[PATCH] powerpc/pseries/iommu: export DMA window data to user space
Vaibhav Jain
vaibhav at linux.ibm.com
Sat Mar 14 03:38:43 AEDT 2026
Hi Gaurav,
Thanks for the patch. Few review comments inline below:
Gaurav Batra <gbatra at linux.ibm.com> writes:
> Export PowerPC DMA window information (both default 2GB and Dynamic
> larger window) to user space via sysfs. Each of these DMA windows has
> attributes like size of the window, page size backing the window, mode,
> etc. Each of these atributes is exported for user space consumption as a
> file.
>
> PowerPC Host Bridge (PHB) can have multiple devices/functions sharing
> the same DMA window. For each PHB, iommu registration creates an iommu
> device under "/sys/devices/virtual/iommu".
>
> These devices will have 2 groups created to export Default and DDW
> attributes.
>
> Reviewed-by: Brian King <brking at linux.ibm.com>
> Signed-off-by: Gaurav Batra <gbatra at linux.ibm.com>
> ---
> .../arch/powerpc/dma_window_attributes.rst | 65 +++++
> arch/powerpc/include/asm/iommu.h | 20 ++
> arch/powerpc/kernel/iommu.c | 235 ++++++++++++++++++
> arch/powerpc/platforms/pseries/iommu.c | 156 ++++++++++++
> 4 files changed, 476 insertions(+)
> create mode 100644 Documentation/arch/powerpc/dma_window_attributes.rst
>
> diff --git a/Documentation/arch/powerpc/dma_window_attributes.rst b/Documentation/arch/powerpc/dma_window_attributes.rst
> new file mode 100644
> index 000000000000..8bd9aec8539d
> --- /dev/null
> +++ b/Documentation/arch/powerpc/dma_window_attributes.rst
> @@ -0,0 +1,65 @@
> +.. SPDX-License-Identifier: GPL-2.0
> +
> +=====================
> +DMA Window Attributes
> +=====================
> +
> +In PowerPC architecture there are 2 types of DMA windows -
> +
> +1. Default 2GB DMA window which is backed by 4K page size
> +2. A bigger Dynamic DMA Window (DDW) which is backed by larger page size
> + (64K or 2MB)
> +
> +A dedicated device will have both the DMA windows instantiated but an SR-IOV
> +device will only have the bigger Dynamic DMA Window.
> +
> +The attributes of these 2 DMA windows are exported to user space via sysfs.
> +Each IOMMU isolation unit will have its directory created under
> +/sys/devices/virtual/iommu.
> +
> +As an exapmple, iommu-phb0001
> +
> +Under each IOMMU isolation unit, there will be a group of attributes for
> +"Default 2GB DMA Window" and "Dynamic DMA Window" - spapr-tce-dma and
> +spapr-tce-ddw respectively.
> +
> +Attributes under each group
> +
> +spapr-tce-ddw:
> +direct_address dynamic_address dynamic_size window_type
> +direct_size dynamic_pages_mapped page_size
> +
> +spapr-tce-dma:
> +dynamic_address dynamic_pages_mapped dynamic_size page_size
> +
> +
> +The bigger Dynamic DMA Window is configured into pre-mapped and/or dynamically
> +allocated TCEs. If the DDW is in "Hybrid" mode, then both the Direct
> +(pre-mapped) and Dynamic part of the DMA window will have valid values. Hybrid
> +mode is valid only for SR-IOV devices.
