[PATCH] md: optimised resync using Bitmap based intent logging

With this patch, the intent to write to some block in the array can be logged
to a bitmap file.  Each bit represents some number of sectors and is set
before any update happens, and only cleared when all writes relating to all
sectors are complete.

After an unclean shutdown, information in this bitmap can be used to optimise
resync - only sectors which could be out-of-sync need to be updated.

Also if a drive is removed and then added back into an array, the recovery can
make use of the bitmap to optimise reconstruction.  This is not implemented in
this patch.

Currently the bitmap is stored in a file which must (obviously) be stored on a
separate device.

The patch only provided infrastructure.  It does not update any personalities
to bitmap intent logging.

Md arrays can still be used with no bitmap file.  This patch has minimal
impact on such arrays.

Signed-off-by: Neil Brown <neilb@cse.unsw.edu.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

3 files changed, 291 insertions, 0 deletions

diff --git a/include/linux/raid/bitmap.h b/include/linux/raid/bitmap.h
new file mode 100644
index 000000000000..f785cf26cbad
--- /dev/null
+++ b/include/linux/raid/bitmap.h
@@ -0,0 +1,280 @@
+/*
+ * bitmap.h: Copyright (C) Peter T. Breuer (ptb@ot.uc3m.es) 2003
+ *
+ * additions: Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.
+ */
+#ifndef BITMAP_H
+#define BITMAP_H 1
+
+#define BITMAP_MAJOR 3
+#define BITMAP_MINOR 38
+
+/*
+ * in-memory bitmap:
+ *
+ * Use 16 bit block counters to track pending writes to each "chunk".
+ * The 2 high order bits are special-purpose, the first is a flag indicating
+ * whether a resync is needed.  The second is a flag indicating whether a
+ * resync is active.
+ * This means that the counter is actually 14 bits:
+ *
+ * +--------+--------+------------------------------------------------+
+ * | resync | resync |               counter                          |
+ * | needed | active |                                                |
+ * |  (0-1) |  (0-1) |              (0-16383)                         |
+ * +--------+--------+------------------------------------------------+
+ *
+ * The "resync needed" bit is set when:
+ *    a '1' bit is read from storage at startup.
+ *    a write request fails on some drives
+ *    a resync is aborted on a chunk with 'resync active' set
+ * It is cleared (and resync-active set) when a resync starts across all drives
+ * of the chunk.
+ *
+ *
+ * The "resync active" bit is set when:
+ *    a resync is started on all drives, and resync_needed is set.
+ *       resync_needed will be cleared (as long as resync_active wasn't already set).
+ * It is cleared when a resync completes.
+ *
+ * The counter counts pending write requests, plus the on-disk bit.
+ * When the counter is '1' and the resync bits are clear, the on-disk
+ * bit can be cleared aswell, thus setting the counter to 0.
+ * When we set a bit, or in the counter (to start a write), if the fields is
+ * 0, we first set the disk bit and set the counter to 1.
+ *
+ * If the counter is 0, the on-disk bit is clear and the stipe is clean
+ * Anything that dirties the stipe pushes the counter to 2 (at least)
+ * and sets the on-disk bit (lazily).
+ * If a periodic sweep find the counter at 2, it is decremented to 1.
+ * If the sweep find the counter at 1, the on-disk bit is cleared and the
+ * counter goes to zero.
+ *
+ * Also, we'll hijack the "map" pointer itself and use it as two 16 bit block
+ * counters as a fallback when "page" memory cannot be allocated:
+ *
+ * Normal case (page memory allocated):
+ *
+ *     page pointer (32-bit)
+ *
+ *     [ ] ------+
+ *               |
+ *               +-------> [   ][   ]..[   ] (4096 byte page == 2048 counters)
+ *                          c1   c2    c2048
+ *
+ * Hijacked case (page memory allocation failed):
+ *
+ *     hijacked page pointer (32-bit)
+ *
+ *     [		  ][		  ] (no page memory allocated)
+ *      counter #1 (16-bit) counter #2 (16-bit)
