Re: [PATCH OLK-5.10] KVM: arm64: Translate logic cluster id to physical cluster id when updating lsudvmbm

On 2023/12/5 14:55, chenxiang wrote:

From: Xiang Chen chenxiang66@hisilicon.com

driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I8HY1B

---

For dvmbm feature, MN requires physical cluster id while it is filled with logic cluster id right now. In some situations which physical cluster id is not equal to logic cluster id such as in PG boards, it will cause issues when enabling dvmbm.

To avoid the issue, translate logic cluster id to physical cluster id when updating lsudvmbm.

Signed-off-by: Xiang Chen chenxiang66@hisilicon.com Signed-off-by: Cai Jian caijian11@h-partners.com

---

arch/arm64/kvm/hisilicon/hisi_virt.c | 166 +++++++++++++++++++++++++-- 1 file changed, 157 insertions(+), 9 deletions(-)

diff --git a/arch/arm64/kvm/hisilicon/hisi_virt.c b/arch/arm64/kvm/hisilicon/hisi_virt.c index 10233b801896..1f4a8e44228e 100644 --- a/arch/arm64/kvm/hisilicon/hisi_virt.c +++ b/arch/arm64/kvm/hisilicon/hisi_virt.c @@ -236,12 +236,113 @@ static int kvm_dvmbm_get_dies_info(struct kvm *kvm, u64 *vm_aff3s, int size) return num; }

+#define MAX_CLUSTERS_PER_DIE 8 +#define TB_MN_BASE 0x00C6067f0000 +#define TA_MN_BASE 0x0046067F0000 +#define CHIP_ADDR_OFFSET(_chip) (((((_chip) >> 3) & 0x1) * 0x80000000000) + \

•    ((((_chip) >> 2) & 0x1) * (0x100000000000)) + \
  

•    (((_chip) & 0x3) * 0x200000000000))
  

+#define MAX_PG_CFG_SKITS 4

"SKITS"是什么意思？

+#define MAX_PG_CFG_DIES 2 +#define MN_ECO0_OFFSET 0xc00 +#define SOCKET_ID_MASK 0xf8 +#define SOCKET_ID_SHIFT 3 +#define DIE_ID_MASK 0x7

跟芯片实现相关的宏，现在都放在hisi_virt.h里面。

+static u32 socket_num, die_num;

+static void kvm_get_socket_num(u32 *socket_num) +{

• int socket_id[MAX_PG_CFG_SKITS], cpu;
• u32 num = 0;
• for_each_cpu(cpu, cpu_possible_mask) {

• bool found = false;
  

• u64 aff3, socket;
  

• int i;
  

• aff3 = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 3);
  

• /* aff3[7:3]: socket ID */
  

• socket = (aff3 & SOCKET_ID_MASK) >> SOCKET_ID_SHIFT;
  

• for (i = 0; i < num; i++) {
  

• 	if (socket_id[i] == socket) {
  

• 		found = true;
  

• 		break;
  

• 	}
  

• }
  

• if (!found)
  

• 	socket_id[num++] = socket;
  

• }
• *socket_num = num;

把socket_nr作为返回值就行吧。不必传入参数

+}

+static void kvm_get_die_num(u32 *die_num) +{

• int die_id[MAX_PG_CFG_DIES], cpu;
• u32 num = 0;
• for_each_cpu(cpu, cpu_possible_mask) {

• bool found = false;
  

• u64 aff3, die;
  

• int i;
  

• aff3 = MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 3);
  

• /* aff3[2:0]: die ID */
  

• die = aff3 & DIE_ID_MASK;
  

• for (i = 0; i < num; i++) {
  

• 	if (die_id[i] == die) {
  

• 		found = true;
  

• 		break;
  

• 	}
  

• }
  

• if (!found)
  

• 	die_id[num++] = die;
  

• }
• *die_num = num;

+}

这个函数的作用是，得到一个socket内的dir_nr？是不是没必要遍历所有CPU？

函数的名字也可以考虑改一下。只针对hip09

+static u32 g_pg_cfg[MAX_PG_CFG_SKITS * MAX_PG_CFG_DIES];

+static u64 kvm_switch_aff2_logic_to_phy(unsigned long pg_cfg, u64 logic_aff2) +{

• u32 pg_num = 0, i, j;
• u32 cluster_pg[MAX_CLUSTERS_PER_DIE];
• u32 pg_flag[MAX_CLUSTERS_PER_DIE];
• for (i = 0; i < MAX_CLUSTERS_PER_DIE; i++) {

• if (test_bit(i, &pg_cfg))
  

• 	pg_num++;
  

• cluster_pg[i] = i;
  

• pg_flag[i] = 0;
  

• }
• for (i = 0; i < MAX_CLUSTERS_PER_DIE - pg_num; i++) {

• if (test_bit(i, &pg_cfg)) {
  

• 	for (j = 0; j < pg_num; j++) {
  

• 		u32 cluster_bak = MAX_CLUSTERS_PER_DIE - pg_num + j;
  

• 		if (!test_bit(cluster_bak, &pg_cfg) &&
  

• 		    !pg_flag[cluster_bak]) {
  

• 			pg_flag[cluster_bak] = 1;
  

• 			cluster_pg[i] = cluster_bak;
  

• 			cluster_pg[cluster_bak] = i;
  

• 			break;
  

• 		}
  

• 	}
  

• }
  

• }

cluster_pg[]表示MPIDR_EL1.aff2到“physical” aff2的映射？这个数组能不能 在kvm初始化的时候就生成？后面直接查表使用就行。

数组名字可以换个不？容易理解点的

• return cluster_pg[logic_aff2];

+}

static void kvm_update_vm_lsudvmbm(struct kvm *kvm) {

• u64 mpidr, aff3, aff2, aff1;

