Multi-Queue Block IO Queueing Mechanism (blk-mq)

The Multi-Queue Block IO Queueing Mechanism is an API to enable fast storage devices to achieve a huge number of input/output operations per second (IOPS) through queueing and submitting IO requests to block devices simultaneously, benefiting from the parallelism offered by modern storage devices.



Magnetic hard disks have been the de facto standard from the beginning of the development of the kernel. The Block IO subsystem aimed to achieve the best performance possible for those devices with a high penalty when doing random access, and the bottleneck was the mechanical moving parts, a lot slower than any layer on the storage stack. One example of such optimization technique involves ordering read/write requests according to the current position of the hard disk head.

However, with the development of Solid State Drives and Non-Volatile Memories without mechanical parts nor random access penalty and capable of performing high parallel access, the bottleneck of the stack had moved from the storage device to the operating system. In order to take advantage of the parallelism in those devices’ design, the multi-queue mechanism was introduced.

The former design had a single queue to store block IO requests with a single lock. That did not scale well in SMP systems due to dirty data in cache and the bottleneck of having a single lock for multiple processors. This setup also suffered with congestion when different processes (or the same process, moving to different CPUs) wanted to perform block IO. Instead of this, the blk-mq API spawns multiple queues with individual entry points local to the CPU, removing the need for a lock. A deeper explanation on how this works is covered in the following section (Operation).


When the userspace performs IO to a block device (reading or writing a file, for instance), blk-mq takes action: it will store and manage IO requests to the block device, acting as middleware between the userspace (and a file system, if present) and the block device driver.

blk-mq has two group of queues: software staging queues and hardware dispatch queues. When the request arrives at the block layer, it will try the shortest path possible: send it directly to the hardware queue. However, there are two cases that it might not do that: if there’s an IO scheduler attached at the layer or if we want to try to merge requests. In both cases, requests will be sent to the software queue.

Then, after the requests are processed by software queues, they will be placed at the hardware queue, a second stage queue where the hardware has direct access to process those requests. However, if the hardware does not have enough resources to accept more requests, blk-mq will place requests on a temporary queue, to be sent in the future, when the hardware is able.

Software staging queues

The block IO subsystem adds requests in the software staging queues (represented by struct blk_mq_ctx) in case that they weren’t sent directly to the driver. A request is one or more BIOs. They arrived at the block layer through the data structure struct bio. The block layer will then build a new structure from it, the struct request that will be used to communicate with the device driver. Each queue has its own lock and the number of queues is defined by a per-CPU or per-node basis.

The staging queue can be used to merge requests for adjacent sectors. For instance, requests for sector 3-6, 6-7, 7-9 can become one request for 3-9. Even if random access to SSDs and NVMs have the same time of response compared to sequential access, grouped requests for sequential access decreases the number of individual requests. This technique of merging requests is called plugging.

Along with that, the requests can be reordered to ensure fairness of system resources (e.g. to ensure that no application suffers from starvation) and/or to improve IO performance, by an IO scheduler.

IO Schedulers

There are several schedulers implemented by the block layer, each one following a heuristic to improve the IO performance. They are “pluggable” (as in plug and play), in the sense of they can be selected at run time using sysfs. You can read more about Linux’s IO schedulers here. The scheduling happens only between requests in the same queue, so it is not possible to merge requests from different queues, otherwise there would be cache trashing and a need to have a lock for each queue. After the scheduling, the requests are eligible to be sent to the hardware. One of the possible schedulers to be selected is the NONE scheduler, the most straightforward one. It will just place requests on whatever software queue the process is running on, without any reordering. When the device starts processing requests in the hardware queue (a.k.a. run the hardware queue), the software queues mapped to that hardware queue will be drained in sequence according to their mapping.

Hardware dispatch queues

The hardware queue (represented by struct blk_mq_hw_ctx) is a struct used by device drivers to map the device submission queues (or device DMA ring buffer), and are the last step of the block layer submission code before the low level device driver taking ownership of the request. To run this queue, the block layer removes requests from the associated software queues and tries to dispatch to the hardware.