+ *
+ */
+
+#ifdef __KERNEL__
+
+#define PAGE_BITS (PAGE_SIZE << 3)
+#define PAGE_BIT_SHIFT (PAGE_SHIFT + 3)
+
+typedef __u16 bitmap_counter_t;
+#define COUNTER_BITS 16
+#define COUNTER_BIT_SHIFT 4
+#define COUNTER_BYTE_RATIO (COUNTER_BITS / 8)
+#define COUNTER_BYTE_SHIFT (COUNTER_BIT_SHIFT - 3)
+
+#define NEEDED_MASK ((bitmap_counter_t) (1 << (COUNTER_BITS - 1)))
+#define RESYNC_MASK ((bitmap_counter_t) (1 << (COUNTER_BITS - 2)))
+#define COUNTER_MAX ((bitmap_counter_t) RESYNC_MASK - 1)
+#define NEEDED(x) (((bitmap_counter_t) x) & NEEDED_MASK)
+#define RESYNC(x) (((bitmap_counter_t) x) & RESYNC_MASK)
+#define COUNTER(x) (((bitmap_counter_t) x) & COUNTER_MAX)
+
+/* how many counters per page? */
+#define PAGE_COUNTER_RATIO (PAGE_BITS / COUNTER_BITS)
+/* same, except a shift value for more efficient bitops */
+#define PAGE_COUNTER_SHIFT (PAGE_BIT_SHIFT - COUNTER_BIT_SHIFT)
+/* same, except a mask value for more efficient bitops */
+#define PAGE_COUNTER_MASK  (PAGE_COUNTER_RATIO - 1)
+
+#define BITMAP_BLOCK_SIZE 512
+#define BITMAP_BLOCK_SHIFT 9
+
+/* how many blocks per chunk? (this is variable) */
+#define CHUNK_BLOCK_RATIO(bitmap) ((bitmap)->chunksize >> BITMAP_BLOCK_SHIFT)
+#define CHUNK_BLOCK_SHIFT(bitmap) ((bitmap)->chunkshift - BITMAP_BLOCK_SHIFT)
+#define CHUNK_BLOCK_MASK(bitmap) (CHUNK_BLOCK_RATIO(bitmap) - 1)
+
+/* when hijacked, the counters and bits represent even larger "chunks" */
+/* there will be 1024 chunks represented by each counter in the page pointers */
+#define PAGEPTR_BLOCK_RATIO(bitmap) \
+			(CHUNK_BLOCK_RATIO(bitmap) << PAGE_COUNTER_SHIFT >> 1)
+#define PAGEPTR_BLOCK_SHIFT(bitmap) \
+			(CHUNK_BLOCK_SHIFT(bitmap) + PAGE_COUNTER_SHIFT - 1)
+#define PAGEPTR_BLOCK_MASK(bitmap) (PAGEPTR_BLOCK_RATIO(bitmap) - 1)
+
+/*
+ * on-disk bitmap:
+ *
+ * Use one bit per "chunk" (block set). We do the disk I/O on the bitmap
+ * file a page at a time. There's a superblock at the start of the file.
+ */
+
+/* map chunks (bits) to file pages - offset by the size of the superblock */
+#define CHUNK_BIT_OFFSET(chunk) ((chunk) + (sizeof(bitmap_super_t) << 3))
+
+#endif
+
+/*
+ * bitmap structures:
+ */
+
+#define BITMAP_MAGIC 0x6d746962
+
+/* use these for bitmap->flags and bitmap->sb->state bit-fields */
+enum bitmap_state {
+	BITMAP_ACTIVE = 0x001, /* the bitmap is in use */
+	BITMAP_STALE  = 0x002  /* the bitmap file is out of date or had -EIO */
+};
+
+/* the superblock at the front of the bitmap file -- little endian */
+typedef struct bitmap_super_s {
+	__u32 magic;        /*  0  BITMAP_MAGIC */
+	__u32 version;      /*  4  the bitmap major for now, could change... */
+	__u8  uuid[16];     /*  8  128 bit uuid - must match md device uuid */
+	__u64 events;       /* 24  event counter for the bitmap (1)*/
+	__u64 events_cleared;/*32  event counter when last bit cleared (2) */
+	__u64 sync_size;    /* 40  the size of the md device's sync range(3) */
+	__u32 state;        /* 48  bitmap state information */
+	__u32 chunksize;    /* 52  the bitmap chunk size in bytes */
+	__u32 daemon_sleep; /* 56  seconds between disk flushes */
+
+	__u8  pad[256 - 60]; /* set to zero */
+} bitmap_super_t;
+
+/* notes:
+ * (1) This event counter is updated before the eventcounter in the md superblock
+ *    When a bitmap is loaded, it is only accepted if this event counter is equal
+ *    to, or one greater than, the event counter in the superblock.
+ * (2) This event counter is updated when the other one is *if*and*only*if* the
+ *    array is not degraded.  As bits are not cleared when the array is degraded,
+ *    this represents the last time that any bits were cleared.
+ *    If a device is being added that has an event count with this value or
+ *    higher, it is accepted as conforming to the bitmap.
+ * (3)This is the number of sectors represented by the bitmap, and is the range that
+ *    resync happens across.  For raid1 and raid5/6 it is the size of individual
+ *    devices.  For raid10 it is the size of the array.
+ */
+
+#ifdef __KERNEL__
+
+/* the in-memory bitmap is represented by bitmap_pages */
