patchreduce progress updates in background

Hi,

I've noticed that when a long-running git operation that generates progress output is suspended and converted to a background process, the terminal still gets spammed with progress updates (to stderr).

Many years ago I fixed a similar issue in the NetBSD ftp progress bar code (which I wrote).

I've experimented around with a couple of different solutions, including:
1. suppress all progress output whilst in the background
2. suppress "in progress" updates whilst in the background,
   but display the "done" message even if in the background.

In both cases, warnings were still output to the terminal.

I've attached a patch that implements (2) above.

If the consensus is that all progress messages should be suppressed, I can provide the (simpler) patch for that.

I've explicitly separated the in_progress_fd() function so that it's easier to (a) reuse elsewhere where appropriate, and (b) make any portability changes to the test if necessary. I also used getpgid(0) versus getpgrp() to avoid portability issues with the signature in the latter with pre-POSIX.

A minor optimisation could be to pass in struct progress * and to cache getpgid(0) in a member of struct progress in start_progress_delay(), since this value shouldn't change during the life of the process.

regards, Luke.

From 843a367bac87674666dafbaf7fdb7d6b0e1660f7 Mon Sep 17 00:00:00 2001
From: Luke Mewburn <luke@mewburn.net>
Date: Mon, 13 Apr 2015 23:30:51 +1000
Subject: [PATCH] progress: no progress in background

Disable the display of the progress if stderr is not the current foreground process. Still display the final result when done.

Signed-off-by: Luke Mewburn <luke@mewburn.net>
---
 progress.c | 23 +++++++++++++++++------
 1 file changed, 17 insertions(+), 6 deletions(-)

On Mon, 13 Apr 2015, Luke Mewburn wrote:

What if you suspend the task and push it into the background? Would be nice to inhibit progress display in that case, and resume it if the task returns to the foreground.

Also the display() function may be called quite a lot without necessarily resulting in a display output. Therefore I'd suggest adding in_progress_fd() to the if condition right before the printf() instead.

Nicolas

On Mon, Apr 13, 2015 at 10:11:09AM -0400, Nicolas Pitre wrote:
  | What if you suspend the task and push it into the background? Would be 
  | nice to inhibit progress display in that case, and resume it if the task 
  | returns to the foreground.

That's what happens; the suppression only occurs if the process is currently background. If I start a long-running operation (such as "git fsck"), the progress is displayed. I then suspend & background, and the progress is suppressed. If I resume the process in the foreground, the progress starts to display again at the appropriate point.

In the proposed patch, the stop_progress display for a given progress
(i.e. the one that ends in ", done.") is displayed even if in the
background so that there's some indication of progress. E.g.
  Checking object directories: 100% (256/256), done.
  Checking objects: 100% (184664/184664), done.
  Checking connectivity: 184667, done.
This is the test 'if (is_foreground || done)'.

I'm not 100% happy with my choice here, and the simpler behaviour of "suppress all background progress output" can be achieved by removing '|| done' from those two tests.

That still doesn't suppress _all_ output whilst in the background. In order to do that, a larger refactor of various warning methods would be required. I would argue that's a separate orthoganal fix.

  | Also the display() function may be called quite a lot without 
  | necessarily resulting in a display output. Therefore I'd suggest adding 
  | in_progress_fd() to the if condition right before the printf() instead.

That's an easy enough change to make (although I speculate that the testing of the foreground status is not that big a performance issue, especially compared the the existing performance "overhead" of printing the progress to stderr then forcing a flush :)

Should I submit a revised patch with (1) call in_progress_fd() just before the fprintf() as requested, and (2) suppress all display output including the "done" call. ?

regards, Luke.

On Tue, 14 Apr 2015, Luke Mewburn wrote:

I agree. I was just comenting on your suggestion about caching the in_progress_fd() result which would prevent that.

Yes.  And I think this is nice.

Sure. But what I'm saying is that progress() may be called a thousand times and only one or two of those calls will result in an actual print-out. So it is best to test the foreground status only at that point.

I'd suggest (1) but not (2).

Nicolas

Updated patch where is_foreground_fd() is only called in display() just before the output is to be displayed.

From d87997509fc631b8cdc7db63f289102d6ddfe933 Mon Sep 17 00:00:00 2001
From: Luke Mewburn <luke@mewburn.net>
Date: Mon, 13 Apr 2015 23:30:51 +1000
Subject: [PATCH] progress: no progress in background

Disable the display of the progress if stderr is not the current foreground process. Still display the final result when done.

Signed-off-by: Luke Mewburn <luke@mewburn.net>
---
 progress.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)

On Tue, 14 Apr 2015, Luke Mewburn wrote:

Acked-by: Nicolas Pitre <nico@fluxnic.net>

Windows does not have process groups. It is, therefore, the simplest to pretend that each process is in its own process group.

While here, move the getppid() stub from its old location (between two sync related functions) next to the two new functions.

Signed-off-by: Johannes Sixt <j6t@kdbg.org>
---
 compat/mingw.h | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

Johannes Sixt <j6t@kdbg.org> writes:

Thanks for a quick update.

The patch should do for now, but I suspect that it may give us a better abstraction to make the "is_foreground_fd(int fd)" or even "is_foreground(void)" the public API that would be implemented as

	int we_are_in_the_foreground(void)
        {
		return getpgid(0) == tcgetpgrp(fileno(stderr));
	}

in POSIX and Windows can implement entirely differently.

Thoughts?

Am 15.04.2015 um 20:48 schrieb Junio C Hamano:

IMHO, this level of abstraction goes too far. It may be that I am not familiar with process groups, and I find the implementation of is_foreground_fd() in progress.c close to where it's used quite educating and enlightening. Hiding a similar implementation miles away in cache.h and/or compat/ would not pay off for me.

