Re: [PATCH v7 15/17] dax: add struct iomap based DAX PMD support

[Aneesh Kumar K.V]

Ross Zwisler <ross.zwisler@xxxxxxxxxxxxxxx> writes:

> DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based
> locking.  This patch allows DAX PMDs to participate in the DAX radix tree
> based locking scheme so that they can be re-enabled using the new struct
> iomap based fault handlers.
>
> There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX
> mappings that have an associated block allocation, and 4k DAX empty
> entries.  The empty entries exist to provide locking for the duration of a
> given page fault.
>
> This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP)
> entries, PMD DAX entries that have associated block allocations, and 2 MiB
> DAX empty entries.
>
> Unlike the 4k case where we insert a struct page* into the radix tree for
> 4k zero pages, for HZP we insert a DAX exceptional entry with the new
> RADIX_DAX_HZP flag set.  This is because we use a single 2 MiB zero page in
> every 2MiB hole mapping, and it doesn't make sense to have that same struct
> page* with multiple entries in multiple trees.  This would cause contention
> on the single page lock for the one Huge Zero Page, and it would break the
> page->index and page->mapping associations that are assumed to be valid in
> many other places in the kernel.
>
> One difficult use case is when one thread is trying to use 4k entries in
> radix tree for a given offset, and another thread is using 2 MiB entries
> for that same offset.  The current code handles this by making the 2 MiB
> user fall back to 4k entries for most cases.  This was done because it is
> the simplest solution, and because the use of 2MiB pages is already
> opportunistic.
>
> If we were to try to upgrade from 4k pages to 2MiB pages for a given range,
> we run into the problem of how we lock out 4k page faults for the entire
> 2MiB range while we clean out the radix tree so we can insert the 2MiB
> entry.  We can solve this problem if we need to, but I think that the cases
> where both 2MiB entries and 4K entries are being used for the same range
> will be rare enough and the gain small enough that it probably won't be
> worth the complexity.
>
> Signed-off-by: Ross Zwisler <ross.zwisler@xxxxxxxxxxxxxxx>
> Reviewed-by: Jan Kara <jack@xxxxxxx>
> ---
>  fs/dax.c            | 378 +++++++++++++++++++++++++++++++++++++++++++++++-----
>  include/linux/dax.h |  55 ++++++--
>  mm/filemap.c        |   3 +-
>  3 files changed, 386 insertions(+), 50 deletions(-)
>
> diff --git a/fs/dax.c b/fs/dax.c
> index 0582c7c..153cfd5 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -76,6 +76,26 @@ static void dax_unmap_atomic(struct block_device *bdev,
>  	blk_queue_exit(bdev->bd_queue);
>  }
>  
> +static int dax_is_pmd_entry(void *entry)
> +{
> +	return (unsigned long)entry & RADIX_DAX_PMD;
> +}
> +
> +static int dax_is_pte_entry(void *entry)
> +{
> +	return !((unsigned long)entry & RADIX_DAX_PMD);
> +}
> +
> +static int dax_is_zero_entry(void *entry)
> +{
> +	return (unsigned long)entry & RADIX_DAX_HZP;
> +}

How about dax_is_pmd_zero_entry() ?


> +
> +static int dax_is_empty_entry(void *entry)
> +{
> +	return (unsigned long)entry & RADIX_DAX_EMPTY;
> +}
> +

-aneesh