* [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions
@ 2026-06-01 20:45 Seth Forshee
2026-06-01 20:45 ` [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size Seth Forshee
` (2 more replies)
0 siblings, 3 replies; 14+ messages in thread
From: Seth Forshee @ 2026-06-01 20:45 UTC (permalink / raw)
To: Sudeep Holla, Sebastian Ene; +Cc: linux-arm-kernel, linux-kernel, Seth Forshee
Commit 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size
during RXTX_MAP") caused a regression for some FF-A implementations. It
has the effect of passing the size of the page-aligned buffers to
FFA_RXTX_MAP, which may be larger than the maximum size supported by the
SPMC. When this happens FFA_RXTX_MAP will fail with INVALID_PARAMETERS.
The following patches deal with two distinct scenarios which lead to
this issue. With FF-A v1.2+, FFA_FEATURES can specify a maximum supported
RX/TX buffer size, so patch 1 decodes this field and honors the maximum if
specified. For FF-A v1.1 and earlier the maximum is unknown, so patch 2
deals with this by first attempting FFA_RXTX_MAP with the page-aligned
buffer size (preserving the behavior introduced by commit 83210251fd70).
If this fails due to invalid parameters it retries with the minimum buffer
size from FFA_FEATURES.
Testing was done with FF-A v1.1 and v1.2 implementations, both of which
reject buffer sizes larger than 4K. Both implementations were tested with
4K, 16K, and 64K pages. Without these patches, probe fails for page sizes
larger than 4K with the message "failed to register FFA RxTx buffers."
With the patches probe succeeds for all page sizes.
The patches are based on for-next/ffa/fixes.
Signed-off-by: Seth Forshee <sforshee@nvidia•com>
---
Seth Forshee (2):
firmware: arm_ffa: Honor maximum RX/TX buffer size
firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails
drivers/firmware/arm_ffa/driver.c | 51 ++++++++++++++++++++++++++++++++++-----
1 file changed, 45 insertions(+), 6 deletions(-)
---
base-commit: a6848a50404eefb6f0b131c21881a2d8d21b31a9
change-id: 20260531-b4-ffa-rxtx-map-fixes-244ede71a935
Best regards,
--
Seth Forshee <sforshee@nvidia•com>
^ permalink raw reply [flat|nested] 14+ messages in thread
* [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size
2026-06-01 20:45 [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions Seth Forshee
@ 2026-06-01 20:45 ` Seth Forshee
2026-06-02 15:56 ` Sudeep Holla
2026-06-01 20:45 ` [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails Seth Forshee
2026-06-02 16:51 ` [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions Sudeep Holla
2 siblings, 1 reply; 14+ messages in thread
From: Seth Forshee @ 2026-06-01 20:45 UTC (permalink / raw)
To: Sudeep Holla, Sebastian Ene; +Cc: linux-arm-kernel, linux-kernel, Seth Forshee
FFA_FEATURES in v1.2+ supports a maximum RXTX_MAP size. This maximum
size is not checked when page-aligning the RX/TX buffer size, and
FFA_RXTX_MAP may fail when passed a size greater than the maximum.
Decode the maximum buffer size returned from FFA_FEATURES and limit the
buffer size based on this value if it is non-zero (zero indicates no
maximum). Include verification that the max returned by the SPMC is
larger than the minimum, otherwise use the minimum.
While there, also update RXTX_MAP_MIN_BUFSZ() to use FIELD_GET() for
consistency.
Fixes: 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size during RXTX_MAP")
Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Seth Forshee <sforshee@nvidia•com>
---
drivers/firmware/arm_ffa/driver.c | 29 +++++++++++++++++++++++++++--
1 file changed, 27 insertions(+), 2 deletions(-)
diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
index b9f17fda7243..dc45724a29ba 100644
--- a/drivers/firmware/arm_ffa/driver.c
+++ b/drivers/firmware/arm_ffa/driver.c
@@ -32,6 +32,7 @@
#include <linux/interrupt.h>
#include <linux/io.h>
#include <linux/kernel.h>
+#include <linux/minmax.h>
#include <linux/module.h>
#include <linux/mm.h>
#include <linux/mutex.h>
@@ -55,7 +56,9 @@
(FIELD_PREP(SENDER_ID_MASK, (s)) | FIELD_PREP(RECEIVER_ID_MASK, (r)))
#define RXTX_MAP_MIN_BUFSZ_MASK GENMASK(1, 0)
-#define RXTX_MAP_MIN_BUFSZ(x) ((x) & RXTX_MAP_MIN_BUFSZ_MASK)
+#define RXTX_MAP_MAX_BUFSZ_MASK GENMASK(31, 16)
+#define RXTX_MAP_MIN_BUFSZ(x) (FIELD_GET(RXTX_MAP_MIN_BUFSZ_MASK, (x)))
+#define RXTX_MAP_MAX_BUFSZ(x) (FIELD_GET(RXTX_MAP_MAX_BUFSZ_MASK, (x)))
#define FFA_MAX_NOTIFICATIONS 64
@@ -2086,11 +2089,13 @@ static void ffa_notifications_setup(void)
ffa_notifications_cleanup();
}
+#define FFA_SUPPORTS_RXTX_MAX_BUFSZ(version) ((version) > FFA_VERSION_1_1)
+
static int __init ffa_init(void)
{
int ret;
u32 buf_sz;
- size_t rxtx_bufsz = SZ_4K;
+ size_t rxtx_bufsz = SZ_4K, rxtx_max_bufsz = 0;
ret = ffa_transport_init(&invoke_ffa_fn);
if (ret)
@@ -2118,9 +2123,29 @@ static int __init ffa_init(void)
rxtx_bufsz = SZ_16K;
else
rxtx_bufsz = SZ_4K;
+
+ if (FFA_SUPPORTS_RXTX_MAX_BUFSZ(drv_info->version)) {
+ rxtx_max_bufsz = (size_t)RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
+ if (rxtx_max_bufsz != 0 && rxtx_max_bufsz < rxtx_bufsz) {
+ /*
+ * Per spec the maximum must be >= the minimum, or
+ * else zero if there is no size limit. If the SPMC
+ * violates this constraint, use the minimum as the
+ * effective maximum.
+ */
+ rxtx_max_bufsz = rxtx_bufsz;
+ }
+ }
}
+ /*
+ * alloc_pages_exact() allocates full pages. Use the full allocated
+ * space up to the max supported by the SPMC.
+ */
rxtx_bufsz = PAGE_ALIGN(rxtx_bufsz);
+ if (rxtx_max_bufsz)
+ rxtx_bufsz = min(rxtx_bufsz, rxtx_max_bufsz);
+
drv_info->rxtx_bufsz = rxtx_bufsz;
drv_info->rx_buffer = alloc_pages_exact(rxtx_bufsz, GFP_KERNEL);
if (!drv_info->rx_buffer) {
--
2.43.0
^ permalink raw reply related [flat|nested] 14+ messages in thread
* [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails
2026-06-01 20:45 [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions Seth Forshee
2026-06-01 20:45 ` [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size Seth Forshee
@ 2026-06-01 20:45 ` Seth Forshee
2026-06-02 15:48 ` Sudeep Holla
2026-06-02 16:51 ` [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions Sudeep Holla
2 siblings, 1 reply; 14+ messages in thread
From: Seth Forshee @ 2026-06-01 20:45 UTC (permalink / raw)
To: Sudeep Holla, Sebastian Ene; +Cc: linux-arm-kernel, linux-kernel, Seth Forshee
FFA_FEATURES only supports a maximum size value for RXTX_MAP in v1.2+.
Older implementations may still reject larger sizes, which can happen
when the minimum size is rounded up to the page size when PAGE_SIZE >
4K.
