On Fri, 19 Aug 2022, Daniele Ceraolo Spurio <daniele.ceraolospurio@xxxxxxxxx> wrote:
> Given that HuC load is delayed on DG2, this patch adds support for a fence
> that can be used to wait for load completion. No waiters are added in this
> patch (they're coming up in the next one), to keep the focus of the
> patch on the tracking logic.
>
> The full HuC loading flow on boot DG2 is as follows:
> 1) i915 exports the GSC as an aux device;
> 2) the mei-gsc driver is loaded on the aux device;
> 3) the mei-pxp component is loaded;
> 4) mei-pxp calls back into i915 and we load the HuC.
>
> Between steps 1 and 2 there can be several seconds of gap, mainly due to
> the kernel doing other work during the boot.
> The resume flow is slightly different, because we don't need to
> re-expose or re-probe the aux device, so we go directly to step 3 once
> i915 and mei-gsc have completed their resume flow.
>
> Here's an example of the boot timing, captured with some logs added to
> i915:
>
> [ 17.908307] [drm] adding GSC device
> [ 17.915717] [drm] i915 probe done
> [ 22.282917] [drm] mei-gsc bound
> [ 22.938153] [drm] HuC authenticated
>
> Also to note is that if something goes wrong during GSC HW init the
> mei-gsc driver will still bind, but steps 3 and 4 will not happen.
>
> The status tracking is done by registering a bus_notifier to receive a
> callback when the mei-gsc driver binds, with a large enough timeout to
> account for delays. Once mei-gsc is bound, we switch to a smaller
> timeout to wait for the mei-pxp component to load.
> The fence is signalled on HuC load complete or if anything goes wrong in
> any of the tracking steps. Timeout are enforced via hrtimer callbacks.
>
> Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@xxxxxxxxx>
> Reviewed-by: Alan Previn <alan.previn.teres.alexis@xxxxxxxxx>
> ---
> drivers/gpu/drm/i915/gt/intel_gsc.c | 22 ++-
> drivers/gpu/drm/i915/gt/uc/intel_huc.c | 198 +++++++++++++++++++++++++
> drivers/gpu/drm/i915/gt/uc/intel_huc.h | 19 +++
> 3 files changed, 236 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/gt/intel_gsc.c b/drivers/gpu/drm/i915/gt/intel_gsc.c
> index 162bea57fbb5..bc9db728cdc8 100644
> --- a/drivers/gpu/drm/i915/gt/intel_gsc.c
> +++ b/drivers/gpu/drm/i915/gt/intel_gsc.c
> @@ -154,8 +154,14 @@ static void gsc_destroy_one(struct drm_i915_private *i915,
> struct intel_gsc_intf *intf = &gsc->intf[intf_id];
>
> if (intf->adev) {
> - auxiliary_device_delete(&intf->adev->aux_dev);
> - auxiliary_device_uninit(&intf->adev->aux_dev);
> + struct auxiliary_device *aux_dev = &intf->adev->aux_dev;
> +
> + if (intf_id == 0)
> + intel_huc_unregister_gsc_notifier(&gsc_to_gt(gsc)->uc.huc,
> + aux_dev->dev.bus);
> +
> + auxiliary_device_delete(aux_dev);
> + auxiliary_device_uninit(aux_dev);
> intf->adev = NULL;
> }
>
> @@ -254,14 +260,24 @@ static void gsc_init_one(struct drm_i915_private *i915, struct intel_gsc *gsc,
> goto fail;
> }
>
> + intf->adev = adev; /* needed by the notifier */
> +
> + if (intf_id == 0)
> + intel_huc_register_gsc_notifier(&gsc_to_gt(gsc)->uc.huc,
> + aux_dev->dev.bus);
> +
> ret = auxiliary_device_add(aux_dev);
> if (ret < 0) {
> drm_err(&i915->drm, "gsc aux add failed %d\n", ret);
> + if (intf_id == 0)
> + intel_huc_unregister_gsc_notifier(&gsc_to_gt(gsc)->uc.huc,
> + aux_dev->dev.bus);
> + intf->adev = NULL;
> +
> /* adev will be freed with the put_device() and .release sequence */
> auxiliary_device_uninit(aux_dev);
> goto fail;
> }
> - intf->adev = adev;
>
> return;
> fail:
> diff --git a/drivers/gpu/drm/i915/gt/uc/intel_huc.c b/drivers/gpu/drm/i915/gt/uc/intel_huc.c
> index 40217fb69824..9a97b8cc90c7 100644
> --- a/drivers/gpu/drm/i915/gt/uc/intel_huc.c
> +++ b/drivers/gpu/drm/i915/gt/uc/intel_huc.c
> @@ -10,6 +10,8 @@
> #include "intel_huc.h"
> #include "i915_drv.h"
>
> +#include <linux/mei_aux.h>
> +
> /**
> * DOC: HuC
> *
> @@ -42,6 +44,164 @@
> * HuC-specific commands.
