File bug-1150021_04-bcache-Bring-bcache-into-sync-with-master-branch.patch of Package lvm2.16468
From b6e6ea2d65785f03f3cee7938be635bcb8ad4944 Mon Sep 17 00:00:00 2001
From: Joe Thornber <ejt@redhat.com>
Date: Thu, 16 Jan 2020 14:20:35 +0000
Subject: [PATCH 2/9] [bcache] Bring bcache into sync with master branch
---
lib/device/bcache.c | 350 ++++++++++++++++++---------------------------
lib/device/bcache.h | 1 -
test/unit/bcache_t.c | 7 +-
test/unit/bcache_utils_t.c | 3 +-
test/unit/unit-test.sh | 2 -
5 files changed, 143 insertions(+), 220 deletions(-)
diff --git a/lib/device/bcache.c b/lib/device/bcache.c
index d487ca2..9f67a27 100644
--- a/lib/device/bcache.c
+++ b/lib/device/bcache.c
@@ -12,9 +12,9 @@
* Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
-#define _GNU_SOURCE
-
#include "bcache.h"
+
+#include "base/data-struct/radix-tree.h"
#include "lvm-logging.h"
#include "log.h"
@@ -67,14 +67,14 @@ struct cb_set {
static struct cb_set *_cb_set_create(unsigned nr)
{
int i;
- struct cb_set *cbs = dm_malloc(sizeof(*cbs));
+ struct cb_set *cbs = malloc(sizeof(*cbs));
if (!cbs)
return NULL;
- cbs->vec = dm_malloc(nr * sizeof(*cbs->vec));
+ cbs->vec = malloc(nr * sizeof(*cbs->vec));
if (!cbs->vec) {
- dm_free(cbs);
+ free(cbs);
return NULL;
}
@@ -97,8 +97,8 @@ static void _cb_set_destroy(struct cb_set *cbs)
return;
}
- dm_free(cbs->vec);
- dm_free(cbs);
+ free(cbs->vec);
+ free(cbs);
}
static struct control_block *_cb_alloc(struct cb_set *cbs, void *context)
@@ -152,7 +152,7 @@ static void _async_destroy(struct io_engine *ioe)
if (r)
log_sys_warn("io_destroy");
- dm_free(e);
+ free(e);
}
static int _last_byte_fd;
@@ -169,7 +169,6 @@ static bool _async_issue(struct io_engine *ioe, enum dir d, int fd,
sector_t offset;
sector_t nbytes;
sector_t limit_nbytes;
- sector_t orig_nbytes;
sector_t extra_nbytes = 0;
if (((uintptr_t) data) & e->page_mask) {
@@ -192,41 +191,11 @@ static bool _async_issue(struct io_engine *ioe, enum dir d, int fd,
return false;
}
- /*
- * If the bcache block offset+len goes beyond where lvm is
- * intending to write, then reduce the len being written
- * (which is the bcache block size) so we don't write past
- * the limit set by lvm. If after applying the limit, the
- * resulting size is not a multiple of the sector size (512
- * or 4096) then extend the reduced size to be a multiple of
- * the sector size (we don't want to write partial sectors.)
- */
if (offset + nbytes > _last_byte_offset) {
limit_nbytes = _last_byte_offset - offset;
-
- if (limit_nbytes % _last_byte_sector_size) {
+ if (limit_nbytes % _last_byte_sector_size)
extra_nbytes = _last_byte_sector_size - (limit_nbytes % _last_byte_sector_size);
- /*
- * adding extra_nbytes to the reduced nbytes (limit_nbytes)
- * should make the final write size a multiple of the
- * sector size. This should never result in a final size
- * larger than the bcache block size (as long as the bcache
- * block size is a multiple of the sector size).