Fall back to the minimum size when RXTX_MAP fails with
INVALID_PARAMETERS to fix the regression on these platforms. This also
has the side effect of falling back for v1.2+ with a maximum of 0 (i.e.
no maximum), where the rounded-up value isn't expected to be rejected,
but there's no harm in retrying with a smaller size and it might even be
of benefit for a non-compliant SPMC.
Note that this may result in passing a size smaller than the allocated
size to free_pages_exact(). This does not result in any leaked memory,
since the size extends into all allocated pages, and free_pages_exact()
correctly handles sizes which are not page aligned.
Fixes: 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size during RXTX_MAP")
Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Seth Forshee <sforshee@nvidia•com>
---
drivers/firmware/arm_ffa/driver.c | 28 +++++++++++++++++++++-------
1 file changed, 21 insertions(+), 7 deletions(-)
diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
index dc45724a29ba..60ad58e229ce 100644
--- a/drivers/firmware/arm_ffa/driver.c
+++ b/drivers/firmware/arm_ffa/driver.c
@@ -2095,7 +2095,7 @@ static int __init ffa_init(void)
{
int ret;
u32 buf_sz;
- size_t rxtx_bufsz = SZ_4K, rxtx_max_bufsz = 0;
+ size_t rxtx_min_bufsz = SZ_4K, rxtx_max_bufsz = 0, rxtx_bufsz;
ret = ffa_transport_init(&invoke_ffa_fn);
if (ret)
@@ -2118,22 +2118,22 @@ static int __init ffa_init(void)
ret = ffa_features(FFA_FN_NATIVE(RXTX_MAP), 0, &buf_sz, NULL);
if (!ret) {
if (RXTX_MAP_MIN_BUFSZ(buf_sz) == 1)
- rxtx_bufsz = SZ_64K;
+ rxtx_min_bufsz = SZ_64K;
else if (RXTX_MAP_MIN_BUFSZ(buf_sz) == 2)
- rxtx_bufsz = SZ_16K;
+ rxtx_min_bufsz = SZ_16K;
else
- rxtx_bufsz = SZ_4K;
+ rxtx_min_bufsz = SZ_4K;
if (FFA_SUPPORTS_RXTX_MAX_BUFSZ(drv_info->version)) {
rxtx_max_bufsz = (size_t)RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
- if (rxtx_max_bufsz != 0 && rxtx_max_bufsz < rxtx_bufsz) {
+ if (rxtx_max_bufsz != 0 && rxtx_max_bufsz < rxtx_min_bufsz) {
/*
* Per spec the maximum must be >= the minimum, or
* else zero if there is no size limit. If the SPMC
* violates this constraint, use the minimum as the
* effective maximum.
*/
- rxtx_max_bufsz = rxtx_bufsz;
+ rxtx_max_bufsz = rxtx_min_bufsz;
}
}
}
@@ -2142,7 +2142,7 @@ static int __init ffa_init(void)
* alloc_pages_exact() allocates full pages. Use the full allocated
* space up to the max supported by the SPMC.
*/
- rxtx_bufsz = PAGE_ALIGN(rxtx_bufsz);
+ rxtx_bufsz = PAGE_ALIGN(rxtx_min_bufsz);
if (rxtx_max_bufsz)
rxtx_bufsz = min(rxtx_bufsz, rxtx_max_bufsz);
@@ -2162,6 +2162,20 @@ static int __init ffa_init(void)
ret = ffa_rxtx_map(virt_to_phys(drv_info->tx_buffer),
virt_to_phys(drv_info->rx_buffer),
rxtx_bufsz / FFA_PAGE_SIZE);
+ if (ret == -EINVAL && rxtx_max_bufsz == 0 && rxtx_min_bufsz < rxtx_bufsz) {
+ /*
+ * FFA_FEATURES only supports maximum buffer size in v1.2+,
+ * and some implementations without it may fail when the
+ * buffer size is rounded up to a larger page size. If
+ * RXTX_MAP fails due to invalid parameters, try again with
+ * the minimum buffer size.
+ */
+ rxtx_bufsz = rxtx_min_bufsz;
+ drv_info->rxtx_bufsz = rxtx_bufsz;
+ ret = ffa_rxtx_map(virt_to_phys(drv_info->tx_buffer),
+ virt_to_phys(drv_info->rx_buffer),
+ rxtx_bufsz / FFA_PAGE_SIZE);
+ }
if (ret) {
pr_err("failed to register FFA RxTx buffers\n");
goto free_pages;
--
2.43.0
^ permalink raw reply related [flat|nested] 14+ messages in thread
* Re: [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails
2026-06-01 20:45 ` [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails Seth Forshee
@ 2026-06-02 15:48 ` Sudeep Holla
2026-06-02 16:21 ` Seth Forshee
0 siblings, 1 reply; 14+ messages in thread
From: Sudeep Holla @ 2026-06-02 15:48 UTC (permalink / raw)
To: Seth Forshee; +Cc: Sebastian Ene, Sudeep Holla, linux-arm-kernel, linux-kernel
On Mon, Jun 01, 2026 at 03:45:12PM -0500, Seth Forshee wrote:
> FFA_FEATURES only supports a maximum size value for RXTX_MAP in v1.2+.
> Older implementations may still reject larger sizes, which can happen
> when the minimum size is rounded up to the page size when PAGE_SIZE >
> 4K.
>
> Fall back to the minimum size when RXTX_MAP fails with
> INVALID_PARAMETERS to fix the regression on these platforms. This also
> has the side effect of falling back for v1.2+ with a maximum of 0 (i.e.
> no maximum), where the rounded-up value isn't expected to be rejected,
> but there's no harm in retrying with a smaller size and it might even be
> of benefit for a non-compliant SPMC.
>
Is this needed on any of your platforms ? I don't want to push a solution
or fallback/workaround if there are no users for it.
--
Regards,
Sudeep
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size
2026-06-01 20:45 ` [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size Seth Forshee
@ 2026-06-02 15:56 ` Sudeep Holla
2026-06-02 16:12 ` Seth Forshee
0 siblings, 1 reply; 14+ messages in thread
From: Sudeep Holla @ 2026-06-02 15:56 UTC (permalink / raw)
To: Seth Forshee; +Cc: Sebastian Ene, Sudeep Holla, linux-arm-kernel, linux-kernel
On Mon, Jun 01, 2026 at 03:45:11PM -0500, Seth Forshee wrote:
> FFA_FEATURES in v1.2+ supports a maximum RXTX_MAP size. This maximum
> size is not checked when page-aligning the RX/TX buffer size, and
> FFA_RXTX_MAP may fail when passed a size greater than the maximum.
>
> Decode the maximum buffer size returned from FFA_FEATURES and limit the
> buffer size based on this value if it is non-zero (zero indicates no
> maximum). Include verification that the max returned by the SPMC is
> larger than the minimum, otherwise use the minimum.
>
> While there, also update RXTX_MAP_MIN_BUFSZ() to use FIELD_GET() for
> consistency.
