vitastor/src/blockstore_stable.cpp

562 rivejä
20 KiB
C++

// Copyright (c) Vitaliy Filippov, 2019+
// License: VNPL-1.1 (see README.md for details)
#include "blockstore_impl.h"
// Stabilize small write:
// 1) Copy data from the journal to the data device
// 2) Increase version on the metadata device and sync it
// 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 + 4K (data) [+ sync?] + 512b (metadata) + sync.
// WA = 2.375. It's not the best, SSD FTL-like redirect-write could probably be lower
// even with defragmentation. But it's fixed and it's still better than in Ceph. :)
// except for HDD-only clusters, because each write results in 3 seeks.
// Stabilize big write:
// 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 :)
// Stabilize delete:
// 1) Remove metadata entry and sync it
// 2) Remove dirty_db entry and clear previous journal entries
// We have 2 problems here:
// - In the cluster environment, we must store the "tombstones" of deleted objects until
// all replicas (not just quorum) agrees about their deletion. That is, "stabilize" is
// not possible for deletes in degraded placement groups
// - With simple "fixed" metadata tables we can't just clear the metadata entry of the latest
// object version. We must clear all previous entries, too.
// FIXME Fix both problems - probably, by switching from "fixed" metadata tables to "dynamic"
// AND We must do it in batches, for the sake of reduced fsync call count
// AND We must know what we stabilize. Basic workflow is like:
// 1) primary OSD receives sync request
// 2) it submits syncs to blockstore and peers
// 3) after everyone acks sync it acks sync to the client
// 4) after a while it takes his synced object list and sends stabilize requests
// to peers and to its own blockstore, thus freeing the old version
struct ver_vector_t
{
obj_ver_id *items = NULL;
uint64_t alloc = 0, size = 0;
};
static void init_versions(ver_vector_t & vec, obj_ver_id *start, obj_ver_id *end, uint64_t len)
{
if (!vec.items)
{
vec.alloc = len;
vec.items = (obj_ver_id*)malloc_or_die(sizeof(obj_ver_id) * vec.alloc);
for (auto sv = start; sv < end; sv++)
{
vec.items[vec.size++] = *sv;
}
}
}
static void append_version(ver_vector_t & vec, obj_ver_id ov)
{
if (vec.size >= vec.alloc)
{
vec.alloc = !vec.alloc ? 4 : vec.alloc*2;
vec.items = (obj_ver_id*)realloc_or_die(vec.items, sizeof(obj_ver_id) * vec.alloc);
}
vec.items[vec.size++] = ov;
}
static bool check_unsynced(std::vector<obj_ver_id> & check, obj_ver_id ov, std::vector<obj_ver_id> & to, int *count)
{
bool found = false;
int j = 0, k = 0;
while (j < check.size())
{
if (check[j] == ov)
found = true;
if (check[j].oid == ov.oid && check[j].version <= ov.version)
{
to.push_back(check[j++]);
if (count)
(*count)--;
}
else
check[k++] = check[j++];
}
check.resize(k);
return found;
}
blockstore_op_t* blockstore_impl_t::selective_sync(blockstore_op_t *op)
{
unsynced_big_write_count -= unsynced_big_writes.size();
unsynced_big_writes.swap(PRIV(op)->sync_big_writes);
unsynced_big_write_count += unsynced_big_writes.size();
unsynced_small_writes.swap(PRIV(op)->sync_small_writes);
// Create a sync operation, insert into the end of the queue
// And move ourselves into the end too!
// Rather hacky but that's what we need...
