[XFS] Fix race in xfs_write() between direct and buffered I/O with DMAPI
[safe/jmp/linux-2.6] / fs / dquot.c
index afa06a8..5e95261 100644 (file)
@@ -9,8 +9,6 @@
  * implementation is based on one of the several variants of the LINUX
  * inode-subsystem with added complexity of the diskquota system.
  * 
- * Version: $Id: dquot.c,v 6.3 1996/11/17 18:35:34 mvw Exp mvw $
- * 
  * Author:     Marco van Wieringen <mvw@planets.elm.net>
  *
  * Fixes:   Dmitry Gorodchanin <pgmdsg@ibi.com>, 11 Feb 96
@@ -69,7 +67,6 @@
 #include <linux/file.h>
 #include <linux/slab.h>
 #include <linux/sysctl.h>
-#include <linux/smp_lock.h>
 #include <linux/init.h>
 #include <linux/module.h>
 #include <linux/proc_fs.h>
 #include <linux/kmod.h>
 #include <linux/namei.h>
 #include <linux/buffer_head.h>
+#include <linux/capability.h>
 #include <linux/quotaops.h>
+#include <linux/writeback.h> /* for inode_lock, oddly enough.. */
+#ifdef CONFIG_QUOTA_NETLINK_INTERFACE
+#include <net/netlink.h>
+#include <net/genetlink.h>
+#endif
 
 #include <asm/uaccess.h>
 
  * operation is just reading pointers from inode (or not using them at all) the
  * read lock is enough. If pointers are altered function must hold write lock
  * (these locking rules also apply for S_NOQUOTA flag in the inode - note that
- * for altering the flag i_sem is also needed).  If operation is holding
+ * for altering the flag i_mutex is also needed).  If operation is holding
  * reference to dquot in other way (e.g. quotactl ops) it must be guarded by
- * dqonoff_sem.
+ * dqonoff_mutex.
  * This locking assures that:
  *   a) update/access to dquot pointers in inode is serialized
  *   b) everyone is guarded against invalidate_dquots()
  *
- * Each dquot has its dq_lock semaphore. Locked dquots might not be referenced
+ * Each dquot has its dq_lock mutex. Locked dquots might not be referenced
  * from inodes (dquot_alloc_space() and such don't check the dq_lock).
  * Currently dquot is locked only when it is being read to memory (or space for
  * it is being allocated) on the first dqget() and when it is being released on
  * spinlock to internal buffers before writing.
  *
  * Lock ordering (including related VFS locks) is the following:
- *   i_sem > dqonoff_sem > iprune_sem > journal_lock > dqptr_sem >
- *   > dquot->dq_lock > dqio_sem
- * i_sem on quota files is special (it's below dqio_sem)
+ *   i_mutex > dqonoff_sem > journal_lock > dqptr_sem > dquot->dq_lock >
+ *   dqio_mutex
+ * i_mutex on quota files is special (it's below dqio_mutex)
  */
 
 static DEFINE_SPINLOCK(dq_list_lock);
@@ -130,7 +133,7 @@ static struct quota_format_type *quota_formats;     /* List of registered formats */
 static struct quota_module_name module_names[] = INIT_QUOTA_MODULE_NAMES;
 
 /* SLAB cache for dquot structures */
-static kmem_cache_t *dquot_cachep;
+static struct kmem_cache *dquot_cachep;
 
 int register_quota_format(struct quota_format_type *fmt)
 {
@@ -249,7 +252,7 @@ static inline struct dquot *find_dquot(unsigned int hashent, struct super_block
 /* Add a dquot to the tail of the free list */
 static inline void put_dquot_last(struct dquot *dquot)
 {
-       list_add(&dquot->dq_free, free_dquots.prev);
+       list_add_tail(&dquot->dq_free, &free_dquots);
        dqstats.free_dquots++;
 }
 
@@ -265,7 +268,7 @@ static inline void put_inuse(struct dquot *dquot)
 {
        /* We add to the back of inuse list so we don't have to restart
         * when traversing this list and we block */
-       list_add(&dquot->dq_inuse, inuse_list.prev);
+       list_add_tail(&dquot->dq_inuse, &inuse_list);
        dqstats.allocated_dquots++;
 }
 
@@ -280,11 +283,19 @@ static inline void remove_inuse(struct dquot *dquot)
 
 static void wait_on_dquot(struct dquot *dquot)
 {
-       down(&dquot->dq_lock);
-       up(&dquot->dq_lock);
+       mutex_lock(&dquot->dq_lock);
+       mutex_unlock(&dquot->dq_lock);
+}
+
+static inline int dquot_dirty(struct dquot *dquot)
+{
+       return test_bit(DQ_MOD_B, &dquot->dq_flags);
 }
 
-#define mark_dquot_dirty(dquot) ((dquot)->dq_sb->dq_op->mark_dirty(dquot))
+static inline int mark_dquot_dirty(struct dquot *dquot)
+{
+       return dquot->dq_sb->dq_op->mark_dirty(dquot);
+}
 
 int dquot_mark_dquot_dirty(struct dquot *dquot)
 {
@@ -320,8 +331,8 @@ int dquot_acquire(struct dquot *dquot)
        int ret = 0, ret2 = 0;
        struct quota_info *dqopt = sb_dqopt(dquot->dq_sb);
 
-       down(&dquot->dq_lock);
-       down(&dqopt->dqio_sem);
+       mutex_lock(&dquot->dq_lock);
+       mutex_lock(&dqopt->dqio_mutex);
        if (!test_bit(DQ_READ_B, &dquot->dq_flags))
                ret = dqopt->ops[dquot->dq_type]->read_dqblk(dquot);
        if (ret < 0)
@@ -342,8 +353,8 @@ int dquot_acquire(struct dquot *dquot)
        }
        set_bit(DQ_ACTIVE_B, &dquot->dq_flags);
 out_iolock:
-       up(&dqopt->dqio_sem);
-       up(&dquot->dq_lock);
+       mutex_unlock(&dqopt->dqio_mutex);
+       mutex_unlock(&dquot->dq_lock);
        return ret;
 }
 
@@ -355,7 +366,7 @@ int dquot_commit(struct dquot *dquot)
        int ret = 0, ret2 = 0;
        struct quota_info *dqopt = sb_dqopt(dquot->dq_sb);
 
-       down(&dqopt->dqio_sem);
+       mutex_lock(&dqopt->dqio_mutex);
        spin_lock(&dq_list_lock);
        if (!clear_dquot_dirty(dquot)) {
                spin_unlock(&dq_list_lock);
@@ -372,7 +383,7 @@ int dquot_commit(struct dquot *dquot)
                        ret = ret2;
        }
 out_sem:
-       up(&dqopt->dqio_sem);
+       mutex_unlock(&dqopt->dqio_mutex);
        return ret;
 }
 
@@ -384,11 +395,11 @@ int dquot_release(struct dquot *dquot)
        int ret = 0, ret2 = 0;
        struct quota_info *dqopt = sb_dqopt(dquot->dq_sb);
 
-       down(&dquot->dq_lock);
+       mutex_lock(&dquot->dq_lock);
        /* Check whether we are not racing with some other dqget() */
        if (atomic_read(&dquot->dq_count) > 1)
                goto out_dqlock;
-       down(&dqopt->dqio_sem);
+       mutex_lock(&dqopt->dqio_mutex);
        if (dqopt->ops[dquot->dq_type]->release_dqblk) {
                ret = dqopt->ops[dquot->dq_type]->release_dqblk(dquot);
                /* Write the info */
@@ -398,31 +409,57 @@ int dquot_release(struct dquot *dquot)
                        ret = ret2;
        }
        clear_bit(DQ_ACTIVE_B, &dquot->dq_flags);
-       up(&dqopt->dqio_sem);
+       mutex_unlock(&dqopt->dqio_mutex);
 out_dqlock:
-       up(&dquot->dq_lock);
+       mutex_unlock(&dquot->dq_lock);
        return ret;
 }
 
 /* Invalidate all dquots on the list. Note that this function is called after
  * quota is disabled and pointers from inodes removed so there cannot be new
- * quota users. Also because we hold dqonoff_sem there can be no quota users
- * for this sb+type at all. */
+ * quota users. There can still be some users of quotas due to inodes being
+ * just deleted or pruned by prune_icache() (those are not attached to any
+ * list). We have to wait for such users.
+ */
 static void invalidate_dquots(struct super_block *sb, int type)
 {
        struct dquot *dquot, *tmp;
 
