2   lru_cache.c
   4   This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
   6   Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
   7   Copyright (C) 2003-2008, Philipp Reisner <>.
   8   Copyright (C) 2003-2008, Lars Ellenberg <>.
  10   drbd is free software; you can redistribute it and/or modify
  11   it under the terms of the GNU General Public License as published by
  12   the Free Software Foundation; either version 2, or (at your option)
  13   any later version.
  15   drbd is distributed in the hope that it will be useful,
  16   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18   GNU General Public License for more details.
  20   You should have received a copy of the GNU General Public License
  21   along with drbd; see the file COPYING.  If not, write to
  22   the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
  24 */
  26#ifndef LRU_CACHE_H
  27#define LRU_CACHE_H
  29#include <linux/list.h>
  30#include <linux/slab.h>
  31#include <linux/bitops.h>
  32#include <linux/string.h> /* for memset */
  33#include <linux/seq_file.h>
  36This header file (and its .c file; kernel-doc of functions see there)
  37  define a helper framework to easily keep track of index:label associations,
  38  and changes to an "active set" of objects, as well as pending transactions,
  39  to persistently record those changes.
  41  We use an LRU policy if it is necessary to "cool down" a region currently in
  42  the active set before we can "heat" a previously unused region.
  44  Because of this later property, it is called "lru_cache".
  45  As it actually Tracks Objects in an Active SeT, we could also call it
  46  toast (incidentally that is what may happen to the data on the
  47  backend storage uppon next resync, if we don't get it right).
  49What for?
  51We replicate IO (more or less synchronously) to local and remote disk.
  53For crash recovery after replication node failure,
  54  we need to resync all regions that have been target of in-flight WRITE IO
  55  (in use, or "hot", regions), as we don't know whether or not those WRITEs
  56  have made it to stable storage.
  58  To avoid a "full resync", we need to persistently track these regions.
  60  This is known as "write intent log", and can be implemented as on-disk
  61  (coarse or fine grained) bitmap, or other meta data.
  63  To avoid the overhead of frequent extra writes to this meta data area,
  64  usually the condition is softened to regions that _may_ have been target of
  65  in-flight WRITE IO, e.g. by only lazily clearing the on-disk write-intent
  66  bitmap, trading frequency of meta data transactions against amount of
  67  (possibly unnecessary) resync traffic.
  69  If we set a hard limit on the area that may be "hot" at any given time, we
  70  limit the amount of resync traffic needed for crash recovery.
  72For recovery after replication link failure,
  73  we need to resync all blocks that have been changed on the other replica
  74  in the mean time, or, if both replica have been changed independently [*],
  75  all blocks that have been changed on either replica in the mean time.
  76  [*] usually as a result of a cluster split-brain and insufficient protection.
  77      but there are valid use cases to do this on purpose.
  79  Tracking those blocks can be implemented as "dirty bitmap".
  80  Having it fine-grained reduces the amount of resync traffic.
  81  It should also be persistent, to allow for reboots (or crashes)
  82  while the replication link is down.
  84There are various possible implementations for persistently storing
  85write intent log information, three of which are mentioned here.
  87"Chunk dirtying"
  88  The on-disk "dirty bitmap" may be re-used as "write-intent" bitmap as well.
  89  To reduce the frequency of bitmap updates for write-intent log purposes,
  90  one could dirty "chunks" (of some size) at a time of the (fine grained)
  91  on-disk bitmap, while keeping the in-memory "dirty" bitmap as clean as
  92  possible, flushing it to disk again when a previously "hot" (and on-disk
  93  dirtied as full chunk) area "cools down" again (no IO in flight anymore,
  94  and none expected in the near future either).
  96"Explicit (coarse) write intent bitmap"
  97  An other implementation could chose a (probably coarse) explicit bitmap,
  98  for write-intent log purposes, additionally to the fine grained dirty bitmap.
 100"Activity log"
 101  Yet an other implementation may keep track of the hot regions, by starting
 102  with an empty set, and writing down a journal of region numbers that have
 103  become "hot", or have "cooled down" again.
 105  To be able to use a ring buffer for this journal of changes to the active
 106  set, we not only record the actual changes to that set, but also record the
 107  not changing members of the set in a round robin fashion. To do so, we use a
 108  fixed (but configurable) number of slots which we can identify by index, and
 109  associate region numbers (labels) with these indices.
