thebestvibedev/tromcho.net
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Remove duplicate journal writing code (and fix it at the same time)

Browse Source
This commit is contained in:
Vitaliy Filippov
2019-11-11 00:28:14 +03:00
parent ff57dd420e
commit 8edb9e9d6f
7 changed files with 151 additions and 168 deletions
Download Patch File Download Diff File Expand all files Collapse all files
blockstore_stable.cpp
+20 -49
View File
@@ -36,52 +36,18 @@ int blockstore::dequeue_stable(blockstore_operation *op)
op->callback(op);
return 1;
}
// FIXME: Try to deduplicate journal entry submission code...
// Check journal space
uint64_t next_pos = journal.next_free;
if (512 - journal.in_sector_pos < sizeof(struct journal_entry_stable))
blockstore_journal_check_t space_check(this);
if (!space_check.check_available(op, 1, sizeof(journal_entry_stable), 0))
{
next_pos = (next_pos+512) < journal.len ? next_pos+512 : 512;
// Also check if we have an unused memory buffer for the journal sector
if (journal.sector_info[((journal.cur_sector + 1) % journal.sector_count)].usage_count > 0)
{
// No memory buffer available. Wait for it.
op->wait_for = WAIT_JOURNAL_BUFFER;
return 0;
}
}
if (next_pos >= journal.used_start)
{
// No space in the journal. Wait for it.
op->wait_for = WAIT_JOURNAL;
op->wait_detail = next_pos;
return 0;
}
// There is sufficient space. Get SQE
BS_SUBMIT_GET_SQE(sqe, data);
// Got SQE. Prepare journal sector write
if (512 - journal.in_sector_pos < sizeof(struct journal_entry_stable))
{
// Move to the next journal sector
// Also select next sector buffer in memory
journal.cur_sector = ((journal.cur_sector + 1) % journal.sector_count);
journal.sector_info[journal.cur_sector].offset = journal.next_free;
journal.in_sector_pos = 0;
journal.next_free = (journal.next_free+512) < journal.len ? journal.next_free + 512 : 512;
memset(journal.sector_buf + 512*journal.cur_sector, 0, 512);
}
journal_entry_stable *je = (journal_entry_stable*)(
journal.sector_buf + 512*journal.cur_sector + journal.in_sector_pos
);
*je = {
.crc32 = 0,
.magic = JOURNAL_MAGIC,
.type = JE_STABLE,
.size = sizeof(struct journal_entry_stable),
.crc32_prev = journal.crc32_last,
.oid = op->oid,
.version = op->version,
};
journal_entry_stable *je = (journal_entry_stable*)prefill_single_journal_entry(journal, JE_STABLE, sizeof(struct journal_entry_stable));
je->oid = op->oid;
je->version = op->version;
je->crc32 = je_crc32((journal_entry*)je);
journal.crc32_last = je->crc32;
data->iov = (struct iovec){ journal.sector_buf + 512*journal.cur_sector, 512 };
@@ -113,25 +79,30 @@ void blockstore::handle_stable_event(ring_data_t *data, blockstore_operation *op
});
if (dirty_it->second.state == ST_J_SYNCED)
{
dirty_it->second.state = ST_J_STABLE;
// Copy data from the journal to the data device
// -> increase version on the metadata device
// -> advance clean_db entry's version and clear previous journal entries
// This makes 1 4K small write look like:
// 1) Copy data from the journal to the data device
// 2) Increase version on the metadata device
// 3) Advance clean_db entry's version, clear previous journal entries
// This makes 1 4K small write+sync look like:
// 512b+4K (journal) + sync + 512b (journal) + sync + 512b (metadata) + 4K (data) + sync.
// WA = 2.375. It's not the best, SSD FTL-like redirect-write with defragmentation
// could probably be lower even with defragmentation. But it's fixed and it's still
// better than in Ceph. :)
dirty_it->second.state = ST_J_STABLE;
// Acknowledge op
op->retval = 0;
op->callback(op);
}
else if (dirty_it->second.state == ST_D_META_SYNCED)
{
dirty_it->second.state = ST_D_STABLE;
// Copy metadata from the journal to the metadata device
// -> move dirty_db entry to clean_db and clear previous journal entries
// This makes 1 128K big write look like:
// 1) Copy metadata from the journal to the metadata device
// 2) Move dirty_db entry to clean_db and clear previous journal entries
// This makes 1 128K big write+sync look like:
// 128K (data) + sync + 512b (journal) + sync + 512b (journal) + sync + 512b (metadata) + sync.
// WA = 1.012. Very good :)
dirty_it->second.state = ST_D_STABLE;
// Acknowledge op
op->retval = 0;
op->callback(op);
}
}
}