From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pl0-x242.google.com (mail-pl0-x242.google.com [IPv6:2607:f8b0:400e:c01::242]) (using TLSv1.2 with cipher ECDHE-RSA-AES128-GCM-SHA256 (128/128 bits)) (No client certificate requested) by lists.ozlabs.org (Postfix) with ESMTPS id 3zdl5V4F4SzF0dt for ; Sat, 10 Feb 2018 19:11:58 +1100 (AEDT) Received: by mail-pl0-x242.google.com with SMTP id g18so2548676plo.7 for ; Sat, 10 Feb 2018 00:11:58 -0800 (PST) From: Nicholas Piggin To: linuxppc-dev@lists.ozlabs.org Cc: Nicholas Piggin , "Aneesh Kumar K . V" , Christophe Leroy Subject: [RFC PATCH 2/5] powerpc/mm/slice: implement a slice mask cache Date: Sat, 10 Feb 2018 18:11:36 +1000 Message-Id: <20180210081139.27236-3-npiggin@gmail.com> In-Reply-To: <20180210081139.27236-1-npiggin@gmail.com> References: <20180210081139.27236-1-npiggin@gmail.com> List-Id: Linux on PowerPC Developers Mail List List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Calculating the slice mask can become a signifcant overhead for get_unmapped_area. This patch adds a struct slice_mask for each page size in the mm_context, and keeps these in synch with the slices psize arrays and slb_addr_limit. This saves about 30% kernel time on a single-page mmap/munmap micro benchmark. Signed-off-by: Nicholas Piggin --- arch/powerpc/include/asm/book3s/64/mmu.h | 20 +++++++++- arch/powerpc/mm/slice.c | 68 ++++++++++++++++++++++++-------- 2 files changed, 71 insertions(+), 17 deletions(-) diff --git a/arch/powerpc/include/asm/book3s/64/mmu.h b/arch/powerpc/include/asm/book3s/64/mmu.h index 0abeb0e2d616..b6d136fd8ffd 100644 --- a/arch/powerpc/include/asm/book3s/64/mmu.h +++ b/arch/powerpc/include/asm/book3s/64/mmu.h @@ -80,6 +80,16 @@ struct spinlock; /* Maximum possible number of NPUs in a system. */ #define NV_MAX_NPUS 8 +/* + * One bit per slice. We have lower slices which cover 256MB segments + * upto 4G range. That gets us 16 low slices. For the rest we track slices + * in 1TB size. + */ +struct slice_mask { + u64 low_slices; + DECLARE_BITMAP(high_slices, SLICE_NUM_HIGH); +}; + typedef struct { mm_context_id_t id; u16 user_psize; /* page size index */ @@ -91,9 +101,17 @@ typedef struct { struct npu_context *npu_context; #ifdef CONFIG_PPC_MM_SLICES + unsigned long slb_addr_limit; u64 low_slices_psize; /* SLB page size encodings */ unsigned char high_slices_psize[SLICE_ARRAY_SIZE]; - unsigned long slb_addr_limit; +# ifdef CONFIG_PPC_64K_PAGES + struct slice_mask mask_64k; +# endif + struct slice_mask mask_4k; +# ifdef CONFIG_HUGETLB_PAGE + struct slice_mask mask_16m; + struct slice_mask mask_16g; +# endif #else u16 sllp; /* SLB page size encoding */ #endif diff --git a/arch/powerpc/mm/slice.c b/arch/powerpc/mm/slice.c index e8f6922d3c9b..837700bb50a9 100644 --- a/arch/powerpc/mm/slice.c +++ b/arch/powerpc/mm/slice.c @@ -37,15 +37,6 @@ #include static DEFINE_SPINLOCK(slice_convert_lock); -/* - * One bit per slice. We have lower slices which cover 256MB segments - * upto 4G range. That gets us 16 low slices. For the rest we track slices - * in 1TB size. - */ -struct slice_mask { - u64 low_slices; - DECLARE_BITMAP(high_slices, SLICE_NUM_HIGH); -}; #ifdef DEBUG int _slice_debug = 1; @@ -144,7 +135,7 @@ static void slice_mask_for_free(struct mm_struct *mm, struct slice_mask *ret, __set_bit(i, ret->high_slices); } -static void slice_mask_for_size(struct mm_struct *mm, int psize, +static void calc_slice_mask_for_size(struct mm_struct *mm, int psize, struct slice_mask *ret, unsigned long high_limit) { @@ -173,6 +164,40 @@ static void slice_mask_for_size(struct mm_struct *mm, int psize, } } +#ifdef CONFIG_PPC_BOOK3S_64 +static void recalc_slice_mask_cache(struct mm_struct *mm) +{ + unsigned