Skip to contentSkip to conversation

KVM: x86: Add a return code and check kvm_page_track_init

PATCH In review [email protected] 3 participants 3 messages Aug 31, 2021

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 KVM: x86: Add a return code and check kvm_page_track_init

From: Haimin Zhang <[email protected]>

We found a null pointer deref by our modified syzkaller.

 KASAN: null-ptr-deref in range [0x0000000000000000-0x0000000000000007]
 CPU: 1 PID: 13993 Comm: syz-executor.0 Kdump: loaded Tainted: 
 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), 
 BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014
 RIP: 0010:rcu_segcblist_enqueue+0xf5/0x1d0 
 RSP: 0018:ffffc90001e1fc10 EFLAGS: 00010046
 RAX: dffffc0000000000 RBX: ffff888135c00080 RCX: ffffffff815ba8a1
 RDX: 0000000000000000 RSI: ffffc90001e1fd00 RDI: ffff888135c00080
 RBP: ffff888135c000a0 R08: 0000000000000004 R09: fffff520003c3f75
 R10: 0000000000000003 R11: fffff520003c3f75 R12: 0000000000000000
 R13: ffff888135c00080 R14: ffff888135c00040 R15: 0000000000000000
 FS:  00007fecc99f1700(0000) GS:ffff888135c00000(0000) knlGS:0000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 0000001b2f225000 CR3: 0000000093d08000 CR4: 0000000000750ee0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 PKRU: 55555554
 Call Trace:
 srcu_gp_start_if_needed+0x158/0xc60 build/../kernel/rcu/srcutree.c:823
 __synchronize_srcu+0x1dc/0x250 build/../kernel/rcu/srcutree.c:929
 kvm_mmu_uninit_vm+0x18/0x30 build/../arch/x86/kvm/mmu/mmu.c:5585
 kvm_arch_destroy_vm+0x43f/0x5c0 build/../arch/x86/kvm/x86.c:11277
 kvm_create_vm build/../arch/x86/kvm/../../../virt/kvm/kvm_main.c:1060 
 kvm_dev_ioctl_create_vm build/../arch/x86/kvm/../../../virt/kvm/kvm_main
 kvm_dev_ioctl+0xdfb/0x1860 build/../arch/x86/kvm/../../../virt/kvm/kvm_main
 vfs_ioctl build/../fs/ioctl.c:51 [inline]
 __do_sys_ioctl build/../fs/ioctl.c:1069 [inline]
 __se_sys_ioctl build/../fs/ioctl.c:1055 [inline]
 __x64_sys_ioctl+0x183/0x210 build/../fs/ioctl.c:1055
 do_syscall_x64 build/../arch/x86/entry/common.c:50 [inline]
 do_syscall_64+0x34/0xb0 build/../arch/x86/entry/common.c:80
 entry_SYSCALL_64_after_hwframe+0x44/0xae

This is because when init_srcu_struct() calls alloc_percpu(struct srcu_data) failed, kvm_page_track_init() didn't check init_srcu_struct return code.

---
 arch/x86/include/asm/kvm_page_track.h | 2 +-
 arch/x86/kvm/mmu/page_track.c         | 8 ++++++--
 arch/x86/kvm/x86.c                    | 7 +++++--
 3 files changed, 12 insertions(+), 5 deletions(-)
arch/x86/include/asm/kvm_page_track.h +1 -1
index 87bd6025d91d..6a5f3acf2b33 100644
@@ -46,7 +46,7 @@ struct kvm_page_track_notifier_node {
 			    struct kvm_page_track_notifier_node *node);
 };
 
-void kvm_page_track_init(struct kvm *kvm);
+int kvm_page_track_init(struct kvm *kvm);
 void kvm_page_track_cleanup(struct kvm *kvm);
 
 void kvm_page_track_free_memslot(struct kvm_memory_slot *slot);
arch/x86/kvm/mmu/page_track.c +6 -2
index 91a9f7e0fd91..44a67a50f6d2 100644
@@ -163,13 +163,17 @@ void kvm_page_track_cleanup(struct kvm *kvm)
 	cleanup_srcu_struct(&head->track_srcu);
 }
 
