On Mon, Oct 12, 2020 at 05:44:38PM +0100, Matthew Wilcox wrote:

On Mon, Oct 12, 2020 at 09:28:29AM -0700, Dave Hansen wrote:

quoted

kmap_atomic() is always preferred over kmap()/kmap_thread().
kmap_atomic() is _much_ more lightweight since its TLB invalidation is
always CPU-local and never broadcast.

So, basically, unless you *must* sleep while the mapping is in place,
kmap_atomic() is preferred.

But kmap_atomic() disables preemption, so the _ideal_ interface would map
it only locally, then on preemption make it global.  I don't even know
if that _can_ be done.  But this email makes it seem like kmap_atomic()
has no downsides.

And that is IIUC what Thomas was trying to solve.

Also, Linus brought up that kmap_atomic() has quirks in nesting.[1]

From what I can see all of these discussions support the need to have something
between kmap() and kmap_atomic().

However, the reason behind converting call sites to kmap_thread() are different
between Thomas' patch set and mine.  Both require more kmap granularity.
However, they do so with different reasons and underlying implementations but
with the _same_ resulting semantics; a thread local mapping which is
preemptable.[2]  Therefore they each focus on changing different call sites.

While this patch set is huge I think it serves a valuable purpose to identify a
large number of call sites which are candidates for this new semantic.

Ira

[1] https://lore.kernel.org/lkml/CAHk-=wgbmwsTOKs23Z=71EBTrULoeaH2U3TNqT2atHEWvkBKdw@mail.gmail.com/ (local)
[2] It is important to note these implementations are not incompatible with
each other.  So I don't see yet another 'kmap_something()' being required.