Re: [PATCH v3 14/14] irqchip/gic-v5: Invalidate the L1 DTE when allocating an L2 device table

From: Lorenzo Pieralisi

Date: Thu Nov 19 2026 - 05:14:04 EST


On Wed, Nov 18, 2026 at 05:35:43PM +0000, Sascha Bischoff wrote:
> Allocating an L2 device table makes its parent L1 DTE valid, but the
> single-device invalidation performed during device registration does
> not invalidate that L1 DTE. An ITS that cached the L1 DTE while it was
> invalid can therefore continue dropping events for DeviceIDs in its
> range.
>
> Invalidate the L1 DTE and wait for completion immediately after
> publishing the new L2 table. This also covers a registration that
> subsequently fails, leaving the L2 table allocated and its parent L1
> DTE valid. Keep single-device invalidations for device registration
> and teardown when updating individual L2 DTEs.
>
> Fixes: 57d72196dfc85 ("irqchip/gic-v5: Add GICv5 ITS support")
> Reported-by: Fuad Tabba <fuad.tabba@xxxxxxxxx>
> Signed-off-by: Sascha Bischoff <sascha.bischoff@xxxxxxx>
> Cc: stable@xxxxxxxxxxxxxxx
> Link: https://lore.kernel.org/r/20261112145814.853424-1-fuad.tabba@xxxxxxxxx
> ---
> drivers/irqchip/irq-gic-v5-its.c | 26 ++++++++++++++++++++------
> 1 file changed, 20 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/irqchip/irq-gic-v5-its.c b/drivers/irqchip/irq-gic-v5-its.c
> index 7fff2db9ae08e..a05417bdd3a19 100644
> --- a/drivers/irqchip/irq-gic-v5-its.c
> +++ b/drivers/irqchip/irq-gic-v5-its.c
> @@ -365,17 +365,17 @@ static __le64 *gicv5_its_device_get_itte_ref(struct gicv5_its_dev *its_dev,
> return &l2_itt[l2_idx];
> }
>
> -static int gicv5_its_device_cache_inv(struct gicv5_its_chip_data *its,
> - struct gicv5_its_dev *its_dev)
> +static int __gicv5_its_device_cache_inv(struct gicv5_its_chip_data *its,
> + u32 device_id, u8 event_id_bits, bool l1)
> {
> u32 devicer;
> u64 didr;
>
> - didr = FIELD_PREP(GICV5_ITS_DIDR_DEVICEID, its_dev->device_id);
> + didr = FIELD_PREP(GICV5_ITS_DIDR_DEVICEID, device_id);
> devicer = FIELD_PREP(GICV5_ITS_INV_DEVICER_I, 0x1) |
> FIELD_PREP(GICV5_ITS_INV_DEVICER_EVENTID_BITS,
> - its_dev->itt_cfg.event_id_bits) |
> - FIELD_PREP(GICV5_ITS_INV_DEVICER_L1, 0x0);
> + event_id_bits) |
> + FIELD_PREP(GICV5_ITS_INV_DEVICER_L1, l1);
>
> guard(raw_spinlock_irqsave)(&its->inv_lock);
>
> @@ -385,6 +385,19 @@ static int gicv5_its_device_cache_inv(struct gicv5_its_chip_data *its,
> return gicv5_its_cache_sync(its);
> }
>
> +static int gicv5_its_device_cache_inv(struct gicv5_its_chip_data *its,
> + struct gicv5_its_dev *its_dev)
> +{
> + return __gicv5_its_device_cache_inv(its, its_dev->device_id,
> + its_dev->itt_cfg.event_id_bits, false);
> +}
> +
> +static int gicv5_its_device_l1_cache_inv(struct gicv5_its_chip_data *its,
> + u32 device_id)
> +{
> + return __gicv5_its_device_cache_inv(its, device_id, 0, true);

I would highlight why event_id_bits is set to 0 (actually it is irrelevant,
let's explicitly say so).

> +}
> +
> /*
> * Allocate a level 2 device table entry, update L1 parent to reference it.
> * Only used for 2-level device tables, and it is called on demand.
> @@ -424,7 +437,8 @@ static int gicv5_its_alloc_l2_devtab(struct gicv5_its_chip_data *its,
> FIELD_PREP(GICV5_DTL1E_VALID, 0x1);
> its_write_table_entry(its, &l1devtab[l1_index], l1dte);
>
> - return 0;
> + /* Invalidate the cached L1 DTE before using the new L2 table. */

Nit: it is not a given that it is cached.

> + return gicv5_its_device_l1_cache_inv(its, l1_index << l2_bits);

Nit: this l1/l2/deviceid bits/index dance is a bit clumsy (I mean, in
current code, without taking into account this patch) but we can rework it
in a follow-up patch to keep this one as simple as possible.

Acked-by: Lorenzo Pieralisi <lpieralisi@xxxxxxxxxx>

> }
>
> static __le64 *gicv5_its_devtab_get_dte_ref(struct gicv5_its_chip_data *its,
>
> --
> 2.34.1
>