Re: [PATCH kernel v9 23/32] powerpc/powernv/ioda: Define and implement DMA table/window management callbacks

From: Alexey Kardashevskiy
Date: Thu Apr 30 2015 - 05:56:34 EST

On 04/30/2015 02:37 PM, David Gibson wrote:
On Wed, Apr 29, 2015 at 07:44:20PM +1000, Alexey Kardashevskiy wrote:
On 04/29/2015 03:30 PM, David Gibson wrote:
On Sat, Apr 25, 2015 at 10:14:47PM +1000, Alexey Kardashevskiy wrote:
This extends iommu_table_group_ops by a set of callbacks to support
dynamic DMA windows management.

create_table() creates a TCE table with specific parameters.
it receives iommu_table_group to know nodeid in order to allocate
TCE table memory closer to the PHB. The exact format of allocated
multi-level table might be also specific to the PHB model (not
the case now though).
This callback calculated the DMA window offset on a PCI bus from @num
and stores it in a just created table.

set_window() sets the window at specified TVT index + @num on PHB.

unset_window() unsets the window from specified TVT.

This adds a free() callback to iommu_table_ops to free the memory
(potentially a tree of tables) allocated for the TCE table.

Doesn't the free callback belong with the previous patch introducing
multi-level tables?

If I did that, you would say "why is it here if nothing calls it" on
"multilevel" patch and "I see the allocation but I do not see memory
release" ;)

Yeah, fair enough ;)

I need some rule of thumb here. I think it is a bit cleaner if the same
patch adds a callback for memory allocation and its counterpart, no?

On further consideration, yes, I think you're right.

create_table() and free() are supposed to be called once per
VFIO container and set_window()/unset_window() are supposed to be
called for every group in a container.

This adds IOMMU capabilities to iommu_table_group such as default
32bit window parameters and others.

Signed-off-by: Alexey Kardashevskiy <aik@xxxxxxxxx>
arch/powerpc/include/asm/iommu.h | 19 ++++++++
arch/powerpc/platforms/powernv/pci-ioda.c | 75 ++++++++++++++++++++++++++---
arch/powerpc/platforms/powernv/pci-p5ioc2.c | 12 +++--
3 files changed, 96 insertions(+), 10 deletions(-)

diff --git a/arch/powerpc/include/asm/iommu.h b/arch/powerpc/include/asm/iommu.h
index 0f50ee2..7694546 100644
--- a/arch/powerpc/include/asm/iommu.h
+++ b/arch/powerpc/include/asm/iommu.h
@@ -70,6 +70,7 @@ struct iommu_table_ops {
/* get() returns a physical address */
unsigned long (*get)(struct iommu_table *tbl, long index);
void (*flush)(struct iommu_table *tbl);
+ void (*free)(struct iommu_table *tbl);

/* These are used by VIO */
@@ -148,6 +149,17 @@ extern struct iommu_table *iommu_init_table(struct iommu_table * tbl,
struct iommu_table_group;

struct iommu_table_group_ops {
+ long (*create_table)(struct iommu_table_group *table_group,
+ int num,
+ __u32 page_shift,
+ __u64 window_size,
+ __u32 levels,
+ struct iommu_table *tbl);
+ long (*set_window)(struct iommu_table_group *table_group,
+ int num,
+ struct iommu_table *tblnew);
+ long (*unset_window)(struct iommu_table_group *table_group,
+ int num);
* Switches ownership from the kernel itself to an external
* user. While onwership is taken, the kernel cannot use IOMMU itself.
@@ -160,6 +172,13 @@ struct iommu_table_group {
struct iommu_group *group;
+ /* Some key properties of IOMMU */
+ __u32 tce32_start;
+ __u32 tce32_size;
+ __u64 pgsizes; /* Bitmap of supported page sizes */
+ __u32 max_dynamic_windows_supported;
+ __u32 max_levels;

With this information, table_group seems even more like a bad name.
"iommu_state" maybe?

Please, no. We will never come to agreement then :( And "iommu_state" is too
general anyway, it won't pass.

struct iommu_table tables[IOMMU_TABLE_GROUP_MAX_TABLES];
struct iommu_table_group_ops *ops;
diff --git a/arch/powerpc/platforms/powernv/pci-ioda.c b/arch/powerpc/platforms/powernv/pci-ioda.c
index cc1d09c..4828837 100644
--- a/arch/powerpc/platforms/powernv/pci-ioda.c
+++ b/arch/powerpc/platforms/powernv/pci-ioda.c
@@ -24,6 +24,7 @@
#include <linux/msi.h>
#include <linux/memblock.h>
#include <linux/iommu.h>
+#include <linux/sizes.h>

#include <asm/sections.h>
#include <asm/io.h>
@@ -1846,6 +1847,7 @@ static struct iommu_table_ops pnv_ioda2_iommu_ops = {
.clear = pnv_ioda2_tce_free,
.get = pnv_tce_get,
+ .free = pnv_pci_free_table,

static void pnv_pci_ioda_setup_opal_tce_kill(struct pnv_phb *phb,
@@ -1936,6 +1938,8 @@ static void pnv_pci_ioda_setup_dma_pe(struct pnv_phb *phb,

tbl->it_ops = &pnv_ioda1_iommu_ops;
+ pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
+ pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
iommu_init_table(tbl, phb->hose->node);

if (pe->flags & PNV_IODA_PE_DEV) {
@@ -1961,7 +1965,7 @@ static void pnv_pci_ioda_setup_dma_pe(struct pnv_phb *phb,

static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
- struct iommu_table *tbl)
+ int num, struct iommu_table *tbl)
struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
@@ -1972,9 +1976,10 @@ static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
const __u64 win_size = tbl->it_size << tbl->it_page_shift;

- pe_info(pe, "Setting up window at %llx..%llx "
+ pe_info(pe, "Setting up window#%d at %llx..%llx "
"pgsize=0x%x tablesize=0x%lx "
"levels=%d levelsize=%x\n",
+ num,
start_addr, start_addr + win_size - 1,
1UL << tbl->it_page_shift, tbl->it_size << 3,
tbl->it_indirect_levels + 1, tbl->it_level_size << 3);
@@ -1987,7 +1992,7 @@ static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
rc = opal_pci_map_pe_dma_window(phb->opal_id,
- pe->pe_number << 1,
+ (pe->pe_number << 1) + num,

Heh, yes, well, that makes it rather clear that only 2 tables are possible.

tbl->it_indirect_levels + 1,
size << 3,
@@ -2000,7 +2005,7 @@ static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,

/* Store fully initialized *tbl (may be external) in PE */
- pe->table_group.tables[0] = *tbl;
+ pe->table_group.tables[num] = *tbl;

I'm a bit confused by this whole set_window thing. Is the idea that
with multiple groups in a container you have multiple table_group s
each with different copies of the iommu_table structures, but pointing
to the same actual TCE entries (it_base)?


It seems to me not terribly
obvious when you "create" a table and when you "set" a window.

A table is not attached anywhere until its address is programmed (in
set_window()) to the hardware, it is just a table in memory. For
POWER8/IODA2, I create a table before I attach any group to a container,
then I program this table to every attached container, right now it is done
in container's attach_group(). So later we can hotplug any host PCI device
to a container - it will program same TCE table to every new group in the

So you "create" once, then "set" it to one or more table_groups? It
seems odd that "create" is a table_group callback in that case.

Where else could it be? ppc_md? We are getting rid of these. Some global function? We do not want VFIO to know about this. I run out of ideas here.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at