From: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
When multiple mappings are present for an input id, linux matches just
the first one. There is a usecase[1] where all the mappings are to be
maintained in parallel for an iommu-map entry of a same input id.
Whether multi-map is needed is reported by the callers through the
callback function passed, which is called for every input id match.
Since the requirement in the usecase[1] is for platform devices, not
sure if it is really clean to maintain this decision on the bus type at
the of_iommu layer or further to be from the respective
iommu_driver->impl_ops().
[1] https://lore.kernel.org/all/20250627-video_cb-v3-0-51e18c0ffbce@quicinc.com/
Signed-off-by: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
Signed-off-by: Vijayanand Jitta <vijayanand.jitta@oss.qualcomm.com>
Signed-off-by: Vikash Garodia <vikash.garodia@oss.qualcomm.com>
---
drivers/iommu/of_iommu.c | 36 ++++++++++++++++++++++++++++--------
drivers/of/base.c | 38 ++++++++++++++++++++++++++++----------
include/linux/of.h | 6 ++++++
3 files changed, 62 insertions(+), 18 deletions(-)
diff --git a/drivers/iommu/of_iommu.c b/drivers/iommu/of_iommu.c
index 768eaddf927b0700b2497b08ea21611b1a1b5688..067bb2298973671e1eaf01bb2ea52df3d2a52a44 100644
--- a/drivers/iommu/of_iommu.c
+++ b/drivers/iommu/of_iommu.c
@@ -16,6 +16,7 @@
#include <linux/pci.h>
#include <linux/slab.h>
#include <linux/fsl/mc.h>
+#include <linux/platform_device.h>
#include "iommu-priv.h"
@@ -41,22 +42,41 @@ static int of_iommu_xlate(struct device *dev,
return ret;
}
+/*
+ * Callback to be called from of_map_id(), that tells if
+ * all the mappings for an input id to be maintained in
+ * parallel. Should this decission be from further layers,
+ * iommu_driver->impl_ops?
+ */
+static int of_iommu_configure_cb(struct of_map_id_arg *arg)
+{
+ struct of_phandle_args *iommu_spec = &arg->map_args;
+ struct device *dev = arg->dev;
+ int err;
+
+ err = of_iommu_xlate(dev, iommu_spec);
+ of_node_put(iommu_spec->np);
+
+ /* !iommu_spec->np may be from the bypassed translations */
+ if (!err)
+ err = (!arg->multi_map || !iommu_spec->np) ? 0 : -EAGAIN;
+
+ return err;
+}
+
static int of_iommu_configure_dev_id(struct device_node *master_np,
struct device *dev,
const u32 *id)
{
struct of_map_id_arg arg = {
.map_args = {},
+ .cb = of_iommu_configure_cb,
+ .dev = dev,
+ /* Should this be pushed to iommu_driver->impl_ops? */
+ .multi_map = dev_is_platform(dev),
};
- int err;
-
- err = of_map_iommu_id(master_np, *id, &arg);
- if (err)
- return err;
- err = of_iommu_xlate(dev, &arg.map_args);
- of_node_put(arg.map_args.np);
- return err;
+ return of_map_iommu_id(master_np, *id, &arg);
}
static int of_iommu_configure_dev(struct device_node *master_np,
diff --git a/drivers/of/base.c b/drivers/of/base.c
index 606bef4f90e7d13bae4f7b0c45acd1755ad89826..a1c3c5954ec7e8eb3753c8fd782a1570f9eb9c17 100644
--- a/drivers/of/base.c
+++ b/drivers/of/base.c
@@ -2122,14 +2122,21 @@ static bool of_check_bad_map(const __be32 *map, int len)
return true;
}
-static int of_map_id_fill_output(struct of_map_id_arg *arg,
- struct device_node *phandle_node, u32 id_or_offset,
- const __be32 *out_base, u32 cells,
- bool bypass)
+/*
+ * Fill the id_out and target for the of_map_id() caller. Also
+ * call the callback passed to the of_map_id() as part of the arg
+ * that decides if to continue further search.