If it’s not possible to send the requests directly to hardware, they will be added to a linked list (hctx->dispatch) of requests. Then, next time the block layer runs a queue, it will send the requests laying at the dispatch list first, to ensure a fairness dispatch with those requests that were ready to be sent first. The number of hardware queues depends on the number of hardware contexts supported by the hardware and its device driver, but it will not be more than the number of cores of the system. There is no reordering at this stage, and each software queue has a set of hardware queues to send requests for.


Neither the block layer nor the device protocols guarantee the order of completion of requests. This must be handled by higher layers, like the filesystem.

Tag-based completion

In order to indicate which request has been completed, every request is identified by an integer, ranging from 0 to the dispatch queue size. This tag is generated by the block layer and later reused by the device driver, removing the need to create a redundant identifier. When a request is completed in the driver, the tag is sent back to the block layer to notify it of the finalization. This removes the need to do a linear search to find out which IO has been completed.

Further reading

Source code documentation

void rq_list_move(struct request **src, struct request **dst, struct request *rq, struct request *prev)

move a struct request from one list to another


struct request **src

The source list rq is currently in

struct request **dst

The destination list that rq will be appended to

struct request *rq

The request to move

struct request *prev

The request preceding rq in src (NULL if rq is the head)

enum blk_eh_timer_return

How the timeout handler should proceed



The block driver completed the command or will complete it at a later time.


Reset the request timer and continue waiting for the request to complete.

struct blk_mq_hw_ctx

State for a hardware queue facing the hardware block device


struct blk_mq_hw_ctx {
    struct {
        spinlock_t lock;
        struct list_head        dispatch;
        unsigned long           state;
    struct delayed_work     run_work;
    cpumask_var_t cpumask;
    int next_cpu;
    int next_cpu_batch;
    unsigned long           flags;
    void *sched_data;
    struct request_queue    *queue;
    struct blk_flush_queue  *fq;
    void *driver_data;
    struct sbitmap          ctx_map;
    struct blk_mq_ctx       *dispatch_from;
    unsigned int            dispatch_busy;
    unsigned short          type;
    unsigned short          nr_ctx;
    struct blk_mq_ctx       **ctxs;
    spinlock_t dispatch_wait_lock;
    wait_queue_entry_t dispatch_wait;
    atomic_t wait_index;
    struct blk_mq_tags      *tags;
    struct blk_mq_tags      *sched_tags;
    unsigned int            numa_node;
    unsigned int            queue_num;
    atomic_t nr_active;
    struct hlist_node       cpuhp_online;
    struct hlist_node       cpuhp_dead;
    struct kobject          kobj;
    struct dentry           *debugfs_dir;
    struct dentry           *sched_debugfs_dir;
    struct list_head        hctx_list;





Protects the dispatch list.


Used for requests that are ready to be dispatched to the hardware but for some reason (e.g. lack of resources) could not be sent to the hardware. As soon as the driver can send new requests, requests at this list will be sent first for a fairer dispatch.


BLK_MQ_S_* flags. Defines the state of the hw queue (active, scheduled to restart, stopped).


Used for scheduling a hardware queue run at a later time.


Map of available CPUs where this hctx can run.


Used by blk_mq_hctx_next_cpu() for round-robin CPU selection from cpumask.


Counter of how many works left in the batch before changing to the next CPU.


BLK_MQ_F_* flags. Defines the behaviour of the queue.


Pointer owned by the IO scheduler attached to a request queue. It’s up to the IO scheduler how to use this pointer.


Pointer to the request queue that owns this hardware context.


Queue of requests that need to perform a flush operation.


Pointer to data owned by the block driver that created this hctx


Bitmap for each software queue. If bit is on, there is a pending request in that software queue.


Software queue to be used when no scheduler was selected.


Number used by blk_mq_update_dispatch_busy() to decide if the hw_queue is busy using Exponential Weighted Moving Average algorithm.


HCTX_TYPE_* flags. Type of hardware queue.


Number of software queues.


Array of software queues.


Lock for dispatch_wait queue.


Waitqueue to put requests when there is no tag available at the moment, to wait for another try in the future.


Index of next available dispatch_wait queue to insert requests.


Tags owned by the block driver. A tag at this set is only assigned when a request is dispatched from a hardware queue.


