Re: [PATCH v4 5/5] iommu/s390: allow larger region tables

[Matthew Rosato <mjrosato@linux.ibm.com>]

On 4/1/25 5:11 AM, Heiko Carstens wrote:
> On Mon, Mar 31, 2025 at 04:21:59PM -0400, Matthew Rosato wrote:
>> Extend the aperture calculation to consider sizes beyond the maximum
>> size of a region third table.  Attempt to always use the smallest
>> table size possible to avoid unnecessary extra steps during translation.
>> Update reserved region calculations to use the appropriate table size.
>>
>> Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
>> ---
>>  arch/s390/include/asm/pci_dma.h |  1 +
>>  drivers/iommu/s390-iommu.c      | 71 ++++++++++++++++++++++++---------
>>  2 files changed, 53 insertions(+), 19 deletions(-)
> 
> ...
> 
>> +	spin_lock_irqsave(&zdev->dom_lock, flags);
>> +	if (zdev->s390_domain->type == IOMMU_DOMAIN_BLOCKED ||
>> +	    zdev->s390_domain->type == IOMMU_DOMAIN_IDENTITY)
>> +		goto out;
>> +
>> +	s390_domain = to_s390_domain(zdev->s390_domain);
>> +	if (zdev->end_dma < max_tbl_size(s390_domain)) {
>> +		end_resv = max_tbl_size(s390_domain) - zdev->end_dma;
>> +		region = iommu_alloc_resv_region(zdev->end_dma + 1, end_resv,
>> +						 0, IOMMU_RESV_RESERVED,
>> +						 GFP_KERNEL);
> 
> GFP_KERNEL allocation while holding a spinlock is not correct.

Looks like I broke this in v3 -- Thanks for catching it.  There's no need to hold this spinlock for as long as we do here, including over the allocation itself; I will re-arrange so we can safely use GFP_KERNEL here.