Virtualization - Apr 2018 - [PATCH] kvmalloc: always use vmalloc if CONFIG_DEBUG_VM

On Mon 23-04-18 20:25:15, Mikulas Patocka wrote:
> 
> 
> On Mon, 23 Apr 2018, Michal Hocko wrote:
> 
> > On Mon 23-04-18 10:06:08, Mikulas Patocka wrote:
> > 
> > > > > He didn't want to fix vmalloc(GFP_NOIO)
> > > > 
> > > > I don't remember that conversation, so I don't know
whether I agree with
> > > > his reasoning or not.  But we are supposed to be moving away
from GFP_NOIO
> > > > towards marking regions with memalloc_noio_save() / restore.
If you do
> > > > that, you won't need vmalloc(GFP_NOIO).
> > > 
> > > He said the same thing a year ago. And there was small progress.
6 out of
> > > 27 __vmalloc calls were converted to memalloc_noio_save in a year
- 5 in
> > > infiniband and 1 in btrfs. (the whole discussion is here 
> > > http://lkml.iu.edu/hypermail/linux/kernel/1706.3/04681.html )
> > 
> > Well this is not that easy. It requires a cooperation from
maintainers.
> > I can only do as much. I've posted patches in the past and
actively
> > bringing up this topic at LSFMM last two years...
> 
> You're right - but you have chosen the uneasy path.
Yes.
> Fixing __vmalloc code 
> is easy and it doesn't require cooperation with maintainers.
But it is a hack against the intention of the scope api. It also alows
maintainers to not care about their broken code.
> > > He refuses 15-line patch to fix GFP_NOIO bug because he believes
that in 4
> > > years, the kernel will be refactored and GFP_NOIO will be
eliminated. Why
> > > does he have veto over this part of the code? I'd much rather
argue with
> > > people who have constructive comments about fixing bugs than with
him.
> > 
> > I didn't NACK the patch AFAIR. I've said it is not a good idea
longterm.
> > I would be much more willing to change my mind if you would back your
> > patch by a real bug report. Hacks are acceptable when we have a real
> > issue in hands. But if we want to fix potential issue then better make
> > it properly.
> 
> Developers should fix bugs in advance, not to wait until a crash hapens, 
> is analyzed and reported.
I agree. But are those existing users broken in the first place? I have
seen so many GFP_NOFS abuses that I would dare to guess that most of
those vmalloc NOFS abusers can be simply turned into GFP_KERNEL. Maybe
that is the reason we haven't heard any complains in years.

-- 
Michal Hocko
SUSE Labs

On Tue, 24 Apr 2018, Michal Hocko wrote:
> On Mon 23-04-18 20:25:15, Mikulas Patocka wrote:
> 
> > Fixing __vmalloc code 
> > is easy and it doesn't require cooperation with maintainers.
> 
> But it is a hack against the intention of the scope api.
It is not! You can fix __vmalloc now and you can convert the kernel to the 
scope API in 4 years. It's not one way or the other.
> It also alows maintainers to not care about their broken code.
Most maintainers don't even know that it's broken. Out of 14 subsystems 
using __vmalloc with GFP_NOIO/NOFS, only 2 realized that its 
implementation is broken and implemented a workaround (me and the XFS 
developers).

Misimplementing a function in a subtle and hard-to-notice way won't drive 
developers away from using it.
> > > > He refuses 15-line patch to fix GFP_NOIO bug because he
believes that in 4
> > > > years, the kernel will be refactored and GFP_NOIO will be
eliminated. Why
> > > > does he have veto over this part of the code? I'd much
rather argue with
> > > > people who have constructive comments about fixing bugs than
with him.
> > > 
> > > I didn't NACK the patch AFAIR. I've said it is not a good
idea longterm.
> > > I would be much more willing to change my mind if you would back
your
> > > patch by a real bug report. Hacks are acceptable when we have a
real
> > > issue in hands. But if we want to fix potential issue then better
make
> > > it properly.
> > 
> > Developers should fix bugs in advance, not to wait until a crash
hapens,
> > is analyzed and reported.
> 
> I agree. But are those existing users broken in the first place? I have
> seen so many GFP_NOFS abuses that I would dare to guess that most of
> those vmalloc NOFS abusers can be simply turned into GFP_KERNEL. Maybe
> that is the reason we haven't heard any complains in years.
alloc_pages reclaims clean pages and most hard work is done by kswapd, so 
GFP_KERNEL doesn't cause much issues with writeback. But cheating isn't 
justified if you can get away with it. Incorrect GFP flags cause real 
problems with shrinkers - because shrinkers are called from alloc_pages 
and they do respond to GFP flags.

I had reported deadlock due to GFP issues (9d28eb12447). And the worst 
thing about these bug reports is that they are totally unreproducible and 
I get nothing, but a stacktrace in bugzilla. I had to guess what happened 
and I couldn't even test if the patch fixed the bug.

I'm not really happy that you are deliberately leaving these issues behind 
and making excuses.

Mikulas

On Tue 24-04-18 11:30:40, Mikulas Patocka wrote:
> 
> 
> On Tue, 24 Apr 2018, Michal Hocko wrote:
> 
> > On Mon 23-04-18 20:25:15, Mikulas Patocka wrote:
> > 
> > > Fixing __vmalloc code 
> > > is easy and it doesn't require cooperation with maintainers.
> > 
> > But it is a hack against the intention of the scope api.
> 
> It is not!
This discussion simply doesn't make much sense it seems. The scope API
is to document the scope of the reclaim recursion critical section. That
certainly is not a utility function like vmalloc.
-- 
Michal Hocko
SUSE Labs