fixed aggressive arc_p increase

Fixes #14120

The original ARC paper calls for an initial 50/50 split between
MRU and MFU caches, and this is included in current code via
some arc_p = arc_c >> 1 assignement. However, both arc_adapt()
and arc_get_data_impl() are too aggressive in letting
arc_p == arc_c, with the result that MFU cache can be totally
evicted by MRU cache until the ghost list adapt code kick in.

This patch restores the original arc_p = arc_c >> 1 intent,
without preventing MRU to be 100% of actual ARC if MFU is empty.

Signed-off-by: Gionatan Danti <[email protected]>

Author: shodanshok

module/zfs/arc.c

@@ -5173,7 +5173,7 @@ arc_adapt(int bytes, arc_state_t *state)
 		atomic_add_64(&arc_c, (int64_t)bytes);
 		if (arc_c > arc_c_max)
 			arc_c = arc_c_max;
-		else if (state == arc_anon)
+		else if (state == arc_anon && arc_p < arc_c >> 1)
 			atomic_add_64(&arc_p, (int64_t)bytes);
 		if (arc_p > arc_c)
 			arc_p = arc_c;
@@ -5386,7 +5386,8 @@ arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
 		if (aggsum_upper_bound(&arc_sums.arcstat_size) < arc_c &&
 		    hdr->b_l1hdr.b_state == arc_anon &&
 		    (zfs_refcount_count(&arc_anon->arcs_size) +
-		    zfs_refcount_count(&arc_mru->arcs_size) > arc_p))
+		    zfs_refcount_count(&arc_mru->arcs_size) > arc_p &&
+		    arc_p < arc_c >> 1))
 			arc_p = MIN(arc_c, arc_p + size);
 	}
 }