Re: [PATCH 0/6] use memcpy_mcsafe() for copy_to_iter()

[Dan Williams <dan.j.williams@intel.com>]

On Tue, May 1, 2018 at 5:09 PM, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
> On Tue, May 1, 2018 at 4:03 PM Dan Williams <dan.j.williams@intel.com>
> wrote:
>
>> I'm confused. Are you talking about getting rid of the block-layer
>> bypass or changing how MCS errors are handled?
>
> The latter.
>
>> If it's the latter, MCS error handling, I don't see how get
>> around something like copy_to_iter_mcsafe().
>
> So the basic issue is that since everybody wants mmap() to be at least an
> option (and preferably one of the _main_ options), I think that the whole
> "MCS errors are fatal" is fundamentally flawed.
>
> Which means that MCS errors can't be fatal.
>
> Which in turn means that the whole "special memcpy" seems very suspect.
>
> Can't we just do
>
>   - use a normal memcpy()
>
>   - basically set an "IO error flag" on MCE.
>
>   - for a user access the IO error flag potentially causes a SIGBUS as you
> mention, but even there it's not 100% clear that's necessarily possible or
> a good idea (I'm assuming that it can be damned hard to figure out _who_
> caused the problem if it was a cached write that causes an MCE much much
> later).

Writes don't trigger MCE. Only consumed poison / media errors trigger
MCE. I.e. even a read-modify-write operation to write-back a partially
dirty cacheline will not trigger an MCE because the read is not
consumed by the core only the cache. We'll get notified when that
happens, but only by CMCI interrupt not an MCE exception.

>   - for the kernel, the "IO error flag" can hopefully be then (again,
> assuming you can correlate the MCE with the right process) be turned into
> EIO.

This is precisely the current implementation / usage of
memcpy_mcsafe(). Reads go through the driver and the driver does the
right / simple thing to turn an MCE into EIO. I'd like to make this
the only model and kill the driver bypass in fs/dax.c so that the vfs
does not need to contend with these low level architecture details.

To be clear I'm not against dax specific optimization that does not go
through the block layer, but it should still be a driver call.

>> You mention mmap. Yes, we want the predominant access model to be
>> dax-mmap for Persistent Memory, but there's still the question about
>> what to do with media errors. To date we are trying to mirror the
>> error handling model for System Memory, i.e. SIGBUS to the process
>> that consumed the error. Is that error handling model also problematic
>> in your view?
>
> See above: if you can handle user space errors "gracefully" (ie with a
> SIGBUS, no crazy "system fatal (reboot)" garbage), then I really don't see
> why you can't do the same for the kernel accesses.
>
> IOW, why do we need that special "copy_to_iter_mcsafe()", when a normal
> "copy_to_iter()" should just work (and basically _has_ to work) anyway?
>
> Put another way: I think the whole basic premis of your patch is wrong,
> because (to quote your original patch descriptor), the fundamental starting
> point is garbage:
>
>     The result of the bypass is that the kernel treats machine checks during
>     read as system fatal (reboot) [..]
>
> See? If you are able to map that memory into user space, and recover, then
> why the whole crazy "system fatal" thing for kernel accesses?

Right, but the only way to make MCE non-fatal is to teach the machine
check handler about recoverable conditions. This patch teaches the
machine check handler how to recover copy_to_iter() errors.

We already have copy_from_iter_flushcache() that is used as a 'struct
dax_operations' op. I can do the same for this copy_to_iter() case so
at least it's up to the driver and not the vfs (fs/dax.c) to decide
how to handle this case.