+ */
+static int of_map_id_fill_arg(struct of_map_id_arg *arg,
+ struct device_node *phandle_node, u32 id_or_offset,
+ const __be32 *out_base, u32 cells,
+ bool bypass, bool *multi_id_map)
{
+ int ret;
+
if (bypass) {
arg->map_args.args[0] = id_or_offset;
- return 0;
+ goto output;
}
if (arg->map_args.np)
@@ -2145,7 +2152,14 @@ static int of_map_id_fill_output(struct of_map_id_arg *arg,
arg->map_args.args_count = cells;
- return 0;
+output:
+ /* pass the output for the callback, callers may further decide */
+ ret = arg->cb ? arg->cb(arg) : 0;
+
+ if (multi_id_map && ret == -EAGAIN)
+ *multi_id_map = true;
+
+ return ret;
}
/**
@@ -2179,6 +2193,7 @@ int of_map_id(const struct device_node *np, u32 id, const char *map_name,
int map_bytes, map_len, offset = 0;
bool bad_map = false;
const __be32 *map = NULL;
+ bool multi_id_map = false;
if (!np || !map_name || !arg)
return -EINVAL;
@@ -2264,23 +2279,26 @@ int of_map_id(const struct device_node *np, u32 id, const char *map_name,
if (masked_id < id_base || id_off >= id_len)
continue;
- ret = of_map_id_fill_output(arg, phandle_node, id_off, out_base, cells, false);
+ ret = of_map_id_fill_arg(arg, phandle_node, id_off, out_base,
+ cells, false, &multi_id_map);
if (ret == -EAGAIN)
continue;
pr_debug("%pOF: %s, using mask %08x, id-base: %08x, out-base: %08x, length: %08x, id: %08x -> %08x\n",
np, map_name, map_mask, id_base, be32_to_cpup(out_base),
id_len, id, id_off + be32_to_cpup(out_base));
- return 0;
+ return ret;
}
+ if (multi_id_map)
+ return 0;
+
pr_info("%pOF: no %s translation for id 0x%x on %pOF\n", np, map_name,
id, arg->map_args.np ? arg->map_args.np : NULL);
bypass_translation:
/* Bypasses translation */
- return of_map_id_fill_output(arg, NULL, id, 0, 0, true);
-
+ return of_map_id_fill_arg(arg, NULL, id, 0, 0, true, NULL);
err_map_len:
pr_err("%pOF: Error: Bad %s length: %d\n", np, map_name, map_bytes);
return -EINVAL;
diff --git a/include/linux/of.h b/include/linux/of.h
index 9efa6f93712c6024f05476f9fd39f3294f942ec1..abab73a76682351f5635c1127a6c899917525050 100644
--- a/include/linux/of.h
+++ b/include/linux/of.h
@@ -25,6 +25,9 @@
typedef u32 phandle;
typedef u32 ihandle;
+struct of_map_id_arg;
+typedef int (*of_map_id_cb)(struct of_map_id_arg *arg);
+
struct property {
char *name;
int length;
@@ -76,6 +79,9 @@ struct of_phandle_args {
struct of_map_id_arg {
struct of_phandle_args map_args;
+ of_map_id_cb cb;
+ struct device *dev;
+ bool multi_map;
};
struct of_phandle_iterator {
--
2.34.1On 26/01/2026 12:25, Vikash Garodia wrote:
> From: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
>
> When multiple mappings are present for an input id, linux matches just
> the first one. There is a usecase[1] where all the mappings are to be
> maintained in parallel for an iommu-map entry of a same input id.
>
> Whether multi-map is needed is reported by the callers through the
> callback function passed, which is called for every input id match.
>
> Since the requirement in the usecase[1] is for platform devices, not
> sure if it is really clean to maintain this decision on the bus type at
> the of_iommu layer or further to be from the respective
> iommu_driver->impl_ops().
>
> [1] https://lore.kernel.org/all/20250627-video_cb-v3-0-51e18c0ffbce@quicinc.com/
>
> Signed-off-by: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
> Signed-off-by: Vijayanand Jitta <vijayanand.jitta@oss.qualcomm.com>
> Signed-off-by: Vikash Garodia <vikash.garodia@oss.qualcomm.com>
> ---
> drivers/iommu/of_iommu.c | 36 ++++++++++++++++++++++++++++--------
> drivers/of/base.c | 38 ++++++++++++++++++++++++++++----------
> include/linux/of.h | 6 ++++++
> 3 files changed, 62 insertions(+), 18 deletions(-)
>
> diff --git a/drivers/iommu/of_iommu.c b/drivers/iommu/of_iommu.c
> index 768eaddf927b0700b2497b08ea21611b1a1b5688..067bb2298973671e1eaf01bb2ea52df3d2a52a44 100644
> --- a/drivers/iommu/of_iommu.c
> +++ b/drivers/iommu/of_iommu.c
> @@ -16,6 +16,7 @@
> #include <linux/pci.h>
> #include <linux/slab.h>
> #include <linux/fsl/mc.h>
> +#include <linux/platform_device.h>
>
> #include "iommu-priv.h"
>
> @@ -41,22 +42,41 @@ static int of_iommu_xlate(struct device *dev,
> return ret;
> }
>
> +/*
> + * Callback to be called from of_map_id(), that tells if
> + * all the mappings for an input id to be maintained in
> + * parallel. Should this decission be from further layers,
> + * iommu_driver->impl_ops?
> + */
> +static int of_iommu_configure_cb(struct of_map_id_arg *arg)
> +{
> + struct of_phandle_args *iommu_spec = &arg->map_args;
> + struct device *dev = arg->dev;
> + int err;
> +
> + err = of_iommu_xlate(dev, iommu_spec);
> + of_node_put(iommu_spec->np);
> +
> + /* !iommu_spec->np may be from the bypassed translations */
> + if (!err)
> + err = (!arg->multi_map || !iommu_spec->np) ? 0 : -EAGAIN;
> +
> + return err;
> +}
> +
> static int of_iommu_configure_dev_id(struct device_node *master_np,
> struct device *dev,
> const u32 *id)
> {
> struct of_map_id_arg arg = {
> .map_args = {},
> + .cb = of_iommu_configure_cb,
> + .dev = dev,
> + /* Should this be pushed to iommu_driver->impl_ops? */
> + .multi_map = dev_is_platform(dev),
> };
> - int err;
> -
> - err = of_map_iommu_id(master_np, *id, &arg);
> - if (err)
> - return err;
>
> - err = of_iommu_xlate(dev, &arg.map_args);
> - of_node_put(arg.map_args.np);
> - return err;
> + return of_map_iommu_id(master_np, *id, &arg);
> }
>
> static int of_iommu_configure_dev(struct device_node *master_np,
> diff --git a/drivers/of/base.c b/drivers/of/base.c
> index 606bef4f90e7d13bae4f7b0c45acd1755ad89826..a1c3c5954ec7e8eb3753c8fd782a1570f9eb9c17 100644
> --- a/drivers/of/base.c
> +++ b/drivers/of/base.c
> @@ -2122,14 +2122,21 @@ static bool of_check_bad_map(const __be32 *map, int len)
> return true;
> }
>
> -static int of_map_id_fill_output(struct of_map_id_arg *arg,
> - struct device_node *phandle_node, u32 id_or_offset,
> - const __be32 *out_base, u32 cells,
> - bool bypass)
> +/*
> + * Fill the id_out and target for the of_map_id() caller. Also
> + * call the callback passed to the of_map_id() as part of the arg
> + * that decides if to continue further search.
