| Message ID | 1551497728-12576-1-git-send-email-yuzenghui@huawei.com (mailing list archive) |
|---|---|
| State | RFC |
| Headers | show |
| Series | [RFC] KVM: arm64: Force a PTE mapping when logging is enabled | expand |
I think there're still some problems in this patch... Details below. On Sat, Mar 2, 2019 at 11:39 AM Zenghui Yu <yuzenghui@huawei.com> wrote: > > The idea behind this is: we don't want to keep tracking of huge pages when > logging_active is true, which will result in performance degradation. We > still need to set vma_pagesize to PAGE_SIZE, so that we can make use of it > to force a PTE mapping. > > Cc: Suzuki K Poulose <suzuki.poulose@arm.com> > Cc: Punit Agrawal <punit.agrawal@arm.com> > Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> > > --- > Atfer looking into https://patchwork.codeaurora.org/patch/647985/ , the > "vma_pagesize = PAGE_SIZE" logic was not intended to be deleted. As far > as I can tell, we used to have "hugetlb" to force the PTE mapping, but > we have "vma_pagesize" currently instead. We should set it properly for > performance reasons (e.g, in VM migration). Did I miss something important? > > --- > virt/kvm/arm/mmu.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c > index 30251e2..7d41b16 100644 > --- a/virt/kvm/arm/mmu.c > +++ b/virt/kvm/arm/mmu.c > @@ -1705,6 +1705,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, > (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && > !force_pte) { > gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; > + } else { > + /* > + * Fallback to PTE if it's not one of the stage2 > + * supported hugepage sizes or the corresponding level > + * doesn't exist, or logging is enabled. First, Instead of "logging is enabled", it should be "force_pte is true", since "force_pte" will be true when: 1) fault_supports_stage2_pmd_mappings() return false; or 2) "logging is enabled" (e.g, in VM migration). Second, fallback some unsupported hugepage sizes (e.g, 64K hugepage with 4K pages) to PTE is somewhat strange. And it will then _unexpectedly_ reach transparent_hugepage_adjust(), though no real adjustment will happen since commit fd2ef358282c ("KVM: arm/arm64: Ensure only THP is candidate for adjustment"). Keeping "vma_pagesize" there as it is will be better, right? So I'd just simplify the logic like: } else if (force_pte) { vma_pagesize = PAGE_SIZE; } Will send a V2 later and waiting for your comments :) thanks, zenghui > + */ > + vma_pagesize = PAGE_SIZE; > } > up_read(¤t->mm->mmap_sem); > > -- > 1.8.3.1 > > > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
Hi Zenghui, On Sun, Mar 03, 2019 at 11:14:38PM +0800, Zenghui Yu wrote: > I think there're still some problems in this patch... Details below. > > On Sat, Mar 2, 2019 at 11:39 AM Zenghui Yu <yuzenghui@huawei.com> wrote: > > > > The idea behind this is: we don't want to keep tracking of huge pages when > > logging_active is true, which will result in performance degradation. We > > still need to set vma_pagesize to PAGE_SIZE, so that we can make use of it > > to force a PTE mapping. Yes, you're right. We are indeed ignoring the force_pte flag. > > > > Cc: Suzuki K Poulose <suzuki.poulose@arm.com> > > Cc: Punit Agrawal <punit.agrawal@arm.com> > > Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> > > > > --- > > Atfer looking into https://patchwork.codeaurora.org/patch/647985/ , the > > "vma_pagesize = PAGE_SIZE" logic was not intended to be deleted. As far > > as I can tell, we used to have "hugetlb" to force the PTE mapping, but > > we have "vma_pagesize" currently instead. We should set it properly for > > performance reasons (e.g, in VM migration). Did I miss something important? > > > > --- > > virt/kvm/arm/mmu.c | 7 +++++++ > > 1 file changed, 7 insertions(+) > > > > diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c > > index 30251e2..7d41b16 100644 > > --- a/virt/kvm/arm/mmu.c > > +++ b/virt/kvm/arm/mmu.c > > @@ -1705,6 +1705,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, > > (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && > > !force_pte) { > > gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; > > + } else { > > + /* > > + * Fallback to PTE if it's not one of the stage2 > > + * supported hugepage sizes or the corresponding level > > + * doesn't exist, or logging is enabled. > > First, Instead of "logging is enabled", it should be "force_pte is true", > since "force_pte" will be true when: > > 1) fault_supports_stage2_pmd_mappings() return false; or > 2) "logging is enabled" (e.g, in VM migration). > > Second, fallback some unsupported hugepage sizes (e.g, 64K hugepage with > 4K pages) to PTE is somewhat strange. And it will then _unexpectedly_ > reach transparent_hugepage_adjust(), though no real adjustment will happen > since commit fd2ef358282c ("KVM: arm/arm64: Ensure only THP is candidate > for adjustment"). Keeping "vma_pagesize" there as it is will be better, > right? > > So I'd just simplify the logic like: We could fix this right in the beginning. See patch below: > > } else if (force_pte) { > vma_pagesize = PAGE_SIZE; > } > > > Will send a V2 later and waiting for your comments :) diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c index 30251e2..529331e 100644 --- a/virt/kvm/arm/mmu.c +++ b/virt/kvm/arm/mmu.c @@ -1693,7 +1693,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, return -EFAULT; } - vma_pagesize = vma_kernel_pagesize(vma); + /* If we are forced to map at page granularity, force the pagesize here */ + vma_pagesize = force_pte ? PAGE_SIZE : vma_kernel_pagesize(vma); + /* * The stage2 has a minimum of 2 level table (For arm64 see * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can @@ -1701,11 +1703,10 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, * As for PUD huge maps, we must make sure that we have at least * 3 levels, i.e, PMD is not folded. */ - if ((vma_pagesize == PMD_SIZE || - (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && - !force_pte) { + if (vma_pagesize == PMD_SIZE || + (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; - } + up_read(¤t->mm->mmap_sem); /* We need minimum second+third level pages */ --- Suzuki
Hi Zenghui, Suzuki, On 04/03/2019 17:13, Suzuki K Poulose wrote: > Hi Zenghui, > > On Sun, Mar 03, 2019 at 11:14:38PM +0800, Zenghui Yu wrote: >> I think there're still some problems in this patch... Details below. >> >> On Sat, Mar 2, 2019 at 11:39 AM Zenghui Yu <yuzenghui@huawei.com> wrote: >>> >>> The idea behind this is: we don't want to keep tracking of huge pages when >>> logging_active is true, which will result in performance degradation. We >>> still need to set vma_pagesize to PAGE_SIZE, so that we can make use of it >>> to force a PTE mapping. > > Yes, you're right. We are indeed ignoring the force_pte flag. > >>> >>> Cc: Suzuki K Poulose <suzuki.poulose@arm.com> >>> Cc: Punit Agrawal <punit.agrawal@arm.com> >>> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> >>> >>> --- >>> Atfer looking into https://patchwork.codeaurora.org/patch/647985/ , the >>> "vma_pagesize = PAGE_SIZE" logic was not intended to be deleted. As far >>> as I can tell, we used to have "hugetlb" to force the PTE mapping, but >>> we have "vma_pagesize" currently instead. We should set it properly for >>> performance reasons (e.g, in VM migration). Did I miss something important? >>> >>> --- >>> virt/kvm/arm/mmu.c | 7 +++++++ >>> 1 file changed, 7 insertions(+) >>> >>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >>> index 30251e2..7d41b16 100644 >>> --- a/virt/kvm/arm/mmu.c >>> +++ b/virt/kvm/arm/mmu.c >>> @@ -1705,6 +1705,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >>> (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && >>> !force_pte) { >>> gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; >>> + } else { >>> + /* >>> + * Fallback to PTE if it's not one of the