Hi André, On 6/5/20 10:55 AM, André Almeida wrote: > Create a documentation providing a background and explanation around the > operation of the Multi-Queue Block IO Queueing Mechanism (blk-mq). > > The reference for writing this documentation was the source code and > "Linux Block IO: Introducing Multi-queue SSD Access on Multi-core > Systems", by Axboe et al. > > Signed-off-by: André Almeida <andrealmeid@xxxxxxxxxxxxx> > --- > Changes from v1: > - Fixed typos > - Reworked blk_mq_hw_ctx > > Hello, > > This commit was tested using "make htmldocs" and the HTML output has > been verified. > > Thanks, > André > --- > Documentation/block/blk-mq.rst | 154 +++++++++++++++++++++++++++++++++ > Documentation/block/index.rst | 1 + > 2 files changed, 155 insertions(+) > create mode 100644 Documentation/block/blk-mq.rst > > diff --git a/Documentation/block/blk-mq.rst b/Documentation/block/blk-mq.rst > new file mode 100644 > index 000000000000..1f702adbc577 > --- /dev/null > +++ b/Documentation/block/blk-mq.rst > @@ -0,0 +1,154 @@ > +.. SPDX-License-Identifier: GPL-2.0 > + > +================================================ > +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. > + > +Introduction > +============ > + > +Background > +---------- > + > +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 more slower > +than any layer on the storage stack. One example of such optimization technique > +involves ordering read/write requests accordingly 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`_). > + > +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 were 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 places 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 (represented by struct > +:c:type:`blk_mq_ctx`) in the software staging queues in case that they weren't > +sent directly to the driver. A request is a collection of BIOs. They arrived at > +the block layer through the data structure struct :c:type:`bio`. The block > +layer will then build a new structure from it, the struct :c:type:`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 > +<https://www.kernel.org/doc/html/latest/block/index.html>`_. 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 NOOP scheduler, the most straightforward one, that implements a > +simple FIFO, without performing any reordering. This is useful in the following > +scenarios: when scheduling will be performed in a next step somewhere in the > +stack, like block device controllers; the actual sector position of blocks are > +transparent for the host, meaning it hasn't enough information to take a proper > +decision; or the overhead of reordering is higher than the handicap of > +non-sequential accesses. > + > +Hardware dispatch queues > +~~~~~~~~~~~~~~~~~~~~~~~~ > + > +The hardware queues (represented by struct :c:type:`blk_mq_hw_ctx`) have a 1:1 > +correspondence to the device driver's submission queues, and are the last step I am not clear with the definition of "submission queues". Is it the device queue with DMA ring buffer? If it is the DMA ring buffer, multiple blk_mq_hw_ctx would map to the same DMA ring buffer, e.g., multiple nvme namespaces would share the same tagset. This is not 1:1 any longer. > +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 (:c:type:`hctx->dispatch`) of requests. Then, > +next time the block layer runs a queue, it will send the requests laying at the > +:c:type:`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. > + > +.. note:: > + > + 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 > +drive, 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. Assume I am a beginner and does not know about blk-mq well. What I expect is to expand this sections to explain the usage of sbitmap to manage tags, e.g., like the comments in block/blk-mq-tag.c or block/blk-mq-tag.h. In addition, I would be interested in that percpu-refcount is used to track the lifecycle of requests. I have no idea how much detail is required for a kernel doc. The is just the feedback from me by assuming the audience is beginner :) Thank you very much! Dongli Zhang
