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

Treat copied buffers as written only after completing the write in client

Browse Source 
SYNC operation fsyncs only completed operations, so treating writes as "eligible
for fsync" before actually completing them is incorrect

It affected SCHEME=ec test_heal.sh (with immediate_commit=none) test - it was
flapping with lost writes - some non-fsynced writes were legitimately lost by
the OSD, but weren't repeated by the client
This commit is contained in:
Vitaliy Filippov
2024-06-20 02:11:53 +03:00
parent e400a851f4
commit ea0d72289c
5 changed files with 34 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/client/cluster_client_wb.cpp
+18 -12
View File
@@ -71,7 +71,7 @@ bool writeback_cache_t::is_merged(const dirty_buf_it_t & dirty_it)
return is_left_merged(dirty_it) || is_right_merged(dirty_it);
}
void writeback_cache_t::copy_write(cluster_op_t *op, int state)
void writeback_cache_t::copy_write(cluster_op_t *op, int state, uint64_t new_flush_id)
{
// Save operation for replay when one of PGs goes out of sync
// (primary OSD drops our connection in this case)
@@ -180,6 +180,7 @@ void writeback_cache_t::copy_write(cluster_op_t *op, int state)
.buf = buf,
.len = op->len,
.state = state,
.flush_id = new_flush_id,
.refcnt = refcnt,
});
if (state == CACHE_DIRTY)
@@ -268,7 +269,7 @@ void writeback_cache_t::flush_buffers(cluster_client_t *cli, dirty_buf_it_t from
writebacks_active++;
op->callback = [this, flush_id](cluster_op_t* op)
{
// Buffer flushes should be always retried, regardless of the error,
// Buffer flushes are always retried, regardless of the error,
// so they should never result in an error here
assert(op->retval == op->len);
for (auto fl_it = flushed_buffers.find(flush_id);
@@ -280,16 +281,7 @@ void writeback_cache_t::flush_buffers(cluster_client_t *cli, dirty_buf_it_t from
}
flushed_buffers.erase(fl_it++);
}
for (auto dirty_it = find_dirty(op->inode, op->offset);
dirty_it != dirty_buffers.end() && dirty_it->first.inode == op->inode &&
dirty_it->first.stripe < op->offset+op->len; dirty_it++)
{
if (dirty_it->second.flush_id == flush_id && dirty_it->second.state == CACHE_REPEATING)
{
dirty_it->second.flush_id = 0;
dirty_it->second.state = CACHE_WRITTEN;
}
}
mark_flush_written(op->inode, op->offset, op->len, flush_id);
delete op;
writebacks_active--;
// We can't call execute_internal because it affects an invalid copy of the list here
@@ -307,6 +299,20 @@ void writeback_cache_t::flush_buffers(cluster_client_t *cli, dirty_buf_it_t from
}
}
void writeback_cache_t::mark_flush_written(uint64_t inode, uint64_t offset, uint64_t len, uint64_t flush_id)
{
for (auto dirty_it = find_dirty(inode, offset);
dirty_it != dirty_buffers.end() && dirty_it->first.inode == inode &&
dirty_it->first.stripe < offset+len; dirty_it++)
{
if (dirty_it->second.flush_id == flush_id && dirty_it->second.state == CACHE_REPEATING)
{
dirty_it->second.flush_id = 0;
dirty_it->second.state = CACHE_WRITTEN;
}
}
}
void writeback_cache_t::start_writebacks(cluster_client_t *cli, int count)
{
if (!writeback_queue.size())