1===================== 2I/O statistics fields 3===================== 4 5The kernel exposes disk statistics via ``/proc/diskstats`` and 6``/sys/block/<device>/stat``. These stats are usually accessed via tools 7such as ``sar`` and ``iostat``. 8 9Here are examples using a disk with two partitions:: 10 11 /proc/diskstats: 12 259 0 nvme0n1 255999 814 12369153 47919 996852 81 36123024 425995 0 301795 580470 0 0 0 0 60602 106555 13 259 1 nvme0n1p1 492 813 17572 96 848 81 108288 210 0 76 307 0 0 0 0 0 0 14 259 2 nvme0n1p2 255401 1 12343477 47799 996004 0 36014736 425784 0 344336 473584 0 0 0 0 0 0 15 16 /sys/block/nvme0n1/stat: 17 255999 814 12369153 47919 996858 81 36123056 426009 0 301809 580491 0 0 0 0 60605 106562 18 19 /sys/block/nvme0n1/nvme0n1p1/stat: 20 492 813 17572 96 848 81 108288 210 0 76 307 0 0 0 0 0 0 21 22Both files contain the same 17 statistics. ``/sys/block/<device>/stat`` 23contains the fields for ``<device>``. In ``/proc/diskstats`` the fields 24are prefixed with the major and minor device numbers and the device 25name. In the example above, the first stat value for ``nvme0n1`` is 26255999 in both files. 27 28The sysfs ``stat`` file is efficient for monitoring a small, known set 29of disks. If you're tracking a large number of devices, 30``/proc/diskstats`` is often the better choice since it avoids the 31overhead of opening and closing multiple files for each snapshot. 32 33All fields are cumulative, monotonic counters, except for field 9, which 34resets to zero as I/Os complete. The remaining fields reset at boot, on 35device reattachment or reinitialization, or when the underlying counter 36overflows. Applications reading these counters should detect and handle 37resets when comparing stat snapshots. 38 39Each set of stats only applies to the indicated device; if you want 40system-wide stats you'll have to find all the devices and sum them all up. 41 42Field 1 -- # of reads completed (unsigned long) 43 This is the total number of reads completed successfully. 44 45Field 2 -- # of reads merged, field 6 -- # of writes merged (unsigned long) 46 Reads and writes which are adjacent to each other may be merged for 47 efficiency. Thus two 4K reads may become one 8K read before it is 48 ultimately handed to the disk, and so it will be counted (and queued) 49 as only one I/O. This field lets you know how often this was done. 50 51Field 3 -- # of sectors read (unsigned long) 52 This is the total number of sectors read successfully. 53 54Field 4 -- # of milliseconds spent reading (unsigned int) 55 This is the total number of milliseconds spent by all reads (as 56 measured from blk_mq_alloc_request() to __blk_mq_end_request()). 57 58Field 5 -- # of writes completed (unsigned long) 59 This is the total number of writes completed successfully. 60 61Field 6 -- # of writes merged (unsigned long) 62 See the description of field 2. 63 64Field 7 -- # of sectors written (unsigned long) 65 This is the total number of sectors written successfully. 66 67Field 8 -- # of milliseconds spent writing (unsigned int) 68 This is the total number of milliseconds spent by all writes (as 69 measured from blk_mq_alloc_request() to __blk_mq_end_request()). 70 71Field 9 -- # of I/Os currently in progress (unsigned int) 72 The only field that should go to zero. Incremented as requests are 73 given to appropriate struct request_queue and decremented as they finish. 74 75Field 10 -- # of milliseconds spent doing I/Os (unsigned int) 76 This field increases so long as field 9 is nonzero. 77 78 Since 5.0 this field counts jiffies when at least one request was 79 started or completed. If request runs more than 2 jiffies then some 80 I/O time might be not accounted in case of concurrent requests. 81 82Field 11 -- weighted # of milliseconds spent doing I/Os (unsigned int) 83 This field is incremented at each I/O start, I/O completion, I/O 84 merge, or read of these stats by the number of I/Os in progress 85 (field 9) times the number of milliseconds spent doing I/O since the 86 last update of this field. This can provide an easy measure of both 87 I/O completion time and the backlog that may be accumulating. 88 89Field 12 -- # of discards completed (unsigned long) 90 This is the total number of discards completed successfully. 91 92Field 13 -- # of discards merged (unsigned long) 93 See the description of field 2 94 95Field 14 -- # of sectors discarded (unsigned long) 96 This is the total number of sectors discarded successfully. 