bash usage of fdescfs [was: Re: amd64/188699: Dev tree]

[Bryan Drewery]

On 2014-04-21 16:40, Konstantin Belousov wrote:
> On Mon, Apr 21, 2014 at 05:23:04PM -0400, John Baldwin wrote:
>> On Monday, April 21, 2014 3:51:33 pm Konstantin Belousov wrote:
>> > On Mon, Apr 21, 2014 at 02:31:12PM -0400, John Baldwin wrote:
>> > > On Thursday, April 17, 2014 2:50:01 pm Konstantin Belousov wrote:
>> > > > The following reply was made to PR amd64/188699; it has been noted by
>> GNATS.
>> > > >
>> > > > From: Konstantin Belousov <kostikbel at gmail.com>
>> > > > To: John Allman <freebsd at hugme.org>
>> > > > Cc: freebsd-gnats-submit at FreeBSD.org
>> > > > Subject: Re: amd64/188699: Dev tree
>> > > > Date: Thu, 17 Apr 2014 21:44:52 +0300
>> > > >
>> > > >  On Wed, Apr 16, 2014 at 05:32:45PM +0000, John Allman wrote:
>> > > >  > This is how to reproduce it:
>> > > >  >
>> > > >  > Fresh install of 10 on AMD 64
>> > > >  > install bash `pkg install bash`
>> > > >  > Switch to bash `bash`
>> > > >  > push a here document into a loop: `while true ; do echo; done< <(echo
>> "123")`
>> > > >  > receive an error: "-su: /dev/fd/62: No such file or directory"
>> > > >  >
>> > > >  > I'm sorry I haven't been able to research this any further. I found
>> how while working on some important matters. As I mentioned the above 
>> works
>> fine in all
>> > > previous versions of FreeBSD up until 10.
>> > > >  > >How-To-Repeat:
>> > > >  > Fresh install
>> > > >  > pkg install bash
>> > > >  > bash
>> > > >  > while true; do echo foo done< <(echo "123")
>> > > >  >
>> > > >  > -su: /dev/fd/62: No such file or directory
>> > > >
>> > > >  So do you have fdescfs mounted on /dev/fd on the machine where the
>> > > >  test fails ?  It works for me on head, and if unmounted, I get the
>> > > >  same failure message as yours.  I very much doubt that it has anything
>> > > >  to do with a system version.
>> > >
>> > > Question I have is why is bash deciding to use /dev/fd/<n> and require
>> > > fdescfs?  On older releases bash uses named pipes for this instead.
>> >
>> > The aclocal.m4 contains the test which verifies the presence and usability
>> > of /dev/fd/n for n>=3 on the _build_ host.  The result of the test
>> > is used on the installation host afterward.
>> >
>> > Such kinds of bugs are endemic in our ports, but apparently upstreams
>> > are guilty too.
>> 
>> Yuck, yuck.  Should we fix our default package builders to not mount 
>> fdescfs?
> 
> IMO, using /dev/fd is more efficient since it avoids pipe inode 
> creation
> for the 'document here' interpretation.  The /dev/fd is also needed for
> fexecve(2) to work (with the shebang scripts).  Also, I believe that
> some other high-profile ports require it (OpenJDK ?).
> 
> That said, the solution is to have fdescfs mounted on /dev/fd.
> This probably should be done by an installer.

Yup, we mount it mostly for the java ports.

-- 
Regards,
Bryan Drewery