Re: [PATCH 0/3] irq/core: Fix and expand the irq affinity descriptor

[Sumit Saxena]

On Wed, Dec 19, 2018 at 6:25 PM Sumit Saxena <sumit.saxena@xxxxxxxxxxxx> wrote:
>
> On Wed, Dec 19, 2018 at 4:23 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >
> > On Tue, 4 Dec 2018, Dou Liyang wrote:
> >
> > > Now,  Spreading the interrupt affinity info by a cpumask pointer is not
> > > enough, meets a problem[1] and hard to expand in the future.
> > >
> > > Fix it by:
> > >
> > >      +-----------------------------------+
> > >      |                                   |
> > >      |     struct cpumask *affinity      |
> > >      |                                   |
> > >      +-----------------------------------+
> > >                        |
> > >     +------------------v-------------------+
> > >     |                                      |
> > >     | struct irq_affinity_desc {           |
> > >     |     struct cpumask   mask;           |
> > >     |     unsigned int     is_managed : 1; |
> > >     | };                                   |
> > >     |                                      |
> > >     +--------------------------------------+
> > >
> >
> > So, I've applied that lot for 4.21 (or whatever number it will be). That's
> > only the first step for solving Kashyap's problem.
> >
> > IIRC, then Kashap wanted to get initial interrupt spreading for these extra
> > magic interrupts as well, but not have them marked managed.
> >
> > That's trivial to do now with the two queued changes in that area:
> >
> >   - The rework above
> >
> >   - The support for interrupt sets from Jens
> >
> > Just adding a small bitfield to struct irq_affinity which allows to tell
> > the core that a particular interrupt set is not managed does the trick.
> >
> > Untested patch below.
> >
> > Kashyap, is that what you were looking for and if so, does it work?
> Thomas,
> We could not test these patches as they did net get applied to latest
> linux-block tree cleanly.
>
> Our requirement is: 1. extra interrupts should be un-managed and 2.
> should be spread to CPUs of local NUMA node.
> If interrupts are un-managed but not spread as per our requirement,
> then still driver/userspace apps can manage by spreading
> them as required by calling API- irq_set_affinity_hint().
>
> Thanks,
> Sumit
I tested this patchset with some minor rework to apply it on latest
linux block tree(4.20-rc7).
It worked as our expectation. For "pre_vectors" IRQs(extra set of
interrupts), "is_managed" flag is set to 0
and later driver can affine these "pre_vectors" to CPUs of local NUMA
node through API- irq_set_affinity_hint().
Regular set of interrupts(not pre_vectors/post_vectors) are managed,
"is_managed" set to 1.

Below are some data from my test setup-

# numactl --hardware

available: 2 nodes (0-1)

node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 36 37 38 39
40 41 42 43 44 45 46 47 48 49 50 51 52 53

node 0 size: 31822 MB

node 0 free: 30241 MB

node 1 cpus: 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 54
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71

node 1 size: 32248 MB

node 1 free: 31960 MB

node distances:

node   0   1

  0:  10  21

  1:  21  10

MegaRAID Controller(PCI device 86:00.0) is attached to node 1.
# find /sys -name *numa_node* | grep "86:00" | xargs cat
1

IRQ-CPU affinity of extra 16 interrupts for PCI  device 86:00.0:
irq 149, cpu list 18-35,54-71
irq 150, cpu list 18-35,54-71
irq 151, cpu list 18-35,54-71
irq 152, cpu list 18-35,54-71
irq 153, cpu list 18-35,54-71
irq 154, cpu list 18-35,54-71
irq 155, cpu list 18-35,54-71
irq 156, cpu list 18-35,54-71
irq 157, cpu list 18-35,54-71
irq 158, cpu list 18-35,54-71
irq 159, cpu list 18-35,54-71
irq 160, cpu list 18-35,54-71
irq 161, cpu list 18-35,54-71
irq 162, cpu list 18-35,54-71
irq 163, cpu list 18-35,54-71
irq 164, cpu list 18-35,54-71
---
# cat /sys/kernel/debug/irq/irqs/164 | grep is_managed
   is_managed:       0