• u64 mpidr, aff3, aff2, aff1, phy_aff2; u64 vm_aff3s[DVMBM_MAX_DIES]; u64 val; int cpu, nr_dies;

• unsigned long pg_cfg;

• u32 socket_id;

  nr_dies = kvm_dvmbm_get_dies_info(kvm, vm_aff3s, DVMBM_MAX_DIES); if (nr_dies > 2) {

@@ -258,8 +359,14 @@ static void kvm_update_vm_lsudvmbm(struct kvm *kvm) mpidr = cpu_logical_map(cpu); aff2 = MPIDR_AFFINITY_LEVEL(mpidr, 2); aff1 = MPIDR_AFFINITY_LEVEL(mpidr, 1);

• 	val |= 1ULL << (aff2 * 4 + aff1);
  

• 	socket_id = (aff3 & SOCKET_ID_MASK) >> SOCKET_ID_SHIFT;
  

• 	if ((vm_aff3s[0] & DIE_ID_MASK) == 0x3) /* TB totem */
  

• 		pg_cfg = g_pg_cfg[0 + socket_id * die_num] & 0xff;
  

• 	else	/* TA totem */
  

• 		pg_cfg = g_pg_cfg[1 + socket_id * die_num] & 0xff;
  

TA/TB totem，这种细节的转化封装到下面的函数里面去吧？

• 	phy_aff2 = kvm_switch_aff2_logic_to_phy(pg_cfg, aff2);
  

• 	val |= 1ULL << (phy_aff2 * 4 + aff1);
  

  }

  goto out_update;

@@ -268,7 +375,7 @@ static void kvm_update_vm_lsudvmbm(struct kvm *kvm) /* nr_dies == 2 */ val = DVMBM_RANGE_TWO_DIES << DVMBM_RANGE_SHIFT | DVMBM_GRAN_CLUSTER << DVMBM_GRAN_SHIFT |

•      vm_aff3s[0] << DVMBM_DIE1_SHIFT		|
  

•      vm_aff3s[0] << DVMBM_DIE1_SHIFT  |
  

这里不用改。

     vm_aff3s[1] << DVMBM_DIE2_SHIFT;

/* and fulfill bits [43:0] */ @@ -276,11 +383,23 @@ static void kvm_update_vm_lsudvmbm(struct kvm *kvm) mpidr = cpu_logical_map(cpu); aff3 = MPIDR_AFFINITY_LEVEL(mpidr, 3); aff2 = MPIDR_AFFINITY_LEVEL(mpidr, 2);

• if (aff3 == vm_aff3s[0])
  

• 	val |= 1ULL << (aff2 + DVMBM_DIE1_CLUSTER_SHIFT);
  

• else
  

• 	val |= 1ULL << (aff2 + DVMBM_DIE2_CLUSTER_SHIFT);
  

• socket_id = (aff3 & SOCKET_ID_MASK) >> SOCKET_ID_SHIFT;
  

• if (aff3 == vm_aff3s[0]) {
  

• 	if ((vm_aff3s[0] & DIE_ID_MASK) == 0x3) /* TB totem */
  

• 		pg_cfg = g_pg_cfg[0 + socket_id * die_num] & 0xff;
  

• 	else	/* TA totem */
  

• 		pg_cfg = g_pg_cfg[1 + socket_id * die_num] & 0xff;
  

• 	phy_aff2 = kvm_switch_aff2_logic_to_phy(pg_cfg, aff2);
  

• 	val |= 1ULL << (phy_aff2 + DVMBM_DIE1_CLUSTER_SHIFT);
  

• } else {
  

• 	if ((vm_aff3s[1] & DIE_ID_MASK) == 0x3) /* TB totem */
  

• 		pg_cfg = g_pg_cfg[0 + socket_id * die_num] & 0xff;
  

• 	else	/* TA totem */
  

• 		pg_cfg = g_pg_cfg[1 + socket_id * die_num] & 0xff;
  

• 	phy_aff2 = kvm_switch_aff2_logic_to_phy(pg_cfg, aff2);
  

• 	val |= 1ULL << (phy_aff2 + DVMBM_DIE2_CLUSTER_SHIFT);
  

• }
  

  }

out_update: @@ -357,6 +476,33 @@ void kvm_hisi_dvmbm_put(struct kvm_vcpu *vcpu) */ }

+static void kvm_get_pg_cfg(u32 *pg_cfgs) +{

• void __iomem *mn_base;
• u32 i, j;
• u64 mn_phy_base;
• u32 val;
• kvm_get_socket_num(&socket_num);
• kvm_get_die_num(&die_num);
• for (i = 0; i < socket_num; i++) {

• for (j = 0; j < die_num; j++) {
  

• 	mn_phy_base = (j == 0) ? TB_MN_BASE : TA_MN_BASE;
  

• 	mn_phy_base += CHIP_ADDR_OFFSET(i);
  

• 	mn_phy_base += MN_ECO0_OFFSET;
  

• 	mn_base = ioremap(mn_phy_base, 4);
  

• 	if (!mn_base) {
  

• 		kvm_info("MN base addr ioremap fail\n");
  

• 		return;
  

• 	}
  

• 	val = readl_relaxed(mn_base);
  

• 	pg_cfgs[j + i * die_num] = val & 0xff;
  

• 	iounmap(mn_base);
  

• }
  

• }

+}

int kvm_hisi_init_dvmbm(struct kvm *kvm) { if (!kvm_dvmbm_support) @@ -367,6 +513,8 @@ int kvm_hisi_init_dvmbm(struct kvm *kvm) if (!kvm->arch.dvm_cpumask) return -ENOMEM;

• kvm_get_pg_cfg(g_pg_cfg);

这个函数，应该放在kvm初始化的时候，不是在创建vm的时候。