Logging which interface was connected to in log_line_prefix

The following review has been posted through the commitfest application:
make installcheck-world: tested, passed
Implements feature: tested, passed
Spec compliant: tested, passed
Documentation: tested, passed

Hi

I did a quick test on this patch and it seems to work as expected. Originally I thought the patch would add the name of "local interface" such as "eth0", "eth1", "lo"... etc as %L log prefix format. Instead, it formats the local interface IP addresses , but I think it is fine too.

I have tested this new addition with various types of IPs including IPv4, IPv4 and IPv6 local loop back addresses, global IPv6 address, linked local IPv6 address with interface specifier, it seems to format these IPs correctly

There is a comment in the patch that states:

/* We do not need clean_ipv6_addr here: just report verbatim */

I am not quite sure what it means, but I am guessing it means that the patch does not need to format the IPv6 addresses in any specific way. For example, removing leading zeros or compressing consecutive zeros to make a IPv6 address shorter. It may not be necessary to indicate this in a comment because In my test, if any of my interface's IPv6 address have consecutive zeroes like this: 2000:0000:0000:0000:0000:0000:200:cafe/64, my network driver (Ubuntu 18.04) will format it as 2000::200:cafe, and the patch of course will read it as 2000::200:cafe, which is ... correct and clean.

thank you
Cary Huang
www.highgo.ca

Thank you for taking the time to review this. I've attached a new rebased
version, which has no significant changes.

There is a comment in the patch that states:

/* We do not need clean_ipv6_addr here: just report verbatim */

I am not quite sure what it means, but I am guessing it means that the
patch does not need to format the IPv6 addresses in any specific way.

Yes, basically correct. There is a kluge (their word, not mine) in
utils/adt/network.c to strip the zone - see the comment for the
clean_ipv6_addr() function in that file. I added the patch comment in case
some future person wonders why we don't "clean up" the ipv6 address, like
other places in the code base do. We don't need to pass it back to anything
else, so we can simply output the correct version, zone and all.

Cheers,
Greg

On 06.03.24 16:59, Greg Sabino Mullane wrote:

Someone on -general was asking about this, as they are listening on
multiple IPs and would like to know which exact one clients were
hitting. I took a quick look and we already have that information, so I
grabbed some stuff from inet_server_addr and added it as part of a "%L"
(for 'local interface'). Quick patch / POC attached.

I was confused by this patch title. This feature does not log the
interface (like "eth0" or "lo"), but the local address. Please adjust
the terminology.

I noticed that for Unix-domain socket connections, %r and %h write
"[local]". I think that should be done for this new placeholder as well.

On 01.05.24 19:04, Greg Sabino Mullane wrote:

Thank you for taking the time to review this. I've attached a new
rebased version, which has no significant changes.

There is a comment in the patch that states:

/* We do not need clean_ipv6_addr here: just report verbatim */

I am not quite sure what it means, but I am guessing it means that
the patch does not need to format the IPv6 addresses in any specific
way.

Yes, basically correct. There is a kluge (their word, not mine) in
utils/adt/network.c to strip the zone - see the comment for the
clean_ipv6_addr() function in that file. I added the patch comment in
case some future person wonders why we don't "clean up" the ipv6
address, like other places in the code base do. We don't need to pass it
back to anything else, so we can simply output the correct version, zone
and all.

clean_ipv6_addr() needs to be called before trying to convert a string
representation into inet/cidr types. This is not what is happening
here. So the comment is not applicable.

Peter, thank you for the feedback. Attached is a new patch with "address"
rather than "interface", plus a new default of "local" if there is no
address. I also removed the questionable comment, and updated the
commitfest title.

Cheers,
Greg

On 5/24/24 22:33, Greg Sabino Mullane wrote:

Peter, thank you for the feedback. Attached is a new patch with
"address" rather than "interface", plus a new default of "local" if
there is no address. I also removed the questionable comment, and
updated the commitfest title.

I tried the updated patch and it behaved as expected with [local] being
logged for peer connections and an IP being logged for host connections.

One thing -- the changes in postgresql.conf.sample should use tabs to
match the other lines. The patch uses spaces.

I also find the formatting in log_status_format() pretty awkward but I
guess that will be handled by pgindent.

Regards,
-David

Thanks for the review. Please find attached a new version with proper tabs
and indenting.

Cheers,
Greg

On 7/11/24 23:09, Greg Sabino Mullane wrote:

Thanks for the review. Please find attached a new version with proper
tabs and indenting.

This looks good to me now. +1 overall for the feature.

Regards,
-David

Hi Greg

On 11.07.24 18:09, Greg Sabino Mullane wrote:

Thanks for the review. Please find attached a new version with proper
tabs and indenting.