- */
- if (limit_nbytes + extra_nbytes > nbytes) {
- log_warn("Skip extending write at %llu len %llu limit %llu extra %llu sector_size %llu",
- (unsigned long long)offset,
- (unsigned long long)nbytes,
- (unsigned long long)limit_nbytes,
- (unsigned long long)extra_nbytes,
- (unsigned long long)_last_byte_sector_size);
- extra_nbytes = 0;
- }
- }
-
- orig_nbytes = nbytes;
-
if (extra_nbytes) {
log_debug("Limit write at %llu len %llu to len %llu rounded to %llu",
(unsigned long long)offset,
@@ -241,22 +210,6 @@ static bool _async_issue(struct io_engine *ioe, enum dir d, int fd,
(unsigned long long)limit_nbytes);
nbytes = limit_nbytes;
}
-
- /*
- * This shouldn't happen, the reduced+extended
- * nbytes value should never be larger than the
- * bcache block size.
- */
- if (nbytes > orig_nbytes) {
- log_error("Invalid adjusted write at %llu len %llu adjusted %llu limit %llu extra %llu sector_size %llu",
- (unsigned long long)offset,
- (unsigned long long)orig_nbytes,
- (unsigned long long)nbytes,
- (unsigned long long)limit_nbytes,
- (unsigned long long)extra_nbytes,
- (unsigned long long)_last_byte_sector_size);
- return false;
- }
}
}
@@ -361,7 +314,7 @@ static unsigned _async_max_io(struct io_engine *e)
struct io_engine *create_async_io_engine(void)
{
int r;
- struct async_engine *e = dm_malloc(sizeof(*e));
+ struct async_engine *e = malloc(sizeof(*e));
if (!e)
return NULL;
@@ -375,14 +328,14 @@ struct io_engine *create_async_io_engine(void)
r = io_setup(MAX_IO, &e->aio_context);
if (r < 0) {
log_debug("io_setup failed %d", r);
- dm_free(e);
+ free(e);
return NULL;
}
e->cbs = _cb_set_create(MAX_IO);
if (!e->cbs) {
log_warn("couldn't create control block set");
- dm_free(e);
+ free(e);
return NULL;
}
@@ -411,7 +364,7 @@ static struct sync_engine *_to_sync(struct io_engine *e)
static void _sync_destroy(struct io_engine *ioe)
{
struct sync_engine *e = _to_sync(ioe);
- dm_free(e);
+ free(e);
}
static bool _sync_issue(struct io_engine *ioe, enum dir d, int fd,
@@ -430,7 +383,6 @@ static bool _sync_issue(struct io_engine *ioe, enum dir d, int fd,
}
where = sb * 512;
-
off = lseek(fd, where, SEEK_SET);
if (off == (off_t) -1) {
log_warn("Device seek error %d for offset %llu", errno, (unsigned long long)where);
@@ -451,7 +403,6 @@ static bool _sync_issue(struct io_engine *ioe, enum dir d, int fd,
uint64_t nbytes = len;
sector_t limit_nbytes = 0;
sector_t extra_nbytes = 0;
- sector_t orig_nbytes = 0;
if (offset > _last_byte_offset) {
log_error("Limit write at %llu len %llu beyond last byte %llu",
@@ -464,30 +415,9 @@ static bool _sync_issue(struct io_engine *ioe, enum dir d, int fd,
if (offset + nbytes > _last_byte_offset) {
limit_nbytes = _last_byte_offset - offset;
-
- if (limit_nbytes % _last_byte_sector_size) {
+ if (limit_nbytes % _last_byte_sector_size)
extra_nbytes = _last_byte_sector_size - (limit_nbytes % _last_byte_sector_size);
- /*
- * adding extra_nbytes to the reduced nbytes (limit_nbytes)
- * should make the final write size a multiple of the
- * sector size. This should never result in a final size
- * larger than the bcache block size (as long as the bcache
- * block size is a multiple of the sector size).
- */
- if (limit_nbytes + extra_nbytes > nbytes) {
- log_warn("Skip extending write at %llu len %llu limit %llu extra %llu sector_size %llu",
- (unsigned long long)offset,
- (unsigned long long)nbytes,
- (unsigned long long)limit_nbytes,
- (unsigned long long)extra_nbytes,
- (unsigned long long)_last_byte_sector_size);
- extra_nbytes = 0;
- }
- }
-
- orig_nbytes = nbytes;
-
if (extra_nbytes) {
log_debug("Limit write at %llu len %llu to len %llu rounded to %llu",
(unsigned long long)offset,
@@ -502,23 +432,6 @@ static bool _sync_issue(struct io_engine *ioe, enum dir d, int fd,
(unsigned long long)limit_nbytes);
nbytes = limit_nbytes;
}
-
- /*
- * This shouldn't happen, the reduced+extended
- * nbytes value should never be larger than the
- * bcache block size.
