by Akinobu Mita

[permalink] [raw]

Subject: Re: [PATCH 0/6] jbd cleanup

On Fri, Sep 09, 2005 at 02:15:22AM -0700, Andrew Morton wrote:
> Akinobu Mita <[email protected]> wrote:
> >
> > The following 6 patches cleanup the jbd code and kill about 200 lines.
> >
>
> Thanks, but I'm not inclined to apply them.
>
> a) Maybe 70-80% of the Linux world uses this filesystem. We need to be
> very cautious in making changes to it.

And we need many eyeballs.
(I've tried to understand how the jbd works several times.
But I always failed.)

> b) A relatively large number of people are carrying quite large
> out-of-tree patches, some of which they're hoping to merge sometime.
> Admittedly more against ext3 than JBD, but there is potential here to
> cause those people trouble.
>
> Plus the switch to list_heads in journal_s has some impact on type safety
> and debuggability - I considered doing it years ago but decided not to
> because I found I _used_ those pointers fairly commonly in development.
> list_heads are a bit of a pain in gdb (kgdb and kernel core dumps), for
> example.

About the debuggability of list_heads, how about adding the kind of
the following gdb macros in .gdbinit?

---

define list_entry
set $ptr=$arg0
p ($arg1 *)((char *)$ptr - (size_t) &(($arg1 *)0)->$arg2)
end

define list_entry_s
set $ptr=$arg0
p (struct $arg1 *)((char *)$ptr - (size_t) &((struct $arg1 *)0)->$arg2)
end

define to_journal_head
list_entry_s $arg0 journal_head b_list
end

2005-09-10 21:59:29

by Andrew Morton

[permalink] [raw]

Subject: Re: [PATCH 0/6] jbd cleanup

Akinobu Mita <[email protected]> wrote:
>
> On Fri, Sep 09, 2005 at 02:15:22AM -0700, Andrew Morton wrote:
> > Akinobu Mita <[email protected]> wrote:
> > >
> > > The following 6 patches cleanup the jbd code and kill about 200 lines.
> > >
> >
> > Thanks, but I'm not inclined to apply them.
> >
> > a) Maybe 70-80% of the Linux world uses this filesystem. We need to be
> > very cautious in making changes to it.
>
> And we need many eyeballs.

True. And the only way to really learn code is to make changes to it.

> (I've tried to understand how the jbd works several times.
> But I always failed.)

It's very hard to reverse engineer the high-level design concepts from the
implementation. And the design concepts in JBD are really complex, which
is a problem fo us.

When I first had to learn the thing 4-5 years back I sat down for a solid
week and wrote a 40-odd page how-it-works document for myself, just to
force it into my head. It was probably about 50% accurate, but it was a
useful exercise.

> About the debuggability of list_heads, how about adding the kind of
> the following gdb macros in .gdbinit?
>
> ---
>
> define list_entry
> set $ptr=$arg0
> p ($arg1 *)((char *)$ptr - (size_t) &(($arg1 *)0)->$arg2)
> end
>
> define list_entry_s
> set $ptr=$arg0
> p (struct $arg1 *)((char *)$ptr - (size_t) &((struct $arg1 *)0)->$arg2)
> end
>
> define to_journal_head
> list_entry_s $arg0 journal_head b_list
> end

Here's mine ;)

# list_entry list type member
define list_entry
set $off = (int)&(((struct $arg1 *)0)->$arg2)
set $addr = (int)$arg0
set $res = $addr - $off
printf "0x%x\n", (struct $arg1 *)$res
end