| Message ID | 20171103055327.21684-4-haozhong.zhang@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 11/03/2017 01:53 PM, Haozhong Zhang wrote: > Some reserved pages, such as those from NVDIMM DAX devices, are > not for MMIO, and can be mapped with cached memory type for better > performance. However, the above check misconceives those pages as > MMIO. Because KVM maps MMIO pages with UC memory type, the > performance of guest accesses to those pages would be harmed. > Therefore, we check the host memory type by lookup_memtype() in > addition and only treat UC/UC- pages as MMIO. > > Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> > Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> > Reported-by: Kumar, Karthik <karthik.kumar@intel.com> > --- > arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- > 1 file changed, 18 insertions(+), 1 deletion(-) > > diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c > index 0b481cc9c725..e9ed0e666a83 100644 > --- a/arch/x86/kvm/mmu.c > +++ b/arch/x86/kvm/mmu.c > @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, > static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) > { > if (pfn_valid(pfn)) > - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); > + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && > + /* > + * Some reserved pages, such as those from > + * NVDIMM DAX devices, are not for MMIO, and > + * can be mapped with cached memory type for > + * better performance. However, the above > + * check misconceives those pages as MMIO. > + * Because KVM maps MMIO pages with UC memory > + * type, the performance of guest accesses to > + * those pages would be harmed. Therefore, we > + * check the host memory type in addition and > + * only treat UC/UC- pages as MMIO. > + * > + * pat_pfn_is_uc() works only when PAT is enabled, > + * so check pat_enabled() as well. > + */ > + (!pat_enabled() || > + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); Can it be compiled if !CONFIG_PAT? It would be better if we move pat_enabled out of kvm as well, please refer to pgprot_writecombine() which is implemented in pat.c and in include\asm-generic\pgtable.h: #ifndef pgprot_writecombine #define pgprot_writecombine pgprot_noncached #endif
On 11/03/17 14:54 +0800, Xiao Guangrong wrote: > > > On 11/03/2017 01:53 PM, Haozhong Zhang wrote: > > Some reserved pages, such as those from NVDIMM DAX devices, are > > not for MMIO, and can be mapped with cached memory type for better > > performance. However, the above check misconceives those pages as > > MMIO. Because KVM maps MMIO pages with UC memory type, the > > performance of guest accesses to those pages would be harmed. > > Therefore, we check the host memory type by lookup_memtype() in > > addition and only treat UC/UC- pages as MMIO. > > > > Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> > > Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> > > Reported-by: Kumar, Karthik <karthik.kumar@intel.com> > > --- > > arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- > > 1 file changed, 18 insertions(+), 1 deletion(-) > > > > diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c > > index 0b481cc9c725..e9ed0e666a83 100644 > > --- a/arch/x86/kvm/mmu.c > > +++ b/arch/x86/kvm/mmu.c > > @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, > > static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) > > { > > if (pfn_valid(pfn)) > > - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); > > + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && > > + /* > > + * Some reserved pages, such as those from > > + * NVDIMM DAX devices, are not for MMIO, and > > + * can be mapped with cached memory type for > > + * better performance. However, the above > > + * check misconceives those pages as MMIO. > > + * Because KVM maps MMIO pages with UC memory > > + * type, the performance of guest accesses to > > + * those pages would be harmed. Therefore, we > > + * check the host memory type in addition and > > + * only treat UC/UC- pages as