Skip to contentSkip to conversation

Use para_fill instead of vmi_get_function for APIC ops

RFC In review Anthony Liguori 3 participants 6 messages Feb 27, 2007

lore
Share & tools
lore
#0 lore raw

RFC 1/1 Use para_fill instead of vmi_get_function for APIC ops

Hi Zach,

It seems to me that the APIC paravirt_ops should be filled by para_fill() instead of vmi_get_function(). vmi_get_function() returns a nop when the relocation type is NONE. para_fill() leaves the native code in place.

The native version of the apic write ops is more or less *(APIC_BASE + reg) = value. APIC_BASE is unknown to the ROM so it's impossible to simulate this in the ROM.

This means that a ROM has no choice but to do APIC emulation (or jump through seriously hairy loops to get the APIC mapped in it's address space). Was this the intention?

N.B. attached patch is just to illustrate the point. Has not even been compile tested.

Regards,

Anthony Liguori

arch/i386/kernel/vmi.c renamed from arch/i386/kernel/vmi.c~ +3 -3
@@ -852,9 +852,9 @@
 #endif
 
 #ifdef CONFIG_X86_LOCAL_APIC
-	paravirt_ops.apic_read = vmi_get_function(VMI_CALL_APICRead);
-	paravirt_ops.apic_write = vmi_get_function(VMI_CALL_APICWrite);
-	paravirt_ops.apic_write_atomic = vmi_get_function(VMI_CALL_APICWrite);
+	para_fill(apic_read, APICRead);
+	para_fill(apic_write, APICWrite);
+	para_fill(apic_write_atomic, APICWrite);
 #endif
 
 	/*
#1 lore raw
Quoted 7 lines · Anthony Liguori

Hi Zach,

It seems to me that the APIC paravirt_ops should be filled by para_fill() instead of vmi_get_function(). vmi_get_function() returns a nop when the relocation type is NONE. para_fill() leaves the native code in place.

The native version of the apic write ops is more or less *(APIC_BASE + reg) = value. APIC_BASE is unknown to the ROM so it's impossible to simulate this in the ROM.

This means that a ROM has no choice but to do APIC emulation (or jump through seriously hairy loops to get the APIC mapped in it's address space). Was this the intention?

No, but certainly the effect. Actually, it is very easy to get the APIC mapped in the ROM address space without jumping through seriously hairy loops - we do it today in our hypervisor.

Quoted 2 lines

N.B. attached patch is just to illustrate the point. Has not even been compile tested.

Patch looks good, thanks. But the whole para_fill / vmi_get_function stuff could probably be done even cleaner. It was just a helper at first to work around the awkward syntax, and it is still a bit ugly, but I haven't come up with a better solution yet, mostly because with the new inlining work Jeremy is doing, we might want to start doing selective inlining, in which case I'll have to go back over the code anyway to clean everything to get the logic right in all cases.

I assume this patch is signed-off-by you? If so, I'll add it to my patch queue.

Zach

#2 lore raw
Quoted 11 lines · Zachary Amsden
Quoted 9 lines · Anthony Liguori

Hi Zach,

It seems to me that the APIC paravirt_ops should be filled by para_fill() instead of vmi_get_function(). vmi_get_function() returns a nop when the relocation type is NONE. para_fill() leaves the native code in place.

The native version of the apic write ops is more or less *(APIC_BASE
+ reg) = value. APIC_BASE is unknown to the ROM so it's impossible to simulate this in the ROM.

This means that a ROM has no choice but to do APIC emulation (or jump through seriously hairy loops to get the APIC mapped in it's address space). Was this the intention?

Quoted 7 lines

No, but certainly the effect. Actually, it is very easy to get the APIC mapped in the ROM address space without jumping through seriously hairy loops - we do it today in our hypervisor.

Quoted 2 lines

N.B. attached patch is just to illustrate the point. Has not even been compile tested.

