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

Move blockstore journal fields to journal_t, implement multiple write buffers for journal sectors

Browse Source
This commit is contained in:
Vitaliy Filippov
2019-11-07 23:42:24 +03:00
parent 84c62840bd
commit 5330461029
7 changed files with 167 additions and 84 deletions
Download Patch File Download Diff File Expand all files Collapse all files
blockstore_write.cpp
+47 -33
View File
@@ -32,29 +32,39 @@ int blockstore::dequeue_write(blockstore_operation *op)
sqe, data_fd, &data->iov, 1, data_offset + (loc << block_order)
);
op->pending_ops = 1;
op->used_journal_sector = 0;
}
else
{
// Small (journaled) write
// First check if the journal has sufficient space
bool two_sqes = false;
uint64_t next_pos = journal_data_pos;
if (512 - journal_sector_pos < sizeof(struct journal_entry_small_write))
// FIXME Always two SQEs for now. Although it's possible to send 1
bool two_sqes = true;
uint64_t next_pos = journal.next_free;
if (512 - journal.in_sector_pos < sizeof(struct journal_entry_small_write))
{
next_pos = next_pos + 512;
if (journal_len - next_pos < op->len)
two_sqes = true;
if (next_pos >= journal_len)
//if (journal.len - next_pos < op->len)
// two_sqes = true;
if (next_pos >= journal.len)
next_pos = 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;
}
}
else if (journal_sector + 512 != journal_data_pos || journal_len - journal_data_pos < op->len)
two_sqes = true;
next_pos = (journal_len - next_pos < op->len ? 512 : next_pos) + op->len;
if (next_pos >= journal_start)
//else if (journal.sector_info[journal.cur_sector].offset + 512 != journal.next_free ||
// journal.len - next_pos < op->len)
// two_sqes = true;
next_pos = (journal.len - next_pos < op->len ? 512 : next_pos) + op->len;
if (next_pos >= journal.used_start)
{
// No space in the journal. Wait until it's available
// No space in the journal. Wait for it.
op->wait_for = WAIT_JOURNAL;
op->wait_detail = next_pos - journal_start;
op->wait_detail = next_pos - journal.used_start;
return 0;
}
// There is sufficient space. Get SQE(s)
@@ -68,50 +78,54 @@ int blockstore::dequeue_write(blockstore_operation *op)
struct ring_data_t *data1 = ((ring_data_t*)sqe1->user_data);
struct ring_data_t *data2 = two_sqes ? ((ring_data_t*)sqe2->user_data) : NULL;
// Got SQEs. Prepare journal sector write
if (512 - journal_sector_pos < sizeof(struct journal_entry_small_write))
if (512 - journal.in_sector_pos < sizeof(struct journal_entry_small_write))
{
// Move to the next journal sector
next_pos = journal_data_pos + 512;
if (next_pos >= journal_len)
next_pos = 512;
journal_sector = journal_data_pos;
journal_sector_pos = 0;
journal_data_pos = next_pos;
memset(journal_sector_buf, 0, 512);
// 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_small_write *je = (struct journal_entry_small_write*)(journal_sector_buf + journal_sector_pos);
journal_entry_small_write *je = (struct journal_entry_small_write*)(
journal.sector_buf + 512*journal.cur_sector + journal.in_sector_pos
);
*je = {
.crc32 = 0,
.magic = JOURNAL_MAGIC,
.type = JE_SMALL_WRITE,
.size = sizeof(struct journal_entry_small_write),
.crc32_prev = journal_crc32_last,
.crc32_prev = journal.crc32_last,
.oid = op->oid,
.version = op->version,
.offset = op->offset,
.len = op->len,
};
je.crc32 = je_crc32((journal_entry*)je);
data1->iov = (struct iovec){ journal_sector_buf, 512 };
je->crc32 = je_crc32((journal_entry*)je);
data1->iov = (struct iovec){ journal.sector_buf + 512*journal.cur_sector, 512 };
data1->op = op;
io_uring_prep_writev(
sqe1, journal_fd, &data1->iov, 1, journal_offset + journal_sector
sqe1, journal.fd, &data1->iov, 1, journal.offset + journal.sector_info[journal.cur_sector].offset
);
// Prepare journal data write
if (journal_len - journal_data_pos < op->len)
journal_data_pos = 512;
if (journal.len - journal.next_free < op->len)
journal.next_free = 512;
data2->iov = (struct iovec){ op->buf, op->len };
data2->op = op;
io_uring_prep_writev(
sqe2, journal_fd, &data2->iov, 1, journal_offset + journal_data_pos
sqe2, journal.fd, &data2->iov, 1, journal.offset + journal.next_free
);
(*dirty_it).location = journal_data_pos;
(*dirty_it).location = journal.next_free;
//(*dirty_it).state = ST_J_SUBMITTED;
// Move journal_data_pos
journal_data_pos += op->len;
if (journal_data_pos >= journal_len)
journal_data_pos = 512;
// Move journal.next_free and save last write for current sector
journal.next_free += op->len;
if (journal.next_free >= journal.len)
journal.next_free = 512;
journal.sector_info[journal.cur_sector].usage_count++;
journal.crc32_last = je->crc32;
op->pending_ops = 2;
op->used_journal_sector = 1 + journal.cur_sector;
}
in_process_ops.insert(op);
int ret = ringloop->submit();