Re: [PATCH] iommu/amd: Do not clear unobserved log overflow status bits

From: Ankit Soni

Date: Mon Nov 16 2026 - 06:42:38 EST


On Mon, Nov 16, 2026 at 11:39:02AM +0800, Guanghui Feng wrote:
> amd_iommu_handle_irq() clears the interrupt status with an unconditional
> writel(mask, MMIO_STATUS_OFFSET), where mask is int_mask | overflow_mask.
> The log overflow status bits (GALOverflow, EventOverflow, PprOverflow)
> are RW1C and, per the AMD IOMMU spec, transitioning them from 0 to 1 does
> not itself generate an interrupt; they can only be observed by an IRQ
> round triggered by the corresponding log-interrupt bit.
>
> If hardware sets an overflow bit after the status register is sampled at
> the top of the loop but before the unconditional writel(), that write
> clears the freshly-set bit while the stale sampled status does not
> contain it, so overflow_handler() is skipped for an overflow that really
> happened. For the GA log this is fatal: amd_iommu_restart_ga_log() is the
> only recovery path, and per spec section 2.7.4 logging stays halted until
> GALogEn is toggled. A bare W1C of GALOverflow is not part of the
> documented recovery, so the GA log stays halted, no further entries are
> written, GAInt never re-asserts and the GA log interrupt line goes
> permanently silent, losing all AVIC non-running vCPU wakeups on that
> IOMMU.
>
> Clear only the bits actually observed this round by writing status & mask
> instead of mask. An overflow bit set concurrently is then left asserted
> and is picked up by the existing ERBT1312 status read-back at the bottom
> of the loop, which re-enters the loop and runs overflow_handler().
>
> Fixes: 2379f3485239 ("iommu/amd: Refactor IOMMU interrupt handling logic for Event, PPR, and GA logs")

This commit only moved the existing clear into amd_iommu_handle_irq(). The
unconditional write already cleared overflow bits before that refactor.
The first commit that put an overflow bit into that write is:

Fixes: 5ce97f4ec5e0 ("iommu/amd: Recover from event log overflow")

The GA and PPR overflow bits were added later, by af47b0a24058
("iommu/amd: Handle GALog overflows") and 274c2218b8b2 ("iommu/amd: Handle
PPR log overflow").

Otherwise, patch looks good to me.

Reviewed-by: Ankit Soni <Ankit.Soni@xxxxxxx>

-Ankit

> Signed-off-by: Guanghui Feng <guanghuifeng@xxxxxxxxxxxxxxxxx>
> ---
> drivers/iommu/amd/iommu.c | 13 +++++++++++--
> 1 file changed, 11 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
> index d0d55c93a480..39254671b924 100644
> --- a/drivers/iommu/amd/iommu.c
> +++ b/drivers/iommu/amd/iommu.c
> @@ -1168,8 +1168,17 @@ static void amd_iommu_handle_irq(void *data, const char *evt_type,
> u32 mask = int_mask | overflow_mask;
>
> while (status & mask) {
> - /* Enable interrupt sources again */
> - writel(mask, iommu->mmio_base + MMIO_STATUS_OFFSET);
> + /*
> + * Clear only the status bits observed in this round. The log
> + * overflow bits (GALOverflow/EventOverflow/PprOverflow) are RW1C
> + * and do not raise an interrupt on their own; clearing them
> + * unconditionally here would drop an overflow that hardware set
> + * between the readl() above and this write, without ever running
> + * overflow_handler(). Leaving such a bit asserted lets the
> + * existing ERBT1312 read-back below re-observe it and re-enter
> + * the loop so overflow_handler() runs.
> + */
> + writel(status & mask, iommu->mmio_base + MMIO_STATUS_OFFSET);
>
> if (int_handler) {
> pr_devel("Processing IOMMU (ivhd%d) %s Log\n",
> --
> 2.43.7
>