1.. SPDX-License-Identifier: GPL-2.0
2
3================================================
4Multi-Queue Block IO Queueing Mechanism (blk-mq)
5================================================
6
7The Multi-Queue Block IO Queueing Mechanism is an API to enable fast storage
8devices to achieve a huge number of input/output operations per second (IOPS)
9through queueing and submitting IO requests to block devices simultaneously,
10benefiting from the parallelism offered by modern storage devices.
11
12Introduction
13============
14
15Background
16----------
17
18Magnetic hard disks have been the de facto standard from the beginning of the
19development of the kernel. The Block IO subsystem aimed to achieve the best
20performance possible for those devices with a high penalty when doing random
21access, and the bottleneck was the mechanical moving parts, a lot slower than
22any layer on the storage stack. One example of such optimization technique
23involves ordering read/write requests according to the current position of the
24hard disk head.
25
26However, with the development of Solid State Drives and Non-Volatile Memories
27without mechanical parts nor random access penalty and capable of performing
28high parallel access, the bottleneck of the stack had moved from the storage
29device to the operating system. In order to take advantage of the parallelism
30in those devices' design, the multi-queue mechanism was introduced.
31
32The former design had a single queue to store block IO requests with a single
33lock. That did not scale well in SMP systems due to dirty data in cache and the
34bottleneck of having a single lock for multiple processors. This setup also
35suffered with congestion when different processes (or the same process, moving
36to different CPUs) wanted to perform block IO. Instead of this, the blk-mq API
37spawns multiple queues with individual entry points local to the CPU, removing
38the need for a lock. A deeper explanation on how this works is covered in the
39following section (`Operation`_).
40
41Operation
42---------
43
44When the userspace performs IO to a block device (reading or writing a file,
45for instance), blk-mq takes action: it will store and manage IO requests to
46the block device, acting as middleware between the userspace (and a file
47system, if present) and the block device driver.
48
49blk-mq has two group of queues: software staging queues and hardware dispatch
50queues. When the request arrives at the block layer, it will try the shortest
51path possible: send it directly to the hardware queue. However, there are two
52cases that it might not do that: if there's an IO scheduler attached at the
53layer or if we want to try to merge requests. In both cases, requests will be
54sent to the software queue.
55
56Then, after the requests are processed by software queues, they will be placed
57at the hardware queue, a second stage queue were the hardware has direct access
58to process those requests. However, if the hardware does not have enough
59resources to accept more requests, blk-mq will places requests on a temporary
60queue, to be sent in the future, when the hardware is able.
61
62Software staging queues
63~~~~~~~~~~~~~~~~~~~~~~~
64
65The block IO subsystem adds requests  in the software staging queues
66(represented by struct :c:type:`blk_mq_ctx`) in case that they weren't sent
67directly to the driver. A request is one or more BIOs. They arrived at the
68block layer through the data structure struct :c:type:`bio`. The block layer
69will then build a new structure from it, the struct :c:type:`request` that will
70be used to communicate with the device driver. Each queue has its own lock and
71the number of queues is defined by a per-CPU or per-node basis.
72
73The staging queue can be used to merge requests for adjacent sectors. For
74instance, requests for sector 3-6, 6-7, 7-9 can become one request for 3-9.
75Even if random access to SSDs and NVMs have the same time of response compared
76to sequential access, grouped requests for sequential access decreases the
77number of individual requests. This technique of merging requests is called
78plugging.
79
80Along with that, the requests can be reordered to ensure fairness of system
81resources (e.g. to ensure that no application suffers from starvation) and/or to
82improve IO performance, by an IO scheduler.
83
84IO Schedulers
85^^^^^^^^^^^^^
86
87There are several schedulers implemented by the block layer, each one following
88a heuristic to improve the IO performance. They are "pluggable" (as in plug
89and play), in the sense of they can be selected at run time using sysfs. You
90can read more about Linux's IO schedulers `here
91<https://www.kernel.org/doc/html/latest/block/index.html>`_. The scheduling
92happens only between requests in the same queue, so it is not possible to merge
93requests from different queues, otherwise there would be cache trashing and a
94need to have a lock for each queue. After the scheduling, the requests are
95eligible to be sent to the hardware. One of the possible schedulers to be
96selected is the NONE scheduler, the most straightforward one. It will just
97place requests on whatever software queue the process is running on, without
98any reordering. When the device starts processing requests in the hardware
99queue (a.k.a. run the hardware queue), the software queues mapped to that
100hardware queue will be drained in sequence according to their mapping.
101
102Hardware dispatch queues
103~~~~~~~~~~~~~~~~~~~~~~~~
104
105The hardware queue (represented by struct :c:type:`blk_mq_hw_ctx`) is a struct
106used by device drivers to map the device submission queues (or device DMA ring
107buffer), and are the last step of the block layer submission code before the
108low level device driver taking ownership of the request. To run this queue, the
109block layer removes requests from the associated software queues and tries to
110dispatch to the hardware.
111
112If it's not possible to send the requests directly to hardware, they will be
113added to a linked list (:c:type:`hctx->dispatch`) of requests. Then,
114next time the block layer runs a queue, it will send the requests laying at the
115:c:type:`dispatch` list first, to ensure a fairness dispatch with those
116requests that were ready to be sent first. The number of hardware queues
117depends on the number of hardware contexts supported by the hardware and its
118device driver, but it will not be more than the number of cores of the system.
119There is no reordering at this stage, and each software queue has a set of
120hardware queues to send requests for.
121
122.. note::
123
124        Neither the block layer nor the device protocols guarantee
125        the order of completion of requests. This must be handled by
126        higher layers, like the filesystem.
127
128Tag-based completion
129~~~~~~~~~~~~~~~~~~~~
130
131In order to indicate which request has been completed, every request is
132identified by an integer, ranging from 0 to the dispatch queue size. This tag
133is generated by the block layer and later reused by the device driver, removing
134the need to create a redundant identifier. When a request is completed in the
135drive, the tag is sent back to the block layer to notify it of the finalization.
136This removes the need to do a linear search to find out which IO has been
137completed.
138
139Further reading
140---------------
141
142- `Linux Block IO: Introducing Multi-queue SSD Access on Multi-core Systems <http://kernel.dk/blk-mq.pdf>`_
143
144- `NOOP scheduler <https://en.wikipedia.org/wiki/Noop_scheduler>`_
145
146- `Null block device driver <https://www.kernel.org/doc/html/latest/block/null_blk.html>`_
147
148Source code documentation
149=========================
150
151.. kernel-doc:: include/linux/blk-mq.h
152
153.. kernel-doc:: block/blk-mq.c
154