Thread (16 messages) flat view 16 messages, 2 authors, 2026-08-05

Re: [PATCH v2 1/4] powerpc/pseries: Limit PVR list to 16 entries for CAS negotiation

From: Mahesh J Salgaonkar <mahesh@linux.ibm.com>
Date: 2026-08-04 06:43:02

On 2026-08-04 00:02:08 Tue, Amit Machhiwal wrote:
On 2026/08/03 09:27 PM, Mahesh Salgaonkar wrote:
quoted
Current Power system firmware caps the PVR list array size at 16 entries
during CAS (Client Architecture Support) negotiation. Passing more than
capped size to older firmware could cause an undefined behaviour and
breaks compatibility.

Future Power system firmware releases will lift this restriction and
support greater than 16 array entries.

Ensure that when running on Power11 or below hardware, the number
of PVR entries passed during CAS negotiation does not exceed the
firmware-imposed limit of 16.

In prom_send_capabilities(), compute start_index to skip the oldest
leading pvrs[] entries when running on Power11 or below hardware, so
that the pointer passed to ibm,client-architecture-support points to
ibm_architecture_vec.pvrs[start_index], presenting exactly 16 entries to
firmware.

Signed-off-by: Mahesh Salgaonkar <mahesh@linux.ibm.com>
Tested-by: Praveen K Pandey <redacted>
---
Change in v2:
- Define a named macro CAS_MAX_PVR_ENTRIES
---
[...]
quoted
@@ -1032,13 +1042,9 @@ static const struct ibm_arch_vec ibm_architecture_vec_template __initconst = {
 			.val  = cpu_to_be32(0x0f000003),
 		},
 		{
-			.mask = cpu_to_be32(0xffffffff), /* all 2.05-compliant */
+			.mask = cpu_to_be32(0xfffffffd), /* all 2.05-compliant */
 			.val  = cpu_to_be32(0x0f000002),
 		},
-		{
-			.mask = cpu_to_be32(0xfffffffe), /* all 2.04-compliant and earlier */
-			.val  = cpu_to_be32(0x0f000001),
-		},
 	},
 
 	.num_vectors = NUM_VECTORS(6),
@@ -1347,7 +1353,6 @@ static void __init prom_check_platform_support(void)
 	};
 	int prop_len = prom_getproplen(prom.chosen,
 				       "ibm,arch-vec-5-platform-support");
-
 
Nit: Looks unintentional whitespace change, please drop.
Hi Maddy,

Let me know if you want ime to send out v3 to handle this unintentional
whitespace change Or do you think you handle this while merging ?

Thanks,
-Mahesh.
quoted
 	/*
 	 * First copy the architecture vec template
 	 *
@@ -1403,6 +1408,22 @@ static void __init prom_send_capabilities(void)
 	ihandle root;
 	prom_arg_t ret;
 	u32 cores;
+	int start_index = 0;
+
+	/*
+	 * Ensure that when running on Power11 or below hardware, the number
+	 * of PVR entries passed during CAS negotiation does not exceed the
+	 * firmware-imposed limit of 16.
+	 *
+	 * Compute the start_index to skip the oldest leading pvrs[] entries
+	 * when running on Power11 or below hardware, so that the pointer
+	 * passed to ibm,client-architecture-support points to
+	 * ibm_architecture_vec.pvrs[start_index], presenting exactly 16
+	 * entries to firmware.
+	 */
+	if ((ARRAY_SIZE(ibm_architecture_vec_template.pvrs) > CAS_MAX_PVR_ENTRIES) &&
+	    (PVR_VER(mfspr(SPRN_PVR)) <= PVR_POWER11))
+		start_index = ARRAY_SIZE(ibm_architecture_vec_template.pvrs) - CAS_MAX_PVR_ENTRIES;
 
 	/* Check ibm,arch-vec-5-platform-support and fixup vec5 if required */
 	prom_check_platform_support();
@@ -1427,7 +1448,7 @@ static void __init prom_send_capabilities(void)
 		if (call_prom_ret("call-method", 3, 2, &ret,
 				  ADDR("ibm,client-architecture-support"),
 				  root,
-				  ADDR(&ibm_architecture_vec)) == 0) {
+				  ADDR(&ibm_architecture_vec.pvrs[start_index])) == 0) {
 			/* the call exists... */
 			if (ret)
 				prom_printf("\nWARNING: ibm,client-architecture"
-- 
2.55.0
-- 
Mahesh J Salgaonkar
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help