| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <ef1665a2-9b54-6a9f-f2eb-06f1e54b9272@loongson.cn>
Date: Tue, 3 Feb 2026 17:37:46 +0800
From: Bibo Mao <maobibo@...ngson.cn>
To: Huacai Chen <chenhuacai@...nel.org>
Cc: WANG Xuerui <kernel@...0n.name>, Tianrui Zhao <zhaotianrui@...ngson.cn>,
loongarch@...ts.linux.dev, linux-kernel@...r.kernel.org, kvm@...r.kernel.org
Subject: Re: [PATCH v3 4/4] LoongArch: KVM: Add FPU delay load support
On 2026/2/3 下午5:17, Huacai Chen wrote:
> On Tue, Feb 3, 2026 at 4:59 PM Bibo Mao <maobibo@...ngson.cn> wrote:
>>
>>
>>
>> On 2026/2/3 下午4:50, Huacai Chen wrote:
>>> On Tue, Feb 3, 2026 at 3:51 PM Bibo Mao <maobibo@...ngson.cn> wrote:
>>>>
>>>>
>>>>
>>>> On 2026/2/3 下午3:34, Huacai Chen wrote:
>>>>> On Tue, Feb 3, 2026 at 2:48 PM Bibo Mao <maobibo@...ngson.cn> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 2026/2/3 下午12:15, Huacai Chen wrote:
>>>>>>> Hi, Bibo,
>>>>>>>
>>>>>>> On Tue, Feb 3, 2026 at 11:31 AM Bibo Mao <maobibo@...ngson.cn> wrote:
>>>>>>>>
>>>>>>>> FPU is lazy enabled with KVM hypervisor. After FPU is enabled and
>>>>>>>> loaded, vCPU can be preempted and FPU will be lost again, there will
>>>>>>>> be unnecessary FPU exception, load and store process. Here FPU is
>>>>>>>> delay load until guest enter entry.
>>>>>>> Calling LSX/LASX as FPU is a little strange, but somewhat reasonable.
>>>>>>> Calling LBT as FPU is very strange. So I still like the V1 logic.
>>>>>> yeap, LBT can use another different BIT and separate with FPU. It is
>>>>>> actually normal use one bit + fpu type variant to represent different
>>>>>> different FPU load requirement, such as
>>>>>> TIF_FOREIGN_FPSTATE/TIF_NEED_FPU_LOAD on other architectures.
>>>>>>
>>>>>> I think it is better to put int fpu_load_type in structure loongarch_fpu.
>>>>>>
>>>>>> And there will be another optimization to avoid load FPU again if FPU HW
>>>>>> is owned by current thread/vCPU, that will add last_cpu int type in
>>>>>> structure loongarch_fpu also.
>>>>>>
>>>>>> Regards
>>>>>> Bibo Mao
>>>>>>>
>>>>>>> If you insist on this version, please rename KVM_REQ_FPU_LOAD to
>>>>>>> KVM_REQ_AUX_LOAD and rename fpu_load_type to aux_type, which is
>>>>>>> similar to aux_inuse.
>>>>> Then why not consider this?
>>>> this can work now. However there is two different structure struct
>>>> loongarch_fpu and struct loongarch_lbt.
>>> Yes, but two structures don't block us from using KVM_REQ_AUX_LOAD and
>>> aux_type to abstract both FPU and LBT, which is similar to aux_inuse.
>>>>
>>>> 1. If kernel wants to use late FPU load, new element fpu_load_type can
>>>> be added in struct loongarch_fpu for both user app/KVM.
>> where aux_type is put for kernel/kvm? Put it in thread structure with
>> kernel late FPU load and vcpu.arch with KVM late FPU load?
> aux_type is renamed from fpu_load_type, so where fpu_load_type is,
> then where aux_type is.
>
>>>>
>>>> 2. With further optimization, FPU HW can own by user app/kernel/KVM,
>>>> there will be another last_cpu int type added in struct loongarch_fpu.
>>> Both loongarch_fpu and loongarch_lbt are register copies, so adding
>>> fpu_load_type/last_cpu is not a good idea.
>> If vCPU using FPU is preempted by kernel thread and kernel thread does
>> not use FPU, HW FPU is the same with SW FPU state, HW FPU load can be
>> skipped.
>>
>> BTW do you ever investigate FPU load/save process on other general
>> architectures except MIPS?
