Re: [EXT] Re: [PATCH 1/2] mmc: Support kmsg dumper based on pstore/blk

[Ulf Hansson]

On Tue, 15 Dec 2020 at 21:37, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> On Tue, Dec 15, 2020 at 12:42:58PM +0100, Ulf Hansson wrote:
> > In principle, for non atomic path, I would rather see that the pstore
> > file system should be able to be mounted on top of any generic block
> > device partition - without requiring the block device driver to
> > implement specific pstore ops.
> > [...]
> > Exactly. That's why I wonder if it's really worth it to support the
> > panic writes at all.
>
> pstore/blk already provides the generic hooking -- but it can't do
> the panic write part (which that's very device/driver-specific). The
> design was for individual backing devices to provide that directly
> (which would needed read/write support too). And for those that don't
> have panic/read/write support, they could still use the generic hooks
> but they wouldn't be able to reliably (or at all?) catch panics (just
> console writes, ftrace, pmsg, etc).

I understand the motivation behind pstore's hook for panic-writes.
It's a special thing and perhaps it's easier to support this via a
specific hook, rather than adopting the regular block device request
path to cope with some special I/O request. On the other hand, in the
discussion I have had with Bhaskara, I have pointed out several severe
implications for mmc to support these panic writes (and believe me,
there are even more than those I have brought up). So I am starting to
think that, perhaps there is a better option.

In any case, I didn't catch *why* pstore needs to force block device
drivers to implement specific pstore hooks to support the pstore file
system. I don't think this is the way it should work, for many
reasons. The pstore file system should be able to be extended, to
support the regular block device request path, no?

Kind regards
Uffe