Re: [PATCH v2 5/5] iommu/amd: Use 128-bit cmpxchg in set_dte_irq_entry()

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Aug 29, 2024 at 06:07:26PM +0000, Suravee Suthikulpanit wrote:
> Interrupt-remapping-related fields are in the top 128-bit of the Device
> Table Entry (DTE), which should be updated using 128-bit write based on the
> AMD I/O Virtualization Techonology (IOMMU) Specification.
> 
> Therefore, modify set_dte_irq_entry() to use 128-bit cmpxchg. Also, use
> struct dev_data->dte_sem to synchronize DTE access.
> 
> Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
> Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
> ---
>  drivers/iommu/amd/iommu.c | 25 ++++++++++++++++---------
>  1 file changed, 16 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
> index a24986c2478b..4eb53bd40487 100644
> --- a/drivers/iommu/amd/iommu.c
> +++ b/drivers/iommu/amd/iommu.c
> @@ -3020,17 +3020,24 @@ static void iommu_flush_irt_and_complete(struct amd_iommu *iommu, u16 devid)
>  static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid,
>  			      struct irq_remap_table *table)
>  {
> -	u64 dte;
> -	struct dev_table_entry *dev_table = get_dev_table(iommu);
> +	u128 new, old;
> +	struct dev_table_entry *dte = &get_dev_table(iommu)[devid];
> +	struct iommu_dev_data *dev_data = search_dev_data(iommu, devid);
> +
> +	if (dev_data)
> +		down_write(&dev_data->dte_sem);
> +
> +	old = new = dte->data128[1];
> +	new &= ~DTE_IRQ_PHYS_ADDR_MASK;
> +	new |= iommu_virt_to_phys(table->table);
> +	new |= DTE_IRQ_REMAP_INTCTL;
> +	new |= DTE_INTTABLEN;
> +	new |= DTE_IRQ_REMAP_ENABLE;
>  
> -	dte	= dev_table[devid].data[2];
> -	dte	&= ~DTE_IRQ_PHYS_ADDR_MASK;
> -	dte	|= iommu_virt_to_phys(table->table);
> -	dte	|= DTE_IRQ_REMAP_INTCTL;
> -	dte	|= DTE_INTTABLEN;
> -	dte	|= DTE_IRQ_REMAP_ENABLE;
> +	WARN_ON(!try_cmpxchg128(&dte->data128[1], &old, new));

This probably doesn't need cmpxchg because it is only touching 64 bit
quanta [2], just a lock is good enough and avoids the "what to do if
cmpxchg fails" question.

> -	dev_table[devid].data[2] = dte;

But this should be 

  WRITE_ONCE(dev_table[devid].data[2], dte);

Beaucse it is writing to memory read back by HW.

Required for all the DTE touches everywhere.

Jason