stage2 >>> + * supported hugepage sizes or the corresponding level >>> + * doesn't exist, or logging is enabled. >> >> First, Instead of "logging is enabled", it should be "force_pte is true", >> since "force_pte" will be true when: >> >> 1) fault_supports_stage2_pmd_mappings() return false; or >> 2) "logging is enabled" (e.g, in VM migration). >> >> Second, fallback some unsupported hugepage sizes (e.g, 64K hugepage with >> 4K pages) to PTE is somewhat strange. And it will then _unexpectedly_ >> reach transparent_hugepage_adjust(), though no real adjustment will happen >> since commit fd2ef358282c ("KVM: arm/arm64: Ensure only THP is candidate >> for adjustment"). Keeping "vma_pagesize" there as it is will be better, >> right? >> >> So I'd just simplify the logic like: > > We could fix this right in the beginning. See patch below: > >> >> } else if (force_pte) { >> vma_pagesize = PAGE_SIZE; >> } >> >> >> Will send a V2 later and waiting for your comments :) > > > diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c > index 30251e2..529331e 100644 > --- a/virt/kvm/arm/mmu.c > +++ b/virt/kvm/arm/mmu.c > @@ -1693,7 +1693,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, > return -EFAULT; > } > > - vma_pagesize = vma_kernel_pagesize(vma); > + /* If we are forced to map at page granularity, force the pagesize here */ > + vma_pagesize = force_pte ? PAGE_SIZE : vma_kernel_pagesize(vma); > + > /* > * The stage2 has a minimum of 2 level table (For arm64 see > * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can > @@ -1701,11 +1703,10 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, > * As for PUD huge maps, we must make sure that we have at least > * 3 levels, i.e, PMD is not folded. > */ > - if ((vma_pagesize == PMD_SIZE || > - (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && > - !force_pte) { > + if (vma_pagesize == PMD_SIZE || > + (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) > gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; > - } > + > up_read(¤t->mm->mmap_sem); > > /* We need minimum second+third level pages */ That's pretty interesting, because this is almost what we already have in the NV code: https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/virt/kvm/arm/mmu.c?h=kvm-arm64/nv-wip-v5.0-rc7#n1752 (note that force_pte is gone in that branch). Thanks, M.
Hi Marc, Suzuki, On 2019/3/5 1:34, Marc Zyngier wrote: > Hi Zenghui, Suzuki, > > On 04/03/2019 17:13, Suzuki K Poulose wrote: >> Hi Zenghui, >> >> On Sun, Mar 03, 2019 at 11:14:38PM +0800, Zenghui Yu wrote: >>> I think there're still some problems in this patch... Details below. >>> >>> On Sat, Mar 2, 2019 at 11:39 AM Zenghui Yu <yuzenghui@huawei.com> wrote: >>>> >>>> The idea behind this is: we don't want to keep tracking of huge pages when >>>> logging_active is true, which will result in performance degradation. We >>>> still need to set vma_pagesize to PAGE_SIZE, so that we can make use of it >>>> to force a PTE mapping. >> >> Yes, you're right. We are indeed ignoring the force_pte flag. >> >>>> >>>> Cc: Suzuki K Poulose <suzuki.poulose@arm.com> >>>> Cc: Punit Agrawal <punit.agrawal@arm.com> >>>> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> >>>> >>>> --- >>>> Atfer looking into https://patchwork.codeaurora.org/patch/647985/ , the >>>> "vma_pagesize = PAGE_SIZE" logic was not intended to be deleted. As far >>>> as I can tell, we used to have "hugetlb" to force the PTE mapping, but >>>> we have "vma_pagesize" currently instead. We should set it properly for >>>> performance reasons (e.g, in VM migration). Did I miss something important? >>>> >>>> --- >>>> virt/kvm/arm/mmu.c | 7 +++++++ >>>> 1 file changed, 7 insertions(+) >>>> >>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >>>> index 30251e2..7d41b16 100644 >>>> --- a/virt/kvm/arm/mmu.c >>>> +++ b/virt/kvm/arm/mmu.c >>>> @@ -1705,6 +1705,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >>>> (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && >>>> !force_pte) { >>>> gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; >>>> + } else { >>>> + /* >>>> + * Fallback to PTE if it's not one of the stage2 >>>> + * supported hugepage sizes or the corresponding level >>>> + * doesn't exist, or logging is enabled. >>> >>> First, Instead of "logging is enabled", it should be "force_pte is true", >>> since "force_pte" will be true when: >>> >>> 1) fault_supports_stage2_pmd_mappings() return false; or >>> 2) "logging is enabled" (e.g, in VM migration). >>> >>> Second, fallback some unsupported hugepage sizes (e.g, 64K hugepage with >>> 4K pages) to PTE is somewhat strange. And it will then _unexpectedly_ >>> reach transparent_hugepage_adjust(), though no real adjustment will happen >>> since commit fd2ef358282c ("KVM: arm/arm64: Ensure only THP is candidate >>> for adjustment"). Keeping "vma_pagesize" there as it is will be better, >>> right? >>> >>> So I'd just simplify the logic like: >> >> We could fix this right in the beginning. See patch below: >> >>> >>> } else if (force_pte) { >>> vma_pagesize = PAGE_SIZE; >>> } >>> >>> >>> Will send a V2 later and waiting for your comments :) >> >> >> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >> index 30251e2..529331e 100644 >> --- a/virt/kvm/arm/mmu.c >> +++ b/virt/kvm/arm/mmu.c >> @@ -1693,7 +1693,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >> return -EFAULT; >> } >> >> - vma_pagesize = vma_kernel_pagesize(vma); >> + /* If we are forced to map at page granularity, force the pagesize here */ >> + vma_pagesize = force_pte ? PAGE_SIZE : vma_kernel_pagesize(vma); >> + >> /* >> * The stage2 has a minimum of 2 level table (For arm64 see >> * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can >> @@ -1701,11 +1703,10 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >> * As for PUD huge maps, we must make sure that we have at least >> * 3 levels, i.e, PMD is not folded. >> */ >> - if ((vma_pagesize == PMD_SIZE || >> - (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && >> - !force_pte) { >> + if (vma_pagesize == PMD_SIZE || >> + (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) >> gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; >> - } >> + >> up_read(¤t->mm->mmap_sem); >> >> /* We need minimum second+third level pages */ A nicer implementation and easier to understand, thanks! > That's pretty interesting, because this is almost what we already have > in the NV code: > > https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/virt/kvm/arm/mmu.c?h=kvm-arm64/nv-wip-v5.0-rc7#n1752 > > (note that force_pte is gone in that branch). haha :-) sorry about that. I haven't looked into the NV code yet, so ... But I'm still wondering: should we fix this wrong mapping size problem before NV is introduced? Since this problem has not much to do with NV, and 5.0 has already been released with this problem (and 5.1 will without fix ...). Just a personal idea, ignore it if unnecessary. thanks, zenghui
Hi Zenghui, On 05/03/2019 11:09, Zenghui Yu wrote: > Hi Marc, Suzuki, > > On 2019/3/5 1:34, Marc Zyngier wrote: >> Hi Zenghui, Suzuki, >> >> On 04/03/2019 17:13, Suzuki K Poulose wrote: >>> Hi Zenghui, >>> >>> On Sun, Mar 03, 2019 at 11:14:38PM +0800, Zenghui Yu wrote: >>>> I think there're still some problems in this patch... Details below. >>>> >>>> On Sat, Mar 2, 2019 at 11:39 AM Zenghui Yu <yuzenghui@huawei.com> wrote: >>>>> >>>>> The idea behind this is: we don't want to keep tracking of huge pages when >>>>> logging_active is true, which will result in performance degradation. We >>>>> still need to set vma_pagesize to PAGE_SIZE, so that we can make use of it >>>>> to force a PTE mapping. >>> >>> Yes, you're right. We are indeed ignoring the force_pte flag. >>> >>>>> >>>>> Cc: Suzuki K Poulose <suzuki.poulose@arm.com> >>>>> Cc: Punit Agrawal <punit.agrawal@arm.com> >>>>> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> >>>>> >>>>> --- >>>>> Atfer looking into https://patchwork.codeaurora.org/patch/647985/ , the >>>>> "vma_pagesize = PAGE_SIZE" logic was not intended to be deleted. As far >>>>> as I can tell, we used to have "hugetlb" to force the PTE mapping, but >>>>> we have "vma_pagesize" currently instead. We should set it properly for >>>>> performance reasons (e.g, in VM migration). Did I miss something important? >>>>> >>>>> --- >>>>> virt/kvm/arm/mmu.c | 7 +++++++ >>>>> 1 file changed, 7 insertions(+) >>>>> >>>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >>>>> index 30251e2..7d41b16 100644 >>>>> --- a/virt/kvm/arm/mmu.c >>>>> +++ b/virt/kvm/arm/mmu.c >>>>> @@ -1705,6 +1705,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >>>>> (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && >>>>> !force_pte) { >>>>> gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; >>>>> + } else { >>>>> + /* >>>>> + * Fallback to PTE if it's not one of the stage2 >>>>> + * supported hugepage sizes or the corresponding level >>>>> + * doesn't exist, or logging is enabled. >>>> >>>> First, Instead of "logging is enabled", it should be "force_pte is true", >>>> since "force_pte" will be true when: >>>> >>>> 1) fault_supports_stage2_pmd_mappings() return false; or >>>> 2) "logging is enabled" (e.g, in VM migration). >>>> >>>> Second, fallback some unsupported hugepage sizes (e.g, 64K hugepage with >>>> 4K pages) to PTE is somewhat strange. And it will then _unexpectedly_ >>>> reach transparent_hugepage_adjust(), though no real adjustment will happen >>>> since commit fd2ef358282c ("KVM: arm/arm64: Ensure only THP is candidate >>>> for adjustment"). Keeping "vma_pagesize" there as it is will be better, >>>> right? >>>> >>>> So I'd just simplify the logic like: >>> >>> We could fix this right in the beginning. See patch below: >>> >>>> >>>> } else if (force_pte) { >>>> vma_pagesize = PAGE_SIZE; >>>> } >>>> >>>> >>>> Will send a V2 later and waiting for your comments :) >>> >>> >>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >>> index 30251e2..529331e 100644 >>> --- a/virt/kvm/arm/mmu.c >>> +++ b/virt/kvm/arm/mmu.c >>> @@ -1693,7 +1693,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >>> return -EFAULT; >>> } >>> >>> - vma_pagesize = vma_kernel_pagesize(vma); >>> + /* If we are forced to map at page granularity, force the pagesize here */ >>> + vma_pagesize = force_pte ? PAGE_SIZE : vma_kernel_pagesize(vma); >>> + >>> /* >>> * The stage2 has a minimum of 2 level table (For arm64 see >>> * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can >>> @@ -1701,11 +1703,10 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >>> * As for PUD huge maps, we must make sure that we have at least >>> * 3 levels, i.e, PMD is not folded. >>> */ >>> - if ((vma_pagesize == PMD_SIZE || >>> - (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && >>> - !force_pte) { >>> + if (vma_pagesize == PMD_SIZE || >>> + (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) >>> gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; >>> - } >>> + >>> up_read(¤t->mm->mmap_sem); >>> >>> /* We need minimum second+third level pages */ > > A nicer implementation and easier to understand, thanks! > >> That's pretty interesting, because this is almost what we already have >> in the NV code: >> >> https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/virt/kvm/arm/mmu.c?h=kvm-arm64/nv-wip-v5.0-rc7#n1752 >> >> (note that force_pte is gone in that branch). > > haha :-) sorry about that. I haven't looked into the NV code yet, so ... > > But I'm still wondering: should we fix this wrong mapping size problem > before NV is introduced? Since this problem has not much to do with NV, > and 5.0 has already been released with this problem (and 5.1 will > without fix ...). Yes, we must fix it. I will soon send out a patch copying on it. Its just that I find some more issues around forcing the PTE mappings with PUD huge pages. I will send something out soon. Cheers Suzuki
