Re: [PATCH 05/17] asm-generic/tlb: Rename HAVE_RCU_TABLE_NO_INVALIDATE

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Dec 16, 2019 at 07:10:30PM +0530, Aneesh Kumar K.V wrote:
> On 12/16/19 6:50 PM, Peter Zijlstra wrote:
> > On Mon, Dec 16, 2019 at 06:43:53PM +0530, Aneesh Kumar K.V wrote:
> > > On 12/16/19 6:07 PM, Peter Zijlstra wrote:
> > 
> > > > I'm confused, are you saing you're happy to have PowerPC eat the extra
> > > > TLB invalidates? I thought you cared about PPC performance :-)
> > > > 
> > > > 
> > > 
> > > Instead can we do
> > > 
> > > static inline void tlb_table_invalidate(struct mmu_gather *tlb)
> > > {
> > > #ifndef CONFIG_MMU_GATHER_RCU_TABLE_FREE
> > > 	 * Invalidate page-table caches used by hardware walkers. Then we still
> > > 	 * need to RCU-sched wait while freeing the pages because software
> > > 	 * walkers can still be in-flight.
> > > 	 */
> > > 	tlb_flush_mmu_tlbonly(tlb);
> > > #endif
> > > }
> > 
> > How does that not break ARM/ARM64/s390 and x86 ?
> > 
> 
> Hmm I missed that usage of RCU_TABLE_NO_INVALIDATE.

What use? Only PPC and SPARC64 use that option. The reason they can use
it is because they don't have a hardware walker (with exception of
PPC-Radix, which I suppose your below patch fudges ?!).

So HAVE_RCU_TABLE_FREE will provide tlb_remove_table() for the use of
freeing page-tables/directories. This is required for all architectures
that have software walkers and !IPI TLBI.

Arm, Arm64, Power, Sparc64, s390 and x86 use this option. While x86
natively has IPI based TLBI, a bunch of the virtualization solutions got
rid of the IPI for performance.

Of those, Arm, Arm64, s390, x86 (and PPC-Radix) also have hardware
walkers on those page-tables, and thus _must_ TLBI in between unhooking
and freeing these pages.

PPC-Hash and Sparc64 OTOH only ever access the linux page-tables through
the software walker and thus can forgo this TLBI, and _THAT_ is what
TABLE_NO_INVALIDATE is about (there actually is a comment that clearly
states this).

> Ok I guess we need to revert this change that went upstream this merge
> window then
> 
> commit 52162ec784fa05f3a4b1d8e84421279998be3773
> Author: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
> Date:   Thu Oct 24 13:28:00 2019 +0530
> 
>     powerpc/mm/book3s64/radix: Use freed_tables instead of need_flush_all
> 

This really looks like you've got PPC-Radix wrong. As soon as you got
hardware walkers on the linux page-tables, you must not use
TABLE_NO_INVALIDATE.