> I investigate nothing, including MIPS. Other architectures may give us
> some inspiration, but that doesn't mean we should copy them, no matter
> X86 or MIPS.
>
> X86 introduced lazy fpu, then others also use lazy fpu; but now X86
> have switched to eager fpu, others should also do the same?
>
> On the other hand, when you use separate FPU/LSX/LASX, I only mention
> the trace functions. Then you changed to centralized FPU/LSX/LASX/LBT.
> Then I suggest you improve centralized FPU/LSX/LASX/LBT, you changed
> to separate FPU/LBT again, where is the end?
OK, I can use aux bit and aux type for kvm only.
In future if there is FPU register skipping loading optimization patch,
then we will discuss it then.
Regards
Bibo Mao
>
>
>
> Huacai
>>
>> Regards
>> Bibo Mao
>>>
>>>
>>> Huacai
>>>>
>>>> Regards
>>>> Bibo Mao
>>>>
>>>> Regards
>>>> Bibo Mao
>>>>
>>>>>
>>>>> Huacai
>>>>>
>>>>>>>
>>>>>>> Huacai
>>>>>>>
>>>>>>>>
>>>>>>>> Signed-off-by: Bibo Mao <maobibo@...ngson.cn>
>>>>>>>> ---
>>>>>>>> arch/loongarch/include/asm/kvm_host.h | 2 ++
>>>>>>>> arch/loongarch/kvm/exit.c | 21 ++++++++++-----
>>>>>>>> arch/loongarch/kvm/vcpu.c | 37 ++++++++++++++++++---------
>>>>>>>> 3 files changed, 41 insertions(+), 19 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/arch/loongarch/include/asm/kvm_host.h b/arch/loongarch/include/asm/kvm_host.h
>>>>>>>> index e4fe5b8e8149..902ff7bc0e35 100644
>>>>>>>> --- a/arch/loongarch/include/asm/kvm_host.h
>>>>>>>> +++ b/arch/loongarch/include/asm/kvm_host.h
>>>>>>>> @@ -37,6 +37,7 @@
>>>>>>>> #define KVM_REQ_TLB_FLUSH_GPA KVM_ARCH_REQ(0)
>>>>>>>> #define KVM_REQ_STEAL_UPDATE KVM_ARCH_REQ(1)
>>>>>>>> #define KVM_REQ_PMU KVM_ARCH_REQ(2)
>>>>>>>> +#define KVM_REQ_FPU_LOAD KVM_ARCH_REQ(3)
>>>>>>>>
>>>>>>>> #define KVM_GUESTDBG_SW_BP_MASK \
>>>>>>>> (KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP)
>>>>>>>> @@ -234,6 +235,7 @@ struct kvm_vcpu_arch {
>>>>>>>> u64 vpid;
>>>>>>>> gpa_t flush_gpa;
>>>>>>>>
>>>>>>>> + int fpu_load_type;
>>>>>>>> /* Frequency of stable timer in Hz */
>>>>>>>> u64 timer_mhz;
>>>>>>>> ktime_t expire;
>>>>>>>> diff --git a/arch/loongarch/kvm/exit.c b/arch/loongarch/kvm/exit.c
>>>>>>>> index 65ec10a7245a..62403c7c6f9a 100644
>>>>>>>> --- a/arch/loongarch/kvm/exit.c
>>>>>>>> +++ b/arch/loongarch/kvm/exit.c
>>>>>>>> @@ -754,7 +754,8 @@ static int kvm_handle_fpu_disabled(struct kvm_vcpu *vcpu, int ecode)
>>>>>>>> return RESUME_HOST;
>>>>>>>> }
>>>>>>>>
>>>>>>>> - kvm_own_fpu(vcpu);
>>>>>>>> + vcpu->arch.fpu_load_type = KVM_LARCH_FPU;
