linux/drivers/md/dm-cache-target.c

2665 lines
64 KiB
C
Raw Normal View History

/*
* Copyright (C) 2012 Red Hat. All rights reserved.
*
* This file is released under the GPL.
*/
#include "dm.h"
#include "dm-bio-prison.h"
dm cache: fix writes to cache device in writethrough mode The dm-cache writethrough strategy introduced by commit e2e74d617eadc15 ("dm cache: fix race in writethrough implementation") issues a bio to the origin device, remaps and then issues the bio to the cache device. This more conservative in-series approach was selected to favor correctness over performance (of the previous parallel writethrough). However, this in-series implementation that reuses the same bio to write both the origin and cache device didn't take into account that the block layer's req_bio_endio() modifies a completing bio's bi_sector and bi_size. So the new writethrough strategy needs to preserve these bio fields, and restore them before submission to the cache device, otherwise nothing gets written to the cache (because bi_size is 0). This patch adds a struct dm_bio_details field to struct per_bio_data, and uses dm_bio_record() and dm_bio_restore() to ensure the bio is restored before reissuing to the cache device. Adding such a large structure to the per_bio_data is not ideal but we can improve this later, for now correctness is the important thing. This problem initially went unnoticed because the dm-cache test-suite uses a linear DM device for the dm-cache device's origin device. Writethrough worked as expected because DM submits a *clone* of the original bio, so the original bio which was reused for the cache was never touched. Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-04-05 18:36:32 +04:00
#include "dm-bio-record.h"
#include "dm-cache-metadata.h"
#include <linux/dm-io.h>
#include <linux/dm-kcopyd.h>
#include <linux/init.h>
#include <linux/mempool.h>
#include <linux/module.h>
#include <linux/slab.h>
#include <linux/vmalloc.h>
#define DM_MSG_PREFIX "cache"
DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
"A percentage of time allocated for copying to and/or from cache");
/*----------------------------------------------------------------*/
/*
* Glossary:
*
* oblock: index of an origin block
* cblock: index of a cache block
* promotion: movement of a block from origin to cache
* demotion: movement of a block from cache to origin
* migration: movement of a block between the origin and cache device,
* either direction
*/
/*----------------------------------------------------------------*/
static size_t bitset_size_in_bytes(unsigned nr_entries)
{
return sizeof(unsigned long) * dm_div_up(nr_entries, BITS_PER_LONG);
}
static unsigned long *alloc_bitset(unsigned nr_entries)
{
size_t s = bitset_size_in_bytes(nr_entries);
return vzalloc(s);
}
static void clear_bitset(void *bitset, unsigned nr_entries)
{
size_t s = bitset_size_in_bytes(nr_entries);
memset(bitset, 0, s);
}
static void free_bitset(unsigned long *bits)
{
vfree(bits);
}
/*----------------------------------------------------------------*/
#define PRISON_CELLS 1024
#define MIGRATION_POOL_SIZE 128
#define COMMIT_PERIOD HZ
#define MIGRATION_COUNT_WINDOW 10
/*
* The block size of the device holding cache data must be >= 32KB
*/
#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
/*
* FIXME: the cache is read/write for the time being.
*/
enum cache_mode {
CM_WRITE, /* metadata may be changed */
CM_READ_ONLY, /* metadata may not be changed */
};
struct cache_features {
enum cache_mode mode;
bool write_through:1;
};
struct cache_stats {
atomic_t read_hit;
atomic_t read_miss;
atomic_t write_hit;
atomic_t write_miss;
atomic_t demotion;
atomic_t promotion;
atomic_t copies_avoided;
atomic_t cache_cell_clash;
atomic_t commit_count;
atomic_t discard_count;
};
struct cache {
struct dm_target *ti;
struct dm_target_callbacks callbacks;
/*
* Metadata is written to this device.
*/
struct dm_dev *metadata_dev;
/*
* The slower of the two data devices. Typically a spindle.
*/
struct dm_dev *origin_dev;
/*
* The faster of the two data devices. Typically an SSD.
*/
struct dm_dev *cache_dev;
/*
* Cache features such as write-through.
*/
struct cache_features features;
/*
* Size of the origin device in _complete_ blocks and native sectors.
*/
dm_oblock_t origin_blocks;
sector_t origin_sectors;
/*
* Size of the cache device in blocks.
*/
dm_cblock_t cache_size;
/*
* Fields for converting from sectors to blocks.
*/
uint32_t sectors_per_block;
int sectors_per_block_shift;
struct dm_cache_metadata *cmd;
spinlock_t lock;
struct bio_list deferred_bios;
struct bio_list deferred_flush_bios;
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
struct bio_list deferred_writethrough_bios;
struct list_head quiesced_migrations;
struct list_head completed_migrations;
struct list_head need_commit_migrations;
sector_t migration_threshold;
atomic_t nr_migrations;
wait_queue_head_t migration_wait;
/*
* cache_size entries, dirty if set
*/
dm_cblock_t nr_dirty;
unsigned long *dirty_bitset;
/*
* origin_blocks entries, discarded if set.
*/
uint32_t discard_block_size; /* a power of 2 times sectors per block */
dm_dblock_t discard_nr_blocks;
unsigned long *discard_bitset;
struct dm_kcopyd_client *copier;
struct workqueue_struct *wq;
struct work_struct worker;
struct delayed_work waker;
unsigned long last_commit_jiffies;
struct dm_bio_prison *prison;
struct dm_deferred_set *all_io_ds;
mempool_t *migration_pool;
struct dm_cache_migration *next_migration;
struct dm_cache_policy *policy;
unsigned policy_nr_args;
bool need_tick_bio:1;
bool sized:1;
bool quiescing:1;
bool commit_requested:1;
bool loaded_mappings:1;
bool loaded_discards:1;
struct cache_stats stats;
/*
* Rather than reconstructing the table line for the status we just
* save it and regurgitate.
*/
unsigned nr_ctr_args;
const char **ctr_args;
};
struct per_bio_data {
bool tick:1;
unsigned req_nr:2;
struct dm_deferred_entry *all_io_entry;
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
/*
* writethrough fields. These MUST remain at the end of this
* structure and the 'cache' member must be the first as it
* is used to determine the offset of the writethrough fields.
*/
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
struct cache *cache;
dm_cblock_t cblock;
bio_end_io_t *saved_bi_end_io;
dm cache: fix writes to cache device in writethrough mode The dm-cache writethrough strategy introduced by commit e2e74d617eadc15 ("dm cache: fix race in writethrough implementation") issues a bio to the origin device, remaps and then issues the bio to the cache device. This more conservative in-series approach was selected to favor correctness over performance (of the previous parallel writethrough). However, this in-series implementation that reuses the same bio to write both the origin and cache device didn't take into account that the block layer's req_bio_endio() modifies a completing bio's bi_sector and bi_size. So the new writethrough strategy needs to preserve these bio fields, and restore them before submission to the cache device, otherwise nothing gets written to the cache (because bi_size is 0). This patch adds a struct dm_bio_details field to struct per_bio_data, and uses dm_bio_record() and dm_bio_restore() to ensure the bio is restored before reissuing to the cache device. Adding such a large structure to the per_bio_data is not ideal but we can improve this later, for now correctness is the important thing. This problem initially went unnoticed because the dm-cache test-suite uses a linear DM device for the dm-cache device's origin device. Writethrough worked as expected because DM submits a *clone* of the original bio, so the original bio which was reused for the cache was never touched. Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-04-05 18:36:32 +04:00
struct dm_bio_details bio_details;
};
struct dm_cache_migration {
struct list_head list;
struct cache *cache;
unsigned long start_jiffies;
dm_oblock_t old_oblock;
dm_oblock_t new_oblock;
dm_cblock_t cblock;
bool err:1;
bool writeback:1;
bool demote:1;
bool promote:1;
struct dm_bio_prison_cell *old_ocell;
struct dm_bio_prison_cell *new_ocell;
};
/*
* Processing a bio in the worker thread may require these memory
* allocations. We prealloc to avoid deadlocks (the same worker thread
* frees them back to the mempool).
