Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock

From: Kirill A. Shutemov
Date: Tue Aug 11 2015 - 16:26:51 EST


On Tue, Aug 11, 2015 at 07:17:12PM +0300, Boaz Harrosh wrote:
> On 08/11/2015 06:28 PM, Kirill A. Shutemov wrote:
> > We also used lock_page() to make sure we shoot out all pages as we don't
> > exclude page faults during truncate. Consider this race:
> >
> > <fault> <truncate>
> > get_block
> > check i_size
> > update i_size
> > unmap
> > setup pte
> >
>
> Please consider this senario then:
>
> <fault> <truncate>
> read_lock(inode)
>
> get_block
> check i_size
>
> read_unlock(inode)
>
> write_lock(inode)
>
> update i_size
> * remove allocated blocks
> unmap
>
> write_unlock(inode)
>
> setup pte
>
> IS what you suppose to do in xfs

Do you realize that you describe a race? :-P

Exactly in this scenario pfn your pte point to is not belong to the file
anymore. Have fun.

> > With normal page cache we make sure that all pages beyond i_size is
> > dropped using lock_page() in truncate_inode_pages_range().
> >
>
> Yes there is no truncate_inode_pages_range() in DAX again radix tree is
> empty.
>
> Please do you have a reproducer I would like to see this race and also
> experiment with xfs (I guess you saw it in ext4)

I don't. And I don't see how race like above can be FS-specific. All
critical points in generic code.

> > For DAX we need a way to stop all page faults to the pgoff range before
> > doing unmap.
> >
>
> Why ?

Because you can end up with ptes pointing to pfns which fs consider not be
part of the file.

<truncate> <fault>
unmap..
fault in pfn which unmap already unmapped
..continue unmap

> >> Because with DAX there is no inode->mapping "mapping" at all. You have the call
> >> into the FS with get_block() to replace "holes" (zero pages) with real allocated
> >> blocks, on WRITE faults, but this conversion should be protected inside the FS
> >> already. Then there is the atomic exchange of the PTE which is fine.
> >> (And vis versa with holes mapping and writes)
> >
> > Having unmap_mapping_range() in PMD fault handling is very unfortunate.
> > Go to rmap just to solve page fault is very wrong.
> > BTW, we need to do it in write path too.
> >
>
> Only the write path and only when we exchange a zero-page (hole) with
> a new allocated (written to) page. Both write fault and/or write-path

No. Always on new BH. We don't have anything (except rmap) to find out if
any other process have zero page for the pgoff.

> > I'm not convinced that all these "let's avoid backing storage allocation"
> > in DAX code is not layering violation. I think the right place to solve
> > this is filesystem. And we have almost all required handles for this in
> > place. We only need to change vm_ops->page_mkwrite() interface to be able
> > to return different page than what was given on input.
> >
>
> What? there is no page returned for DAX page_mkwrite(), it is all insert_mixed
> with direct pmd.

That was bogus idea, please ignore.

> > Hm. Where does XFS take this read-write lock in fault path?
> >
> > IIUC, truncation vs. page fault serialization relies on i_size being
> > updated before doing truncate_pagecache() and checking i_size under
> > page_lock() on fault side. We don't have i_size fence for punch hole.
> >
>
> again truncate_pagecache() is NONE.
> And yes the read-write locking will protect punch-hole just as truncate
> see my locking senario above.

Do you mean as racy as your truncate scenario? ;-P

--
Kirill A. Shutemov
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/