MMIO. > > + * > > + * pat_pfn_is_uc() works only when PAT is enabled, > > + * so check pat_enabled() as well. > > + */ > > + (!pat_enabled() || > > + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); > > Can it be compiled if !CONFIG_PAT? Yes. What I check via pat_enabled() is not only whether PAT support is compiled, but also whether PAT is enabled at runtime. > > It would be better if we move pat_enabled out of kvm as well, Surely I can combine them in one function like bool pat_pfn_is_uc(pfn_t pfn) { enum page_cache_mode cm; if (!pat_enabled()) return false; cm = lookup_memtype(pfn_t_to_phys(pfn)); return cm == _PAGE_CACHE_MODE_UC || cm == _PAGE_CACHE_MODE_UC_MINUS; } but I need a good name to make its semantics clear, or is it enough to just leave a comment like? /* * Check via PAT whether the cache mode of a page if UC or UC-. * * Returns true, if PAT is enabled and the cache mode is UC or UC-. * Returns false otherwise. */ > please refer > to pgprot_writecombine() which is implemented in pat.c and in > include\asm-generic\pgtable.h: > > #ifndef pgprot_writecombine > #define pgprot_writecombine pgprot_noncached > #endif >
On 11/03/17 16:51 +0800, Haozhong Zhang wrote: > On 11/03/17 14:54 +0800, Xiao Guangrong wrote: > > > > > > On 11/03/2017 01:53 PM, Haozhong Zhang wrote: > > > Some reserved pages, such as those from NVDIMM DAX devices, are > > > not for MMIO, and can be mapped with cached memory type for better > > > performance. However, the above check misconceives those pages as > > > MMIO. Because KVM maps MMIO pages with UC memory type, the > > > performance of guest accesses to those pages would be harmed. > > > Therefore, we check the host memory type by lookup_memtype() in > > > addition and only treat UC/UC- pages as MMIO. > > > > > > Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> > > > Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> > > > Reported-by: Kumar, Karthik <karthik.kumar@intel.com> > > > --- > > > arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- > > > 1 file changed, 18 insertions(+), 1 deletion(-) > > > > > > diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c > > > index 0b481cc9c725..e9ed0e666a83 100644 > > > --- a/arch/x86/kvm/mmu.c > > > +++ b/arch/x86/kvm/mmu.c > > > @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, > > > static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) > > > { > > > if (pfn_valid(pfn)) > > > - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); > > > + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && > > > + /* > > > + * Some reserved pages, such as those from > > > + * NVDIMM DAX devices, are not for MMIO, and > > > + * can be mapped with cached memory type for > > > + * better performance. However, the above > > > + * check misconceives those pages as MMIO. > > > + * Because KVM maps MMIO pages with UC memory > > > + * type, the performance of guest accesses to > > > + * those pages would be harmed. Therefore, we > > > + * check the host memory type in addition and > > > + * only treat UC/UC- pages as MMIO. > > > + * > > > + * pat_pfn_is_uc() works only when PAT is enabled, > > > + * so check pat_enabled() as well. > > > + */ > > > + (!pat_enabled() || > > > + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); > > > > Can it be compiled if !CONFIG_PAT? > > Yes. > > What I check via pat_enabled() is not only whether PAT support is > compiled, but also whether PAT is enabled at runtime. > > > > > It would be better if we move pat_enabled out of kvm as well, > > Surely I can combine them in one function like > > bool pat_pfn_is_uc(pfn_t pfn) > { > enum page_cache_mode cm; > > if (!pat_enabled()) > return false; > > cm = lookup_memtype(pfn_t_to_phys(pfn)); > > return cm == _PAGE_CACHE_MODE_UC || cm == _PAGE_CACHE_MODE_UC_MINUS; > } In addition, I think it's better to split this function into pat_pfn_is_uc() and pat_pfn_is_uc_minus() to avoid additional confusion. Haozhong > > but I need a good name to make its semantics clear, or is it enough to > just leave a comment like? > > /* > * Check via PAT whether the cache mode of a page if UC or UC-. > * > * Returns true, if PAT is enabled and the cache mode is UC or UC-. > * Returns false otherwise. > */ > > > > please refer > > to pgprot_writecombine() which is implemented in pat.c and in > > include\asm-generic\pgtable.h: > > > > #ifndef pgprot_writecombine > > #define pgprot_writecombine pgprot_noncached > > #endif > > > > >
