2 * bitmap.h: Copyright (C) Peter T. Breuer (ptb@ot.uc3m.es) 2003
4 * additions: Copyright (C) 2003-2004, Paul Clements, SteelEye Technology, Inc.
9 #define BITMAP_MAJOR_LO 3
10 /* version 4 insists the bitmap is in little-endian order
11 * with version 3, it is host-endian which is non-portable
13 #define BITMAP_MAJOR_HI 4
14 #define BITMAP_MAJOR_HOSTENDIAN 3
16 #define BITMAP_MINOR 39
21 * Use 16 bit block counters to track pending writes to each "chunk".
22 * The 2 high order bits are special-purpose, the first is a flag indicating
23 * whether a resync is needed. The second is a flag indicating whether a
25 * This means that the counter is actually 14 bits:
27 * +--------+--------+------------------------------------------------+
28 * | resync | resync | counter |
29 * | needed | active | |
30 * | (0-1) | (0-1) | (0-16383) |
31 * +--------+--------+------------------------------------------------+
33 * The "resync needed" bit is set when:
34 * a '1' bit is read from storage at startup.
35 * a write request fails on some drives
36 * a resync is aborted on a chunk with 'resync active' set
37 * It is cleared (and resync-active set) when a resync starts across all drives
41 * The "resync active" bit is set when:
42 * a resync is started on all drives, and resync_needed is set.
43 * resync_needed will be cleared (as long as resync_active wasn't already set).
44 * It is cleared when a resync completes.
46 * The counter counts pending write requests, plus the on-disk bit.
47 * When the counter is '1' and the resync bits are clear, the on-disk
48 * bit can be cleared aswell, thus setting the counter to 0.
49 * When we set a bit, or in the counter (to start a write), if the fields is
50 * 0, we first set the disk bit and set the counter to 1.
52 * If the counter is 0, the on-disk bit is clear and the stipe is clean
53 * Anything that dirties the stipe pushes the counter to 2 (at least)
54 * and sets the on-disk bit (lazily).
55 * If a periodic sweep find the counter at 2, it is decremented to 1.
56 * If the sweep find the counter at 1, the on-disk bit is cleared and the
57 * counter goes to zero.
59 * Also, we'll hijack the "map" pointer itself and use it as two 16 bit block
60 * counters as a fallback when "page" memory cannot be allocated:
62 * Normal case (page memory allocated):
64 * page pointer (32-bit)
68 * +-------> [ ][ ]..[ ] (4096 byte page == 2048 counters)
71 * Hijacked case (page memory allocation failed):
73 * hijacked page pointer (32-bit)
75 * [ ][ ] (no page memory allocated)
76 * counter #1 (16-bit) counter #2 (16-bit)
82 #define PAGE_BITS (PAGE_SIZE << 3)
83 #define PAGE_BIT_SHIFT (PAGE_SHIFT + 3)
85 typedef __u16 bitmap_counter_t
;
86 #define COUNTER_BITS 16
87 #define COUNTER_BIT_SHIFT 4
88 #define COUNTER_BYTE_RATIO (COUNTER_BITS / 8)
89 #define COUNTER_BYTE_SHIFT (COUNTER_BIT_SHIFT - 3)
91 #define NEEDED_MASK ((bitmap_counter_t) (1 << (COUNTER_BITS - 1)))
92 #define RESYNC_MASK ((bitmap_counter_t) (1 << (COUNTER_BITS - 2)))
93 #define COUNTER_MAX ((bitmap_counter_t) RESYNC_MASK - 1)
94 #define NEEDED(x) (((bitmap_counter_t) x) & NEEDED_MASK)
95 #define RESYNC(x) (((bitmap_counter_t) x) & RESYNC_MASK)
96 #define COUNTER(x) (((bitmap_counter_t) x) & COUNTER_MAX)
98 /* how many counters per page? */
