[PATCH v2 0/3] iommu/s390: add support for IOMMU passthrough

[PATCH v2 0/3] iommu/s390: add support for IOMMU passthrough

[Matthew Rosato]

This series introduces the ability for certain devices on s390 to bypass
a layer of IOMMU via the iommu.passthrough=1 option.  In order to enable
this, the concept of an identity domain is added to s390-iommu.  On s390,
IOMMU passthrough is only allowed if indicated via a special bit in s390
CLP data for the associated device group, otherwise we must fall back to
dma-iommu.

Changes for v2:
- Remove ARCH_HAS_PHYS_TO_DMA, use bus_dma_region
- Remove use of def_domain_type, use 1 of 2 ops chosen at init

Matthew Rosato (3):
  s390/pci: check for relaxed translation capability
  s390/pci: store DMA offset in bus_dma_region
  iommu/s390: implement iommu passthrough via identity domain

 arch/s390/include/asm/pci.h     |  2 +-
 arch/s390/include/asm/pci_clp.h |  4 +-
 arch/s390/pci/pci.c             |  6 ++-
 arch/s390/pci/pci_bus.c         | 18 +++++++
 arch/s390/pci/pci_clp.c         |  1 +
 drivers/iommu/s390-iommu.c      | 95 +++++++++++++++++++++++++--------
 6 files changed, 99 insertions(+), 27 deletions(-)

-- 
2.47.1

[PATCH v2 1/3] s390/pci: check for relaxed translation capability

[Matthew Rosato]

For each zdev, record whether or not CLP indicates relaxed translation
capability for the associated device group.

Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
---
 arch/s390/include/asm/pci.h     | 2 +-
 arch/s390/include/asm/pci_clp.h | 4 +++-
 arch/s390/pci/pci_clp.c         | 1 +
 3 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/arch/s390/include/asm/pci.h b/arch/s390/include/asm/pci.h
index 474e1f8d1d3c..8fe4c7a72c0b 100644
--- a/arch/s390/include/asm/pci.h
+++ b/arch/s390/include/asm/pci.h
@@ -144,7 +144,7 @@ struct zpci_dev {
 	u8		util_str_avail	: 1;
 	u8		irqs_registered	: 1;
 	u8		tid_avail	: 1;
-	u8		reserved	: 1;
+	u8		rtr_avail	: 1; /* Relaxed translation allowed */
 	unsigned int	devfn;		/* DEVFN part of the RID*/
 
 	u8 pfip[CLP_PFIP_NR_SEGMENTS];	/* pci function internal path */
diff --git a/arch/s390/include/asm/pci_clp.h b/arch/s390/include/asm/pci_clp.h
index 3fff2f7095c8..7ebff39c84b3 100644
--- a/arch/s390/include/asm/pci_clp.h
+++ b/arch/s390/include/asm/pci_clp.h
@@ -156,7 +156,9 @@ struct clp_rsp_query_pci_grp {
 	u16			:  4;
 	u16 noi			: 12;	/* number of interrupts */
 	u8 version;
-	u8			:  6;
+	u8			:  2;
+	u8 rtr			:  1;	/* Relaxed translation requirement */
+	u8			:  3;
 	u8 frame		:  1;
 	u8 refresh		:  1;	/* TLB refresh mode */
 	u16			:  3;
diff --git a/arch/s390/pci/pci_clp.c b/arch/s390/pci/pci_clp.c
index 14bf7e8d06b7..27248686e588 100644
--- a/arch/s390/pci/pci_clp.c
+++ b/arch/s390/pci/pci_clp.c
@@ -112,6 +112,7 @@ static void clp_store_query_pci_fngrp(struct zpci_dev *zdev,
 	zdev->version = response->version;
 	zdev->maxstbl = response->maxstbl;
 	zdev->dtsm = response->dtsm;
+	zdev->rtr_avail = response->rtr;
 
 	switch (response->version) {
 	case 1:
-- 
2.47.1

[PATCH v2 2/3] s390/pci: store DMA offset in bus_dma_region

[Matthew Rosato]

PCI devices on s390 have a DMA offset that is reported via CLP.  In
preparation for allowing identity domains, setup the bus_dma_region
for all PCI devices using the reported CLP value.

Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
---
 arch/s390/pci/pci_bus.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/arch/s390/pci/pci_bus.c b/arch/s390/pci/pci_bus.c
index d5ace00d10f0..14527687d0f2 100644
--- a/arch/s390/pci/pci_bus.c
+++ b/arch/s390/pci/pci_bus.c
@@ -19,6 +19,7 @@
 #include <linux/jump_label.h>
 #include <linux/pci.h>
 #include <linux/printk.h>
+#include <linux/dma-direct.h>
 
 #include <asm/pci_clp.h>
 #include <asm/pci_dma.h>
@@ -284,10 +285,27 @@ static struct zpci_bus *zpci_bus_alloc(int topo, bool topo_is_tid)
 	return zbus;
 }
 
+static void pci_dma_range_setup(struct pci_dev *pdev)
+{
+	struct zpci_dev *zdev = to_zpci(pdev);
+	struct bus_dma_region *map;
+
+	map = kzalloc(sizeof(*map), GFP_KERNEL);
+	if (!map)
+		return;
+
+	map->cpu_start = 0;
+	map->dma_start = PAGE_ALIGN(zdev->start_dma);
+	map->size = (u64)virt_to_phys(high_memory);
+	pdev->dev.dma_range_map = map;
+}
+
 void pcibios_bus_add_device(struct pci_dev *pdev)
 {
 	struct zpci_dev *zdev = to_zpci(pdev);
 
+	pci_dma_range_setup(pdev);
+
 	/*
 	 * With pdev->no_vf_scan the common PCI probing code does not
 	 * perform PF/VF linking.
-- 
2.47.1

[PATCH v2 3/3] iommu/s390: implement iommu passthrough via identity domain

[Matthew Rosato]

Enabled via the kernel command-line 'iommu.passthrough=1' option.

Introduce the concept of identity domains to s390-iommu, which relies on
the bus_dma_region to offset identity mappings to the start of the DMA
aperture advertized by CLP.

Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
---
 arch/s390/pci/pci.c        |  6 ++-
 drivers/iommu/s390-iommu.c | 95 +++++++++++++++++++++++++++++---------
 2 files changed, 76 insertions(+), 25 deletions(-)

diff --git a/arch/s390/pci/pci.c b/arch/s390/pci/pci.c
index 88f72745fa59..758b23331754 100644
--- a/arch/s390/pci/pci.c
+++ b/arch/s390/pci/pci.c
@@ -124,14 +124,16 @@ int zpci_register_ioat(struct zpci_dev *zdev, u8 dmaas,
 	struct zpci_fib fib = {0};
 	u8 cc;
 
-	WARN_ON_ONCE(iota & 0x3fff);
 	fib.pba = base;
 	/* Work around off by one in ISM virt device */
 	if (zdev->pft == PCI_FUNC_TYPE_ISM && limit > base)
 		fib.pal = limit + (1 << 12);
 	else
 		fib.pal = limit;
-	fib.iota = iota | ZPCI_IOTA_RTTO_FLAG;
+	if (iota == 0)
+		fib.iota = iota;
+	else
+		fib.iota = iota | ZPCI_IOTA_RTTO_FLAG;
 	fib.gd = zdev->gisa;
 	cc = zpci_mod_fc(req, &fib, status);
 	if (cc)
diff --git a/drivers/iommu/s390-iommu.c b/drivers/iommu/s390-iommu.c
index fbdeded3d48b..3d93a9644fca 100644
--- a/drivers/iommu/s390-iommu.c
+++ b/drivers/iommu/s390-iommu.c
@@ -16,7 +16,7 @@
 
 #include "dma-iommu.h"
 
