[XFS] Concurrent Multi-File Data Streams

In media spaces, video is often stored in a frame-per-file format. When
dealing with uncompressed realtime HD video streams in this format, it is
crucial that files do not get fragmented and that multiple files a placed
contiguously on disk.

When multiple streams are being ingested and played out at the same time,
it is critical that the filesystem does not cross the streams and
interleave them together as this creates seek and readahead cache miss
latency and prevents both ingest and playout from meeting frame rate
targets.

This patch set creates a "stream of files" concept into the allocator to
place all the data from a single stream contiguously on disk so that RAID
array readahead can be used effectively. Each additional stream gets
placed in different allocation groups within the filesystem, thereby
ensuring that we don't cross any streams. When an AG fills up, we select a
new AG for the stream that is not in use.

The core of the functionality is the stream tracking - each inode that we
create in a directory needs to be associated with the directories' stream.
Hence every time we create a file, we look up the directories' stream
object and associate the new file with that object.

Once we have a stream object for a file, we use the AG that the stream
object point to for allocations. If we can't allocate in that AG (e.g. it
is full) we move the entire stream to another AG. Other inodes in the same
stream are moved to the new AG on their next allocation (i.e. lazy
update).

Stream objects are kept in a cache and hold a reference on the inode.
Hence the inode cannot be reclaimed while there is an outstanding stream
reference. This means that on unlink we need to remove the stream
association and we also need to flush all the associations on certain
events that want to reclaim all unreferenced inodes (e.g. filesystem
freeze).

SGI-PV: 964469
SGI-Modid: xfs-linux-melb:xfs-kern:29096a

Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Barry Naujok <bnaujok@sgi.com>
Signed-off-by: Donald Douwsma <donaldd@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Vlad Apostolov <vapo@sgi.com>

1 files changed, 608 insertions, 0 deletions

diff --git a/fs/xfs/xfs_mru_cache.c b/fs/xfs/xfs_mru_cache.c
new file mode 100644
index 00000000000..7deb9e3cbbd
--- /dev/null
+++ b/fs/xfs/xfs_mru_cache.c
@@ -0,0 +1,608 @@
+/*
+ * Copyright (c) 2006-2007 Silicon Graphics, Inc.
+ * All Rights Reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it would be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write the Free Software Foundation,
+ * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
+ */
+#include "xfs.h"
+#include "xfs_mru_cache.h"
+
+/*
+ * The MRU Cache data structure consists of a data store, an array of lists and
+ * a lock to protect its internal state.  At initialisation time, the client
+ * supplies an element lifetime in milliseconds and a group count, as well as a
+ * function pointer to call when deleting elements.  A data structure for
+ * queueing up work in the form of timed callbacks is also included.
+ *
+ * The group count controls how many lists are created, and thereby how finely
+ * the elements are grouped in time.  When reaping occurs, all the elements in
+ * all the lists whose time has expired are deleted.
+ *
+ * To give an example of how this works in practice, consider a client that
+ * initialises an MRU Cache with a lifetime of ten seconds and a group count of
+ * five.  Five internal lists will be created, each representing a two second
+ * period in time.  When the first element is added, time zero for the data
+ * structure is initialised to the current time.
+ *
+ * All the elements added in the first two seconds are appended to the first
+ * list.  Elements added in the third second go into the second list, and so on.
+ * If an element is accessed at any point, it is removed from its list and
+ * inserted at the head of the current most-recently-used list.
+ *
+ * The reaper function will have nothing to do until at least twelve seconds
+ * have elapsed since the first element was added.  The reason for this is that
+ * if it were called at t=11s, there could be elements in the first list that
+ * have only been inactive for nine seconds, so it still does nothing.  If it is
+ * called anywhere between t=12 and t=14 seconds, it will delete all the
+ * elements that remain in the first list.  It's therefore possible for elements
+ * to remain in the data store even after they've been inactive for up to
+ * (t + t/g) seconds, where t is the inactive element lifetime and g is the
+ * number of groups.