 110  For each transaction recording a change to the active set, we record the
 111  change itself (index: -old_label, +new_label), and which index is associated
 112  with which label (index: current_label) within a certain sliding window that
 113  is moved further over the available indices with each such transaction.
 115  Thus, for crash recovery, if the ringbuffer is sufficiently large, we can
 116  accurately reconstruct the active set.
 118  Sufficiently large depends only on maximum number of active objects, and the
 119  size of the sliding window recording "index: current_label" associations within
 120  each transaction.
 122  This is what we call the "activity log".
 124  Currently we need one activity log transaction per single label change, which
 125  does not give much benefit over the "dirty chunks of bitmap" approach, other
 126  than potentially less seeks.
 128  We plan to change the transaction format to support multiple changes per
 129  transaction, which then would reduce several (disjoint, "random") updates to
 130  the bitmap into one transaction to the activity log ring buffer.
 133/* this defines an element in a tracked set
 134 * .colision is for hash table lookup.
 135 * When we process a new IO request, we know its sector, thus can deduce the
 136 * region number (label) easily.  To do the label -> object lookup without a
 137 * full list walk, we use a simple hash table.
 138 *
 139 * .list is on one of three lists:
 140 *  in_use: currently in use (refcnt > 0, lc_number != LC_FREE)
 141 *     lru: unused but ready to be reused or recycled
 142 *          (lc_refcnt == 0, lc_number != LC_FREE),
 143 *    free: unused but ready to be recycled
 144 *          (lc_refcnt == 0, lc_number == LC_FREE),
 145 *
 146 * an element is said to be "in the active set",
 147 * if either on "in_use" or "lru", i.e. lc_number != LC_FREE.
 148 *
 149 * DRBD currently (May 2009) only uses 61 elements on the resync lru_cache
 150 * (total memory usage 2 pages), and up to 3833 elements on the act_log
 151 * lru_cache, totalling ~215 kB for 64bit architecture, ~53 pages.
 152 *
 153 * We usually do not actually free these objects again, but only "recycle"
 154 * them, as the change "index: -old_label, +LC_FREE" would need a transaction
 155 * as well.  Which also means that using a kmem_cache to allocate the objects
 156 * from wastes some resources.
 157 * But it avoids high order page allocations in kmalloc.
 158 */
 159struct lc_element {
 160        struct hlist_node colision;
 161        struct list_head list;           /* LRU list or free list */
 162        unsigned refcnt;
 163        /* back "pointer" into lc_cache->element[index],
 164         * for paranoia, and for "lc_element_to_index" */
 165        unsigned lc_index;
 166        /* if we want to track a larger set of objects,
 167         * it needs to become arch independend u64 */
 168        unsigned lc_number;
 169        /* special label when on free list */
 170#define LC_FREE (~0U)
 172        /* for pending changes */
 173        unsigned lc_new_number;
 176struct lru_cache {
 177        /* the least recently used item is kept at lru->prev */
 178        struct list_head lru;
 179        struct list_head free;
 180        struct list_head in_use;
 181        struct list_head to_be_changed;
 183        /* the pre-created kmem cache to allocate the objects from */
 184        struct kmem_cache *lc_cache;
 186        /* size of tracked objects, used to memset(,0,) them in lc_reset */
 187        size_t element_size;
 188        /* offset of struct lc_element member in the tracked object */
 189        size_t element_off;
 191        /* number of elements (indices) */
 192        unsigned int nr_elements;
 193        /* Arbitrary limit on maximum tracked objects. Practical limit is much
 194         * lower due to allocation failures, probably. For typical use cases,
 195         * nr_elements should be a few thousand at most.
 196         * This also limits the maximum value of lc_element.lc_index, allowing the
 197         * 8 high bits of .lc_index to be overloaded with flags in the future. */
 198#define LC_MAX_ACTIVE   (1<<24)
 200        /* allow to accumulate a few (index:label) changes,
 201         * but no more than max_pending_changes */
 202        unsigned int max_pending_changes;
 203        /* number of elements currently on to_be_changed list */
 204        unsigned int pending_changes;
 206        /* statistics */
 207        unsigned used; /* number of elements currently on in_use list */
 208        unsigned long hits, misses, starving, locked, changed;
 210        /* see below: flag-bits for lru_cache */
 211        unsigned long flags;
 214        void  *lc_private;
 215        const char *name;
 217        /* nr_elements there */
 218        struct hlist_head *lc_slot;
 219        struct lc_element **lc_element;
 223/* flag-bits for lru_cache */
 224enum {
 225        /* debugging aid, to catch concurrent access early.
