Loading...
1#ifndef _RAID5_H
2#define _RAID5_H
3
4#include <linux/raid/xor.h>
5#include <linux/dmaengine.h>
6
7/*
8 *
9 * Each stripe contains one buffer per device. Each buffer can be in
10 * one of a number of states stored in "flags". Changes between
11 * these states happen *almost* exclusively under the protection of the
12 * STRIPE_ACTIVE flag. Some very specific changes can happen in bi_end_io, and
13 * these are not protected by STRIPE_ACTIVE.
14 *
15 * The flag bits that are used to represent these states are:
16 * R5_UPTODATE and R5_LOCKED
17 *
18 * State Empty == !UPTODATE, !LOCK
19 * We have no data, and there is no active request
20 * State Want == !UPTODATE, LOCK
21 * A read request is being submitted for this block
22 * State Dirty == UPTODATE, LOCK
23 * Some new data is in this buffer, and it is being written out
24 * State Clean == UPTODATE, !LOCK
25 * We have valid data which is the same as on disc
26 *
27 * The possible state transitions are:
28 *
29 * Empty -> Want - on read or write to get old data for parity calc
30 * Empty -> Dirty - on compute_parity to satisfy write/sync request.(RECONSTRUCT_WRITE)
31 * Empty -> Clean - on compute_block when computing a block for failed drive
32 * Want -> Empty - on failed read
33 * Want -> Clean - on successful completion of read request
34 * Dirty -> Clean - on successful completion of write request
35 * Dirty -> Clean - on failed write
36 * Clean -> Dirty - on compute_parity to satisfy write/sync (RECONSTRUCT or RMW)
37 *
38 * The Want->Empty, Want->Clean, Dirty->Clean, transitions
39 * all happen in b_end_io at interrupt time.
40 * Each sets the Uptodate bit before releasing the Lock bit.
41 * This leaves one multi-stage transition:
42 * Want->Dirty->Clean
43 * This is safe because thinking that a Clean buffer is actually dirty
44 * will at worst delay some action, and the stripe will be scheduled
45 * for attention after the transition is complete.
46 *
47 * There is one possibility that is not covered by these states. That
48 * is if one drive has failed and there is a spare being rebuilt. We
49 * can't distinguish between a clean block that has been generated
50 * from parity calculations, and a clean block that has been
51 * successfully written to the spare ( or to parity when resyncing).
52 * To distingush these states we have a stripe bit STRIPE_INSYNC that
53 * is set whenever a write is scheduled to the spare, or to the parity
54 * disc if there is no spare. A sync request clears this bit, and
55 * when we find it set with no buffers locked, we know the sync is
56 * complete.
57 *
58 * Buffers for the md device that arrive via make_request are attached
59 * to the appropriate stripe in one of two lists linked on b_reqnext.
60 * One list (bh_read) for read requests, one (bh_write) for write.
61 * There should never be more than one buffer on the two lists
62 * together, but we are not guaranteed of that so we allow for more.
63 *
64 * If a buffer is on the read list when the associated cache buffer is
65 * Uptodate, the data is copied into the read buffer and it's b_end_io
66 * routine is called. This may happen in the end_request routine only
67 * if the buffer has just successfully been read. end_request should
68 * remove the buffers from the list and then set the Uptodate bit on
69 * the buffer. Other threads may do this only if they first check
70 * that the Uptodate bit is set. Once they have checked that they may
71 * take buffers off the read queue.
72 *
73 * When a buffer on the write list is committed for write it is copied
74 * into the cache buffer, which is then marked dirty, and moved onto a
75 * third list, the written list (bh_written). Once both the parity
76 * block and the cached buffer are successfully written, any buffer on
77 * a written list can be returned with b_end_io.
78 *
79 * The write list and read list both act as fifos. The read list,
80 * write list and written list are protected by the device_lock.
81 * The device_lock is only for list manipulations and will only be
82 * held for a very short time. It can be claimed from interrupts.
83 *
84 *
85 * Stripes in the stripe cache can be on one of two lists (or on
86 * neither). The "inactive_list" contains stripes which are not
87 * currently being used for any request. They can freely be reused
88 * for another stripe. The "handle_list" contains stripes that need
89 * to be handled in some way. Both of these are fifo queues. Each
90 * stripe is also (potentially) linked to a hash bucket in the hash
91 * table so that it can be found by sector number. Stripes that are
92 * not hashed must be on the inactive_list, and will normally be at
93 * the front. All stripes start life this way.
94 *
95 * The inactive_list, handle_list and hash bucket lists are all protected by the
96 * device_lock.
97 * - stripes have a reference counter. If count==0, they are on a list.
98 * - If a stripe might need handling, STRIPE_HANDLE is set.
99 * - When refcount reaches zero, then if STRIPE_HANDLE it is put on
100 * handle_list else inactive_list
101 *
102 * This, combined with the fact that STRIPE_HANDLE is only ever
103 * cleared while a stripe has a non-zero count means that if the
104 * refcount is 0 and STRIPE_HANDLE is set, then it is on the
105 * handle_list and if recount is 0 and STRIPE_HANDLE is not set, then
106 * the stripe is on inactive_list.
107 *
108 * The possible transitions are:
109 * activate an unhashed/inactive stripe (get_active_stripe())
110 * lockdev check-hash unlink-stripe cnt++ clean-stripe hash-stripe unlockdev
111 * activate a hashed, possibly active stripe (get_active_stripe())
112 * lockdev check-hash if(!cnt++)unlink-stripe unlockdev
113 * attach a request to an active stripe (add_stripe_bh())
114 * lockdev attach-buffer unlockdev
115 * handle a stripe (handle_stripe())
116 * setSTRIPE_ACTIVE, clrSTRIPE_HANDLE ...
117 * (lockdev check-buffers unlockdev) ..
118 * change-state ..
