mm/damon/core: make charge_addr_from aware of end-address exclusivity

Sashiko AI review found 3 potential issue(s):
- [High] The remainder of the last region of a target is unconditionally skipped during quota resumption.
- [High] Quota state is not saved when the remaining quota is less than min_region_sz, breaking round-robin fairness.
- [Medium] Setting charge_addr_from = r->ar.end causes resumption failure when a region covers the end of the address space.

--

commit 50bbd8d578080e27beac4fe1ad0cecbb013a8fd7
Author: SeongJae Park <[email protected]>

mm/damon/core: make charge_addr_from aware of end-address exclusivity

This commit updates the DAMOS quota charge assignment to use the exclusive region end address rather than assuming it is inclusive. This addresses a bug where the action might skip the next memory block up to min_region_sz.

Link https://lore.kernel.org/[email protected] open here

mm/damon/core.c static void damos_apply_scheme(struct damon_ctx *c, struct damon_target *t, · line 2109

Quoted 10 lines

mm/damon/core.c +1 -1

index 3dbbbfdeff71..901ffdaefb7f 100644
@@ -2106,7 +2106,7 @@ static void damos_apply_scheme(struct damon_ctx *c, struct damon_target *t,
 		if (damos_quota_is_set(quota) &&
 				quota->charged_sz >= quota->esz) {
 			quota->charge_target_from = t;
-			quota->charge_addr_from = r->ar.end + 1;
+			quota->charge_addr_from = r->ar.end;

Does this cause a resumption failure if the region covers the absolute end of the address space?

If r->ar.end wraps to 0 (e.g., ULONG_MAX + 1), charge_addr_from is assigned
0. Since DAMON uses 0 as a sentinel value for an uninitialized state, the checks for quota->charge_addr_from in damos_skip_charged_region() will evaluate to false.

Will this completely fail to identify the resumption point, prematurely clear the state, and restart from the first region instead of resuming?

Additionally, this isn't a bug introduced by this patch, but does the resuming logic unconditionally skip the remainder of the last region of a target?

In damos_skip_charged_region(), the damon_last_region() check executes before evaluating charge_addr_from:

mm/damon/core.c:damos_skip_charged_region() {
    ...
	if (quota->charge_target_from) {
		if (t != quota->charge_target_from)
			return true;
		if (r == damon_last_region(t)) {
			quota->charge_target_from = NULL;
			quota->charge_addr_from = 0;
			return true;
		}
        ...
}

If a target's last region is partially charged and split during a window, the uncharged remainder becomes the new damon_last_region. In the next quota window, when the loop reaches this remainder, will it immediately match the last region check, clear the state, and permanently starve the end of targets by returning true (skip)?

Finally, this isn't a bug introduced by this patch, but is the quota state lost when the remaining quota is less than min_region_sz?

Earlier in damos_apply_scheme(), if the remaining quota is small:

mm/damon/core.c:damos_apply_scheme() {
    ...
		if (damos_quota_is_set(quota) &&
				quota->charged_sz + sz > quota->esz) {
			sz = ALIGN_DOWN(quota->esz - quota->charged_sz,
					c->min_region_sz);
			if (!sz)
				goto update_stat;
    ...
}

Does the goto update_stat bypass incrementing quota->charged_sz and saving charge_target_from and charge_addr_from?

Because charged_sz never reaches esz, the loop will uselessly evaluate all remaining regions. Because the resume state is never saved, will the next interval incorrectly restart from the very first target and region, breaking round-robin fairness and starving later regions whenever the dynamically calculated quota is not perfectly aligned?