VFS: hot data tracking
cover VFS: hot data tracking
From: Zhi Yong Wu <[email protected]>
HI, folks
I have pushed the patchset to my kernel dev git tree:
[email protected]:wuzhy/kernel.git
Also, you can review it via
https://github.com/wuzhy/kernel/commits/hottrack
NOTE:
The patchset still has a lot of bugfix and cleanup to do. It is post out mainly to make sure it is going in the correct direction and hope to get some helpful comments from other guys.
TODO List:
1.) Need to do scalability or performance tests.
2.) Fix up bugs.
3.) Strictly split this patchset to keep them in order
This patchset is in RFC state, i haven't strictly split it
When it is in PATCH state, i will strictly split it and let them in order.
4.) Turn some Micro in to tunables
TIME_TO_KICK, and HEAT_UPDATE_DELAY
5.) Rafactor hot_hash_is_aging()
If you just made the timeout value a timespec and compared
the _timespecs_, you would be doing a lot fewer conversions.
6.) Cleanup some unnecessary lock protect
7.) Add more comments to explain how to calc temperature
Ben Chociej, Matt Lupfer and Conor Scott originally wrote this code to
be very btrfs-specific. I've taken their code and attempted to
make it more generic and integrate it at the VFS level.
INTRODUCTION:
Essentially, this means maintaining some key stats
(like number of reads/writes, last read/write time, frequency of reads/writes), then distilling those numbers down to a single "temperature" value that reflects what data is "hot," and using that temperature to move data to SSDs.
The long-term goal of these patches is to allow some FSs, e.g. Btrfs to intelligently utilize SSDs in a heterogenous volume. Incidentally, this project has been motivated by
the Project Ideas page on the Btrfs wiki.
Of course, users are warned not to run this code outside of development environments. These patches are EXPERIMENTAL, and as such they might eat your data and/or memory. That said, the code should be relatively safe when the hottrack mount option are disabled.
MOTIVATION:
The overall goal of enabling hot data relocation to SSD has been motivated by the Project Ideas page on the Btrfs wiki at
<https://btrfs.wiki.kernel.org/index.php/Project_ideas>.
It will divide into two steps. VFS provide hot data tracking function while specific FS will provide hot data relocation function. So as the first step of this goal, it is hoped that the patchset for hot data tracking will eventually mature into VFS.
This is essentially the traditional cache argument: SSD is fast and expensive; HDD is cheap but slow. ZFS, for example, can already take advantage of SSD caching. Btrfs should also be able to take advantage of hybrid storage without many broad, sweeping changes to existing code.
SUMMARY:
- Hooks in existing vfs functions to track data access frequency
- New rbtrees for tracking access frequency of inodes and sub-file ranges (hot_rb.c)
The relationship between super_block and rbtree is as below:
super_block->s_hotinfo.hot_inode_tree
In include/linux/fs.h, one struct hot_info s_hotinfo is added to
super_block struct. Each FS instance can find hot tracking info s_hotinfo via its super_block. In this hot_info, it store a lot of hot tracking info such as hot_inode_tree, inode and range hash list, etc.
- A hash list for indexing data by its temperature (hot_hash.c)
- A debugfs interface for dumping data from the rbtrees (hot_debugfs.c)
- A background kthread for updating inode heat info
- Mount options for enabling temperature tracking(-o hottrack, default mean disabled) (hot_track.c)
- An ioctl to retrieve the frequency information collected for a certain file
- Ioctls to enable/disable frequency tracking per inode.
Usage syntax:
root@debian-i386:~# mount -o hottrack /dev/sdb /mnt [ 1505.894078] device label test devid 1 transid 29 /dev/sdb [ 1505.952977] btrfs: disk space caching is enabled [ 1506.069678] vfs: turning on hot data tracking root@debian-i386:~# mount -t debugfs none /sys/kernel/debug root@debian-i386:~# ls -l /sys/kernel/debug/vfs_hotdata/ total 0 drwxr-xr-x 2 root root 0 Aug 8 04:40 sdb root@debian-i386:~# ls -l /sys/kernel/debug/vfs_hotdata/sdb total 0 -rw-r--r-- 1 root root 0 Aug 8 04:40 inode_data -rw-r--r-- 1 root root 0 Aug 8 04:40 range_data root@debian-i386:~# vi /mnt/file root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/inode_data inode #279, reads 0, writes 1, avg read time 18446744073709551615, avg write time 5251566408153596, temp 109 root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/range_data inode #279, range start 0 (range len 1048576) reads 0, writes 1, avg read time 18446744073709551615, avg write time 1128690176623144209, temp 64 root@debian-i386:~# echo "hot data tracking test" >> /mnt/file root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/inode_data inode #279, reads 0, writes 2, avg read time 18446744073709551615, avg write time 4923343766042451, temp 109 root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/range_data inode #279, range start 0 (range len 1048576) reads 0, writes 2, avg read time 18446744073709551615, avg write time 1058147040842596150, temp 64 root@debian-i386:~#
Zhi Yong Wu (11):
vfs: introduce one structure hot_info
vfs: introduce one rb tree - hot_inode_tree
vfs: introduce 2 rb tree items - inode and range
vfs: add support for updating access frequency
vfs: add one new mount option -o hottrack
vfs: add init and exit support
vfs: introduce one hash table
vfs: enable hot data tracking
vfs: fork one private kthread to update temperature info
vfs: add 3 new ioctl interfaces
vfs: add debugfs support
fs/Makefile | 3 +- fs/compat_ioctl.c | 8 + fs/dcache.c | 2 + fs/direct-io.c | 10 + fs/hot_debugfs.c | 488 ++++++++++++++++++++++++++++++++++ fs/hot_debugfs.h | 60 +++++ fs/hot_hash.c | 382 ++++++++++++++++++++++++++ fs/hot_hash.h | 112 ++++++++ fs/hot_rb.c | 648 +++++++++++++++++++++++++++++++++++++++++++++ fs/hot_rb.h | 81 ++++++ fs/hot_track.c | 85 ++++++ fs/hot_track.h | 23 ++ fs/ioctl.c | 132 +++++++++ fs/namespace.c | 10 + fs/super.c | 11 + include/linux/fs.h | 15 + include/linux/hot_track.h | 169 ++++++++++++ mm/filemap.c | 8 + mm/page-writeback.c | 21 ++ mm/readahead.c | 9 + 20 files changed, 2276 insertions(+), 1 deletions(-) create mode 100644 fs/hot_debugfs.c create mode 100644 fs/hot_debugfs.h create mode 100644 fs/hot_hash.c create mode 100644 fs/hot_hash.h create mode 100644 fs/hot_rb.c create mode 100644 fs/hot_rb.h create mode 100644 fs/hot_track.c create mode 100644 fs/hot_track.h create mode 100644 include/linux/hot_track.h
Signature · 1 line
1.7.6.5
RFC 1/11 vfs: introduce one structure hot_info
From: Zhi Yong Wu <[email protected]>
One root structure hot_info is defined, is hooked
up in super_block, and will be used to hold rb trees root, hash list root and some other information, etc.