On 11/03/2017 04:51 PM, Haozhong Zhang wrote: > On 11/03/17 14:54 +0800, Xiao Guangrong wrote: >> >> >> On 11/03/2017 01:53 PM, Haozhong Zhang wrote: >>> Some reserved pages, such as those from NVDIMM DAX devices, are >>> not for MMIO, and can be mapped with cached memory type for better >>> performance. However, the above check misconceives those pages as >>> MMIO. Because KVM maps MMIO pages with UC memory type, the >>> performance of guest accesses to those pages would be harmed. >>> Therefore, we check the host memory type by lookup_memtype() in >>> addition and only treat UC/UC- pages as MMIO. >>> >>> Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> >>> Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> >>> Reported-by: Kumar, Karthik <karthik.kumar@intel.com> >>> --- >>> arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- >>> 1 file changed, 18 insertions(+), 1 deletion(-) >>> >>> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c >>> index 0b481cc9c725..e9ed0e666a83 100644 >>> --- a/arch/x86/kvm/mmu.c >>> +++ b/arch/x86/kvm/mmu.c >>> @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, >>> static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) >>> { >>> if (pfn_valid(pfn)) >>> - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); >>> + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && >>> + /* >>> + * Some reserved pages, such as those from >>> + * NVDIMM DAX devices, are not for MMIO, and >>> + * can be mapped with cached memory type for >>> + * better performance. However, the above >>> + * check misconceives those pages as MMIO. >>> + * Because KVM maps MMIO pages with UC memory >>> + * type, the performance of guest accesses to >>> + * those pages would be harmed. Therefore, we >>> + * check the host memory type in addition and >>> + * only treat UC/UC- pages as MMIO. >>> + * >>> + * pat_pfn_is_uc() works only when PAT is enabled, >>> + * so check pat_enabled() as well. >>> + */ >>> + (!pat_enabled() || >>> + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); >> >> Can it be compiled if !CONFIG_PAT? > > Yes. > > What I check via pat_enabled() is not only whether PAT support is > compiled, but also whether PAT is enabled at runtime. The issue is about pat_pfn_is_uc() which is implemented only if CONFIG_PAT is enabled, but you used it here unconditionally. I am not sure if gcc is smart enough to omit pat_pfn_is_uc() completely under this case. If you really have done the test to compile kernel and KVM module with CONFIG_PAT disabled, it is fine.
On 11/03/2017 05:02 PM, Haozhong Zhang wrote: > On 11/03/17 16:51 +0800, Haozhong Zhang wrote: >> On 11/03/17 14:54 +0800, Xiao Guangrong wrote: >>> >>> >>> On 11/03/2017 01:53 PM, Haozhong Zhang wrote: >>>> Some reserved pages, such as those from NVDIMM DAX devices, are >>>> not for MMIO, and can be mapped with cached memory type for better >>>> performance. However, the above check misconceives those pages as >>>> MMIO. Because KVM maps MMIO pages with UC memory type, the >>>> performance of guest accesses to those pages would be harmed. >>>> Therefore, we check the host memory type by lookup_memtype() in >>>> addition and only treat UC/UC- pages as MMIO. >>>> >>>> Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> >>>> Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> >>>> Reported-by: Kumar, Karthik <karthik.kumar@intel.com> >>>> --- >>>> arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- >>>> 1 file changed, 18 insertions(+), 1 deletion(-) >>>> >>>> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c >>>> index 0b481cc9c725..e9ed0e666a83 