Re: memory leak in do_eventfd

From: Rustam Kovhaev
Date: Fri Aug 07 2020 - 03:03:44 EST


On Thu, Jun 04, 2020 at 09:24:02PM -0700, Eric Biggers wrote:
> [+Cc kvm mailing list]
>
> On Wed, May 20, 2020 at 06:12:17PM -0700, syzbot wrote:
> > Hello,
> >
> > syzbot found the following crash on:
> >
> > HEAD commit: 5a9ffb95 Merge tag '5.7-rc5-smb3-fixes' of git://git.samba..
> > git tree: upstream
> > console output: https://syzkaller.appspot.com/x/log.txt?x=10b72a02100000
> > kernel config: https://syzkaller.appspot.com/x/.config?x=f8295ae5b3f8268d
> > dashboard link: https://syzkaller.appspot.com/bug?extid=f196caa45793d6374707
> > compiler: gcc (GCC) 9.0.0 20181231 (experimental)
> > syz repro: https://syzkaller.appspot.com/x/repro.syz?x=17585b76100000
> > C reproducer: https://syzkaller.appspot.com/x/repro.c?x=12500a02100000
> >
> > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> > Reported-by: syzbot+f196caa45793d6374707@xxxxxxxxxxxxxxxxxxxxxxxxx
> >
> > BUG: memory leak
> > unreferenced object 0xffff888117169ac0 (size 64):
> > comm "syz-executor012", pid 6609, jiffies 4294942172 (age 13.720s)
> > hex dump (first 32 bytes):
> > 01 00 00 00 ff ff ff ff 00 00 00 00 00 c9 ff ff ................
> > d0 9a 16 17 81 88 ff ff d0 9a 16 17 81 88 ff ff ................
> > backtrace:
> > [<00000000351bb234>] kmalloc include/linux/slab.h:555 [inline]
> > [<00000000351bb234>] do_eventfd+0x35/0xf0 fs/eventfd.c:418
> > [<00000000c2f69a77>] __do_sys_eventfd fs/eventfd.c:443 [inline]
> > [<00000000c2f69a77>] __se_sys_eventfd fs/eventfd.c:441 [inline]
> > [<00000000c2f69a77>] __x64_sys_eventfd+0x14/0x20 fs/eventfd.c:441
> > [<0000000086d6f989>] do_syscall_64+0x6e/0x220 arch/x86/entry/common.c:295
> > [<000000006c5bcb63>] entry_SYSCALL_64_after_hwframe+0x44/0xa9
> >
> > BUG: memory leak
> > unreferenced object 0xffff888117169100 (size 64):
> > comm "syz-executor012", pid 6609, jiffies 4294942172 (age 13.720s)
> > hex dump (first 32 bytes):
> > e8 99 dd 00 00 c9 ff ff e8 99 dd 00 00 c9 ff ff ................
> > 00 00 00 20 00 00 00 00 00 00 00 00 00 00 00 00 ... ............
> > backtrace:
> > [<00000000436d2955>] kmalloc include/linux/slab.h:555 [inline]
> > [<00000000436d2955>] kzalloc include/linux/slab.h:669 [inline]
> > [<00000000436d2955>] kvm_assign_ioeventfd_idx+0x4f/0x270 arch/x86/kvm/../../../virt/kvm/eventfd.c:798
> > [<00000000e89390cc>] kvm_assign_ioeventfd arch/x86/kvm/../../../virt/kvm/eventfd.c:934 [inline]
> > [<00000000e89390cc>] kvm_ioeventfd+0xbb/0x194 arch/x86/kvm/../../../virt/kvm/eventfd.c:961
> > [<00000000ba9f6732>] kvm_vm_ioctl+0x1e6/0x1030 arch/x86/kvm/../../../virt/kvm/kvm_main.c:3670
> > [<000000005da94937>] vfs_ioctl fs/ioctl.c:47 [inline]
> > [<000000005da94937>] ksys_ioctl+0xa6/0xd0 fs/ioctl.c:771
> > [<00000000a583d097>] __do_sys_ioctl fs/ioctl.c:780 [inline]
> > [<00000000a583d097>] __se_sys_ioctl fs/ioctl.c:778 [inline]
> > [<00000000a583d097>] __x64_sys_ioctl+0x1a/0x20 fs/ioctl.c:778
> > [<0000000086d6f989>] do_syscall_64+0x6e/0x220 arch/x86/entry/common.c:295
> > [<000000006c5bcb63>] entry_SYSCALL_64_after_hwframe+0x44/0xa9
> >

i had to slightly change syzbot reproducer to reproduce this bug in my
lab:
root@syzkaller:~# diff repro_fixed.c repro_syz.c
371c371
< inject_fault(2);
---
> inject_fault(4);

the memleak happens when kmalloc() fails in kvm_io_bus_unregister_dev()
i also confirmed the leak by running ftrace - eventfd_free_ctx() was
not executed when we got to kvm_vm_release()

it seems like we should call kvm_iodevice_destructor(), because there
could be other devices linked to the bus that's being removed when that
particular kmalloc() failure happens

i did not want to modify callers of kvm_io_bus_unregister_dev() and i
tested the patch below that fixed the leak, but i am not sure if it
completely breaks other things in KVM, i would really appreciate if
someone can review it:

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 0a68c9d3d3ab..b1996989d576 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -4268,7 +4268,7 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
struct kvm_io_device *dev)
{
- int i;
+ int i, j;
struct kvm_io_bus *new_bus, *bus;

bus = kvm_get_bus(kvm, bus_idx);
@@ -4298,6 +4298,13 @@ void kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
broken:
rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
synchronize_srcu_expedited(&kvm->srcu);
+ if (!new_bus) {
+ for (j = 0; j < bus->dev_count; j++) {
+ if (j == i)
+ continue;
+ kvm_iodevice_destructor(bus->range[j].dev);
+ }
+ }
kfree(bus);
return;
}