> */
>
> +/*
> + * MEI-GSC load is an async process. The probing of the exposed aux device
> + * (see intel_gsc.c) usually happens a few seconds after i915 probe, depending
> + * on when the kernel schedules it. Unless something goes terribly wrong, we're
> + * guaranteed for this to happen during boot, so the big timeout is a safety net
> + * that we never expect to need.
> + * MEI-PXP + HuC load usually takes ~300ms, but if the GSC needs to be resumed
> + * and/or reset, this can take longer.
> + */
> +#define GSC_INIT_TIMEOUT_MS 10000
> +#define PXP_INIT_TIMEOUT_MS 2000
> +
> +static int sw_fence_dummy_notify(struct i915_sw_fence *sf,
> + enum i915_sw_fence_notify state)
> +{
> + return NOTIFY_DONE;
> +}
> +
> +static void __delayed_huc_load_complete(struct intel_huc *huc)
> +{
> + if (!i915_sw_fence_done(&huc->delayed_load.fence))
> + i915_sw_fence_complete(&huc->delayed_load.fence);
> +}
> +
> +static void delayed_huc_load_complete(struct intel_huc *huc)
> +{
> + hrtimer_cancel(&huc->delayed_load.timer);
> + __delayed_huc_load_complete(huc);
> +}
> +
> +static void __gsc_init_error(struct intel_huc *huc)
> +{
> + huc->delayed_load.status = INTEL_HUC_DELAYED_LOAD_ERROR;
> + __delayed_huc_load_complete(huc);
> +}
> +
> +static void gsc_init_error(struct intel_huc *huc)
> +{
> + hrtimer_cancel(&huc->delayed_load.timer);
> + __gsc_init_error(huc);
> +}
> +
> +static void gsc_init_done(struct intel_huc *huc)
> +{
> + hrtimer_cancel(&huc->delayed_load.timer);
> +
> + /* MEI-GSC init is done, now we wait for MEI-PXP to bind */
> + huc->delayed_load.status = INTEL_HUC_WAITING_ON_PXP;
> + if (!i915_sw_fence_done(&huc->delayed_load.fence))
> + hrtimer_start(&huc->delayed_load.timer,
> + ms_to_ktime(PXP_INIT_TIMEOUT_MS),
> + HRTIMER_MODE_REL);
> +}
> +
> +static enum hrtimer_restart huc_delayed_load_timer_callback(struct hrtimer *hrtimer)
> +{
> + struct intel_huc *huc = container_of(hrtimer, struct intel_huc, delayed_load.timer);
> +
> + if (!intel_huc_is_authenticated(huc)) {
> + drm_err(&huc_to_gt(huc)->i915->drm,
> + "timed out waiting for GSC init to load HuC\n");
> +
> + __gsc_init_error(huc);
> + }
> +
> + return HRTIMER_NORESTART;
> +}
> +
> +static void huc_delayed_load_start(struct intel_huc *huc)
> +{
> + ktime_t delay;
> +
> + GEM_BUG_ON(intel_huc_is_authenticated(huc));
> +
> + /*
> + * On resume we don't have to wait for MEI-GSC to be re-probed, but we
> + * do need to wait for MEI-PXP to reset & re-bind
> + */
> + switch (huc->delayed_load.status) {
> + case INTEL_HUC_WAITING_ON_GSC:
> + delay = ms_to_ktime(GSC_INIT_TIMEOUT_MS);
> + break;
> + case INTEL_HUC_WAITING_ON_PXP:
> + delay = ms_to_ktime(PXP_INIT_TIMEOUT_MS);
> + break;
> + default:
> + gsc_init_error(huc);
> + return;
> + }
> +
> + /*
> + * This fence is always complete unless we're waiting for the
> + * GSC device to come up to load the HuC. We arm the fence here
> + * and complete it when we confirm that the HuC is loaded from
> + * the PXP bind callback.
