On Thu, 27 Mar 2025 09:58:48 -0700
Eric Grosse <[email protected]> wrote:
> With your patch, system has been running fine with various loads
> applied. I'll update in a couple weeks to confirm this resolved the
> issue. Thanks! (And I recognize that this is only a temporary patch.)
To get back my 500M of ram, I removed my old patch and made this new
patch, which uses an atomic compare-and-swap to check if another cpu
allocated the same thing. My goal is the same, to protect the
kernel_pmap without a vp lock. This patch only affects allocations
of structs pmapvp1 and pmapvp2, not pte_desc.
With this patch, I am trying to fix a problem that never seems to
happen to me. I am less than sure of this fix.
--gkoehler
Index: arch/powerpc64/powerpc64/pmap.c
===================================================================
RCS file: /cvs/src/sys/arch/powerpc64/powerpc64/pmap.c,v
diff -u -p -r1.64 pmap.c
--- arch/powerpc64/powerpc64/pmap.c 28 Nov 2024 18:54:36 -0000 1.64
+++ arch/powerpc64/powerpc64/pmap.c 8 Apr 2025 01:04:58 -0000
@@ -598,6 +598,10 @@ pmap_vp_enter(pmap_t pm, vaddr_t va, str
struct slb_desc *slbd;
struct pmapvp1 *vp1;
struct pmapvp2 *vp2;
+#ifdef MULTIPROCESSOR
+ struct pmapvp1 *old_vp1;
+ struct pmapvp2 *old_vp2;
+#endif
slbd = pmap_slbd_lookup(pm, va);
if (slbd == NULL) {
@@ -617,7 +621,15 @@ pmap_vp_enter(pmap_t pm, vaddr_t va, str
panic("%s: unable to allocate L1", __func__);
return ENOMEM;
}
+#ifdef MULTIPROCESSOR
+ old_vp1 = atomic_cas_ptr(&slbd->slbd_vp, NULL, vp1);
+ if (old_vp1 != NULL) {
+ pool_put(&pmap_vp_pool, vp1);
+ vp1 = old_vp1;
+ }
+#else
slbd->slbd_vp = vp1;
+#endif
}
vp2 = vp1->vp[VP_IDX1(va)];
@@ -628,7 +640,15 @@ pmap_vp_enter(pmap_t pm, vaddr_t va, str
panic("%s: unable to allocate L2", __func__);
return ENOMEM;
}
+#ifdef MULTIPROCESSOR
+ old_vp2 = atomic_cas_ptr(&vp1->vp[VP_IDX1(va)], NULL, vp2);
+ if (old_vp2 != NULL) {
+ pool_put(&pmap_vp_pool, vp2);
+ vp2 = old_vp2;
+ }
+#else
vp1->vp[VP_IDX1(va)] = vp2;
+#endif
}
vp2->vp[VP_IDX2(va)] = pted;