+restart:
        spin_lock(&dq_list_lock);
        list_for_each_entry_safe(dquot, tmp, &inuse_list, dq_inuse) {
                if (dquot->dq_sb != sb)
                        continue;
                if (dquot->dq_type != type)
                        continue;
-#ifdef __DQUOT_PARANOIA
-               if (atomic_read(&dquot->dq_count))
-                       BUG();
-#endif
-               /* Quota now has no users and it has been written on last dqput() */
+               /* Wait for dquot users */
+               if (atomic_read(&dquot->dq_count)) {
+                       DEFINE_WAIT(wait);
+
+                       atomic_inc(&dquot->dq_count);
+                       prepare_to_wait(&dquot->dq_wait_unused, &wait,
+                                       TASK_UNINTERRUPTIBLE);
+                       spin_unlock(&dq_list_lock);
+                       /* Once dqput() wakes us up, we know it's time to free
+                        * the dquot.
+                        * IMPORTANT: we rely on the fact that there is always
+                        * at most one process waiting for dquot to free.
+                        * Otherwise dq_count would be > 1 and we would never
+                        * wake up.
+                        */
+                       if (atomic_read(&dquot->dq_count) > 1)
+                               schedule();
+                       finish_wait(&dquot->dq_wait_unused, &wait);
+                       dqput(dquot);
+                       /* At this moment dquot() need not exist (it could be
+                        * reclaimed by prune_dqcache(). Hence we must
+                        * restart. */
+                       goto restart;
+               }
+               /*
+                * Quota now has no users and it has been written on last
+                * dqput()
+                */
                remove_dquot_hash(dquot);
                remove_free_dquot(dquot);
                remove_inuse(dquot);
@@ -438,7 +475,7 @@ int vfs_quota_sync(struct super_block *sb, int type)
        struct quota_info *dqopt = sb_dqopt(sb);
        int cnt;
 
-       down(&dqopt->dqonoff_sem);
+       mutex_lock(&dqopt->dqonoff_mutex);
        for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
                if (type != -1 && cnt != type)
                        continue;
@@ -447,7 +484,7 @@ int vfs_quota_sync(struct super_block *sb, int type)
                spin_lock(&dq_list_lock);
                dirty = &dqopt->info[cnt].dqi_dirty_list;
                while (!list_empty(dirty)) {
-                       dquot = list_entry(dirty->next, struct dquot, dq_dirty);
+                       dquot = list_first_entry(dirty, struct dquot, dq_dirty);
                        /* Dirty and inactive can be only bad dquot... */
                        if (!test_bit(DQ_ACTIVE_B, &dquot->dq_flags)) {
                                clear_dquot_dirty(dquot);
@@ -473,7 +510,7 @@ int vfs_quota_sync(struct super_block *sb, int type)
        spin_lock(&dq_list_lock);
        dqstats.syncs++;
        spin_unlock(&dq_list_lock);
-       up(&dqopt->dqonoff_sem);
+       mutex_unlock(&dqopt->dqonoff_mutex);
 
        return 0;
 }
@@ -511,13 +548,20 @@ static int shrink_dqcache_memory(int nr, gfp_t gfp_mask)
        return (dqstats.free_dquots / 100) * sysctl_vfs_cache_pressure;
 }
 
+static struct shrinker dqcache_shrinker = {
+       .shrink = shrink_dqcache_memory,
+       .seeks = DEFAULT_SEEKS,
+};
+
 /*
  * Put reference to dquot
  * NOTE: If you change this function please check whether dqput_blocks() works right...
- * MUST be called with either dqptr_sem or dqonoff_sem held
+ * MUST be called with either dqptr_sem or dqonoff_mutex held
  */
 static void dqput(struct dquot *dquot)
 {
+       int ret;
+
        if (!dquot)
                return;
 #ifdef __DQUOT_PARANOIA
@@ -539,6 +583,10 @@ we_slept:
        if (atomic_read(&dquot->dq_count) > 1) {
                /* We have more than one user... nothing to do */
                atomic_dec(&dquot->dq_count);
+               /* Releasing dquot during quotaoff phase? */
+               if (!sb_has_quota_enabled(dquot->dq_sb, dquot->dq_type) &&
+                   atomic_read(&dquot->dq_count) == 1)
+                       wake_up(&dquot->dq_wait_unused);
                spin_unlock(&dq_list_lock);
                return;
        }
@@ -546,7 +594,19 @@ we_slept:
        if (test_bit(DQ_ACTIVE_B, &dquot->dq_flags) && dquot_dirty(dquot)) {
                spin_unlock(&dq_list_lock);
                /* Commit dquot before releasing */
-               dquot->dq_sb->dq_op->write_dquot(dquot);
+               ret = dquot->dq_sb->dq_op->write_dquot(dquot);
+               if (ret < 0) {
+                       printk(KERN_ERR "VFS: cannot write quota structure on "
+                               "device %s (error %d). Quota may get out of "
+                               "sync!\n", dquot->dq_sb->s_id, ret);
+                       /*
+                        * We clear dirty bit anyway, so that we avoid
+                        * infinite loop here
+                        */
+                       spin_lock(&dq_list_lock);
+                       clear_dquot_dirty(dquot);
+                       spin_unlock(&dq_list_lock);
+               }
                goto we_slept;
        }
        /* Clear flag in case dquot was inactive (something bad happened) */
@@ -559,8 +619,7 @@ we_slept:
        atomic_dec(&dquot->dq_count);
 #ifdef __DQUOT_PARANOIA
        /* sanity check */
-       if (!list_empty(&dquot->dq_free))
-               BUG();
+       BUG_ON(!list_empty(&dquot->dq_free));
 #endif
        put_dquot_last(dquot);
        spin_unlock(&dq_list_lock);
@@ -570,16 +629,16 @@ static struct dquot *get_empty_dquot(struct super_block *sb, int type)
 {
        struct dquot *dquot;
 
-       dquot = kmem_cache_alloc(dquot_cachep, SLAB_NOFS);
+       dquot = kmem_cache_zalloc(dquot_cachep, GFP_NOFS);
        if(!dquot)
                return NODQUOT;
 
-       memset((caddr_t)dquot, 0, sizeof(struct dquot));
-       sema_init(&dquot->dq_lock, 1);
+       mutex_init(&dquot->dq_lock);
        INIT_LIST_HEAD(&dquot->dq_free);
        INIT_LIST_HEAD(&dquot->dq_inuse);
        INIT_HLIST_NODE(&dquot->dq_hash);
        INIT_LIST_HEAD(&dquot->dq_dirty);
+       init_waitqueue_head(&dquot->dq_wait_unused);
        dquot->dq_sb = sb;
        dquot->dq_type = type;
        atomic_set(&dquot->dq_count, 1);
@@ -589,7 +648,7 @@ static struct dquot *get_empty_dquot(struct super_block *sb, int type)
 
 /*
  * Get reference to dquot
- * MUST be called with either dqptr_sem or dqonoff_sem held
+ * MUST be called with either dqptr_sem or dqonoff_mutex held
  */
 static struct dquot *dqget(struct super_block *sb, unsigned int id, int type)
 {
@@ -634,8 +693,7 @@ we_slept:
                return NODQUOT;
        }
 #ifdef __DQUOT_PARANOIA
-       if (!dquot->dq_sb)      /* Has somebody invalidated entry under us? */
-               BUG();
+       BUG_ON(!dquot->dq_sb);  /* Has somebody invalidated entry under us? */
 #endif
 
        return dquot;
@@ -655,26 +713,35 @@ static int dqinit_needed(struct inode *inode, int type)
        return 0;
 }
 
-/* This routine is guarded by dqonoff_sem semaphore */
+/* This routine is guarded by dqonoff_mutex mutex */
 static void add_dquot_ref(struct super_block *sb, int type)
 {
-       struct list_head *p;
+       struct inode *inode, *old_inode = NULL;
 