*/
struct prealloc {
struct dm_cache_migration *mg;
struct dm_bio_prison_cell *cell1;
struct dm_bio_prison_cell *cell2;
};
static void wake_worker(struct cache *cache)
{
queue_work(cache->wq, &cache->worker);
}
/*----------------------------------------------------------------*/
static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
{
/* FIXME: change to use a local slab. */
return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
}
static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
{
dm_bio_prison_free_cell(cache->prison, cell);
}
static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
{
if (!p->mg) {
p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
if (!p->mg)
return -ENOMEM;
}
if (!p->cell1) {
p->cell1 = alloc_prison_cell(cache);
if (!p->cell1)
return -ENOMEM;
}
if (!p->cell2) {
p->cell2 = alloc_prison_cell(cache);
if (!p->cell2)
return -ENOMEM;
}
return 0;
}
static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
{
if (p->cell2)
free_prison_cell(cache, p->cell2);
if (p->cell1)
free_prison_cell(cache, p->cell1);
if (p->mg)
mempool_free(p->mg, cache->migration_pool);
}
static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
{
struct dm_cache_migration *mg = p->mg;
BUG_ON(!mg);
p->mg = NULL;
return mg;
}
/*
* You must have a cell within the prealloc struct to return. If not this
* function will BUG() rather than returning NULL.
*/
static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
{
struct dm_bio_prison_cell *r = NULL;
if (p->cell1) {
r = p->cell1;
p->cell1 = NULL;
} else if (p->cell2) {
r = p->cell2;
p->cell2 = NULL;
} else
BUG();
return r;
}
/*
* You can't have more than two cells in a prealloc struct. BUG() will be
* called if you try and overfill.
*/
static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
{
if (!p->cell2)
p->cell2 = cell;
else if (!p->cell1)
p->cell1 = cell;
else
BUG();
}
/*----------------------------------------------------------------*/
static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
{
key->virtual = 0;
key->dev = 0;
key->block = from_oblock(oblock);
}
/*
* The caller hands in a preallocated cell, and a free function for it.
* The cell will be freed if there's an error, or if it wasn't used because
* a cell with that key already exists.
*/
typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
static int bio_detain(struct cache *cache, dm_oblock_t oblock,
struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
cell_free_fn free_fn, void *free_context,
struct dm_bio_prison_cell **cell_result)
{
int r;
struct dm_cell_key key;
build_key(oblock, &key);
r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
if (r)
free_fn(free_context, cell_prealloc);
return r;
}
static int get_cell(struct cache *cache,
dm_oblock_t oblock,
struct prealloc *structs,
struct dm_bio_prison_cell **cell_result)
{
int r;
struct dm_cell_key key;
struct dm_bio_prison_cell *cell_prealloc;
cell_prealloc = prealloc_get_cell(structs);
build_key(oblock, &key);
r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
if (r)
prealloc_put_cell(structs, cell_prealloc);
return r;
}
/*----------------------------------------------------------------*/
static bool is_dirty(struct cache *cache, dm_cblock_t b)
{
return test_bit(from_cblock(b), cache->dirty_bitset);
}
static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
{
if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) + 1);
policy_set_dirty(cache->policy, oblock);
}
}
static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
{
if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
policy_clear_dirty(cache->policy, oblock);
cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) - 1);
if (!from_cblock(cache->nr_dirty))
dm_table_event(cache->ti->table);
}
}
/*----------------------------------------------------------------*/
static bool block_size_is_power_of_two(struct cache *cache)
{
return cache->sectors_per_block_shift >= 0;
}
static dm_block_t block_div(dm_block_t b, uint32_t n)
{
do_div(b, n);
return b;
}
static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
{
uint32_t discard_blocks = cache->discard_block_size;
dm_block_t b = from_oblock(oblock);
if (!block_size_is_power_of_two(cache))
discard_blocks = discard_blocks / cache->sectors_per_block;
else
discard_blocks >>= cache->sectors_per_block_shift;
b = block_div(b, discard_blocks);
return to_dblock(b);
}
static void set_discard(struct cache *cache, dm_dblock_t b)
{
unsigned long flags;
atomic_inc(&cache->stats.discard_count);
spin_lock_irqsave(&cache->lock, flags);
set_bit(from_dblock(b), cache->discard_bitset);
spin_unlock_irqrestore(&cache->lock, flags);
}
static void clear_discard(struct cache *cache, dm_dblock_t b)
{
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
clear_bit(from_dblock(b), cache->discard_bitset);
spin_unlock_irqrestore(&cache->lock, flags);
}
static bool is_discarded(struct cache *cache, dm_dblock_t b)
{
int r;
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
r = test_bit(from_dblock(b), cache->discard_bitset);
spin_unlock_irqrestore(&cache->lock, flags);
return r;
}
static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
{
int r;
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
cache->discard_bitset);
spin_unlock_irqrestore(&cache->lock, flags);
return r;
}
/*----------------------------------------------------------------*/
static void load_stats(struct cache *cache)
{
struct dm_cache_statistics stats;
dm_cache_metadata_get_stats(cache->cmd, &stats);
atomic_set(&cache->stats.read_hit, stats.read_hits);
atomic_set(&cache->stats.read_miss, stats.read_misses);
atomic_set(&cache->stats.write_hit, stats.write_hits);
atomic_set(&cache->stats.write_miss, stats.write_misses);
}
static void save_stats(struct cache *cache)
{
struct dm_cache_statistics stats;
stats.read_hits = atomic_read(&cache->stats.read_hit);
stats.read_misses = atomic_read(&cache->stats.read_miss);
stats.write_hits = atomic_read(&cache->stats.write_hit);
stats.write_misses = atomic_read(&cache->stats.write_miss);
dm_cache_metadata_set_stats(cache->cmd, &stats);
}
/*----------------------------------------------------------------
* Per bio data
*--------------------------------------------------------------*/
/*
* If using writeback, leave out struct per_bio_data's writethrough fields.
*/
#define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
#define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
static size_t get_per_bio_data_size(struct cache *cache)
{
return cache->features.write_through ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
}
static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
{
struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
BUG_ON(!pb);
return pb;
}
static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
{
struct per_bio_data *pb = get_per_bio_data(bio, data_size);
pb->tick = false;
pb->req_nr = dm_bio_get_target_bio_nr(bio);
pb->all_io_entry = NULL;
return pb;
}
/*----------------------------------------------------------------
* Remapping
*--------------------------------------------------------------*/
static void remap_to_origin(struct cache *cache, struct bio *bio)
{
bio->bi_bdev = cache->origin_dev->bdev;
}
static void remap_to_cache(struct cache *cache, struct bio *bio,
dm_cblock_t cblock)
{
sector_t bi_sector = bio->bi_sector;
bio->bi_bdev = cache->cache_dev->bdev;
if (!block_size_is_power_of_two(cache))
bio->bi_sector = (from_cblock(cblock) * cache->sectors_per_block) +
sector_div(bi_sector, cache->sectors_per_block);
else
bio->bi_sector = (from_cblock(cblock) << cache->sectors_per_block_shift) |
(bi_sector & (cache->sectors_per_block - 1));
}
static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
{
unsigned long flags;
size_t pb_data_size = get_per_bio_data_size(cache);
struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
spin_lock_irqsave(&cache->lock, flags);
if (cache->need_tick_bio &&
!(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
pb->tick = true;
cache->need_tick_bio = false;
}
spin_unlock_irqrestore(&cache->lock, flags);
}
static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
dm_oblock_t oblock)
{
check_if_tick_bio_needed(cache, bio);
remap_to_origin(cache, bio);
if (bio_data_dir(bio) == WRITE)
clear_discard(cache, oblock_to_dblock(cache, oblock));
}
static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
dm_oblock_t oblock, dm_cblock_t cblock)
{
remap_to_cache(cache, bio, cblock);
if (bio_data_dir(bio) == WRITE) {
set_dirty(cache, oblock, cblock);
clear_discard(cache, oblock_to_dblock(cache, oblock));
}
}
static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
{
sector_t block_nr = bio->bi_sector;
if (!block_size_is_power_of_two(cache))
(void) sector_div(block_nr, cache->sectors_per_block);
else
block_nr >>= cache->sectors_per_block_shift;
return to_oblock(block_nr);
}
static int bio_triggers_commit(struct cache *cache, struct bio *bio)
{
return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
}
static void issue(struct cache *cache, struct bio *bio)
{
unsigned long flags;
if (!bio_triggers_commit(cache, bio)) {
generic_make_request(bio);
return;
}
/*
* Batch together any bios that trigger commits and then issue a
* single commit for them in do_worker().
