On Thu, Nov 4, 2021 at 11:34 AM Jane Chu [off-list ref] wrote:

Thanks for the enlightening discussion here, it's so helpful!

Please allow me to recap what I've caught up so far -

1. recovery write at page boundary due to NP setting in poisoned
    page to prevent undesirable prefetching
2. single interface to perform 3 tasks:
      { clear-poison, update error-list, write }
    such as an API in pmem driver.
    For CPUs that support MOVEDIR64B, the 'clear-poison' and 'write'
    task can be combined (would need something different from the
    existing _copy_mcsafe though) and 'update error-list' follows
    closely behind;
    For CPUs that rely on firmware call to clear posion, the existing
    pmem_clear_poison() can be used, followed by the 'write' task.
3. if user isn't given RWF_RECOVERY_FLAG flag, then dax recovery
    would be automatic for a write if range is page aligned;
    otherwise, the write fails with EIO as usual.
    Also, user mustn't have punched out the poisoned page in which
    case poison repairing will be a lot more complicated.
4. desirable to fetch as much data as possible from a poisoned range.

If this understanding is in the right direction, then I'd like to
propose below changes to
   dax_direct_access(), dax_copy_to/from_iter(), pmem_copy_to/from_iter()
   and the dm layer copy_to/from_iter, dax_iomap_iter().

1. dax_iomap_iter() rely on dax_direct_access() to decide whether there
    is likely media error: if the API without DAX_F_RECOVERY returns
    -EIO, then switch to recovery-read/write code.  In recovery code,
    supply DAX_F_RECOVERY to dax_direct_access() in order to obtain
    'kaddr', and then call dax_copy_to/from_iter() with DAX_F_RECOVERY.

I like it. It allows for an atomic write+clear implementation on
capable platforms and coordinates with potentially unmapped pages. The
best of both worlds from the dax_clear_poison() proposal and my "take
a fault and do a slow-path copy".

2. the _copy_to/from_iter implementation would be largely the same
    as in my recent patch, but some changes in Christoph's
    'dax-devirtualize' maybe kept, such as DAX_F_VIRTUAL, obviously
    virtual devices don't have the ability to clear poison, so no need
    to complicate them.  And this also means that not every endpoint
    dax device has to provide dax_op.copy_to/from_iter, they may use the
    default.

Did I miss this series or are you talking about this one?
https://lore.kernel.org/all/20211018044054.1779424-1-hch@lst.de/ (local)

I'm not sure about nova and others, if they use different 'write' other
than via iomap, does that mean there will be need for a new set of
dax_op for their read/write?

No, they're out-of-tree they'll adjust to the same interface that xfs
and ext4 are using when/if they go upstream.

the 3-in-1 binding would always be
required though. Maybe that'll be an ongoing discussion?

Yeah, let's cross that bridge when we come to it.

Comments? Suggestions?

It sounds great to me!