Tested-by: Sumit Saxena <sumit.saxena@xxxxxxxxxxxx>

> >
> > Thanks,
> >
> >         tglx
> >
> > 8<-----------------
> >
> > Subject: genirq/affinity: Add support for non-managed affinity sets
> > From: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Date: Tue, 18 Dec 2018 16:46:47 +0100
> >
> > Some drivers need an extra set of interrupts which are not marked managed,
> > but should get initial interrupt spreading.
> >
> > Add a bitmap to struct irq_affinity which allows the driver to mark a
> > particular set of interrupts as non managed. Check the bitmap during
> > spreading and use the result to mark the interrupts in the sets
> > accordingly.
> >
> > The unmanaged interrupts get initial spreading, but user space can change
> > their affinity later on.
> >
> > Usage example:
> >
> >       struct irq_affinity affd = { .pre_vectors = 2 };
> >       int sets[2];
> >
> >       /* Fill in sets[] */
> >
> >       affd.nr_sets = 2;
> >       affd.sets = &sets;
> >       affd.unmanaged_sets = 0x02;
> >
> >       ......
> >
> > So both sets are properly spread out, but the second set is not marked
> > managed.
> >
> > Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > ---
> >  include/linux/interrupt.h |   10 ++++++----
> >  kernel/irq/affinity.c     |   24 ++++++++++++++----------
> >  2 files changed, 20 insertions(+), 14 deletions(-)
> >
> > --- a/kernel/irq/affinity.c
> > +++ b/kernel/irq/affinity.c
> > @@ -99,7 +99,8 @@ static int __irq_build_affinity_masks(co
> >                                       cpumask_var_t *node_to_cpumask,
> >                                       const struct cpumask *cpu_mask,
> >                                       struct cpumask *nmsk,
> > -                                     struct irq_affinity_desc *masks)
> > +                                     struct irq_affinity_desc *masks,
> > +                                     bool managed)
> >  {
> >         int n, nodes, cpus_per_vec, extra_vecs, done = 0;
> >         int last_affv = firstvec + numvecs;
> > @@ -154,6 +155,7 @@ static int __irq_build_affinity_masks(co
> >                         }
> >                         irq_spread_init_one(&masks[curvec].mask, nmsk,
> >                                                 cpus_per_vec);
> > +                       masks[curvec].is_managed = managed;
> >                 }
> >
> >                 done += v;
> > @@ -176,7 +178,8 @@ static int __irq_build_affinity_masks(co
> >  static int irq_build_affinity_masks(const struct irq_affinity *affd,
> >                                     int startvec, int numvecs, int firstvec,
> >                                     cpumask_var_t *node_to_cpumask,
> > -                                   struct irq_affinity_desc *masks)
> > +                                   struct irq_affinity_desc *masks,
> > +                                   bool managed)
> >  {
> >         int curvec = startvec, nr_present, nr_others;
> >         int ret = -ENOMEM;
> > @@ -196,7 +199,8 @@ static int irq_build_affinity_masks(cons
> >         /* Spread on present CPUs starting from affd->pre_vectors */
> >         nr_present = __irq_build_affinity_masks(affd, curvec, numvecs,
> >                                                 firstvec, node_to_cpumask,
> > -                                               cpu_present_mask, nmsk, masks);
> > +                                               cpu_present_mask, nmsk, masks,
> > +                                               managed);
> >
> >         /*
> >          * Spread on non present CPUs starting from the next vector to be
> > @@ -211,7 +215,7 @@ static int irq_build_affinity_masks(cons
> >         cpumask_andnot(npresmsk, cpu_possible_mask, cpu_present_mask);
> >         nr_others = __irq_build_affinity_masks(affd, curvec, numvecs,
> >                                                firstvec, node_to_cpumask,
> > -                                              npresmsk, nmsk, masks);
> > +                                              npresmsk, nmsk, masks, managed);
> >         put_online_cpus();
> >
> >         if (nr_present < numvecs)
> > @@ -268,10 +272,11 @@ irq_create_affinity_masks(int nvecs, con
> >
> >         for (i = 0, usedvecs = 0; i < nr_sets; i++) {
> >                 int this_vecs = affd->sets ? affd->sets[i] : affvecs;
> > +               bool managed = !test_bit(i, &affd->unmanaged_sets);
> >                 int ret;
> >
> > -               ret = irq_build_affinity_masks(affd, curvec, this_vecs,
> > -                                               curvec, node_to_cpumask, masks);
> > +               ret = irq_build_affinity_masks(affd, curvec, this_vecs, curvec,
> > +                                              node_to_cpumask, masks, managed);
> >                 if (ret) {
> >                         kfree(masks);
> >                         masks = NULL;
> > @@ -289,10 +294,6 @@ irq_create_affinity_masks(int nvecs, con
> >         for (; curvec < nvecs; curvec++)
> >                 cpumask_copy(&masks[curvec].mask, irq_default_affinity);
> >
> > -       /* Mark the managed interrupts */
> > -       for (i = affd->pre_vectors; i < nvecs - affd->post_vectors; i++)
> > -               masks[i].is_managed = 1;
> > -
> >  outnodemsk:
> >         free_node_to_cpumask(node_to_cpumask);
> >         return masks;
> > @@ -316,6 +317,9 @@ int irq_calc_affinity_vectors(int minvec
> >         if (affd->nr_sets) {
> >                 int i;
> >
> > +               if (WARN_ON_ONCE(affd->nr_sets > BITS_PER_LONG))
> > +                       return 0;
> > +
> >                 for (i = 0, set_vecs = 0;  i < affd->nr_sets; i++)
> >                         set_vecs += affd->sets[i];
> >         } else {
> > --- a/include/linux/interrupt.h
> > +++ b/include/linux/interrupt.h
> > @@ -249,12 +249,14 @@ struct irq_affinity_notify {
> >   *                     the MSI(-X) vector space
> >   * @nr_sets:           Length of passed in *sets array
> >   * @sets:              Number of affinitized sets
> > + * @unmanaged_sets:    Bitmap to mark members of @sets as unmanaged
> >   */
> >  struct irq_affinity {
> > -       int     pre_vectors;
> > -       int     post_vectors;
> > -       int     nr_sets;
> > -       int     *sets;
> > +       int             pre_vectors;
> > +       int             post_vectors;
> > +       int             nr_sets;
> > +       int             *sets;
> > +       unsigned long   unmanaged_sets;
> >  };
> >
> >  /**