Patch looks good, thanks. But the whole para_fill / vmi_get_function stuff could probably be done even cleaner. It was just a helper at first to work around the awkward syntax, and it is still a bit ugly, but I haven't come up with a better solution yet, mostly because with the new inlining work Jeremy is doing, we might want to start doing selective inlining, in which case I'll have to go back over the code anyway to clean everything to get the logic right in all cases.

It would be really great if one could write a ROM by just specifying a GetRelocationInfo function that always returns rel.type == NONE. Right now, there are a half a dozen or so ops that have to be implemented b/c of the vmi_get_function stuff.

Quoted 1 line

I assume this patch is signed-off-by you? If so, I'll add it to my patch queue.

Yeah, but please make sure to test it. I haven't at all.

Thanks,

Anthony Liguori

Quoted 2 lines

Zach

#3 lore raw
Quoted 2 lines · Anthony Liguori

It would be really great if one could write a ROM by just specifying a GetRelocationInfo function that always returns rel.type == NONE. Right now, there are a half a dozen or so ops that have to be implemented b/c of the vmi_get_function stuff.

Yes, I need to clean this up. There are a couple other places where I took liberties and just did things that way because this was how our VMware ROM was implemented. Just be sure, if you are going to implement a ROM, it has to be GPL'd now, otherwise the VMI code won't accept it.

Quoted 4 lines
Quoted 2 lines

I assume this patch is signed-off-by you? If so, I'll add it to my patch queue.

Yeah, but please make sure to test it. I haven't at all.

It passes test level 1, which is good enough to make it into my patchset.

Zach

#4 lore raw
In reply to Zachary Amsden
Quoted 7 lines · Zachary Amsden

Patch looks good, thanks. But the whole para_fill / vmi_get_function
stuff could probably be done even cleaner. It was just a helper at first to work around the awkward syntax, and it is still a bit ugly, but I haven't come up with a better solution yet, mostly because with the new inlining work Jeremy is doing, we might want to start doing selective inlining, in which case I'll have to go back over the code anyway to clean everything to get the logic right in all cases.

Yes, my patching updates make pretty much all the pv_ops patchable, including the apic ones. The simple thing is to fill out paravirt_ops with the appropriate pointers, then set the .patch operation to paravirt_default_patch, which will basically turn them all into direct calls. If you want to start inlining stuff, then you can do that too.

J

#5 lore raw
In reply to Zachary Amsden
Quoted 5 lines · Zachary Amsden
Quoted 4 lines · Anthony Liguori

Hi Zach,

It seems to me that the APIC paravirt_ops should be filled by para_fill() instead of vmi_get_function(). vmi_get_function() returns a nop when the relocation type is NONE. para_fill() leaves the native code in place.

Quoted 7 lines
Quoted 5 lines

The native version of the apic write ops is more or less *(APIC_BASE
+ reg) = value. APIC_BASE is unknown to the ROM so it's impossible to simulate this in the ROM.

This means that a ROM has no choice but to do APIC emulation (or jump through seriously hairy loops to get the APIC mapped in it's address space). Was this the intention?

No, but certainly the effect. Actually, it is very easy to get the APIC mapped in the ROM address space without jumping through seriously hairy loops - we do it today in our hypervisor.

I neglected to mention that I didn't want to use a memory hole. One could allocate a small one to map the APIC but that seems to defeat the purpose of having a native ROM.

Regards,

Anthony Liguori

Quoted 10 lines
Quoted 2 lines

N.B. attached patch is just to illustrate the point. Has not even been compile tested.

Patch looks good, thanks. But the whole para_fill / vmi_get_function stuff could probably be done even cleaner. It was just a helper at first to work around the awkward syntax, and it is still a bit ugly, but I haven't come up with a better solution yet, mostly because with the new inlining work Jeremy is doing, we might want to start doing selective inlining, in which case I'll have to go back over the code anyway to clean everything to get the logic right in all cases.

I assume this patch is signed-off-by you? If so, I'll add it to my patch queue.

Zach