> + */
> +static int of_map_id_fill_arg(struct of_map_id_arg *arg,
> + struct device_node *phandle_node, u32 id_or_offset,
> + const __be32 *out_base, u32 cells,
> + bool bypass, bool *multi_id_map)
> {
> + int ret;
> +
> if (bypass) {
> arg->map_args.args[0] = id_or_offset;
> - return 0;
> + goto output;
> }
>
> if (arg->map_args.np)
> @@ -2145,7 +2152,14 @@ static int of_map_id_fill_output(struct of_map_id_arg *arg,
>
> arg->map_args.args_count = cells;
>
> - return 0;
> +output:
> + /* pass the output for the callback, callers may further decide */
> + ret = arg->cb ? arg->cb(arg) : 0;
> +
> + if (multi_id_map && ret == -EAGAIN)
> + *multi_id_map = true;
> +
> + return ret;
> }
>
> /**
> @@ -2179,6 +2193,7 @@ int of_map_id(const struct device_node *np, u32 id, const char *map_name,
> int map_bytes, map_len, offset = 0;
> bool bad_map = false;
> const __be32 *map = NULL;
> + bool multi_id_map = false;
>
> if (!np || !map_name || !arg)
> return -EINVAL;
> @@ -2264,23 +2279,26 @@ int of_map_id(const struct device_node *np, u32 id, const char *map_name,
> if (masked_id < id_base || id_off >= id_len)
> continue;
>
> - ret = of_map_id_fill_output(arg, phandle_node, id_off, out_base, cells, false);
> + ret = of_map_id_fill_arg(arg, phandle_node, id_off, out_base,
> + cells, false, &multi_id_map);
> if (ret == -EAGAIN)
> continue;
>
> pr_debug("%pOF: %s, using mask %08x, id-base: %08x, out-base: %08x, length: %08x, id: %08x -> %08x\n",
> np, map_name, map_mask, id_base, be32_to_cpup(out_base),
> id_len, id, id_off + be32_to_cpup(out_base));
> - return 0;
> + return ret;
> }
>
> + if (multi_id_map)
> + return 0;
> +
> pr_info("%pOF: no %s translation for id 0x%x on %pOF\n", np, map_name,
> id, arg->map_args.np ? arg->map_args.np : NULL);
>
> bypass_translation:
> /* Bypasses translation */
> - return of_map_id_fill_output(arg, NULL, id, 0, 0, true);
> -
> + return of_map_id_fill_arg(arg, NULL, id, 0, 0, true, NULL);
> err_map_len:
> pr_err("%pOF: Error: Bad %s length: %d\n", np, map_name, map_bytes);
> return -EINVAL;
> diff --git a/include/linux/of.h b/include/linux/of.h
> index 9efa6f93712c6024f05476f9fd39f3294f942ec1..abab73a76682351f5635c1127a6c899917525050 100644
> --- a/include/linux/of.h
> +++ b/include/linux/of.h
> @@ -25,6 +25,9 @@
> typedef u32 phandle;
> typedef u32 ihandle;
>
> +struct of_map_id_arg;
> +typedef int (*of_map_id_cb)(struct of_map_id_arg *arg);
> +
> struct property {
> char *name;
> int length;
> @@ -76,6 +79,9 @@ struct of_phandle_args {
>
> struct of_map_id_arg {
> struct of_phandle_args map_args;
> + of_map_id_cb cb;
> + struct device *dev;
> + bool multi_map;
> };
>
> struct of_phandle_iterator {
>
I think at a minimum this and the previous patch should be separated
into its/their own series ∵ you really require this to be applied before
proceeding on with the rest of the submission.
Get these two patches through iommu@lists.linux.dev in isolation and
then submit the driver changes to consume.
---
bod
On 2/2/2026 8:27 PM, Bryan O'Donoghue wrote:
> On 26/01/2026 12:25, Vikash Garodia wrote:
>> From: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
>>
>> When multiple mappings are present for an input id, linux matches just
>> the first one. There is a usecase[1] where all the mappings are to be
>> maintained in parallel for an iommu-map entry of a same input id.
>>
>> Whether multi-map is needed is reported by the callers through the
>> callback function passed, which is called for every input id match.
>>
>> Since the requirement in the usecase[1] is for platform devices, not
>> sure if it is really clean to maintain this decision on the bus type at
>> the of_iommu layer or further to be from the respective
>> iommu_driver->impl_ops().