--- include/linux/fs.h | 4 ++++ include/linux/hot_track.h | 26 ++++++++++++++++++++++++++ 2 files changed, 30 insertions(+), 0 deletions(-) create mode 100644 include/linux/hot_track.h
include/linux/fs.h +4 -0
@@ -415,6 +415,7 @@ struct inodes_stat_t {#include <linux/migrate_mode.h> #include <linux/uidgid.h> #include <linux/lockdep.h> +#include <linux/hot_track.h> #include <asm/byteorder.h>
@@ -1578,6 +1579,9 @@ struct super_block {/* Being remounted read-only */ int s_readonly_remount; + +/* Hot data tracking info*/ +struct hot_info s_hotinfo; };/* superblock cache pruning functions */
include/linux/hot_track.h new file +26 -0
@@ -0,0 +1,26 @@ +/* +* include/linux/hot_track.h +* +* This file has definitions for VFS hot data tracking +* structures etc. +* +* Copyright (C) 2012 IBM Corp. All rights reserved. +* Written by Zhi Yong Wu <[email protected]> +* Ben Chociej <[email protected]> +* +* This program is free software; you can redistribute it and/or +* modify it under the terms of the GNU General Public +* License v2 as published by the Free Software Foundation. +*/ + +#ifndef _LINUX_HOTTRACK_H +#define _LINUX_HOTTRACK_H + +#include <linux/types.h> +#include <linux/rbtree.h> +#include <linux/kref.h> + +struct hot_info { +}; + +#endif /* _LINUX_HOTTRACK_H */
Signature · 1 line
1.7.6.5
RFC 2/11 vfs: introduce one rb tree - hot_inode_tree
In reply to [email protected]From: Zhi Yong Wu <[email protected]>
Adds hot_inode_tree struct to keep track of
frequently accessed files, and be keyed by {inode, offset}. Trees contain hot_inode_items representing those files and ranges.
Having these trees means that vfs can quickly determine the
temperature of some data by doing some calculations on the hot_freq_data struct that hangs off of the tree item.
--- fs/Makefile | 3 ++- fs/hot_rb.c | 30 ++++++++++++++++++++++++++++++ fs/hot_rb.h | 21 +++++++++++++++++++++ include/linux/hot_track.h | 9 +++++++++ 4 files changed, 62 insertions(+), 1 deletions(-) create mode 100644 fs/hot_rb.c create mode 100644 fs/hot_rb.h
fs/Makefile +2 -1
@@ -11,7 +11,8 @@ obj-y := open.o read_write.o file_table.o super.o \ attr.o bad_inode.o file.o filesystems.o namespace.o \ seq_file.o xattr.o libfs.o fs-writeback.o \ pnode.o drop_caches.o splice.o sync.o utimes.o \ - stack.o fs_struct.o statfs.o + stack.o fs_struct.o statfs.o \ + hot_rb.o ifeq ($(CONFIG_BLOCK),y) obj-y += buffer.o bio.o block_dev.o direct-io.o mpage.o ioprio.o
fs/hot_rb.c new file +30 -0
@@ -0,0 +1,30 @@ +/* +* fs/hot_rb.c +* +* Copyright (C) 2012 IBM Corp. All rights reserved. +* Written by Zhi Yong Wu <[email protected]> +* Ben Chociej <[email protected]> +* +* This program is free software; you can redistribute it and/or +* modify it under the terms of the GNU General Public +* License v2 as published by the Free Software Foundation. +*/ + +#include <linux/err.h> +#include <linux/slab.h> +#include <linux/module.h> +#include <linux/spinlock.h> +#include <linux/hardirq.h> +#include <linux/blkdev.h> +#include "hot_rb.h" +#include "hot_hash.h" + +/* +* Initialize the inode tree. Should be called for each new inode +* access or other user of the hot_inode interface. +*/ +void hot_rb_inode_tree_init (struct hot_inode_tree *tree) +{ + tree->map =RB_ROOT ; +rwlock_init (&tree->lock); +}
fs/hot_rb.h new file +21 -0
@@ -0,0 +1,21 @@ +/* +* fs/hot_rb.h +* +* Copyright (C) 2012 IBM Corp. All rights reserved. +* Written by Zhi Yong Wu <[email protected]> +* Ben Chociej <[email protected]> +* +* This program is free software; you can redistribute it and/or +* modify it under the terms of the GNU General Public +* License v2 as published by the Free Software Foundation. +*/ + +#ifndef __HOT_MAP__ +#define __HOT_MAP__ + +#include <linux/rbtree.h> +#include <linux/hot_track.h> + +void hot_rb_inode_tree_init (struct hot_inode_tree *tree); + +#endif /* __HOT_MAP__ */
include/linux/hot_track.h +9 -0
@@ -20,7 +20,16 @@#include <linux/rbtree.h> #include <linux/kref.h> +/* A tree that sits on the hot_info */ +struct hot_inode_tree { +struct rb_root map; +rwlock_t lock; +}; +struct hot_info { + +/* red-black tree that keeps track of fs-wide hot data */ +struct hot_inode_tree hot_inode_tree; };#endif /* _LINUX_HOTTRACK_H */
Signature · 1 line
1.7.6.5
RFC 3/11 vfs: introduce 2 rb tree items - inode and range
In reply to [email protected]From: Zhi Yong Wu <[email protected]>
Define two items hot_inode_item and hot_range_item,
one of them represents one tracked file
to keep track of its access frequency and the tree of ranges in this file, while the latter represents a file range of one inode.
