Add 2 tests for intent writes

This commit is contained in:
Vitaliy Filippov
2025-12-02 01:52:12 +03:00
parent 57d2f30303
commit e0d2705294
8 changed files with 204 additions and 54 deletions
+32 -25
View File
@@ -56,6 +56,7 @@ journal_flusher_t::~journal_flusher_t()
journal_flusher_co::~journal_flusher_co()
{
free_buffers();
if (csum_buf)
{
free(csum_buf);
@@ -235,7 +236,7 @@ resume_1:
#endif
flusher->active_flushers++;
// Scan versions to flush
read_vec.clear();
free_buffers();
for (auto wr = begin_wr; wr != end_wr; wr = wr->next())
{
bs->prepare_read(read_vec, cur_obj, wr, 0, bs->dsk.data_block_size);
@@ -295,18 +296,20 @@ resume_8:
else if (res == ENOENT || res == EDOM)
{
// Abort compaction
goto release_oid;
flusher->active_flushers--;
goto resume_0;
}
assert(res == 0);
// Submit data writes
copy_count = 0;
for (i = 0; i < read_vec.size(); i++)
{
if (read_vec[i].copy_flags & COPY_BUF_JOURNAL)
if ((read_vec[i].copy_flags & COPY_BUF_JOURNAL) &&
!(read_vec[i].copy_flags & COPY_BUF_COALESCED))
{
assert(read_vec[i].buf);
await_sqe(9);
data->iov = (struct iovec){ read_vec[i].buf, (size_t)read_vec[i].len };
data->iov = (struct iovec){ (bs->dsk.inmemory_journal ? bs->buffer_area + read_vec[i].disk_offset : read_vec[i].buf), (size_t)read_vec[i].len };
data->callback = simple_callback_w;
io_uring_prep_writev(sqe, bs->dsk.data_fd, &data->iov, 1, bs->dsk.data_offset + clean_loc + read_vec[i].offset);
wait_count++;
@@ -328,7 +331,7 @@ resume_10:
if (!cur_obj)
{
// Abort compaction
goto release_oid;
goto resume_0;
}
calc_block_checksums();
if (read_to_fill_incomplete)
@@ -341,8 +344,7 @@ resume_24:
}
}
bs->heap->mark_object_compacted(cur_obj, compact_lsn);
// Done, free all buffers
free_buffers();
// Done
#ifdef BLOCKSTORE_DEBUG
printf("Compacted %jx:%jx l%ju (%d writes)\n", cur_oid.inode, cur_oid.stripe, compact_lsn, copy_count);
#endif
@@ -360,7 +362,6 @@ resume_15:
if (!trim_lsn(11))
return false;
}
release_oid:
if (should_repeat)
{
// Flush the same object again
@@ -374,20 +375,23 @@ void journal_flusher_co::iterate_partial_overwrites(std::function<int(int, uint3
{
int prev = 0;
uint32_t prev_begin = 0, prev_end = 0;
for (int i = 0; i < read_vec.size() && !(read_vec[i].copy_flags & COPY_BUF_CSUM_FILL); i++)
for (int i = 0; i < read_vec.size(); i++)
{
if (read_vec[i].copy_flags != (COPY_BUF_JOURNAL|COPY_BUF_COALESCED))
if (!(read_vec[i].copy_flags & COPY_BUF_COALESCED))
{
if (read_vec[i].offset > prev_end)
{
i += cb(prev, prev_begin, prev_end);
if (prev_end > prev_begin && ((prev_begin % bs->dsk.csum_block_size) || (prev_end % bs->dsk.csum_block_size)))
{
i += cb(prev, prev_begin, prev_end);
}
prev = i;
prev_begin = read_vec[i].offset;
}
prev_end = read_vec[i].offset + read_vec[i].len;
}
}
if (prev_end > prev_begin)
if (prev_end > prev_begin && ((prev_begin % bs->dsk.csum_block_size) || (prev_end % bs->dsk.csum_block_size)))
{
cb(prev, prev_begin, prev_end);
}
@@ -401,14 +405,12 @@ void journal_flusher_co::iterate_checksum_holes(std::function<void(int, uint32_t
if ((prev_begin % bs->dsk.csum_block_size) &&
(prev_begin / bs->dsk.csum_block_size) != (prev_end / bs->dsk.csum_block_size))
{
cb(pos, prev_begin, prev_begin + bs->dsk.csum_block_size - prev_begin%bs->dsk.csum_block_size);
cb(pos, prev_begin - prev_begin%bs->dsk.csum_block_size, prev_begin);
r++;
}
if ((prev_end % bs->dsk.csum_block_size) ||
(prev_begin % bs->dsk.csum_block_size) &&
(prev_end / bs->dsk.csum_block_size) == (prev_begin / bs->dsk.csum_block_size))
