Skip to contentSkip to conversation

KVM: TDX: Use struct_size and simplify tdx_get_capabilities

PATCH RESEND In review Thorsten Blum 3 participants 8 messages Nov 12, 2025 – Nov 13, 2025

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH RESEND 1/1 KVM: TDX: Use struct_size and simplify tdx_get_capabilities

Retrieve the number of user entries with get_user() first and return -E2BIG early if 'user_caps' is too small to fit 'caps'.

Allocate memory for 'caps' only after checking the user buffer's number of entries, thus removing two gotos and the need for premature freeing.

Use struct_size() instead of manually calculating the number of bytes to allocate for 'caps', including the nested flexible array.

Finally, copy 'caps' to user space with a single copy_to_user() call.

---
 arch/x86/kvm/vmx/tdx.c | 32 ++++++++++++--------------------
 1 file changed, 12 insertions(+), 20 deletions(-)
arch/x86/kvm/vmx/tdx.c +12 -20
index 0a49c863c811..23d638b4a003 100644
@@ -2282,37 +2282,29 @@ static int tdx_get_capabilities(struct kvm_tdx_cmd *cmd)
 	if (cmd->flags)
 		return -EINVAL;
 
-	caps = kzalloc(sizeof(*caps) +
-		       sizeof(struct kvm_cpuid_entry2) * td_conf->num_cpuid_config,
-		       GFP_KERNEL);
-	if (!caps)
-		return -ENOMEM;
-
 	user_caps = u64_to_user_ptr(cmd->data);
-	if (get_user(nr_user_entries, &user_caps->cpuid.nent)) {
-		ret = -EFAULT;
-		goto out;
-	}
+	ret = get_user(nr_user_entries, &user_caps->cpuid.nent);
+	if (ret)
+		return ret;
 
-	if (nr_user_entries < td_conf->num_cpuid_config) {
-		ret = -E2BIG;
-		goto out;
-	}
+	if (nr_user_entries < td_conf->num_cpuid_config)
+		return -E2BIG;
+
+	caps = kzalloc(struct_size(caps, cpuid.entries,
+				   td_conf->num_cpuid_config), GFP_KERNEL);
+	if (!caps)
+		return -ENOMEM;
 
 	ret = init_kvm_tdx_caps(td_conf, caps);
 	if (ret)
 		goto out;
 
-	if (copy_to_user(user_caps, caps, sizeof(*caps))) {
+	if (copy_to_user(user_caps, caps, struct_size(caps, cpuid.entries,
+						      caps->cpuid.nent))) {
 		ret = -EFAULT;
 		goto out;
 	}
 
-	if (copy_to_user(user_caps->cpuid.entries, caps->cpuid.entries,
-			 caps->cpuid.nent *
-			 sizeof(caps->cpuid.entries[0])))
-		ret = -EFAULT;
-
 out:
 	/* kfree() accepts NULL. */
 	kfree(caps);
Signature · 1 line
2.51.1
maintainer Tested-by on v1 1/1
#1 lore raw

On Wed, 2025-11-12 at 18:16 +0100, Thorsten Blum wrote:

kvm x86 logs are suggested to start with a short summary of the patch. Maybe:

Simplify the logic for copying the KVM_TDX_CAPABILITIES struct to userspace.

It looks like you are conducting a treewide pattern matching cleanup?

Quoted 10 lines
Quoted 10 lines

Retrieve the number of user entries with get_user() first and return -E2BIG early if 'user_caps' is too small to fit 'caps'.

Allocate memory for 'caps' only after checking the user buffer's number of entries, thus removing two gotos and the need for premature freeing.

Use struct_size() instead of manually calculating the number of bytes to allocate for 'caps', including the nested flexible array.

Finally, copy 'caps' to user space with a single copy_to_user() call.

In the handling of get_user(nr_user_entries, &user_caps->cpuid.nent), the old code forced -EFAULT, this patch doesn't. But it leaves the copy_to_user()'s to still force EFAULT. Why?

Quoted 60 lines
Quoted 60 lines
---
 arch/x86/kvm/vmx/tdx.c | 32 ++++++++++++--------------------
 1 file changed, 12 insertions(+), 20 deletions(-)
arch/x86/kvm/vmx/tdx.c +12 -20
index 0a49c863c811..23d638b4a003 100644
@@ -2282,37 +2282,29 @@ static int tdx_get_capabilities(struct kvm_tdx_cmd *cmd)
 	if (cmd->flags)
 		return -EINVAL;
 
-	caps = kzalloc(sizeof(*caps) +
-		       sizeof(struct kvm_cpuid_entry2) * td_conf->num_cpuid_config,
-		       GFP_KERNEL);
-	if (!caps)
-		return -ENOMEM;
-
 	user_caps = u64_to_user_ptr(cmd->data);
-	if (get_user(nr_user_entries, &user_caps->cpuid.nent)) {
-		ret = -EFAULT;
-		goto out;
-	}
+	ret = get_user(nr_user_entries, &user_caps->cpuid.nent);
+	if (ret)
+		return ret;
 
-	if (nr_user_entries < td_conf->num_cpuid_config) {
-		ret = -E2BIG;
-		goto out;
-	}
+	if (nr_user_entries < td_conf->num_cpuid_config)
+		return -E2BIG;
+
+	caps = kzalloc(struct_size(caps, cpuid.entries,
+				   td_conf->num_cpuid_config), GFP_KERNEL);
+	if (!caps)
+		return -ENOMEM;
 
 	ret = init_kvm_tdx_caps(td_conf, caps);
 	if (ret)
 		goto out;
 