>>
>> [1] https://lore.kernel.org/all/20250627-video_cb-v3-0-51e18c0ffbce@quicinc.com/
>>
>> Signed-off-by: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
>> Signed-off-by: Vijayanand Jitta <vijayanand.jitta@oss.qualcomm.com>
>> Signed-off-by: Vikash Garodia <vikash.garodia@oss.qualcomm.com>
>> ---
>> drivers/iommu/of_iommu.c | 36 ++++++++++++++++++++++++++++--------
>> drivers/of/base.c | 38 ++++++++++++++++++++++++++++----------
>> include/linux/of.h | 6 ++++++
>> 3 files changed, 62 insertions(+), 18 deletions(-)
>>
>> diff --git a/drivers/iommu/of_iommu.c b/drivers/iommu/of_iommu.c
>> index 768eaddf927b0700b2497b08ea21611b1a1b5688..067bb2298973671e1eaf01bb2ea52df3d2a52a44 100644
>> --- a/drivers/iommu/of_iommu.c
>> +++ b/drivers/iommu/of_iommu.c
>> @@ -16,6 +16,7 @@
>> #include <linux/pci.h>
>> #include <linux/slab.h>
>> #include <linux/fsl/mc.h>
>> +#include <linux/platform_device.h>
>> #include "iommu-priv.h"
>> @@ -41,22 +42,41 @@ static int of_iommu_xlate(struct device *dev,
>> return ret;
>> }
>> +/*
>> + * Callback to be called from of_map_id(), that tells if
>> + * all the mappings for an input id to be maintained in
>> + * parallel. Should this decission be from further layers,
>> + * iommu_driver->impl_ops?
>> + */
>> +static int of_iommu_configure_cb(struct of_map_id_arg *arg)
>> +{
>> + struct of_phandle_args *iommu_spec = &arg->map_args;
>> + struct device *dev = arg->dev;
>> + int err;
>> +
>> + err = of_iommu_xlate(dev, iommu_spec);
>> + of_node_put(iommu_spec->np);
>> +
>> + /* !iommu_spec->np may be from the bypassed translations */
>> + if (!err)
>> + err = (!arg->multi_map || !iommu_spec->np) ? 0 : -EAGAIN;
>> +
>> + return err;
>> +}
>> +
>> static int of_iommu_configure_dev_id(struct device_node *master_np,
>> struct device *dev,
>> const u32 *id)
>> {
>> struct of_map_id_arg arg = {
>> .map_args = {},
>> + .cb = of_iommu_configure_cb,
>> + .dev = dev,
>> + /* Should this be pushed to iommu_driver->impl_ops? */
>> + .multi_map = dev_is_platform(dev),
>> };
>> - int err;
>> -
>> - err = of_map_iommu_id(master_np, *id, &arg);
>> - if (err)
>> - return err;
>> - err = of_iommu_xlate(dev, &arg.map_args);
>> - of_node_put(arg.map_args.np);
>> - return err;
>> + return of_map_iommu_id(master_np, *id, &arg);
>> }
>> static int of_iommu_configure_dev(struct device_node *master_np,
>> diff --git a/drivers/of/base.c b/drivers/of/base.c
>> index 606bef4f90e7d13bae4f7b0c45acd1755ad89826..a1c3c5954ec7e8eb3753c8fd782a1570f9eb9c17 100644
>> --- a/drivers/of/base.c
>> +++ b/drivers/of/base.c
>> @@ -2122,14 +2122,21 @@ static bool of_check_bad_map(const __be32 *map, int len)
>> return true;
>> }
>> -static int of_map_id_fill_output(struct of_map_id_arg *arg,
>> - struct device_node *phandle_node, u32 id_or_offset,
>> - const __be32 *out_base, u32 cells,
>> - bool bypass)
>> +/*
>> + * Fill the id_out and target for the of_map_id() caller. Also
>> + * call the callback passed to the of_map_id() as part of the arg
>> + * that decides if to continue further search.
>> + */
>> +static int of_map_id_fill_arg(struct of_map_id_arg *arg,
>> + struct device_node *phandle_node, u32 id_or_offset,
>> + const __be32 *out_base, u32 cells,
>> + bool bypass, bool *multi_id_map)
>> {
>> + int ret;
>> +
>> if (bypass) {
>> arg->map_args.args[0] = id_or_offset;
>> - return 0;
>> + goto output;
>> }
>> if (arg->map_args.np)
>> @@ -2145,7 +2152,14 @@ static int of_map_id_fill_output(struct of_map_id_arg *arg,
>> arg->map_args.args_count = cells;
>> - return 0;
>> +output:
>> + /* pass the output for the callback, callers may further decide */
>> + ret = arg->cb ? arg->cb(arg) : 0;
>> +
>> + if (multi_id_map && ret == -EAGAIN)
>> + *multi_id_map = true;
>> +
>> + return ret;
>> }
>> /**
>> @@ -2179,6 +2193,7 @@ int of_map_id(const struct device_node *np, u32 id, const char *map_name,
>> int map_bytes, map_len, offset = 0;
>> bool bad_map = false;
>> const __be32 *map = NULL;
>> + bool multi_id_map = false;
>> if (!np || !map_name || !arg)
>> return -EINVAL;
>> @@ -2264,23 +2279,26 @@ int of_map_id(const struct device_node *np, u32 id, const char *map_name,
>> if (masked_id < id_base || id_off >= id_len)
>> continue;
>> - ret = of_map_id_fill_output(arg, phandle_node, id_off, out_base, cells, false);
>> + ret = of_map_id_fill_arg(arg, phandle_node, id_off, out_base,
>> + cells, false, &multi_id_map);
>> if (ret == -EAGAIN)
>> continue;
>> pr_debug("%pOF: %s, using mask %08x, id-base: %08x, out-base: %08x, length: %08x, id: %08x -> %08x\n",
>> np, map_name, map_mask, id_base, be32_to_cpup(out_base),
>> id_len, id, id_off + be32_to_cpup(out_base));
>> - return 0;
>> + return ret;
>> }
>> + if (multi_id_map)
>> + return 0;
>> +
>> pr_info("%pOF: no %s translation for id 0x%x on %pOF\n", np, map_name,
>> id, arg->map_args.np ? arg->map_args.np : NULL);
>> bypass_translation:
>> /* Bypasses translation */
>> - return of_map_id_fill_output(arg, NULL, id, 0, 0, true);
>> -
>> + return of_map_id_fill_arg(arg, NULL, id, 0, 0, true, NULL);
>> err_map_len:
>> pr_err("%pOF: Error: Bad %s length: %d\n", np, map_name, map_bytes);
>> return -EINVAL;
>> diff --git a/include/linux/of.h b/include/linux/of.h
>> index 9efa6f93712c6024f05476f9fd39f3294f942ec1..abab73a76682351f5635c1127a6c899917525050 100644
>> --- a/include/linux/of.h
>> +++ b/include/linux/of.h
>> @@ -25,6 +25,9 @@
>> typedef u32 phandle;
>> typedef u32 ihandle;
>> +struct of_map_id_arg;
>> +typedef int (*of_map_id_cb)(struct of_map_id_arg *arg);
>> +
>> struct property {
>> char *name;
>> int length;
>> @@ -76,6 +79,9 @@ struct of_phandle_args {
>> struct of_map_id_arg {
>> struct of_phandle_args map_args;
>> + of_map_id_cb cb;
>> + struct device *dev;
>> + bool multi_map;
>> };
>> struct of_phandle_iterator {
>>
>
> I think at a minimum this and the previous patch should be separated into its/their own series ∵ you really require this to be applied before proceeding on with the rest of the submission.