if (prev_end % bs->dsk.csum_block_size)
{
cb(i, prev_end - (prev_end % bs->dsk.csum_block_size ? (prev_end % bs->dsk.csum_block_size) : bs->dsk.csum_block_size), prev_end);
cb(i, prev_end, prev_end - (prev_end % bs->dsk.csum_block_size) + bs->dsk.csum_block_size);
r++;
}
return r;
@@ -420,17 +422,18 @@ void journal_flusher_co::fill_partial_checksum_blocks()
iterate_checksum_holes([&](int vec_pos, uint32_t hole_start, uint32_t hole_end)
{
read_to_fill_incomplete = true;
int out_pos = read_vec.size();
bs->prepare_disk_read(read_vec, out_pos, cur_obj, end_wr,
bs->prepare_disk_read(read_vec, read_vec.size(), cur_obj, end_wr,
hole_start - hole_start % bs->dsk.csum_block_size, hole_start - hole_start % bs->dsk.csum_block_size + bs->dsk.csum_block_size,
hole_start - hole_start % bs->dsk.csum_block_size, hole_start - hole_start % bs->dsk.csum_block_size + bs->dsk.csum_block_size,
COPY_BUF_CSUM_FILL | (bs->padded_csum_update ? 0 : COPY_BUF_SKIP_CSUM));
out_pos--;
auto & vec = read_vec[read_vec.size()-1];
if (!vec.buf)
vec.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, vec.disk_len);
read_vec.insert(read_vec.begin()+vec_pos, (copy_buffer_t){
.copy_flags = COPY_BUF_JOURNAL|COPY_BUF_COALESCED,
.offset = hole_start,
.len = hole_end-hole_start,
.buf = read_vec[out_pos].buf + hole_start - read_vec[out_pos].offset,
.buf = vec.buf + hole_start - vec.offset,
});
});
}
@@ -458,7 +461,7 @@ int journal_flusher_co::check_and_punch_checksums()
return 0;
}
// Verify data checksums
cur_obj = bs->heap->read_locked_entry(cur_oid, copy_id);
cur_obj = bs->heap->read_locked_entry(cur_oid, copy_id); // FIXME locks can be removed from flusher
bool csum_ok = true;
for (int i = 0; i < read_vec.size(); i++)
{
@@ -469,9 +472,11 @@ int journal_flusher_co::check_and_punch_checksums()
while (wr && wr->lsn != vec.wr_lsn)
wr = wr->next();
assert(wr);
uint32_t *csums = (uint32_t*)(wr->get_checksums(bs->heap)
+ (vec.offset/bs->dsk.csum_block_size)*(bs->dsk.data_csum_type & 0xFF)
- (((wr->flags & BS_HEAP_TYPE) == BS_HEAP_BIG_WRITE) ? 0 : (wr->offset/bs->dsk.csum_block_size)*(bs->dsk.data_csum_type & 0xFF)));
bs->heap->calc_block_checksums(
(uint32_t*)((uint8_t*)wr->get_checksums(bs->heap) + vec.offset/bs->dsk.csum_block_size*(bs->dsk.data_csum_type & 0xFF)),
vec.buf, wr->get_int_bitmap(bs->heap), vec.offset, vec.offset+vec.len, false,
csums, vec.buf, wr->get_int_bitmap(bs->heap), vec.offset, vec.offset+vec.len, false,
[&](uint32_t mismatch_pos, uint32_t expected_csum, uint32_t real_csum)
{
printf("Checksum mismatch in object %jx:%jx v%ju in %s area at offset 0x%jx: got %08x, expected %08x\n",
@@ -546,7 +551,7 @@ void journal_flusher_co::calc_block_checksums()
for (auto it = read_vec.begin(); it != read_vec.end(); it++)
{
if (it->copy_flags & COPY_BUF_CSUM_FILL)
break;
continue;
if (block_done == 0)
{
// `read_vec` should contain aligned items, possibly split into pieces
@@ -637,6 +642,8 @@ bool journal_flusher_co::read_buffered(int wait_base)
{
await_sqe(0);
auto & vec = read_vec[i];
if (!vec.buf)
vec.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, vec.disk_len);
data->iov = (struct iovec){ vec.buf, (size_t)vec.disk_len };
wait_count++;
io_uring_prep_readv(
+2
View File
@@ -1366,7 +1366,9 @@ int blockstore_heap_t::update_object(uint32_t block_num, heap_object_t *obj, hea
else if ((wr->flags & BS_HEAP_TYPE) == BS_HEAP_INTENT_WRITE &&
(first_wr->flags & BS_HEAP_TYPE) == BS_HEAP_INTENT_WRITE)