-	if (copy_to_user(user_caps, caps, sizeof(*caps))) {
+	if (copy_to_user(user_caps, caps, struct_size(caps, cpuid.entries,
+						      caps->cpuid.nent))) {
 		ret = -EFAULT;
 		goto out;
 	}
 
-	if (copy_to_user(user_caps->cpuid.entries, caps->cpuid.entries,
-			 caps->cpuid.nent *
-			 sizeof(caps->cpuid.entries[0])))
-		ret = -EFAULT;
-
 out:
 	/* kfree() accepts NULL. */
 	kfree(caps);
#2 lore raw
Quoted 5 lines · Rick P Edgecombe

On Wed, 2025-11-12 at 18:16 +0100, Thorsten Blum wrote:

kvm x86 logs are suggested to start with a short summary of the patch. Maybe:

Simplify the logic for copying the KVM_TDX_CAPABILITIES struct to userspace.

Yeah, I have this locally as two separate patches:

KVM: TDX: Use struct_size to simplify tdx_get_capabilities()
KVM: TDX: Check size of user's kvm_tdx_capabilities array before allocating

Your CI caught me just in time; I applied this locally last week, but haven't fully pushed it to kvm-x86 yet. :-)

Quoted 8 lines

It looks like you are conducting a treewide pattern matching cleanup?

Quoted 6 lines
Quoted 6 lines

Retrieve the number of user entries with get_user() first and return -E2BIG early if 'user_caps' is too small to fit 'caps'.

Allocate memory for 'caps' only after checking the user buffer's number of entries, thus removing two gotos and the need for premature freeing.

Quoted 8 lines
Quoted 4 lines
Quoted 4 lines

Use struct_size() instead of manually calculating the number of bytes to allocate for 'caps', including the nested flexible array.

Finally, copy 'caps' to user space with a single copy_to_user() call.

In the handling of get_user(nr_user_entries, &user_caps->cpuid.nent), the old code forced -EFAULT, this patch doesn't. But it leaves the copy_to_user()'s to still force EFAULT. Why?

I'll tweak it to explicitly return -EFAULT. Doesn't matter terribly, but KVM's standard pattern is to explicitly return -EFAULT.

Quoted 1 line
#7 lore raw
In reply to Edgecombe, Rick P
Quoted 1 line · Rick P

It looks like you are conducting a treewide pattern matching cleanup?

Just a few instances here and there, but not really treewide.

Quoted 3 lines

In the handling of get_user(nr_user_entries, &user_caps->cpuid.nent), the old code forced -EFAULT, this patch doesn't. But it leaves the copy_to_user()'s to still force EFAULT. Why?

get_user() already returns -EFAULT and the error can just be forwarded, whereas copy_to_user() returns the number of bytes that could not be copied and we must return -EFAULT manually.

Quoted 1 line

Thanks,
Thorsten

#3 lore raw
In reply to Sean Christopherson
Quoted 2 lines · Sean Christopherson

Your CI caught me just in time; I applied this locally last week, but haven't fully pushed it to kvm-x86 yet. :-)

The TDX CI tracks some upstream branches. Is there one in kvm_x86 tree that would be useful? It's not foolproof enough to warrant sending out automated mails. But we monitor it and might notice TDX specific issues. Ideally we would not be chasing generic bugs in like scratch code not headed upstream or something.

#4 lore raw
Quoted 9 lines · Rick P Edgecombe
Quoted 2 lines · Sean Christopherson

Your CI caught me just in time; I applied this locally last week, but haven't fully pushed it to kvm-x86 yet. :-)

The TDX CI tracks some upstream branches. Is there one in kvm_x86 tree that would be useful? It's not foolproof enough to warrant sending out automated mails. But we monitor it and might notice TDX specific issues. Ideally we would not be chasing generic bugs in like scratch code not headed upstream or something.

Assuming you're tracking linux-next, I wouldn't bother adding kvm-x86 as kvm-x86/next is fed into linux-next. I do push to topic branches, e.g. kvm-x86/tdx, before merging to kvm-x86/next, but at best you might "gain" a day or two, and the entire reason I do "half" pushes is so that I can run everything through my testing before "officially" publishing it to the world.

All in all, explicitly tracking anything kvm-x86 would likely be a net negative.

#5 lore raw
Quoted 7 lines · Sean Christopherson

Assuming you're tracking linux-next, I wouldn't bother adding kvm-x86 as kvm-x86/next is fed into linux-next.  I do push to topic branches, e.g. kvm-x86/tdx, before merging to kvm-x86/next, but at best you might "gain" a day or two, and the entire reason I do "half" pushes is so that I can run everything through my testing before "officially" publishing it to the world.

All in all, explicitly tracking anything kvm-x86 would likely be a net negative.

Yea, linux-next and Linus releases. Ok, we'll leave it. I was just thinking about your lack of TDX testing setup, and wondering if it could help. All good.

#6 lore raw
Quoted 11 lines · Rick P Edgecombe
Quoted 7 lines · Sean Christopherson

Assuming you're tracking linux-next, I wouldn't bother adding kvm-x86 as kvm-x86/next is fed into linux-next.  I do push to topic branches, e.g. kvm-x86/tdx, before merging to kvm-x86/next, but at best you might "gain" a day or two, and the entire reason I do "half" pushes is so that I can run everything through my testing before "officially" publishing it to the world.

All in all, explicitly tracking anything kvm-x86 would likely be a net negative.

Yea, linux-next and Linus releases. Ok, we'll leave it. I was just thinking about your lack of TDX testing setup, and wondering if it could help. All good.

Heh, I appreciate the offer, but you probably shouldn't encourage my laziness at this point :-)