> + */
> + GEM_BUG_ON(!i915_sw_fence_done(&huc->delayed_load.fence));
> + i915_sw_fence_fini(&huc->delayed_load.fence);
> + i915_sw_fence_reinit(&huc->delayed_load.fence);
> + i915_sw_fence_await(&huc->delayed_load.fence);
> + i915_sw_fence_commit(&huc->delayed_load.fence);
> +
> + hrtimer_start(&huc->delayed_load.timer, delay, HRTIMER_MODE_REL);
> +}
> +
> +static int gsc_notifier(struct notifier_block *nb, unsigned long action, void *data)
> +{
> + struct device *dev = data;
> + struct intel_huc *huc = container_of(nb, struct intel_huc, delayed_load.nb);
> + struct intel_gsc_intf *intf = &huc_to_gt(huc)->gsc.intf[0];
> +
> + if (!intf->adev || (&intf->adev->aux_dev.dev != dev))
> + return 0;
> +
> + switch (action) {
> + case BUS_NOTIFY_BOUND_DRIVER: /* mei driver bound to aux device */
> + gsc_init_done(huc);
> + break;
> +
> + case BUS_NOTIFY_DRIVER_NOT_BOUND: /* mei driver fails to be bound */
> + case BUS_NOTIFY_UNBIND_DRIVER: /* mei driver about to be unbound */
> + drm_info(&huc_to_gt(huc)->i915->drm,
> + "mei driver not bound, disabling HuC load\n");
> + gsc_init_error(huc);
> + break;
> + }
> +
> + return 0;
> +}
> +
> +void intel_huc_register_gsc_notifier(struct intel_huc *huc, struct bus_type *bus)
> +{
> + int ret;
> +
> + if (!intel_huc_is_loaded_by_gsc(huc))
> + return;
> +
> + huc->delayed_load.nb.notifier_call = gsc_notifier;
> + ret = bus_register_notifier(bus, &huc->delayed_load.nb);
> + if (ret) {
> + drm_err(&huc_to_gt(huc)->i915->drm,
> + "failed to register GSC notifier\n");
> + huc->delayed_load.nb.notifier_call = NULL;
> + gsc_init_error(huc);
> + }
> +}
> +
> +void intel_huc_unregister_gsc_notifier(struct intel_huc *huc, struct bus_type *bus)
> +{
> + if (!huc->delayed_load.nb.notifier_call)
> + return;
> +
> + delayed_huc_load_complete(huc);
> +
> + bus_unregister_notifier(bus, &huc->delayed_load.nb);
> + huc->delayed_load.nb.notifier_call = NULL;
> +}
> +
> void intel_huc_init_early(struct intel_huc *huc)
> {
> struct drm_i915_private *i915 = huc_to_gt(huc)->i915;
> @@ -57,6 +217,17 @@ void intel_huc_init_early(struct intel_huc *huc)
> huc->status.mask = HUC_FW_VERIFIED;
> huc->status.value = HUC_FW_VERIFIED;
> }
> +
> + /*
> + * Initialize fence to be complete as this is expected to be complete
> + * unless there is a delayed HuC reload in progress.
> + */
> + i915_sw_fence_init(&huc->delayed_load.fence,
> + sw_fence_dummy_notify);
> + i915_sw_fence_commit(&huc->delayed_load.fence);
> +
> + hrtimer_init(&huc->delayed_load.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
> + huc->delayed_load.timer.function = huc_delayed_load_timer_callback;
> }
>
> #define HUC_LOAD_MODE_STRING(x) (x ? "GSC" : "legacy")
> @@ -122,9 +293,25 @@ void intel_huc_fini(struct intel_huc *huc)
> if (!intel_uc_fw_is_loadable(&huc->fw))
> return;
>
> + delayed_huc_load_complete(huc);
> +
> + i915_sw_fence_fini(&huc->delayed_load.fence);
> intel_uc_fw_fini(&huc->fw);
> }
>
> +void intel_huc_suspend(struct intel_huc *huc)
> +{
> + if (!intel_uc_fw_is_loadable(&huc->fw))
> + return;
> +
> + /*
> + * in the unlikely case that we're suspending before the GSC has
> + * completed its loading sequence, just stop waiting. We'll restart
> + * on resume.
