| Message ID | 20190719204110.18306-1-sean.j.christopherson@intel.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | KVM: VMX: Optimize VMX instrs error/fault handling | expand |
Paolo, Any chance this can be picked up for 5.4? Josh's kvm_spurious_fault() changes went into 5.3, so this can be taken through the KVM tree, which probably makes the most sense since the non-KVM change is a one line deletion. On Fri, Jul 19, 2019 at 01:41:05PM -0700, Sean Christopherson wrote: > A recent commit reworked __kvm_handle_fault_on_reboot() to play nice with > objtool. An unfortunate side effect is that JMP is now inserted after > most VMX instructions so that the reboot macro can use an actual CALL to > kvm_spurious_fault() instead of a funky PUSH+JMP facsimile in .fixup. > > Rework the low level VMX instruction helpers to handle unexpected faults > manually instead of relying on the "fault on reboot" macro. By using > asm-goto, most helpers can branch directly to an in-function call to > kvm_spurious_fault(), which can then be optimized by compilers to reside > out-of-line at the end of the function instead of inline as done by > "fault on reboot". > > The net impact relative to the current code base is more or less a nop > when building with a compiler that supports __GCC_ASM_FLAG_OUTPUTS__. > A bunch of code that was previously in .fixup gets moved into the slow > paths of functions, but the fast paths are more basically unchanged. > > Without __GCC_ASM_FLAG_OUTPUTS__, manually coding the Jcc is a net > positive as CC_SET() without compiler support almost always generates a > SETcc+CMP+Jcc sequence, which is now replaced with a single Jcc. > > A small bonus is that the Jcc instrs are hinted to predict that the VMX > instr will be successful. > > v2: > - Rebased to x86/master, commit eceffd88ca20 ("Merge branch 'x86/urgent'") > - Reworded changelogs to reference the commit instead lkml link for > the recent changes to __kvm_handle_fault_on_reboot(). > - Added Paolo's acks for patch 1-4 > - Added patch 5 to do more cleanup, which was made possible by rebasing > on top of the __kvm_handle_fault_on_reboot() changes. > > Sean Christopherson (5): > objtool: KVM: x86: Check kvm_rebooting in kvm_spurious_fault() > KVM: VMX: Optimize VMX instruction error and fault handling > KVM: VMX: Add error handling to VMREAD helper > KVM: x86: Drop ____kvm_handle_fault_on_reboot() > KVM: x86: Don't check kvm_rebooting in __kvm_handle_fault_on_reboot() > > arch/x86/include/asm/kvm_host.h | 16 ++---- > arch/x86/kvm/vmx/ops.h | 93 ++++++++++++++++++++------------- > arch/x86/kvm/vmx/vmx.c | 42 +++++++++++++++ > arch/x86/kvm/x86.c | 3 +- > tools/objtool/check.c | 1 - > 5 files changed, 104 insertions(+), 51 deletions(-) > > -- > 2.22.0 >
On 19/07/19 22:41, Sean Christopherson wrote: > A recent commit reworked __kvm_handle_fault_on_reboot() to play nice with > objtool. An unfortunate side effect is that JMP is now inserted after > most VMX instructions so that the reboot macro can use an actual CALL to > kvm_spurious_fault() instead of a funky PUSH+JMP facsimile in .fixup. > > Rework the low level VMX instruction helpers to handle unexpected faults > manually instead of relying on the "fault on reboot" macro. By using > asm-goto, most helpers can branch directly to an in-function call to > kvm_spurious_fault(), which can then be optimized by compilers to reside > out-of-line at the end of the function instead of inline as done by > "fault on reboot". > > The net impact relative to the current code base is more or less a nop > when building with a compiler that supports __GCC_ASM_FLAG_OUTPUTS__. > A