Each of the two structures contains a hot_freq_data
struct with its frequency of access metrics (number of {reads, writes}, last {read,write} time, frequency of {reads,writes}).
Also, each hot_inode_item contains one hot_range_tree
struct which is keyed by {inode, offset, length} and used to keep track of all the ranges in this file.
--- fs/Makefile | 2 +- fs/dcache.c | 2 + fs/hot_rb.c | 74 +++++++++++++++++++++++++++++++++++++++++++++ fs/hot_rb.h | 9 +++++ fs/hot_track.c | 26 ++++++++++++++++ fs/hot_track.h | 20 ++++++++++++ include/linux/hot_track.h | 62 +++++++++++++++++++++++++++++++++++++ 7 files changed, 194 insertions(+), 1 deletions(-) create mode 100644 fs/hot_track.c create mode 100644 fs/hot_track.h
fs/Makefile +1 -1
@@ -12,7 +12,7 @@ obj-y := open.o read_write.o file_table.o super.o \ seq_file.o xattr.o libfs.o fs-writeback.o \ pnode.o drop_caches.o splice.o sync.o utimes.o \ stack.o fs_struct.o statfs.o \ - hot_rb.o + hot_rb.o hot_track.o ifeq ($(CONFIG_BLOCK),y) obj-y += buffer.o bio.o block_dev.o direct-io.o mpage.o ioprio.o
fs/dcache.c +2 -0
@@ -39,6 +39,7 @@#include <linux/ratelimit.h> #include "internal.h" #include "mount.h" +#include "hot_track.h" /* * Usage:
@@ -3164,6 +3165,7 @@ void __init vfs_caches_init(unsigned long mempages)inode_init ();files_init (mempages);mnt_init (); +hot_track_item_cache_init ();bdev_cache_init ();chrdev_init (); }
fs/hot_rb.c +74 -0
83 lines not on this page · raw message
fs/hot_rb.h +9 -0
@@ -18,4 +18,13 @@void hot_rb_inode_tree_init (struct hot_inode_tree *tree); +/* values for hot_freq_data flags */ +#define FREQ_DATA_TYPE_INODE (1 << 0)/* freq data struct is for an inode */ +#define FREQ_DATA_TYPE_RANGE (1 << 1)/* freq data struct is for a range */ + +void hot_rb_inode_item_init (void *_item); +void hot_rb_range_item_init (void *_item); + +int __init hot_rb_item_cache_init (void ); +#endif /* __HOT_MAP__ */
fs/hot_track.c new file +26 -0
@@ -0,0 +1,26 @@ +/* +* fs/hot_track.c +* +* Copyright (C) 2012 IBM Corp. All rights reserved. +* Written by Zhi Yong Wu <[email protected]> +* Ben Chociej <[email protected]> +* +* This program is free software; you can redistribute it and/or +* modify it under the terms of the GNU General Public +* License v2 as published by the Free Software Foundation. +*/ + +#include <linux/slab.h> +#include <linux/gfp.h> +#include <linux/parser.h> +#include "hot_track.h" + +/* +* Initialize hot_inode_item, hot_range_item +* and hot_hash_node kmem cache +*/ +void __init hot_track_item_cache_init (void ) +{ +if (hot_rb_item_cache_init ()) +return ; +}
fs/hot_track.h new file +20 -0
@@ -0,0 +1,20 @@ +/* +* fs/hot_track.h +* +* Copyright (C) 2012 IBM Corp. All rights reserved. +* Written by Zhi Yong Wu <[email protected]> +* Ben Chociej <[email protected]> +* +* This program is free software; you can redistribute it and/or +* modify it under the terms of the GNU General Public +* License v2 as published by the Free Software Foundation. +*/ + +#ifndef __HOT_TRACK__ +#define __HOT_TRACK__ + +#include "hot_rb.h" + +void __init hot_track_item_cache_init (void ); + +#endif /* __HOT_TRACK__ */
include/linux/hot_track.h +62 -0
68 lines not on this page · raw message
Signature · 1 line
1.7.6.5
RFC 4/11 vfs: add support for updating access frequency
In reply to [email protected]From: Zhi Yong Wu <[email protected]>
Add some utils helpers to update access frequencies
for one file or its range.
