Re: No fsync on history file? I lost my history

[Mikael Magnusson <mikachu@xxxxxxxxx>]

On Sun, Sep 23, 2018 at 5:25 PM, lilydjwg <lilydjwg@xxxxxxxxx> wrote:
> I'm sending an updated patch.
>
> On Sun, Sep 23, 2018 at 02:46:51PM +0000, Daniel Shahaf wrote:
>> fsync() is in POSIX.  I assume we can just call it, but if somebody complains
>> we'll need to use an HAVE_FSYNC guard.
>
> I don't know how to add a HAVE_FSYNC macro to the build system, sorry.
>
>> > +++ b/Src/hist.c
>> > @@ -2933,6 +2933,9 @@ savehistfile(char *fn, int err, int writeflags)
>> >             lasthist.text = ztrdup(start);
>> >         }
>> >     }
>> > +   fflush(out); /* need to flush before fsync */
>>
>> Isn't the fflush() on line 2927 sufficient?  (Even if it isn't, I would have
>> expected a ret>=0 guard around this call.)
>
> It should call write(2) to write out the buffered data. Then the kernel
> can fsync the data to disk. A guard has been added.
>
>> > +   if (fsync(fileno(out)) < 0 && ret >= 0)
>> > +       ret = -1;
>>
>> fileno() can return -1.
>
> It shouldn't matter, fsync will return EBADF for -1. Other parts of the
> code don't check for this either, and I can't think a case when fileno
> would fail after so many successful I/O operations on it (corrupted memory?)
>
>> Shouldn't the ret>=0 check happen before the calls to fileno() and fsync()?
>
> Yes, I've changed that.
>
> --
> Best regards,
> lilydjwg
>
> From 3c6c07733f12176c737d1f610f0dceafd07437df Mon Sep 17 00:00:00 2001
> From: lilydjwg <lilydjwg@xxxxxxxxx>
> Date: Sun, 23 Sep 2018 22:12:56 +0800
> Subject: [PATCH] Call fsync after writing out new histfile
>
> to ensure the data is on disk before the rename in case of a system crash.
> ---
>  Src/hist.c | 4 ++++
>  1 file changed, 4 insertions(+)
>
> diff --git a/Src/hist.c b/Src/hist.c
> index dbdc1e4e5..d3370252d 100644
> --- a/Src/hist.c
> +++ b/Src/hist.c
> @@ -2933,6 +2933,10 @@ savehistfile(char *fn, int err, int writeflags)
>                 lasthist.text = ztrdup(start);
>             }
>         }
> +       if (ret >= 0)
> +           ret = fflush(out); /* need to flush before fsync */
> +       if (ret >= 0 && fsync(fileno(out)) < 0)
> +           ret = -1;
>         if (fclose(out) < 0 && ret >= 0)
>             ret = -1;
>         if (ret >= 0) {

Please only do this for the tmpfile case, or interactive usage is
going to be unbearable under load with incappendhistory.

-- 
Mikael Magnusson