+ *
+ * The above example assumes that the reaper function gets called at least once
+ * every (t/g) seconds.  If it is called less frequently, unused elements will
+ * accumulate in the reap list until the reaper function is eventually called.
+ * The current implementation uses work queue callbacks to carefully time the
+ * reaper function calls, so this should happen rarely, if at all.
+ *
+ * From a design perspective, the primary reason for the choice of a list array
+ * representing discrete time intervals is that it's only practical to reap
+ * expired elements in groups of some appreciable size.  This automatically
+ * introduces a granularity to element lifetimes, so there's no point storing an
+ * individual timeout with each element that specifies a more precise reap time.
+ * The bonus is a saving of sizeof(long) bytes of memory per element stored.
+ *
+ * The elements could have been stored in just one list, but an array of
+ * counters or pointers would need to be maintained to allow them to be divided
+ * up into discrete time groups.  More critically, the process of touching or
+ * removing an element would involve walking large portions of the entire list,
+ * which would have a detrimental effect on performance.  The additional memory
+ * requirement for the array of list heads is minimal.
+ *
+ * When an element is touched or deleted, it needs to be removed from its
+ * current list.  Doubly linked lists are used to make the list maintenance
+ * portion of these operations O(1).  Since reaper timing can be imprecise,
+ * inserts and lookups can occur when there are no free lists available.  When
+ * this happens, all the elements on the LRU list need to be migrated to the end
+ * of the reap list.  To keep the list maintenance portion of these operations
+ * O(1) also, list tails need to be accessible without walking the entire list.
+ * This is the reason why doubly linked list heads are used.
+ */
+
+/*
+ * An MRU Cache is a dynamic data structure that stores its elements in a way
+ * that allows efficient lookups, but also groups them into discrete time
+ * intervals based on insertion time.  This allows elements to be efficiently
+ * and automatically reaped after a fixed period of inactivity.
+ *
+ * When a client data pointer is stored in the MRU Cache it needs to be added to
+ * both the data store and to one of the lists.  It must also be possible to
+ * access each of these entries via the other, i.e. to:
+ *
+ *    a) Walk a list, removing the corresponding data store entry for each item.
+ *    b) Look up a data store entry, then access its list entry directly.
+ *
+ * To achieve both of these goals, each entry must contain both a list entry and
+ * a key, in addition to the user's data pointer.  Note that it's not a good
+ * idea to have the client embed one of these structures at the top of their own
+ * data structure, because inserting the same item more than once would most
+ * likely result in a loop in one of the lists.  That's a sure-fire recipe for
+ * an infinite loop in the code.
+ */
+typedef struct xfs_mru_cache_elem
+{
+	struct list_head list_node;
+	unsigned long	key;
+	void		*value;
+} xfs_mru_cache_elem_t;
+
+static kmem_zone_t		*xfs_mru_elem_zone;
+static struct workqueue_struct	*xfs_mru_reap_wq;
+
+/*
+ * When inserting, destroying or reaping, it's first necessary to update the
+ * lists relative to a particular time.  In the case of destroying, that time
+ * will be well in the future to ensure that all items are moved to the reap
+ * list.  In all other cases though, the time will be the current time.
+ *
+ * This function enters a loop, moving the contents of the LRU list to the reap
+ * list again and again until either a) the lists are all empty, or b) time zero
+ * has been advanced sufficiently to be within the immediate element lifetime.
+ *
+ * Case a) above is detected by counting how many groups are migrated and
+ * stopping when they've all been moved.  Case b) is detected by monitoring the
+ * time_zero field, which is updated as each group is migrated.
+ *
+ * The return value is the earliest time that more migration could be needed, or
+ * zero if there's no need to schedule more work because the lists are empty.
+ */
+STATIC unsigned long
+_xfs_mru_cache_migrate(
+	xfs_mru_cache_t	*mru,
+	unsigned long	now)
+{
+	unsigned int	grp;
+	unsigned int	migrated = 0;
+	struct list_head *lru_list;
+
+	/* Nothing to do if the data store is empty. */
+	if (!mru->time_zero)
+		return 0;
+
+	/* While time zero is older than the time spanned by all the lists. */
+	while (mru->time_zero <= now - mru->grp_count * mru->grp_time) {
+
+		/*
+		 * If the LRU list isn't empty, migrate its elements to the tail
+		 * of the reap list.
