/proc/self/io · Linux Labs

Lesson

0 of 12

Where the kernel counts

1 cell ≈ 1/32 of the file

process 4242 user space

waiting for your first command

system call boundary rcharwcharsyscrsyscw

/dev/null

Takes the bytes and drops them. Nothing below it.

pipe

A 64 KiB buffer in RAM between two fds.

page cache 4 KiB pages in RAM

write_bytes · a page turns dirty cancelled_write_bytes · a dirty page is thrown away

block layer read_bytes · a read had to go to the disk

disk ext4 on a block device

cached, clean cached, dirty (not on disk yet) on disk not cached

pid 4242 · every command runs in this one process ·

iolab[4242]$

This is the live file for pid 4242. The chips show what the last command changed. Reading this panel is free. A cat of the real file is not, as step 3 shows.

Counted in fs/proc/base.c, do_io_accounting()

Field	Counted when
`rchar`	A read-family syscall returns bytes, from any fd: file, pipe, socket, tty, `/dev/zero`.
`wchar`	A write-family syscall accepts bytes, to any fd.
`syscr` / `syscw`	Each read or write call, whatever its size. A read at end of file still counts.
`read_bytes`	The block layer submits a read for this process because the page cache didn't have the data.
`write_bytes`	This process makes a clean or new page dirty. Writing it to disk later costs nothing extra.
`cancelled_write_bytes`	This process truncates or deletes a file while its pages are still dirty.

The counts cover the whole thread group, plus children the process has reaped with wait(). The kernel needs CONFIG_TASK_IO_ACCOUNTING=y, otherwise the file doesn't exist.

Kernel 6.8.0 · ext4 · iolab.c

What the process did	What moved
64 KiB to `/dev/null`, 1 byte per write	wchar +65536 · syscw +65536
Same 64 KiB in one write	wchar +65536 · syscw +1
11 bytes through a pipe	rchar +11 · wchar +11
Write an 8 MiB file, no fsync	wchar +8388608 · syscw +128 write_bytes +8388608
Then `fsync()` it	nothing
Read it back from the cache	rchar +8388608 · syscr +129 read_bytes 0
Evict it, read it again	rchar +8388608 · read_bytes +8388608
Write 4 MiB to a new file, delete it	write_bytes +4194304 cancelled_write_bytes +4194304
Rewrite an existing file, then delete it	cancelled_write_bytes 0 (ext4 flushed it on close)

Each snapshot also reads the proc file, so every row has about 125 extra bytes of rchar and one extra syscr. They're left out of the table. You can run this yourself with linux-labs/01-proc-self-io/run.sh.