Fix false-positive -EDOM checksum errors on overlapping unaligned reads #79

Closed
changwei.ye wants to merge 1 commits from changwei.ye/vitastor:fix-intent-write-corruption into master
2 changed files with 60 additions and 1 deletions

View File

@@ -75,10 +75,11 @@ int blockstore_impl_t::dequeue_read(blockstore_op_t *op)
return false;
}
}
if (need_skip && wr->type() == BS_HEAP_SMALL_WRITE &&
if (need_skip && (wr->type() == BS_HEAP_SMALL_WRITE || wr->type() == BS_HEAP_INTENT_WRITE) &&
wr->small().offset < blk_end && wr->small().offset+wr->small().len > blk_start)
{
// Small write may mutate big write checksums during flush
// Intent write directly mutates physical data in-place
skip_csum = COPY_BUF_SKIP_CSUM;
}
return true;

View File

@@ -1091,6 +1091,63 @@ static void test_list_limit()
free(op.buf);
}
static void test_intent_write_corruption()
{
printf("\n-- test_intent_write_corruption\n");
bs_test_t test;
test.default_cfg();
test.config["csum_block_size"] = "16384";
test.config["bitmap_granularity"] = "4096";
test.config["perfect_csum_update"] = "false";
test.init();
// 1. Write BIG_WRITE (0-48K)
blockstore_op_t op;
op.opcode = BS_OP_WRITE_STABLE;
op.oid = { .inode = 1, .stripe = 0 };
op.version = 1;
op.offset = 0;
op.len = 49152; // 48K
op.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, op.len);
memset(op.buf, 0xaa, op.len);
test.exec_op(&op);
assert(op.retval == op.len);
// 2. Write unaligned INTENT_WRITE B (8K-12K)
free(op.buf);
op.opcode = BS_OP_WRITE_STABLE;
op.oid = { .inode = 1, .stripe = 0 };
op.version = 2;
op.offset = 8192; // 8K
op.len = 4096; // 4K
op.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, op.len);
memset(op.buf, 0xbb, op.len);
test.exec_op(&op);
assert(op.retval == op.len);
// 3. Read 12-40K
// This will hit the hole logic where B provided 8-12K, but we don't even request it.
// The request is for 12-40K.
// 12K is not aligned to 16K, so A must pad by reading 0-16K.
// However, 8-12K of the physical block has been overwritten by B.
// A will try to verify the 0-16K block with its old checksum and fail.
free(op.buf);
op.opcode = BS_OP_READ;
op.oid = { .inode = 1, .stripe = 0 };
op.version = UINT64_MAX; // latest version
op.offset = 12288; // 12K
op.len = 28672; // 28K (so end = 40K)
op.buf = (uint8_t*)memalign_or_die(MEM_ALIGNMENT, op.len);
test.exec_op(&op);
// Assert that the read succeeds without checksum corruption errors
assert(op.retval == op.len);
free(op.buf);
}
// FIXME Add a simple intent_write / big_intent test
int main(int narg, char *args[])
@@ -1100,6 +1157,7 @@ int main(int narg, char *args[])
test_fsync(true);
test_fsync_meta_before_complete();
test_intent_over_unstable();
test_intent_write_corruption();
test_padded_csum_intent(false);
test_padded_csum_intent(true);
test_padded_csum_sparse_leading_hole();