Trace block i/o events as they occur. uses linux ftrace.
iosnoop [-hQst] [-d device] [-i iotype] [-p pid] [-n name] [duration]
iosnoop prints block device I/O events as they happen, with useful details such as PID, device, I/O type, block number, I/O size, and latency.
This traces disk I/O at the block device interface, using the block: tracepoints. This can help characterize the I/O requested for the storage devices and their resulting performance. I/O completions can also be studied event-by-event for debugging disk and controller I/O scheduling issues.
NOTE: Use of a duration buffers I/O, which reduces overheads, but this also introduces a limit to the number of I/O that will be captured. See the duration section in OPTIONS.
Since this uses ftrace, only the root user can use this tool.
FTRACE CONFIG, and the tracepoints block:block_rq_insert, block:block_rq_issue, and block:block_rq_complete, which you may already have enabled and available on recent Linux kernels. And awk.
Only show I/O issued by this device. (eg, "202,1"). This matches the DEV column in the iosnoop output, and is filtered in-kernel.
Only show I/O issued that matches this I/O type. This matches the TYPE column in the iosnoop output, and wildcards ("*") can be used at the beginning or end (only). Eg, "*R*" matches all reads. This is filtered in-kernel.
Only show I/O issued by this PID. This filters in-kernel. Note that I/O may be issued indirectly; for example, as the result of a memory allocation, causing dirty buffers (maybe from another PID) to be written to storage.
Only show I/O issued by processes with this name. Partial strings and regular expressions are allowed. This is a post-filter, so all I/O is traced and then filtered in user space. As with PID, this includes indirectly issued I/O.
Print usage message.
Use the queue insertion as the I/O start time (block:block_rq_insert), instead of the issue time (block:block_rq_issue). This affects the STARTs and LATms columns.
Include a column for the start time (issue time) of the I/O, in seconds. If the -Q option is used, this is the time the I/O is inserted on the block I/O queue.
Include a column for the completion time of the I/O, in seconds.
Set the duration of tracing, in seconds. Trace output will be buffered and printed at the end. This also reduces overheads by buffering in-kernel, instead of printing events as they occur.
The ftrace buffer has a fixed size per-CPU (see /sys/kernel/debug/tracing/buffer_size_kb). If you think events are missing, try increasing that size (the bufsize_kb setting in iosnoop). With the default setting (4 Mbytes), I'd expect this to happen around 50k I/O.
Default output, print I/O activity as it occurs:
Buffer for 5 seconds (lower overhead) and write to a file:
# iosnoop 5 > outfile
Trace reads only:
# iosnoop -i '*R*'
Trace I/O issued to device 202,1 only:
# iosnoop -d 202,1
Include I/O start and completion timestamps:
# iosnoop -ts
Include block I/O queueing time:
# iosnoop -Q Include I/O queueing and completion timestamps: # iosnop -Qts
Trace I/O issued when PID 181 was on-CPU only:
# iosnoop -p 181
Process name (command) for the PID that was on-CPU when the I/O was issued. See PID.
Process ID which was on-CPU when the I/O was issued. This will usually be the process directly requesting I/O, however, it may also include indirect I/O. For example, a memory allocation by this PID which causes dirty memory from another PID to be flushed to disk.
Type of I/O. R=read, W=write, M=metadata, S=sync, A=readahead, F=flush or FUA (force unit access), D=discard, E=secure, N=null (not RWFD).
Storage device ID.
Disk block for the operation (location, relative to this device).
Size of the I/O, in bytes.
Latency (time) for the I/O, in milliseconds.
By default, iosnoop works without buffering, printing I/O events as they happen (uses trace_pipe), context switching and consuming CPU to do so. This has a limit of about 10,000 IOPS (depending on your platform), at which point iosnoop will be consuming 1 CPU. The duration mode uses buffering, and can handle much higher IOPS rates, however, the buffer has a limit of about 50,000 I/O, after which events will be dropped. You can tune this with bufsize_kb, which is per-CPU. Also note that the "-n" option is currently post-filtered, so all events are traced.
The overhead may be acceptable in many situations. If it isn't, this tool can be reimplemented in C, or using a different tracer (eg, perf_events, SystemTap, ktap.)
This is from the perf-tools collection.
Also look under the examples directory for a text file containing example usage, output, and commentary for this tool.
Unstable - in development.