Re: [PATCH 5/6] drm/i915: dynamically allocate forcewake domains

[Daniele Ceraolo Spurio <daniele.ceraolospurio@xxxxxxxxx>]

On 6/18/19 2:23 AM, Tvrtko Ursulin wrote:
> On 17/06/2019 19:09, Daniele Ceraolo Spurio wrote:
> > We'd like to introduce a display uncore with no forcewake domains, so
> > let's avoid wasting memory and be ready to allocate only what we need.
> > Even without multiple uncore, we still don't need all the domains on all
> > gens.
>
> Looks good in principle, only a few conversation points below.
>
> > Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@xxxxxxxxx>
> > ---
> >   drivers/gpu/drm/i915/intel_uncore.c | 155 ++++++++++++++++++----------
> >   drivers/gpu/drm/i915/intel_uncore.h |  13 +--
> >   2 files changed, 102 insertions(+), 66 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/intel_uncore.c 
> > b/drivers/gpu/drm/i915/intel_uncore.c
> > index c0f5567ee096..7f311827c3ef 100644
> > --- a/drivers/gpu/drm/i915/intel_uncore.c
> > +++ b/drivers/gpu/drm/i915/intel_uncore.c
> > @@ -344,7 +344,7 @@ intel_uncore_fw_release_timer(struct hrtimer *timer)
> >   {
> >       struct intel_uncore_forcewake_domain *domain =
> >              container_of(timer, struct intel_uncore_forcewake_domain, 
> > timer);
> > -    struct intel_uncore *uncore = forcewake_domain_to_uncore(domain);
> > +    struct intel_uncore *uncore = domain->uncore;
> >       unsigned long irqflags;
> >       assert_rpm_device_not_suspended(uncore->rpm);
> > @@ -1291,23 +1291,23 @@ do { \
> >       (uncore)->funcs.read_fw_domains = x##_reg_read_fw_domains; \
> >   } while (0)
> > -static void fw_domain_init(struct intel_uncore *uncore,
> > +static int fw_domain_init(struct intel_uncore *uncore,
> >                  enum forcewake_domain_id domain_id,
> >                  i915_reg_t reg_set,
> >                  i915_reg_t reg_ack)
> >   {
> >       struct intel_uncore_forcewake_domain *d;
> > -    if (WARN_ON(domain_id >= FW_DOMAIN_ID_COUNT))
> > -        return;
> > -
> > -    d = &uncore->fw_domain[domain_id];
> > +    GEM_BUG_ON(domain_id >= FW_DOMAIN_ID_COUNT);
> > -    WARN_ON(d->wake_count);
>
> I wonder what was this for.
>
> Do you want to add some protection against double init of the same domain?

just a GEM_BUG_ON maybe? It shouldn't really ever happen unless we're 
moving something around.

> > +    d = kzalloc(sizeof(*d), GFP_KERNEL);
> > +    if (!d)
> > +        return -ENOMEM;
> >       WARN_ON(!i915_mmio_reg_valid(reg_set));
> >       WARN_ON(!i915_mmio_reg_valid(reg_ack));
> > +    d->uncore = uncore;
> >       d->wake_count = 0;
> >       d->reg_set = uncore->regs + i915_mmio_reg_offset(reg_set);
> >       d->reg_ack = uncore->regs + i915_mmio_reg_offset(reg_ack);
> > @@ -1324,7 +1324,6 @@ static void fw_domain_init(struct intel_uncore 
> > *uncore,
> >       BUILD_BUG_ON(FORCEWAKE_MEDIA_VEBOX0 != (1 << 
> > FW_DOMAIN_ID_MEDIA_VEBOX0));
> >       BUILD_BUG_ON(FORCEWAKE_MEDIA_VEBOX1 != (1 << 
> > FW_DOMAIN_ID_MEDIA_VEBOX1));
> > -
> >       d->mask = BIT(domain_id);
> >       hrtimer_init(&d->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
> > @@ -1333,6 +1332,10 @@ static void fw_domain_init(struct intel_uncore 
> > *uncore,
> >       uncore->fw_domains |= BIT(domain_id);
> >       fw_domain_reset(d);
> > +
> > +    uncore->fw_domain[domain_id] = d;
> > +
> > +    return 0;
> >   }
> >   static void fw_domain_fini(struct intel_uncore *uncore,
> > @@ -1340,77 +1343,92 @@ static void fw_domain_fini(struct intel_uncore 
> > *uncore,
> >   {
> >       struct intel_uncore_forcewake_domain *d;
> > -    if (WARN_ON(domain_id >= FW_DOMAIN_ID_COUNT))
> > -        return;
> > +    GEM_BUG_ON(domain_id >= FW_DOMAIN_ID_COUNT);
> > -    d = &uncore->fw_domain[domain_id];
> > +    d = fetch_and_zero(&uncore->fw_domain[domain_id]);
>
> Maybe early if !d here and function flow just carries on?

