Re: [MINI-OS PATCH v2 2/2] x86: don't use a memory page for mapping the shared info page

[Jürgen Groß]

On 29.07.25 11:00, Jan Beulich wrote:

On 29.07.2025 10:38, Juergen Gross wrote:

--- a/arch/x86/x86_64.S
+++ b/arch/x86/x86_64.S
@@ -33,13 +33,8 @@ _start:
  stack_start:
          .quad stack+(2*__STACK_SIZE)

-        /* Unpleasant -- the PTE that maps this page is actually overwritten */
-        /* to map the real shared-info page! :-)                             */
          .align __PAGE_SIZE
-shared_info:
-        .fill __PAGE_SIZE,1,0
-
+.globl hypercall_page

-.globl shared_info, hypercall_page

While touching this line, may I suggest to indent this directive to match all
other directives in context? Even if assemblers accept them for most targets,
directives starting in the first column strictly speaking are misplaced.

--- a/hypervisor.c
+++ b/hypervisor.c
@@ -27,8 +27,10 @@

  #include <mini-os/lib.h>
+#include <mini-os/e820.h>
  #include <mini-os/hypervisor.h>
  #include <mini-os/events.h>
+#include <mini-os/mm.h>
  #include <xen/memory.h>

@@ -37,7 +39,8 @@ EXPORT_SYMBOL(hypercall_page);
      ((sh)->evtchn_pending[idx] & ~(sh)->evtchn_mask[idx])

-extern shared_info_t shared_info;
+static unsigned long shinfo_pfn;
+static unsigned long shinfo_va;

  {
@@ -69,24 +72,31 @@ shared_info_t *map_shared_info(void)
  {
      struct xen_add_to_physmap xatp;

      xatp.domid = DOMID_SELF;
      xatp.idx = 0;
      xatp.space = XENMAPSPACE_shared_info;
-    xatp.gpfn = virt_to_pfn(&shared_info);
+    xatp.gpfn = shinfo_pfn;
      if ( HYPERVISOR_memory_op(XENMEM_add_to_physmap, &xatp) != 0 )
          BUG();
+    if ( !shinfo_va )
+        shinfo_va = alloc_virt_kernel(1);
+    if ( !shinfo_va || map_frame_rw(shinfo_va, shinfo_pfn) )
+        BUG();

#include <mini-os/os.h> EXPORT_SYMBOL(hypercall_page); #ifndef CONFIG_PARAVIRT int hvm_get_parameter(int idx, uint64_t *value) + shinfo_pfn = e820_get_reserved_pfns(1);

Now there's a new asymmetry: Here you check whether alloc_virt_kernel()
(appears to have) failed, whereas in the PV variant you don't. And it's
really only "appears to", as the function won't return 0 in the failure
case, afaics. I therefore think that extra condition simply wants
dropping here. Then

Oh, right. First I had

if ( !shinfo_va )
    shinfo_va = map_frame_virt(shinfo_pfn);
else
    ret = map_frame_rw(shinfo_va, shinfo_pfn);
if ( ret || !shinfo_va )
    BUG();

which I then simplified to above sequence, but without dropping the test for
shinfo_va being not 0.

Reviewed-by: Jan Beulich <jbeul...@suse.com>

As for the other patch, happy to make both adjustments while committing.
As long as you agree, of course.

Yes, I agree, thanks.


Juergen

OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

OpenPGP_signature.asc
Description: OpenPGP digital signature