{
// FIXME: All other types of writes should also purge&merge the intent write
auto second_wr = first_wr->next();
second_wr->version = first_wr->version;
bitmap_set(second_wr->get_int_bitmap(this), first_wr->offset, first_wr->len, dsk->bitmap_granularity);
if (dsk->csum_block_size)
{
+3 -6
View File
@@ -76,6 +76,7 @@ struct blockstore_op_private_t
class blockstore_impl_t: public blockstore_i
{
public:
blockstore_disk_t dsk;
/******* OPTIONS *******/
@@ -134,10 +135,6 @@ class blockstore_impl_t: public blockstore_i
return ringloop->get_sqe();
}
friend class blockstore_init_meta;
friend class journal_flusher_t;
friend class journal_flusher_co;
void open_data();
void open_meta();
void open_journal();
@@ -159,7 +156,7 @@ class blockstore_impl_t: public blockstore_i
uint32_t prepare_read_with_bitmaps(std::vector<copy_buffer_t> & read_vec, heap_object_t *obj, heap_write_t *wr, uint32_t start, uint32_t end);
uint32_t prepare_read_zero(std::vector<copy_buffer_t> & read_vec, uint32_t start, uint32_t end);
uint32_t prepare_read_simple(std::vector<copy_buffer_t> & read_vec, heap_object_t *obj, heap_write_t *wr, uint32_t start, uint32_t end);
void prepare_disk_read(std::vector<copy_buffer_t> & read_vec, int & pos, heap_object_t *obj, heap_write_t *wr,
void prepare_disk_read(std::vector<copy_buffer_t> & read_vec, int pos, heap_object_t *obj, heap_write_t *wr,
uint32_t blk_start, uint32_t blk_end, uint32_t start, uint32_t end, uint32_t copy_flags);
void find_holes(std::vector<copy_buffer_t> & read_vec, uint32_t item_start, uint32_t item_end,
std::function<void(int&, uint32_t, uint32_t)> callback);
@@ -186,7 +183,7 @@ class blockstore_impl_t: public blockstore_i
// List
void process_list(blockstore_op_t *op);
public:
/*public:*/
blockstore_impl_t(blockstore_config_t & config, ring_loop_i *ringloop, timerfd_manager_t *tfd, bool mock_mode = false);
~blockstore_impl_t();
+7 -7
View File
@@ -41,10 +41,10 @@
// Suspend operation until there are <wait_detail> bytes of free space in the journal on disk
#define WAIT_COMPACTION 2
#define COPY_BUF_JOURNAL 1
#define COPY_BUF_DATA 2
#define COPY_BUF_ZERO 4
#define COPY_BUF_CSUM_FILL 8
#define COPY_BUF_COALESCED 16
#define COPY_BUF_PADDED 32
#define COPY_BUF_SKIP_CSUM 64
#define COPY_BUF_JOURNAL 0x01
#define COPY_BUF_DATA 0x02
#define COPY_BUF_ZERO 0x04
#define COPY_BUF_CSUM_FILL 0x08
#define COPY_BUF_COALESCED 0x10
#define COPY_BUF_PADDED 0x20
#define COPY_BUF_SKIP_CSUM 0x40
+10 -10
View File
@@ -195,7 +195,7 @@ uint32_t blockstore_impl_t::prepare_read_simple(std::vector<copy_buffer_t> & rea
else if (dsk.csum_block_size <= dsk.bitmap_granularity)
{
// simple disk read
prepare_disk_read(read_vec, pos, obj, wr, start, end, start, end, 0);
prepare_disk_read(read_vec, pos++, obj, wr, start, end, start, end, 0);
}
else
{
@@ -221,7 +221,7 @@ uint32_t blockstore_impl_t::prepare_read_simple(std::vector<copy_buffer_t> & rea
blk_end == end && blk_start == start)
{
// single block, two partial blocks, or any number of full blocks
prepare_disk_read(read_vec, pos, obj, wr, blk_start, blk_end, start, end, skip_csum);
prepare_disk_read(read_vec, pos++, obj, wr, blk_start, blk_end, start, end, skip_csum);
}
else
{
@@ -229,18 +229,18 @@ uint32_t blockstore_impl_t::prepare_read_simple(std::vector<copy_buffer_t> & rea
uint32_t full_start = (blk_start != start ? blk_start+dsk.csum_block_size : blk_start);
uint32_t full_end = (blk_end != end ? blk_end-dsk.csum_block_size : blk_end);
