Created attachment 303502 [details] c.img.xz: Guest software to reproduce this bug (xz compressed) Host CPU model: 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz Host kernel version: 6.0.14-300.fc37.x86_64 Host kernel arch: x86_64 Guest: a system level software (called LHV) I wrote myself, 32-bits, compressed and attached as c.img.xz. QEMU command line: qemu-system-i386 -m 256M -smp 4 -enable-kvm -serial stdio -drive media=disk,file=c.img The problem does not go away if using -machine kernel_irqchip=off The problem goes away if -accel tcg is used (also remove -enable-kvm) Actual behavior: serial port shows a few lines, then stops. Example: ... counts: count0 count1 count2 count0-count1+count2 counts: 1 0 1 0 counts: 2 1 1 0 counts: 3 2 1 0 counts: 4 2 3 -1 counts: 5 3 4 -2 counts: 6 3 5 -2 counts: 7 4 6 -3 counts: 8 4 7 -3 counts: 9 4 8 -3 counts: 10 5 8 -3 (no more outputs due to deadlock) Usually the output is shorter. Sometimes only the table header. Expected behaivor (reproducible using TCG): serial port shows: ... counts: count0 count1 count2 count0-count1+count2 counts: 1 0 1 0 counts: 2 1 1 0 counts: 3 1 2 0 counts: 4 2 2 0 counts: 5 2 3 0 counts: 6 3 3 0 counts: 7 3 4 0 counts: 8 4 4 0 counts: 9 5 4 0 counts: 10 5 5 0 counts: 11 5 6 0 counts: 12 6 6 0 counts: 13 7 6 0 ... Explanation: See the following for source code, line 5 - 89: https://github.com/lxylxy123456/uberxmhf/blob/b5935eaf8aab38ce1933da1c1be22dcf1b992eaf/xmhf/src/xmhf-core/xmhf-runtime/xmhf-startup/lhv.c#L5 My code performs the following experiment repeatedly on 3 CPUs: * Initially, "ptr" at address 0xb8000 (VGA memory mapped I/O) is set to 0 * CPU 0 writes 0x12345678 to ptr, then increases counter "count0". * In an infinite loop, CPU 1 tries exchanges ptr with register EAX (contains 0) using the XCHG instruction. If CPU 1 sees 0x12345678, it increases counter "count1". * CPU 2's behavior is similar to CPU 1, except it increases counter "count2" when it sees 0x12345678. Ideally, after each experiment there should always be count1 + count2 = count0. However, in KVM, there may be count1 + count2 > count0. This because CPU 0 writes 0x12345678 to ptr once, but CPU 1 and CPU 2 both get 0x12345678 in XCHG. Note that XCHG instruction always implements the locking protocol. There is also a deadlock after running the experiment a few times. However I am not trying to explain it for now. Guessed cause: I guess that KVM emulates the XCHG instruction that accesses 0xb8000. The call stack should be: ... x86_emulate_instruction (arch/x86/kvm/x86.c) x86_emulate_insn (arch/x86/kvm/emulate.c) writeback (arch/x86/kvm/emulate.c) segmented_cmpxchg (arch/x86/kvm/emulate.c) emulator_cmpxchg_emulated (arch/x86/kvm/x86.c, ->cmpxchg_emulated) emulator_try_cmpxchg_user (arch/x86/kvm/x86.c) ... CMPXCHG instruction Suppose CPU 2 wants to write 0 to ptr using writeback(), and expecting ptr to already contain 0x13245678. However, CPU 1 changes the content of ptr to 0. So * The CMPXCHG instruction fails (clears ZF). * emulator_try_cmpxchg_user returns 1. * emulator_cmpxchg_emulated() returns X86EMUL_CMPXCHG_FAILED. * segmented_cmpxchg() returns X86EMUL_CMPXCHG_FAILED. * writeback() returns X86EMUL_CMPXCHG_FAILED. * x86_emulate_insn() returns EMULATION_OK. Thus, I think the root cause of this bug is that x86_emulate_insn() ignores the X86EMUL_CMPXCHG_FAILED error. The correct behavior should be retrying the emulation using the updated value (similar to load-linked/store-conditional).