> +
> +DMA Window properties:
> +
> +direct_address Starting address of the pre-mapped DMA window
> +direct_size Size of the pre-mapped DMA Window
> +dynamic_address Starting address of the dynamic allocations
> +dynamic_size Size of the dynamic allocation window
> +dynamic_pages_mapped Pages mapped for DMA by dynamic allocations
> +page_size Page size backing the DMA window
> +window_type Type of the DMA Window (Direct/Dynamic/Hybrid)
> +
> +
> +An example of DDW attributes for an SR-IOV device::
> +
> + $ cd /sys/devices/virtual/iommu/iommu-phb0001/spapr-tce-ddw
> +
> + $ grep . *
> +
> + direct_address:0x800000000000000 <-- Starting addr of pre-mapped Window
> + direct_size:137438953472 <-- Size of pre-mapped Window (128GB)
> + dynamic_address:0x800002000000000 <-- Starting addr of Dynamic allocations
> + dynamic_size:412316860416 <-- Size of dynamic allocation window (384GB)
> + dynamic_pages_mapped:270 <-- Pages mapped by dynamic allocations
> + page_size:2097152 <-- DMA window page size (2MB)
> + window_type:Hybrid <-- window has both pre-mapped and
> + dynamic sections
Since sysfs is ABI can you propose appropriate entries under Documentation/ABI/testing
> diff --git a/arch/powerpc/include/asm/iommu.h b/arch/powerpc/include/asm/iommu.h
> index eafdd63cd6c4..e644c6e95301 100644
> --- a/arch/powerpc/include/asm/iommu.h
> +++ b/arch/powerpc/include/asm/iommu.h
> @@ -90,6 +90,7 @@ struct iommu_pool {
> unsigned long start;
> unsigned long end;
> unsigned long hint;
> + unsigned long inuse;
> spinlock_t lock;
> } ____cacheline_aligned_in_smp;
>
Review-comment from Shivaprasad:
Instead of maintaining a counter in iommu_pool can you just 'weigh' the it_map
bitmap. That way you wont have to introduce a new counter. Please look
into how iommu_debugfs_weight_get() does this.
> @@ -319,5 +320,24 @@ extern unsigned long iommu_direction_to_tce_perm(enum dma_data_direction dir);
>
> extern const struct dma_map_ops dma_iommu_ops;
>
> +/* used by sysfs when querying Dynamic/Default DMA Window data */
> +struct dma_win_data {
> + u32 win_pgsize;
> + u64 direct_addr;
> + u64 direct_size;
> + u64 dynamic_addr;
> + u64 dynamic_size;
> + u32 dynamic_tces_inuse;
> + char win_type[15];
> +};
> +
> +#define SPAPR_SUCCESS 0
> +#define SPAPR_NODMAWIN -1
> +#define SPAPR_NODDWWIN -2
> +#define SPAPR_ERROR -3
> +
> +extern int gather_ddw_info(struct device *dev, struct dma_win_data *data);
> +extern int gather_dma_info(struct device *dev, struct dma_win_data *data);
> +
> #endif /* __KERNEL__ */
> #endif /* _ASM_IOMMU_H */
> diff --git a/arch/powerpc/kernel/iommu.c b/arch/powerpc/kernel/iommu.c
> index 0ce71310b7d9..e3cf3701dd6e 100644
> --- a/arch/powerpc/kernel/iommu.c
> +++ b/arch/powerpc/kernel/iommu.c
> @@ -339,6 +339,9 @@ static unsigned long iommu_range_alloc(struct device *dev,
> if (handle)
> *handle = end;
>
> + /* update use count */
> + pool->inuse += npages;
> +
See the review comment above. This counter can be done away with.