-static const struct iommu_ops s390_iommu_ops;
+static const struct iommu_ops s390_iommu_ops, s390_iommu_rtr_ops;
 
 static struct kmem_cache *dma_region_table_cache;
 static struct kmem_cache *dma_page_table_cache;
@@ -392,9 +392,11 @@ static int blocking_domain_attach_device(struct iommu_domain *domain,
 		return 0;
 
 	s390_domain = to_s390_domain(zdev->s390_domain);
-	spin_lock_irqsave(&s390_domain->list_lock, flags);
-	list_del_rcu(&zdev->iommu_list);
-	spin_unlock_irqrestore(&s390_domain->list_lock, flags);
+	if (zdev->dma_table) {
+		spin_lock_irqsave(&s390_domain->list_lock, flags);
+		list_del_rcu(&zdev->iommu_list);
+		spin_unlock_irqrestore(&s390_domain->list_lock, flags);
+	}
 
 	zpci_unregister_ioat(zdev, 0);
 	zdev->dma_table = NULL;
@@ -723,7 +725,13 @@ int zpci_init_iommu(struct zpci_dev *zdev)
 	if (rc)
 		goto out_err;
 
-	rc = iommu_device_register(&zdev->iommu_dev, &s390_iommu_ops, NULL);
+	if (zdev->rtr_avail) {
+		rc = iommu_device_register(&zdev->iommu_dev,
+					   &s390_iommu_rtr_ops, NULL);
+	} else {
+		rc = iommu_device_register(&zdev->iommu_dev, &s390_iommu_ops,
+					   NULL);
+	}
 	if (rc)
 		goto out_sysfs;
 
@@ -787,6 +795,39 @@ static int __init s390_iommu_init(void)
 }
 subsys_initcall(s390_iommu_init);
 
+static int s390_attach_dev_identity(struct iommu_domain *domain,
+				    struct device *dev)
+{
+	struct zpci_dev *zdev = to_zpci_dev(dev);
+	u8 status;
+	int cc;
+
+	blocking_domain_attach_device(&blocking_domain, dev);
+
+	/* If we fail now DMA remains blocked via blocking domain */
+	cc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma,
+				0, &status);
+	/*
+	 * If the device is undergoing error recovery the reset code
+	 * will re-establish the new domain.
+	 */
+	if (cc && status != ZPCI_PCI_ST_FUNC_NOT_AVAIL)
+		return -EIO;
+
+	zdev_s390_domain_update(zdev, domain);
+
+	return 0;
+}
+
+static const struct iommu_domain_ops s390_identity_ops = {
+	.attach_dev = s390_attach_dev_identity,
+};
+
+static struct iommu_domain s390_identity_domain = {
+	.type = IOMMU_DOMAIN_IDENTITY,
+	.ops = &s390_identity_ops,
+};
+
 static struct iommu_domain blocking_domain = {
 	.type = IOMMU_DOMAIN_BLOCKED,
 	.ops = &(const struct iommu_domain_ops) {
@@ -794,23 +835,31 @@ static struct iommu_domain blocking_domain = {
 	}
 };
 