-restart:
-       file_list_lock();
-       list_for_each(p, &sb->s_files) {
-               struct file *filp = list_entry(p, struct file, f_u.fu_list);
-               struct inode *inode = filp->f_dentry->d_inode;
-               if (filp->f_mode & FMODE_WRITE && dqinit_needed(inode, type)) {
-                       struct dentry *dentry = dget(filp->f_dentry);
-                       file_list_unlock();
-                       sb->dq_op->initialize(inode, type);
-                       dput(dentry);
-                       /* As we may have blocked we had better restart... */
-                       goto restart;
-               }
-       }
-       file_list_unlock();
+       spin_lock(&inode_lock);
+       list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
+               if (!atomic_read(&inode->i_writecount))
+                       continue;
+               if (!dqinit_needed(inode, type))
+                       continue;
+               if (inode->i_state & (I_FREEING|I_WILL_FREE))
+                       continue;
+
+               __iget(inode);
+               spin_unlock(&inode_lock);
+
+               iput(old_inode);
+               sb->dq_op->initialize(inode, type);
+               /* We hold a reference to 'inode' so it couldn't have been
+                * removed from s_inodes list while we dropped the inode_lock.
+                * We cannot iput the inode now as we can be holding the last
+                * reference and we cannot iput it under inode_lock. So we
+                * keep the reference and iput it later. */
+               old_inode = inode;
+               spin_lock(&inode_lock);
+       }
+       spin_unlock(&inode_lock);
+       iput(old_inode);
 }
 
 /* Return 0 if dqput() won't block (note that 1 doesn't necessarily mean blocking) */
@@ -687,7 +754,8 @@ static inline int dqput_blocks(struct dquot *dquot)
 
 /* Remove references to dquots from inode - add dquot to list for freeing if needed */
 /* We can't race with anybody because we hold dqptr_sem for writing... */
-int remove_inode_dquot_ref(struct inode *inode, int type, struct list_head *tofree_head)
+static int remove_inode_dquot_ref(struct inode *inode, int type,
+                                 struct list_head *tofree_head)
 {
        struct dquot *dquot = inode->i_dquot[type];
 
@@ -726,19 +794,30 @@ static void put_dquot_list(struct list_head *tofree_head)
        }
 }
 
+static void remove_dquot_ref(struct super_block *sb, int type,
+               struct list_head *tofree_head)
+{
+       struct inode *inode;
+
+       spin_lock(&inode_lock);
+       list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
+               if (!IS_NOQUOTA(inode))
+                       remove_inode_dquot_ref(inode, type, tofree_head);
+       }
+       spin_unlock(&inode_lock);
+}
+
 /* Gather all references from inodes and drop them */
 static void drop_dquot_ref(struct super_block *sb, int type)
 {
        LIST_HEAD(tofree_head);
 
-       /* We need to be guarded against prune_icache to reach all the
-        * inodes - otherwise some can be on the local list of prune_icache */
-       down(&iprune_sem);
-       down_write(&sb_dqopt(sb)->dqptr_sem);
-       remove_dquot_ref(sb, type, &tofree_head);
-       up_write(&sb_dqopt(sb)->dqptr_sem);
-       up(&iprune_sem);
-       put_dquot_list(&tofree_head);
+       if (sb->dq_op) {
+               down_write(&sb_dqopt(sb)->dqptr_sem);
+               remove_dquot_ref(sb, type, &tofree_head);
+               up_write(&sb_dqopt(sb)->dqptr_sem);
+               put_dquot_list(&tofree_head);
+       }
 }
 
 static inline void dquot_incr_inodes(struct dquot *dquot, unsigned long number)
@@ -773,6 +852,19 @@ static inline void dquot_decr_space(struct dquot *dquot, qsize_t number)
        clear_bit(DQ_BLKS_B, &dquot->dq_flags);
 }
 
+static int warning_issued(struct dquot *dquot, const int warntype)
+{
+       int flag = (warntype == QUOTA_NL_BHARDWARN ||
+               warntype == QUOTA_NL_BSOFTLONGWARN) ? DQ_BLKS_B :
+               ((warntype == QUOTA_NL_IHARDWARN ||
+               warntype == QUOTA_NL_ISOFTLONGWARN) ? DQ_INODES_B : 0);
+
+       if (!flag)
+               return 0;
+       return test_and_set_bit(flag, &dquot->dq_flags);
+}
+
+#ifdef CONFIG_PRINT_QUOTA_WARNING
 static int flag_print_warnings = 1;
 
 static inline int need_print_warning(struct dquot *dquot)
@@ -789,61 +881,137 @@ static inline int need_print_warning(struct dquot *dquot)
        return 0;
 }
 
-/* Values of warnings */
-#define NOWARN 0
-#define IHARDWARN 1
-#define ISOFTLONGWARN 2
-#define ISOFTWARN 3
-#define BHARDWARN 4
-#define BSOFTLONGWARN 5
-#define BSOFTWARN 6
-
 /* Print warning to user which exceeded quota */
-static void print_warning(struct dquot *dquot, const char warntype)
+static void print_warning(struct dquot *dquot, const int warntype)
 {
        char *msg = NULL;
-       int flag = (warntype == BHARDWARN || warntype == BSOFTLONGWARN) ? DQ_BLKS_B :
-         ((warntype == IHARDWARN || warntype == ISOFTLONGWARN) ? DQ_INODES_B : 0);
+       struct tty_struct *tty;
 
-       if (!need_print_warning(dquot) || (flag && test_and_set_bit(flag, &dquot->dq_flags)))
+       if (warntype == QUOTA_NL_IHARDBELOW ||
+           warntype == QUOTA_NL_ISOFTBELOW ||
+           warntype == QUOTA_NL_BHARDBELOW ||
+           warntype == QUOTA_NL_BSOFTBELOW || !need_print_warning(dquot))
                return;
 
-       tty_write_message(current->signal->tty, dquot->dq_sb->s_id);
-       if (warntype == ISOFTWARN || warntype == BSOFTWARN)
-               tty_write_message(current->signal->tty, ": warning, ");
+       tty = get_current_tty();
+       if (!tty)
+               return;
+       tty_write_message(tty, dquot->dq_sb->s_id);
+       if (warntype == QUOTA_NL_ISOFTWARN || warntype == QUOTA_NL_BSOFTWARN)
+               tty_write_message(tty, ": warning, ");
        else
-               tty_write_message(current->signal->tty, ": write failed, ");
-       tty_write_message(current->signal->tty, quotatypes[dquot->dq_type]);
+               tty_write_message(tty, ": write failed, ");
+       tty_write_message(tty, quotatypes[dquot->dq_type]);
        switch (warntype) {
-               case IHARDWARN:
+               case QUOTA_NL_IHARDWARN:
                        msg = " file limit reached.\r\n";
                        break;
-               case ISOFTLONGWARN:
+               case QUOTA_NL_ISOFTLONGWARN:
                        msg = " file quota exceeded too long.\r\n";
                        break;
-               case ISOFTWARN:
+               case QUOTA_NL_ISOFTWARN:
                        msg = " file quota exceeded.\r\n";
                        break;
-               case BHARDWARN:
+               case QUOTA_NL_BHARDWARN:
                        msg = " block limit reached.\r\n";
                        break;
-               case BSOFTLONGWARN:
+               case QUOTA_NL_BSOFTLONGWARN:
                        msg = " block quota exceeded too long.\r\n";
                        break;
-               case BSOFTWARN:
+               case QUOTA_NL_BSOFTWARN:
                        msg = " block quota exceeded.\r\n";
                        break;
        }
-       tty_write_message(current->signal->tty, msg);
+       tty_write_message(tty, msg);
+       tty_kref_put(tty);
 }
+#endif
 
