This page copies verbatim the contents of
Documentation/filesystems/proc.rst, added wholesale in
commit f9c99463b0cd05603d125c915e2886d55a686b82 ("[PATCH] Documentation
for io-accounting / reporting via procfs") in 2007.
As such, it mirrors the sensibilities of the time ‒
writing "successful read returns" as "data pulled from storage. actually
just the data the process gave to read(). this also means from non-regular
files! whether the data was pulled from storage doesn't matter actually
(obligatory cache mention)"
for the modern reader this is just a lot of waffling
(note also that processes give no data to read()!)
‒ and sensibilities of the sheepish implementer in kernel documentation ‒
"an attempt" for a well-defined kernel behaviour, mentioning the
"current implementation", consistent mentions of specific kernel-internal
caching mechanisms, "the big inaccuracy here".
Re-write to be more useful and less misleading as documentation;
the syscall enumeration is accurate for kernel v6.8, but the sysc? stats
are also bumped by kernel_{read,write}(), which is sometimes used by too
many syscalls in too many scenarios to usefully enumerate.
Signed-off-by: Ahelenia Ziemiańska <nabijaczleweli@xxxxxxxxxxxxxxxxxx>
---
Hi!
On Sun, Mar 17, 2024 at 01:15:18PM +0100, Alejandro Colomar wrote:
> On Sun, Mar 17, 2024 at 12:01:41PM +0100, наб wrote:
> > -The number of bytes
> > -which this task and its waited-for children
> > -have caused to be read from storage.
> > -This is simply the sum of bytes which this process passed to
> > +The number of bytes returned by successful
> In this case, I think I prefer to break before "returned".
> What would you do?
this is a "meh" moment imo; in running text sure, maybe,
but this is broken up by the .BR
so it starts to devolve into 3-word-line territory which is worse
> > -Attempt to count the number of read I/O operations\[em]that is,
> > -system calls such as
> > +The number of "I/O read" system calls\[em]those from the
> From I/O, read only accounts for the I. :)
> Should we just say "read"?
Yeah, "I/O" is overloaded here; "file read" works better.
> > .BR read (2)
> > +family
> > +(including when invoked used by the kernel as part of other syscalls),
> This parenthesis being there seems to imply that if the kernel calls
> sendfile internally for $reasons (even if it doesn't at the moment), it
> won't be counted. I think it makes more sense at the end of the list,
> no?
Well, as-is it doesn't, and I reduced this to the narrowest definition
I can prove, but I guess so, yes.
Also just noticed "invoked used".
man5/proc_pid_io.5 | 67 +++++++++++++++++++++-------------------------
1 file changed, 30 insertions(+), 37 deletions(-)
diff --git a/man5/proc_pid_io.5 b/man5/proc_pid_io.5
index dc75a91de..7f840f3bb 100644
--- a/man5/proc_pid_io.5
+++ b/man5/proc_pid_io.5
@@ -33,63 +33,56 @@ .SH DESCRIPTION
.TP
.IR rchar ": characters read"
The number of bytes
-which this task and its waited-for children
-have caused to be read from storage.
-This is simply the sum of bytes which this process passed to
+returned by successful
.BR read (2)
and similar system calls.
-It includes things such as terminal I/O and
-is unaffected by whether or not actual
-physical disk I/O was required (the read might have been satisfied from
-pagecache).
.TP
.IR wchar ": characters written"
The number of bytes
-which this task and its waited-for children
-have caused, or shall cause to be written to disk.
-Similar caveats apply here as with
-.IR rchar .
+returned by successful
+.BR write (2)
+and similar system calls.
.TP
.IR syscr ": read syscalls"
-Attempt to count the number of read I/O operations\[em]that is,
-system calls such as
+The number of "file read" system calls\[em]those from the
.BR read (2)
+family
+.BR sendfile (2),
+.BR copy_file_range (2),
and
-.BR pread (2).
+.BR ioctl (2)
+.BR BTRFS_IOC_ENCODED_READ [ _32 ]
+(including when invoked by the kernel as part of other syscalls),
.TP
.IR syscw ": write syscalls"
-Attempt to count the number of write I/O operations\[em]that is,
-system calls such as
+The number of "file write" system calls\[em]those from the
.BR write (2)
+family
+.BR sendfile (2),
+.BR copy_file_range (2),
and
-.BR pwrite (2).
+.BR ioctl (2)
+.BR BTRFS_IOC_ENCODED_WRITE [ _32 ]
+(including when invoked by the kernel as part of other syscalls),
.TP
.IR read_bytes ": bytes read"
-Attempt to count the number of bytes
-which this process and its waited-for children
-really did cause to be fetched from the storage layer.
+The number of bytes really fetched from the storage layer.
This is accurate for block-backed filesystems.
.TP
.IR write_bytes ": bytes written"
-Attempt to count the number of bytes
-which this process and its waited-for children
-caused to be sent to the storage layer.
+The number of bytes really sent to the storage layer.
.TP
.IR cancelled_write_bytes :
-The big inaccuracy here is truncate.
-If a process writes 1 MB to a file and then deletes the file,
-it will in fact perform no writeout.
-But it will have been accounted as having caused 1 MB of write.
-In other words:
-this field represents the number of bytes
-which this process and its waited-for children
-caused to not happen, by truncating pagecache.
-A task can cause "negative" I/O too.
-If this task truncates some dirty pagecache,
-some I/O which another task has been accounted for
-(in its
-.IR write_bytes )
-will not be happening.
+The above statistics fail to account for truncation:
+if a process writes 1 MB to a regular file and then removes it,
+said 1 MB will not be written, but
+.I will
+have nevertheless been accounted as a 1 MB write.
+This field represents the number of bytes "saved" from I/O writeback.
+This can yield to having done negative I/O
+if caches dirtied by another process are truncated.
+This figure applies to I/O already accounted-for by
+.IR write_bytes .
.RE
.IP
.IR Note :
--
2.39.2
Attachment:
signature.asc
Description: PGP signature