>>>>>>>> + kvm_make_request(KVM_REQ_FPU_LOAD, vcpu);
>>>>>>>>
>>>>>>>> return RESUME_GUEST;
>>>>>>>> }
>>>>>>>> @@ -794,8 +795,10 @@ static int kvm_handle_lsx_disabled(struct kvm_vcpu *vcpu, int ecode)
>>>>>>>> {
>>>>>>>> if (!kvm_guest_has_lsx(&vcpu->arch))
>>>>>>>> kvm_queue_exception(vcpu, EXCCODE_INE, 0);
>>>>>>>> - else
>>>>>>>> - kvm_own_lsx(vcpu);
>>>>>>>> + else {
>>>>>>>> + vcpu->arch.fpu_load_type = KVM_LARCH_LSX;
>>>>>>>> + kvm_make_request(KVM_REQ_FPU_LOAD, vcpu);
>>>>>>>> + }
>>>>>>>>
>>>>>>>> return RESUME_GUEST;
>>>>>>>> }
>>>>>>>> @@ -812,8 +815,10 @@ static int kvm_handle_lasx_disabled(struct kvm_vcpu *vcpu, int ecode)
>>>>>>>> {
>>>>>>>> if (!kvm_guest_has_lasx(&vcpu->arch))
>>>>>>>> kvm_queue_exception(vcpu, EXCCODE_INE, 0);
>>>>>>>> - else
>>>>>>>> - kvm_own_lasx(vcpu);
>>>>>>>> + else {
>>>>>>>> + vcpu->arch.fpu_load_type = KVM_LARCH_LASX;
>>>>>>>> + kvm_make_request(KVM_REQ_FPU_LOAD, vcpu);
>>>>>>>> + }
>>>>>>>>
>>>>>>>> return RESUME_GUEST;
>>>>>>>> }
>>>>>>>> @@ -822,8 +827,10 @@ static int kvm_handle_lbt_disabled(struct kvm_vcpu *vcpu, int ecode)
>>>>>>>> {
>>>>>>>> if (!kvm_guest_has_lbt(&vcpu->arch))
>>>>>>>> kvm_queue_exception(vcpu, EXCCODE_INE, 0);
>>>>>>>> - else
>>>>>>>> - kvm_own_lbt(vcpu);
>>>>>>>> + else {
>>>>>>>> + vcpu->arch.fpu_load_type = KVM_LARCH_LBT;
>>>>>>>> + kvm_make_request(KVM_REQ_FPU_LOAD, vcpu);
>>>>>>>> + }
>>>>>>>>
>>>>>>>> return RESUME_GUEST;
>>>>>>>> }
>>>>>>>> diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c
>>>>>>>> index 995461d724b5..d05fe6c8f456 100644
>>>>>>>> --- a/arch/loongarch/kvm/vcpu.c
>>>>>>>> +++ b/arch/loongarch/kvm/vcpu.c
>>>>>>>> @@ -232,6 +232,31 @@ static void kvm_late_check_requests(struct kvm_vcpu *vcpu)
>>>>>>>> kvm_flush_tlb_gpa(vcpu, vcpu->arch.flush_gpa);
>>>>>>>> vcpu->arch.flush_gpa = INVALID_GPA;
>>>>>>>> }
>>>>>>>> +
>>>>>>>> + if (kvm_check_request(KVM_REQ_FPU_LOAD, vcpu)) {
>>>>>>>> + switch (vcpu->arch.fpu_load_type) {
>>>>>>>> + case KVM_LARCH_FPU:
>>>>>>>> + kvm_own_fpu(vcpu);
>>>>>>>> + break;
>>>>>>>> +
>>>>>>>> + case KVM_LARCH_LSX:
>>>>>>>> + kvm_own_lsx(vcpu);
>>>>>>>> + break;
>>>>>>>> +
>>>>>>>> + case KVM_LARCH_LASX:
>>>>>>>> + kvm_own_lasx(vcpu);
>>>>>>>> + break;
>>>>>>>> +
>>>>>>>> + case KVM_LARCH_LBT:
>>>>>>>> + kvm_own_lbt(vcpu);
>>>>>>>> + break;
>>>>>>>> +
>>>>>>>> + default:
>>>>>>>> + break;
>>>>>>>> + }
>>>>>>>> +
>>>>>>>> + vcpu->arch.fpu_load_type = 0;
>>>>>>>> + }
>>>>>>>> }
>>>>>>>>
>>>>>>>> /*
>>>>>>>> @@ -1286,13 +1311,11 @@ int kvm_arch_vcpu_ioctl_set_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