-static inline void flush_warnings(struct dquot **dquots, char *warntype)
+#ifdef CONFIG_QUOTA_NETLINK_INTERFACE
+
+/* Netlink family structure for quota */
+static struct genl_family quota_genl_family = {
+       .id = GENL_ID_GENERATE,
+       .hdrsize = 0,
+       .name = "VFS_DQUOT",
+       .version = 1,
+       .maxattr = QUOTA_NL_A_MAX,
+};
+
+/* Send warning to userspace about user which exceeded quota */
+static void send_warning(const struct dquot *dquot, const char warntype)
+{
+       static atomic_t seq;
+       struct sk_buff *skb;
+       void *msg_head;
+       int ret;
+       int msg_size = 4 * nla_total_size(sizeof(u32)) +
+                      2 * nla_total_size(sizeof(u64));
+
+       /* We have to allocate using GFP_NOFS as we are called from a
+        * filesystem performing write and thus further recursion into
+        * the fs to free some data could cause deadlocks. */
+       skb = genlmsg_new(msg_size, GFP_NOFS);
+       if (!skb) {
+               printk(KERN_ERR
+                 "VFS: Not enough memory to send quota warning.\n");
+               return;
+       }
+       msg_head = genlmsg_put(skb, 0, atomic_add_return(1, &seq),
+                       &quota_genl_family, 0, QUOTA_NL_C_WARNING);
+       if (!msg_head) {
+               printk(KERN_ERR
+                 "VFS: Cannot store netlink header in quota warning.\n");
+               goto err_out;
+       }
+       ret = nla_put_u32(skb, QUOTA_NL_A_QTYPE, dquot->dq_type);
+       if (ret)
+               goto attr_err_out;
+       ret = nla_put_u64(skb, QUOTA_NL_A_EXCESS_ID, dquot->dq_id);
+       if (ret)
+               goto attr_err_out;
+       ret = nla_put_u32(skb, QUOTA_NL_A_WARNING, warntype);
+       if (ret)
+               goto attr_err_out;
+       ret = nla_put_u32(skb, QUOTA_NL_A_DEV_MAJOR,
+               MAJOR(dquot->dq_sb->s_dev));
+       if (ret)
+               goto attr_err_out;
+       ret = nla_put_u32(skb, QUOTA_NL_A_DEV_MINOR,
+               MINOR(dquot->dq_sb->s_dev));
+       if (ret)
+               goto attr_err_out;
+       ret = nla_put_u64(skb, QUOTA_NL_A_CAUSED_ID, current->user->uid);
+       if (ret)
+               goto attr_err_out;
+       genlmsg_end(skb, msg_head);
+
+       ret = genlmsg_multicast(skb, 0, quota_genl_family.id, GFP_NOFS);
+       if (ret < 0 && ret != -ESRCH)
+               printk(KERN_ERR
+                       "VFS: Failed to send notification message: %d\n", ret);
+       return;
+attr_err_out:
+       printk(KERN_ERR "VFS: Not enough space to compose quota message!\n");
+err_out:
+       kfree_skb(skb);
+}
+#endif
+
+static inline void flush_warnings(struct dquot * const *dquots, char *warntype)
 {
        int i;
 
        for (i = 0; i < MAXQUOTAS; i++)
-               if (dquots[i] != NODQUOT && warntype[i] != NOWARN)
+               if (dquots[i] != NODQUOT && warntype[i] != QUOTA_NL_NOWARN &&
+                   !warning_issued(dquots[i], warntype[i])) {
+#ifdef CONFIG_PRINT_QUOTA_WARNING
                        print_warning(dquots[i], warntype[i]);
+#endif
+#ifdef CONFIG_QUOTA_NETLINK_INTERFACE
+                       send_warning(dquots[i], warntype[i]);
+#endif
+               }
 }
 
 static inline char ignore_hardlimit(struct dquot *dquot)
@@ -857,14 +1025,14 @@ static inline char ignore_hardlimit(struct dquot *dquot)
 /* needs dq_data_lock */
 static int check_idq(struct dquot *dquot, ulong inodes, char *warntype)
 {
-       *warntype = NOWARN;
+       *warntype = QUOTA_NL_NOWARN;
        if (inodes <= 0 || test_bit(DQ_FAKE_B, &dquot->dq_flags))
                return QUOTA_OK;
 
        if (dquot->dq_dqb.dqb_ihardlimit &&
           (dquot->dq_dqb.dqb_curinodes + inodes) > dquot->dq_dqb.dqb_ihardlimit &&
             !ignore_hardlimit(dquot)) {
-               *warntype = IHARDWARN;
+               *warntype = QUOTA_NL_IHARDWARN;
                return NO_QUOTA;
        }
 
@@ -872,14 +1040,14 @@ static int check_idq(struct dquot *dquot, ulong inodes, char *warntype)
           (dquot->dq_dqb.dqb_curinodes + inodes) > dquot->dq_dqb.dqb_isoftlimit &&
            dquot->dq_dqb.dqb_itime && get_seconds() >= dquot->dq_dqb.dqb_itime &&
             !ignore_hardlimit(dquot)) {
-               *warntype = ISOFTLONGWARN;
+               *warntype = QUOTA_NL_ISOFTLONGWARN;
                return NO_QUOTA;
        }
 
        if (dquot->dq_dqb.dqb_isoftlimit &&
           (dquot->dq_dqb.dqb_curinodes + inodes) > dquot->dq_dqb.dqb_isoftlimit &&
            dquot->dq_dqb.dqb_itime == 0) {
-               *warntype = ISOFTWARN;
+               *warntype = QUOTA_NL_ISOFTWARN;
                dquot->dq_dqb.dqb_itime = get_seconds() + sb_dqopt(dquot->dq_sb)->info[dquot->dq_type].dqi_igrace;
        }
 
@@ -889,7 +1057,7 @@ static int check_idq(struct dquot *dquot, ulong inodes, char *warntype)
 /* needs dq_data_lock */
 static int check_bdq(struct dquot *dquot, qsize_t space, int prealloc, char *warntype)
 {
-       *warntype = 0;
+       *warntype = QUOTA_NL_NOWARN;
        if (space <= 0 || test_bit(DQ_FAKE_B, &dquot->dq_flags))
                return QUOTA_OK;
 
@@ -897,7 +1065,7 @@ static int check_bdq(struct dquot *dquot, qsize_t space, int prealloc, char *war
           toqb(dquot->dq_dqb.dqb_curspace + space) > dquot->dq_dqb.dqb_bhardlimit &&
             !ignore_hardlimit(dquot)) {
                if (!prealloc)
-                       *warntype = BHARDWARN;
+                       *warntype = QUOTA_NL_BHARDWARN;
                return NO_QUOTA;
        }
 
@@ -906,7 +1074,7 @@ static int check_bdq(struct dquot *dquot, qsize_t space, int prealloc, char *war
            dquot->dq_dqb.dqb_btime && get_seconds() >= dquot->dq_dqb.dqb_btime &&
             !ignore_hardlimit(dquot)) {
                if (!prealloc)
-                       *warntype = BSOFTLONGWARN;
+                       *warntype = QUOTA_NL_BSOFTLONGWARN;
                return NO_QUOTA;
        }
 
@@ -914,7 +1082,7 @@ static int check_bdq(struct dquot *dquot, qsize_t space, int prealloc, char *war
           toqb(dquot->dq_dqb.dqb_curspace + space) > dquot->dq_dqb.dqb_bsoftlimit &&
            dquot->dq_dqb.dqb_btime == 0) {
                if (!prealloc) {
-                       *warntype = BSOFTWARN;
+                       *warntype = QUOTA_NL_BSOFTWARN;
                        dquot->dq_dqb.dqb_btime = get_seconds() + sb_dqopt(dquot->dq_sb)->info[dquot->dq_type].dqi_bgrace;
                }
                else
@@ -928,6 +1096,35 @@ static int check_bdq(struct dquot *dquot, qsize_t space, int prealloc, char *war
        return QUOTA_OK;
 }
 
+static int info_idq_free(struct dquot *dquot, ulong inodes)
+{
+       if (test_bit(DQ_FAKE_B, &dquot->dq_flags) ||
+           dquot->dq_dqb.dqb_curinodes <= dquot->dq_dqb.dqb_isoftlimit)
+               return QUOTA_NL_NOWARN;
+
+       if (dquot->dq_dqb.dqb_curinodes - inodes <= dquot->dq_dqb.dqb_isoftlimit)
+               return QUOTA_NL_ISOFTBELOW;
+       if (dquot->dq_dqb.dqb_curinodes >= dquot->dq_dqb.dqb_ihardlimit &&
+           dquot->dq_dqb.dqb_curinodes - inodes < dquot->dq_dqb.dqb_ihardlimit)
+               return QUOTA_NL_IHARDBELOW;
+       return QUOTA_NL_NOWARN;
+}
+
+static int info_bdq_free(struct dquot *dquot, qsize_t space)
+{
+       if (test_bit(DQ_FAKE_B, &dquot->dq_flags) ||
+           toqb(dquot->dq_dqb.dqb_curspace) <= dquot->dq_dqb.dqb_bsoftlimit)
+               return QUOTA_NL_NOWARN;
+
+       if (toqb(dquot->dq_dqb.dqb_curspace - space) <=
+           dquot->dq_dqb.dqb_bsoftlimit)
+               return QUOTA_NL_BSOFTBELOW;
+       if (toqb(dquot->dq_dqb.dqb_curspace) >= dquot->dq_dqb.dqb_bhardlimit &&
+           toqb(dquot->dq_dqb.dqb_curspace - space) <
+                                               dquot->dq_dqb.dqb_bhardlimit)
+               return QUOTA_NL_BHARDBELOW;
+       return QUOTA_NL_NOWARN;
+}
 /*
  *     Initialize quota pointers in inode
  *     Transaction must be started at entry
@@ -937,8 +1134,8 @@ int dquot_initialize(struct inode *inode, int type)
        unsigned int id = 0;
        int cnt, ret = 0;
 