+		 */
+		lru_list = mru->lists + mru->lru_grp;
+		if (!list_empty(lru_list))
+			list_splice_init(lru_list, mru->reap_list.prev);
+
+		/*
+		 * Advance the LRU group number, freeing the old LRU list to
+		 * become the new MRU list; advance time zero accordingly.
+		 */
+		mru->lru_grp = (mru->lru_grp + 1) % mru->grp_count;
+		mru->time_zero += mru->grp_time;
+
+		/*
+		 * If reaping is so far behind that all the elements on all the
+		 * lists have been migrated to the reap list, it's now empty.
+		 */
+		if (++migrated == mru->grp_count) {
+			mru->lru_grp = 0;
+			mru->time_zero = 0;
+			return 0;
+		}
+	}
+
+	/* Find the first non-empty list from the LRU end. */
+	for (grp = 0; grp < mru->grp_count; grp++) {
+
+		/* Check the grp'th list from the LRU end. */
+		lru_list = mru->lists + ((mru->lru_grp + grp) % mru->grp_count);
+		if (!list_empty(lru_list))
+			return mru->time_zero +
+			       (mru->grp_count + grp) * mru->grp_time;
+	}
+
+	/* All the lists must be empty. */
+	mru->lru_grp = 0;
+	mru->time_zero = 0;
+	return 0;
+}
+
+/*
+ * When inserting or doing a lookup, an element needs to be inserted into the
+ * MRU list.  The lists must be migrated first to ensure that they're
+ * up-to-date, otherwise the new element could be given a shorter lifetime in
+ * the cache than it should.
+ */
+STATIC void
+_xfs_mru_cache_list_insert(
+	xfs_mru_cache_t		*mru,
+	xfs_mru_cache_elem_t	*elem)
+{
+	unsigned int	grp = 0;
+	unsigned long	now = jiffies;
+
+	/*
+	 * If the data store is empty, initialise time zero, leave grp set to
+	 * zero and start the work queue timer if necessary.  Otherwise, set grp
+	 * to the number of group times that have elapsed since time zero.
+	 */
+	if (!_xfs_mru_cache_migrate(mru, now)) {
+		mru->time_zero = now;
+		if (!mru->next_reap)
+			mru->next_reap = mru->grp_count * mru->grp_time;
+	} else {
+		grp = (now - mru->time_zero) / mru->grp_time;
+		grp = (mru->lru_grp + grp) % mru->grp_count;
+	}
+
+	/* Insert the element at the tail of the corresponding list. */
+	list_add_tail(&elem->list_node, mru->lists + grp);
+}
+
+/*
+ * When destroying or reaping, all the elements that were migrated to the reap
+ * list need to be deleted.  For each element this involves removing it from the
+ * data store, removing it from the reap list, calling the client's free
+ * function and deleting the element from the element zone.
+ */
+STATIC void
+_xfs_mru_cache_clear_reap_list(
+	xfs_mru_cache_t		*mru)
+{
+	xfs_mru_cache_elem_t	*elem, *next;
+	struct list_head	tmp;
+
+	INIT_LIST_HEAD(&tmp);
+	list_for_each_entry_safe(elem, next, &mru->reap_list, list_node) {
+
+		/* Remove the element from the data store. */
+		radix_tree_delete(&mru->store, elem->key);
+
+		/*
+		 * remove to temp list so it can be freed without
+		 * needing to hold the lock
+		 */
+		list_move(&elem->list_node, &tmp);
+	}
+	mutex_spinunlock(&mru->lock, 0);
+
+	list_for_each_entry_safe(elem, next, &tmp, list_node) {
+
+		/* Remove the element from the reap list. */
+		list_del_init(&elem->list_node);
+
+		/* Call the client's free function with the key and value pointer. */
+		mru->free_func(elem->key, elem->value);
+
+		/* Free the element structure. */
+		kmem_zone_free(xfs_mru_elem_zone, elem);
+	}
+
+	mutex_spinlock(&mru->lock);
+}
+
+/*
+ * We fire the reap timer every group expiry interval so
+ * we always have a reaper ready to run. This makes shutdown
+ * and flushing of the reaper easy to do. Hence we need to
+ * keep when the next reap must occur so we can determine
+ * at each interval whether there is anything we need to do.
