Re: [PATCH v3 0/2] Align mmap address for DAX pmd mappings

[Toshi Kani]

On 4/21/2016 3:06 AM, Matthew Wilcox wrote:
> On Wed, Apr 20, 2016 at 11:10:25PM -0400, Toshi Kani wrote:
> > How about moving the function (as is) to mm/huge_memory.c, rename it to
> > get_hugepage_unmapped_area(), which is defined to NULL in huge_mm.h
> > when TRANSPARENT_HUGEPAGE is unset?
> Great idea.  Perhaps it should look something like this?

Yes, it looks good! I will use it. :-)

> unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
>                  unsigned long len, unsigned long pgoff, unsigned long flags)
> {
>          loff_t off, off_end, off_pmd;
>          unsigned long len_pmd, addr_pmd;
>
>          if (addr)
>                  goto out;
>          if (IS_DAX(filp->f_mapping->host) && !IS_ENABLED(CONFIG_FS_DAX_PMD))
>                  goto out;
>          /* Kirill, please fill in the right condition here for THP pagecache */
>
>          off = (loff_t)pgoff << PAGE_SHIFT;
>          off_end = off + len;
>          off_pmd = round_up(off, PMD_SIZE);      /* pmd-aligned start offset */
>
>          if ((off_end <= off_pmd) || ((off_end - off_pmd) < PMD_SIZE))
>                  goto out;
>
>          len_pmd = len + PMD_SIZE;
>          if ((off + len_pmd) < off)
>                  goto out;
>
>          addr_pmd = current->mm->get_unmapped_area(filp, NULL, len_pmd,
>                                                  pgoff, flags);
>          if (!IS_ERR_VALUE(addr_pmd)) {
>                  addr_pmd += (off - addr_pmd) & (PMD_SIZE - 1);
>                  return addr_pmd;
>          }
>   out:
>          return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
> }
>
>   - I deleted the check for filp == NULL.  It can't be NULL ... this is a
>     file_operation ;-)

Right.

>   - Why is len_pmd len + PMD_SIZE instead of round_up(len, PMD_SIZE)?

The length is padded with an extra-PMD size so that any assigned address 
'addr_pmd'
can be aligned by PMD.  IOW, it does not make an assumption that 
addr_pmd is aligned
by the length.

>   - I'm still in two minds about passing 'addr' to the first call to
>     get_unmapped_area() instead of NULL.

When 'addr' is specified, we need to use 'len' since user may be 
managing free VMA
range by itself.  So, I think falling back with the original args is 
correct.

Thanks,
-Toshi