- */
- if (nbytes > orig_nbytes) {
- log_error("Invalid adjusted write at %llu len %llu adjusted %llu limit %llu extra %llu sector_size %llu",
- (unsigned long long)offset,
- (unsigned long long)orig_nbytes,
- (unsigned long long)nbytes,
- (unsigned long long)limit_nbytes,
- (unsigned long long)extra_nbytes,
- (unsigned long long)_last_byte_sector_size);
- free(io);
- return false;
- }
}
where = offset;
@@ -580,7 +493,7 @@ static bool _sync_wait(struct io_engine *ioe, io_complete_fn fn)
dm_list_iterate_items_safe(io, tmp, &e->complete) {
fn(io->context, 0);
dm_list_del(&io->list);
- dm_free(io);
+ free(io);
}
return true;
@@ -593,7 +506,7 @@ static unsigned _sync_max_io(struct io_engine *e)
struct io_engine *create_sync_io_engine(void)
{
- struct sync_engine *e = dm_malloc(sizeof(*e));
+ struct sync_engine *e = malloc(sizeof(*e));
if (!e)
return NULL;
@@ -673,12 +586,7 @@ struct bcache {
struct dm_list clean;
struct dm_list io_pending;
- /*
- * Hash table.
- */
- unsigned nr_buckets;
- unsigned hash_mask;
- struct dm_list *buckets;
+ struct radix_tree *rtree;
/*
* Statistics
@@ -693,75 +601,50 @@ struct bcache {
//----------------------------------------------------------------
-/* 2^63 + 2^61 - 2^57 + 2^54 - 2^51 - 2^18 + 1 */
-#define GOLDEN_RATIO_PRIME_64 0x9e37fffffffc0001ULL
+struct key_parts {
+ uint32_t fd;
+ uint64_t b;
+} __attribute__ ((packed));
-static unsigned _hash(struct bcache *cache, int fd, uint64_t i)
-{
- uint64_t h = (i << 10) & fd;
- h *= GOLDEN_RATIO_PRIME_64;
- return h & cache->hash_mask;
-}
+union key {
+ struct key_parts parts;
+ uint8_t bytes[12];
+};
-static struct block *_hash_lookup(struct bcache *cache, int fd, uint64_t i)
+static struct block *_block_lookup(struct bcache *cache, int fd, uint64_t i)
{
- struct block *b;
- unsigned h = _hash(cache, fd, i);
-
- dm_list_iterate_items_gen (b, cache->buckets + h, hash)
- if (b->fd == fd && b->index == i)
- return b;
+ union key k;
+ union radix_value v;
- return NULL;
-}
+ k.parts.fd = fd;
+ k.parts.b = i;
-static void _hash_insert(struct block *b)
-{
- unsigned h = _hash(b->cache, b->fd, b->index);
- dm_list_add_h(b->cache->buckets + h, &b->hash);
-}
+ if (radix_tree_lookup(cache->rtree, k.bytes, k.bytes + sizeof(k.bytes), &v))
+ return v.ptr;
-static inline void _hash_remove(struct block *b)
-{
- dm_list_del(&b->hash);
+ return NULL;
}
-/*
- * Must return a power of 2.
