How to introduce Kernel into Linux Block layer Multi-queue

This article focuses on "how to introduce the kernel into Linux block-layer multi-queue". Interested friends may wish to take a look. The method introduced in this paper is simple, fast and practical. Let's let the editor take you to learn "how to introduce the kernel into the Linux block layer multi-queue".

First, take a look at the multi-queue architecture:

Take read IO as an example, single queue and multiple queues have the same execution path:

Read_pages () {... Blk_start_plug () / * process prepares current storage * / mapping- > BLK_MAX_REQUEST_COUNT-> readpages () / * current storage * / blk_finish_plug () / * process starts to drain * /...} io_schedule () process waits for io to complete (in the blk_mq_make_request () function, the number of request is greater than or equal to 16 request_count > = BLK_MAX_REQUEST_COUNT does not need to call io_schedule () Directly drain to the block device driver)

Mapping- > astatops-> readpages () will always call Q-> make_request_fn ()

Generic_make_request () Q-> make_request_fn () calls blk_queue_bio () or multi-queue blk_queue_make_request () _ _ elv_add_request ()

Why multiple queues are introduced: compared with single queues, there is a soft queue (represented by blk_mq_ctx structure) on each cpu to avoid using spinlock locks when inserting request, and today's high-speed storage devices, such as ssd that supports nvme (I just bought one, is really fast), the access latency is very small. And the hardware itself supports multi-queues (the introduced multi-queues replace request_queue- > delay_work with each hardware queue hctx- > delayed_work) the previous single-queue architecture can no longer drain its performance and has become an encumbrance. When a single queue inserts request and leaks to a block device driver, there is always a global spinlock lock on the request_queue, which makes people want to directly bypass the block layer.

The global spinlock lock on request_queue is used when a single queue is inserted into request

Blk_queue_bio () {... Spin_lock_irq (Q-> queue_lock); elv_merge () spin_lock_irq (Q-> queue_lock);...}

The global spinlock lock on request_queue is also used when a single queue is leaked to a block device driver:

Struct request_queue * blk_alloc_queue_node () INIT_DELAYED_WORK (& Q-> delay_work, blk_delay_work); blk_delay_work () _ blk_run_queue () Q-> request_fn (Q)

The _ _ blk_run_queue () function must be in the queue lock, that is, spin_lock_irq (Q-> queue_lock)

Blk_run_queue-run a single device queue 282 * @ Q: The queue to run 283 * 284 * Description: 285 * See @ blk_run_queue. This variant must be called with the queue lock 286 * held and interrupts disabled. 287 * / 288 void _ blk_run_queue (struct request_queue * Q) 289 {290 if (unlikely (blk_queue_stopped (Q) 291 return; 292 293 _ blk_run_queue_uncond (Q); 294}

The spinlock lock is not used when multiple queues are inserted into request:

Blk_mq_insert_requests () _ blk_mq_insert_request () struct blk_mq_ctx * ctx = rq- > mq_ctx; (blk_mq_ctx on each cpu) list_add_tail (& rq- > queuelist, & ctx- > rq_list)

The multi-queue drain to the block device driver also does not use spinlock locks:

Static int blk_mq_init_hw_queues () INIT_DELAYED_WORK (& hctx- > delayed_work, blk_mq_work_fn); 708 static void blk_mq_work_fn (struct work_struct * work) 709 {710 struct blk_mq_hw_ctx * hctx; 711712 hctx = container_of (work, struct blk_mq_hw_ctx, delayed_work.work) 713 _ _ blk_mq_run_hw_queue (hctx); 714} _ blk_mq_run_hw_queue () has no spinlock lock Q-> mq_ops- > queue_rq (hctx, rq); execute the-> queue_rq () callback function on multiple queues

The performance improvement of the system after using multiple queues can be seen from the following figure:

(I have not tested the performance myself, so it should be consistent with the following figure in terms of objective imagination:)

At this point, I believe you have a deeper understanding of "how to introduce the kernel into Linux block-layer multi-queue". You might as well do it in practice. Here is the website, more related content can enter the relevant channels to inquire, follow us, continue to learn!