_deleted: what to return for deleted files
  *
- * Convert a dentry into an ASCII path name. If the entry has been deleted,
- * then if @fail_deleted is true, ERR_PTR(-ENOENT) is returned. Otherwise,
- * the the string " (deleted)" is appended. Note that this is ambiguous.
+ * Convert a dentry into an ASCII path name. If the entry has been deleted
+ * the string " (deleted)" is appended. Note that this is ambiguous.
+ *
+ * Returns the buffer or an error code if the path was too long.
  *
- * Returns the buffer or an error code.
+ * "buflen" should be positive. Caller holds the dcache_lock.
  */
-static char *__d_path(struct dentry *dentry, struct vfsmount *vfsmnt,
-                     struct dentry *root, struct vfsmount *rootmnt,
-                     char *buffer, int buflen, int fail_deleted)
+static char * __d_path( struct dentry *dentry, struct vfsmount *vfsmnt,
+                       struct dentry *root, struct vfsmount *rootmnt,
+                       char *buffer, int buflen)
 {
-       int namelen, is_slash;
+       char * end = buffer+buflen;
+       char * retval;
+       int namelen;
 
-       if (buflen < 2)
-               return ERR_PTR(-ENAMETOOLONG);
-       buffer += --buflen;
-       *buffer = '\0';
-
-       spin_lock(&dcache_lock);
+       *--end = '\0';
+       buflen--;
        if (!IS_ROOT(dentry) && d_unhashed(dentry)) {
-               if (fail_deleted) {
-                       buffer = ERR_PTR(-ENOENT);
-                       goto out;
-               }
-               if (buflen < 10)
-                       goto Elong;
                buflen -= 10;
-               buffer -= 10;
-               memcpy(buffer, " (deleted)", 10);
+               end -= 10;
+               if (buflen < 0)
+                       goto Elong;
+               memcpy(end, " (deleted)", 10);
        }
-       while (dentry != root || vfsmnt != rootmnt) {
+
+       if (buflen < 1)
+               goto Elong;
+       /* Get '/' right */
+       retval = end-1;
+       *retval = '/';
+
+       for (;;) {
                struct dentry * parent;
 
+               if (dentry == root && vfsmnt == rootmnt)
+                       break;
                if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) {
+                       /* Global root? */
                        spin_lock(&vfsmount_lock);
                        if (vfsmnt->mnt_parent == vfsmnt) {
                                spin_unlock(&vfsmount_lock);
@@ -1787,75 +1825,81 @@ static char *__d_path(struct dentry *dentry, struct vfsmount *vfsmnt,
                parent = dentry->d_parent;
                prefetch(parent);
                namelen = dentry->d_name.len;
-               if (buflen <= namelen)
-                       goto Elong;
                buflen -= namelen + 1;
-               buffer -= namelen;
-               memcpy(buffer, dentry->d_name.name, namelen);
-               *--buffer = '/';
+               if (buflen < 0)
+                       goto Elong;
+               end -= namelen;
+               memcpy(end, dentry->d_name.name, namelen);
+               *--end = '/';
+               retval = end;
                dentry = parent;
        }
-       /* Get '/' right */
-       if (*buffer != '/')
-               *--buffer = '/';
 
-out:
-       spin_unlock(&dcache_lock);
-       return buffer;
+       return retval;
 
 global_root:
-       /*
-        * We went past the (vfsmount, dentry) we were looking for and have
-        * either hit a root dentry, a lazily unmounted dentry, an
-        * unconnected dentry, or the file is on a pseudo filesystem.
-        */
        namelen = dentry->d_name.len;
-       is_slash = (namelen == 1 && *dentry->d_name.name == '/');
-       if (is_slash || (dentry->d_sb->s_flags & MS_NOUSER)) {
-               /*
-                * Make sure we won't return a pathname starting with '/'.
-                *
-                * Historically, we also glue together the root dentry and
-                * remaining name for pseudo filesystems like pipefs, which
-                * have the MS_NOUSER flag set. This results in pathnames
-                * like "pipe:[439336]".
-                */
-               if (*buffer == '/') {
-                       buffer++;
-                       buflen++;
-               }
-               if (is_slash)
-                       goto out;
-       }
-       if (buflen < namelen)
+       buflen -= namelen;
+       if (buflen < 0)
                goto Elong;
-       buffer -= namelen;
-       memcpy(buffer, dentry->d_name.name, namelen);
-       goto out;
-
+       retval -= namelen-1;    /* hit the slash */
+       memcpy(retval, dentry->d_name.name, namelen);
+       return retval;
 Elong:
-       buffer = ERR_PTR(-ENAMETOOLONG);
-       goto out;
+       return ERR_PTR(-ENAMETOOLONG);
 }
 