Cheers,
Greg

I'm testing this new log prefix and I'm wondering whether the following
behaviour is expected. The value of '%L' is different in the following
cases:

postgres=# SHOW log_line_prefix;
 log_line_prefix
-----------------
 %m [%p] -> %L
(1 row)

--

postgres=# SELECT 1/0;

2024-11-18 16:00:42.720 CET [3135117] -> 192.168.178.27 ERROR:  division
by zero
2024-11-18 16:00:42.720 CET [3135117] -> 192.168.178.27 STATEMENT: 
SELECT 1/0;

--

postgres=# SELECT pg_reload_conf();

2024-11-18 16:01:23.273 CET [3114980] -> [local] LOG:  received SIGHUP,
reloading configuration files

--

postgres=# CHECKPOINT;

2024-11-18 16:01:46.758 CET [3114981] -> [local] LOG:  checkpoint
starting: immediate force wait
2024-11-18 16:01:46.769 CET [3114981] -> [local] LOG:  checkpoint
complete: wrote 0 buffers (0.0%), wrote 0 SLRU buffers; 0 WAL file(s)
added, 0 removed, 0 recycled; write=0.001 s, sync=0.001 s, total=0.012
s; sync files=0, longest=0.000 s, average=0.000 s; distance=0 kB,
estimate=25924 kB; lsn=0/26166430, redo lsn=0/261663D8

Is it supposed to be like this?

Thanks for the patch!

--
Jim

On Mon, Nov 18, 2024 at 10:07 AM Jim Jones <jim.jones@uni-muenster.de>
wrote:

2024-11-18 16:00:42.720 CET [3135117] -> 192.168.178.27 STATEMENT:
...
2024-11-18 16:01:23.273 CET [3114980] -> [local] LOG: received SIGHUP,
...

2024-11-18 16:01:46.769 CET [3114981] -> [local] LOG: checkpoint

Is it supposed to be like this?

Great question. I think "supposed to" is a bit of a stretch, but I presume
it's the difference between a client connecting and using its connection
information versus an already existing backend process, which is always
going to be "local".

Overall this makes sense, as that checkpoint example above is coming from
the checkpointer background process at 3114981, not the backend process
that happened to trigger it. And 3114981 has no way of knowing the details
of the caller's connection.

FWIW, the patch still applies cleanly to head as of 2/27/2025, so no rebase
needed.

Cheers,
Greg

--
Crunchy Data - https://www.crunchydata.com
Enterprise Postgres Software Products & Tech Support

On 27.02.25 14:54, Greg Sabino Mullane wrote:

Great question. I think "supposed to" is a bit of a stretch, but I
presume it's the difference between a client connecting and using its
connection information versus an already existing backend process,
which is always going to be "local".

Overall this makes sense, as that checkpoint example above is coming
from the checkpointer background process at 3114981, not the backend
process that happened to trigger it. And 3114981 has no way of knowing
the details of the caller's connection.

In that case, it LGTM.

I revisited this patch and tested it with two different computers (for
client and server).

Initially, I was momentarily confused by the logged address format,
which varies depending on the client's format. However, I found that %h
behaves just like this, so I guess it is ok.

postgres=# SHOW log_line_prefix;
    log_line_prefix    
-----------------------
 %m [%p]:  L=%L, h=%h
(1 row)

2025-03-02 18:19:07.859 CET [2246150]:  L=192.168.178.27,
h=192.168.178.79 ERROR:  division by zero
2025-03-02 18:19:07.859 CET [2246150]:  L=192.168.178.27,
h=192.168.178.79 STATEMENT:  SELECT 1/0

2025-03-02 18:19:19.327 CET [2246291]:  L=2a02:...:7591, h=2a02:...:4a7
ERROR:  division by zero
2025-03-02 18:19:19.327 CET [2246291]:  L=2a02:...:7591, h=2a02:...:4a7
STATEMENT:  SELECT 1/0

Best, Jim

Jim Jones <jim.jones@uni-muenster.de> writes:

In that case, it LGTM.

I looked at 0002 briefly. I don't have any particular objection to
the proposed feature, but I'm quite concerned about the potential
performance implications of doing a new pg_getnameinfo_all() call
for every line of log output. I think that needs to be cached
somehow. It's a little tricky since we may pass through this function
one or more times before the socket info has been filled in, but it
seems do-able.

"Local address" doesn't convey a lot to my mind, and Jim's confusion
about what "[local]" means seems tied to that. Can we think of a
different explanation for the docs? Also, in %h we use "[local]" to
signify a Unix socket, but this patch prints that for both the
Unix-socket case and the case of no client connection at all.
I think "[none]" or some such would be a better idea than "[local]"
for background processes.

The patch pays no attention to the "padding" feature that
all the other switch cases honor.