On 2019/3/5 19:13, Suzuki K Poulose wrote: > Hi Zenghui, > > On 05/03/2019 11:09, Zenghui Yu wrote: >> Hi Marc, Suzuki, >> >> On 2019/3/5 1:34, Marc Zyngier wrote: >>> Hi Zenghui, Suzuki, >>> >>> On 04/03/2019 17:13, Suzuki K Poulose wrote: >>>> Hi Zenghui, >>>> >>>> On Sun, Mar 03, 2019 at 11:14:38PM +0800, Zenghui Yu wrote: >>>>> I think there're still some problems in this patch... Details below. >>>>> >>>>> On Sat, Mar 2, 2019 at 11:39 AM Zenghui Yu <yuzenghui@huawei.com> >>>>> wrote: >>>>>> >>>>>> The idea behind this is: we don't want to keep tracking of huge >>>>>> pages when >>>>>> logging_active is true, which will result in performance >>>>>> degradation. We >>>>>> still need to set vma_pagesize to PAGE_SIZE, so that we can make >>>>>> use of it >>>>>> to force a PTE mapping. >>>> >>>> Yes, you're right. We are indeed ignoring the force_pte flag. >>>> >>>>>> >>>>>> Cc: Suzuki K Poulose <suzuki.poulose@arm.com> >>>>>> Cc: Punit Agrawal <punit.agrawal@arm.com> >>>>>> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> >>>>>> >>>>>> --- >>>>>> Atfer looking into https://patchwork.codeaurora.org/patch/647985/ >>>>>> , the >>>>>> "vma_pagesize = PAGE_SIZE" logic was not intended to be deleted. >>>>>> As far >>>>>> as I can tell, we used to have "hugetlb" to force the PTE mapping, >>>>>> but >>>>>> we have "vma_pagesize" currently instead. We should set it >>>>>> properly for >>>>>> performance reasons (e.g, in VM migration). Did I miss something >>>>>> important? >>>>>> >>>>>> --- >>>>>> virt/kvm/arm/mmu.c | 7 +++++++ >>>>>> 1 file changed, 7 insertions(+) >>>>>> >>>>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >>>>>> index 30251e2..7d41b16 100644 >>>>>> --- a/virt/kvm/arm/mmu.c >>>>>> +++ b/virt/kvm/arm/mmu.c >>>>>> @@ -1705,6 +1705,13 @@ static int user_mem_abort(struct kvm_vcpu >>>>>> *vcpu, phys_addr_t fault_ipa, >>>>>> (vma_pagesize == PUD_SIZE && >>>>>> kvm_stage2_has_pmd(kvm))) && >>>>>> !force_pte) { >>>>>> gfn = (fault_ipa & >>>>>> huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; >>>>>> + } else { >>>>>> + /* >>>>>> + * Fallback to PTE if it's not one of the stage2 >>>>>> + * supported hugepage sizes or the corresponding >>>>>> level >>>>>> + * doesn't exist, or logging is enabled. >>>>> >>>>> First, Instead of "logging is enabled", it should be "force_pte is >>>>> true", >>>>> since "force_pte" will be true when: >>>>> >>>>> 1) fault_supports_stage2_pmd_mappings() return false; or >>>>> 2) "logging is enabled" (e.g, in VM migration). >>>>> >>>>> Second, fallback some unsupported hugepage sizes (e.g, 64K hugepage >>>>> with >>>>> 4K pages) to PTE is somewhat strange. And it will then _unexpectedly_ >>>>> reach transparent_hugepage_adjust(), though no real adjustment will >>>>> happen >>>>> since commit fd2ef358282c ("KVM: arm/arm64: Ensure only THP is >>>>> candidate >>>>> for adjustment"). Keeping "vma_pagesize" there as it is will be >>>>> better, >>>>> right? >>>>> >>>>> So I'd just simplify the logic like: >>>> >>>> We could fix this right in the beginning. See patch below: >>>> >>>>> >>>>> } else if (force_pte) { >>>>> vma_pagesize = PAGE_SIZE; >>>>> } >>>>> >>>>> >>>>> Will send a V2 later and waiting for your comments :) >>>> >>>> >>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >>>> index 30251e2..529331e 100644 >>>> --- a/virt/kvm/arm/mmu.c >>>> +++ b/virt/kvm/arm/mmu.c >>>> @@ -1693,7 +1693,9 @@ static int user_mem_abort(struct kvm_vcpu >>>> *vcpu, phys_addr_t fault_ipa, >>>> return -EFAULT; >>>> } >>>> - vma_pagesize = vma_kernel_pagesize(vma); >>>> + /* If we are forced to map at page granularity, force the >>>> pagesize here */ >>>> + vma_pagesize = force_pte ? PAGE_SIZE : vma_kernel_pagesize(vma); >>>> + >>>> /* >>>> * The stage2 has a minimum of 2 level table (For arm64 see >>>> * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can >>>> @@ -1701,11 +1703,10 @@ static int user_mem_abort(struct kvm_vcpu >>>> *vcpu, phys_addr_t fault_ipa, >>>> * As for PUD huge maps, we must make sure that we have at least >>>> * 3 levels, i.e, PMD is not folded. >>>> */ >>>> - if ((vma_pagesize == PMD_SIZE || >>>> - (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && >>>> - !force_pte) { >>>> + if (vma_pagesize == PMD_SIZE || >>>> + (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) >>>> gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> >>>> PAGE_SHIFT; >>>> - } >>>> + >>>> up_read(¤t->mm->mmap_sem); >>>> /* We need minimum second+third level pages */ >> >> A nicer implementation and easier to understand, thanks! >> >>> That's pretty interesting, because this is almost what we already have >>> in the NV code: >>> >>> https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/virt/kvm/arm/mmu.c?h=kvm-arm64/nv-wip-v5.0-rc7#n1752 >>> >>> >>> (note that force_pte is gone in that branch). >> >> haha :-) sorry about that. I haven't looked into the NV code yet, so ... >> >> But I'm still wondering: should we fix this wrong mapping size problem >> before NV is introduced? Since this problem has not much to do with NV, >> and 5.0 has already been released with this problem (and 5.1 will >> without fix ...). > > Yes, we must fix it. I will soon send out a patch copying on it. > Its just that I find some more issues around forcing the PTE > mappings with PUD huge pages. I will send something out soon. Sounds good! zenghui
On 05/03/2019 11:09, Zenghui Yu wrote: > Hi Marc, Suzuki, > > On 2019/3/5 1:34, Marc Zyngier wrote: >> Hi Zenghui, Suzuki, >> >> On 04/03/2019 17:13, Suzuki K Poulose wrote: >>> Hi Zenghui, >>> >>> On Sun, Mar 03, 2019 at 11:14:38PM +0800, Zenghui Yu wrote: >>>> I think there're still some problems in this patch... Details below. >>>> >>>> On Sat, Mar 2, 2019 at 11:39 AM Zenghui Yu <yuzenghui@huawei.com> wrote: >>>>> >>>>> The idea behind this is: we don't want to keep tracking of huge pages when >>>>> logging_active is true, which will result in performance degradation. We >>>>> still need to set vma_pagesize to PAGE_SIZE, so that we can make use of it >>>>> to force a PTE mapping. >>> >>> Yes, you're right. We are indeed ignoring the force_pte flag. >>> >>>>> >>>>> Cc: Suzuki K Poulose <suzuki.poulose@arm.com> >>>>> Cc: Punit Agrawal <punit.agrawal@arm.com> >>>>> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> >>>>> >>>>> --- >>>>> Atfer looking into https://patchwork.codeaurora.org/patch/647985/ , the >>>>> "vma_pagesize = PAGE_SIZE" logic was not intended to be deleted. As far >>>>> as I can tell, we used to have "hugetlb" to force the PTE mapping, but >>>>> we have "vma_pagesize" currently instead. We should set it properly for >>>>> performance reasons (e.g, in VM migration). Did I miss something important? >>>>> >>>>> --- >>>>> virt/kvm/arm/mmu.c | 7 +++++++ >>>>> 1 file changed, 7 insertions(+) >>>>> >>>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >>>>> index 30251e2..7d41b16 100644 >>>>> --- a/virt/kvm/arm/mmu.c >>>>> +++ b/virt/kvm/arm/mmu.c >>>>> @@ -1705,6 +1705,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >>>>> (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && >>>>> !force_pte) { >>>>> gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; >>>>> + } else { >>>>> + /* >>>>> + * Fallback to PTE if it's not one of the stage2 >>>>> + * supported hugepage sizes or the corresponding level >>>>> + * doesn't exist, or logging is enabled. >>>> >>>> First, Instead of "logging is enabled", it should be "force_pte is true", >>>> since "force_pte" will be true when: >>>> >>>> 1) fault_supports_stage2_pmd_mappings() return false; or >>>> 2) "logging is enabled" (e.g, in VM migration). >>>> >>>> Second, fallback some unsupported hugepage sizes (e.g, 64K hugepage with >>>> 4K pages) to PTE is somewhat strange. And it will then _unexpectedly_ >>>> reach