答复: Found a memory leak in kvm module

[柳菁峰]

I had tried to retest with your patch,and I think the memory leak was resolved by it.
Thanks!

-----邮件原件-----
发件人: Sean Christopherson [mailto:seanjc@xxxxxxxxxx] 
发送时间: 2022年12月16日 2:17
收件人: 柳菁峰 <liujingfeng@xxxxxxxxxxx>
抄送: linux-kernel@xxxxxxxxxxxxxxx; kvm@xxxxxxxxxxxxxxx; pbonzini@xxxxxxxxxx; syzkaller@xxxxxxxxxxxxxxxx
主题: Re: Found a memory leak in kvm module

On Mon, Dec 12, 2022, 柳菁峰 wrote:
> Hello,I have found a memory leak bug in kvm module by syzkaller.It was 
> found in linux-5.4 but it also could be reproduced in the latest linux version.

Ah, I assume by "linux-5.4" you mean "stable v5.4.x kernels that contain commit
7d1bc32d6477 ("KVM: Stop looking for coalesced MMIO zones if the bus is destroyed")", because without that fix I can't see any bug that would affect both 5.4 and the upstream kernel.

If my assumption is correct, then I'm 99% certain the issue is that the target device isn't destroyed if allocating the new bus fails.  I haven't had luck with the automatic fault injection, but was able to confirm a leak with this hack.

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c index 13e88297f999..22d9ab1b5c25 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -5424,7 +5424,7 @@ int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
                              struct kvm_io_device *dev)  {
        int i, j;
-       struct kvm_io_bus *new_bus, *bus;
+       struct kvm_io_bus *new_bus = NULL, *bus;
 
        lockdep_assert_held(&kvm->slots_lock);
 
@@ -5441,6 +5441,7 @@ int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
        if (i == bus->dev_count)
                return 0;
 
+       if (!IS_ENABLED(CONFIG_X86_64))
        new_bus = kmalloc(struct_size(bus, range, bus->dev_count - 1),
                          GFP_KERNEL_ACCOUNT);
        if (new_bus) {


The fix is to destroy the target device before bailing.  I'll send a proper patch either way, but it would be nice to get confirmation that this is the same bug that you hit with "linux-5.4".

Thanks!

---
 virt/kvm/coalesced_mmio.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/virt/kvm/coalesced_mmio.c b/virt/kvm/coalesced_mmio.c index 0be80c213f7f..5ef88f5a0864 100644
--- a/virt/kvm/coalesced_mmio.c
+++ b/virt/kvm/coalesced_mmio.c
@@ -187,15 +187,17 @@ int kvm_vm_ioctl_unregister_coalesced_mmio(struct kvm *kvm,
 			r = kvm_io_bus_unregister_dev(kvm,
 				zone->pio ? KVM_PIO_BUS : KVM_MMIO_BUS, &dev->dev);
 
+			kvm_iodevice_destructor(&dev->dev);
+
 			/*
 			 * On failure, unregister destroys all devices on the
 			 * bus _except_ the target device, i.e. coalesced_zones
-			 * has been modified.  No need to restart the walk as
-			 * there aren't any zones left.
+			 * has been modified.  Bail after destroying the target
+			 * device, there's no need to restart the walk as there
+			 * aren't any zones left.
 			 */
 			if (r)
 				break;
-			kvm_iodevice_destructor(&dev->dev);
 		}
 	}
 

base-commit: 0f30b25edea48433eb32448990557364436818e6
--