-       /* First test before acquiring semaphore - solves deadlocks when we
-         * re-enter the quota code and are already holding the semaphore */
+       /* First test before acquiring mutex - solves deadlocks when we
+         * re-enter the quota code and are already holding the mutex */
        if (IS_NOQUOTA(inode))
                return 0;
        down_write(&sb_dqopt(inode->i_sb)->dqptr_sem);
@@ -984,6 +1181,28 @@ int dquot_drop(struct inode *inode)
        return 0;
 }
 
+/* Wrapper to remove references to quota structures from inode */
+void vfs_dq_drop(struct inode *inode)
+{
+       /* Here we can get arbitrary inode from clear_inode() so we have
+        * to be careful. OTOH we don't need locking as quota operations
+        * are allowed to change only at mount time */
+       if (!IS_NOQUOTA(inode) && inode->i_sb && inode->i_sb->dq_op
+           && inode->i_sb->dq_op->drop) {
+               int cnt;
+               /* Test before calling to rule out calls from proc and such
+                 * where we are not allowed to block. Note that this is
+                * actually reliable test even without the lock - the caller
+                * must assure that nobody can come after the DQUOT_DROP and
+                * add quota pointers back anyway */
+               for (cnt = 0; cnt < MAXQUOTAS; cnt++)
+                       if (inode->i_dquot[cnt] != NODQUOT)
+                               break;
+               if (cnt < MAXQUOTAS)
+                       inode->i_sb->dq_op->drop(inode);
+       }
+}
+
 /*
  * Following four functions update i_blocks+i_bytes fields and
  * quota information (together with appropriate checks)
@@ -1001,15 +1220,15 @@ int dquot_alloc_space(struct inode *inode, qsize_t number, int warn)
        int cnt, ret = NO_QUOTA;
        char warntype[MAXQUOTAS];
 
-       /* First test before acquiring semaphore - solves deadlocks when we
-         * re-enter the quota code and are already holding the semaphore */
+       /* First test before acquiring mutex - solves deadlocks when we
+         * re-enter the quota code and are already holding the mutex */
        if (IS_NOQUOTA(inode)) {
 out_add:
                inode_add_bytes(inode, number);
                return QUOTA_OK;
        }
        for (cnt = 0; cnt < MAXQUOTAS; cnt++)
-               warntype[cnt] = NOWARN;
+               warntype[cnt] = QUOTA_NL_NOWARN;
 
        down_read(&sb_dqopt(inode->i_sb)->dqptr_sem);
        if (IS_NOQUOTA(inode)) {        /* Now we can do reliable test... */
@@ -1050,12 +1269,12 @@ int dquot_alloc_inode(const struct inode *inode, unsigned long number)
        int cnt, ret = NO_QUOTA;
        char warntype[MAXQUOTAS];
 
-       /* First test before acquiring semaphore - solves deadlocks when we
-         * re-enter the quota code and are already holding the semaphore */
+       /* First test before acquiring mutex - solves deadlocks when we
+         * re-enter the quota code and are already holding the mutex */
        if (IS_NOQUOTA(inode))
                return QUOTA_OK;
        for (cnt = 0; cnt < MAXQUOTAS; cnt++)
-               warntype[cnt] = NOWARN;
+               warntype[cnt] = QUOTA_NL_NOWARN;
        down_read(&sb_dqopt(inode->i_sb)->dqptr_sem);
        if (IS_NOQUOTA(inode)) {
                up_read(&sb_dqopt(inode->i_sb)->dqptr_sem);
@@ -1082,7 +1301,7 @@ warn_put_all:
                for (cnt = 0; cnt < MAXQUOTAS; cnt++)
                        if (inode->i_dquot[cnt])
                                mark_dquot_dirty(inode->i_dquot[cnt]);
-       flush_warnings((struct dquot **)inode->i_dquot, warntype);
+       flush_warnings(inode->i_dquot, warntype);
        up_read(&sb_dqopt(inode->i_sb)->dqptr_sem);
        return ret;
 }
@@ -1093,14 +1312,16 @@ warn_put_all:
 int dquot_free_space(struct inode *inode, qsize_t number)
 {
        unsigned int cnt;
+       char warntype[MAXQUOTAS];
 
-       /* First test before acquiring semaphore - solves deadlocks when we
-         * re-enter the quota code and are already holding the semaphore */
+       /* First test before acquiring mutex - solves deadlocks when we
+         * re-enter the quota code and are already holding the mutex */
        if (IS_NOQUOTA(inode)) {
 out_sub:
                inode_sub_bytes(inode, number);
                return QUOTA_OK;
        }
+
        down_read(&sb_dqopt(inode->i_sb)->dqptr_sem);
        /* Now recheck reliably when holding dqptr_sem */
        if (IS_NOQUOTA(inode)) {
@@ -1111,6 +1332,7 @@ out_sub:
        for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
                if (inode->i_dquot[cnt] == NODQUOT)
                        continue;
+               warntype[cnt] = info_bdq_free(inode->i_dquot[cnt], number);
                dquot_decr_space(inode->i_dquot[cnt], number);
        }
        inode_sub_bytes(inode, number);
@@ -1119,6 +1341,7 @@ out_sub:
        for (cnt = 0; cnt < MAXQUOTAS; cnt++)
                if (inode->i_dquot[cnt])
                        mark_dquot_dirty(inode->i_dquot[cnt]);
+       flush_warnings(inode->i_dquot, warntype);
        up_read(&sb_dqopt(inode->i_sb)->dqptr_sem);
        return QUOTA_OK;
 }
@@ -1129,11 +1352,13 @@ out_sub:
 int dquot_free_inode(const struct inode *inode, unsigned long number)
 {
        unsigned int cnt;
+       char warntype[MAXQUOTAS];
 
-       /* First test before acquiring semaphore - solves deadlocks when we
-         * re-enter the quota code and are already holding the semaphore */
+       /* First test before acquiring mutex - solves deadlocks when we
+         * re-enter the quota code and are already holding the mutex */
        if (IS_NOQUOTA(inode))
                return QUOTA_OK;
+
        down_read(&sb_dqopt(inode->i_sb)->dqptr_sem);
        /* Now recheck reliably when holding dqptr_sem */
        if (IS_NOQUOTA(inode)) {
@@ -1144,6 +1369,7 @@ int dquot_free_inode(const struct inode *inode, unsigned long number)
        for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
                if (inode->i_dquot[cnt] == NODQUOT)
                        continue;
+               warntype[cnt] = info_idq_free(inode->i_dquot[cnt], number);
                dquot_decr_inodes(inode->i_dquot[cnt], number);
        }
        spin_unlock(&dq_data_lock);
@@ -1151,6 +1377,7 @@ int dquot_free_inode(const struct inode *inode, unsigned long number)
        for (cnt = 0; cnt < MAXQUOTAS; cnt++)
                if (inode->i_dquot[cnt])
                        mark_dquot_dirty(inode->i_dquot[cnt]);
+       flush_warnings(inode->i_dquot, warntype);
        up_read(&sb_dqopt(inode->i_sb)->dqptr_sem);
        return QUOTA_OK;
 }
@@ -1168,16 +1395,17 @@ int dquot_transfer(struct inode *inode, struct iattr *iattr)
        struct dquot *transfer_to[MAXQUOTAS];
        int cnt, ret = NO_QUOTA, chuid = (iattr->ia_valid & ATTR_UID) && inode->i_uid != iattr->ia_uid,
            chgid = (iattr->ia_valid & ATTR_GID) && inode->i_gid != iattr->ia_gid;
-       char warntype[MAXQUOTAS];
+       char warntype_to[MAXQUOTAS];
+       char warntype_from_inodes[MAXQUOTAS], warntype_from_space[MAXQUOTAS];
 
-       /* First test before acquiring semaphore - solves deadlocks when we
-         * re-enter the quota code and are already holding the semaphore */
+       /* First test before acquiring mutex - solves deadlocks when we
+         * re-enter the quota code and are already holding the mutex */
        if (IS_NOQUOTA(inode))
                return QUOTA_OK;
        /* Clear the arrays */
        for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
                transfer_to[cnt] = transfer_from[cnt] = NODQUOT;
-               warntype[cnt] = NOWARN;
+               warntype_to[cnt] = QUOTA_NL_NOWARN;
        }
        down_write(&sb_dqopt(inode->i_sb)->dqptr_sem);
        /* Now recheck reliably when holding dqptr_sem */
@@ -1209,8 +1437,9 @@ int dquot_transfer(struct inode *inode, struct iattr *iattr)
                if (transfer_to[cnt] == NODQUOT)
                        continue;
                transfer_from[cnt] = inode->i_dquot[cnt];
-               if (check_idq(transfer_to[cnt], 1, warntype+cnt) == NO_QUOTA ||
-                   check_bdq(transfer_to[cnt], space, 0, warntype+cnt) == NO_QUOTA)
+               if (check_idq(transfer_to[cnt], 1, warntype_to + cnt) ==
+                   NO_QUOTA || check_bdq(transfer_to[cnt], space, 0,
+                   warntype_to + cnt) == NO_QUOTA)
                        goto warn_put_all;
        }
 
