[PATCH] lpm: fix SVE bulk lookup loop bound for small bursts

[Rupesh Chiluka]

In __rte_lpm_lookup_vec() (SVE), the main loop condition was

    for (i = 0; i < n - vl; i += vl)

where n is uint32_t and vl is uint64_t. When n < vl the subtraction
is done in 64-bit unsigned arithmetic and wraps to a very large
value, so the loop does not terminate at the end of the input array.
It reads past the end of ips[] resulting in a segmentation fault.

Rewrite the condition as "i + vl < n" so no unsigned wrap occurs. The
existing tail handling with svwhilelt_b32() covers the remaining
elements, including n == 0.

Fixes: b29ccbea24 ("lpm: use SVE for bulk lookup")
Cc: stable at dpdk.org

Signed-off-by: Rupesh Chiluka <rchiluka at marvell.com>
---
 lib/lpm/rte_lpm_sve.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/lib/lpm/rte_lpm_sve.h b/lib/lpm/rte_lpm_sve.h
index 52b8bd4940..6113269120 100644
--- a/lib/lpm/rte_lpm_sve.h
+++ b/lib/lpm/rte_lpm_sve.h
@@ -26,7 +26,7 @@ __rte_lpm_lookup_vec(const struct rte_lpm *lpm, const uint32_t *ips,
 	for (i = 0; i < n; i++)
 		next_hops[i] = 0;
 
-	for (i = 0; i < n - vl; i += vl) {
+	for (i = 0; i + vl < n; i += vl) {
 		v_ip = svld1(pg, &ips[i]);
 		/* Get indices for tbl24[] */
 		v_idx = svlsr_x(pg, v_ip, 8);
-- 
2.43.0