99 #define PAGE_COUNTER_RATIO (PAGE_BITS / COUNTER_BITS)
100 /* same, except a shift value for more efficient bitops */
101 #define PAGE_COUNTER_SHIFT (PAGE_BIT_SHIFT - COUNTER_BIT_SHIFT)
102 /* same, except a mask value for more efficient bitops */
103 #define PAGE_COUNTER_MASK (PAGE_COUNTER_RATIO - 1)
105 #define BITMAP_BLOCK_SIZE 512
106 #define BITMAP_BLOCK_SHIFT 9
108 /* how many blocks per chunk? (this is variable) */
109 #define CHUNK_BLOCK_RATIO(bitmap) ((bitmap)->chunksize >> BITMAP_BLOCK_SHIFT)
110 #define CHUNK_BLOCK_SHIFT(bitmap) ((bitmap)->chunkshift - BITMAP_BLOCK_SHIFT)
111 #define CHUNK_BLOCK_MASK(bitmap) (CHUNK_BLOCK_RATIO(bitmap) - 1)
113 /* when hijacked, the counters and bits represent even larger "chunks" */
114 /* there will be 1024 chunks represented by each counter in the page pointers */
115 #define PAGEPTR_BLOCK_RATIO(bitmap) \
116 (CHUNK_BLOCK_RATIO(bitmap) << PAGE_COUNTER_SHIFT >> 1)
117 #define PAGEPTR_BLOCK_SHIFT(bitmap) \
118 (CHUNK_BLOCK_SHIFT(bitmap) + PAGE_COUNTER_SHIFT - 1)
119 #define PAGEPTR_BLOCK_MASK(bitmap) (PAGEPTR_BLOCK_RATIO(bitmap) - 1)
124 * Use one bit per "chunk" (block set). We do the disk I/O on the bitmap
125 * file a page at a time. There's a superblock at the start of the file.
128 /* map chunks (bits) to file pages - offset by the size of the superblock */
129 #define CHUNK_BIT_OFFSET(chunk) ((chunk) + (sizeof(bitmap_super_t) << 3))
137 #define BITMAP_MAGIC 0x6d746962
139 /* use these for bitmap->flags and bitmap->sb->state bit-fields */
141 BITMAP_ACTIVE
= 0x001, /* the bitmap is in use */
142 BITMAP_STALE
= 0x002 /* the bitmap file is out of date or had -EIO */
145 /* the superblock at the front of the bitmap file -- little endian */
146 typedef struct bitmap_super_s
{
147 __u32 magic
; /* 0 BITMAP_MAGIC */
148 __u32 version
; /* 4 the bitmap major for now, could change... */
149 __u8 uuid
[16]; /* 8 128 bit uuid - must match md device uuid */
150 __u64 events
; /* 24 event counter for the bitmap (1)*/
151 __u64 events_cleared
;/*32 event counter when last bit cleared (2) */
152 __u64 sync_size
; /* 40 the size of the md device's sync range(3) */
153 __u32 state
; /* 48 bitmap state information */
154 __u32 chunksize
; /* 52 the bitmap chunk size in bytes */
155 __u32 daemon_sleep
; /* 56 seconds between disk flushes */
156 __u32 write_behind
; /* 60 number of outstanding write-behind writes */
158 __u8 pad
[256 - 64]; /* set to zero */
162 * (1) This event counter is updated before the eventcounter in the md superblock
163 * When a bitmap is loaded, it is only accepted if this event counter is equal
164 * to, or one greater than, the event counter in the superblock.
165 * (2) This event counter is updated when the other one is *if*and*only*if* the
166 * array is not degraded. As bits are not cleared when the array is degraded,
167 * this represents the last time that any bits were cleared.
168 * If a device is being added that has an event count with this value or
169 * higher, it is accepted as conforming to the bitmap.
170 * (3)This is the number of sectors represented by the bitmap, and is the range that
171 * resync happens across. For raid1 and raid5/6 it is the size of individual
172 * devices. For raid10 it is the size of the array.
