fileslower [-h] [-p PID] [-a] [min_ms]
This script uses kernel dynamic tracing of synchronous reads and writes at the VFS interface, to identify slow file reads and writes for any file system.
This version traces __vfs_read() and __vfs_write() and only showing synchronous I/O (the path to new_sync_read() and new_sync_write()), and I/O with filenames. This approach provides a view of just two file system request types: file reads and writes. There are typically many others: asynchronous I/O, directory operations, file handle operations, file open()s, fflush(), etc.
WARNING: See the Overhead section.
By default, a minimum millisecond threshold of 10 is used.
Since this works by tracing various kernel __vfs_*() functions using dynamic tracing, it will need updating to match any changes to these functions. A future version should switch to using FS tracepoints instead.
Since this uses BPF, only the root user can use this tool.
CONFIG_BPF and bcc.
-p PID Trace this PID only.
Include non-regular file types in output (sockets, FIFOs, etc).
Minimum I/O latency (duration) to trace, in milliseconds. Default is 10 ms.
- Trace synchronous file reads and writes slower than 10 ms:
- Trace slower than 1 ms:
# fileslower 1
- Trace slower than 1 ms, for PID 181 only:
# fileslower -p 181 1
Time of I/O completion since the first I/O seen, in seconds.
Direction of I/O. R == read, W == write.
Size of I/O, in bytes.
Latency (duration) of I/O, measured from when the application issued it to VFS to when it completed. This time is inclusive of block device I/O, file system CPU cycles, file system locks, run queue latency, etc. It's a more accurate measure of the latency suffered by applications performing file system I/O, than to measure this down at the block device interface.
A cached kernel file name (comes from dentry->d_name.name).
Depending on the frequency of application reads and writes, overhead can become severe, in the worst case slowing applications by 2x. In the best case, the overhead is negligible. Hopefully for real world workloads the overhead is often at the lower end of the spectrum -- test before use. The reason for high overhead is that this traces VFS reads and writes, which includes FS cache reads and writes, and can exceed one million events per second if the application is I/O heavy. While the instrumentation is extremely lightweight, and uses in-kernel eBPF maps for efficient timing and filtering, multiply that cost by one million events per second and that cost becomes a million times worse. You can get an idea of the possible cost by just counting the instrumented events using the bcc funccount tool, eg:
# ./funccount.py -i 1 -r '^__vfs_(read|write)$'
This also costs overhead, but is somewhat less than fileslower.
If the overhead is prohibitive for your workload, I'd recommend moving down-stack a little from VFS into the file system functions (ext4, xfs, etc). Look for updates to bcc for specific file system tools that do this. The advantage of a per-file system approach is that we can trace post-cache, greatly reducing events and overhead. The disadvantage is needing custom tracing approaches for each different file system (whereas VFS is generic).
This is from bcc.
Also look in the bcc distribution for a companion _examples.txt file containing example usage, output, and commentary for this tool.
Unstable - in development.