>
> Get these two patches through iommu@lists.linux.dev in isolation and then submit the driver changes to consume.
>
> ---
> bod
Sure, I’ll split the first two patches into a seperate series and submit them.
Thanks,
Vijay
On Mon, Jan 26, 2026 at 05:55:46PM +0530, Vikash Garodia wrote:
> From: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
>
> When multiple mappings are present for an input id, linux matches just
> the first one. There is a usecase[1] where all the mappings are to be
> maintained in parallel for an iommu-map entry of a same input id.
This contradicts the IOMMU idealogy (at least as far as I understood it
fom the maintainers): the device (driver) doesn't control which IOMMUs
are getting used. Instead _all_ defined entries should get used. For
iommu-map it means that if the map defines several entries for a single
function, then all entries should always get mapped.
>
> Whether multi-map is needed is reported by the callers through the
> callback function passed, which is called for every input id match.
>
> Since the requirement in the usecase[1] is for platform devices, not
> sure if it is really clean to maintain this decision on the bus type at
> the of_iommu layer or further to be from the respective
> iommu_driver->impl_ops().
This doesn't tell us, why do you want to control, which entries of the
map get used.
>
> [1] https://lore.kernel.org/all/20250627-video_cb-v3-0-51e18c0ffbce@quicinc.com/
>
> Signed-off-by: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
> Signed-off-by: Vijayanand Jitta <vijayanand.jitta@oss.qualcomm.com>
> Signed-off-by: Vikash Garodia <vikash.garodia@oss.qualcomm.com>
> ---
> drivers/iommu/of_iommu.c | 36 ++++++++++++++++++++++++++++--------
> drivers/of/base.c | 38 ++++++++++++++++++++++++++++----------
> include/linux/of.h | 6 ++++++
> 3 files changed, 62 insertions(+), 18 deletions(-)
>
> diff --git a/drivers/iommu/of_iommu.c b/drivers/iommu/of_iommu.c
> index 768eaddf927b0700b2497b08ea21611b1a1b5688..067bb2298973671e1eaf01bb2ea52df3d2a52a44 100644
> --- a/drivers/iommu/of_iommu.c
> +++ b/drivers/iommu/of_iommu.c
> @@ -16,6 +16,7 @@
> #include <linux/pci.h>
> #include <linux/slab.h>
> #include <linux/fsl/mc.h>
> +#include <linux/platform_device.h>
>
> #include "iommu-priv.h"
>
> @@ -41,22 +42,41 @@ static int of_iommu_xlate(struct device *dev,
> return ret;
> }
>
> +/*
> + * Callback to be called from of_map_id(), that tells if
> + * all the mappings for an input id to be maintained in
> + * parallel. Should this decission be from further layers,
> + * iommu_driver->impl_ops?
> + */
> +static int of_iommu_configure_cb(struct of_map_id_arg *arg)
> +{
> + struct of_phandle_args *iommu_spec = &arg->map_args;
> + struct device *dev = arg->dev;
> + int err;
> +
> + err = of_iommu_xlate(dev, iommu_spec);
> + of_node_put(iommu_spec->np);
> +
> + /* !iommu_spec->np may be from the bypassed translations */
> + if (!err)
> + err = (!arg->multi_map || !iommu_spec->np) ? 0 : -EAGAIN;
> +
> + return err;
> +}
> +
> static int of_iommu_configure_dev_id(struct device_node *master_np,
> struct device *dev,
> const u32 *id)
> {
> struct of_map_id_arg arg = {
> .map_args = {},
> + .cb = of_iommu_configure_cb,
> + .dev = dev,
> + /* Should this be pushed to iommu_driver->impl_ops? */
> + .multi_map = dev_is_platform(dev),
> };
> - int err;
> -
> - err = of_map_iommu_id(master_np, *id, &arg);
> - if (err)
> - return err;
>
> - err = of_iommu_xlate(dev, &arg.map_args);
> - of_node_put(arg.map_args.np);
> - return err;
> + return of_map_iommu_id(master_np, *id, &arg);
> }
>
> static int of_iommu_configure_dev(struct device_node *master_np,
> diff --git a/drivers/of/base.c b/drivers/of/base.c
> index 606bef4f90e7d13bae4f7b0c45acd1755ad89826..a1c3c5954ec7e8eb3753c8fd782a1570f9eb9c17 100644
> --- a/drivers/of/base.c
> +++ b/drivers/of/base.c
> @@ -2122,14 +2122,21 @@ static bool of_check_bad_map(const __be32 *map, int len)
> return true;
> }
>
> -static int of_map_id_fill_output(struct of_map_id_arg *arg,
> - struct device_node *phandle_node, u32 id_or_offset,
> - const __be32 *out_base, u32 cells,
> - bool bypass)
> +/*
> + * Fill the id_out and target for the of_map_id() caller. Also
> + * call the callback passed to the of_map_id() as part of the arg
> + * that decides if to continue further search.
