| Message ID | 20200423135656.2712-6-yezhenyu2@huawei.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | arm64: tlb: add support for TTL feature | expand |
On Thu, Apr 23, 2020 at 09:56:55PM +0800, Zhenyu Ye wrote: > diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c > index 3d7c01e76efc..3eff199d3507 100644 > --- a/mm/pgtable-generic.c > +++ b/mm/pgtable-generic.c > @@ -101,6 +101,28 @@ pte_t ptep_clear_flush(struct vm_area_struct *vma, unsigned long address, > > #ifdef CONFIG_TRANSPARENT_HUGEPAGE > > +#ifndef __HAVE_ARCH_FLUSH_PMD_TLB_RANGE > + > +#define FLUSH_Pxx_TLB_RANGE(_pxx) \ > +void flush_##_pxx##_tlb_range(struct vm_area_struct *vma, \ > + unsigned long addr, unsigned long end) \ > +{ \ > + struct mmu_gather tlb; \ > + \ > + tlb_gather_mmu(&tlb, vma->vm_mm, addr, end); \ > + tlb_start_vma(&tlb, vma); \ > + tlb_flush_##_pxx##_range(&tlb, addr, end - addr); \ > + tlb_end_vma(&tlb, vma); \ > + tlb_finish_mmu(&tlb, addr, end); \ > +} I may have confused myself (flush_p??_tlb_* vs. tlb_flush_p??_*) but do actually need this whole tlb_gather thing here? IIUC (by grep'ing), flush_p?d_tlb_range() is only called on huge pages, so we should know the level already.
On 2020/5/22 23:42, Catalin Marinas wrote: > On Thu, Apr 23, 2020 at 09:56:55PM +0800, Zhenyu Ye wrote: >> diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c >> index 3d7c01e76efc..3eff199d3507 100644 >> --- a/mm/pgtable-generic.c >> +++ b/mm/pgtable-generic.c >> @@ -101,6 +101,28 @@ pte_t ptep_clear_flush(struct vm_area_struct *vma, unsigned long address, >> >> #ifdef CONFIG_TRANSPARENT_HUGEPAGE >> >> +#ifndef __HAVE_ARCH_FLUSH_PMD_TLB_RANGE >> + >> +#define FLUSH_Pxx_TLB_RANGE(_pxx) \ >> +void flush_##_pxx##_tlb_range(struct vm_area_struct *vma, \ >> + unsigned long addr, unsigned long end) \ >> +{ \ >> + struct mmu_gather tlb; \ >> + \ >> + tlb_gather_mmu(&tlb, vma->vm_mm, addr, end); \ >> + tlb_start_vma(&tlb, vma); \ >> + tlb_flush_##_pxx##_range(&tlb, addr, end - addr); \ >> + tlb_end_vma(&tlb, vma); \ >> + tlb_finish_mmu(&tlb, addr, end); \ >> +} > > I may have confused myself (flush_p??_tlb_* vs. tlb_flush_p??_*) but do > actually need this whole tlb_gather thing here? IIUC (by grep'ing), > flush_p?d_tlb_range() is only called on huge pages, so we should know > the level already. > tlb_flush_##_pxx##_range() is used to set tlb->cleared_*, flush_##_pxx##_tlb_range() will actually flush the TLB entry. In arch64, tlb_flush_p?d_range() is defined as: #define flush_pmd_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) #define flush_pud_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) So even if we know the level here, we can not pass the value to tlbi instructions (flush_tlb_range() is a common kernel interface and retro-fit it needs lots of changes), according to Peter's suggestion, I finally decide to pass the value of TTL by the tlb_gather_* frame.[1] [1] https://lore.kernel.org/linux-arm-kernel/20200331142927.1237-1-yezhenyu2@huawei.com/ Thanks, Zhenyu
On Mon, May 25, 2020 at 03:19:42PM +0800, Zhenyu Ye wrote: > On 2020/5/22 23:42, Catalin Marinas wrote: > > On Thu, Apr 23, 2020 at 09:56:55PM +0800, Zhenyu Ye wrote: > >> diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c > >> index 3d7c01e76efc..3eff199d3507 100644 > >> --- a/mm/pgtable-generic.c > >> +++ b/mm/pgtable-generic.c > >> @@ -101,6 +101,28 @@ pte_t ptep_clear_flush(struct vm_area_struct *vma, unsigned long address, > >> > >> #ifdef CONFIG_TRANSPARENT_HUGEPAGE > >> > >> +#ifndef __HAVE_ARCH_FLUSH_PMD_TLB_RANGE > >> + > >> +#define FLUSH_Pxx_TLB_RANGE(_pxx) \ > >> +void flush_##_pxx##_tlb_range(struct vm_area_struct *vma, \ > >> + unsigned long addr, unsigned long end) \ > >> +{ \ > >> + struct mmu_gather tlb; \ > >> + \ > >> + tlb_gather_mmu(&tlb, vma->vm_mm, addr, end); \ > >> + tlb_start_vma(&tlb, vma); \ > >> + tlb_flush_##_pxx##_range(&tlb, addr, end - addr); \ > >> + tlb_end_vma(&tlb, vma); \ > >> + tlb_finish_mmu(&tlb, addr, end); \ > >> +} > > > > I may have confused myself (flush_p??_tlb_* vs. tlb_flush_p??