*/
spin_lock_irqsave(&cache->lock, flags);
cache->commit_requested = true;
bio_list_add(&cache->deferred_flush_bios, bio);
spin_unlock_irqrestore(&cache->lock, flags);
}
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
{
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
bio_list_add(&cache->deferred_writethrough_bios, bio);
spin_unlock_irqrestore(&cache->lock, flags);
wake_worker(cache);
}
static void writethrough_endio(struct bio *bio, int err)
{
struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
bio->bi_end_io = pb->saved_bi_end_io;
if (err) {
bio_endio(bio, err);
return;
}
dm cache: fix writes to cache device in writethrough mode The dm-cache writethrough strategy introduced by commit e2e74d617eadc15 ("dm cache: fix race in writethrough implementation") issues a bio to the origin device, remaps and then issues the bio to the cache device. This more conservative in-series approach was selected to favor correctness over performance (of the previous parallel writethrough). However, this in-series implementation that reuses the same bio to write both the origin and cache device didn't take into account that the block layer's req_bio_endio() modifies a completing bio's bi_sector and bi_size. So the new writethrough strategy needs to preserve these bio fields, and restore them before submission to the cache device, otherwise nothing gets written to the cache (because bi_size is 0). This patch adds a struct dm_bio_details field to struct per_bio_data, and uses dm_bio_record() and dm_bio_restore() to ensure the bio is restored before reissuing to the cache device. Adding such a large structure to the per_bio_data is not ideal but we can improve this later, for now correctness is the important thing. This problem initially went unnoticed because the dm-cache test-suite uses a linear DM device for the dm-cache device's origin device. Writethrough worked as expected because DM submits a *clone* of the original bio, so the original bio which was reused for the cache was never touched. Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-04-05 18:36:32 +04:00
dm_bio_restore(&pb->bio_details, bio);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
remap_to_cache(pb->cache, bio, pb->cblock);
/*
* We can't issue this bio directly, since we're in interrupt
* context. So it gets put on a bio list for processing by the
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
* worker thread.
*/
defer_writethrough_bio(pb->cache, bio);
}
/*
* When running in writethrough mode we need to send writes to clean blocks
* to both the cache and origin devices. In future we'd like to clone the
* bio and send them in parallel, but for now we're doing them in
* series as this is easier.
*/
static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
dm_oblock_t oblock, dm_cblock_t cblock)
{
struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
pb->cache = cache;
pb->cblock = cblock;
pb->saved_bi_end_io = bio->bi_end_io;
dm cache: fix writes to cache device in writethrough mode The dm-cache writethrough strategy introduced by commit e2e74d617eadc15 ("dm cache: fix race in writethrough implementation") issues a bio to the origin device, remaps and then issues the bio to the cache device. This more conservative in-series approach was selected to favor correctness over performance (of the previous parallel writethrough). However, this in-series implementation that reuses the same bio to write both the origin and cache device didn't take into account that the block layer's req_bio_endio() modifies a completing bio's bi_sector and bi_size. So the new writethrough strategy needs to preserve these bio fields, and restore them before submission to the cache device, otherwise nothing gets written to the cache (because bi_size is 0). This patch adds a struct dm_bio_details field to struct per_bio_data, and uses dm_bio_record() and dm_bio_restore() to ensure the bio is restored before reissuing to the cache device. Adding such a large structure to the per_bio_data is not ideal but we can improve this later, for now correctness is the important thing. This problem initially went unnoticed because the dm-cache test-suite uses a linear DM device for the dm-cache device's origin device. Writethrough worked as expected because DM submits a *clone* of the original bio, so the original bio which was reused for the cache was never touched. Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-04-05 18:36:32 +04:00
dm_bio_record(&pb->bio_details, bio);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
bio->bi_end_io = writethrough_endio;
remap_to_origin_clear_discard(pb->cache, bio, oblock);
}
/*----------------------------------------------------------------
* Migration processing
*
* Migration covers moving data from the origin device to the cache, or
* vice versa.
*--------------------------------------------------------------*/
static void free_migration(struct dm_cache_migration *mg)
{
mempool_free(mg, mg->cache->migration_pool);
}
static void inc_nr_migrations(struct cache *cache)
{
atomic_inc(&cache->nr_migrations);
}
static void dec_nr_migrations(struct cache *cache)
{
atomic_dec(&cache->nr_migrations);
/*
* Wake the worker in case we're suspending the target.
*/
wake_up(&cache->migration_wait);
}
static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
bool holder)
{
(holder ? dm_cell_release : dm_cell_release_no_holder)
(cache->prison, cell, &cache->deferred_bios);
free_prison_cell(cache, cell);
}
static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
bool holder)
{
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
__cell_defer(cache, cell, holder);
spin_unlock_irqrestore(&cache->lock, flags);
wake_worker(cache);
}
static void cleanup_migration(struct dm_cache_migration *mg)
{
dec_nr_migrations(mg->cache);
free_migration(mg);
}
static void migration_failure(struct dm_cache_migration *mg)
{
struct cache *cache = mg->cache;
if (mg->writeback) {
DMWARN_LIMIT("writeback failed; couldn't copy block");
set_dirty(cache, mg->old_oblock, mg->cblock);
cell_defer(cache, mg->old_ocell, false);
} else if (mg->demote) {
DMWARN_LIMIT("demotion failed; couldn't copy block");
policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
if (mg->promote)
cell_defer(cache, mg->new_ocell, 1);
} else {
DMWARN_LIMIT("promotion failed; couldn't copy block");
policy_remove_mapping(cache->policy, mg->new_oblock);
cell_defer(cache, mg->new_ocell, 1);
}
cleanup_migration(mg);
}
static void migration_success_pre_commit(struct dm_cache_migration *mg)
{
unsigned long flags;
struct cache *cache = mg->cache;
if (mg->writeback) {
cell_defer(cache, mg->old_ocell, false);
clear_dirty(cache, mg->old_oblock, mg->cblock);
cleanup_migration(mg);
return;
} else if (mg->demote) {
if (dm_cache_remove_mapping(cache->cmd, mg->cblock)) {
DMWARN_LIMIT("demotion failed; couldn't update on disk metadata");
policy_force_mapping(cache->policy, mg->new_oblock,
mg->old_oblock);
if (mg->promote)
cell_defer(cache, mg->new_ocell, true);
cleanup_migration(mg);
return;
}
} else {
if (dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock)) {
DMWARN_LIMIT("promotion failed; couldn't update on disk metadata");
policy_remove_mapping(cache->policy, mg->new_oblock);
cleanup_migration(mg);
return;
}
}
spin_lock_irqsave(&cache->lock, flags);
list_add_tail(&mg->list, &cache->need_commit_migrations);
cache->commit_requested = true;
spin_unlock_irqrestore(&cache->lock, flags);
}
static void migration_success_post_commit(struct dm_cache_migration *mg)
{
unsigned long flags;
struct cache *cache = mg->cache;
if (mg->writeback) {
DMWARN("writeback unexpectedly triggered commit");
return;
} else if (mg->demote) {
cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
if (mg->promote) {
mg->demote = false;
spin_lock_irqsave(&cache->lock, flags);
list_add_tail(&mg->list, &cache->quiesced_migrations);
spin_unlock_irqrestore(&cache->lock, flags);
} else
cleanup_migration(mg);
} else {
cell_defer(cache, mg->new_ocell, true);
clear_dirty(cache, mg->new_oblock, mg->cblock);
cleanup_migration(mg);
}
}
static void copy_complete(int read_err, unsigned long write_err, void *context)
{
unsigned long flags;
struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
struct cache *cache = mg->cache;
if (read_err || write_err)
mg->err = true;
spin_lock_irqsave(&cache->lock, flags);
list_add_tail(&mg->list, &cache->completed_migrations);
spin_unlock_irqrestore(&cache->lock, flags);