>
> Fixes: 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size during RXTX_MAP")
> Assisted-by: Claude:claude-opus-4-8
> Signed-off-by: Seth Forshee <sforshee@nvidia•com>
> ---
> drivers/firmware/arm_ffa/driver.c | 29 +++++++++++++++++++++++++++--
> 1 file changed, 27 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
> index b9f17fda7243..dc45724a29ba 100644
> --- a/drivers/firmware/arm_ffa/driver.c
> +++ b/drivers/firmware/arm_ffa/driver.c
> @@ -32,6 +32,7 @@
> #include <linux/interrupt.h>
> #include <linux/io.h>
> #include <linux/kernel.h>
> +#include <linux/minmax.h>
> #include <linux/module.h>
> #include <linux/mm.h>
> #include <linux/mutex.h>
> @@ -55,7 +56,9 @@
> (FIELD_PREP(SENDER_ID_MASK, (s)) | FIELD_PREP(RECEIVER_ID_MASK, (r)))
>
> #define RXTX_MAP_MIN_BUFSZ_MASK GENMASK(1, 0)
> -#define RXTX_MAP_MIN_BUFSZ(x) ((x) & RXTX_MAP_MIN_BUFSZ_MASK)
> +#define RXTX_MAP_MAX_BUFSZ_MASK GENMASK(31, 16)
> +#define RXTX_MAP_MIN_BUFSZ(x) (FIELD_GET(RXTX_MAP_MIN_BUFSZ_MASK, (x)))
> +#define RXTX_MAP_MAX_BUFSZ(x) (FIELD_GET(RXTX_MAP_MAX_BUFSZ_MASK, (x)))
>
> #define FFA_MAX_NOTIFICATIONS 64
>
> @@ -2086,11 +2089,13 @@ static void ffa_notifications_setup(void)
> ffa_notifications_cleanup();
> }
>
> +#define FFA_SUPPORTS_RXTX_MAX_BUFSZ(version) ((version) > FFA_VERSION_1_1)
> +
> static int __init ffa_init(void)
> {
> int ret;
> u32 buf_sz;
> - size_t rxtx_bufsz = SZ_4K;
> + size_t rxtx_bufsz = SZ_4K, rxtx_max_bufsz = 0;
>
> ret = ffa_transport_init(&invoke_ffa_fn);
> if (ret)
> @@ -2118,9 +2123,29 @@ static int __init ffa_init(void)
> rxtx_bufsz = SZ_16K;
> else
> rxtx_bufsz = SZ_4K;
> +
> + if (FFA_SUPPORTS_RXTX_MAX_BUFSZ(drv_info->version)) {
This is not even compile tested ? I don't think this is defined anyway.
Anyways I don't think we need this condition as v1.1 or below had it
MBZ, so it is fine to drop it.
> + rxtx_max_bufsz = (size_t)RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
The cast is unnecessary, its is 16 bit, so max 0xFFFF << 12.
> + if (rxtx_max_bufsz != 0 && rxtx_max_bufsz < rxtx_bufsz) {
> + /*
> + * Per spec the maximum must be >= the minimum, or
> + * else zero if there is no size limit. If the SPMC
> + * violates this constraint, use the minimum as the
> + * effective maximum.
> + */
I don't see any text implying the above from the spec, drop it. Please point
it to me if you find. Even otherwise it seems obvious here and doesn't need
the note.
--
Regards,
Sudeep
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size
2026-06-02 15:56 ` Sudeep Holla
@ 2026-06-02 16:12 ` Seth Forshee
2026-06-02 16:27 ` Sudeep Holla
0 siblings, 1 reply; 14+ messages in thread
From: Seth Forshee @ 2026-06-02 16:12 UTC (permalink / raw)
To: Sudeep Holla; +Cc: Sebastian Ene, linux-arm-kernel, linux-kernel
On Tue, Jun 02, 2026 at 04:56:48PM +0100, Sudeep Holla wrote:
> On Mon, Jun 01, 2026 at 03:45:11PM -0500, Seth Forshee wrote:
> > FFA_FEATURES in v1.2+ supports a maximum RXTX_MAP size. This maximum
> > size is not checked when page-aligning the RX/TX buffer size, and
> > FFA_RXTX_MAP may fail when passed a size greater than the maximum.
> >
> > Decode the maximum buffer size returned from FFA_FEATURES and limit the
> > buffer size based on this value if it is non-zero (zero indicates no
> > maximum). Include verification that the max returned by the SPMC is
> > larger than the minimum, otherwise use the minimum.
> >
> > While there, also update RXTX_MAP_MIN_BUFSZ() to use FIELD_GET() for
> > consistency.
> >
> > Fixes: 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size during RXTX_MAP")
> > Assisted-by: Claude:claude-opus-4-8
> > Signed-off-by: Seth Forshee <sforshee@nvidia•com>
> > ---
> > drivers/firmware/arm_ffa/driver.c | 29 +++++++++++++++++++++++++++--
> > 1 file changed, 27 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
> > index b9f17fda7243..dc45724a29ba 100644
> > --- a/drivers/firmware/arm_ffa/driver.c
> > +++ b/drivers/firmware/arm_ffa/driver.c
> > @@ -32,6 +32,7 @@
> > #include <linux/interrupt.h>
> > #include <linux/io.h>
> > #include <linux/kernel.h>
> > +#include <linux/minmax.h>
> > #include <linux/module.h>
> > #include <linux/mm.h>
> > #include <linux/mutex.h>
> > @@ -55,7 +56,9 @@
> > (FIELD_PREP(SENDER_ID_MASK, (s)) | FIELD_PREP(RECEIVER_ID_MASK, (r)))
> >
> > #define RXTX_MAP_MIN_BUFSZ_MASK GENMASK(1, 0)
> > -#define RXTX_MAP_MIN_BUFSZ(x) ((x) & RXTX_MAP_MIN_BUFSZ_MASK)
> > +#define RXTX_MAP_MAX_BUFSZ_MASK GENMASK(31, 16)
> > +#define RXTX_MAP_MIN_BUFSZ(x) (FIELD_GET(RXTX_MAP_MIN_BUFSZ_MASK, (x)))
> > +#define RXTX_MAP_MAX_BUFSZ(x) (FIELD_GET(RXTX_MAP_MAX_BUFSZ_MASK, (x)))
> >
> > #define FFA_MAX_NOTIFICATIONS 64
> >
> > @@ -2086,11 +2089,13 @@ static void ffa_notifications_setup(void)
> > ffa_notifications_cleanup();
> > }
> >
> > +#define FFA_SUPPORTS_RXTX_MAX_BUFSZ(version) ((version) > FFA_VERSION_1_1)
> > +
> > static int __init ffa_init(void)
> > {
> > int ret;
> > u32 buf_sz;
> > - size_t rxtx_bufsz = SZ_4K;
> > + size_t rxtx_bufsz = SZ_4K, rxtx_max_bufsz = 0;
> >
> > ret = ffa_transport_init(&invoke_ffa_fn);
> > if (ret)
> > @@ -2118,9 +2123,29 @@ static int __init ffa_init(void)
> > rxtx_bufsz = SZ_16K;
> > else
> > rxtx_bufsz = SZ_4K;
> > +
> > + if (FFA_SUPPORTS_RXTX_MAX_BUFSZ(drv_info->version)) {
>
> This is not even compile tested ? I don't think this is defined anyway.
> Anyways I don't think we need this condition as v1.1 or below had it
> MBZ, so it is fine to drop it.
It is added just above ffa_init(). The code as been compile and run
tested on multiple platforms, as noted in the cover letter.
I hadn't checked older spec versions, so if they have it as MBZ then I
agree we can drop the check.
>
> > + rxtx_max_bufsz = (size_t)RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
>
> The cast is unnecessary, its is 16 bit, so max 0xFFFF << 12.
Correct, though linters complain about it which is why I added it. I can
drop it though if you prefer.
>
>
> > + if (rxtx_max_bufsz != 0 && rxtx_max_bufsz < rxtx_bufsz) {
> > + /*
> > + * Per spec the maximum must be >= the minimum, or
> > + * else zero if there is no size limit. If the SPMC
> > + * violates this constraint, use the minimum as the
> > + * effective maximum.
> > + */
>
> I don't see any text implying the above from the spec, drop it. Please point
> it to me if you find. Even otherwise it seems obvious here and doesn't need
> the note.