119 * record io/ops needed clearSTRIPE_ACTIVE schedule io/ops
120 * release an active stripe (release_stripe())
121 * lockdev if (!--cnt) { if STRIPE_HANDLE, add to handle_list else add to inactive-list } unlockdev
122 *
123 * The refcount counts each thread that have activated the stripe,
124 * plus raid5d if it is handling it, plus one for each active request
125 * on a cached buffer, and plus one if the stripe is undergoing stripe
126 * operations.
127 *
128 * The stripe operations are:
129 * -copying data between the stripe cache and user application buffers
130 * -computing blocks to save a disk access, or to recover a missing block
131 * -updating the parity on a write operation (reconstruct write and
132 * read-modify-write)
133 * -checking parity correctness
134 * -running i/o to disk
135 * These operations are carried out by raid5_run_ops which uses the async_tx
136 * api to (optionally) offload operations to dedicated hardware engines.
137 * When requesting an operation handle_stripe sets the pending bit for the
138 * operation and increments the count. raid5_run_ops is then run whenever
139 * the count is non-zero.
140 * There are some critical dependencies between the operations that prevent some
141 * from being requested while another is in flight.
142 * 1/ Parity check operations destroy the in cache version of the parity block,
143 * so we prevent parity dependent operations like writes and compute_blocks
144 * from starting while a check is in progress. Some dma engines can perform
145 * the check without damaging the parity block, in these cases the parity
146 * block is re-marked up to date (assuming the check was successful) and is
147 * not re-read from disk.
148 * 2/ When a write operation is requested we immediately lock the affected
149 * blocks, and mark them as not up to date. This causes new read requests
150 * to be held off, as well as parity checks and compute block operations.
151 * 3/ Once a compute block operation has been requested handle_stripe treats
152 * that block as if it is up to date. raid5_run_ops guaruntees that any
153 * operation that is dependent on the compute block result is initiated after
154 * the compute block completes.
155 */
156
157/*
158 * Operations state - intermediate states that are visible outside of
159 * STRIPE_ACTIVE.
160 * In general _idle indicates nothing is running, _run indicates a data
161 * processing operation is active, and _result means the data processing result
162 * is stable and can be acted upon. For simple operations like biofill and
163 * compute that only have an _idle and _run state they are indicated with
164 * sh->state flags (STRIPE_BIOFILL_RUN and STRIPE_COMPUTE_RUN)
165 */
166/**
167 * enum check_states - handles syncing / repairing a stripe
168 * @check_state_idle - check operations are quiesced
169 * @check_state_run - check operation is running
170 * @check_state_result - set outside lock when check result is valid
171 * @check_state_compute_run - check failed and we are repairing
172 * @check_state_compute_result - set outside lock when compute result is valid
173 */
174enum check_states {
175 check_state_idle = 0,
176 check_state_run, /* xor parity check */
177 check_state_run_q, /* q-parity check */
178 check_state_run_pq, /* pq dual parity check */
179 check_state_check_result,
180 check_state_compute_run, /* parity repair */
181 check_state_compute_result,
182};
183
184/**
185 * enum reconstruct_states - handles writing or expanding a stripe
186 */
187enum reconstruct_states {
188 reconstruct_state_idle = 0,
189 reconstruct_state_prexor_drain_run, /* prexor-write */
190 reconstruct_state_drain_run, /* write */
191 reconstruct_state_run, /* expand */
192 reconstruct_state_prexor_drain_result,
193 reconstruct_state_drain_result,
194 reconstruct_state_result,
195};
196
197struct stripe_head {
198 struct hlist_node hash;
199 struct list_head lru; /* inactive_list or handle_list */
200 struct raid5_private_data *raid_conf;
201 short generation; /* increments with every
202 * reshape */
203 sector_t sector; /* sector of this row */
204 short pd_idx; /* parity disk index */
205 short qd_idx; /* 'Q' disk index for raid6 */
206 short ddf_layout;/* use DDF ordering to calculate Q */
207 unsigned long state; /* state flags */
208 atomic_t count; /* nr of active thread/requests */
209 int bm_seq; /* sequence number for bitmap flushes */
210 int disks; /* disks in stripe */
211 enum check_states check_state;
212 enum reconstruct_states reconstruct_state;
213 /**
214 * struct stripe_operations
215 * @target - STRIPE_OP_COMPUTE_BLK target
216 * @target2 - 2nd compute target in the raid6 case
217 * @zero_sum_result - P and Q verification flags
218 * @request - async service request flags for raid_run_ops
219 */
220 struct stripe_operations {
221 int target, target2;
222 enum sum_check_flags zero_sum_result;
223 #ifdef CONFIG_MULTICORE_RAID456
224 unsigned long request;
225 wait_queue_head_t wait_for_ops;
226 #endif
227 } ops;
228 struct r5dev {
229 struct bio req;
230 struct bio_vec vec;
231 struct page *page;
232 struct bio *toread, *read, *towrite, *written;
233 sector_t sector; /* sector of this page */
234 unsigned long flags;
235 } dev[1]; /* allocated with extra space depending of RAID geometry */
236};
237
238/* stripe_head_state - collects and tracks the dynamic state of a stripe_head
239 * for handle_stripe.