-void kvm_page_track_init(struct kvm *kvm)
+int kvm_page_track_init(struct kvm *kvm)
 {
+	int r = -ENOMEM;
 	struct kvm_page_track_notifier_head *head;
 
 	head = &kvm->arch.track_notifier_head;
-	init_srcu_struct(&head->track_srcu);
+	r = init_srcu_struct(&head->track_srcu);
+	if (r)
+		return r;
 	INIT_HLIST_HEAD(&head->track_notifier_list);
+	return r;
 }
 
 /*
arch/x86/kvm/x86.c +5 -2
index e5d5c5ed7dd4..5da76f989207 100644
@@ -11086,8 +11086,9 @@ void kvm_arch_free_vm(struct kvm *kvm)
 
 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 {
+	int r = -EINVAL;
 	if (type)
-		return -EINVAL;
+		return r;
 
 	INIT_HLIST_HEAD(&kvm->arch.mask_notifier_list);
 	INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
@@ -11121,7 +11122,9 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 
 	kvm_apicv_init(kvm);
 	kvm_hv_init_vm(kvm);
-	kvm_page_track_init(kvm);
+	r = kvm_page_track_init(kvm);
+	if (r)
+		return r;
 	kvm_mmu_init_vm(kvm);
 
 	return static_call(kvm_x86_vm_init)(kvm);
Signature · 1 line
2.27.0
#1 lore raw

For the shortlog, describe what is being fixed instead of the literal code change, otherwise the shortlog doesn't help explain _why_ a change is being made.

Quoted 27 lines · On Tue, Aug 31, 2021, [email protected] wrote:

From: Haimin Zhang <[email protected]>

We found a null pointer deref by our modified syzkaller.

 KASAN: null-ptr-deref in range [0x0000000000000000-0x0000000000000007]
 CPU: 1 PID: 13993 Comm: syz-executor.0 Kdump: loaded Tainted: 
 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), 
 BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014
 RIP: 0010:rcu_segcblist_enqueue+0xf5/0x1d0 
 RSP: 0018:ffffc90001e1fc10 EFLAGS: 00010046
 RAX: dffffc0000000000 RBX: ffff888135c00080 RCX: ffffffff815ba8a1
 RDX: 0000000000000000 RSI: ffffc90001e1fd00 RDI: ffff888135c00080
 RBP: ffff888135c000a0 R08: 0000000000000004 R09: fffff520003c3f75
 R10: 0000000000000003 R11: fffff520003c3f75 R12: 0000000000000000
 R13: ffff888135c00080 R14: ffff888135c00040 R15: 0000000000000000
 FS:  00007fecc99f1700(0000) GS:ffff888135c00000(0000) knlGS:0000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 0000001b2f225000 CR3: 0000000093d08000 CR4: 0000000000750ee0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 PKRU: 55555554
 Call Trace:
 srcu_gp_start_if_needed+0x158/0xc60 build/../kernel/rcu/srcutree.c:823
 __synchronize_srcu+0x1dc/0x250 build/../kernel/rcu/srcutree.c:929
 kvm_mmu_uninit_vm+0x18/0x30 build/../arch/x86/kvm/mmu/mmu.c:5585
 kvm_arch_destroy_vm+0x43f/0x5c0 build/../arch/x86/kvm/x86.c:11277
 kvm_create_vm build/../arch/x86/kvm/../../../virt/kvm/kvm_main.c:1060 
 kvm_dev_ioctl_create_vm build/../arch/x86/kvm/../../../virt/kvm/kvm_main
Quoted 8 lines
 kvm_dev_ioctl+0xdfb/0x1860 build/../arch/x86/kvm/../../../virt/kvm/kvm_main
 vfs_ioctl build/../fs/ioctl.c:51 [inline]
 __do_sys_ioctl build/../fs/ioctl.c:1069 [inline]
 __se_sys_ioctl build/../fs/ioctl.c:1055 [inline]
 __x64_sys_ioctl+0x183/0x210 build/../fs/ioctl.c:1055
 do_syscall_x64 build/../arch/x86/entry/common.c:50 [inline]
 do_syscall_64+0x34/0xb0 build/../arch/x86/entry/common.c:80
 entry_SYSCALL_64_after_hwframe+0x44/0xae

Newline here to make it easier to differentiate between the splat and the explanation. Though I would say hoist the explanation of the "why" to the top, e.g.