I found it in v1.3 of the spec in Table 13.15: "Size must be greater or
equal to the minimum buffer size, else MBZ if there is no size limit." I
will remove the comment.
Thanks,
Seth
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails
2026-06-02 15:48 ` Sudeep Holla
@ 2026-06-02 16:21 ` Seth Forshee
2026-06-02 16:29 ` Sudeep Holla
0 siblings, 1 reply; 14+ messages in thread
From: Seth Forshee @ 2026-06-02 16:21 UTC (permalink / raw)
To: Sudeep Holla; +Cc: Sebastian Ene, linux-arm-kernel, linux-kernel
On Tue, Jun 02, 2026 at 04:48:11PM +0100, Sudeep Holla wrote:
> On Mon, Jun 01, 2026 at 03:45:12PM -0500, Seth Forshee wrote:
> > FFA_FEATURES only supports a maximum size value for RXTX_MAP in v1.2+.
> > Older implementations may still reject larger sizes, which can happen
> > when the minimum size is rounded up to the page size when PAGE_SIZE >
> > 4K.
> >
> > Fall back to the minimum size when RXTX_MAP fails with
> > INVALID_PARAMETERS to fix the regression on these platforms. This also
> > has the side effect of falling back for v1.2+ with a maximum of 0 (i.e.
> > no maximum), where the rounded-up value isn't expected to be rejected,
> > but there's no harm in retrying with a smaller size and it might even be
> > of benefit for a non-compliant SPMC.
> >
>
> Is this needed on any of your platforms ? I don't want to push a solution
> or fallback/workaround if there are no users for it.
Yes, we have platforms for which commit 83210251fd70 causes RXTX_MAP
failures with page sizes larger than 4K.
Thanks,
Seth
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size
2026-06-02 16:12 ` Seth Forshee
@ 2026-06-02 16:27 ` Sudeep Holla
2026-06-02 17:05 ` Seth Forshee
0 siblings, 1 reply; 14+ messages in thread
From: Sudeep Holla @ 2026-06-02 16:27 UTC (permalink / raw)
To: Seth Forshee; +Cc: Sebastian Ene, Sudeep Holla, linux-arm-kernel, linux-kernel
On Tue, Jun 02, 2026 at 11:12:03AM -0500, Seth Forshee wrote:
> On Tue, Jun 02, 2026 at 04:56:48PM +0100, Sudeep Holla wrote:
> > On Mon, Jun 01, 2026 at 03:45:11PM -0500, Seth Forshee wrote:
> > > FFA_FEATURES in v1.2+ supports a maximum RXTX_MAP size. This maximum
> > > size is not checked when page-aligning the RX/TX buffer size, and
> > > FFA_RXTX_MAP may fail when passed a size greater than the maximum.
> > >
> > > Decode the maximum buffer size returned from FFA_FEATURES and limit the
> > > buffer size based on this value if it is non-zero (zero indicates no
> > > maximum). Include verification that the max returned by the SPMC is
> > > larger than the minimum, otherwise use the minimum.
> > >
> > > While there, also update RXTX_MAP_MIN_BUFSZ() to use FIELD_GET() for
> > > consistency.
> > >
> > > Fixes: 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size during RXTX_MAP")
> > > Assisted-by: Claude:claude-opus-4-8
> > > Signed-off-by: Seth Forshee <sforshee@nvidia•com>
> > > ---
> > > drivers/firmware/arm_ffa/driver.c | 29 +++++++++++++++++++++++++++--
> > > 1 file changed, 27 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
> > > index b9f17fda7243..dc45724a29ba 100644
> > > --- a/drivers/firmware/arm_ffa/driver.c
> > > +++ b/drivers/firmware/arm_ffa/driver.c
> > > @@ -32,6 +32,7 @@
> > > #include <linux/interrupt.h>
> > > #include <linux/io.h>
> > > #include <linux/kernel.h>
> > > +#include <linux/minmax.h>
> > > #include <linux/module.h>
> > > #include <linux/mm.h>
> > > #include <linux/mutex.h>
> > > @@ -55,7 +56,9 @@
> > > (FIELD_PREP(SENDER_ID_MASK, (s)) | FIELD_PREP(RECEIVER_ID_MASK, (r)))
> > >
> > > #define RXTX_MAP_MIN_BUFSZ_MASK GENMASK(1, 0)
> > > -#define RXTX_MAP_MIN_BUFSZ(x) ((x) & RXTX_MAP_MIN_BUFSZ_MASK)
> > > +#define RXTX_MAP_MAX_BUFSZ_MASK GENMASK(31, 16)
> > > +#define RXTX_MAP_MIN_BUFSZ(x) (FIELD_GET(RXTX_MAP_MIN_BUFSZ_MASK, (x)))
> > > +#define RXTX_MAP_MAX_BUFSZ(x) (FIELD_GET(RXTX_MAP_MAX_BUFSZ_MASK, (x)))
> > >
> > > #define FFA_MAX_NOTIFICATIONS 64
> > >
> > > @@ -2086,11 +2089,13 @@ static void ffa_notifications_setup(void)
> > > ffa_notifications_cleanup();
> > > }
> > >
> > > +#define FFA_SUPPORTS_RXTX_MAX_BUFSZ(version) ((version) > FFA_VERSION_1_1)
> > > +
> > > static int __init ffa_init(void)
> > > {
> > > int ret;
> > > u32 buf_sz;
> > > - size_t rxtx_bufsz = SZ_4K;
> > > + size_t rxtx_bufsz = SZ_4K, rxtx_max_bufsz = 0;
> > >
> > > ret = ffa_transport_init(&invoke_ffa_fn);
> > > if (ret)
> > > @@ -2118,9 +2123,29 @@ static int __init ffa_init(void)
> > > rxtx_bufsz = SZ_16K;
> > > else
> > > rxtx_bufsz = SZ_4K;
> > > +
> > > + if (FFA_SUPPORTS_RXTX_MAX_BUFSZ(drv_info->version)) {
> >
> > This is not even compile tested ? I don't think this is defined anyway.
> > Anyways I don't think we need this condition as v1.1 or below had it
> > MBZ, so it is fine to drop it.
>
> It is added just above ffa_init(). The code as been compile and run
> tested on multiple platforms, as noted in the cover letter.
>
Ah sorry, my bad. It failed to apply and I must have messed it up.
I have some changes in linux-next which conflicts and I dropped it when
resolving. Sorry for that.
> I hadn't checked older spec versions, so if they have it as MBZ then I
> agree we can drop the check.
>
> >
> > > + rxtx_max_bufsz = (size_t)RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
> >
> > The cast is unnecessary, its is 16 bit, so max 0xFFFF << 12.
>
> Correct, though linters complain about it which is why I added it. I can
> drop it though if you prefer.
>
Can you be more specific ? Any way to trigger it ?
> >
> >
> > > + if (rxtx_max_bufsz != 0 && rxtx_max_bufsz < rxtx_bufsz) {
> > > + /*
> > > + * Per spec the maximum must be >= the minimum, or
> > > + * else zero if there is no size limit. If the SPMC
> > > + * violates this constraint, use the minimum as the
> > > + * effective maximum.
> > > + */
> >
> > I don't see any text implying the above from the spec, drop it. Please point
> > it to me if you find. Even otherwise it seems obvious here and doesn't need
> > the note.
>
> I found it in v1.3 of the spec in Table 13.15: "Size must be greater or
> equal to the minimum buffer size, else MBZ if there is no size limit." I
> will remove the comment.
>
Ah v1.3, sorry I was just looking at v1.2.
