Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Laurent Dufour <ldufour@linux.ibm.com>
To:     linuxppc-dev@lists.ozlabs.org, linux-kernel@vger.kernel.org
Cc:     mpe@ellerman.id.au, nathanl@linux.ibm.com, cheloha@linux.ibm.com,
        benh@kernel.crashing.org
Subject: [PATCH v2] powerpc/drmem: Don't compute the NUMA node for each LMB
Date:   Wed,  5 Aug 2020 15:35:02 +0200
Message-Id: <20200805133502.33723-1-ldufour@linux.ibm.com>
In-Reply-To: <202008051807.Vi8NDJtX%lkp@intel.com>
References: <202008051807.Vi8NDJtX%lkp@intel.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

All the LMB from the same set of ibm,dynamic-memory-v2 property are
sharing the same NUMA node. Don't compute that node for each one.

Tested on a system with 1022 LMBs spread on 4 NUMA nodes, only 4 calls to
lmb_set_nid() have been made instead of 1022.

This should prevent some soft lockups when starting large guests

Code has meaning only if CONFIG_MEMORY_HOTPLUG is set, otherwise the nid
field is not present in the drmem_lmb structure.

Signed-off-by: Laurent Dufour <ldufour@linux.ibm.com>
---
 arch/powerpc/mm/drmem.c | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

diff --git a/arch/powerpc/mm/drmem.c b/arch/powerpc/mm/drmem.c
index b2eeea39684c..c11b6ec99ea3 100644
--- a/arch/powerpc/mm/drmem.c
+++ b/arch/powerpc/mm/drmem.c
@@ -402,6 +402,9 @@ static void __init init_drmem_v2_lmbs(const __be32 *prop)
 	const __be32 *p;
 	u32 i, j, lmb_sets;
 	int lmb_index;
+#ifdef CONFIG_MEMORY_HOTPLUG
+	struct drmem_lmb *first = NULL;
+#endif
 
 	lmb_sets = of_read_number(prop++, 1);
 	if (lmb_sets == 0)
@@ -426,6 +429,15 @@ static void __init init_drmem_v2_lmbs(const __be32 *prop)
 	for (i = 0; i < lmb_sets; i++) {
 		read_drconf_v2_cell(&dr_cell, &p);
 
+#ifdef CONFIG_MEMORY_HOTPLUG
+		/*
+		 * Fetch the NUMA node id for the fist set or if the
+		 * associativity index is different from the previous set.
+		 */
+		if (first && dr_cell.aa_index != first->aa_index)
+			first = NULL;
+#endif
+
 		for (j = 0; j < dr_cell.seq_lmbs; j++) {
 			lmb = &drmem_info->lmbs[lmb_index++];
 
@@ -438,7 +450,18 @@ static void __init init_drmem_v2_lmbs(const __be32 *prop)
 			lmb->aa_index = dr_cell.aa_index;
 			lmb->flags = dr_cell.flags;
 
-			lmb_set_nid(lmb);
+#ifdef CONFIG_MEMORY_HOTPLUG
+			/*
+			 * All the LMB in the set share the same NUMA
+			 * associativity property. So read that node only once.
+			 */
+			if (!first) {
+				lmb_set_nid(lmb);
+				first = lmb;
+			} else {
+				lmb->nid = first->nid;
+			}
+#endif
 		}
 	}
 }
-- 
2.28.0