Tags owned by I/O scheduler. If there is an I/O scheduler associated with a request queue, a tag is assigned when that request is allocated. Else, this member is not used.


NUMA node the storage adapter has been connected to.


Index of this hardware queue.


Number of active requests. Only used when a tag set is shared across request queues.


List to store request if CPU is going to die


List to store request if some CPU die.


Kernel object for sysfs.


debugfs directory for this hardware queue. Named as cpu<cpu_number>.


debugfs directory for the scheduler.


if this hctx is not in use, this is an entry in q->unused_hctx_list.

struct blk_mq_queue_map

Map software queues to hardware queues


struct blk_mq_queue_map {
    unsigned int *mq_map;
    unsigned int nr_queues;
    unsigned int queue_offset;



CPU ID to hardware queue index map. This is an array with nr_cpu_ids elements. Each element has a value in the range [queue_offset, queue_offset + nr_queues).


Number of hardware queues to map CPU IDs onto.


First hardware queue to map onto. Used by the PCIe NVMe driver to map each hardware queue type (enum hctx_type) onto a distinct set of hardware queues.

enum hctx_type

Type of hardware queue



All I/O not otherwise accounted for.


Just for READ I/O.


Polled I/O of any kind.


Number of types of hctx.

struct blk_mq_tag_set

tag set that can be shared between request queues


struct blk_mq_tag_set {
    const struct blk_mq_ops *ops;
    struct blk_mq_queue_map map[HCTX_MAX_TYPES];
    unsigned int            nr_maps;
    unsigned int            nr_hw_queues;
    unsigned int            queue_depth;
    unsigned int            reserved_tags;
    unsigned int            cmd_size;
    int numa_node;
    unsigned int            timeout;
    unsigned int            flags;
    void *driver_data;
    struct blk_mq_tags      **tags;
    struct blk_mq_tags      *shared_tags;
    struct mutex            tag_list_lock;
    struct list_head        tag_list;
    struct srcu_struct      *srcu;



Pointers to functions that implement block driver behavior.


One or more ctx -> hctx mappings. One map exists for each hardware queue type (enum hctx_type) that the driver wishes to support. There are no restrictions on maps being of the same size, and it’s perfectly legal to share maps between types.


Number of elements in the map array. A number in the range [1, HCTX_MAX_TYPES].


Number of hardware queues supported by the block driver that owns this data structure.


Number of tags per hardware queue, reserved tags included.


Number of tags to set aside for BLK_MQ_REQ_RESERVED tag allocations.


Number of additional bytes to allocate per request. The block driver owns these additional bytes.


NUMA node the storage adapter has been connected to.


Request processing timeout in jiffies.


Zero or more BLK_MQ_F_* flags.


Pointer to data owned by the block driver that created this tag set.


Tag sets. One tag set per hardware queue. Has nr_hw_queues elements.


Shared set of tags. Has nr_hw_queues elements. If set, shared by all tags.


Serializes tag_list accesses.


List of the request queues that use this tag set. See also request_queue.tag_set_list.


Use as lock when type of the request queue is blocking (BLK_MQ_F_BLOCKING).

struct blk_mq_queue_data

Data about a request inserted in a queue


struct blk_mq_queue_data {
    struct request *rq;
    bool last;



Request pointer.


If it is the last request in the queue.

struct blk_mq_ops

Callback functions that implements block driver behaviour.


struct blk_mq_ops {
    blk_status_t (*queue_rq)(struct blk_mq_hw_ctx *, const struct blk_mq_queue_data *);
    void (*commit_rqs)(struct blk_mq_hw_ctx *);
    void (*queue_rqs)(struct request **rqlist);
    int (*get_budget)(struct request_queue *);
    void (*put_budget)(struct request_queue *, int);
    void (*set_rq_budget_token)(struct request *, int);
    int (*get_rq_budget_token)(struct request *);
    enum blk_eh_timer_return (*timeout)(struct request *);
    int (*poll)(struct blk_mq_hw_ctx *, struct io_comp_batch *);
    void (*complete)(struct request *);
    int (*init_hctx)(struct blk_mq_hw_ctx *, void *, unsigned int);
    void (*exit_hctx)(struct blk_mq_hw_ctx *, unsigned int);
    int (*init_request)(struct blk_mq_tag_set *set, struct request *, unsigned int, unsigned int);
    void (*exit_request)(struct blk_mq_tag_set *set, struct request *, unsigned int);
    void (*cleanup_rq)(struct request *);
    bool (*busy)(struct request_queue *);
    void (*map_queues)(struct blk_mq_tag_set *set);
    void (*show_rq)(struct seq_file *m, struct request *rq);



Queue a new request from block IO.