KVM: x86: Handle SRCU initialization failure during page track init

Check the return of init_srcu_struct(), which can fail due to OOM, when initializing the page track mechanism. Lack of checking leads to a NULL pointer deref found by a modified syzkaller.

<splat goes here>

Quoted 30 lines

This is because when init_srcu_struct() calls alloc_percpu(struct srcu_data) failed, kvm_page_track_init() didn't check init_srcu_struct return code.

---
 arch/x86/include/asm/kvm_page_track.h | 2 +-
 arch/x86/kvm/mmu/page_track.c         | 8 ++++++--
 arch/x86/kvm/x86.c                    | 7 +++++--
 3 files changed, 12 insertions(+), 5 deletions(-)
arch/x86/include/asm/kvm_page_track.h +1 -1
index 87bd6025d91d..6a5f3acf2b33 100644
@@ -46,7 +46,7 @@ struct kvm_page_track_notifier_node {
 			    struct kvm_page_track_notifier_node *node);
 };
 
-void kvm_page_track_init(struct kvm *kvm);
+int kvm_page_track_init(struct kvm *kvm);
 void kvm_page_track_cleanup(struct kvm *kvm);
 
 void kvm_page_track_free_memslot(struct kvm_memory_slot *slot);
arch/x86/kvm/mmu/page_track.c +0 -0
index 91a9f7e0fd91..44a67a50f6d2 100644

arch/x86/kvm/mmu/page_track.c void kvm_page_track_cleanup(struct kvm *kvm) · line 168

Quoted 7 lines
arch/x86/kvm/mmu/page_track.c +2 -1
@@ -163,13 +163,17 @@ void kvm_page_track_cleanup(struct kvm *kvm)
 	cleanup_srcu_struct(&head->track_srcu);
 }
 
-void kvm_page_track_init(struct kvm *kvm)
+int kvm_page_track_init(struct kvm *kvm)
 {
+	int r = -ENOMEM;

Unnecessary initialization.

arch/x86/kvm/mmu/page_track.c void kvm_page_track_cleanup(struct kvm *kvm) · line 176

Quoted 9 lines
arch/x86/kvm/mmu/page_track.c +4 -1
 	struct kvm_page_track_notifier_head *head;
 
 	head = &kvm->arch.track_notifier_head;
-	init_srcu_struct(&head->track_srcu);
+	r = init_srcu_struct(&head->track_srcu);
+	if (r)
+		return r;
 	INIT_HLIST_HEAD(&head->track_notifier_list);
+	return r;

Just do "return 0", which is guaranteed by the above. Or even better, I would vote for returning init_srcu_struct() directly, the ordering doesn't matter and obviously failure is a very rare occurence.

	@@ -175,8 +175,8 @@ void kvm_page_track_init(struct kvm *kvm)
        struct kvm_page_track_notifier_head *head;
 
        head = &kvm->arch.track_notifier_head;
-       init_srcu_struct(&head->track_srcu);
        INIT_HLIST_HEAD(&head->track_notifier_list);

+ return init_srcu_struct(&head->track_srcu);
}

arch/x86/kvm/x86.c void kvm_arch_free_vm(struct kvm *kvm) · line 11089

Quoted 12 lines
arch/x86/kvm/mmu/page_track.c +0 -0
 }
 
 /*
arch/x86/kvm/x86.c +1 -0
index e5d5c5ed7dd4..5da76f989207 100644
@@ -11086,8 +11086,9 @@ void kvm_arch_free_vm(struct kvm *kvm)
 
 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 {
+	int r = -EINVAL;

Unnecessary initialization.

arch/x86/kvm/x86.c void kvm_arch_free_vm(struct kvm *kvm) · line 11091

Quoted 3 lines
arch/x86/kvm/x86.c +1 -1
 	if (type)
-		return -EINVAL;
+		return r;