> + */
> +static int of_map_id_fill_arg(struct of_map_id_arg *arg,
> + struct device_node *phandle_node, u32 id_or_offset,
> + const __be32 *out_base, u32 cells,
> + bool bypass, bool *multi_id_map)
> {
> + int ret;
> +
> if (bypass) {
> arg->map_args.args[0] = id_or_offset;
> - return 0;
> + goto output;
> }
>
> if (arg->map_args.np)
> @@ -2145,7 +2152,14 @@ static int of_map_id_fill_output(struct of_map_id_arg *arg,
>
> arg->map_args.args_count = cells;
>
> - return 0;
> +output:
> + /* pass the output for the callback, callers may further decide */
> + ret = arg->cb ? arg->cb(arg) : 0;
> +
> + if (multi_id_map && ret == -EAGAIN)
> + *multi_id_map = true;
> +
> + return ret;
> }
>
> /**
> @@ -2179,6 +2193,7 @@ int of_map_id(const struct device_node *np, u32 id, const char *map_name,
> int map_bytes, map_len, offset = 0;
> bool bad_map = false;
> const __be32 *map = NULL;
> + bool multi_id_map = false;
>
> if (!np || !map_name || !arg)
> return -EINVAL;
> @@ -2264,23 +2279,26 @@ int of_map_id(const struct device_node *np, u32 id, const char *map_name,
> if (masked_id < id_base || id_off >= id_len)
> continue;
>
> - ret = of_map_id_fill_output(arg, phandle_node, id_off, out_base, cells, false);
> + ret = of_map_id_fill_arg(arg, phandle_node, id_off, out_base,
> + cells, false, &multi_id_map);
> if (ret == -EAGAIN)
> continue;
>
> pr_debug("%pOF: %s, using mask %08x, id-base: %08x, out-base: %08x, length: %08x, id: %08x -> %08x\n",
> np, map_name, map_mask, id_base, be32_to_cpup(out_base),
> id_len, id, id_off + be32_to_cpup(out_base));
> - return 0;
> + return ret;
> }
>
> + if (multi_id_map)
> + return 0;
> +
> pr_info("%pOF: no %s translation for id 0x%x on %pOF\n", np, map_name,
> id, arg->map_args.np ? arg->map_args.np : NULL);
>
> bypass_translation:
> /* Bypasses translation */
> - return of_map_id_fill_output(arg, NULL, id, 0, 0, true);
> -
> + return of_map_id_fill_arg(arg, NULL, id, 0, 0, true, NULL);
> err_map_len:
> pr_err("%pOF: Error: Bad %s length: %d\n", np, map_name, map_bytes);
> return -EINVAL;
> diff --git a/include/linux/of.h b/include/linux/of.h
> index 9efa6f93712c6024f05476f9fd39f3294f942ec1..abab73a76682351f5635c1127a6c899917525050 100644
> --- a/include/linux/of.h
> +++ b/include/linux/of.h
> @@ -25,6 +25,9 @@
> typedef u32 phandle;
> typedef u32 ihandle;
>
> +struct of_map_id_arg;
> +typedef int (*of_map_id_cb)(struct of_map_id_arg *arg);
> +
> struct property {
> char *name;
> int length;
> @@ -76,6 +79,9 @@ struct of_phandle_args {
>
> struct of_map_id_arg {
> struct of_phandle_args map_args;
> + of_map_id_cb cb;
> + struct device *dev;
> + bool multi_map;
> };
>
> struct of_phandle_iterator {
>
> --
> 2.34.1
>
--
With best wishes
Dmitry
On 2026-01-27 11:45 am, Dmitry Baryshkov wrote: > On Mon, Jan 26, 2026 at 05:55:46PM +0530, Vikash Garodia wrote: >> From: Charan Teja Kalla <charan.kalla@oss.qualcomm.com> >> >> When multiple mappings are present for an input id, linux matches just >> the first one. There is a usecase[1] where all the mappings are to be >> maintained in parallel for an iommu-map entry of a same input id. > > This contradicts the IOMMU idealogy (at least as far as I understood it > fom the maintainers): the device (driver) doesn't control which IOMMUs > are getting used. Instead _all_ defined entries should get used. For > iommu-map it means that if the map defines several entries for a single > function, then all entries should always get mapped. Indeed there is no concept of "multi-map" - if a single input ID represents more than one thing then that notion of "input ID" is fundamentally wrong. A single *device* may have multiple IDs, as in the case of PCI bridge aliasing, but in that case there are multiple things to map. Thanks, Robin.
On 1/27/2026 7:50 PM, Robin Murphy wrote:
> On 2026-01-27 11:45 am, Dmitry Baryshkov wrote:
>> On Mon, Jan 26, 2026 at 05:55:46PM +0530, Vikash Garodia wrote:
>>> From: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
>>>
>>> When multiple mappings are present for an input id, linux matches just
>>> the first one. There is a usecase[1] where all the mappings are to be
>>> maintained in parallel for an iommu-map entry of a same input id.