> spin_unlock_irqrestore(&(pool->lock), flags);
>
> return n;
> @@ -452,6 +455,7 @@ static void __iommu_free(struct iommu_table *tbl, dma_addr_t dma_addr,
> tbl->it_ops->clear(tbl, entry, npages);
>
> spin_lock_irqsave(&(pool->lock), flags);
> + pool->inuse -= npages;
Ditto as above
> bitmap_clear(tbl->it_map, free_entry, npages);
> spin_unlock_irqrestore(&(pool->lock), flags);
> }
> @@ -759,6 +763,7 @@ struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid,
> p->start = tbl->poolsize * i;
> p->hint = p->start;
> p->end = p->start + tbl->poolsize;
> + p->inuse = 0;
Ditto as above
> }
>
> p = &tbl->large_pool;
> @@ -766,6 +771,7 @@ struct iommu_table *iommu_init_table(struct iommu_table *tbl, int nid,
> p->start = tbl->poolsize * i;
> p->hint = p->start;
> p->end = tbl->it_size;
> + p->inuse = 0;
>
> iommu_table_clear(tbl);
>
> @@ -1269,6 +1275,233 @@ static const struct iommu_ops spapr_tce_iommu_ops = {
> .device_group = spapr_tce_iommu_device_group,
> };
>
> +static inline const char *dma_win_error(int err)
> +{
> + switch (err) {
> + case SPAPR_ERROR:
> + return "Error";
> + case SPAPR_NODMAWIN:
> + return "No Default DMA Window Found";
> + case SPAPR_NODDWWIN:
> + return "No Dynamic DMA Window Found";
> + default:
> + return "Unknown Result";
> + }
> +}
> +
> +static ssize_t ddw_direct_address_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_ddw_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%#llx\n", data.direct_addr);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
Instead of returning success from these *_show() functions despite dma
window not available, can you just return an error (e.g ENOENT) so that
userspace know the error instantly instead of having to parse the sysfs
contents.
> +static ssize_t ddw_dynamic_address_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_ddw_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%#llx\n", data.dynamic_addr);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t ddw_direct_size_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_ddw_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%lld\n", data.direct_size);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t ddw_dynamic_size_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_ddw_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%lld\n", data.dynamic_size);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t ddw_page_size_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_ddw_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%d\n", data.win_pgsize);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t ddw_window_type_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_ddw_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%s\n", data.win_type);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t ddw_dynamic_pages_mapped_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_ddw_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%d\n", data.dynamic_tces_inuse);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t dma_dynamic_address_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_dma_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%#llx\n", data.dynamic_addr);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t dma_dynamic_size_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_dma_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%lld\n", data.dynamic_size);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t dma_page_size_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_dma_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%d\n", data.win_pgsize);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
> +
> +static ssize_t dma_dynamic_pages_mapped_show(struct device *dev,
> + struct device_attribute *attr,
> + char *buf)
> +{
> + int rc = 0;
> + struct dma_win_data data;
> +
> + rc = gather_dma_info(dev, &data);
> +
> + if (rc == SPAPR_SUCCESS)
> + return sysfs_emit(buf, "%d\n", data.dynamic_tces_inuse);
> + else
> + return sysfs_emit(buf, "%s\n", dma_win_error(rc));
> +}
All the *_show() functions above share same template. Please convert
them to macros expansion below to reduce code volume.
> +
> +#define DEVICE_ATTR_DDW(_name) \
> + struct device_attribute dev_attr_ddw_##_name = \
> + __ATTR(_name, 0444, ddw_##_name##_show, NULL)
