// SPDX-License-Identifier: GPL-2.0 /* * background writeback - scan btree for dirty data and write it to the backing * device * * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com> * Copyright 2012 Google, Inc.
*/
/* * This is the size of the cache, minus the amount used for * flash-only devices
*/
uint64_t cache_sectors = c->nbuckets * c->cache->sb.bucket_size -
atomic_long_read(&c->flash_dev_dirty_sectors);
/* * Unfortunately there is no control of global dirty data. If the * user states that they want 10% dirty data in the cache, and has, * e.g., 5 backing volumes of equal size, we try and ensure each * backing volume uses about 2% of the cache for dirty data.
*/
uint32_t bdev_share =
div64_u64(bdev_nr_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
c->cached_dev_sectors);
staticvoid __update_writeback_rate(struct cached_dev *dc)
{ /* * PI controller: * Figures out the amount that should be written per second. * * First, the error (number of sectors that are dirty beyond our * target) is calculated. The error is accumulated (numerically * integrated). * * Then, the proportional value and integral value are scaled * based on configured values. These are stored as inverses to * avoid fixed point math and to make configuration easy-- e.g. * the default value of 40 for writeback_rate_p_term_inverse * attempts to write at a rate that would retire all the dirty * blocks in 40 seconds. * * The writeback_rate_i_inverse value of 10000 means that 1/10000th * of the error is accumulated in the integral term per second. * This acts as a slow, long-term average that is not subject to * variations in usage like the p term.
*/
int64_t target = __calc_target_rate(dc);
int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
int64_t error = dirty - target;
int64_t proportional_scaled =
div_s64(error, dc->writeback_rate_p_term_inverse);
int64_t integral_scaled;
uint32_t new_rate;
/* * We need to consider the number of dirty buckets as well * when calculating the proportional_scaled, Otherwise we might * have an unreasonable small writeback rate at a highly fragmented situation * when very few dirty sectors consumed a lot dirty buckets, the * worst case is when dirty buckets reached cutoff_writeback_sync and * dirty data is still not even reached to writeback percent, so the rate * still will be at the minimum value, which will cause the write * stuck at a non-writeback mode.
*/ struct cache_set *c = dc->disk.c;
if (c->gc_stats.in_use <= BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID) {
fp_term = (int64_t)dc->writeback_rate_fp_term_low *
(c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW);
} elseif (c->gc_stats.in_use <= BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH) {
fp_term = (int64_t)dc->writeback_rate_fp_term_mid *
(c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID);
} else {
fp_term = (int64_t)dc->writeback_rate_fp_term_high *
(c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH);
}
fps = div_s64(dirty, dirty_buckets) * fp_term; if (fragment > 3 && fps > proportional_scaled) { /* Only overrite the p when fragment > 3 */
proportional_scaled = fps;
}
}
if ((error < 0 && dc->writeback_rate_integral > 0) ||
(error > 0 && time_before64(local_clock(),
dc->writeback_rate.next + NSEC_PER_MSEC))) { /* * Only decrease the integral term if it's more than * zero. Only increase the integral term if the device * is keeping up. (Don't wind up the integral * ineffectively in either case). * * It's necessary to scale this by * writeback_rate_update_seconds to keep the integral * term dimensioned properly.
*/
dc->writeback_rate_integral += error *
dc->writeback_rate_update_seconds;
}
staticbool idle_counter_exceeded(struct cache_set *c)
{ int counter, dev_nr;
/* * If c->idle_counter is overflow (idel for really long time), * reset as 0 and not set maximum rate this time for code * simplicity.