 /* write full pathname into buffer and return start of pathname */
-char *d_path(struct dentry *dentry, struct vfsmount *vfsmnt, char *buf,
-            int buflen)
+char * d_path(struct dentry *dentry, struct vfsmount *vfsmnt,
+                               char *buf, int buflen)
 {
        char *res;
        struct vfsmount *rootmnt;
        struct dentry *root;
 
+       /*
+        * We have various synthetic filesystems that never get mounted.  On
+        * these filesystems dentries are never used for lookup purposes, and
+        * thus don't need to be hashed.  They also don't need a name until a
+        * user wants to identify the object in /proc/pid/fd/.  The little hack
+        * below allows us to generate a name for these objects on demand:
+        */
+       if (dentry->d_op && dentry->d_op->d_dname)
+               return dentry->d_op->d_dname(dentry, buf, buflen);
+
        read_lock(&current->fs->lock);
        rootmnt = mntget(current->fs->rootmnt);
        root = dget(current->fs->root);
        read_unlock(&current->fs->lock);
-       res = __d_path(dentry, vfsmnt, root, rootmnt, buf, buflen, 0);
+       spin_lock(&dcache_lock);
+       res = __d_path(dentry, vfsmnt, root, rootmnt, buf, buflen);
+       spin_unlock(&dcache_lock);
        dput(root);
        mntput(rootmnt);
        return res;
 }
 
+/*
+ * Helper function for dentry_operations.d_dname() members
+ */
+char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen,
+                       const char *fmt, ...)
+{
+       va_list args;
+       char temp[64];
+       int sz;
+
+       va_start(args, fmt);
+       sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1;
+       va_end(args);
+
+       if (sz > sizeof(temp) || sz > buflen)
+               return ERR_PTR(-ENAMETOOLONG);
+
+       buffer += buflen - sz;
+       return memcpy(buffer, temp, sz);
+}
+
 /*
  * NOTE! The user-level library version returns a
  * character pointer. The kernel system call just
@@ -1876,10 +1920,10 @@ char *d_path(struct dentry *dentry, struct vfsmount *vfsmnt, char *buf,
  */
 asmlinkage long sys_getcwd(char __user *buf, unsigned long size)
 {
-       int error, len;
+       int error;
        struct vfsmount *pwdmnt, *rootmnt;
        struct dentry *pwd, *root;
-       char *page = (char *) __get_free_page(GFP_USER), *cwd;
+       char *page = (char *) __get_free_page(GFP_USER);
 
        if (!page)
                return -ENOMEM;
@@ -1891,18 +1935,29 @@ asmlinkage long sys_getcwd(char __user *buf, unsigned long size)
        root = dget(current->fs->root);
        read_unlock(&current->fs->lock);
 
-       cwd = __d_path(pwd, pwdmnt, root, rootmnt, page, PAGE_SIZE, 1);
-       error = PTR_ERR(cwd);
-       if (IS_ERR(cwd))
-               goto out;
+       error = -ENOENT;
+       /* Has the current directory has been unlinked? */
+       spin_lock(&dcache_lock);
+       if (pwd->d_parent == pwd || !d_unhashed(pwd)) {
+               unsigned long len;
+               char * cwd;
 
-       error = -ERANGE;
-       len = PAGE_SIZE + page - cwd;
-       if (len <= size) {
-               error = len;
-               if (copy_to_user(buf, cwd, len))
-                       error = -EFAULT;
-       }
+               cwd = __d_path(pwd, pwdmnt, root, rootmnt, page, PAGE_SIZE);
+               spin_unlock(&dcache_lock);
+
+               error = PTR_ERR(cwd);
+               if (IS_ERR(cwd))
+                       goto out;
+
+               error = -ERANGE;
+               len = PAGE_SIZE + page - cwd;
+               if (len <= size) {
+                       error = len;
+                       if (copy_to_user(buf, cwd, len))
+                               error = -EFAULT;
+               }
+       } else
+               spin_unlock(&dcache_lock);
 
 out:
        dput(pwd);
@@ -2062,14 +2117,10 @@ static void __init dcache_init(unsigned long mempages)
         * but it is probably not worth it because of the cache nature
         * of the dcache. 
         */
-       dentry_cache = kmem_cache_create("dentry_cache",
-                                        sizeof(struct dentry),
-                                        0,
-                                        (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
-                                        SLAB_MEM_SPREAD),
-                                        NULL, NULL);
+       dentry_cache = KMEM_CACHE(dentry,
+               SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
        
-       set_shrinker(DEFAULT_SEEKS, shrink_dcache_memory);
+       register_shrinker(&dcache_shrinker);
 
        /* Hash may have been set up in dcache_init_early */
        if (!hashdist)
@@ -2114,10 +2165,10 @@ void __init vfs_caches_init(unsigned long mempages)
        mempages -= reserve;
 
        names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0,
-                       SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
+                       SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
 
        filp_cachep = kmem_cache_create("filp", sizeof(struct file), 0,
-                       SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
+                       SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
 
        dcache_init(mempages);
        inode_init(mempages);