wake_worker(cache);
}
static void issue_copy_real(struct dm_cache_migration *mg)
{
int r;
struct dm_io_region o_region, c_region;
struct cache *cache = mg->cache;
o_region.bdev = cache->origin_dev->bdev;
o_region.count = cache->sectors_per_block;
c_region.bdev = cache->cache_dev->bdev;
c_region.sector = from_cblock(mg->cblock) * cache->sectors_per_block;
c_region.count = cache->sectors_per_block;
if (mg->writeback || mg->demote) {
/* demote */
o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
} else {
/* promote */
o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
}
if (r < 0)
migration_failure(mg);
}
static void avoid_copy(struct dm_cache_migration *mg)
{
atomic_inc(&mg->cache->stats.copies_avoided);
migration_success_pre_commit(mg);
}
static void issue_copy(struct dm_cache_migration *mg)
{
bool avoid;
struct cache *cache = mg->cache;
if (mg->writeback || mg->demote)
avoid = !is_dirty(cache, mg->cblock) ||
is_discarded_oblock(cache, mg->old_oblock);
else
avoid = is_discarded_oblock(cache, mg->new_oblock);
avoid ? avoid_copy(mg) : issue_copy_real(mg);
}
static void complete_migration(struct dm_cache_migration *mg)
{
if (mg->err)
migration_failure(mg);
else
migration_success_pre_commit(mg);
}
static void process_migrations(struct cache *cache, struct list_head *head,
void (*fn)(struct dm_cache_migration *))
{
unsigned long flags;
struct list_head list;
struct dm_cache_migration *mg, *tmp;
INIT_LIST_HEAD(&list);
spin_lock_irqsave(&cache->lock, flags);
list_splice_init(head, &list);
spin_unlock_irqrestore(&cache->lock, flags);
list_for_each_entry_safe(mg, tmp, &list, list)
fn(mg);
}
static void __queue_quiesced_migration(struct dm_cache_migration *mg)
{
list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
}
static void queue_quiesced_migration(struct dm_cache_migration *mg)
{
unsigned long flags;
struct cache *cache = mg->cache;
spin_lock_irqsave(&cache->lock, flags);
__queue_quiesced_migration(mg);
spin_unlock_irqrestore(&cache->lock, flags);
wake_worker(cache);
}
static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
{
unsigned long flags;
struct dm_cache_migration *mg, *tmp;
spin_lock_irqsave(&cache->lock, flags);
list_for_each_entry_safe(mg, tmp, work, list)
__queue_quiesced_migration(mg);
spin_unlock_irqrestore(&cache->lock, flags);
wake_worker(cache);
}
static void check_for_quiesced_migrations(struct cache *cache,
struct per_bio_data *pb)
{
struct list_head work;
if (!pb->all_io_entry)
return;
INIT_LIST_HEAD(&work);
if (pb->all_io_entry)
dm_deferred_entry_dec(pb->all_io_entry, &work);
if (!list_empty(&work))
queue_quiesced_migrations(cache, &work);
}
static void quiesce_migration(struct dm_cache_migration *mg)
{
if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
queue_quiesced_migration(mg);
}
static void promote(struct cache *cache, struct prealloc *structs,
dm_oblock_t oblock, dm_cblock_t cblock,
struct dm_bio_prison_cell *cell)
{
struct dm_cache_migration *mg = prealloc_get_migration(structs);
mg->err = false;
mg->writeback = false;
mg->demote = false;
mg->promote = true;
mg->cache = cache;
mg->new_oblock = oblock;
mg->cblock = cblock;
mg->old_ocell = NULL;
mg->new_ocell = cell;
mg->start_jiffies = jiffies;
inc_nr_migrations(cache);
quiesce_migration(mg);
}
static void writeback(struct cache *cache, struct prealloc *structs,
dm_oblock_t oblock, dm_cblock_t cblock,
struct dm_bio_prison_cell *cell)
{
struct dm_cache_migration *mg = prealloc_get_migration(structs);
mg->err = false;
mg->writeback = true;
mg->demote = false;
mg->promote = false;
mg->cache = cache;
mg->old_oblock = oblock;
mg->cblock = cblock;
mg->old_ocell = cell;
mg->new_ocell = NULL;
mg->start_jiffies = jiffies;
inc_nr_migrations(cache);
quiesce_migration(mg);
}
static void demote_then_promote(struct cache *cache, struct prealloc *structs,
dm_oblock_t old_oblock, dm_oblock_t new_oblock,
dm_cblock_t cblock,
struct dm_bio_prison_cell *old_ocell,
struct dm_bio_prison_cell *new_ocell)
{
struct dm_cache_migration *mg = prealloc_get_migration(structs);
mg->err = false;
mg->writeback = false;
mg->demote = true;
mg->promote = true;
mg->cache = cache;
mg->old_oblock = old_oblock;
mg->new_oblock = new_oblock;
mg->cblock = cblock;
mg->old_ocell = old_ocell;
mg->new_ocell = new_ocell;
mg->start_jiffies = jiffies;
inc_nr_migrations(cache);
quiesce_migration(mg);
}
/*----------------------------------------------------------------
* bio processing
*--------------------------------------------------------------*/
static void defer_bio(struct cache *cache, struct bio *bio)
{
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
bio_list_add(&cache->deferred_bios, bio);
spin_unlock_irqrestore(&cache->lock, flags);
wake_worker(cache);
}
static void process_flush_bio(struct cache *cache, struct bio *bio)
{
size_t pb_data_size = get_per_bio_data_size(cache);
struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
BUG_ON(bio->bi_size);
if (!pb->req_nr)
remap_to_origin(cache, bio);
else
remap_to_cache(cache, bio, 0);
issue(cache, bio);
}
/*
* People generally discard large parts of a device, eg, the whole device
* when formatting. Splitting these large discards up into cache block
* sized ios and then quiescing (always neccessary for discard) takes too
* long.
*
* We keep it simple, and allow any size of discard to come in, and just
* mark off blocks on the discard bitset. No passdown occurs!
*
* To implement passdown we need to change the bio_prison such that a cell
* can have a key that spans many blocks.
*/
static void process_discard_bio(struct cache *cache, struct bio *bio)
{
dm_block_t start_block = dm_sector_div_up(bio->bi_sector,
cache->discard_block_size);
dm_block_t end_block = bio->bi_sector + bio_sectors(bio);
dm_block_t b;
end_block = block_div(end_block, cache->discard_block_size);
for (b = start_block; b < end_block; b++)
set_discard(cache, to_dblock(b));
bio_endio(bio, 0);
}
static bool spare_migration_bandwidth(struct cache *cache)
{
sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
cache->sectors_per_block;
return current_volume < cache->migration_threshold;
}
static bool is_writethrough_io(struct cache *cache, struct bio *bio,
dm_cblock_t cblock)
{
return bio_data_dir(bio) == WRITE &&
cache->features.write_through && !is_dirty(cache, cblock);
}
static void inc_hit_counter(struct cache *cache, struct bio *bio)
{
atomic_inc(bio_data_dir(bio) == READ ?
&cache->stats.read_hit : &cache->stats.write_hit);
}
static void inc_miss_counter(struct cache *cache, struct bio *bio)
{
atomic_inc(bio_data_dir(bio) == READ ?
&cache->stats.read_miss : &cache->stats.write_miss);
}
static void process_bio(struct cache *cache, struct prealloc *structs,
struct bio *bio)
{
int r;
bool release_cell = true;
dm_oblock_t block = get_bio_block(cache, bio);
struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
struct policy_result lookup_result;
size_t pb_data_size = get_per_bio_data_size(cache);
struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
bool discarded_block = is_discarded_oblock(cache, block);
bool can_migrate = discarded_block || spare_migration_bandwidth(cache);
/*
* Check to see if that block is currently migrating.
*/
cell_prealloc = prealloc_get_cell(structs);
r = bio_detain(cache, block, bio, cell_prealloc,
(cell_free_fn) prealloc_put_cell,
structs, &new_ocell);
if (r > 0)
return;
r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
bio, &lookup_result);
if (r == -EWOULDBLOCK)
/* migration has been denied */
lookup_result.op = POLICY_MISS;
switch (lookup_result.op) {
case POLICY_HIT:
inc_hit_counter(cache, bio);
pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
if (is_writethrough_io(cache, bio, lookup_result.cblock))
remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
else
remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
issue(cache, bio);
break;
case POLICY_MISS:
inc_miss_counter(cache, bio);
pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
remap_to_origin_clear_discard(cache, bio, block);
issue(cache, bio);
break;
case POLICY_NEW:
atomic_inc(&cache->stats.promotion);
promote(cache, structs, block, lookup_result.cblock, new_ocell);
release_cell = false;
break;
case POLICY_REPLACE:
cell_prealloc = prealloc_get_cell(structs);
r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
(cell_free_fn) prealloc_put_cell,
structs, &old_ocell);
if (r > 0) {
/*
* We have to be careful to avoid lock inversion of
* the cells. So we back off, and wait for the
* old_ocell to become free.
