Обсуждение: BUG #15076: postmaster crashes unexpectedly when using up arrow keyin psql command

On Wednesday, February 21, 2018, PG Bug reporting form <noreply@postgresql.org> wrote:

The following bug has been logged on the website:

Bug reference:      15076
Logged by:          Mingchun Zhao
Email address:      mzhao@insight-tec.co.jp
PostgreSQL version: 9.3.20
Operating system:   Windows
Description:

steps to reproduce:

(1) run psql.exe in a Windows Command Prompt
(2) execute a SQL with more than one page of results, e.g. `show all;`
(3) press Ctrl + C to break it on the first page
(4) press up arrow key to try to execute the previous SQL of the step(2)
again
(5) Occasionally, it will show you nothing about the SQL, then you can try
to input a character ';' and press Enter key, the PostgreSQL will crash
unexpectedly

It would help to explain what you actually see in step 5 instead of just saying crash.  In particular is the "failure" in the server or client?

David J. 

On Thu, Feb 22, 2018 at 03:57:53PM +0900, 趙明春 wrote:
> You mean that the up-arrow key operation in psql.exe may send a SIGINT(or
> SIGTERM) signal to server(postmaster) unexpectedly?

No.  What Tom means here is that you are likely mistaking what you think
is a server crash by a client crash, and that the version of libreadline
your psql version is linked to is buggy and unable to handle that
properly.  psql has no control on the server itself.
--
Michael

趙明春 <mzhao@insight-tec.co.jp> writes:
> So, Let me change my question:
> If I started the PostgreSQL by pg_ctl command in the same Windows command
> prompt,
> Is it possible that pressing Ctrl + C on psql to break query(not session)
> will send SIGINT to postmaster process unexpectedly?

I don't know much about Windows signal processing, but the log entries you
show certainly look like that's what happened.

On Unix-ish machines it's normal for control-C to send a SIGINT to every
process that's in the terminal's process group --- *not* just the
foreground process.  So if you launch the postmaster manually you'd better
make sure that it gets dissociated from the terminal, or you're at risk
of killing it accidentally.  (Or, if you want to leave it connected to
that terminal, you use another window to run psql etc.)  It seems that
something similar is happening on your Windows setup.  I don't know
the rules for process/terminal association on Windows, so I can't suggest
what's the right way to do it.  I will say though that most people
run the postmaster as a background service on Windows, rather than
launching it manually.

On Thu, Feb 22, 2018 at 4:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:

趙明春 <mzhao@insight-tec.co.jp> writes:
> So, Let me change my question:
> If I started the PostgreSQL by pg_ctl command in the same Windows command
> prompt,
> Is it possible that pressing Ctrl + C on psql to break query(not session)
> will send SIGINT to postmaster process unexpectedly?

I don't know much about Windows signal processing, but the log entries you
show certainly look like that's what happened.

Yes, this is unfortunately exactly how it works on Windows.

There is on way to intercept the Ctrl-C and have it not delivered to *all* processes out of the same session. It can only be added or removed for the current process (which psql does). So it gets delivered to the backends which call pg_console_handler, which emulates SIGINT.

On Unix-ish machines it's normal for control-C to send a SIGINT to every
process that's in the terminal's process group --- *not* just the
foreground process.  So if you launch the postmaster manually you'd better
make sure that it gets dissociated from the terminal, or you're at risk
of killing it accidentally.  (Or, if you want to leave it connected to
that terminal, you use another window to run psql etc.)  It seems that
something similar is happening on your Windows setup.  I don't know
the rules for process/terminal association on Windows, so I can't suggest
what's the right way to do it.  I will say though that most people
run the postmaster as a background service on Windows, rather than
launching it manually.

AFAIC Windows has no way to disassociate a process with a console window in a similar way. 

--

 Magnus Hagander
 Me: https://www.hagander.net/
 Work: https://www.redpill-linpro.com/

Hi, Tom and Magnus,

Thanks for your information for this issue.

> AFAIC Windows has no way to disassociate a process with a console window in a similar way.

Can this be a solution for this issue to add an option CREATE_NEW_PROCESS_GROUP to CreateProcess() function into pg_ctl.exe to start postgres.exe process?

>  I will say though that most people run the postmaster as a background service on Windows, rather than launching it manually.

Unfortunately, we also met this issue, with poor reproducibility, when we running the postmaster as Windows Service.

Maybe this is another story than launching it manually by pg_ctl.exe.

I will update you if I get any new information.

On Fri, Feb 23, 2018 at 3:16 AM, 趙明春 <mzhao@insight-tec.co.jp> wrote:

Hi, Tom and Magnus,

Thanks for your information for this issue.

> AFAIC Windows has no way to disassociate a process with a console window in a similar way.

Can this be a solution for this issue to add an option CREATE_NEW_PROCESS_GROUP to CreateProcess() function into pg_ctl.exe to start postgres.exe process?

That definitely sounds like it should be worth a try. 

>  I will say though that most people run the postmaster as a background service on Windows, rather than launching it manually.

Unfortunately, we also met this issue, with poor reproducibility, when we running the postmaster as Windows Service.

Maybe this is another story than launching it manually by pg_ctl.exe.

I will update you if I get any new information.

That sounds very strange. I'm not saying you're not running into an issue there, but my guess would be you're actually running into a different problem (which may have similar symptoms).

//Magnus