*/
counter = atomic_inc_return(&c->idle_counter); if (counter <= 0) {
atomic_set(&c->idle_counter, 0); returnfalse;
}
dev_nr = atomic_read(&c->attached_dev_nr); if (dev_nr == 0) returnfalse;
/* * c->idle_counter is increased by writeback thread of all * attached backing devices, in order to represent a rough * time period, counter should be divided by dev_nr. * Otherwise the idle time cannot be larger with more backing * device attached. * The following calculation equals to checking * (counter / dev_nr) < (dev_nr * 6)
*/ if (counter < (dev_nr * dev_nr * 6)) returnfalse;
returntrue;
}
/* * Idle_counter is increased every time when update_writeback_rate() is * called. If all backing devices attached to the same cache set have * identical dc->writeback_rate_update_seconds values, it is about 6 * rounds of update_writeback_rate() on each backing device before * c->at_max_writeback_rate is set to 1, and then max wrteback rate set * to each dc->writeback_rate.rate. * In order to avoid extra locking cost for counting exact dirty cached * devices number, c->attached_dev_nr is used to calculate the idle * throushold. It might be bigger if not all cached device are in write- * back mode, but it still works well with limited extra rounds of * update_writeback_rate().
*/ staticbool set_at_max_writeback_rate(struct cache_set *c, struct cached_dev *dc)
{ /* Don't sst max writeback rate if it is disabled */ if (!c->idle_max_writeback_rate_enabled) returnfalse;
/* Don't set max writeback rate if gc is running */ if (!c->gc_mark_valid) returnfalse;
if (!idle_counter_exceeded(c)) returnfalse;
if (atomic_read(&c->at_max_writeback_rate) != 1)
atomic_set(&c->at_max_writeback_rate, 1);
/* keep writeback_rate_target as existing value */
dc->writeback_rate_proportional = 0;
dc->writeback_rate_integral_scaled = 0;
dc->writeback_rate_change = 0;
/* * In case new I/O arrives during before * set_at_max_writeback_rate() returns.
*/ if (!idle_counter_exceeded(c) ||
!atomic_read(&c->at_max_writeback_rate)) returnfalse;
/* * should check BCACHE_DEV_RATE_DW_RUNNING before calling * cancel_delayed_work_sync().
*/
set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags); /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
smp_mb__after_atomic();
/* * CACHE_SET_IO_DISABLE might be set via sysfs interface, * check it here too.
*/ if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags); /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
smp_mb__after_atomic(); return;
}
/* * If the whole cache set is idle, set_at_max_writeback_rate() * will set writeback rate to a max number. Then it is * unncessary to update writeback rate for an idle cache set * in maximum writeback rate number(s).
*/ if (atomic_read(&dc->has_dirty) && dc->writeback_percent &&
!set_at_max_writeback_rate(c, dc)) { do { if (!down_read_trylock((&dc->writeback_lock))) {
dc->rate_update_retry++; if (dc->rate_update_retry <=
BCH_WBRATE_UPDATE_MAX_SKIPS) break;
down_read(&dc->writeback_lock);
dc->rate_update_retry = 0;
}
__update_writeback_rate(dc);
update_gc_after_writeback(c);
up_read(&dc->writeback_lock);
} while (0);
}
/* * CACHE_SET_IO_DISABLE might be set via sysfs interface, * check it here too.
*/ if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
!test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
schedule_delayed_work(&dc->writeback_rate_update,
dc->writeback_rate_update_seconds * HZ);
}
/* * should check BCACHE_DEV_RATE_DW_RUNNING before calling * cancel_delayed_work_sync().
*/
clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags); /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
smp_mb__after_atomic();
}
if (atomic_read(&dc->writeback_sequence_next) != io->sequence) { /* Not our turn to write; wait for a write to complete */
closure_wait(&dc->writeback_ordering_wait, cl);
if (atomic_read(&dc->writeback_sequence_next) == io->sequence) { /* * Edge case-- it happened in indeterminate order * relative to when we were added to wait list..
*/
closure_wake_up(&dc->writeback_ordering_wait);
}
/* * IO errors are signalled using the dirty bit on the key. * If we failed to read, we should not attempt to write to the * backing device. Instead, immediately go to write_dirty_finish * to clean up.
*/ if (KEY_DIRTY(&w->key)) {
dirty_init(w);
io->bio.bi_opf = REQ_OP_WRITE;
io->bio.bi_iter.bi_sector = KEY_START(&w->key);
bio_set_dev(&io->bio, io->dc->bdev);
io->bio.bi_end_io = dirty_endio;
/* I/O request sent to backing device */
closure_bio_submit(io->dc->disk.c, &io->bio, cl);
}
/* * XXX: if we error, background writeback just spins. Should use some * mempools.