*/
policy_force_mapping(cache->policy, block,
lookup_result.old_oblock);
atomic_inc(&cache->stats.cache_cell_clash);
break;
}
atomic_inc(&cache->stats.demotion);
atomic_inc(&cache->stats.promotion);
demote_then_promote(cache, structs, lookup_result.old_oblock,
block, lookup_result.cblock,
old_ocell, new_ocell);
release_cell = false;
break;
default:
DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
(unsigned) lookup_result.op);
bio_io_error(bio);
}
if (release_cell)
cell_defer(cache, new_ocell, false);
}
static int need_commit_due_to_time(struct cache *cache)
{
return jiffies < cache->last_commit_jiffies ||
jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
}
static int commit_if_needed(struct cache *cache)
{
if (dm_cache_changed_this_transaction(cache->cmd) &&
(cache->commit_requested || need_commit_due_to_time(cache))) {
atomic_inc(&cache->stats.commit_count);
cache->last_commit_jiffies = jiffies;
cache->commit_requested = false;
return dm_cache_commit(cache->cmd, false);
}
return 0;
}
static void process_deferred_bios(struct cache *cache)
{
unsigned long flags;
struct bio_list bios;
struct bio *bio;
struct prealloc structs;
memset(&structs, 0, sizeof(structs));
bio_list_init(&bios);
spin_lock_irqsave(&cache->lock, flags);
bio_list_merge(&bios, &cache->deferred_bios);
bio_list_init(&cache->deferred_bios);
spin_unlock_irqrestore(&cache->lock, flags);
while (!bio_list_empty(&bios)) {
/*
* If we've got no free migration structs, and processing
* this bio might require one, we pause until there are some
* prepared mappings to process.
*/
if (prealloc_data_structs(cache, &structs)) {
spin_lock_irqsave(&cache->lock, flags);
bio_list_merge(&cache->deferred_bios, &bios);
spin_unlock_irqrestore(&cache->lock, flags);
break;
}
bio = bio_list_pop(&bios);
if (bio->bi_rw & REQ_FLUSH)
process_flush_bio(cache, bio);
else if (bio->bi_rw & REQ_DISCARD)
process_discard_bio(cache, bio);
else
process_bio(cache, &structs, bio);
}
prealloc_free_structs(cache, &structs);
}
static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
{
unsigned long flags;
struct bio_list bios;
struct bio *bio;
bio_list_init(&bios);
spin_lock_irqsave(&cache->lock, flags);
bio_list_merge(&bios, &cache->deferred_flush_bios);
bio_list_init(&cache->deferred_flush_bios);
spin_unlock_irqrestore(&cache->lock, flags);
while ((bio = bio_list_pop(&bios)))
submit_bios ? generic_make_request(bio) : bio_io_error(bio);
}
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
static void process_deferred_writethrough_bios(struct cache *cache)
{
unsigned long flags;
struct bio_list bios;
struct bio *bio;
bio_list_init(&bios);
spin_lock_irqsave(&cache->lock, flags);
bio_list_merge(&bios, &cache->deferred_writethrough_bios);
bio_list_init(&cache->deferred_writethrough_bios);
spin_unlock_irqrestore(&cache->lock, flags);
while ((bio = bio_list_pop(&bios)))
generic_make_request(bio);
}
static void writeback_some_dirty_blocks(struct cache *cache)
{
int r = 0;
dm_oblock_t oblock;
dm_cblock_t cblock;
struct prealloc structs;
struct dm_bio_prison_cell *old_ocell;
memset(&structs, 0, sizeof(structs));
while (spare_migration_bandwidth(cache)) {
if (prealloc_data_structs(cache, &structs))
break;
r = policy_writeback_work(cache->policy, &oblock, &cblock);
if (r)
break;
r = get_cell(cache, oblock, &structs, &old_ocell);
if (r) {
policy_set_dirty(cache->policy, oblock);
break;
}
writeback(cache, &structs, oblock, cblock, old_ocell);
}
prealloc_free_structs(cache, &structs);
}
/*----------------------------------------------------------------
* Main worker loop
*--------------------------------------------------------------*/
static void start_quiescing(struct cache *cache)
{
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
cache->quiescing = 1;
spin_unlock_irqrestore(&cache->lock, flags);
}
static void stop_quiescing(struct cache *cache)
{
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
cache->quiescing = 0;
spin_unlock_irqrestore(&cache->lock, flags);
}
static bool is_quiescing(struct cache *cache)
{
int r;
unsigned long flags;
spin_lock_irqsave(&cache->lock, flags);
r = cache->quiescing;
spin_unlock_irqrestore(&cache->lock, flags);
return r;
}
static void wait_for_migrations(struct cache *cache)
{
wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
}
static void stop_worker(struct cache *cache)
{
cancel_delayed_work(&cache->waker);
flush_workqueue(cache->wq);
}
static void requeue_deferred_io(struct cache *cache)
{
struct bio *bio;
struct bio_list bios;
bio_list_init(&bios);
bio_list_merge(&bios, &cache->deferred_bios);
bio_list_init(&cache->deferred_bios);
while ((bio = bio_list_pop(&bios)))
bio_endio(bio, DM_ENDIO_REQUEUE);
}
static int more_work(struct cache *cache)
{
if (is_quiescing(cache))
return !list_empty(&cache->quiesced_migrations) ||
!list_empty(&cache->completed_migrations) ||
!list_empty(&cache->need_commit_migrations);
else
return !bio_list_empty(&cache->deferred_bios) ||
!bio_list_empty(&cache->deferred_flush_bios) ||
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
!bio_list_empty(&cache->deferred_writethrough_bios) ||
!list_empty(&cache->quiesced_migrations) ||
!list_empty(&cache->completed_migrations) ||
!list_empty(&cache->need_commit_migrations);
}
static void do_worker(struct work_struct *ws)
{
struct cache *cache = container_of(ws, struct cache, worker);
do {
if (!is_quiescing(cache))
process_deferred_bios(cache);
process_migrations(cache, &cache->quiesced_migrations, issue_copy);
process_migrations(cache, &cache->completed_migrations, complete_migration);
writeback_some_dirty_blocks(cache);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
process_deferred_writethrough_bios(cache);
if (commit_if_needed(cache)) {
process_deferred_flush_bios(cache, false);
/*
* FIXME: rollback metadata or just go into a
* failure mode and error everything
*/
} else {
process_deferred_flush_bios(cache, true);
process_migrations(cache, &cache->need_commit_migrations,
migration_success_post_commit);
}
} while (more_work(cache));
}
/*
* We want to commit periodically so that not too much
* unwritten metadata builds up.
*/
static void do_waker(struct work_struct *ws)
{
struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
policy_tick(cache->policy);
wake_worker(cache);
queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
}
/*----------------------------------------------------------------*/
static int is_congested(struct dm_dev *dev, int bdi_bits)
{
struct request_queue *q = bdev_get_queue(dev->bdev);
return bdi_congested(&q->backing_dev_info, bdi_bits);
}
static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
{
struct cache *cache = container_of(cb, struct cache, callbacks);
return is_congested(cache->origin_dev, bdi_bits) ||
is_congested(cache->cache_dev, bdi_bits);
}
/*----------------------------------------------------------------
* Target methods
*--------------------------------------------------------------*/
/*
* This function gets called on the error paths of the constructor, so we
* have to cope with a partially initialised struct.
*/
static void destroy(struct cache *cache)
{
unsigned i;
if (cache->next_migration)
mempool_free(cache->next_migration, cache->migration_pool);
if (cache->migration_pool)
mempool_destroy(cache->migration_pool);
if (cache->all_io_ds)
dm_deferred_set_destroy(cache->all_io_ds);
if (cache->prison)
dm_bio_prison_destroy(cache->prison);
if (cache->wq)
destroy_workqueue(cache->wq);
if (cache->dirty_bitset)
free_bitset(cache->dirty_bitset);
if (cache->discard_bitset)
free_bitset(cache->discard_bitset);
if (cache->copier)
dm_kcopyd_client_destroy(cache->copier);
if (cache->cmd)
dm_cache_metadata_close(cache->cmd);
if (cache->metadata_dev)
dm_put_device(cache->ti, cache->metadata_dev);
if (cache->origin_dev)
dm_put_device(cache->ti, cache->origin_dev);
if (cache->cache_dev)
dm_put_device(cache->ti, cache->cache_dev);
if (cache->policy)
dm_cache_policy_destroy(cache->policy);
for (i = 0; i < cache->nr_ctr_args ; i++)
kfree(cache->ctr_args[i]);
kfree(cache->ctr_args);
kfree(cache);
}
static void cache_dtr(struct dm_target *ti)
{
struct cache *cache = ti->private;
destroy(cache);
}
static sector_t get_dev_size(struct dm_dev *dev)
{
return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
}
/*----------------------------------------------------------------*/
/*
* Construct a cache device mapping.
*
* cache <metadata dev> <cache dev> <origin dev> <block size>
* <#feature args> [<feature arg>]*
* <policy> <#policy args> [<policy arg>]*
*
* metadata dev : fast device holding the persistent metadata
* cache dev : fast device holding cached data blocks
* origin dev : slow device holding original data blocks
* block size : cache unit size in sectors
*
* #feature args : number of feature arguments passed
* feature args : writethrough. (The default is writeback.)