@@ -1226,6 +1455,10 @@ int dquot_transfer(struct inode *inode, struct iattr *iattr)
 
                /* Due to IO error we might not have transfer_from[] structure */
                if (transfer_from[cnt]) {
+                       warntype_from_inodes[cnt] =
+                               info_idq_free(transfer_from[cnt], 1);
+                       warntype_from_space[cnt] =
+                               info_bdq_free(transfer_from[cnt], space);
                        dquot_decr_inodes(transfer_from[cnt], 1);
                        dquot_decr_space(transfer_from[cnt], space);
                }
@@ -1245,7 +1478,9 @@ warn_put_all:
                if (transfer_to[cnt])
                        mark_dquot_dirty(transfer_to[cnt]);
        }
-       flush_warnings(transfer_to, warntype);
+       flush_warnings(transfer_to, warntype_to);
+       flush_warnings(transfer_from, warntype_from_inodes);
+       flush_warnings(transfer_from, warntype_from_space);
        
        for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
                if (ret == QUOTA_OK && transfer_from[cnt] != NODQUOT)
@@ -1257,6 +1492,18 @@ warn_put_all:
        return ret;
 }
 
+/* Wrapper for transferring ownership of an inode */
+int vfs_dq_transfer(struct inode *inode, struct iattr *iattr)
+{
+       if (sb_any_quota_enabled(inode->i_sb) && !IS_NOQUOTA(inode)) {
+               vfs_dq_init(inode);
+               if (inode->i_sb->dq_op->transfer(inode, iattr) == NO_QUOTA)
+                       return 1;
+       }
+       return 0;
+}
+
+
 /*
  * Write info of quota file to disk
  */
@@ -1265,9 +1512,9 @@ int dquot_commit_info(struct super_block *sb, int type)
        int ret;
        struct quota_info *dqopt = sb_dqopt(sb);
 
-       down(&dqopt->dqio_sem);
+       mutex_lock(&dqopt->dqio_mutex);
        ret = dqopt->ops[type]->write_file_info(sb, type);
-       up(&dqopt->dqio_sem);
+       mutex_unlock(&dqopt->dqio_mutex);
        return ret;
 }
 
@@ -1294,45 +1541,73 @@ static inline void set_enable_flags(struct quota_info *dqopt, int type)
        switch (type) {
                case USRQUOTA:
                        dqopt->flags |= DQUOT_USR_ENABLED;
+                       dqopt->flags &= ~DQUOT_USR_SUSPENDED;
                        break;
                case GRPQUOTA:
                        dqopt->flags |= DQUOT_GRP_ENABLED;
+                       dqopt->flags &= ~DQUOT_GRP_SUSPENDED;
                        break;
        }
 }
 
-static inline void reset_enable_flags(struct quota_info *dqopt, int type)
+static inline void reset_enable_flags(struct quota_info *dqopt, int type,
+                                     int remount)
 {
        switch (type) {
                case USRQUOTA:
                        dqopt->flags &= ~DQUOT_USR_ENABLED;
+                       if (remount)
+                               dqopt->flags |= DQUOT_USR_SUSPENDED;
+                       else
+                               dqopt->flags &= ~DQUOT_USR_SUSPENDED;
                        break;
                case GRPQUOTA:
                        dqopt->flags &= ~DQUOT_GRP_ENABLED;
+                       if (remount)
+                               dqopt->flags |= DQUOT_GRP_SUSPENDED;
+                       else
+                               dqopt->flags &= ~DQUOT_GRP_SUSPENDED;
                        break;
        }
 }
 
+
 /*
  * Turn quota off on a device. type == -1 ==> quotaoff for all types (umount)
  */
-int vfs_quota_off(struct super_block *sb, int type)
+int vfs_quota_off(struct super_block *sb, int type, int remount)
 {
-       int cnt;
+       int cnt, ret = 0;
        struct quota_info *dqopt = sb_dqopt(sb);
        struct inode *toputinode[MAXQUOTAS];
-       struct vfsmount *toputmnt[MAXQUOTAS];
 
        /* We need to serialize quota_off() for device */
-       down(&dqopt->dqonoff_sem);
+       mutex_lock(&dqopt->dqonoff_mutex);
+
+       /*
+        * Skip everything if there's nothing to do. We have to do this because
+        * sometimes we are called when fill_super() failed and calling
+        * sync_fs() in such cases does no good.
+        */
+       if (!sb_any_quota_enabled(sb) && !sb_any_quota_suspended(sb)) {
+               mutex_unlock(&dqopt->dqonoff_mutex);
+               return 0;
+       }
        for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
                toputinode[cnt] = NULL;
-               toputmnt[cnt] = NULL;
                if (type != -1 && cnt != type)
                        continue;
+               /* If we keep inodes of quota files after remount and quotaoff
+                * is called, drop kept inodes. */
+               if (!remount && sb_has_quota_suspended(sb, cnt)) {
+                       iput(dqopt->files[cnt]);
+                       dqopt->files[cnt] = NULL;
+                       reset_enable_flags(dqopt, cnt, 0);
+                       continue;
+               }
                if (!sb_has_quota_enabled(sb, cnt))
                        continue;
-               reset_enable_flags(dqopt, cnt);
+               reset_enable_flags(dqopt, cnt, remount);
 
                /* Note: these are blocking operations */
                drop_dquot_ref(sb, cnt);
@@ -1348,20 +1623,16 @@ int vfs_quota_off(struct super_block *sb, int type)
                put_quota_format(dqopt->info[cnt].dqi_format);
 
                toputinode[cnt] = dqopt->files[cnt];
-               toputmnt[cnt] = dqopt->mnt[cnt];
-               dqopt->files[cnt] = NULL;
-               dqopt->mnt[cnt] = NULL;
+               if (!remount)
+                       dqopt->files[cnt] = NULL;
                dqopt->info[cnt].dqi_flags = 0;
                dqopt->info[cnt].dqi_igrace = 0;
                dqopt->info[cnt].dqi_bgrace = 0;
                dqopt->ops[cnt] = NULL;
        }
-       up(&dqopt->dqonoff_sem);
+       mutex_unlock(&dqopt->dqonoff_mutex);
        /* Sync the superblock so that buffers with quota data are written to
-        * disk (and so userspace sees correct data afterwards).
-        * The reference to vfsmnt we are still holding protects us from
-        * umount (we don't have it only when quotas are turned on/off for
-        * journal replay but in that case we are guarded by the fs anyway). */
+        * disk (and so userspace sees correct data afterwards). */
        if (sb->s_op->sync_fs)
                sb->s_op->sync_fs(sb, 1);
        sync_blockdev(sb->s_bdev);
@@ -1372,27 +1643,30 @@ int vfs_quota_off(struct super_block *sb, int type)
         * changes done by userspace on the next quotaon() */
        for (cnt = 0; cnt < MAXQUOTAS; cnt++)
                if (toputinode[cnt]) {
-                       down(&dqopt->dqonoff_sem);
+                       mutex_lock(&dqopt->dqonoff_mutex);
                        /* If quota was reenabled in the meantime, we have
                         * nothing to do */
                        if (!sb_has_quota_enabled(sb, cnt)) {
-                               down(&toputinode[cnt]->i_sem);
+                               mutex_lock_nested(&toputinode[cnt]->i_mutex, I_MUTEX_QUOTA);
                                toputinode[cnt]->i_flags &= ~(S_IMMUTABLE |
                                  S_NOATIME | S_NOQUOTA);
                                truncate_inode_pages(&toputinode[cnt]->i_data, 0);
-                               up(&toputinode[cnt]->i_sem);
+                               mutex_unlock(&toputinode[cnt]->i_mutex);
                                mark_inode_dirty(toputinode[cnt]);
-                               iput(toputinode[cnt]);
                        }
-                       up(&dqopt->dqonoff_sem);
-                       /* We don't hold the reference when we turned on quotas
-                        * just for the journal replay... */
-                       if (toputmnt[cnt])
-                               mntput(toputmnt[cnt]);
+                       mutex_unlock(&dqopt->dqonoff_mutex);
+                       /* On remount RO, we keep the inode pointer so that we
+                        * can reenable quota on the subsequent remount RW.
+                        * But we have better not keep inode pointer when there
+                        * is pending delete on the quota file... */
+                       if (!remount)
+                               iput(toputinode[cnt]);
+                       else if (!toputinode[cnt]->i_nlink)
+                               ret = -EBUSY;
                }
        if (sb->s_bdev)
-               invalidate_bdev(sb->s_bdev, 0);
-       return 0;
+               invalidate_bdev(sb->s_bdev);
+       return ret;
 }
 