>>>>>>>> #ifdef CONFIG_CPU_HAS_LBT
>>>>>>>> int kvm_own_lbt(struct kvm_vcpu *vcpu)
>>>>>>>> {
>>>>>>>> - preempt_disable();
>>>>>>>> if (!(vcpu->arch.aux_inuse & KVM_LARCH_LBT)) {
>>>>>>>> set_csr_euen(CSR_EUEN_LBTEN);
>>>>>>>> _restore_lbt(&vcpu->arch.lbt);
>>>>>>>> vcpu->arch.aux_inuse |= KVM_LARCH_LBT;
>>>>>>>> }
>>>>>>>> - preempt_enable();
>>>>>>>>
>>>>>>>> return 0;
>>>>>>>> }
>>>>>>>> @@ -1335,8 +1358,6 @@ static inline void kvm_check_fcsr_alive(struct kvm_vcpu *vcpu) { }
>>>>>>>> /* Enable FPU and restore context */
>>>>>>>> void kvm_own_fpu(struct kvm_vcpu *vcpu)
>>>>>>>> {
>>>>>>>> - preempt_disable();
>>>>>>>> -
>>>>>>>> /*
>>>>>>>> * Enable FPU for guest
>>>>>>>> * Set FR and FRE according to guest context
>>>>>>>> @@ -1347,16 +1368,12 @@ void kvm_own_fpu(struct kvm_vcpu *vcpu)
>>>>>>>> kvm_restore_fpu(&vcpu->arch.fpu);
>>>>>>>> vcpu->arch.aux_inuse |= KVM_LARCH_FPU;
>>>>>>>> trace_kvm_aux(vcpu, KVM_TRACE_AUX_RESTORE, KVM_TRACE_AUX_FPU);
>>>>>>>> -
>>>>>>>> - preempt_enable();
>>>>>>>> }
>>>>>>>>
>>>>>>>> #ifdef CONFIG_CPU_HAS_LSX
>>>>>>>> /* Enable LSX and restore context */
>>>>>>>> int kvm_own_lsx(struct kvm_vcpu *vcpu)
>>>>>>>> {
>>>>>>>> - preempt_disable();
>>>>>>>> -
>>>>>>>> /* Enable LSX for guest */
>>>>>>>> kvm_check_fcsr(vcpu, vcpu->arch.fpu.fcsr);
>>>>>>>> set_csr_euen(CSR_EUEN_LSXEN | CSR_EUEN_FPEN);
>>>>>>>> @@ -1378,7 +1395,6 @@ int kvm_own_lsx(struct kvm_vcpu *vcpu)
>>>>>>>>
>>>>>>>> trace_kvm_aux(vcpu, KVM_TRACE_AUX_RESTORE, KVM_TRACE_AUX_LSX);
>>>>>>>> vcpu->arch.aux_inuse |= KVM_LARCH_LSX | KVM_LARCH_FPU;
>>>>>>>> - preempt_enable();
>>>>>>>>
>>>>>>>> return 0;
>>>>>>>> }
>>>>>>>> @@ -1388,8 +1404,6 @@ int kvm_own_lsx(struct kvm_vcpu *vcpu)
>>>>>>>> /* Enable LASX and restore context */
>>>>>>>> int kvm_own_lasx(struct kvm_vcpu *vcpu)
>>>>>>>> {
>>>>>>>> - preempt_disable();
>>>>>>>> -
>>>>>>>> kvm_check_fcsr(vcpu, vcpu->arch.fpu.fcsr);
>>>>>>>> set_csr_euen(CSR_EUEN_FPEN | CSR_EUEN_LSXEN | CSR_EUEN_LASXEN);
>>>>>>>> switch (vcpu->arch.aux_inuse & (KVM_LARCH_FPU | KVM_LARCH_LSX)) {
>>>>>>>> @@ -1411,7 +1425,6 @@ int kvm_own_lasx(struct kvm_vcpu *vcpu)
>>>>>>>>
>>>>>>>> trace_kvm_aux(vcpu, KVM_TRACE_AUX_RESTORE, KVM_TRACE_AUX_LASX);
>>>>>>>> vcpu->arch.aux_inuse |= KVM_LARCH_LASX | KVM_LARCH_LSX | KVM_LARCH_FPU;
>>>>>>>> - preempt_enable();
>>>>>>>>
>>>>>>>> return 0;
>>>>>>>> }
>>>>>>>> --
>>>>>>>> 2.39.3
>>>>>>>>
>>>>>>>>
>>>>>>
>>>>>>
>>>>
>>>>
>>
>>
Powered by blists - more mailing lists