[PATCH] eal: fix DMA mask validation inconsistency in IOVA VA

[PATCH] eal: fix DMA mask validation inconsistency in IOVA VA

[Shani Peretz]

When --iova-mode is explicitly specified in command line, DMA mask
constraints were not being validated, leading to potential runtime
failures when device DMA capabilities are exceeded.

The issue occurred because rte_bus_get_iommu_class() was only called
during IOVA mode auto-detection, but this function has the important
side effect of triggering DMA mask detection (e.g., Intel IOMMU
address width checking via pci_device_iommu_support_va()).

This created an inconsistency, when choosing explicit mode,
the DMA checks are bypassed, but when choosing auto-detection mode,
the constraints are checked and enforced.

The fix moves rte_bus_get_iommu_class() outside the conditional logic
to ensure it's always called during EAL initialization.

Fixes: 4374ebc24bc1 ("malloc: modify error message for DMA mask check")
Cc: stable@dpdk.org

Signed-off-by: Shani Peretz <shperetz@nvidia.com>
---
 lib/eal/freebsd/eal.c | 6 +++++-
 lib/eal/linux/eal.c   | 5 ++++-
 lib/eal/windows/eal.c | 5 ++++-
 3 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/lib/eal/freebsd/eal.c b/lib/eal/freebsd/eal.c
index c1ab8d86d2..0f957919d3 100644
--- a/lib/eal/freebsd/eal.c
+++ b/lib/eal/freebsd/eal.c
@@ -670,12 +670,16 @@ rte_eal_init(int argc, char **argv)
 	 * with a message describing the cause.
 	 */
 	has_phys_addr = internal_conf->no_hugetlbfs == 0;
+
+	/* Always call rte_bus_get_iommu_class() to trigger DMA mask detection and validation */
+	enum rte_iova_mode bus_iova_mode = rte_bus_get_iommu_class();
+
 	iova_mode = internal_conf->iova_mode;
 	if (iova_mode == RTE_IOVA_DC) {
 		EAL_LOG(DEBUG, "Specific IOVA mode is not requested, autodetecting");
 		if (has_phys_addr) {
 			EAL_LOG(DEBUG, "Selecting IOVA mode according to bus requests");
-			iova_mode = rte_bus_get_iommu_class();
+			iova_mode = bus_iova_mode;
 			if (iova_mode == RTE_IOVA_DC) {
 				if (!RTE_IOVA_IN_MBUF) {
 					iova_mode = RTE_IOVA_VA;
diff --git a/lib/eal/linux/eal.c b/lib/eal/linux/eal.c
index 52efb8626b..3a0c9c9db6 100644
--- a/lib/eal/linux/eal.c
+++ b/lib/eal/linux/eal.c
@@ -1042,10 +1042,13 @@ rte_eal_init(int argc, char **argv)
 
 	phys_addrs = rte_eal_using_phys_addrs() != 0;
 
+	/* Always call rte_bus_get_iommu_class() to trigger DMA mask detection and validation */
+	enum rte_iova_mode bus_iova_mode = rte_bus_get_iommu_class();
+
 	/* if no EAL option "--iova-mode=<pa|va>", use bus IOVA scheme */
 	if (internal_conf->iova_mode == RTE_IOVA_DC) {
 		/* autodetect the IOVA mapping mode */
-		enum rte_iova_mode iova_mode = rte_bus_get_iommu_class();
+		enum rte_iova_mode iova_mode = bus_iova_mode;
 
 		if (iova_mode == RTE_IOVA_DC) {
 			EAL_LOG(DEBUG, "Buses did not request a specific IOVA mode.");
diff --git a/lib/eal/windows/eal.c b/lib/eal/windows/eal.c
index 4f0a164d9b..2502ec3c3d 100644
--- a/lib/eal/windows/eal.c
+++ b/lib/eal/windows/eal.c
@@ -348,12 +348,15 @@ rte_eal_init(int argc, char **argv)
 		has_phys_addr = false;
 	}
 
