On Fri, Jul 25, 2014 at 03:44:50PM -0400, Matthew Wilcox wrote:
> On Wed, Jul 23, 2014 at 06:55:00PM +0300, Kirill A. Shutemov wrote:
> > > update_hiwater_rss(mm);
> >
> > No: you cannot end up with lower rss after replace, iiuc.
>
> Actually, you can ... when we replace a real page with a PFN, our rss
> decreases.
Okay.
> > Do you mean you pointed to new file all the time? O_CREAT doesn't truncate
> > file if it exists, iirc.
>
> It was pointing to a new file. Still not sure why that one failed to trigger
> the problem. The slightly modified version attached triggered the problem
> *just fine* :-)
>
> I've attached all the patches in my tree so far. For the v9 patch kit,
> I'll keep patch 3 as a separate patch, but roll patches 1, 2 and 4 into
> other patches.
>
> I am seeing something odd though. When I run double-map with debugging
> printks inserted in strategic spots in the kernel, I see four calls to
> do_dax_fault(). The first two, as expected, are the loads from the two
> mapped addresses. The third is via mkwrite, but then the fourth time
> I get a regular page fault for write, and I don't understand why I get it.
>
> Any ideas?
unmap_mapping_range() clears pte you've just set by vm_replace_mixed() on
third fault.
And locking looks wrong: it seems you need to hold i_mmap_mutex while
replacing hole page with pfn. Your VM_BUG_ON() in zap_pte_single()
triggers on my setup.
> +static void zap_pte_single(struct vm_area_struct *vma, pte_t *pte,
> + unsigned long addr)
> +{
> + struct mm_struct *mm = vma->vm_mm;
> + int force_flush = 0;
> + int rss[NR_MM_COUNTERS];
> +
> + VM_BUG_ON(!mutex_is_locked(&vma->vm_file->f_mapping->i_mmap_mutex));
It's wrong place for VM_BUG_ON(): zap_pte_single() on anon mapping should
work fine)
> +
> + init_rss_vec(rss);
Vector to commit single update to mm counters? What about inline counters
update for rss == NULL case?
> + update_hiwater_rss(mm);
> + flush_cache_page(vma, addr, pte_pfn(*pte));
> + zap_pte(NULL, vma, pte, addr, NULL, rss, &force_flush);
> + flush_tlb_page(vma, addr);
> + add_mm_rss_vec(mm, rss);
> +}
> +
--
Kirill A. Shutemov
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