--- fs/hot_rb.c | 359 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ fs/hot_rb.h | 28 +++++ 2 files changed, 387 insertions(+), 0 deletions(-)
fs/hot_rb.c +359 -0
@@ -102,3 +102,362 @@ void hot_rb_range_item_init(void *_item) hr->hot_freq_data.avg_delta_writes = (u64) -1 ; hr->hot_freq_data.flags =FREQ_DATA_TYPE_RANGE ; } + +/* +* Drops the reference out on hot_inode_item by one and free the structure +* if the reference count hits zero +*/ +void hot_rb_free_hot_inode_item (struct hot_inode_item *he) +{ +if (!he) +return ; + +if (atomic_dec_and_test (&he->refs.refcount)) { +WARN_ON (he->in_tree); +kmem_cache_free (hot_inode_item_cache, he); + } +} + +/* +* Drops the reference out on hot_range_item by one and free the structure +* if the reference count hits zero +*/ +void hot_rb_free_hot_range_item (struct hot_range_item *hr) +{ +if (!hr) +return ; + +if (atomic_dec_and_test (&hr->refs.refcount)) { +WARN_ON (hr->in_tree); +kmem_cache_free (hot_range_item_cache, hr); + } +} + +static struct rb_node *hot_rb_insert_hot_inode_item (struct rb_root *root, +unsigned long inode_num, +struct rb_node *node) +{ +struct rb_node **p = &root->rb_node; +struct rb_node *parent =NULL ; +struct hot_inode_item *entry; + +/* walk tree to find insertion point */ +while (*p) { + parent = *p; + entry = rb_entry(parent, struct hot_inode_item, rb_node); + +if (inode_num < entry->i_ino) + p = &(*p)->rb_left; +else if (inode_num > entry->i_ino) + p = &(*p)->rb_right; +else +return parent; + } + + entry = rb_entry(node, struct hot_inode_item, rb_node); + entry->in_tree =1 ; +rb_link_node (node, parent, p); +rb_insert_color (node, root); + +return NULL ; +} + +static u64 hot_rb_range_end (struct hot_range_item *hr) +{ +if (hr->start + hr->len < hr->start) +return (u64)-1 ; + +return hr->start + hr->len -1 ; +} + +static struct rb_node *hot_rb_insert_hot_range_item (struct rb_root *root, +u64 start, +struct rb_node *node) +{ +struct rb_node **p = &root->rb_node; +struct rb_node *parent =NULL ; +struct hot_range_item *entry; + +/* ensure start is on a range boundary */ + start = start &RANGE_SIZE_MASK ; +/* walk tree to find insertion point */ +while (*p) { + parent = *p; + entry = rb_entry(parent, struct hot_range_item, rb_node); + +if (start < entry->start) + p = &(*p)->rb_left; +else if (start >=hot_rb_range_end (entry)) + p = &(*p)->rb_right; +else +return parent; + } + + entry = rb_entry(node, struct hot_range_item, rb_node); + entry->in_tree =1 ; +rb_link_node (node, parent, p); +rb_insert_color (node, root); + +return NULL ; +} + +/* +* Add a hot_inode_item to a hot_inode_tree. If the tree already contains +* an item with the index given, return -EEXIST +*/ +int hot_rb_add_hot_inode_item (struct hot_inode_tree *tree, +struct hot_inode_item *he) +{ +int ret =0 ; +struct rb_node *rb; + + rb =hot_rb_insert_hot_inode_item ( + &tree->map, he->i_ino, &he->rb_node); +if (rb) { + ret = -EEXIST ; + goto out; + } + +kref_get (&he->refs); + +out: +return ret; +} + +/* +* Add a hot_range_item to a hot_range_tree. If the tree already contains +* an item with the index given, return -EEXIST +* +* Also optionally aggresively merge ranges (currently disabled) +*/ +int hot_rb_add_hot_range_item (struct hot_range_tree *tree, +struct hot_range_item *hr) +{ +int ret =0 ; +struct rb_node *rb; + + rb =hot_rb_insert_hot_range_item ( + &tree->map, hr->start, &hr->rb_node); +if (rb) { + ret = -EEXIST ; + goto out; + } + +kref_get (&hr->refs); + +out: +return ret; +} + +/* +* Lookup a hot_inode_item in the hot_inode_tree with the given index +* (inode_num) +*/ +struct hot_inode_item +*hot_rb_lookup_hot_inode_item (struct hot_inode_tree *tree, +unsigned long inode_num) +{ +struct rb_node **p = &(tree->map.rb_node); +struct rb_node *parent =NULL ; +struct hot_inode_item *entry; + +while (*p) { + parent = *p; + entry = rb_entry(parent, struct hot_inode_item, rb_node); + +if (inode_num < entry->i_ino) + p = &(*p)->rb_left; +else if (inode_num > entry->i_ino) + p = &(*p)->rb_right; +else { +kref_get (&entry->refs); +return entry; + } + } + +return NULL ; +} + +/* +* Lookup a hot_range_item in a hot_range_tree with the given index +* (start, offset) +*/ +struct hot_range_item +*hot_rb_lookup_hot_range_item (struct hot_range_tree *tree, +u64 start) +{ +struct rb_node **p = &(tree->map.rb_node); +struct rb_node *parent =NULL ; +struct hot_range_item *entry; + +/* ensure start is on a range boundary */ + start = start &RANGE_SIZE_MASK ; +while (*p) { + parent = *p; + entry = rb_entry(parent, struct hot_range_item, rb_node); + +if (start < entry->start) + p = &(*p)->rb_left; +else if (start >hot_rb_range_end (entry)) + p = &(*p)->rb_right; +else { +kref_get (&entry->refs); +return entry; + } + } + +return NULL ; +} + +/* Update inode frequency struct */ +static struct hot_inode_item *hot_rb_update_inode_freq (struct inode *inode, +int rw) +{ +struct hot_info *root = &(inode->i_sb->s_hotinfo); +struct hot_inode_tree *hitree = &(root->hot_inode_tree); +struct hot_inode_item *he; + +read_lock (&hitree->lock); + he =hot_rb_lookup_hot_inode_item (hitree, inode->i_ino); +read_unlock (&hitree->lock); + +if (!he) { + he =kmem_cache_alloc (hot_inode_item_cache, +GFP_KERNEL |GFP_NOFS ); +if (!he) + goto out; + +write_lock (&hitree->lock); +hot_rb_inode_item_init (he); + he->i_ino = inode->i_ino; +hot_rb_add_hot_inode_item (hitree, he); +write_unlock (&hitree->lock); + } + +spin_lock (&he->lock); +hot_rb_update_freq (&he->hot_freq_data, rw); +spin_unlock (&he->lock); + +out: +return he; +} + +/* Update range frequency struct */ +static bool hot_rb_update_range_freq (struct hot_inode_item *he, +u64 off,u64 len,int rw, +struct hot_info *root) +{ +struct hot_range_tree *hrtree = &(he->hot_range_tree); +struct hot_range_item *hr =NULL ; +u64 start_off = off &RANGE_SIZE_MASK ; +u64 end_off = (off + len -1 ) &RANGE_SIZE_MASK ; +u64 cur; +int ret = true; + +if (len ==0 ) +return false; + +/* +* Align ranges on RANGE_SIZE boundary to prevent proliferation +* of range structs +*/ +for (cur = start_off; cur <= end_off; cur +=RANGE_SIZE ) { +read_lock (&hrtree->lock); + hr =hot_rb_lookup_hot_range_item (hrtree, cur); +read_unlock (&hrtree->lock); + +if (!hr) { + hr =kmem_cache_alloc (hot_range_item_cache, +GFP_KERNEL |GFP_NOFS ); +if (!hr) { + ret = false; + goto out; + } + +write_lock (&hrtree->lock); +hot_rb_range_item_init (hr); + hr->start = cur &RANGE_SIZE_MASK ; + hr->len =RANGE_SIZE ; + hr->hot_inode = he; +hot_rb_add_hot_range_item (hrtree, hr); +write_unlock (&hrtree->lock); + } + +spin_lock (&hr->lock); +hot_rb_update_freq (&hr->hot_freq_data, rw); +spin_unlock (&hr->lock); +hot_rb_free_hot_range_item (hr); + } + +out: +return ret; +} + +/* +* This function does the actual work of updating the frequency numbers, +* whatever they turn out to be. FREQ_POWER determines how many atime +* deltas we keep track of (as a power of 2). So, setting it to anything above +* 16ish is probably overkill. Also, the higher the power, the more bits get +* right shifted out of the timestamp, reducing precision, so take note of that +* as well. +* +* The caller should have already locked freq_data's parent's spinlock. +* +* FREQ_POWER, defined immediately below, determines how heavily to weight +* the current frequency numbers against the newest access. For example, a value +* of 4 means that the new access information will be weighted 1/16th (ie 2^-4) +* as heavily as the existing frequency info. In essence, this is a kludged- +* together version of a weighted average, since we can't afford to keep all of +* the information that it would take to get a _real_ weighted average. +*/ +void hot_rb_update_freq (struct hot_freq_data *freq_data,int rw) +{ +struct timespec old_atime; +struct timespec current_time; +struct timespec delta_ts; +u64 new_avg; +u64 new_delta; + +if (unlikely (rw)) { + old_atime = freq_data->last_write_time; + freq_data->nr_writes +=1 ; + new_avg = freq_data->avg_delta_writes; + }else { + old_atime = freq_data->last_read_time; + freq_data->nr_reads +=1 ; + new_avg = freq_data->avg_delta_reads; + } + + current_time =current_kernel_time (); + delta_ts =timespec_sub (current_time, old_atime); + new_delta =timespec_to_ns (&delta_ts) >>FREQ_POWER ; + + new_avg = (new_avg <<FREQ_POWER ) - new_avg + new_delta; + new_avg = new_avg >>FREQ_POWER ; + +if (unlikely (rw)) { + freq_data->last_write_time = current_time; + freq_data->avg_delta_writes = new_avg; + }else { + freq_data->last_read_time = current_time; + freq_data->avg_delta_reads = new_avg; + } +} + +/* main function to update access frequency from read/writepage(s) hooks */ +void hot_rb_update_freqs (struct inode *inode,u64 start, +u64 len,int rw) +{ +struct hot_inode_item *he; + + he =hot_rb_update_inode_freq (inode, rw); + +WARN_ON (!he); + +if (he) { +hot_rb_update_range_freq (he, start, len, + rw, &(inode->i_sb->s_hotinfo)); + +hot_rb_free_hot_inode_item (he); + } +}
fs/hot_rb.h +28 -0
@@ -21,10 +21,38 @@ void hot_rb_inode_tree_init(struct hot_inode_tree *tree);/* values for hot_freq_data flags */ #define FREQ_DATA_TYPE_INODE (1 << 0)/* freq data struct is for an inode */ #define FREQ_DATA_TYPE_RANGE (1 << 1)/* freq data struct is for a range */ +/* size of sub-file ranges */ +#define RANGE_SIZE (1<<20) +#define RANGE_SIZE_MASK (~((u64)(RANGE_SIZE - 1))) + +#define FREQ_POWER 4 + +struct hot_info ; +struct inode ;void hot_rb_inode_item_init (void *_item);void hot_rb_range_item_init (void *_item); +struct hot_range_item +*hot_rb_lookup_hot_range_item (struct hot_range_tree *tree, +u64 start); + +struct hot_inode_item +*hot_rb_lookup_hot_inode_item (struct hot_inode_tree *tree, +unsigned long inode_num); + +int hot_rb_add_hot_inode_item (struct hot_inode_tree *tree, +struct hot_inode_item *he); +int hot_rb_add_hot_range_item (struct hot_range_tree *tree, +struct hot_range_item *hr); + +void hot_rb_free_hot_inode_item (struct hot_inode_item *he); +void hot_rb_free_hot_range_item (struct hot_range_item *hr); +int __init hot_rb_item_cache_init (void ); +void hot_rb_update_freq (struct hot_freq_data *freq_data,int rw); +void hot_rb_update_freqs (struct inode *inode,u64 start,u64 len, +int rw); +#endif /* __HOT_MAP__ */
Signature · 1 line
1.7.6.5
RFC 5/11 vfs: add one new mount option -o hottrack
In reply to [email protected]From: Zhi Yong Wu <[email protected]>
Introduce one new mount option '-o hottrack',
and add its parsing support.