+	/* Always call rte_bus_get_iommu_class() to trigger DMA mask detection and validation */
+	enum rte_iova_mode bus_iova_mode = rte_bus_get_iommu_class();
+
 	iova_mode = internal_conf->iova_mode;
 	if (iova_mode == RTE_IOVA_DC) {
 		EAL_LOG(DEBUG, "Specific IOVA mode is not requested, autodetecting");
 		if (has_phys_addr) {
 			EAL_LOG(DEBUG, "Selecting IOVA mode according to bus requests");
-			iova_mode = rte_bus_get_iommu_class();
+			iova_mode = bus_iova_mode;
 			if (iova_mode == RTE_IOVA_DC) {
 				if (!RTE_IOVA_IN_MBUF) {
 					iova_mode = RTE_IOVA_VA;
-- 
2.34.1

Re: [PATCH] eal: fix DMA mask validation inconsistency in IOVA VA

[Burakov, Anatoly]

On 9/8/2025 6:34 PM, Shani Peretz wrote:
> When --iova-mode is explicitly specified in command line, DMA mask
> constraints were not being validated, leading to potential runtime
> failures when device DMA capabilities are exceeded.
> 
> The issue occurred because rte_bus_get_iommu_class() was only called
> during IOVA mode auto-detection, but this function has the important
> side effect of triggering DMA mask detection (e.g., Intel IOMMU
> address width checking via pci_device_iommu_support_va()).
> 
> This created an inconsistency, when choosing explicit mode,
> the DMA checks are bypassed, but when choosing auto-detection mode,
> the constraints are checked and enforced.
> 
> The fix moves rte_bus_get_iommu_class() outside the conditional logic
> to ensure it's always called during EAL initialization.
> 
> Fixes: 4374ebc24bc1 ("malloc: modify error message for DMA mask check")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Shani Peretz <shperetz@nvidia.com>
> ---

Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>

-- 
Thanks,
Anatoly

Re: [PATCH] eal: fix DMA mask validation inconsistency in IOVA VA

[Patrick Robb]

[-- Attachment #1: Type: text/plain, Size: 4654 bytes --]

FYI The DTS failure for this series on patchwork for the MTU testsuite is
coming from a different patch than this one. Once I have that resolved, I
will put in a retest for this series which will clear up that MTU testsuite
failure.

On Mon, Sep 8, 2025 at 12:35 PM Shani Peretz <shperetz@nvidia.com> wrote:

> When --iova-mode is explicitly specified in command line, DMA mask
> constraints were not being validated, leading to potential runtime
> failures when device DMA capabilities are exceeded.
>
> The issue occurred because rte_bus_get_iommu_class() was only called
> during IOVA mode auto-detection, but this function has the important
> side effect of triggering DMA mask detection (e.g., Intel IOMMU
> address width checking via pci_device_iommu_support_va()).
>
> This created an inconsistency, when choosing explicit mode,
> the DMA checks are bypassed, but when choosing auto-detection mode,
> the constraints are checked and enforced.
>
> The fix moves rte_bus_get_iommu_class() outside the conditional logic
> to ensure it's always called during EAL initialization.
>
> Fixes: 4374ebc24bc1 ("malloc: modify error message for DMA mask check")
> Cc: stable@dpdk.org
>
> Signed-off-by: Shani Peretz <shperetz@nvidia.com>
> ---
>  lib/eal/freebsd/eal.c | 6 +++++-
>  lib/eal/linux/eal.c   | 5 ++++-
>  lib/eal/windows/eal.c | 5 ++++-
>  3 files changed, 13 insertions(+), 3 deletions(-)
>
> diff --git a/lib/eal/freebsd/eal.c b/lib/eal/freebsd/eal.c
> index c1ab8d86d2..0f957919d3 100644
> --- a/lib/eal/freebsd/eal.c
> +++ b/lib/eal/freebsd/eal.c
> @@ -670,12 +670,16 @@ rte_eal_init(int argc, char **argv)
>          * with a message describing the cause.
>          */
>         has_phys_addr = internal_conf->no_hugetlbfs == 0;
> +
> +       /* Always call rte_bus_get_iommu_class() to trigger DMA mask
> detection and validation */
> +       enum rte_iova_mode bus_iova_mode = rte_bus_get_iommu_class();
> +
>         iova_mode = internal_conf->iova_mode;
>         if (iova_mode == RTE_IOVA_DC) {
>                 EAL_LOG(DEBUG, "Specific IOVA mode is not requested,
> autodetecting");
>                 if (has_phys_addr) {
>                         EAL_LOG(DEBUG, "Selecting IOVA mode according to
> bus requests");
> -                       iova_mode = rte_bus_get_iommu_class();
> +                       iova_mode = bus_iova_mode;
>                         if (iova_mode == RTE_IOVA_DC) {
>                                 if (!RTE_IOVA_IN_MBUF) {
>                                         iova_mode = RTE_IOVA_VA;
> diff --git a/lib/eal/linux/eal.c b/lib/eal/linux/eal.c
> index 52efb8626b..3a0c9c9db6 100644
> --- a/lib/eal/linux/eal.c
> +++ b/lib/eal/linux/eal.c
> @@ -1042,10 +1042,13 @@ rte_eal_init(int argc, char **argv)
>
>         phys_addrs = rte_eal_using_phys_addrs() != 0;
>
> +       /* Always call rte_bus_get_iommu_class() to trigger DMA mask
> detection and validation */
> +       enum rte_iova_mode bus_iova_mode = rte_bus_get_iommu_class();
> +
>         /* if no EAL option "--iova-mode=<pa|va>", use bus IOVA scheme */
>         if (internal_conf->iova_mode == RTE_IOVA_DC) {
>                 /* autodetect the IOVA mapping mode */
> -               enum rte_iova_mode iova_mode = rte_bus_get_iommu_class();
> +               enum rte_iova_mode iova_mode = bus_iova_mode;
>
>                 if (iova_mode == RTE_IOVA_DC) {
>                         EAL_LOG(DEBUG, "Buses did not request a specific
> IOVA mode.");
> diff --git a/lib/eal/windows/eal.c b/lib/eal/windows/eal.c
> index 4f0a164d9b..2502ec3c3d 100644
> --- a/lib/eal/windows/eal.c
> +++ b/lib/eal/windows/eal.c
> @@ -348,12 +348,15 @@ rte_eal_init(int argc, char **argv)
>                 has_phys_addr = false;
>         }
>
> +       /* Always call rte_bus_get_iommu_class() to trigger DMA mask
> detection and validation */
> +       enum rte_iova_mode bus_iova_mode = rte_bus_get_iommu_class();
> +
>         iova_mode = internal_conf->iova_mode;
>         if (iova_mode == RTE_IOVA_DC) {
>                 EAL_LOG(DEBUG, "Specific IOVA mode is not requested,
> autodetecting");
>                 if (has_phys_addr) {
>                         EAL_LOG(DEBUG, "Selecting IOVA mode according to
> bus requests");
> -                       iova_mode = rte_bus_get_iommu_class();
> +                       iova_mode = bus_iova_mode;
>                         if (iova_mode == RTE_IOVA_DC) {
>                                 if (!RTE_IOVA_IN_MBUF) {
>                                         iova_mode = RTE_IOVA_VA;
> --
> 2.34.1
>
>

[-- Attachment #2: Type: text/html, Size: 5651 bytes --]

Re: [PATCH] eal: fix DMA mask validation inconsistency in IOVA VA

[Thomas Monjalon]

Recheck-request: rebase=main, iol-mellanox-Functional, iol-compile-amd64-testing, iol-compile-arm64-testing, iol-unit-amd64-testing, iol-unit-arm64-testing