 /*
@@ -1427,10 +1701,11 @@ static int vfs_quota_on_inode(struct inode *inode, int type, int format_id)
         * we see all the changes from userspace... */
        write_inode_now(inode, 1);
        /* And now flush the block cache so that kernel sees the changes */
-       invalidate_bdev(sb->s_bdev, 0);
-       down(&inode->i_sem);
-       down(&dqopt->dqonoff_sem);
-       if (sb_has_quota_enabled(sb, type)) {
+       invalidate_bdev(sb->s_bdev);
+       mutex_lock(&inode->i_mutex);
+       mutex_lock(&dqopt->dqonoff_mutex);
+       if (sb_has_quota_enabled(sb, type) ||
+                       sb_has_quota_suspended(sb, type)) {
                error = -EBUSY;
                goto out_lock;
        }
@@ -1453,18 +1728,19 @@ static int vfs_quota_on_inode(struct inode *inode, int type, int format_id)
 
        dqopt->ops[type] = fmt->qf_ops;
        dqopt->info[type].dqi_format = fmt;
+       dqopt->info[type].dqi_fmt_id = format_id;
        INIT_LIST_HEAD(&dqopt->info[type].dqi_dirty_list);
-       down(&dqopt->dqio_sem);
+       mutex_lock(&dqopt->dqio_mutex);
        if ((error = dqopt->ops[type]->read_file_info(sb, type)) < 0) {
-               up(&dqopt->dqio_sem);
+               mutex_unlock(&dqopt->dqio_mutex);
                goto out_file_init;
        }
-       up(&dqopt->dqio_sem);
-       up(&inode->i_sem);
+       mutex_unlock(&dqopt->dqio_mutex);
+       mutex_unlock(&inode->i_mutex);
        set_enable_flags(dqopt, type);
 
        add_dquot_ref(sb, type);
-       up(&dqopt->dqonoff_sem);
+       mutex_unlock(&dqopt->dqonoff_mutex);
 
        return 0;
 
@@ -1472,7 +1748,7 @@ out_file_init:
        dqopt->files[type] = NULL;
        iput(inode);
 out_lock:
-       up(&dqopt->dqonoff_sem);
+       mutex_unlock(&dqopt->dqonoff_mutex);
        if (oldflags != -1) {
                down_write(&dqopt->dqptr_sem);
                /* Set the flags back (in the case of accidental quotaon()
@@ -1481,35 +1757,68 @@ out_lock:
                inode->i_flags |= oldflags;
                up_write(&dqopt->dqptr_sem);
        }
-       up(&inode->i_sem);
+       mutex_unlock(&inode->i_mutex);
 out_fmt:
        put_quota_format(fmt);
 
        return error; 
 }
 
-/* Actual function called from quotactl() */
-int vfs_quota_on(struct super_block *sb, int type, int format_id, char *path)
+/* Reenable quotas on remount RW */
+static int vfs_quota_on_remount(struct super_block *sb, int type)
 {
-       struct nameidata nd;
-       int error;
+       struct quota_info *dqopt = sb_dqopt(sb);
+       struct inode *inode;
+       int ret;
 
-       error = path_lookup(path, LOOKUP_FOLLOW, &nd);
-       if (error < 0)
-               return error;
-       error = security_quota_on(nd.dentry);
+       mutex_lock(&dqopt->dqonoff_mutex);
+       if (!sb_has_quota_suspended(sb, type)) {
+               mutex_unlock(&dqopt->dqonoff_mutex);
+               return 0;
+       }
+       BUG_ON(sb_has_quota_enabled(sb, type));
+
+       inode = dqopt->files[type];
+       dqopt->files[type] = NULL;
+       reset_enable_flags(dqopt, type, 0);
+       mutex_unlock(&dqopt->dqonoff_mutex);
+
+       ret = vfs_quota_on_inode(inode, type, dqopt->info[type].dqi_fmt_id);
+       iput(inode);
+
+       return ret;
+}
+
+int vfs_quota_on_path(struct super_block *sb, int type, int format_id,
+                     struct path *path)
+{
+       int error = security_quota_on(path->dentry);
        if (error)
-               goto out_path;
+               return error;
        /* Quota file not on the same filesystem? */
-       if (nd.mnt->mnt_sb != sb)
+       if (path->mnt->mnt_sb != sb)
                error = -EXDEV;
-       else {
-               error = vfs_quota_on_inode(nd.dentry->d_inode, type, format_id);
-               if (!error)
-                       sb_dqopt(sb)->mnt[type] = mntget(nd.mnt);
+       else
+               error = vfs_quota_on_inode(path->dentry->d_inode, type,
+                                          format_id);
+       return error;
+}
+
+/* Actual function called from quotactl() */
+int vfs_quota_on(struct super_block *sb, int type, int format_id, char *name,
+                int remount)
+{
+       struct path path;
+       int error;
+
+       if (remount)
+               return vfs_quota_on_remount(sb, type);
+
+       error = kern_path(name, LOOKUP_FOLLOW, &path);
+       if (!error) {
+               error = vfs_quota_on_path(sb, type, format_id, &path);
+               path_put(&path);
        }
-out_path:
-       path_release(&nd);
        return error;
 }
 
@@ -1527,14 +1836,36 @@ int vfs_quota_on_mount(struct super_block *sb, char *qf_name,
        if (IS_ERR(dentry))
                return PTR_ERR(dentry);
 
+       if (!dentry->d_inode) {
+               error = -ENOENT;
+               goto out;
+       }
+
        error = security_quota_on(dentry);
        if (!error)
                error = vfs_quota_on_inode(dentry->d_inode, type, format_id);
 
+out:
        dput(dentry);
        return error;
 }
 
+/* Wrapper to turn on quotas when remounting rw */
+int vfs_dq_quota_on_remount(struct super_block *sb)
+{
+       int cnt;
+       int ret = 0, err;
+
+       if (!sb->s_qcop || !sb->s_qcop->quota_on)
+               return -ENOSYS;
+       for (cnt = 0; cnt < MAXQUOTAS; cnt++) {
+               err = sb->s_qcop->quota_on(sb, cnt, 0, NULL, 1);
+               if (err < 0 && !ret)
+                       ret = err;
+       }
+       return ret;
+}
+
 /* Generic routine for getting common part of quota structure */
 static void do_get_dqblk(struct dquot *dquot, struct if_dqblk *di)
 {
@@ -1557,22 +1888,31 @@ int vfs_get_dqblk(struct super_block *sb, int type, qid_t id, struct if_dqblk *d
 {
        struct dquot *dquot;
 
-       down(&sb_dqopt(sb)->dqonoff_sem);
+       mutex_lock(&sb_dqopt(sb)->dqonoff_mutex);
        if (!(dquot = dqget(sb, id, type))) {
-               up(&sb_dqopt(sb)->dqonoff_sem);
+               mutex_unlock(&sb_dqopt(sb)->dqonoff_mutex);
                return -ESRCH;
        }
        do_get_dqblk(dquot, di);
        dqput(dquot);
-       up(&sb_dqopt(sb)->dqonoff_sem);
+       mutex_unlock(&sb_dqopt(sb)->dqonoff_mutex);
        return 0;
 }
 