will do.

> > +    uncore->fw_domains &= ~BIT(domain_id);
> > -    WARN_ON(d->wake_count);
> > -    WARN_ON(hrtimer_cancel(&d->timer));
> > -    memset(d, 0, sizeof(*d));
> > +    if (d) {
> > +        WARN_ON(d->wake_count);
> > +        WARN_ON(hrtimer_cancel(&d->timer));
> > +        kfree(d);
> > +    }
> > +}
> > -    uncore->fw_domains &= ~BIT(domain_id);
> > +static void intel_uncore_fw_domains_fini(struct intel_uncore *uncore)
> > +{
> > +    struct intel_uncore_forcewake_domain *d;
> > +    int tmp;
> > +
> > +    for_each_fw_domain(d, uncore, tmp)
> > +        fw_domain_fini(uncore, d->id);
> >   }
> > -static void intel_uncore_fw_domains_init(struct intel_uncore *uncore)
> > +static int intel_uncore_fw_domains_init(struct intel_uncore *uncore)
> >   {
> >       struct drm_i915_private *i915 = uncore->i915;
> > +    int ret;
> >       GEM_BUG_ON(!intel_uncore_has_forcewake(uncore));
> > +#define __fw_domain_init(id__, set__, ack__) \
> > +    ret = fw_domain_init(uncore, (id__), (set__), (ack__)); \
> > +    if (ret) \
> > +        goto out_clean;
>
> Hidden control flow is slightly objectionable but I don't offer any nice 
> alternatives so I guess will have to pass. Or maybe accumulate the error 
> (err |= fw_domain_init(..)) as you go and then cleanup at the end if any 
> failed?

I'd prefer to avoid accumulating the error since it'd just cause us to 
having to unroll more domains when we could've stopped early.

> On the other hand the idea slightly conflicts with my other suggestion 
> to rename existing fw_domain_init to __fw_domain_init and call the macro 
> fw_domain_init and avoid all the churn below.

I'll pick this suggestion among the 2, unless there is another 
suggestion on how to avoid the hidden goto.