- */
-static unsigned _calc_nr_buckets(unsigned nr_blocks)
+static bool _block_insert(struct block *b)
{
- unsigned r = 8;
- unsigned n = nr_blocks / 4;
+ union key k;
+ union radix_value v;
- if (n < 8)
- n = 8;
+ k.parts.fd = b->fd;
+ k.parts.b = b->index;
+ v.ptr = b;
- while (r < n)
- r <<= 1;
-
- return r;
+ return radix_tree_insert(b->cache->rtree, k.bytes, k.bytes + sizeof(k.bytes), v);
}
-static bool _hash_table_init(struct bcache *cache, unsigned nr_entries)
+static void _block_remove(struct block *b)
{
- unsigned i;
+ union key k;
- cache->nr_buckets = _calc_nr_buckets(nr_entries);
- cache->hash_mask = cache->nr_buckets - 1;
- cache->buckets = dm_malloc(cache->nr_buckets * sizeof(*cache->buckets));
- if (!cache->buckets)
- return false;
+ k.parts.fd = b->fd;
+ k.parts.b = b->index;
- for (i = 0; i < cache->nr_buckets; i++)
- dm_list_init(cache->buckets + i);
-
- return true;
-}
-
-static void _hash_table_exit(struct bcache *cache)
-{
- dm_free(cache->buckets);
+ radix_tree_remove(b->cache->rtree, k.bytes, k.bytes + sizeof(k.bytes));
}
//----------------------------------------------------------------
@@ -777,7 +660,7 @@ static bool _init_free_list(struct bcache *cache, unsigned count, unsigned pgsiz
if (!data)
return false;
- cache->raw_blocks = dm_malloc(count * sizeof(*cache->raw_blocks));
+ cache->raw_blocks = malloc(count * sizeof(*cache->raw_blocks));
if (!cache->raw_blocks) {
free(data);
return false;
@@ -797,8 +680,8 @@ static bool _init_free_list(struct bcache *cache, unsigned count, unsigned pgsiz
static void _exit_free_list(struct bcache *cache)
{
- dm_free(cache->raw_data);
- dm_free(cache->raw_blocks);
+ free(cache->raw_data);
+ free(cache->raw_blocks);
}
static struct block *_alloc_block(struct bcache *cache)
@@ -809,6 +692,11 @@ static struct block *_alloc_block(struct bcache *cache)
return dm_list_struct_base(_list_pop(&cache->free), struct block, list);
}
+static void _free_block(struct block *b)
+{
+ dm_list_add(&b->cache->free, &b->list);
+}
+
/*----------------------------------------------------------------
* Clean/dirty list management.
* Always use these methods to ensure nr_dirty_ is correct.
@@ -963,7 +851,7 @@ static struct block *_find_unused_clean_block(struct bcache *cache)
dm_list_iterate_items (b, &cache->clean) {
if (!b->ref_count) {
_unlink_block(b);
- _hash_remove(b);
+ _block_remove(b);
return b;
}
}
@@ -993,29 +881,18 @@ static struct block *_new_block(struct bcache *cache, int fd, block_address i, b
if (b) {
dm_list_init(&b->list);
- dm_list_init(&b->hash);
b->flags = 0;
b->fd = fd;
b->index = i;
b->ref_count = 0;
b->error = 0;
- _hash_insert(b);
- }
-
-#if 0
- if (!b) {
- log_error("bcache no new blocks for fd %d index %u "
- "clean %u free %u dirty %u pending %u nr_data_blocks %u nr_cache_blocks %u",
- fd, (uint32_t) i,
- dm_list_size(&cache->clean),
- dm_list_size(&cache->free),
- dm_list_size(&cache->dirty),
- dm_list_size(&cache->io_pending),
- (uint32_t)cache->nr_data_blocks,
- (uint32_t)cache->nr_cache_blocks);
+ if (!_block_insert(b)) {
+ log_error("bcache unable to insert block in radix tree (OOM?)");
+ _free_block(b);
+ return NULL;
+ }
}
-#endif
return b;
}
@@ -1054,7 +931,7 @@ static struct block *_lookup_or_read_block(struct bcache *cache,
int fd, block_address i,