-- Hannes

Hi Junio,

On 2015-04-15 20:48, Junio C Hamano wrote:

I really like it. We already require a couple of workarounds to be able to use `mintty` (which is a replacement terminal emulator that is more flexible than the default Windows terminal window, but it comes at the price that most Win32 programs think they are not running interactively inside a `mintty` session). I would really like to avoid having to finagle some really ugly code into `getpgid()` to make that test work.

In general, I find it rewarding not only from a portability point of view, but *especially* from a readability one if the code contains functions that are named semantically, i.e. they describe *why* they are called, not *how* they answer the question.

In this case, I would prefer the `is_foreground_fd(int fd)` one, but I am sure I can make the other signature work for us, too.

Ciao, Dscho

Johannes Schindelin <johannes.schindelin@gmx.de> writes:

Yeah, that was the rationale behind the suggestion (i.e. how may be different depending on the platform, and the main code flow cares more about why and not how).

I'll queue Luke's patch with J6t's compat/ for now, but if you find more suitable implementation for the higher-level abstraction, please do send a follow-up patch to update it.

Two Johannes'es may need to talk between themselves to agree what the right level of abstraction is, though. I trust two of you a lot more than my gut feeling when it comes to POSIX vs Windows API differences ;-)

Thanks.

On Wed, Apr 15, 2015 at 8:29 PM, Johannes Sixt <j6t@kdbg.org> wrote:

Windows does have some concept of process groups, but probably not quite what you want:

https://msdn.microsoft.com/en-us/library/windows/desktop/ms682083%28v=vs.85%29.aspx

Hi kusma,

On 2015-04-15 21:43, Erik Faye-Lund wrote:

Yes, and we actually need that in Git for Windows anyway because shooting down a process does not kill its child processes:

https://github.com/git-for-windows/msys2-runtime/commit/15f209511985092588b171703e5046eba937b47b#diff-8753cda163376cee6c80aab11eb8701fR402

However, using this code for `getppid()` would be serious overkill (not to mention an unbearable performance hit because you have to enumerate *all* processes to get that information).

Ciao, Dscho

On Thursday, April 16, 2015 at 2:45:11 PM UTC+2, Johannes Schindelin wrote:

is the windows "JobObject" similar to processgroup? at least killing the parent process in a jobobject will kill all childs as well: https://msdn.microsoft.com/en-us/library/windows/desktop/ms681949%28v=vs.85%29.aspx

best,
rupert
 

On Thursday, April 23, 2015 at 9:25:49 PM UTC+2, rupert thurner wrote:

here some discussion of windows console process group, CREATE_NEW_PROCESS_GROUP, GenerateConsoleCtrlEvent, and job object from the last century: https://www.microsoft.com/msj/0698/win320698.aspx

rupert

Hi Rupert,

On 2015-04-23 21:25, rupert thurner wrote:

Reading this page carefully reveals that you have to construct JobObjects explicitly. So you cannot get from a process ID to a JobObject; there is probably none. Or there is one. Or many.

Ciao, Johannes

On Wed, Apr 15, 2015 at 08:29:30PM +0200, Johannes Sixt wrote:
  | Windows does not have process groups. It is, therefore, the simplest
  | to pretend that each process is in its own process group.
  | 
  |  [...]
  | 
  | diff --git a/compat/mingw.h b/compat/mingw.h
  | index 7b523cf..a552026 100644
  | @@ -118,6 +116,12 @@ static inline int sigaddset(sigset_t *set, int signum)
  |  #define SIG_UNBLOCK 0
  |  static inline int sigprocmask(int how, const sigset_t *set, sigset_t *oldset)
  |  { return 0; }
  | +static inline pid_t getppid(void)
  | +{ return 1; }
  | +static inline pid_t getpgid(pid_t pid)
  | +{ return pid == 0 ? getpid() : pid; }
  | +static inline pid_t tcgetpgrp(int fd)
  | +{ return getpid(); }

This appears to be similar to the approach that tcsh uses too; return the current process ID for the process group ID. See https://github.com/tcsh-org/tcsh/blob/master/win32/ntport.h for tcsh's implementation of getpgrp() (a variation of getpgid()) and tcgetpgrp().

regards, Luke.

On Mon, Apr 13, 2015 at 11:01:04AM -0400, Nicolas Pitre wrote:
  | > That's what happens; the suppression only occurs if the process is
  | > currently background.  If I start a long-running operation (such as "git
  | > fsck"), the progress is displayed. I then suspend & background, and the
  | > progress is suppressed.  If I resume the process in the foreground, the
  | > progress starts to display again at the appropriate point.
  | 
  | I agree. I was just comenting on your suggestion about caching the 
  | in_progress_fd() result which would prevent that.

Ahh. My suggestion about is_foreground_fd() result caching within struct progress was only about caching the getpgid(0) portion of the test (as that's not expected to change for the life of the process), and not the tcgetpgrp(fd) portion. I.e, add 'int curpgid' to struct progress, set that to getpgid(0) in start_progress_display(), and compare tcgetpgrp(fd) against progress->curpgid.

In any case, I think it's a micro optimisation not worth worrying about at this point, given is_foreground_fd() is only called each time the output would change, per your feedback.

regards, Luke.

On Mon, Apr 13, 2015 at 11:48:50PM +1000, Luke Mewburn wrote:

I like this patch. It's simple and seems like a sensible change, and I appreciated the opportunity to learn about tcgetpgrp(3). The Windows folks will probably need to stub that function out, but they're no worse off than they were before.

I do agree with Nicolas that optimizing the code to avoid calling in_progress_fd as much as possible is a good idea, since system calls can be expensive on some systems.

Hi Brian,

On 2015-04-14 05:12, brian m. carlson wrote:

Thanks for thinking of us!

Ciao, Dscho