240 */
241struct stripe_head_state {
242 int syncing, expanding, expanded;
243 int locked, uptodate, to_read, to_write, failed, written;
244 int to_fill, compute, req_compute, non_overwrite;
245 int failed_num[2];
246 int p_failed, q_failed;
247 int dec_preread_active;
248 unsigned long ops_request;
249
250 struct bio *return_bi;
251 mdk_rdev_t *blocked_rdev;
252 int handle_bad_blocks;
253};
254
255/* Flags */
256#define R5_UPTODATE 0 /* page contains current data */
257#define R5_LOCKED 1 /* IO has been submitted on "req" */
258#define R5_OVERWRITE 2 /* towrite covers whole page */
259/* and some that are internal to handle_stripe */
260#define R5_Insync 3 /* rdev && rdev->in_sync at start */
261#define R5_Wantread 4 /* want to schedule a read */
262#define R5_Wantwrite 5
263#define R5_Overlap 7 /* There is a pending overlapping request on this block */
264#define R5_ReadError 8 /* seen a read error here recently */
265#define R5_ReWrite 9 /* have tried to over-write the readerror */
266
267#define R5_Expanded 10 /* This block now has post-expand data */
268#define R5_Wantcompute 11 /* compute_block in progress treat as
269 * uptodate
270 */
271#define R5_Wantfill 12 /* dev->toread contains a bio that needs
272 * filling
273 */
274#define R5_Wantdrain 13 /* dev->towrite needs to be drained */
275#define R5_WantFUA 14 /* Write should be FUA */
276#define R5_WriteError 15 /* got a write error - need to record it */
277#define R5_MadeGood 16 /* A bad block has been fixed by writing to it*/
278/*
279 * Write method
280 */
281#define RECONSTRUCT_WRITE 1
282#define READ_MODIFY_WRITE 2
283/* not a write method, but a compute_parity mode */
284#define CHECK_PARITY 3
285/* Additional compute_parity mode -- updates the parity w/o LOCKING */
286#define UPDATE_PARITY 4
287
288/*
289 * Stripe state
290 */
291enum {
292 STRIPE_ACTIVE,
293 STRIPE_HANDLE,
294 STRIPE_SYNC_REQUESTED,
295 STRIPE_SYNCING,
296 STRIPE_INSYNC,
297 STRIPE_PREREAD_ACTIVE,
298 STRIPE_DELAYED,
299 STRIPE_DEGRADED,
300 STRIPE_BIT_DELAY,
301 STRIPE_EXPANDING,
302 STRIPE_EXPAND_SOURCE,
303 STRIPE_EXPAND_READY,
304 STRIPE_IO_STARTED, /* do not count towards 'bypass_count' */
305 STRIPE_FULL_WRITE, /* all blocks are set to be overwritten */
306 STRIPE_BIOFILL_RUN,
307 STRIPE_COMPUTE_RUN,
308 STRIPE_OPS_REQ_PENDING,
309};
310
311/*
312 * Operation request flags
313 */
314#define STRIPE_OP_BIOFILL 0
315#define STRIPE_OP_COMPUTE_BLK 1
316#define STRIPE_OP_PREXOR 2
317#define STRIPE_OP_BIODRAIN 3
318#define STRIPE_OP_RECONSTRUCT 4
319#define STRIPE_OP_CHECK 5
320
321/*
322 * Plugging:
323 *
324 * To improve write throughput, we need to delay the handling of some
325 * stripes until there has been a chance that several write requests
326 * for the one stripe have all been collected.
327 * In particular, any write request that would require pre-reading
328 * is put on a "delayed" queue until there are no stripes currently
329 * in a pre-read phase. Further, if the "delayed" queue is empty when
330 * a stripe is put on it then we "plug" the queue and do not process it
331 * until an unplug call is made. (the unplug_io_fn() is called).
332 *
333 * When preread is initiated on a stripe, we set PREREAD_ACTIVE and add
334 * it to the count of prereading stripes.
335 * When write is initiated, or the stripe refcnt == 0 (just in case) we
336 * clear the PREREAD_ACTIVE flag and decrement the count
337 * Whenever the 'handle' queue is empty and the device is not plugged, we
338 * move any strips from delayed to handle and clear the DELAYED flag and set
339 * PREREAD_ACTIVE.
340 * In stripe_handle, if we find pre-reading is necessary, we do it if
341 * PREREAD_ACTIVE is set, else we set DELAYED which will send it to the delayed queue.
342 * HANDLE gets cleared if stripe_handle leaves nothing locked.
343 */
344
345
346struct disk_info {
347 mdk_rdev_t *rdev;
348};
349
350struct raid5_private_data {
351 struct hlist_head *stripe_hashtbl;
352 mddev_t *mddev;
353 struct disk_info *spare;
354 int chunk_sectors;
355 int level, algorithm;
356 int max_degraded;
357 int raid_disks;
358 int max_nr_stripes;
359
360 /* reshape_progress is the leading edge of a 'reshape'
361 * It has value MaxSector when no reshape is happening
362 * If delta_disks < 0, it is the last sector we started work on,
363 * else is it the next sector to work on.
364 */
365 sector_t reshape_progress;
366 /* reshape_safe is the trailing edge of a reshape. We know that
367 * before (or after) this address, all reshape has completed.
368 */
369 sector_t reshape_safe;
370 int previous_raid_disks;
371 int prev_chunk_sectors;
372 int prev_algo;
373 short generation; /* increments with every reshape */
374 unsigned long reshape_checkpoint; /* Time we last updated
375 * metadata */
376
377 struct list_head handle_list; /* stripes needing handling */
378 struct list_head hold_list; /* preread ready stripes */
379 struct list_head delayed_list; /* stripes that have plugged requests */
380 struct list_head bitmap_list; /* stripes delaying awaiting bitmap update */
381 struct bio *retry_read_aligned; /* currently retrying aligned bios */
382 struct bio *retry_read_aligned_list; /* aligned bios retry list */
383 atomic_t preread_active_stripes; /* stripes with scheduled io */
384 atomic_t active_aligned_reads;
385 atomic_t pending_full_writes; /* full write backlog */
386 int bypass_count; /* bypassed prereads */
387 int bypass_threshold; /* preread nice */
388 struct list_head *last_hold; /* detect hold_list promotions */
389
390 atomic_t reshape_stripes; /* stripes with pending writes for reshape */
391 /* unfortunately we need two cache names as we temporarily have
392 * two caches.
