Block Layer Subsystem Details

Block Layer Subsystem Details

Queue Freezing Synchronization

blk_mq_freeze_queue() ensures all in-flight bios complete by waiting for q->q_usage_counter to reach zero. Every bio holds a reference to q->q_usage_counter from submission (blk_try_enter_queue) to completion (blk_queue_exit), so teardown functions protected by a freeze cannot race with bio completion paths.

Bio Merging

Merging combines bios into requests at the queue level. The merging rules establish which bios are allowed to be combined. When investigating a bug related to merged bios, first verify the merging rules permit that combination to exist. All bio operation types can pass through merge paths.

Bio Operation Type Safety

Accessing bio data fields on a bio that has no data buffers (e.g., discard, flush) causes a NULL pointer dereference. Code that handles bios must check the operation type before accessing data fields.

Data field accesses that require guards:

• Direct: bio->bi_io_vec, bio->bi_vcnt, bio->bi_iter.bi_bvec_done
• Indirect: bio_get_first_bvec(), bio_get_last_bvec(), bio_for_each_bvec(), bio_for_each_segment()

Required guard: bio_has_data() before accessing any data field. Note that op_is_write() is NOT a valid guard — it returns true for DISCARD, WRITE_ZEROES, and SECURE_ERASE (which have no data). Merge functions must also guard data field accesses since all bio operation types can reach them.

Bio Mempool Allocation Guarantees

Mempool-backed bio allocations (bio_alloc, bio_alloc_bioset) with GFP_NOIO/GFP_NOFS cannot fail (mempool guarantee via __GFP_DIRECT_RECLAIM). ENOMEM paths are only reachable with GFP_ATOMIC/GFP_NOWAIT. Note that bio_kmalloc() does NOT use a mempool and CAN fail regardless of GFP flags.

• bio_alloc_bioset() — mempool alloc, cannot fail with GFP_NOIO
• bvec_alloc() — falls back to mempool with __GFP_DIRECT_RECLAIM
• bio_integrity_prep() — uses GFP_NOIO, cannot fail
• bio_integrity_alloc_buf() — tries kmalloc with GFP_NOIO (without direct reclaim), falls back to mempool with GFP_NOFS; cannot fail