 226         * user needs to guarantee exclusive access by proper locking! */
 227        __LC_PARANOIA,
 229        /* annotate that the set is "dirty", possibly accumulating further
 230         * changes, until a transaction is finally triggered */
 231        __LC_DIRTY,
 233        /* Locked, no further changes allowed.
 234         * Also used to serialize changing transactions. */
 235        __LC_LOCKED,
 237        /* if we need to change the set, but currently there is no free nor
 238         * unused element available, we are "starving", and must not give out
 239         * further references, to guarantee that eventually some refcnt will
 240         * drop to zero and we will be able to make progress again, changing
 241         * the set, writing the transaction.
 242         * if the statistics say we are frequently starving,
 243         * nr_elements is too small. */
 244        __LC_STARVING,
 246#define LC_PARANOIA (1<<__LC_PARANOIA)
 247#define LC_DIRTY    (1<<__LC_DIRTY)
 248#define LC_LOCKED   (1<<__LC_LOCKED)
 249#define LC_STARVING (1<<__LC_STARVING)
 251extern struct lru_cache *lc_create(const char *name, struct kmem_cache *cache,
 252                unsigned max_pending_changes,
 253                unsigned e_count, size_t e_size, size_t e_off);
 254extern void lc_reset(struct lru_cache *lc);
 255extern void lc_destroy(struct lru_cache *lc);
 256extern void lc_set(struct lru_cache *lc, unsigned int enr, int index);
 257extern void lc_del(struct lru_cache *lc, struct lc_element *element);
 259extern struct lc_element *lc_try_get(struct lru_cache *lc, unsigned int enr);
 260extern struct lc_element *lc_find(struct lru_cache *lc, unsigned int enr);
 261extern struct lc_element *lc_get(struct lru_cache *lc, unsigned int enr);
 262extern unsigned int lc_put(struct lru_cache *lc, struct lc_element *e);
 263extern void lc_committed(struct lru_cache *lc);
 265struct seq_file;
 266extern size_t lc_seq_printf_stats(struct seq_file *seq, struct lru_cache *lc);
 268extern void lc_seq_dump_details(struct seq_file *seq, struct lru_cache *lc, char *utext,
 269                                void (*detail) (struct seq_file *, struct lc_element *));
 272 * lc_try_lock_for_transaction - can be used to stop lc_get() from changing the tracked set
 273 * @lc: the lru cache to operate on
 274 *
 275 * Allows (expects) the set to be "dirty".  Note that the reference counts and
 276 * order on the active and lru lists may still change.  Used to serialize
 277 * changing transactions.  Returns true if we aquired the lock.
 278 */
 279static inline int lc_try_lock_for_transaction(struct lru_cache *lc)
 281        return !test_and_set_bit(__LC_LOCKED, &lc->flags);
 285 * lc_try_lock - variant to stop lc_get() from changing the tracked set
 286 * @lc: the lru cache to operate on
 287 *
 288 * Note that the reference counts and order on the active and lru lists may
 289 * still change.  Only works on a "clean" set.  Returns true if we aquired the
 290 * lock, which means there are no pending changes, and any further attempt to
 291 * change the set will not succeed until the next lc_unlock().
 292 */
 293extern int lc_try_lock(struct lru_cache *lc);
 296 * lc_unlock - unlock @lc, allow lc_get() to change the set again
 297 * @lc: the lru cache to operate on
 298 */
 299static inline void lc_unlock(struct lru_cache *lc)
 301        clear_bit(__LC_DIRTY, &lc->flags);
 302        clear_bit_unlock(__LC_LOCKED, &lc->flags);
 305extern bool lc_is_used(struct lru_cache *lc, unsigned int enr);
 307#define lc_entry(ptr, type, member) \
 308        container_of(ptr, type, member)
 310extern struct lc_element *lc_element_by_index(struct lru_cache *lc, unsigned i);
 311extern unsigned int lc_index_of(struct lru_cache *lc, struct lc_element *e);
 314 kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.