 /* Generic routine for setting common part of quota structure */
-static void do_set_dqblk(struct dquot *dquot, struct if_dqblk *di)
+static int do_set_dqblk(struct dquot *dquot, struct if_dqblk *di)
 {
        struct mem_dqblk *dm = &dquot->dq_dqb;
        int check_blim = 0, check_ilim = 0;
+       struct mem_dqinfo *dqi = &sb_dqopt(dquot->dq_sb)->info[dquot->dq_type];
+
+       if ((di->dqb_valid & QIF_BLIMITS &&
+            (di->dqb_bhardlimit > dqi->dqi_maxblimit ||
+             di->dqb_bsoftlimit > dqi->dqi_maxblimit)) ||
+           (di->dqb_valid & QIF_ILIMITS &&
+            (di->dqb_ihardlimit > dqi->dqi_maxilimit ||
+             di->dqb_isoftlimit > dqi->dqi_maxilimit)))
+               return -ERANGE;
 
        spin_lock(&dq_data_lock);
        if (di->dqb_valid & QIF_SPACE) {
@@ -1604,7 +1944,7 @@ static void do_set_dqblk(struct dquot *dquot, struct if_dqblk *di)
                        clear_bit(DQ_BLKS_B, &dquot->dq_flags);
                }
                else if (!(di->dqb_valid & QIF_BTIME))  /* Set grace only if user hasn't provided his own... */
-                       dm->dqb_btime = get_seconds() + sb_dqopt(dquot->dq_sb)->info[dquot->dq_type].dqi_bgrace;
+                       dm->dqb_btime = get_seconds() + dqi->dqi_bgrace;
        }
        if (check_ilim) {
                if (!dm->dqb_isoftlimit || dm->dqb_curinodes < dm->dqb_isoftlimit) {
@@ -1612,7 +1952,7 @@ static void do_set_dqblk(struct dquot *dquot, struct if_dqblk *di)
                        clear_bit(DQ_INODES_B, &dquot->dq_flags);
                }
                else if (!(di->dqb_valid & QIF_ITIME))  /* Set grace only if user hasn't provided his own... */
-                       dm->dqb_itime = get_seconds() + sb_dqopt(dquot->dq_sb)->info[dquot->dq_type].dqi_igrace;
+                       dm->dqb_itime = get_seconds() + dqi->dqi_igrace;
        }
        if (dm->dqb_bhardlimit || dm->dqb_bsoftlimit || dm->dqb_ihardlimit || dm->dqb_isoftlimit)
                clear_bit(DQ_FAKE_B, &dquot->dq_flags);
@@ -1620,21 +1960,24 @@ static void do_set_dqblk(struct dquot *dquot, struct if_dqblk *di)
                set_bit(DQ_FAKE_B, &dquot->dq_flags);
        spin_unlock(&dq_data_lock);
        mark_dquot_dirty(dquot);
+
+       return 0;
 }
 
 int vfs_set_dqblk(struct super_block *sb, int type, qid_t id, struct if_dqblk *di)
 {
        struct dquot *dquot;
+       int rc;
 
-       down(&sb_dqopt(sb)->dqonoff_sem);
+       mutex_lock(&sb_dqopt(sb)->dqonoff_mutex);
        if (!(dquot = dqget(sb, id, type))) {
-               up(&sb_dqopt(sb)->dqonoff_sem);
+               mutex_unlock(&sb_dqopt(sb)->dqonoff_mutex);
                return -ESRCH;
        }
-       do_set_dqblk(dquot, di);
+       rc = do_set_dqblk(dquot, di);
        dqput(dquot);
-       up(&sb_dqopt(sb)->dqonoff_sem);
-       return 0;
+       mutex_unlock(&sb_dqopt(sb)->dqonoff_mutex);
+       return rc;
 }
 
 /* Generic routine for getting common part of quota file information */
@@ -1642,9 +1985,9 @@ int vfs_get_dqinfo(struct super_block *sb, int type, struct if_dqinfo *ii)
 {
        struct mem_dqinfo *mi;
   
-       down(&sb_dqopt(sb)->dqonoff_sem);
+       mutex_lock(&sb_dqopt(sb)->dqonoff_mutex);
        if (!sb_has_quota_enabled(sb, type)) {
-               up(&sb_dqopt(sb)->dqonoff_sem);
+               mutex_unlock(&sb_dqopt(sb)->dqonoff_mutex);
                return -ESRCH;
        }
        mi = sb_dqopt(sb)->info + type;
@@ -1654,7 +1997,7 @@ int vfs_get_dqinfo(struct super_block *sb, int type, struct if_dqinfo *ii)
        ii->dqi_flags = mi->dqi_flags & DQF_MASK;
        ii->dqi_valid = IIF_ALL;
        spin_unlock(&dq_data_lock);
-       up(&sb_dqopt(sb)->dqonoff_sem);
+       mutex_unlock(&sb_dqopt(sb)->dqonoff_mutex);
        return 0;
 }
 
@@ -1663,9 +2006,9 @@ int vfs_set_dqinfo(struct super_block *sb, int type, struct if_dqinfo *ii)
 {
        struct mem_dqinfo *mi;
 
-       down(&sb_dqopt(sb)->dqonoff_sem);
+       mutex_lock(&sb_dqopt(sb)->dqonoff_mutex);
        if (!sb_has_quota_enabled(sb, type)) {
-               up(&sb_dqopt(sb)->dqonoff_sem);
+               mutex_unlock(&sb_dqopt(sb)->dqonoff_mutex);
                return -ESRCH;
        }
        mi = sb_dqopt(sb)->info + type;
@@ -1680,7 +2023,7 @@ int vfs_set_dqinfo(struct super_block *sb, int type, struct if_dqinfo *ii)
        mark_info_dirty(sb, type);
        /* Force write to disk */
        sb->dq_op->write_info(sb, type);
-       up(&sb_dqopt(sb)->dqonoff_sem);
+       mutex_unlock(&sb_dqopt(sb)->dqonoff_mutex);
        return 0;
 }
 
@@ -1759,6 +2102,7 @@ static ctl_table fs_dqstats_table[] = {
                .mode           = 0444,
                .proc_handler   = &proc_dointvec,
        },
+#ifdef CONFIG_PRINT_QUOTA_WARNING
        {
                .ctl_name       = FS_DQ_WARNINGS,
                .procname       = "warnings",
@@ -1767,6 +2111,7 @@ static ctl_table fs_dqstats_table[] = {
                .mode           = 0644,
                .proc_handler   = &proc_dointvec,
        },
+#endif
        { .ctl_name = 0 },
 };
 
@@ -1797,12 +2142,13 @@ static int __init dquot_init(void)
 
        printk(KERN_NOTICE "VFS: Disk quotas %s\n", __DQUOT_VERSION__);
 
-       register_sysctl_table(sys_table, 0);
+       register_sysctl_table(sys_table);
 
-       dquot_cachep = kmem_cache_create("dquot", 
+       dquot_cachep = kmem_cache_create("dquot",
                        sizeof(struct dquot), sizeof(unsigned long) * 4,
-                       SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT|SLAB_PANIC,
-                       NULL, NULL);
+                       (SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT|
+                               SLAB_MEM_SPREAD|SLAB_PANIC),
+                       NULL);
 
        order = 0;
        dquot_hash = (struct hlist_head *)__get_free_pages(GFP_ATOMIC, order);
@@ -1825,7 +2171,12 @@ static int __init dquot_init(void)
        printk("Dquot-cache hash table entries: %ld (order %ld, %ld bytes)\n",
                        nr_hash, order, (PAGE_SIZE << order));
 
-       set_shrinker(DEFAULT_SEEKS, shrink_dqcache_memory);
+       register_shrinker(&dqcache_shrinker);
+
+#ifdef CONFIG_QUOTA_NETLINK_INTERFACE
+       if (genl_register_family(&quota_genl_family) != 0)
+               printk(KERN_ERR "VFS: Failed to create quota netlink interface.\n");
+#endif
 
        return 0;
 }
@@ -1836,6 +2187,7 @@ EXPORT_SYMBOL(unregister_quota_format);
 EXPORT_SYMBOL(dqstats);
 EXPORT_SYMBOL(dq_data_lock);
 EXPORT_SYMBOL(vfs_quota_on);
+EXPORT_SYMBOL(vfs_quota_on_path);
 EXPORT_SYMBOL(vfs_quota_on_mount);
 EXPORT_SYMBOL(vfs_quota_off);
 EXPORT_SYMBOL(vfs_quota_sync);
@@ -1850,8 +2202,11 @@ EXPORT_SYMBOL(dquot_release);
 EXPORT_SYMBOL(dquot_mark_dquot_dirty);
 EXPORT_SYMBOL(dquot_initialize);
 EXPORT_SYMBOL(dquot_drop);
+EXPORT_SYMBOL(vfs_dq_drop);
 EXPORT_SYMBOL(dquot_alloc_space);
 EXPORT_SYMBOL(dquot_alloc_inode);
 EXPORT_SYMBOL(dquot_free_space);
 EXPORT_SYMBOL(dquot_free_inode);
 EXPORT_SYMBOL(dquot_transfer);
+EXPORT_SYMBOL(vfs_dq_transfer);
+EXPORT_SYMBOL(vfs_dq_quota_on_remount);