If a driver uses bd->last to judge when to submit requests to hardware, it must define this function. In case of errors that make us stop issuing further requests, this hook serves the purpose of kicking the hardware (which the last request otherwise would have done).


Queue a list of new requests. Driver is guaranteed that each request belongs to the same queue. If the driver doesn’t empty the rqlist completely, then the rest will be queued individually by the block layer upon return.


Reserve budget before queue request, once .queue_rq is run, it is driver’s responsibility to release the reserved budget. Also we have to handle failure case of .get_budget for avoiding I/O deadlock.


Release the reserved budget.


store rq’s budget token


retrieve rq’s budget token


Called on request timeout.


Called to poll for completion of a specific tag.


Mark the request as complete.


Called when the block layer side of a hardware queue has been set up, allowing the driver to allocate/init matching structures.


Ditto for exit/teardown.


Called for every command allocated by the block layer to allow the driver to set up driver specific data.

Tag greater than or equal to queue_depth is for setting up flush request.


Ditto for exit/teardown.


Called before freeing one request which isn’t completed yet, and usually for freeing the driver private data.


If set, returns whether or not this queue currently is busy.


This allows drivers specify their own queue mapping by overriding the setup-time function that builds the mq_map.


Used by the debugfs implementation to show driver-specific information about a request.

enum mq_rq_state blk_mq_rq_state(struct request *rq)

read the current MQ_RQ_* state of a request


struct request *rq

target request.

struct request *blk_mq_rq_from_pdu(void *pdu)

cast a PDU to a request


void *pdu

the PDU (Protocol Data Unit) to be casted




Driver command data is immediately after the request. So subtract request size to get back to the original request.

void *blk_mq_rq_to_pdu(struct request *rq)

cast a request to a PDU


struct request *rq

the request to be casted


pointer to the PDU


Driver command data is immediately after the request. So add request to get the PDU.

void blk_mq_wait_quiesce_done(struct blk_mq_tag_set *set)

wait until in-progress quiesce is done


struct blk_mq_tag_set *set

tag_set to wait on


it is driver’s responsibility for making sure that quiesce has been started on or more of the request_queues of the tag_set. This function only waits for the quiesce on those request_queues that had the quiesce flag set using blk_mq_quiesce_queue_nowait.

void blk_mq_quiesce_queue(struct request_queue *q)

wait until all ongoing dispatches have finished


struct request_queue *q

request queue.


this function does not prevent that the struct request end_io() callback function is invoked. Once this function is returned, we make sure no dispatch can happen until the queue is unquiesced via blk_mq_unquiesce_queue().

bool blk_update_request(struct request *req, blk_status_t error, unsigned int nr_bytes)

Complete multiple bytes without completing the request


struct request *req

the request being processed

blk_status_t error

block status code

unsigned int nr_bytes

number of bytes to complete for req


Ends I/O on a number of bytes attached to req, but doesn’t complete the request structure even if req doesn’t have leftover. If req has leftover, sets it up for the next range of segments.

Passing the result of blk_rq_bytes() as nr_bytes guarantees false return from this function.


The RQF_SPECIAL_PAYLOAD flag is ignored on purpose in this function except in the consistency check at the end of this function.


false - this request doesn’t have any more data true - this request has more data

void blk_mq_complete_request(struct request *rq)

end I/O on a request


struct request *rq

the request being processed


Complete a request by scheduling the ->complete_rq operation.

void blk_mq_start_request(struct request *rq)

Start processing a request


struct request *rq

Pointer to request to be started


Function used by device drivers to notify the block layer that a request is going to be processed now, so blk layer can do proper initializations such as starting the timeout timer.

void blk_execute_rq_nowait(struct request *rq, bool at_head)

insert a request to I/O scheduler for execution


struct request *rq

request to insert

bool at_head

insert request at head or tail of queue


Insert a fully prepared request at the back of the I/O scheduler queue for execution. Don’t wait for completion.