393 */
394 int active_name;
395 char cache_name[2][32];
396 struct kmem_cache *slab_cache; /* for allocating stripes */
397
398 int seq_flush, seq_write;
399 int quiesce;
400
401 int fullsync; /* set to 1 if a full sync is needed,
402 * (fresh device added).
403 * Cleared when a sync completes.
404 */
405 int recovery_disabled;
406 /* per cpu variables */
407 struct raid5_percpu {
408 struct page *spare_page; /* Used when checking P/Q in raid6 */
409 void *scribble; /* space for constructing buffer
410 * lists and performing address
411 * conversions
412 */
413 } __percpu *percpu;
414 size_t scribble_len; /* size of scribble region must be
415 * associated with conf to handle
416 * cpu hotplug while reshaping
417 */
418#ifdef CONFIG_HOTPLUG_CPU
419 struct notifier_block cpu_notify;
420#endif
421
422 /*
423 * Free stripes pool
424 */
425 atomic_t active_stripes;
426 struct list_head inactive_list;
427 wait_queue_head_t wait_for_stripe;
428 wait_queue_head_t wait_for_overlap;
429 int inactive_blocked; /* release of inactive stripes blocked,
430 * waiting for 25% to be free
431 */
432 int pool_size; /* number of disks in stripeheads in pool */
433 spinlock_t device_lock;
434 struct disk_info *disks;
435
436 /* When taking over an array from a different personality, we store
437 * the new thread here until we fully activate the array.
438 */
439 struct mdk_thread_s *thread;
440};
441
442typedef struct raid5_private_data raid5_conf_t;
443
444/*
445 * Our supported algorithms
446 */
447#define ALGORITHM_LEFT_ASYMMETRIC 0 /* Rotating Parity N with Data Restart */
448#define ALGORITHM_RIGHT_ASYMMETRIC 1 /* Rotating Parity 0 with Data Restart */
449#define ALGORITHM_LEFT_SYMMETRIC 2 /* Rotating Parity N with Data Continuation */
450#define ALGORITHM_RIGHT_SYMMETRIC 3 /* Rotating Parity 0 with Data Continuation */
451
452/* Define non-rotating (raid4) algorithms. These allow
453 * conversion of raid4 to raid5.
454 */
455#define ALGORITHM_PARITY_0 4 /* P or P,Q are initial devices */
456#define ALGORITHM_PARITY_N 5 /* P or P,Q are final devices. */
457
458/* DDF RAID6 layouts differ from md/raid6 layouts in two ways.
459 * Firstly, the exact positioning of the parity block is slightly
460 * different between the 'LEFT_*' modes of md and the "_N_*" modes
461 * of DDF.
462 * Secondly, or order of datablocks over which the Q syndrome is computed
463 * is different.
464 * Consequently we have different layouts for DDF/raid6 than md/raid6.
465 * These layouts are from the DDFv1.2 spec.
466 * Interestingly DDFv1.2-Errata-A does not specify N_CONTINUE but
467 * leaves RLQ=3 as 'Vendor Specific'
468 */
469
470#define ALGORITHM_ROTATING_ZERO_RESTART 8 /* DDF PRL=6 RLQ=1 */
471#define ALGORITHM_ROTATING_N_RESTART 9 /* DDF PRL=6 RLQ=2 */
472#define ALGORITHM_ROTATING_N_CONTINUE 10 /*DDF PRL=6 RLQ=3 */
473
474
475/* For every RAID5 algorithm we define a RAID6 algorithm
476 * with exactly the same layout for data and parity, and
477 * with the Q block always on the last device (N-1).
478 * This allows trivial conversion from RAID5 to RAID6
479 */
480#define ALGORITHM_LEFT_ASYMMETRIC_6 16
481#define ALGORITHM_RIGHT_ASYMMETRIC_6 17
482#define ALGORITHM_LEFT_SYMMETRIC_6 18
483#define ALGORITHM_RIGHT_SYMMETRIC_6 19
484#define ALGORITHM_PARITY_0_6 20
485#define ALGORITHM_PARITY_N_6 ALGORITHM_PARITY_N
486
487static inline int algorithm_valid_raid5(int layout)
488{
489 return (layout >= 0) &&
490 (layout <= 5);
491}
492static inline int algorithm_valid_raid6(int layout)
493{
494 return (layout >= 0 && layout <= 5)
495 ||
496 (layout >= 8 && layout <= 10)
497 ||
498 (layout >= 16 && layout <= 20);
499}
500
501static inline int algorithm_is_DDF(int layout)
502{
503 return layout >= 8 && layout <= 10;
504}
505
506extern int md_raid5_congested(mddev_t *mddev, int bits);
507extern void md_raid5_kick_device(raid5_conf_t *conf);
508extern int raid5_set_cache_size(mddev_t *mddev, int size);
509#endif
1#ifndef _RAID5_H
2#define _RAID5_H
3
4#include <linux/raid/xor.h>
5#include <linux/dmaengine.h>
6
7/*
8 *
9 * Each stripe contains one buffer per device. Each buffer can be in
10 * one of a number of states stored in "flags". Changes between
11 * these states happen *almost* exclusively under the protection of the
12 * STRIPE_ACTIVE flag. Some very specific changes can happen in bi_end_io, and
13 * these are not protected by STRIPE_ACTIVE.
14 *
15 * The flag bits that are used to represent these states are:
16 * R5_UPTODATE and R5_LOCKED
17 *
18 * State Empty == !UPTODATE, !LOCK
19 * We have no data, and there is no active request
20 * State Want == !UPTODATE, LOCK
21 * A read request is being submitted for this block
22 * State Dirty == UPTODATE, LOCK
23 * Some new data is in this buffer, and it is being written out
24 * State Clean == UPTODATE, !LOCK
25 * We have valid data which is the same as on disc
26 *
27 * The possible state transitions are:
28 *
29 * Empty -> Want - on read or write to get old data for parity calc
30 * Empty -> Dirty - on compute_parity to satisfy write/sync request.