>>
>> This contradicts the IOMMU idealogy (at least as far as I understood it
>> fom the maintainers): the device (driver) doesn't control which IOMMUs
>> are getting used. Instead _all_ defined entries should get used. For
>> iommu-map it means that if the map defines several entries for a single
>> function, then all entries should always get mapped.
>
> Indeed there is no concept of "multi-map" - if a single input ID represents more than one thing then that notion of "input ID" is fundamentally wrong. A single *device* may have multiple IDs, as in the case of PCI bridge aliasing, but in that case there are multiple things to map.
>
vpu hardware do have video decode and encode usecases that would generate multiple Stream ID's.
So, all these Stream ID's would need to be represented using single input id as mentioned
in dt binding.
Referring patch [1/7] in this series
iommu-map = <0x100 &apps_smmu 0x1940 0x0 0x1>,
<0x100 &apps_smmu 0x1a20 0x0 0x1>,
<0x100 &apps_smmu 0x1944 0x0 0x1>;
Here, IRIS_CB_NON_SECURE_NON_PIXEL [1] is the input id.
enum iris_iommu_map_function_id {
IRIS_CB_NON_SECURE_NON_PIXEL = 0x100,
[1] https://lore.kernel.org/all/20260126-kaanapali-iris-v1-5-e2646246bfc1@oss.qualcomm.com/
Thanks,
Vijay
> Thanks,
> Robin.
Hi,
Le mardi 27 janvier 2026 à 13:45 +0200, Dmitry Baryshkov a écrit :
> On Mon, Jan 26, 2026 at 05:55:46PM +0530, Vikash Garodia wrote:
> > From: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
> >
> > When multiple mappings are present for an input id, linux matches just
> > the first one. There is a usecase[1] where all the mappings are to be
> > maintained in parallel for an iommu-map entry of a same input id.
>
> This contradicts the IOMMU idealogy (at least as far as I understood it
> fom the maintainers): the device (driver) doesn't control which IOMMUs
> are getting used. Instead _all_ defined entries should get used. For
> iommu-map it means that if the map defines several entries for a single
> function, then all entries should always get mapped.
>
> >
> > Whether multi-map is needed is reported by the callers through the
> > callback function passed, which is called for every input id match.
> >
> > Since the requirement in the usecase[1] is for platform devices, not
> > sure if it is really clean to maintain this decision on the bus type at
> > the of_iommu layer or further to be from the respective
> > iommu_driver->impl_ops().
>
> This doesn't tell us, why do you want to control, which entries of the
> map get used.
>
> >
> > [1]
> > https://lore.kernel.org/all/20250627-video_cb-v3-0-51e18c0ffbce@quicinc.com/
Since its for an M2M V4L2 device, and its all about having more address space
according to this link, you may want to use a different iommu domain per m2m
instances, rather then per device (what is done implicitly). On top of which, it
will ensure memory isolation between instances, making this driver suitable for
virtualisation. 4Gb per instance (which is also per stream) seems like a lot of
memory to me. But like Dmitry says, there is little clarity in what you are
trying to achieve.
Nicolas
> >
> > Signed-off-by: Charan Teja Kalla <charan.kalla@oss.qualcomm.com>
> > Signed-off-by: Vijayanand Jitta <vijayanand.jitta@oss.qualcomm.com>
> > Signed-off-by: Vikash Garodia <vikash.garodia@oss.qualcomm.com>
> > ---
> > drivers/iommu/of_iommu.c | 36 ++++++++++++++++++++++++++++--------
> > drivers/of/base.c | 38 ++++++++++++++++++++++++++++----------
> > include/linux/of.h | 6 ++++++
> > 3 files changed, 62 insertions(+), 18 deletions(-)
> >
> > diff --git a/drivers/iommu/of_iommu.c b/drivers/iommu/of_iommu.c
> > index
> > 768eaddf927b0700b2497b08ea21611b1a1b5688..067bb2298973671e1eaf01bb2ea52df3d2
> > a52a44 100644
> > --- a/drivers/iommu/of_iommu.c
> > +++ b/drivers/iommu/of_iommu.c
> > @@ -16,6 +16,7 @@
> > #include <linux/pci.h>
> > #include <linux/slab.h>
> > #include <linux/fsl/mc.h>
> > +#include <linux/platform_device.h>
> >
> > #include "iommu-priv.h"
> >
> > @@ -41,22 +42,41 @@ static int of_iommu_xlate(struct device *dev,
> > return ret;
> > }
> >
> > +/*
> > + * Callback to be called from of_map_id(), that tells if
> > + * all the mappings for an input id to be maintained in
> > + * parallel. Should this decission be from further layers,
> > + * iommu_driver->impl_ops?