> > +
> >       if (INTEL_GEN(i915) >= 11) {
> >           int i;
> > -        uncore->funcs.force_wake_get =
> > -            fw_domains_get_with_fallback;
> > +        uncore->funcs.force_wake_get = fw_domains_get_with_fallback;
> >           uncore->funcs.force_wake_put = fw_domains_put;
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_RENDER,
> > -                   FORCEWAKE_RENDER_GEN9,
> > -                   FORCEWAKE_ACK_RENDER_GEN9);
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_BLITTER,
> > -                   FORCEWAKE_BLITTER_GEN9,
> > -                   FORCEWAKE_ACK_BLITTER_GEN9);
> > +        __fw_domain_init(FW_DOMAIN_ID_RENDER,
> > +                 FORCEWAKE_RENDER_GEN9,
> > +                 FORCEWAKE_ACK_RENDER_GEN9);
> > +        __fw_domain_init(FW_DOMAIN_ID_BLITTER,
> > +                 FORCEWAKE_BLITTER_GEN9,
> > +                 FORCEWAKE_ACK_BLITTER_GEN9);
> > +
> >           for (i = 0; i < I915_MAX_VCS; i++) {
> >               if (!HAS_ENGINE(i915, _VCS(i)))
> >                   continue;
> > -            fw_domain_init(uncore, FW_DOMAIN_ID_MEDIA_VDBOX0 + i,
> > -                       FORCEWAKE_MEDIA_VDBOX_GEN11(i),
> > -                       FORCEWAKE_ACK_MEDIA_VDBOX_GEN11(i));
> > +            __fw_domain_init(FW_DOMAIN_ID_MEDIA_VDBOX0 + i,
> > +                     FORCEWAKE_MEDIA_VDBOX_GEN11(i),
> > +                     FORCEWAKE_ACK_MEDIA_VDBOX_GEN11(i));
> >           }
> >           for (i = 0; i < I915_MAX_VECS; i++) {
> >               if (!HAS_ENGINE(i915, _VECS(i)))
> >                   continue;
> > -            fw_domain_init(uncore, FW_DOMAIN_ID_MEDIA_VEBOX0 + i,
> > -                       FORCEWAKE_MEDIA_VEBOX_GEN11(i),
> > -                       FORCEWAKE_ACK_MEDIA_VEBOX_GEN11(i));
> > +            __fw_domain_init(FW_DOMAIN_ID_MEDIA_VEBOX0 + i,
> > +                     FORCEWAKE_MEDIA_VEBOX_GEN11(i),
> > +                     FORCEWAKE_ACK_MEDIA_VEBOX_GEN11(i));
> >           }
> >       } else if (IS_GEN_RANGE(i915, 9, 10)) {
> > -        uncore->funcs.force_wake_get =
> > -            fw_domains_get_with_fallback;
> > +        uncore->funcs.force_wake_get = fw_domains_get_with_fallback;
> >           uncore->funcs.force_wake_put = fw_domains_put;
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_RENDER,
> > -                   FORCEWAKE_RENDER_GEN9,
> > -                   FORCEWAKE_ACK_RENDER_GEN9);
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_BLITTER,
> > -                   FORCEWAKE_BLITTER_GEN9,
> > -                   FORCEWAKE_ACK_BLITTER_GEN9);
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_MEDIA,
> > -                   FORCEWAKE_MEDIA_GEN9, FORCEWAKE_ACK_MEDIA_GEN9);
> > +        __fw_domain_init(FW_DOMAIN_ID_RENDER,
> > +                 FORCEWAKE_RENDER_GEN9,
> > +                 FORCEWAKE_ACK_RENDER_GEN9);
> > +        __fw_domain_init(FW_DOMAIN_ID_BLITTER,
> > +                 FORCEWAKE_BLITTER_GEN9,
> > +                 FORCEWAKE_ACK_BLITTER_GEN9);
> > +        __fw_domain_init(FW_DOMAIN_ID_MEDIA,
> > +                 FORCEWAKE_MEDIA_GEN9,
> > +                 FORCEWAKE_ACK_MEDIA_GEN9);
> >       } else if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) {
> >           uncore->funcs.force_wake_get = fw_domains_get;
> >           uncore->funcs.force_wake_put = fw_domains_put;
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_RENDER,
> > -                   FORCEWAKE_VLV, FORCEWAKE_ACK_VLV);
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_MEDIA,
> > -                   FORCEWAKE_MEDIA_VLV, FORCEWAKE_ACK_MEDIA_VLV);
> > +        __fw_domain_init(FW_DOMAIN_ID_RENDER,
> > +                 FORCEWAKE_VLV, FORCEWAKE_ACK_VLV);
> > +        __fw_domain_init(FW_DOMAIN_ID_MEDIA,
> > +                 FORCEWAKE_MEDIA_VLV, FORCEWAKE_ACK_MEDIA_VLV);
> >       } else if (IS_HASWELL(i915) || IS_BROADWELL(i915)) {
> >           uncore->funcs.force_wake_get =
> >               fw_domains_get_with_thread_status;
> >           uncore->funcs.force_wake_put = fw_domains_put;
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_RENDER,
> > -                   FORCEWAKE_MT, FORCEWAKE_ACK_HSW);
> > +        __fw_domain_init(FW_DOMAIN_ID_RENDER,
> > +                 FORCEWAKE_MT, FORCEWAKE_ACK_HSW);
> >       } else if (IS_IVYBRIDGE(i915)) {
> >           u32 ecobus;
> > @@ -1437,8 +1455,8 @@ static void intel_uncore_fw_domains_init(struct 
> > intel_uncore *uncore)
> >           __raw_uncore_write32(uncore, FORCEWAKE, 0);
> >           __raw_posting_read(uncore, ECOBUS);
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_RENDER,
> > -                   FORCEWAKE_MT, FORCEWAKE_MT_ACK);
> > +        __fw_domain_init(FW_DOMAIN_ID_RENDER,