> Thanks,
> Seth
--
Regards,
Sudeep
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails
2026-06-02 16:21 ` Seth Forshee
@ 2026-06-02 16:29 ` Sudeep Holla
2026-06-02 17:12 ` Seth Forshee
0 siblings, 1 reply; 14+ messages in thread
From: Sudeep Holla @ 2026-06-02 16:29 UTC (permalink / raw)
To: Seth Forshee; +Cc: Sebastian Ene, Sudeep Holla, linux-arm-kernel, linux-kernel
On Tue, Jun 02, 2026 at 11:21:33AM -0500, Seth Forshee wrote:
> On Tue, Jun 02, 2026 at 04:48:11PM +0100, Sudeep Holla wrote:
> > On Mon, Jun 01, 2026 at 03:45:12PM -0500, Seth Forshee wrote:
> > > FFA_FEATURES only supports a maximum size value for RXTX_MAP in v1.2+.
> > > Older implementations may still reject larger sizes, which can happen
> > > when the minimum size is rounded up to the page size when PAGE_SIZE >
> > > 4K.
> > >
> > > Fall back to the minimum size when RXTX_MAP fails with
> > > INVALID_PARAMETERS to fix the regression on these platforms. This also
> > > has the side effect of falling back for v1.2+ with a maximum of 0 (i.e.
> > > no maximum), where the rounded-up value isn't expected to be rejected,
> > > but there's no harm in retrying with a smaller size and it might even be
> > > of benefit for a non-compliant SPMC.
> > >
> >
> > Is this needed on any of your platforms ? I don't want to push a solution
> > or fallback/workaround if there are no users for it.
>
> Yes, we have platforms for which commit 83210251fd70 causes RXTX_MAP
> failures with page sizes larger than 4K.
>
And they are FF-A v1.1 or doesn't set max_buf size ? Sorry for being
pedantic, just want to be sure it is useful on real platforms before
pushing it as fix/workaround.
--
Regards,
Sudeep
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions
2026-06-01 20:45 [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions Seth Forshee
2026-06-01 20:45 ` [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size Seth Forshee
2026-06-01 20:45 ` [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails Seth Forshee
@ 2026-06-02 16:51 ` Sudeep Holla
2026-06-02 17:19 ` Seth Forshee
2 siblings, 1 reply; 14+ messages in thread
From: Sudeep Holla @ 2026-06-02 16:51 UTC (permalink / raw)
To: Seth Forshee; +Cc: Sebastian Ene, Sudeep Holla, linux-arm-kernel, linux-kernel
On Mon, Jun 01, 2026 at 03:45:10PM -0500, Seth Forshee wrote:
> Commit 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size
> during RXTX_MAP") caused a regression for some FF-A implementations. It
> has the effect of passing the size of the page-aligned buffers to
> FFA_RXTX_MAP, which may be larger than the maximum size supported by the
> SPMC. When this happens FFA_RXTX_MAP will fail with INVALID_PARAMETERS.
>
> The following patches deal with two distinct scenarios which lead to
> this issue. With FF-A v1.2+, FFA_FEATURES can specify a maximum supported
> RX/TX buffer size, so patch 1 decodes this field and honors the maximum if
> specified. For FF-A v1.1 and earlier the maximum is unknown, so patch 2
> deals with this by first attempting FFA_RXTX_MAP with the page-aligned
> buffer size (preserving the behavior introduced by commit 83210251fd70).
> If this fails due to invalid parameters it retries with the minimum buffer
> size from FFA_FEATURES.
>
> Testing was done with FF-A v1.1 and v1.2 implementations, both of which
> reject buffer sizes larger than 4K. Both implementations were tested with
> 4K, 16K, and 64K pages. Without these patches, probe fails for page sizes
> larger than 4K with the message "failed to register FFA RxTx buffers."
> With the patches probe succeeds for all page sizes.
>
> The patches are based on for-next/ffa/fixes.
for-next/ffa/updates has patches queued for v7.2.
>
> Signed-off-by: Seth Forshee <sforshee@nvidia•com>
> ---
> Seth Forshee (2):
> firmware: arm_ffa: Honor maximum RX/TX buffer size
> firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails
>
> drivers/firmware/arm_ffa/driver.c | 51 ++++++++++++++++++++++++++++++++++-----
> 1 file changed, 45 insertions(+), 6 deletions(-)
Wondering if these 2 fixes can be merged into one and simplified something
like(untested) patch below(rebased on linux-next or for-next/ffa/updates)
Regards,
Sudeep
-->8
diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
index 0f468362c288..5ffe21c568b7 100644
--- a/drivers/firmware/arm_ffa/driver.c
+++ b/drivers/firmware/arm_ffa/driver.c
@@ -32,6 +32,7 @@
#include <linux/interrupt.h>
#include <linux/io.h>
#include <linux/kernel.h>
+#include <linux/minmax.h>
#include <linux/module.h>
#include <linux/mm.h>
#include <linux/mutex.h>
@@ -59,7 +60,9 @@
(FIELD_PREP(SENDER_ID_MASK, (s)) | FIELD_PREP(RECEIVER_ID_MASK, (r)))
#define RXTX_MAP_MIN_BUFSZ_MASK GENMASK(1, 0)
-#define RXTX_MAP_MIN_BUFSZ(x) ((x) & RXTX_MAP_MIN_BUFSZ_MASK)
+#define RXTX_MAP_MAX_BUFSZ_MASK GENMASK(31, 16)
+#define RXTX_MAP_MIN_BUFSZ(x) (FIELD_GET(RXTX_MAP_MIN_BUFSZ_MASK, (x)))
+#define RXTX_MAP_MAX_BUFSZ(x) (FIELD_GET(RXTX_MAP_MAX_BUFSZ_MASK, (x)))
#define FFA_MAX_NOTIFICATIONS 64
@@ -2101,7 +2104,7 @@ static int ffa_probe(struct platform_device *pdev)
{
int ret;
u32 buf_sz;
- size_t rxtx_bufsz = SZ_4K;
+ size_t rxtx_min_bufsz = SZ_4K, rxtx_bufsz, rxtx_max_bufsz = 0;
if (IS_BUILTIN(CONFIG_ARM_FFA_TRANSPORT) &&
is_protected_kvm_enabled() && !is_pkvm_initialized())
@@ -2132,15 +2135,18 @@ static int ffa_probe(struct platform_device *pdev)
ret = ffa_features(FFA_FN_NATIVE(RXTX_MAP), 0, &buf_sz, NULL);
if (!ret) {
if (RXTX_MAP_MIN_BUFSZ(buf_sz) == 1)
- rxtx_bufsz = SZ_64K;
+ rxtx_min_bufsz = SZ_64K;
else if (RXTX_MAP_MIN_BUFSZ(buf_sz) == 2)
- rxtx_bufsz = SZ_16K;
+ rxtx_min_bufsz = SZ_16K;
else
- rxtx_bufsz = SZ_4K;
+ rxtx_min_bufsz = SZ_4K;
+
+ rxtx_max_bufsz = RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
+ if (rxtx_max_bufsz != 0 && rxtx_max_bufsz < rxtx_min_bufsz)
+ rxtx_max_bufsz = rxtx_min_bufsz;
}
- rxtx_bufsz = PAGE_ALIGN(rxtx_bufsz);
- drv_info->rxtx_bufsz = rxtx_bufsz;
+ rxtx_bufsz = min_not_zero(PAGE_ALIGN(rxtx_min_bufsz), rxtx_max_bufsz);
drv_info->rx_buffer = alloc_pages_exact(rxtx_bufsz, GFP_KERNEL);
if (!drv_info->rx_buffer) {
ret = -ENOMEM;
@@ -2156,10 +2162,17 @@ static int ffa_probe(struct platform_device *pdev)
ret = ffa_rxtx_map(virt_to_phys(drv_info->tx_buffer),
virt_to_phys(drv_info->rx_buffer),
rxtx_bufsz / FFA_PAGE_SIZE);
+ if (ret == -EINVAL && !rxtx_max_bufsz && rxtx_min_bufsz < rxtx_bufsz) {
+ rxtx_bufsz = rxtx_min_bufsz;
+ ret = ffa_rxtx_map(virt_to_phys(drv_info->tx_buffer),
+ virt_to_phys(drv_info->rx_buffer),
+ rxtx_bufsz / FFA_PAGE_SIZE);
+ }
if (ret) {
pr_err("failed to register FFA RxTx buffers\n");
goto free_pages;
}
+ drv_info->rxtx_bufsz = rxtx_bufsz;
mutex_init(&drv_info->rx_lock);
mutex_init(&drv_info->tx_lock);
^ permalink raw reply related [flat|nested] 14+ messages in thread
* Re: [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size
2026-06-02 16:27 ` Sudeep Holla
@ 2026-06-02 17:05 ` Seth Forshee
2026-06-02 21:58 ` Seth Forshee
0 siblings, 1 reply; 14+ messages in thread
From: Seth Forshee @ 2026-06-02 17:05 UTC (permalink / raw)
To: Sudeep Holla; +Cc: Sebastian Ene, linux-arm-kernel, linux-kernel
On Tue, Jun 02, 2026 at 05:27:48PM +0100, Sudeep Holla wrote:
> On Tue, Jun 02, 2026 at 11:12:03AM -0500, Seth Forshee wrote:
> > On Tue, Jun 02, 2026 at 04:56:48PM +0100, Sudeep Holla wrote:
> > > On Mon, Jun 01, 2026 at 03:45:11PM -0500, Seth Forshee wrote:
> > > > FFA_FEATURES in v1.2+ supports a maximum RXTX_MAP size. This maximum
> > > > size is not checked when page-aligning the RX/TX buffer size, and
> > > > FFA_RXTX_MAP may fail when passed a size greater than the maximum.