Its usage looks like:
mount -o hottrack
mount -o nouser,hottrack
mount -o nouser,hottrack,loop
mount -o hottrack,nouser
--- fs/hot_track.c | 34 ++++++++++++++++++++++++++++++++++ fs/hot_track.h | 1 + fs/super.c | 5 +++++ include/linux/hot_track.h | 6 ++++++ 4 files changed, 46 insertions(+), 0 deletions(-)
fs/hot_track.c +34 -0
@@ -16,6 +16,40 @@#include "hot_track.h" /* +* Regular mount options parser for -hottrack option. +* return false if no -hottrack is specified; +* otherwise return true. And the -hottrack will be +* removed from options. +*/ +bool hot_track_parse_options (char *options) +{ +long len; +char *p; +static char opts_hot[] ="hottrack" ; + +if (!options) +return false; + + p =strstr (options, opts_hot); +if (!p) +return false; + +while (p) { + len = options +strlen (options) - (p +strlen (opts_hot)); +if (len ==0 ) { + options[0 ] ='\0' ; +break ; + } + +memmove (p, p +strlen (opts_hot) +1 , len); + p =strstr (options, opts_hot); + } + +printk (KERN_INFO "vfs: turning on hot data tracking\n" ); +return true; +} + +/* * Initialize hot_inode_item, hot_range_item * and hot_hash_node kmem cache */
fs/hot_track.h +1 -0
@@ -15,6 +15,7 @@#include "hot_rb.h" +bool hot_track_parse_options (char *options);void __init hot_track_item_cache_init (void );#endif /* __HOT_TRACK__ */
fs/super.c +5 -0
@@ -35,6 +35,7 @@#include <linux/fsnotify.h> #include <linux/lockdep.h> #include "internal.h" +#include "hot_track.h" LIST_HEAD (super_blocks);
@@ -1125,6 +1126,7 @@ mount_fs(struct file_system_type *type, int flags, const char *name, void *data)struct dentry *root;struct super_block *sb;char *secdata =NULL ; +bool hottrack = false;int error = -ENOMEM ;if (data && !(type->fs_flags &FS_BINARY_MOUNTDATA )) {
@@ -1137,6 +1139,9 @@ mount_fs(struct file_system_type *type, int flags, const char *name, void *data) goto out_free_secdata; } +if (data &&hot_track_parse_options (data)) + hottrack = true; + root = type->mount (type, flags, name, data);if (IS_ERR (root)) { error =PTR_ERR (root);
include/linux/hot_track.h +6 -0
@@ -20,6 +20,11 @@#include <linux/rbtree.h> #include <linux/kref.h> +/* +* Flags for hot data tracking mount options. +*/ +#define HOT_MOUNT_HOT_TRACK (1 << 0) +/* A tree that sits on the hot_info */ struct hot_inode_tree {struct rb_root map;
@@ -89,6 +94,7 @@ struct hot_range_item { };struct hot_info { +unsigned long mount_opt;/* red-black tree that keeps track of fs-wide hot data */ struct hot_inode_tree hot_inode_tree;
Signature · 1 line
1.7.6.5
Sorry, forgot CCed to Ted.
Quoted 176 lines · On Tue, Sep 11, 2012 at 10:27 PM, <[email protected]> wrote:
From: Zhi Yong Wu <[email protected]>
HI, folks
I have pushed the patchset to my kernel dev git tree:
[email protected]:wuzhy/kernel.gitAlso, you can review it via
https://github.com/wuzhy/kernel/commits/hottrackNOTE:
The patchset still has a lot of bugfix and cleanup to do. It is post out mainly to make sure it is going in the correct direction and hope to get some helpful comments from other guys.
TODO List:
1.) Need to do scalability or performance tests.
2.) Fix up bugs.
3.) Strictly split this patchset to keep them in order
This patchset is in RFC state, i haven't strictly split it
When it is in PATCH state, i will strictly split it and let them in order.
4.) Turn some Micro in to tunables
TIME_TO_KICK, andHEAT_UPDATE_DELAY
5.) Rafactorhot_hash_is_aging()
If you just made the timeout value a timespec and compared
the _timespecs_, you would be doing a lot fewer conversions.
6.) Cleanup some unnecessary lock protect
7.) Add more comments to explain how to calc temperatureBen Chociej, Matt Lupfer and Conor Scott originally wrote this code to
be very btrfs-specific. I've taken their code and attempted to
make it more generic and integrate it at the VFS level.INTRODUCTION:
Essentially, this means maintaining some key stats
(like number of reads/writes, last read/write time, frequency of reads/writes), then distilling those numbers down to a single "temperature" value that reflects what data is "hot," and using that temperature to move data to SSDs.The long-term goal of these patches is to allow some FSs, e.g. Btrfs to intelligently utilize SSDs in a heterogenous volume. Incidentally, this project has been motivated by
the Project Ideas page on the Btrfs wiki.Of course, users are warned not to run this code outside of development environments. These patches are EXPERIMENTAL, and as such they might eat your data and/or memory. That said, the code should be relatively safe when the hottrack mount option are disabled.
MOTIVATION:
The overall goal of enabling hot data relocation to SSD has been motivated by the Project Ideas page on the Btrfs wiki at
<https://btrfs.wiki.kernel.org/index.php/Project_ideas>.
It will divide into two steps. VFS provide hot data tracking function while specific FS will provide hot data relocation function. So as the first step of this goal, it is hoped that the patchset for hot data tracking will eventually mature into VFS.This is essentially the traditional cache argument: SSD is fast and expensive; HDD is cheap but slow. ZFS, for example, can already take advantage of SSD caching. Btrfs should also be able to take advantage of hybrid storage without many broad, sweeping changes to existing code.
