Skip to main content

Write Ordering Control

Historically, the Linux® kernel block I/O stack (i.e., the block layer and the SCSI layer) has never guaranteed the exact execution order of block I/O requests. The exact execution order of block I/O requests cannot be guaranteed due to the asynchronous nature of the execution in the kernel of block I/O requests and the necessity of a fine-granularity lock model for the device request queue (to minimize lock-contention overhead when multiple contexts issue I/O requests to a block device at the same time).

A direct result of this design is the inability to give guarantees to a well- behaving ZBD-compliant application that write commands for a zone will be delivered in increasing LBA order (matching the zone sequential write constraint).

To address this problem, the kernel ZBD support adds zone write locking to ensure that write requests are processed in order per zone.

Zone Write Locking#

Zone write locking implements a per-zone write lock to serialize the execution of write requests that target the same zone. This feature does not guarantee that write commands are always issued at the location of the zone write pointer: this is the responsibility of the write I/O issuer. Zone write locking guarantees only that the order in which write commands are issued by an application, file system, or device mapper target will be respected by the block I/O stack. A well-behaved user of zoned block devices will thus avoid unaligned write command failures.

Zone write locking does not affect read commands in any way. Read requests are not serialized and processed in exactly the same manner as with regular block devices.

Initial Implementation#

Zone write locking was first implemented in kernel 4.10 in the SCSI disk driver (below the block layer), operating on requests already passed to the device dispatch queue by the block I/O scheduler.

This early implementation relied on the fact that the SCSI layer could delay issuing any request to the device. By maintaining a bitmap with one bit per zone, the SCSI disk driver marked a zone as locked whenever it saw a write command. This algorithm is presented here in more detail:

  1. If the next command to be dispatched to the device is not a write command, then the command is dispatched immediately.

  2. If the next command to by dispatched is a write command, then the zone write lock bit for the target zone of the command is inspected.

    1. If the target zone of the write command is not write locked (i.e., the bit is not set), then the zone is locked and the write command issued to the device. Both operations are atomically executed under the device request queue spinlock.
    2. If the target zone is already locked (i.e., the bit is set), then the SCSI disk driver temporarily delays issuing the command to the device until the zone write lock is released.
    3. When a write command completes, the zone write lock for the target zone of the command is released and the dispatch process is resumed. This means that if the command at the head of the dispatch queue targets the same zone, then the command can be issued (when the write command completes) (step 2.a).
note

Zone write locking that is implemented as shown above also prevents the unintended reordering of commands by the SAS HBAs or SATA adapters. The AHCI specifications do not define a clear order for issuing commands to devices. As a result, many chipsets are unable to guarantee the proper order of commands.

Although this implementation provides write-ordering guarantees for the legacy single-queue block I/O path and is not dependent upon any particular HBA, it has several limitations:

  • Potential performance degradation Any write command to any zone results in the command dispatch processing being stalled. This prevents all other command from being dispatched, including read commands. This can limit performance benefits that can be obtained with device-level command reordering when operating the device at high queue depth. The extreme case is an application issuing a write stream to a zone with asynchronous I/O system calls (e.g. io_submit()). In this case, the sequential write commands would be queued in sequence in the device dispatch queue, resulting in the drive being operated at a queue depth of one, one write command at a time.

  • No support for the block multi-queue I/O path Unlike the legacy single queue block I/O interface, the multi-queue block I/O implementation does not heavily rely on the device queue spin-lock to process block I/O requests issued by the disk users (applications or kernel components). This results in potential block I/O request reordering happening before requests are passed on to the device dispatch queue and the ineffectiveness of zone write locking.

These limitations led to the development of a new implementation of zone write locking at a higher level in the I/O stack, using the block layer I/O schedulers.

Improved Implementation: Block I/O Scheduler#

By moving zone write locking implementation higher up in the I/O stack, the block multi-queue (and SCSI multi-queue) infrastructure can also be supported. This improvement was added with kernel version 4.16 and the SCSI layer implementation of zone write locking was removed.

This new implementation of zone write locking relies on the block layer deadline and mq-deadline I/O schedulers and also addresses the performance limitations of the previous implementation. In detail, the new algorithm is as follows.

note

The deadline and mq-deadline schedulers operate by grouping commands per type (reads vs writes) and always processsing these two groups of commands separately, e.g. first issuing many reads, then many writes. This improves performance by taking advantage of hardware features such as device-level read-ahead.