_*) but do > > actually need this whole tlb_gather thing here? IIUC (by grep'ing), > > flush_p?d_tlb_range() is only called on huge pages, so we should know > > the level already. > > tlb_flush_##_pxx##_range() is used to set tlb->cleared_*, > flush_##_pxx##_tlb_range() will actually flush the TLB entry. > > In arch64, tlb_flush_p?d_range() is defined as: > > #define flush_pmd_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) > #define flush_pud_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) Currently, flush_p??_tlb_range() are generic and defined as above. I think in the generic code they can remain an alias for flush_tlb_range(). On arm64, we can redefine them as: #define flush_pte_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 3) #define flush_pmd_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 2) #define flush_pud_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 1) #define flush_p4d_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 0) (unless the compiler optimises away all the mmu_gather stuff in your macro above but they don't look trivial to me) Also, I don't see the new flush_pte_* and flush_p4d_* macros used anywhere and I don't think they are needed. The pte equivalent is flush_tlb_page() (we need to make sure it's not used on a pmd in the hugetlb context). > So even if we know the level here, we can not pass the value to tlbi > instructions (flush_tlb_range() is a common kernel interface and retro-fit it > needs lots of changes), according to Peter's suggestion, I finally decide to > pass the value of TTL by the tlb_gather_* frame.[1] My comment was about the generic implementation using mmu_gather as you are proposing. We don't need to change the flush_tlb_range() interface, nor do we need to rewrite flush_p??_tlb_range().
Hi Catalin, Sorry for taking so long to reply to you. On 2020/5/26 22:52, Catalin Marinas wrote: > On Mon, May 25, 2020 at 03:19:42PM +0800, Zhenyu Ye wrote: >> >> tlb_flush_##_pxx##_range() is used to set tlb->cleared_*, >> flush_##_pxx##_tlb_range() will actually flush the TLB entry. >> >> In arch64, tlb_flush_p?d_range() is defined as: >> >> #define flush_pmd_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) >> #define flush_pud_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) > > Currently, flush_p??_tlb_range() are generic and defined as above. I > think in the generic code they can remain an alias for > flush_tlb_range(). > > On arm64, we can redefine them as: > > #define flush_pte_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 3) > #define flush_pmd_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 2) > #define flush_pud_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 1) > #define flush_p4d_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 0) > > (unless the compiler optimises away all the mmu_gather stuff in your > macro above but they don't look trivial to me) > I changed generic code before considering that other structures may also use this feature, such as Power9. And Peter may want to replace all flush_tlb_range() by tlb_flush() in the future, see [1] for details. If only enable this feature on aarch64, your codes are better. [1] https://lore.kernel.org/linux-arm-kernel/20200402163849.GM20713@hirez.programming.kicks-ass.net/ > Also, I don't see the new flush_pte_* and flush_p4d_* macros used > anywhere and I don't think they are needed. The pte equivalent is > flush_tlb_page() (we need to make sure it's not used on a pmd in the > hugetlb context). > flush_tlb_page() is used to flush only one page. If we add the flush_pte_tlb_range(), then we can use it to flush a range of pages in the future. But flush_pte_* and flush_p4d_* macros are really not used anywhere. I will remove them in next version of series, and add them if someone needs. >> So even if we know the level here, we can not pass the value to tlbi >> instructions (flush_tlb_range() is a common kernel interface and retro-fit it >> needs lots of changes), according to Peter's suggestion, I finally decide to >> pass the value of TTL by the tlb_gather_* frame.[1] > > My comment was about the generic implementation using mmu_gather as you > are proposing. We don't need to change the flush_tlb_range() interface, > nor do we need to rewrite flush_p??_tlb_range(). > Thanks, Zhenyu