-static const struct iommu_ops s390_iommu_ops = {
-	.blocked_domain		= &blocking_domain,
-	.release_domain		= &blocking_domain,
-	.capable = s390_iommu_capable,
-	.domain_alloc_paging = s390_domain_alloc_paging,
-	.probe_device = s390_iommu_probe_device,
-	.device_group = generic_device_group,
-	.pgsize_bitmap = SZ_4K,
-	.get_resv_regions = s390_iommu_get_resv_regions,
-	.default_domain_ops = &(const struct iommu_domain_ops) {
-		.attach_dev	= s390_iommu_attach_device,
-		.map_pages	= s390_iommu_map_pages,
-		.unmap_pages	= s390_iommu_unmap_pages,
-		.flush_iotlb_all = s390_iommu_flush_iotlb_all,
-		.iotlb_sync      = s390_iommu_iotlb_sync,
-		.iotlb_sync_map  = s390_iommu_iotlb_sync_map,
-		.iova_to_phys	= s390_iommu_iova_to_phys,
-		.free		= s390_domain_free,
+#define S390_IOMMU_COMMON_OPS() \
+	.blocked_domain		= &blocking_domain, \
+	.release_domain		= &blocking_domain, \
+	.capable = s390_iommu_capable, \
+	.domain_alloc_paging = s390_domain_alloc_paging, \
+	.probe_device = s390_iommu_probe_device, \
+	.device_group = generic_device_group, \
+	.pgsize_bitmap = SZ_4K, \
+	.get_resv_regions = s390_iommu_get_resv_regions, \
+	.default_domain_ops = &(const struct iommu_domain_ops) { \
+		.attach_dev	= s390_iommu_attach_device, \
+		.map_pages	= s390_iommu_map_pages, \
+		.unmap_pages	= s390_iommu_unmap_pages, \
+		.flush_iotlb_all = s390_iommu_flush_iotlb_all, \
+		.iotlb_sync      = s390_iommu_iotlb_sync, \
+		.iotlb_sync_map  = s390_iommu_iotlb_sync_map, \
+		.iova_to_phys	= s390_iommu_iova_to_phys, \
+		.free		= s390_domain_free, \
 	}
+
+static const struct iommu_ops s390_iommu_ops = {
+	S390_IOMMU_COMMON_OPS()
+};
+
+static const struct iommu_ops s390_iommu_rtr_ops = {
+	.identity_domain	= &s390_identity_domain,
+	S390_IOMMU_COMMON_OPS()
 };
-- 
2.47.1

Re: [PATCH v2 2/3] s390/pci: store DMA offset in bus_dma_region

[Niklas Schnelle]

On Fri, 2024-12-13 at 17:49 -0500, Matthew Rosato wrote:
> PCI devices on s390 have a DMA offset that is reported via CLP.  In
> preparation for allowing identity domains, setup the bus_dma_region
> for all PCI devices using the reported CLP value.
> 
> Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
> ---
>  arch/s390/pci/pci_bus.c | 18 ++++++++++++++++++
>  1 file changed, 18 insertions(+)
> 
> diff --git a/arch/s390/pci/pci_bus.c b/arch/s390/pci/pci_bus.c
> index d5ace00d10f0..14527687d0f2 100644
> --- a/arch/s390/pci/pci_bus.c
> +++ b/arch/s390/pci/pci_bus.c
> @@ -19,6 +19,7 @@
>  #include <linux/jump_label.h>
>  #include <linux/pci.h>
>  #include <linux/printk.h>
> +#include <linux/dma-direct.h>
>  
>  #include <asm/pci_clp.h>
>  #include <asm/pci_dma.h>
> @@ -284,10 +285,27 @@ static struct zpci_bus *zpci_bus_alloc(int topo, bool topo_is_tid)
>  	return zbus;
>  }
>  
> +static void pci_dma_range_setup(struct pci_dev *pdev)
> +{
> +	struct zpci_dev *zdev = to_zpci(pdev);
> +	struct bus_dma_region *map;
> +
> +	map = kzalloc(sizeof(*map), GFP_KERNEL);
> +	if (!map)
> +		return;
> +
> +	map->cpu_start = 0;
> +	map->dma_start = PAGE_ALIGN(zdev->start_dma);
> +	map->size = (u64)virt_to_phys(high_memory);

I don't think we should restrict the size here to the size of memory.
Instead I think it should be zdev->end_dma - zdev->start_dma.

Since we handle the restriction to memory size as reserved regions I
think that should be compatible. Also I think otherwise this might
break the admittedly odd s390_iommu_aperture=X kernel parameter on
LPARs.


> +	pdev->dev.dma_range_map = map;
> +}
> +
>  void pcibios_bus_add_device(struct pci_dev *pdev)
>  {
>  	struct zpci_dev *zdev = to_zpci(pdev);
>  
> +	pci_dma_range_setup(pdev);
> +
>  	/*
>  	 * With pdev->no_vf_scan the common PCI probing code does not
>  	 * perform PF/VF linking.