> +#define DEVICE_ATTR_DMA(_name) \
> + struct device_attribute dev_attr_dma_##_name = \
> + __ATTR(_name, 0444, dma_##_name##_show, NULL)
> +
> +static DEVICE_ATTR_DDW(direct_address);
> +static DEVICE_ATTR_DDW(direct_size);
> +static DEVICE_ATTR_DDW(page_size);
> +static DEVICE_ATTR_DDW(window_type);
> +static DEVICE_ATTR_DDW(dynamic_address);
> +static DEVICE_ATTR_DDW(dynamic_size);
> +static DEVICE_ATTR_DDW(dynamic_pages_mapped);
> +static DEVICE_ATTR_DMA(dynamic_address);
> +static DEVICE_ATTR_DMA(dynamic_size);
> +static DEVICE_ATTR_DMA(page_size);
> +static DEVICE_ATTR_DMA(dynamic_pages_mapped);
> +
> +static struct attribute *spapr_tce_ddw_attrs[] = {
> + &dev_attr_ddw_direct_address.attr,
> + &dev_attr_ddw_direct_size.attr,
> + &dev_attr_ddw_page_size.attr,
> + &dev_attr_ddw_window_type.attr,
> + &dev_attr_ddw_dynamic_address.attr,
> + &dev_attr_ddw_dynamic_size.attr,
> + &dev_attr_ddw_dynamic_pages_mapped.attr,
> + NULL,
> +};
> +
> +static struct attribute *spapr_tce_dma_attrs[] = {
> + &dev_attr_dma_dynamic_address.attr,
> + &dev_attr_dma_dynamic_size.attr,
> + &dev_attr_dma_page_size.attr,
> + &dev_attr_dma_dynamic_pages_mapped.attr,
> + NULL,
> +};
> +
> +static struct attribute_group spapr_tce_ddw_group = {
> + .name = "spapr-tce-ddw",
> + .attrs = spapr_tce_ddw_attrs,
> +};
> +
> +static struct attribute_group spapr_tce_dma_group = {
> + .name = "spapr-tce-dma",
> + .attrs = spapr_tce_dma_attrs,
> +};
> +
These attributes are PSeries specific but they are being setup in ppc
generic iommu code at arch/powerpc/kernel/iommu.c . Can you move these
attributes to arch/powerpc/platforms/pseries/iommu.c
> static struct attribute *spapr_tce_iommu_attrs[] = {
> NULL,
> };
> @@ -1280,6 +1513,8 @@ static struct attribute_group spapr_tce_iommu_group = {
>
> static const struct attribute_group *spapr_tce_iommu_groups[] = {
> &spapr_tce_iommu_group,
> + &spapr_tce_ddw_group,
> + &spapr_tce_dma_group,
> NULL,
> };
>
> diff --git a/arch/powerpc/platforms/pseries/iommu.c b/arch/powerpc/platforms/pseries/iommu.c
> index 5497b130e026..5d04b50ae265 100644
> --- a/arch/powerpc/platforms/pseries/iommu.c
> +++ b/arch/powerpc/platforms/pseries/iommu.c
> @@ -837,6 +837,162 @@ static struct device_node *pci_dma_find(struct device_node *dn,
> return rdn;
> }
>
> +static unsigned long iommu_table_inuse_tces(struct iommu_table *tbl)
> +{
> + struct iommu_pool *pool;
> + unsigned long ntces = 0;
> +
> + /* Number of TCEs in-use */
> + for (int i = 0; i < tbl->nr_pools; i++) {
> + pool = &tbl->pools[i];
> + ntces += pool->inuse;
> + }
> +
> + pool = &tbl->large_pool;
> + ntces += pool->inuse;
> +
> + return ntces;
> +}
It would be better to use this functions as a callback in
iommu_table_ops which can be implemented by pseries and powernv code
differently.
> +
> +/* Get DDW information for the device */
> +int gather_ddw_info(struct device *dev, struct dma_win_data *data)
> +{
> + struct iommu_device *iommu;
> + struct pci_controller *phb;
> + struct device_node *dn;
> + struct pci_dn *pci;
> + const __be32 *prop = NULL;
> + bool ddw_direct = false;
> + bool found = false;
> + struct iommu_table *tbl;
> + u32 pgshift;
> + struct dynamic_dma_window_prop *p;
> +
> + memset(data, 0, sizeof(*data));
> +
> + iommu = dev_get_drvdata(dev);
> + phb = container_of(iommu, struct pci_controller, iommu);
> + dn = phb->dn;
> +
> + if (!dn)
> + return SPAPR_ERROR;
> +
> + pci = PCI_DN(dn);
> + if (!pci || !pci->table_group)
> + return SPAPR_ERROR;
> +
> + /* Find DDW */
> + prop = of_get_property(dn, DIRECT64_PROPNAME, NULL);
> + if (prop) {
> + ddw_direct = true;
> + found = true;
> + } else {
> + prop = of_get_property(dn, DMA64_PROPNAME, NULL);
> + if (prop)
> + found = true;
> + }
> +
> + /* NO DDW */
> + if (!found)
> + return SPAPR_NODDWWIN;
> +
> + p = (struct dynamic_dma_window_prop *)prop;
> +
> + pgshift = be32_to_cpu(p->tce_shift);
> + if (pgshift != 0xc && pgshift != 0x10 && pgshift != 0x15)
> + data->win_pgsize = 0;
> + else
> + data->win_pgsize = 1 << pgshift;
> +
> + /* Check if DDW has table associated with it. Having a table associated with
> + * DDW is indicative that is has some dynamic TCE allocations. In this case the
> + * DDW can be fully Dynamic or in Hybrid mode. For SR-IOV DDW is on index 0,
> + * for dedicated adapter on index 1.