31 * Empty -> Clean - on compute_block when computing a block for failed drive
32 * Want -> Empty - on failed read
33 * Want -> Clean - on successful completion of read request
34 * Dirty -> Clean - on successful completion of write request
35 * Dirty -> Clean - on failed write
36 * Clean -> Dirty - on compute_parity to satisfy write/sync (RECONSTRUCT or RMW)
37 *
38 * The Want->Empty, Want->Clean, Dirty->Clean, transitions
39 * all happen in b_end_io at interrupt time.
40 * Each sets the Uptodate bit before releasing the Lock bit.
41 * This leaves one multi-stage transition:
42 * Want->Dirty->Clean
43 * This is safe because thinking that a Clean buffer is actually dirty
44 * will at worst delay some action, and the stripe will be scheduled
45 * for attention after the transition is complete.
46 *
47 * There is one possibility that is not covered by these states. That
48 * is if one drive has failed and there is a spare being rebuilt. We
49 * can't distinguish between a clean block that has been generated
50 * from parity calculations, and a clean block that has been
51 * successfully written to the spare ( or to parity when resyncing).
52 * To distinguish these states we have a stripe bit STRIPE_INSYNC that
53 * is set whenever a write is scheduled to the spare, or to the parity
54 * disc if there is no spare. A sync request clears this bit, and
55 * when we find it set with no buffers locked, we know the sync is
56 * complete.
57 *
58 * Buffers for the md device that arrive via make_request are attached
59 * to the appropriate stripe in one of two lists linked on b_reqnext.
60 * One list (bh_read) for read requests, one (bh_write) for write.
61 * There should never be more than one buffer on the two lists
62 * together, but we are not guaranteed of that so we allow for more.
63 *
64 * If a buffer is on the read list when the associated cache buffer is
65 * Uptodate, the data is copied into the read buffer and it's b_end_io
66 * routine is called. This may happen in the end_request routine only
67 * if the buffer has just successfully been read. end_request should
68 * remove the buffers from the list and then set the Uptodate bit on
69 * the buffer. Other threads may do this only if they first check
70 * that the Uptodate bit is set. Once they have checked that they may
71 * take buffers off the read queue.
72 *
73 * When a buffer on the write list is committed for write it is copied
74 * into the cache buffer, which is then marked dirty, and moved onto a
75 * third list, the written list (bh_written). Once both the parity
76 * block and the cached buffer are successfully written, any buffer on
77 * a written list can be returned with b_end_io.
78 *
79 * The write list and read list both act as fifos. The read list,
80 * write list and written list are protected by the device_lock.
81 * The device_lock is only for list manipulations and will only be
82 * held for a very short time. It can be claimed from interrupts.
83 *
84 *
85 * Stripes in the stripe cache can be on one of two lists (or on
86 * neither). The "inactive_list" contains stripes which are not
87 * currently being used for any request. They can freely be reused
88 * for another stripe. The "handle_list" contains stripes that need
89 * to be handled in some way. Both of these are fifo queues. Each
90 * stripe is also (potentially) linked to a hash bucket in the hash
91 * table so that it can be found by sector number. Stripes that are
92 * not hashed must be on the inactive_list, and will normally be at
93 * the front. All stripes start life this way.
94 *
95 * The inactive_list, handle_list and hash bucket lists are all protected by the
96 * device_lock.
97 * - stripes have a reference counter. If count==0, they are on a list.
98 * - If a stripe might need handling, STRIPE_HANDLE is set.
99 * - When refcount reaches zero, then if STRIPE_HANDLE it is put on
100 * handle_list else inactive_list
101 *
102 * This, combined with the fact that STRIPE_HANDLE is only ever
103 * cleared while a stripe has a non-zero count means that if the
104 * refcount is 0 and STRIPE_HANDLE is set, then it is on the
105 * handle_list and if recount is 0 and STRIPE_HANDLE is not set, then
106 * the stripe is on inactive_list.
107 *
108 * The possible transitions are:
109 * activate an unhashed/inactive stripe (get_active_stripe())
110 * lockdev check-hash unlink-stripe cnt++ clean-stripe hash-stripe unlockdev
111 * activate a hashed, possibly active stripe (get_active_stripe())
112 * lockdev check-hash if(!cnt++)unlink-stripe unlockdev
113 * attach a request to an active stripe (add_stripe_bh())
114 * lockdev attach-buffer unlockdev
115 * handle a stripe (handle_stripe())
116 * setSTRIPE_ACTIVE, clrSTRIPE_HANDLE ...
117 * (lockdev check-buffers unlockdev) ..
118 * change-state ..
119 * record io/ops needed clearSTRIPE_ACTIVE schedule io/ops
120 * release an active stripe (release_stripe())
121 * lockdev if (!--cnt) { if STRIPE_HANDLE, add to handle_list else add to inactive-list } unlockdev
122 *
123 * The refcount counts each thread that have activated the stripe,
124 * plus raid5d if it is handling it, plus one for each active request
125 * on a cached buffer, and plus one if the stripe is undergoing stripe
126 * operations.
127 *
128 * The stripe operations are:
129 * -copying data between the stripe cache and user application buffers
130 * -computing blocks to save a disk access, or to recover a missing block
131 * -updating the parity on a write operation (reconstruct write and
132 * read-modify-write)
133 * -checking parity correctness
134 * -running i/o to disk
135 * These operations are carried out by raid5_run_ops which uses the async_tx
136 * api to (optionally) offload operations to dedicated hardware engines.
137 * When requesting an operation handle_stripe sets the pending bit for the
138 * operation and increments the count. raid5_run_ops is then run whenever
139 * the count is non-zero.
