On Fri, Jun 28, 2024 at 09:18:16PM +0000, Oliver Upton wrote:

Hi Seb,

On Fri, Jun 21, 2024 at 12:32:29PM +0000, Sebastian Ene wrote:

quoted

Define a set of attributes used by the ptdump parser to display the
properties of a guest memory region covered by a pagetable descriptor.
Build a description of the pagetable levels and initialize the parser
with this configuration.

Signed-off-by: Sebastian Ene <redacted>

This patch should come *before* patch 4, no point in exposing the
debugfs file if we aren't ready to handle it yet.

This is true but this patch doesn't make sense without 4 because here I
add a bunch of functions which will not be invoked (they are invoked
from the debugfs calls).

IMO we can squash them (4 and 5) but it will be a bit harder to follow.
Let me know what you think, thanks.

Seb

quoted

---
 arch/arm64/kvm/ptdump.c | 143 ++++++++++++++++++++++++++++++++++++++--
 1 file changed, 137 insertions(+), 6 deletions(-)

diff --git a/arch/arm64/kvm/ptdump.c b/arch/arm64/kvm/ptdump.c
index 36dc7662729f..cc1d4fdddc6e 100644
--- a/arch/arm64/kvm/ptdump.c
+++ b/arch/arm64/kvm/ptdump.c

@@ -14,6 +14,61 @@
 #include <kvm_ptdump.h>
 
 
+#define MARKERS_LEN		(2)
+#define KVM_PGTABLE_MAX_LEVELS	(KVM_PGTABLE_LAST_LEVEL + 1)
+
+struct kvm_ptdump_guest_state {
+	struct kvm		*kvm;
+	struct pg_state		parser_state;
+	struct addr_marker	ipa_marker[MARKERS_LEN];
+	struct pg_level		level[KVM_PGTABLE_MAX_LEVELS];
+	struct ptdump_range	range[MARKERS_LEN];
+};
+
+static const struct prot_bits stage2_pte_bits[] = {
+	{
+		.mask	= PTE_VALID,
+		.val	= PTE_VALID,
+		.set	= " ",
+		.clear	= "F",
+	}, {
+		.mask	= KVM_PTE_LEAF_ATTR_HI_S2_XN | PTE_VALID,
+		.val	= KVM_PTE_LEAF_ATTR_HI_S2_XN | PTE_VALID,
+		.set	= "XN",
+		.clear	= "  ",
+	}, {
+		.mask	= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R | PTE_VALID,
+		.val	= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R | PTE_VALID,
+		.set	= "R",
+		.clear	= " ",
+	}, {
+		.mask	= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W | PTE_VALID,
+		.val	= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W | PTE_VALID,
+		.set	= "W",
+		.clear	= " ",
+	}, {
+		.mask	= KVM_PTE_LEAF_ATTR_LO_S2_AF | PTE_VALID,
+		.val	= KVM_PTE_LEAF_ATTR_LO_S2_AF | PTE_VALID,
+		.set	= "AF",
+		.clear	= "  ",

<snip>

quoted

+	}, {
+		.mask	= PTE_NG,
+		.val	= PTE_NG,
+		.set	= "FnXS",
+		.clear	= "  ",
+	}, {
+		.mask	= PTE_CONT | PTE_VALID,
+		.val	= PTE_CONT | PTE_VALID,
+		.set	= "CON",
+		.clear	= "   ",
+	}, {

</snip>

Neither of these bits are used at stage-2, why have descriptors for
them?

quoted

+static int kvm_ptdump_build_levels(struct pg_level *level, u32 start_lvl)
+{
+	static const char * const level_names[] = {"PGD", "PUD", "PMD", "PTE"};
+	u32 i = 0;
+	u64 mask = 0;
+
+	if (start_lvl > 2) {
+		pr_err("invalid start_lvl %u\n", start_lvl);
+		return -EINVAL;
+	}

	if (WARN_ON_ONCE(start_lvl >= KVM_PGTABLE_LAST_LEVEL))
		return -EINVAL;

quoted

+	for (i = 0; i < ARRAY_SIZE(stage2_pte_bits); i++)
+		mask |= stage2_pte_bits[i].mask;
+
+	for (i = start_lvl; i < KVM_PGTABLE_MAX_LEVELS; i++) {
+		strscpy(level[i].name, level_names[i], sizeof(level[i].name));
+
+		level[i].num	= ARRAY_SIZE(stage2_pte_bits);
+		level[i].bits	= stage2_pte_bits;
+		level[i].mask	= mask;
+	}
+
+	if (start_lvl > 0)
+		strscpy(level[start_lvl].name, level_names[0], sizeof(level_names[0]));

This should pass the size of @dst, not the source. This becomes slightly
more self-documenting if you use a literal for "PGD" here too.

		strscpy(level[start_lvl].name, "PGD", sizeof(level[start_lvl].name));

quoted

+	return 0;
+}
+
+static struct kvm_ptdump_guest_state
+*kvm_ptdump_parser_init(struct kvm *kvm)
+{
+	struct kvm_ptdump_guest_state *st;
+	struct kvm_s2_mmu *mmu = &kvm->arch.mmu;
+	struct kvm_pgtable *pgtable = mmu->pgt;
+	int ret;
+
+	st = kzalloc(sizeof(struct kvm_ptdump_guest_state), GFP_KERNEL_ACCOUNT);
+	if (!st)
+		return NULL;
+
+	ret = kvm_ptdump_build_levels(&st->level[0], pgtable->start_level);
+	if (ret)
+		goto free_with_state;

I don't see any value in the use of goto here, as there isn't any sort
of cascading initialization / cleanup. This also presents an opportunity
to get an error back out to the caller.

	if (ret) {
		kfree(st);
		return ERR_PTR(ret);
	}

quoted

@@ -57,22 +176,34 @@ static int kvm_ptdump_guest_show(struct seq_file *m, void *unused)
 static int kvm_ptdump_guest_open(struct inode *m, struct file *file)
 {
 	struct kvm *kvm = m->i_private;
+	struct kvm_ptdump_guest_state *st;
 	int ret;
 
 	if (!kvm_get_kvm_safe(kvm))
 		return -ENOENT;
 
-	ret = single_open(file, kvm_ptdump_guest_show, m->i_private);
-	if (ret < 0)
-		kvm_put_kvm(kvm);
+	st = kvm_ptdump_parser_init(kvm);
+	if (!st) {
+		ret = -ENOMEM;
+		goto free_with_kvm_ref;
+	}

(with the earlier suggestion)

	st = kvm_ptdump_parser_init(kvm);
	if (IS_ERR(st)) {
		ret = PTR_ERR(st);
		goto free_with_kvm_ref;
	}

Otherwise genuine KVM bugs (-EINVAL) are getting lumped into ENOMEM.

-- 
Thanks,
Oliver