+ */
+STATIC void
+_xfs_mru_cache_reap(
+	struct work_struct	*work)
+{
+	xfs_mru_cache_t		*mru = container_of(work, xfs_mru_cache_t, work.work);
+	unsigned long		now;
+
+	ASSERT(mru && mru->lists);
+	if (!mru || !mru->lists)
+		return;
+
+	mutex_spinlock(&mru->lock);
+	now = jiffies;
+	if (mru->reap_all ||
+	    (mru->next_reap && time_after(now, mru->next_reap))) {
+		if (mru->reap_all)
+			now += mru->grp_count * mru->grp_time * 2;
+		mru->next_reap = _xfs_mru_cache_migrate(mru, now);
+		_xfs_mru_cache_clear_reap_list(mru);
+	}
+
+	/*
+	 * the process that triggered the reap_all is responsible
+	 * for restating the periodic reap if it is required.
+	 */
+	if (!mru->reap_all)
+		queue_delayed_work(xfs_mru_reap_wq, &mru->work, mru->grp_time);
+	mru->reap_all = 0;
+	mutex_spinunlock(&mru->lock, 0);
+}
+
+int
+xfs_mru_cache_init(void)
+{
+	xfs_mru_elem_zone = kmem_zone_init(sizeof(xfs_mru_cache_elem_t),
+	                                 "xfs_mru_cache_elem");
+	if (!xfs_mru_elem_zone)
+		return ENOMEM;
+
+	xfs_mru_reap_wq = create_singlethread_workqueue("xfs_mru_cache");
+	if (!xfs_mru_reap_wq) {
+		kmem_zone_destroy(xfs_mru_elem_zone);
+		return ENOMEM;
+	}
+
+	return 0;
+}
+
+void
+xfs_mru_cache_uninit(void)
+{
+	destroy_workqueue(xfs_mru_reap_wq);
+	kmem_zone_destroy(xfs_mru_elem_zone);
+}
+
+/*
+ * To initialise a struct xfs_mru_cache pointer, call xfs_mru_cache_create()
+ * with the address of the pointer, a lifetime value in milliseconds, a group
+ * count and a free function to use when deleting elements.  This function
+ * returns 0 if the initialisation was successful.
+ */
+int
+xfs_mru_cache_create(
+	xfs_mru_cache_t		**mrup,
+	unsigned int		lifetime_ms,
+	unsigned int		grp_count,
+	xfs_mru_cache_free_func_t free_func)
+{
+	xfs_mru_cache_t	*mru = NULL;
+	int		err = 0, grp;
+	unsigned int	grp_time;
+
+	if (mrup)
+		*mrup = NULL;
+
+	if (!mrup || !grp_count || !lifetime_ms || !free_func)
+		return EINVAL;
+
+	if (!(grp_time = msecs_to_jiffies(lifetime_ms) / grp_count))
+		return EINVAL;
+
+	if (!(mru = kmem_zalloc(sizeof(*mru), KM_SLEEP)))
+		return ENOMEM;
+
+	/* An extra list is needed to avoid reaping up to a grp_time early. */
+	mru->grp_count = grp_count + 1;
+	mru->lists = kmem_alloc(mru->grp_count * sizeof(*mru->lists), KM_SLEEP);
+
+	if (!mru->lists) {
+		err = ENOMEM;
+		goto exit;
+	}
+
+	for (grp = 0; grp < mru->grp_count; grp++)
+		INIT_LIST_HEAD(mru->lists + grp);
+
+	/*
+	 * We use GFP_KERNEL radix tree preload and do inserts under a
+	 * spinlock so GFP_ATOMIC is appropriate for the radix tree itself.