> > +                 FORCEWAKE_MT, FORCEWAKE_MT_ACK);
> >           spin_lock_irq(&uncore->lock);
> >           fw_domains_get_with_thread_status(uncore, FORCEWAKE_RENDER);
> > @@ -1449,19 +1467,27 @@ static void 
> > intel_uncore_fw_domains_init(struct intel_uncore *uncore)
> >           if (!(ecobus & FORCEWAKE_MT_ENABLE)) {
> >               DRM_INFO("No MT forcewake available on Ivybridge, this 
> > can result in issues\n");
> >               DRM_INFO("when using vblank-synced partial screen 
> > updates.\n");
> > -            fw_domain_init(uncore, FW_DOMAIN_ID_RENDER,
> > -                       FORCEWAKE, FORCEWAKE_ACK);
> > +            __fw_domain_init(FW_DOMAIN_ID_RENDER,
> > +                     FORCEWAKE, FORCEWAKE_ACK);
> >           }
> >       } else if (IS_GEN(i915, 6)) {
> >           uncore->funcs.force_wake_get =
> >               fw_domains_get_with_thread_status;
> >           uncore->funcs.force_wake_put = fw_domains_put;
> > -        fw_domain_init(uncore, FW_DOMAIN_ID_RENDER,
> > -                   FORCEWAKE, FORCEWAKE_ACK);
> > +        __fw_domain_init(FW_DOMAIN_ID_RENDER,
> > +                 FORCEWAKE, FORCEWAKE_ACK);
> >       }
> > +#undef __fw_domain_init
> > +
> >       /* All future platforms are expected to require complex power 
> > gating */
> >       WARN_ON(uncore->fw_domains == 0);
> > +
> > +    return 0;
> > +
> > +out_clean:
> > +    intel_uncore_fw_domains_fini(uncore);
> > +    return ret;
> >   }
> >   #define ASSIGN_FW_DOMAINS_TABLE(uncore, d) \
> > @@ -1562,13 +1588,17 @@ static void uncore_raw_init(struct 
> > intel_uncore *uncore)
> >       }
> >   }
> > -static void uncore_forcewake_init(struct intel_uncore *uncore)
> > +static int uncore_forcewake_init(struct intel_uncore *uncore)
> >   {
> >       struct drm_i915_private *i915 = uncore->i915;
> > +    int ret;
> >       GEM_BUG_ON(!intel_uncore_has_forcewake(uncore));
> > -    intel_uncore_fw_domains_init(uncore);
> > +    ret = intel_uncore_fw_domains_init(uncore);
> > +    if (ret)
> > +        return ret;
> > +
> >       forcewake_early_sanitize(uncore, 0);
> >       if (IS_GEN_RANGE(i915, 6, 7)) {
> > @@ -1601,6 +1631,8 @@ static void uncore_forcewake_init(struct 
> > intel_uncore *uncore)
> >       uncore->pmic_bus_access_nb.notifier_call = 
> > i915_pmic_bus_access_notifier;
> >       
> > iosf_mbi_register_pmic_bus_access_notifier(&uncore->pmic_bus_access_nb);
> > +
> > +    return 0;
> >   }
> >   int intel_uncore_init_mmio(struct intel_uncore *uncore)
> > @@ -1619,10 +1651,13 @@ int intel_uncore_init_mmio(struct intel_uncore 
> > *uncore)
> >       uncore->unclaimed_mmio_check = 1;
> > -    if (!intel_uncore_has_forcewake(uncore))
> > +    if (!intel_uncore_has_forcewake(uncore)) {
> >           uncore_raw_init(uncore);
> > -    else
> > -        uncore_forcewake_init(uncore);
> > +    } else {
> > +        ret = uncore_forcewake_init(uncore);
> > +        if (ret)
> > +            goto out_mmio_cleanup;
> > +    }
> >       /* make sure fw funcs are set if and only if we have fw*/
> >       GEM_BUG_ON(intel_uncore_has_forcewake(uncore) != 
> > !!uncore->funcs.force_wake_get);
> > @@ -1644,6 +1679,11 @@ int intel_uncore_init_mmio(struct intel_uncore 
> > *uncore)
> >           DRM_DEBUG("unclaimed mmio detected on uncore init, 
> > clearing\n");
> >       return 0;
> > +
> > +out_mmio_cleanup:
> > +    uncore_mmio_cleanup(uncore);
> > +
> > +    return ret;
> >   }
> >   /*
> > @@ -1689,6 +1729,7 @@ void intel_uncore_fini_mmio(struct intel_uncore 
> > *uncore)
> >           iosf_mbi_unregister_pmic_bus_access_notifier_unlocked(
> >               &uncore->pmic_bus_access_nb);
> >           intel_uncore_forcewake_reset(uncore);
> > +        intel_uncore_fw_domains_fini(uncore);
> >           iosf_mbi_punit_release();
> >       }
> > diff --git a/drivers/gpu/drm/i915/intel_uncore.h 
> > b/drivers/gpu/drm/i915/intel_uncore.h
> > index 879252735bba..bbff281b880d 100644
> > --- a/drivers/gpu/drm/i915/intel_uncore.h
> > +++ b/drivers/gpu/drm/i915/intel_uncore.h
> > @@ -126,6 +126,7 @@ struct intel_uncore {
> >       enum forcewake_domains fw_domains_saved; /* user domains saved 
> > for S3 */
> >       struct intel_uncore_forcewake_domain {
> > +        struct intel_uncore *uncore;
> >           enum forcewake_domain_id id;
> >           enum forcewake_domains mask;
> >           unsigned int wake_count;
> > @@ -133,7 +134,7 @@ struct intel_uncore {
> >           struct hrtimer timer;
> >           u32 __iomem *reg_set;
> >           u32 __iomem *reg_ack;
> > -    } fw_domain[FW_DOMAIN_ID_COUNT];
> > +    } *fw_domain[FW_DOMAIN_ID_COUNT];
>
> The array itself could also be made dynamic, no? To much hassle for very 
> little gain?