unsigned flags)
{
- struct block *b = _hash_lookup(cache, fd, i);
+ struct block *b = _block_lookup(cache, fd, i);
if (b) {
// FIXME: this is insufficient. We need to also catch a read
@@ -1125,8 +1002,8 @@ struct bcache *bcache_create(sector_t block_sectors, unsigned nr_cache_blocks,
unsigned max_io = engine->max_io(engine);
long pgsize = sysconf(_SC_PAGESIZE);
- if ((pgsize = sysconf(_SC_PAGESIZE)) < 0) {
- log_warn("bcache cannot read pagesize.");
+ if (pgsize < 0) {
+ log_warn("WARNING: _SC_PAGESIZE returns negative value.");
return NULL;
}
@@ -1145,7 +1022,7 @@ struct bcache *bcache_create(sector_t block_sectors, unsigned nr_cache_blocks,
return NULL;
}
- cache = dm_malloc(sizeof(*cache));
+ cache = malloc(sizeof(*cache));
if (!cache)
return NULL;
@@ -1163,9 +1040,10 @@ struct bcache *bcache_create(sector_t block_sectors, unsigned nr_cache_blocks,
dm_list_init(&cache->clean);
dm_list_init(&cache->io_pending);
- if (!_hash_table_init(cache, nr_cache_blocks)) {
+ cache->rtree = radix_tree_create(NULL, NULL);
+ if (!cache->rtree) {
cache->engine->destroy(cache->engine);
- dm_free(cache);
+ free(cache);
return NULL;
}
@@ -1178,8 +1056,8 @@ struct bcache *bcache_create(sector_t block_sectors, unsigned nr_cache_blocks,
if (!_init_free_list(cache, nr_cache_blocks, pgsize)) {
cache->engine->destroy(cache->engine);
- _hash_table_exit(cache);
- dm_free(cache);
+ radix_tree_destroy(cache->rtree);
+ free(cache);
return NULL;
}
@@ -1192,12 +1070,12 @@ void bcache_destroy(struct bcache *cache)
log_warn("some blocks are still locked");
if (!bcache_flush(cache))
- log_warn("cache flushing failed.");
+ stack;
_wait_all(cache);
_exit_free_list(cache);
- _hash_table_exit(cache);
+ radix_tree_destroy(cache->rtree);
cache->engine->destroy(cache->engine);
- dm_free(cache);
+ free(cache);
}
sector_t bcache_block_sectors(struct bcache *cache)
@@ -1217,7 +1095,7 @@ unsigned bcache_max_prefetches(struct bcache *cache)
void bcache_prefetch(struct bcache *cache, int fd, block_address i)
{
- struct block *b = _hash_lookup(cache, fd, i);
+ struct block *b = _block_lookup(cache, fd, i);
if (!b) {
if (cache->nr_io_pending < cache->max_io) {
@@ -1230,11 +1108,13 @@ void bcache_prefetch(struct bcache *cache, int fd, block_address i)
}
}
+//----------------------------------------------------------------
+
static void _recycle_block(struct bcache *cache, struct block *b)
{
_unlink_block(b);
- _hash_remove(b);
- dm_list_add(&cache->free, &b->list);
+ _block_remove(b);
+ _free_block(b);
}
bool bcache_get(struct bcache *cache, int fd, block_address i,
@@ -1268,6 +1148,8 @@ bool bcache_get(struct bcache *cache, int fd, block_address i,
return false;
}
+//----------------------------------------------------------------
+
static void _put_ref(struct block *b)
{
if (!b->ref_count) {
@@ -1288,6 +1170,8 @@ void bcache_put(struct block *b)
_preemptive_writeback(b->cache);
}
+//----------------------------------------------------------------
+
bool bcache_flush(struct bcache *cache)
{
// Only dirty data is on the errored list, since bad read blocks get
@@ -1310,6 +1194,7 @@ bool bcache_flush(struct bcache *cache)
return dm_list_empty(&cache->errored);
}
+//----------------------------------------------------------------
/*
* You can safely call this with a NULL block.
*/
@@ -1342,29 +1227,72 @@ static bool _invalidate_block(struct bcache *cache, struct block *b)
bool bcache_invalidate(struct bcache *cache, int fd, block_address i)
{
- return _invalidate_block(cache, _hash_lookup(cache, fd, i));
+ return _invalidate_block(cache, _block_lookup(cache, fd, i));
+}
+
+//----------------------------------------------------------------
+
+struct invalidate_iterator {
+ bool success;
+ struct radix_tree_iterator it;
+};
+
+static bool _writeback_v(struct radix_tree_iterator *it,
+ uint8_t *kb, uint8_t *ke, union radix_value v)
+{
+ struct block *b = v.ptr;
+
+ if (_test_flags(b, BF_DIRTY))
+ _issue_write(b);
+
+ return true;
+}
+
+static bool _invalidate_v(struct radix_tree_iterator *it,
+ uint8_t *kb, uint8_t *ke, union radix_value v)
+{
+ struct block *b = v.ptr;
+ struct invalidate_iterator *iit = container_of(it, struct invalidate_iterator, it);
+
+ if (b->error || _test_flags(b, BF_DIRTY)) {
+ log_warn("bcache_invalidate: block (%d, %llu) still dirty",
+ b->fd, (unsigned long long) b->index);
+ iit->success = false;
+ return true;
+ }
+
+ if (b->ref_count) {
+ log_warn("bcache_invalidate: block (%d, %llu) still held",
+ b->fd, (unsigned long long) b->index);
+ iit->success = false;
+ return true;
+ }
+
+ _unlink_block(b);
+ _free_block(b);
+
+ // We can't remove the block from the radix tree yet because
+ // we're in the middle of an iteration.
+ return true;
}
-// FIXME: switch to a trie, or maybe 1 hash table per fd? To save iterating
-// through the whole cache.
bool bcache_invalidate_fd(struct bcache *cache, int fd)
{
- struct block *b, *tmp;
- bool r = true;
+ union key k;
+ struct invalidate_iterator it;
- // Start writing back any dirty blocks on this fd.
- dm_list_iterate_items_safe (b, tmp, &cache->dirty)
- if (b->fd == fd)
- _issue_write(b);
+ k.parts.fd = fd;
- _wait_all(cache);
+ it.it.visit = _writeback_v;
+ radix_tree_iterate(cache->rtree, k.bytes, k.bytes + sizeof(k.parts.fd), &it.it);
- // Everything should be in the clean list now.
- dm_list_iterate_items_safe (b, tmp, &cache->clean)
- if (b->fd == fd)
- r = _invalidate_block(cache, b) && r;
+ _wait_all(cache);
- return r;
+ it.success = true;
+ it.it.visit = _invalidate_v;
+ radix_tree_iterate(cache->rtree, k.bytes, k.bytes + sizeof(k.parts.fd), &it.it);
+ radix_tree_remove_prefix(cache->rtree, k.bytes, k.bytes + sizeof(k.parts.fd));
+ return it.success;
}
//----------------------------------------------------------------
diff --git a/lib/device/bcache.h b/lib/device/bcache.h
index cb902ef..790d0fe 100644
--- a/lib/device/bcache.h
+++ b/lib/device/bcache.h
@@ -61,7 +61,6 @@ struct block {
struct bcache *cache;
struct dm_list list;
- struct dm_list hash;
unsigned flags;
unsigned ref_count;
diff --git a/test/unit/bcache_t.c b/test/unit/bcache_t.c
index 925b95d..92c2d57 100644
--- a/test/unit/bcache_t.c
+++ b/test/unit/bcache_t.c
@@ -12,15 +12,14 @@
* Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
+#include "units.h"
+#include "lib/device/bcache.h"
+
#include <errno.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
-#include "bcache.h"
-#include "framework.h"
-#include "units.h"
-
#define SHOW_MOCK_CALLS 0
/*----------------------------------------------------------------
diff --git a/test/unit/bcache_utils_t.c b/test/unit/bcache_utils_t.c
index 9ddc194..2e08320 100644
--- a/test/unit/bcache_utils_t.c
+++ b/test/unit/bcache_utils_t.c
@@ -14,9 +14,8 @@
#define _GNU_SOURCE
-#include "bcache.h"
-#include "framework.h"
#include "units.h"
+#include "lib/device/bcache.h"
#include <errno.h>
#include <stdio.h>
diff --git a/test/unit/unit-test.sh b/test/unit/unit-test.sh
index e8332d6..f545f14 100644
--- a/test/unit/unit-test.sh
+++ b/test/unit/unit-test.sh
@@ -13,8 +13,6 @@
SKIP_WITH_LVMLOCKD=1
SKIP_WITH_LVMPOLLD=1
-SKIP_WITH_LVMETAD=1
-SKIP_WITH_CLVMD=1
SKIP_ROOT_DM_CHECK=1
--
1.8.3.1
