1=====================
2I/O statistics fields
3=====================
4
5Since 2.4.20 (and some versions before, with patches), and 2.5.45,
6more extensive disk statistics have been introduced to help measure disk
7activity. Tools such as ``sar`` and ``iostat`` typically interpret these and do
8the work for you, but in case you are interested in creating your own
9tools, the fields are explained here.
10
11In 2.4 now, the information is found as additional fields in
12``/proc/partitions``.  In 2.6 and upper, the same information is found in two
13places: one is in the file ``/proc/diskstats``, and the other is within
14the sysfs file system, which must be mounted in order to obtain
15the information. Throughout this document we'll assume that sysfs
16is mounted on ``/sys``, although of course it may be mounted anywhere.
17Both ``/proc/diskstats`` and sysfs use the same source for the information
18and so should not differ.
19
20Here are examples of these different formats::
21
22   2.4:
23      3     0   39082680 hda 446216 784926 9550688 4382310 424847 312726 5922052 19310380 0 3376340 23705160
24      3     1    9221278 hda1 35486 0 35496 38030 0 0 0 0 0 38030 38030
25
26   2.6+ sysfs:
27      446216 784926 9550688 4382310 424847 312726 5922052 19310380 0 3376340 23705160
28      35486    38030    38030    38030
29
30   2.6+ diskstats:
31      3    0   hda 446216 784926 9550688 4382310 424847 312726 5922052 19310380 0 3376340 23705160
32      3    1   hda1 35486 38030 38030 38030
33
34   4.18+ diskstats:
35      3    0   hda 446216 784926 9550688 4382310 424847 312726 5922052 19310380 0 3376340 23705160 0 0 0 0
36
37On 2.4 you might execute ``grep 'hda ' /proc/partitions``. On 2.6+, you have
38a choice of ``cat /sys/block/hda/stat`` or ``grep 'hda ' /proc/diskstats``.
39
40The advantage of one over the other is that the sysfs choice works well
41if you are watching a known, small set of disks.  ``/proc/diskstats`` may
42be a better choice if you are watching a large number of disks because
43you'll avoid the overhead of 50, 100, or 500 or more opens/closes with
44each snapshot of your disk statistics.
45
46In 2.4, the statistics fields are those after the device name. In
47the above example, the first field of statistics would be 446216.
48By contrast, in 2.6+ if you look at ``/sys/block/hda/stat``, you'll
49find just the eleven fields, beginning with 446216.  If you look at
50``/proc/diskstats``, the eleven fields will be preceded by the major and
51minor device numbers, and device name.  Each of these formats provides
52eleven fields of statistics, each meaning exactly the same things.
53All fields except field 9 are cumulative since boot.  Field 9 should
54go to zero as I/Os complete; all others only increase (unless they
55overflow and wrap).  Yes, these are (32-bit or 64-bit) unsigned long
56(native word size) numbers, and on a very busy or long-lived system they
57may wrap. Applications should be prepared to deal with that; unless
58your observations are measured in large numbers of minutes or hours,
59they should not wrap twice before you notice them.
60
61Each set of stats only applies to the indicated device; if you want
62system-wide stats you'll have to find all the devices and sum them all up.
63
64Field  1 -- # of reads completed
65    This is the total number of reads completed successfully.
66
67Field  2 -- # of reads merged, field 6 -- # of writes merged
68    Reads and writes which are adjacent to each other may be merged for
69    efficiency.  Thus two 4K reads may become one 8K read before it is
70    ultimately handed to the disk, and so it will be counted (and queued)
71    as only one I/O.  This field lets you know how often this was done.
72
73Field  3 -- # of sectors read
74    This is the total number of sectors read successfully.
75
76Field  4 -- # of milliseconds spent reading
77    This is the total number of milliseconds spent by all reads (as
78    measured from __make_request() to end_that_request_last()).
79
80Field  5 -- # of writes completed
81    This is the total number of writes completed successfully.
82
83Field  6 -- # of writes merged
84    See the description of field 2.
85
86Field  7 -- # of sectors written
87    This is the total number of sectors written successfully.
88
89Field  8 -- # of milliseconds spent writing
90    This is the total number of milliseconds spent by all writes (as
91    measured from __make_request() to end_that_request_last()).