> > > >
> > > > Decode the maximum buffer size returned from FFA_FEATURES and limit the
> > > > buffer size based on this value if it is non-zero (zero indicates no
> > > > maximum). Include verification that the max returned by the SPMC is
> > > > larger than the minimum, otherwise use the minimum.
> > > >
> > > > While there, also update RXTX_MAP_MIN_BUFSZ() to use FIELD_GET() for
> > > > consistency.
> > > >
> > > > Fixes: 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size during RXTX_MAP")
> > > > Assisted-by: Claude:claude-opus-4-8
> > > > Signed-off-by: Seth Forshee <sforshee@nvidia•com>
> > > > ---
> > > > drivers/firmware/arm_ffa/driver.c | 29 +++++++++++++++++++++++++++--
> > > > 1 file changed, 27 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
> > > > index b9f17fda7243..dc45724a29ba 100644
> > > > --- a/drivers/firmware/arm_ffa/driver.c
> > > > +++ b/drivers/firmware/arm_ffa/driver.c
> > > > @@ -32,6 +32,7 @@
> > > > #include <linux/interrupt.h>
> > > > #include <linux/io.h>
> > > > #include <linux/kernel.h>
> > > > +#include <linux/minmax.h>
> > > > #include <linux/module.h>
> > > > #include <linux/mm.h>
> > > > #include <linux/mutex.h>
> > > > @@ -55,7 +56,9 @@
> > > > (FIELD_PREP(SENDER_ID_MASK, (s)) | FIELD_PREP(RECEIVER_ID_MASK, (r)))
> > > >
> > > > #define RXTX_MAP_MIN_BUFSZ_MASK GENMASK(1, 0)
> > > > -#define RXTX_MAP_MIN_BUFSZ(x) ((x) & RXTX_MAP_MIN_BUFSZ_MASK)
> > > > +#define RXTX_MAP_MAX_BUFSZ_MASK GENMASK(31, 16)
> > > > +#define RXTX_MAP_MIN_BUFSZ(x) (FIELD_GET(RXTX_MAP_MIN_BUFSZ_MASK, (x)))
> > > > +#define RXTX_MAP_MAX_BUFSZ(x) (FIELD_GET(RXTX_MAP_MAX_BUFSZ_MASK, (x)))
> > > >
> > > > #define FFA_MAX_NOTIFICATIONS 64
> > > >
> > > > @@ -2086,11 +2089,13 @@ static void ffa_notifications_setup(void)
> > > > ffa_notifications_cleanup();
> > > > }
> > > >
> > > > +#define FFA_SUPPORTS_RXTX_MAX_BUFSZ(version) ((version) > FFA_VERSION_1_1)
> > > > +
> > > > static int __init ffa_init(void)
> > > > {
> > > > int ret;
> > > > u32 buf_sz;
> > > > - size_t rxtx_bufsz = SZ_4K;
> > > > + size_t rxtx_bufsz = SZ_4K, rxtx_max_bufsz = 0;
> > > >
> > > > ret = ffa_transport_init(&invoke_ffa_fn);
> > > > if (ret)
> > > > @@ -2118,9 +2123,29 @@ static int __init ffa_init(void)
> > > > rxtx_bufsz = SZ_16K;
> > > > else
> > > > rxtx_bufsz = SZ_4K;
> > > > +
> > > > + if (FFA_SUPPORTS_RXTX_MAX_BUFSZ(drv_info->version)) {
> > >
> > > This is not even compile tested ? I don't think this is defined anyway.
> > > Anyways I don't think we need this condition as v1.1 or below had it
> > > MBZ, so it is fine to drop it.
> >
> > It is added just above ffa_init(). The code as been compile and run
> > tested on multiple platforms, as noted in the cover letter.
> >
>
> Ah sorry, my bad. It failed to apply and I must have messed it up.
> I have some changes in linux-next which conflicts and I dropped it when
> resolving. Sorry for that.
>
> > I hadn't checked older spec versions, so if they have it as MBZ then I
> > agree we can drop the check.
> >
> > >
> > > > + rxtx_max_bufsz = (size_t)RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
> > >
> > > The cast is unnecessary, its is 16 bit, so max 0xFFFF << 12.
> >
> > Correct, though linters complain about it which is why I added it. I can
> > drop it though if you prefer.
> >
>
> Can you be more specific ? Any way to trigger it ?
I've been trying out various tools for reviwing/analyzing patches, so I
don't remember now which one(s) complained about this. But I'll run the
patches back through them after I make the updates you've requested and
let you know.
Seth
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails
2026-06-02 16:29 ` Sudeep Holla
@ 2026-06-02 17:12 ` Seth Forshee
0 siblings, 0 replies; 14+ messages in thread
From: Seth Forshee @ 2026-06-02 17:12 UTC (permalink / raw)
To: Sudeep Holla; +Cc: Sebastian Ene, linux-arm-kernel, linux-kernel
On Tue, Jun 02, 2026 at 05:29:37PM +0100, Sudeep Holla wrote:
> On Tue, Jun 02, 2026 at 11:21:33AM -0500, Seth Forshee wrote:
> > On Tue, Jun 02, 2026 at 04:48:11PM +0100, Sudeep Holla wrote:
> > > On Mon, Jun 01, 2026 at 03:45:12PM -0500, Seth Forshee wrote:
> > > > FFA_FEATURES only supports a maximum size value for RXTX_MAP in v1.2+.
> > > > Older implementations may still reject larger sizes, which can happen
> > > > when the minimum size is rounded up to the page size when PAGE_SIZE >
> > > > 4K.
> > > >
> > > > Fall back to the minimum size when RXTX_MAP fails with
> > > > INVALID_PARAMETERS to fix the regression on these platforms. This also
> > > > has the side effect of falling back for v1.2+ with a maximum of 0 (i.e.