With the current approach we don't know how many domains we have in 
advance, so we'd have to either add a table for that or realloc the 
array as we increase the number of domains. Both options don't seem 
worth the hassle IMO.

Thanks,
Daniele

> >       struct {
> >           unsigned int count;
> > @@ -147,18 +148,12 @@ struct intel_uncore {
> >   /* Iterate over initialised fw domains */
> >   #define for_each_fw_domain_masked(domain__, mask__, uncore__, tmp__) \
> > -    for (tmp__ = (mask__); \
> > -         tmp__ ? (domain__ = 
> > &(uncore__)->fw_domain[__mask_next_bit(tmp__)]), 1 : 0;)
> > +    for (tmp__ = (mask__); tmp__ ;) \
> > +        for_each_if(domain__ = 
> > (uncore__)->fw_domain[__mask_next_bit(tmp__)])
> >   #define for_each_fw_domain(domain__, uncore__, tmp__) \
> >       for_each_fw_domain_masked(domain__, (uncore__)->fw_domains, 
> > uncore__, tmp__)
> > -static inline struct intel_uncore *
> > -forcewake_domain_to_uncore(const struct intel_uncore_forcewake_domain 
> > *d)
> > -{
> > -    return container_of(d, struct intel_uncore, fw_domain[d->id]);
> > -}
> > -
> >   static inline bool
> >   intel_uncore_has_forcewake(const struct intel_uncore *uncore)
> >   {
>
> Regards,
>
> Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx