KVM: kvm_io_bus_unregister_dev() should never fail

No caller currently checks the return value of kvm_io_bus_unregister_dev(). This is evil, as all callers silently go on freeing their device. A stale reference will remain in the io_bus, getting at least used again, when the iobus gets teared down on kvm_destroy_vm() - leading to use after free errors.

There is nothing the callers could do, except retrying over and over again.

So let's simply remove the bus altogether, print an error and make sure no one can access this broken bus again (returning -ENOMEM on any attempt to access it).

Fixes e93f8a0f821e ("KVM: convert io_bus to SRCU")

Cc [email protected] # 3.4+

Reported-by Dmitry Vyukov <[email protected]>

Signed-off-by David Hildenbrand <[email protected]>

---

Based on kvm/queue, where we just got 2a108a4e7c1 ("KVM: x86: clear bus pointer when destroyed"), which added a check we need here.

v1 -> v2:
- added a check in kvm_destroy_vm()
- added a check in virt/kvm/eventfd.c

Using 'git grep -C 4 "kvm->buses"' should help to find all users. The other user in virt/kvm/eventfd.c should be fine.

---
 include/linux/kvm_host.h |  4 ++--
 virt/kvm/eventfd.c       |  3 ++-
 virt/kvm/kvm_main.c      | 42 +++++++++++++++++++++++++-----------------
 3 files changed, 29 insertions(+), 20 deletions(-)

No caller currently checks the return value of kvm_io_bus_unregister_dev(). This is evil, as all callers silently go on freeing their device. A stale reference will remain in the io_bus, getting at least used again, when the iobus gets teared down on kvm_destroy_vm() - leading to use after free errors.

There is nothing the callers could do, except retrying over and over again.

So let's simply remove the bus altogether, print an error and make sure no one can access this broken bus again (returning -ENOMEM on any attempt to access it).

Fixes e93f8a0f821e ("KVM: convert io_bus to SRCU")

Cc [email protected] # 3.4+

Reported-by Dmitry Vyukov <[email protected]>

Signed-off-by David Hildenbrand <[email protected]>

---

Based on kvm/queue, where we just got 2a108a4e7c1 ("KVM: x86: clear bus pointer when destroyed"), which added a check we need here.

v1 -> v2:
- added a check in kvm_destroy_vm()
- added a check in virt/kvm/eventfd.c

Using 'git grep -C 4 "kvm->buses"' should help to find all users. The other user in virt/kvm/eventfd.c should be fine.

---
 include/linux/kvm_host.h |  4 ++--
 virt/kvm/eventfd.c       |  3 ++-
 virt/kvm/kvm_main.c      | 42 +++++++++++++++++++++++++-----------------
 3 files changed, 29 insertions(+), 20 deletions(-)

include/linux/kvm_host.h +2 -2

index 2c14ad9..d025074 100644
@@ -162,8 +162,8 @@ int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
 		    int len, void *val);
 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
 			    int len, struct kvm_io_device *dev);
-int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
-			      struct kvm_io_device *dev);
+void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
+			       struct kvm_io_device *dev);
 struct kvm_io_device *kvm_io_bus_get_dev(struct kvm *kvm, enum kvm_bus bus_idx,
 					 gpa_t addr);
 

virt/kvm/eventfd.c +2 -1

index a29786d..4d28a9d 100644
@@ -870,7 +870,8 @@ kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 			continue;
 
 		kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
-		kvm->buses[bus_idx]->ioeventfd_count--;
+		if (kvm->buses[bus_idx])
+			kvm->buses[bus_idx]->ioeventfd_count--;
 		ioeventfd_release(p);
 		ret = 0;
 		break;

virt/kvm/kvm_main.c +15 -14

index 7445566..ef1aa7f 100644
@@ -728,7 +728,8 @@ static void kvm_destroy_vm(struct kvm *kvm)
 	spin_unlock(&kvm_lock);
 	kvm_free_irq_routing(kvm);
 	for (i = 0; i < KVM_NR_BUSES; i++) {
-		kvm_io_bus_destroy(kvm->buses[i]);
+		if (kvm->buses[i])
+			kvm_io_bus_destroy(kvm->buses[i]);
 		kvm->buses[i] = NULL;
 	}
 	kvm_coalesced_mmio_free(kvm);

@@ -3476,6 +3477,8 @@ int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
 	};
 
 	bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+	if (!bus)
+		return -ENOMEM;
 	r = __kvm_io_bus_write(vcpu, bus, &range, val);
 	return r < 0 ? r : 0;
 }

@@ -3493,6 +3496,8 @@ int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx,
 	};
 
 	bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+	if (!bus)
+		return -ENOMEM;
 
 	/* First try the device referenced by cookie. */
 	if ((cookie >= 0) && (cookie < bus->dev_count) &&

@@ -3543,6 +3548,8 @@ int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
 	};
 
 	bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+	if (!bus)
+		return -ENOMEM;
 	r = __kvm_io_bus_read(vcpu, bus, &range, val);
 	return r < 0 ? r : 0;
 }

@@ -3555,6 +3562,9 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
 	struct kvm_io_bus *new_bus, *bus;
 
 	bus = kvm->buses[bus_idx];
+	if (!bus)
+		return -ENOMEM;
+
 	/* exclude ioeventfd which is limited by maximum fd */
 	if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
 		return -ENOSPC;

@@ -3574,45 +3584,41 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
 }
 
 /* Caller must hold slots_lock. */
-int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
-			      struct kvm_io_device *dev)
+void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
+			       struct kvm_io_device *dev)
 {
-	int i, r;
+	int i;
 	struct kvm_io_bus *new_bus, *bus;
 
 	bus = kvm->buses[bus_idx];
-
-	/*
-	 * It's possible the bus being released before hand. If so,
-	 * we're done here.
-	 */
 	if (!bus)
-		return 0;
+		return;
 
-	r = -ENOENT;
 	for (i = 0; i < bus->dev_count; i++)
 		if (bus->range[i].dev == dev) {
-			r = 0;
 			break;
 		}
 
-	if (r)
-		return r;

Quoted 147 lines · David Hildenbrand

No caller currently checks the return value of kvm_io_bus_unregister_dev(). This is evil, as all callers silently go on freeing their device. A stale reference will remain in the io_bus, getting at least used again, when the iobus gets teared down on kvm_destroy_vm() - leading to use after free errors.

There is nothing the callers could do, except retrying over and over again.

So let's simply remove the bus altogether, print an error and make sure no one can access this broken bus again (returning -ENOMEM on any attempt to access it).

Fixes e93f8a0f821e ("KVM: convert io_bus to SRCU")

Cc [email protected] # 3.4+

Reported-by Dmitry Vyukov <[email protected]>

Signed-off-by David Hildenbrand <[email protected]>

---

Based on kvm/queue, where we just got 2a108a4e7c1 ("KVM: x86: clear bus pointer when destroyed"), which added a check we need here.

v1 -> v2:
- added a check in kvm_destroy_vm()
- added a check in virt/kvm/eventfd.c

Using 'git grep -C 4 "kvm->buses"' should help to find all users. The other user in virt/kvm/eventfd.c should be fine.

---
 include/linux/kvm_host.h |  4 ++--
 virt/kvm/eventfd.c       |  3 ++-
 virt/kvm/kvm_main.c      | 42 +++++++++++++++++++++++++-----------------
 3 files changed, 29 insertions(+), 20 deletions(-)

include/linux/kvm_host.h +2 -2

index 2c14ad9..d025074 100644
@@ -162,8 +162,8 @@ int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
                  int len, void *val);
 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
                          int len, struct kvm_io_device *dev);
-int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
-                           struct kvm_io_device *dev);
+void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
+                            struct kvm_io_device *dev);
 struct kvm_io_device *kvm_io_bus_get_dev(struct kvm *kvm, enum kvm_bus bus_idx,
                                       gpa_t addr);
 

virt/kvm/eventfd.c +2 -1

index a29786d..4d28a9d 100644
@@ -870,7 +870,8 @@ kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
                      continue;
 
              kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
-             kvm->buses[bus_idx]->ioeventfd_count--;
+             if (kvm->buses[bus_idx])
+                     kvm->buses[bus_idx]->ioeventfd_count--;
              ioeventfd_release(p);
              ret = 0;
              break;

virt/kvm/kvm_main.c +15 -14

index 7445566..ef1aa7f 100644
@@ -728,7 +728,8 @@ static void kvm_destroy_vm(struct kvm *kvm)
      spin_unlock(&kvm_lock);
      kvm_free_irq_routing(kvm);
      for (i = 0; i < KVM_NR_BUSES; i++) {
-             kvm_io_bus_destroy(kvm->buses[i]);
+             if (kvm->buses[i])
+                     kvm_io_bus_destroy(kvm->buses[i]);
              kvm->buses[i] = NULL;
      }
      kvm_coalesced_mmio_free(kvm);

@@ -3476,6 +3477,8 @@ int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
      };
 
      bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+     if (!bus)
+             return -ENOMEM;
      r = __kvm_io_bus_write(vcpu, bus, &range, val);
      return r < 0 ? r : 0;
 }

@@ -3493,6 +3496,8 @@ int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx,
      };
 
      bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+     if (!bus)
+             return -ENOMEM;
 
      /* First try the device referenced by cookie. */
      if ((cookie >= 0) && (cookie < bus->dev_count) &&

@@ -3543,6 +3548,8 @@ int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
      };
 
      bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+     if (!bus)
+             return -ENOMEM;
      r = __kvm_io_bus_read(vcpu, bus, &range, val);
      return r < 0 ? r : 0;
 }

@@ -3555,6 +3562,9 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
      struct kvm_io_bus *new_bus, *bus;
 
      bus = kvm->buses[bus_idx];
+     if (!bus)
+             return -ENOMEM;
+
      /* exclude ioeventfd which is limited by maximum fd */
      if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
              return -ENOSPC;

@@ -3574,45 +3584,41 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
 }
 
 /* Caller must hold slots_lock. */
-int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
-                           struct kvm_io_device *dev)
+void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
+                            struct kvm_io_device *dev)
 {
-     int i, r;
+     int i;
      struct kvm_io_bus *new_bus, *bus;
 
      bus = kvm->buses[bus_idx];
-
-     /*
-      * It's possible the bus being released before hand. If so,
-      * we're done here.
-      */
      if (!bus)
-             return 0;
+             return;
 
-     r = -ENOENT;
      for (i = 0; i < bus->dev_count; i++)
              if (bus->range[i].dev == dev) {
-                     r = 0;
                      break;
              }
 
-     if (r)
-             return r;

virt/kvm/kvm_main.c int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr, · line 3609

Quoted 10 lines

virt/kvm/kvm_main.c +5 -2

+     if (i == bus->dev_count)
+             return;
 
      new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count - 1) *
                        sizeof(struct kvm_io_range)), GFP_KERNEL);
-     if (!new_bus)
-             return -ENOMEM;
+     if (!new_bus)  {
+             pr_err("kvm: failed to shrink bus, removing it completely\n");
+             goto broken;

The guest will fail in mysterious ways, if you do this (and io_bus_unregister_dev can be called during runtime): in-kernel device accesses will fail with unknown behaviour in the guest.

Can't you retry a handful of times with GFP_KERNEL before switching to GFP_ATOMIC? (which in case fails the machine is likely to be crashing soon).

Quoted 13 lines

Quoted 12 lines

virt/kvm/kvm_main.c +5 -3

-             return r;
+     if (i == bus->dev_count)
+             return;
 
      new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count - 1) *
                        sizeof(struct kvm_io_range)), GFP_KERNEL);
-     if (!new_bus)
-             return -ENOMEM;
+     if (!new_bus)  {
+             pr_err("kvm: failed to shrink bus, removing it completely\n");
+             goto broken;

Quoted 3 lines

The guest will fail in mysterious ways, if you do this (and io_bus_unregister_dev can be called during runtime): in-kernel device accesses will fail with unknown behaviour in the guest.

Quoted 3 lines

Can't you retry a handful of times with GFP_KERNEL before switching to GFP_ATOMIC? (which in case fails the machine is likely to be crashing soon).

The process can run in a cgroup, then kmalloc failure has nothing to do with overall memory consumption. Machine can be perfectly fine. Also, this very process can be chosen as an OOM kill target, then it needs to gracefully deal with kmalloc failure and proceed to a termination point.
Generally retrying something in a loop does not look like a solid plan to deal with errors.

Quoted 13 lines

Quoted 13 lines

Quoted 12 lines

virt/kvm/kvm_main.c +5 -3

-             return r;
+     if (i == bus->dev_count)
+             return;
 
      new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count - 1) *
                        sizeof(struct kvm_io_range)), GFP_KERNEL);
-     if (!new_bus)
-             return -ENOMEM;
+     if (!new_bus)  {
+             pr_err("kvm: failed to shrink bus, removing it completely\n");
+             goto broken;

Quoted 3 lines

Quoted 3 lines

The guest will fail in mysterious ways, if you do this (and io_bus_unregister_dev can be called during runtime): in-kernel device accesses will fail with unknown behaviour in the guest.

Actually, the next access to the BUS should result in -ENOMEM. And the error message should be enough to then figure out what went wrong.

However, to hit this scenario at all feels very unlikely. So I would like to avoid advanced allocation schemes.

Quoted 163 lines · Marcelo Tosatti

Quoted 147 lines · David Hildenbrand

No caller currently checks the return value of kvm_io_bus_unregister_dev(). This is evil, as all callers silently go on freeing their device. A stale reference will remain in the io_bus, getting at least used again, when the iobus gets teared down on kvm_destroy_vm() - leading to use after free errors.

There is nothing the callers could do, except retrying over and over again.

So let's simply remove the bus altogether, print an error and make sure no one can access this broken bus again (returning -ENOMEM on any attempt to access it).

Fixes e93f8a0f821e ("KVM: convert io_bus to SRCU")

Cc [email protected] # 3.4+

Reported-by Dmitry Vyukov <[email protected]>

Signed-off-by David Hildenbrand <[email protected]>

---

Based on kvm/queue, where we just got 2a108a4e7c1 ("KVM: x86: clear bus pointer when destroyed"), which added a check we need here.

v1 -> v2:
- added a check in kvm_destroy_vm()
- added a check in virt/kvm/eventfd.c

Using 'git grep -C 4 "kvm->buses"' should help to find all users. The other user in virt/kvm/eventfd.c should be fine.

---
 include/linux/kvm_host.h |  4 ++--
 virt/kvm/eventfd.c       |  3 ++-
 virt/kvm/kvm_main.c      | 42 +++++++++++++++++++++++++-----------------
 3 files changed, 29 insertions(+), 20 deletions(-)

include/linux/kvm_host.h +2 -2

index 2c14ad9..d025074 100644
@@ -162,8 +162,8 @@ int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
                  int len, void *val);
 int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
                          int len, struct kvm_io_device *dev);
-int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
-                           struct kvm_io_device *dev);
+void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
+                            struct kvm_io_device *dev);
 struct kvm_io_device *kvm_io_bus_get_dev(struct kvm *kvm, enum kvm_bus bus_idx,
                                       gpa_t addr);
 

virt/kvm/eventfd.c +2 -1

index a29786d..4d28a9d 100644
@@ -870,7 +870,8 @@ kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
                      continue;
 
              kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
-             kvm->buses[bus_idx]->ioeventfd_count--;
+             if (kvm->buses[bus_idx])
+                     kvm->buses[bus_idx]->ioeventfd_count--;
              ioeventfd_release(p);
              ret = 0;
              break;

virt/kvm/kvm_main.c +15 -14

index 7445566..ef1aa7f 100644
@@ -728,7 +728,8 @@ static void kvm_destroy_vm(struct kvm *kvm)
      spin_unlock(&kvm_lock);
      kvm_free_irq_routing(kvm);
      for (i = 0; i < KVM_NR_BUSES; i++) {
-             kvm_io_bus_destroy(kvm->buses[i]);
+             if (kvm->buses[i])
+                     kvm_io_bus_destroy(kvm->buses[i]);
              kvm->buses[i] = NULL;
      }
      kvm_coalesced_mmio_free(kvm);

