Re: [PATCH 1/1] arm64/hugetlb: clear PG_dcache_clean if the page is dirty when munmap

[Catalin Marinas]

On Mon, Aug 22, 2016 at 12:19:04PM +0800, Leizhen (ThunderTown) wrote:
> On 2016/7/20 17:19, Catalin Marinas wrote:
> > On Wed, Jul 20, 2016 at 10:46:27AM +0800, Leizhen (ThunderTown) wrote:
> >>>>>> On 2016/7/8 21:54, Catalin Marinas wrote:
> >>>>>>> ------------8<----------------
> >>>>>>> diff --git a/arch/arm64/mm/flush.c b/arch/arm64/mm/flush.c
> >>>>>>> index dbd12ea8ce68..c753fa804165 100644
> >>>>>>> --- a/arch/arm64/mm/flush.c
> >>>>>>> +++ b/arch/arm64/mm/flush.c
> >>>>>>> @@ -75,7 +75,8 @@ void __sync_icache_dcache(pte_t pte, unsigned long addr)
> >>>>>>>  	if (!page_mapping(page))
> >>>>>>>  		return;
> >>>>>>>  
> >>>>>>> -	if (!test_and_set_bit(PG_dcache_clean, &page->flags))
> >>>>>>> +	if (!test_and_set_bit(PG_dcache_clean, &page->flags) ||
> >>>>>>> +	    PageDirty(page))
> >>>>>>>  		sync_icache_aliases(page_address(page),
> >>>>>>>  				    PAGE_SIZE << compound_order(page));
> >>>>>>>  	else if (icache_is_aivivt())
> >>>>>>> ----------------8<---------------------
> >>
> >> Do you plan to send this patch? My colleagues told me that if our
> >> patches are quite different, it should be Signed-off-by you.
> > 
> > The reason I'm not sending it is that I don't fully understand how it
> > solves the problem for a shared file mmap(), not just hugetlbfs. As I
> > said in an earlier email: after an msync() in user space we
> > should flush the pages to disk via write_cache_pages(). This function
> Hi Catalin:
>    I'm so sorry for my fault. The previous small pages test result I actually ran on ramfs.
> Today, I ran the case on harddisk fs, it worked well without this patch.
> 
> Summarized as follows:
> small pages on ramfs: need this patch
> small pages on harddisk fs: no need this patch
> hugetlbfs: need this patch

I would add:

small pages over nfs: fails with or without this patch

(tested on Juno, Cortex-A57; seems to be fixed if I remove the
PG_dcache_clean test altogether but, well, we end up over-flushing)

I assume that when using a hard drive, it goes through the block I/O
layer and we may have a flush_dcache_page() called when the kernel is
about to read a page that has been mapped in user space. This would
clear the PG_dcache_clean bit and subsequent __sync_icache_dcache()
would perform cache maintenance.

Could you try on your system the test case without the msync() call? I'm
not sure whether munmap() would trigger an immediate write-back, in
which case we may see the issue even with the filesystem on a hard
drive.

-- 
Catalin