Re: [PATCH 2/2] mm: Add kvmalloc_ab_c and kvzalloc_struct

From: Julia Lawall
Date: Tue Mar 13 2018 - 13:20:04 EST




On Thu, 8 Mar 2018, Matthew Wilcox wrote:

> On Thu, Mar 08, 2018 at 07:24:47AM +0100, Julia Lawall wrote:
> > On Wed, 7 Mar 2018, Matthew Wilcox wrote:
> > > On Wed, Mar 07, 2018 at 10:18:21PM +0100, Julia Lawall wrote:
> > > > > Otherwise, yes, please. We could build a coccinelle rule for
> > > > > additional replacements...
> > > >
> > > > A potential semantic patch and the changes it generates are attached
> > > > below. Himanshu Jha helped with its development. Working on this
> > > > uncovered one bug, where the allocated array is too large, because the
> > > > size provided for it was a structure size, but actually only pointers to
> > > > that structure were to be stored in it.
> > >
> > > This is cool! Thanks for doing the coccinelle patch! Diffstat:
> > >
> > > 50 files changed, 81 insertions(+), 124 deletions(-)
> > >
> > > I find that pretty compelling. I'll repost the kvmalloc_struct patch
> > > imminently.
> >
> > Thanks. So it's OK to replace kmalloc and kzalloc, even though they
> > didn't previously consider vmalloc and even though kmalloc doesn't zero?
>
> We'll also need to replace the corresponding places where those structs
> are freed with kvfree(). Can coccinelle handle that too?

Is the use of vmalloc a necessary part of the design? Or could there be a
non vmalloc versions for call sites that are already ok with that?

julia

> > There are a few other cases that use GFP_NOFS and GFP_NOWAIT, but I didn't
> > transform those because the comment says that the flags should be
> > GFP_KERNEL based. Should those be transformed too?
>
> The problem with non-GFP_KERNEL allocations is that vmalloc may have to
> allocate page tables, which is always done with an implicit GFP_KERNEL
> allocation. There's an intent to get rid of GFP_NOFS, but that's not
> been realised yet (and I'm not sure of our strategy to eliminate it ...
> I'll send a separate email about that). I'm not sure why anything's
> trying to allocate with GFP_NOWAIT; can you send a list of those places?
>