On gioved? 14 aprile 2022 09:03:40 CEST Julia Lawall wrote:
>
> On Wed, 13 Apr 2022, Ira Weiny wrote:
>
> > On Wed, Apr 13, 2022 at 05:44:54PM -0700, Alison Schofield wrote:
> > > On Thu, Apr 14, 2022 at 12:55:31AM +0200, Fabio M. De Francesco
wrote:
> > > > The use of kmap() is being deprecated in favor of kmap_local_page()
> > > > where it is feasible. The same is true for kmap_atomic().
> > > >
> > > > In file pci/hmm/hmm.c, function hmm_store() test if we are in
atomic
> > > > context and, if so, it calls kmap_atomic(), if not, it calls
kmap().
> > > >
> > > > First of all, in_atomic() shouldn't be used in drivers. This macro
> > > > cannot always detect atomic context; in particular, it cannot know
> > > > about held spinlocks in non-preemptible kernels.
> > > >
> > > > Notwithstanding what it is said above, this code doesn't need to
care
> > > > whether or not it is executing in atomic context. It can simply use
> > > > kmap_local_page() / kunmap_local() that can instead do the mapping
/
> > > > unmapping regardless of the context.
> > > >
> > > > With kmap_local_page(), the mapping is per thread, CPU local and
not
> > > > globally visible. Therefore, hmm_store()() is a function where the
use
> > > > of kmap_local_page() in place of both kmap() and kmap_atomic() is
> > > > correctly suited.
> > > >
> > > > Convert the calls of kmap() / kunmap() and kmap_atomic() /
> > > > kunmap_atomic() to kmap_local_page() / kunmap_local() and drop the
> > > > unnecessary tests which test if the code is in atomic context.
> > > >
> > >
> > > Not specifically about this patch, but more generally about all
> > > such conversions - is there a 'proof' that shows this just works
> >
> > Just code inspection. Most of them that I have done have been compile
tested
> > only. Part of the key is that des is a local variable and is not
aliased by
> > anything outside this function.
>
> Typically, the concern about being in atomic context has to do with
> whether GFP_KERNEL or GFP_ATOMIC should be used, ie whether allocation
> can sleep.
I'd add that the concern about being in atomic context has mainly to do
with calling whatever function that may sleep.
Some time ago I analyzed a calls chain which, under spinlocks and with
IRQ's disabled, led to console_lock() which is annotated with
might_sleep(). It took about 8000 ms to recover when executing in a 4 CPU /
8 SMT System. Linus T. suggested to make this work asynchronous (commit
1ee33b1ca2b8 ("tty: n_hdlc: make n_hdlc_tty_wakeup() asynchronous")).
> It doesn't have to do with whether some data can be shared.
Yes, FWIW I agree with you.
> Is that the concern here?
The concern here is about the locality of the pointer variable to which the
struct page has been mapped to. In atomic context we are not allowed to
kmap() (this is why in the code we had that in_atomic() test), instead we
can kmap_local_page() or kmap_atomic(). The latter is strongly discouraged
in favor of the former.
Furthermore, Alison was asking if we can prove that these kinds of
conversions can actually work when we have not the hardware for testing. As
Ira wrote, code inspection is sufficient to prove it.
Thanks,
Fabio M. De Francesco