> > > > no maximum), where the rounded-up value isn't expected to be rejected,
> > > > but there's no harm in retrying with a smaller size and it might even be
> > > > of benefit for a non-compliant SPMC.
> > > >
> > >
> > > Is this needed on any of your platforms ? I don't want to push a solution
> > > or fallback/workaround if there are no users for it.
> >
> > Yes, we have platforms for which commit 83210251fd70 causes RXTX_MAP
> > failures with page sizes larger than 4K.
> >
>
> And they are FF-A v1.1 or doesn't set max_buf size ? Sorry for being
> pedantic, just want to be sure it is useful on real platforms before
> pushing it as fix/workaround.
Yes, sorry, that's information I should have included. Some were v1.1,
don't set max buf size, and did fail with 64K pages. While I tested the
fix with 16K pages for thoroughness, I did not try to reproduce the
problem with that page size, so I can't say whether or not that would
succeed. But either way there's still a problem, and I figure that going
with the minimum size in that case is the surest bet.
Seth
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions
2026-06-02 16:51 ` [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions Sudeep Holla
@ 2026-06-02 17:19 ` Seth Forshee
0 siblings, 0 replies; 14+ messages in thread
From: Seth Forshee @ 2026-06-02 17:19 UTC (permalink / raw)
To: Sudeep Holla; +Cc: Sebastian Ene, linux-arm-kernel, linux-kernel
On Tue, Jun 02, 2026 at 05:51:10PM +0100, Sudeep Holla wrote:
> On Mon, Jun 01, 2026 at 03:45:10PM -0500, Seth Forshee wrote:
> > Commit 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size
> > during RXTX_MAP") caused a regression for some FF-A implementations. It
> > has the effect of passing the size of the page-aligned buffers to
> > FFA_RXTX_MAP, which may be larger than the maximum size supported by the
> > SPMC. When this happens FFA_RXTX_MAP will fail with INVALID_PARAMETERS.
> >
> > The following patches deal with two distinct scenarios which lead to
> > this issue. With FF-A v1.2+, FFA_FEATURES can specify a maximum supported
> > RX/TX buffer size, so patch 1 decodes this field and honors the maximum if
> > specified. For FF-A v1.1 and earlier the maximum is unknown, so patch 2
> > deals with this by first attempting FFA_RXTX_MAP with the page-aligned
> > buffer size (preserving the behavior introduced by commit 83210251fd70).
> > If this fails due to invalid parameters it retries with the minimum buffer
> > size from FFA_FEATURES.
> >
> > Testing was done with FF-A v1.1 and v1.2 implementations, both of which
> > reject buffer sizes larger than 4K. Both implementations were tested with
> > 4K, 16K, and 64K pages. Without these patches, probe fails for page sizes
> > larger than 4K with the message "failed to register FFA RxTx buffers."
> > With the patches probe succeeds for all page sizes.
> >
> > The patches are based on for-next/ffa/fixes.
>
> for-next/ffa/updates has patches queued for v7.2.
>
> >
> > Signed-off-by: Seth Forshee <sforshee@nvidia•com>
> > ---
> > Seth Forshee (2):
> > firmware: arm_ffa: Honor maximum RX/TX buffer size
> > firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails
> >
> > drivers/firmware/arm_ffa/driver.c | 51 ++++++++++++++++++++++++++++++++++-----
> > 1 file changed, 45 insertions(+), 6 deletions(-)
>
> Wondering if these 2 fixes can be merged into one and simplified something
> like(untested) patch below(rebased on linux-next or for-next/ffa/updates)
From a quick review, yes, I think that looks like it should work. I'll
give it a closer look and some testing.
Seth
>
> Regards,
> Sudeep
>
> -->8
> diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
> index 0f468362c288..5ffe21c568b7 100644
> --- a/drivers/firmware/arm_ffa/driver.c
> +++ b/drivers/firmware/arm_ffa/driver.c
> @@ -32,6 +32,7 @@
> #include <linux/interrupt.h>
> #include <linux/io.h>
> #include <linux/kernel.h>
> +#include <linux/minmax.h>
> #include <linux/module.h>
> #include <linux/mm.h>
> #include <linux/mutex.h>
> @@ -59,7 +60,9 @@
> (FIELD_PREP(SENDER_ID_MASK, (s)) | FIELD_PREP(RECEIVER_ID_MASK, (r)))
>
> #define RXTX_MAP_MIN_BUFSZ_MASK GENMASK(1, 0)
> -#define RXTX_MAP_MIN_BUFSZ(x) ((x) & RXTX_MAP_MIN_BUFSZ_MASK)
> +#define RXTX_MAP_MAX_BUFSZ_MASK GENMASK(31, 16)
> +#define RXTX_MAP_MIN_BUFSZ(x) (FIELD_GET(RXTX_MAP_MIN_BUFSZ_MASK, (x)))
> +#define RXTX_MAP_MAX_BUFSZ(x) (FIELD_GET(RXTX_MAP_MAX_BUFSZ_MASK, (x)))
>
> #define FFA_MAX_NOTIFICATIONS 64
>
> @@ -2101,7 +2104,7 @@ static int ffa_probe(struct platform_device *pdev)
> {
> int ret;
> u32 buf_sz;
> - size_t rxtx_bufsz = SZ_4K;
> + size_t rxtx_min_bufsz = SZ_4K, rxtx_bufsz, rxtx_max_bufsz = 0;
>
> if (IS_BUILTIN(CONFIG_ARM_FFA_TRANSPORT) &&
> is_protected_kvm_enabled() && !is_pkvm_initialized())
> @@ -2132,15 +2135,18 @@ static int ffa_probe(struct platform_device *pdev)
> ret = ffa_features(FFA_FN_NATIVE(RXTX_MAP), 0, &buf_sz, NULL);
> if (!ret) {
> if (RXTX_MAP_MIN_BUFSZ(buf_sz) == 1)
> - rxtx_bufsz = SZ_64K;
> + rxtx_min_bufsz = SZ_64K;
> else if (RXTX_MAP_MIN_BUFSZ(buf_sz) == 2)
> - rxtx_bufsz = SZ_16K;
> + rxtx_min_bufsz = SZ_16K;
> else
> - rxtx_bufsz = SZ_4K;
> + rxtx_min_bufsz = SZ_4K;
> +
> + rxtx_max_bufsz = RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
> + if (rxtx_max_bufsz != 0 && rxtx_max_bufsz < rxtx_min_bufsz)
> + rxtx_max_bufsz = rxtx_min_bufsz;
> }
>
> - rxtx_bufsz = PAGE_ALIGN(rxtx_bufsz);
> - drv_info->rxtx_bufsz = rxtx_bufsz;
> + rxtx_bufsz = min_not_zero(PAGE_ALIGN(rxtx_min_bufsz), rxtx_max_bufsz);
> drv_info->rx_buffer = alloc_pages_exact(rxtx_bufsz, GFP_KERNEL);
> if (!drv_info->rx_buffer) {
> ret = -ENOMEM;
> @@ -2156,10 +2162,17 @@ static int ffa_probe(struct platform_device *pdev)
> ret = ffa_rxtx_map(virt_to_phys(drv_info->tx_buffer),
> virt_to_phys(drv_info->rx_buffer),
> rxtx_bufsz / FFA_PAGE_SIZE);
> + if (ret == -EINVAL && !rxtx_max_bufsz && rxtx_min_bufsz < rxtx_bufsz) {
> + rxtx_bufsz = rxtx_min_bufsz;
> + ret = ffa_rxtx_map(virt_to_phys(drv_info->tx_buffer),
> + virt_to_phys(drv_info->rx_buffer),
> + rxtx_bufsz / FFA_PAGE_SIZE);
> + }
> if (ret) {
> pr_err("failed to register FFA RxTx buffers\n");
> goto free_pages;
> }
> + drv_info->rxtx_bufsz = rxtx_bufsz;
>
> mutex_init(&drv_info->rx_lock);
> mutex_init(&drv_info->tx_lock);
>
>
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size
2026-06-02 17:05 ` Seth Forshee
@ 2026-06-02 21:58 ` Seth Forshee
0 siblings, 0 replies; 14+ messages in thread
From: Seth Forshee @ 2026-06-02 21:58 UTC (permalink / raw)
To: Sudeep Holla; +Cc: Sebastian Ene, linux-arm-kernel, linux-kernel
On Tue, Jun 02, 2026 at 12:05:47PM -0500, Seth Forshee wrote:
> On Tue, Jun 02, 2026 at 05:27:48PM +0100, Sudeep Holla wrote:
> > On Tue, Jun 02, 2026 at 11:12:03AM -0500, Seth Forshee wrote:
> > > On Tue, Jun 02, 2026 at 04:56:48PM +0100, Sudeep Holla wrote:
> > > > On Mon, Jun 01, 2026 at 03:45:11PM -0500, Seth Forshee wrote:
> > > > > FFA_FEATURES in v1.2+ supports a maximum RXTX_MAP size. This maximum
> > > > > size is not checked when page-aligning the RX/TX buffer size, and
> > > > > FFA_RXTX_MAP may fail when passed a size greater than the maximum.
