On Tue, Nov 1, 2022 at 2:42 PM Marcelo Tosatti <mtosatti@xxxxxxxxxx> wrote:
>
> On Tue, Oct 11, 2022 at 03:20:39PM -0300, Leonardo Brás wrote:
> > On Fri, 2022-10-07 at 18:42 -0300, Leonardo Brás wrote:
> > > On Wed, 2022-10-05 at 09:57 -0300, Marcelo Tosatti wrote:
> > > > On Tue, Oct 04, 2022 at 03:25:37AM -0300, Leonardo Bras wrote:
> > > > > Currently pcrypt_aead_init_tfm() will pick callback cpus (ctx->cb_cpu)
> > > > > from any online cpus. Later padata_reorder() will queue_work_on() the
> > > > > chosen cb_cpu.
> > > > >
> > > > > This is undesired if the chosen cb_cpu is listed as isolated (i.e. using
> > > > > isolcpus=... kernel parameter), since the work queued will interfere with
> > > > > the workload on the isolated cpu.
> > > > >
> > > > > Make sure isolated cpus are not used for pcrypt.
> > > > >
> > > > > Signed-off-by: Leonardo Bras <leobras@xxxxxxxxxx>
> > > > > ---
> > > > > crypto/pcrypt.c | 10 +++++++---
> > > > > 1 file changed, 7 insertions(+), 3 deletions(-)
> > > > >
> > > > > diff --git a/crypto/pcrypt.c b/crypto/pcrypt.c
> > > > > index 9d10b846ccf73..9017d08c91a8d 100644
> > > > > --- a/crypto/pcrypt.c
> > > > > +++ b/crypto/pcrypt.c
> > > > > @@ -16,6 +16,7 @@
> > > > > #include <linux/kobject.h>
> > > > > #include <linux/cpu.h>
> > > > > #include <crypto/pcrypt.h>
> > > > > +#include <linux/sched/isolation.h>
> > > > >
> > > > > static struct padata_instance *pencrypt;
> > > > > static struct padata_instance *pdecrypt;
> > > > > @@ -175,13 +176,16 @@ static int pcrypt_aead_init_tfm(struct crypto_aead *tfm)
> > > > > struct pcrypt_instance_ctx *ictx = aead_instance_ctx(inst);
> > > > > struct pcrypt_aead_ctx *ctx = crypto_aead_ctx(tfm);
> > > > > struct crypto_aead *cipher;
> > > > > + struct cpumask non_isolated;
> > > > > +
> > > > > + cpumask_and(&non_isolated, cpu_online_mask, housekeeping_cpumask(HK_TYPE_DOMAIN));
> > > >
> > > > Since certain systems do not use isolcpus=domain, so please use a flag
> > > > that is setup by nohz_full=, for example HK_FLAG_MISC:
> > > >
> > > > static int __init housekeeping_nohz_full_setup(char *str)
> > > > {
> > > > unsigned long flags;
> > > >
> > > > flags = HK_FLAG_TICK | HK_FLAG_WQ | HK_FLAG_TIMER | HK_FLAG_RCU |
> > > > HK_FLAG_MISC | HK_FLAG_KTHREAD;
> > > >
> > > > return housekeeping_setup(str, flags);
> > > > }
> > > > __setup("nohz_full=", housekeeping_nohz_full_setup);
> > >
> > > Oh, sure.
> > > Since we are talking about WorkQueues, I think it makes sense to pick
> > > HK_FLAG_WQ.
> > >
> > > >
> > > > Also, shouldnt you use cpumask_t ?/
> > >
> > > Yeah, I think so.
> > > I was quick to choose the 'struct cpumask' because all functions would operate
> > > in this variable type, but yeah, I think it makes sense to have this variable
> > > being opaque here.
> >
> > In fact, it seems neither 'struct cpumask' nor 'cpumask_t' are recommended to be
> > used allocated in the stack, due to the large size it can get (up to 1kB).
> >
> > At include/linux/cpumask.h we have:
> > 'cpumask_var_t: struct cpumask for stack usage'
> > which should work better at least for init functions like this.
> >
> > In other cases, I see 'static cpumask_t' being used to avoid the allocation
> > overhead, but it's probably due to the functions being called in very specific
> > scenarios. It could mean trouble if multiple cpus try to use it at once.
> >
> > What do you recommend on it?
>
> Sorry for the delay. I suppose allocating and freeing is OK in this context, since
> its initialization time and not a hot path?
Yeah, it makes sense this way. I will allocate as suggested!
(Unless my other change gets in, so this new variable and a lot of
overhead can be avoided.)
Thanks Marcelo!