177 /* the in-memory bitmap is represented by bitmap_pages */
180 * map points to the actual memory page
184 * in emergencies (when map cannot be alloced), hijack the map
185 * pointer and use it as two counters itself
187 unsigned int hijacked
;
189 * count of dirty bits on the page
194 /* keep track of bitmap file pages that have pending writes on them */
196 struct list_head list
;
200 /* the main bitmap structure - one per mddev */
202 struct bitmap_page
*bp
;
203 unsigned long pages
; /* total number of pages in the bitmap */
204 unsigned long missing_pages
; /* number of pages not yet allocated */
206 mddev_t
*mddev
; /* the md device that the bitmap is for */
208 int counter_bits
; /* how many bits per block counter */
210 /* bitmap chunksize -- how much data does each bit represent? */
211 unsigned long chunksize
;
212 unsigned long chunkshift
; /* chunksize = 2^chunkshift (for bitops) */
213 unsigned long chunks
; /* total number of data chunks for the array */
215 /* We hold a count on the chunk currently being synced, and drop
216 * it when the last block is started. If the resync is aborted
217 * midway, we need to be able to drop that count, so we remember
218 * the counted chunk..
220 unsigned long syncchunk
;
222 __u64 events_cleared
;
224 /* bitmap spinlock */
227 struct file
*file
; /* backing disk file */
228 struct page
*sb_page
; /* cached copy of the bitmap file superblock */
229 struct page
**filemap
; /* list of cache pages for the file */
230 unsigned long *filemap_attr
; /* attributes associated w/ filemap pages */
231 unsigned long file_pages
; /* number of pages in the file */
236 * the bitmap daemon - periodically wakes up and sweeps the bitmap
237 * file, cleaning up bits and flushing out pages to disk as necessary
239 mdk_thread_t
*daemon
;
240 unsigned long daemon_sleep
; /* how many seconds between updates? */
243 * bitmap write daemon - this daemon performs writes to the bitmap file
244 * this thread is only needed because of a limitation in ext3 (jbd)
245 * that does not allow a task to have two journal transactions ongoing
246 * simultaneously (even if the transactions are for two different
247 * filesystems) -- in the case of bitmap, that would be the filesystem
248 * that the bitmap file resides on and the filesystem that is mounted
249 * on the md device -- see current->journal_info in jbd/transaction.c
251 mdk_thread_t
*write_daemon
;
252 mdk_thread_t
*writeback_daemon
;
253 spinlock_t write_lock
;
254 struct semaphore write_ready
;
255 struct semaphore write_done
;
256 unsigned long writes_pending
;
257 wait_queue_head_t write_wait
;
258 struct list_head write_pages
;
259 struct list_head complete_pages
;
260 mempool_t
*write_pool
;
265 /* these are used only by md/bitmap */
266 int bitmap_create(mddev_t
*mddev
);
267 void bitmap_destroy(mddev_t
*mddev
);
268 int bitmap_active(struct bitmap
*bitmap
);
270 char *file_path(struct file
*file
, char *buf
, int count
);
271 void bitmap_print_sb(struct bitmap
*bitmap
);
272 int bitmap_update_sb(struct bitmap
*bitmap
);
274 int bitmap_setallbits(struct bitmap
*bitmap
);
276 /* these are exported */
277 void bitmap_startwrite(struct bitmap
*bitmap
, sector_t offset
, unsigned long sectors
);
278 void bitmap_endwrite(struct bitmap
*bitmap
, sector_t offset
, unsigned long sectors
,
280 int bitmap_start_sync(struct bitmap
*bitmap
, sector_t offset
, int *blocks
);
281 void bitmap_end_sync(struct bitmap
*bitmap
, sector_t offset
, int *blocks
, int aborted
);
282 void bitmap_close_sync(struct bitmap
*bitmap
);
284 int bitmap_unplug(struct bitmap
*bitmap
);