140 * There are some critical dependencies between the operations that prevent some
141 * from being requested while another is in flight.
142 * 1/ Parity check operations destroy the in cache version of the parity block,
143 * so we prevent parity dependent operations like writes and compute_blocks
144 * from starting while a check is in progress. Some dma engines can perform
145 * the check without damaging the parity block, in these cases the parity
146 * block is re-marked up to date (assuming the check was successful) and is
147 * not re-read from disk.
148 * 2/ When a write operation is requested we immediately lock the affected
149 * blocks, and mark them as not up to date. This causes new read requests
150 * to be held off, as well as parity checks and compute block operations.
151 * 3/ Once a compute block operation has been requested handle_stripe treats
152 * that block as if it is up to date. raid5_run_ops guaruntees that any
153 * operation that is dependent on the compute block result is initiated after
154 * the compute block completes.
155 */
156
157/*
158 * Operations state - intermediate states that are visible outside of
159 * STRIPE_ACTIVE.
160 * In general _idle indicates nothing is running, _run indicates a data
161 * processing operation is active, and _result means the data processing result
162 * is stable and can be acted upon. For simple operations like biofill and
163 * compute that only have an _idle and _run state they are indicated with
164 * sh->state flags (STRIPE_BIOFILL_RUN and STRIPE_COMPUTE_RUN)
165 */
166/**
167 * enum check_states - handles syncing / repairing a stripe
168 * @check_state_idle - check operations are quiesced
169 * @check_state_run - check operation is running
170 * @check_state_result - set outside lock when check result is valid
171 * @check_state_compute_run - check failed and we are repairing
172 * @check_state_compute_result - set outside lock when compute result is valid
173 */
174enum check_states {
175 check_state_idle = 0,
176 check_state_run, /* xor parity check */
177 check_state_run_q, /* q-parity check */
178 check_state_run_pq, /* pq dual parity check */
179 check_state_check_result,
180 check_state_compute_run, /* parity repair */
181 check_state_compute_result,
182};
183
184/**
185 * enum reconstruct_states - handles writing or expanding a stripe
186 */
187enum reconstruct_states {
188 reconstruct_state_idle = 0,
189 reconstruct_state_prexor_drain_run, /* prexor-write */
190 reconstruct_state_drain_run, /* write */
191 reconstruct_state_run, /* expand */
192 reconstruct_state_prexor_drain_result,
193 reconstruct_state_drain_result,
194 reconstruct_state_result,
195};
196
197struct stripe_head {
198 struct hlist_node hash;
199 struct list_head lru; /* inactive_list or handle_list */
200 struct llist_node release_list;
201 struct r5conf *raid_conf;
202 short generation; /* increments with every
203 * reshape */
204 sector_t sector; /* sector of this row */
205 short pd_idx; /* parity disk index */
206 short qd_idx; /* 'Q' disk index for raid6 */
207 short ddf_layout;/* use DDF ordering to calculate Q */
208 short hash_lock_index;
209 unsigned long state; /* state flags */
210 atomic_t count; /* nr of active thread/requests */
211 int bm_seq; /* sequence number for bitmap flushes */
212 int disks; /* disks in stripe */
213 enum check_states check_state;
214 enum reconstruct_states reconstruct_state;
215 spinlock_t stripe_lock;
216 int cpu;
217 struct r5worker_group *group;
218 /**
219 * struct stripe_operations
220 * @target - STRIPE_OP_COMPUTE_BLK target
221 * @target2 - 2nd compute target in the raid6 case
222 * @zero_sum_result - P and Q verification flags
223 * @request - async service request flags for raid_run_ops
224 */
225 struct stripe_operations {
226 int target, target2;
227 enum sum_check_flags zero_sum_result;
228 } ops;
229 struct r5dev {
230 /* rreq and rvec are used for the replacement device when
231 * writing data to both devices.
232 */
233 struct bio req, rreq;
234 struct bio_vec vec, rvec;
235 struct page *page;
236 struct bio *toread, *read, *towrite, *written;
237 sector_t sector; /* sector of this page */
238 unsigned long flags;
239 } dev[1]; /* allocated with extra space depending of RAID geometry */
240};
241
242/* stripe_head_state - collects and tracks the dynamic state of a stripe_head
243 * for handle_stripe.
244 */
245struct stripe_head_state {
246 /* 'syncing' means that we need to read all devices, either
247 * to check/correct parity, or to reconstruct a missing device.
248 * 'replacing' means we are replacing one or more drives and
249 * the source is valid at this point so we don't need to
250 * read all devices, just the replacement targets.
251 */
252 int syncing, expanding, expanded, replacing;
253 int locked, uptodate, to_read, to_write, failed, written;
254 int to_fill, compute, req_compute, non_overwrite;
255 int failed_num[2];
256 int p_failed, q_failed;
257 int dec_preread_active;
258 unsigned long ops_request;
259
260 struct bio *return_bi;
261 struct md_rdev *blocked_rdev;
262 int handle_bad_blocks;
263};
264
265/* Flags for struct r5dev.flags */
266enum r5dev_flags {
267 R5_UPTODATE, /* page contains current data */
268 R5_LOCKED, /* IO has been submitted on "req" */
269 R5_DOUBLE_LOCKED,/* Cannot clear R5_LOCKED until 2 writes complete */
270 R5_OVERWRITE, /* towrite covers whole page */
271/* and some that are internal to handle_stripe */
272 R5_Insync, /* rdev && rdev->in_sync at start */
273 R5_Wantread, /* want to schedule a read */
274 R5_Wantwrite,
275 R5_Overlap, /* There is a pending overlapping request
276 * on this block */
277 R5_ReadNoMerge, /* prevent bio from merging in block-layer */
278 R5_ReadError, /* seen a read error here recently */
279 R5_ReWrite, /* have tried to over-write the readerror */
280
281 R5_Expanded, /* This block now has post-expand data */
282 R5_Wantcompute, /* compute_block in progress treat as
283 * uptodate
284 */
285 R5_Wantfill, /* dev->toread contains a bio that needs
286 * filling
287 */
288 R5_Wantdrain, /* dev->towrite needs to be drained */
289 R5_WantFUA, /* Write should be FUA */
290 R5_SyncIO, /* The IO is sync */
291 R5_WriteError, /* got a write error - need to record it */
292 R5_MadeGood, /* A bad block has been fixed by writing to it */
293 R5_ReadRepl, /* Will/did read from replacement rather than orig */
294 R5_MadeGoodRepl,/* A bad block on the replacement device has been
295 * fixed by writing to it */
296 R5_NeedReplace, /* This device has a replacement which is not
297 * up-to-date at this stripe. */
298 R5_WantReplace, /* We need to update the replacement, we have read
299 * data in, and now is a good time to write it out.