+struct bitmap_page {
+	/*
+	 * map points to the actual memory page
+	 */
+	char *map;
+	/*
+	 * in emergencies (when map cannot be alloced), hijack the map
+	 * pointer and use it as two counters itself
+	 */
+	unsigned int hijacked:1;
+	/*
+	 * count of dirty bits on the page
+	 */
+	unsigned int  count:31;
+};
+
+/* keep track of bitmap file pages that have pending writes on them */
+struct page_list {
+	struct list_head list;
+	struct page *page;
+};
+
+/* the main bitmap structure - one per mddev */
+struct bitmap {
+	struct bitmap_page *bp;
+	unsigned long pages; /* total number of pages in the bitmap */
+	unsigned long missing_pages; /* number of pages not yet allocated */
+
+	mddev_t *mddev; /* the md device that the bitmap is for */
+
+	int counter_bits; /* how many bits per block counter */
+
+	/* bitmap chunksize -- how much data does each bit represent? */
+	unsigned long chunksize;
+	unsigned long chunkshift; /* chunksize = 2^chunkshift (for bitops) */
+	unsigned long chunks; /* total number of data chunks for the array */
+
+	/* We hold a count on the chunk currently being synced, and drop
+	 * it when the last block is started.  If the resync is aborted
+	 * midway, we need to be able to drop that count, so we remember
+	 * the counted chunk..
+	 */
+	unsigned long syncchunk;
+
+	__u64	events_cleared;
+
+	/* bitmap spinlock */
+	spinlock_t lock;
+
+	struct file *file; /* backing disk file */
+	struct page *sb_page; /* cached copy of the bitmap file superblock */
+	struct page **filemap; /* list of cache pages for the file */
+	unsigned long *filemap_attr; /* attributes associated w/ filemap pages */
+	unsigned long file_pages; /* number of pages in the file */
+
+	unsigned long flags;
+
+	/*
+	 * the bitmap daemon - periodically wakes up and sweeps the bitmap
+	 * file, cleaning up bits and flushing out pages to disk as necessary
+	 */
+	unsigned long daemon_lastrun; /* jiffies of last run */
+	unsigned long daemon_sleep; /* how many seconds between updates? */
+
+	/*
+	 * bitmap write daemon - this daemon performs writes to the bitmap file
+	 * this thread is only needed because of a limitation in ext3 (jbd)
+	 * that does not allow a task to have two journal transactions ongoing
+	 * simultaneously (even if the transactions are for two different
+	 * filesystems) -- in the case of bitmap, that would be the filesystem
+	 * that the bitmap file resides on and the filesystem that is mounted
+	 * on the md device -- see current->journal_info in jbd/transaction.c
+	 */
+	mdk_thread_t *writeback_daemon;
+	spinlock_t write_lock;
+	struct semaphore write_ready;
+	struct semaphore write_done;
+	unsigned long writes_pending;
+	wait_queue_head_t write_wait;
+	struct list_head write_pages;
+	struct list_head complete_pages;
+	mempool_t *write_pool;
+};
+
+/* the bitmap API */
+
+/* these are used only by md/bitmap */
+int  bitmap_create(mddev_t *mddev);
+void bitmap_destroy(mddev_t *mddev);
+int  bitmap_active(struct bitmap *bitmap);
+
+char *file_path(struct file *file, char *buf, int count);
+void bitmap_print_sb(struct bitmap *bitmap);
+int bitmap_update_sb(struct bitmap *bitmap);
+
+int  bitmap_setallbits(struct bitmap *bitmap);
+
+/* these are exported */
+int bitmap_startwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors);
+void bitmap_endwrite(struct bitmap *bitmap, sector_t offset, unsigned long sectors,
+		     int success);
+int bitmap_start_sync(struct bitmap *bitmap, sector_t offset, int *blocks);
+void bitmap_end_sync(struct bitmap *bitmap, sector_t offset, int *blocks, int aborted);
+void bitmap_close_sync(struct bitmap *bitmap);
+
+int bitmap_unplug(struct bitmap *bitmap);
+int bitmap_daemon_work(struct bitmap *bitmap);
+#endif
+
+#endif
diff --git a/include/linux/raid/md_k.h b/include/linux/raid/md_k.h
index bce0032decff..16e94a9f0f8c 100644
--- a/include/linux/raid/md_k.h
+++ b/include/linux/raid/md_k.h
@@ -267,6 +267,9 @@ struct mddev_s
 	atomic_t			writes_pending; 
 	request_queue_t			*queue;	/* for plugging ... */
 