+	 */
+	INIT_RADIX_TREE(&mru->store, GFP_ATOMIC);
+	INIT_LIST_HEAD(&mru->reap_list);
+	spinlock_init(&mru->lock, "xfs_mru_cache");
+	INIT_DELAYED_WORK(&mru->work, _xfs_mru_cache_reap);
+
+	mru->grp_time  = grp_time;
+	mru->free_func = free_func;
+
+	/* start up the reaper event */
+	mru->next_reap = 0;
+	mru->reap_all = 0;
+	queue_delayed_work(xfs_mru_reap_wq, &mru->work, mru->grp_time);
+
+	*mrup = mru;
+
+exit:
+	if (err && mru && mru->lists)
+		kmem_free(mru->lists, mru->grp_count * sizeof(*mru->lists));
+	if (err && mru)
+		kmem_free(mru, sizeof(*mru));
+
+	return err;
+}
+
+/*
+ * Call xfs_mru_cache_flush() to flush out all cached entries, calling their
+ * free functions as they're deleted.  When this function returns, the caller is
+ * guaranteed that all the free functions for all the elements have finished
+ * executing.
+ *
+ * While we are flushing, we stop the periodic reaper event from triggering.
+ * Normally, we want to restart this periodic event, but if we are shutting
+ * down the cache we do not want it restarted. hence the restart parameter
+ * where 0 = do not restart reaper and 1 = restart reaper.
+ */
+void
+xfs_mru_cache_flush(
+	xfs_mru_cache_t		*mru,
+	int			restart)
+{
+	if (!mru || !mru->lists)
+		return;
+
+	cancel_rearming_delayed_workqueue(xfs_mru_reap_wq, &mru->work);
+
+	mutex_spinlock(&mru->lock);
+	mru->reap_all = 1;
+	mutex_spinunlock(&mru->lock, 0);
+
+	queue_work(xfs_mru_reap_wq, &mru->work.work);
+	flush_workqueue(xfs_mru_reap_wq);
+
+	mutex_spinlock(&mru->lock);
+	WARN_ON_ONCE(mru->reap_all != 0);
+	mru->reap_all = 0;
+	if (restart)
+		queue_delayed_work(xfs_mru_reap_wq, &mru->work, mru->grp_time);
+	mutex_spinunlock(&mru->lock, 0);
+}
+
+void
+xfs_mru_cache_destroy(
+	xfs_mru_cache_t		*mru)
+{
+	if (!mru || !mru->lists)
+		return;
+
+	/* we don't want the reaper to restart here */
+	xfs_mru_cache_flush(mru, 0);
+
+	kmem_free(mru->lists, mru->grp_count * sizeof(*mru->lists));
+	kmem_free(mru, sizeof(*mru));
+}
+
+/*
+ * To insert an element, call xfs_mru_cache_insert() with the data store, the
+ * element's key and the client data pointer.  This function returns 0 on
+ * success or ENOMEM if memory for the data element couldn't be allocated.
+ */
+int
+xfs_mru_cache_insert(
+	xfs_mru_cache_t	*mru,
+	unsigned long	key,
+	void		*value)
+{
+	xfs_mru_cache_elem_t *elem;
+
+	ASSERT(mru && mru->lists);
+	if (!mru || !mru->lists)
+		return EINVAL;
+
+	elem = kmem_zone_zalloc(xfs_mru_elem_zone, KM_SLEEP);
+	if (!elem)
+		return ENOMEM;
+
+	if (radix_tree_preload(GFP_KERNEL)) {
+		kmem_zone_free(xfs_mru_elem_zone, elem);
+		return ENOMEM;
+	}
+
+	INIT_LIST_HEAD(&elem->list_node);
+	elem->key = key;
+	elem->value = value;
+
+	mutex_spinlock(&mru->lock);
+
+	radix_tree_insert(&mru->store, key, elem);
+	radix_tree_preload_end();
+	_xfs_mru_cache_list_insert(mru, elem);
+
+	mutex_spinunlock(&mru->lock, 0);
+
+	return 0;
+}
+
+/*
+ * To remove an element without calling the free function, call
+ * xfs_mru_cache_remove() with the data store and the element's key.  On success
+ * the client data pointer for the removed element is returned, otherwise this
+ * function will return a NULL pointer.
