kvm_regs vs kvm_sregs?

DISCUSSION Hollis Blanchard 2 participants 3 messages Mar 16, 2007 – Mar 23, 2007 1 message not in the archive KVM

Share & tools

Hollis Blanchard

author

2007-03-16 15:00 UTC #0 lore raw

What is the distinction between kvm_regs and kvm_sregs? As far as I can see, kvm_regs is only used when emulating IO, emulating MMIO, and emulating CPUID, where guest GPRs are directly modified. kvm_sregs is only used for full CPU state save (for later restore).

When the kernel had to use copy_to_user() to transfer state to userland, I can see that this split resulted in less memory copying. However, now that userland can directly map register state without a copy, why not combine the two structures?

I guess it takes longer to copy state out of the VMCS into kvm_sregs, so why bother if userspace isn't going to use it?

-Hollis

-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV

Avi Kivity

2007-03-18 05:13 UTC #1 lore raw

Quoted 5 lines · Hollis Blanchard

What is the distinction between kvm_regs and kvm_sregs? As far as I can see, kvm_regs is only used when emulating IO, emulating MMIO, and emulating CPUID, where guest GPRs are directly modified. kvm_sregs is only used for full CPU state save (for later restore).

After the userspace interface changes, neither of them are used except for migration, savevm/loadvm, and debugging. Theoretically they could be unified, practically SET_SREGS is a dangerous operation on Intel cpus due to kvm's very imperfect real mode support.

Quoted 5 lines

When the kernel had to use copy_to_user() to transfer state to userland, I can see that this split resulted in less memory copying. However, now that userland can directly map register state without a copy, why not combine the two structures?

We don't map the register state as that would require us to sync it on every exit.

Quoted 3 lines

I guess it takes longer to copy state out of the VMCS into kvm_sregs, so why bother if userspace isn't going to use it?

When it's needed, it's really needed. Or did I misunderstand the question?

Signature · 8 lines

Do not meddle in the internals of kernels, for they are subtle and quick to panic.


-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV

A message that is not in the archive (sent off-list or not yet received): 45FCCA6D.3070207-atKUWr5tajBWk0Htik3J/[email protected]

Hollis Blanchard

author

2007-03-23 01:17 UTC #3 lore raw

Quoted 18 lines · Avi Kivity

Quoted 5 lines · Hollis Blanchard
What is the distinction between kvm_regs and kvm_sregs? As far as I can see, kvm_regs is only used when emulating IO, emulating MMIO, and emulating CPUID, where guest GPRs are directly modified. kvm_sregs is only used for full CPU state save (for later restore).
After the userspace interface changes, neither of them are used except for migration, savevm/loadvm, and debugging. Theoretically they could be unified, practically SET_SREGS is a dangerous operation on Intel cpus due to kvm's very imperfect real mode support.
Quoted 5 lines
When the kernel had to use copy_to_user() to transfer state to userland, I can see that this split resulted in less memory copying. However, now that userland can directly map register state without a copy, why not combine the two structures?

Quoted 8 lines

We don't map the register state as that would require us to sync it on every exit.
Quoted 3 lines
I guess it takes longer to copy state out of the VMCS into kvm_sregs, so why bother if userspace isn't going to use it?
When it's needed, it's really needed. Or did I misunderstand the question?

I guess I misunderstood. I thought the mmap was for userspace to access kvm_regs, which previously was needed when emulating IO.

Apparently this is no longer the case: instead the data is copied somewhere into mmap space and then complete_pio() is responsible for loading that raw data into the guest register state.

Seems a bit more convoluted, but I guess you've managed to divorce user and kernel data structures, if that was the goal.

-Hollis