blockstore_op_t *sync_op = new blockstore_op_t;
sync_op->opcode = BS_OP_SYNC;
sync_op->buf = NULL;
sync_op->callback = [](blockstore_op_t *sync_op)
{
delete sync_op;
};
init_op(sync_op);
int sync_res = continue_sync(sync_op);
if (sync_res != 2)
{
// Put SYNC into the queue if it's not finished yet
submit_queue.push_back(sync_op);
}
// Restore unsynced_writes
unsynced_small_writes.swap(PRIV(op)->sync_small_writes);
unsynced_big_write_count -= unsynced_big_writes.size();
unsynced_big_writes.swap(PRIV(op)->sync_big_writes);
unsynced_big_write_count += unsynced_big_writes.size();
if (sync_res == 2)
{
// Sync is immediately completed
return NULL;
}
return sync_op;
}
// Returns: 2 = stop processing and dequeue, 0 = stop processing and do not dequeue, 1 = proceed with op itself
int blockstore_impl_t::split_stab_op(blockstore_op_t *op, std::function<int(obj_ver_id v)> decider)
{
bool add_sync = false;
ver_vector_t good_vers, bad_vers;
obj_ver_id* v;
int i, todo = 0;
for (i = 0, v = (obj_ver_id*)op->buf; i < op->len; i++, v++)
{
int action = decider(*v);
if (action < 0)
{
// Rollback changes
for (auto & ov: PRIV(op)->sync_big_writes)
{
unsynced_big_writes.push_back(ov);
unsynced_big_write_count++;
}
for (auto & ov: PRIV(op)->sync_small_writes)
{
unsynced_small_writes.push_back(ov);
}
free(good_vers.items);
good_vers.items = NULL;
free(bad_vers.items);
bad_vers.items = NULL;
// Error
op->retval = action;
FINISH_OP(op);
return 2;
}
else if (action == STAB_SPLIT_DONE)
{
// Already done
init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
}
else if (action == STAB_SPLIT_WAIT)
{
// Already in progress, we just have to wait until it finishes
init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
append_version(bad_vers, *v);
}
else if (action == STAB_SPLIT_SYNC)
{
// Needs a SYNC, we have to send a SYNC if not already in progress
//
// If the object is not present in unsynced_(big|small)_writes then
// it's currently being synced. If it's present then we can initiate
// its sync ourselves.
init_versions(good_vers, (obj_ver_id*)op->buf, v, op->len);
append_version(bad_vers, *v);
if (!add_sync)
{
PRIV(op)->sync_big_writes.clear();
PRIV(op)->sync_small_writes.clear();
add_sync = true;
}
check_unsynced(unsynced_small_writes, *v, PRIV(op)->sync_small_writes, NULL);
check_unsynced(unsynced_big_writes, *v, PRIV(op)->sync_big_writes, &unsynced_big_write_count);
}
else /* if (action == STAB_SPLIT_TODO) */
{
if (good_vers.items)
{
// If we're selecting versions then append it
// Main idea is that 99% of the time all versions passed to BS_OP_STABLE are synced
// And we don't want to select/allocate anything in that optimistic case
append_version(good_vers, *v);
}
todo++;
}
}
// In a pessimistic scenario, an operation may be split into 3:
// - Stabilize synced entries
// - Sync unsynced entries
// - Continue for unsynced entries after sync
add_sync = add_sync && (PRIV(op)->sync_big_writes.size() || PRIV(op)->sync_small_writes.size());
if (!todo && !bad_vers.size)
{
// Already stable
op->retval = 0;
FINISH_OP(op);
return 2;
}
op->retval = 0;
if (!todo && !add_sync)
{
// Only wait for inflight writes or current in-progress syncs
return 0;
}
blockstore_op_t *sync_op = NULL, *split_stab_op = NULL;
if (add_sync)
{
// Initiate a selective sync for PRIV(op)->sync_(big|small)_writes
sync_op = selective_sync(op);
}
if (bad_vers.size)
{
// Split part of the request into a separate operation
split_stab_op = new blockstore_op_t;
split_stab_op->opcode = op->opcode;
split_stab_op->buf = bad_vers.items;
split_stab_op->len = bad_vers.size;
init_op(split_stab_op);
submit_queue.push_back(split_stab_op);
}
if (sync_op || split_stab_op || good_vers.items)
{
void *orig_buf = op->buf;
if (good_vers.items)
{
op->buf = good_vers.items;
op->len = good_vers.size;
}
// Make a wrapped callback
int *split_op_counter = (int*)malloc_or_die(sizeof(int));
*split_op_counter = (sync_op ? 1 : 0) + (split_stab_op ? 1 : 0) + (todo ? 1 : 0);
auto cb = [op, good_items = good_vers.items,
bad_items = bad_vers.items, split_op_counter,
orig_buf, real_cb = op->callback](blockstore_op_t *split_op)
{
if (split_op->retval != 0)
op->retval = split_op->retval;
(*split_op_counter)--;
assert((*split_op_counter) >= 0);
if (op != split_op)