100644 >>>> --- a/arch/x86/kvm/mmu.c >>>> +++ b/arch/x86/kvm/mmu.c >>>> @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, >>>> static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) >>>> { >>>> if (pfn_valid(pfn)) >>>> - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); >>>> + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && >>>> + /* >>>> + * Some reserved pages, such as those from >>>> + * NVDIMM DAX devices, are not for MMIO, and >>>> + * can be mapped with cached memory type for >>>> + * better performance. However, the above >>>> + * check misconceives those pages as MMIO. >>>> + * Because KVM maps MMIO pages with UC memory >>>> + * type, the performance of guest accesses to >>>> + * those pages would be harmed. Therefore, we >>>> + * check the host memory type in addition and >>>> + * only treat UC/UC- pages as MMIO. >>>> + * >>>> + * pat_pfn_is_uc() works only when PAT is enabled, >>>> + * so check pat_enabled() as well. >>>> + */ >>>> + (!pat_enabled() || >>>> + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); >>> >>> Can it be compiled if !CONFIG_PAT? >> >> Yes. >> >> What I check via pat_enabled() is not only whether PAT support is >> compiled, but also whether PAT is enabled at runtime. >> >>> >>> It would be better if we move pat_enabled out of kvm as well, >> >> Surely I can combine them in one function like >> >> bool pat_pfn_is_uc(pfn_t pfn) >> { >> enum page_cache_mode cm; >> >> if (!pat_enabled()) >> return false; >> >> cm = lookup_memtype(pfn_t_to_phys(pfn)); >> >> return cm == _PAGE_CACHE_MODE_UC || cm == _PAGE_CACHE_MODE_UC_MINUS; >> } > > In addition, I think it's better to split this function into > pat_pfn_is_uc() and pat_pfn_is_uc_minus() to avoid additional > confusion. Why not use pat_pfn_is_uc_or_uc_minus(). :)
On 11/03/17 17:10 +0800, Xiao Guangrong wrote: > > > On 11/03/2017 04:51 PM, Haozhong Zhang wrote: > > On 11/03/17 14:54 +0800, Xiao Guangrong wrote: > > > > > > > > > On 11/03/2017 01:53 PM, Haozhong Zhang wrote: > > > > Some reserved pages, such as those from NVDIMM DAX devices, are > > > > not for MMIO, and can be mapped with cached memory type for better > > > > performance. However, the above check misconceives those pages as > > > > MMIO. Because KVM maps MMIO pages with UC memory type, the > > > > performance of guest accesses to those pages would be harmed. > > > > Therefore, we check the host memory type by lookup_memtype() in > > > > addition and only treat UC/UC- pages as MMIO. > > > > > > > > Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> > > > > Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> > > > > Reported-by: Kumar, Karthik <karthik.kumar@intel.com> > > > > --- > > > > arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- > > > > 1 file changed, 18 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c > > > > index 0b481cc9c725..e9ed0e666a83 100644 > > > > --- a/arch/x86/kvm/mmu.c > > > > +++ b/arch/x86/kvm/mmu.c > > > > @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, > > > > static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) > > > > { > > > > if (pfn_valid(pfn)) > > > > - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); > > > > + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && > > > > + /* > > > > + * Some reserved pages, such as those from > > > > + * NVDIMM DAX devices, are not for MMIO, and > > > > + * can be mapped with cached memory type for > > > > + * better performance. However, the above > > > > + * check misconceives those pages as MMIO. > > > > + * Because KVM maps MMIO pages with UC memory > > > > + * type, the performance of guest accesses to > > > > + * those pages would be harmed. Therefore, we > > > > + * check the host memory type in addition and > > > > + * only treat UC/UC- pages as MMIO. > > > > + * > > > > + * pat_pfn_is_uc() works only when PAT is enabled, > > > > + * so check pat_enabled() as well. > > > > + */ > > > > + (!pat_enabled() || > > > > + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); > > > > > > Can