huynp911
/
vitastor
forked from vitalif/vitastor
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Send STABLE entry to the journal

blocking-uring-test
Vitaliy Filippov 2019-11-10 15:00:01 +03:00
parent 890335bff6
commit 69581e6d0c
3 changed files with 86 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
blockstore_read.cpp
View File

@ -87,6 +87,7 @@ int blockstore::dequeue_read(blockstore_operation *read_op)
read_op->callback(read_op);
return 1;
}
// FIXME track fulfilled and stop when it is equal to read_op->len
uint64_t fulfilled = 0;
if (dirty_found)
{

84
blockstore_stable.cpp
View File

@ -36,7 +36,63 @@ int blockstore::dequeue_stable(blockstore_operation *op)
op->callback(op);
return 1;
}
return 0;
// 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))
{
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,
};
je->crc32 = je_crc32((journal_entry*)je);
journal.crc32_last = je->crc32;
data->iov = (struct iovec){ journal.sector_buf + 512*journal.cur_sector, 512 };
data->op = op;
io_uring_prep_writev(
sqe, journal.fd, &data->iov, 1, journal.offset + journal.sector_info[journal.cur_sector].offset
);
journal.sector_info[journal.cur_sector].usage_count++;
op->pending_ops = 1;
op->min_used_journal_sector = op->max_used_journal_sector = 1 + journal.cur_sector;
return 1;
}
void blockstore::handle_stable_event(ring_data_t *data, blockstore_operation *op)
@ -50,6 +106,32 @@ void blockstore::handle_stable_event(ring_data_t *data, blockstore_operation *op
op->pending_ops--;
if (op->pending_ops == 0)
{
// Mark dirty_db entry as stable
auto dirty_it = dirty_db.find((obj_ver_id){
.oid = op->oid,
.version = op->version,
});
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 object_db entry's version and clear previous journal entries
// This makes 1 4K small write 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. :)
}
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 object_db and clear previous journal entries
// This makes 1 128K big write look like:
// 128K (data) + sync + 512b (journal) + sync + 512b (journal) + sync + 512b (metadata) + sync.
// WA = 1.012. Very good :)
}
}
}

5
blockstore_write.cpp
View File

@ -98,7 +98,6 @@ int blockstore::dequeue_write(blockstore_operation *op)
return 0;
}
// There is sufficient space. Get SQE(s)
unsigned prev_sqe_pos = ringloop->ring->sq.sqe_tail;
BS_SUBMIT_GET_SQE(sqe1, data1);
BS_SUBMIT_GET_SQE(sqe2, data2);
// Got SQEs. Prepare journal sector write
@ -133,6 +132,7 @@ int blockstore::dequeue_write(blockstore_operation *op)
io_uring_prep_writev(
sqe1, journal.fd, &data1->iov, 1, journal.offset + journal.sector_info[journal.cur_sector].offset
);
journal.sector_info[journal.cur_sector].usage_count++;
// Prepare journal data write
journal.next_free = (journal.next_free + op->len) < journal.len ? journal.next_free + op->len : 512;
data2->iov = (struct iovec){ op->buf, op->len };
@ -142,9 +142,8 @@ int blockstore::dequeue_write(blockstore_operation *op)
);
dirty_it->second.location = journal.next_free;
dirty_it->second.state = ST_J_SUBMITTED;
// Move journal.next_free and save last write for current sector
// Move journal.next_free
journal.next_free += op->len;
journal.sector_info[journal.cur_sector].usage_count++;
op->pending_ops = 2;
op->min_used_journal_sector = op->max_used_journal_sector = 1 + journal.cur_sector;
}