Unrelated and unnecessary change.

arch/x86/kvm/x86.c int kvm_arch_init_vm(struct kvm *kvm, unsigned long type) · line 11127

Quoted 11 lines
arch/x86/kvm/x86.c +3 -1
 	INIT_HLIST_HEAD(&kvm->arch.mask_notifier_list);
 	INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
@@ -11121,7 +11122,9 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 
 	kvm_apicv_init(kvm);
 	kvm_hv_init_vm(kvm);
-	kvm_page_track_init(kvm);
+	r = kvm_page_track_init(kvm);
+	if (r)
+		return r;

Hmm, so I don't see anything above this that needs to be unwound, but I'm still worried this will be hard to audit/maintain.

As an alternative "fix", about dropping kvm->arch.track_notifier_head.track_srcu and using kvm->srcu? kvm_page_track_write() pretty much _has_ to hold that since the caller is writing guest memory, and conversely kvm_page_track_flush_slot() _can't_ hold it because the caller is modifying memslots and thus would deadlock if it held kvm->srcu for read. In other words, kvm_page_track_write() can rely (assert?) on vcpu->srcu_idx, and kvm_page_track_flush_slot() can take and release kvm->srcu.

Practially speaking, (Un)Registering is going to happen only at VM creation so waiting all kvm->srcu readers instead of just page track readers should not be a problem.

Quoted 6 lines
arch/x86/kvm/x86.c +0 -0
 	kvm_mmu_init_vm(kvm);
 
 	return static_call(kvm_x86_vm_init)(kvm);
Signature · 1 line
2.27.0
#2 lore raw
In reply to [email protected]
Quoted 65 lines · [email protected] writes:

From: Haimin Zhang <[email protected]>

We found a null pointer deref by our modified syzkaller.

 KASAN: null-ptr-deref in range [0x0000000000000000-0x0000000000000007]
 CPU: 1 PID: 13993 Comm: syz-executor.0 Kdump: loaded Tainted: 
 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), 
 BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014
 RIP: 0010:rcu_segcblist_enqueue+0xf5/0x1d0 
 RSP: 0018:ffffc90001e1fc10 EFLAGS: 00010046
 RAX: dffffc0000000000 RBX: ffff888135c00080 RCX: ffffffff815ba8a1
 RDX: 0000000000000000 RSI: ffffc90001e1fd00 RDI: ffff888135c00080
 RBP: ffff888135c000a0 R08: 0000000000000004 R09: fffff520003c3f75
 R10: 0000000000000003 R11: fffff520003c3f75 R12: 0000000000000000
 R13: ffff888135c00080 R14: ffff888135c00040 R15: 0000000000000000
 FS:  00007fecc99f1700(0000) GS:ffff888135c00000(0000) knlGS:0000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 0000001b2f225000 CR3: 0000000093d08000 CR4: 0000000000750ee0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 PKRU: 55555554
 Call Trace:
 srcu_gp_start_if_needed+0x158/0xc60 build/../kernel/rcu/srcutree.c:823
 __synchronize_srcu+0x1dc/0x250 build/../kernel/rcu/srcutree.c:929
 kvm_mmu_uninit_vm+0x18/0x30 build/../arch/x86/kvm/mmu/mmu.c:5585
 kvm_arch_destroy_vm+0x43f/0x5c0 build/../arch/x86/kvm/x86.c:11277
 kvm_create_vm build/../arch/x86/kvm/../../../virt/kvm/kvm_main.c:1060 
 kvm_dev_ioctl_create_vm build/../arch/x86/kvm/../../../virt/kvm/kvm_main
 kvm_dev_ioctl+0xdfb/0x1860 build/../arch/x86/kvm/../../../virt/kvm/kvm_main
 vfs_ioctl build/../fs/ioctl.c:51 [inline]
 __do_sys_ioctl build/../fs/ioctl.c:1069 [inline]
 __se_sys_ioctl build/../fs/ioctl.c:1055 [inline]
 __x64_sys_ioctl+0x183/0x210 build/../fs/ioctl.c:1055
 do_syscall_x64 build/../arch/x86/entry/common.c:50 [inline]
 do_syscall_64+0x34/0xb0 build/../arch/x86/entry/common.c:80
 entry_SYSCALL_64_after_hwframe+0x44/0xae

This is because when init_srcu_struct() calls alloc_percpu(struct srcu_data) failed, kvm_page_track_init() didn't check init_srcu_struct return code.

