Re: [PATCH v8 00/13] fold per-CPU vmstats remotely

[Marcelo Tosatti <mtosatti@redhat.com>]

Hi Christoph,

On Tue, May 16, 2023 at 10:09:02AM +0200, Christoph Lameter wrote:
> The patchset still modifies the semantics of this_cpu operations semantics
> replacing the lockless RMV operations with locked ones. 

It does that to follow the pre-existing kernel convention:

function-name			LOCK prefix
cmpxchg				YES
cmpxchg_local			NO

So the patchset introduces:

function-name			LOCK prefix
this_cpu_cmpxchg		YES
this_cpu_cmpxchg_local		NO

> One of the
> rationales for the use  this_cpu operations is their efficiency since
> locked RMV atomics are avoided. 

And there is the freedom to choose between this_cpu_cmpxchg and
this_cpu_cmpxchg_local (depending on intended usage).

> This patchset destroys that functionality.

Patch 6 is

Subject: [PATCH v8 06/13] add this_cpu_cmpxchg_local and asm-generic definitions

Which adds this_cpu_cmpxchg_local

Patch 7 converts all other this_cmpxchg users
(except the vmstat ones)

[PATCH v8 07/13] convert this_cpu_cmpxchg users to this_cpu_cmpxchg_local

So the non-LOCK'ed behaviour is maintained for existing users.

> If you want locked RMV semantics then use them through cmpxchg() and
> friends. Do not modify this_cpu operations by changing the implementation
> in the arch code.

But then it would be necessary to disable preemption here:

static inline void mod_zone_state(struct zone *zone, enum zone_stat_item item,
                                  long delta, int overstep_mode)
{
        struct per_cpu_zonestat __percpu *pcp = zone->per_cpu_zonestats;
        s32 __percpu *p = pcp->vm_stat_diff + item;
        long o, n, t, z;

        do {
                z = 0;  /* overflow to zone counters */

                /*
                 * The fetching of the stat_threshold is racy. We may apply
                 * a counter threshold to the wrong the cpu if we get
                 * rescheduled while executing here. However, the next
                 * counter update will apply the threshold again and
                 * therefore bring the counter under the threshold again.
                 *
                 * Most of the time the thresholds are the same anyways
                 * for all cpus in a zone.
                 */
                t = this_cpu_read(pcp->stat_threshold);

                o = this_cpu_read(*p);
                n = delta + o;

                if (abs(n) > t) {
                        int os = overstep_mode * (t >> 1);

                        /* Overflow must be added to zone counters */
                        z = n + os;
                        n = -os;
                }
        } while (this_cpu_cmpxchg(*p, o, n) != o);
		 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

        if (z)
                zone_page_state_add(z, zone, item);
}

Earlier you objected to disabling preemption on this codepath
(which is what led to this patchset in the first place):

"Using preemption is a way to make this work correctly. However, 
doing so would sacrifice the performance, low impact and the
scalability of the vm counters."

So it seems a locked, this_cpu function which does lock cmxpchg
is desired.

Perhaps you disagree with the this_cpu_cmpxchg_local/this_cpu_cmpxchg
naming?