> > + */
> > +static int of_iommu_configure_cb(struct of_map_id_arg *arg)
> > +{
> > + struct of_phandle_args *iommu_spec = &arg->map_args;
> > + struct device *dev = arg->dev;
> > + int err;
> > +
> > + err = of_iommu_xlate(dev, iommu_spec);
> > + of_node_put(iommu_spec->np);
> > +
> > + /* !iommu_spec->np may be from the bypassed translations */
> > + if (!err)
> > + err = (!arg->multi_map || !iommu_spec->np) ? 0 : -EAGAIN;
> > +
> > + return err;
> > +}
> > +
> > static int of_iommu_configure_dev_id(struct device_node *master_np,
> > struct device *dev,
> > const u32 *id)
> > {
> > struct of_map_id_arg arg = {
> > .map_args = {},
> > + .cb = of_iommu_configure_cb,
> > + .dev = dev,
> > + /* Should this be pushed to iommu_driver->impl_ops? */
> > + .multi_map = dev_is_platform(dev),
> > };
> > - int err;
> > -
> > - err = of_map_iommu_id(master_np, *id, &arg);
> > - if (err)
> > - return err;
> >
> > - err = of_iommu_xlate(dev, &arg.map_args);
> > - of_node_put(arg.map_args.np);
> > - return err;
> > + return of_map_iommu_id(master_np, *id, &arg);
> > }
> >
> > static int of_iommu_configure_dev(struct device_node *master_np,
> > diff --git a/drivers/of/base.c b/drivers/of/base.c
> > index
> > 606bef4f90e7d13bae4f7b0c45acd1755ad89826..a1c3c5954ec7e8eb3753c8fd782a1570f9
> > eb9c17 100644
> > --- a/drivers/of/base.c
> > +++ b/drivers/of/base.c
> > @@ -2122,14 +2122,21 @@ static bool of_check_bad_map(const __be32 *map, int
> > len)
> > return true;
> > }
> >
> > -static int of_map_id_fill_output(struct of_map_id_arg *arg,
> > - struct device_node *phandle_node, u32
> > id_or_offset,
> > - const __be32 *out_base, u32 cells,
> > - bool bypass)
> > +/*
> > + * Fill the id_out and target for the of_map_id() caller. Also
> > + * call the callback passed to the of_map_id() as part of the arg
> > + * that decides if to continue further search.
> > + */
> > +static int of_map_id_fill_arg(struct of_map_id_arg *arg,
> > + struct device_node *phandle_node, u32
> > id_or_offset,
> > + const __be32 *out_base, u32 cells,
> > + bool bypass, bool *multi_id_map)
> > {
> > + int ret;
> > +
> > if (bypass) {
> > arg->map_args.args[0] = id_or_offset;
> > - return 0;
> > + goto output;
> > }
> >
> > if (arg->map_args.np)
> > @@ -2145,7 +2152,14 @@ static int of_map_id_fill_output(struct of_map_id_arg
> > *arg,
> >
> > arg->map_args.args_count = cells;
> >
> > - return 0;
> > +output:
> > + /* pass the output for the callback, callers may further decide */
> > + ret = arg->cb ? arg->cb(arg) : 0;
> > +
> > + if (multi_id_map && ret == -EAGAIN)
> > + *multi_id_map = true;
> > +
> > + return ret;
> > }
> >
> > /**
> > @@ -2179,6 +2193,7 @@ int of_map_id(const struct device_node *np, u32 id,
> > const char *map_name,
> > int map_bytes, map_len, offset = 0;
> > bool bad_map = false;
> > const __be32 *map = NULL;
> > + bool multi_id_map = false;
> >
> > if (!np || !map_name || !arg)
> > return -EINVAL;
> > @@ -2264,23 +2279,26 @@ int of_map_id(const struct device_node *np, u32 id,
> > const char *map_name,
> > if (masked_id < id_base || id_off >= id_len)
> > continue;
> >
> > - ret = of_map_id_fill_output(arg, phandle_node, id_off,
> > out_base, cells, false);
> > + ret = of_map_id_fill_arg(arg, phandle_node, id_off,
> > out_base,
> > + cells, false, &multi_id_map);
> > if (ret == -EAGAIN)
> > continue;
> >
> > pr_debug("%pOF: %s, using mask %08x, id-base: %08x, out-
> > base: %08x, length: %08x, id: %08x -> %08x\n",
> > np, map_name, map_mask, id_base,
> > be32_to_cpup(out_base),
> > id_len, id, id_off + be32_to_cpup(out_base));
> > - return 0;
> > + return ret;
> > }
> >
> > + if (multi_id_map)
> > + return 0;
> > +
> > pr_info("%pOF: no %s translation for id 0x%x on %pOF\n", np,
> > map_name,
> > id, arg->map_args.np ? arg->map_args.np : NULL);
> >
> > bypass_translation:
> > /* Bypasses translation */
> > - return of_map_id_fill_output(arg, NULL, id, 0, 0, true);
> > -
> > + return of_map_id_fill_arg(arg, NULL, id, 0, 0, true, NULL);
> > err_map_len:
> > pr_err("%pOF: Error: Bad %s length: %d\n", np, map_name,
> > map_bytes);
> > return -EINVAL;
> > diff --git a/include/linux/of.h b/include/linux/of.h
> > index
> > 9efa6f93712c6024f05476f9fd39f3294f942ec1..abab73a76682351f5635c1127a6c899917
> > 525050 100644
> > --- a/include/linux/of.h
> > +++ b/include/linux/of.h
> > @@ -25,6 +25,9 @@
> > typedef u32 phandle;
> > typedef u32 ihandle;
> >
> > +struct of_map_id_arg;
> > +typedef int (*of_map_id_cb)(struct of_map_id_arg *arg);
> > +
> > struct property {
> > char *name;
> > int length;
> > @@ -76,6 +79,9 @@ struct of_phandle_args {
> >
> > struct of_map_id_arg {
> > struct of_phandle_args map_args;
> > + of_map_id_cb cb;
> > + struct device *dev;
> > + bool multi_map;
> > };
> >
> > struct of_phandle_iterator {
> >
> > --
> > 2.34.1
> >
© 2016 - 2026 Red Hat, Inc.