long l = mm->context.slb_addr_limit; + calc_slice_mask_for_size(mm, MMU_PAGE_4K, &mm->context.mask_4k, l); +#ifdef CONFIG_PPC_64K_PAGES + calc_slice_mask_for_size(mm, MMU_PAGE_64K, &mm->context.mask_64k, l); +#endif +#ifdef CONFIG_HUGETLB_PAGE + calc_slice_mask_for_size(mm, MMU_PAGE_16M, &mm->context.mask_16m, l); + calc_slice_mask_for_size(mm, MMU_PAGE_16G, &mm->context.mask_16g, l); +#endif +} + +static const struct slice_mask *slice_mask_for_size(struct mm_struct *mm, int psize) +{ +#ifdef CONFIG_PPC_64K_PAGES + if (psize == MMU_PAGE_64K) + return &mm->context.mask_64k; +#endif + if (psize == MMU_PAGE_4K) + return &mm->context.mask_4k; +#ifdef CONFIG_HUGETLB_PAGE + if (psize == MMU_PAGE_16M) + return &mm->context.mask_16m; + if (psize == MMU_PAGE_16G) + return &mm->context.mask_16g; +#endif + BUG(); +} +#else +#error "Must define the slice masks for page sizes supported by the platform" +#endif + static int slice_check_fit(struct mm_struct *mm, const struct slice_mask *mask, const struct slice_mask *available) @@ -246,6 +271,8 @@ static void slice_convert(struct mm_struct *mm, (unsigned long)mm->context.low_slices_psize, (unsigned long)mm->context.high_slices_psize); + recalc_slice_mask_cache(mm); + spin_unlock_irqrestore(&slice_convert_lock, flags); copro_flush_all_slbs(mm); @@ -448,7 +475,14 @@ unsigned long slice_get_unmapped_area(unsigned long addr, unsigned long len, } if (high_limit > mm->context.slb_addr_limit) { + unsigned long flags; + mm->context.slb_addr_limit = high_limit; + + spin_lock_irqsave(&slice_convert_lock, flags); + recalc_slice_mask_cache(mm); + spin_unlock_irqrestore(&slice_convert_lock, flags); + on_each_cpu(slice_flush_segments, mm, 1); } @@ -487,7 +521,7 @@ unsigned long slice_get_unmapped_area(unsigned long addr, unsigned long len, /* First make up a "good" mask of slices that have the right size * already */ - slice_mask_for_size(mm, psize, &good_mask, high_limit); + good_mask = *slice_mask_for_size(mm, psize); slice_print_mask(" good_mask", &good_mask); /* @@ -512,7 +546,7 @@ unsigned long slice_get_unmapped_area(unsigned long addr, unsigned long len, #ifdef CONFIG_PPC_64K_PAGES /* If we support combo pages, we can allow 64k pages in 4k slices */ if (psize == MMU_PAGE_64K) { - slice_mask_for_size(mm, MMU_PAGE_4K, &compat_mask, high_limit); + compat_mask = *slice_mask_for_size(mm, MMU_PAGE_4K); if (fixed) slice_or_mask(&good_mask, &compat_mask); } @@ -693,7 +727,7 @@ void slice_set_user_psize(struct mm_struct *mm, unsigned int psize) goto bail; mm->context.user_psize = psize; - wmb(); + wmb(); /* Why? */ lpsizes = mm->context.low_slices_psize; for (i = 0; i < SLICE_NUM_LOW; i++) @@ -720,6 +754,9 @@ void slice_set_user_psize(struct mm_struct *mm, unsigned int psize) (unsigned long)mm->context.low_slices_psize, (unsigned long)mm->context.high_slices_psize); + recalc_slice_mask_cache(mm); + spin_unlock_irqrestore(&slice_convert_lock, flags); + return; bail: spin_unlock_irqrestore(&slice_convert_lock, flags); } @@ -760,18 +797,17 @@ int is_hugepage_only_range(struct mm_struct *mm, unsigned long addr, { struct slice_mask mask, available; unsigned int psize = mm->context.user_psize; - unsigned long high_limit = mm->context.slb_addr_limit; if (radix_enabled()) return 0; slice_range_to_mask(addr, len, &mask); - slice_mask_for_size(mm, psize, &available, high_limit); + available = *slice_mask_for_size(mm, psize); #ifdef CONFIG_PPC_64K_PAGES /* We need to account for 4k slices too */ if (psize == MMU_PAGE_64K) { struct slice_mask compat_mask; - slice_mask_for_size(mm, MMU_PAGE_4K, &compat_mask, high_limit); + compat_mask = *slice_mask_for_size(mm, MMU_PAGE_4K); slice_or_mask(&available, &compat_mask); } #endif -- 2.15.1