]> xenbits.xensource.com Git - xen.git/commit
adjust system domain creation (and call it earlier on x86)
authorJan Beulich <jbeulich@suse.com>
Thu, 6 Jun 2019 09:16:57 +0000 (11:16 +0200)
committerJan Beulich <jbeulich@suse.com>
Thu, 6 Jun 2019 09:16:57 +0000 (11:16 +0200)
commit7177f589ba84433e1ca8bb97a5d074545133c49c
tree5f5129dcfe19c9586e4240da2aadd9d1a2a09358
parentf3f862f9b720abd6795986651a5b10e1a7e05f81
adjust system domain creation (and call it earlier on x86)

Split out this mostly arch-independent code into a common-code helper
function. (This does away with Arm's arch_init_memory() altogether.)

On x86 this needs to happen before acpi_boot_init(): Commit 9fa94e1058
("x86/ACPI: also parse AMD IOMMU tables early") only appeared to work
fine - it's really broken, and doesn't crash (on non-EFI AMD systems)
only because of there being a mapping of linear address 0 during early
boot. On EFI there is:

 Early fatal page fault at e008:ffff82d08024d58e (cr2=0000000000000220, ec=0000)
 ----[ Xen-4.13-unstable  x86_64  debug=y   Not tainted ]----
 CPU:    0
 RIP:    e008:[<ffff82d08024d58e>] pci.c#_pci_hide_device+0x17/0x3a
 RFLAGS: 0000000000010046   CONTEXT: hypervisor
 rax: 0000000000000000   rbx: 0000000000006000   rcx: 0000000000000000
 rdx: ffff83104f2ee9b0   rsi: ffff82e0209e5d48   rdi: ffff83104f2ee9a0
 rbp: ffff82d08081fce0   rsp: ffff82d08081fcb8   r8:  0000000000000000
 r9:  8000000000000000   r10: 0180000000000000   r11: 7fffffffffffffff
 r12: ffff83104f2ee9a0   r13: 0000000000000002   r14: ffff83104f2ee4b0
 r15: 0000000000000064   cr0: 0000000080050033   cr4: 00000000000000a0
 cr3: 000000009f614000   cr2: 0000000000000220
 fsb: 0000000000000000   gsb: 0000000000000000   gss: 0000000000000000
 ds: 0000   es: 0000   fs: 0000   gs: 0000   ss: 0000   cs: e008
 Xen code around <ffff82d08024d58e> (pci.c#_pci_hide_device+0x17/0x3a):
  48 89 47 38 48 8d 57 10 <48> 8b 88 20 02 00 00 48 89 51 08 48 89 4f 10 48
 Xen stack trace from rsp=ffff82d08081fcb8:
[...]
 Xen call trace:
    [<ffff82d08024d58e>] pci.c#_pci_hide_device+0x17/0x3a
[   [<                >] pci_ro_device+...]
    [<ffff82d080617fe1>] amd_iommu_detect_one_acpi+0x161/0x249
    [<ffff82d0806186ac>] iommu_acpi.c#detect_iommu_acpi+0xb5/0xe7
    [<ffff82d08061cde0>] acpi_table_parse+0x61/0x90
    [<ffff82d080619e7d>] amd_iommu_detect_acpi+0x17/0x19
    [<ffff82d08061790b>] acpi_ivrs_init+0x20/0x5b
    [<ffff82d08062e838>] acpi_boot_init+0x301/0x30f
    [<ffff82d080628b10>] __start_xen+0x1daf/0x28a2

 Pagetable walk from 0000000000000220:
  L4[0x000] = 000000009f44f063 ffffffffffffffff
  L3[0x000] = 000000009f44b063 ffffffffffffffff
  L2[0x000] = 0000000000000000 ffffffffffffffff

 ****************************************
 Panic on CPU 0:
 FATAL TRAP: vector = 14 (page fault)
 [error_code=0000] , IN INTERRUPT CONTEXT
 ****************************************

Of course the bug would nevertheless have lead to post-boot crashes as
soon as the list would actually get traversed.

Take the opportunity and
- convert BUG_ON()s being moved to panic(),
- add __read_mostly annotations to the dom_* definitions.

Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Julien Grall <julien.grall@arm.com>
xen/arch/arm/mm.c
xen/arch/arm/setup.c
xen/arch/x86/mm.c
xen/arch/x86/setup.c
xen/common/domain.c
xen/include/asm-arm/mm.h
xen/include/asm-arm/setup.h
xen/include/asm-x86/mm.h
xen/include/xen/domain.h
xen/include/xen/mm.h