Re: [PATCH v2 2/3] s390/pci: store DMA offset in bus_dma_region

[Niklas Schnelle]

On Mon, 2024-12-16 at 10:29 +0100, Niklas Schnelle wrote:
> On Fri, 2024-12-13 at 17:49 -0500, Matthew Rosato wrote:
> > PCI devices on s390 have a DMA offset that is reported via CLP.  In
> > preparation for allowing identity domains, setup the bus_dma_region
> > for all PCI devices using the reported CLP value.
> > 
> > Signed-off-by: Matthew Rosato <mjrosato@linux.ibm.com>
> > ---
> >  arch/s390/pci/pci_bus.c | 18 ++++++++++++++++++
> >  1 file changed, 18 insertions(+)
> > 
> > diff --git a/arch/s390/pci/pci_bus.c b/arch/s390/pci/pci_bus.c
> > index d5ace00d10f0..14527687d0f2 100644
> > --- a/arch/s390/pci/pci_bus.c
> > +++ b/arch/s390/pci/pci_bus.c
> > @@ -19,6 +19,7 @@
> >  #include <linux/jump_label.h>
> >  #include <linux/pci.h>
> >  #include <linux/printk.h>
> > +#include <linux/dma-direct.h>
> >  
> >  #include <asm/pci_clp.h>
> >  #include <asm/pci_dma.h>
> > @@ -284,10 +285,27 @@ static struct zpci_bus *zpci_bus_alloc(int topo, bool topo_is_tid)
> >  	return zbus;
> >  }
> >  
> > +static void pci_dma_range_setup(struct pci_dev *pdev)
> > +{
> > +	struct zpci_dev *zdev = to_zpci(pdev);
> > +	struct bus_dma_region *map;
> > +
> > +	map = kzalloc(sizeof(*map), GFP_KERNEL);
> > +	if (!map)
> > +		return;
> > +
> > +	map->cpu_start = 0;
> > +	map->dma_start = PAGE_ALIGN(zdev->start_dma);
> > +	map->size = (u64)virt_to_phys(high_memory);
> 
> I don't think we should restrict the size here to the size of memory.
> Instead I think it should be zdev->end_dma - zdev->start_dma.
> 
> Since we handle the restriction to memory size as reserved regions I
> think that should be compatible. Also I think otherwise this might
> break the admittedly odd s390_iommu_aperture=X kernel parameter on
> LPARs.

Correction, zdev->end_dma - zdev->start_dma + 1 because zdev->end_dma
is inclusive ;-)

> 
> 
> > +	pdev->dev.dma_range_map = map;
> > +}
> > +
> >  void pcibios_bus_add_device(struct pci_dev *pdev)
> >  {
> >  	struct zpci_dev *zdev = to_zpci(pdev);
> >  
> > +	pci_dma_range_setup(pdev);
> > +
> >  	/*
> >  	 * With pdev->no_vf_scan the common PCI probing code does not
> >  	 * perform PF/VF linking.
> 

Re: [PATCH v2 2/3] s390/pci: store DMA offset in bus_dma_region

[Matthew Rosato]

>>> +	map->cpu_start = 0;
>>> +	map->dma_start = PAGE_ALIGN(zdev->start_dma);
>>> +	map->size = (u64)virt_to_phys(high_memory);
>>
>> I don't think we should restrict the size here to the size of memory.
>> Instead I think it should be zdev->end_dma - zdev->start_dma.
>>
>> Since we handle the restriction to memory size as reserved regions I
>> think that should be compatible. Also I think otherwise this might
>> break the admittedly odd s390_iommu_aperture=X kernel parameter on
>> LPARs.
> 
> Correction, zdev->end_dma - zdev->start_dma + 1 because zdev->end_dma
> is inclusive ;-)
> 

Forgot about that parameter, thanks...  OK, will change to:

map->size = zdev->end_dma - zdev->start_dma + 1;