L3 entry updates aren't specified to take immediate effect in PAE mode:
On bare metal, only the next CR3 load actually loads the PDPTEs, and a
32-bit Xen also wouldn't immediately propagate new entries into the
PDPTEs. Them taking immediate effect (leaving aside the need to flush
the TLB) on 64-bit Xen is merely to not complicate the hypervisor
implementation more than necessary. Guests cannot depend on such
behavior, and hence this test shouldn't either.

Insert the hypercall equivalent of a CR3 reload into the multicall.

Signed-off-by: Jan Beulich <jbeulich@suse.com>
---
With this, cb199cc7de98 ('Revert "x86/PV32: avoid TLB flushing after
mod_l3_entry()" and "x86/PV: restrict TLB flushing after
mod_l[234]_entry()"') should imo be reverted from the Xen tree. The
claim was wrong that the test was correct and hypervisor code flawed.

--- a/tests/xsa-286/main.c
+++ b/tests/xsa-286/main.c
@@ -128,9 +128,18 @@ void test_main(void)
          *
          * - update_va_mapping(addr, 0, INLVPG)
          * - mmu_update(&l3t[slot], l2t2)
+         * - (PAE only) new_baseptr(cr3)
          * - update_va_mapping(addr, gfn0 | AD|WR|P, INLVPG)
          */
         mu[0].val = pte_from_virt(l2t2, PF_SYM(AD, RW, P));
+#ifdef __i386__
+        mmuext_op_t mx[] = {
+            {
+                .cmd = MMUEXT_NEW_BASEPTR,
+                .arg1.mfn = read_cr3() >> PAGE_SHIFT,
+            },
+        };
+#endif
         intpte_t nl1e = pte_from_gfn(pfn_to_mfn(0), PF_SYM(AD, RW, P));
         multicall_entry_t multi[] = {
             {
@@ -153,6 +162,17 @@ void test_main(void)
                     DOMID_SELF,
                 },
             },
+#ifdef __i386__
+            {
+                .op = __HYPERVISOR_mmuext_op,
+                .args = {
+                    _u(mx),
+                    ARRAY_SIZE(mx),
+                    _u(NULL),
+                    DOMID_SELF,
+                },
+            },
+#endif
             {
                 .op = __HYPERVISOR_update_va_mapping,
                 .args = {

On 17.05.2021 15:22, Jan Beulich wrote:
> L3 entry updates aren't specified to take immediate effect in PAE mode:
> On bare metal, only the next CR3 load actually loads the PDPTEs, and a
> 32-bit Xen also wouldn't immediately propagate new entries into the
> PDPTEs. Them taking immediate effect (leaving aside the need to flush
> the TLB) on 64-bit Xen is merely to not complicate the hypervisor
> implementation more than necessary. Guests cannot depend on such
> behavior, and hence this test shouldn't either.
> 
> Insert the hypercall equivalent of a CR3 reload into the multicall.
> 
> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> ---
> With this, cb199cc7de98 ('Revert "x86/PV32: avoid TLB flushing after
> mod_l3_entry()" and "x86/PV: restrict TLB flushing after
> mod_l[234]_entry()"') should imo be reverted from the Xen tree. The
> claim was wrong that the test was correct and hypervisor code flawed.

Ping?

I know you've expressed your disagreement with things stated above,
but I'm afraid you're still on the hook to actually prove some of
the claims you've made.

Thanks, Jan

> --- a/tests/xsa-286/main.c
> +++ b/tests/xsa-286/main.c
> @@ -128,9 +128,18 @@ void test_main(void)
>           *
>           * - update_va_mapping(addr, 0, INLVPG)
>           * - mmu_update(&l3t[slot], l2t2)
> +         * - (PAE only) new_baseptr(cr3)
>           * - update_va_mapping(addr, gfn0 | AD|WR|P, INLVPG)
>           */
>          mu[0].val = pte_from_virt(l2t2, PF_SYM(AD, RW, P));
> +#ifdef __i386__
> +        mmuext_op_t mx[] = {
> +            {
> +                .cmd = MMUEXT_NEW_BASEPTR,
> +                .arg1.mfn = read_cr3() >> PAGE_SHIFT,
> +            },
> +        };
> +#endif
>          intpte_t nl1e = pte_from_gfn(pfn_to_mfn(0), PF_SYM(AD, RW, P));
>          multicall_entry_t multi[] = {
>              {
> @@ -153,6 +162,17 @@ void test_main(void)
>                      DOMID_SELF,
>                  },
>              },
> +#ifdef __i386__
> +            {
> +                .op = __HYPERVISOR_mmuext_op,
> +                .args = {
> +                    _u(mx),
> +                    ARRAY_SIZE(mx),
> +                    _u(NULL),
> +                    DOMID_SELF,
> +                },
> +            },
> +#endif
>              {
>                  .op = __HYPERVISOR_update_va_mapping,
>                  .args = {
>