delete split_op;
if (!*split_op_counter)
{
free(good_items);
free(bad_items);
free(split_op_counter);
op->buf = orig_buf;
real_cb(op);
}
};
if (sync_op)
{
sync_op->callback = cb;
}
if (split_stab_op)
{
split_stab_op->callback = cb;
}
op->callback = cb;
}
if (!todo)
{
// All work is postponed
op->callback = NULL;
return 2;
}
return 1;
}
int blockstore_impl_t::dequeue_stable(blockstore_op_t *op)
{
if (PRIV(op)->op_state)
{
return continue_stable(op);
}
int r = split_stab_op(op, [this](obj_ver_id ov)
{
auto dirty_it = dirty_db.find(ov);
if (dirty_it == dirty_db.end())
{
auto & clean_db = clean_db_shard(ov.oid);
auto clean_it = clean_db.find(ov.oid);
if (clean_it == clean_db.end() || clean_it->second.version < ov.version)
{
// No such object version
printf("Error: %jx:%jx v%ju not found while stabilizing\n", ov.oid.inode, ov.oid.stripe, ov.version);
return -ENOENT;
}
else
{
// Already stable
return STAB_SPLIT_DONE;
}
}
else if (IS_STABLE(dirty_it->second.state))
{
// Already stable
return STAB_SPLIT_DONE;
}
while (true)
{
if (IS_IN_FLIGHT(dirty_it->second.state))
{
// Object write is still in progress. Wait until the write request completes
return STAB_SPLIT_WAIT;
}
else if (!IS_SYNCED(dirty_it->second.state))
{
// Object not synced yet - sync it
// In previous versions we returned EBUSY here and required
// the caller (OSD) to issue a global sync first. But a global sync
// waits for all writes in the queue including inflight writes. And
// inflight writes may themselves be blocked by unstable writes being
// still present in the journal and not flushed away from it.
// So we must sync specific objects here.
//
// Even more, we have to process "stabilize" request in parts. That is,
// we must stabilize all objects which are already synced. Otherwise
// they may block objects which are NOT synced yet.
return STAB_SPLIT_SYNC;
}
else if (IS_STABLE(dirty_it->second.state))
{
break;
}
// Check previous versions too
if (dirty_it == dirty_db.begin())
{
break;
}
dirty_it--;
if (dirty_it->first.oid != ov.oid)
{
break;
}
}
return STAB_SPLIT_TODO;
});
if (r != 1)
{
return r;
}
// Check journal space
blockstore_journal_check_t space_check(this);
if (!space_check.check_available(op, op->len, sizeof(journal_entry_stable), 0))
{
return 0;
}
// There is sufficient space. Check SQEs
BS_SUBMIT_CHECK_SQES(space_check.sectors_to_write);
// Prepare and submit journal entries
int s = 0;
auto v = (obj_ver_id*)op->buf;
for (int i = 0; i < op->len; i++, v++)
{
if (!journal.entry_fits(sizeof(journal_entry_stable)) &&
journal.sector_info[journal.cur_sector].dirty)
{
prepare_journal_sector_write(journal.cur_sector, op);
s++;
}
journal_entry_stable *je = (journal_entry_stable*)
prefill_single_journal_entry(journal, JE_STABLE, sizeof(journal_entry_stable));
je->oid = v->oid;
je->version = v->version;
je->crc32 = je_crc32((journal_entry*)je);
journal.crc32_last = je->crc32;
}
prepare_journal_sector_write(journal.cur_sector, op);
s++;
assert(s == space_check.sectors_to_write);
PRIV(op)->op_state = 1;
return 1;
}
int blockstore_impl_t::continue_stable(blockstore_op_t *op)
{
if (PRIV(op)->op_state == 2)
goto resume_2;
else if (PRIV(op)->op_state == 4)
goto resume_4;
else
return 1;
resume_2:
if (!disable_journal_fsync)
{
BS_SUBMIT_GET_SQE(sqe, data);
my_uring_prep_fsync(sqe, dsk.journal_fd, IORING_FSYNC_DATASYNC);
data->iov = { 0 };
data->callback = [this, op](ring_data_t *data) { handle_write_event(data, op); };
PRIV(op)->min_flushed_journal_sector = PRIV(op)->max_flushed_journal_sector = 0;
PRIV(op)->pending_ops = 1;
PRIV(op)->op_state = 3;
return 1;
}
resume_4:
// Mark dirty_db entries as stable, acknowledge op completion
obj_ver_id* v;
int i;
for (i = 0, v = (obj_ver_id*)op->buf; i < op->len; i++, v++)
{
// Mark all dirty_db entries up to op->version as stable
#ifdef BLOCKSTORE_DEBUG
printf("Stabilize %jx:%jx v%ju\n", v->oid.inode, v->oid.stripe, v->version);
#endif
mark_stable(*v);
}
// Acknowledge op
op->retval = 0;
FINISH_OP(op);
return 2;
}
void blockstore_impl_t::mark_stable(obj_ver_id v, bool forget_dirty)
{
auto dirty_it = dirty_db.find(v);
if (dirty_it != dirty_db.end())
{
if (IS_INSTANT(dirty_it->second.state))
{
// 'Instant' (non-EC) operations may complete and try to become stable out of order. Prevent it.