*
* policy : the replacement policy to use
* #policy args : an even number of policy arguments corresponding
* to key/value pairs passed to the policy
* policy args : key/value pairs passed to the policy
* E.g. 'sequential_threshold 1024'
* See cache-policies.txt for details.
*
* Optional feature arguments are:
* writethrough : write through caching that prohibits cache block
* content from being different from origin block content.
* Without this argument, the default behaviour is to write
* back cache block contents later for performance reasons,
* so they may differ from the corresponding origin blocks.
*/
struct cache_args {
struct dm_target *ti;
struct dm_dev *metadata_dev;
struct dm_dev *cache_dev;
sector_t cache_sectors;
struct dm_dev *origin_dev;
sector_t origin_sectors;
uint32_t block_size;
const char *policy_name;
int policy_argc;
const char **policy_argv;
struct cache_features features;
};
static void destroy_cache_args(struct cache_args *ca)
{
if (ca->metadata_dev)
dm_put_device(ca->ti, ca->metadata_dev);
if (ca->cache_dev)
dm_put_device(ca->ti, ca->cache_dev);
if (ca->origin_dev)
dm_put_device(ca->ti, ca->origin_dev);
kfree(ca);
}
static bool at_least_one_arg(struct dm_arg_set *as, char **error)
{
if (!as->argc) {
*error = "Insufficient args";
return false;
}
return true;
}
static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
char **error)
{
int r;
sector_t metadata_dev_size;
char b[BDEVNAME_SIZE];
if (!at_least_one_arg(as, error))
return -EINVAL;
r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
&ca->metadata_dev);
if (r) {
*error = "Error opening metadata device";
return r;
}
metadata_dev_size = get_dev_size(ca->metadata_dev);
if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
return 0;
}
static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
char **error)
{
int r;
if (!at_least_one_arg(as, error))
return -EINVAL;
r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
&ca->cache_dev);
if (r) {
*error = "Error opening cache device";
return r;
}
ca->cache_sectors = get_dev_size(ca->cache_dev);
return 0;
}
static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
char **error)
{
int r;
if (!at_least_one_arg(as, error))
return -EINVAL;
r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
&ca->origin_dev);
if (r) {
*error = "Error opening origin device";
return r;
}
ca->origin_sectors = get_dev_size(ca->origin_dev);
if (ca->ti->len > ca->origin_sectors) {
*error = "Device size larger than cached device";
return -EINVAL;
}
return 0;
}
static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
char **error)
{
unsigned long tmp;
if (!at_least_one_arg(as, error))
return -EINVAL;
if (kstrtoul(dm_shift_arg(as), 10, &tmp) || !tmp ||
tmp < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
tmp & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
*error = "Invalid data block size";
return -EINVAL;
}
if (tmp > ca->cache_sectors) {
*error = "Data block size is larger than the cache device";
return -EINVAL;
}
ca->block_size = tmp;
return 0;
}
static void init_features(struct cache_features *cf)
{
cf->mode = CM_WRITE;
cf->write_through = false;
}
static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
char **error)
{
static struct dm_arg _args[] = {
{0, 1, "Invalid number of cache feature arguments"},
};
int r;
unsigned argc;
const char *arg;
struct cache_features *cf = &ca->features;
init_features(cf);
r = dm_read_arg_group(_args, as, &argc, error);
if (r)
return -EINVAL;
while (argc--) {
arg = dm_shift_arg(as);
if (!strcasecmp(arg, "writeback"))
cf->write_through = false;
else if (!strcasecmp(arg, "writethrough"))
cf->write_through = true;
else {
*error = "Unrecognised cache feature requested";
return -EINVAL;
}
}
return 0;
}
static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
char **error)
{
static struct dm_arg _args[] = {
{0, 1024, "Invalid number of policy arguments"},
};
int r;
if (!at_least_one_arg(as, error))
return -EINVAL;
ca->policy_name = dm_shift_arg(as);
r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
if (r)
return -EINVAL;
ca->policy_argv = (const char **)as->argv;
dm_consume_args(as, ca->policy_argc);
return 0;
}
static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
char **error)
{
int r;
struct dm_arg_set as;
as.argc = argc;
as.argv = argv;
r = parse_metadata_dev(ca, &as, error);
if (r)
return r;
r = parse_cache_dev(ca, &as, error);
if (r)
return r;
r = parse_origin_dev(ca, &as, error);
if (r)
return r;
r = parse_block_size(ca, &as, error);
if (r)
return r;
r = parse_features(ca, &as, error);
if (r)
return r;
r = parse_policy(ca, &as, error);
if (r)
return r;
return 0;
}
/*----------------------------------------------------------------*/
static struct kmem_cache *migration_cache;
#define NOT_CORE_OPTION 1
static int process_config_option(struct cache *cache, const char *key, const char *value)
{
unsigned long tmp;
if (!strcasecmp(key, "migration_threshold")) {
if (kstrtoul(value, 10, &tmp))
return -EINVAL;
cache->migration_threshold = tmp;
return 0;
}
return NOT_CORE_OPTION;
}
static int set_config_value(struct cache *cache, const char *key, const char *value)
{
int r = process_config_option(cache, key, value);
if (r == NOT_CORE_OPTION)
r = policy_set_config_value(cache->policy, key, value);
if (r)
DMWARN("bad config value for %s: %s", key, value);
return r;
}
static int set_config_values(struct cache *cache, int argc, const char **argv)
{
int r = 0;
if (argc & 1) {
DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
return -EINVAL;
}
while (argc) {
r = set_config_value(cache, argv[0], argv[1]);
if (r)
break;
argc -= 2;
argv += 2;
}
return r;
}
static int create_cache_policy(struct cache *cache, struct cache_args *ca,
char **error)
{
cache->policy = dm_cache_policy_create(ca->policy_name,
cache->cache_size,
cache->origin_sectors,
cache->sectors_per_block);
if (!cache->policy) {
*error = "Error creating cache's policy";
return -ENOMEM;
}
return 0;
}
/*
* We want the discard block size to be a power of two, at least the size
* of the cache block size, and have no more than 2^14 discard blocks
* across the origin.
