Re: + x86-mm-handle-mm_fault_error-in-kernel-space.patch added to -mm tree

[Oleg Nesterov <oleg@redhat.com>]

(add cc's)

> Subject: x86/mm: handle mm_fault_error() in kernel space
> From: Andrey Vagin <avagin@openvz.org>
>
> mm_fault_error() should not execute oom-killer, if page fault occurs in
> kernel space.  E.g.  in copy_from_user/copy_to_user.

Why? I don't understand this part.

> This would happen if we find ourselves in OOM on a copy_to_user(), or a
> copy_from_user() which faults.
>
> Without this patch, the kernels hangs up in copy_from_user, because OOM
> killer sends SIG_KILL to current process,

This depends. OOM can choose another victim, and if it does we shouldn't
return -EFAULT.

> but it can't handle a signal
> while in syscall, then the kernel returns to copy_from_user, reexcute
> current command and provokes page_fault again.

Yes. This is buggy.

> --- a/arch/x86/mm/fault.c~x86-mm-handle-mm_fault_error-in-kernel-space
> +++ a/arch/x86/mm/fault.c
> @@ -827,6 +827,13 @@ mm_fault_error(struct pt_regs *regs, uns
>  	       unsigned long address, unsigned int fault)
>  {
>  	if (fault & VM_FAULT_OOM) {
> +		/* Kernel mode? Handle exceptions or die: */
> +		if (!(error_code & PF_USER)) {
> +			up_read(&current->mm->mmap_sem);
> +			no_context(regs, error_code, address);
> +			return;
> +		}
> +

At first glance, this is not optimal...

Perhaps I missed something, but afaics it is better to call
out_of_memory() first, then check if current was killed. In this case
no_context() is fine, we are not going to return to the user-mode.

IOW, what do you think about the (untested/uncompiled) patch below?

Oleg.

--- x/arch/x86/mm/fault.c
+++ x/arch/x86/mm/fault.c
@@ -829,6 +829,11 @@ mm_fault_error(struct pt_regs *regs, uns
 {
 	if (fault & VM_FAULT_OOM) {
 		out_of_memory(regs, error_code, address);
+
+		if (!(error_code & PF_USER) && fatal_signal_pending(current)) {
+			no_context(regs, error_code, address);
+			return;
+		}
 	} else {
 		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
 			     VM_FAULT_HWPOISON_LARGE))