Also, the coding style is randomly unlike project style in various
details, notably paren placement and the use of "0 == foo" to mean
"!foo".

regards, tom lane

Oh, one other thing: pg_getnameinfo_all is perfectly capable of
dealing with a Unix-socket address, so I think you should drop
the tests on ss_family and let pg_getnameinfo_all be in charge
of producing "[local]" for that case.

regards, tom lane

Thanks for all the feedback. Please find attached a version which prints
"[none]" as the default value, "[local]" for a socket, and otherwise
whatever pg_getnameinfo_all spits out. I cleaned up the coding, respected
padding, removed the family checks, and expanded the docs a tiny bit to
give the reader more context as to what "local address" means. I also
looked into alternatives to the term "local address" but that still seems
the most correct and commonly used term.

I have not attempted the caching change yet.

Cheers,
Greg

--
Crunchy Data - https://www.crunchydata.com
Enterprise Postgres Software Products & Tech Support

Greg Sabino Mullane <htamfids@gmail.com> writes:

I have not attempted the caching change yet.

After some thought I concluded that caching the local-address string
in MyProcPort itself would be the most robust way of making that work.
Otherwise you need some way to update the cache when MyProcPort is
created (in case the process already emitted some messages before
that), and the patch starts to spread into other places.

I think 0004 attached is about committable, but there is one
definitional point that is troubling me slightly: our choice to
emit "[none]" when there's no port isn't consistent with the
log_line_prefix documentation's statement that

... Some escapes are only recognized by session processes,
and will be treated as empty by background processes such as
the main server process.

Since we've marked %L as "Session only" = yes, this implies
that it should print as an empty string not "[none]".
We could either

1. Ignore the inconsistency, commit 0004 as-is.

2. Change the output to be an empty string in background processes.
This is consistent, but it goes against our upthread feeling
that "[none]" would avoid confusion.

3. Mark %L as "Session only" = no. This seems a little weird,
but it'd also be consistent.

4. Add something to the above-quoted text about %L being an exception.

I don't really care for #3 or #4, but I'm ambivalent between #1 and
#2. I think the worry about confusion originated when the patch
would print "[local]" for either a Unix socket or a background
process, and that certainly was confusing. "[local]" versus
an empty string is not so ambiguous, so maybe it's fine.

Thoughts?

regards, tom lane

On Sun, Apr 06, 2025 at 06:01:01PM -0400, Tom Lane wrote:

I don't really care for #3 or #4, but I'm ambivalent between #1 and
#2. I think the worry about confusion originated when the patch
would print "[local]" for either a Unix socket or a background
process, and that certainly was confusing. "[local]" versus
an empty string is not so ambiguous, so maybe it's fine.

I'd suggest the addition of this data to csvlog.c and jsonlog.c,
perhaps only adding this information if local_host[0] is not '\0'
rather than assigning a default "[none]" all the time to save some
space in the entries generated.

config.sgml would also need a refresh:
- runtime-config-logging-jsonlog-keys-values for the new key/value
pair.
- runtime-config-logging-csvlog for the CREATE TABLE and the list of
columns.

Perhaps it would be time to switch the list of columns to use a proper
table instead of a raw list for the CSV docs, for clarity.
--
Michael

Michael Paquier <michael@paquier.xyz> writes:

I'd suggest the addition of this data to csvlog.c and jsonlog.c,
perhaps only adding this information if local_host[0] is not '\0'
rather than assigning a default "[none]" all the time to save some
space in the entries generated.

I think that's completely impractical for csvlog: changing the row
type for CSV output is a serious compatibility break, and there is
exactly zero evidence that the local address info is worth that.

JSON has less of a compatibility problem, but I still doubt that
the local address info is worth the space it'd take, for about
99% of users. Also note that we have no good way for the user
to specify what fields she wants in jsonlog, otherwise we could
make it appear only if asked for.

regards, tom lane

On Sun, Apr 6, 2025 at 6:01 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:

1. Ignore the inconsistency, commit 0004 as-is.

2. Change the output to be an empty string in background processes.
This is consistent, but it goes against our upthread feeling that
"[none]" would avoid confusion.

I lean for #1. Yes, there is some inconsistency, but it feels like the
right thing to do, and this is a feature I suspect not many people will use
anyway.

--
Cheers,
Greg

--
Crunchy Data - https://www.crunchydata.com
Enterprise Postgres Software Products & Tech Support

Greg Sabino Mullane <htamfids@gmail.com> writes:

On Sun, Apr 6, 2025 at 6:01 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:

1. Ignore the inconsistency, commit 0004 as-is.

2. Change the output to be an empty string in background processes.
This is consistent, but it goes against our upthread feeling that
"[none]" would avoid confusion.

I lean for #1. Yes, there is some inconsistency, but it feels like the
right thing to do, and this is a feature I suspect not many people will use
anyway.

Hearing no other comments, I pushed 0004 as-is. It's trivial enough
to replace "[none]" with "" if somebody makes the case for that.

regards, tom lane