*/
#define MAX_DISCARD_BLOCKS (1 << 14)
static bool too_many_discard_blocks(sector_t discard_block_size,
sector_t origin_size)
{
(void) sector_div(origin_size, discard_block_size);
return origin_size > MAX_DISCARD_BLOCKS;
}
static sector_t calculate_discard_block_size(sector_t cache_block_size,
sector_t origin_size)
{
sector_t discard_block_size;
discard_block_size = roundup_pow_of_two(cache_block_size);
if (origin_size)
while (too_many_discard_blocks(discard_block_size, origin_size))
discard_block_size *= 2;
return discard_block_size;
}
#define DEFAULT_MIGRATION_THRESHOLD 2048
static int cache_create(struct cache_args *ca, struct cache **result)
{
int r = 0;
char **error = &ca->ti->error;
struct cache *cache;
struct dm_target *ti = ca->ti;
dm_block_t origin_blocks;
struct dm_cache_metadata *cmd;
bool may_format = ca->features.mode == CM_WRITE;
cache = kzalloc(sizeof(*cache), GFP_KERNEL);
if (!cache)
return -ENOMEM;
cache->ti = ca->ti;
ti->private = cache;
ti->num_flush_bios = 2;
ti->flush_supported = true;
ti->num_discard_bios = 1;
ti->discards_supported = true;
ti->discard_zeroes_data_unsupported = true;
cache->features = ca->features;
ti->per_bio_data_size = get_per_bio_data_size(cache);
cache->callbacks.congested_fn = cache_is_congested;
dm_table_add_target_callbacks(ti->table, &cache->callbacks);
cache->metadata_dev = ca->metadata_dev;
cache->origin_dev = ca->origin_dev;
cache->cache_dev = ca->cache_dev;
ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
/* FIXME: factor out this whole section */
origin_blocks = cache->origin_sectors = ca->origin_sectors;
origin_blocks = block_div(origin_blocks, ca->block_size);
cache->origin_blocks = to_oblock(origin_blocks);
cache->sectors_per_block = ca->block_size;
if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
r = -EINVAL;
goto bad;
}
if (ca->block_size & (ca->block_size - 1)) {
dm_block_t cache_size = ca->cache_sectors;
cache->sectors_per_block_shift = -1;
cache_size = block_div(cache_size, ca->block_size);
cache->cache_size = to_cblock(cache_size);
} else {
cache->sectors_per_block_shift = __ffs(ca->block_size);
cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
}
r = create_cache_policy(cache, ca, error);
if (r)
goto bad;
cache->policy_nr_args = ca->policy_argc;
cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
if (r) {
*error = "Error setting cache policy's config values";
goto bad;
}
cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
ca->block_size, may_format,
dm_cache_policy_get_hint_size(cache->policy));
if (IS_ERR(cmd)) {
*error = "Error creating metadata object";
r = PTR_ERR(cmd);
goto bad;
}
cache->cmd = cmd;
spin_lock_init(&cache->lock);
bio_list_init(&cache->deferred_bios);
bio_list_init(&cache->deferred_flush_bios);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
bio_list_init(&cache->deferred_writethrough_bios);
INIT_LIST_HEAD(&cache->quiesced_migrations);
INIT_LIST_HEAD(&cache->completed_migrations);
INIT_LIST_HEAD(&cache->need_commit_migrations);
atomic_set(&cache->nr_migrations, 0);
init_waitqueue_head(&cache->migration_wait);
r = -ENOMEM;
cache->nr_dirty = 0;
cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
if (!cache->dirty_bitset) {
*error = "could not allocate dirty bitset";
goto bad;
}
clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
cache->discard_block_size =
calculate_discard_block_size(cache->sectors_per_block,
cache->origin_sectors);
cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
if (!cache->discard_bitset) {
*error = "could not allocate discard bitset";
goto bad;
}
clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
if (IS_ERR(cache->copier)) {
*error = "could not create kcopyd client";
r = PTR_ERR(cache->copier);
goto bad;
}
cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
if (!cache->wq) {
*error = "could not create workqueue for metadata object";
goto bad;
}
INIT_WORK(&cache->worker, do_worker);
INIT_DELAYED_WORK(&cache->waker, do_waker);
cache->last_commit_jiffies = jiffies;
cache->prison = dm_bio_prison_create(PRISON_CELLS);
if (!cache->prison) {
*error = "could not create bio prison";
goto bad;
}
cache->all_io_ds = dm_deferred_set_create();
if (!cache->all_io_ds) {
*error = "could not create all_io deferred set";
goto bad;
}
cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
migration_cache);
if (!cache->migration_pool) {
*error = "Error creating cache's migration mempool";
goto bad;
}
cache->next_migration = NULL;
cache->need_tick_bio = true;
cache->sized = false;
cache->quiescing = false;
cache->commit_requested = false;
cache->loaded_mappings = false;
cache->loaded_discards = false;
load_stats(cache);
atomic_set(&cache->stats.demotion, 0);
atomic_set(&cache->stats.promotion, 0);
atomic_set(&cache->stats.copies_avoided, 0);
atomic_set(&cache->stats.cache_cell_clash, 0);
atomic_set(&cache->stats.commit_count, 0);
atomic_set(&cache->stats.discard_count, 0);
*result = cache;
return 0;
bad:
destroy(cache);
return r;
}
static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
{
unsigned i;
const char **copy;
copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
if (!copy)
return -ENOMEM;
for (i = 0; i < argc; i++) {
copy[i] = kstrdup(argv[i], GFP_KERNEL);
if (!copy[i]) {
while (i--)
kfree(copy[i]);
kfree(copy);
return -ENOMEM;
}
}
cache->nr_ctr_args = argc;
cache->ctr_args = copy;
return 0;
}
static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
{
int r = -EINVAL;
struct cache_args *ca;
struct cache *cache = NULL;
ca = kzalloc(sizeof(*ca), GFP_KERNEL);
if (!ca) {
ti->error = "Error allocating memory for cache";
return -ENOMEM;
}
ca->ti = ti;
r = parse_cache_args(ca, argc, argv, &ti->error);
if (r)
goto out;
r = cache_create(ca, &cache);
if (r)
goto out;
r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
if (r) {
destroy(cache);
goto out;
}
ti->private = cache;
out:
destroy_cache_args(ca);
return r;
}
static int cache_map(struct dm_target *ti, struct bio *bio)
{
struct cache *cache = ti->private;
int r;
dm_oblock_t block = get_bio_block(cache, bio);
size_t pb_data_size = get_per_bio_data_size(cache);
bool can_migrate = false;
bool discarded_block;
struct dm_bio_prison_cell *cell;
struct policy_result lookup_result;
struct per_bio_data *pb;
if (from_oblock(block) > from_oblock(cache->origin_blocks)) {
/*
* This can only occur if the io goes to a partial block at
* the end of the origin device. We don't cache these.
* Just remap to the origin and carry on.
*/
remap_to_origin_clear_discard(cache, bio, block);
return DM_MAPIO_REMAPPED;
}
pb = init_per_bio_data(bio, pb_data_size);
if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
defer_bio(cache, bio);
return DM_MAPIO_SUBMITTED;
}
/*
* Check to see if that block is currently migrating.
*/
cell = alloc_prison_cell(cache);
if (!cell) {
defer_bio(cache, bio);
return DM_MAPIO_SUBMITTED;
}
r = bio_detain(cache, block, bio, cell,
(cell_free_fn) free_prison_cell,
cache, &cell);
if (r) {
if (r < 0)
defer_bio(cache, bio);
return DM_MAPIO_SUBMITTED;
}
discarded_block = is_discarded_oblock(cache, block);
r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
bio, &lookup_result);
if (r == -EWOULDBLOCK) {
cell_defer(cache, cell, true);
return DM_MAPIO_SUBMITTED;
} else if (r) {
DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
bio_io_error(bio);
return DM_MAPIO_SUBMITTED;
}
switch (lookup_result.op) {
case POLICY_HIT:
inc_hit_counter(cache, bio);
pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
if (is_writethrough_io(cache, bio, lookup_result.cblock))
remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
else
remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
dm cache: fix race in writethrough implementation We have found a race in the optimisation used in the dm cache writethrough implementation. Currently, dm core sends the cache target two bios, one for the origin device and one for the cache device and these are processed in parallel. This patch avoids the race by changing the code back to a simpler (slower) implementation which processes the two writes in series, one after the other, until we can develop a complete fix for the problem. When the cache is in writethrough mode it needs to send WRITE bios to both the origin and cache devices. Previously we've been implementing this by having dm core query the cache target on every write to find out how many copies of the bio it wants. The cache will ask for two bios if the block is in the cache, and one otherwise. Then main problem with this is it's racey. At the time this check is made the bio hasn't yet been submitted and so isn't being taken into account when quiescing a block for migration (promotion or demotion). This means a single bio may be submitted when two were needed because the block has since been promoted to the cache (catastrophic), or two bios where only one is needed (harmless). I really don't want to start entering bios into the quiescing system (deferred_set) in the get_num_write_bios callback. Instead this patch simplifies things; only one bio is submitted by the core, this is first written to the origin and then the cache device in series. Obviously this will have a latency impact. deferred_writethrough_bios is introduced to record bios that must be later issued to the cache device from the worker thread. This deferred submission, after the origin bio completes, is required given that we're in interrupt context (writethrough_endio). Signed-off-by: Joe Thornber <ejt@redhat.com> Signed-off-by: Mike Snitzer <snitzer@redhat.com> Signed-off-by: Alasdair G Kergon <agk@redhat.com>
2013-03-20 21:21:27 +04:00
cell_defer(cache, cell, false);
break;
case POLICY_MISS:
inc_miss_counter(cache, bio);
pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
if (pb->req_nr != 0) {
/*
* This is a duplicate writethrough io that is no
* longer needed because the block has been demoted.