+	struct bitmap                   *bitmap; /* the bitmap for the device */
+	struct file			*bitmap_file; /* the bitmap file */
+
 	struct list_head		all_mddevs;
 };
 
@@ -341,6 +344,7 @@ typedef struct mdk_thread_s {
 	unsigned long           flags;
 	struct completion	*event;
 	struct task_struct	*tsk;
+	unsigned long		timeout;
 	const char		*name;
 } mdk_thread_t;
 
diff --git a/include/linux/raid/md_u.h b/include/linux/raid/md_u.h
index a2df5c2a42af..81da20ccec4d 100644
--- a/include/linux/raid/md_u.h
+++ b/include/linux/raid/md_u.h
@@ -23,6 +23,7 @@
 #define GET_DISK_INFO		_IOR (MD_MAJOR, 0x12, mdu_disk_info_t)
 #define PRINT_RAID_DEBUG	_IO (MD_MAJOR, 0x13)
 #define RAID_AUTORUN		_IO (MD_MAJOR, 0x14)
+#define GET_BITMAP_FILE		_IOR (MD_MAJOR, 0x15, mdu_bitmap_file_t)
 
 /* configuration */
 #define CLEAR_ARRAY		_IO (MD_MAJOR, 0x20)
@@ -36,6 +37,7 @@
 #define HOT_ADD_DISK		_IO (MD_MAJOR, 0x28)
 #define SET_DISK_FAULTY		_IO (MD_MAJOR, 0x29)
 #define HOT_GENERATE_ERROR	_IO (MD_MAJOR, 0x2a)
+#define SET_BITMAP_FILE		_IOW (MD_MAJOR, 0x2b, int)
 
 /* usage */
 #define RUN_ARRAY		_IOW (MD_MAJOR, 0x30, mdu_param_t)
@@ -106,6 +108,11 @@ typedef struct mdu_start_info_s {
 
 } mdu_start_info_t;
 
+typedef struct mdu_bitmap_file_s
+{
+	char pathname[4096];
+} mdu_bitmap_file_t;
+
 typedef struct mdu_param_s
 {
 	int			personality;	/* 1,2,3,4 */