Without announcing hvc0 as preferred it won't get used as long as tty0
gets registered earlier. This is particularly problematic with there not
being any screen output for PVH Dom0 when the screen is in graphics
mode, as the necessary information doesn't get conveyed yet from the
hypervisor.
Follow PV's model, but be conservative and do this for Dom0 only for
now.
Signed-off-by: Jan Beulich <[email protected]>
---
Prior to 418492ba40b2 ("x86/virt/xen: Use guest_late_init to detect Xen
PVH guest") x86_init.oem.arch_setup was already used by PVH, so I assume
the use of this hook is acceptable here.
Seeing that change, I wonder in how far setting xen_pvh to true only in
xen_hvm_guest_late_init() can really work: This hook, as its name says,
gets called pretty late; any decision taken earlier might have been
wrong. One such wrong decision is what gets added here - preferred
consoles won't be registered when taking that path. While adding a 2nd
call there might work, aiui they would better be registered prior to
parse_early_param(), i.e. before "earlyprintk=" gets evaluated.
I also consider tying "detecting" PVH mode to the no-VGA and no-CMOS-RTC
FADT flags as problematic looking forward: There may conceivably be
"legacy free" HVM guests down the road, yet they shouldn't be mistaken
for being PVH. Most of the XEN_X86_EMU_* controlled functionality would
seem unsuitable for the same reason; presence/absence of
XENFEAT_hvm_pirqs (tied to XEN_X86_EMU_USE_PIRQ) might be sufficiently
reliable an indicator. Question there is whether the separation
introduced by Xen commit b96b50004804 ("x86: remove XENFEAT_hvm_pirqs
for PVHv2 guests") came early enough in the process of enabling PVHv2.
Plus I'm not sure a HVM guest without pass-through enabled couldn't be
run with this off (i.e. by relaxing emulation_flags_ok() and having the
tool stack not request this emulation in such cases).
I think the approach here might be equally applicable for DomU, albeit
potentially pointless (i.e. dropping the conditional might make sense
even if simply benign there): A PVH DomU ought to never come with a VGA
console. Yet even then a dummy one may still get registered and would
take precedence over hvc?
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -3,6 +3,7 @@
#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
#include <linux/memblock.h>
#endif
+#include <linux/console.h>
#include <linux/cpu.h>
#include <linux/kexec.h>
#include <linux/slab.h>
@@ -18,6 +19,7 @@
#include <asm/xen/hypervisor.h>
#include <asm/cpu.h>
#include <asm/e820/api.h>
+#include <asm/setup.h>
#include "xen-ops.h"
#include "smp.h"
@@ -274,6 +276,16 @@ bool xen_running_on_version_or_later(uns
return false;
}
+void __init xen_add_preferred_consoles(void)
+{
+ add_preferred_console("xenboot", 0, NULL);
+ if (!boot_params.screen_info.orig_video_isVGA)
+ add_preferred_console("tty", 0, NULL);
+ add_preferred_console("hvc", 0, NULL);
+ if (boot_params.screen_info.orig_video_isVGA)
+ add_preferred_console("tty", 0, NULL);
+}
+
void xen_reboot(int reason)
{
struct sched_shutdown r = { .reason = reason };
--- a/arch/x86/xen/enlighten_pv.c
+++ b/arch/x86/xen/enlighten_pv.c
@@ -28,7 +28,6 @@
#include <linux/mm.h>
#include <linux/page-flags.h>
#include <linux/highmem.h>
-#include <linux/console.h>
#include <linux/pci.h>
#include <linux/gfp.h>
#include <linux/edd.h>
@@ -1382,12 +1381,7 @@ asmlinkage __visible void __init xen_sta
#endif
}
- add_preferred_console("xenboot", 0, NULL);
- if (!boot_params.screen_info.orig_video_isVGA)
- add_preferred_console("tty", 0, NULL);
- add_preferred_console("hvc", 0, NULL);
- if (boot_params.screen_info.orig_video_isVGA)
- add_preferred_console("tty", 0, NULL);
+ xen_add_preferred_consoles();
#ifdef CONFIG_PCI
/* PCI BIOS service won't work from a PV guest. */
--- a/arch/x86/xen/enlighten_pvh.c
+++ b/arch/x86/xen/enlighten_pvh.c
@@ -36,6 +36,9 @@ void __init xen_pvh_init(struct boot_par
pfn = __pa(hypercall_page);
wrmsr_safe(msr, (u32)pfn, (u32)(pfn >> 32));
+ if (xen_initial_domain())
+ x86_init.oem.arch_setup = xen_add_preferred_consoles;
+
xen_efi_init(boot_params);
}
--- a/arch/x86/xen/xen-ops.h
+++ b/arch/x86/xen/xen-ops.h
@@ -118,6 +118,8 @@ static inline void __init xen_init_vga(c
}
#endif
+void xen_add_preferred_consoles(void);
+
void __init xen_init_apic(void);
#ifdef CONFIG_XEN_EFI
On 07.09.21 12:10, Jan Beulich wrote:
> Without announcing hvc0 as preferred it won't get used as long as tty0
> gets registered earlier. This is particularly problematic with there not
> being any screen output for PVH Dom0 when the screen is in graphics
> mode, as the necessary information doesn't get conveyed yet from the
> hypervisor.
>
> Follow PV's model, but be conservative and do this for Dom0 only for
> now.
>
> Signed-off-by: Jan Beulich <[email protected]>
Reviewed-by: Juergen Gross <[email protected]>
> ---
> Prior to 418492ba40b2 ("x86/virt/xen: Use guest_late_init to detect Xen
> PVH guest") x86_init.oem.arch_setup was already used by PVH, so I assume
> the use of this hook is acceptable here.
Yes, I think so.
> Seeing that change, I wonder in how far setting xen_pvh to true only in
> xen_hvm_guest_late_init() can really work: This hook, as its name says,
> gets called pretty late; any decision taken earlier might have been
> wrong. One such wrong decision is what gets added here - preferred
> consoles won't be registered when taking that path. While adding a 2nd
> call there might work, aiui they would better be registered prior to
> parse_early_param(), i.e. before "earlyprintk=" gets evaluated.
>
> I also consider tying "detecting" PVH mode to the no-VGA and no-CMOS-RTC
> FADT flags as problematic looking forward: There may conceivably be
> "legacy free" HVM guests down the road, yet they shouldn't be mistaken
> for being PVH. Most of the XEN_X86_EMU_* controlled functionality would
> seem unsuitable for the same reason; presence/absence of
> XENFEAT_hvm_pirqs (tied to XEN_X86_EMU_USE_PIRQ) might be sufficiently
> reliable an indicator. Question there is whether the separation
> introduced by Xen commit b96b50004804 ("x86: remove XENFEAT_hvm_pirqs
> for PVHv2 guests") came early enough in the process of enabling PVHv2.
Yes, it did. The boot path not using the PVH specific entry point was
enabled with Xen 4.11, while commit b96b50004804 was in 4.9.
Juergen