Messages in this thread | | | From | "Alexander van Heukelum" <> | Subject | Re: [PATCH] reserve end-of-conventional-memory to 1MB on 32-bit | Date | Fri, 29 Feb 2008 19:38:43 +0100 |
| |
On Fri, 29 Feb 2008 17:14:07 +0000, "Mark McLoughlin" <markmc@redhat.com> said: > On Fri, 2008-02-29 at 12:49 +0100, Alexander van Heukelum wrote: > > This patch adds explicit detection of the EBDA and reservation > > of the rom and adapter address space 0xa0000-0x100000 to the > > i386 kernels. It uses reserve_bootmem instead of reserve_early, > > because reserve_early is not yet available on i386. > > > > Before this patch, the EBDA size was hardcoded as 4Kb. Also, the > > reservation of the adapter range was done by modifying the e820 > > map which is now not necessary any longer, and the code is removed > > from copy_e820_map. > > > > The changes in e820_64.c are only a change in the comment above > > copy_e820_map, and some changes of the types of local variables > > in that function such that the 32 and 64 bit versions become equal. > > > > Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm> > > > > --- > > > > On Thu, Feb 28, 2008 at 09:09:56PM +0000, Ian Campbell wrote: > > > On Thu, 2008-02-28 at 14:13 +0100, Alexander van Heukelum wrote: > > > > The 32-bit code still uses reserve_bootmem, so this is not really > > > > a unification with the 64-bit version of the ebda reservation code, > > > > but at least it provides the same detection logic and reserves the > > > > same areas. > > > > > > > > This does not crash immediately on qemu. No further testing was > > > > done! Otherwise: > > > > > > I haven't tested extensively either but it does seem to solve the > > > problem for Xen. > > > > > > Thanks! > > > Ian > > > > Thank you! > > > > Ingo, > > > > I think this is ready for -x86#testing. > > It boots to a small userspace in qemu (i386). > > If I should separate the cleanups, let me know. > > I haven't investigated in any detail, but with 2.6.25-rc3 and your > patch I'm seeing a Xen guest hit this BUG: > > void __init smp_alloc_memory(void) > { > trampoline_base = alloc_bootmem_low_pages(PAGE_SIZE); > /* > * Has to be in very low memory so we can execute > * real-mode AP code. > */ > if (__pa(trampoline_base) >= 0x9F000) > BUG(); > } > > Stack looks like: > > [<c137ef97>] smp_alloc_memory+0x25 <-- > [<c137ef97>] smp_alloc_memory+0x25 > [<c137a500>] setup_arch+0x28e > [<c13735f7>] start_kernel+0x7a > [<c1379240>] xen_start_kernel+0x300
Ouch.
My first guess is that the BIOS data area is completely non-existent for Xen. Is it guaranteed that the memory is zeroed out on boot? In that case we can special-case it easily:
change: /* Paranoia: should never happen, but... */ if (lowmem >= 0x100000) lowmem = 0xa0000;
into: /* Strange case, like Xen ;) */ if (lowmem == 0 || lowmem >= 0x100000) lowmem = 0x9f000;
Can you test that?
Greetings, Alexander -- Alexander van Heukelum heukelum@fastmail.fm
-- http://www.fastmail.fm - Send your email first class
| |