300 */
301 R5_Discard, /* Discard the stripe */
302};
303
304/*
305 * Stripe state
306 */
307enum {
308 STRIPE_ACTIVE,
309 STRIPE_HANDLE,
310 STRIPE_SYNC_REQUESTED,
311 STRIPE_SYNCING,
312 STRIPE_INSYNC,
313 STRIPE_REPLACED,
314 STRIPE_PREREAD_ACTIVE,
315 STRIPE_DELAYED,
316 STRIPE_DEGRADED,
317 STRIPE_BIT_DELAY,
318 STRIPE_EXPANDING,
319 STRIPE_EXPAND_SOURCE,
320 STRIPE_EXPAND_READY,
321 STRIPE_IO_STARTED, /* do not count towards 'bypass_count' */
322 STRIPE_FULL_WRITE, /* all blocks are set to be overwritten */
323 STRIPE_BIOFILL_RUN,
324 STRIPE_COMPUTE_RUN,
325 STRIPE_OPS_REQ_PENDING,
326 STRIPE_ON_UNPLUG_LIST,
327 STRIPE_DISCARD,
328 STRIPE_ON_RELEASE_LIST,
329};
330
331/*
332 * Operation request flags
333 */
334enum {
335 STRIPE_OP_BIOFILL,
336 STRIPE_OP_COMPUTE_BLK,
337 STRIPE_OP_PREXOR,
338 STRIPE_OP_BIODRAIN,
339 STRIPE_OP_RECONSTRUCT,
340 STRIPE_OP_CHECK,
341};
342/*
343 * Plugging:
344 *
345 * To improve write throughput, we need to delay the handling of some
346 * stripes until there has been a chance that several write requests
347 * for the one stripe have all been collected.
348 * In particular, any write request that would require pre-reading
349 * is put on a "delayed" queue until there are no stripes currently
350 * in a pre-read phase. Further, if the "delayed" queue is empty when
351 * a stripe is put on it then we "plug" the queue and do not process it
352 * until an unplug call is made. (the unplug_io_fn() is called).
353 *
354 * When preread is initiated on a stripe, we set PREREAD_ACTIVE and add
355 * it to the count of prereading stripes.
356 * When write is initiated, or the stripe refcnt == 0 (just in case) we
357 * clear the PREREAD_ACTIVE flag and decrement the count
358 * Whenever the 'handle' queue is empty and the device is not plugged, we
359 * move any strips from delayed to handle and clear the DELAYED flag and set
360 * PREREAD_ACTIVE.
361 * In stripe_handle, if we find pre-reading is necessary, we do it if
362 * PREREAD_ACTIVE is set, else we set DELAYED which will send it to the delayed queue.
363 * HANDLE gets cleared if stripe_handle leaves nothing locked.
364 */
365
366
367struct disk_info {
368 struct md_rdev *rdev, *replacement;
369};
370
371/* NOTE NR_STRIPE_HASH_LOCKS must remain below 64.
372 * This is because we sometimes take all the spinlocks
373 * and creating that much locking depth can cause
374 * problems.
375 */
376#define NR_STRIPE_HASH_LOCKS 8
377#define STRIPE_HASH_LOCKS_MASK (NR_STRIPE_HASH_LOCKS - 1)
378
379struct r5worker {
380 struct work_struct work;
381 struct r5worker_group *group;
382 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
383 bool working;
384};
385
386struct r5worker_group {
387 struct list_head handle_list;
388 struct r5conf *conf;
389 struct r5worker *workers;
390 int stripes_cnt;
391};
392
393struct r5conf {
394 struct hlist_head *stripe_hashtbl;
395 /* only protect corresponding hash list and inactive_list */
396 spinlock_t hash_locks[NR_STRIPE_HASH_LOCKS];
397 struct mddev *mddev;
398 int chunk_sectors;
399 int level, algorithm;
400 int max_degraded;
401 int raid_disks;
402 int max_nr_stripes;
403
404 /* reshape_progress is the leading edge of a 'reshape'
405 * It has value MaxSector when no reshape is happening
406 * If delta_disks < 0, it is the last sector we started work on,
407 * else is it the next sector to work on.
408 */
409 sector_t reshape_progress;
410 /* reshape_safe is the trailing edge of a reshape. We know that
411 * before (or after) this address, all reshape has completed.
412 */
413 sector_t reshape_safe;
414 int previous_raid_disks;
415 int prev_chunk_sectors;
416 int prev_algo;
417 short generation; /* increments with every reshape */
418 seqcount_t gen_lock; /* lock against generation changes */
419 unsigned long reshape_checkpoint; /* Time we last updated
420 * metadata */
421 long long min_offset_diff; /* minimum difference between
422 * data_offset and
423 * new_data_offset across all
424 * devices. May be negative,
425 * but is closest to zero.