97 98Field 15 -- # of milliseconds spent discarding (unsigned int) 99 This is the total number of milliseconds spent by all discards (as 100 measured from blk_mq_alloc_request() to __blk_mq_end_request()). 101 102Field 16 -- # of flush requests completed 103 This is the total number of flush requests completed successfully. 104 105 Block layer combines flush requests and executes at most one at a time. 106 This counts flush requests executed by disk. Not tracked for partitions. 107 108Field 17 -- # of milliseconds spent flushing 109 This is the total number of milliseconds spent by all flush requests. 110 111To avoid introducing performance bottlenecks, no locks are held while 112modifying these counters. This implies that minor inaccuracies may be 113introduced when changes collide, so (for instance) adding up all the 114read I/Os issued per partition should equal those made to the disks ... 115but due to the lack of locking it may only be very close. 116 117In 2.6+, there are counters for each CPU, which make the lack of locking 118almost a non-issue. When the statistics are read, the per-CPU counters 119are summed (possibly overflowing the unsigned long variable they are 120summed to) and the result given to the user. There is no convenient 121user interface for accessing the per-CPU counters themselves. 122 123Since 4.19 request times are measured with nanoseconds precision and 124truncated to milliseconds before showing in this interface. 125 126Disks vs Partitions 127------------------- 128 129There were significant changes between 2.4 and 2.6+ in the I/O subsystem. 130As a result, some statistic information disappeared. The translation from 131a disk address relative to a partition to the disk address relative to 132the host disk happens much earlier. All merges and timings now happen 133at the disk level rather than at both the disk and partition level as 134in 2.4. Consequently, you'll see a different statistics output on 2.6+ for 135partitions from that for disks. There are only *four* fields available 136for partitions on 2.6+ machines. This is reflected in the examples above. 137 138Field 1 -- # of reads issued 139 This is the total number of reads issued to this partition. 140 141Field 2 -- # of sectors read 142 This is the total number of sectors requested to be read from this 143 partition. 144 145Field 3 -- # of writes issued 146 This is the total number of writes issued to this partition. 147 148Field 4 -- # of sectors written 149 This is the total number of sectors requested to be written to 150 this partition. 151 152Note that since the address is translated to a disk-relative one, and no 153record of the partition-relative address is kept, the subsequent success 154or failure of the read cannot be attributed to the partition. In other 155words, the number of reads for partitions is counted slightly before time 156of queuing for partitions, and at completion for whole disks. This is 157a subtle distinction that is probably uninteresting for most cases. 158 159More significant is the error induced by counting the numbers of 160reads/writes before merges for partitions and after for disks. Since a 161typical workload usually contains a lot of successive and adjacent requests, 162the number of reads/writes issued can be several times higher than the 163number of reads/writes completed. 164 165In 2.6.25, the full statistic set is again available for partitions and 166disk and partition statistics are consistent again. Since we still don't 167keep record of the partition-relative address, an operation is attributed to 168the partition which contains the first sector of the request after the 169eventual merges. As requests can be merged across partition, this could lead 170to some (probably insignificant) inaccuracy. 171 172Additional notes 173---------------- 174 175In 2.6+, sysfs is not mounted by default. If your distribution of 176Linux hasn't added it already, here's the line you'll want to add to 177your ``/etc/fstab``:: 178 179 none /sys sysfs defaults 0 0 180 181 182In 2.6+, all disk statistics were removed from ``/proc/stat``. In 2.4, they 183appear in both ``/proc/partitions`` and ``/proc/stat``, although the ones in 184``/proc/stat`` take a very different format from those in ``/proc/partitions`` 185(see proc(5), if your system has it.) 186 187-- ricklind@us.ibm.com 188