transparent_hugepage_adjust(), though no real adjustment will happen >>>> since commit fd2ef358282c ("KVM: arm/arm64: Ensure only THP is candidate >>>> for adjustment"). Keeping "vma_pagesize" there as it is will be better, >>>> right? >>>> >>>> So I'd just simplify the logic like: >>> >>> We could fix this right in the beginning. See patch below: >>> >>>> >>>> } else if (force_pte) { >>>> vma_pagesize = PAGE_SIZE; >>>> } >>>> >>>> >>>> Will send a V2 later and waiting for your comments :) >>> >>> >>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c >>> index 30251e2..529331e 100644 >>> --- a/virt/kvm/arm/mmu.c >>> +++ b/virt/kvm/arm/mmu.c >>> @@ -1693,7 +1693,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >>> return -EFAULT; >>> } >>> >>> - vma_pagesize = vma_kernel_pagesize(vma); >>> + /* If we are forced to map at page granularity, force the pagesize here */ >>> + vma_pagesize = force_pte ? PAGE_SIZE : vma_kernel_pagesize(vma); >>> + >>> /* >>> * The stage2 has a minimum of 2 level table (For arm64 see >>> * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can >>> @@ -1701,11 +1703,10 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, >>> * As for PUD huge maps, we must make sure that we have at least >>> * 3 levels, i.e, PMD is not folded. >>> */ >>> - if ((vma_pagesize == PMD_SIZE || >>> - (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && >>> - !force_pte) { >>> + if (vma_pagesize == PMD_SIZE || >>> + (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) >>> gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; >>> - } >>> + >>> up_read(¤t->mm->mmap_sem); >>> >>> /* We need minimum second+third level pages */ > > A nicer implementation and easier to understand, thanks! > >> That's pretty interesting, because this is almost what we already have >> in the NV code: >> >> https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/tree/virt/kvm/arm/mmu.c?h=kvm-arm64/nv-wip-v5.0-rc7#n1752 >> >> (note that force_pte is gone in that branch). > > haha :-) sorry about that. I haven't looked into the NV code yet, so ... > > But I'm still wondering: should we fix this wrong mapping size problem > before NV is introduced? Since this problem has not much to do with NV, > and 5.0 has already been released with this problem (and 5.1 will > without fix ...). > > Just a personal idea, ignore it if unnecessary. We definitely want to fix it now, and have the fix backported to older versions. I can always rebase the NV branch on top of the current state of mainline. Thanks, M.
diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c index 30251e2..7d41b16 100644 --- a/virt/kvm/arm/mmu.c +++ b/virt/kvm/arm/mmu.c @@ -1705,6 +1705,13 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa, (vma_pagesize == PUD_SIZE && kvm_stage2_has_pmd(kvm))) && !force_pte) { gfn = (fault_ipa & huge_page_mask(hstate_vma(vma))) >> PAGE_SHIFT; + } else { + /* + * Fallback to PTE if it's not one of the stage2 + * supported hugepage sizes or the corresponding level + * doesn't exist, or logging is enabled. + */ + vma_pagesize = PAGE_SIZE; } up_read(¤t->mm->mmap_sem);
The idea behind this is: we don't want to keep tracking of huge pages when logging_active is true, which will result in performance degradation. We still need to set vma_pagesize to PAGE_SIZE, so that we can make use of it to force a PTE mapping. Cc: Suzuki K Poulose <suzuki.poulose@arm.com> Cc: Punit Agrawal <punit.agrawal@arm.com> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> --- Atfer looking into https://patchwork.codeaurora.org/patch/647985/ , the "vma_pagesize = PAGE_SIZE" logic was not intended to be deleted. As far as I can tell, we used to have "hugetlb" to force the PTE mapping, but we have "vma_pagesize" currently instead. We should set it properly for performance reasons (e.g, in VM migration). Did I miss something important? --- virt/kvm/arm/mmu.c | 7 +++++++ 1 file changed, 7 insertions(+)