Re: [PATCH] mm/gup_benchmark: GUP_BENCHMARK depends on DEBUG_FS

[John Hubbard]

On 11/7/20 6:58 PM, Song Bao Hua (Barry Song) wrote:
> > On 11/7/20 2:20 PM, Randy Dunlap wrote:
> > > On 11/7/20 11:16 AM, John Hubbard wrote:
> > > > On 11/7/20 11:05 AM, Song Bao Hua (Barry Song) wrote:
> > > > > > From: John Hubbard [mailto:jhubbard@xxxxxxxxxx]
> > > > ...
> > But if you really disagree, then I'd go with, just drop the patch entirely, because
> > it doesn't really make things better as written...IMHO anyway. :)
>
> Just imagine a case, we don't enable DEBUG_FS but we enable GUP_TEST, we will
> get an image with totally useless code section since GUP_TEST depends on debugfs
> entry to perform any useful functionality.

Looking at the choices, from the user's (usually kernel developer's) experience:

a) The user enables GUP_TEST, then boots up, runs, and is briefly surprised by a
runtime failure. But it's a very quick diagnosis: "open: No such file or directory",
when trying to make that ioctl call. The path indicates that it's a debug fs path,
so the solution is pretty clear, at least for the main audience.

b) The other choice: the user *never even sees* GUP_TEST as a choice. This especially
bothers me because sometimes you find things by poking around in the menu, although
of course "you should already know about it"...but there's a lot to "already know"
in a large kernel.

From a user experience, it's way better to simply see what you want, and select it
in the menu. Or, at least get some prompt that you need to pre-select something else.


> The difference between "depends on" and "select" for this case is like:
> depends on: if we want to use GUP_TEST, we have to enable DEBUG_FS first;
> select: if we enable GUP_TEST, Kconfig will enable DEBUG_FS automatically.
>
> To me, I am 60% inclined to "depends on" as I think "DEBUG_FS" is more
> of a pre-condition of GUP_TEST than an internal part of GUP_TEST. So people
> should realize the pre-condition must be met before using GUP_TEST and


Right, but first of course they must read every single line of the test code
carefully. And while it is true the you often *do* end up reading most or
all of the test code, there are situations in which you don't need to. We'd
be taking away some of those situations. :)


> they must manually enable it if they haven't. That's why I think this patch is
> making things better.

...which makes things a little bit worse.


thanks,
--
John Hubbard
NVIDIA