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