On Sat, Apr 07, 2018 at 12:38:24PM -0700, Dan Williams wrote:
> [ adding Paul and Josh ]
>
> On Wed, Apr 4, 2018 at 2:46 AM, Jan Kara <jack@xxxxxxx> wrote:
> > On Fri 30-03-18 21:03:30, Dan Williams wrote:
> >> Background:
> >>
> >> get_user_pages() in the filesystem pins file backed memory pages for
> >> access by devices performing dma. However, it only pins the memory pages
> >> not the page-to-file offset association. If a file is truncated the
> >> pages are mapped out of the file and dma may continue indefinitely into
> >> a page that is owned by a device driver. This breaks coherency of the
> >> file vs dma, but the assumption is that if userspace wants the
> >> file-space truncated it does not matter what data is inbound from the
> >> device, it is not relevant anymore. The only expectation is that dma can
> >> safely continue while the filesystem reallocates the block(s).
> >>
> >> Problem:
> >>
> >> This expectation that dma can safely continue while the filesystem
> >> changes the block map is broken by dax. With dax the target dma page
> >> *is* the filesystem block. The model of leaving the page pinned for dma,
> >> but truncating the file block out of the file, means that the filesytem
> >> is free to reallocate a block under active dma to another file and now
> >> the expected data-incoherency situation has turned into active
> >> data-corruption.
> >>
> >> Solution:
> >>
> >> Defer all filesystem operations (fallocate(), truncate()) on a dax mode
> >> file while any page/block in the file is under active dma. This solution
> >> assumes that dma is transient. Cases where dma operations are known to
> >> not be transient, like RDMA, have been explicitly disabled via
> >> commits like 5f1d43de5416 "IB/core: disable memory registration of
> >> filesystem-dax vmas".
> >>
> >> The dax_layout_busy_page() routine is called by filesystems with a lock
> >> held against mm faults (i_mmap_lock) to find pinned / busy dax pages.
> >> The process of looking up a busy page invalidates all mappings
> >> to trigger any subsequent get_user_pages() to block on i_mmap_lock.
> >> The filesystem continues to call dax_layout_busy_page() until it finally
> >> returns no more active pages. This approach assumes that the page
> >> pinning is transient, if that assumption is violated the system would
> >> have likely hung from the uncompleted I/O.
> >>
> >> Cc: Jan Kara <jack@xxxxxxx>
> >> Cc: Jeff Moyer <jmoyer@xxxxxxxxxx>
> >> Cc: Dave Chinner <david@xxxxxxxxxxxxx>
> >> Cc: Matthew Wilcox <mawilcox@xxxxxxxxxxxxx>
> >> Cc: Alexander Viro <viro@xxxxxxxxxxxxxxxxxx>
> >> Cc: "Darrick J. Wong" <darrick.wong@xxxxxxxxxx>
> >> Cc: Ross Zwisler <ross.zwisler@xxxxxxxxxxxxxxx>
> >> Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> >> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> >> Reported-by: Christoph Hellwig <hch@xxxxxx>
> >> Reviewed-by: Christoph Hellwig <hch@xxxxxx>
> >> Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> >> ---
> >> drivers/dax/super.c | 2 +
> >> fs/dax.c | 92 +++++++++++++++++++++++++++++++++++++++++++++++++++
> >> include/linux/dax.h | 25 ++++++++++++++
> >> mm/gup.c | 5 +++
> >> 4 files changed, 123 insertions(+), 1 deletion(-)
> >
> > ...
> >
> >> +/**
> >> + * dax_layout_busy_page - find first pinned page in @mapping
> >> + * @mapping: address space to scan for a page with ref count > 1
> >> + *
> >> + * DAX requires ZONE_DEVICE mapped pages. These pages are never
> >> + * 'onlined' to the page allocator so they are considered idle when
> >> + * page->count == 1. A filesystem uses this interface to determine if
> >> + * any page in the mapping is busy, i.e. for DMA, or other
> >> + * get_user_pages() usages.
> >> + *
> >> + * It is expected that the filesystem is holding locks to block the
> >> + * establishment of new mappings in this address_space. I.e. it expects
> >> + * to be able to run unmap_mapping_range() and subsequently not race
> >> + * mapping_mapped() becoming true. It expects that get_user_pages() pte
> >> + * walks are performed under rcu_read_lock().
> >> + */
> >> +struct page *dax_layout_busy_page(struct address_space *mapping)
> >> +{
> >> + pgoff_t indices[PAGEVEC_SIZE];
> >> + struct page *page = NULL;
> >> + struct pagevec pvec;
> >> + pgoff_t index, end;
> >> + unsigned i;
> >> +
> >> + /*
> >> + * In the 'limited' case get_user_pages() for dax is disabled.
> >> + */
> >> + if (IS_ENABLED(CONFIG_FS_DAX_LIMITED))
> >> + return NULL;
> >> +
> >> + if (!dax_mapping(mapping) || !mapping_mapped(mapping))
> >> + return NULL;
> >> +
> >> + pagevec_init(&pvec);
> >> + index = 0;
> >> + end = -1;
> >> + /*
> >> + * Flush dax_layout_lock() sections to ensure all possible page
> >> + * references have been taken, or otherwise arrange for faults
> >> + * to block on the filesystem lock that is taken for
> >> + * establishing new mappings.
> >> + */
> >> + unmap_mapping_range(mapping, 0, 0, 1);
> >> + synchronize_rcu();
> >
> > So I still don't like the use of RCU for this. It just seems as an abuse to
> > use RCU like that. Furthermore it has a hefty latency cost for the truncate
> > path. A trivial test to truncate 100 times the last page of a 16k file that
> > is mmaped (only the first page):
> >
> > DAX+your patches 3.899s
> > non-DAX 0.015s
> >
> > So you can see synchronize_rcu() increased time to run truncate(2) more
> > than 200 times (the process is indeed sitting in __wait_rcu_gp all the
> > time). IMHO that's just too costly.
>
> I wonder if this can be trivially solved by using srcu. I.e. we don't
> need to wait for a global quiescent state, just a
> get_user_pages_fast() quiescent state. ...or is that an abuse of the
> srcu api?
>From what I can see (not that I claim to understand DAX), SRCU
is worth trying. Another thing to try (as a test) is to replace the
synchronize_rcu() above with synchronize_rcu_expedited(), which might
get you an order of magnitude or thereabouts.
Thanx, Paul