@@ -3476,6 +3477,8 @@ int kvm_io_bus_write(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
      };
 
      bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+     if (!bus)
+             return -ENOMEM;
      r = __kvm_io_bus_write(vcpu, bus, &range, val);
      return r < 0 ? r : 0;
 }

@@ -3493,6 +3496,8 @@ int kvm_io_bus_write_cookie(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx,
      };
 
      bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+     if (!bus)
+             return -ENOMEM;
 
      /* First try the device referenced by cookie. */
      if ((cookie >= 0) && (cookie < bus->dev_count) &&

@@ -3543,6 +3548,8 @@ int kvm_io_bus_read(struct kvm_vcpu *vcpu, enum kvm_bus bus_idx, gpa_t addr,
      };
 
      bus = srcu_dereference(vcpu->kvm->buses[bus_idx], &vcpu->kvm->srcu);
+     if (!bus)
+             return -ENOMEM;
      r = __kvm_io_bus_read(vcpu, bus, &range, val);
      return r < 0 ? r : 0;
 }

@@ -3555,6 +3562,9 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
      struct kvm_io_bus *new_bus, *bus;
 
      bus = kvm->buses[bus_idx];
+     if (!bus)
+             return -ENOMEM;
+
      /* exclude ioeventfd which is limited by maximum fd */
      if (bus->dev_count - bus->ioeventfd_count > NR_IOBUS_DEVS - 1)
              return -ENOSPC;

@@ -3574,45 +3584,41 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
 }
 
 /* Caller must hold slots_lock. */
-int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
-                           struct kvm_io_device *dev)
+void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
+                            struct kvm_io_device *dev)
 {
-     int i, r;
+     int i;
      struct kvm_io_bus *new_bus, *bus;
 
      bus = kvm->buses[bus_idx];
-
-     /*
-      * It's possible the bus being released before hand. If so,
-      * we're done here.
-      */
      if (!bus)
-             return 0;
+             return;
 
-     r = -ENOENT;
      for (i = 0; i < bus->dev_count; i++)
              if (bus->range[i].dev == dev) {
-                     r = 0;
                      break;
              }
 
-     if (r)
-             return r;

virt/kvm/kvm_main.c int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr, · line 3609

Quoted 10 lines

virt/kvm/kvm_main.c +5 -2

+     if (i == bus->dev_count)
+             return;
 
      new_bus = kmalloc(sizeof(*bus) + ((bus->dev_count - 1) *
                        sizeof(struct kvm_io_range)), GFP_KERNEL);
-     if (!new_bus)
-             return -ENOMEM;
+     if (!new_bus)  {
+             pr_err("kvm: failed to shrink bus, removing it completely\n");
+             goto broken;

The guest will fail in mysterious ways, if you do this (and io_bus_unregister_dev can be called during runtime): in-kernel device accesses will fail with unknown behaviour in the guest.

Quoted 2 lines

Can't you retry a handful of times with GFP_KERNEL before switching to GFP_ATOMIC? (which in case fails the machine is likely to be crashing soon).

The process can run in a cgroup, then kmalloc failure has nothing to do with overall memory consumption. Machine can be perfectly fine. Also, this very process can be chosen as an OOM kill target, then it needs to gracefully deal with kmalloc failure and proceed to a termination point.
Generally retrying something in a loop does not look like a solid plan to deal with errors.

Still dislike the "unregister all bus devices" as that renders the guest unuseable. Can't you fail gracefully? Say force qemu error.