it be compiled if !CONFIG_PAT? > > > > Yes. > > > > What I check via pat_enabled() is not only whether PAT support is > > compiled, but also whether PAT is enabled at runtime. > > The issue is about pat_pfn_is_uc() which is implemented only if CONFIG_PAT is > enabled, but you used it here unconditionally. > > I am not sure if gcc is smart enough to omit pat_pfn_is_uc() completely under > this case. If you really have done the test to compile kernel and KVM module > with CONFIG_PAT disabled, it is fine. > I've done the test and it can compile. arch/x86/mm/Makefile shows pat.c is compiled regardless of CONFIG_X86_PAT, and pat_pfn_is_uc() is defined out of #ifdef CONFIG_X86_PAT ... #endif. Haozhong
On 11/03/17 17:24 +0800, Xiao Guangrong wrote: > > > On 11/03/2017 05:02 PM, Haozhong Zhang wrote: > > On 11/03/17 16:51 +0800, Haozhong Zhang wrote: > > > On 11/03/17 14:54 +0800, Xiao Guangrong wrote: > > > > > > > > > > > > On 11/03/2017 01:53 PM, Haozhong Zhang wrote: > > > > > Some reserved pages, such as those from NVDIMM DAX devices, are > > > > > not for MMIO, and can be mapped with cached memory type for better > > > > > performance. However, the above check misconceives those pages as > > > > > MMIO. Because KVM maps MMIO pages with UC memory type, the > > > > > performance of guest accesses to those pages would be harmed. > > > > > Therefore, we check the host memory type by lookup_memtype() in > > > > > addition and only treat UC/UC- pages as MMIO. > > > > > > > > > > Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> > > > > > Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> > > > > > Reported-by: Kumar, Karthik <karthik.kumar@intel.com> > > > > > --- > > > > > arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- > > > > > 1 file changed, 18 insertions(+), 1 deletion(-) > > > > > > > > > > diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c > > > > > index 0b481cc9c725..e9ed0e666a83 100644 > > > > > --- a/arch/x86/kvm/mmu.c > > > > > +++ b/arch/x86/kvm/mmu.c > > > > > @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, > > > > > static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) > > > > > { > > > > > if (pfn_valid(pfn)) > > > > > - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); > > > > > + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && > > > > > + /* > > > > > + * Some reserved pages, such as those from > > > > > + * NVDIMM DAX devices, are not for MMIO, and > > > > > + * can be mapped with cached memory type for > > > > > + * better performance. However, the above > > > > > + * check misconceives those pages as MMIO. > > > > > + * Because KVM maps MMIO pages with UC memory > > > > > + * type, the performance of guest accesses to > > > > > + * those pages would be harmed. Therefore, we > > > > > + * check the host memory type in addition and > > > > > + * only treat UC/UC- pages as MMIO. > > > > > + * > > > > > + * pat_pfn_is_uc() works only when PAT is enabled, > > > > > + * so check pat_enabled() as well. > > > > > + */ > > > > > + (!pat_enabled() || > > > > > + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); > > > > > > > > Can it be compiled if !CONFIG_PAT? > > > > > > Yes. > > > > > > What I check via pat_enabled() is not only whether PAT support is > > > compiled, but also whether PAT is enabled at runtime. > > > > > > > > > > > It would be better if we move pat_enabled out of kvm as well, > > > > > > Surely I can combine them in one function like > > > > > > bool pat_pfn_is_uc(pfn_t pfn) > > > { > > > enum page_cache_mode cm; > > > > > > if (!pat_enabled()) > > > return false; > > > > > > cm = lookup_memtype(pfn_t_to_phys(pfn)); > > > > > > return cm == _PAGE_CACHE_MODE_UC || cm == _PAGE_CACHE_MODE_UC_MINUS; > > > } > > > > In addition, I think it's better to split this function into > > pat_pfn_is_uc() and pat_pfn_is_uc_minus() to avoid additional > > confusion. > > Why not use pat_pfn_is_uc_or_uc_minus(). :) Just in case that other places other than KVM do not need both of them.