+ */
+void *
+xfs_mru_cache_remove(
+	xfs_mru_cache_t	*mru,
+	unsigned long	key)
+{
+	xfs_mru_cache_elem_t *elem;
+	void		*value = NULL;
+
+	ASSERT(mru && mru->lists);
+	if (!mru || !mru->lists)
+		return NULL;
+
+	mutex_spinlock(&mru->lock);
+	elem = radix_tree_delete(&mru->store, key);
+	if (elem) {
+		value = elem->value;
+		list_del(&elem->list_node);
+	}
+
+	mutex_spinunlock(&mru->lock, 0);
+
+	if (elem)
+		kmem_zone_free(xfs_mru_elem_zone, elem);
+
+	return value;
+}
+
+/*
+ * To remove and element and call the free function, call xfs_mru_cache_delete()
+ * with the data store and the element's key.
+ */
+void
+xfs_mru_cache_delete(
+	xfs_mru_cache_t	*mru,
+	unsigned long	key)
+{
+	void		*value = xfs_mru_cache_remove(mru, key);
+
+	if (value)
+		mru->free_func(key, value);
+}
+
+/*
+ * To look up an element using its key, call xfs_mru_cache_lookup() with the
+ * data store and the element's key.  If found, the element will be moved to the
+ * head of the MRU list to indicate that it's been touched.
+ *
+ * The internal data structures are protected by a spinlock that is STILL HELD
+ * when this function returns.  Call xfs_mru_cache_done() to release it.  Note
+ * that it is not safe to call any function that might sleep in the interim.
+ *
+ * The implementation could have used reference counting to avoid this
+ * restriction, but since most clients simply want to get, set or test a member
+ * of the returned data structure, the extra per-element memory isn't warranted.
+ *
+ * If the element isn't found, this function returns NULL and the spinlock is
+ * released.  xfs_mru_cache_done() should NOT be called when this occurs.
+ */
+void *
+xfs_mru_cache_lookup(
+	xfs_mru_cache_t	*mru,
+	unsigned long	key)
+{
+	xfs_mru_cache_elem_t *elem;
+
+	ASSERT(mru && mru->lists);
+	if (!mru || !mru->lists)
+		return NULL;
+
+	mutex_spinlock(&mru->lock);
+	elem = radix_tree_lookup(&mru->store, key);
+	if (elem) {
+		list_del(&elem->list_node);
+		_xfs_mru_cache_list_insert(mru, elem);
+	}
+	else
+		mutex_spinunlock(&mru->lock, 0);
+
+	return elem ? elem->value : NULL;
+}
+
+/*
+ * To look up an element using its key, but leave its location in the internal
+ * lists alone, call xfs_mru_cache_peek().  If the element isn't found, this
+ * function returns NULL.
+ *
+ * See the comments above the declaration of the xfs_mru_cache_lookup() function
+ * for important locking information pertaining to this call.
+ */
+void *
+xfs_mru_cache_peek(
+	xfs_mru_cache_t	*mru,
+	unsigned long	key)
+{
+	xfs_mru_cache_elem_t *elem;
+
+	ASSERT(mru && mru->lists);
+	if (!mru || !mru->lists)
+		return NULL;
+
+	mutex_spinlock(&mru->lock);
+	elem = radix_tree_lookup(&mru->store, key);
+	if (!elem)
+		mutex_spinunlock(&mru->lock, 0);
+
+	return elem ? elem->value : NULL;
+}
+
+/*
+ * To release the internal data structure spinlock after having performed an
+ * xfs_mru_cache_lookup() or an xfs_mru_cache_peek(), call xfs_mru_cache_done()
+ * with the data store pointer.
+ */
+void
+xfs_mru_cache_done(
+	xfs_mru_cache_t	*mru)
+{
+	mutex_spinunlock(&mru->lock, 0);
+}