> + */
> + found = false;
> + for (int i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
> + tbl = pci->table_group->tables[i];
> +
> + if (tbl && tbl->it_index == be32_to_cpu(p->liobn)) {
> + found = true;
> + break;
> + }
> + }
> +
> + /* set the parameters depnding on the DDW type */
> + if (ddw_direct && found) { /* Hybrid */
> + data->direct_addr = be64_to_cpu(p->dma_base);
> + data->dynamic_size = (u64)(tbl->it_size << tbl->it_page_shift);
> +
> + data->dynamic_addr = data->direct_addr
> + + (u64)(1UL << be32_to_cpu(p->window_shift))
> + - data->dynamic_size;
> +
> + data->direct_size = data->dynamic_addr - data->direct_addr;
> + data->dynamic_tces_inuse = iommu_table_inuse_tces(tbl);
> +
> + sprintf(data->win_type, "%s", "Hybrid");
> + } else if (ddw_direct && !found) { /* Direct */
> + data->direct_addr = be64_to_cpu(p->dma_base);
> + data->direct_size = (u64)(1UL << be32_to_cpu(p->window_shift));
> +
> + sprintf(data->win_type, "%s", "Direct");
> + } else { /* Dynamic */
> + data->dynamic_addr = be64_to_cpu(p->dma_base);
> + data->dynamic_size = (u64)(1UL << be32_to_cpu(p->window_shift));
> + data->dynamic_tces_inuse = iommu_table_inuse_tces(tbl);
> +
> + sprintf(data->win_type, "%s", "Dynamic");
> + }
> +
> + return SPAPR_SUCCESS;
> +}
> +
> +/* Get DDW information for the device */
> +int gather_dma_info(struct device *dev, struct dma_win_data *data)
> +{
> + struct iommu_device *iommu;
> + struct pci_controller *phb;
> + struct device_node *dn;
> + struct pci_dn *pci;
> + const __be32 *prop = NULL;
> + struct iommu_table *tbl;
> + unsigned long offset, size, liobn;
> +
> + memset(data, 0, sizeof(*data));
> +
> + iommu = dev_get_drvdata(dev);
> + phb = container_of(iommu, struct pci_controller, iommu);
> + dn = phb->dn;
> +
> + if (!dn)
> + return SPAPR_ERROR;
> +
> + pci = PCI_DN(dn);
> + if (!pci || !pci->table_group)
> + return SPAPR_ERROR;
> +
> + /* search for default DMA window */
> + prop = of_get_property(dn, "ibm,dma-window", NULL);
> +
> + if (!prop)
> + return SPAPR_NODMAWIN;
> +
> + /* default DMA Window is always at index 0 */
> + tbl = pci->table_group->tables[0];
> + if (!tbl)
> + return SPAPR_ERROR;
> +
> + of_parse_dma_window(dn, prop, &liobn, &offset, &size);
> +
> + data->dynamic_addr = offset;
> + data->dynamic_size = size;
> + data->win_pgsize = 1ULL << IOMMU_PAGE_SHIFT_4K;
> + data->dynamic_tces_inuse = iommu_table_inuse_tces(tbl);
> +
> + return SPAPR_SUCCESS;
> +}
> +
> static void pci_dma_bus_setup_pSeriesLP(struct pci_bus *bus)
> {
> struct iommu_table *tbl;
>
> base-commit: 192c0159402e6bfbe13de6f8379546943297783d
> --
> 2.39.3
>
--
Cheers
~ Vaibhav
More information about the Linuxppc-dev
mailing list