auto back_it = dirty_it;
while (back_it != dirty_db.begin())
{
back_it--;
if (back_it->first.oid != v.oid)
{
break;
}
if (!IS_STABLE(back_it->second.state))
{
// There are preceding unstable versions, can't flush <v>
return;
}
}
while (true)
{
dirty_it++;
if (dirty_it == dirty_db.end() || dirty_it->first.oid != v.oid ||
!IS_SYNCED(dirty_it->second.state))
{
dirty_it--;
break;
}
v.version = dirty_it->first.version;
}
}
while (1)
{
bool was_stable = IS_STABLE(dirty_it->second.state);
if ((dirty_it->second.state & BS_ST_WORKFLOW_MASK) == BS_ST_SYNCED)
{
dirty_it->second.state = (dirty_it->second.state & ~BS_ST_WORKFLOW_MASK) | BS_ST_STABLE;
// Allocations and deletions are counted when they're stabilized
if (IS_BIG_WRITE(dirty_it->second.state))
{
int exists = -1;
if (dirty_it != dirty_db.begin())
{
auto prev_it = dirty_it;
prev_it--;
if (prev_it->first.oid == v.oid)
{
exists = IS_DELETE(prev_it->second.state) ? 0 : 1;
}
}
if (exists == -1)
{
auto & clean_db = clean_db_shard(v.oid);
auto clean_it = clean_db.find(v.oid);
exists = clean_it != clean_db.end() ? 1 : 0;
}
if (!exists)
{
uint64_t space_id = dirty_it->first.oid.inode;
if (no_inode_stats[dirty_it->first.oid.inode >> (64-POOL_ID_BITS)])
space_id = space_id & ~(((uint64_t)1 << (64-POOL_ID_BITS)) - 1);
inode_space_stats[space_id] += dsk.data_block_size;
used_blocks++;
}
big_to_flush++;
}
else if (IS_DELETE(dirty_it->second.state))
{
uint64_t space_id = dirty_it->first.oid.inode;
if (no_inode_stats[dirty_it->first.oid.inode >> (64-POOL_ID_BITS)])
space_id = space_id & ~(((uint64_t)1 << (64-POOL_ID_BITS)) - 1);
auto & sp = inode_space_stats[space_id];
if (sp > dsk.data_block_size)
sp -= dsk.data_block_size;
else
inode_space_stats.erase(space_id);
used_blocks--;
big_to_flush++;
}
}
else if (IS_IN_FLIGHT(dirty_it->second.state))
{
// mark_stable should never be called for in-flight or submitted writes
printf(
"BUG: Attempt to mark_stable object %jx:%jx v%ju state of which is %x\n",
dirty_it->first.oid.inode, dirty_it->first.oid.stripe, dirty_it->first.version,
dirty_it->second.state
);
exit(1);
}
if (forget_dirty && (IS_BIG_WRITE(dirty_it->second.state) ||
IS_DELETE(dirty_it->second.state)))
{
// Big write overrides all previous dirty entries
auto erase_end = dirty_it;
while (dirty_it != dirty_db.begin())
{
dirty_it--;
if (dirty_it->first.oid != v.oid)
{
dirty_it++;
break;
}
}
auto & clean_db = clean_db_shard(v.oid);
auto clean_it = clean_db.find(v.oid);
uint64_t clean_loc = clean_it != clean_db.end()
? clean_it->second.location : UINT64_MAX;
erase_dirty(dirty_it, erase_end, clean_loc);
break;
}
if (was_stable || dirty_it == dirty_db.begin())
{
break;
}
dirty_it--;
if (dirty_it->first.oid != v.oid)
{
break;
}
}
flusher->enqueue_flush(v);
}
auto unstab_it = unstable_writes.find(v.oid);
if (unstab_it != unstable_writes.end() &&
unstab_it->second <= v.version)
{
unstable_writes.erase(unstab_it);
}
}