Enhanced partition statistics: documentation update
Jerome Marchand [Fri, 8 Feb 2008 10:10:56 +0000 (11:10 +0100)]
Update the documentation to reflect the change in userspace interface.

Signed-off-by: Jerome Marchand <jmarchan@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

Documentation/ABI/testing/procfs-diskstats [new file with mode: 0644]
Documentation/ABI/testing/sysfs-block [new file with mode: 0644]
Documentation/iostats.txt

diff --git a/Documentation/ABI/testing/procfs-diskstats b/Documentation/ABI/testing/procfs-diskstats
new file mode 100644 (file)
index 0000000..9923390
--- /dev/null
@@ -0,0 +1,22 @@
+What:          /proc/diskstats
+Date:          February 2008
+Contact:       Jerome Marchand <jmarchan@redhat.com>
+Description:
+               The /proc/diskstats file displays the I/O statistics
+               of block devices. Each line contains the following 14
+               fields:
+                1 - major number
+                2 - minor mumber
+                3 - device name
+                4 - reads completed succesfully
+                5 - reads merged
+                6 - sectors read
+                7 - time spent reading (ms)
+                8 - writes completed
+                9 - writes merged
+               10 - sectors written
+               11 - time spent writing (ms)
+               12 - I/Os currently in progress
+               13 - time spent doing I/Os (ms)
+               14 - weighted time spent doing I/Os (ms)
+               For more details refer to Documentation/iostats.txt
diff --git a/Documentation/ABI/testing/sysfs-block b/Documentation/ABI/testing/sysfs-block
new file mode 100644 (file)
index 0000000..4bd9ea5
--- /dev/null
@@ -0,0 +1,28 @@
+What:          /sys/block/<disk>/stat
+Date:          February 2008
+Contact:       Jerome Marchand <jmarchan@redhat.com>
+Description:
+               The /sys/block/<disk>/stat files displays the I/O
+               statistics of disk <disk>. They contain 11 fields:
+                1 - reads completed succesfully
+                2 - reads merged
+                3 - sectors read
+                4 - time spent reading (ms)
+                5 - writes completed
+                6 - writes merged
+                7 - sectors written
+                8 - time spent writing (ms)
+                9 - I/Os currently in progress
+               10 - time spent doing I/Os (ms)
+               11 - weighted time spent doing I/Os (ms)
+               For more details refer Documentation/iostats.txt
+
+
+What:          /sys/block/<disk>/<part>/stat
+Date:          February 2008
+Contact:       Jerome Marchand <jmarchan@redhat.com>
+Description:
+               The /sys/block/<disk>/<part>/stat files display the
+               I/O statistics of partition <part>. The format is the
+               same as the above-written /sys/block/<disk>/stat
+               format.
index b963c3b..5925c3c 100644 (file)
@@ -58,7 +58,7 @@ they should not wrap twice before you notice them.
 Each set of stats only applies to the indicated device; if you want
 system-wide stats you'll have to find all the devices and sum them all up.
 
-Field  1 -- # of reads issued
+Field  1 -- # of reads completed
     This is the total number of reads completed successfully.
 Field  2 -- # of reads merged, field 6 -- # of writes merged
     Reads and writes which are adjacent to each other may be merged for
@@ -132,6 +132,19 @@ words, the number of reads for partitions is counted slightly before time
 of queuing for partitions, and at completion for whole disks.  This is
 a subtle distinction that is probably uninteresting for most cases.
 
+More significant is the error induced by counting the numbers of
+reads/writes before merges for partitions and after for disks. Since a
+typical workload usually contains a lot of successive and adjacent requests,
+the number of reads/writes issued can be several times higher than the
+number of reads/writes completed.
+
+In 2.6.25, the full statistic set is again available for partitions and
+disk and partition statistics are consistent again. Since we still don't
+keep record of the partition-relative address, an operation is attributed to
+the partition which contains the first sector of the request after the
+eventual merges. As requests can be merged across partition, this could lead
+to some (probably insignificant) innacuracy.
+
 Additional notes
 ----------------