92
93Field  9 -- # of I/Os currently in progress
94    The only field that should go to zero. Incremented as requests are
95    given to appropriate struct request_queue and decremented as they finish.
96
97Field 10 -- # of milliseconds spent doing I/Os
98    This field increases so long as field 9 is nonzero.
99
100    Since 5.0 this field counts jiffies when at least one request was
101    started or completed. If request runs more than 2 jiffies then some
102    I/O time will not be accounted unless there are other requests.
103
104Field 11 -- weighted # of milliseconds spent doing I/Os
105    This field is incremented at each I/O start, I/O completion, I/O
106    merge, or read of these stats by the number of I/Os in progress
107    (field 9) times the number of milliseconds spent doing I/O since the
108    last update of this field.  This can provide an easy measure of both
109    I/O completion time and the backlog that may be accumulating.
110
111Field 12 -- # of discards completed
112    This is the total number of discards completed successfully.
113
114Field 13 -- # of discards merged
115    See the description of field 2
116
117Field 14 -- # of sectors discarded
118    This is the total number of sectors discarded successfully.
119
120Field 15 -- # of milliseconds spent discarding
121    This is the total number of milliseconds spent by all discards (as
122    measured from __make_request() to end_that_request_last()).
123
124To avoid introducing performance bottlenecks, no locks are held while
125modifying these counters.  This implies that minor inaccuracies may be
126introduced when changes collide, so (for instance) adding up all the
127read I/Os issued per partition should equal those made to the disks ...
128but due to the lack of locking it may only be very close.
129
130In 2.6+, there are counters for each CPU, which make the lack of locking
131almost a non-issue.  When the statistics are read, the per-CPU counters
132are summed (possibly overflowing the unsigned long variable they are
133summed to) and the result given to the user.  There is no convenient
134user interface for accessing the per-CPU counters themselves.
135
136Disks vs Partitions
137-------------------
138
139There were significant changes between 2.4 and 2.6+ in the I/O subsystem.
140As a result, some statistic information disappeared. The translation from
141a disk address relative to a partition to the disk address relative to
142the host disk happens much earlier.  All merges and timings now happen
143at the disk level rather than at both the disk and partition level as
144in 2.4.  Consequently, you'll see a different statistics output on 2.6+ for
145partitions from that for disks.  There are only *four* fields available
146for partitions on 2.6+ machines.  This is reflected in the examples above.
147
148Field  1 -- # of reads issued
149    This is the total number of reads issued to this partition.
150
151Field  2 -- # of sectors read
152    This is the total number of sectors requested to be read from this
153    partition.
154
155Field  3 -- # of writes issued
156    This is the total number of writes issued to this partition.
157
158Field  4 -- # of sectors written
159    This is the total number of sectors requested to be written to
160    this partition.
161
162Note that since the address is translated to a disk-relative one, and no
163record of the partition-relative address is kept, the subsequent success
164or failure of the read cannot be attributed to the partition.  In other
165words, the number of reads for partitions is counted slightly before time
166of queuing for partitions, and at completion for whole disks.  This is
167a subtle distinction that is probably uninteresting for most cases.
168
169More significant is the error induced by counting the numbers of
170reads/writes before merges for partitions and after for disks. Since a
171typical workload usually contains a lot of successive and adjacent requests,
172the number of reads/writes issued can be several times higher than the
173number of reads/writes completed.
174
175In 2.6.25, the full statistic set is again available for partitions and
176disk and partition statistics are consistent again. Since we still don't
177keep record of the partition-relative address, an operation is attributed to
178the partition which contains the first sector of the request after the
179eventual merges. As requests can be merged across partition, this could lead
180to some (probably insignificant) inaccuracy.
181
182Additional notes
183----------------
184
185In 2.6+, sysfs is not mounted by default.  If your distribution of
186Linux hasn't added it already, here's the line you'll want to add to
187your ``/etc/fstab``::
188
189	none /sys sysfs defaults 0 0
190
191
192In 2.6+, all disk statistics were removed from ``/proc/stat``.  In 2.4, they
193appear in both ``/proc/partitions`` and ``/proc/stat``, although the ones in
194``/proc/stat`` take a very different format from those in ``/proc/partitions``
195(see proc(5), if your system has it.)
196
197-- ricklind@us.ibm.com
198