Ira,

On Fri, Oct 09 2020 at 12:49, ira weiny wrote:

From: Ira Weiny <ira.weiny@intel.com>

To correctly support the semantics of kmap() with Kernel protection keys
(PKS), kmap() may be required to set the protections on multiple
processors (globally).  Enabling PKS globally can be very expensive
depending on the requested operation.  Furthermore, enabling a domain
globally reduces the protection afforded by PKS.

Most kmap() (Aprox 209 of 229) callers use the map within a single thread and
have no need for the protection domain to be enabled globally.  However, the
remaining callers do not follow this pattern and, as best I can tell, expect
the mapping to be 'global' and available to any thread who may access the
mapping.[1]

We don't anticipate global mappings to pmem, however in general there is a
danger in changing the semantics of kmap().  Effectively, this would cause an
unresolved page fault with little to no information about why the failure
occurred.

To resolve this a number of options were considered.

1) Attempt to change all the thread local kmap() calls to kmap_atomic()[2]
2) Introduce a flags parameter to kmap() to indicate if the mapping should be
   global or not
3) Change ~20 call sites to 'kmap_global()' to indicate that they require a
   global enablement of the pages.
4) Change ~209 call sites to 'kmap_thread()' to indicate that the mapping is to
   be used within that thread of execution only

Option 1 is simply not feasible.  Option 2 would require all of the call sites
of kmap() to change.  Option 3 seems like a good minimal change but there is a
danger that new code may miss the semantic change of kmap() and not get the
behavior the developer intended.  Therefore, #4 was chosen.

There is Option #5:

Convert the thread local kmap() invocations to the proposed kmap_local()
interface which is coming along [1].

That solves a couple of issues:

 1) It relieves the current kmap_atomic() usage sites from the implict
    pagefault/preempt disable semantics which apply even when
    CONFIG_HIGHMEM is disabled. kmap_local() still can be invoked from
    atomic context.

 2) Due to #1 it allows to replace the conditional usage of kmap() and
    kmap_atomic() for purely thread local mappings.

 3) It puts the burden on the HIGHMEM inflicted systems

 4) It is actually more efficient for most of the pure thread local use
    cases on HIGHMEM inflicted systems because it avoids the overhead of
    the global lock and the potential kmap slot exhaustion. A potential
    preemption will be more expensive, but that's not really the case we
    want to optimize for.

 5) It solves the RT issue vs. kmap_atomic()

So instead of creating yet another variety of kmap() which is just
scratching the particular PKRS itch, can we please consolidate all of
that on the wider reaching kmap_local() approach?

Thanks,

        tglx
     
[1] https://lore.kernel.org/lkml/20201103092712.714480842@linutronix.de/ (local)