> + */
> + delayed_huc_load_complete(huc);
> +}
> +
> int intel_huc_wait_for_auth_complete(struct intel_huc *huc)
> {
> struct intel_gt *gt = huc_to_gt(huc);
> @@ -136,6 +323,9 @@ int intel_huc_wait_for_auth_complete(struct intel_huc *huc)
> huc->status.value,
> 2, 50, NULL);
>
> + /* mark the load process as complete even if the wait failed */
> + delayed_huc_load_complete(huc);
> +
> if (ret) {
> drm_err(>->i915->drm,"HuC: Firmware not verified %d\n", ret);
> intel_uc_fw_change_status(&huc->fw, INTEL_UC_FIRMWARE_LOAD_FAIL);
> @@ -239,6 +429,12 @@ int intel_huc_check_status(struct intel_huc *huc)
> return intel_huc_is_authenticated(huc);
> }
>
> +static bool huc_has_delayed_load(struct intel_huc *huc)
> +{
> + return intel_huc_is_loaded_by_gsc(huc) &&
> + (huc->delayed_load.status != INTEL_HUC_DELAYED_LOAD_ERROR);
> +}
> +
> void intel_huc_update_auth_status(struct intel_huc *huc)
> {
> if (!intel_uc_fw_is_loadable(&huc->fw))
> @@ -247,6 +443,8 @@ void intel_huc_update_auth_status(struct intel_huc *huc)
> if (intel_huc_is_authenticated(huc))
> intel_uc_fw_change_status(&huc->fw,
> INTEL_UC_FIRMWARE_RUNNING);
> + else if (huc_has_delayed_load(huc))
> + huc_delayed_load_start(huc);
> }
>
> /**
> diff --git a/drivers/gpu/drm/i915/gt/uc/intel_huc.h b/drivers/gpu/drm/i915/gt/uc/intel_huc.h
> index 51f9d96a3ca3..49374f306a7f 100644
> --- a/drivers/gpu/drm/i915/gt/uc/intel_huc.h
> +++ b/drivers/gpu/drm/i915/gt/uc/intel_huc.h
> @@ -10,6 +10,14 @@
> #include "intel_uc_fw.h"
> #include "intel_huc_fw.h"
>
> +#include <linux/device/bus.h>
struct bus_type;
is enough.
BR,
Jani.
> +
> +enum intel_huc_delayed_load_status {
> + INTEL_HUC_WAITING_ON_GSC = 0,
> + INTEL_HUC_WAITING_ON_PXP,
> + INTEL_HUC_DELAYED_LOAD_ERROR,
> +};
> +
> struct intel_huc {
> /* Generic uC firmware management */
> struct intel_uc_fw fw;
> @@ -20,17 +28,28 @@ struct intel_huc {
> u32 mask;
> u32 value;
> } status;
> +
> + struct {
> + struct i915_sw_fence fence;
> + struct hrtimer timer;
> + struct notifier_block nb;
> + enum intel_huc_delayed_load_status status;
> + } delayed_load;
> };
>
> void intel_huc_init_early(struct intel_huc *huc);
> int intel_huc_init(struct intel_huc *huc);
> void intel_huc_fini(struct intel_huc *huc);
> +void intel_huc_suspend(struct intel_huc *huc);
> int intel_huc_auth(struct intel_huc *huc);
> int intel_huc_wait_for_auth_complete(struct intel_huc *huc);
> int intel_huc_check_status(struct intel_huc *huc);
> void intel_huc_update_auth_status(struct intel_huc *huc);
> bool intel_huc_is_authenticated(struct intel_huc *huc);
>
> +void intel_huc_register_gsc_notifier(struct intel_huc *huc, struct bus_type *bus);
> +void intel_huc_unregister_gsc_notifier(struct intel_huc *huc, struct bus_type *bus);
> +
> static inline int intel_huc_sanitize(struct intel_huc *huc)
> {
> intel_uc_fw_sanitize(&huc->fw);
--
Jani Nikula, Intel Open Source Graphics Center