bunch of code that was previously in .fixup gets moved into the slow > paths of functions, but the fast paths are more basically unchanged. > > Without __GCC_ASM_FLAG_OUTPUTS__, manually coding the Jcc is a net > positive as CC_SET() without compiler support almost always generates a > SETcc+CMP+Jcc sequence, which is now replaced with a single Jcc. > > A small bonus is that the Jcc instrs are hinted to predict that the VMX > instr will be successful. > > v2: > - Rebased to x86/master, commit eceffd88ca20 ("Merge branch 'x86/urgent'") > - Reworded changelogs to reference the commit instead lkml link for > the recent changes to __kvm_handle_fault_on_reboot(). > - Added Paolo's acks for patch 1-4 > - Added patch 5 to do more cleanup, which was made possible by rebasing > on top of the __kvm_handle_fault_on_reboot() changes. > > Sean Christopherson (5): > objtool: KVM: x86: Check kvm_rebooting in kvm_spurious_fault() > KVM: VMX: Optimize VMX instruction error and fault handling > KVM: VMX: Add error handling to VMREAD helper > KVM: x86: Drop ____kvm_handle_fault_on_reboot() > KVM: x86: Don't check kvm_rebooting in __kvm_handle_fault_on_reboot() > > arch/x86/include/asm/kvm_host.h | 16 ++---- > arch/x86/kvm/vmx/ops.h | 93 ++++++++++++++++++++------------- > arch/x86/kvm/vmx/vmx.c | 42 +++++++++++++++ > arch/x86/kvm/x86.c | 3 +- > tools/objtool/check.c | 1 - > 5 files changed, 104 insertions(+), 51 deletions(-) > Queued, thanks. Paolo
A recent commit reworked __kvm_handle_fault_on_reboot() to play nice with objtool. An unfortunate side effect is that JMP is now inserted after most VMX instructions so that the reboot macro can use an actual CALL to kvm_spurious_fault() instead of a funky PUSH+JMP facsimile in .fixup. Rework the low level VMX instruction helpers to handle unexpected faults manually instead of relying on the "fault on reboot" macro. By using asm-goto, most helpers can branch directly to an in-function call to kvm_spurious_fault(), which can then be optimized by compilers to reside out-of-line at the end of the function instead of inline as done by "fault on reboot". The net impact relative to the current code base is more or less a nop when building with a compiler that supports __GCC_ASM_FLAG_OUTPUTS__. A bunch of code that was previously in .fixup gets moved into the slow paths of functions, but the fast paths are more basically unchanged. Without __GCC_ASM_FLAG_OUTPUTS__, manually coding the Jcc is a net positive as CC_SET() without compiler support almost always generates a SETcc+CMP+Jcc sequence, which is now replaced with a single Jcc. A small bonus is that the Jcc instrs are hinted to predict that the VMX instr will be successful. v2: - Rebased to x86/master, commit eceffd88ca20 ("Merge branch 'x86/urgent'") - Reworded changelogs to reference the commit instead lkml link for the recent changes to __kvm_handle_fault_on_reboot(). - Added Paolo's acks for patch 1-4 - Added patch 5 to do more cleanup, which was made possible by rebasing on top of the __kvm_handle_fault_on_reboot() changes. Sean Christopherson (5): objtool: KVM: x86: Check kvm_rebooting in kvm_spurious_fault() KVM: VMX: Optimize VMX instruction error and fault handling KVM: VMX: Add error handling to VMREAD helper KVM: x86: Drop ____kvm_handle_fault_on_reboot() KVM: x86: Don't check kvm_rebooting in __kvm_handle_fault_on_reboot() arch/x86/include/asm/kvm_host.h | 16 ++---- arch/x86/kvm/vmx/ops.h | 93 ++++++++++++++++++++------------- arch/x86/kvm/vmx/vmx.c | 42 +++++++++++++++ arch/x86/kvm/x86.c | 3 +- tools/objtool/check.c | 1 - 5 files changed, 104 insertions(+), 51 deletions(-)