Hi Zhenyu, On Sat, May 30, 2020 at 06:24:21PM +0800, Zhenyu Ye wrote: > On 2020/5/26 22:52, Catalin Marinas wrote: > > On Mon, May 25, 2020 at 03:19:42PM +0800, Zhenyu Ye wrote: > >> tlb_flush_##_pxx##_range() is used to set tlb->cleared_*, > >> flush_##_pxx##_tlb_range() will actually flush the TLB entry. > >> > >> In arch64, tlb_flush_p?d_range() is defined as: > >> > >> #define flush_pmd_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) > >> #define flush_pud_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) > > > > Currently, flush_p??_tlb_range() are generic and defined as above. I > > think in the generic code they can remain an alias for > > flush_tlb_range(). > > > > On arm64, we can redefine them as: > > > > #define flush_pte_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 3) > > #define flush_pmd_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 2) > > #define flush_pud_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 1) > > #define flush_p4d_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 0) > > > > (unless the compiler optimises away all the mmu_gather stuff in your > > macro above but they don't look trivial to me) > > I changed generic code before considering that other structures may also > use this feature, such as Power9. And Peter may want to replace all > flush_tlb_range() by tlb_flush() in the future, see [1] for details. > > If only enable this feature on aarch64, your codes are better. > > [1] https://lore.kernel.org/linux-arm-kernel/20200402163849.GM20713@hirez.programming.kicks-ass.net/ But we change the semantics slightly if we implement these as mmu_gather. For example, tlb_end_vma() -> tlb_flush_mmu_tlbonly() ends up calling mmu_notifier_invalidate_range() which it didn't before. I think we end up invoking the notifier unnecessarily in some cases (see the comment in __split_huge_pmd()) or we end up calling the notifier twice (e.g. pmdp_huge_clear_flush_notify()). > > Also, I don't see the new flush_pte_* and flush_p4d_* macros used > > anywhere and I don't think they are needed. The pte equivalent is > > flush_tlb_page() (we need to make sure it's not used on a pmd in the > > hugetlb context). > > flush_tlb_page() is used to flush only one page. If we add the > flush_pte_tlb_range(), then we can use it to flush a range of pages in > the future. If we know flush_tlb_page() is only called on a small page, could we add TTL information here as well? > But flush_pte_* and flush_p4d_* macros are really not used anywhere. I > will remove them in next version of series, and add them if someone > needs. I think it makes sense.
On 2020/6/1 19:56, Catalin Marinas wrote: > Hi Zhenyu, > > On Sat, May 30, 2020 at 06:24:21PM +0800, Zhenyu Ye wrote: >> On 2020/5/26 22:52, Catalin Marinas wrote: >>> On Mon, May 25, 2020 at 03:19:42PM +0800, Zhenyu Ye wrote: >>>> tlb_flush_##_pxx##_range() is used to set tlb->cleared_*, >>>> flush_##_pxx##_tlb_range() will actually flush the TLB entry. >>>> >>>> In arch64, tlb_flush_p?d_range() is defined as: >>>> >>>> #define flush_pmd_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) >>>> #define flush_pud_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) >>> >>> Currently, flush_p??_tlb_range() are generic and defined as above. I >>> think in the generic code they can remain an alias for >>> flush_tlb_range(). >>> >>> On arm64, we can redefine them as: >>> >>> #define flush_pte_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 3) >>> #define flush_pmd_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 2) >>> #define flush_pud_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 1) >>> #define flush_p4d_tlb_range(vma, addr, end) __flush_tlb_range(vma, addr, end, 0) >>> >>> (unless the compiler optimises away all the mmu_gather stuff in your >>> macro above but they don't look trivial to me) >> >> I changed generic code before considering that other structures may also >> use this feature, such as Power9. And Peter may want to replace all >> flush_tlb_range() by tlb_flush() in the future, see [1] for