On Tue, Nov 02, 2021 at 09:03:55AM -0700, Dan Williams wrote:

quoted

why devices are built to handle them.  It is just the Intel-style
pmem interface to handle them which is completely broken.

No, any media can report checksum / parity errors. NVME also seems to
do a poor job with multi-bit ECC errors consumed from DRAM. There is
nothing "pmem" or "Intel" specific here.

If you do get data corruption from NVMe (which yes can happen despite
the typical very good UBER rate) you just write over it again.  You
don't need to magically whack the underlying device.  Same for hard
drives.

quoted

Well, my point is doing recovery from bit errors is by definition not
the fast path.  Which is why I'd rather keep it away from the pmem
read/write fast path, which also happens to be the (much more important)
non-pmem read/write path.

I would expect this interface to be useful outside of pmem as a
"failfast" or "try harder to recover" flag for reading over media
errors.

Maybe we need to sit down and define useful semantics then?  The problem
on the write side isn't really that the behavior with the flag is
undefined, it is more that writes without the flag have horrible
semantics if they don't just clear the error.