> > > > >
> > > > > Decode the maximum buffer size returned from FFA_FEATURES and limit the
> > > > > buffer size based on this value if it is non-zero (zero indicates no
> > > > > maximum). Include verification that the max returned by the SPMC is
> > > > > larger than the minimum, otherwise use the minimum.
> > > > >
> > > > > While there, also update RXTX_MAP_MIN_BUFSZ() to use FIELD_GET() for
> > > > > consistency.
> > > > >
> > > > > Fixes: 83210251fd70 ("firmware: arm_ffa: Use the correct buffer size during RXTX_MAP")
> > > > > Assisted-by: Claude:claude-opus-4-8
> > > > > Signed-off-by: Seth Forshee <sforshee@nvidia•com>
> > > > > ---
> > > > > drivers/firmware/arm_ffa/driver.c | 29 +++++++++++++++++++++++++++--
> > > > > 1 file changed, 27 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/drivers/firmware/arm_ffa/driver.c b/drivers/firmware/arm_ffa/driver.c
> > > > > index b9f17fda7243..dc45724a29ba 100644
> > > > > --- a/drivers/firmware/arm_ffa/driver.c
> > > > > +++ b/drivers/firmware/arm_ffa/driver.c
> > > > > @@ -32,6 +32,7 @@
> > > > > #include <linux/interrupt.h>
> > > > > #include <linux/io.h>
> > > > > #include <linux/kernel.h>
> > > > > +#include <linux/minmax.h>
> > > > > #include <linux/module.h>
> > > > > #include <linux/mm.h>
> > > > > #include <linux/mutex.h>
> > > > > @@ -55,7 +56,9 @@
> > > > > (FIELD_PREP(SENDER_ID_MASK, (s)) | FIELD_PREP(RECEIVER_ID_MASK, (r)))
> > > > >
> > > > > #define RXTX_MAP_MIN_BUFSZ_MASK GENMASK(1, 0)
> > > > > -#define RXTX_MAP_MIN_BUFSZ(x) ((x) & RXTX_MAP_MIN_BUFSZ_MASK)
> > > > > +#define RXTX_MAP_MAX_BUFSZ_MASK GENMASK(31, 16)
> > > > > +#define RXTX_MAP_MIN_BUFSZ(x) (FIELD_GET(RXTX_MAP_MIN_BUFSZ_MASK, (x)))
> > > > > +#define RXTX_MAP_MAX_BUFSZ(x) (FIELD_GET(RXTX_MAP_MAX_BUFSZ_MASK, (x)))
> > > > >
> > > > > #define FFA_MAX_NOTIFICATIONS 64
> > > > >
> > > > > @@ -2086,11 +2089,13 @@ static void ffa_notifications_setup(void)
> > > > > ffa_notifications_cleanup();
> > > > > }
> > > > >
> > > > > +#define FFA_SUPPORTS_RXTX_MAX_BUFSZ(version) ((version) > FFA_VERSION_1_1)
> > > > > +
> > > > > static int __init ffa_init(void)
> > > > > {
> > > > > int ret;
> > > > > u32 buf_sz;
> > > > > - size_t rxtx_bufsz = SZ_4K;
> > > > > + size_t rxtx_bufsz = SZ_4K, rxtx_max_bufsz = 0;
> > > > >
> > > > > ret = ffa_transport_init(&invoke_ffa_fn);
> > > > > if (ret)
> > > > > @@ -2118,9 +2123,29 @@ static int __init ffa_init(void)
> > > > > rxtx_bufsz = SZ_16K;
> > > > > else
> > > > > rxtx_bufsz = SZ_4K;
> > > > > +
> > > > > + if (FFA_SUPPORTS_RXTX_MAX_BUFSZ(drv_info->version)) {
> > > >
> > > > This is not even compile tested ? I don't think this is defined anyway.
> > > > Anyways I don't think we need this condition as v1.1 or below had it
> > > > MBZ, so it is fine to drop it.
> > >
> > > It is added just above ffa_init(). The code as been compile and run
> > > tested on multiple platforms, as noted in the cover letter.
> > >
> >
> > Ah sorry, my bad. It failed to apply and I must have messed it up.
> > I have some changes in linux-next which conflicts and I dropped it when
> > resolving. Sorry for that.
> >
> > > I hadn't checked older spec versions, so if they have it as MBZ then I
> > > agree we can drop the check.
> > >
> > > >
> > > > > + rxtx_max_bufsz = (size_t)RXTX_MAP_MAX_BUFSZ(buf_sz) * SZ_4K;
> > > >
> > > > The cast is unnecessary, its is 16 bit, so max 0xFFFF << 12.
> > >
> > > Correct, though linters complain about it which is why I added it. I can
> > > drop it though if you prefer.
> > >
> >
> > Can you be more specific ? Any way to trigger it ?
>
> I've been trying out various tools for reviwing/analyzing patches, so I
> don't remember now which one(s) complained about this. But I'll run the
> patches back through them after I make the updates you've requested and
> let you know.
Nothing flagged it this time, but I'm sure I only added it originally
after a tool flagged it. However it got there, it's removed in v2.
Seth
^ permalink raw reply [flat|nested] 14+ messages in thread
end of thread, other threads:[~2026-06-02 21:58 UTC | newest]
Thread overview: 14+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-06-01 20:45 [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions Seth Forshee
2026-06-01 20:45 ` [PATCH 1/2] firmware: arm_ffa: Honor maximum RX/TX buffer size Seth Forshee
2026-06-02 15:56 ` Sudeep Holla
2026-06-02 16:12 ` Seth Forshee
2026-06-02 16:27 ` Sudeep Holla
2026-06-02 17:05 ` Seth Forshee
2026-06-02 21:58 ` Seth Forshee
2026-06-01 20:45 ` [PATCH 2/2] firmware: arm_ffa: Fall back to minimum buffer size if RXTX_MAP fails Seth Forshee
2026-06-02 15:48 ` Sudeep Holla
2026-06-02 16:21 ` Seth Forshee
2026-06-02 16:29 ` Sudeep Holla
2026-06-02 17:12 ` Seth Forshee
2026-06-02 16:51 ` [PATCH 0/2] firmware: arm_ffa: Fix RXTX_MAP buffer size regressions Sudeep Holla
2026-06-02 17:19 ` Seth Forshee
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox