[Qemu-devel] [kqemu patch] get Open/NetBSD to work with the kqemu accelerator

[Qemu-devel] [kqemu patch] get Open/NetBSD to work with the kqemu accelerator

[Enache Adrian]

[sorry if this is the wrong list, but I haven't figured out any public
 address where I could send kqemu bug reports and patches]

Currently, both NetBSD and OpenBSD are hanging or crashing when running
on qemu with the kqemu accelerator enabled.

This happens because both systems are using a weird scheme where they
are loading the GDT table with LGDT up-front (with the limit set to
the maximum), but are growing the table and actually mapping the memory
behind it only when needed.
(see src/sys/arch/i386/i386/gdt.c in both source trees)

That is causing the kqemu accelerator to generate a page fault in
update_dt_cache() when trying to fill its 'soft' tlb with pages that
are beyond the real end of the GDT table.

With this diff applied, NetBSD and OpenBSD seem to run fine with
kqemu + user-only virtualization (I've tried netbsd-4.0-rc2 and
openbsd 4.2).

Full virtualization (-kernel-kqemu) doesn't work yet for different
reasons (I think).

Regards,
Adi

--- xx/kqemu-1.3.0pre11/common/monitor.c	Tue Feb  6 23:02:00 2007
+++ kqemu-1.3.0pre11/common/monitor.c	Mon Nov  5 18:59:58 2007
@@ -990,7 +990,8 @@ static void *map_vaddr(struct kqemu_state *s, unsigned
     e = &s->soft_tlb[(addr >> PAGE_SHIFT) & (SOFT_TLB_SIZE - 1)];
  redo:
     if (e->vaddr[(is_user << 1) + is_write] != (addr & PAGE_MASK)) {
-        soft_tlb_fill(s, addr, is_write, is_user);
+        if(cpu_x86_handle_mmu_fault(s, addr, is_write, is_user, 1))
+                return NULL;
         goto redo;
     } else {
         taddr = e->addend + addr;
@@ -1802,6 +1803,11 @@ static void update_dt_cache(struct kqemu_state *s, int
             page_end = dt_end;
         sel2 = sel + (page_end - dt_ptr);
         ptr = map_vaddr(s, dt_ptr, 0, 0);
+        if(!ptr)
+                /* Open/NetBSD have a 'dynamic' GDT, but they load the gdt
+                   register with LGDT only once and with a limit far beyond
+                   the end of the memory actually mapped for the table */
+                goto skip_the_rest;
         ram_addr = ram_ptr_to_ram_addr(s, ptr);
         if (dt_changed || 
             s->dt_ram_addr[dt_type][pindex] != ram_addr ||
@@ -1818,7 +1824,7 @@ static void update_dt_cache(struct kqemu_state *s, int
     sel_end = (s->dt_limit[dt_type] + 1) & ~7;
     if (sel < sel_end)
         reset_dt_entries(s, dt_type, sel, sel_end);
-
+skip_the_rest:
     s->dt_base[dt_type] = base;
     s->dt_limit[dt_type] = limit;
 }

Re: [Qemu-devel] [kqemu patch] get Open/NetBSD to work with the kqemu accelerator

[Sean Donner]

I just wanted to comment and say this patch works great. I too was
having problems with getting OpenBSD 4.2 working without using the
-no-kqemu option.  Recompiling with this patch I was able to 
confirm that kqemu was working by running 'info kqemu' in the qemu
monitor.  I would really like to see this patch (or a functionally
similar one) incorporated into the next release.

>[sorry if this is the wrong list, but I haven't figured out any public
> address where I could send kqemu bug reports and patches]
>
>Currently, both NetBSD and OpenBSD are hanging or crashing when running
>on qemu with the kqemu accelerator enabled.
>
>This happens because both systems are using a weird scheme where they
>are loading the GDT table with LGDT up-front (with the limit set to
>the maximum), but are growing the table and actually mapping the memory
>behind it only when needed.
>(see src/sys/arch/i386/i386/gdt.c in both source trees)
>
>That is causing the kqemu accelerator to generate a page fault in
>update_dt_cache() when trying to fill its 'soft' tlb with pages that
>are beyond the real end of the GDT table.
>
>With this diff applied, NetBSD and OpenBSD seem to run fine with
>kqemu + user-only virtualization (I've tried netbsd-4.0-rc2 and
>openbsd 4.2).
>
>Full virtualization (-kernel-kqemu) doesn't work yet for different
>reasons (I think).
>
>Regards,
>Adi
>
>--- xx/kqemu-1.3.0pre11/common/monitor.c        Tue Feb  6 23:02:00 2007
>+++ kqemu-1.3.0pre11/common/monitor.c   Mon Nov  5 18:59:58 2007
>@@ -990,7 +990,8 @@ static void *map_vaddr(struct kqemu_state *s, unsigned
>     e = &s->soft_tlb[(addr >> PAGE_SHIFT) & (SOFT_TLB_SIZE - 1)];
>  redo:
>     if (e->vaddr[(is_user << 1) + is_write] != (addr & PAGE_MASK)) {
>-        soft_tlb_fill(s, addr, is_write, is_user);
>+        if(cpu_x86_handle_mmu_fault(s, addr, is_write, is_user, 1))
>+                return NULL;
>         goto redo;
>     } else {
>         taddr = e->addend + addr;
>@@ -1802,6 +1803,11 @@ static void update_dt_cache(struct kqemu_state *s, int
>             page_end = dt_end;
>         sel2 = sel + (page_end - dt_ptr);
>         ptr = map_vaddr(s, dt_ptr, 0, 0);
>+        if(!ptr)
>+                /* Open/NetBSD have a 'dynamic' GDT, but they load the gdt
>+                   register with LGDT only once and with a limit far beyond
>+                   the end of the memory actually mapped for the table */
>+                goto skip_the_rest;
>         ram_addr = ram_ptr_to_ram_addr(s, ptr);
>         if (dt_changed || 
>             s->dt_ram_addr[dt_type][pindex] != ram_addr ||
>@@ -1818,7 +1824,7 @@ static void update_dt_cache(struct kqemu_state *s, int
>     sel_end = (s->dt_limit[dt_type] + 1) & ~7;
>     if (sel < sel_end)
>         reset_dt_entries(s, dt_type, sel, sel_end);
>-
>+skip_the_rest:
>     s->dt_base[dt_type] = base;
>     s->dt_limit[dt_type] = limit;
> }