  1. If the scheduler is processing read commands...

    1. ...the first command queued in the list of read commands is allowed to proceed and is submitted to the device dispatch queue.
    2. If no read commands are available, activate write processing (step 2).
    3. If the read-command processing time limit is reached, write-command processing (step 2) is activated to avoid write-command starvation.
    4. If read commands are still available, restart at step 1.
  2. When processing write commands, the list of write commands queued in the scheduler is scanned in order starting with the command at the head of the LBA ordered list or the first command in the arrival-time ordered list (when there is a risk of starving commands).

    1. If the target zone of the first write command is not write locked (zone bitmap bit not set), then the zone is locked and the write command is issued to the device. Both operations are atomically executed under a spinlock maintained by the scheduler.
    2. If the target zone is already locked (bit set), the command is skipped and the first write command that targets a different zone is searched for in the LBA ordered list of write commands. If such a command is found, step 2 is executed again.
    3. If all queued write commands target locked zones, the scheduler operation mode (batch mode) is switched to *read* and step 1 is called.
  3. When a write command completes, the zone write lock for the target zone of the command is released and the scheduler is activated. Operation is resumed at step 1 or 2 depending on the current batch mode.

From this algorithm, it is clear that the device can now be operated at higher queue depth and that only sequential writes that target the same zone will be throttled. All read commands can proceed, always, and write commands that target different zones do not impact each other.

note

This new implementation does not guarantee overall command ordering. Guarantees are given only for write commands that target the same zone. The dispatch order of write commands that target different zones may be changed by the scheduler. For any single sequential zone, at any time, there is always at most a single write command in-flight being executed. Overall disk operation at high queue depth is possible when there are read accesses and if multiple zones are being written simultaneously.

Block I/O Scheduler Configuration#

The deadline and mq-deadline schedulers must be enabled in the kernel compilation configuration. Refer to the Write Ordering Control section for details.

note

The legacy single queue block I/O path was removed from the kernel in version 5.0. As of kernel version 5.0, the deadline scheduler cannot be enabled. The mq-deadline scheduler is the only ZBD compliant scheduler.

Manual Configuration#

A system may define a default I/O scheduler other than deadline or mq-deadline. The block I/O scheduler for a zoned block device can be checked with the following command.

# cat /sys/block/sdb/queue/scheduler
[none] mq-deadline kyber bfq

If the block I/O scheduler selected is not deadline nor mq-deadline as in the example above, the scheduler can be changed with the following command.

# echo mq-deadline > /sys/block/sdb/queue/scheduler
# cat sys/block/sdb/queue/scheduler
[mq-deadline] kyber bfq none

deadline is an alias for the mq-deadline scheduler. The following command can therefore be used to get the same results in environments that use the legacy single queue I/O path (kernels 4.16 to 4.20) and environments that use the block multi-queue infrastructure (the sole possibility as of kernel version 5.0).

# echo deadline > /sys/block/sdb/queue/scheduler
# cat sys/block/sdb/queue/scheduler
[mq-deadline] kyber bfq none

Automatic Persistent Configuration#

Automatically configuring the deadline scheduler at system boot time can also be done using a udev rule. The procedure for defining a new udev rule varies slightly between distributions. Refer to your distribution documentation for details.

ACTION=="add|change", KERNEL=="sd*[!0-9]", ATTRS{queue/zoned}=="host-managed", ATTR{queue/scheduler}="deadline"

This rule sets up the deadline scheduler for any host-managed zoned block device found in the system. A host-aware zoned block disk can accept random writes and thus tolerate occasional write reordering within a zone sequential write stream. Nevertheless, write ordering can be maintained for these disks too by using the deadline scheduler. The above udev rule modified will automatically enable this.

ACTION=="add|change", KERNEL=="sd*[!0-9]", ATTRS{queue/zoned}=="host-aware", ATTR{queue/scheduler}="deadline"

The following single rule enables the deadline scheduler for any zoned block device, regardless of the device zone model.

ACTION=="add|change", KERNEL=="sd*[!0-9]", ATTRS{queue/zoned}!="none", ATTR{queue/scheduler}="deadline"