SUMMARY:
- Hooks in existing vfs functions to track data access frequency
- New rbtrees for tracking access frequency of inodes and sub-file ranges (
hot_rb.c)
The relationship betweensuper_blockand rbtree is as below:
super_block->s_hotinfo.hot_inode_tree
Ininclude/linux/fs.h, one structhot_infos_hotinfois added to
super_blockstruct. Each FS instance can find hot tracking infos_hotinfovia itssuper_block. In thishot_info, it store a lot of hot tracking info such ashot_inode_tree, inode and range hash list, etc.- A hash list for indexing data by its temperature (
hot_hash.c)- A debugfs interface for dumping data from the rbtrees (
hot_debugfs.c)- A background kthread for updating inode heat info
- Mount options for enabling temperature
tracking(-o hottrack, default mean disabled)(hot_track.c)- An ioctl to retrieve the frequency information collected for a certain file
- Ioctls to enable/disable frequency tracking per inode.
Usage syntax:
root@debian-i386:~# mount -o hottrack /dev/sdb /mnt [ 1505.894078] device label test devid 1 transid 29 /dev/sdb [ 1505.952977] btrfs: disk space caching is enabled [ 1506.069678] vfs: turning on hot data tracking root@debian-i386:~# mount -t debugfs none /sys/kernel/debug root@debian-i386:~# ls -l /sys/kernel/debug/vfs_hotdata/ total 0 drwxr-xr-x 2 root root 0 Aug 8 04:40 sdb root@debian-i386:~# ls -l /sys/kernel/debug/vfs_hotdata/sdb total 0(quoted diff) +0
-2-rw-r--r-- 1 root root 0 Aug 8 04:40 inode_data -rw-r--r-- 1 root root 0 Aug 8 04:40 range_dataroot@debian-i386:~# vi /mnt/file root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/inode_data inode #279, reads 0, writes 1, avg read time 18446744073709551615, avg write time 5251566408153596, temp 109 root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/range_data inode #279, range start 0 (range len 1048576) reads 0, writes 1, avg read time 18446744073709551615, avg write time 1128690176623144209, temp 64 root@debian-i386:~# echo "hot data tracking test" >> /mnt/file root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/inode_data inode #279, reads 0, writes 2, avg read time 18446744073709551615, avg write time 4923343766042451, temp 109 root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/range_data inode #279, range start 0 (range len 1048576) reads 0, writes 2, avg read time 18446744073709551615, avg write time 1058147040842596150, temp 64 root@debian-i386:~#Zhi Yong Wu (11):
vfs: introduce one structurehot_info
vfs: introduce one rb tree -hot_inode_tree
vfs: introduce 2 rb tree items - inode and range
vfs: add support for updating access frequency
vfs: add one new mount option -o hottrack
vfs: add init and exit support
vfs: introduce one hash table
vfs: enable hot data tracking
vfs: fork one private kthread to update temperature info
vfs: add 3 new ioctl interfaces
vfs: add debugfs supportfs/Makefile | 3 +- fs/compat_ioctl.c | 8 + fs/dcache.c | 2 + fs/direct-io.c | 10 + fs/hot_debugfs.c | 488 ++++++++++++++++++++++++++++++++++ fs/hot_debugfs.h | 60 +++++ fs/hot_hash.c | 382 ++++++++++++++++++++++++++ fs/hot_hash.h | 112 ++++++++ fs/hot_rb.c | 648 +++++++++++++++++++++++++++++++++++++++++++++ fs/hot_rb.h | 81 ++++++ fs/hot_track.c | 85 ++++++ fs/hot_track.h | 23 ++ fs/ioctl.c | 132 +++++++++ fs/namespace.c | 10 + fs/super.c | 11 + include/linux/fs.h | 15 + include/linux/hot_track.h | 169 ++++++++++++ mm/filemap.c | 8 + mm/page-writeback.c | 21 ++ mm/readahead.c | 9 + 20 files changed, 2276 insertions(+), 1 deletions(-) create mode 100644 fs/hot_debugfs.c create mode 100644 fs/hot_debugfs.h create mode 100644 fs/hot_hash.c create mode 100644 fs/hot_hash.h create mode 100644 fs/hot_rb.c create mode 100644 fs/hot_rb.h create mode 100644 fs/hot_track.c create mode 100644 fs/hot_track.h create mode 100644 include/linux/hot_track.h--
1.7.6.5--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Signature · 3 lines
Regards, Zhi Yong Wu
hi, all maintainers.
ping? any comments are appreciated, thanks.
Quoted 186 lines · Zhi Yong Wu
Sorry, forgot CCed to Ted.
Quoted 169 lines · On Tue, Sep 11, 2012 at 10:27 PM, <[email protected]> wrote:
From: Zhi Yong Wu <[email protected]>
HI, folks
I have pushed the patchset to my kernel dev git tree:
[email protected]:wuzhy/kernel.gitAlso, you can review it via
https://github.com/wuzhy/kernel/commits/hottrackNOTE:
The patchset still has a lot of bugfix and cleanup to do. It is post out mainly to make sure it is going in the correct direction and hope to get some helpful comments from other guys.
TODO List:
1.) Need to do scalability or performance tests.
2.) Fix up bugs.
3.) Strictly split this patchset to keep them in order
This patchset is in RFC state, i haven't strictly split it
When it is in PATCH state, i will strictly split it and let them in order.
4.) Turn some Micro in to tunables
TIME_TO_KICK, andHEAT_UPDATE_DELAY
5.) Rafactorhot_hash_is_aging()
If you just made the timeout value a timespec and compared
the _timespecs_, you would be doing a lot fewer conversions.