426 */
427
428 struct list_head handle_list; /* stripes needing handling */
429 struct list_head hold_list; /* preread ready stripes */
430 struct list_head delayed_list; /* stripes that have plugged requests */
431 struct list_head bitmap_list; /* stripes delaying awaiting bitmap update */
432 struct bio *retry_read_aligned; /* currently retrying aligned bios */
433 struct bio *retry_read_aligned_list; /* aligned bios retry list */
434 atomic_t preread_active_stripes; /* stripes with scheduled io */
435 atomic_t active_aligned_reads;
436 atomic_t pending_full_writes; /* full write backlog */
437 int bypass_count; /* bypassed prereads */
438 int bypass_threshold; /* preread nice */
439 struct list_head *last_hold; /* detect hold_list promotions */
440
441 atomic_t reshape_stripes; /* stripes with pending writes for reshape */
442 /* unfortunately we need two cache names as we temporarily have
443 * two caches.
444 */
445 int active_name;
446 char cache_name[2][32];
447 struct kmem_cache *slab_cache; /* for allocating stripes */
448
449 int seq_flush, seq_write;
450 int quiesce;
451
452 int fullsync; /* set to 1 if a full sync is needed,
453 * (fresh device added).
454 * Cleared when a sync completes.
455 */
456 int recovery_disabled;
457 /* per cpu variables */
458 struct raid5_percpu {
459 struct page *spare_page; /* Used when checking P/Q in raid6 */
460 void *scribble; /* space for constructing buffer
461 * lists and performing address
462 * conversions
463 */
464 } __percpu *percpu;
465 size_t scribble_len; /* size of scribble region must be
466 * associated with conf to handle
467 * cpu hotplug while reshaping
468 */
469#ifdef CONFIG_HOTPLUG_CPU
470 struct notifier_block cpu_notify;
471#endif
472
473 /*
474 * Free stripes pool
475 */
476 atomic_t active_stripes;
477 struct list_head inactive_list[NR_STRIPE_HASH_LOCKS];
478 atomic_t empty_inactive_list_nr;
479 struct llist_head released_stripes;
480 wait_queue_head_t wait_for_stripe;
481 wait_queue_head_t wait_for_overlap;
482 int inactive_blocked; /* release of inactive stripes blocked,
483 * waiting for 25% to be free
484 */
485 int pool_size; /* number of disks in stripeheads in pool */
486 spinlock_t device_lock;
487 struct disk_info *disks;
488
489 /* When taking over an array from a different personality, we store
490 * the new thread here until we fully activate the array.
491 */
492 struct md_thread *thread;
493 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
494 struct r5worker_group *worker_groups;
495 int group_cnt;
496 int worker_cnt_per_group;
497};
498
499/*
500 * Our supported algorithms
501 */
502#define ALGORITHM_LEFT_ASYMMETRIC 0 /* Rotating Parity N with Data Restart */
503#define ALGORITHM_RIGHT_ASYMMETRIC 1 /* Rotating Parity 0 with Data Restart */
504#define ALGORITHM_LEFT_SYMMETRIC 2 /* Rotating Parity N with Data Continuation */
505#define ALGORITHM_RIGHT_SYMMETRIC 3 /* Rotating Parity 0 with Data Continuation */
506
507/* Define non-rotating (raid4) algorithms. These allow
508 * conversion of raid4 to raid5.
509 */
510#define ALGORITHM_PARITY_0 4 /* P or P,Q are initial devices */
511#define ALGORITHM_PARITY_N 5 /* P or P,Q are final devices. */
512
513/* DDF RAID6 layouts differ from md/raid6 layouts in two ways.
514 * Firstly, the exact positioning of the parity block is slightly
515 * different between the 'LEFT_*' modes of md and the "_N_*" modes
516 * of DDF.
517 * Secondly, or order of datablocks over which the Q syndrome is computed
518 * is different.
519 * Consequently we have different layouts for DDF/raid6 than md/raid6.
520 * These layouts are from the DDFv1.2 spec.
521 * Interestingly DDFv1.2-Errata-A does not specify N_CONTINUE but
522 * leaves RLQ=3 as 'Vendor Specific'
523 */
524
525#define ALGORITHM_ROTATING_ZERO_RESTART 8 /* DDF PRL=6 RLQ=1 */
526#define ALGORITHM_ROTATING_N_RESTART 9 /* DDF PRL=6 RLQ=2 */
527#define ALGORITHM_ROTATING_N_CONTINUE 10 /*DDF PRL=6 RLQ=3 */
528
529
530/* For every RAID5 algorithm we define a RAID6 algorithm
531 * with exactly the same layout for data and parity, and
532 * with the Q block always on the last device (N-1).
533 * This allows trivial conversion from RAID5 to RAID6
534 */
535#define ALGORITHM_LEFT_ASYMMETRIC_6 16
536#define ALGORITHM_RIGHT_ASYMMETRIC_6 17
537#define ALGORITHM_LEFT_SYMMETRIC_6 18
538#define ALGORITHM_RIGHT_SYMMETRIC_6 19
539#define ALGORITHM_PARITY_0_6 20
540#define ALGORITHM_PARITY_N_6 ALGORITHM_PARITY_N
541
542static inline int algorithm_valid_raid5(int layout)
543{
544 return (layout >= 0) &&
545 (layout <= 5);
546}
547static inline int algorithm_valid_raid6(int layout)
548{
549 return (layout >= 0 && layout <= 5)
550 ||
551 (layout >= 8 && layout <= 10)
552 ||
553 (layout >= 16 && layout <= 20);
554}
555
556static inline int algorithm_is_DDF(int layout)
557{
558 return layout >= 8 && layout <= 10;
559}
560
561extern int md_raid5_congested(struct mddev *mddev, int bits);
562extern void md_raid5_kick_device(struct r5conf *conf);
563extern int raid5_set_cache_size(struct mddev *mddev, int size);
564#endif