[PATCH] bus/dpaa: defer bus initialization to probe

[PATCH] bus/dpaa: defer bus initialization to probe

[Prashant Gupta]

In the DPAA bus initialization: it builds the
device list, loads the QMAN/BMAN drivers, registers the platform mbuf
pool ops and populates the PA to VA translation table. All of it needs
the DPDK heap or a memzone for multi process scenarios.

get_iommu_class() for this bus only checks two sysfs paths, so the
initialization is not needed in scan to report the IOVA mode. Move it
back to probe.

Fixes: cdefd2e980bd ("drivers/bus: initialize NXP bus specifics in scan")

Signed-off-by: Prashant Gupta <prashant.gupta_3@nxp.com>
---
 drivers/bus/dpaa/dpaa_bus.c | 31 +++++++++++++++++++++++++------
 1 file changed, 25 insertions(+), 6 deletions(-)

diff --git a/drivers/bus/dpaa/dpaa_bus.c b/drivers/bus/dpaa/dpaa_bus.c
index 62dcc79534..85f6957466 100644
--- a/drivers/bus/dpaa/dpaa_bus.c
+++ b/drivers/bus/dpaa/dpaa_bus.c
@@ -740,10 +740,8 @@ static int rte_dpaa_setup_intr(struct rte_intr_handle *intr_handle)
 static int
 rte_dpaa_bus_scan(void)
 {
-	struct rte_dpaa_device *dev;
 	FILE *svr_file = NULL;
 	uint32_t svr_ver;
-	static int process_once;
 	char *penv;
 	int ret;
 
@@ -806,9 +804,29 @@ rte_dpaa_bus_scan(void)
 			dpaa_bus.max_push_rxq_num = DPAA_MAX_PUSH_MODE_QUEUE;
 	}
 
+	return 0;
+}
+
+/* Bus initialization needs the DPDK heap and a memzone, which EAL only sets
+ * up after the bus scan, so it is done here.
+ */
+static int
+rte_dpaa_bus_probe(struct rte_bus *bus)
+{
+	static int process_once;
+	struct rte_dpaa_device *dev;
+	int ret;
+
+	if (!dpaa_bus.detected)
+		return 0;
+
 	/* Device list creation is only done once */
 	if (!process_once) {
-		rte_dpaa_bus_dev_build();
+		ret = rte_dpaa_bus_dev_build();
+		if (ret) {
+			DPAA_BUS_ERR("Unable to build device list. (%d)", ret);
+			return ret;
+		}
 		/* One time load of Qman/Bman drivers */
 		ret = qman_global_init();
 		if (ret) {
@@ -822,8 +840,8 @@ rte_dpaa_bus_scan(void)
 				     ret);
 			return ret;
 		}
+		process_once = 1;
 	}
-	process_once = 1;
 
 	/* If no device present on DPAA bus nothing needs to be done */
 	if (TAILQ_EMPTY(&rte_dpaa_bus.device_list))
@@ -846,7 +864,8 @@ rte_dpaa_bus_scan(void)
 	dpaax_iova_table_populate();
 
 	dpaa_bus_global_init = 1;
-	return 0;
+
+	return rte_bus_generic_probe(bus);
 }
 
 /*
@@ -928,7 +947,7 @@ RTE_FINI_PRIO(dpaa_cleanup, 102)
 
 static struct rte_bus rte_dpaa_bus = {
 	.scan = rte_dpaa_bus_scan,
-	.probe = rte_bus_generic_probe,
+	.probe = rte_dpaa_bus_probe,
 	.parse = rte_dpaa_bus_parse,
 	.dev_compare = dpaa_bus_dev_compare,
 	.find_device = rte_bus_generic_find_device,
-- 
2.43.0

Re: [PATCH] bus/dpaa: defer bus initialization to probe

[David Marchand]

Hello,

On Thu, 27 Aug 2026 at 09:37, Prashant Gupta <prashant.gupta_3@nxp.com> wrote:
>
> In the DPAA bus initialization: it builds the
> device list, loads the QMAN/BMAN drivers, registers the platform mbuf
> pool ops and populates the PA to VA translation table. All of it needs
> the DPDK heap or a memzone for multi process scenarios.
>
> get_iommu_class() for this bus only checks two sysfs paths, so the
> initialization is not needed in scan to report the IOVA mode. Move it
> back to probe.
>
> Fixes: cdefd2e980bd ("drivers/bus: initialize NXP bus specifics in scan")
>
> Signed-off-by: Prashant Gupta <prashant.gupta_3@nxp.com>

While the change looks correct, I don't understand what is being fixed.
Could you elaborate on the issue?


-- 
David Marchand

Re: [PATCH] bus/dpaa: defer bus initialization to probe

[Hemant Agrawal]

On 27-08-2026 14:24, David Marchand wrote:
> Hello,
>
> On Thu, 27 Aug 2026 at 09:37, Prashant Gupta <prashant.gupta_3@nxp.com> wrote:
>> In the DPAA bus initialization: it builds the
>> device list, loads the QMAN/BMAN drivers, registers the platform mbuf
>> pool ops and populates the PA to VA translation table. All of it needs
>> the DPDK heap or a memzone for multi process scenarios.
>>
>> get_iommu_class() for this bus only checks two sysfs paths, so the
>> initialization is not needed in scan to report the IOVA mode. Move it
>> back to probe.
>>
>> Fixes: cdefd2e980bd ("drivers/bus: initialize NXP bus specifics in scan")
>>
>> Signed-off-by: Prashant Gupta <prashant.gupta_3@nxp.com>
> While the change looks correct, I don't understand what is being fixed.
> Could you elaborate on the issue?
>
In my brief discussion about this issue, I was told that some of the 
these QMAN/BMAN related initialization should be done using 
shared/Hugepage memory for multi-process sharing;

Hugepages are not available during scan, they become available later.

Acked-by: Hemant Agrawal <hemant.agrawal@nxp.com>

>