if (blk_start != start)
prepare_disk_read(read_vec, pos, obj, wr, blk_start, full_start, start, full_start, skip_csum);
prepare_disk_read(read_vec, pos++, obj, wr, blk_start, full_start, start, full_start, skip_csum);
if (full_start > full_end)
prepare_disk_read(read_vec, pos, obj, wr, full_start, full_end, full_start, full_end, skip_csum);
prepare_disk_read(read_vec, pos++, obj, wr, full_start, full_end, full_start, full_end, skip_csum);
if (blk_end != end)
prepare_disk_read(read_vec, pos, obj, wr, full_end, blk_end, full_end, end, skip_csum);
prepare_disk_read(read_vec, pos++, obj, wr, full_end, blk_end, full_end, end, skip_csum);
}
}
});
return res;
}
void blockstore_impl_t::prepare_disk_read(std::vector<copy_buffer_t> & read_vec, int & pos, heap_object_t *obj, heap_write_t *wr,
void blockstore_impl_t::prepare_disk_read(std::vector<copy_buffer_t> & read_vec, int pos, heap_object_t *obj, heap_write_t *wr,
uint32_t blk_start, uint32_t blk_end, uint32_t start, uint32_t end, uint32_t copy_flags)
{
// Only one INTENT_WRITE is allowed at a time
@@ -256,19 +256,19 @@ void blockstore_impl_t::prepare_disk_read(std::vector<copy_buffer_t> & read_vec,
if (blk_start != start || blk_end != end)
{
vec.copy_flags |= COPY_BUF_PADDED;
if (read_vec.size() > pos && (read_vec[pos].copy_flags & ~COPY_BUF_CSUM_FILL) == vec.copy_flags &&
read_vec[pos].offset >= blk_start && read_vec[pos].offset+read_vec[pos].len <= blk_end)
if (pos > 0 && read_vec.size() >= pos && (read_vec[pos-1].copy_flags & ~COPY_BUF_CSUM_FILL) == vec.copy_flags &&
read_vec[pos-1].offset >= blk_start && read_vec[pos-1].offset+read_vec[pos-1].len <= blk_end)
{
// This is the same block as the previous one, we can read it only once
vec.copy_flags |= COPY_BUF_COALESCED;
vec.buf = read_vec[pos].buf;
vec.buf = read_vec[pos-1].buf;
}
else
{
vec.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, blk_end-blk_start);
}
}
read_vec.insert(read_vec.begin() + (pos++), vec);
read_vec.insert(read_vec.begin() + pos, vec);
}
void blockstore_impl_t::find_holes(std::vector<copy_buffer_t> & read_vec,
+1
View File
@@ -17,6 +17,7 @@ int blockstore_impl_t::dequeue_stable(blockstore_op_t *op)
// Modify in-memory state and assign contiguous LSNs
priv->stab_pos = 0;
priv->lsn = priv->to_lsn = 0;
op->retval = 0;
while (priv->stab_pos < op->len)
{
uint32_t modified_block = 0;
+7 -6
View File
@@ -177,13 +177,14 @@ int blockstore_impl_t::dequeue_write(blockstore_op_t *op)
{
// Direct intent-write
BS_SUBMIT_CHECK_SQES(1);
for (auto wr = obj->get_writes(); wr; wr = wr->next())
if ((obj->get_writes()->flags & BS_HEAP_TYPE) == BS_HEAP_BIG_WRITE)
{
assert(wr->flags != BS_HEAP_BIG_WRITE);
if (wr->flags == (BS_HEAP_BIG_WRITE|BS_HEAP_STABLE))
{
PRIV(op)->location = wr->location;
}
PRIV(op)->location = obj->get_writes()->location;
}
else
{
assert((obj->get_writes()->next()->flags & BS_HEAP_TYPE) == BS_HEAP_BIG_WRITE);
PRIV(op)->location = obj->get_writes()->next()->location;
}
process_intent:
uint8_t wr_buf[heap->get_max_write_entry_size()];
+142
View File
@@ -132,6 +132,14 @@ struct bs_test_t
}
};
static bool memcheck(uint8_t *buf, uint8_t byte, size_t len)
{
for (size_t i = 0; i < len; i++)
if (buf[i] != byte)
return false;
return true;
}
static void test_simple()
{
printf("\n-- test_simple\n");
@@ -274,10 +282,144 @@ static void test_fsync(bool separate_meta)
free(op2.buf);
}
static void test_intent_over_unstable()
{
printf("\n-- test_intent_over_unstable\n");
bs_test_t test;
test.default_cfg();
test.init();
// Write
printf("writing\n");
blockstore_op_t op;
op.opcode = BS_OP_WRITE;
op.oid = { .inode = 1, .stripe = 0 };
op.version = 1;
op.offset = 20480;
op.len = 4096;
op.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, 4096);
memset(op.buf, 0xaa, 4096);
test.exec_op(&op);
assert(op.retval == op.len);
// Write again
printf("writing again\n");
op.version = 2;
op.offset = 28*1024;
test.exec_op(&op);
assert(op.retval == op.len);
free(op.buf);
}
static void test_padded_csum_intent()
{
printf("\n-- test_padded_csum_intent\n");
bs_test_t test;
test.default_cfg();
test.config["csum_block_size"] = "16384";
test.init();
// Write
printf("writing\n");
blockstore_op_t op;
op.opcode = BS_OP_WRITE;
op.oid = { .inode = 1, .stripe = 0 };
op.version = 1;
op.offset = 8192;
op.len = 4096;
op.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, 4096);
memset(op.buf, 0xaa, 4096);
test.exec_op(&op);
assert(op.retval == op.len);
// Read
printf("reading\n");
blockstore_op_t op2;
op2.opcode = BS_OP_READ;
op2.oid = { .inode = 1, .stripe = 0 };
op2.version = UINT64_MAX;
op2.offset = 0;
op2.len = 128*1024;
op2.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, 128*1024);
test.exec_op(&op2);
assert(op2.retval == op2.len);
assert(is_zero(op2.buf, 8*1024));
assert(memcmp(op2.buf+8*1024, op.buf, 4*1024) == 0);
assert(is_zero(op2.buf+12*1024, 116*1024));
// Write again (intent)
printf("writing (intent)\n");
op.version = 2;
op.offset = 28*1024;
memset(op.buf, 0xbb, 4096);
test.exec_op(&op);
assert(op.retval == op.len);
// Write again (small because uncompactable)
printf("writing (small)\n");
op.version = 3;
op.offset = 60*1024;
memset(op.buf, 0xcc, 4096);
test.exec_op(&op);
assert(op.retval == op.len);
// Check that these are really big+intent+small writes
// (intent is not collapsible because of csum_block_size > bitmap_granularity)
heap_object_t *obj = test.bs->heap->read_entry((object_id){ .inode = 1, .stripe = 0 }, NULL);
assert(obj);
assert(!obj->get_writes()->next()->next()->next());
assert(obj->get_writes()->flags == BS_HEAP_SMALL_WRITE);
assert(obj->get_writes()->next()->flags == BS_HEAP_INTENT_WRITE);
assert(obj->get_writes()->next()->next()->flags == BS_HEAP_BIG_WRITE);
// Commit
printf("commit version 3\n");
op.opcode = BS_OP_STABLE;
op.len = 1;
*((obj_ver_id*)op.buf) = {
.oid = { .inode = 1, .stripe = 0 },
.version = 3,
};
test.exec_op(&op);
assert(op.retval == 0);
assert(test.bs->heap->get_compact_queue_size());
// Trigger & wait compaction
test.bs->flusher->request_trim();
// FIXME: Не зацикливаться при обломе
while (test.bs->heap->get_compact_queue_size())
test.ringloop->loop();
test.bs->flusher->release_trim();
// Read again and check
printf("reading compacted\n");
op2.version = UINT64_MAX;
test.exec_op(&op2);
assert(op2.retval == op2.len);
assert(memcheck(op2.buf, 0, 8*1024));
assert(memcheck(op2.buf+8*1024, 0xaa, 4*1024));
assert(memcheck(op2.buf+12*1024, 0, 16*1024));
assert(memcheck(op2.buf+28*1024, 0xbb, 4*1024));
assert(memcheck(op2.buf+32*1024, 0, 28*1024));
assert(memcheck(op2.buf+60*1024, 0xcc, 4*1024));
assert(memcheck(op2.buf+64*1024, 0, 64*1024));
obj = test.bs->heap->read_entry((object_id){ .inode = 1, .stripe = 0 }, NULL);
assert(!obj->get_writes()->next());
free(op.buf);
free(op2.buf);
}
int main(int narg, char *args[])
{
test_simple();
test_fsync(false);
test_fsync(true);
test_intent_over_unstable();
test_padded_csum_intent();
return 0;
}