*/
next = bch_keybuf_next(&dc->writeback_keys);
while (!kthread_should_stop() &&
!test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
next) {
size = 0;
nk = 0;
do {
BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
/* * Don't combine too many operations, even if they * are all small.
*/ if (nk >= MAX_WRITEBACKS_IN_PASS) break;
/* * If the current operation is very large, don't * further combine operations.
*/ if (size >= MAX_WRITESIZE_IN_PASS) break;
/* * Operations are only eligible to be combined * if they are contiguous. * * TODO: add a heuristic willing to fire a * certain amount of non-contiguous IO per pass, * so that we can benefit from backing device * command queueing.
*/ if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
&START_KEY(&next->key))) break;
if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL)) goto err_free;
trace_bcache_writeback(&w->key);
down(&dc->in_flight);
/* * We've acquired a semaphore for the maximum * simultaneous number of writebacks; from here * everything happens asynchronously.
*/
closure_call(&io->cl, read_dirty_submit, NULL, &cl);
}
/* * Returns true if we scanned the entire disk
*/ staticbool refill_dirty(struct cached_dev *dc)
{ struct keybuf *buf = &dc->writeback_keys; struct bkey start = KEY(dc->disk.id, 0, 0); struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0); struct bkey start_pos;
/* * make sure keybuf pos is inside the range for this disk - at bringup * we might not be attached yet so this disk's inode nr isn't * initialized then
*/ if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
bkey_cmp(&buf->last_scanned, &end) > 0)
buf->last_scanned = start;
if (dc->partial_stripes_expensive) {
refill_full_stripes(dc); if (array_freelist_empty(&buf->freelist)) returnfalse;
}
if (bkey_cmp(&buf->last_scanned, &end) < 0) returnfalse;
/* * If we get to the end start scanning again from the beginning, and * only scan up to where we initially started scanning from:
*/
buf->last_scanned = start;
bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
while (!kthread_should_stop() &&
!test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
down_write(&dc->writeback_lock);
set_current_state(TASK_INTERRUPTIBLE); /* * If the bache device is detaching, skip here and continue * to perform writeback. Otherwise, if no dirty data on cache, * or there is dirty data on cache but writeback is disabled, * the writeback thread should sleep here and wait for others * to wake up it.
*/ if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
(!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
up_write(&dc->writeback_lock);
if (kthread_should_stop() ||
test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
set_current_state(TASK_RUNNING); break;
}
if (searched_full_index &&
RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
atomic_set(&dc->has_dirty, 0);
SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
bch_write_bdev_super(dc, NULL); /* * If bcache device is detaching via sysfs interface, * writeback thread should stop after there is no dirty * data on cache. BCACHE_DEV_DETACHING flag is set in * bch_cached_dev_detach().
*/ if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) { struct closure cl;
/* * When dirty data rate is high (e.g. 50%+), there might * be heavy buckets fragmentation after writeback * finished, which hurts following write performance. * If users really care about write performance they * may set BCH_ENABLE_AUTO_GC via sysfs, then when * BCH_DO_AUTO_GC is set, garbage collection thread * will be wake up here. After moving gc, the shrunk * btree and discarded free buckets SSD space may be * helpful for following write requests.
*/ if (c->gc_after_writeback ==
(BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
force_wake_up_gc(c);
}
}
up_write(&dc->writeback_lock);
read_dirty(dc);
if (searched_full_index) { unsignedint delay = dc->writeback_delay * HZ;
/* * The op may be added to cache_set's btree_cache_wait * in mca_cannibalize(), must ensure it is removed from * the list and release btree_cache_alloc_lock before * free op memory. * Otherwise, the btree_cache_wait will be damaged.
*/
bch_cannibalize_unlock(c);
finish_wait(&c->btree_cache_wait, &(&op.op)->wait);
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.