On 11/03/17 16:51 +0800, Haozhong Zhang wrote: > On 11/03/17 14:54 +0800, Xiao Guangrong wrote: > > > > > > On 11/03/2017 01:53 PM, Haozhong Zhang wrote: > > > Some reserved pages, such as those from NVDIMM DAX devices, are > > > not for MMIO, and can be mapped with cached memory type for better > > > performance. However, the above check misconceives those pages as > > > MMIO. Because KVM maps MMIO pages with UC memory type, the > > > performance of guest accesses to those pages would be harmed. > > > Therefore, we check the host memory type by lookup_memtype() in > > > addition and only treat UC/UC- pages as MMIO. > > > > > > Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> > > > Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> > > > Reported-by: Kumar, Karthik <karthik.kumar@intel.com> > > > --- > > > arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- > > > 1 file changed, 18 insertions(+), 1 deletion(-) > > > > > > diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c > > > index 0b481cc9c725..e9ed0e666a83 100644 > > > --- a/arch/x86/kvm/mmu.c > > > +++ b/arch/x86/kvm/mmu.c > > > @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, > > > static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) > > > { > > > if (pfn_valid(pfn)) > > > - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); > > > + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && > > > + /* > > > + * Some reserved pages, such as those from > > > + * NVDIMM DAX devices, are not for MMIO, and > > > + * can be mapped with cached memory type for > > > + * better performance. However, the above > > > + * check misconceives those pages as MMIO. > > > + * Because KVM maps MMIO pages with UC memory > > > + * type, the performance of guest accesses to > > > + * those pages would be harmed. Therefore, we > > > + * check the host memory type in addition and > > > + * only treat UC/UC- pages as MMIO. > > > + * > > > + * pat_pfn_is_uc() works only when PAT is enabled, > > > + * so check pat_enabled() as well. > > > + */ > > > + (!pat_enabled() || > > > + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); > > > > Can it be compiled if !CONFIG_PAT? > > Yes. > > What I check via pat_enabled() is not only whether PAT support is > compiled, but also whether PAT is enabled at runtime. > > > > > It would be better if we move pat_enabled out of kvm as well, > > Surely I can combine them in one function like > > bool pat_pfn_is_uc(pfn_t pfn) > { > enum page_cache_mode cm; > > if (!pat_enabled()) > return false; I made a mistake: it should return true here. Then the semantics of this function is confused. I think it's still better to leave !pat_enabled() check in KVM. Haozhong > > cm = lookup_memtype(pfn_t_to_phys(pfn)); > > return cm == _PAGE_CACHE_MODE_UC || cm == _PAGE_CACHE_MODE_UC_MINUS; > } > > but I need a good name to make its semantics clear, or is it enough to > just leave a comment like? > > /* > * Check via PAT whether the cache mode of a page if UC or UC-. > * > * Returns true, if PAT is enabled and the cache mode is UC or UC-. > * Returns false otherwise. > */ > > > > please refer > > to pgprot_writecombine() which is implemented in pat.c and in > > include\asm-generic\pgtable.h: > > > > #ifndef pgprot_writecombine > > #define pgprot_writecombine pgprot_noncached > > #endif > > > > >
On 11/03/2017 05:29 PM, Haozhong Zhang wrote: > On 11/03/17 17:24 +0800, Xiao Guangrong wrote: >> >> >> On 11/03/2017 05:02 PM, Haozhong Zhang wrote: >>> On 11/03/17 16:51 +0800, Haozhong Zhang wrote: >>>> On 11/03/17 14:54 +0800, Xiao Guangrong wrote: >>>>> >>>>> >>>>> On 11/03/2017 01:53 PM, Haozhong Zhang wrote: >>>>>> Some reserved pages, such as those from NVDIMM DAX devices, are >>>>>> not for MMIO, and can be mapped with cached memory type for better >>>>>> performance. However, the above check misconceives those pages as >>>>>> MMIO. Because KVM maps MMIO pages with UC memory type, the >>>>>> performance of guest accesses to those pages would be harmed. >>>>>> Therefore, we check the host memory type by lookup_memtype() in >>>>>> addition and only treat UC/UC- pages as MMIO. >>>>>> >>>>>> Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> >>>>>> Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> >>>>>> Reported-by: Kumar, Karthik <karthik.kumar@intel.com> >>>>>> --- >>>>>> arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- >>>>>> 1 file changed, 18 insertions(+), 1 deletion(-) >>>>>> >>>>>> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c >>>>>> index 0b481cc9c725..e9ed0e666a83 100644 >>>>>> --- a/arch/x86/kvm/mmu.c >>>>>> +++ b/arch/x86/kvm/mmu.c >>>>>> @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, >>>>>> static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) >>>>>> { >>>>>> if (pfn_valid(pfn)) >>>>>> - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); >>>>>> + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && >>>>>> + /* >>>>>> + * Some reserved pages, such as those from >>>>>> + * NVDIMM DAX devices, are not for MMIO, and >>>>>> + * can be mapped with cached memory type for >>>>>> + * better performance. However, the above >>>>>> + * check misconceives those pages as MMIO. >>>>>> + * Because KVM maps MMIO pages with UC memory >>>>>> + * type, the performance of guest accesses to >>>>>> + * those pages would be harmed. Therefore, we >>>>>> + * check the host memory type in addition and >>>>>> + * only treat UC/UC- pages as MMIO. >>>>>> + * >>>>>> + * pat_pfn_is_uc() works only when PAT is enabled, >>>>>> + * so check pat_enabled() as well. >>>>>> + */ >>>>>> + (!pat_enabled() || >>>>>> + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); >>>>> >>>>> Can it be compiled if !CONFIG_PAT? >>>> >>>> Yes. >>>> >>>> What I check via pat_enabled() is not only whether PAT support is >>>> compiled, but also whether PAT is enabled at runtime. >>>> >>>>> >>>>> It would be better if we move pat_enabled out of kvm as well, >>>> >>>> Surely I can combine them in one function like >>>> >>>> bool pat_pfn_is_uc(pfn_t pfn) >>>> { >>>> enum page_cache_mode cm; >>>> >>>> if (!pat_enabled()) >>>> return false; >>>> >>>> cm = lookup_memtype(pfn_t_to_phys(pfn)); >>>> >>>> return cm == _PAGE_CACHE_MODE_UC || cm == _PAGE_CACHE_MODE_UC_MINUS; >>>> } >>> >>> In addition, I think it's better to split this function into >>> pat_pfn_is_uc() and pat_pfn_is_uc_minus() to avoid additional >>> confusion. >> >> Why not use pat_pfn_is_uc_or_uc_minus(). :) > > Just in case that other places other than KVM do not need both of them. > I think we need not care the future case too much, it is not ABI anyway. i.e, it can be easily adjusted if it is really needed in the future. But it is not a big deal...
diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c index 0b481cc9c725..e9ed0e666a83 100644 --- a/arch/x86/kvm/mmu.c +++ b/arch/x86/kvm/mmu.c @@ -2708,7 +2708,24 @@ static bool mmu_need_write_protect(struct kvm_vcpu *vcpu, gfn_t gfn, static bool kvm_is_mmio_pfn(kvm_pfn_t pfn) { if (pfn_valid(pfn)) - return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)); + return !is_zero_pfn(pfn) && PageReserved(pfn_to_page(pfn)) && + /* + * Some reserved pages, such as those from + * NVDIMM DAX devices, are not for MMIO, and + * can be mapped with cached memory type for + * better performance. However, the above + * check misconceives those pages as MMIO. + * Because KVM maps MMIO pages with UC memory + * type, the performance of guest accesses to + * those pages would be harmed. Therefore, we + * check the host memory type in addition and + * only treat UC/UC- pages as MMIO. + * + * pat_pfn_is_uc() works only when PAT is enabled, + * so check pat_enabled() as well. + */ + (!pat_enabled() || + pat_pfn_is_uc(kvm_pfn_t_to_pfn_t(pfn))); return true; }
Some reserved pages, such as those from NVDIMM DAX devices, are not for MMIO, and can be mapped with cached memory type for better performance. However, the above check misconceives those pages as MMIO. Because KVM maps MMIO pages with UC memory type, the performance of guest accesses to those pages would be harmed. Therefore, we check the host memory type by lookup_memtype() in addition and only treat UC/UC- pages as MMIO. Signed-off-by: Haozhong Zhang <haozhong.zhang@intel.com> Reported-by: Cuevas Escareno, Ivan D <ivan.d.cuevas.escareno@intel.com> Reported-by: Kumar, Karthik <karthik.kumar@intel.com> --- arch/x86/kvm/mmu.c | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-)