This function will invoke done directly if the queue is dead.

blk_status_t blk_execute_rq(struct request *rq, bool at_head)

insert a request into queue for execution


struct request *rq

request to insert

bool at_head

insert request at head or tail of queue


Insert a fully prepared request at the back of the I/O scheduler queue for execution and wait for completion.


The blk_status_t result provided to blk_mq_end_request().

void blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)

Run a hardware queue asynchronously.


struct blk_mq_hw_ctx *hctx

Pointer to the hardware queue to run.

unsigned long msecs

Milliseconds of delay to wait before running the queue.


Run a hardware queue asynchronously with a delay of msecs.

void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)

Start to run a hardware queue.


struct blk_mq_hw_ctx *hctx

Pointer to the hardware queue to run.

bool async

If we want to run the queue asynchronously.


Check if the request queue is not in a quiesced state and if there are pending requests to be sent. If this is true, run the queue to send requests to hardware.

void blk_mq_run_hw_queues(struct request_queue *q, bool async)

Run all hardware queues in a request queue.


struct request_queue *q

Pointer to the request queue to run.

bool async

If we want to run the queue asynchronously.

void blk_mq_delay_run_hw_queues(struct request_queue *q, unsigned long msecs)

Run all hardware queues asynchronously.


struct request_queue *q

Pointer to the request queue to run.

unsigned long msecs

Milliseconds of delay to wait before running the queues.

void blk_mq_request_bypass_insert(struct request *rq, blk_insert_t flags)

Insert a request at dispatch list.


struct request *rq

Pointer to request to be inserted.

blk_insert_t flags



Should only be used carefully, when the caller knows we want to bypass a potential IO scheduler on the target device.

void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx, struct request *rq)

Try to send a request directly to device driver.


struct blk_mq_hw_ctx *hctx

Pointer of the associated hardware queue.

struct request *rq

Pointer to request to be sent.


If the device has enough resources to accept a new request now, send the request directly to device driver. Else, insert at hctx->dispatch queue, so we can try send it another time in the future. Requests inserted at this queue have higher priority.

void blk_mq_submit_bio(struct bio *bio)

Create and send a request to block device.


struct bio *bio

Bio pointer.


Builds up a request structure from q and bio and send to the device. The request may not be queued directly to hardware if: * This request can be merged with another one * We want to place request at plug queue for possible future merging * There is an IO scheduler active at this queue

It will not queue the request if there is an error with the bio, or at the request creation.

blk_status_t blk_insert_cloned_request(struct request *rq)

Helper for stacking drivers to submit a request


struct request *rq

the request being queued

void blk_rq_unprep_clone(struct request *rq)

Helper function to free all bios in a cloned request


struct request *rq

the clone request to be cleaned up


Free all bios in rq for a cloned request.

int blk_rq_prep_clone(struct request *rq, struct request *rq_src, struct bio_set *bs, gfp_t gfp_mask, int (*bio_ctr)(struct bio*, struct bio*, void*), void *data)

Helper function to setup clone request


struct request *rq

the request to be setup

struct request *rq_src

original request to be cloned

struct bio_set *bs

bio_set that bios for clone are allocated from

gfp_t gfp_mask

memory allocation mask for bio

int (*bio_ctr)(struct bio *, struct bio *, void *)

setup function to be called for each clone bio. Returns 0 for success, non 0 for failure.

void *data

private data to be passed to bio_ctr


Clones bios in rq_src to rq, and copies attributes of rq_src to rq. Also, pages which the original bios are pointing to are not copied and the cloned bios just point same pages. So cloned bios must be completed before original bios, which means the caller must complete rq before rq_src.

void blk_mq_destroy_queue(struct request_queue *q)

shutdown a request queue


struct request_queue *q

request queue to shutdown


This shuts down a request queue allocated by blk_mq_alloc_queue(). All future requests will be failed with -ENODEV. The caller is responsible for dropping the reference from blk_mq_alloc_queue() by calling blk_put_queue().


can sleep