*/
bio_endio(bio, 0);
cell_defer(cache, cell, false);
return DM_MAPIO_SUBMITTED;
} else {
remap_to_origin_clear_discard(cache, bio, block);
cell_defer(cache, cell, false);
}
break;
default:
DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
(unsigned) lookup_result.op);
bio_io_error(bio);
return DM_MAPIO_SUBMITTED;
}
return DM_MAPIO_REMAPPED;
}
static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
{
struct cache *cache = ti->private;
unsigned long flags;
size_t pb_data_size = get_per_bio_data_size(cache);
struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
if (pb->tick) {
policy_tick(cache->policy);
spin_lock_irqsave(&cache->lock, flags);
cache->need_tick_bio = true;
spin_unlock_irqrestore(&cache->lock, flags);
}
check_for_quiesced_migrations(cache, pb);
return 0;
}
static int write_dirty_bitset(struct cache *cache)
{
unsigned i, r;
for (i = 0; i < from_cblock(cache->cache_size); i++) {
r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
is_dirty(cache, to_cblock(i)));
if (r)
return r;
}
return 0;
}
static int write_discard_bitset(struct cache *cache)
{
unsigned i, r;
r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
cache->discard_nr_blocks);
if (r) {
DMERR("could not resize on-disk discard bitset");
return r;
}
for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
r = dm_cache_set_discard(cache->cmd, to_dblock(i),
is_discarded(cache, to_dblock(i)));
if (r)
return r;
}
return 0;
}
static int save_hint(void *context, dm_cblock_t cblock, dm_oblock_t oblock,
uint32_t hint)
{
struct cache *cache = context;
return dm_cache_save_hint(cache->cmd, cblock, hint);
}
static int write_hints(struct cache *cache)
{
int r;
r = dm_cache_begin_hints(cache->cmd, cache->policy);
if (r) {
DMERR("dm_cache_begin_hints failed");
return r;
}
r = policy_walk_mappings(cache->policy, save_hint, cache);
if (r)
DMERR("policy_walk_mappings failed");
return r;
}
/*
* returns true on success
*/
static bool sync_metadata(struct cache *cache)
{
int r1, r2, r3, r4;
r1 = write_dirty_bitset(cache);
if (r1)
DMERR("could not write dirty bitset");
r2 = write_discard_bitset(cache);
if (r2)
DMERR("could not write discard bitset");
save_stats(cache);
r3 = write_hints(cache);
if (r3)
DMERR("could not write hints");
/*
* If writing the above metadata failed, we still commit, but don't
* set the clean shutdown flag. This will effectively force every
* dirty bit to be set on reload.
*/
r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
if (r4)
DMERR("could not write cache metadata. Data loss may occur.");
return !r1 && !r2 && !r3 && !r4;
}
static void cache_postsuspend(struct dm_target *ti)
{
struct cache *cache = ti->private;
start_quiescing(cache);
wait_for_migrations(cache);
stop_worker(cache);
requeue_deferred_io(cache);
stop_quiescing(cache);
(void) sync_metadata(cache);
}
static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
bool dirty, uint32_t hint, bool hint_valid)
{
int r;
struct cache *cache = context;
r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
if (r)
return r;
if (dirty)
set_dirty(cache, oblock, cblock);
else
clear_dirty(cache, oblock, cblock);
return 0;
}
static int load_discard(void *context, sector_t discard_block_size,
dm_dblock_t dblock, bool discard)
{
struct cache *cache = context;
/* FIXME: handle mis-matched block size */
if (discard)
set_discard(cache, dblock);
else
clear_discard(cache, dblock);
return 0;
}
static int cache_preresume(struct dm_target *ti)
{
int r = 0;
struct cache *cache = ti->private;
sector_t actual_cache_size = get_dev_size(cache->cache_dev);
(void) sector_div(actual_cache_size, cache->sectors_per_block);
/*
* Check to see if the cache has resized.
*/
if (from_cblock(cache->cache_size) != actual_cache_size || !cache->sized) {
cache->cache_size = to_cblock(actual_cache_size);
r = dm_cache_resize(cache->cmd, cache->cache_size);
if (r) {
DMERR("could not resize cache metadata");
return r;
}
cache->sized = true;
}
if (!cache->loaded_mappings) {
r = dm_cache_load_mappings(cache->cmd, cache->policy,
load_mapping, cache);
if (r) {
DMERR("could not load cache mappings");
return r;
}
cache->loaded_mappings = true;
}
if (!cache->loaded_discards) {
r = dm_cache_load_discards(cache->cmd, load_discard, cache);
if (r) {
DMERR("could not load origin discards");
return r;
}
cache->loaded_discards = true;
}
return r;
}
static void cache_resume(struct dm_target *ti)
{
struct cache *cache = ti->private;
cache->need_tick_bio = true;
do_waker(&cache->waker.work);
}
/*
* Status format:
*
* <#used metadata blocks>/<#total metadata blocks>
* <#read hits> <#read misses> <#write hits> <#write misses>
* <#demotions> <#promotions> <#blocks in cache> <#dirty>
* <#features> <features>*
* <#core args> <core args>
* <#policy args> <policy args>*
*/
static void cache_status(struct dm_target *ti, status_type_t type,
unsigned status_flags, char *result, unsigned maxlen)
{
int r = 0;
unsigned i;
ssize_t sz = 0;
dm_block_t nr_free_blocks_metadata = 0;
dm_block_t nr_blocks_metadata = 0;
char buf[BDEVNAME_SIZE];
struct cache *cache = ti->private;
dm_cblock_t residency;
switch (type) {
case STATUSTYPE_INFO:
/* Commit to ensure statistics aren't out-of-date */
if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
r = dm_cache_commit(cache->cmd, false);
if (r)
DMERR("could not commit metadata for accurate status");
}
r = dm_cache_get_free_metadata_block_count(cache->cmd,
&nr_free_blocks_metadata);
if (r) {
DMERR("could not get metadata free block count");
goto err;
}
r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
if (r) {
DMERR("could not get metadata device size");
goto err;
}
residency = policy_residency(cache->policy);
DMEMIT("%llu/%llu %u %u %u %u %u %u %llu %u ",
(unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
(unsigned long long)nr_blocks_metadata,
(unsigned) atomic_read(&cache->stats.read_hit),
(unsigned) atomic_read(&cache->stats.read_miss),
(unsigned) atomic_read(&cache->stats.write_hit),
(unsigned) atomic_read(&cache->stats.write_miss),
(unsigned) atomic_read(&cache->stats.demotion),
(unsigned) atomic_read(&cache->stats.promotion),
(unsigned long long) from_cblock(residency),
cache->nr_dirty);
if (cache->features.write_through)
DMEMIT("1 writethrough ");
else
DMEMIT("0 ");
DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
if (sz < maxlen) {
r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
if (r)
DMERR("policy_emit_config_values returned %d", r);
}
break;
case STATUSTYPE_TABLE:
format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
DMEMIT("%s ", buf);
format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
DMEMIT("%s ", buf);
format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
DMEMIT("%s", buf);
for (i = 0; i < cache->nr_ctr_args - 1; i++)
DMEMIT(" %s", cache->ctr_args[i]);
if (cache->nr_ctr_args)
DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
}
return;
err:
DMEMIT("Error");
}
/*
* Supports <key> <value>.
*
* The key migration_threshold is supported by the cache target core.
*/
static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
{
struct cache *cache = ti->private;
if (argc != 2)
return -EINVAL;
return set_config_value(cache, argv[0], argv[1]);
}
static int cache_iterate_devices(struct dm_target *ti,
iterate_devices_callout_fn fn, void *data)
{
int r = 0;
struct cache *cache = ti->private;
r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
if (!r)
r = fn(ti, cache->origin_dev, 0, ti->len, data);
return r;
}
/*
* We assume I/O is going to the origin (which is the volume
* more likely to have restrictions e.g. by being striped).
* (Looking up the exact location of the data would be expensive
* and could always be out of date by the time the bio is submitted.)
*/
static int cache_bvec_merge(struct dm_target *ti,
struct bvec_merge_data *bvm,
struct bio_vec *biovec, int max_size)
{
struct cache *cache = ti->private;
struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
if (!q->merge_bvec_fn)
return max_size;
bvm->bi_bdev = cache->origin_dev->bdev;
return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
}
static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
{
/*
* FIXME: these limits may be incompatible with the cache device
*/
limits->max_discard_sectors = cache->discard_block_size * 1024;
limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
}
static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
{
struct cache *cache = ti->private;
blk_limits_io_min(limits, 0);
blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
set_discard_limits(cache, limits);
}
/*----------------------------------------------------------------*/
static struct target_type cache_target = {
.name = "cache",
.version = {1, 1, 1},
.module = THIS_MODULE,
.ctr = cache_ctr,
.dtr = cache_dtr,
.map = cache_map,
.end_io = cache_end_io,
.postsuspend = cache_postsuspend,
.preresume = cache_preresume,
.resume = cache_resume,
.status = cache_status,
.message = cache_message,
.iterate_devices = cache_iterate_devices,
.merge = cache_bvec_merge,
.io_hints = cache_io_hints,
};
static int __init dm_cache_init(void)
{
int r;
r = dm_register_target(&cache_target);
if (r) {
DMERR("cache target registration failed: %d", r);
return r;
}
migration_cache = KMEM_CACHE(dm_cache_migration, 0);
if (!migration_cache) {
dm_unregister_target(&cache_target);
return -ENOMEM;
}
return 0;
}
static void __exit dm_cache_exit(void)
{
dm_unregister_target(&cache_target);
kmem_cache_destroy(migration_cache);
}
module_init(dm_cache_init);
module_exit(dm_cache_exit);
MODULE_DESCRIPTION(DM_NAME " cache target");
MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
MODULE_LICENSE("GPL");