details. >> >> If only enable this feature on aarch64, your codes are better. >> >> [1] https://lore.kernel.org/linux-arm-kernel/20200402163849.GM20713@hirez.programming.kicks-ass.net/ > > But we change the semantics slightly if we implement these as > mmu_gather. For example, tlb_end_vma() -> tlb_flush_mmu_tlbonly() ends > up calling mmu_notifier_invalidate_range() which it didn't before. I > think we end up invoking the notifier unnecessarily in some cases (see > the comment in __split_huge_pmd()) or we end up calling the notifier > twice (e.g. pmdp_huge_clear_flush_notify()). > Yes, so only enable this feature on aarch64 may be better. I will change this in V4 of this series. [the v3 only has some minor changes and can be ignored :)] >>> Also, I don't see the new flush_pte_* and flush_p4d_* macros used >>> anywhere and I don't think they are needed. The pte equivalent is >>> flush_tlb_page() (we need to make sure it's not used on a pmd in the >>> hugetlb context). >> >> flush_tlb_page() is used to flush only one page. If we add the >> flush_pte_tlb_range(), then we can use it to flush a range of pages in >> the future. > > If we know flush_tlb_page() is only called on a small page, could we add > TTL information here as well? > Yes, we could. I will add this in flush_tlb_page(). >> But flush_pte_* and flush_p4d_* macros are really not used anywhere. I >> will remove them in next version of series, and add them if someone >> needs. > > I think it makes sense. > Thanks, Zhenyu
diff --git a/include/asm-generic/pgtable.h b/include/asm-generic/pgtable.h index 329b8c8ca703..8c92122ded9b 100644 --- a/include/asm-generic/pgtable.h +++ b/include/asm-generic/pgtable.h @@ -1161,11 +1161,19 @@ static inline int pmd_free_pte_page(pmd_t *pmd, unsigned long addr) * invalidate the entire TLB which is not desitable. * e.g. see arch/arc: flush_pmd_tlb_range */ -#define flush_pmd_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) -#define flush_pud_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) +extern void flush_pte_tlb_range(struct vm_area_struct *vma, + unsigned long addr, unsigned long end); +extern void flush_pmd_tlb_range(struct vm_area_struct *vma, + unsigned long addr, unsigned long end); +extern void flush_pud_tlb_range(struct vm_area_struct *vma, + unsigned long addr, unsigned long end); +extern void flush_p4d_tlb_range(struct vm_area_struct *vma, + unsigned long addr, unsigned long end); #else +#define flush_pte_tlb_range(vma, addr, end) BUILD_BUG() #define flush_pmd_tlb_range(vma, addr, end) BUILD_BUG() #define flush_pud_tlb_range(vma, addr, end) BUILD_BUG() +#define flush_p4d_tlb_range(vma, addr, end) BUILD_BUG() #endif #endif diff --git a/mm/pgtable-generic.c b/mm/pgtable-generic.c index 3d7c01e76efc..3eff199d3507 100644 --- a/mm/pgtable-generic.c +++ b/mm/pgtable-generic.c @@ -101,6 +101,28 @@ pte_t ptep_clear_flush(struct vm_area_struct *vma, unsigned long address, #ifdef CONFIG_TRANSPARENT_HUGEPAGE +#ifndef __HAVE_ARCH_FLUSH_PMD_TLB_RANGE + +#define FLUSH_Pxx_TLB_RANGE(_pxx) \ +void flush_##_pxx##_tlb_range(struct vm_area_struct *vma, \ + unsigned long addr, unsigned long end) \ +{ \ + struct mmu_gather tlb; \ + \ + tlb_gather_mmu(&tlb, vma->vm_mm, addr, end); \ + tlb_start_vma(&tlb, vma); \ + tlb_flush_##_pxx##_range(&tlb, addr, end - addr); \ + tlb_end_vma(&tlb, vma); \ + tlb_finish_mmu(&tlb, addr, end); \ +} + +FLUSH_Pxx_TLB_RANGE(pte) +FLUSH_Pxx_TLB_RANGE(pmd) +FLUSH_Pxx_TLB_RANGE(pud) +FLUSH_Pxx_TLB_RANGE(p4d) + +#endif /* __HAVE_ARCH_FLUSH_PMD_TLB_RANGE */ + #ifndef __HAVE_ARCH_PMDP_SET_ACCESS_FLAGS int pmdp_set_access_flags(struct vm_area_struct *vma, unsigned long address, pmd_t *pmdp,
This patch provides flush_{pte|pmd|pud|p4d}_tlb_range() in generic code, which are expressed through the mmu_gather APIs. These interface set tlb->cleared_* and finally call tlb_flush(), so we can do the tlb invalidation according to the information in struct mmu_gather. Signed-off-by: Zhenyu Ye <yezhenyu2@huawei.com> --- include/asm-generic/pgtable.h | 12 ++++++++++-- mm/pgtable-generic.c | 22 ++++++++++++++++++++++ 2 files changed, 32 insertions(+), 2 deletions(-)