On Fri, Dec 30, 2022, bugzilla-daemon@kernel.org wrote: > My code performs the following experiment repeatedly on 3 CPUs: > > * Initially, "ptr" at address 0xb8000 (VGA memory mapped I/O) is set to 0 > * CPU 0 writes 0x12345678 to ptr, then increases counter "count0". > * In an infinite loop, CPU 1 tries exchanges ptr with register EAX (contains > 0) > using the XCHG instruction. If CPU 1 sees 0x12345678, it increases counter > "count1". > * CPU 2's behavior is similar to CPU 1, except it increases counter "count2" > when it sees 0x12345678. > > Ideally, after each experiment there should always be count1 + count2 = > count0. > However, in KVM, there may be count1 + count2 > count0. This because CPU 0 > writes 0x12345678 to ptr once, but CPU 1 and CPU 2 both get 0x12345678 in > XCHG. > Note that XCHG instruction always implements the locking protocol. > > There is also a deadlock after running the experiment a few times. However I > am > not trying to explain it for now. Is the suspect deadlock in userspace, the guest, or in the host kernel? > Guessed cause: > > I guess that KVM emulates the XCHG instruction that accesses 0xb8000. The > call > stack should be: > > ... > x86_emulate_instruction (arch/x86/kvm/x86.c) > x86_emulate_insn (arch/x86/kvm/emulate.c) > writeback (arch/x86/kvm/emulate.c) > segmented_cmpxchg (arch/x86/kvm/emulate.c) > emulator_cmpxchg_emulated (arch/x86/kvm/x86.c, ->cmpxchg_emulated) > emulator_try_cmpxchg_user (arch/x86/kvm/x86.c) > ... > CMPXCHG instruction > > Suppose CPU 2 wants to write 0 to ptr using writeback(), and expecting ptr to > already contain 0x13245678. However, CPU 1 changes the content of ptr to 0. > So > * The CMPXCHG instruction fails (clears ZF). > * emulator_try_cmpxchg_user returns 1. > * emulator_cmpxchg_emulated() returns X86EMUL_CMPXCHG_FAILED. > * segmented_cmpxchg() returns X86EMUL_CMPXCHG_FAILED. > * writeback() returns X86EMUL_CMPXCHG_FAILED. > * x86_emulate_insn() returns EMULATION_OK. > > Thus, I think the root cause of this bug is that x86_emulate_insn() ignores > the > X86EMUL_CMPXCHG_FAILED error. The correct behavior should be retrying the > emulation using the updated value (similar to load-linked/store-conditional). KVM does retry the emulation, albeit in a very roundabout and non-robust way. On X86EMUL_CMPXCHG_FAILED, x86_emulate_insn() skips the EIP update and doesn't writeback GPRs. x86_emulate_instruction() is flawed and emulates single-step, but the "eip" written should be the original RIP, i.e. shouldn't advance past the instructions being emulated. The single-step mess should be fixed, but I doubt that's the root cause here. Is there a memslot for 0xb8000? I assume not since KVM is emulating (have you actually verified that, e.g. with tracepoints?). KVM's ABI doesn't support atomic MMIO operations, i.e. if there's no memslot, KVM will effectively drop the LOCK semantics. If that's indeed what's happening, you should see kvm: emulating exchange as write in the host dmesg (just once though).
在 2023-01-03星期二的 22:05 +0000，bugzilla-daemon@kernel.org写道： > https://bugzilla.kernel.org/show_bug.cgi?id=216867 > > --- Comment #1 from Sean Christopherson (seanjc@google.com) --- > On Fri, Dec 30, 2022, bugzilla-daemon@kernel.org wrote: > > > My code performs the following experiment repeatedly on 3 CPUs: > > > > * Initially, "ptr" at address 0xb8000 (VGA memory mapped I/O) is > > set to 0 > > * CPU 0 writes 0x12345678 to ptr, then increases counter "count0". > > * In an infinite loop, CPU 1 tries exchanges ptr with register EAX > > (contains > > 0) > > using the XCHG instruction. If CPU 1 sees 0x12345678, it increases > > counter > > "count1". > > * CPU 2's behavior is similar to CPU 1, except it increases counter > > "count2" > > when it sees 0x12345678. > > > > Ideally, after each experiment there should always be count1 + > > count2 = > > count0. > > However, in KVM, there may be count1 + count2 > count0. This > > because CPU 0 > > writes 0x12345678 to ptr once, but CPU 1 and CPU 2 both get > > 0x12345678 in > > XCHG. > > Note that XCHG instruction always implements the locking protocol. > > > > There is also a deadlock after running the experiment a few times. > > However I > > am > > not trying to explain it for now. > > Is the suspect deadlock in userspace, the guest, or in the host > kernel? > The deadlock happens in the guest. It is due to how my experiment is implemented. It is not directly related to KVM. > > Guessed cause: > > > > I guess that KVM emulates the XCHG instruction that accesses > > 0xb8000. The > > call > > stack should be: > > > > ... > > x86_emulate_instruction (arch/x86/kvm/x86.c) > > x86_emulate_insn (arch/x86/kvm/emulate.c) > > writeback (arch/x86/kvm/emulate.c) > > segmented_cmpxchg (arch/x86/kvm/emulate.c) > > emulator_cmpxchg_emulated (arch/x86/kvm/x86.c, - > > >cmpxchg_emulated) > > emulator_try_cmpxchg_user (arch/x86/kvm/x86.c) > > ... > > CMPXCHG instruction > > > > Suppose CPU 2 wants to write 0 to ptr using writeback(), and > > expecting ptr to > > already contain 0x13245678. However, CPU 1 changes the content of > > ptr to 0. > > So > > * The CMPXCHG instruction fails (clears ZF). > > * emulator_try_cmpxchg_user returns 1. > > * emulator_cmpxchg_emulated() returns X86EMUL_CMPXCHG_FAILED. > > * segmented_cmpxchg() returns X86EMUL_CMPXCHG_FAILED. > > * writeback() returns X86EMUL_CMPXCHG_FAILED. > > * x86_emulate_insn() returns EMULATION_OK. > > > > Thus, I think the root cause of this bug is that x86_emulate_insn() > > ignores > > the > > X86EMUL_CMPXCHG_FAILED error. The correct behavior should be > > retrying the > > emulation using the updated value (similar to load-linked/store- > > conditional). > > KVM does retry the emulation, albeit in a very roundabout and non- > robust way. > On X86EMUL_CMPXCHG_FAILED, x86_emulate_insn() skips the EIP update > and doesn't > writeback GPRs. x86_emulate_instruction() is flawed and emulates > single-step, > but > the "eip" written should be the original RIP, i.e. shouldn't advance > past the > instructions being emulated. The single-step mess should be fixed, > but I doubt > that's the root cause here. > I see, thanks for the explanation. Now the retrying code looks correct to me (though I agree that the code could have been written in a better way). > Is there a memslot for 0xb8000? I assume not since KVM is emulating > (have you > actually verified that, e.g. with tracepoints?). KVM's ABI doesn't > support > atomic MMIO operations, i.e. if there's no memslot, KVM will > effectively drop > the LOCK semantics. If that's indeed what's happening, you should > see > > kvm: emulating exchange as write > > in the host dmesg (just once though). > You are right. I see "kvm: emulating exchange as write" when I run the guest I wrote. Looks like this is the check that causes KVM to drop LOCK on VGA MMIO: > gpa = kvm_mmu_gva_to_gpa_write(vcpu, addr, NULL); > > if (gpa == INVALID_GPA || > (gpa & PAGE_MASK) == APIC_DEFAULT_PHYS_BASE) > goto emul_write; Closing this bug since LOCK on MMIO is not supported by KVM's ABI.
Just FYI: VMware Workstation Player 16.2.4 has the same behavior, which is reported in https://communities.vmware.com/t5/VMware-Workstation-Player/LOCK-instruction-atomicity-broken-on-VGA-memory-mapped-IO/m-p/2946505#M39982 I tested on the following hypervisors / emulators, and do not see this bug: * Virtual Box (version 6.1.40) * QEMU TCG (version 5.2.0) * Hyper-V