On 4/7/20 3:59 AM, zhe.he@xxxxxxxxxxxxx wrote:
> From: He Zhe <zhe.he@xxxxxxxxxxxxx>
>
> commit b5e683d5cab8 ("eventfd: track eventfd_signal() recursion depth")
> introduces a percpu counter that tracks the percpu recursion depth and
> warn if it greater than one, to avoid potential deadlock and stack
> overflow.
>
> However sometimes different eventfds may be used in parallel.
> Specifically, when high network load goes through kvm and vhost, working
> as below, it would trigger the following call trace.
>
> - 100.00%
> - 66.51%
> ret_from_fork
> kthread
> - vhost_worker
> - 33.47% handle_tx_kick
> handle_tx
> handle_tx_copy
> vhost_tx_batch.isra.0
> vhost_add_used_and_signal_n
> eventfd_signal
> - 33.05% handle_rx_net
> handle_rx
> vhost_add_used_and_signal_n
> eventfd_signal
> - 33.49%
> ioctl
> entry_SYSCALL_64_after_hwframe
> do_syscall_64
> __x64_sys_ioctl
> ksys_ioctl
> do_vfs_ioctl
> kvm_vcpu_ioctl
> kvm_arch_vcpu_ioctl_run
> vmx_handle_exit
> handle_ept_misconfig
> kvm_io_bus_write
> __kvm_io_bus_write
> eventfd_signal
>
> 001: WARNING: CPU: 1 PID: 1503 at fs/eventfd.c:73 eventfd_signal+0x85/0xa0
> ---- snip ----
> 001: Call Trace:
> 001: vhost_signal+0x15e/0x1b0 [vhost]
> 001: vhost_add_used_and_signal_n+0x2b/0x40 [vhost]
> 001: handle_rx+0xb9/0x900 [vhost_net]
> 001: handle_rx_net+0x15/0x20 [vhost_net]
> 001: vhost_worker+0xbe/0x120 [vhost]
> 001: kthread+0x106/0x140
> 001: ? log_used.part.0+0x20/0x20 [vhost]
> 001: ? kthread_park+0x90/0x90
> 001: ret_from_fork+0x35/0x40
> 001: ---[ end trace 0000000000000003 ]---
>
> This patch moves the percpu counter into eventfd control structure and
> does the clean-ups, so that eventfd can still be protected from deadlock
> while allowing different ones to work in parallel.
>
> As to potential stack overflow, we might want to figure out a better
> solution in the future to warn when the stack is about to overflow so it
> can be better utilized, rather than break the working flow when just the
> second one comes.
This doesn't work for the infinite recursion case, the state has to be
global, or per thread.
--
Jens Axboe
