This was triggered by investigation of a deadlock after OOM killer invocation, see [1] for more details. Looks like our handling of fatal signal in do_page_fault() has some issues: 1. We only want to do special (read "early") handling of fatal signal if handle_mm_fault() returned VM_FAULT_RETRY so that we don't loop in retry loop endlessly, otherwise we'll handle that signal normally on exit from exception handler. 2. up_read() is not needed as indeed it will be done in __lock_page_or_retry() in mm/filemap.c. With above comments in mind simplified version should be like that: ------------------------------->8--------------------------- if (fatal_signal_pending(current) if (fault & VM_FAULT_RETRY) if (user_mode(regs)) return; ------------------------------->8--------------------------- But looks like there's a room for improvement, see [2]. Instead of proceeding forward and then inevitably hitting retry path we short-cut right to kernel fix-up code in no_context. [1] http://lists.infradead.org/pipermail/linux-snps-arc/2018-February/003403.html [2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=746a272e44141af24a02f6c9b0f65f4c4598ed42 Signed-off-by: Alexey Brodkin <abrodkin at synopsys.com> --- arch/arc/mm/fault.c | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/arch/arc/mm/fault.c b/arch/arc/mm/fault.c index a0b7bd6d030d..17ed78e2f5eb 100644 --- a/arch/arc/mm/fault.c +++ b/arch/arc/mm/fault.c @@ -139,12 +139,16 @@ void do_page_fault(unsigned long address, struct pt_regs *regs) */ fault = handle_mm_fault(vma, address, flags); - /* If Pagefault was interrupted by SIGKILL, exit page fault "early" */ + /* If we need to retry but a fatal signal is pending, handle the + * signal first. We do not need to release the mmap_sem because + * it would already be released in __lock_page_or_retry in + * mm/filemap.c. */ if (unlikely(fatal_signal_pending(current))) { - if ((fault & VM_FAULT_ERROR) && !(fault & VM_FAULT_RETRY)) - up_read(&mm->mmap_sem); - if (user_mode(regs)) + if (fault & VM_FAULT_RETRY) { + if (!user_mode(regs)) + goto no_context; return; + } } perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address); -- 2.17.1