6.) Cleanup some unnecessary lock protect
7.) Add more comments to explain how to calc temperatureBen Chociej, Matt Lupfer and Conor Scott originally wrote this code to
be very btrfs-specific. I've taken their code and attempted to
make it more generic and integrate it at the VFS level.INTRODUCTION:
Essentially, this means maintaining some key stats
(like number of reads/writes, last read/write time, frequency of reads/writes), then distilling those numbers down to a single "temperature" value that reflects what data is "hot," and using that temperature to move data to SSDs.The long-term goal of these patches is to allow some FSs, e.g. Btrfs to intelligently utilize SSDs in a heterogenous volume. Incidentally, this project has been motivated by
the Project Ideas page on the Btrfs wiki.Of course, users are warned not to run this code outside of development environments. These patches are EXPERIMENTAL, and as such they might eat your data and/or memory. That said, the code should be relatively safe when the hottrack mount option are disabled.
MOTIVATION:
The overall goal of enabling hot data relocation to SSD has been motivated by the Project Ideas page on the Btrfs wiki at
<https://btrfs.wiki.kernel.org/index.php/Project_ideas>.
It will divide into two steps. VFS provide hot data tracking function while specific FS will provide hot data relocation function. So as the first step of this goal, it is hoped that the patchset for hot data tracking will eventually mature into VFS.This is essentially the traditional cache argument: SSD is fast and expensive; HDD is cheap but slow. ZFS, for example, can already take advantage of SSD caching. Btrfs should also be able to take advantage of hybrid storage without many broad, sweeping changes to existing code.
SUMMARY:
- Hooks in existing vfs functions to track data access frequency
- New rbtrees for tracking access frequency of inodes and sub-file ranges (
hot_rb.c)
The relationship betweensuper_blockand rbtree is as below:
super_block->s_hotinfo.hot_inode_tree
Ininclude/linux/fs.h, one structhot_infos_hotinfois added to
super_blockstruct. Each FS instance can find hot tracking infos_hotinfovia itssuper_block. In thishot_info, it store a lot of hot tracking info such ashot_inode_tree, inode and range hash list, etc.- A hash list for indexing data by its temperature (
hot_hash.c)- A debugfs interface for dumping data from the rbtrees (
hot_debugfs.c)- A background kthread for updating inode heat info
- Mount options for enabling temperature
tracking(-o hottrack, default mean disabled)(hot_track.c)- An ioctl to retrieve the frequency information collected for a certain file
- Ioctls to enable/disable frequency tracking per inode.
Usage syntax:
root@debian-i386:~# mount -o hottrack /dev/sdb /mnt [ 1505.894078] device label test devid 1 transid 29 /dev/sdb [ 1505.952977] btrfs: disk space caching is enabled [ 1506.069678] vfs: turning on hot data tracking root@debian-i386:~# mount -t debugfs none /sys/kernel/debug root@debian-i386:~# ls -l /sys/kernel/debug/vfs_hotdata/ total 0 drwxr-xr-x 2 root root 0 Aug 8 04:40 sdb root@debian-i386:~# ls -l /sys/kernel/debug/vfs_hotdata/sdb total 0(quoted diff) +0
-2-rw-r--r-- 1 root root 0 Aug 8 04:40 inode_data -rw-r--r-- 1 root root 0 Aug 8 04:40 range_dataroot@debian-i386:~# vi /mnt/file root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/inode_data inode #279, reads 0, writes 1, avg read time 18446744073709551615, avg write time 5251566408153596, temp 109 root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/range_data inode #279, range start 0 (range len 1048576) reads 0, writes 1, avg read time 18446744073709551615, avg write time 1128690176623144209, temp 64 root@debian-i386:~# echo "hot data tracking test" >> /mnt/file root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/inode_data inode #279, reads 0, writes 2, avg read time 18446744073709551615, avg write time 4923343766042451, temp 109 root@debian-i386:~# cat /sys/kernel/debug/hot_track/sdb/range_data inode #279, range start 0 (range len 1048576) reads 0, writes 2, avg read time 18446744073709551615, avg write time 1058147040842596150, temp 64 root@debian-i386:~#Zhi Yong Wu (11):
vfs: introduce one structurehot_info
vfs: introduce one rb tree -hot_inode_tree
vfs: introduce 2 rb tree items - inode and range
vfs: add support for updating access frequency
vfs: add one new mount option -o hottrack
vfs: add init and exit support
vfs: introduce one hash table
vfs: enable hot data tracking
vfs: fork one private kthread to update temperature info
vfs: add 3 new ioctl interfaces
vfs: add debugfs supportfs/Makefile | 3 +- fs/compat_ioctl.c | 8 + fs/dcache.c | 2 + fs/direct-io.c | 10 + fs/hot_debugfs.c | 488 ++++++++++++++++++++++++++++++++++ fs/hot_debugfs.h | 60 +++++ fs/hot_hash.c | 382 ++++++++++++++++++++++++++ fs/hot_hash.h | 112 ++++++++ fs/hot_rb.c | 648 +++++++++++++++++++++++++++++++++++++++++++++ fs/hot_rb.h | 81 ++++++ fs/hot_track.c | 85 ++++++ fs/hot_track.h | 23 ++ fs/ioctl.c | 132 +++++++++ fs/namespace.c | 10 + fs/super.c | 11 + include/linux/fs.h | 15 + include/linux/hot_track.h | 169 ++++++++++++ mm/filemap.c | 8 + mm/page-writeback.c | 21 ++ mm/readahead.c | 9 + 20 files changed, 2276 insertions(+), 1 deletions(-) create mode 100644 fs/hot_debugfs.c create mode 100644 fs/hot_debugfs.h create mode 100644 fs/hot_hash.c create mode 100644 fs/hot_hash.h create mode 100644 fs/hot_rb.c create mode 100644 fs/hot_rb.h create mode 100644 fs/hot_track.c create mode 100644 fs/hot_track.h create mode 100644 include/linux/hot_track.hQuoted 7 lines
--
1.7.6.5--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html--
Regards,Zhi Yong Wu
Signature · 3 lines
Regards, Zhi Yong Wu