---
 arch/x86/include/asm/kvm_page_track.h | 2 +-
 arch/x86/kvm/mmu/page_track.c         | 8 ++++++--
 arch/x86/kvm/x86.c                    | 7 +++++--
 3 files changed, 12 insertions(+), 5 deletions(-)
arch/x86/include/asm/kvm_page_track.h +1 -1
index 87bd6025d91d..6a5f3acf2b33 100644
@@ -46,7 +46,7 @@ struct kvm_page_track_notifier_node {
 			    struct kvm_page_track_notifier_node *node);
 };
 
-void kvm_page_track_init(struct kvm *kvm);
+int kvm_page_track_init(struct kvm *kvm);
 void kvm_page_track_cleanup(struct kvm *kvm);
 
 void kvm_page_track_free_memslot(struct kvm_memory_slot *slot);
arch/x86/kvm/mmu/page_track.c +0 -0
index 91a9f7e0fd91..44a67a50f6d2 100644

arch/x86/kvm/mmu/page_track.c void kvm_page_track_cleanup(struct kvm *kvm) · line 168

Quoted 7 lines
arch/x86/kvm/mmu/page_track.c +2 -1
@@ -163,13 +163,17 @@ void kvm_page_track_cleanup(struct kvm *kvm)
 	cleanup_srcu_struct(&head->track_srcu);
 }
 
-void kvm_page_track_init(struct kvm *kvm)
+int kvm_page_track_init(struct kvm *kvm)
 {
+	int r = -ENOMEM;

Nitpick: pointless initializer, the value is always overwritten by init_srcu_struct()'s return value.

Quoted 17 lines
arch/x86/kvm/mmu/page_track.c +4 -1
 	struct kvm_page_track_notifier_head *head;
 
 	head = &kvm->arch.track_notifier_head;
-	init_srcu_struct(&head->track_srcu);
+	r = init_srcu_struct(&head->track_srcu);
+	if (r)
+		return r;
 	INIT_HLIST_HEAD(&head->track_notifier_list);
+	return r;
 }
 
 /*
arch/x86/kvm/x86.c +0 -0
index e5d5c5ed7dd4..5da76f989207 100644

arch/x86/kvm/x86.c void kvm_arch_free_vm(struct kvm *kvm) · line 11089

Quoted 4 lines
arch/x86/kvm/x86.c +1 -0
@@ -11086,8 +11086,9 @@ void kvm_arch_free_vm(struct kvm *kvm)
 
 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 {
+	int r = -EINVAL;

Blank line missing here.

arch/x86/kvm/x86.c void kvm_arch_free_vm(struct kvm *kvm) · line 11091

Quoted 3 lines
arch/x86/kvm/x86.c +1 -1
 	if (type)
-		return -EINVAL;
+		return r;

I'd keep this code as-is and dropped then-pointless initializer to -EINVAL, it's OK to return directly:

	int ret;

	if (type)
		return -EINVAL;

        ...

        ret = kvm_page_track_init(kvm);
	if (ret)
		return ret;
Quoted 14 lines
arch/x86/kvm/x86.c +3 -1
 	INIT_HLIST_HEAD(&kvm->arch.mask_notifier_list);
 	INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
@@ -11121,7 +11122,9 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 
 	kvm_apicv_init(kvm);
 	kvm_hv_init_vm(kvm);
-	kvm_page_track_init(kvm);
+	r = kvm_page_track_init(kvm);
+	if (r)
+		return r;
 	kvm_mmu_init_vm(kvm);
 
 	return static_call(kvm_x86_vm_init)(kvm);
Signature · 1 line
Vitaly